Bandit Algorithms

Tor Lattimore and Csaba Szepesvári

This is the (free) online edition. The content is the same as the print edition, published
by Cambridge University Press, except that minor typos are corrected here. There are
also font and other typographical differences that mean the page numbers do not
match between the versions.
 Contents

Contents page ii

Part I Bandits, Probability and Concentration 7

1 Introduction 8
1.1 The Language of Bandits 10
1.2 Applications 13
1.3 Notes 16
1.4 Bibliographic Remarks 16

2 Foundations of Probability ( ) 18
2.1 Probability Spaces and Random Elements 18
2.2 σ-Algebras and Knowledge 26
2.3 Conditional Probabilities 28
2.4 Independence 29
2.5 Integration and Expectation 30
2.6 Conditional Expectation 33
2.7 Notes 37
2.8 Bibliographic Remarks 41
2.9 Exercises 42

3 Stochastic Processes and Markov Chains ( ) 46

3.1 Stochastic Processes 47
3.2 Markov Chains 48
3.3 Martingales and Stopping Times 49
3.4 Notes 52
3.5 Bibliographic Remarks 53
3.6 Exercises 54

4 Stochastic Bandits 56
4.1 Core Assumptions 56
4.2 The Learning Objective 57
4.3 Knowledge and Environment Classes 57
 CONTENTS iii

4.4 The Regret 60

4.5 Decomposing the Regret 62
4.6 The Canonical Bandit Model ( ) 63
4.7 The Canonical Bandit Model for Uncountable Action Sets ( ) 66
4.8 Notes 66
4.9 Bibliographical Remarks 68
4.10 Exercises 69

5 Concentration of Measure 73
5.1 Tail Probabilities 73
5.2 The Inequalities of Markov and Chebyshev 74
5.3 The Cramér-Chernoff Method and Subgaussian Random Variables 76
5.4 Notes 78
5.5 Bibliographical Remarks 80
5.6 Exercises 80

Part II Stochastic Bandits with Finitely Many Arms 88

6 The Explore-Then-Commit Algorithm 90

6.1 Algorithm and Regret Analysis 90
6.2 Notes 94
6.3 Bibliographical Remarks 94
6.4 Exercises 95

7 The Upper Confidence Bound Algorithm 101

7.1 The Optimism Principle 101
7.2 Notes 109
7.3 Bibliographical Remarks 110
7.4 Exercises 110

8 The Upper Confidence Bound Algorithm: Asymptotic Optimality 116

8.1 Asymptotically Optimal UCB 116
8.2 Notes 119
8.3 Bibliographic Remarks 119
8.4 Exercises 120

9 The Upper Confidence Bound Algorithm: Minimax Optimality ( ) 122

9.1 The MOSS Algorithm 122
9.2 Two Problems 126
9.3 Notes 127
9.4 Bibliographic Remarks 128
9.5 Exercises 128

10 The Upper Confidence Bound Algorithm: Bernoulli Noise ( ) 132

 CONTENTS iv

10.1 Concentration for Sums of Bernoulli Random Variables 132

10.2 The KL-UCB Algorithm 136
10.3 Notes 139
10.4 Bibliographic Remarks 140
10.5 Exercises 140

Part III Adversarial Bandits with Finitely Many Arms 143

11 The Exp3 Algorithm 147

11.1 Adversarial Bandit Environments 147
11.2 Importance-Weighted Estimators 149
11.3 The Exp3 Algorithm 151
11.4 Regret Analysis 152
11.5 Notes 156
11.6 Bibliographic Remarks 159
11.7 Exercises 159

12 The Exp3-IX Algorithm 164

12.1 The Exp3-IX Algorithm 164
12.2 Regret Analysis 166
12.3 Notes 170
12.4 Bibliographic Remarks 171
12.5 Exercises 171

Part IV Lower Bounds for Bandits with Finitely Many Arms 176

13 Lower Bounds: Basic Ideas 179

13.1 Main Ideas Underlying Minimax Lower Bounds 180
13.2 Notes 182
13.3 Bibliographic Remarks 183
13.4 Exercises 183

14 Foundations of Information Theory ( ) 185

14.1 Entropy and Optimal Coding 185
14.2 Relative Entropy 187
14.3 Notes 190
14.4 Bibliographic Remarks 193
14.5 Exercises 193

15 Minimax Lower Bounds 197

15.1 Relative Entropy Between Bandits 197
15.2 Minimax Lower Bounds 198
15.3 Notes 200
15.4 Bibliographic Remarks 202
 CONTENTS v

15.5 Exercises 202

16 Instance-Dependent Lower Bounds 205

16.1 Asymptotic Bounds 206
16.2 Finite-Time Bounds 208
16.3 Notes 209
16.4 Bibliographic Remarks 210
16.5 Exercises 210

17 High-Probability Lower Bounds 214

17.1 Stochastic Bandits 215
17.2 Adversarial Bandits 217
17.3 Notes 219
17.4 Bibliographic Remarks 219
17.5 Exercises 220

Part V Contextual and Linear Bandits 221

18 Contextual Bandits 223

18.1 Contextual Bandits: One Bandit per Context 223
18.2 Bandits with Expert Advice 225
18.3 Exp4 228
18.4 Regret Analysis 228
18.5 Notes 230
18.6 Bibliographic Remarks 232
18.7 Exercises 233

19 Stochastic Linear Bandits 237

19.1 Stochastic Contextual Bandits 237
19.2 Stochastic Linear Bandits 239
19.3 Regret Analysis 242
19.4 Notes 244
19.5 Bibliographic Remarks 246
19.6 Exercises 247

20 Confidence Bounds for Least Squares Estimators 253

20.1 Martingales and the Method of Mixtures 255
20.2 Notes 260
20.3 Bibliographic Remarks 261
20.4 Exercises 261

21 Optimal Design for Least Squares Estimators 266

21.1 The Kiefer–Wolfowitz Theorem 266
21.2 Notes 269
 CONTENTS vi

21.3 Bibliographic Remarks 270

21.4 Exercises 270

22 Stochastic Linear Bandits with Finitely Many Arms 272

22.1 Notes 273
22.2 Bibliographic Remarks 274
22.3 Exercises 274

23 Stochastic Linear Bandits with Sparsity 276

23.1 Sparse Linear Stochastic Bandits 276
23.2 Elimination on the Hypercube 277
23.3 Online to Confidence Set Conversion 281
23.4 Sparse Online Linear Prediction 284
23.5 Notes 285
23.6 Bibliographical Remarks 286
23.7 Exercises 286

24 Minimax Lower Bounds for Stochastic Linear Bandits 287

24.1 Hypercube 288
24.2 Unit Ball 289
24.3 Sparse Parameter Vectors 291
24.4 Misspecified Models 292
24.5 Notes 294
24.6 Bibliographic Remarks 294
24.7 Exercises 294

25 Asymptotic Lower Bounds for Stochastic Linear Bandits 295

25.1 An Asymptotic Lower Bound for Fixed Action Sets 295
25.2 Clouds Looming for Optimism 299
25.3 Notes 301
25.4 Bibliographic Remarks 301
25.5 Exercises 301

Part VI Adversarial Linear Bandits 303

26 Foundations of Convex Analysis ( ) 305

26.1 Convex Sets and Functions 305
26.2 Jensen’s Inequality 307
26.3 Bregman Divergence 307
26.4 Legendre Functions 309
26.5 Optimisation 311
26.6 Projections 312
26.7 Notes 313
26.8 Bibliographic Remarks 313
 CONTENTS vii

26.9 Exercises 313

27 Exp3 for Adversarial Linear Bandits 317

27.1 Exponential Weights for Linear Bandits 317
27.2 Regret Analysis 319
27.3 Continuous Exponential Weights 320
27.4 Notes 322
27.5 Bibliographic Remarks 323
27.6 Exercises 324

28 Follow-the-Regularised-Leader and Mirror Descent 326

28.1 Online Linear Optimisation 326
28.2 Regret Analysis 330
28.3 Application to Linear Bandits 334
28.4 Linear Bandits on the Unit Ball 336
28.5 Notes 339
28.6 Bibliographic Remarks 342
28.7 Exercises 343

29 The Relation between Adversarial and Stochastic Linear Bandits 349

29.1 Unified View 349
29.2 Reducing Stochastic Linear Bandits to Adversarial Linear Bandits 350
29.3 Stochastic Linear Bandits with Parameter Noise 351
29.4 Contextual Linear Bandits 353
29.5 Notes 354
29.6 Bibliographic Remarks 355
29.7 Exercises 355

Part VII Other Topics 357

30 Combinatorial Bandits 361

30.1 Notation and Assumptions 362
30.2 Applications 362
30.3 Bandit Feedback 363
30.4 Semi-bandit Feedback and Mirror Descent 364
30.5 Follow-the-Perturbed-Leader 366
30.6 Notes 371
30.7 Bibliographic Remarks 373
30.8 Exercises 374

31 Non-stationary Bandits 377

31.1 Adversarial Bandits 377
31.2 Stochastic Bandits 380
31.3 Notes 382
 CONTENTS viii

31.4 Bibliographic Remarks 384

31.5 Exercises 385

32 Ranking 387
32.1 Click Models 388
32.2 Policy 391
32.3 Regret Analysis 393
32.4 Notes 397
32.5 Bibliographic Remarks 399
32.6 Exercises 400

33 Pure Exploration 402

33.1 Simple Regret 402
33.2 Best-Arm Identification with a Fixed Confidence 404
33.3 Best-Arm Identification with a Budget 411
33.4 Notes 412
33.5 Bibliographical Remarks 414
33.6 Exercises 416

34 Foundations of Bayesian Learning 420

34.1 Statistical Decision Theory and Bayesian Learning 420
34.2 Bayesian Learning and the Posterior Distribution 421
34.3 Conjugate Pairs, Conjugate Priors and the Exponential Family 425
34.4 The Bayesian Bandit Environment 429
34.5 Posterior Distributions in Bandits 430
34.6 Bayesian Regret 431
34.7 Notes 432
34.8 Bibliographic Remarks 435
34.9 Exercises 435

35 Bayesian Bandits 439

35.1 Bayesian Optimal Regret for k-Armed Stochastic Bandits 439
35.2 Optimal Stopping ( ) 440
35.3 One-armed Bayesian Bandits 442
35.4 Gittins Index 446
35.5 Computing the Gittins Index 452
35.6 Notes 453
35.7 Bibliographical Remarks 455
35.8 Exercises 456

36 Thompson Sampling 459

36.1 Finite-Armed Bandits 460
36.2 Frequentist Analysis 461
36.3 Linear Bandits 465
 CONTENTS ix

36.4 Information Theoretic Analysis 467

36.5 Notes 470
36.6 Bibliographic Remarks 473
36.7 Exercises 474

Part VIII Beyond Bandits 477

37 Partial Monitoring 478

37.1 Finite Adversarial Partial Monitoring Problems 479
37.2 The Structure of Partial Monitoring 482
37.3 Classification of Finite Adversarial Partial Monitoring 486
37.4 Lower Bounds 487
37.5 Policy and Upper Bounds 491
37.6 Proof of Theorem 37.16 496
37.7 Proof of Theorem 37.17 497
37.8 Proof of the Classification Theorem 502
37.9 Notes 502
37.10 Bibliographical Remarks 506
37.11 Exercises 507

38 Markov Decision Processes 511

38.1 Problem Set-Up 511
38.2 Optimal Policies and the Bellman Optimality Equation 515
38.3 Finding an Optimal Policy ( ) 518
38.4 Learning in Markov Decision Processes 521
38.5 Upper Confidence Bounds for Reinforcement Learning 522
38.6 Proof of Upper Bound 525
38.7 Proof of Lower Bound 528
38.8 Notes 531
38.9 Bibliographical Remarks 535
38.10 Exercises 537

Bibliography 548

Index 582
Preface

Multi-armed bandits have now been studied for nearly a century. While research
in the beginning was quite meandering, there is now a large community publishing
hundreds of articles every year. Bandit algorithms are also finding their way into
practical applications in industry, especially in on-line platforms where data is
readily available and automation is the only way to scale.
We had hoped to write a comprehensive book, but the literature is now so vast
that many topics have been excluded. In the end we settled on the more modest
goal of equipping our readers with enough expertise to explore the specialised
literature by themselves, and to adapt existing algorithms to their applications.
This latter point is important. Problems in theory are all alike; every application is
different. A practitioner seeking to apply a bandit algorithm needs to understand
which assumptions in the theory are important and how to modify the algorithm
when the assumptions change. We hope this book can provide that understanding.
What is covered in the book is covered in some depth. The focus is on the
mathematical analysis of algorithms for bandit problems, but this is not a
traditional mathematics book, where lemmas are followed by proofs, theorems
and more lemmas. We worked hard to include guiding principles for designing
algorithms and intuition for their analysis. Many algorithms are accompanied by
empirical demonstrations that further aid intuition.
We expect our readers to be familiar with basic analysis and calculus and
some linear algebra. The book uses the notation of measure-theoretic probability
theory, but does not rely on any deep results. A dedicated chapter is included to
introduce the notation and provide intuitions for the basic results we need. This
chapter is unusual for an introduction to measure theory in that it emphasises the
reasons to use σ-algebras beyond the standard technical justifications. We hope
this will convince the reader that measure theory is an important and intuitive
tool. Some chapters use techniques from information theory and convex analysis,
and we devote a short chapter to each.
Most chapters are short and should be readable in an afternoon or presented in
a single lecture. Some components of the book contain content that is not really
about bandits. These can be skipped by knowledgeable readers, or otherwise
referred to when necessary. They are marked with a ( ) because ‘Skippy the
 Preface 2

Kangaroo’ skips things.1 The same mark is used for those parts that contain
useful, but perhaps overly specific information for the first-time reader. Later
parts will not build on these chapters in any substantial way. Most chapters end
with a list of notes and exercises. These are intended to deepen intuition and
highlight the connections between various subsections and the literature. There
is a table of notation at the end of this preface.

Thanks
We’re indebted to our many collaborators and feel privileged that there are
too many of you to name. The University of Alberta, Indiana University and
DeepMind have all provided outstanding work environments and supported the
completion of this book. The book has benefited enormously from the proofreading
efforts of a large number of our friends and colleagues. We are sorry for all the
mistakes introduced after your hard work. Alphabetically, they are: Aaditya
Ramdas, Abbas Mehrabian, Aditya Gopalan, Ambuj Tewari, András György,
Arnoud den Boer, Branislav Kveton, Brendan Patch, Chao Tao, Chao Qin,
Christoph Dann, Claire Vernade, Emilie Kaufmann, Eugene Ji, Gellért Weisz,
Gergely Neu, Johannes Kirschner, Julian Zimmert, Kwang-Sung Jun, Lalit Jain,
Laurent Orseau, Marcus Hutter, Michal Valko, Omar Rivasplata, Pierre Menard,
Ramana Kumar, Roman Pogodin, Ronald Ortner, Ronan Fruit, Ruihao Zhu,
Shuai Li, Toshiyuki Tanaka, Wei Chen, Yoan Russac, Yufei Yi and Zhu Xiaohu.
We are especially grateful to Gábor Balázs and Wouter Koolen, who both read
almost the entire book. Thanks to Lauren Cowels and Cambridge University
Press for providing free books for our proofreaders, tolerating the delays and
for supporting a freely available PDF version. Réka Szepesvári is responsible for
converting some of our primary school figures to their current glory. Last of all,
our families have endured endless weekends of editing and multiple false promises
of ‘done by Christmas’. Rosina and Beáta, it really is done now!

1 Taking inspiration from Tor’s grandfather-in-law, John Dillon [Anderson et al., 1977].
Notation

Some sections are marked with special symbols, which are listed and described
below.

This symbol is a note. Usually this is a remark that is slightly tangential to

the topic at hand.

A warning to the reader.

Something important.

An experiment.

Nomenclature and Conventions

A sequence (an )∞ n=1 is increasing if an+1 ≥ an for all n ≥ 1 and
decreasing if an+1 ≤ an . When the inequalities are strict, we say strictly
increasing/decreasing. The same terminology holds for functions. We will
not be dogmatic about what is the range of argmin/argmax. Sometimes they
return sets, sometimes arbitrary elements of those sets and, where stated, specific
elements of those sets. We will be specific when it is non-obvious/matters. The
infimum of the empty set is inf ∅ = ∞ and the supremum is sup ∅ = −∞. The
P Q
empty sum is i∈∅ ai = 0 and the empty product is i∈∅ ai = 1.

Landau Notation
We make frequent use of the Bachmann–Landau notation. Both were nineteenth
century mathematicians who could have never expected their notation to be
adopted so enthusiastically by computer scientists. Given functions f, g : N →
 Notation 4

[0, ∞), define

f (n)
f (n) = O(g(n)) ⇔ lim sup < ∞,
n→∞ g(n)
f (n)
f (n) = o(g(n)) ⇔ lim = 0,
n→∞ g(n)

f (n)
f (n) = Ω(g(n)) ⇔ lim inf > 0,
n→∞ g(n)

f (n)
f (n) = ω(g(n)) ⇔ lim inf = ∞,
n→∞ g(n)

f (n) = Θ(g(n)) ⇔ f (n) = O(g(n)) and f (n) = Ω(g(n)) .

We make use of the (Bachmann–)Landau notation in two contexts. First, in
proofs where limiting arguments are made, we sometimes write lower-order terms
√ √
using Landau notation. For example, we might write that f (n) = n + o( n), by
√
which we mean that limn→∞ f (n)/ n = 1. In this case we use the mathematical
definitions as envisaged by Bachmann and Landau. The second usage is to
informally describe a result without the clutter of uninteresting constants. For
better or worse, this usage is often a little √
imprecise. For example, we will often
write expressions of the form: Rn = O(m dn). Almost always what is meant
by this is that there exists a universal constant c > 0 (a constant√that does
not depend on either of the quantities involved) such that Rn ≤ cm dn for all
(reasonable) choices of m, d and n. In this context we are careful not to use Landau
notation to hide large lower-order terms. For example, if f (x) = x2 + 10100 x, we
will not write f (x) = O(x2 ), although this would be true.

Bandits
At action in round t
k number of arms/actions
n time horizon
Xt reward in round t
Yt loss in round t
π a policy
ν a bandit
µi mean reward of arm i
Sets
∅ empty set
N, N+ natural numbers, N = {0, 1, 2, . . .} and N+ = N \ {0}
R real numbers
R̄ R ∪ {−∞, ∞}
[n] {1, 2, 3, . . . , n − 1, n}
2A the power set of set A (the set of all subsets of A)
S∞
A∗ set of finite sequences over A, A∗ = i=0 Ai
B2d d-dimensional unit ball, {x ∈ Rd : kxk2 ≤ 1}
Pd probability simplex, {x ∈ [0, 1]d+1 : kxk1 = 1}
 Notation 5

P(A) set of distributions over set A

B(A) Borel σ-algebra on A
[x, y] convex hull of vectors or real values x and y

Functions, Operators and Operations

|A| the cardinality (number of elements) of the finite set A
(x)+ max(x, 0)
a mod b remainder when natural number a is divided by b
bxc, dxe floor and ceiling functions of x
dom(f ) domain of function f
E expectation
V variance
Supp support of distribution or random variable
∇f (x) gradient of f at x
∇v f (x) directional derivative of f at x in direction v
∇2 f (x) Hessian of f at x
∨, ∧ maximum
Rx and minimum, a∨b = max(a, b) and a∧b = min(a, b)
erf(x) √2 exp(−y 2
)dy
π 0
erfc(x) 1 − erf(x) R∞
Γ(z) Gamma function, Γ(z) = 0 xz−1 exp(−x)dx
φA (x) support function φA (x) = supy∈A hx, yi
f ∗
(y) convex conjugate, f ∗ (y) = supx∈A hx, yi − f (x)
n
k binomial coefficient
argmaxx f (x) maximiser or maximisers of f
argminx f (x) minimiser or minimisers of f
Iφ indicator function: converts Boolean φ into binary
IB indicator of set B
D(P, Q) Relative entropy between probability distributions P and Q
d(p, q) Relative entropy between B(p) and B(q)

Linear Algebra
e1 , . . . , e d standard basis vectors of the d-dimensional Euclidean space
0, 1 vectors whose elements are all zeros and all ones, respectively
det(A) determinant of matrix A
trace(A) trace of matrix A
im(A) image of matrix A
ker(A) kernel of matrix A
span(v1 , . . . , vd ) span of vectors v1 , . . . , vd
λmin (G) minimum eigenvalue of matrix G
P
hx, yi inner product, hx, yi = i xi yi
kxkp p-norm of vector x
kxk2G x> Gx for positive definite G ∈ Rd×d and x ∈ Rd
 Notation 6

≺,  Loewner partial order of positive semidefinite matrices: A  B

(A ≺ B) if B − A is positive semidefinite (respectively, definite).
Distributions
N (µ, σ 2 ) Normal distribution with mean µ and variance σ 2
B(p) Bernoulli distribution with mean p
U(a, b) uniform distribution supported on [a, b]
Beta(α, β) Beta distribution with parameters α, β > 0
δx Dirac distribution with point mass at x

Topological
cl(A) closure of set A
int(A) interior of set A
∂A boundary of a set A, ∂A = cl(A) \ int(A)
co(A) convex hull of A
aff(A) affine hull of A
ri(A) relative interior of A
Part I
Bandits, Probability and
Concentration
1 Introduction

Bandit problems were introduced by William R. Thompson in an article

published in 1933 in Biometrika. Thompson was interested in medical
trials and the cruelty of running a trial blindly, without adapting the
treatment allocations on the fly as the drug appears more or less effective.
The name comes from the 1950s, when
Frederick Mosteller and Robert Bush decided
to study animal learning and ran trials on
mice and then on humans. The mice faced
the dilemma of choosing to go left or right
after starting in the bottom of a T-shaped
maze, not knowing each time at which end
they would find food. To study a similar
learning setting in humans, a ‘two-armed
bandit’ machine was commissioned where
humans could choose to pull either the left or
Figure 1.1 Mouse learning a T-maze.
the right arm of the machine, each giving a
random pay-off with the distribution of pay-
offs for each arm unknown to the human player. The machine was called a
‘two-armed bandit’ in homage to the one-armed bandit, an old-fashioned name
for a lever-operated slot machine (‘bandit’ because they steal your money).
There are many reasons to care about bandit problems. Decision-making with
uncertainty is a challenge we all face, and bandits provide a simple model of
this dilemma. Bandit problems also have practical applications. We already
mentioned clinical trial design, which researchers have used to motivate their
work for 80 years. We can’t point to an example where bandits have actually
been used in clinical trials, but adaptive experimental design is gaining popularity
and is actively encouraged by the US Food and Drug Administration, with the
justification that not doing so can lead to the withholding of effective drugs until
long after a positive effect has been established.
While clinical trials are an important application for the future, there are
applications where bandit algorithms are already in use. Major tech companies
use bandit algorithms for configuring web interfaces, where applications include
news recommendation, dynamic pricing and ad placement. A bandit algorithm
 Introduction 9

plays a role in Monte Carlo Tree Search, an algorithm made famous by the recent
success of AlphaGo.
Finally, the mathematical formulation of bandit problems leads to a rich
structure with connections to other branches of mathematics. In writing this
book (and previous papers), we have read books on convex analysis/optimisation,
Brownian motion, probability theory, concentration analysis, statistics, differential
geometry, information theory, Markov chains, computational complexity and more.
What fun!
A combination of all these factors has led to an enormous growth in research
over the last two decades. Google Scholar reports less than 1000, then 2700 and
7000 papers when searching for the phrase ‘bandit algorithm’ for the periods of
2001–5, 2006–10, and 2011–15, respectively, and the trend just seems to have
strengthened since then, with 5600 papers coming up for the period of 2016 to
the middle of 2018. Even if these numbers are somewhat overblown, they are
indicative of a rapidly growing field. This could be a fashion, or maybe there is
something interesting happening here. We think that the latter is true.

A Classical Dilemma
Imagine you are playing a two-armed bandit machine and you already pulled
each lever five times, resulting in the following pay-offs (in dollars):

Round 1 2 3 4 5 6 7 8 9 10

left 0 10 0 0 10

right 10 0 0 0 0

The left arm appears to be doing slightly better. The

average pay-off for this arm is $4, while the average for the Figure 1.2 Two-
right arm is only $2. Let’s say you have 10 more trials (pulls) armed bandit
altogether. What is your strategy? Will you keep pulling
the left arm, ignoring the right? Or would you attribute the poor performance of
the right arm to bad luck and try it a few more times? How many more times?
This illustrates one of the main interests in bandit problems. They capture the
fundamental dilemma a learner faces when choosing between uncertain options.
Should one explore an option that looks inferior or exploit by going with the
option that looks best currently? Finding the right balance between exploration
and exploitation is at the heart of all bandit problems.
 1.1 The Language of Bandits 10

1.1 The Language of Bandits

A bandit problem is a sequential game between a learner and an environment.

The game is played over n rounds, where n is a positive natural number called
the horizon. In each round t ∈ [n], the learner first chooses an action At from a
given set A, and the environment then reveals a reward Xt ∈ R.

In the literature, actions are often also called ‘arms’. We talk about k-armed
bandits when the number of actions is k, and about multi-armed bandits
when the number of arms is at least two and the actual number is immaterial
to the discussion. If there are multi-armed bandits, there are also one-armed
bandits, which are really two-armed bandits where the pay-off of one of the
arms is a known fixed deterministic number.

Of course the learner cannot peek into the future when choosing their
actions, which means that At should only depend on the history Ht−1 =
(A1 , X1 , . . . , At−1 , Xt−1 ). A policy is a mapping from histories to actions: A
learner adopts a policy to interact with an environment. An environment is a
mapping from history sequences ending in actions to rewards. Both the learner
and the environment may randomise their decisions, but this detail is not so
important for now. The most common objective of the learner is to choose actions
that lead to the largest possible cumulative reward over all n rounds, which is
Pn
t=1 Xt .
The fundamental challenge in bandit problems is that the environment is
unknown to the learner. All the learner knows is that the true environment
lies in some set E called the environment class. Most of this book is about
designing policies for different kinds of environment classes, though in some cases
the framework is extended to include side observations as well as actions and
rewards.
The next question is how to evaluate a learner. We discuss several performance
measures throughout the book, but most of our efforts are devoted to
understanding the regret. There are several ways to define this quantity. To avoid
getting bogged down in details, we start with a somewhat informal definition.

Definition 1.1. The regret of the learner relative to a policy π (not necessarily
that followed by the learner) is the difference between the total expected reward
using policy π for n rounds and the total expected reward collected by the learner
over n rounds. The regret relative to a set of policies Π is the maximum regret
relative to any policy π ∈ Π in the set.

The set Π is often called the competitor class. Another way of saying all this
is that the regret measures the performance of the learner relative to the best
policy in the competitor class. We usually measure the regret relative to a set of
policies Π that is large enough to include the optimal policy for all environments
 1.1 The Language of Bandits 11

in E. In this case, the regret measures the loss suffered by the learner relative to
the optimal policy.
Example 1.2. Suppose the action set is A = {1, 2, . . . , k}. An environment is
called a stochastic Bernoulli bandit if the reward Xt ∈ {0, 1} is binary valued
and there exists a vector µ ∈ [0, 1]k such that the probability that Xt = 1 given
the learner chose action At = a is µa . The class of stochastic Bernoulli bandits is
the set of all such bandits, which are characterised by their mean vectors. If you
knew the mean vector associated with the environment, then the optimal policy
is to play the fixed action a∗ = argmaxa∈A µa . This means that for this problem
the natural competitor class is the set of k constant polices Π = {π1 , . . . , πk },
where πi chooses action i in every round. The regret over n rounds becomes
" n #
X
Rn = n max µa − E Xt ,
a∈A
t=1

where the expectation is with respect to the randomness in the environment and
policy. The first term in this expression is the maximum expected reward using
any policy. The second term is the expected reward collected by the learner.
For a fixed policy and competitor class, the regret depends on the environment.
The environments where the regret is large are those where the learner is behaving
worse. Of course the ideal case is that the regret be small for all environments.
The worst-case regret is the maximum regret over all possible environments.
One of the core questions in the study of bandits is to understand the growth
rate of the regret as n grows. A good learner achieves sublinear regret. Letting Rn
denote the regret over n rounds, this means that Rn = o(n) or equivalently that
limn→∞ Rn /n = 0. Of course one can ask for more. Under what circumstances is
√
Rn = O( n) or Rn = O(log(n))? And what are the leading constants? How does
the regret depend on the specific environment in which the learner finds itself?
We will discover eventually that for the environment class in Example 1.2, the
√
worst-case regret for any policy is at least Ω( n) and that there exist policies for
√
which Rn = O( n).

A large environment class corresponds to less knowledge by the learner. A

large competitor class means the regret is a more demanding criteria. Some
care is sometimes required to choose these sets appropriately so that (a)
guarantees on the regret are meaningful and (b) there exist policies that
make the regret small.

The framework is general enough to model almost anything by using a rich

enough environment class. This cannot be bad, but with too much generality it
becomes impossible to say much. For this reason, we usually restrict our attention
to certain kinds of environment classes and competitor classes.
A simple problem setting is that of stochastic stationary bandits. In this
 1.1 The Language of Bandits 12

case the environment is restricted to generate the reward in response to each action
from a distribution that is specific to that action and independent of the previous
action choices and rewards. The environment class in Example 1.2 satisfies these
conditions, but there are many alternatives. For example, the rewards could follow
a Gaussian distribution rather than Bernoulli. This relatively mild difference does
not change the nature of the problem in a significant way. A more drastic change
is to assume the action set A is a subset of Rd and that the mean reward for
choosing some action a ∈ A follows a linear model, Xt = ha, θi + ηt for θ ∈ Rd
and ηt a standard Gaussian (zero mean, unit variance). The unknown quantity
in this case is θ, and the environment class corresponds to its possible values
(E = Rd ).
For some applications, the assumption that the rewards are stochastic and
stationary may be too restrictive. The world mostly appears deterministic, even
if it is hard to predict and often chaotic looking. Of course, stochasticity has
been enormously successful in explaining patterns in data, and this may be
sufficient reason to keep it as the modelling assumption. But what if the stochastic
assumptions fail to hold? What if they are violated for a single round? Or just for
one action, at some rounds? Will our best algorithms suddenly perform poorly?
Or will the algorithms developed be robust to smaller or larger deviations from
the modelling assumptions?
An extreme idea is to drop all assumptions on how the rewards are generated,
except that they are chosen without knowledge of the learner’s actions and lie
in a bounded set. If these are the only assumptions, we get what is called the
setting of adversarial bandits. The trick to say something meaningful in this
setting is to restrict the competitor class. The learner is not expected to find
the best sequence of actions, which may be like finding a needle in a haystack.
Instead, we usually choose Π to be the set of constant policies and demand that
the learner is not much worse than any of these. By defining the regret in this
way, the stationarity assumption is transported into the definition of regret rather
than constraining the environment.
Of course there are all shades of grey between these two extremes. Sometimes
we consider the case where the rewards are stochastic, but not stationary. Or
one may analyse the robustness of an algorithm for stochastic bandits to small
adversarial perturbations. Another idea is to isolate exactly which properties of
the stochastic assumption are really exploited by a policy designed for stochastic
bandits. This kind of inverse analysis can help explain the strong performance of
policies when facing environments that clearly violate the assumptions they were
designed for.

1.1.1 Other Learning Objectives

We already mentioned that the regret can be defined in several ways, each
capturing slightly different aspects of the behaviour of a policy. Because the
regret depends on the environment, it becomes a multi-objective criterion: ideally,
 1.2 Applications 13

we want to keep the regret small across all possible environments. One way to
convert a multi-objective criterion into a single number is to take averages. This
corresponds to the Bayesian viewpoint where the objective is to minimise the
average cumulative regret with respect to a prior on the environment class.
Maximising the sum of rewards is not always the objective. Sometimes the
learner just wants to find a near-optimal policy after n rounds, but the actual
rewards accumulated over those rounds are unimportant. We will see examples
of this shortly.

1.1.2 Limitations of the Bandit Framework

One of the distinguishing features of all bandit problems studied in this book
is that the learner never needs to plan for the future. More precisely, we will
invariably make the assumption that the learner’s available choices and rewards
tomorrow are not affected by their decisions today. Problems that do require
this kind of long-term planning fall into the realm of reinforcement learning,
which is the topic of the final chapter. Another limitation of the bandit framework
is the assumption that the learner observes the reward in every round. The setting
where the reward is not observed is called partial monitoring and is the topic
of Chapter 37. Finally, often, the environment itself consists of strategic agents,
which the learner needs to take into account. This problem is studied in game
theory and would need a book on its own.

1.2 Applications

After this short preview, and as an appetiser before the hard work, we briefly
describe the formalisations of a variety of applications.

A/B Testing
The designers of a company website are trying to decide whether the ‘buy it now’
button should be placed at the top of the product page or at the bottom. In
the old days, they would commit to a trial of each version by splitting incoming
users into two groups of 10 000. Each group would be shown a different version
of the site, and a statistician would examine the data at the end to decide which
version was better. One problem with this approach is the non-adaptivity of the
test. For example, if the effect size is large, then the trial could be stopped early.
One way to apply bandits to this problem is to view the two versions of the
site as actions. Each time t a user makes a request, a bandit algorithm is used
to choose an action At ∈ A = {SiteA, SiteB}, and the reward is Xt = 1 if the
user purchases the product and Xt = 0 otherwise.
 1.2 Applications 14

In traditional A/B testing, the objective of the statistician is to decide which

website is better. When using a bandit algorithm, there is no need to end
the trial. The algorithm automatically decides when one version of the site
should be shown more often than another. Even if the real objective is to
identify the best site, then adaptivity or early stopping can be added to the
A/B process using techniques from bandit theory. While this is not the focus
of this book, some of the basic ideas are explained in Chapter 33.

Advert Placement
In advert placement, each round corresponds to a user visiting a website, and
the set of actions A is the set of all available adverts. One could treat this as
a standard multi-armed bandit problem, where in each round a policy chooses
At ∈ A, and the reward is Xt = 1 if the user clicked on the advert and Xt = 0
otherwise. This might work for specialised websites where the adverts are all
likely to be appropriate. But for a company like Amazon, the advertising should
be targeted. A user that recently purchased rock-climbing shoes is much more
likely to buy a harness than another user. Clearly an algorithm should take this
into account.
The standard way to incorporate this additional knowledge is to use the
information about the user as context. In its simplest formulation, this might
mean clustering users and implementing a separate bandit algorithm for each
cluster. Much of this book is devoted to the question of how to use side information
to improve the performance of a learner.
This is a good place to emphasise that the world is messy. The set of available
adverts is changing from round to round. The feedback from the user can be
delayed for many rounds. Finally, the real objective is rarely just to maximise
clicks. Other metrics such as user satisfaction, diversity, freshness and fairness,
just to mention a few, are important too. These are the kinds of issues that make
implementing bandit algorithms in the real world a challenge. This book will not
address all these issues in detail. Instead we focus on the foundations and hope
this provides enough understanding that you can invent solutions for whatever
peculiar challenges arise in your problem.

Recommendation Services
Netflix has to decide which movies to place most prominently in your ‘Browse’
page. Like in advert placement, users arrive at the page sequentially, and the
reward can be measured as some function of (a) whether or not you watched a
movie and (b) whether or not you rated it positively. There are many challenges.
First of all, Netflix shows a long list of movies, so the set of possible actions
is combinatorially large. Second, each user watches relatively few movies, and
individual users are different. This suggests approaches such as low-rank matrix
factorisation (a popular approach in ‘collaborative filtering’). But notice this is
 1.2 Applications 15

not an offline problem. The learning algorithm gets to choose what users see and
this affects the data. If the users are never recommended the AlphaGo movie,
then few users will watch it, and the amount of data about this film will be
scarce.

Network Routing
Another problem with an interesting structure is network routing, where the
learner tries to direct internet traffic through the shortest path on a network. In
each round the learner receives the start/end destinations for a packet of data.
The set of actions is the set of all paths starting and ending at the appropriate
points on some known graph. The feedback in this case is the time it takes for
the packet to be received at its destination, and the reward is the negation of
this value. Again the action set is combinatorially large. Even relatively small
graphs have an enormous number of paths. The routing problem can obviously
be applied to more physical networks such as transportation systems used in
operations research.

Dynamic Pricing
In dynamic pricing, a company is trying to automatically optimise the price of
some product. Users arrive sequentially, and the learner sets the price. The user
will only purchase the product if the price is lower than their valuation. What
makes this problem interesting is (a) the learner never actually observes the
valuation of the product, only the binary signal that the price was too low/too
high, and (b) there is a monotonicity structure in the pricing. If a user purchased
an item priced at $10, then they would surely purchase it for $5, but whether or
not it would sell when priced at $11 is uncertain. Also, the set of possible actions
is close to continuous.

Waiting Problems
Every day you travel to work, either by bus or by walking. Once you get on the
bus, the trip only takes 5 minutes, but the timetable is unreliable, and the bus
arrival time is unknown and stochastic. Sometimes the bus doesn’t come at all.
Walking, on the other hand, takes 30 minutes along a beautiful river away from
the road. The problem is to devise a policy for choosing how long to wait at
the bus stop before giving up and walking to minimise the time to get to your
workplace. Walk too soon, and you miss the bus and gain little information. But
waiting too long also comes at a price.
While waiting for a bus is not a problem we all face, there are other applications
of this setting. For example, deciding the amount of inactivity required before
putting a hard drive into sleep mode or powering off a car engine at traffic lights.
The statistical part of the waiting problem concerns estimating the cumulative
distribution function of the bus arrival times from data. The twist is that the
data is censored on the days you chose to walk before the bus arrived, which
is a problem analysed in the subfield of statistics called survival analysis. The
 1.3 Notes 16

interplay between the statistical estimation problem and the challenge of balancing
exploration and exploitation is what makes this and the other problems studied
in this book interesting.

Resource Allocation
A large part of operations research is focussed on designing strategies for allocating
scarce resources. When the dynamics of demand or supply are uncertain, the
problem has elements reminiscent of a bandit problem. Allocating too few
resources reveals only partial information about the true demand, but allocating
too many resources is wasteful. Of course, resource allocation is broad, and many
problems exhibit structure that is not typical of bandit problems, like the need
for long-term planning.

Tree Search
The UCT algorithm is a tree search algorithm commonly used in perfect-
information game-playing algorithms. The idea is to iteratively build a search
tree where in each iteration the algorithm takes three steps: (1) chooses a path
from the root to a leaf; (2) expands the leaf (if possible); (3) performs a Monte
Carlo roll-out to the end of the game. The contribution of a bandit algorithm is in
selecting the path from the root to the leaves. At each node in the tree, a bandit
algorithm is used to select the child based on the series of rewards observed
through that node so far. The resulting algorithm can be analysed theoretically,
but more importantly has demonstrated outstanding empirical performance in
game-playing problems.

1.3 Notes

1 The reader may find it odd that at one point we identified environments with
maps from histories to rewards, while we used the language that a learner
‘adopts a policy’ (a map from histories to actions). The reason is part historical
and part because policies and their design are at the center of the book, while
the environment strategies will mostly be kept fixed (and relatively simple).
On this note, strategy is also a word that sometimes used interchangeably with
policy.

1.4 Bibliographic Remarks

As we mentioned in the very beginning, the first paper on bandits was by

Thompson [1933]. The experimentation on mice and humans that led to the
name comes from the paper by Bush and Mosteller [1953]. Much credit for the
popularisation of the field must go to famous mathematician and statistician,
Herbert Robbins, whose name appears on many of the works that we reference,
 1.4 Bibliographic Remarks 17

with the earliest being: [Robbins, 1952]. Another early pioneer is Herman Chernoff,
who wrote papers with titles like ‘Sequential Decisions in the Control of a
Spaceship’ [Bather and Chernoff, 1967].
Besides these seminal papers, there are already a number of books on bandits
that may serve as useful additional reading. The most recent (and also most
related) is by Bubeck and Cesa-Bianchi [2012] and is freely available online. This is
an excellent book and is warmly recommended. The main difference between their
book and ours is that (a) we have the benefit of seven years of additional research
in a fast-moving field and (b) our longer page limit permits more depth. Another
relatively recent book is Prediction, Learning and Games by Cesa-Bianchi and
Lugosi [2006]. This is a wonderful book, and quite comprehensive. But its scope
is ‘all of’ online learning, which is so broad that bandits are not covered in great
depth. We should mention there is also a recent book on bandits by Slivkins
[2019]. Conveniently it covers some topics not covered in this book (notably
Lipschitz bandits and bandits with knapsacks). The reverse is also true, which
should not be surprising since our book is currently 400 pages longer. There are
also four books on sequential design and multi-armed bandits in the Bayesian
setting, which we will address only a little. These are based on relatively old
material, but are still useful references for this line of work and are well worth
reading [Chernoff, 1959, Berry and Fristedt, 1985, Presman and Sonin, 1990,
Gittins et al., 2011].
Without trying to be exhaustive, here are a few articles applying bandit
algorithms; a recent survey is by Bouneffouf and Rish [2019]. The papers
themselves will contain more useful pointers to the vast literature. We mentioned
AlphaGo already [Silver et al., 2016]. The tree search algorithm that drives its
search uses a bandit algorithm at each node [Kocsis and Szepesvári, 2006]. Le et al.
[2014] apply bandits to wireless monitoring, where the problem is challenging
due to the large action space. Lei et al. [2017] design specialised contextual
bandit algorithms for just-in-time adaptive interventions in mobile health: in
the typical application the user is prompted with the intention of inducing a
long-term beneficial behavioural change. See also the article by Greenewald et al.
[2017]. Rafferty et al. [2018] apply Thompson sampling to educational software
and note the trade-off between knowledge and reward. Sadly, by 2015, bandit
algorithms still have not been used in clinical trials, as explicitly mentioned
by Villar et al. [2015]. Microsoft offers a ‘Decision Service’ that uses bandit
algorithms to automate decision-making [Agarwal et al., 2016].
2 Foundations of Probability ( )

This chapter covers the fundamental concepts of measure-theoretic probability,

on which the remainder of this book relies. Readers familiar with this topic can
safely skip the chapter, but perhaps a brief reading would yield some refreshing
perspectives. Measure-theoretic probability is often viewed as a necessary evil,
to be used when a demand for rigour combined with continuous spaces breaks
the simple approach we know and love from high school. We claim that measure-
theoretic probability offers more than annoying technical machinery. In this
chapter we attempt to prove this by providing a non-standard introduction.
Rather than a long list of definitions, we demonstrate the intuitive power of
the notation and tools. For those readers with little prior experience in measure
theory this chapter will no doubt be a challenging read. We think the investment
is worth the effort, but a great deal of the book can be read without it, provided
one is willing to take certain results on faith.

2.1 Probability Spaces and Random Elements

The thrill of gambling comes from the fact that the bet is placed on future
outcomes that are uncertain at the time of the gamble. A central question in
gambling is the fair value of a game. This can be difficult to answer for all but
the simplest games. As an illustrative example, imagine the following moderately
complex game: I throw a dice. If the result is four, I throw two more dice; otherwise
I throw one dice only. Looking at each newly thrown dice (one or two), I repeat
the same, for a total of three rounds. Afterwards, I pay you the sum of the values
on the faces of the dice. How much are you willing to pay to play this game with
me?
Many examples of practical interest exhibit a complex random interdependency
between outcomes. The cornerstone of modern probability as proposed by
Kolmogorov aims to remove this complexity by separating the randomness from
the mechanism that produces the outcome.
Instead of rolling the dice one by one, imagine that sufficiently many dice were
rolled before the game has even started. For our game we need to roll seven
dice, because this is the maximum number that might be required (one in the
first round, two in the second round and four in the third round. See Fig. 2.1).
 2.1 Probability Spaces and Random Elements 19

X1 := throw()

No Yes
X1 = 4?

X21 := throw() X21 := throw() X22 := throw()

Figure 2.1 The initial phase of a gambling game with a random number of dice rolls.
Depending on the outcome of a dice roll, one or two dice are rolled for a total of three
rounds. The number of dice used will then be random in the range of three to seven.

After all the dice are rolled, the game can be emulated by ordering the dice and
revealing the outcomes sequentially. Then the value of the first dice in the chosen
ordering is the outcome of the dice in the first round. If we see a four, we look at
the next two dice in the ordering; otherwise we look at the single next dice.
By taking this approach, we get a simple calculus for the probabilities of all
kinds of events. Rather than directly calculating the likelihood of each pay-off,
we first consider the probability of any single outcome of the dice. Since there
are seven dice, the set of all possible outcomes is Ω = {1, . . . , 6}7 . Because
all outcomes are equally probable, the probability of any ω ∈ Ω is (1/6)7 . The
probability of the game pay-off taking value v can then be evaluated by calculating
the total probability assigned to all those outcomes ω ∈ Ω that would result
in the value of v. In principle, this is trivial to do thanks to the separation of
everything that is probabilistic from the rest. The set Ω is called the outcome
space, and its elements are the outcomes. Fig. 2.2 illustrates this idea. Random
outcomes are generated on the left, while on the right, various mechanisms are
used to arrive at values; some of these values may be observed and some not.
There will be much benefit from being a little more formal about how we
come up with the value of our artificial game. For this, note that the process by
which the game gets its value is a function X that maps Ω to the reals (simply,
X : Ω → R). We find it ironic that functions of this type (from the outcome
space to subsets of the reals) are called random variables. They are neither
random nor variables in a programming language sense. The randomness is in
the argument that X is acting on, producing randomly changing results. Later
we will put a little more structure on random variables, but for now it suffices to
think of them as maps from the outcome space to the reals.
 2.1 Probability Spaces and Random Elements 20

Outcomes

Randomising device
Mechanisms
all randomness

Figure 2.2 A key idea in probability theory is the separation of sources of randomness
from game mechanisms. A mechanism creates values from the elementary random
outcomes, some of which are visible for observers, while others may remain hidden.

We follow the standard convention in probability theory where random

variables are denoted by capital letters. Be warned that capital letters are
also used for other purposes as demanded by different conventions.

Pick some number v ∈ N. What is the probability of seeing X = v? As

described above, this probability is (1/6)7 times the size of the set X −1 (v) =
{ω ∈ Ω : X(ω) = v}. The set X −1 (v) is called the preimage of v under X. More
generally, the probability that X takes its value in some set A ⊆ N is given by
(1/6)7 times the cardinality of X −1 (A) = {ω ∈ Ω : X(ω) ∈ A}, where we have
overloaded the definition of X −1 to set-valued inputs.
Notice in the previous paragraph we only needed probabilities assigned to
subsets of Ω, regardless of the question asked. To make this a bit more general,
let us introduce a map P that assigns probabilities to certain subsets of Ω. The
intuitive meaning of P is as follows. Random outcomes are generated in Ω. The
probability that an outcome falls into a set A ⊂ Ω is P (A). If A is not in the
domain of P, then there is no answer to the question of the probability of the
outcome falling in A. But let’s postpone the discussion of why P should be
restricted to only certain subsets of Ω later. In the above example with the dice,
the set of subsets in the domain of P is not restricted and, in particular, for any
subset A ⊆ Ω, P (A) = (1/6)7 |A|.


The probability of seeing X taking the value of v is thus P X −1 (v) . To
minimise clutter, the more readable notation for this is P (X = v). But
always
keep in mind that this familiar form is just a shorthand for P X −1 (v) . More
generally, we also use

P (predicate(U, V, . . . )) = P ({ω ∈ Ω : predicate(U (ω), V (ω), . . . ) is true})

 2.1 Probability Spaces and Random Elements 21

with any predicate (an expression evaluating to true or false) where U, V, . . . are
functions with domain Ω.
What properties should P satisfy? Since Ω is the set of all possible outcomes,
it seems reasonable to expect that P is defined for Ω and P(Ω) = 1 and since ∅
contains no outcomes, P(∅) = 0 is also expected to hold. Furthermore, probabilities
should be non-negative so P(A) ≥ 0 for any A ⊂ Ω on which P is defined. Let
Ac = Ω \ A be the complement of A. Then we should expect that P is defined
for A exactly when it is defined for Ac and P(Ac ) = 1 − P(A) (negation rule).
Finally, if A, B are disjoint so that A ∩ B = ∅ and P(A), P(B) and P(A ∪ B) are
all defined, then P(A ∪ B) = P(A) + P(B). This is called the finite additivity
property.
Let F be the set of subsets of Ω on which P is defined. It would seem silly if
A ∈ F and Ac ∈ / F, since P(Ac ) could simply be defined by P(Ac ) = 1 − P(A).
Similarly, if P is defined on disjoint sets A and B, then it makes sense if A∪B ∈ F.
We will also require the additivity property to hold (i) regardless of whether
the sets are disjoint and (ii) even for countably infinitely many sets. If {Ai }i
is a collection of sets and Ai ∈ F for all i ∈ N, then ∪i Ai ∈ F, and if these
P
sets are pairwise disjoint, P(∪i Ai ) = i P(Ai ). A set of subsets that satisfies all
these properties is called a σ-algebra, which is pronounced ‘sigma-algebra’ and
sometimes also called a σ-field (see Note 1).
Definition 2.1 (σ-algebra and probability measures). A set F ⊆ 2Ω is a σ-
algebra if Ω ∈ F and Ac ∈ F for all A ∈ F and ∪i Ai ∈ F for all {Ai }i with
Ai ∈ F for all i ∈ N. That is, it should include the whole outcome space and
be closed under complementation and countable unions. A function P : F → R
is a probability measure if P(Ω) = 1 and for all A ∈ F, P(A) ≥ 0 and
P
P(Ac ) = 1 − P(A) and P(∪i Ai ) = i P(Ai ) for all countable collections of disjoint
sets {Ai }i with Ai ∈ F for all i. If F is a σ-algebra and G ⊂ F is also a σ-algebra,
then we say G is a sub-σ-algebra of F. If P is a measure defined on F, then
the restriction of P to G is a measure P|G on G defined by P|G (A) = P(A) for
all A ∈ G.
At this stage, the reader may rightly wonder about why we introduced the notion
of sub-σ-algebras. The answer should become clear quite soon. The elements
of F are called measurable sets. They are measurable in the sense that P
assigns values to them. The pair (Ω, F) alone is called a measurable space,
while the triplet (Ω, F, P) is called a probability space. If the condition that
P(Ω) = 1 is lifted, then P is called a measure. If the condition that P(A) ≥ 0
is also lifted, then P is called a signed measure. For measures and signed
measures, it would be unusual to use the symbol P, which is mostly reserved for
probabilities. Probability measures are also called probability distributions,
or just distributions.
Random variables lead to new probability

measures. In particular, in the
example above PX (A) = P X −1 (A)
is a probability measure defined for all the
subsets A of R for which P X −1 (A) is defined. More generally, for a random
 2.1 Probability Spaces and Random Elements 22

variable X, the probability measure PX is called the law of X, or the push-

forward measure of P under X.

The significance of the push-forward measure PX is that any probabilistic

question concerning X can be answered from the knowledge of PX alone.
Even Ω and the details of the map X are not needed. This is often used as
an excuse to not even mention the underlying probability space (Ω, F, P).

If we keep X fixed but change P (for example, by switching to loaded dice),

then the measure induced by X changes. We will often use arguments that do
exactly this, especially when proving lower bounds on the limits of how well
bandit algorithms can perform.
The astute reader would have noticed that we skipped over some details.
Measures are defined as functions from a σ-algebra to R, so if we want to call
PX a measure, then its domain {A ⊂ R : X −1 (A) ∈ F} better be a σ-algebra.
This holds in great generality. You will show in Exercise 2.3 that for functions
X : Ω → X with X arbitrary, the collection {A ⊂ X : X −1 (A) ∈ F} is a
σ-algebra.
It will be useful to generalise our example a little by allowing X to take on
values in sets other than the reals. For example, the range could be vectors
or abstract objects like sequences. Let (Ω, F) be a measurable space, X be an
arbitrary set and G ⊆ 2X . A function X : Ω → X is called an F /G-measurable
map if X −1 (A) ∈ F for all A ∈ G. Note that G need not be a σ-algebra.
When F and G are obvious from the context, X is called a measurable map.
What are the typical choices for G? When X is real-valued, it is usual to let
G = {(a, b) : a < b with a, b ∈ R} be the set of all open intervals. The reader can
verify that if X is F/G-measurable, then it is also F/σ(G)-measurable, where
σ(G) is the smallest σ-algebra that contains G. This smallest σ-algebra can be
shown to exist. Furthermore, it contains exactly those sets A that are in every
σ-algebra that contains G (see Exercise 2.5). When G is the set of open intervals,
σ(G) is usually denoted by B or B(R) and is called the Borel σ-algebra of R.
This definition is extended to Rk by replacing open intervals with open rectangles
Qk
of the form i=1 (ai , bi ), where a < b ∈ Rk . If G is the set of all such open
rectangles, then σ(G) is the Borel σ-algebra: B(Rk ). More generally, the Borel
σ-algebra of a topological space X is the σ-algebra generated by the open sets of
X.

Definition 2.2 (Random variables and elements). A random variable

(random vector) on measurable space (Ω, F) is a F/B(R)-measurable function
X : Ω → R (respectively F/B(Rk )-measurable function X : Ω → Rk ). A random
element between measurable spaces (Ω, F) and (X , G) is a F/G-measurable
function X : Ω → X .

Thus, random vectors are random elements where the range space is
 2.1 Probability Spaces and Random Elements 23

(Rk , B(Rk )), and random vectors are random variables when k = 1. Random
elements generalise random variables and vectors to functions that do not
take values in Rk . The push-forward measure (or law) can be defined for
any random element. Furthermore, random variables and vectors work nicely
together. If X1 , . . . , Xk are k random variables on the same domain (Ω, F),
then X(ω) = (X1 (ω), . . . , Xk (ω)) is an Rk -valued random vector, and vice versa
(Exercise 2.2). Multiple random variables X1 , . . . , Xk from the same measurable
space can thus be viewed as a random vector X = (X1 , . . . , Xk ).
Given a map X : Ω → X between measurable spaces (Ω, F) and (X , G), we let
σ(X) = {X −1 (A) : A ∈ G} be the σ-algebra generated by X. The map X is
F/G-measurable if and only if σ(X) ⊆ F. By checking the definitions one can
show that σ(X) is a sub-σ-algebra of F and in fact is the smallest sub-σ-algebra
for which X is measurable. If G = σ(A) itself is generated by a set system
A ⊂ 2X , then to check the F/G-measurability of X, it suffices to check whether
X −1 (A) = {X −1 (A) : A ∈ A} is a subset of F. The reason this is sufficient is
because σ(X −1 (A)) = X −1 (σ(A)), and by definition the latter is σ(X). In fact,
to check whether a map is measurable, either one uses the composition rule or
checks X −1 (A) ⊂ F for a ‘generator’ A of G.
Random elements can be combined to produce new random elements by
composition. One can show that if f is F/G-measurable and g is G/H-measurable
for σ-algebras F, G and H over appropriate spaces, then their composition g ◦ f
is F/H-measurable (Exercise 2.1). This is used most often for Borel functions,
which is a special name for B(Rm )/B(Rn )-measurable functions from Rm to
Rn . These functions are also called Borel measurable. The reader will find it
pleasing that all familiar functions are Borel. First and foremost, all continuous
functions are Borel, which includes elementary operations such as addition and
multiplication. Continuity is far from essential, however. In fact one is hard-
pressed to construct a function that is not Borel. This means the usual operations
are ‘safe’ when working with random variables.

Indicator Functions
Given an arbitrary set Ω and A ⊆ Ω, the indicator function of A is
IA : Ω → {0, 1} given by
(
1 , if ω ∈ A ;
IA (ω) =
0 , otherwise .

Sometimes A has a complicated description, and it becomes convenient to abuse

notation by writing I {ω ∈ A} instead of IA (ω). Similarly, we will often write
I {predicate(X, Y, . . .)} to mean the indicator function of the subset of Ω on
which the predicate is true. It is easy to check that an indicator function IA is a
random variable on (Ω, F) if and only if A is measurable: A ∈ F.
 2.1 Probability Spaces and Random Elements 24

Why So Complicated?
You may be wondering why we did not define P on the power set of Ω, which
is equivalent to declaring that all sets are measurable. In many cases this is a
perfectly reasonable thing to do, including the example game where nothing
prevents us from defining F = 2Ω . However, beyond this example, there are two
justifications not to have F = 2Ω , the first technical and the second conceptual.
The technical reason is highlighted by the following surprising theorem
according to which there does not exist a uniform probability distribution on
Ω = [0, 1] if F is chosen to be the power set of Ω (a uniform probability distribution
over [0, 1], if existed, would have the property of assigning its length to every
interval). In other words, if you want to be able to define the uniform measure,
then F cannot be too large. By contrast, the uniform measure can be defined
over the Borel σ-algebra, though proving this is not elementary.
Theorem 2.3. Let Ω = [0, 1], and F be the power set of Ω. Then there does not
exist a measure P on (Ω, F) such that P([a, b]) = b − a for all 0 ≤ a ≤ b ≤ 1.
The main conceptual reason of why not to have F = 2Ω is because then we
can use σ-algebras to represent information. This is especially useful in the study
of bandits where the learner is interacting with an environment and is slowly
gaining knowledge. One useful way to represent this is by using a sequence of
nested σ-algebras, as we explain in the next section. One might also be worried
that the Borel σ-algebra does not contain enough measurable sets. Rest assured
that this is not a problem and you will not easily find a non-measurable set. For
completeness, an example of a non-measurable set will still be given in the notes,
along with a little more discussion on this topic.
A second technical reason to prefer the measure-theoretic approach to
probabilities is that this approach allows for the unification of distributions
on discrete spaces and densities on continuous ones (the uninitiated reader will
find the definitions of these later). This unification can be necessary when dealing
with random variables that combine elements of both, e.g. a random variable
that is zero with probability 1/2 and otherwise behaves like a standard Gaussian.
Random variables like this give rise to so-called “mixed continuous and discrete
distributions”, which seem to require special treatment in a naive approach
to probabilities, yet dealing with random variables like these are nothing but
ordinary under the measure-theoretic approach.

From Laws to Probability Spaces and Random Variables

A big ‘conspiracy’ in probability theory is that probability spaces are seldom
mentioned in theorem statements, despite the fact that a measure cannot be
defined without one. Statements are instead given in terms of random elements
and constraints on their joint probabilities. For example, suppose that X and Y
are random variables such that
|A ∩ [6]| |B ∩ [2]|
P (X ∈ A, Y ∈ B) = · for all A, B ∈ B(R) , (2.1)
6 2
 2.1 Probability Spaces and Random Elements 25

which represents the joint distribution for the values of a dice (X ∈ [6]) and coin
(Y ∈ [2]). The formula describes some constraints on the probabilistic interactions
between the outputs of X and Y , but says nothing about their domain. In a way,
the domain is an unimportant detail. Nevertheless, one must ask whether or not
an appropriate domain exists at all. More generally, one may ask whether an
appropriate probability space exists given some constraints on the joint law of a
collection X1 , . . . , Xk of random variables. For this to make sense, the constraints
should not contradict each other, which means there is a probability measure
µ on B(Rk ) such that µ satisfies the postulated constraints. But then we can
choose Ω = Rk , F = B(Rk ), P = µ and Xi : Ω → R to be the ith coordinate
map: Xi (ω) = ωi . The push-forward of P under X = (X1 , . . . , Xk ) is µ, which by
definition is compatible with the constraints.
A more specific question is whether for a particular set of constraints on the
joint law there exists a measure µ compatible with the constraints. Very often the
constraints are specified for elements of the cartesian product of finitely many
σ-algebras, like in Eq. (2.1). If (Ω1 , F1 ), . . . , (Ωn , Fn ) are measurable spaces, then
the cartesian product of F1 , . . . Fn is
F1 × · · · × Fn = {A1 × · · · × An : A1 ∈ F1 , . . . , An ∈ Fn } ⊆ 2Ω1 ×···×Ωn .
Elements of this set are known as measurable rectangles in Ω1 × · · · × Ωn .
Theorem 2.4 (Carathéodory’s extension theorem). Let (Ω1 , F1 ), . . . , (Ωn , Fn )
be measurable spaces and µ̄ : F1 × · · · × Fn → [0, 1] be a function such that
(a) µ̄(Ω1 × · · · × Ωn ) = 1; and
P∞
k=1 Ak ) =
(b) µ̄(∪∞ k=1 µ̄(Ak ) for all sequences of disjoint sets with Ak ∈
F1 × · · · × Fn .
Let Ω = Ω1 × · · · × Ωn and F = σ(F1 × · · · × Fn ). Then there exists a unique
probability measure µ on (Ω, F) such that µ agrees with µ̄ on F1 × · · · × Fn .
The theorem is applied by letting Ωk = R and Fk = B(R). Then the values of
a measure on all cartesian products uniquely determines its value everywhere.

It is not true that F1 ×F2 = σ(F1 ×F2 ). Take, for example, F1 = F2 = 2{1,2} .
Then, |F1 × F2 | = 1 + 3 × 3 = 10 (because ∅ × X = ∅), while, since
F1 × F2 includes the singletons of 2{1,2}×{1,2} , σ(F1 × F2 ) = 2{1,2}×{1,2} .
Hence, six sets are missing from F1 × F2 . For example, {(1, 1), (2, 2)} ∈
σ(F1 × F2 ) \ F1 × F2 .

The σ-algebra σ(F1 × · · · × Fn ) is called the product σ-algebra of (Fk )k∈[n]

and is also denoted by F1 ⊗ · · · ⊗ Fn . The product operation turns out to be
associative: (F1 ⊗ F2 ) ⊗ F3 = F1 ⊗ (F2 ⊗ F3 ), which justifies writing F1 ⊗ F2 ⊗ F3 .
As it turns out, things work out well again with Borel σ-algebras: for p, q ∈ N+ ,
B(Rp+q ) = B(Rp ) ⊗ B(Rq ). Needless to say, the same holds when there are more
 2.2 σ-Algebras and Knowledge 26

than two terms in the product. The n-fold product σ-algebra of F is denoted by
F ⊗n .

2.2 σ-algebras and knowledge

One of the conceptual advantages of measure-theoretic probability is the

relationship between σ-algebras and the intuitive idea of ‘knowledge’. Although
the relationship is useful and intuitive, it is regrettably not quite perfect. Let
(Ω, F), (X , G) and (Y, H) be measurable spaces and X : Ω → X and Y : Ω → Y
be random elements. Having observed the value of X (‘knowing X’), one might
wonder what this entails about the value of Y . Even more simplistically, under
what circumstances can the value of Y be determined exactly having observed X?
The situation is illustrated in Fig. 2.3. As it turns out, with some restrictions, the
answer can be given in terms of the σ-algebras generated by X and Y . Except

X
(Ω, F) (X , G)

f
Y
(Y, H)

Figure 2.3 The factorisation problem asks whether there exists a (measurable) function
f that makes the diagram commute.

for a technical assumption on (Y, H), the following result shows that Y is a
measurable function of X if and only if Y is σ(X)/H-measurable. The technical
assumption mentioned requires (Y, H) to be a Borel space, which is true of all
probability spaces considered in this book, including (Rk , B(Rk )). We leave the
exact definition of Borel spaces to the next chapter.
Lemma 2.5 (Factorisation lemma). Assume that (Y, H) is a Borel space. Then Y
is σ(X)-measurable (σ(Y ) ⊆ σ(X)) if and only if there exists a G/H-measurable
map f : X → Y such that Y = f ◦ X.
In this sense σ(X) contains all the information that can be extracted from X
via measurable functions. This is not the same as saying that Y can be deduced
from X if and only if Y is σ(X)-measurable because the set of X → Y maps
can be much larger than the set of G/H-measurable functions. When G is coarse,
there are not many G/H-measurable functions with the extreme case occurring
when G = {X , ∅}. In cases like this, the intuition that σ(X) captures all there
is to know about X is not true anymore (Exercise 2.6). The issue is that σ(X)
does not only depend on X, but also on the σ-algebra of (X , G) and that if G is
coarse-grained, then σ(X) can also be coarse-grained and not many functions
will be σ(X)-measurable. If X is a random variable, then by definition X = R
 2.2 σ-Algebras and Knowledge 27

and G = B(R), which is relatively fine-grained, and the requirement that f

be measurable is less restrictive. Nevertheless, even in the nicest setting where
Ω = X = Y = R and F = G = H = B(R), it can still occur that Y = f ◦ X for
some non-measurable f . In other words, all the information about Y exists in X
but cannot be extracted in a measurable way. These problems only occur when
X maps measurable sets in Ω to non-measurable sets in X . Fortunately, while
such random variables exist, they are never encountered in applications, which
provides the final justification for thinking of σ(X) as containing all that there is
to know about any random variable X that one may ever expect to encounter.

Filtrations
In the study of bandits and other online settings, information is revealed to the
learner sequentially. Let X1 , . . . , Xn be a collection of random variables on a
common measurable space (Ω, F). We imagine a learner is sequentially observing
the values of these random variables. First X1 , then X2 and so on. The learner
needs to make a prediction, or act, based on the available observations. Say, a
prediction or an act must produce a real-valued response. Then, having observed
.
X1:t = (X1 , . . . , Xt ), the set of maps f ◦ X1:t where f : Rt → R is Borel, captures
all the possible ways the learner can respond. By Lemma 2.5, this set contains
exactly the σ(X1:t )/B(R)-measurable maps. Thus, if we need to reason about
the set of Ω → R maps available after observing X1:t , it suffices to concentrate
on the σ-algebra Ft = σ(X1:t ). Conveniently, Ft is independent of the space of
possible responses, and being a subset of F, it also hides details about the range
space of X1:t . It is easy to check that F0 ⊆ F1 ⊆ F2 ⊆ · · · ⊆ Fn ⊆ F, which
means that more and more functions are becoming Ft -measurable as t increases,
which corresponds to increasing knowledge (note that F0 = {∅, Ω}, and the set
of F0 -measurable functions is the set of constant functions on Ω).
Bringing these a little further, we will often find it useful to talk about increasing
sequences of σ-algebras without constructing them in terms of random variables
as above. Given a measurable space (Ω, F), a filtration is a sequence (Ft )nt=0 of
sub-σ-algebras of F where Ft ⊆ Ft+1 for all t < n. We also allow n = ∞, and in
this case we define
∞
!
[
F∞ = σ Ft
t=0

to be the smallest σ-algebra containing the union of all Ft . Filtrations can also
be defined in continuous time, but we have no need for that here. A sequence
of random variables (Xt )nt=1 is adapted to filtration F = (Ft )nt=0 if Xt is Ft -
measurable for each t. We also say in this case that (Xt )t is F-adapted. The
same nomenclature applies if n is infinite. Finally, (Xt )t is F-predictable if Xt
is Ft−1 -measurable for each t ∈ [n]. Intuitively we may think of an F-predictable
process X = (Xt )t as one that has the property that Xt can be known (or
‘predicted’) based on Ft−1 , while a F-adapted process is one that has the property
that Xt can be known based on Ft only. Since Ft−1 ⊆ Ft , a predictable process
 2.3 Conditional Probabilities 28

is also adapted. A filtered probability space is the tuple (Ω, F, F, P), where
(Ω, F, P) is a probability space and F = (Ft )t is filtration of F.

2.3 Conditional Probabilities

Conditional probabilities are introduced so that we can talk about how

probabilities should be updated when one gains some partial knowledge about a
random outcome. Let (Ω, F, P) be a probability space, and let A, B ∈ F be such
that P (B) > 0. The conditional probability P (A | B) of A given B is defined
as
P (A ∩ B)
P (A | B) = .
P (B)
We can think about the outcome ω ∈ Ω as the result of throwing a many-sided
dice. The question asked is the probability that the dice landed so that ω ∈ A
given that it landed with ω ∈ B. The meaning of the condition ω ∈ B is that we
focus on dice rolls when ω ∈ B is true. All dice rolls when ω ∈ B does not hold
are discarded. Intuitively, what should matter in the conditional probability of A
given B is how large the portion of A is that lies in B, and this is indeed what
the definition means.

The importance of conditional probabilities is that they define a calculus of

how probabilities are to be updated in the presence of extra information.

The probability P (A | B) is also called the a posteriori (‘after the fact’)

probability of A given B. The a priori probability is P (A). Note that P (A | B) is
defined for every A ∈ F as long as P (B) > 0. In fact, A 7→ P (A | B) is a probability
measure over the measure space (Ω, F) called the a posteriori probability measure
given B (see Exercise 2.7). In a way the temporal characteristics attached to
the words ‘a posteriori’ and ‘a priori’ can be a bit misleading. Probabilities are
concerned with predictions. They express the degrees of uncertainty one assigns
to future events. The conditional probability of A given B is a prediction of
certain properties of the outcome of the random experiment that results in ω
given a certain condition. Everything is related to a future hypothetical outcome.
Once the dice is rolled, ω gets fixed, and either ω ∈ A, B or not. There is no
uncertainty left: predictions are trivial after an experiment is done.
Bayes rule states that provided events A, B ∈ F both occur with positive
probability,
P (B | A) P (A)
P (A | B) = . (2.2)
P (B)
Bayes rule is useful because it allows one to obtain P (A | B) based on information
about the quantities on the right-hand side. Remarkably, this happens to be
 2.4 Independence 29

the case quite often, explaining why this simple formula has quite a status in
probability and statistics. Exercise 2.8 asks the reader to verify this law.

2.4 Independence

Independence is another basic concept of probability that relates to

knowledge/information. In its simplest form, independence is a relation that
holds between events on a probability space (Ω, F, P). Two events A, B ∈ F are
independent if

P (A ∩ B) = P (A) P (B) . (2.3)

How is this related to knowledge? Assuming that P (B) > 0, dividing both sides
by P (B) and using the definition of conditional probability, we get that the above
is equivalent to

P (A | B) = P (A) . (2.4)

Of course, we also have that if P (A) > 0, (2.3) is equivalent to P (B | A) = P (B).

Both of the latter relations express that A and B are independent if the probability
assigned to A (or B) remains the same regardless of whether it is known that B
(respectively, A) occurred.
We hope our readers will find the definition of independence in terms of a ‘lack
of influence’ to be sensible. The reason not to use Eq. (2.4) as the definition is
mostly for the sake of convenience. If we started with (2.4), we would need to
separately discuss the case of P (B) = 0, which would be cumbersome. A second
reason is that (2.4) suggests an asymmetric relationship, but intuitively we expect
independence to be symmetric.
Uncertain outcomes are often generated part by part with no interaction
between the processes, which naturally leads to an independence structure (think
of rolling multiple dice with no interactions between the rolls). Once we discover
some independence structure, calculations with probabilities can be immensely
simplified. In fact, independence is often used as a way of constructing probability
measures of interest (cf. Eq. (2.1), Theorem 2.4 and Exercise 2.9). Independence
can also appear serendipitously in the sense that a probability space may hold
many more independent events than its construction may suggest (Exercise 2.10).

You should always carefully judge whether assumptions about independence

are really justified. This is part of the modelling and hence is not
mathematical in nature. Instead you have to think about the physical
process being modelled.

A collection of events G ⊂ F is said to be pairwise independent if any two

distinct elements of G are independent of each other. The events in G are said
 2.5 Integration and Expectation 30

to be mutually independent if for any n > 0 integer and A1 , . . . , An distinct

Qn
elements of G, P (A1 ∩ · · · ∩ An ) = i=1 P (Ai ). This is a stronger restriction
than pairwise independence. In the case of mutually independent events, the
knowledge of joint occurrence of any finitely many events from the collection will
not change our prediction of whether some other event in the collection happens.
But this may not be the case when the events are only pairwise independent
(Exercise 2.10). Two collections of events G1 , G2 are said to be independent of
each other if for any A ∈ G1 and B ∈ G2 it holds that A and B are independent.
This definition is often applied to σ-algebras.
When the σ-algebras are induced by random variables, this leads to the
definition of independence between random variables. Two random
variables X and Y are independent if σ(X) and σ(Y ) are independent of each
other. The notions of pairwise and mutual independence can also be naturally
extended to apply to collections of random variables. All these concepts can be
and are in fact extended to random elements.
The default meaning of independence when multiple events or random variables
are involved is mutual independence.

When we say that X1 , . . . , Xn are independent random variables, we mean

that they are mutually independent. Independence is always relative to
some probability measure, even when a probability measure is not explicitly
mentioned. In such cases the identity of the probability measure should be
clear from the context.

2.5 Integration and Expectation

A key quantity in probability theory is the expectation of a random variable. Fix

a probability space (Ω, F, P) and random variable X : Ω → R. The expectation X
is often denoted by E [X]. This notation unfortunately obscures the dependence
on the measure P. When the underlying measure is not obvious from context, we
write EP to indicate the expectation with respect to P. Mathematically, we define
the expected value of X as its Lebesgue integral with respect to P:
Z
E [X] = X(ω) dP(ω) .
Ω
R
The right-hand side is also often abbreviated to X dP. The integral on the
right-hand side is constructed to satisfy the following two key properties:

(a) The integral of indicators is the probability of the underlying

R event. If X(ω) =
I {ω ∈ A} is an indicator function for some A ∈ F, then XdP = P (A).
(b) Integrals are linear. For all random variables X1 , X2 and reals α1 , α2 such
 2.5 Integration and Expectation 31

R R R
that X1 dP and X2 dP are defined, (α1 X1 + α2 X2 )dP is defined and
satisfies
Z Z Z
(α1 X1 + α2 X2 ) dP = α1 X1 dP + α2 X2 dP . (2.5)
Ω Ω Ω
Pn
These two properties together tell us that whenever X(ω) = i=1 αi I {ω ∈ Ai }
for some n, αi ∈ R and Ai ∈ F, i = 1, . . . , n, then
Z X
XdP = αi P (Ai ) . (2.6)
Ω i

Functions of the form X are called simple functions.

In defining the Lebesgue integral of some random variable X, we use (2.6) as
the definition of the integral when X is a simple function. The next step is to
extend the definition to non-negative random variables. Let X : Ω → [0, ∞) be
measurable. The idea is to approximate X from below using simple functions
and take the largest value that can be obtained this way:
Z Z 
XdP = sup h dP : h is simple and 0 ≤ h ≤ X . (2.7)
Ω Ω

The meaning of U ≤ V for random variables U, V is that U (ω) ≤ V (ω) for all
ω ∈ Ω. The supremum on the right-hand side could be infinite, in which case we
say the integral of X is not defined. Whenever the integral of X is defined, we
say that X is integrable or, if the identity of the measure P is unclear, that X
is integrable with respect to P.
Integrals for arbitrary random variables are defined by decomposing the
random variable into positive and negative parts. Let X : Ω → R be any
measurable function. Then define X + (ω) = X(ω)I {X(ω) > 0} and X − (ω) =
−X(ω)I {X(ω) < 0} so that X(ω) = X + (ω) − X − (ω). Now X + and X − are
both non-negative random variables called the positive and negative parts of
X. Provided that both X + and X − are integrable, we define
Z Z Z
XdP = X + dP − X − dP .
Ω Ω Ω

Note that X is integrable if and only if the non-negative-valued random variable

|X| is integrable (Exercise 2.12).

None of what we have done depends on P being a probability measure. The

definitions hold for any measure, though for signed measures it is necessary to
split Ω into disjoint measurable sets on which the measure is positive/negative,
an operation that is possible by the Hahn decomposition theorem. We
will never need signed measures in this book, however.

A particularly interesting case is when Ω = R is the real line, F = B(R) is

the Borel σ-algebra and the measure is the Lebesgue measure λ, which is the
 2.5 Integration and Expectation 32

unique measure on B(R) such that λ((a, b)) = b − a for any a ≤ b. In this scenario,
if f : R → R is a Borel-measurable function, then we can write the Lebesgue
integral of f with respect to the Lebesgue measure as
Z
f dλ .
R

Perhaps unsurprisingly, this almost always coincides

R∞ with the improper Riemann
integral of f , which is normally written as −∞ f (x)dx. Precisely, if |f | is both
Lebesgue integrable and Riemann integrable, then the integrals are equal.

There exist functions that are Riemann integrable and not Lebesgue
integrable, and also the other way around (although examples of the former
are more exotic than the latter).

The Lebesgue measure and its relation to Riemann integration is mentioned

because when it comes to actually calculating the value of an expectation or
integral, this is often reduced to calculating integrals over the real line with
respect to the Lebesgue measure. The calculation is then performed by evaluating
the Riemann integral, thereby circumventing the need to rederive the integral
of many elementary functions. Integrals (and thus expectations) have a number
of important properties. By far the most important is their linearity, which was
postulated above as the second property in (2.5). To practice using the notation
with expectations, we restate the first half of this property. In fact, the statement
is slightly more general than what we demanded for integrals above.

Proposition 2.6. Let (Xi )i be a (possibly infinite) sequence of random variables

on the same probability space and assume that E [Xi ] exists for all i and
P P
furthermore that X = i Xi and E [ i |Xi |] also exist. Then
X
E [X] = E [Xi ] .
i

This exchange of expectations and summation is the source of much magic

in probability theory because it holds even if Xi are not independent. This
means that (unlike probabilities) we can very often decouple the expectations of
dependent random variables, which often proves extremely useful (a collection
of random variables is dependent if they are not independent). You will prove
Proposition 2.6 in Exercise 2.14. The other requirement for linearity is that if
c ∈ R is a constant, then E [cX] = c E [X] (Exercise 2.15).
Another important statement is concerned with independent random variables.

Proposition 2.7. If X and Y are independent, then E [XY ] = E [X] E [Y ].

In general E [XY ] =
6 E [X] E [Y ] (Exercise 2.18). Finally, an important simple
result connects expectations of non-negative random variables to their tail
probabilities.
 2.6 Conditional Expectation 33

Proposition 2.8. If X ≥ 0 is a non-negative random variable, then

Z ∞
E [X] = P (X > x) dx .
0

The integrand in Proposition 2.8 is called the tail probability function

x 7→ P (X > x) of X. This is also known as the complementary cumulative
distribution function of X. The cumulative distribution function (CDF) of
X is defined as x 7→ P (X ≤ x) and is usually denoted by FX . These functions
are defined for all random variables, not just non-negative ones. One can check
that FX : R → [0, 1] is increasing, right continuous and limx→−∞ FX (x) = 0 and
limx→∞ FX (x) = 1. The CDF of a random variable captures every aspect of the
probability measure PX induced by X, while still being just a function on the real
line, a property that makes it a little more human friendly than PX . One can also
generalise CDFs to random vectors: if X is an Rk -valued random vector, then its
CDF is defined as the FX : Rk → [0, 1] function that satisfies FX (x) = P (X ≤ x),
where, in line with our conventions, X ≤ x means that all components of X are
less than or equal to the respective component of x. The pushforward PX of a
random element is an alternative way to summarise the distribution of X. In
particular, for any real-valued, f : X → R measurable function,
Z
E [f (X)] = f (x)dPX (x)
X

provided that either the right-hand side, or the left-hand side exist.

2.6 Conditional Expectation

Conditional expectation allows us to talk about the expectation of a random

variable given the value of another random variable, or more generally, given
some σ-algebra.

Example 2.9. Let (Ω, F, P) model the outcomes of an unloaded dice: Ω = [6],
F = 2Ω and P(A) = |A|/6. Define two random variables X and Y by
Y (ω) = I {ω > 3} and X(ω) = ω. Suppose we are interested in the expectation
of X given a specific value of Y . Arguing intuitively, we might notice that Y = 1
means that the unobserved X must be either 4, 5 or 6, and that each of these
outcomes is equally likely, and so the expectation of X given Y = 1 should
be (4 + 5 + 6)/3 = 5. Similarly, the expectation of X given Y = 0 should be
(1 + 2 + 3)/3 = 2. If we want a concise summary, we can just write that ‘the
expectation of X given Y ’ is 5Y + 2(1 − Y ). Notice how this is a random variable
itself.

The notation for this conditional expectation is E [X | Y ]. Using this notation,

in Example 2.9 we can concisely write E [X | Y ] = 5Y + 2(1 − Y ). A little more
generally, if X : Ω → X and Y : Ω → Y with X , Y ⊂ R and |X |, |Y| < ∞, then
 2.6 Conditional Expectation 34

E[X | Y ] : Ω → R is the random variable given by E[X | Y ](ω) = E[X | Y = Y (ω)],

where
X X x P (X = x, Y = y)
E[X | Y = y] = x P (X = x | Y = y) = . (2.8)
P (Y = y)
x∈X x∈X

This is undefined when P(Y = y) = 0 so that E[X | Y ](ω) is undefined on the

measure zero set {ω : P(Y = Y (ω)) = 0}.
Eq. (2.8) does not generalise to continuous random variables because P (Y = y)
in the denominator might be zero for all y. For example, let Y be a random
variable taking values on [0, 1] according to a uniform distribution and X ∈ {0, 1}
be Bernoulli with biasRY . This means that the joint measure on X and Y is
P (X = 1, Y ∈ [p, q]) = p xdx for 0 ≤ p < q ≤ 1. Intuitively it seems like E[X | Y ]
q

should be equal to Y , but how to define it? The mean of a Bernoulli random
variable is equal to its bias so the definition of conditional probability shows that
for 0 ≤ p < q ≤ 1,

E[X = 1 | Y ∈ [p, q]] = P (X = 1 | Y ∈ [p, q])

P (X = 1, Y ∈ [p, q])
=
P (Y ∈ [p, q])
q − p2
2
=
2(q − p)
p+q
= .
2
This calculation is not well defined when p = q because P (Y ∈ [p, p]) = 0.
Nevertheless, letting q = p + ε for ε > 0 and taking the limit as ε tends to zero
seems like a reasonable way to argue that P (X = 1 | Y = p) = p. Unfortunately
this approach does not generalise to abstract spaces because there is no canonical
way of taking limits towards a set of measure zero, and different choices lead to
different answers.
Instead we use Eq. (2.8) as the starting point for an abstract definition of
conditional expectation as a random variable satisfying two requirements. First,
from Eq. (2.8) we see that E[X | Y ](ω) should only depend on Y (ω) and so
should be measurable with respect to σ(Y ). The second requirement is called the
‘averaging property’. For measurable A ⊆ Y, Eq. (2.8) shows that
X
E[IY −1 (A) E[X | Y ]] = P (Y = y) E[X | Y = y]
y∈A
XX
= x P (X = x, Y = y)
y∈A x∈X

= E[IY −1 (A) X] .

This can be viewed as putting a set of linear constraints on E[X | Y ] with one
constraint for each measurable A ⊆ Y. By treating E[X | Y ] as an unknown
σ(Y )-measurable random variable, we can attempt to solve this linear system. As
 2.6 Conditional Expectation 35

it turns out, this can always be done: the linear constraints and the measurability
restriction on E [X | Y ] completely determine E[X | Y ] except for a set of measure
zero. Notice that both conditions only depend on σ(Y ) ⊆ F. The abstract
definition of conditional expectation takes these properties as the definition and
replaces the role of Y with a sub-σ-algebra.
Definition 2.10 (Conditional expectation). Let (Ω, F, P) be a probability space
and X : Ω → R be random variable and H be a sub-σ-algebra of F. The
conditional expectation of X given H is denoted by E[X | H] and defined to be
any H-measurable random variable on Ω such that for all H ∈ H,
Z Z
E[X | H]dP = XdP . (2.9)
H H

Given a random variable Y , the conditional expectation of X given Y is

E [X | Y ] = E [X | σ(Y )].
Theorem 2.11. Given any probability space (Ω, F, P), a sub-σ-algebra H of F
and a P-integrable random variable X : Ω → R, there exists an H-measurable
function f : Ω → R that satisfies (2.9). Further, any two H-measurable functions
f1 , f2 : Ω → R that satisfy (2.9) are equal with probability one: P(f1 = f2 ) = 1.
When random variables X and Y agree with P-probability one, we say they
are P-almost surely equal, which is often abbreviated to ‘X = Y P-a.s.’, or
‘X = Y a.s.’ when the measure is clear from context. A related useful notion is
the concept of null sets: U ∈ F is a null set of P, or a P-null set if P(U ) = 0.
Thus, X = Y P-a.s. if and only if X = Y agree except on a P-null set.

The reader may find it odd that E[X | Y ] is a random variable on Ω rather
than the range of Y . Lemma 2.5 and the fact that E[X | σ(Y )] is σ(Y )-
measurable shows there exists a measurable function f : (R, B(R)) →
(R, B(R)) such that E[X | σ(Y )](ω) = (f ◦ Y )(ω) (see Fig. 2.4). In this sense
E[X | Y ](ω) only depends on Y (ω), and occasionally we write E[X | Y ](y).

(Ω, F)

Y E[X | Y ]

(R, B(R)) (R, B(R))

f

Figure 2.4 Factorisation of conditional expectation. When there is no confusion, we

occasionally write E[X | Y ](y) in place of f (y).

Returning to Example 2.9, we see that E [X | Y ] = E [X | σ(Y )] and σ(Y ) =

 2.6 Conditional Expectation 36

{{1, 2, 3}, {4, 5, 6}, ∅, Ω}. Denote this set-system by H for brevity. The condition
that E[X | H] is H-measurable can only be satisfied if E[X | H](ω) is constant on
{1, 2, 3} and {4, 5, 6}. Then (2.9) immediately implies that
(
2, if ω ∈ {1, 2, 3} ;
E [X | H] (ω) =
5, if ω ∈ {4, 5, 6} .
While the definition of conditional expectations given above is non-constructive
and E[X | H] is uniquely defined only up to events of P-measure zero, none of
this should be of a significant concern. First, we will rarely need closed-form
expressions for conditional expectations, but we rather need how they relate to
other expectations, conditional or not. This is also the reason why it should not
be concerning that they are only determined up to zero probability events: usually,
conditional expectations appear in other expectations or in statements that are
concerned with how probable some event is, making the difference between the
different ‘versions’ of conditional expectations disappear.
We close the section by summarising some additional important properties of
conditional expectations. These follow from the definition directly, and the reader
is invited to prove them in Exercise 2.20.
Theorem 2.12. Let (Ω, F, P) be a probability space, G, G1 , G2 ⊂ F be sub-σ-
algebras of F and X, Y integrable random variables on (Ω, F, P). The following
hold true:

1 If X ≥ 0, then E [X | G] ≥ 0 almost surely.

2 E [1 | G] = 1 almost surely.
3 E [X + Y | G] = E [X | G] + E [Y | G] almost surely.
4 E [XY | G] = Y E [X | G] almost surely if E [XY ] exists and Y is G-measurable.
5 If G1 ⊂ G2 , then E [X | G1 ] = E [E [X | G2 ] | G1 ] almost surely.
6 If σ(X) is independent of G2 given G1 , then E [X | σ(G1 ∪ G2 )] = E [X | G1 ]
almost surely.
7 If G = {∅, Ω} is the trivial σ-algebra, then E [X | G] = E [X] almost surely.
Properties 1 and 2 are self-explanatory. Property 3 generalises the linearity of
expectation. Property 4 shows that a measurable quantity can be pulled outside
of a conditional expectation and corresponds to the property that for constants
c, E [cX] = cE [X]. Property 5 is called the tower rule or the law of total
expectations. It says that the fineness of E[X | G2 ] is obliterated when taking the
conditional expectation with respect to G1 . Property 6 relates independence and
conditional expectations, and it says that conditioning on independent quantities
does not give further information on expectations. Here, the two event systems A
and B are said to be conditionally independent of each other given a σ-algebra
F if for all A ∈ A and B ∈ B, P (A ∩ B | F) = P (A | F) P (B | F) holds almost
surely. We also often say that A is conditionally independent of B given F, but
of course, this relation is symmetric. This property is often applied with random
variables: X is said to be conditionally independent of Y given Z, if σ(X) is
 2.7 Notes 37

conditionally independent of σ(Y ) given σ(Z). In this case, E [X | Y, Z] = E [X | Z]

holds almost surely. Property 7 states that conditioning on no information gives
the same expectation as not conditioning at all.

The above list of abstract properties will be used over and over again. We
encourage the reader to study the list carefully and convince yourself that
all items are intuitive. Playing around with discrete random variables can
be invaluable for this. Eventually it will all become second nature.

2.7 Notes

1 The Greek letter σ is often used by mathematicians in association with

countable infinities. Hence the term σ-algebra (and σ-field). Note that countable
additivity is often called σ-additivity. The requirement that additivity should
hold for systems of countably infinitely many sets is made so that probabilities
of (interesting) limiting events are guaranteed to exist.
2 Measure theory is concerned with measurable spaces, measures and with
their properties. An obvious distinction between probability theory and measure
theory is that in probability theory, one is (mostly) concerned with probability
measures. But the distinction does not stop here. In probability theory, the
emphasis is on the probability measures and their relations to each other. The
measurable spaces are there in the background, but are viewed as part of the
technical toolkit rather than the topic of main interest. Also, in probability
theory, independence is often at the center of attention, while independence is
not a property measure-theorists care much about.
3 In our toy example, instead of Ω = [6]7 , we could have chosen Ω = [6]8
(considering rolling eight dice instead of seven, one dice never used). There are
many other possibilities. We can consider coin flips instead of dice rolls (think
about how this could be done). To make this easy, we could use weighted coins
(for example, a coin that lands on heads with probability 1/6), but we don’t
actually need weighted coins (this may be a little tricky to see). The main
point is that there are many ways to emulate one randomisation device by
using another. The difference between these is the set Ω. What makes a choice
of Ω viable is if we can emulate the game mechanism on the top of Ω so that
in the end the probability of seeing any particular value remains the same. But
the main point is that the choice of Ω is far from unique. The same is true for
the way we calculate the value of the game! For example, the dice could be
reordered, if we stay with the first construction. This was noted already, but it
cannot be repeated frequently enough: the biggest conspiracy in all probability
theory is that we first make a big fuss about introducing Ω, and then it turns
out that the actual construction of Ω does not matter.
 2.7 Notes 38

4 All Riemann-integrable functions on a bounded domain are Lebesgue integrable.

Difficulties
R∞ only arise when taking improper integrals. A standard example
is 0 sin(x)dx
x , which is an improper
R Riemann integrable function, but is
not Lebesgue integrable because (0,∞) | sin(x)/x|dx = ∞. The situation is
analogous to the difference between conditionally and absolutely convergent
series, with the Lebesgue integral only defined in the latter case.
5 Can you think of a set that is not Borel measurable? Such sets exist, but do not
arise naturally in applications. The classic example is the Vitali set, which is
formed by taking the quotient group G = R/Q and then applying the axiom
of choice to choose a representative in [0, 1] from each equivalence class in G.
Non-measurable functions are so unusual that you do not have to worry much
about whether or not functions X : R → R are measurable. With only a few
exceptions, questions of measurability arising in this book are not related to
the fine details of the Borel σ-algebra. Much more frequently they are related
to filtrations and the notion of knowledge available having observed certain
random elements.
6 There is a lot to say about why the sum, or the product of random variables
are also random variables. Or why inf n Xn , supn Xn , lim inf n Xn , lim supn Xn
are measurable when Xn are. The key point is to show that the composition of
measurable maps is a measurable map and that continuous maps are measurable
and then apply these results (Exercise 2.1). For lim supn Xn , just rewrite it as
limm→∞ supn≥m Xn ; note that supn≥m Xn is decreasing (we take suprema of
smaller sets as m increases), hence lim supn Xn = inf m supn≥m Xn , reducing
the question to studying inf n Xn and supn Xn . Finally, for inf n Xn note that
it suffices if {ω : inf n Xn ≥ t} is measurable for any t real. Now, inf n Xn ≥ t
if and only if Xn ≥ t for all n. Hence, {ω : inf n Xn ≥ t} = ∩n {ω : Xn ≥ t},
which is a countable intersection of measurable sets, hence measurable (this
latter follows by the elementary identity (∩i Ai )c = ∪i Aci ).
7 The factorisation lemma, Lemma 2.5, is attributed to Joseph Doob and Eugene
Dynkin. The lemma sneakily uses the properties of real numbers (think about
why), which is another reason why what we said about σ-algebras containing
all information is not entirely true. The lemma has extensions to more general
random elements [Taraldsen, 2018, for example]. The key requirement in a
way is that the σ-algebra associated with the range space of Y should be rich
enough.
8 We did not talk about basic results like Lebesgue’s dominated/monotone
convergence theorems, Fatou’s lemma or Jensen’s inequality. We will definitely
use the last of these, which is explained in a dedicated chapter on convexity
(Chapter 26). The other results can be found in the texts we cite. They are
concerned with infinite sequences of random variables and conditions under
which their limits can be interchanged with Lebesgue integrals. In this book
we rarely encounter problems related to such sequences and hope you forgive
us on the few occasions they are necessary (the reason is simply because we
 2.7 Notes 39

mostly focus on finite time results or take expectations before taking limits
when dealing with asymptotics).
9 You might be surprised that we have not mentioned densities. For most of
us, our first exposure to probability on continuous spaces was by studying the
normal distribution and its density
1
p(x) = √ exp(−x2 /2) , (2.10)
2π
which can be integrated over intervals to obtain the probability that a Gaussian
random variable will take a value in that interval. The reader should notice
that p : R → R is Borel measurable and that the Gaussian measure associated
with this density is P on (R, B(R)) defined by
Z
P(A) = p dλ .
A

Here the integral is with respect to the Lebesgue measure λ on (R, B(R)). The
notion of a density can be generalised beyond this simple setup. Let P and Q
be measures (not necessarily probability measures) on arbitrary measurable
space (Ω, F). The Radon–Nikodym derivative of P with respect to Q is
an F-measurable random variable dQ dP
: Ω → [0, ∞) such that
Z
dP
P (A) = dQ for all A ∈ F . (2.11)
A dQ
R R
We can also write this in the form IA dP = IA dQ dP
dQ, A ∈ F, from which we
R R dP
may realise that for any X P -integrable random variable, XdP = X dQ dQ
must also hold. This is often called the change-of-measure formula. Another
word for the Radon–Nikodym derivative dQ dP
is the density of P with respect to
Q. It is not hard to find examples where the density does not exist. We say that
P is absolutely continuous with respect to Q if Q(A) = 0 =⇒ P (A) = 0
for all A ∈ F. When dQ dP
exists, it follows immediately that P is absolutely
continuous with respect to Q by Eq. (2.11). Except for some pathological cases,
it turns out that this is both necessary and sufficient for the existence of dP/dQ.
The measure Q is σ-finite if there exists a countable covering {Ai } of Ω with
F-measurable sets such that Q(Ai ) < ∞ for each i.
Theorem 2.13. Let P, Q be measures on a common measurable space (Ω, F)
and assume that Q is σ-finite. Then the density of P with respect to Q, dQ dP
,
exists if and only if P is absolutely continuous with respect to Q. Furthermore,
dQ is uniquely defined up to a Q-null set so that for any f1 , f2 satisfying (2.11),
dP

f1 = f2 holds Q-almost surely.

Densities work as expected. Suppose that Z is a standard Gaussian random
variable. We usually write its density as in Eq. (2.10), which we now know
is the Radon–Nikodym derivative of the Gaussian measure with respect to
the Lebesgue measure. The densities of ‘classical’ continuous distributions are
almost always defined with respect to the Lebesgue measure.
 2.7 Notes 40

10 In line with the literature, we will use P  Q to denote that P is absolutely

continuous with respect to Q. When P is absolutely continuous with respect
to Q, we also say that Q dominates P .
11 A useful result for Radon–Nikodym derivatives is the chain rule, which states
that if P  Q  S, then dQ dS = dS . The proof of this result follows from our
dP dQ dP
R R dQ
earlier observation that f dQ = f dS dS for any Q-integrable f . Indeed, the
chain rule is obtained from this by taking f = IA dQ dP
with A ∈ F and noting
R R
that this is indeed Q-integrable and IA dQ dQ = IA dQ. The chain rule is
dP

often used to reduce the calculation of densities to calculation with known

densities.
12 The Radon–Nikodym derivative unifies the notions of distribution (for discrete
spaces) and density (for continuous spaces). Let Ω be discrete (finite or
countable) and let ρ be the counting measure on (Ω, 2Ω ), which is defined
by ρ(A) = |A|. For any P on (Ω, F), it is easy to see that P  ρ and
dρ (i) = P ({i}), which is sometimes called the distribution function of P .
dP

13 The Radon–Nikodym derivative provides another way to define the conditional

expectation. Let X be an integrable random variable on (Ω, F, P) and H ⊂ F
be a sub-σ-algebra and P|RH be the restriction of P to (Ω, H). Define measure
µ on (Ω, H) by µ(A) = A XdP|H . It is easy to check that µ  P|H and
that E[X | H] = dP| dµ
H
satisfies Eq. (2.9). We note that the proof of the
Radon–Nikodym theorem is nontrivial and that the existence of conditional
expectations are more easily guaranteed via an ‘elementary’ but abstract
argument using functional analysis.
14 The Fubini–Tonelli theorem, which we will also refer to as Fubini’s theorem,
is a powerful result that allows one to exchange the order of integrations. This
result is needed for example for proving Proposition 2.8 (Exercise 2.19). To state
it, we need to introduce product measures. These work as expected: given two
probability spaces, (Ω1 , F1 , P1 ) and (Ω2 , F2 , P2 ), the product measure P of P1
and P2 is defined as any measure on (Ω1 ×Ω2 , F1 ⊗F2 ) that satisfies P(A1 , A2 ) =
P1 (A1 )P2 (A2 ) for all (A1 , A2 ) ∈ F1 × F2 (recall that F1 ⊗ F2 = σ(F1 × F2 ) is
the product σ-algebra of F1 and F2 ). Theorem 2.4 implies that this product
measure, which is often denoted by P1 × P2 (or P1 ⊗ P2 ) is uniquely defined.
(Think about what this product measure has to do with independence.) The
Fubini–Tonelli theorem (often just ‘Fubini’) states the following: let (Ω1 , F1 , P1 )
and (Ω2 , F2 , P2 ) be two probability spaces and consider a random variable
X on the product probability R space (Ω, F, P) R R= (Ω1 × Ω2 , F1 ⊗ F2 , P1 × P2 ).
If
R any
R of the three integrals |X(ω)| dP(ω), ( |X(ω1 , ω2 )| dP1 (ω1 )) dP2 (ω2 ),
( |X(ω1 , ω2 )| dP2 (ω2 )) dP1 (ω1 ) is finite, then
Z Z Z 
X(ω) dP(ω) = X(ω1 , ω2 ) dP1 (ω1 ) dP2 (ω2 )
Z Z 
= X(ω1 , ω2 ) dP2 (ω2 ) dP1 (ω1 ) .
 2.8 Bibliographic Remarks 41

15 For topological space X, the support of a measure µ on (X, B(X)) is

Supp(µ) = {x ∈ X : µ(U ) > 0 for all neighborhoods U of x} .
When X is discrete, this reduces to Supp(µ) = {x : µ({x}) > 0}.
16 Let X be a topological space. The weak* topology on the space of probability
R
measures P(X) on (X, B(X)) is the coarsest topology such that µ 7→ f dµ
is continuous for all bounded continuous functions f : X → R. In particular,
a sequence of probability
R measuresR (µn )n=1 converges to µ in this topology
∞

if and only if limn→∞ f dµn = f dµ for all bounded continuous functions

f : X → R.
Theorem 2.14. When X is compact and Hausdorff and P(X) is the space of
regular probability measures on (X, B(X)) with the weak* topology, then P(X)
is compact.
17 Mathematical terminology can be a bit confusing sometimes. Since E maps
(certain) functions to real values, it is also called the expectation operator.
‘Operator’ is just a fancy name for a function. In operator theory, the study
of operators, the focus is on operators whose domain is infinite dimensional,
hence the distinct name. However, most results of operator theory do not
hinge upon this property. If the image space is the set of reals, we talk about
functionals. The properties of functionals are studied in yet another subfield of
mathematics, functional analysis. The expectation operator is a functional
that maps the set of P-integrable functions (often denoted by L1 (Ω, P) or
L1 (P)) to reals. Its most important property is linearity, which was stated as
a requirement for integrals that define the expectation operator (Eq. (2.5)).
In line with the previous comment, when we use E, more often than not, the
probability space remains hidden. As such, the symbol E is further abused.

2.8 Bibliographic Remarks

Much of this chapter draws inspiration from David Pollard’s A user’s guide to
measure theoretic probability [Pollard, 2002]. We like this book because the author
takes a rigourous approach, but still explains the ‘why’ and ‘how’ with great
care. The book gets quite advanced quite fast, concentrating on the big picture
rather than getting lost in the details. Other useful references include the book by
Billingsley [2008], which has many good exercises and is quite comprehensive in
terms of its coverage of the ‘basics’. These books are both quite detailed. For an
outstanding shorter introduction to measure-theoretic probability, see the book
by Williams [1991], which has an enthusiastic style and a pleasant bias towards
martingales. We also like the book by Kallenberg [2002], which is recommended
for the mathematically inclined readers who already have a good understanding of
the basics. The author has put a major effort into organising the material so that
redundancy is minimised and generality is maximised. This reorganisation resulted
 2.9 Exercises 42

in quite a few original proofs, and the book is comprehensive. The factorisation
lemma (Lemma 2.5) is stated in the book by Kallenberg [2002] (Lemma 1.13
there). Kallenberg calls this lemma the ‘functional representation’ lemma and
attributes it to Joseph Doob. Theorem 2.4 is a corollary of Carathéodory’s
extension theorem, which says that probability measures defined on semi-rings of
sets have a unique extension to the generated σ-algebra. The remaining results can
be found in either of the three books mentioned above. Theorem 2.14 appears as
theorem 8.9.3 in the two-volume book by Bogachev [2007]. Finally, for something
older and less technical, we recommend the philosophical essays on probability
by Pierre Laplace, which was recently reprinted [Laplace, 2012].

2.9 Exercises

2.1 (Composing random elements) Show that if f is F/G-measurable and g

is G/H-measurable for sigma algebras F, G and H over appropriate spaces, then
their composition, g ◦ f (defined the usual way: (g ◦ f )(ω) = g(f (ω)), ω ∈ Ω), is
F/H-measurable.

2.2 Let X1 , . . . , Xn be random variables on (Ω, F). Prove that (X1 , . . . , Xn ) is

a random vector.

2.3 (Random variable induced σ-algebra) Let U be an arbitrary set and

(V, Σ) a measurable space and X : U → V an arbitrary function. Show that
ΣX = {X −1 (A) : A ∈ Σ} is a σ-algebra over U.

2.4 Let (Ω, F) be a measurable space and A ⊆ Ω and F|A = {A ∩ B : B ∈ F}.

(a) Show that (A, F|A ) is a measurable space.
(b) Show that if A ∈ F, then F|A = {B : B ∈ F, B ⊆ A}.

2.5 Let G ⊆ 2Ω be a non-empty collection of sets and define σ(G) as the smallest
σ-algebra that contains G. By ‘smallest’ we mean that F ∈ 2Ω is smaller than
F 0 ∈ 2Ω if F ⊂ F 0 .
(a) Show that σ(G) exists and contains exactly those sets A that are in every
σ-algebra that contains G.
(b) Suppose (Ω0 , F) is a measurable space and X : Ω0 → Ω be F/G-measurable.
Show that X is also F/σ(G)-measurable. (We often use this result to simplify
the job of checking whether a random variable satisfies some measurability
property).
(c) Prove that if A ∈ F where F is a σ-algebra, then I {A} is F-measurable.

2.6 (Knowledge and σ-algebras: a pathological example) In the context

of Lemma 2.5, show an example where Y = X and yet Y is not σ(X) measurable.

Hint As suggested after the lemma, this can be arranged by choosing

 2.9 Exercises 43

Ω = Y = X = R, X(ω) = Y (ω) = ω, F = H = B(R) and G = {∅, R} to

be the trivial σ-algebra.
2.7 Let (Ω, F, P) be a probability space, B ∈ F be such that P (B) > 0. Prove
that A 7→ P (A | B) is a probability measure over (Ω, F).

2.8 (Bayes law) Verify (2.2).

2.9 Consider the standard probability space (Ω, F, P) generated by two standard,
unbiased, six-sided dice that are thrown independently of each other. Thus,
Ω = {1, . . . , 6}2 , F = 2Ω and P(A) = |A|/62 for any A ∈ F so that Xi (ω) = ωi
represents the outcome of throwing dice i ∈ {1, 2}.

(a) Show that the events ‘X1 < 2’ and ‘X2 is even’ are independent of each
other.
(b) More generally, show that for any two events, A ∈ σ(X1 ) and B ∈ σ(X2 ),
are independent of each other.

2.10 (Serendipitous independence) The point of this exercise is to understand

independence more deeply. Solve the following problems:

(a) Let (Ω, F, P) be a probability space. Show that ∅ and Ω (which are events)
are independent of any other event. What is the intuitive meaning of this?
(b) Continuing the previous part, show that any event A ∈ F with P (A) ∈ {0, 1}
is independent of any other event.
(c) What can we conclude about an event A ∈ F that is independent of its
complement, Ac = Ω \ A? Does your conclusion make intuitive sense?
(d) What can we conclude about an event A ∈ F that is independent of itself?
Does your conclusion make intuitive sense?
(e) Consider the probability space generated by two independent flips of unbiased
coins with the smallest possible σ-algebra. Enumerate all pairs of events
A, B such that A and B are independent of each other.
(f) Consider the probability space generated by the independent rolls of two
unbiased three-sided dice. Call the possible outcomes of the individual dice
rolls 1, 2 and 3. Let Xi be the random variable that corresponds to the
outcome of the ith dice roll (i ∈ {1, 2}). Show that the events {X1 ≤ 2} and
{X1 = X2 } are independent of each other.
(g) The probability space of the previous example is an example when the
probability measure is uniform on a finite outcome space (which happens to
have a product structure). Now consider any n-element, finite outcome space
with the uniform measure. Show that A and B are independent of each other
if and only if the cardinalities |A|, |B|, |A ∩ B| satisfy n|A ∩ B| = |A| · |B|.
(h) Continuing with the previous problem, show that if n is prime, then no
non-trivial events are independent (an event A is trivial if P (A) ∈ {0, 1}).
(i) Construct an example showing that pairwise independence does not imply
mutual independence.
 2.9 Exercises 44

(j) Is it true or not that A, B, C are mutually independent if and only if

P (A ∩ B ∩ C) = P (A) P (B) P (C)? Prove your claim.

2.11 (Independence and random elements) Solve the following problems:

(a) Let X be a constant random element (that is, X(ω) = x for any ω ∈ Ω over
the outcome space over which X is defined). Show that X is independent of
any other random variable.
(b) Show that the above continues to hold if X is almost surely constant (that
is, P (X = x) = 1 for an appropriate value x).
(c) Show that two events are independent if and only if their indicator random
variables are independent (that is, A, B are independent if and only if
X(ω) = I {ω ∈ A} and Y (ω) = I {ω ∈ B} are independent of each other).
(d) Generalise the result of the previous item to pairwise and mutual
independence for collections of events and their indicator random variables.

2.12 Our goal in this exercise is to show that X is integrable if and only if |X|
is integrable. This is broken down into multiple steps. The first issue is to deal
with the measurability of |X|. While a direct calculation can also show this, it
may be worthwhile to follow a more general path:

(a) Any f : R → R continuous function is Borel measurable.

(b) Conclude that for any random variable X, |X| is also a random variable.
(c) Prove that for any random variable X, X is integrable if and only if |X|
is integrable. (The statement makes sense since |X| is a random variable
whenever X is).

Hint For (b) recall Exercise 2.1. For (c) examine the relationship between
|X| and (X)+ and (X)− .
2.13 (Infinite-valued integrals) Can we consistently extend the definition of
integrals so that for non-negative random variables, the integral is always defined
(it may be infinite)? Defend your view by either constructing an example (if you
are arguing against) or by proving that your definition is consistent with the
requirements we have for integrals.

2.14 Prove Proposition 2.6.

Hint You may find it useful to use Lebesgue’s dominated/monotone convergence

theorems.
2.15 Prove that if c ∈ R is a constant, then E [cX] = cE [X] (as long as X is
integrable).

2.16 Prove Proposition 2.7.

Hint Follow the ‘inductive’ definition of Lebesgue integrals, starting with

 2.9 Exercises 45

simple functions, then non-negative functions and finally arbitrary independent

random variables.
2.17 Suppose that G1 ⊂ G2 and prove that E[X | G1 ] = E[E[X | G1 ] | G2 ] almost
surely.

2.18 Demonstrate using an example that in general, for dependent random

variables, E [XY ] = E [X] E [Y ] does not hold.

2.19 Prove Proposition 2.8.

R
Hint Argue that X(ω) = [0,∞) I {[0, X(ω)]} (x) dx and exchange the integrals.
Use the Fubini–Tonelli theorem to justify the exchange of integrals.
2.20 Prove Theorem 2.12.
3 Stochastic Processes and Markov
Chains ( )

The measure-theoretic probability in the previous chapter covers almost all the
definitions required. Occasionally, however, infinite sequences of random variables
arise, and for these a little more machinery is needed. We expect most readers
will skip this chapter on the first reading, perhaps referring to it when necessary.
Before one can argue about the properties of infinite sequences of random
variables, it must be demonstrated that such sequences exist under certain
constraints on their joint distributions. For example, does there exist an infinite
sequence of random variables such that any finite subset of the random variables
are independent and distributed like a standard Gaussian? The first theorem
provides conditions under which questions like this can be answered positively.
This allows us to write, for example, ‘let (Xn )∞ n=1 be an infinite sequence of
independent standard Gaussian random variables’ and be comfortable knowing
there exists a probability space on which these random variables can be defined.
To state the theorem, we need the concept of Borel spaces.
Two measurable spaces (X , F) and (Y, G) are said to be isomorphic if there
exists a bijective function f : X → Y such that f is F/G-measurable and f −1 is
G/F-measurable. A Borel space is a measurable space (X , F) that is isomorphic
to (A, B(A)) with A ∈ B(R) a Borel measurable subset of the reals. This is not
a very strong assumption. For example, (Rn , B(Rn )) is a Borel space, along with
all of its measurable subsets.

Theorem 3.1. Let µ be a probability measure on a Borel space S and λ be the

Lebesgue measure on ([0, 1], B([0, 1]). Then there exists a sequence of independent
random elements X1 , X2 , . . . on ([0, 1], B([0, 1]), λ) such that the law λXt = µ for
all t.

We give a sketch of the proof because, although it is not really relevant for
the material in this book, it illustrates the general picture and dispels some of
the mystic about what is really going on. Exercise 3.1 asks you to provide the
missing steps from the proof.

Proof sketch of Theorem 3.1 For simplicity we consider only the case that
S = ([0, 1], B([0, 1])) and µ is the Lebesgue measure. For any x ∈ [0, 1], let
F1 (x), F2 (x), . . . be the binary expansion of x, which is the unique binary-valued
 3.1 Stochastic Processes 47

infinite sequence such that

∞
X
x= Ft (x)2−t .
t=1

We can view F1 , F2 , . . . as (binary-valued) random variables over the probability

space ([0, 1], B([0, 1]), λ). Viewed as such, a direct calculation shows that
F1 , F2 , . . . are independent. From this we can create an infinite sequence of
uniform random variables by reversing the process. To do this, we rearrange the
(Ft )∞
t=1 sequence into a grid. For example:

F1 , F2 , F4 , F7 , · · ·
F3 , F5 , F8 , · · ·
F6 , F9 , · · ·
F10 , · · ·
..
.
Letting Xm,t be the tth entry in the mth row of this grid, we define Xm =
P∞ −t
t=1 2 Xm,t , and again one can easily check that with this choice the sequence
X1 , X2 , . . . is independent and λXt = µ is uniform for each t.

3.1 Stochastic Processes

Let T be an arbitrary set. A stochastic process on probability space (Ω, F, P)

is a collection of random variables {Xt : t ∈ T }. In this book T will always
be countable, and so in the following we restrict ourselves to T = N. The first
theorem is not the most general, but suffices for our purposes and is more easily
stated than more generic alternatives.
Theorem 3.2. For each n ∈ N+ , let (Ωn , Fn ) be a Borel space and µn be a
measure on (Ω1 × · · · × Ωn , F1 ⊗ · · · ⊗ Fn ) and assume that µn and µn+1 are
related through
µn+1 (A × Ωn+1 ) = µn (A) for all A ∈ Ω1 ⊗ · · · ⊗ Ωn . (3.1)
Then there exists a probability space (Ω, F, P) and random elements X1 , X2 , . . .
with Xt : Ω → Ωt such that PX1 ,...,Xn = µn for all n.

Sequences of measures (µn )n satisfying Eq. (3.1) are called projective.

Theorem 3.1 follows immediately from Theorem 3.2. By assumption a random

variable takes values in (R, B(R)), which is Borel. Then let µn = ⊗nt=1 µ be
the n-fold product measure of µ with itself. That this sequence of measures is
projective is clear, and the theorem does the rest.
 3.2 Markov Chains 48

3.2 Markov Chains

A Markov chain is an infinite sequence of random elements (Xt )∞ t=1 where the
conditional distribution of Xt+1 given X1 , . . . , Xt is the same as the conditional
distribution of Xt+1 given Xt . The sequence has the property that given the last
element, the history is irrelevant to ‘predict’ the future. Such random sequences
appear throughout probability theory and have many applications besides. The
theory is too rich to explain in detail, so we give the basics and point towards the
literature for more details at the end. The focus here is mostly on the definition
and existence of Markov chains.
Let (X , F) and (Y, G) be measurable spaces. A probability kernel or Markov
kernel from (X , F) to (Y, G) is a function K : X × G → [0, 1] such that
(a) K(x, ·) is a measure for all x ∈ X ; and
(b) K(·, A) is F-measurable for all A ∈ G.
The idea here is that K describes a stochastic transition. Having arrived at x, a
process’s next state is sampled Y ∼ K(x, ·). Occasionally, we will use the notation
Kx (A) or K(A | x) rather than K(x, A).
If K1 is a (X , F) → (Y, G) probability kernel and K2 is a (Y, G) → (Z, H)
probability kernel, then the product kernel K1 ⊗ K2 is the probability kernel
from (X , F) → (Y × Z, G ⊗ H) defined by
Z Z
(K1 ⊗ K2 )(x, A) = IA ((y, z))K2 (y, dz)K1 (x, dy) .
Y Z

When P is a measure on (X , F) and K is a kernel from X to Y, then P ⊗ K is a

measure on (X × Y, F ⊗ G) defined by
Z Z
(P ⊗ K)(A) = IA ((x, y))K(x, dy)dP (x) .
X Y

These operations can be composed. When P is a probability measure on X and

K1 a kernel from X to Y and K2 a kernel from X × Y to Z, then P ⊗ K1 ⊗ K2
is a probability measure on X × Y × Z. The following provides a counterpart of
Theorem 3.2.
Theorem 3.3 (Ionescu–Tulcea). Let (Ωn , Fn )∞
n=1 be a sequence of measurable
spaces and K1 be a probability measure on (Ω1 , F1 ). For n ≥ 2, let Kn
Qn−1
be a probability kernel from t=1 Ωt to Ωn . Then there exists a probability
space (Ω, F, P) and random elements (Xt )∞
t=1 with Xt : Ω → Ωt such that
Nn
PX1 ,...,Xn = t=1 Kt for all n ∈ N+ .
A homogeneous Markov chain is a sequence of random elements (Xt )∞
t=1
taking values in state space S = (X , F) and with
P (Xt+1 ∈ · | X1 , . . . , Xt ) = P (Xt+1 ∈ · | Xt ) = µ(Xt , ·) almost surely ,
where µ is a probability kernel from (X , F) to (X , F) and we assume that
P (X1 ∈ ·) = µ0 (·) for some measure µ0 on (X , F).
 3.3 Martingales and Stopping Times 49

The word ‘homogeneous’ refers to the fact that the probability kernel does
not change with time. Accordingly, sometimes one writes ‘time homogeneous’
instead of homogeneous. The reader can no doubt see how to define a Markov
chain where µ depends on t, though doing so is purely cosmetic since the
state space can always be augmented to include a time component.

Note that if µ(x | ·) = µ0 (·) for all x ∈ X , then Theorem 3.3 is yet another
way to prove the existence of an infinite sequence of independent and identically
distributed random variables. The basic questions in Markov chains resolve around
understanding the evolution of Xt in terms of the probability kernel. For example,
assuming that Ωt = Ω1 for all t ∈ N+ , does the law of Xt converge to some fixed
distribution as t → ∞, and if so, how fast is this convergence? For now we make
do with the definitions, but in the special case that X is finite, we will discuss
some of these topics much later in Chapters 37 and 38.

3.3 Martingales and Stopping Times

Let X1 , X2 , . . . be a sequence of random variables on (Ω, F, P) and F = (Ft )nt=1

a filtration of F and where we allow n = ∞. Recall that the sequence (Xt )nt=1 is
F-adapted if Xt is Ft -measurable for all 1 ≤ t ≤ n.

Definition 3.4. A F-adapted sequence of random variables (Xt )t∈N+ is a F-

adapted martingale if

(a) E[Xt | Ft−1 ] = Xt−1 almost surely for all t ∈ {2, 3, . . .}; and
(b) Xt is integrable.

If the equality is replaced with a less-than (greater-than), then we call (Xt )t a

supermartingale (respectively, a submartingale).

The time index t need not run over N+ . Very often t starts at zero instead.

Example 3.5. A gambler repeatedly throws a coin, winning a dollar for each
heads and losing a dollar for each tails. Their total winnings over time is a
martingale. To model this situation, let Y1 , Y2 , . . . be a sequence of independent
Rademacher distributions, which means that P (Yt = 1) = P (Yt = −1) = 1/2.
Pt
The winnings after t rounds is St = s=1 Ys , which is a martingale adapted to
the filtration (Ft )∞
t=1 given by Ft = σ(Y1 , . . . , Yt ). The definition of super/sub-
martingales (the direction of inequality) can be remembered by remembering
that the definition favors the casino, not the gambler.

Can a gambler increase its expected winning by stopping cleverly? Precisely,

 3.3 Martingales and Stopping Times 50

the gambler at the end of round t can decide to stop (δt = 1) or continue (δt = 0)
based on the information available to them. Denoting by τ = min{t : δt = 1}
the time when the gambler stops, the question is whether by a clever choice of
(δt )t∈N , E [Sτ ] can be made positive. Here, (δt )t∈N , a sequence of binary, F-adapted
random variables, is called a stopping rule, while τ is a stopping time with
respect F.

Note that the stopping rule is not allowed to inject additional randomness
beyond what is already there in F.

Definition 3.6. Let F = (Ft )t∈N be a filtration. A random variable τ with values
in N ∪ {∞} is a stopping time with respect to F if I {τ ≤ t} is Ft -measurable
for all t ∈ N. The σ-algebra at stopping time τ is
Fτ = {A ∈ F∞ : A ∩ {τ ≤ t} ∈ Ft for all t} .

The filtration is usually indicated by writing ‘τ is a F-stopping time’. When

the underlying filtration is obvious from context, it may be omitted. This is
also true for martingales.

Using the interpretation of σ-algebras encoding information, if (Ft )t is thought

of as the knowledge available at time t, Fτ is the information available at the
random time τ . Exercise 3.7 asks you to explore properties of stopped σ-algebras;
amongst other things, it asks you to show that Fτ is in fact a σ-algebra.
Example 3.7. In the gambler example, the first time when the gambler’s
winnings hits 100 is a stopping time: τ = min{t : St = 100}. On the other
hand, τ = min{t : St+1 = −1} is not a stopping time because I {τ = t} is not
Ft -measurable.
Whether or not E [Sτ ] can be made positive by a clever choice of a stopping
time τ is answered in the negative by a fundamental theorem of Doob:
Theorem 3.8 (Doob’s optional stopping). Let F = (Ft )t∈N be a filtration and
(Xt )t∈N be an F-adapted martingale and τ an F-stopping time such that at least
one of the following holds:
(a) There exists an n ∈ N such that P (τ > n) = 0.
(b) E[τ ] < ∞, and there exists a constant c ∈ R such that for all t ∈ N,
E[|Xt+1 − Xt | | Ft ] ≤ c almost surely on the event that τ > t.
(c) There exists a constant c such that |Xt∧τ | ≤ c almost surely for all t ∈ N.
Then Xτ is almost surely well defined, and E[Xτ ] = E[X0 ]. Furthermore, when
(Xt ) is a super/sub-martingale rather than a martingale, then equality is replaced
with less/greater-than, respectively.
 3.3 Martingales and Stopping Times 51

The theorem implies that if Sτ is almost-surely well defined then either

E [τ ] = ∞ or E [Sτ ] = 0. Gamblers trying to outsmart the casino would need to
live a very long life! One application of Doob’s optional stopping theorem is a
useful and a priori surprising generalisation of Markov’s inequality to non-negative
supermartingales.

Theorem 3.9 (Maximal inequality). Let (Xt )∞ t=0 be a supermartingale with

Xt ≥ 0 almost surely for all t. Then for any ε > 0,
 
E[X0 ]
P sup Xt ≥ ε ≤ .
t∈N ε

Proof Let An be the event that supt≤n Xt ≥ ε and τ = (n + 1) ∧ min{t ≤

n : Xt ≥ ε}, where the minimum of an empty set is assumed to be infinite so
that τ = n + 1 if Xt < ε for all 0 ≤ t ≤ n. Clearly τ is a stopping time and
P (τ ≤ n + 1) = 1. Then by Theorem 3.8 and elementary calculation,

E[X0 ] ≥ E[Xτ ] ≥ E[Xτ I {τ ≤ n}] ≥ E[εI {τ ≤ n}] = εP (τ ≤ n) = εP (An ) ,

where the second inequality uses the definition of the stopping time and the non-
negativity of the supermartingale. Rearranging shows that P (An ) ≤ E[X0 ]/ε for
all n ∈ N. Since A1 ⊆ A2 ⊆ . . ., it follows that P (supt∈N Xt ≥ ε) = P (∪n∈N An ) ≤
E[X0 ]/ε.

Markov’s inequality (which we will cover in the next chapter) combined with
the definition of a supermartingale shows that
E[X0 ]
P (Xn ≥ ε) ≤ . (3.2)
ε
In fact, in the above we have effectively applied Markov’s inequality to the
random variable Xτ (the need for the proof arises when the conditions of
Doob’s optional sampling theorem are not met). The maximal inequality is
a strict improvement over Eq. (3.2) by replacing Xn with supt∈N Xt at no
cost whatsoever.

A similar theorem holds for submartingales. You will provide a proof in

Exercise 3.8.

Theorem 3.10. Let (Xt )nt=0 be a submartingale with Xt ≥ 0 almost surely for
all t. Then for any ε > 0,
 
E[Xn ]
P max Xt ≥ ε ≤ .
t∈{0,1,...,n} ε
 3.4 Notes 52

3.4 Notes

1 Some authors include in the definition of a stopping time τ that P (τ < ∞) = 1

and call random times without this property Markov times. We do not adopt
this convention and allow stopping times to be infinite with non-zero probability.
Stopping times are also called optional times.
2 There are several notations for probability kernels depending on the application.
The following are commonly seen and equivalent: K(x, A) = K(A | x) = Kx (A).
For example, in statistics a parametric family is often given by {Pθ : θ ∈ Θ},
where Θ is the parameter space and Pθ is a measure on some measurable space
(Ω, F). This notation is often more convenient than writing P(θ, ·). In Bayesian
statistics the posterior is a probability kernel from the observation space to
the parameter space, and this is often written as P(· | x).
3 There is some disagreement about whether or not a Markov chain on an
uncountable state space should instead be called a Markov process. In this
book we use Markov chain for arbitrary state spaces and discrete time. When
time is continuous (which it never is in this book), there is general agreement
that ‘process’ is more appropriate. For more history on this debate, see [Meyn
and Tweedie, 2012, preface].
4 A topological space X is Polish if it is separable and there exists a metric
d that is compatible with the topology that makes (X , d) a complete metric
space. All Polish spaces are Borel spaces. We follow Kallenberg [2002], but
many authors use standard Borel space rather than Borel space, and define
it as the σ-algebra generated by the open sets of a Polish space.
5 In Theorem 3.2 it was assumed that each µn was defined on a Borel space.
No such assumption was required for Theorem 3.3, however. One can derive
Theorem 3.2 from Theorem 3.3 by using the existence of regular conditional
probability measures when conditioning on random elements taking values
in a Borel space (see the next note). Topological assumptions often creep
into foundational questions relating to the existence of probability measures
satisfying certain conditions, and pathological examples show these assumptions
cannot be removed completely. Luckily, in this book we have no reason to
consider random elements that do not take values in a Borel space.
6 The fact that conditional expectation is only unique almost surely can be
problematic when you want a conditional distribution. Given random elements
X and Y on the same probability space, it seems reasonable to hope that
P (X ∈ · | Y ) is a probability kernel from the space of Y to that of X. A
version of the conditional distributions that satisfies this is called a regular
version. In general, there is no guarantee that such a regular version exist. The
basic properties of conditional expectation only guarantee that for any fixed
measurable A, P (X ∈ A | Y ) is unique up to a set of measure zero. The set of
measure zero can depend on A, which causes problems when there are ‘too
many’ measurable sets in the space of X. Assuming X lives in a Borel space,
the following theorem guarantees the existence of a conditional distribution.
 3.5 Bibliographic Remarks 53

Theorem 3.11 (Regular conditional distributions). Let X and Y be random

elements on the same probability space (Ω, F, P) taking values in measurable
spaces X and Y, and assume that X is Borel. Then there exists a probability
kernel K from Y to X such that K(· | Y ) = P (X ∈ · | Y ) P-almost surely.
Furthermore, K is unique in the sense that for any kernels K1 and K2 satisfying
this condition, it holds that K1 (· | y) = K2 (· | y) for all y in some set of PY -
measure one.

The theorem implies the useful relation that PX,Y = PY ⊗K (cf. Exercise 3.9)
where recall that for a random variable Z, PZ denotes its pushforward under
P. To make the origin K clear, we often write PX|Y instead of K. With this,
the above equality becomes PX,Y = PY ⊗ PX|Y , which can be viewed as the
converse of the Ionescu–Tulcea theorem (Theorem 3.3). Sometimes this is called
the chain rule of probabilities measures.
You can also condition on a σ-algebra G ⊂ F, in which case K is a probability
kernel from (Ω, G) to X . The condition that X be Borel is sufficient, but not
necessary. Some conditions are required, however. An example where no regular
version exists can be found in [Halmos, 1976, p210]. Regular versions play
a role in the following useful theorem for decomposing random variables on
product spaces.

Theorem 3.12 (Disintegration). Let X and Y be random elements on the

same probability space taking values in measurable spaces X and Y. Let f be
a random variable on X × Y so that E[|f (X, Y )|] < ∞. Suppose that K is a
regular version of P (X ∈ · | G) and Y is G-measurable. Then,
Z
E[f (X, Y ) | G] = f (x, Y )K(dx | ·) almost surely .
X

In many applications
R G = σ(Y ), in which case the theorem says that
E[f (X, Y ) | Y ] = X f (x, Y )K(dx | Y ) almost surely. Proofs of both theorems
appear in chapter 6 of Kallenberg [2002].

3.5 Bibliographic Remarks

There are many places to find the construction of a stochastic process. Like before,
we recommend Kallenberg [2002] for readers who want to refresh their memory
and Billingsley [2008] for a more detailed account. For Markov chains the recent
book by Levin and Peres [2017] provides a wonderful introduction. After reading
that, you might like the tome by Meyn and Tweedie [2012]. Theorem 3.1 can be
found as theorem 3.19 in the book by Kallenberg [2002], where the reader can also
find its proof. Theorem 3.2 is credited to Percy John Daniell by Kallenberg [2002]
(see Aldrich 2007). More general versions of this theorem exist. Readers looking
for these should look up Kolmogorov’s extension theorem [Kallenberg, 2002,
theorem 6.16]. The theorem of Ionescu–Tulcea (Theorem 3.3) is attributed to him
 3.6 Exercises 54

[Ionescu Tulcea, 1949–50] with a modern proof in the book by Kallenberg [2002,
theorem 6.17]. There are lots of minor variants of the optional stopping theorem,
most of which can be found in any probability book featuring martingales. The
most historically notable source is by the man himself [Doob, 1953]. A more
modern book that also gives the maximal inequalities is the book on optimal
stopping by Peskir and Shiryaev [2006].

3.6 Exercises

3.1 Fill in the details of Theorem 3.1:

(a) Prove that Ft ∈ {0, 1} is a Bernoulli random variable for all t ≥ 1.
(b) In what follows, equip S with P = λ, the uniform probability measure. Show
that for any t ≥ 1, Ft is uniformly distributed: P (Ft = 0) = P (Ft = 1) = 1/2.
(c) Show that (Ft )∞t=1 are independent.
(d) Show that (Xm,t )∞ t=1 is an independent sequence of Bernoulli random
variables that are uniformly distributed.
P∞
(e) Show that Xt = t=1 Xm,t 2−t is uniformly distributed on [0, 1].
(f) Show that (Xt )∞ t=1 are independent.

3.2 (Martingales and optional stopping) Let (Xt )∞ t=1 be an infinite

Pt
sequence of independent Rademacher random variables and St = s=1 Xs 2s−1 .
(a) Show that (St )∞
t=0 is a martingale.
(b) Let τ = min{t : St = 1} and show that P (τ < ∞) = 1.
(c) What is E[Sτ ]?
(d) Explain why this does not contradict Doob’s optional stopping theorem.

3.3 (Martingales and optional stopping (ii)) Give an example of a

martingale (Sn )∞
n=0 and stopping time τ such that

lim E[Sτ ∧n ] 6= E[Sτ ] .

n→∞

3.4 (Maximal inequality fails without non-negativity) Show that

Theorem 3.9 does not hold in general for supermartingales if the assumption that
it be non-negative is dropped.

3.5 Let (Ω, F) and (X , G) be measurable spaces, X : X → R be a random variable

and K : Ω × G → [0, 1] aR probability kernel from (Ω, F) to (X , G). Define function
U : Ω → R by U (ω) = X X(x)K(ω, dx) and assume that U (ω) exists for all ω.
Prove that U is F-measurable.

3.6 (Limits of increasing stopping times are stopping times) Let (τn )∞ n=1
be an almost surely increasing sequence of F-stopping times on probability space
(Ω, F, P) with filtration F = (Fn )∞
n=1 , which means that τn (ω) ≤ τn+1 (ω) for all
n ≥ 1 almost surely. Prove that τ (ω) = limn→∞ τn (ω) is a F-stopping time.
 3.6 Exercises 55

3.7 (Properties of stopping times) Let F = (Ft )t∈N be a filtration, and

τ, τ1 , τ2 be stopping times with respect to F. Show the following:
(a) Fτ is a σ-algebra.
(b) If τ = k for some k ≥ 1, then Fτ = Fk .
(c) If τ1 ≤ τ2 , then Fτ1 ⊂ Fτ2 .
(d) τ is Fτ -measurable.
(e) If (Xt ) is F-adapted, then Xτ is Fτ -measurable.
(f) Fτ is the smallest σ-algebra such that all F-adapted sequences (Xt ) satisfy
Xτ is Fτ -measurable.

3.8 Prove Theorem 3.10.

3.9 (Decomposing joint distributions) Let X and Y be random elements

on the same probability space (Ω, F, P) taking values in measurable spaces X
and Y respectively and assume that X is Borel. Show that P(X,Y ) = PY ⊗ PX|Y
where PX|Y denotes a regular conditional distribution of X and Y (the existence
of which is guaranteed by Theorem 3.11).
4 Stochastic Bandits

The goal of this chapter is to formally introduce stochastic bandits. The model
introduced here provides the foundation for the remaining chapters that treat
stochastic bandits. While the topic seems a bit mundane, it is important to be
clear about the assumptions and definitions. The chapter also introduces and
motivates the learning objectives, and especially the regret. Besides the definitions,
the main result in this chapter is the regret decomposition, which is presented in
Section 4.5.

4.1 Core Assumptions

A stochastic bandit is a collection of distributions ν = (Pa : a ∈ A), where A is

the set of available actions. The learner and the environment interact sequentially
over n rounds. In each round t ∈ {1, . . . , n}, the learner chooses an action At ∈ A,
which is fed to the environment. The environment then samples a reward Xt ∈ R
from distribution PAt and reveals Xt to the learner. The interaction between
the learner (or policy) and environment induces a probability measure on the
sequence of outcomes A1 , X1 , A2 , X2 , . . . , An , Xn . Usually the horizon n is finite,
but sometimes we allow the interaction to continue indefinitely (n = ∞). The
sequence of outcomes should satisfy the following assumptions:

(a) The conditional distribution of reward Xt given A1 , X1 , . . . , At−1 , Xt−1 , At

is PAt , which captures the intuition that the environment samples Xt from
PAt in round t.
(b) The conditional law of action At given A1 , X1 , . . . , At−1 , Xt−1 is
πt (· | A1 , X1 , . . . , At−1 , Xt−1 ), where π1 , π2 , . . . is a sequence of probability
kernels that characterise the learner. The most important element of this
assumption is the intuitive fact that the learner cannot use the future
observations in current decisions.

A mathematician might ask whether there even exists a probability space carrying
these random elements such that (a) and (b) hold. Specific constructions showing
this in the affirmative are given in Section 4.6. These constructions are also
valuable because they teach us important lessons about equivalent models. For
now, however, we move on.
 4.2 The Learning Objective 57

4.2 The Learning Objective

Pn
The learner’s goal is to maximise the total reward Sn = t=1 Xt , which is a
random quantity that depends on the actions of the learner and the rewards
sampled by the environment. This is not an optimisation problem for three
reasons:
1 What is the value of n for which we are maximising? Occasionally prior
knowledge of the horizon is reasonable, but very often the learner does not
know ahead of time how many rounds are to be played.
2 The cumulative reward is a random quantity. Even if the reward distributions
were known, then we require a measure of utility on distributions of Sn .
3 The learner does not know the distributions that govern the rewards for each
arm.
Of these points, the last is fundamental to the bandit problem and is discussed
in the next section. The lack of knowledge of the horizon is usually not a serious
issue. Generally speaking it is possible to first design a policy assuming the
horizon is known and then adapt it to account for the unknown horizon while
proving that the loss in performance is minimal. This is almost always quite easy,
and there exist generic approaches for making the conversion.
Assigning a utility to distributions of Sn is more challenging. Suppose
that Sn is the revenue of your company. Fig. 4.1 shows the distribution of
Sn for two different learners; call them A and B. Suppose you can choose
between learners A and B. Which one would you choose? One choice is to
go with the learner whose reward distribution has the larger expected value.
This will be our default choice for
stochastic bandits, but it bears remem-
bering that there are other consider- A B
ations, including the variance or tail
behaviour of the cumulative reward,
Density

which we will discuss occasionally. In

particular, in the situation shown on
in Fig. 4.1, learner B achieves a higher
expected reward than A. However B
has a reasonable probability of earning
Reward
less than the least amount that A can
earn, so a risk-sensitive user may prefer Figure 4.1 Alternative revenue distributions
learner A.

4.3 Knowledge and Environment Classes

Even if the horizon is known in advance and we commit to maximising the expected
value of Sn , there is still the problem that the bandit instance ν = (Pa : a ∈ A) is
 4.3 Knowledge and Environment Classes 58

unknown. A policy that maximises the expectation of Sn for one bandit instance
may behave quite badly on another. The learner usually has partial information
about ν, which we represent by defining a set of bandits E for which ν ∈ E is
guaranteed. The set E is called the environment class. We distinguish between
structured and unstructured bandits.

Unstructured Bandits
An environment class E is unstructured if A is finite and there exist sets of
distributions Ma for each a ∈ A such that

E = {ν = (Pa : a ∈ A) : Pa ∈ Ma for all a ∈ A} ,

or, in short, E = ×a∈A Ma . The product structure means that by playing action
a the learner cannot deduce anything about the distributions of actions b 6= a.
Some typical choices of unstructured bandits are listed in Table 4.1. Of course,
these are not the only choices, and the reader can no doubt find ways to construct
more, e.g. by allowing some arms to be Bernoulli and some Gaussian, or have
rewards being exponentially distributed, or Gumbel distributed, or belonging to
your favourite (non-)parametric family.
The Bernoulli, Gaussian and uniform distributions are often used as examples
for illustrating some specific property of learning in stochastic bandit problems.
The Bernoulli distribution is actually a natural choice. Think of applications like
maximising click-through rates in a web-based environment. A bandit problem
is often called a ‘distribution bandit’, where ‘distribution’ is replaced by the
underlying distribution from which the pay-offs are sampled. Some examples
are: Gaussian bandit, Bernoulli bandit or subgaussian bandit. Similarly we say
‘bandits with X’, where ‘X’ is a property of the underlying distribution from
which the pay-offs are sampled. For example, we can talk about bandits with
finite variance, meaning the bandit environment where the a priori knowledge of
the learner is that all pay-off distributions are such that their underlying variance
is finite.
Some environment classes, like Bernoulli bandits, are parametric, while others,
like subgaussian bandits, are non-parametric. The distinction is the number of
degrees of freedom needed to describe an element of the environment class. When
the number of degrees of freedom is finite, it is parametric, and otherwise it is
non-parametric. Of course, if a learner is designed for a specific environment class
E, then we might expect that it has good performance on all bandits ν ∈ E. Some
environment classes are subsets of other classes. For example, Bernoulli bandits
are a special case of bandits with a finite variance, or bandits with bounded
support. Something to keep in mind is that we expect that it will be harder to
achieve a good performance in a larger class. In a way, the theory of finite-armed
stochastic bandits tries to quantify this expectation in a rigourous fashion.
 4.3 Knowledge and Environment Classes 59

Name Symbol Definition

Bernoulli EBk {(B(µi ))i : µ ∈ [0, 1]k }

Uniform EUk {(U(ai , bi ))i : a, b ∈ Rk with ai ≤ bi for all i}

Gaussian (known var.) k

EN (σ 2 ) {(N (µi , σ 2 ))i : µ ∈ Rk }

Gaussian (unknown var.) k

EN {(N (µi , σi2 ))i : µ ∈ Rk and σ 2 ∈ [0, ∞)k }

Finite variance EVk (σ 2 ) {(Pi )i : VX∼Pi [X] ≤ σ 2 for all i}

Finite kurtosis k
EKurt (κ) {(Pi )i : KurtX∼Pi [X] ≤ κ for all i}

Bounded support k
E[a,b] {(Pi )i : Supp(Pi ) ⊆ [a, b]}

Subgaussian k
ESG (σ 2 ) {(Pi )i : Pi is σ-subgaussian for all i}

Table 4.1 Typical environment classes for stochastic bandits. Supp(P ) is the (topological)
support of distribution P . The kurtosis of a random variable X is a measure of its tail
behaviour and is defined by E[(X − E[X])4 ]/V[X]2 . Subgaussian distributions have similar
properties to the Gaussian and will be defined in Chapter 5.

Structured Bandits
Environment classes that are not unstructured are called structured. Relaxing the
requirement that the environment class is a product set makes structured bandit
problems much richer than the unstructured set-up. The following examples
illustrate the flexibility.

Example 4.1. Let A = {1, 2} and E = {(B(θ), B(1 − θ)) : θ ∈ [0, 1]}. In this
environment class, the learner does not know the mean of either arm, but can
learn the mean of both arms by playing just one. The knowledge of this structure
dramatically changes the difficulty of learning in this problem.

Example 4.2 (Stochastic linear bandit). Let A ⊂ Rd and θ ∈ Rd and

νθ = (N (ha, θi, 1) : a ∈ A) and E = {νθ : θ ∈ Rd } .

In this environment class, the reward of an action is Gaussian, and its mean is given
by the inner product between the action and some unknown parameter. Notice
that even if A is extremely large, the learner can deduce the true environment
by playing just d actions that span Rd .

Example 4.3. Consider an undirected graph G with vertices V = {1, . . . , |V |}

and edges E = {1, . . . , |E|}. In each round the learner chooses a path from
vertex 1 to vertex |V |. Then each edge e ∈ [E] is removed from the graph with
probability 1 − θe for unknown θ ∈ [0, 1]|E| . The learner succeeds in reaching
their destination if all the edges in their chosen path are present. This problem
 4.4 The Regret 60

can be formalised by letting A be the set of paths and

! !
Y
νθ = B θe : a ∈ A and E = {νθ : θ ∈ [0, 1]|E| } .
e∈a

An important feature of structured bandits is that the learner can often

obtain information about some actions while never playing them.

4.4 The Regret

In Chapter 1 we informally defined the regret as being the deficit suffered by the
learner relative to the optimal policy. Let ν = (Pa : a ∈ A) be a stochastic bandit
and define
Z ∞
µa (ν) = x dPa (x) .
−∞

Then let µ (ν) = maxa∈A µa (ν) be the largest mean of all the arms.
∗

We assume throughout that µa (ν) exists and is finite for all actions and
that argmaxa∈A µa (ν) is non-empty. The latter assumption could be relaxed
by carefully adapting all arguments using nearly optimal actions, but in
practice this is never required.

The regret of policy π on bandit instance ν is

" n #
X
Rn (π, ν) = nµ (ν) − E
∗
Xt , (4.1)
t=1

where the expectation is taken with respect to the probability measure on

outcomes induced by the interaction of π and ν. Minimising the regret is equivalent
to maximising the expectation of Sn , but the normalisation inherent in the
definition of the regret is useful when stating results, which would otherwise need
to be stated relative to the optimal action.

If the context is clear, we will often drop the dependence on ν and π in various
Pn
quantities. For example, by writing Rn = nµ∗ − E[ t=1 Xt ]. Similarly, the
limits in sums and maxima are abbreviated when we think you can work
out ranges of symbols in a unique way, e.g. µ∗ = maxi µi .

The regret is always non-negative, and for every bandit ν, there exists a policy
π for which the regret vanishes.
 4.4 The Regret 61

Lemma 4.4. Let ν be a stochastic bandit environment. Then,

(a) Rn (π, ν) ≥ 0 for all policies π;
(b) the policy π choosing At ∈ argmaxa µa for all t satisfies Rn (π, ν) = 0; and
(c) if Rn (π, ν) = 0 for some policy π, then P (µAt = µ∗ ) = 1 for all t ∈ [n].
We leave the proof for the reader (Exercise 4.1). Part (b) of Lemma 4.4 shows
that for every bandit ν, there exists a policy for which the regret is zero (the best
possible outcome). According to Part (c), achieving zero is possible if and only if
the learner knows which bandit it is facing (or at least, what is the optimal arm).
In general, however, the learner only knows that ν ∈ E for some environment
class E. So what can we hope for? A relatively weak objective is to find a policy
π with sublinear regret on all ν ∈ E. Formally, this objective is to find a policy π
such that
Rn (π, ν)
for all ν ∈ E , lim = 0.
n→∞ n
If the above holds, then at least the learner is choosing the optimal action almost
all of the time as the horizon tends to infinity. One might hope for much more,
however, for example, that for some specific choice of C > 0 and p < 1 that
for all ν ∈ E , Rn (π, ν) ≤ Cnp . (4.2)
Yet another alternative is to find a function C : E → [0, ∞) and f : N → [0, ∞)
such that
for all n ∈ N, ν ∈ E , Rn (π, ν) ≤ C(ν)f (n) . (4.3)
This factorisation of the regret into a function of the instance and a function
of the horizon is not uncommon in learning theory and appears in particular in
supervised learning.
We will spend a lot of time in the following chapters finding policies satisfying
Eq. (4.2) and Eq. (4.3) for different choices of E. The form of Eq. (4.3) is quite
general, so much time is also spent discovering what are the possibilities for f and
C, both of which should be ‘as small as possible’. All of the policies are inspired
by the simple observation that in order to make the regret small, the algorithm
must discover the action/arm with the largest mean. Usually this means the
algorithm should play each arm some number of times to form an estimate of
the mean of that arm, and subsequently play the arm with the largest estimated
mean. The question essentially boils down to discovering exactly how often the
learner must play each arm in order to have reasonable statistical certainty that
it has found the optimal arm.
There is another candidate objective called the Bayesian regret. If Q is a
prior probability measure on E (which must be equipped with a σ-algebra F),
then the Bayesian regret is the average of the regret with respect to the prior Q.
Z
BRn (π, Q) = Rn (π, ν) dQ(ν) , (4.4)
E
 4.5 Decomposing the Regret 62

which is only defined by assuming (or proving) that the regret is a measurable
function with respect to F. An advantage of the Bayesian approach is that
having settled on a prior and horizon, the problem of finding a policy that
minimises the Bayesian regret is just an optimisation problem. Most of this
book is devoted to analyzing the ‘frequentist’ regret in Eq. (4.1), which does not
integrate over all environments as Eq. (4.4) does. Bayesian methods are covered
in Chapters 34 to 36, where we also discuss the strengths and weaknesses of the
Bayesian approach.

4.5 Decomposing the Regret

We now present a lemma that forms the basis of almost every proof for
stochastic bandits. Let ν = (Pa : a ∈ A) be a stochastic bandit and define
∆a (ν) = µ∗ (ν) − µa (ν), which is called the suboptimality gap or action gap
or immediate regret of action a. Further, let
t
X
Ta (t) = I {As = a}
s=1

be the number of times action a was chosen by the learner after the end of round
t. In general, Ta (n) is random, which may seem surprising if we think about a
deterministic policy that chooses the same action for any fixed history. So why
is Ta (n) random in this case? The reason is because for all rounds t except for
the first, the action At depends on the rewards observed in rounds 1, 2, . . . , t − 1,
which are random, hence At will also inherit their randomness. We are now ready
to state the second and last lemmas of the chapter. In the statement of the lemma,
we use our convention that the dependence of the various quantities involved on
the policy π and the environment ν is suppressed.
Lemma 4.5 (Regret decomposition lemma). For any policy π and stochastic
bandit environment ν with A finite or countable and horizon n ∈ N, the regret
Rn of policy π in ν satisfies
X
Rn = ∆a E [Ta (n)] . (4.5)
a∈A

The lemma decomposes the regret in terms of the loss due to using each of the
arms. It is useful because it tells us that to keep the regret small, the learner
should try to minimise the weighted sum of expected action counts, where the
weights are the respective suboptimality gaps, (∆a )a∈A .

Lemma 4.5 tells us that a learner should aim to use an arm with a larger
suboptimality gap proportionally fewer times.

Note that the suboptimality gap for optimal arm(s) is zero.

 4.6 The Canonical Bandit Model ( ) 63

Proof of Lemma 4.5 Since Rn is based on summing over rounds, and the right-
hand side of the lemma statement is based on summing over actions, to convert one
sum into the other one, we introduce indicators. In particular, note that for any
P P P P
fixed t we have a∈A I {At = a} = 1. Hence Sn = t Xt = t a Xt I {At = a},
and thus
n
XX
Rn = nµ∗ − E [Sn ] = E [(µ∗ − Xt )I {At = a}] . (4.6)
a∈A t=1

The expected reward in round t conditioned on At is µAt , which means that

E [(µ∗ − Xt )I {At = a} | At ] = I {At = a} E [µ∗ − Xt | At ]

= I {At = a} (µ∗ − µAt )
= I {At = a} (µ∗ − µa )
= I {At = a} ∆a .

The result is completed by plugging this into Eq. (4.6) and using the definition
of Ta (n).

The argument fails when A is uncountable because you cannot introduce the
sum over actions. Of course the solution is to use an integral, but for this we need
to assume (A, G) is a measurable space. Given a bandit ν and policy π define
measure G on (A, G) by
" n #
X
G(U ) = E I {At ∈ U } ,
t=1

where the expectation is taken with respect to the measure on outcomes induced
by the interaction of π and ν.

Lemma 4.6. Provided that everything is well defined and appropriately measurable,
" n # Z
X
Rn = E ∆At = ∆a dG(a) .
t=1 A

For those worried about how to ensure everything is well defined, see Section 4.7.

4.6 The Canonical Bandit Model ( )

In most cases the underlying probability space that supports the random rewards
and actions is never mentioned. Occasionally, however, it becomes convenient to
choose a specific probability space, which we call the canonical bandit model.
 4.6 The Canonical Bandit Model ( ) 64

Finite Horizon
Let n ∈ N be the horizon. A policy and bandit interact to produce the outcome,
which is the tuple of random variables Hn = (A1 , X1 , . . . , An , Xn ). The first step
towards constructing a probability space that carries these random variables
is to choose the measurable space. For each t ∈ [n], let Ωt = ([k] × R)t ⊂ R2t
and Ft = B(Ωt ). The random variables A1 , X1 , . . . , An , Xn that make up the
outcome are defined by their coordinate projections:
At (a1 , x1 , . . . , an , xn ) = at and Xt (a1 , x1 , . . . , an , xn ) = xt .
The probability measure on (Ωn , Fn ) depends on both the environment and the
policy. Our informal definition of a policy is not quite sufficient now.
Definition 4.7. A policy π is a sequence (πt )nt=1 , where πt is a probability
kernel from (Ωt−1 , Ft−1 ) to ([k], 2[k] ). Since [k] is discrete, we adopt the notational
convention that for i ∈ [k],
πt (i | a1 , x1 , . . . , at−1 , xt−1 ) = πt ({i} | a1 , x1 , . . . , at−1 , xt−1 ) .
Let ν = (Pi )ki=1 be a stochastic bandit where each Pi is a probability measure
on (R, B(R)). We want to define a probability measure on (Ωn , Fn ) that respects
our understanding of the sequential nature of the interaction between the learner
and a stationary stochastic bandit. Since we only care about the law of the
random variables (Xt ) and (At ), the easiest way to enforce this is to directly list
our expectations, which are
(a) the conditional distribution of action At given A1 , X1 , . . . , At−1 , Xt−1 is
πt ( · | A1 , X1 , . . . , At−1 , Xt−1 ) almost surely.
(b) the conditional distribution of reward Xt given A1 , X1 , . . . , At is PAt almost
surely.
The sufficiency of these assumptions is asserted by the following proposition,
which we ask you to prove in Exercise 4.2.
Proposition 4.8. Suppose that P and Q are probability measures on an arbitrary
measurable space (Ω, F) and A1 , X1 , . . . , An , Xn are random variables on Ω, where
At ∈ [k] and Xt ∈ R. If both P and Q satisfy (a) and (b), then the law of the
outcome under P is the same as under Q:
PA1 ,X1 ,...,An ,Xn = QA1 ,X1 ,...,An ,Xn .
Next we construct a probability measure on (Ωn , Fn ) that satisfies (a) and
(b). To emphasise that what follows is intuitively not complicated, imagine that
Xt ∈ {0, 1} is Bernoulli, which means the set of possible outcomes is finite and
we can define the measure in terms of a distribution. Let pi (0) = Pi ({0}) and
pi (1) = 1 − pi (0) and define
n
Y
pνπ (a1 , x1 , . . . , an , xn ) = π(at | a1 , x1 , . . . , at−1 , xt−1 )pat (xt ) .
t=1
 4.6 The Canonical Bandit Model ( ) 65

The reader can check that pνπ is a distribution on ([k] × {0, 1})n and that the
associated measure satisfies (a) and (b) above. Making this argument rigourous
when (Pi ) are not discrete requires the use of Radon–Nikodym derivatives. Let λ
be a σ-finite measure on (R, B(R)) for which Pi is absolutely continuous with
respect to λ for all i. Next, let pi = dPi /dλ be the Radon–Nikodym
R derivative of
Pi with respect to λ, which is a function pi : R → R such that B pi dλ = Pi (B)
for all B ∈ B(R). Letting ρ be the counting measure with ρ(B) = |B|, the density
pνπ : Ω → R can now be defined with respect to the product measure (ρ × λ)n
by

n
Y
pνπ (a1 , x1 , . . . , an , xn ) = π(at | a1 , x1 , . . . , at−1 , xt−1 )pat (xt ) . (4.7)
t=1

The reader can again check (more abstractly) that (a) and (b) are satisfied by
the probability measure Pνπ defined by
Z
Pνπ (B) = pνπ (ω)(ρ × λ)n (dω) for all B ∈ Fn .
B

It is important to emphasise that this choice of (Ωn , Fn , Pνπ ) is not unique. Instead,
all that this shows is that a suitable probability space does exist. Furthermore, if
some quantity of interest depends on the law of Hn , by Proposition 4.8, there is
no loss in generality in choosing (Ωn , Fn , Pνπ ) as the probability space.

Pk
A choice of λ such that Pi  λ for all i always exists since λ = i=1 Pi
satisfies this condition. For direct calculations, another choice is usually
more convenient, e.g. the counting measure when (Pi ) are discrete and the
Lebesgue measure for continuous (Pi ).

There is another way to define the probability space, which can be useful.
Define a collection of independent random variables (Xsi )s∈[n],i∈[k] such that the
law of Xti is Pi . By Theorem 2.4 these random variables may be defined on
(Ω, F), where Ω = Rnk and F = B(Rnk ). Then let Xt = XtAt , where the actions
At are Ft−1 -measurable with Ft−1 = σ(A1 , X1 , . . . , At−1 , Xt−1 ). We call this the
random table model. Yet another way is to define (Xsi )s,i as above but let
Xt = XTAt (t),At . This corresponds to sampling a stack of rewards for each arm
at the beginning of the game, giving rise to the reward-stack model. Each time
the learner chooses an action, they receive the reward on top of the stack. All of
these models are convenient from time to time. The important thing is that it
does not matter which model we choose because the quantity of ultimate interest
(usually the regret) only depends on the law of A1 , X1 , . . . , An , Xn , and this is
the same for all choices.
 4.7 The Canonical Bandit Model for Uncountable Action Sets ( ) 66

Infinite Horizon
We never need the canonical bandit model for the case that n = ∞. It is comforting
to know, however, that there does exist a probability space (Ω, F, Pνπ ) and infinite
sequences of random variables X1 , X2 , . . . and A1 , A2 , . . . satisfying (a) and (b).
The result follows directly from the theorem of Ionescu–Tulcea (Theorem 3.3).

4.7 The Canonical Bandit Model for Uncountable Action Sets ( )

For uncountable action sets, a little more machinery is necessary to make things
rigourous. The first requirement is that the action set must be a measurable
space (A, G) and the collection of distribution ν = (Pa : a ∈ A) that defines a
bandit environment must be a probability kernel from (A, G) to (R, B(R)). A
policy is a sequence (πt )nt=1 , where πt is a probability kernel from (Ωt−1 , Ft−1 )
to (A, G) with
t
Y t
O
Ωt = (A × R) and Ft = (G ⊗ B(R)) .
s=1 s=1

The canonical bandit model is the probability measure Pνπ on (Ωn , Fn )

obtained by taking the product of the probability kernels π1 , P1 , . . . πn , Pn and
using Ionescu–Tulcea (Theorem 3.3), where Pt is the probability kernel from
(Ωt−1 × A, Ft ⊗ G) to (R, B(R)) given by Pt ( · | a1 , x1 , . . . , at−1 , xt−1 , at ) = Pat (·).

We did not define Pνπ in terms of a density because there may not exist a
common dominating measure for either (Pa : a ∈ A) or the policy. When
such measures exist, as they usually do, then Pνπ may be defined in terms
of a density in the same manner as the previous section.

You will check in Exercise 4.5 that the assumptions on ν and π in this section
are sufficient to ensure the quantities in Lemma 4.6 are well defined and that
Proposition 4.8 continues to hold in this setting without modification. Finally, in
none of the definitions above do we require that n be finite.

4.8 Notes

1 It is not obvious why the expected value is a good summary of the reward
distribution. Decision makers who base their decisions on expected values are
called risk-neutral. In the example shown on the figure above, a risk-averse
decision maker may actually prefer the distribution labelled as A because
occasionally distribution B may incur a very small (even negative) reward.
Risk-seeking decision makers, if they exist at all, would prefer distributions
with occasional large rewards to distributions that give mediocre rewards only.
 4.8 Notes 67

There is a formal theory of what makes a decision maker rational (a decision

maker in a nutshell is rational if they do not contradict themself). Rational
decision makers compare stochastic alternatives based on the alternatives’
expected utilities, according to the von-Neumann–Morgenstern utility theorem.
Humans are known not to do this. We are irrational. No surprise here.
2 The study of utility and risk has a long history, going right back to (at least)
the beginning of probability [Bernoulli, 1954, translated from the original
Latin, 1738]. The research can broadly be categorised into two branches. The
first deals with describing how people actually make choices (descriptive
theories), while the second is devoted to characterising how a rational decision
maker should make decisions (prescriptive theories). A notable example
of the former type is ‘prospect theory’ [Kahneman and Tversky, 1979], which
models how people handle probabilities (especially small ones) and earned
Daniel Kahneman a Nobel Prize (after the death of his long-time collaborator,
Amos Tversky). Further descriptive theories concerned with alternative aspects
of human decision-making include bounded rationality, choice strategies,
recognition-primed decision-making and image theory [Adelman, 2013].
3 The most famous example of a prescriptive theory is the von Neumann–
Morgenstern expected utility theorem, which states that under (reasonable)
axioms of rational behaviour under uncertainty, a rational decision maker
must choose amongst alternatives by computing the expected utility of the
outcomes [Neumann and Morgenstern, 1944]. Thus, rational decision makers,
under the chosen axioms, differ only in terms of how they assign utility to
outcomes (i.e. rewards). Finance is another field where attitudes towards
uncertainty and risk are important. Markowitz [1952] argues against expected
return as a reasonable metric that investors would use. His argument is based
on the (simple) observation that portfolios maximising expected returns will
tend to have a single stock only (unless there are multiple stocks with equal
expected returns, a rather unlikely outcome). He argues that such a complete
lack of diversification is unreasonable. He then proposes that investors should
minimise the variance of the portfolio’s return subject to a constraint on the
portfolio’s expected return, leading to the so-called mean-variance optimal
portfolio choice theory. Under this criteria, portfolios will indeed tend to
be diversified (and in a meaningful way: correlations between returns are taken
into account). This theory eventually won him a Nobel Prize in economics
(shared with two others). Closely related to the mean-variance criterion are the
‘value-at-risk’ (VaR) and the ‘conditional value-at-risk’, the latter of which has
been introduced and promoted by Rockafellar and Uryasev [2000] due to its
superior optimisation properties. The distinction between the prescriptive and
descriptive theories is important: human decision makers are in many ways
violating rules of rationality in their attitudes towards risk.
4 We defined the regret as an expectation, which makes it unusable in conjunction
with measures of risk because the randomness has been eliminated by the
expectation. When using a risk measure in a bandit setting, we can either base
 4.9 Bibliographical Remarks 68

this on the random regret or pseudo-regret defined by

n
X
R̂n = nµ∗ − Xt . (random regret)
t=1
Xn
R̄n = nµ∗ − µAt . (pseudo-regret)
t=1

While R̂n is influenced by the noise Xt − µAt in the rewards, the pseudo-regret
filters this out, which arguably makes it a better basis for measuring the ‘skill’
of a bandit policy. As these random regret measures tend to be highly skewed,
using variance to assess risk suffers not only from the problem of penalising
upside risk, but also from failing to capture the skew of the distribution.
5 What happens if the distributions of the arms are changing with time?
Such bandits are unimaginatively called non-stationary bandits. With no
assumptions, there is not much to be done. Because of this, it is usual to
assume the distributions change infrequently or drift slowly. We’ll eventually
see that techniques for stationary bandits can be adapted to this set-up (see
Chapter 31).
6 The rigourous models introduced in Sections 4.6 and 4.7 are easily extended to
more sophisticated settings. For example, the environment sometimes produces
side information as well as rewards or the set of available actions may change
with time. You are asked to formalise an example in Exercise 4.6.

4.9 Bibliographical Remarks

There is now a huge literature on stochastic bandits, much of which we will

discuss in detail in the chapters that follow. The earliest reference that we know
of is by Thompson [1933], who proposed an algorithm that forms the basis
of many of the currently practical approaches in use today. Thompson was a
pathologist who published broadly and apparently did not pursue bandits much
further. Sadly his approach was not widely circulated, and the algorithm (now
called Thompson sampling) did not become popular until very recently. Two
decades after Thompson, the bandit problem was formally restated in a short but
influential paper by Robbins [1952], an American statistician now most famous
for his work on empirical Bayes. Robbins introduced the notion of regret and
minimax regret in his 1952 paper. The regret decomposition (Lemma 4.5) has
been used in practically every work on stochastic bandits, and its origin is hard
to pinpoint. All we can say for sure is that it does not appear in the paper by
Robbins [1952], but does appear in the work of Lai and Robbins [1985]. Denardo
et al. [2007] considers risk in a (complicated) Bayesian setting. Sani et al. [2012]
consider a mean-variance approach to risk, while Maillard [2013] considers so-
called coherence risk measures (CVaR, is one example of such a risk measure),
 4.10 Exercises 69

and with an approach where the regret itself is redefined. VaR is considered in
the context of a specific bandit policy family by Audibert et al. [2007, 2009].

4.10 Exercises

4.1 (Positivity of the regret) Prove Lemma 4.4.

4.2 (Uniqueness of law) Prove Proposition 4.8.

4.3 (Definition of canonical probability measure) Prove that the measure

defined in terms of the density in Eq. (4.7) satisfies the conditions (a) and (b) in
Section 4.6.

Hint Use the properties of the Radon–Nikodym derivative in combination with

Fubini’s theorem.
4.4 (Mixing policies) Fix a horizon n and k. Let Π be a finite set of policies
for k-armed bandits on horizon n and p ∈ P(Π) be a distribution over Π. Show
there exists a policy π ◦ such that for any k-armed stochastic bandit ν,
X
Pνπ◦ = p(π)Pνπ .
π∈Π

Proof For action/reward sequence a1 , x1 , . . . , an , xn , syntactically abbreviate

ht = a1 , x1 , . . . , at , xt . Then define
P Qt
π∈Π p(π) s=1 πs (as | hs−1 )
πt (at | ht−1 ) = P
◦
Qt−1 .
π∈Π p(π) s=1 πs (as | hs−1 )

By the definition of the canonical probability space and the product of probability
kernels,
Xk Z Xk Z
Pνπ◦ (B) = ··· IB (hn )νan (dxn )πn◦ (an | hn−1 ) · · · νa1 (dx1 )π1◦ (a1 )
a1 =1 R an =1 R

X k Z
X k Z
X
= p(π) ··· IB (hn )νan (dxn )πn (an | hn−1 ) · · · νa1 (dx1 )π1 (a1 )
π∈Π a1 =1 R an =1 R
X
= p(π)Pνπ (B) ,
π∈Π

where the second equality follows by substituting the definition of π ◦ and

induction.
4.5 (Regret decomposition and canonical model for large action
spaces) Let ν be a bandit on measurable action space (A, G) and π1 , . . . , πn be
a policy satisfying the conditions in Section 4.7.
(a) Show that all quantities in Lemma 4.6 are appropriately defined and
measurable.
 4.10 Exercises 70

(b) Prove Lemma 4.6.

(c) Prove that Proposition 4.8 continues to hold.

4.6 (Canonical model for contextual bandit) Let A and C be finite sets.
A stochastic contextual bandit is like a normal stochastic bandit, but in each
round the learner first observes a context Ct ∈ C. They then choose an action
At ∈ A and receive a reward Xt ∼ PAt ,Ct .

(a) Suppose that C1 , . . . , Cn is sampled independently from distribution

ξ on C. Construct the canonical probability space that carries
C1 , A1 , X1 , . . . , Cn , An , Xn .
(b) What changes when Ct is allowed to depend on C1 , A1 , X1 , . . . , Ct−1 , At−1 , Xt−1 ?

4.7 (Bernoulli environment implementation) Implement a Bernoulli bandit

environment in Python using the code snippet below (or adapt to your favourite
language).
class B er no ul li B an di t :
# accepts a list of K >= 2 floats , each lying in [0 ,1]
def __init__ ( self , means ) :
pass

# F u n c t i o n should return the number of arms

def K ( self ) :
pass

# Accepts a p a r a m e t e r 0 <= a <= K -1 and returns the

# r e a l i s a t i o n of random v a r i a b l e X with P ( X = 1) being
# the mean of the ( a +1) th arm .
def pull ( self , a ) :
pass

# Returns the regret i n c u r r e d so far .

def regret ( self ) :
pass

4.8 (Follow-the-leader implementation) Implement the following simple

algorithm called ‘follow-the-leader’, which chooses each action once and
subsequently chooses the action with the largest average observed so far. Ties
should be broken randomly.
def F ol lo wT he L ea de r ( bandit , n ) :
# i m p l e m e n t the Follow - the - Leader a l g o r i t h m by r e p l a c i n g
# the code below that just plays the first arm in every round
for t in range ( n ) :
bandit . pull (0)

Depending on the literature you are reading, follow-the-leader may be called

‘stay with the winner’ or the ‘greedy algorithm’.
 4.10 Exercises 71

4.9 Suppose ν is a finite-armed stochastic bandit and π is a policy such that

Rn (π, ν)
lim = 0.
n→∞ n
Pn
Let T ∗ (n) = t=1 I {µAt = µ∗ } be the number of times an optimal arm is chosen.
Prove or disprove each of the following statements:
(a) limn→∞ E[T ∗ (n)]/n = 1.
(b) limn→∞ P (∆At > 0) = 0.

4.10 (one-armed bandits) Let M1 be a set of distributions on (R, B(R)) with

finite means and M2 = {δµ2 } be the singleton set with a Dirac at µ2 ∈ R. The
set of bandits E = M1 × M2 is called a one-armed bandit because, although
there are two arms, the second arm always yields a known reward of µ2 . A policy
π = (πt )t is called a retirement policy if once action 2 has been played once,
it is played until the end of the game. Precisely, if at = 2, then
πt+1 (2 | a1 , x1 , . . . , at , xt ) = 1 for all (as )t−1
s=1 and (xs )s=1 .
t

(a) Let n be fixed and π = (πt )nt=1 be any policy. Prove there exists a retirement
policy π 0 = (πt0 )nt=1 such that for all ν ∈ E.
Rn (π 0 , ν) ≤ Rn (π, ν) .
(b) Let M1 = {B(µ1 ) : µ1 ∈ [0, 1]} and suppose that π = (πt )∞
t=1 is a retirement
policy. Prove there exists a bandit ν ∈ E such that
Rn (π, ν)
lim sup > 0.
n→∞ n
4.11 (Failure of follow-the-leader (i)) Consider a Bernoulli bandit with
two arms and means µ1 = 0.5 and µ2 = 0.6.
(a) Using a horizon of n = 100, run 1000 simulations of your implementation
of follow-the-leader on the Bernoulli bandit above and record the (random)
Pn
pseudo regret, nµ∗ − t=1 µAt , in each simulation.
(b) Plot the results using a histogram. Your figure should resemble Fig. 4.2.
(c) Explain the results in the figure.

4.12 (Failure of follow-the-leader (ii)) Consider the same Bernoulli

bandit as used in the previous question.
(a) Run 1000 simulations of your implementation of follow-the-leader for each
horizon n ∈ {100, 200, 300, . . . , 1000}.
(b) Plot the average regret obtained as a function of n (see Fig. 4.3). Because the
average regret is an estimator of the expected regret, you should generally
include error bars to indicate the uncertainty in the estimation.
(c) Explain the plot. Do you think follow-the-leader is a good algorithm?
Why/why not?
 4.10 Exercises 72

Follow-the-leader

400
Frequency

200

0 2 4 6 8 10
Regret

Figure 4.2 Histogram of regret for follow-the-leader over 1000 trials on a Bernoulli bandit
with means µ1 = 0.5, µ2 = 0.6

Follow-the-leader
50

40
Expected Regret

30

20

10

200 400 600 800 1,000

n

Figure 4.3 The regret for Follow-the-leader over 1000 trials on Bernoulli bandit with
means µ1 = 0.5, µ2 = 0.6 and horizons ranging from n = 100 to n = 1000.
5 Concentration of Measure

Before we can start designing and analysing algorithms, we need one more tool
from probability theory, called concentration of measure. Recall that the
optimal action is the one with the largest mean. Since the mean pay-offs are
initially unknown, they must be learned from data. How long does it take to
learn about the mean reward of an action? In this section, after introducing
the notion of tail probabilities, we look at ways of obtaining upper bounds on
them. The main point is to introduce subgaussian random variables and the
Cramér–Chernoff exponential tail inequalities, which will play a central role in
the design and analysis of the various bandit algorithms.

5.1 Tail Probabilities

Suppose that X, X1 , X2 , . . . , Xn is a sequence of independent and identically

distributed random variables, and assume that the mean µ = E[X] and variance
σ 2 = V[X] exist. Having observed X1 , X2 , . . . , Xn , we would like to estimate the
common mean µ. The most natural estimator is

1X
n
µ̂ = Xi ,
n i=1

which is called the sample mean or empirical mean. Linearity of expectation

(Proposition 2.6) shows that E[µ̂] = µ, which means that µ̂ is an unbiased
estimator of µ. How far from µ do we expect µ̂ to be? A simple measure
of the spread
 of the distribution of a random variable Z is its variance,
V [Z] = E (Z − E [Z])2 . A quick calculation using independence shows that

  σ2
V [µ̂] = E (µ̂ − µ)2 = , (5.1)
n
which means that we expect the squared distance between µ and µ̂ to shrink as
n grows large at a rate of 1/n and scale linearly with the variance of X. While
the expected squared error is important, it does not tell us very much about the
distribution of the error. To do this we usually analyse the probability that µ̂
overestimates or underestimates µ by more than some value ε > 0. Precisely, how
 5.2 The Inequalities of Markov and Chebyshev 74

Figure 5.1 The figure shows a probability density, with the tails shaded indicating the
regions where X is at least ε away from the mean µ.

do the following quantities depend on ε?

P (µ̂ ≥ µ + ε) and P (µ̂ ≤ µ − ε) .
The expressions above (as a function of ε) are called the tail probabilities of
µ̂ − µ (Fig. 5.1). Specifically, the first is called the upper tail probability and the
second the lower tail probability. Analogously, P (|µ̂ − µ| ≥ ε) is called a two-sided
tail probability.

5.2 The Inequalities of Markov and Chebyshev

The most straightforward way to bound the tails is by using Chebyshev’s

inequality, which is itself a corollary of Markov’s inequality. The latter is
one of the golden hammers of probability theory, and so we include it for the
sake of completeness.
Lemma 5.1. For any random variable X and ε > 0, the following holds:
E [|X|]
(a) (Markov): P (|X| ≥ ε) ≤ .
ε
V [X]
(b) (Chebyshev): P (|X − E [X] | ≥ ε) ≤ .
ε2
We leave the proof of Lemma 5.1 as an exercise for the reader. By combining
(5.1) with Chebyshev’s inequality, we can bound the two-sided tail directly in
terms of the variance by
σ2
P (|µ̂ − µ| ≥ ε) ≤ . (5.2)
nε2
This result is nice because it was so easily bought and relied on no assumptions
other than the existence of the mean and variance. The downside is that when X is
well behaved, the inequality is rather loose. By assuming that higher moments of
X exist, Chebyshev’s inequality can be improved by applying Markov’s inequality
to |µ̂ − µ|k , with the positive integer k to be chosen so that the resulting bound is
 5.2 The Inequalities of Markov and Chebyshev 75

optimised. This is a bit cumbersome, and thus instead we present the continuous
analog of this, known as the Cramér-Chernoff method.
To calibrate our expectations on what improvement to expect relative to
Chebyshev’s inequality, let us start by recalling the central limit theorem
Pn
(CLT). Let Sn = t=1 (Xt − µ). The CLT says that under no additional
assumptions
√ than the existence of the variance, the limiting distribution of
Sn / nσ as n → ∞ is a Gaussian with mean zero and unit variance. If
2

Z ∼ N (0, 1), then

Z  2
∞
1 x
P (Z ≥ u) = √ exp − dx .
u 2π 2

The integral has no closed-form solution, but is easy to bound:

Z  2 Z ∞  2
∞
1 x 1 x
√ exp − dx ≤ √ x exp − dx
u 2π 2 u 2π u 2
r  2
1 u
= exp − , (5.3)
2πu2 2

which gives

 √ p   p 
P (µ̂ ≥ µ + ε) = P Sn / σ 2 n ≥ ε n/σ 2 ≈ P Z ≥ ε n/σ 2
r  
σ2 nε2
≤ exp − . (5.4)
2πnε2 2σ 2

This always improves on what we obtained with Chebyshev’s inequality, usually

by an enormous margin (Exercise 5.3). In particular, the bound on the right-hand
side of (5.4) decays slightly faster than the negative exponential of nε2 /(2σ 2 ),
which means that µ̂ rapidly concentrates around its mean.

An oft-taught rule of thumb is that the CLT provides a reasonable

approximation for n ≥ 30. We advise caution. Suppose that X1 , . . . , Xn
are independent Bernoulli with bias p = 1/n. As n tends to infinity the
Pn
distribution of t=1 Xt converges to a Poisson distribution with parameter
1, which does not look Gaussian at all.

The asymptotic nature of the CLT makes it unsuitable for designing bandit
algorithms. In the next section, we derive finite-time analogs, which are only
possible by making additional assumptions.
 5.3 The Cramér-Chernoff Method and Subgaussian Random Variables 76

5.3 The Cramér-Chernoff Method and Subgaussian Random

Variables

For the sake of moving rapidly towards bandits, we start with a straightforward
and relatively fundamental assumption on the distribution of X, known as the
subgaussian assumption.
Definition 5.2 (Subgaussianity). A random variable
X is σ-subgaussian if for
all λ ∈ R, it holds that E [exp(λX)] ≤ exp λ2 σ 2 /2 .
An alternative way to express the subgaussianity condition uses the moment-
generating function of X, which is a function MX : R → R defined by
MX (λ) = E [exp(λX)]. The condition in the definition can be written as
1 2 2
ψX (λ) = log MX (λ) ≤ λ σ for all λ ∈ R .
2
The function ψX is called the cumulant-generating function. It is not hard
to see that MX (or ψX ) need not exist for all random variables over the whole
range of real numbers. For example, if X is exponentially distributed and λ ≥ 1,
then
Z ∞
E [exp(λX)] = exp(−x) × exp(λx)dx = ∞ .
0 | {z }
density of exponential

The moment-generating function of X ∼ N (0, σ 2 ) satisfies MX (λ) = exp(λ2 σ 2 /2),

and so X is σ-subgaussian.

A random variable X is heavy tailed if MX (λ) = ∞ for all λ > 0. Otherwise

it is light tailed.

The following theorem explains the origin of the term ‘subgaussian’. The tails
of a σ-subgaussian random variable decay approximately as fast as that of a
Gaussian with zero mean and the same variance.
Theorem 5.3. If X is σ-subgaussian, then for any ε ≥ 0,
 
ε2
P (X ≥ ε) ≤ exp − 2 . (5.5)
2σ
Proof We take a generic approach called the Cramér–Chernoff method. Let
λ > 0 be some constant to be tuned later. Then
P (X ≥ ε) = P (exp (λX) ≥ exp (λε))
≤ E [exp (λX)] exp (−λε) (Markov’s inequality)
 2 2 
λ σ
≤ exp − λε . (Def. of subgaussianity)
2
Choosing λ = ε/σ 2 completes the proof.
5.3 The Cramér-Chernoff Method and Subgaussian Random Variables 77

A similar inequality holds for the left tail. By using the union bound
P (A ∪ B) ≤ P (A) + P (B), we also find that P (|X| ≥ ε) ≤ 2 exp(−ε2 /(2σ 2 )).
An equivalent form of these bounds is
 p   p 
P X ≥ 2σ 2 log(1/δ) ≤ δ P |X| ≥ 2σ 2 log(2/δ) ≤ δ .

This form is often more convenient and especially the latter, which for small δ
shows that with overwhelming probability X takes values in the interval
 p p 
− 2σ 2 log(2/δ), 2σ 2 log(2/δ) .

To study the tail behaviour of µ̂ − µ, we need one more lemma.

Lemma 5.4. Suppose that X is σ-subgaussian and X1 and X2 are independent

and σ1 and σ2 -subgaussian, respectively, then:

(a) E[X] = 0 and V [X] ≤ σ 2 .

(b) cX is |c|σ-subgaussian
p for all c ∈ R.
(c) X1 + X2 is σ12 + σ22 -subgaussian.

The proof of the lemma is left to the reader (Exercise 5.7). Combining
Lemma 5.4 and Theorem 5.3 leads to a straightforward bound on the tails
of µ̂ − µ.

Corollary 5.5. Assume that Xi − µ are independent, σ-subgaussian random

variables. Then for any ε ≥ 0,
   
nε2 nε2
P (µ̂ ≥ µ + ε) ≤ exp − 2 and P (µ̂ ≤ µ − ε) ≤ exp − 2 ,
2σ 2σ
1
Pn
where µ̂ = n t=1 Xt .
Pn √
Proof By Lemma 5.4, it holds that µ̂−µ = i=1 (Xi −µ)/n is σ/ n-subgaussian.
Then apply Theorem 5.3.

For x > 0, it holds that exp(−x) ≤ 1/(ex), which shows that the above
inequality is stronger than what we obtained via Chebyshev’s inequality except
when ε is very small. It is exponentially smaller if nε2 is large relative to σ 2 . The
deviation form of the above result says that under the conditions of the result,
for any δ ∈ [0, 1], with probability at least 1 − δ,
r
2σ 2 log(1/δ)
µ ≤ µ̂ + . (5.6)
n
Symmetrically, it also follows that with probability at least 1 − δ,
r
2σ 2 log(1/δ)
µ ≥ µ̂ − . (5.7)
n
Again, one can use a union bound to derive a two-sided inequality.

Example 5.6. The following random variables are subgaussian:

 5.4 Notes 78

(a) If X is Gaussian with mean zero and variance σ 2 , then X is σ-subgaussian.

(b) If X has mean zero and |X| ≤ B almost surely for B ≥ 0, then X is
B-subgaussian.

(c) If X has mean zero and X ∈ [a, b] almost surely, then X is (b − a)/2-
subgaussian.

If X is exponentially distributed with rate λ > 0, then X is not σ-subgaussian

for any σ ∈ R.

For random variables that are not centred (E [X] 6= 0), we abuse notation
by saying that X is σ-subgaussian if the noise X − E [X] is σ-subgaussian.
A distribution is called σ-subgaussian if a random variable drawn from that
distribution is σ-subgaussian. Subgaussianity is really a property of both a
random variable and the measure on the space on which it is defined, so the
nomenclature is doubly abused.

5.4 Notes

1 The Berry–Esseen theorem (independently discovered by Berry [1941] and

Esseen [1942]) quantifies the speed of convergence in the CLT. It essentially
says that the distance between the Gaussian and the actual distribution decays
√
at a rate of 1/ n under some mild assumptions (see Exercise 5.5). This is
known to be tight for the class of probability distributions that appear in the
Berry–Esseen result. However, this is a vacuous result when the tail probabilities
√
themselves are much smaller than 1/ n. Hence the need for concrete finite-time
results.

2 Theorem 5.3 shows that subgaussian random variables have tails that decay
almost as fast as a Gaussian. A version of the converse is also possible. That
is, if a centered random has tails that behave in a similar way to a Gaussian,
then it is subgaussian. In particular, the following holds: let X be a centered
random variable (E[X] = 0) with P (|X| ≥ ε) ≤ 2 exp(−ε2 /2). Then X is
 5.4 Notes 79

√
5-subgaussian:
" ∞
# ∞  i 
X λi X i X λ |X|i
E[exp(λX)] = E ≤1+ E
i=0
i! i=2
i!
Z  
i!1/i 1/i
∞
X ∞
≤1+ P |X| ≥ x dx (Exercise 2.19)
i=2 0
λ
∞ Z ∞  
X i!2/i x2/i
≤1+2 exp − dx (by assumption)
i=2 0
2λ2
    
√ λ
= 1 + 2πλ exp(λ2 /2) 1 + erf √ −1 (by Mathematica)
2
 2
5λ
≤ exp .
2
This bound is surely loose. At the same time, there is little room for
improvement: if X has density p(x) √ = |x| exp(−x2 /2)/2, then P (|X| ≥ ε) =
exp(−ε2 /2). And yet X is at best 2-subgaussian, so some degree of slack is
required (see Exercise 5.4).
3 We saw in (5.4) that if X1 , X2 , . . . , Xn are independent standard Gaussian
Pn
random variables and µ̂ = n1 t=1 , then
r  
σ2 nε2
P (µ̂ ≥ ε) ≤ exp − 2 .
2πnε2 2σ

If nε2 /σ 2 is relatively large, then this bound is marginally stronger than

exp(−nε2 /(2σ 2 )), which follows from the subgaussian analysis. One might ask
whether or not a similar improvement is possible more generally. And Talagrand
[1995] will tell you: yes! At least for bounded random variables (details in the
paper).
4 Hoeffding’s lemma states that for a zero-mean random variable X such that
X ∈ [a, b] almost surely for real values a < b, then MX (λ) ≤ exp(λ2 (b − a)2 /8).
Applying the Cramér–Chernoff method shows that if X1 , X2 , . . . , Xn are
independent and Xt ∈ [at , bt ] almost surely with at < bt for all t, then
!  
1X −2n2 ε2
n
P (Xt − E[Xt ]) ≥ ε ≤ exp Pn . (5.8)
t=1 (bt − at )
n t=1 2

The above is called Hoeffding’s inequality. For details see Exercise 5.11.
There are many variants of this result that provide tighter bounds when X
satisfies certain additional distributional properties like small variance (see
Exercise 5.14).
5 The Cramér–Chernoff method is applicable beyond the subgaussian case, even
when the moment-generating function is not defined globally. One example
where this occurs is when X1 , X2 , . . . , Xn are independent standard Gaussian
Pn
and Y = i=1 Xi2 . Then Y has a χ2 -distribution with n degrees of freedom.
 5.5 Bibliographical Remarks 80

An easy calculation shows that MY (λ) = (1 − 2λ)−n/2 for λ ∈ [0, 1/2) and
MY (λ) is undefined for λ ≥ 1/2. By the Cramér–Chernoff method, we have

P (Y ≥ n + ε) ≤ inf Mλ (Y ) exp(−λ(n + ε))

λ∈[0,1/2)
  n2
1
≤ inf exp(−λ(n + ε))
λ∈[0,1/2) 1 − 2λ

n

Choosing λ = 12 − 2(n+ε)
n
leads to P (Y ≥ n + ε) ≤ 1 + nε 2 exp − 2ε , which
turns out to be about the best you can do [Laurent and Massart, 2000].
6 The subgaussian concept provides a large class of distributions for which
concentration is easily analysed. As mentioned, however, many distributions
are not subgaussian, like the exponential and χ2 -distribution. There are other
general notions based on bounds on the moment generating function that
generalise these kinds of distributions. For more on these ideas, you should
look for keywords subexponential and subgamma.

5.5 Bibliographical Remarks

We return to concentration of measure many times, but note here that it is an

interesting (and still active) topic of research. What we have seen is only the tip
of the iceberg. Readers who want to learn more about this exciting field might
enjoy the book by Boucheron et al. [2013]. For matrix versions of many standard
results, there is a recent book by Tropp [2015]. The survey of McDiarmid [1998]
has many of the classic results. There is a useful type of concentration bound
that are ‘self-normalised’ by the variance. A nice book on this is by de la Peña
et al. [2008]. Another tool that is occasionally useful for deriving concentration
bounds in more unusual set-ups is called empirical process theory. There are
several references for this, including those by van de Geer [2000] or Dudley [2014].

5.6 Exercises

There are too many candidate exercises to list. We heartily recommend all the
exercises in chapter 2 of the book by Boucheron et al. [2013].
5.1 (Variance of average) Let X1 , X2 , . . . , Xn be a sequence of independent
and identically distributed random variables with mean µ and variance σ 2 < ∞.
Pn
Let µ̂ = n1 t=1 Xt and show that V[µ̂] = E[(µ̂ − µ)2 ] = σ 2 /n.

5.2 (Markov’s inequality) Prove Markov’s inequality (Lemma 5.1).

5.3 Prove that the Gaussian tail probability bound on the right-hand side
of Eq. (5.4) is smaller than the bound obtained with Chebyshev’s inequality
 5.6 Exercises 81

Eq. (5.2). In what regime is the improvement most dramatic and in what regime
are both bounds trivial?

5.4 Let X be a random variable on R with density with respect to the Lebesgue
measure of p(x) = |x| exp(−x2 /2)/2. Show the following:
2
(a) P (|X| ≥ ε)
p = exp(−ε /2).
(b) X is not (2 − ε)-subgaussian for any ε > 0.

5.5 (Berry–Esseen inequality) Let X1 , X2 , . . . , Xn be a sequence of

independent and identically distributed random variables with mean µ, variance
σ 2 and bounded third absolute moment:
ρ = E[|X1 − µ|3 ] < ∞ .
Pn
Let Sn = t=1 (Xt − µ)/σ. The Berry–Esseen theorem shows that
 
 
   Z x 
 Sn 1  Cρ
sup P √ ≤ x − √ exp(−y 2 /2)dy  ≤ √ ,
x  n 2π −∞ n
 | {z }
Φ(x)

where C < 1/2 is a universal constant.

Pn
(a) Let µ̂n = n1 t=1 Xt and derive a tail bound from the Berry–Esseen theorem.
That is, give a bound of the form P (µ̂n ≥ µ + ε) for positive values of ε.
(b) Compare your bound with the one that can be obtained from the Cramér–
Chernoff method. Argue pro- and contra- for the superiority of one over the
other.

5.6 (Central limit theorem) We mentioned that invoking the CLT to

approximate the distribution of sums of independent Bernoulli random variables
using a Gaussian can be a bad idea. Let X1 , . . . , Xn ∼ B(p) be independent
Bernoulli random variables with common mean p = pn = λ/n, where λ ∈ (0, 1).
For x ∈ N natural number, let Pn (x) = P (X1 + · · · + Xn = x).

(a) Show that limn→∞ Pn (x) = e−λ λx /(x!), which is a Poisson distribution with
parameter λ.
(b) Explain why this does not contradict the CLT, and discuss the implications
of the Berry–Esseen.
(c) In what way does this show that the CLT is indeed a poor approximation in
some cases?
(d) Based on Monte Carlo simulations, plot the distribution of X1 + · · · + Xn
for n = 30 and some well-chosen values of λ. Compare the distribution to
what you would get from the CLT. What can you conclude?

5.7 (Properties of subgaussian random variables (i)) Prove Lemma 5.4.

Hint Use Taylor series.

 5.6 Exercises 82

5.8 (Properties of subgaussian random variables (ii)) Let Xi be σi -

subgaussian for i ∈ {1, 2} with σi ≥ 0. Prove that X1 +X2 is (σ1 +σ2 )-subgaussian.
Do not assume independence of X1 and X2 .

5.9 (Properties of moment/cumulative-generating functions) Let X

be a real-valued random variable and let MX (λ) = E [exp(λX)] be its moment-
generating function defined over dom(MX ) ⊂ R, where the expectation takes on
finite values. Show that the following properties hold:

(a) MX is convex, and in particular dom(MX ) is an interval containing zero.

(b) MX (λ) ≥ eλE[X] for all λ ∈ dom(MX ).
(c) For any λ in the interior of dom(MX ), MX is infinitely many times
differentiable.
(k)
(d) Let MX (λ)  = dλ dk
k MX (λ). Then, for λ in the interior of dom(MX ),

M (k) (λ) = E X k exp(λX) .
(k)  
(e) Assuming 0 is in the interior of dom(MX ), MX (0) = E X k (hence the
name of MX ).
(f) ψX is convex (that is, MX is log-convex).

Hint For part (a), use the convexity of x 7→ ex .

5.10 (Large deviation theory) Let X, X1 , X2 , . . . , Xn be a sequence of
independent and identically distributed random variables with zero mean and
Pn
moment-generating function MX with dom(MX ) = R. Let µ̂n = n1 t=1 Xt .

(a) Show that for any ε > 0,

1
log P (µ̂n ≥ ε) ≤ −ψX
∗
(ε) = − sup (λε − log MX (λ)) . (5.9)
n λ

(b) Show that when X is a Rademacher variable (P (X = −1) = P (X = 1) =

1/2), ψX∗
(ε) = 1+ε 1−ε
2 log(1 + ε) + 2 log(1 − ε) when |ε| < 1 and ψX (ε) = +∞,
∗

otherwise.
(c) Show that when X is a centered Bernoulli random variable with parameter
p (that is, P (X = −p) = 1 − p and P (X = 1 − p) = p) then ψX ∗
(ε) = ∞
when ε is such that p + ε > 1 and ψX (ε) = d(p + ε, p) otherwise, where
∗

d(p, q) = p log(p/q) + (1 − p) log((1 − p)/(1 − q)) is the relative entropy

between the distributions B(p) and B(q).
(d) Show that when X ∼ N (0, σ 2 ) then ψX ∗
(ε) = ε2 /(2σ 2 ).
(e) Let σ 2 = V[X]. The (strong form of the) central limit theorem says that
  r  
 n 
lim sup P µ̂n
 ≥ x − (1 − Φ(x)) = 0 ,
n→∞ x∈R  σ2
Rx
where Φ(x) = √12π −∞ exp(−y 2 /2)dy is the cumulative distribution of the
 5.6 Exercises 83

standard Gaussian. Let Z be a random variable distributed like a standard

Gaussian. A careless application of this result might suggest that
 r 
1 ? 1 n
lim log P (µ̂n ≥ ε) = lim log P Z ≥ ε .
n→∞ n n→∞ n σ2
Evaluate the right-hand side. In light of the previous parts, what can you
conclude about the validity of the question-marked equality? What goes
wrong with the careless application of the central limit theorem? What do
you conclude about the accuracy of this theorem?

Hint For Part ((e)), consider using Eq. (13.4).

As it happens, the inequality in (5.9) may be replaced by an equality

as n → ∞. The assumption that the moment-generating function exists
everywhere may be relaxed significantly. We refer the interested reader to
the classic text by Dembo and Zeitouni [2009]. The function ψX ∗
is called the
Legendre transform, convex conjugate or Fenchel dual of the convex
function ψX . In probability theory, ψX
∗
is also called the Cramér transform
and is also known as a rate function. Convexity and the Fenchel dual will
play a role in some of the later chapters and will be discussed in more detail
in Chapter 26 and later.

The name “large deviation” originates from rewriting the tail probabilities in
terms of the partial sum Sn = X1 +· · ·+Xn , we see that the inequality in (5.9)
bounds the probability of the deviation of Sn from its mean (which is zero by
assumption) at a scale of Θ(n): P (µ̂n ≥ ε) = P (Sn ≥ nε). In contrast, the
central-limit theorem (CLT) gives the (limiting) probability of the deviation
√ √ √
of Sn from its mean at the scales of Θ( n): P (µ̂n n ≥ ε) = P (Sn ≥ nε).
√
Compared to nε, nε is thought of as a “large” deviation. The deviation
probabilities at this scale can decay to zero faster than what the CLT
predicts, as also showcased in the last part of the last exercise. But what
happens at intermediate scales? That is, when deviations are of size nα ε with
1/2 < α < n? This is studied on the formulaic name of moderate deviations.
As it turns out, in this case, the ruthless use of the large deviation formula
gives correct answers. The reader who wants to learn more about large
deviation theory can check out the lecture notes by Swart [2017].

5.11 (Hoeffding’s lemma) Suppose that X is zero mean and X ∈ [a, b] almost
surely for constants a < b.

(a) Show that X is (b − a)/2-subgaussian.

(b) Prove Hoeffding’s inequality (5.8).
 5.6 Exercises 84

Hint For part (a), it suffices to prove that ψX (λ) ≤ λ2 (b − a)2 /4. By Taylor’s
theorem, for some λ0 between 0 and λ, ψX (λ) = ψX (0) + ψX 0
(0)λ + ψX00
(λ0 )λ2 /2.
To bound the last term, introduce the distribution Pλ for λ ∈ R arbitrary:
Pλ (dz) = e−ψX (λ) eλz P (dz). Show that Ψ00X (λ) = V[Z], where Z ∼ Pλ . Now,
since Z ∈ [a, b] with probability one, argue (without relying on E [Z]) that
V[Z] ≤ (b − a)2 /4.

5.12 (Subgaussianity of Bernoulli distribution) Let X be a random

variable with Bernoulli distribution with mean p. That is X ∼ B(p): P (X = 1) = p
and P (X = 0) = 1 − p.

(a) Show that X is 1/2-subgaussian for all p.

q
1−2p
(b) Let Q : [0, 1] → [0, 1/2] be the function given by Q(p) = 2 ln((1−p)/p)
where undefined points are defined in terms of their limits. Show that X is
Q(p)-subgaussian.
(c) The subgaussianity constant of a random variable X is the smallest value of
σ such that X is σ-subgaussian. Show that the subgaussianity constant of
X ∼ B(p) is Q(p).
p p
(d) Plot Q(p) as a function of p. How does it compare to V[X] = p(1 − p)?
(e) Show that for λ ≥ 0 and p ≥ 1/2, Eexp(λX) ≤ exp(p(1 − p)λ2 /2). Think of
how these inequalities are used for bounding tails. What do you conclude?

Readers looking for a hint to parts (b), (c) and (e) in the previous exercise
might like to look at the papers by Berend and Kontorovich [2013] and
Ostrovsky and Sirota [2014]. The result that the subgaussianity constant of
X ∼ B(p) is upper bounded by Q(p) is known as the Kearn-Saul inequality
and is due to Kearns and Saul [1998].

5.13 (Central limit theorem for sums of Bernoulli random variables)

In this question we try to understand the concentration of the empirical mean
for Bernoulli random variables. Let X1 , X2 , . . . , Xn be independent Bernoulli
Pn
random variables with mean p ∈ [0, 1] and p̂n = t=1 Xt /n. Let Zn be normally
distributed random variable with mean p and variance p(1 − p)/n.

(a) Write down expressions for E[p̂n ] and V[p̂n ].

(b) What does the central limit theorem say about the relationship between p̂n
and Zn as n gets large?
(c) For each p ∈ {1/10, 1/2} and δ = 1/100 and ∆ = 1/10, find the minimum n
such that P (p̂n ≥ p + ∆) ≤ δ.
(d) Let p = 1/10 and ∆ = 1/10 and

nBer (δ, p, ∆) = min {n : P (p̂n ≥ p + ∆) ≤ δ} ,

nGauss (δ, p, ∆) = min {n : P (Zn ≥ p + ∆) ≤ δ} .
 5.6 Exercises 85

(i) Evaluate analytically the value of

nBer (δ, 1/10, 1/10)

lim .
δ→0 nGauss (δ, 1/10, 1/10)

(ii) In light of the central limit theorem, explain why the answer you got in (i)
was not 1.

Hint For Part (d.i) use large deviation theory (Exercise 5.10).
5.14 (Bernstein’s inequality) Let X1 , . . . , Xn be a sequence of independent
Pn
random variables with Xt − E[Xt ] ≤ b almost surely and S = t=1 (Xt − E[Xt ])
Pn
and v = t=1 V[Xt ].

1 x2
(a) Show that g(x) = 2 + x
3! + 4! + · · · = (exp(x) − 1 − x)/x2 is increasing.
(b) Let X be a random variable with E[X] = 0 and X ≤ b almost surely. Show
that E[exp(X)] ≤ 1 + g(b)V[X].
2
3α
(c) Prove that (1 + α) log(1 + α) − α ≥ 6+2α for all α ≥ 0. Prove that this is
the best possible approximation in the sense that the 2 in the denominator
cannot be increased.
(d) Let ε > 0 and α = bε/v and prove that
 v 
P (S ≥ ε) ≤ exp − 2 ((1 + α) log(1 + α) − α) (5.10)
b !
ε2
≤ exp −
. (5.11)
2v 1 + 3v
bε

(e) Use the previous result to show that

s    !
1 2b 1
P S≥ 2v log + log ≤ δ.
δ 3 δ

(f) Let X1 , X2 , . . . , Xn be a sequence of random variables adapted to filtration

F = (Ft )t . Abbreviate Et [·] = E[· | Ft ] and µt = Et−1 [Xt ]. Define S =
Pn Pn
X − µt and let V = t=1 Et−1 [(Xt − µt )2 ] be the predictable variation
Pp t
t=1
of ( t=1 Xt − µt )p . Show that if Xt − µt ≤ b holds almost surely for all
t ∈ [n] then with α = bε/v,
 v 
P (S ≥ ε, V ≤ v) ≤ exp − 2 ((1 + α) log(1 + α) − α) .
b

Note that the right-hand side of this inequality is the same as that shown in
Eq. (5.10).
 5.6 Exercises 86

The bound in Eq. (5.10) is called Bennett’s inequality and the one
in Eq. (5.11) is called Bernstein’s inequality. There are several
generalisations, the most notable of which is the martingale version that
slightly relaxes the independence assumption and which was presented in
Part (f). Martingale techniques appear in Chapter 20. Another useful variant
(under slightly different conditions) replaces the actual variance with the
empirical variance. This is useful when the variance is unknown. For more,
see the papers by Audibert et al. [2007], Mnih et al. [2008], Maurer and
Pontil [2009].

5.15 (Another Bernstein-type inequality) Let X1 , X2 , . . . , Xn be a

sequence of random variables adapted to the filtration F = (Ft )t . Abbreviate
Et [·] = E[· | Ft ] and µt = Et−1 [Xt ]. Prove the following

(a) If η > 0 and η(Xt − µt ) ≤ 1 almost surely, then

 !
1 1
Xn Xn
2
P (Xt − µt ) ≥ η Et−1 [(Xt − µt ) ] + log ≤ δ.
t=1 t=1
η δ

(b) If η > 0 and ηXt ≤ 1 almost surely, then

 !
1 1
Xn Xn
2
P (Xt − µt ) ≥ η Et−1 [Xt ] + log ≤ δ.
t=1 t=1
η δ

Hint Use the Cramér–Chernoff method and the fact that exp(x) ≤ 1 + x + x2
for all x ≤ 1 and exp(x) ≥ 1 + x for all x.

Pt
Let (Mt ) be the martingale defined by Mt = s=1 (Xs −µs ). The inequalities
in Exercise 5.15 can be viewed as a kind of Bernstein’s inequality because
they bound the tail of the martingale (Mt ) in terms of the predictable
Pn 2
variation of the martingale (Mt ), which is V = t=1 Et−1 [(Xt − µt ) ].
The main difference relative to well-known results is that the analysis has
stopped early. The next step is usually to choose η to minimise the bound
in some sense. Either by assuming bounds on the predictable variation,
union bounding or using the method of mixtures [de la Peña et al., 2008].
These techniques are covered in Chapter 20. Note, optimising η directly is
not possible because the bounds hold for any fixed η, but minimising the
right-hand side inside the probability with respect to η would lead to a
random η. For more martingale results with this flavour, see the notes by
McDiarmid [1998].

5.16 Let X1 , . . . , Xn be independent random variables with P (Xt ≤ x) ≤ x for

 5.6 Exercises 87

each x ∈ [0, 1] and t ∈ [n]. Prove for any ε > 0 that

!
Xn  ε n
P log(1/Xt ) ≥ ε ≤ exp(n − ε) .
t=1
n

5.17 (Concentration for categorical distributions) Let X1 , . . . , Xn be

an independent and identically distributed sequence taking values in [m]. For
Pn
i ∈ [m], let p(i) = P (X1 = i) and p̂(i) = n1 t=1 I {Xt = i}. Show that for any
δ ∈ (0, 1),
 s 

2 log 1δ + m log(2)
P kp − p̂k1 ≥  ≤ δ. (5.12)
n

Hint Use the fact that kp − p̂k1 = maxλ∈{−1,1}m hλ, p − p̂i.

5.18 (Expectation of maximum) Let X1 , . . . , Xn be a sequence of σ-
subgaussian random variables (possibly dependent) and Z = maxt∈[n] Xt . Prove
that
p
 ≤ p2σ log(n). 
(a) E[Z] 2

(b) P Z ≥ 2σ 2 log(n/δ) ≤ δ for any δ ∈ (0, 1).

Hint Use Jensen’s inequality to show that exp(λE[Z]) ≤ E[exp(λZ)], and then
provide a naive bound on the moment-generating function of Z.
5.19 (Almost surely bounded sums) Let X1 , X2 , . . . , Xn be a sequence of non-
Pn
negative random variables adapted to filtration (Ft )nt=0 such that t=1 Xt ≤ 1
almost surely. Prove that for all x > 1,
!  
X n  n−x n−1 , if x < n ;
.
P E[Xt | Ft−1 ] ≥ x ≤ fn (x) = n−1
0 , if x ≥ n ,
t=1

where the equality serves as the definition of fn (x).

Hint This problem does not use the techniques introduced in the chapter.
Prove that Bernoulli random variables are the worst case and use backwards
induction. Although this result is new to our knowledge, a weaker version was
derived by Kirschner and Krause [2018] for the analysis of information-directed
sampling. The bound is tight in the sense that there exists a sequence of random
variables and filtration for which equality holds.
Part II
Stochastic Bandits with
Finitely Many Arms
 89

Over the next few chapters, we introduce the fundamental algorithms and
tools of analysis for unstructured stochastic bandits with finitely many actions.
The keywords here are finite, unstructured and stochastic. The first of these
just means that the number of actions available is finite. The second is more
ambiguous, but roughly means that choosing one action yields no information
about the mean pay-off of the other arms. A bandit is stochastic if the sequence
of rewards associated with each action is independent and identically distributed
according to some distribution. This latter assumption will be relaxed in Part III.
There are several reasons to study this class of bandit problems. First, their
simplicity makes them relatively easy to analyse and permits a deep understanding
of the trade-off between exploration and exploitation. Second, many of the
algorithms designed for finite-armed bandits, and the principle underlying them,
can be generalised to other settings. Finally, finite-armed bandits already have
applications – notably as a replacement to A/B testing, as discussed in the
introduction.
6 The Explore-Then-Commit
Algorithm

The first bandit algorithm of the book is called explore-then-commit (ETC),

which explores by playing each arm a fixed number of times and then exploits by
committing to the arm that appeared best during exploration.

For this chapter, as well as Chapters 7 to 9, we assume that all bandit

instances are in ESG
k
(1), which means the reward distribution for all arms is
1-subgaussian.

The focus on subgaussian distributions is mainly for simplicity. Many of the

techniques in the chapters that follow can be applied to other stochastic bandits
such as those listed in Table 4.1. The key difference is that new concentration
analysis is required that exploits the different assumptions. The Bernoulli case is
covered in Chapter 10, where other situations are discussed along with references
to the literature. Notice that the subgaussian assumption restricts the subgaussian
constant to σ = 1, which saves us from endlessly writing σ. All results hold for
other subgaussian constants by scaling the rewards (see Lemma 5.4). Two points
are obscured by this simplification:

(a) All the algorithms that follow rely on the knowledge of σ.

(b) It may happen that Pi is subgaussian for all arms, but with a different
subgaussian constant for each arm. Algorithms are easily adapted to this
situation if the subgaussian constants are known, as you will investigate
in Exercise 7.2. The situation is more complicated when the subgaussian
constant is unknown (Exercise 7.7).

6.1 Algorithm and Regret Analysis

ETC is characterised by the number of times it explores each arm, denoted by a

natural number m. Because there are k actions, the algorithm will explore for mk
rounds before choosing a single action for the remaining rounds. Let µ̂i (t) be the
 6.1 Algorithm and Regret Analysis 91

average reward received from arm i after round t, which is written formally as

1 X
t
µ̂i (t) = I {As = i} Xs ,
Ti (t) s=1
Pt
where Ti (t) = s=1 I {As = i} is the number of times action i has been played
after round t. The ETC policy is given in Algorithm 1 below.

1: Input m.
2: In round t choose action
(
(t mod k) + 1 , if t ≤ mk ;
At =
argmaxi µ̂i (mk) , t > mk .
(ties in the argmax are broken arbitrarily)
Algorithm 1: Explore-then-commit.

Recall that µi is the mean reward when playing action i and ∆i = µ∗ − µi is

suboptimality gap between the mean of action i and the optimal action.

Theorem 6.1. When ETC is interacting with any 1-subgaussian bandit and
1 ≤ m ≤ n/k,
 
m∆2i
k
X k
X
Rn ≤ m ∆i + (n − mk) ∆i exp − .
i=1 i=1
4

Proof Assume without loss of generality that the first arm is optimal, which
means that µ1 = µ∗ = maxi µi . By the decomposition given in Lemma 4.5, the
regret can be written as
k
X
Rn = ∆i E [Ti (n)] . (6.1)
i=1

In the first mk rounds, the policy is deterministic, choosing each action exactly
m times. Subsequently it chooses a single action maximising the average reward
during exploration. Thus,

E [Ti (n)] = m + (n − mk)P (Amk+1 = i)

 
≤ m + (n − mk)P µ̂i (mk) ≥ max µ̂j (mk) . (6.2)
j6=i

The probability on the right-hand side is bounded by

 
P µ̂i (mk) ≥ max µ̂j (mk) ≤ P (µ̂i (mk) ≥ µ̂1 (mk))
j6=i

= P (µ̂i (mk) − µi − (µ̂1 (mk) − µ1 ) ≥ ∆i ) .

p
The next step is to check that µ̂i (mk) − µi − (µ̂1 (mk) − µ1 ) is 2/m-subgaussian,
 6.1 Algorithm and Regret Analysis 92

which by the properties of subgaussian random variables follows from the

definitions of (µ̂j )j and the algorithm. Hence by Corollary 5.5,
 
m∆2i
P (µ̂i (mk) − µi − µ̂1 (mk) + µ1 ≥ ∆i ) ≤ exp − . (6.3)
4
Substituting Eq. (6.3) into Eq. (6.2) and the regret decomposition (Eq. (6.1))
gives the result.
The bound in Theorem 6.1 illustrates the trade-off between exploration and
exploitation. If m is large, then the policy explores for too long, and the first
term will be large. On the other hand, if m is too small, then the probability
that the algorithm commits to the wrong arm will grow, and the second term
becomes large. The question is how to choose m. Assume that k = 2 and that
the first arm is optimal so that ∆1 = 0, and abbreviate ∆ = ∆2 . Then the bound
in Theorem 6.1 simplifies to
   
m∆2 m∆2
Rn ≤ m∆ + (n − 2m)∆ exp − ≤ m∆ + n∆ exp − . (6.4)
4 4
For large n the quantity on the right-hand side of Eq. (6.4) is minimised up to a
possible rounding error by
   
4 n∆2
m = max 1, log , (6.5)
∆2 4
and for this choice and any n, the regret is bounded by
    
4 n∆2
Rn ≤ min n∆, ∆ + 1 + max 0, log . (6.6)
∆ 4
In Exercise 6.2 you will show that Eq. (6.6) implies that
√
Rn ≤ ∆ + C n , (6.7)
where C > 0 is a universal constant. In particular, when ∆ ≤ 1 as is often
assumed, we get
√
Rn ≤ 1 + C n ,
Bounds of this type are called worst-case, problem free or problem
independent (see Eq. (4.2) or Eq. (4.3)). The reason is that the bound only
depends on the horizon and class of bandits for which the algorithm is designed,
and not the specific instance within that class. Because the suboptimality gap does
not appear, bounds like this are sometimes called gap-free. In contrast, bounds
like the one in Eq. (6.6) are called gap/problem/distribution/instance
dependent.
Note that without the condition ∆ ≤ 1, the worst-case bound for ETC is
infinite. In fact, without a bound on the reward range, the worst-case bound of
all reasonable algorithms (that try each action at least once) will also be infinite.
With the understanding that Eq. (6.7) gives rise to a meaningful worst-case
 6.1 Algorithm and Regret Analysis 93

bound for bandits with bounded reward range, we take the liberty and will also
call bounds like that in Eq. (6.7) a worst-case bound.
The bound in (6.6) is close to optimal (see Part IV), but there is a caveat. The
choice of m that defines the policy and leads to this bound depends on both the
suboptimality gap and the horizon. While the horizon is sometimes known in
advance, it is seldom reasonable to assume knowledge of the suboptimality gap.
You will show in Exercise 6.5 that there is a choice of m depending only on n, for
which Rn = O(n2/3 ) regardless of the value of ∆. Alternatively, the number of
plays before commitment can be made data dependent, which means the learner
plays arms alternately until it decides based on its observations to commit to
a single arm for the remainder (Exercise 6.5). ETC also has the property that
its immediate expected regret per time step is monotonically decreasing as time
goes by, though not in a nice smooth fashion. This monotone decreasing property
is a highly desirable property. In later chapters we will see policies where the
decrease is smoother.

Experiment 6.1 Fig. 6.1 shows the expected regret of ETC when playing a
Gaussian bandit with k = 2 and means µ1 = 0 and µ2 = −∆. The horizon is set
to n = 1000, and the suboptimality gap ∆ is varied between 0 and 1. Each data
point is the average of 105 simulations, which makes the error bars invisible. The
results show that the theoretical upper bound provided by Theorem 6.1 is quite
close to the actual performance.

Upper bound in (6.6)

80 ETC with m in (6.5)
Expected regret

60

40

20

0 0.2 0.4 0.6 0.8 1

∆

Figure 6.1 The expected regret of ETC and the upper bound in Eq. (6.6).
 6.2 Notes 94

6.2 Notes

1 An algorithm is called anytime if it does not require advance knowledge of

the horizon n. ETC is not anytime because the choice of commitment time
depends on the horizon. This limitation can be addressed by the doubling
trick, which is a simple way to convert a horizon-dependent algorithm into
an anytime algorithm (Exercise 6.6).
2 By allowing the exploration time m to be a data-dependent random variable,
it is possible to recover near-optimal regret without knowing the suboptimality
gap. For more details see Exercise 6.5. Another idea is to use an elimination
algorithm that acts in phases and eliminates arms using increasingly sensitive
hypothesis tests (Exercise 6.8). Elimination algorithms are often easy to analyse
and can work well in practice, but they also have inherent limitations, just like
ETC algorithms, as will be commented on later.
3 The ε-greedy algorithm is a randomised relative of ETC that in round t
plays the empirically best arm with probability 1 − εt and otherwise explores
uniformly at random. You will analyse this algorithm in Exercise 6.7.

6.3 Bibliographical Remarks

ETC has a long history. Robbins [1952] considered ‘certainty equivalence with
forcing’, which chooses the arm with the largest sample mean except at a fixed
set of times Ti ⊂ N when arm i is chosen for i ∈ [k]. By choosing the set
of times carefully, it is shown that this policy enjoys sublinear regret. While
ETC performs all the exploration at the beginning, Robbins’s policy spreads
the exploration over time. This is advantageous if the horizon is not known,
but disadvantageous otherwise. Anscombe [1963] considered exploration and
commitment in the context of medical trials or other experimental set-ups. He
already largely solves the problem in the Gaussian case and highlights many of
the important considerations. Besides this, the article is beautifully written and
well worth reading. Strategies based on exploration and commitment are simple
to implement and analyse. They can also generalise well to more complex settings.
For example, Langford and Zhang [2008] consider this style of policy under the
name ‘epoch-greedy’ for contextual bandits (the idea of exploring then exploiting
in epochs, or intervals, is essentially what Robbins [1952] suggested). We’ll return
to contextual bandits in Chapter 18. Abbasi-Yadkori et al. [2009], Abbasi-Yadkori
[2009b] and Rusmevichientong and Tsitsiklis [2010] consider ETC-style policies
under the respective names of ‘forced exploration’ and ‘phased exploration and
greedy exploitation’ (PEGE) in the context of linear bandits (which we shall meet
in Chapter 19). Other names include ‘forced sampling’, ‘explore-first’, ‘explore-
then-exploit’. Garivier et al. [2016b] have shown that ETC policies are necessarily
suboptimal in the limit of infinite data in a way that is made precise in Chapter 16.
This comment also applies to elimination-based strategies, which are described in
 6.4 Exercises 95

Exercise 6.8. The history of ε-greedy is unclear, but it is a popular and widely used
and known algorithm in reinforcement learning [Sutton and Barto, 1998]. Auer
et al. [2002a] analyse the regret of ε-greedy with slowly decreasing exploration
probabilities. There are other kinds of randomised exploration as well, including
Thompson sampling [1933] and Boltzmann exploration analysed recently by
Cesa-Bianchi et al. [2017].

6.4 Exercises

6.1 (Subgaussian empirical estimates) Let π be the policy of ETC and

P1 , . . . , Pk be the 1-subgaussian distributions associated with the k arms. Provide
a fully rigourous proof of the claim that

µ̂i (mk) − µi − µ̂1 (mk) + µ1

p
is 2/m-subgaussian. You should only use the definitions and the interaction
protocol, which states that

(a) P (At ∈ · | A1 , X1 , . . . , At−1 , Xt−1 ) = π(· | A1 , X1 , . . . , At−1 , Xt−1 ) a.s.

(b) P (Xt ∈ · | A1 , X1 , . . . , At−1 , Xt−1 , At ) = PAt (·) a.s.

6.2 (Minimax regret) Show that Eq. (6.6) implies the regret of an optimally
√
tuned ETC for subgaussian two-armed bandits satisfies Rn ≤ ∆ + C n where
C > 0 is a universal constant.

6.3 (High-probability bounds (i)) Assume that k = 2, and let δ ∈ (0, 1).
Modify the ETC algorithm to depend on δ and prove a bound on the pseudo-
Pn
regret R̄n = nµ∗ − t=1 µAt of ETC that holds with probability 1 − δ. The
algorithm is allowed to use the action suboptimality gaps.

6.4 (High-probability bounds (ii)) Repeat the previous exercise, but now
Pn
prove a high probability bound on the random regret: R̂n = nµ∗ − t=1 Xt .
Compare this to the bound derived for the pseudo-regret in the previous exercise.
What can you conclude?

6.5 (Adaptive commitment times) Suppose that ETC interacts with a two-
armed 1-subgaussian bandit ν ∈ E with means µ1 , µ2 ∈ R and ∆ν = |µ1 − µ2 |.

(a) Find a choice of m that only depends on the horizon n and not ∆ such that
there exists a constant C > 0 such that for any n and for any ν ∈ E, the
regret Rn (ν) of Algorithm 1 is bounded by

Rn (ν) ≤ (∆ν + C)n2/3 .

Furthermore, show that there is no C > 0 such that for any problem instance
ν and n ≥ 1, Rn (ν) ≤ ∆ν + Cn2/3 holds.
 6.4 Exercises 96

(b) Now suppose the commitment time is allowed to be data dependent, which
means the algorithm explores each arm alternately until some condition is
met and then commits to a single arm for the remainder. Design a condition
such that the regret of the resulting algorithm can be bounded by
C log n
Rn (ν) ≤ ∆ν + , (6.8)
∆ν
where C is a universal constant. Your condition should only depend on the
observed rewards and the time horizon. It should not depend on µ1 , µ2 or
∆ν .
(c) Show that
p any algorithm for which (6.8) holds also satisfies Rn (ν) ≤
∆ν + C n log(n) for any n ≥ 1 and ν ∈ E and a suitably chosen universal
constant C > 0.
(d) As for (b), but now the objective is to design a condition such that for any
n ≥ 1 and ν ∈ E, the regret of the resulting algorithm is bounded by

C log max e, n∆2ν
Rn (ν) ≤ ∆ν + . (6.9)
∆ν

(e) Show that any algorithm for which (6.9) holds also satisfies that for any
√
n ≥ 1 and ν ∈ E, Rn (ν) ≤ ∆ν + C n for suitably chosen universal constant
C > 0.

Hint For (a) start from Rn ≤ m∆ + n∆ exp(−m∆2 /2) and show an upper
bound on the second term which is independent of ∆. Then, choose m. For
(b) think about the simplest stopping policy and then make it robust by using
confidence intervals. Tune the failure probability. For (c) note that the regret
can never be larger than n∆.
6.6 (Doubling trick) The purpose of this exercise is to analyse a meta-
algorithm based on the so-called doubling trick that converts a policy depending
on the horizon to a policy with similar guarantees that does not. Let E be an
arbitrary set of bandits. Suppose you are given a policy π = π(n) designed for E
that accepts the horizon n as a parameter and has a regret guarantee of

max Rt (π(n), ν) ≤ fn (ν) , ∀ν ∈ E ,

1≤t≤n

where fn : E → [0, ∞) is a sequence of functions. Let n1 < n2 < n3 < · · · be a

fixed sequence of integers and consider the policy that runs π with horizon n1
until round t = min{n, n1 }, then runs π with horizon n2 until t = min{n, n1 +n2 },
and then restarts again with horizon n3 until t = min{n, n1 + n2 + n3 } and so-on.
Note that t is the real-time counter and is not reset on each restart. Let π ∗ be the
resulting policy. When n`+1 = 2n` , the length of periods when π is used double
with each phase, hence the name ‘doubling trick’.
P`
(a) Let n > 0 be arbitrary, `max = min{` : i=1 ni ≥ n}. Prove that for any
 6.4 Exercises 97

ν ∈ E, the n-horizon regret of π ∗ on ν is at most

`X
max

Rn (π ∗ , ν) ≤ fn` (ν) . (6.10)

`=1
√
(b) Suppose that fn (ν) ≤ n. Show that if n` = 2`−1 , then for any ν ∈ E and
horizon n the regret of π ∗ is at most
√
Rn (π ∗ , ν) ≤ C n ,
where C > 0 is a carefully chosen universal constant.
(c) Suppose that fn (ν) = g(ν) log(n) for some function g : E → [0, ∞). What is
the regret of π ∗ if n` = 2`−1 ? Can you find a better choice of (n` )` ?
(d) In light of this idea, should we bother trying to design algorithms that do not
depend on the horizon? Are there any disadvantages to using the doubling
trick? If so, what are they? Write a short summary of the pros and cons of
the doubling trick.

According to Besson and Kaufmann [2018], the doubling trick was first
applied to bandits by Auer et al. [1995]. Note, nowhere in this exercise did
we use that the bandit is stochastic. Nothing changes in the adversarial or
contextual settings studied later in the book.

6.7 (ε-greedy) For this exercise assume the rewards are 1-subgaussian and
there are k ≥ 2 arms. The ε-greedy algorithm depends on a sequence of
parameters ε1 , ε2 , . . .. First it chooses each arm once and subsequently chooses
At = argmaxi µ̂i (t − 1) with probability 1 − εt and otherwise chooses an arm
uniformly at random.
k
Rn εX
(a) Prove that if εt = ε > 0, then lim = ∆i .
n→∞ n k i=1
n o
(b) Let ∆min = min {∆i : ∆i > 0} and let εt = min 1, t∆Ck 2 , where C > 0 is
min
a sufficiently large universal constant. Prove that there exists a universal
C 0 > 0 such that
k   
X ∆i n∆2min
Rn ≤ C 0 ∆i + 2 log max e, .
i=1
∆min k

6.8 (Elimination algorithm) A simple way to generalise the ETC policy to

multiple arms and overcome the problem of tuning the commitment time is to
use an elimination algorithm. The algorithm operates in phases and maintains
an active set of arms that could be optimal. In the `th phase, the algorithm aims
to eliminate from the active set all arms i for which ∆i ≥ 2−` .
Without loss of generality, assume that arm 1 is an optimal arm. You may
assume that the horizon n is known.
 6.4 Exercises 98

1: Input: k and sequence (m` )`

2: A1 = {1, 2, . . . , k}
3: for ` = 1, 2, 3, . . . do
4: Choose each arm i ∈ A` exactly m` times
5: Let µ̂i,` be the average reward for arm i from this phase only
6: Update active set:
 
A`+1 = i : µ̂i,` + 2−` ≥ max µ̂j,`
j∈A`

7: end for
Algorithm 2: Phased elimination for finite-armed bandits

(a) Show that for any ` ≥ 1,

 
m` 2−2`
P (1 ∈
/ A`+1 , 1 ∈ A` ) ≤ k exp − .
4

(b) Show that if i ∈ [k] and ` ≥ 1 are such that ∆i ≥ 2−` , then
 
m` (∆i − 2−` )2
P (i ∈ A`+1 , 1 ∈ A` , i ∈ A` ) ≤ exp − .
4

(c) Let `i = min ` ≥ 1 : 2−` ≤ ∆i /2 . Choose m` in such a way that
P (exists ` : 1 ∈
/ A` ) ≤ 1/n and P (i ∈ A`i +1 ) ≤ 1/n.
(d) Show that your algorithm has regret at most
X  1

Rn ≤ C ∆i + log(n) ,
∆i
i:∆i >0

where C > 0 is a carefully chosen universal constant.

(e) Modify your choice of m` and show that the regret of the resulting algorithm
satisfies
X  1 

Rn ≤ C ∆i + log max e, n∆2i .
∆i
i:∆i >0

(f) Show that with an appropriate universal constant C 0 > 0, the regret satisfies
X p
Rn ≤ ∆i + C 0 nk log(k) .
i

Algorithm 2 is due to Auer and Ortner [2010]. The log(k) term in Part (f) can
be removed by modifying the algorithm to use the refined confidence intervals
in Chapter 9, but we would not recommend this for the reasons discussed
in Section 9.2 of that chapter. You could also use a more sophisticated
confidence level [Lattimore, 2018].
 6.4 Exercises 99

ETC

70

Expected regret 60

50

0 100 200 300 400

m

Figure 6.2 Expected regret for ETC over 105 trials on a Gaussian bandit with means
µ1 = 0, µ2 = −1/10

6.9 (Empirical study) In this exercise you will investigate the empirical
behaviour of ETC on a two-armed Gaussian bandit with means µ1 = 0 and
µ2 = −∆. Let
n
X
R̄n = ∆A t ,
t=1

which is chosen so that Rn = E[R̄n ]. Complete the following:

(a) Using programming language of your choice, write a function that accepts
an integer n and ∆ > 0 and returns the value of m that exactly minimises
the expected regret.
(b) Reproduce Fig. 6.1.
(c) Fix ∆ = 1/10 and plot the expected regret as a function of m with n = 2000.
Your plot should resemble Fig. 6.2.
(d) Plot the standard deviation V[R̄n ]1/2 as a function of m for the same bandit
as above. Your plot should resemble Fig. 6.3.
(e) Explain the shape of the curves you observed in Parts (b), (c) and (d) and
reconcile what you see with the theoretical results.
(f) Think, experiment and plot. Is it justified to plot V[R̄n ]1/2 as a summary of
how R̄n is distributed? Explain your thinking.
 6.4 Exercises 100

100
ETC
Standard deviation of the regret

80

60

40

0 100 200 300 400

m

Figure 6.3 Standard deviation of the regret for ETC over 105 trials on a Gaussian bandit
with means µ1 = 0, µ2 = −1/10
7 The Upper Confidence Bound
Algorithm

The upper confidence bound (UCB) algorithm offers several advantages over the
explore-then-commit (ETC) algorithm introduced in the last chapter.

(a) It does not depend on advance knowledge of the suboptimality gaps.

(b) It behaves well when there are more than two arms.
(c) The version introduced here depends on the horizon n, but in the next
chapter, we will see how to eliminate that as well.

The algorithm has many different forms, depending on the distributional

assumptions on the noise. Like in the previous chapter, we assume the noise is
1-subgaussian. A serious discussion of other options is delayed until Chapter 10.

7.1 The Optimism Principle

The UCB algorithm is based on the principle of optimism in the face of

uncertainty, which states that one should act as if the environment is as nice as
plausibly possible. As we shall see in later chapters, the principle is applicable
beyond the finite-armed stochastic bandit problem.
Imagine visiting a new country and making a choice between sampling the local
cuisine or visiting a well-known multinational chain. Taking an optimistic view of
the unknown local cuisine leads to exploration because without data, it could be
amazing. After trying the new option a few times, you can update your statistics
and make a more informed decision. On the other hand, taking a pessimistic
view of the new option discourages exploration, and you may suffer significant
regret if the local options are delicious. Just how optimistic you should be is a
difficult decision, which we explore for the rest of the chapter in the context of
finite-armed bandits.
For bandits, the optimism principle means using the data observed so far to
assign to each arm a value, called the upper confidence bound that with high
probability is an overestimate of the unknown mean. The intuitive reason why
this leads to sublinear regret is simple. Assuming the upper confidence bound
assigned to the optimal arm is indeed an overestimate, then another arm can only
be played if its upper confidence bound is larger than that of the optimal arm,
which in turn is larger than the mean of the optimal arm. And yet this cannot
 7.1 The Optimism Principle 102

happen too often because the additional data provided by playing a suboptimal
arm means that the upper confidence bound for this arm will eventually fall
below that of the optimal arm.
In order to make this argument more precise, we need to define the upper
confidence bound. Let (Xt )nt=1 be a sequence of independent 1-subgaussian random
Pn
variables with mean µ and µ̂ = n1 t=1 Xt . By Eq. (5.6),
r !
2 log(1/δ)
P µ ≥ µ̂ + ≤δ for all δ ∈ (0, 1) . (7.1)
n

When considering its options in round t, the learner has observed Ti (t − 1)

samples from arm i and received rewards from that arm with an empirical mean
of µ̂i (t − 1). Then a reasonable candidate for ‘as large as plausibly possible’ for
the unknown mean of the ith arm is

∞ if Ti (t − 1) = 0
UCBi (t − 1, δ) = q
2 log(1/δ)
(7.2)
µ̂i (t − 1) + otherwise .
Ti (t−1)

Great care is required when comparing (7.1) and (7.2) because in the former the
number of samples is the constant n, but in the latter it is a random variable
Ti (t − 1). By and large, however, this is merely an annoying technicality, and the
intuition remains that δ is approximately an upper bound on the probability of
the event that the above quantity is an underestimate of the true mean. More
details are given in Exercise 7.1.
At last we have everything we need to state a version of the UCB algorithm,
which takes as input the number of arms and the error probability δ.

1: Input k and δ
2: for t ∈ 1, . . . , n do
3: Choose action At = argmaxi UCBi (t − 1, δ)
4: Observe reward Xt and update upper confidence bounds
5: end for
Algorithm 3: UCB(δ).

Although there are many versions of the UCB algorithm, we often do not
distinguish them by name and hope the context is clear. For the rest of this
chapter, we’ll usually call UCB(δ) just UCB.

The value inside the argmax is called the index of arm i. Generally speaking,
an index algorithm chooses the arm in each round that maximises some value
(the index), which usually only depends on the current time step and the samples
from that arm. In the case of UCB, the index is the sum of the empirical mean
 7.1 The Optimism Principle 103

of rewards experienced so far and the exploration bonus, which is also known
as the confidence width.
Besides the slightly vague ‘optimism guarantees optimality or learning’ intuition
we gave before, it is worth exploring other intuitions for the choice of index. At
a very basic level, an algorithm should explore arms more often if they are (a)
promising because µ̂i (t − 1) is large or (b) not well explored because Ti (t − 1) is
small. As one can plainly see, the definition in Eq. (7.2) exhibits this behaviour.
This explanation is not completely satisfying, however, because it does not explain
why the form of the functions is just so.
A more refined explanation comes from thinking of what we expect of any
reasonable algorithm. Suppose at the start of round t the first arm has been
played much more frequently than the rest. If we did a good job designing our
algorithm, we would hope this is the optimal arm, and because it has been played
so often, we expect that µ̂1 (t − 1) ≈ µ1 . To confirm the hypothesis that arm 1 is
optimal, the algorithm had better be highly confident that other arms are indeed
worse. This leads quite naturally to the idea of using upper confidence bounds.
The learner can be reasonably certain that arm i is worse than arm 1 if
s s
2 log(1/δ) 2 log(1/δ)
µ̂i (t − 1) + ≤ µ1 ≈ µ̂1 (t − 1) + , (7.3)
Ti (t − 1) T1 (t − 1)

where δ is called the confidence level and quantifies the degree of certainty.
This means that choosing the arm with the largest upper confidence bound leads
to a situation where arms are only chosen if their true mean could reasonably be
larger than those of arms that have been played often. That this rule is indeed a
good one depends on two factors. The first is whether the width of the confidence
interval at a given confidence level can be significantly decreased, and the second
is whether the confidence level is chosen in a reasonable fashion. For now, we
will take a leap of faith and assume that the width of confidence intervals for
subgaussian bandits cannot be significantly improved from what we use here
(we shall see that this holds in later chapters), and concentrate on choosing the
confidence level now.

Choosing the confidence level is a delicate problem, and we will analyse a

number of choices in future chapters. The basic difficulty is that δ should
be small enough to ensure optimism with high probability, but not so large
that suboptimal arms are explored excessively.

Nevertheless, as a first cut, the choice of this parameter can be guided by

the following considerations. If the confidence interval fails and the index of an
optimal arm drops below its true mean, then it could happen that the algorithm
stops playing the optimal arm and suffers linear regret. This suggests we might
choose δ ≈ 1/n so that the contribution to the regret of this failure case is
relatively small. Unfortunately things are not quite this simple. As we have
 7.1 The Optimism Principle 104

already alluded to, one of the main difficulties is that the number of samples
Ti (t − 1) in the index (7.2) is a random variable, and so our concentration results
cannot be immediately applied. For this reason we will see that (at least naively)
δ should be chosen a bit smaller than 1/n.
Theorem 7.1. Consider UCB as shown in Algorithm 3 on a stochastic k-armed
1-subgaussian bandit problem. For any horizon n, if δ = 1/n2 , then
k
X X 16 log(n)
Rn ≤ 3 ∆i + .
i=1
∆i
i:∆i >0

Before the proof we need a little more notation. Let (Xti )t∈[n],i∈[k] be a collection
of independent random variables with the law of Xti equal to Pi . Then define
Ps
µ̂is = 1s u=1 Xui to be the empirical mean based on the first s samples. We
make use of the third model in Section 4.6 by assuming that the reward in round
t is
Xt = XTAt (t)At .
Then we define µ̂i (t) = µ̂iTi (t) to be the empirical mean of the ith arm after round
t. The proof of Theorem 7.1 relies on the basic regret decomposition identity,
k
X
Rn = ∆i E [Ti (n)] . (Lemma 4.5)
i=1

The theorem will follow by showing that E [Ti (n)] is not too large for suboptimal
arms i. The key observation is that after the initial period where the algorithm
chooses each action once, action i can only be chosen if its index is higher than
that of an optimal arm. This can only happen if at least one of the following is
true:
(a) The index of action i is larger than the true mean of a specific optimal arm.
(b) The index of a specific optimal arm is smaller than its true mean.
Since with reasonably high probability the index of any arm is an upper bound
on its mean, we don’t expect the index of the optimal arm to be below its
mean. Furthermore, if the suboptimal arm i is played sufficiently often, then its
exploration bonus becomes small and simultaneously the empirical estimate of
its mean converges to the true value, putting an upper bound on the expected
total number of times when its index stays above the mean of the optimal arm.
The proof that follows is typical for the analysis of algorithms like UCB, and
hence we provide quite a bit of detail so that readers can later construct their
own proofs.
Proof of Theorem 7.1 Without loss of generality, we assume the first arm is
optimal so that µ1 = µ∗ . As noted above,
k
X
Rn = ∆i E [Ti (n)] . (7.4)
i=1
 7.1 The Optimism Principle 105

The theorem will be proven by bounding E[Ti (n)] for each suboptimal arm i. We
make use of a relatively standard idea, which is to decouple the randomness from
the behaviour of the UCB algorithm. Let Gi be the ‘good’ event defined by
s  
n o n 2 1 o
Gi = µ1 < min UCB1 (t, δ) ∩ µ̂iui + log < µ1 ,
t∈[n] ui δ

where ui ∈ [n] is a constant to be chosen later. So Gi is the event when µ1 is

never underestimated by the upper confidence bound of the first arm, while at the
same time the upper confidence bound for the mean of arm i after ui observations
are taken from this arm is below the pay-off of the optimal arm. We will show
two things:

1 If Gi occurs, then arm i will be played at most ui times: Ti (n) ≤ ui .

2 The complement event Gci occurs with low probability (governed in some way
yet to be discovered by ui ).

Because Ti (n) ≤ n no matter what, this will mean that

E [Ti (n)] = E [I {Gi } Ti (n)] + E [I {Gci } Ti (n)] ≤ ui + P (Gci ) n . (7.5)

The next step is to complete our promise by showing that Ti (n) ≤ ui on Gi and
that P (Gci ) is small. Let us first assume that Gi holds and show that Ti (n) ≤ ui ,
which we do by contradiction. Suppose that Ti (n) > ui . Then arm i was played
more than ui times over the n rounds, and so there must exist a round t ∈ [n]
where Ti (t − 1) = ui and At = i. Using the definition of Gi ,
s
2 log(1/δ)
UCBi (t − 1, δ) = µ̂i (t − 1) + (definition of UCBi (t − 1, δ))
Ti (t − 1)
s
2 log(1/δ)
= µ̂iui + (since Ti (t − 1) = ui )
ui
< µ1 (definition of Gi )
< UCB1 (t − 1, δ) . (definition of Gi )

Hence At = argmaxj UCBj (t − 1, δ) 6= i, which is a contradiction. Therefore if

Gi occurs, then Ti (n) ≤ ui . Let us now turn to upper bounding P (Gci ). By its
definition,
 s 
   
2 log(1/δ)
Gci = µ1 ≥ min UCB1 (t, δ) ∪ µ̂iui + ≥ µ1 . (7.6)
t∈[n]  ui 
 7.1 The Optimism Principle 106

The first of these sets is decomposed using the definition of UCB1 (t, δ),
  ( r )
2 log(1/δ)
µ1 ≥ min UCB1 (t, δ) ⊂ µ1 ≥ min µ̂1s +
t∈[n] s∈[n] s
( r )
[ 2 log(1/δ)
= µ1 ≥ µ̂1s + .
s
s∈[n]

Then using a union bound and the concentration bound for sums of independent
subgaussian random variables in Corollary 5.5, we obtain:
 ( r )
  [ 2 log(1/δ)
P µ1 ≥ min UCB1 (t, δ) ≤ P  µ1 ≥ µ̂1s + 
t∈[n] s
s∈[n]
r !
2 log(1/δ)
Xn
≤ P µ1 ≥ µ̂1s + ≤ nδ . (7.7)
s=1
s

The next step is to bound the probability of the second set in (7.6). Assume that
ui is chosen large enough that
s
2 log(1/δ)
∆i − ≥ c∆i (7.8)
ui

for some c ∈ (0, 1) to be chosen later. Then, since µ1 = µi + ∆i , and using

Corollary 5.5,
 s   s 
2 log(1/δ) 2 log(1/δ)
P µ̂iui + ≥ µ1  = P µ̂iui − µi ≥ ∆i − 
ui ui
 
ui c2 ∆2i
≤ P (µ̂iui − µi ≥ c∆i ) ≤ exp − .
2
Taking this together with (7.7) and (7.6), we have
 
ui c2 ∆2i
P (Gi ) ≤ nδ + exp −
c
.
2
When substituted into Eq. (7.5), we obtain
  
ui c2 ∆2i
E [Ti (n)] ≤ ui + n nδ + exp − . (7.9)
2
It remains to choose ui ∈ [n] satisfying (7.8). A natural choice is the smallest
integer for which (7.8) holds, which is
 
2 log(1/δ)
ui = .
(1 − c)2 ∆2i
This choice of ui can be larger than n, but in this case Eq. (7.9) holds trivially
 7.1 The Optimism Principle 107

since Ti (n) ≤ n. Then, using the assumption that δ = 1/n2 and this choice of ui
leads via (7.9) to
 
2 2 2 log(n2 ) 2 2
E[Ti (n)] ≤ ui + 1 + n1−2c /(1−c) = + 1 + n1−2c /(1−c) . (7.10)
(1 − c)2 ∆2i
All that remains is to choose c ∈ (0, 1). The second term will contribute a
polynomial dependence on n unless 2c2 /(1 − c)2 ≥ 1. However, if c is chosen too
close to 1, then the first term blows up. Somewhat arbitrarily we choose c = 1/2,
which leads to
16 log(n)
E [Ti (n)] ≤ 3 + .
∆2i
The result follows by substituting the above display in Eq. (7.4).
As we saw for the ETC strategy, the regret bound in Theorem 7.1 depends
on the reciprocal of the gaps, which may be meaningless when even a single
suboptimal action has a very small suboptimality gap. As before, one can also
prove a sublinear regret bound that does not depend on the reciprocal of the
gaps.
Theorem 7.2. If δ = 1/n2 , then the regret of UCB, as defined in Algorithm 3,
on any ν ∈ ESG
k
(1) environment, is bounded by
p k
X
Rn ≤ 8 nk log(n) + 3 ∆i .
i=1

Proof Let ∆ > 0 be some value to be tuned subsequently, and recall from the
proof of Theorem 7.1 that for each suboptimal arm i, we can bound
16 log(n)
E[Ti (n)] ≤ 3 + .
∆2i
Therefore, using the basic regret decomposition again (Lemma 4.5), we have
k
X X X
Rn = ∆i E [Ti (n)] = ∆i E [Ti (n)] + ∆i E [Ti (n)]
i=1 i:∆i <∆ i:∆i ≥∆
X  16 log(n)

16k log(n) X
≤ n∆ + 3∆i + ≤ n∆ + +3 ∆i
∆i ∆ i
i:∆i ≥∆

p k
X
≤8 nk log(n) + 3 ∆i ,
i=1
P
where the first inequality follows because i:∆i <∆ Ti (n) ≤ n and the last line by
p
choosing ∆ = 16k log(n)/n.
P
The additive i ∆i term is unavoidable because no reasonable algorithm can
avoid playing each arm once (try to work out what would happen if it did not).
In any case, this term does not grow with the horizon n and is typically negligible.
 7.1 The Optimism Principle 108

100 ETC (m = 25)

ETC (m = 50)
Expected regret 80 ETC (m = 75)
ETC (m = 100)
60 ETC (optimal m)
UCB
40

20

0 0.2 0.4 0.6 0.8 1

∆

Figure 7.1 Experiment showing universality of UCB relative to fixed instances of ETC

As it happens, Theorem 7.2 is close to optimal.

√ We will see in Chapter 15k that
no algorithm can enjoy regret smaller than O( nk) over all problems in ESG (1).
In Chapter 9 we will also see a more complicated variant of Algorithm 3 that
shaves the logarithmic term from the upper bound given above.
Experiment 7.1 We promised that UCB would overcome the limitations
of ETC by achieving the same guarantees but without prior knowledge of
the suboptimality gaps. The theory supports this claim, but just because two
algorithms have similar theoretical guarantees does not mean they perform the
same empirically. The theoretical analysis might be loose for one algorithm and
maybe not the other, or by a different margin. For this reason it is always wise to
prove lower bounds (which we do later) and compare the empirical performance,
which we do (very briefly) now.
The set-up is the same as in Fig. 6.1, which has n = 1000 and k = 2 and
unit variance Gaussian rewards with means 0 and −∆ respectively. The plot in
Fig. 7.1 shows the expected regret of UCB relative to ETC for a variety of choices
of commitment time m. The expected regret of ETC with the optimal choice of
m (which depends on the knowledge of ∆ and that the pay-offs are Gaussian, cf.
Fig. 6.1) is also shown.

The results demonstrate a common phenomenon. If ETC is tuned with the

optimal choice of commitment time for each choice of ∆, then it outperforms
the parameter-free UCB, though only by a relatively small margin. If,
however, the commitment time must be chosen without the knowledge of
∆, then ETC will usually not outperform UCB. As it happens, a variant of
UCB introduced in the next chapter actually outperforms even the optimally
tuned ETC.
 7.2 Notes 109

7.2 Notes

1 The choice of δ = 1/n2 led to an easy analysis, but comes with two
disadvantages. First of all, it turns out that a slightly smaller value of δ
improves the regret (and empirical performance). Secondly, the dependence on
n means the horizon must be known in advance, which is often not reasonable.
Both of these issues are resolved in the next chapter, where δ is chosen to be
smaller and to depend on the current round t rather than n. Nonetheless – as
promised – Algorithm 3 with δ = 1/n2 does achieve a regret bound similar to
the ETC strategy, but without requiring knowledge of the gaps.
2 The assumption that the rewards generated by each arm are independent can
be relaxed significantly. All of the results would go through by assuming there
exists a mean reward vector µ ∈ Rk such that

E[Xt | X1 , A1 , . . . , At−1 , Xt−1 , At ] = µAt a.s. . (7.11)

2
E[exp(λ(Xt − µAt )) | X1 , A1 , . . . , At−1 , Xt−1 , At ] ≤ exp(λ /2) a.s. . (7.12)

Eq. (7.11) is just saying that the conditional mean of the reward in round t
only depends on the chosen action. Eq. (7.12) ensures that the tails of Xt are
conditionally subgaussian. That everything still goes through is proven using
martingale techniques, which we develop in detail in Chapter 20.
3 So is the optimism principle universal? Does it always lead to policies with
strong guarantees in more complicated settings? Unfortunately the answer turns
out to be no. The optimism principle usually leads to reasonable algorithms
when (i) any action gives feedback about the quality of that action and (ii) no
action gives feedback about the value of other actions. When (i) is violated, even
sublinear regret may not be guaranteed. When (ii) is violated, an optimistic
algorithm may avoid actions that lead to large information gain and low reward,
even when this trade-off is optimal. An example where this occurs is provided
in Chapter 25 on linear bandits. Optimism can work in more complex models as
well, but sometimes fails to appropriately balance exploration and exploitation.
4 When thinking about future outcomes, humans and some animals often have
higher expectations than are warranted by past experience or conditions of the
environment. This phenomenon, a form of cognitive bias, is known as the
optimism bias in the psychology and behavioural economics literature and is
in fact ‘one of the most consistent, prevalent, and robust biases documented in
psychology and behavioral economics’ [Sharot, 2011a]. While much has been
written about this bias in these fields, and one of the current explanations
of why the optimism bias is so prevalent is that it helps exploration, to our
best knowledge, the connection to the deeper mathematical justification of
optimism, pursued here and in other parts of this book, has so far escaped the
attention of researchers in all the relevant fields.
 7.3 Bibliographical Remarks 110

7.3 Bibliographical Remarks

The use of confidence bounds and the idea of optimism first appeared in the work
by Lai and Robbins [1985]. They analysed the asymptotics for various parametric
bandit problems (see the next chapter for more details on this). The first version
of UCB is by Lai [1987]. Other early work is by Katehakis and Robbins [1995],
who gave a very straightforward analysis for the Gaussian case, and Agrawal
[1995], who noticed that all that was needed is an appropriate sequence of
upper confidence bounds on the unknown means. In this way, their analysis is
significantly more general than what we have done here. These researchers also
focused on the asymptotics, which at the time was the standard approach in
the statistics literature. The UCB algorithm was independently discovered by
Kaelbling [1993], although with no regret analysis or clear advice on how to tune
the confidence parameter. The version of UCB discussed here is most similar to
that analysed by Auer et al. [2002a] under the name UCB1, but that algorithm
used t rather than n in the confidence level (see the next chapter). Like us, they
prove a finite-time regret bound. However, rather than considering 1-subgaussian
environments, Auer et al. [2002a] considers bandits where the pay-offs are confined
to the [0, 1] interval, which are ensured to be 1/2-subgaussian. See Exercise 7.2
for hints on what must change in this situation. The basic structure of the proof
of our Theorem 7.1 is essentially the same as that of theorem 1 of Auer et al.
[2002a]. The worst-case bound in Theorem 7.2 appeared in the book by Bubeck
and Cesa-Bianchi [2012], which also popularised the subgaussian set-up. We did
not have time to discuss the situation where the subgaussian constant is unknown.
There have been several works exploring this direction. If the variance is unknown,
but the noise is bounded, then one can replace the subgaussian concentration
bounds with an empirical Bernstein inequality [Audibert et al., 2007]. For details,
see Exercise 7.6. If the noise has heavy tails, then a more serious modification is
required, as discussed in Exercise 7.7 and the note that follows.
We found the article by Sharot [2011a] on optimism bias from the psychology
literature quite illuminating. Readers looking to dive deeper into this literature
may enjoy the book by the same author [Sharot, 2011b]. Optimism bias is also
known as ‘unrealistic optimism’, a term that is most puzzling to us – what bias
is ever realistic? The background of this is explained by Jefferson et al. [2017].

7.4 Exercises

7.1 (Concentration for sequences of random length) In this exercise,

we investigate one of the more annoying challenges when analyzing sequential
algorithms. Let X1 , X2 , . . . be a sequence of independent standard Gaussian
random variables defined on probability space (Ω, F, P). Suppose that T : Ω →
PT
{1, 2, 3, . . .} is another random variable, and let µ̂ = t=1 Xt /T be the empirical
mean based on T samples.
 7.4 Exercises 111

(a) Show that if T is independent from Xt for all t, then

r !
2 log(1/δ)
P µ̂ − µ ≥ ≤ δ.
T

(b) Now relax the assumption that T is independent from (Xt )t . Let Et =
I {T = t} be the event that T = t and Ft = σ(X1 , . . . , Xt ) be the σ-algebra
generated by the first t samples. Let δ ∈ (0, 1) and show there exists a T
such that for all t ∈ {1, 2, 3, . . .} it holds that Et is Ft -measurable and
r !
2 log(1/δ)
P µ̂ − µ ≥ = 1.
T

(c) Show that

r !
2 log(T (T + 1)/δ)
P µ̂ − µ ≥ ≤ δ. (7.13)
T

Hint For part (b) above, you may find it useful to apply the law of the iterated
logarithm, which says if X1 , X2 , . . . is a sequence of independent and identically
distributed random variables with zero mean and unit variance, then
Pn
t=1 Xt
lim sup √ =1 almost surely .
n→∞ 2n log log n
This result is especially remarkable because it relies on no assumptions other
than zero mean and unit variance. You might wonder if Eq. (7.13) might continue
to hold if log(T (T + 1)/δ) were replaced by log(log(T )/δ). It almost does, but
the proof of this fact is more sophisticated. For more details, see the paper by
Garivier [2013] or Exercise 20.9.
7.2 (Relaxing the subgaussian assumption) In this chapter, we assumed
the pay-off distributions were 1-subgaussian. The purpose of this exercise is to
relax this assumption.
(a) First suppose that σ 2 > 0 is a known constant and that ν ∈ ESG
k
(σ 2 ). Modify
the UCB algorithm and state and prove an analogue of Theorems 7.1 and 7.2
for this case.
(b) Now suppose that ν = (Pi )ki=1 is chosen so that Pi is σi -subgaussian where
(σi2 )ki=1 are known. Modify the UCB algorithm and state and prove an
analogue of Theorems 7.1 and 7.2 for this case.
(c) If you did things correctly, the regret bound in the previous part should not
depend on the values of {σi2 : ∆i = 0}. Explain why not.

7.3 (High-probability bounds) Recall from Chapter 4 that the pseudo-regret

is defined to be the random variable
n
X
R̄n = ∆A t .
t=1
 7.4 Exercises 112

The UCB policy in Algorithm 3 depends on confidence parameter δ ∈ (0, 1] that

determines the level of optimism. State and prove a bound on the pseudo-regret
of this algorithm that holds with probability 1 − f (n, k)δ, where f (n, k) is a
function, that depends on n and k only. More precisely show that for bandit
k
ν ∈ ESG (1) that


P R̄n ≥ g(n, ν, δ) ≤ f (n, k)δ ,

where g and f should be as small as possible (there are trade-offs – try and come
up with a natural choice).

7.4 (Phased UCB (i)) Fix a 1-subgaussian k-armed bandit environment and a
horizon n. Consider the version of UCB that works in phases of exponentially
increasing length of 1, 2, 4, . . .. In each phase, the algorithm uses the action that
would have been chosen by UCB at the beginning of the phase (see Algorithm 4
below).

(a) State and prove a bound on the regret for this version of UCB.
(b) Compare your result with Theorem 7.1.
 
(c) How would the result change if the `th phase had a length of α` with
α > 1?

1: Input k and δ
2: Choose each arm once
3: for ` = 1, 2, . . . do
4: Compute A` = argmaxi UCBi (t − 1, δ)
5: Choose arm A` exactly 2` times
6: end for
Algorithm 4: A phased version of UCB.

7.5 (Phased UCB (ii)) Let α > 1 and consider the version of UCB that first
plays each arm once. Thereafter it operates in the same way as UCB, but rather
than playing the chosen arm just once, it plays it until the number of plays of
that arm is a factor of α larger (see Algorithm 5 below).

(a) State and prove a bound on the regret for version of UCB with α = 2
(doubling counts).
(b) Compare with the result of the previous exercise and with Theorem 7.1.
What can you conclude?
(c) Repeat the analysis for α > 1. What is the role of α?
(d) Implement these algorithms and compare them empirically to UCB(δ).
 7.4 Exercises 113

1: Input k and δ
2: Choose each arm once
3: for ` = 1, 2, . . . do
4: Let t` = t
5: Compute A` = argmaxi UCBi (t` − 1, δ)
6: Choose arm A` until round t such that Ti (t) ≥ αTi (t` − 1)
7: end for
Algorithm 5: A phased version of UCB.

The algorithms of the last two exercises may seem ridiculous. Why would
you wait before updating empirical estimates and choosing a new action?
There are at least two reasons:
(a) It can happen that the algorithm does not observe its rewards
immediately, but rather they appear asynchronously after some delay.
Alternatively many bandits algorithms may be operating simultaneously
and the results must be communicated at some cost.
(b) If the feedback model has a more complicated structure than what we
examined so far, then even computing the upper confidence bound just
once can be quite expensive. In these circumstances, it’s comforting to
know that the loss of performance by updating the statistics only rarely
is not too severe.

7.6 (Adapting to reward variance in bandits with bounded rewards)

Let X1 , X2 , . . . , Xn be a sequence of independent and identically distributed
random variables with mean µ and variance σ 2 and bounded support so that
Pn Pn
Xt ∈ [0, b] almost surely. Let µ̂ = t=1 Xt /n and σ̂ 2 = t=1 (µ̂ − Xt )2 /n. The
empirical Bernstein inequality says that for any δ ∈ (0, 1),
s    !
2σ̂ 2 3 3b 3
P |µ̂ − µ| ≥ log + log ≤ δ.
n δ n δ

1
Pn
(a) Show that σ̂ 2 = n t=1 (Xt − µ)
2
− (µ̂ − µ)2 .
2 2 2
(b) Show that V[(Xt − µ) ] ≤ b σ .
(c) Use Bernstein’s inequality (Exercise 5.14) to show that
s    !
2 2 2b2 σ 2 1 2b2 1
P σ̂ ≥ σ + log + log ≤ δ.
n δ 3n δ

(d) Suppose that ν = (νi )ki=1 is a bandit where Supp(νi ) ⊂ [0, b] and the variance
of the ith arm is σi2 (with our earlier notation, ν ∈ E[0,b]
k
). Design a policy
 7.4 Exercises 114

that depends on b, but not σi2 such that

X  
σi2
 
Rn ≤ C ∆i + b + log(n) , (7.14)
∆i
i:∆i >0

where C > 0 is a universal constant.

If you did things correctly, then the policy you derived in Exercise 7.6
should resemble UCB-V by Audibert et al. [2007]. The proof of the empirical
Bernstein also appears there or (with slightly better constants) in the papers
by Mnih et al. [2008] and Maurer and Pontil [2009].

It is worth comparing (7.14) to the result of Theorem 7.1. In particular,

recall that if the rewards are bounded by b, the reward distributions are
b-subgaussian. The regret of UCB which adjusts the confidence intervals
P
accordingly can then be shown to be Rn = O( i:∆i >0 b log(n) ∆i ). Thus, the
main advantage of the policy of the previous exercise is the replacement of
σ2
b/∆i in this bound with b + ∆ii . In Exercise 16.7, you will show that this is
essentially unimprovable.

7.7 (Median of means and bandits with known finite variance)

Let n ∈ N+ and (Ai )m i=1 be a partition of [n] so that ∪i=1 Ai = [n] and
m

Ai ∩ Aj = ∅ for all i = 6 j. Suppose that δ ∈ (0, 1) and X1 , X2 , . . . , Xn is a

sequence of independent random variables with mean µ and variance σ 2 . The
median-of-means estimator µ̂M of µ is the median of µ̂1 , µ̂2 , . . . , µ̂m , where
P
µ̂i = t∈Ai Xt /|Ai | is the mean of the data in the ith block.

j n  1/8 ok
(a) Show that if m = min n2 , 8 log e δ and Ai are chosen as equally
sized as possible, then
s   !
192σ 2 e1/8
P µ̂M + log ≤µ ≤ δ.
n δ

(b) Use the median-of-means estimator to design an upper confidence bound

algorithm such that for all ν ∈ EVk (σ 2 ),

X  σ 2 log(n)

Rn ≤ C ∆i + ,
∆i
i:∆i >0

where C > 0 is a universal constant.

 7.4 Exercises 115

This exercise shows that the subgaussian assumption can be relaxed to

requiring only finite variance at the price of increased constant factors. The
result is only possible by replacing the standard empirical estimator with
something more robust. The median-of-means estimator is only one way to
do this. In fact, the empirical estimator can be made robust by truncating
the observed rewards and applying the empirical Bernstein concentration
inequality. The disadvantage of this approach is that choosing the location
of truncation requires prior knowledge about the approximate location of
the mean. Another approach is Catoni’s estimator, which also exhibits
excellent asymptotic properties [Catoni, 2012]. Yet another idea is to minimise
the Huber loss [Sun et al., 2017]. This latter paper is focussing on linear
models, but the results still apply in one dimension. The application of these
ideas to bandits was first made by Bubeck et al. [2013a], where the reader
will find more interesting results. Most notably, that things can still be made
to work even if the variance does not exist. In this case, however, there is a
price to be paid in terms of the regret. The median-of-means estimator is due
to Alon et al. [1996]. In case the variance is also unknown, then it may be
estimated by assuming a known bound on the kurtosis, which covers many
classes of bandits (Gaussian with arbitrary variance, exponential and many
more), but not some simple cases (Bernoulli). The policy that results from
this procedure has the benefit of being invariant under the transformations
of shifting or scaling the losses [Lattimore, 2017].

7.8 (Empirical comparison)

(a) Implement Algorithm 3.
(b) Reproduce Fig. 7.1.
(c) Explain the shape of the curves for ETC. In particular, when m = 50, we
see a bump, a dip and then a linear asymptote as ∆ grows. Why does the
curve look like this?
(d) Design an experiment to determine the practical effect of the choice of δ.
(e) Explain your results.
8 The Upper Confidence Bound
Algorithm: Asymptotic Optimality

The algorithm analysed in the previous chapter is not anytime. This shortcoming
is resolved via a slight modification and a refinement of the analysis. The improved
analysis leads to constant factors in the dominant logarithmic term that match a
lower bound provided later in Chapter 16.

8.1 Asymptotically Optimal UCB

The algorithm studied is shown in Algorithm 6. It differs from the one analysed
in the previous section (Algorithm 3) only by the choice of the confidence level,
the choice of which is dictated by the analysis of its regret.

1: Input k
2: Choose each arm once
3: Subsequently choose
s !
2 log f (t)
At = argmaxi µ̂i (t − 1) +
Ti (t − 1)

where f (t) = 1 + t log2 (t)

Algorithm 6: Asymptotically optimal UCB.

The regret bound for Algorithm 6 is more complicated than the bound for
Algorithm 3 (see Theorem 7.1). The dominant terms in the two results have the
same order, but the gain here is that in this result the leading constant, governing
the asymptotic rate of growth of regret, is smaller.

Theorem 8.1. For any 1-subgaussian bandit, the regret of Algorithm 6 satisfies
  p 
X 5 2 log f (n) + π log f (n) + 1
Rn ≤ inf ∆i 1 + 2 + . (8.1)
ε∈(0,∆i ) ε (∆i − ε)2
i:∆i >0
 8.1 Asymptotically Optimal UCB 117

Furthermore,
Rn X 2
lim sup ≤ . (8.2)
n→∞ log(n) ∆i
i:∆i >0

Choosing ε = ∆i /2 inside the sum shows that

X  1  p 
Rn ≤ ∆i + 8 log f (n) + 8 π log f (n) + 28 . (8.3)
∆i
i:∆i >0

Even more concretely, there exists some universal constant C > 0 such that
X  log(n)

Rn ≤ C ∆i + ,
∆i
i:∆i >0

which by the same argument as

pin the proof of Theorem 7.2 leads a worst-case
Pk
bound of Rn ≤ C i=1 ∆i + 2 Cnk log(n).

Taking the limit of the ratio of the bound in (8.3) and log(n) does not result
in the same constant as in the theorem, which is the main justification for
introducing the more complicated regret bound. You will see in Chapter 15
that the asymptotic bound on the regret given in (8.2) is unimprovable in a
strong sense.

We start with a useful lemma to bound the number of times the index of a
suboptimal arm will be larger than some threshold above its mean.
Lemma 8.2. Let X1 , . . . , Xn be a sequence of independent 1-subgaussian random
Pt
variables, µ̂t = 1t s=1 Xs , ε > 0, a > 0 and
( r ) ( r )
2a 2a
Xn X n
κ= I µ̂t + ≥ε , κ =u+
0
I µ̂t + ≥ε ,
t=1
t t
t=due

2 √
where u = 2aε−2 . Then it holds E[κ] ≤ E[κ0 ] ≤ 1 + 2
(a + πa + 1).
ε
The intuition for this result is as follows. Since the Xi are
p1-subgaussian and
independent we have E[µ̂t ] = 0, so we cannot expect µ̂t + 2a/t to be smaller
than ε until t is at least 2a/ε2 . The lemma confirms that this is the right order
as an estimate for E [κ].
Proof By Corollary 5.5 we have
  q 2 
r ! 2a
n
X 2a
n
X  t ε − t 
E[κ] ≤ E[κ0 ] = u + P µ̂t + ≥ε ≤u+ exp − 
t 2
t=due t=due
  q 2 
Z 2a
 t ε− 2
∞ √
t 
≤1+u+ exp −  dt = 1 + 2 (a + πa + 1) ,
u 2 ε
 8.1 Asymptotically Optimal UCB 118

√ √
where the final equality follows by making the substitution s = ε t − 2a and
substituting the value of u from the lemma statement.

Proof of Theorem 8.1 As usual, the starting point is the fundamental regret
decomposition (Lemma 4.5),

X
Rn = ∆i E[Ti (n)] .
i:∆i >0

The rest of the proof revolves around bounding E[Ti (n)]. Let i be a suboptimal
arm. The main idea is to decompose Ti (n) into two terms. The first measures the
number of times the index of the optimal arm is less than µ1 − ε. The second term
measures the number of times that At = i and its index is larger than µ1 − ε.

( s )
2 log f (t)
n
X n
X
Ti (n) = I {At = i} ≤ I µ̂1 (t − 1) + ≤ µ1 − ε
t=1 t=1
T1 (t − 1)
( s )
2 log f (t)
n
X
+ I µ̂i (t − 1) + ≥ µ1 − ε and At = i . (8.4)
t=1
Ti (t − 1)

The proof of the first part of the theorem is completed by bounding the expectation
of each of these two sums. Starting with the first, we again use Corollary 5.5:

" ( s )#
2 log f (t)
n
X
E I µ̂1 (t − 1) + ≤ µ1 − ε
t=1
T1 (t − 1)
r !
2 log f (t)
Xn X n
≤ P µ̂1s + ≤ µ1 − ε
t=1 s=1
s
 q 2 
2 log f (t)
Xn X n  s s +ε 
 
≤ exp − 
t=1 s=1
 2 

 
1 X sε2 5
Xn n
≤ exp − ≤ 2.
t=1
f (t) s=1 2 ε

The first inequality follows from the union bound over all possible values of
T1 (t − 1). The last inequality is an algebraic exercise (Exercise 8.1). The function
f (t) was chosen precisely so this bound would hold. For the second term in (8.4)
 8.2 Notes 119

we use Lemma 8.2 to get

" n ( s )#
X 2 log f (t)
E I µ̂i (t − 1) + ≥ µ1 − ε and At = i
t=1
Ti (t − 1)
" n ( s )#
X 2 log f (n)
≤E I µ̂i (t − 1) + ≥ µ1 − ε and At = i
t=1
Ti (t − 1)
" n ( r )#
X 2 log f (n)
≤E I µ̂is + ≥ µ1 − ε
s=1
s
" n ( r )#
X 2 log f (n)
=E I µ̂is − µi + ≥ ∆i − ε
s=1
s
2  p 
≤1+ log f (n) + π log f (n) + 1 .
(∆i − ε) 2

The first part of the theorem follows by substituting the results of the previous
two displays into (8.4). The second part follows by choosing ε = log−1/4 (n) and
taking the limit as n tends to infinity.

8.2 Notes

1 The improvement to the constants comes from making the confidence interval
slightly smaller, which is made possible by a more careful analysis. The main
trick is the observation that we do not need to show that µ̂1s ≥ µ1 for all s
with high probability, but instead that µ̂1s ≥ µ1 − ε for small ε.
2 The choice of f (t) = 1 + t log2 (t) looks quite odd. With a slightly messier
calculation we could have chosen f (t) = t logα (t) for any α > 0. If the rewards
are actually Gaussian, then a more careful concentration analysis allows one
to choose f (t) = t or even some slightly slower-growing function [Katehakis
and Robbins, 1995, Lattimore, 2016a, Garivier et al., 2016b].
3 The asymptotic regret is often indicative of finite-time performance. The reader
is advised to be cautious, however. The lower-order terms obscured by the
asymptotics can be dominant in all practical regimes.

8.3 Bibliographic Remarks

Lai and Robbins [1985] designed policies for which Eq. (8.2) holds. They also
proved a lower bound showing that no ‘reasonable’ policy can improve on this
bound for any problem, where ‘reasonable’ means that they suffer subpolynomial
regret on all problems (see Part IV). The policy proposed by Lai and Robbins
[1985] was based on upper confidence bounds, but was not a variant of UCB. The
asymptotics for variants of the policy presented here were given first by Lai [1987],
 8.4 Exercises 120

Katehakis and Robbins [1995] and Agrawal [1995]. None of these articles gave
finite-time bounds like what was presented here. When the reward distributions
lie in an exponential family, then asymptotic and finite-time bounds with the
same flavor to what is presented here are given by Cappé et al. [2013]. There are
now a huge variety of asymptotically optimal policies in a wide range of settings.
Burnetas and Katehakis [1996] study the general case and give conditions for
a version of UCB to be asymptotically optimal. Honda and Takemura [2010,
2011] analyse an algorithm called DMED, proving asymptotic optimality for noise
models where the support is bounded or semi-bounded. Kaufmann et al. [2012b]
prove asymptotic optimality for Thompson sampling (see Chapter 36) when
the rewards are Bernoulli, which is generalised to single-parameter exponential
families by Korda et al. [2013]. Kaufmann [2018] proves asymptotic optimality
for the BayesUCB class of algorithms for single-parameter exponential families.
Ménard and Garivier [2017] prove asymptotic optimality and minimax optimality
for exponential families (more discussion in Chapter 9).

8.4 Exercises

8.1 Do the algebra needed at the end of the proof of Theorem 8.1. Precisely,
show that
 
1 X sε2 5
X n n
exp − ≤ 2,
t=1
f (t) s=1 2 ε

where f (t) = 1 + t log2 (t).

Pn
Hint First bound F = s=1 exp(−sε2 /2) using a geometric series. Then show
that exp(−a)/(1 − exp(−a)) ≤ 1/a holds for any a > 0 and conclude that F ≤ ε22 .
Pn
Finish by bounding t=1 1/f (t) using the fact that 1/f (t) ≤ 1/(t log(t)2 ) and
bounding a sum by an integral.
8.2 (One-armed bandits) Consider the one-armed bandit problem: E =
{N (µ1 , 1) : µ1 ∈ R} × {N (0, 1)}. Suppose that ν = (P1 , P2 ) ∈ E and P1 has mean
µ1 = 1. Evaluate
Rn (π, ν)
lim sup ,
n→∞ log(n)
where π is the policy of Algorithm 6.

8.3 (One-armed bandits (ii)) Consider the setting of Exercise 8.2 and define
a policy by
 q
1 if µ̂ (t − 1) + 2 log f (t) ≥ 0
1 T1 (t−1)
At = (8.5)
2 otherwise .
 8.4 Exercises 121

Suppose that ν = (P1 , P2 ) where P1 = N (µ1 , 1) and P2 = N (0, 1). Prove that
for the modified policy,
(
Rn (ν) 0 if µ1 ≥ 0
lim sup ≤
n→∞ log(n)
2
−µ 1
if µ1 < 0 .

Hint Follow the analysis for UCB, but carefully adapt the proof by using the
fact that the index of the second arm is always zero.

The strategy proposed in the above exercise is based on the idea that
optimism is used to overcome uncertainty in the estimates of the quality of
an arm, but for one-armed bandits the mean of the second arm is known in
advance.

8.4 (One-armed bandits (iii)) The purpose of this question is to compare

UCB and the modified version in (8.5).
(a) Implement a simulator for the one-armed bandit problem and two algorithms:
UCB and the modified version analysed in Exercise 8.3.
(b) Use your simulator to estimate the expected regret of each algorithm for a
horizon of n = 1000 and µ1 ∈ [−1, 1].
(c) Plot your results with µ1 on the x-axis and the estimated expected regret on
the y-axis. Don’t forget to label the axis and include error bars and a legend.
(d) Explain the results. Why do the curves look the way they do?
(e) In your plot, for what values of µ1 does the worst-case expected regret
for each algorithm occur? What is the worst-case expected regret for each
algorithm?

8.5 (Different subgaussian constants) Let σ 2 ∈ [0, ∞)k be known and

2
suppose that the reward is Xt ∼ N (µAt , σA t
). Design an algorithm (that depends
2
on σ ) for which the asymptotic regret is
Rn X 2σ 2
lim sup = i
.
n→∞ log(n) ∆i
i:∆i >0
9 The Upper Confidence Bound
Algorithm: Minimax Optimality ( )

We proved that the variants p of UCB analysed in the last two chapters have a
worst-case regret of Rn = O( kn log(n)). Further, p in Exercise 6.8 you showed
that an elimination algorithm achieves Rn = O( kn log(k)). By modifying the
confidence levels of the algorithm it is possible to remove the log factor entirely.
Building on UCB, the directly named ‘minimax optimal strategy in the stochastic
case’ (MOSS) algorithm was the first to make this modification and is presented
below. MOSS again depends on prior knowledge of the horizon, a requirement
that may be relaxed, as we explain in the notes.

The term minimax is used because, except for constant factors, the worst-
case bound proven in this chapter cannot be improved on by any algorithm.
The lower bounds are deferred to Part IV.

9.1 The MOSS Algorithm

Algorithm 7 shows the pseudocode of MOSS, which is again an instance of the

UCB family. The main novelty is that the confidence level is chosen based on the
number of plays of the individual arms, as well as n and k.

1: Input n and k
2: Choose each arm once
3: Subsequently choose
s  
4 n
At = argmaxi µ̂i (t − 1) + log+ ,
Ti (t − 1) kTi (t − 1)
where log+ (x) = log max {1, x} .

Algorithm 7: MOSS.
 9.1 The MOSS Algorithm 123

Theorem 9.1. For any 1-subgaussian bandit, the regret of Algorithm 7 satisfies

√ k
X
Rn ≤ 39 kn + ∆i .
i=1

Before the proof we state and prove a strengthened version of Corollary 5.5.

Theorem 9.2. Let X1 , X2 , . . . , Xn be a sequence of independent σ-subgaussian

Pt
random variables and St = s=1 Xs . Then, for any ε > 0,
 
ε2
P (exists t ≤ n : St ≥ ε) ≤ exp − . (9.1)
2nσ 2

The bound in Eq. (9.1) is the same as the bound on P (Sn ≥ ε) that appears
in a simple reformulation of Corollary 5.5, so this new result is strictly stronger.

Proof From the definition of subgaussian random variables and Lemma 5.4,
 
nσ 2 λ2
E [exp (λSn )] ≤ exp .
2

Then, choosing λ = ε/(nσ 2 ) leads to

 
P (exists t ≤ n : St ≥ ε) = P max exp (λSt ) ≥ exp (λε)
t≤n
 2 2   
E [exp (λSn )] nσ λ ε2
≤ ≤ exp − λε = exp − .
exp (λε) 2 2nσ 2

The novel step is the first inequality, which follows from Doob’s submartingale
inequality (Theorem 3.10) and the fact that that exp(λSt ) is a submartingale
with respect to the filtration generated by X1 , X2 , . . . , Xn (Exercise 9.1).

Before the proof of Theorem 9.1, we need one more lemma to bound the
probability that the index of the optimal arm ever drops too far below the actual
mean of the optimal arm. The proof of this lemma relies on a tool called the
peeling device, which is an important technique in probability theory and has
many applications beyond bandits. For example, it can be used to prove the
celebrated law of the iterated logarithm.

Lemma 9.3. Let δ ∈ (0, 1) and X1 , X2 , . . . be independent and 1-subgaussian and

Pt
µ̂t = 1t s=1 Xs . Then, for any ∆ > 0,
s   !
4 + 1 15δ
P exists s ≥ 1 : µ̂s + log +∆≤0 ≤ 2 .
s sδ ∆
 9.1 The MOSS Algorithm 124

Proof Let St = tµ̂t . Then

s   !
4 + 1
P exists s ≥ 1 : µ̂s + log +∆≤0
s sδ
s   !
+ 1
= P exists s ≥ 1 : Ss + 4s log + s∆ ≤ 0
sδ
s   !
+ 1
X∞
≤ P exists s ∈ [2 , 2 ] : Ss + 4s log
j j+1
+ s∆ ≤ 0
j=0
sδ
s   !
1
X∞
+
≤ P exists s ≤ 2 j+1
: Ss + 4 · 2 log j +2 ∆≤0
j

j=0
2j+1 δ
 q
2 
∞ 2 j+2 log+ 1
+ 2 j
∆
X  2j+1 δ 
≤ exp − .
j=0
2j+2

The first inequality follows from a union bound over a geometric grid. The second
step is straightforward but important because it sets up to apply Theorem 9.2.
The rest is purely algebraic:
 q
2 
∞ 2 j+2 log+ 1
+ 2 j
∆ ∞
X  2 j+1 δ  X

exp −  ≤ δ 2j+1 exp −∆2 2j−2
j=0
2j+2
j=0
Z
8δ ∞
15δ
≤ +δ 2s+1 exp −∆2 2s−2 ds ≤ 2 .
e∆2 0 ∆
Above, the first inequality follows since (a + b)2 ≥ a2 + b2 for a, b ≥ 0, and
the second last step follows by noting that the integrand is unimodal and
2
has a maximum value of 8δ/(e∆ R b ). For such functions f , one has the bound
Pb
j=a f (j) ≤ maxs∈[a,b] f (s) + a f (s)ds.

Proof of Theorem 9.1 As usual, we assume without loss of generality that the
first arm is optimal, so µ1 = µ∗ . Arguing that the optimal arm is sufficiently
optimistic with high probability is no longer satisfactory because in this refined
analysis, the probability that an arm is played linearly often needs to depend
on its suboptimality gap. A way around this difficulty is to make an argument
in terms of the expected amount of optimism. Define a random variable ∆ that
measures how far below the index of the optimal arm drops below its true mean.
r !!+
4 n
∆ = µ1 − min µ̂1s + log+ .
s≤n s ks

Arms with suboptimality gaps much larger than ∆ will not be played too often,
while arms with suboptimality gaps smaller than ∆ may be played linearly often,
but ∆ is sufficiently small in expectation that this price is small. Using the basic
 9.1 The MOSS Algorithm 125

regret decomposition (Lemma 4.5) and splitting the actions based on whether or
not their suboptimality gap is smaller or larger than 2∆ leads to

X
Rn = ∆i E[Ti (n)]
i:∆i >0
 
X
≤ E 2n∆ + ∆i Ti (n)
i:∆i >2∆
 
 √ X 
≤ E 2n∆ + 8 kn + ∆i Ti (n) .
 √
i:∆i >max 2∆,8 k/n

The first term is easily bounded using Proposition 2.8 and Lemma 9.3:

Z Z  
∞ ∞
15k √
E[2n∆] = 2nE[∆] = 2n P (∆ ≥ x) dx ≤ 2n min 1, 2 dx ≤ 16 kn .
0 0 nx

For suboptimal arm i, define

( r )
n
X 4 n
κi = I µ̂is + log+ ≥ µi + ∆i /2 .
s=1
s ks

The reason for choosing κi in this way is that for arms i with ∆i > 2∆, it holds
that the index of the optimal arm is always larger than µi + ∆i /2, so κi is an
upper bound on the number of times arm i is played, Ti (n). If ∆i ≥ 8(k/n)1/2 ,
then the expectation of ∆i κi is bounded using Lemma 8.2 by

" n ( s   )#
1 X 4 2
+ n∆i
∆i E[κi ] ≤ + ∆i E I µ̂is + log ≥ µi + ∆i /2
∆i s=1
s k
 2
 s  2
 !
1 8 + n∆ + n∆i
≤ + ∆i + 2 log i
+ 2π log +1
∆i ∆i k k
r r   r
1 n n p n
≤ + ∆i + 4 log 8 + 2 π log 8 + 1 ≤ ∆i + 15 ,
8 k k k

where the first inequality follows by replacing the s in the logarithm with 1/∆2i
and adding the ∆i × 1/∆2i correction term to compensate for the first ∆−2 i
rounds where this fails to hold. Then we use Lemma 8.2 and the monotonicity of
√
x 7→ x−1−p log+ (ax2 ) for p ∈ [0, 1], positive a and x ≥ e/ a. The last inequality
 9.2 Two Problems 126

√
follows by naively bounding 1/8 + 4 log 8 + 2 π log 8 + 1 ≤ 15. Then
   
 X   X 
E ∆i Ti (n) ≤ E  ∆i κi 
 √ √
i:∆i >max 2∆,8 k/n i:∆i >8 k/n
 r 
X n
≤ ∆i + 15
√ k
i:∆i >8 k/n

√ k
X
≤ 15 nk + ∆i .
i=1
√ Pk
Combining all the results we have Rn ≤ 39 kn + i=1 ∆i .

9.2 Two Problems

MOSS is not the ultimate algorithm. Here we highlight two drawbacks.

Suboptimality Relative to UCB

Although MOSS is nearly asymptotically optimal (Note 1), all versions of MOSS
can be arbitrarily worse than UCB in some regimes. This unpleasantness is hidden
by both the minimax and asymptotic optimality criteria, which highlights the
importance of fully finite-time upper and lower bounds. The counter-example
witnessing the failure is quite simple. Let the rewards
p for all arms be Gaussian
with unit variance and n = k 3 , µ1 = 0, µ2 = − k/n and µi = −1 for all i > 2.
From Theorem 8.1, we have that

RnUCB = O(k log k) ,

while it turns out that MOSS has a regret of

√
RnMOSS = Ω( kn) = Ω(k 2 ) .

A rigourous proof of this claim is quite delicate, but we encourage readers to try
to understand why it holds intuitively.

Instability
There is a hidden cost of pushing too hard to reduce the expected regret, which
is that the distribution of the regret is less well-behaved. Consider a two-armed
Gaussian bandit with suboptimality gap ∆. The random (pseudo) regret is
Pn
R̂n = t=1 ∆At , which for a carefully tuned algorithm has a roughly bimodal
distribution:
(
n∆ with probability δ
R̂n ≈ 1 1


∆ log δ otherwise ,
 9.3 Notes 127

where δ is a parameter of the policy that determines the likelihood that the
optimal arm is misidentified. Integrating, one has
  
1 1
Rn = E[R̂n ] = O n∆δ + log ,
∆ δ
The choice of δ that minimises the expected regret depends on ∆ and is
approximately 1/(n∆2 ). With this choice, the regret is
 
1 2



Rn = O 1 + log n∆ .
∆
Of course ∆ is not known in advance, but it can be estimated online so that the
above bound is actually realisable by an adaptive policy that does not know ∆
in advance (Exercise 9.3). Let F be the (informal) event that R̂n = Ω(n∆). The
problem is that when δ = 1/(n∆2 ) is chosen to minimise the expected regret,
then the second moment due to failure is
E[IF R̂n2 ] = Ω(n) .
On the other hand, by choosing δ = (n∆)−2 , the regret increases only slightly to
  
1 1

Rn = O + log n2 ∆2 .
∆ n
The second moment of the regret due to failure, however, is E[IF R̂n2 ] = O(1).

9.3 Notes

1 MOSS is quite close to asymptotically optimal. You can prove that

Rn X 4
lim sup ≤ .
n→∞ log(n) ∆i
i:∆i >0

By modifying the algorithm slightly, it is even possible to replace the four

with a two and recover the optimal asymptotic regret. The trick is to increase
g slightly and replace the four in the exploration bonus by two. The major
task is then to re-prove Lemma 9.3, which is done by replacing the intervals
[2j , 2j+1 ] with smaller intervals [ξ j , ξ j+1 ], where ξ is tuned subsequently to be
fractionally larger than one. This procedure is explained in detail by Garivier
[2013]. When the reward distributions are actually Gaussian, there is a more
elegant technique that avoids peeling altogether (Exercise 9.4).
2 One way to mitigate the issues raised in Section 9.2 is to replace the index
used by MOSS with a less aggressive confidence level:
s  
4 + n
µ̂i (t − 1) + log . (9.2)
Ti (t − 1) Ti (t − 1)
The resulting algorithm is never worse than UCB, and you will show in
 9.4 Bibliographic Remarks 128

p
Exercise 9.3 that it has a distribution free regret of O( nk log(k)). An
algorithm that does almost the same thing in disguise is called ‘improved
UCB’, which operates in phases and eliminates arms for which the upper
confidence bound drops below a lower confidence bound for some arm [Auer
and Ortner, 2010]. This algorithm was the topic of Exercise 6.8.
3 Overcoming the failure of MOSS to be instance optimal without sacrificing
minimax optimality is possible by using an adaptive confidence level that tunes
the amount of optimism to match the instance. One of the authors has proposed
two ways to do this, using one of the following indices:
s
2(1 + ε) n
µ̂i (t − 1) + log , or (9.3)
Ti (t − 1) t
v !
u
u 2 n
µ̂i (t − 1) + t log Pk p .
Ti (t − 1) j=1 min{Ti (t − 1), Ti (t − 1)Tj (t − 1)}

The first of these algorithms is called the ‘optimally confident UCB’ [Lattimore,
2015b] while the second is AdaUCB [Lattimore, 2018]. Both algorithms are
minimax optimal up to constant factors and never worse than UCB. The
latter is also asymptotically optimal. If the horizon is unknown, then AdaUCB
can be modified by replacing n with t. It remains a challenge to provide a
straightforward analysis for these algorithms.

9.4 Bibliographic Remarks

MOSS is due to Audibert and Bubeck [2009], while an anytime modification

is by Degenne and Perchet [2016]. The proof that a modified version of MOSS
is asymptotically optimal may be found in the article by Ménard and Garivier
[2017]. There is also a variant of MOSS that adapts to the variance for rewards
bounded in [0, 1] [Mukherjee et al., 2018]. AdaUCB and its friends are by one of
the authors [Lattimore, 2015b, 2016b, 2018]. The idea to modify the confidence
level has been seen in several places, with the earliest by Lai [1987] and more
recently by Honda and Takemura [2010]. Kaufmann [2018] also used a confidence
level like in Eq. (9.2) to derive an algorithm based on Bayesian upper confidence
bounds.

9.5 Exercises

9.1 (Submartingale property) Let X1 , X2 , . . . , Xn be adapted to filtration

Pt
F = (Ft )t with E[Xt | Ft−1 ] = 0 almost surely. Prove that Mt = exp(λ s=1 Xs )
is a F-submartingale for any λ ∈ R.

9.2 (Problem-dependent bound) Let ∆min = mini:∆i >0 ∆i . Show there exists
 9.5 Exercises 129

a universal constant C > 0 such that the regret of MOSS is bounded by

  X
n∆2min
k
Ck
Rn ≤ log+ + ∆i .
∆min k i=1

9.3 (UCB*) Suppose we modify the index used by MOSS to be

s  
4 + n
µ̂i (t − 1) + log .
Ti (t − 1) Ti (t − 1)

(a) Show that for all 1-subgaussian bandits, this new policy suffers regret at
most
 
X 1
Rn ≤ C  ∆i + log+ (n∆2i ) ,
∆i
i:∆i >0

where C > 0 is a universal constant.

(b) Under the same conditions as the previous part, show there exists a universal
constant C > 0 such that
p k
X
Rn ≤ C kn log(k) + ∆i .
i=1

(c) Repeat parts (a) and (b) using the index

s  
4 t
µ̂i (t − 1) + log+ .
Ti (t − 1) Ti (t − 1)

9.4 (Gaussian noise and the tangent approximation) Let g(t) = at + b

with b > 0 and
 2  
1 x 1 (x − 2b)2
u(x, t) = √ exp − −√ exp −2ab − .
2πt 2t 2πt 2t
(a) Show that u(x, t) > 0 for x ∈ (−∞, g(t)) and u(x, t) = 0 for x = g(t).
(b) Show that u(x, t) satisfies the heat equation:
1 2
∂t u(x, t) = ∂ u(x, t) .
2 x
(c) Let Bt be a standard Brownian motion, which for any fixed t has density
with respect to the Lebesgue measure.
 2
1 x
p(x, t) = √ exp − .
2πt 2t
Define τ = min{t : Bt = g(t)} as the first time the Brownian motion hits the
boundary. Put on your physicists hat (or work hard) to argue that
Z g(t)
P (τ ≥ t) = u(x, t)dx .
−∞
 9.5 Exercises 130

(d) Let v(t) be the density

Rt of time τ with respect to the Lebesgue measure so
that P (τg ≤ t) = 0 v(t)dt. Show that
 
b g(t)2
v(t) = √ exp −
2πt3 2t
(e) In the last part, you established the exact density of the hitting time of a
Brownian motion approaching a linear boundary. We now generalise this
to nonlinear boundaries, but at the cost that now we only have a bound.
Suppose that f : [0, ∞) → [0, ∞) is concave and differentiable, and let
λ : R → R be the intersection of the tangent to f at t with the y-axis given
by λ(t) = f (t) − tf 0 (t). Let τ = min{t : Bt = f (t)} and v(t) be the density
of τ . Show that for t > 0,
 
λ(t) f (t)2
v(t) ≤ √ exp − .
2πt3 2t
(f) Suppose that X1 , X2 , . . . is a sequence of independent standard Gaussian
random variables. Show that
! Z  
f (t)2
X t n
λ(t)
P exists t ≤ n : Xs ≥ f (t) ≤ √ exp − dt .
s=1 0 2πt3 2t

(g) Let
p h : (0, ∞) → (1, ∞) be a concave increasing function p such that
log(h(a))/h(a) ≤ c/a for some constant c > 0 and f (t) = 2t log h(1/tδ)+
t∆. Show that
!
2cδ
Xt
P exists t : Xs ≥ f (t) ≤ √ 2 .
s=1
π∆
p
(h) Show that h(a) = 1 + (1 + a) log(1 + a) satisfies the requirements of the
previous part with c = 11/10.
(i) Use your results to modify MOSS for the case when the rewards are Gaussian.
Compare the algorithms empirically.
(j) Prove for your modified algorithm that
Rn X 2
lim sup ≤ .
n→∞ log(n) ∆i
i:∆i >0

Hint The above exercise has several challenging components and assumes
prior knowledge of Brownian motion and its interpretation in terms of the heat
equation. We recommend the book by Lerche [1986] as a nice reference on hitting
times for Brownian motion against concave barriers. The equation you derived in
Part (d) is called the Bachelier–Lévy formula , and the technique for doing
so is the method of images. The use of this theory in bandits was introduced
by one of the authors [Lattimore, 2018], which readers might find useful when
working through these questions.
9.5 (Asymptotic optimality and subgaussian noise) In the last exercise,
 9.5 Exercises 131

you modified MOSS to show asymptotic optimality when the noise is Gaussian.
This is also possible for subgaussian noise. Follow the advice in the notes of this
chapter to adapt MOSS so that for all 1-subgaussian bandits, it holds that
Rn X 2
lim sup ≤ ,
n→∞ log(n) ∆i
i:∆i >0
√
while maintaining the property that Rn ≤ C kn for universal constant C > 0.
10 The Upper Confidence Bound
Algorithm: Bernoulli Noise ( )

In previous chapters we assumed that the noise of the rewards was σ-subgaussian
for some known σ > 0. This has the advantage of simplicity and relative generality,
but stronger assumptions are sometimes justified and often lead to stronger results.
In this chapter the rewards are assumed to be Bernoulli, which just means that
Xt ∈ {0, 1}. This is a fundamental setting found in many applications. For
example, in click-through prediction, the user either clicks on the link or not. A
Bernoulli bandit is characterised by the mean pay-off vector µ ∈ [0, 1]k and the
reward observed in round t is Xt ∼ B(µAt ).
The Bernoulli distribution is 1/2-subgaussian regardless of its mean
(Exercise 5.12). Hence the results of the previous chapters are applicable, and an
appropriately tuned UCB enjoys logarithmic regret. The additional knowledge
that the rewards are Bernoulli is not being fully exploited by these algorithms,
however. The reason is essentially that the variance of a Bernoulli random
variable depends on its mean, and when the variance is small, the empirical mean
concentrates faster, a fact that should be used to make the confidence intervals
smaller.

10.1 Concentration for Sums of Bernoulli Random Variables

The first step when designing a new optimistic algorithm is to construct confidence
sets for the unknown parameters. For Bernoulli bandits, this corresponds to
analysing the concentration of the empirical mean for sums of Bernoulli random
variables. For this, the following definition will prove useful:

Definition 10.1 (Relative entropy between Bernoulli distributions). The

relative entropy between Bernoulli distributions with parameters p, q ∈ [0, 1] is

d(p, q) = p log(p/q) + (1 − p) log((1 − p)/(1 − q)) ,

where singularities are defined by taking limits: d(0, q) = log(1/(1 − q)) and
d(1, q) = log(1/q) for q ∈ [0, 1] and d(p, 0) = 0 if p = 0 and ∞ otherwise and
d(p, 1) = 0 if p = 1 and ∞ otherwise.
 10.1 Concentration for Sums of Bernoulli Random Variables 133

More generally, the relative entropy or Kullback–Leibler divergence

is a measure of similarity between distributions. See Chapter 14 for a generic
definition, interpretation and discussion.

Lemma 10.2. Let p, q, ε ∈ [0, 1]. The following hold:

(a) The functions d(·, q) and d(p, ·) are convex and have unique minimisers at q
and p, respectively.
(b) d(p, q) ≥ 2(p − q)2 (Pinsker’s inequality).
(c) If p ≤ q − ε ≤ q, then d(p, q − ε) ≤ d(p, q) − d(q − ε, q) ≤ d(p, q) − 2ε2 .

Proof We assume that p, q ∈ (0, 1). The corner cases are easily checked
separately. Part (a): d(·, q) is the sum of the negative binary entropy function
h(p) = p log p + (1 − p) log(1 − p) and a linear function. The second derivative
of h is h00 (p) = 1/p + 1/(1 − p), which is positive, and hence h is convex. For
fixed p the function d(p, ·) is the sum of h(p) and convex functions p log(1/q) and
(1 − p) log(1/(1 − q)). Hence d(p, ·) is convex. The minimiser property follows
because d(p, q) > 0 unless p = q in which case d(p, p) = d(q, q) = 0. A more
general version of (b) is given in Chapter 15. A proof of the simple version here
follows by considering the function g(x) = d(p, p + x) − 2x2 , which obviously
satisfies g(0) = 0. The proof is finished by showing that this is the unique
minimiser of g over the interval [−p, 1 − p]. The details are left to Exercise 10.1.
For (c), notice that
q 1−q
h(p) = d(p, q − ε) − d(p, q) = p log + (1 − p) log .
q−ε 1−q+ε
It is easy to see then that h is linear and increasing in its argument. Therefore,
since p ≤ q − ε,

h(p) ≤ h(q − ε) = −d(q − ε, q) ,

as required for the first inequality of (c). The second inequality follows by using
the result in (b).

The next lemma controls the concentration of the sample mean of a sequence
of independent and identically distributed Bernoulli random variables.

Lemma 10.3 (Chernoff’s bound). Let X1 , X2 , . . . , Xn be a sequence of independent

random variables that are Bernoulli distributed with mean µ, and let µ̂ =
1
Pn
n t=1 Xt be the sample mean. Then, for ε ∈ [0, 1 − µ], it holds that

P (µ̂ ≥ µ + ε) ≤ exp (−nd(µ + ε, µ)) (10.1)

and for ε ∈ [0, µ],

P (µ̂ ≤ µ − ε) ≤ exp (−nd(µ − ε, µ)) . (10.2)

 10.1 Concentration for Sums of Bernoulli Random Variables 134

Proof We will again use the Cramér–Chernoff method. Let λ > 0 be some
constant to be chosen later. Then,
n
! !
X
P (µ̂ ≥ µ + ε) = P exp λ (Xt − µ) ≥ exp (λnε)
t=1
Pn
E [exp (λ t=1 (Xt − µ))]
≤
exp (λnε)
= (µ exp(λ(1 − µ − ε)) + (1 − µ) exp(−λ(µ + ε))) .
n

This expression is minimised by λ = log (µ+ε)(1−µ)

µ(1−µ−ε) . Therefore,

P (µ̂ ≥ µ + ε)
 1−µ−ε  −µ−ε !n
(µ + ε)(1 − µ) (µ + ε)(1 − µ)
≤ µ + (1 − µ)
µ(1 − µ − ε) µ(1 − µ − ε)
 1−µ−ε !n
µ (µ + ε)(1 − µ)
=
µ+ε µ(1 − µ − ε)
= exp (−nd(µ + ε, µ)) .
The bound on the left tail is proven identically.
Using Pinsker’s inequality, it follows that P (µ̂ ≥ µ + ε) , P (µ̂ ≤ µ − ε) ≤
exp(−2nε2 ), which is the same as what can be obtained from Hoeffding’s lemma
(see (5.8)). Solving exp(−2nε2 ) = δ, we recover the usual 1 − δ confidence upper
bound. In fact, this cannot be improved when µ ≈ 1/2, but the Chernoff bound
is much stronger when µ is close to either zero or one. Can we invert the Chernoff
tail bound to get confidence intervals that get tighter automatically as µ (or µ̂)
approaches zero or one? The following corollary shows how to do this.
Corollary 10.4. Let µ, µ̂, n be as above. Then, for any a ≥ 0,
P (d(µ̂, µ) ≥ a, µ̂ ≤ µ) ≤ exp(−na) , (10.3)
and P (d(µ̂, µ) ≥ a, µ̂ ≥ µ) ≤ exp(−na) . (10.4)
Furthermore, defining
U (a) = max{u ∈ [0, 1] : d(µ̂, u) ≤ a} ,
and L(a) = min{u ∈ [0, 1] : d(µ̂, u) ≤ a} .
Then, P (µ ≥ U (a)) ≤ exp(−na) and P (µ ≤ L(a)) ≤ exp(−na).
Proof First, we prove (10.3). Note that d(·, µ) is decreasing on [0, µ], and thus,
for 0 ≤ a ≤ d(0, µ), {d(µ̂, µ) ≥ a, µ̂ ≤ µ} = {µ̂ ≤ µ − x, µ̂ ≤ µ} = {µ̂ ≤ µ − x},
where x is the unique solution to d(µ − x, µ) = a on [0, µ]. Hence, by Eq. (10.2)
of Lemma 10.3, P (d(µ̂, µ) ≥ a, µ̂ ≤ µ) ≤ exp(−na). When a ≥ d(0, µ), the
inequality trivially holds. The proof of (10.4) is entirely analogous and hence
is omitted. For the second part of the corollary, fix a and let U = U (a).
 10.1 Concentration for Sums of Bernoulli Random Variables 135

First, notice that U ≥ µ̂ and d(µ̂, ·) is strictly increasing on [µ̂, 1]. Hence,
{µ ≥ U } = {µ ≥ U, µ ≥ µ̂} = {d(µ̂, µ) ≥ d(µ̂, U ), µ ≥ µ̂} = {d(µ̂, µ) ≥ a, µ ≥ µ̂},
where the last equality follows by d(µ̂, U ) = a, which holds by the definition
of U . Taking probabilities and using the first part of the corollary shows that
P (µ ≥ U ) ≤ exp(−na). The statement concerning L = L(a) follows with a similar
reasoning.

Note that for δ ∈ (0, 1), U = U (log(1/δ)/n) and L = L(log(1/δ)/n) are upper
and lower confidence bounds for µ. Although the relative entropy has no closed-
form inverse, the optimisation problem that defines U and L can be solved to a
high degree of accuracy using Newton’s method (the relative entropy d is convex
in its second argument). The advantage of this confidence interval relative to
the one derived from Hoeffding’s bound is now clear. As µ̂ approaches one, the
width of the interval U (a) − µ̂ approaches zero,
p whereas the width of the interval
provided by Hoeffding’s bound stays at log(1/δ)/(2n). The same holds for
µ̂ − L(a) as µ̂ → 0.

Example 10.5. Fig. 10.1 shows a plot of d(3/4, x) and the lower bound given
by Pinsker’s inequality. The approximation degrades as |x − 3/4| grows large,
especially for x > 3/4. As explained in Corollary 10.4, the graph of d(µ̂, ·) can
be used to derive confidence bounds by solving for d(µ̂, x) = a = log(1/δ)/n.
Assuming µ̂ = 3/4 is observed, a confidence level of 90 per cent with n = 10,
a ≈ 0.23. The confidence interval can be read out from the figure by finding
those values where the horizontal dashed black line intersects the solid blue line.
The resulting confidence interval will be highly asymmetric. Note that in this
scenario, the lower confidence bounds produced by both Hoeffding’s inequality
and Chernoff’s bound are similar, while the upper bound provided by Hoeffding’s
bound is vacuous.

d(3/4, x)
0.6 2(x − 3/4)2
a = 0.23
0.4

0.2

0
0 0.25 0.5 0.75 1
x

Figure 10.1 Relative entropy and Pinsker’s inequality

 10.2 The KL-UCB Algorithm 136

10.2 The KL-UCB Algorithm

The difference between KL-UCB and UCB is that Chernoff’s bound is used to
define the upper confidence bound instead of Lemma 5.5.

1: Input k
2: Choose each arm once
3: Subsequently choose
 
log f (t)
At = argmaxi max µ̃ ∈ [0, 1] : d(µ̂i (t − 1), µ̃) ≤ ,
Ti (t − 1)
where f (t) = 1 + t log2 (t) .

Algorithm 8: KL-UCB.

Theorem 10.6. If the reward in round t is Xt ∼ B(µAt ), then the regret of

Algorithm 8 is bounded by
 
X log(f (n)) 1 2
Rn ≤ inf ∆i + + 2 .
ε1 ,ε2 >0 d(µi + ε1 , µ∗ − ε2 ) 2ε21 ε2
i:∆i >0 ε1 +ε2 ∈(0,∆i )

Rn X ∆i
Furthermore, lim sup ≤ .
n→∞ log(n) d(µi , µ∗ )
i:∆i >0

Comparing the regret in Theorem 10.6 to what would be obtained when using
UCB from Chapter 8, which for subgaussian constant σ = 1/2 satisfies
Rn X 1
lim sup ≤ .
n→∞ log(n) 2∆i
i:∆i >0

By Pinsker’s inequality (part (b) of Lemma 10.2) we see that d(µi , µ∗ ) ≥

2(µ∗ − µi )2 = 2∆2i , which means that the asymptotic regret of KL-UCB is
never worse than that of UCB. On the other hand, a Taylor’s expansion shows
that when µi and µ∗ are close (the hard case in the asymptotic regime),
∆2i
d(µi , µ∗ ) = + o(∆2i ) ,
2µi (1 − µi )
indicating that the regret of KL-UCB is approximately
Rn X 2µi (1 − µi )
lim sup ≈ . (10.5)
n→∞ log(n) ∆i
i:∆i >0

Notice that µi (1 − µi ) is the variance of a Bernoulli distribution with mean µi .

The approximation indicates that KL-UCB will improve on UCB in regimes
where µi is close to zero or one.
The proof of Theorem 10.6 relies on two lemmas. The first is used to show
that the index of the optimal arm is never too far below its true value, while the
 10.2 The KL-UCB Algorithm 137

second shows that the index of any other arm is not often much larger than the
same value. These results mirror those given for UCB, but things are complicated
by the non-symmetric and hard-to-invert divergence function.
For the next results, we define d(p, q) = d(p, q)I {p ≤ q}.

Lemma 10.7. Let X1 , X2 , . . . , Xn be independent Bernoulli random variables with

mean µ ∈ [0, 1], ε > 0 and
 
log f (t)
τ = min t : max d(µ̂s , µ − ε) − ≤0 .
1≤s≤n s
2
Then, E[τ ] ≤ .
ε2
Proof We start with a high-probability bound and then integrate to control the
expectation.
 
log f (t)
P (τ > t) ≤ P ∃1 ≤ s ≤ n : d(µ̂s , µ − ε) >
s
 
log f (t)
Xn
≤ P d(µ̂s , µ − ε) >
s=1
s
 
log f (t)
Xn
= P d(µ̂s , µ − ε) > , µ̂s < µ − ε
s=1
s
 
log f (t)
Xn
≤ P d(µ̂s , µ) > + 2ε2 , µ̂s < µ ((c) of Lemma 10.2)
s=1
s
  
log f (t)
Xn
2
≤ exp −s 2ε + (Eq. (10.3) of Corollary 10.4)
s=1
s
1 X
n


≤ exp −2sε2
f (t) s=1
1
≤ .
2f (t)ε2
To finish, we integrate the tail,
Z ∞ Z ∞
1 dt 2
E[τ ] = P (τ ≥ t) dt ≤ 2 ≤ 2.
0 2ε 0 f (t) ε
Lemma 10.8. Let X1 , X2 , . . . , Xn be independent Bernoulli random variables with
mean µ. Further, let ∆ > 0, a > 0 and define
Xn n ao
κ= I d(µ̂s , µ + ∆) ≤ .
s=1
s
 
a 1
Then, E[κ] ≤ inf + .
ε∈(0,∆) d(µ + ε, µ + ∆) 2ε2
 10.2 The KL-UCB Algorithm 138

Proof Let ε ∈ (0, ∆) and u = a/d(µ + ε, µ + ∆). Then,

Xn  a
E[κ] = P d(µ̂s , µ + ∆) ≤
s=1
s
Xn  a
≤ P µ̂s ≥ µ + ε or d(µ + ε, µ + ∆) ≤
s=1
s
(d(·, µ + ∆) is decreasing on [0, µ + ∆])
n
X
≤u+ P (µ̂s ≥ µ + ε)
s=due
∞
X
≤u+ exp (−sd(µ + ε, µ)) (Lemma 10.3)
s=1
a 1
≤ +
d(µ + ε, µ + ∆) d(µ + ε, µ)
a 1
≤ + 2 (Pinsker’s inequality/Lemma 10.2(b))
d(µ + ε, µ + ∆) 2ε
as required.

Proof of Theorem 10.6 As in other proofs, we assume without loss of generality

that µ1 = µ∗ and bound E[Ti (n)] for suboptimal arms i. To this end, fix a
suboptimal arm i and let ε1 + ε2 ∈ (0, ∆i ) with both ε1 and ε2 positive. Define
 
log f (t)
τ = min t : max d(µ̂1s , µ1 − ε2 ) − ≤ 0 , and
1≤s≤n s
 
log f (n)
Xn
κ= I d(µ̂is , µi + ∆i − ε2 ) ≤ .
s=1
s

Using a similar argument as in the proof of Theorem 8.1,

" n #
X
E[Ti (n)] = E I {At = i}
t=1
" n
#
X
≤ E[τ ] + E I {At = i}
t=τ +1
" n  #
X log f (t)
≤ E[τ ] + E I At = i and d(µ̂i,Ti (t−1) , µ1 − ε2 ) ≤
t=1
Ti (t − 1)
≤ E[τ ] + E[κ]
2 f (n) 1
≤ 2 + + 2,
ε2 d(µi + ε1 , µ − ε2 ) 2ε1
∗

where the second inequality follows, since by the definition of τ , if t > τ , then
the index of the optimal arm is at least as large as µ1 − ε2 . The third inequality
follows from the definition of κ as in the proof of Theorem 8.1. The final inequality
 10.3 Notes 139

follows from Lemmas 10.7 and 10.8. The first claim of the theorem is completed
by substituting the above into the standard regret decomposition
k
X
Rn = ∆i E[Ti (n)] .
i=1

The asymptotic claim for you in Exercise 10.2.

10.3 Notes

1 The new concentration inequality (Lemma 10.3) holds more generally for
any sequence of independent and identically distributed random variables
X1 , X2 , . . . , Xn for which Xt ∈ [0, 1] almost surely. Therefore all results in
this section also hold if the assumption that the noise is Bernoulli is relaxed
to the case where it is simply supported in [0, 1] (or other bounded sets by
shifting/scaling).
2 Expanding on the previous note, all that is required is a bound on the moment-
generating function for random variables X where, X ∈ [0, 1] almost surely.
Garivier and Cappé [2011, Lemma 9] noted that f (x) = exp(λx) − x(exp(λ) −
1) − 1 is negative on [0, 1], and so
E [exp(λX)] ≤ E [X(exp(λ) − 1) + 1] = µ exp(λ) + 1 − µ ,
which is precisely the moment-generating function of the Bernoulli distribution
with mean µ. Then the remainder of the proof of Lemma 10.3 goes through
unchanged. This shows that for any bandit ν = (Pi )i with Supp(Pi ) ∈ [0, 1] for
all i the regret of the policy in Algorithm 8 satisfies
Rn X ∆i
lim sup ≤ .
n→∞ log(n) d(µi , µ∗ )
i:∆i >0

3 The bounds obtained using the argument in the previous note are not quite
tight. Specifically one can show there exists an algorithm such that for all
bandits ν = (Pi )i with Pi , the reward distribution of the ith arm supported on
[0, 1], then
Rn X ∆i
lim sup = , where
n→∞ log(n) di
i:∆i >0

di = inf{D(Pi , P ) : µ(P ) > µ∗ and Supp(P ) ⊂ [0, 1]}

and D(P, Q) is the relative entropy between measures Pi and P , which we
define in Chapter 14. The quantity di is never smaller than d(µi , µ∗ ). For details
on this, see the paper by Honda and Takemura [2010].
4 The approximation in Eq. (10.5) was used to show that the regret for KL-UCB
is closely related to the variance of the Bernoulli distribution. It is natural to
ask whether or not this result could be derived, at least asymptotically, by
 10.4 Bibliographic Remarks 140

appealing to the central limit theorem. The answer is no. First, the quality of
the approximation in Eq. (10.5) does not depend on n, so asymptotically it is
not true that the Bernoulli bandit behaves like a Gaussian bandit with variances
tuned to match. The reason is that as n tends to infinity, the confidence level
should be chosen so that the risk of failure also tends to zero. But the central
limit theorem does not provide information about the tails with probability
mass less than O(n−1/2 ). See Note 1 in Chapter 5.
5 The analysis in this chapter is easily generalised to a wide range of alternative
noise models. You will do this for single-parameter exponential families in
Exercises 10.4, 10.5 and 34.5.
6 Chernoff credits Lemma 10.3 to his friend Herman Rubin [Chernoff, 2014], but
the name seems to have stuck.

10.4 Bibliographic Remarks

Several authors have worked on Bernoulli bandits, and the asymptotics have
been well understood since the article by Lai and Robbins [1985]. The earliest
version of the algorithm presented in this chapter is due to Lai [1987], who
provided asymptotic analysis. The finite-time analysis of KL-UCB was given by
two groups simultaneously (and published in the same conference) by Garivier
and Cappé [2011] and Maillard et al. [2011] (see also the combined journal article:
Cappé et al. 2013). Two alternatives are the DMED [Honda and Takemura, 2010]
and IMED [Honda and Takemura, 2015] algorithms. These works go after the
problem of understanding the asymptotic regret for the more general situation
where the rewards lie in a bounded interval (see Note 3). The latter work
covers even the semi-bounded case where the rewards are almost surely upper-
bounded. Both algorithms are asymptotically optimal. Ménard and Garivier
[2017] combined MOSS and KL-UCB to derive an algorithm that is minimax
optimal and asymptotically optimal for single-parameter exponential families.
While the subgaussian and Bernoulli examples are very fundamental, there has
also been work on more generic set-ups where the unknown reward distribution for
each arm is known to lie in some class F. The article by Burnetas and Katehakis
[1996] gives the most generic (albeit, asymptotic) results. These generic set-ups
remain wide open for further work.

10.5 Exercises

10.1 (Pinsker’s inequality) Prove Lemma 10.2(b).

Hint Consider the function g(x) = d(p, p + x) − 2x2 over the [−p, 1 − p] interval.
By taking derivatives, show that g ≥ 0.
10.2 (Asymptotic optimality) Prove the asymptotic claim in Theorem 10.6.
 10.5 Exercises 141

Hint Choose ε1 , ε2 to decrease slowly with n and use the first part of the
theorem.

10.3 (Concentration for bounded random variables) Let F = (Ft )t be a

filtration, (Xt )t be [0, 1]-valued, F-adapted sequence, such that E [Xt | Ft−1 ] = µt
Pn
for some µ1 , . . . , µn ∈ [0, 1] non-random numbers. Define µ = n1 t=1 µt ,
P
µ̂ = n1 t=1 Xt . Prove that the conclusion of Lemma 10.3 still holds.
n

Hint Read Note 2 at the end of this chapter. Let g(·, µ) be the cumulant-
generating function of the µ-parameter Bernoulli distribution. For X ∼ B(µ),
λ ∈ R, g(λ, µ) = log E [exp(λX)]. Show that g(λ, ·) is concave. Next, use this and
the tower rule to show that E [exp(λn(µ̂ − µ))] ≤ g(λ, µ)n .

The bound of the previous exercise is most useful when all µt are either all
close to zero or they are all close to one. When half of the {µt } are close to
zero and the other half close to one, then the bound degrades to Hoeffding’s
bound.

10.4 (KL-UCB for exponential families) Let M = {Pθ : θ ∈ Θ} be a

regular non-singular exponential family with sufficient statistic S(x) = x and
E = {(Pθi )ki=1 : θ ∈ Θk } be the set of bandits with reward distributions in M.
Design a policy π such that for all ν ∈ E, it holds that

Rn (π, ν) X ∆i
lim ≤ ,
n→∞ log(n) di,inf
i:∆i >0

R
where µ(θ) = R xdPθ (x) is the mean of Pθ and di,inf = inf{d(θ, φ) : µ(φ) >
µ∗ , φ ∈ Θ}, with d(θ, φ) the relative entropy between Pθ and Pφ .

Hint Readers not familiar with exponential families should skip ahead to
Section 34.3.1 and then do Exercise 34.5. For the exercise, repeat the proof of
Theorem 10.6, adapting as necessary. See also the paper by Cappé et al. [2013].

10.5 (KL-UCB for non-canonical exponential families) Repeat the

previous exercise, but relax the assumption that S(x) = x.

Hint This is a subtle problem. You should adapt the algorithm so that if there
are ties in the upper confidence bounds, then an arm with the largest number of
plays is chosen. A solution is available. Korda et al. [2013] analysed
R Thompson
sampling in this setting. Their result only holds when θ 7→ R xpθ (x)dh(x) is
invertible, which does not always hold.
 10.5 Exercises 142

In the analysis of KL-UCB for canonical exponential families, the asymptotic

rate is a good indicator of the finite-time regret in the sense that the o(log(n))
term hidden by the asymptotics has roughly the same leading constant as
the dominant term. By contrast, the analysis here indicates that
log(n) 1
E[Ti (n)] ≈ + ,
di,inf di,min
where di,min = di,min (0). Although the latter term is negligible asymptotically,
it may be the dominant term for all reasonable n.

10.6 (Comparison to UCB) In this exercise, you compare KL-UCB and UCB
empirically.
(a) Implement Algorithm 8 and Algorithm 6, where the latter algorithm should
be tuned for 1/2-subgaussian bandits so that
s
log(f (t))
At = argmaxi∈[k] µ̂i (t − 1) + .
2Ti (t − 1)
(b) Let n = 10000 and k = 2. Plot the expected regret of each algorithm as a
function of ∆ when µ1 = 1/2 and µ2 = 1/2 + ∆.
(c) Repeat the above experiment with µ1 = 1/10 and µ1 = 9/10.
(d) Discuss your results.
Part III
Adversarial Bandits with
Finitely Many Arms
 144

Statistician George E. P. Box is famous for writing that ‘all models are wrong,
but some are useful’. In the stochastic bandit model the reward is sampled from
a distribution that depends only on the chosen action. It does not take much
thought to realise this model is almost always wrong. At the macroscopic level
typically considered in bandit problems, there is not much that is stochastic
about the world. And even if there were, it is hard to rule out the existence of
other factors influencing the rewards.
The quotation suggests we should not care whether or not the stochastic bandit
model is right, only whether it is useful. In science, models are used for predicting
the outcomes of future experiments, and their usefulness is measured by the
quality of the predictions. But how can this be applied to bandit problems? What
predictions can be made based on bandit models? In this respect, we postulate
the following:

The point of bandit models is to facilitate predicting the performance of

bandit algorithms on future problem instances that one encounters in their
practice.

A model can fail in two fundamentally different ways. It can be too specific,
imposing assumptions so detached from reality that a catastrophic mismatch
between actual and predicted performance may arise. The second mode of failure
occurs when a model is too general, which makes the algorithms designed to do
well on the bandit model overly cautious, which can harm performance.
Not all assumptions are equally important. It is a critical assumption in
stochastic bandits that the mean reward of individual arms does not change
(significantly) over time. On the other hand, the assumption that a single, arm-
dependent distribution generates the rewards for a given arm plays a relatively
insignificant role. The reader is encouraged to think of cases when the constancy
of arm distributions plays no role, and also of cases when it does – furthermore, to
decide to what extent the algorithms can tolerate deviations from the assumption
that the means of arms stay the same. Stochastic bandits where the means of
the arms are changing over time are called non-stationary and are the topic of
Chapter 31.
If a highly specialised model is actually correct, then the resulting algorithms
usually dominate algorithms derived for a more general model. This is a general
manifestation of the bias-variance trade-off, well known in supervised learning
and statistics. The holy grail is to find algorithms that work ‘optimally’ across
a range of models. The reader should think about examples from the previous
chapters that illustrate these points.
The usefulness of the stochastic model depends on the setting. In particular,
the designer of the bandit algorithm must carefully evaluate whether stochasticity,
stability of the mean and independence are reasonable assumptions. For some
applications, the answer will probably be yes, while in others the practitioner
 145

may seek something more robust. This latter situation is the topic of the next
few chapters.

Adversarial Bandits

The adversarial bandit model abandons almost all the assumptions on how
the rewards are generated, so much so that the environment is often called the
adversary. The adversary has a great deal of power in this model, including the
ability to examine the code of the proposed algorithms and choose the rewards
accordingly. All that is kept from the previous chapters is that the objective will
be framed in terms of how well a policy is able to compete with the best action
in hindsight.
At first sight, it seems remarkable that one can say anything at all about such
a general model. And yet it turns out that this model is not much harder than
the stochastic bandit problem. Why this holds and how to design algorithms that
achieve these guarantees will be explained in the following chapters.
To give you a glimmer of hope, imagine playing the following simple bandit
game with a friend. The horizon is n = 1, and you have two actions. The game
proceeds as follows:
1 You tell your friend your strategy for choosing an action.
2 Your friend secretly chooses rewards x1 ∈ {0, 1} and x2 ∈ {0, 1}.
3 You implement your strategy to select A ∈ {1, 2} and receive reward xA .
4 The regret is R = max{x1 , x2 } − xA .
Clearly, if your friend chooses x1 = x2 , then your regret is zero no matter what.
Now let’s suppose you implement the deterministic strategy A = 1. Then your
friend can choose x1 = 0 and x2 = 1, and your regret is R = 1. The trick to
improve on this is to randomise. If you tell your friend, ‘I will choose A = 1 with
probability one half’, then the best she can do is choose x1 = 1 and x2 = 0 (or
reversed), and your expected regret is R = 1/2. You are forgiven if you did not
settle on this solution yourself because we did not tell you that a strategy may
be randomised. With such a short horizon, you cannot do better than this, but
for longer games the relative advantage of the adversary decreases, as we shall
see soon.
In the next two chapters, we investigate the k-armed adversarial model in detail,
providing both algorithms and regret analysis. Like the stochastic model, the
adversarial model has many generalisations, which we’ll visit in future chapters.

Bibliographic Remarks

The quote by George Box was used several times with different phrasings [Box,
1976, 1979]. The adversarial framework has its roots in game theory, with familiar
 146

names like Hannan [1957] and Blackwell [1954] producing some of the early
work. The non-statistical approach has enjoyed enormous popularity since the
1990’s and has been adopted wholeheartedly by the theoretical computer science
community [Vovk, 1990, Littlestone and Warmuth, 1994, and many, many others].
The earliest work on adversarial bandits is by Auer et al. [1995]. There is now a
big literature on adversarial bandits, which we will cover in more depth in the
chapters that follow. There has been a lot of effort to move away from stochastic
assumptions. An important aspect of this is to define a sense of regularity for
individual sequences. We refer the reader to some of the classic papers by Martin-
Löf [1966] and Levin [1973] and the more recent paper by Ivanenko and Labkovsky
[2013].
11 The Exp3 Algorithm

In this chapter we first introduce the formal model of adversarial bandit

environments and discuss the relationship to the stochastic bandit model. This is
followed by the discussion of importance-weighted estimation, the Exp3 algorithm
that uses this technique and the analysis of the regret of Exp3.

11.1 Adversarial Bandit Environments

Let k > 1 be the number of arms. A k-armed

adversarial bandit is an arbitrary sequence of
reward vectors (xt )nt=1 , where xt ∈ [0, 1]k . In each
round, the learner chooses a distribution over the
actions Pt ∈ Pk−1 . Then the action At ∈ [k] is
sampled from Pt , and the learner receives reward
xtAt . The interaction protocol is summarised in
Fig. 11.2.
A policy in this setting is a function π : ([k] ×
[0, 1])∗ → Pk−1 mapping history sequences to dis-
tributions over actions (regardless of measurability). Figure 11.1 Would you play
The performance of a policy π in environment x is with this multi-armed bandit?
measured by the expected regret, which is the expected loss in revenue of the
policy relative to the best fixed action in hindsight.
n
" n #
X X
Rn (π, x) = max xti − E xtAt , (11.1)
i∈[k]
t=1 t=1

Adversary secretly chooses rewards (xt )n

t=1 with xt ∈ [0, 1]
k

For rounds t = 1, 2, . . . , n:
Learner selects distribution Pt ∈ Pk−1 and samples At from Pt .
Learner observes reward Xt = xtAt .

Figure 11.2 Interaction protocol for k-armed adversarial bandits.

 11.1 Adversarial Bandit Environments 148

where the expectation is over the randomness of the learner’s actions. The
arguments π and x are omitted from the regret when they are clear from context.

The only source of randomness in the regret comes from the randomness in
the actions of the learner. Of course the interaction with the environment
means the action chosen in round t may depend on actions s < t as well as
the observed rewards until round t. As we noted, unlike the case of stochastic
bandits, here, there is no measurability restriction on the learner’s policy π.
This is actually by choice, see Note 12 for details.

The worst-case regret over all environments is

Rn∗ (π) = sup Rn (π, x) .
x∈[0,1]n×k

The main question is whether or not there exist policies π for which Rn∗ (π) is
sublinear in n. In Exercise 11.2 you will show that for deterministic policies
Rn∗ (π) ≥ n(1 − 1/k), which follows by constructing a bandit so that xtAt = 0 for
all t and xti = 1 for i 6= At . Because of this, sublinear worst-case regret is only
possible by using a randomised policy.

Readers familiar with game theory will not be surprised by the need for
randomisation. The interaction between learner and adversarial bandit can be
framed as a two-player zero-sum game between the learner and environment.
The moves for the environment are the possible reward sequences, and for
the player they are the policies. The pay-off for the environment/learner is
the regret and its negation respectively. Since the player goes first, the only
way to avoid being exploited is to choose a randomised policy.

While stochastic and adversarial bandits seem quite different, it turns out that the
optimal worst-case regret is the same up to constant factors and that lower bounds
for adversarial bandits are invariably derived in the same manner as for stochastic
bandits (see Part IV). In this chapter, we present a simple algorithm for which
the worst-case regret is suboptimal by just a logarithmic factor. First, however,
we explore the differences and similarities between stochastic and adversarial
environments.
We already noted that deterministic strategies will have linear regret for
some adversarial bandit. Since strategies in Part II like UCB and ‘Explore-then-
Commit’ were deterministic, they are not well suited for the adversarial setting.
This immediately implies that policies that are good for stochastic bandit can
be very suboptimal in the adversarial setting. What about the other direction?
Will an adversarial bandit strategy have small expected regret in the stochastic
setting? Let π be an adversarial bandit policy and ν = (ν1 , . . . , νk ) be a stochastic
bandit with Supp(νi ) ⊆ [0, 1] for all i. Next, let Xti be sampled from νi for each
 11.2 Importance-Weighted Estimators 149

i ∈ [k] and t ∈ [n], and assume these random variables are mutually independent.
By Jensen’s inequality and convexity of the maximum function, we have
" n #
X
Rn (π, ν) = max E (Xti − XtAt )
i∈[k]
t=1
" n
#
X
≤ E max (Xti − XtAt )
i∈[k]
t=1

= E [Rn (π, X)] ≤ Rn∗ (π) , (11.2)

where the regret in the first line is the stochastic regret (using the random table
model), and in the last it is the adversarial regret. Therefore the worst-case
stochastic regret is upper-bounded by the worst-case adversarial regret. Going
the other way, the above inequality also implies that the worst-case regret for
adversarial problems is lower-bounded by the worst-case regret on stochastic
problems with rewards bounded in [0, 1]. In Chapter 15, we √ prove that the worst-
case regret for stochastic Bernoulli bandits is at least c nk, where c > 0 is a
universal constant (Exercise 15.4). And so for the same universal constant, the
minimax regret for adversarial bandits satisfies
√
Rn∗ = inf sup Rn (π, x) ≥ c nk .
π x∈[0,1]n×k

There is a little subtlety here. In order to define the expectations in the stochastic
regret, the policy should be appropriately measurable. This can be resolved by
noting that lower bounds can be proven using Bernoulli bandits. For details, see
again Note 12.

11.2 Importance-Weighted Estimators

A key ingredient of all adversarial bandit algorithms is a mechanism for estimating

the reward of unplayed arms. Recall that Pt is the conditional distribution of the
action played in round t, and so for i ∈ [k], Pti is the conditional probability

Pti = P (At = i | A1 , X1 , . . . , At−1 , Xt−1 ) .

In what follows, we assume that for all t and i, Pti > 0 almost surely. As we
shall see later, this will be true for all policies considered in this chapter. The
importance-weighted estimator of xti is
I {At = i} Xt
X̂ti = . (11.3)
Pti
Let Et [·] = E[· | A1 , X1 , . . . , At , Xt ] denote the conditional expectation given the
history up to time t. The conditional mean of X̂ti satisfies

Et−1 [X̂ti ] = xti , (11.4)

 11.2 Importance-Weighted Estimators 150

which means that X̂ti is an unbiased estimate of xti conditioned on the history
observed after t − 1 rounds. To see why Eq. (11.4) holds, let Ati = I {At = i} so
that Xt Ati = xti Ati and
Ati
X̂ti = xti .
Pti
Now, Et−1 [Ati ] = Pti , and since Pti is σ(A1 , X1 , . . . , At−1 , Xt−1 )-measurable,
 
Ati xti xti
Et−1 [X̂ti ] = Et−1 xti = Et−1 [Ati ] = Pti = xti .
Pti Pti Pti
Being unbiased is a good start, but the variance of an estimator is also important.
For arbitrary random variable U , the conditional variance Vt−1 [U ] is the random
variable
 
Vt−1 [U ] = Et−1 (U − Et−1 [U ])2 .

So Vt−1 [X̂ti ] is a random variable that measures the variance of X̂ti conditioned
on the past. Calculating the conditional variance using the definition of X̂ti and
Eq. (11.4) shows that
 
2 Ati x2ti x2 (1 − Pti )
Vt−1 [X̂ti ] = Et−1 [X̂ti ] − x2ti = Et−1 2 − x2ti = ti . (11.5)
Pti Pti
This can be extremely large when Pti is small and xti is bounded away from zero.
In the notes and exercises, we shall see to what extent this can cause trouble.
The estimator in (11.3) is the first that comes to mind, but there are alternatives.
For example,
I {At = i}
X̂ti = 1 − (1 − Xt ) . (11.6)
Pti
This estimator is still unbiased. Rewriting the formula in terms of yti = 1 − xti
and Yt = 1 − Xt and Ŷti = 1 − X̂ti leads to
I {At = i}
Ŷti = Yt .
Pti
This is the same as (11.3) except that Yt has replaced Xt . The terms yti , Yt and
Ŷti should be interpreted as losses. Had we started with losses to begin with, then
this would have been the estimator that first came to mind. For obvious reasons,
the estimator in Eq. (11.6) is called the loss-based importance-weighted
estimator. The conditional variance of this estimator is essentially the same as
Eq. (11.5):
2 1 − Pti
Vt [X̂ti ] = Vt [Ŷti ] = yti .
Pti
2
The only difference is that the variance now depends on yti rather than x2ti . Which
is better depends on the rewards for arm i, with smaller rewards suggesting the
superiority of the first estimator and larger rewards (or small losses) suggesting
the superiority of the second estimator. Can we change the estimator (either one
 11.3 The Exp3 Algorithm 151

of them) so that it is more accurate for actions whose reward is close to some
specific value v? Of course! Just change the estimator so that v is subtracted
from the observed reward (or loss), then use the importance-sampling formula,
and subsequently add back v. The problem is that the optimal value of v depends
on the unknown quantity being estimated. Also note that the dependence of the
variance on Pti is the same for both estimators, and since the rewards are bounded,
it is this term that usually contributes most significantly. In Exercise 11.5, we ask
you to show that all unbiased estimators in this setting are importance-weighted
estimators.

Although the two estimators seem quite similar, it should be noted that the
first estimator takes values in [0, ∞) while the second takes values in (−∞, 1].
Soon we will see that this difference has a big impact on the usefulness of
these estimators when used in the Exp3 algorithm.

11.3 The Exp3 Algorithm

The simplest algorithm for adversarial bandits is called Exp3, which stands
for ‘exponential-weight algorithm for exploration and exploitation’. The reason
for this name will become clear after the explanation of the algorithm. Let
Pt
Ŝti = s=1 X̂si be the total estimated reward by the end of round t, where X̂si is
given in Eq. (11.6). It seems natural to play actions with larger estimated reward
with higher probability. While there are many ways to map Ŝti into probabilities,
a simple and popular choice is called exponential weighting, which for tuning
parameter η > 0 sets

exp(η Ŝt−1,i )
Pti = Pk . (11.7)
j=1 exp(η Ŝt−1,j )

The parameter η is called the learning rate. When the learning rate is large, Pt
concentrates about the arm with the largest estimated reward and the resulting
algorithm exploits aggressively. For small learning rates, Pt is more uniform,
and the algorithm explores more frequently. Note that as Pt concentrates, the
variance of the importance-weighted estimators for poorly performing arms
increases dramatically. There are many ways to tune the learning rate, including
allowing it to vary with time. In this chapter we restrict our attention to the
simplest case by choosing η to depend only on the number of actions k and the
horizon n. Since the algorithm depends on η, this means that the horizon must
be known in advance, a requirement that can be relaxed (see Note 10).
 11.4 Regret Analysis 152

1: Input: n, k, η
2: Set Ŝ0i = 0 for all i
3: for t = 1, . . . , n do
4: Calculate the sampling distribution Pt :
 
exp η Ŝt−1,i
Pti = P  
exp
k
j=1 η Ŝt−1,j

5: Sample At ∼ Pt and observe reward Xt

6: Calculate Ŝti :
I {At = i} (1 − Xt )
Ŝti = Ŝt−1,i + 1 −
Pti
7: end for
Algorithm 9: Exp3.

11.4 Regret Analysis

We are now ready to bound the expected regret of Exp3.

Theorem 11.1. Letp x ∈ [0, 1]
n×k
and π be the policy of Exp3 (Algorithm 9) with
learning rate η = log(k)/(nk). Then,
p
Rn (π, x) ≤ 2 nk log(k) .
As we will prove many variants of this result with various tools, here we give a
short algebraic proof, saving the development of intuition for later.
Proof For any arm i, define
n
" n #
X X
Rni = xti − E Xt ,
t=1 t=1

which is the expected regret relative to using action i in all the rounds. The
result will follow by bounding Rni for all i, including the optimal arm. For the
remainder of the proof, let i be some fixed arm. By the unbiasedness property of
the importance-weighted estimator X̂ti ,
n
X k
X k
X
E[Ŝni ] = xti and also Et−1 [Xt ] = Pti xti = Pti Et−1 [X̂ti ] .
t=1 i=1 i=1
(11.8)
The tower rule says that for any random variable X, E[Et−1 [X]] = E[X], which
together with the linearity of expectation and Eq. (11.8) means that
" n k #
h i XX h i
Rni = E Ŝni − E Pti X̂ti = E Ŝni − Ŝn , (11.9)
t=1 i=1
 11.4 Regret Analysis 153

P P
where the last equality serves as the definition of Ŝn = t i Pti X̂ti . To bound
the right-hand side of Eq. (11.9), let
k
X  
Wt = exp η Ŝtj .
j=1

By convention an empty sum is zero, which means that Ŝ0j = 0 and W0 = k.

Then,
k
X Yn
W1 Wn Wt
exp(η Ŝni ) ≤ exp(η Ŝnj ) = Wn = W0 ... =k . (11.10)
j=1
W0 Wn−1 t=1
Wt−1

The ratio in the product can be rewritten in terms of Pt by

Wt X exp(η Ŝt−1,j )
k X k
= exp(η X̂tj ) = Ptj exp(η X̂tj ) . (11.11)
Wt−1 j=1
Wt−1 j=1

We need the following facts:

exp(x) ≤ 1 + x + x2 for all x ≤ 1 and 1 + x ≤ exp(x) for all x ∈ R .

Using these two inequalities leads to

Xk Xk
Wt 2 2
≤1+η Ptj X̂tj + η Ptj X̂tj
Wt−1 j=1 j=1
 
Xk Xk
≤ exp η Ptj X̂tj + η 2 2
Ptj X̂tj . (11.12)
j=1 j=1

Notice that this was only possible because X̂tj is defined by Eq. (11.6), which
ensures that X̂tj ≤ 1 and would not have been true had we used Eq. (11.3).
Combining Eq. (11.12) and Eq. (11.10),
 
  n X
X k
exp η Ŝni ≤ k exp η Ŝn + η 2 2
Ptj X̂tj .
t=1 j=1

Taking the logarithm of both sides, dividing by η > 0 and reordering gives

log(k) Xn X k
2
Ŝni − Ŝn ≤ +η Ptj X̂tj . (11.13)
η t=1 j=1

As noted earlier, the expectation of the left-hand side is Rni . The first term on
the right-hand side is a constant, which leaves us to bound the expectation of the
second term. Letting ytj = 1 − xtj and Yt = 1 − Xt and expanding the definition
 11.4 Regret Analysis 154

2
of X̂tj leads to
   
 2
=
Xk X k
2 I {A j} y
= E Ptj 1 −
t tj
E Ptj X̂tj 
j=1 j=1
Ptj
 !
2
X k
I {A = j} y I {At = j} ytj
= E Ptj 1 − 2 +
t tj 
j=1
Ptj Ptj2
  
2
X k
I {At = j} ytj
= E 1 − 2Yt + Et−1  
j=1
Ptj
 
Xk
2 
= E 1 − 2Yt + ytj
j=1
 
X
= E (1 − Yt )2 + 2 
ytj
j6=At

≤ k.

Summing over t, and then substituting into Eq. (11.13), we get

log(k) p
Rni ≤ + ηnk = 2 nk log(k) ,
η
p
where the equality follows by substituting η = log(k)/(nk), which was chosen
to optimise this bound.

At the heart of the proof are the inequalities:

1 + x ≤ exp(x) for all x ∈ R and exp(x) ≤ 1 + x + x2 for x ≤ 1 .

The former of these inequalities is an ansatz derived from the first-order Taylor
expansion of exp(x) about x = 0. The latter, however, is not the second-order
Taylor expansion, which would be 1 + x + x2 /2. The problem is that the second-
order Taylor series is not an upper bound on exp(x) for x ≤ 1, but only for
x ≤ 0:
1
exp(x) ≤ 1 + x + x2 for all x ≤ 0 . (11.14)
2
But it is nearly an upper bound, and this can be exploited to improve the bound
in Theorem 11.1. The mentioned upper and lower bounds on exp(x) are shown
in Fig. 11.3, from which it is quite obvious that the bound in Eq. (11.14) is
significantly tighter when x ≤ 0.
Let us now put Eq. (11.14) to use in proving the following√improved version of
Theorem 11.1, for which the regret is smaller by a factor of 2. The algorithm is
unchanged except for a slightly increased learning rate.
 11.4 Regret Analysis 155

exp(x) − (1 + x)
exp(x) − (1 + x + x2 )
0.1
exp(x) − (1 + x + x2 /2)

−0.1

−0.5 0 0.5
x

Figure 11.3 Approximations for exp(x) on [−1/2, 1/2].

Theorem 11.2. Let x ∈ [0, 1]p

n×k
be an adversarial bandit and π be the policy of
Exp3 with learning rate η = 2 log(k)/(nk). Then,
p
Rn (π, x) ≤ 2nk log(k) .

Proof By construction, X̂tj ≤ 1. Therefore,

   
exp η X̂tj = exp(η) exp η(X̂tj − 1)
 
η2
≤ exp(η) 1 + η(X̂tj − 1) + (X̂tj − 1)2 .
2
P
Using the fact that j Ptj = 1 and the inequality 1 + x ≤ exp(x), we get
 
Xk Xk 2 Xk
Wt η
= Ptj exp(η X̂tj ) ≤ exp η Ptj X̂tj + Ptj (X̂tj − 1)2  ,
Wt−1 j=1 j=1
2 j=1

where the equality is from Eq. (11.11). We see that here we need to bound
P 2
j Ptj (X̂tj − 1) . Let Ŷtj = 1 − X̂tj . Then,

Ptj (X̂tj − 1)2 = Ptj Ŷtj Ŷtj = I {At = j} ytj Ŷtj ≤ Ŷtj ,

where the last inequality used Ŷtj ≥ 0 and ytj ≤ 1. Thus,

k
X k
X
Ptj (X̂tj − 1)2 ≤ Ŷtj .
j=1 j=1

With the same calculations as before, we get

log(k) η X X
n k
Ŝni − Ŝn ≤ + Ŷtj . (11.15)
η 2 t=1 j=1
 11.5 Notes 156

P
The result is completed by taking expectations of both sides, using E t,j Ŷtj =
P P
E t,j Et−1 Ŷtj = E t,j ytj ≤ nk and substituting the learning rate.

The reader may wonder about the somewhat ad hoc proof. The best we
can do for now is to point out a few things about the proof. It is natural to
replace the true rewards with the estimated ones. Then, to prove a regret
bound in terms of the estimated rewards, an alternative to the proof is
to start with the the trivial inequality that states that for any x = (xj )
P
vector and positive quantity η, the inequality xi ≤ η1 log j exp(ηxj ) holds.
Applying this with x = (Ŝni ) gives
1 X 1
Ŝni ≤ log( exp(η Ŝnj )) = log(Wn ) ,
η j
η

from where the proof can be continued by introducing the telescoping

argument.

11.5 Notes

1 Exp3 is nearly optimal in the sense that its expected regret cannot be improved
significantly in the worst case. The distribution of its regret, however, is very far
from optimal. Define the random regret to be the random variable measuring
the actual deficit of the learner relative to the best arm in hindsight:
n
X n
X n
X n
X
R̂n = max xti − Xt = Yt − min yti .
i∈[k] i∈[k]
t=1 t=1 t=1 t=1
| {z } | {z }
in terms of rewards in terms of losses

In Exercise 11.6 you will show that for all large enough n and reasonable
choices of η, there exists a bandit such that the random regret of Exp3 satisfies
P(R̂n ≥ n/4) > 1/131. In the same exercise, you should explain why this does
not contradict the upper bound. That Exp3 has such a high variance is a
serious limitation, which we address in the next chapter.
2 What happens when the range of the rewards is unbounded? This has been
studied by Allenberg et al. [2006], where some (necessarily much weaker)
positive results are presented.
3 In the full information setting, the learner observes the whole vector
xt ∈ [0, 1]k at the end of round t, but the reward is still xtAt . This setting is
also called prediction with expert advice. Exponential weighting is still
a good idea, but the estimated rewards can now be replaced by the actual
rewards. The resulting algorithm is sometimes called Hedge or the exponential
weights algorithm. The proof as written goes through in almost the same way,
 11.5 Notes 157

but one should replace the polynomial upper

p bound on exp(x) with Hoeffding’s
lemma. This analysis gives a regret of n log(k)/2, which is optimal in an
asymptotic sense [Cesa-Bianchi and Lugosi, 2006].
4 We assumed that the adversary chooses the rewards at the start of the game.
Such adversaries are called oblivious. An adversary is called reactive or
non-oblivious if xt is allowed to depend on the history x1 , A1 , . . . , xt−1 , At−1 .
Despite the fact that this is clearly a harder problem, the result we obtained
can be generalised to this setting without changing the analysis. It is another
question whether the definition of regret makes sense for reactive environments.
p
5 A more sophisticated algorithm and analysis shaves a factor of log(k) from
the regret upper bound in Theorem 11.2 [Audibert and Bubeck, 2009, 2010a,
Bubeck and Cesa-Bianchi, 2012]. It turns out that this algorithm, just like
Exp3, is an instantiation of mirror descent from convex optimisation, which
we present in Chapter 28. More details are in p Exercise 28.15. Interestingly,
this algorithm not only shaves off the extra log(k) factor from the regret,
but also achieves O(log(n))-regret
√ in the stochastic setting provided that one
uses a learning rate of 1/ t in round t [Zimmert and Seldin, 2019]. This
remarkable result improves in an elegant way on many previous attempts to
design algorithms for stochastic and adversarial bandits [Bubeck and Slivkins,
2012, Seldin and Slivkins, 2014, Auer and Chiang, 2016, Seldin and Lugosi,
2017]. There are some complications, however, depending on whether or not the
adversary is oblivious. The situation is best summarised by Auer and Chiang
[2016], where the authors present upper and lower bounds on what is possible
in various scenarios.
6 The initial distribution (the ‘prior’) P1 does not have to be uniform. By biasing
the prior towards a specific action, the regret can be reduced when the favoured
action turns out to be optimal. There is an unavoidable price for this, however,
if the optimal arm is not favoured [Lattimore, 2015a].
7 Building on the previous note, suppose the reward in round t is Xt =
ft (A1 , . . . , At ) and f1 , . . . , fn are a sequence of functions chosen in advance by
the adversary with ft : [k]t → [0, 1]. Let Π ⊂ [k]n be a set of action sequences.
Then the expected policy regret with respect to Π is
n
" n #
X X
max ft (a1 , . . . , at ) − E ft (A1 , . . . , At ) .
a1 ,...,an ∈Π
t=1 t=1

Even if Π only consists of constant sequences, there still does not exist a policy
guaranteeing sublinear regret. The reason is simple. Consider the two candidate
choices of f1 , . . . , fn . In the first choice, ft (a1 , . . . , at ) = I {a1 = 1}, and in
the second we have ft (a1 , . . . , at ) = I {a1 = 2}. Clearly the learner must suffer
linear regret in at least one of these two reactive bandit environments. The
problem is that the learner’s decision in the first round determines the rewards
available in all subsequent rounds, and there is no time for learning. By making
 11.5 Notes 158

additional assumptions, sublinear regret is possible, however – e.g. by assuming

the adversary has limited memory [Arora et al., 2012].
8 There is a common misconception that the adversarial framework is a good fit
for non-stationary environments. While the framework does not assume the
rewards are stationary, the regret concept used in this chapter has stationarity
built in. A policy designed for minimising the regret relative to the best action
in hindsight is seldom suitable for non-stationary bandits, where the whole point
is to adapt to changes in the optimal arm. In such cases a better benchmark is
to compete with a sequence of actions. For more on non-stationary bandits,
see Chapter 31.
9 The estimators in Eq. (11.3) and Eq. (11.6) both have conditional variance
Vt [X̂ti ] ≈ 1/Pti , which blows up for small Pti . It is instructive to think about
whether and how Pti can take on very small values. Consider the loss-based
estimator given by (11.6). For this estimator, when PtAt and Xt are both
small, X̂tAt can take on a large negative value. Through the update formula
(11.7), this then translates into Pt+1,At being squashed aggressively towards
zero. A similar issue arises with the reward-based estimator given by (11.3).
The difference is that now it will be a ‘positive surprise’ (PtAt small, Xt
large) that pushes the probabilities towards zero. But note that in this case,
Pt+1,i is pushed towards zero for all i 6= At . This means that dangerously
small probabilities are expected to be more frequent for the gains estimator
Eq. (11.3).
10 Exp3 requires advance knowledge of the horizon. The doubling trick can be
used to overcome this issue, but a more elegant solution is to use a decreasing
learning rate. The analysis in this chapter can be adapted to this case. More
discussion is provided in the notes and exercises of Chapter 28, where we give
a more generic solution to this problem (Exercise 28.13).
11 The calculation in Eq. (11.2) is a reduction, showing that algorithms with low
regret on finite-armed adversarial bandits also have low regret on stochastic
bandits where the reward distributions have appropriately bounded support.
Reductions play an important role throughout the bandit literature and we will
see many more examples. The reader should be careful not to generalise the idea
that adversarial algorithms work well on stochastic problems. The assumptions
must be checked (like boundedness of the support), and for different models
there can be subtleties. The whole of Chapter 29 is devoted to the linear case.
12 As we mentioned, a policy for k-armed adversarial bandits is defined by any
function π : ([k] × [0, 1])∗ → Pk−1 . There is no need to assume that π is
measurable because the actions are discrete and the rewards are deterministic.
The relations between the stochastic and adversarial regret are only well defined
for policies that are probability kernels as defined in Definition 4.7. You might
be worried that lower bounds for stochastic bandits only imply lower bounds for
measurable adversarial policies. Fortunately, the lower bounds are easily proven
for Bernoulli bandits, and in this case the space of reward sequences is finite
and measurability is no longer problematic. Later we study adversarial bandits
 11.6 Bibliographic Remarks 159

with an infinite action set A, which is equipped with a σ-algebra G. In this

case the reward vectors are replaced by functions (xt )nt=1 , where xt : A → [0, 1]
is G-measurable. Then, the measurability condition on the policy is that for all
choices of the adversary and all B ∈ B(A),
π(B | a1 , x1 (a1 ), . . . , at−1 , xt−1 (at−1 ))
must be measurable as a function of a1 , . . . , at−1 . In practice, of course, all the
policies you might ever propose would also be measurable as a function of the
rewards.

11.6 Bibliographic Remarks

Exponential weighting has been a standard tool in online learning since the
papers by Vovk [1990] and Littlestone and Warmuth [1994]. Exp3 and several
variations were introduced by Auer et al. [1995], which was also the first paper to
study bandits in the adversarial framework. The algorithm and analysis presented
here differs slightly because we do not add any additional exploration, while the
version of Exp3 in that paper explores uniformly with low probability. The fact
that additional exploration is not required was observed by Stoltz [2005].

11.7 Exercises

11.1 (Sampling from a multinomial) In order to implement Exp3, you need

a way to sample from the exponential weights distribution. Many programming
languages provide a standard way to do this. For example, in Python you can use
the Numpy library and numpy.random.multinomial. In more basic languages,
however, you only have access to a function rand() that returns a floating point
number ‘uniformly’ distributed in [0, 1]. Describe an algorithm that takes as input
a probability vector p ∈ Pk−1 and uses a single call to rand() to return X ∈ [k]
with P (X = i) = pi .

On most computers, rand() will return a pseudo-random number, and since

there are only finitely many floating point numbers, the resulting distribution
will not really be uniform on [0, 1]. Thinking about these issues is a worthy
endeavour, and sometimes it really matters. For this exercise you may ignore
these issues, however.

11.2 (Linear regret for deterministic policies) Show that for any
deterministic policy π there exists an environment x ∈ [0, 1]n×k such that
Rn (π, x) ≥ n(1 − 1/k). What does your result say about the policies designed in
Part II?
 11.7 Exercises 160

11.3 (Maximum and expectations) Show that the first inequality in (11.2)
holds: Moving the maximum inside the expectation increases the value of the
expectation.

11.4 (Alternative regret definition) Suppose we had defined the regret by

" n n
#
X X
Rntrack (π, x) = E max xti − xtAt .
i∈[k]
t=1 t=1

At first sight this definition seems like the right thing because it measures what
you actually care about. Unfortunately, however, it gives the adversary too much
power. Show that for any policy π (randomised or not), there exists a x ∈ [0, 1]n×k
such that
 
track 1
Rn (π, x) ≥ n 1 − .
k
11.5 (Unbiased estimators are importance weighted) Let P ∈ Pk−1
be a probability vector with nonzero components and let A ∼ P . Suppose
X̂ : [k] × R → R is a function such that for all x ∈ Rk ,
k
X
E[X̂(A, xA )] = Pi X̂(i, xi ) = x1 .
i=1

Show that there exists an a ∈ Rk such that ha, P i = 0 and for all i and z in their
I {i = 1} z
respective domains, X̂(i, z) = ai + .
P1
11.6 (Variance of Exp3) In this exercise, you will show that if η ∈ [n−p , 1]
for some p ∈ (0, 1), then for sufficiently large n, there exists a bandit on which
Exp3 has a constant probability of suffering linear regret. We work with losses so
that given a bandit y ∈ [0, 1]n×k , the learner samples At from Pt given by
 P 
exp −η s=1 Ŷsi
t−1

Pti = P  P ,
j=1 exp −η
k t−1
s=1 Ŷsj

where Ŷti = Ati yti /Pti . Let α ∈ [1/4, 1/2] be a constant to be tuned subsequently
and define a two-armed adversarial bandit in terms of its losses by
( (
0 if t ≤ n/2 α if t ≤ n/2
yt1 = and yt2 =
1 otherwise 0 otherwise .
For simplicity you may assume that n is even.

(a) Define the sequence of real-valued functions q1 , . . . , qn on domain [1/4, 1/2]

inductively by q0 (α) = 1/2 and
qs (α) exp(−ηα/qs (α))
qs+1 (α) = .
1 − qs (α) + qs (α) exp(−ηα/qs (α))
 11.7 Exercises 161

4,000

3,000

2,000
Regret

1,000

−1,000
8

85

0.3
0.2

0.2

9
0.2

0.2
α

Figure 11.4 Exp3 instability: Box and whisker plot of the distribution of the regret of
Exp3 for different values of α over a horizon of n = 104 with m = 500 repetitions for the
example of Exercise 11.6. The boxes represent the quartiles of the empirical distribution,
the diamond shows the average; the median is equal to the upper quartile (and thus
cannot be seen), while the dots show values outside of the “interquartile range”.

Pt
Show for t ≤ 1 + n/2 that Pt2 = qT2 (t−1) (α), where T2 (t) = s=1 As2 .
(b) Show that for sufficiently large n there exists an α ∈ [1/4, 1/2] and s ∈ N
such that

1 1
s−1
X n
qs (α) = and ≤ .
8n u=0
qu (α) 8

(c) Prove that P(T2 (n/2) ≥ s + 1) ≥ 1/65.

(d) Prove that P(R̂n ≥ n/4) ≥ (1 − n exp(−ηn)/2)/65.
(e) The previous part shows that the regret is linear with constant probability
for sufficiently large n. On the other hand, a dubious application of Markov’s
inequality and Theorem 11.1 shows that

4E[R̂n ]
P(R̂n ≥ n/4) ≤ = O(n−1/2 ) .
n
Explain the apparent contradiction.
(f) Validate the theoretical results of this exercise in an experimental fashion:
Implement Exp3 with the loss sequence suggested to reproduce Fig. 11.4.
The
p learning rate is set to the value computed in Theorem 11.2: η =
2 log(k)/(nk). Compare the figure with the theoretical results: Is there an
agreement between theory and the empirical results?
 11.7 Exercises 162

11.7 (Gumbel trick) Let a1 , . . . , ak be positive real values and U1 , . . . , Uk be

a sequence of independent and identically distributed uniform random variables
on [0, 1]. Then let Gi = − log(− log(Ui )), which follows a standard Gumbel
distribution. Prove that
 
ai
P log(ai ) + Gi = max(log(aj ) + Gj ) = Pk .
j=1 aj
j∈[k]

11.8 (Exp3 as follow-the-perturbed-leader) Let (Zti )ti be a collection

of independent and identically distributed random variables. The follow-the-
perturbed-leader algorithm chooses
t−1
!
X
At = argmaxi∈[k] Zti − η Ŷsi .
s=1

Show that if Zti is a standard Gumbel, then follow-the-perturbed-leader is the

same as Exp3.

11.9 (Exp3 on stochastic bandits) In this exercise we compare UCB and

Exp3 on stochastic data. Suppose we have a two-armed stochastic Bernoulli
bandit with µ1 = 0.5 and µ2 = µ1 + ∆ with ∆ = 0.05.
(a) Plot the regret of UCB and Exp3 on the same plot as a function of the
horizon n using the learning rate from Theorem 11.2.
(b) Now fix the horizon to n = 105 and plot the regret as a function of the
learning rate. Your plot should look like Fig. 11.5.
(c) Investigate how the shape of this graph changes as you change ∆.
(d) Find empirically the choice of η that minimises the worst-case regret over all
reasonable choices of ∆, and compare to the value proposed by the theory.
(e) What can you conclude from all this? Tell an interesting story.

Hint The performance of UCB depends greatly on which version you use. For
best results, remember that Bernoulli distributions are 1/2-subgaussian or use
the KL-UCB algorithm from Chapter 10.
 11.7 Exercises 163

Exp3
250

200
Expected regret

150

100

0 0.1
η

Figure 11.5 Expected regret for Exp3 for different learning rates over n = 105 rounds
on a Bernoulli bandit with means µ1 = 0.5 and µ2 = 0.55.
12 The Exp3-IX Algorithm

In the last chapter, we proved a sublinear bound on the expected regret of Exp3,
but with a dishearteningly large variance. The objective of this chapter is to
modify Exp3 so that the regret stays small in expectation and is simultaneously
well concentrated about its mean. Such results are called high-probability
bounds. By slightly modifying the algorithm, we show that for each δ ∈ (0, 1),
there exists an algorithm such that with probability at least 1 − δ,
s  !
Xn
k
R̂n = max (ytAt − yta ) = O nk log .
a∈A
t=1
δ
The poor behaviour of Exp3 occurs because the variance of the importance-
weighted estimators can become very large. In this chapter we modify the reward
estimates to control the variance at the price of introducing some bias.

12.1 The Exp3-IX Algorithm

We start by summarising what we know about the behaviour of the random regret
of Exp3. Because we want to use the loss-based estimator, it is more convenient
to switch to losses, which we do for the remainder of the chapter. Rewriting
Eq. (11.15) in terms of losses,

log(k) η X
k
L̂n − L̂ni ≤ + L̂nj , (12.1)
η 2 j=1

where L̂n and L̂ni are defined using the loss estimator Ŷtj by
n X
X k n
X
L̂n = Ptj Ŷtj and L̂ni = Ŷti .
t=1 j=1 t=1

Eq. (12.1) holds no matter how the loss estimators are chosen, provided
they satisfy 0 ≤ Ŷti ≤ 1/Pti for all t and i. Of course, the left-hand side of
Eq. (12.1) is not close to the regret unless Ŷti is a reasonable estimator of
the loss yti .
 12.1 The Exp3-IX Algorithm 165

We also need to define the sum of losses observed by the learner and for each
fixed action, which are
n
X n
X
L̃n = ytAt and Lni = yti
t=1 t=1

Like in the previous chapter, we need to define the (random) regret with respect
to a given arm i as follows:
n
X n
X
R̂ni = xti − Xt = L̃n − Lni . (12.2)
t=1 t=1

By substituting the above definitions into Eq. (12.1) and rearranging, the regret
with respect to arm i is bounded by

R̂ni = L̃n − Lni = (L̃n − L̂n ) + (L̂n − L̂ni ) + (L̂ni − Lni )

log(k)
k
η X
≤ + (L̃n − L̂n ) + (L̂ni − Lni ) + L̂nj . (12.3)
η 2 j=1

This means the random regret can be bounded by controlling L̃n − L̂n , L̂nj − Lnj
and L̂nj for each j. As promised we now modify the loss estimate. Let γ > 0 be
a small constant to be chosen later and define the biased estimator
I {At = i} Yt
Ŷti = . (12.4)
Pti + γ

First, note that Ŷti still satisfies 0 ≤ Ŷti ≤ 1/Pti , so (12.3) is still valid. As γ
increases, the predictable variance decreases, but the bias increases. The optimal
choice of γ depends on finding the sweet spot, which we will do once the dust
has settled in the analysis. When Eq. (12.4) is used in the exponential update in
Exp3, the resulting algorithm is called Exp3-IX (Algorithm 10). The suffix ‘IX’
stands for implicit exploration, a name justified by the following argument. A
simple calculation shows that
Pti yti γyti
Et [Ŷti ] = = yti − ≤ yti .
Pti + γ Pti + γ
Since small losses correspond to large rewards, the estimator is optimistically
biased. The effect is a smoothing of Pt so that actions with large losses for which
Exp3 would assign negligible probability are still chosen occasionally. In fact, the
smaller is Pti , the larger the bias is. As a result, Exp3-IX will explore more than
the standard Exp3 algorithm (see Exercise 12.5).

The reason for calling the exploration implicit is because the algorithm
explores more as a consequence of modifying the reward estimates, rather
than directly alternating Pt .
 12.2 Regret Analysis 166

1: Input: n, k, η, γ
2: Set L̂0i = 0 for all i
3: for t = 1, . . . , n do
4: Calculate the sampling distribution Pt :
 
exp −η L̂t−1,i
Pti = P  
exp
k
j=1 −η L̂ t−1,j

5: Sample At ∼ Pt and observe reward Xt

I {At = i} (1 − Xt )
6: Calculate L̂ti = L̂t−1,i +
Pti + γ
7: end for
Algorithm 10: Exp3-IX.

12.2 Regret Analysis

We now prove the following theorem bounding the random regret of Exp3-IX
with high probability.
Theorem 12.1. Let δ ∈ (0, 1) and define
r s
2 log(k + 1) δ )
log(k) + log( k+1
η1 = and η2 = .
nk nk
The following statements hold:
1 If Exp3-IX is run with parameters η = η1 and γ = η/2, then
s    !
p nk 1 k+1
P R̂n ≥ 8nk log(k + 1) + log + log ≤ δ.
2 log(k + 1) δ δ
(12.5)
2 If Exp3-IX is run with parameters η = η2 and γ = η/2, then
  
p k+1
P R̂n ≥ 2 (2 log(k + 1) + log(1/δ))nk + log ≤ δ. (12.6)
δ

The value of η1 is independent of δ, which means that using this choice of

learning rate leads to a single algorithm with a high-probability bound for
all δ. On the other hand, η2 does depend on δ, so the user must choose
a confidence level from the beginning. The advantage is that the bound
is improved, but only for the specified confidence level. We will show in
Chapter 17 that this trade-off is unavoidable.

The proof follows by bounding each of the terms in Eq. (12.3), which we do
via a series of lemmas. The first of these lemmas is a new concentration bound.
 12.2 Regret Analysis 167

To state the lemma, we recall two useful notions: Recall that given a filtration
F = (Ft )nt=0 , (Zt )nt=1 is F-adapted if for t ∈ [n], Zt is Ft -measurable and (Zt )nt=1
is F-predictable, if for t ∈ [n], Zt is Ft−1 -measurable.
Lemma 12.2. Let F = (Ft )nt=0 be a filtration and for i ∈ [k] let (Ỹti )t be F-adapted
such that:
Q  
 S ⊂[k] with |S| > 1, E i∈S Ỹti Ft−1 ≤ 0; and
1 for any 
2 E Ỹti  Ft−1 = yti for all t ∈ [n] and i ∈ [k].
Furthermore, let (αti )ti and (λti )ti be real-valued F-predictable random sequences
such that for all t, i it holds that 0 ≤ αti Ỹti ≤ 2λti . Then, for all δ ∈ (0, 1),
   !
1
Xn X k
Ỹti
P αti − yti ≥ log ≤ δ.
t=1 i=1
1 + λ ti δ

The proof relies on the Cramér–Chernoff method and is deferred until the
end of the chapter. Condition 1 states that the variables {Ỹti }i are negatively
correlated, and it helps us save a factor of k. Equipped with this result, we can
easily bound the terms L̂ni − Lni .
Lemma 12.3 (Concentration – variance). Let δ ∈ (0, 1). With probability at least
1 − δ, the following inequalities hold simultaneously:
  log( k+1 ) k 
X  log( k+1 )
max L̂ni − Lni ≤ δ
and L̂ni − Lni ≤ δ
. (12.7)
i∈[k] 2γ i=1
2γ

Proof Fix δ 0 ∈ (0, 1) to be chosen later and let Ati = I {At = i} as before. Then
X k Xn X k  
Ati yti
(L̂ni − Lni ) = − yti
i=1 t=1 i=1
Pti + γ
!
1 XX 1
n k
Ati yti
= 2γ − yti .
2γ t=1 i=1 1 + Pγti Pti

Introduce λti = Pγti , Ỹti = APtitiyti and αti = 2γ. Notice that the conditions of
Lemma 12.2 are now satisfied. In particular, for any S ⊂ [k] with |S| > 1, it holds
Q Q
that i∈S Ati = 0 and hence i∈S Ỹti = 0. Therefore,
!
log(1/δ 0 )
Xk
P (L̂ni − Lni ) ≥ ≤ δ0 . (12.8)
i=1
2γ

Similarly, for any fixed i,

 
log(1/δ 0 )
P L̂ni − Lni ≥ ≤ δ0 . (12.9)
2γ
To see this, use the previous argument with αtj = I {j = i} 2γ. The result follows
by choosing δ 0 = δ/(k + 1) and the union bound.
Pk
Lemma 12.4 (Bias). L̃n − L̂n = γ j=1 L̂nj .
 12.2 Regret Analysis 168

P
Proof Let Ati = I {At = i} as before. Writing Yt = j Atj ytj , we calculate
k
X k 
X  k
X Xk
Ptj Atj
Yt − Ptj Ŷtj = 1− Atj ytj = γ ytj = γ Ŷtj .
j=1 j=1
Ptj + γ j=1
Ptj + γ j=1
Pk
Therefore L̃n − L̂n = γ j=1 L̂nj as required.

Proof of Theorem 12.1 By Eq. (12.3) and Lemma 12.4, we have

log(k)
k
ηX
R̂n ≤ + (L̃n − L̂n ) + max(L̂ni − Lni ) + L̂nj
η i∈[k] 2 j=1

log(k) η X k
= + max(L̂ni − Lni ) + +γ L̂nj .
η i∈[k] 2 j=1

Therefore, by Lemma 12.3, with probability at least 1 − δ,

 

log(k) log k+1 η  X k
log k+1
R̂n ≤ + δ
+ γ+  Lnj + δ 
η 2γ 2 j=1
2γ

log(k)  η   η  log k+1

≤ + γ+ nk + γ + + 1 δ
,
η 2 2 2γ
where the second inequality follows since Lnj ≤ n for all j. The result follows by
substituting the definitions of η ∈ {η1 , η2 } and γ = η/2.

The attentive reader may be wondering whether proving the new

concentration inequality of Lemma 12.2, which looks a bit ad hoc, was
really necessary to get the bounds on L̂ni that were stated in the next
lemma. After all, we had a number of concentration inequalities available
to us that could be applied. As it turns out, one could also use Bernstein
inequality to get a result that only loses a factor of two compared to the
specialised lemma. The details are in Exercise 12.1. There are two important
lessons that are the basis of both proofs. The first is that since E[L̂ni ] < Lni
and the gap between these two quantities is large enough in a manner that
we make precise in Exercise 12.1, the deviation L̂ni − Lni can p be bounded
independently of Pti , Ati and yti . The price is that instead of log(1/δ), the
bound scales linearly with the generally larger quantity log(1/δ). The factor
1/γ here is the maximum scale of the individual summands in L̂ni . The
P
second lesson is specific to how in bounding i L̂ni − Lni a union bound over
i is avoided: this works because for a fixed time index t, (Ati )i are negatively
correlated. Negative dependence/association/correlation are known to be
good substitutes for independence, and by exploiting such properties one
can often demonstrate better concentration.
 12.2 Regret Analysis 169

12.2.1 Proof of Lemma 12.2

We start with a technical inequality:
 
x
Lemma 12.5. For any 0 ≤ x ≤ 2λ it holds that exp ≤ 1 + x.
1+λ
Note that 1+x ≤ exp(x). What the lemma shows is that by slightly discounting
the argument of the exponential function, in a bounded neighbourhood of zero,
1 + x can be an upper bound for the resulting function. Or, equivalently, slightly
inflating the linear term in 1 + x, the linear lower bound becomes an upper bound.
Proof of Lemma 12.5 We have
   
x x
exp ≤ exp ≤ 1 + x,
1+λ 1 + x/2
where the first inequality is because λ 7→ exp( 1+λx
) is decreasing in λ, and the
2u
second is because 1+u ≤ log(1 + 2u) holds for all u ≥ 0. This latter inequality
can be seen to hold by noting that for u = 0, the two sides are equal, while the
derivative of the left-hand side is smaller than that of the right-hand side at any
u ≥ 0.
Proof of Lemma 12.2 Fix t ∈ [n] and let Et [·] = E[· | Ft ] denote the conditional
expectation with respect to Ft . By Lemma 12.5 and the assumption that
0 ≤ αti Ỹti ≤ 2λti , we have
 
αti Ỹti
exp ≤ 1 + αti Ỹti .
1 + λti
Taking the product of these inequalities over i,
" k
!# " k # " k #
X αti Ỹti Y X
Et−1 exp ≤ Et−1 (1 + αti Ỹti ) ≤ 1 + Et−1 αti Ỹti
i=1
1 + λti i=1 i=1
k k
!
X X
=1+ αti yti ≤ exp αti yti , (12.10)
i=1 i=1
Qk P Qk
where the second inequality follows from i=1 (1 + ai ) = b∈{0,1}k i=1 abi i and
Q
the assumption that for S ⊂ [k] with |S| > 1, Et−1 [ i∈S Ỹti ] ≤ 0, the third one
follows from the assumption that Et−1 [Ỹti ] = yti , while the last one follows from
1 + x ≤ exp(x). Define
X k  !
Ỹti
Zt = exp αti − yti
i=1
1 + λti

and let Mt = Z1 . . . Zt , t ∈ [n] with M0 = 1. By (12.10), Et−1 [Zt ] ≤ 1. Therefore

E[Mt ] = E[Et−1 [Mt ]] = E[Mt−1 Et−1 [Zt ]] ≤ E[Mt−1 ] ≤ · · · ≤ E[M0 ] = 1 .
Setting t = n and combining the above display with Markov’s inequality leads to
P (log(Mn ) ≥ log(1/δ)) = P (Mn δ ≥ 1) ≤ E [Mn ] δ ≤ δ.
 12.3 Notes 170

12.3 Notes

1 An alternative to the somewhat custom-made Lemma 12.2 is to use a Bernstein-

type bound that simply bounds the deviation of a martingale from its mean
in terms of its quadratic variation. The slight disadvantage of this is that this
way we lose a factor of two. If this is not a concern, one may even prefer this
approach due to its greater transparency. For details, see Exercise 12.1.
2 An upper bound on the expected regret of Exp3-IX can be obtained by
integrating the tail:
Z ∞   Z ∞  
+ +
Rn ≤ E[(R̂n ) ] = P (R̂n ) ≥ x dx ≤ P R̂n ≥ x dx ,
0 0

where the first equality follows from Proposition 2.8. The result is completed
using either the high-probability bound in Theorem 12.1 and by straightforward
integration. We leave the details to the reader in Exercise 12.7.
3 The analysis presented here uses a fixed
p learning rate that depends on the
horizon. Replacing η and γ with ηt = log(k)/(kt) and γt = ηt /2 leads to an
anytime algorithm with about the same regret [Kocák et al., 2014, Neu, 2015a].
4 There is another advantage of the modified importance-weighted estimators
used by Exp3-IX, which leads to an improved regret in the special case that
one of the arms has small losses. Specifically, it is possible to show that
!
r
Rn = O k min Lni log(k) .
i∈[k]

In the worst case, Lni is linear in n and the usual bound is recovered. But
if the optimal arm enjoys low cumulative regret, then the above can be a
big improvement over the bounds given in Theorem 12.1. Bounds of this
kind are called first-order bounds. We refer the interested reader to the
papers by Allenberg et al. [2006], Abernethy et al. [2012] and Neu [2015b] and
Exercise 28.14.
5 Another situation where one might hope to have a smaller regret is when the
rewards/losses for each arm do not deviate too far from their averages. Define
the quadratic variation by
1X
n
X n
Qn = kxt − µk2 , where µ = xt .
t=1
n t=1
√
Hazan and Kale [2011] gave an algorithm for which Rn = O(k 2 Qn ), which can
be better than the worst-case bound of Exp3 or Exp3-IX when the quadratic
variation is very small. The factor of k 2 is suboptimal and can be removed
using a careful instantiation of the mirror descent algorithm [Bubeck et al.,
2018]. We do not cover this exact algorithm in this book, but the techniques
based on mirror descent are presented in Chapter 28.
6 An alternative to the algorithm presented here is to mix the probability
distribution computed using exponential weights with the uniform distribution,
 12.4 Bibliographic Remarks 171

while biasing the estimates. This leads to the Exp3.P algorithm due to Auer
et al. [2002b], who considered the case where δ is given and derived a bound that
is similar to Eq. (12.6) of Theorem 12.1. With an appropriate modification of
their proof, it is possible to derive a weaker bound similar to Eq. (12.5), where
the knowledge of δ is not needed by the algorithm. This has been explored by
Beygelzimer et al. [2010] in the context of a related algorithm, which will be
considered in Chapter 18. One advantage of this approach is that it generalises
to the case where the loss estimators are sometimes negative, a situation that
can arise in more complicated settings. For technical details, we advise the
reader to work through Exercise 12.3.

12.4 Bibliographic Remarks

The Exp3-IX algorithm is due to Kocák et al. [2014], who also introduced the
biased loss estimators. The focus of that paper was to improve algorithms for
more complex models with potentially large action sets and side information,
though their analysis can still be applied to the model studied in this chapter. The
observation that this algorithm also leads to high-probability bounds appeared in
a follow-up paper by Neu [2015a]. High-probability bounds for adversarial bandits
were first provided by Auer et al. [2002b] and explored in a more generic way by
Abernethy and Rakhlin [2009]. The idea to reduce the variance of importance-
weighted estimators is not new and seems to have been applied in various forms
[Uchibe and Doya, 2004, Wawrzynski and Pacut, 2007, Ionides, 2008, Bottou
et al., 2013]. All of these papers are based on truncating the estimators, which
makes the resulting estimator less smooth. Surprisingly, the variance-reduction
technique used in this chapter seems to be recent [Kocák et al., 2014].

12.5 Exercises

12.1 (Bernstein-type inequality and Lemma 12.3) Using the Berstein-type

inequality stated in Exercise 5.15, show the following:

(a) For any δ ∈ (0, 1), with probability at least 1 − δ, L̂ni − Lni < γ1 log(1/δ).
P P
(b) For any δ ∈ (0, 1), with probability at least 1 − δ, i L̂ni − i Lni <
1
γ log(1/δ).

12.2 Prove the claims made in Note 3.

Hint The source for this exercise is theorem 1 of the paper by Neu [2015a].
You can also read ahead and use the techniques from Exercise 28.13.
12.3 (Exp3.P) In this exercise we ask you to analyse the Exp3.P algorithm,
which as we mentioned in the notes is another way to obtain high probability
 12.5 Exercises 172

bounds. The idea is to modify Exp3 by biasing the estimators and introducing
some forced exploration. Let Ŷti = Ati yti /Pti − β/Pti be a biased version of the
loss-based importance-weighted estimator that was used in the previous chapter.
Pt
Define L̂ti = s=1 Ŷsi and consider the policy that samples At ∼ Pt , where
 
γ exp −η L̂t−1,i
Pti = (1 − γ)P̃ti + with P̃ti = P  .
k exp −η L̂
k
j=1 t−1,j

(a) Let δ ∈ (0, 1) and i ∈ [k]. Show that with probability 1 − δ, the random
regret R̂ni against i (cf. (12.2)) satisfies
r
n log(1/δ)
X n Xk Xn
β
R̂ni < nγ + (1 − γ) P̃tj (Ŷtj − yti ) + + .
t=1 j=1
P
t=1 tAt
2

(b) Show that

n X
X k n X
X k n
X
P̃tj (Ŷtj − yti ) = P̃tj (Ŷtj − Ŷti ) + (Ŷti − yti ) .
t=1 j=1 t=1 j=1 t=1

(c) Show that

log(k)
n X
X k Xn X k
P̃tj (Ŷtj − Ŷti ) ≤ +η P̃tj Ŷtj2 .
t=1 j=1
η t=1 j=1

(d) Show that

nk 2 β 2 X 1
n X
X k n
P̃tj Ŷtj2 ≤ + .
t=1 j=1
γ P
t=1 tAt

(e) Suppose that γ = kη and η = β. Apply the result of Exercise 5.15 to show
that for any δ ∈ (0, 1), the following hold:
 !
1 1
Xn
k
P ≥ 2nk + log ≤ δ.
P
t=1 tAt
γ δ
 !
1 1
Xn
P Ŷti − yti ≥ log ≤ δ.
t=1
β δ

(f) Combining the previous steps, show that there exists a universal constant
C > 0 such that for any δ ∈ (0, 1), for an appropriate choice of η, γ and β,
with probability at least 1 − δ it holds that the random regret R̂n of Exp3.P
satisfies
p
R̂n ≤ C nk log(k/δ) .
(g) In which step did you use the modified estimators?
(h) Show a bound where the algorithm parameters η, γ, β can only depend on
n, k, but not on δ.
(i) Compare the bounds with the analogous bounds for Exp3-IX in Theorem 12.1.
 12.5 Exercises 173

12.4 (Generic Exp3.P analysis) This exercise is concerned with a

generalisation of the core idea underlying Exp3.P of the previous exercise in that
rather than giving explicit expressions for the biased loss estimates, we focus
on the key properties required by the analysis of Exp3.P. To reduce clutter, we
assume for the remainder that t ranges in [n] and a ∈ [k]. Let (Ω, F, F, P) be a
filtered probability space with F = (Ft )nt=0 . Let (Zt ), (Ẑt ), (Z̃t ), (βt ) be sequences
of random elements in Rk , where Z̃t = Ẑt − βt and (Zt ), (βt ) are F-predictable,
whereas (Ẑt ) and therefore also (Z̃t ) are F-adapted. You should think of Ẑt as
the estimate of Zt that uses randomisation, and βt is the bias as in the previous
exercise. Given positive constant η, define the probability vector Pt ∈ Pk−1 by
 P 
exp −η s=1 Z̃sa
t−1

Pta = P  P .
b=1 exp −η
k t−1
s=1 Z̃sb

Let Et−1 [·] = E [· | Ft−1 ]. Assume the following hold for all a ∈ [k]:

(a) η|Ẑta | ≤ 1 , (b) ηβta ≤ 1 ,

2
(c) ηEt−1 [Ẑta ] ≤ βta almost surely , (d) Et−1 [Ẑta ] = Zta almost surely .

n X
X k
Pn
Let A∗ = argmina∈[k] t=1 Zta and Rn = Pta (Zta − ZtA∗ ).
t=1 a=1

(a) Show that

n X
X k
Pta (Zta − ZtA∗ )
t=1 a=1
n X
X k n X
X k n
X
= Pta (Z̃ta − Z̃tA∗ ) + Pta (Zta − Z̃ta ) + (Z̃tA∗ − ZtA∗ ) .
t=1 a=1 t=1 a=1 t=1
| {z } | {z } | {z }
(A) (B) (C)

(b) Show that

log(k) Xn X k Xn X k
2
(A) ≤ +η Pta Ẑta +3 Pta βta .
η t=1 a=1 t=1 a=1

(c) Show that with probability at least 1 − δ,

log(1/δ)
n X
X k
(B) ≤ 2 Pta βta + .
t=1 a=1
η

(d) Show that with probability at least 1 − kδ,

log(1/δ)
(C) ≤ .
η
 12.5 Exercises 174

(e) Conclude that for any δ ≤ 1/(k + 1), with probability at least 1 − (k + 1)δ,

3 log(1/δ) Xn X k Xn X k
2
Rn ≤ +η Pta Ẑta +5 Pta βta .
η t=1 a=1 t=1 a=1

Hint This is a long and challenging exercise. You may find it helpful to use
the result in Exercise 5.15. The solution is also available.
12.5 (Implementation) Consider the Bernoulli bandit with k = 5 arms and
n = 104 with means µ1 = 1/2 and µi = 1/2 − ∆ for i > 1. Plot the regret of Exp3
and Exp3-IX for ∆ ∈ [0, 1/2]. You should get a plot similar to that of Fig. 12.1.
Does the result surprise you?

300 Exp3
Exp3-IX

250
Regret

200

150

100
0 0.1 0.2 0.3 0.4 0.5
∆

Figure 12.1 Comparison between Exp3 and Exp3-IX on Bernoulli bandit

12.6 (Implementation: Variance of Exp3-IX) Repeat the experiment that

led to Fig. 11.4 but with Exp3 swapped to Exp3-IX. Use the confidence parameter
independent value of η and γ from Theorem 12.1. You should get a figure similar
to Fig. 12.2. Compare the new and the old figures and summarise your findings,
including the outcome of the results of Exercise 12.5.

12.7 (Expected regret of Exp3-IX) In this exercise, you will complete the
steps explained in Note 2 to prove a bound on the expected regret of Exp3-IX.
(a) Find a choice of η and universal constant C > 0 such that
p
Rn ≤ C kn log(k) .
(b) What happens as η grows? Write a bound on the expected regret of Exp3-IX
in terms of η and k and n.
 12.5 Exercises 175

−400

−500

−600
Regret

−700

−800
8

85

95

0.3
0.2

0.2
0.2

0.2

Figure 12.2 Box and whisker plot of the regret of Exp3-IX for the same setting as those
used to produce Fig. 11.4. For details of the experimental settings, see the text of
Exercise 11.6.
Part IV
Lower Bounds for Bandits
with Finitely Many Arms
 177

Until now, we have indulged ourselves by presenting algorithms and upper

bounds on their regret. As satisfying as this is, the real truth of a problem is
usually to be found in the lower bounds. There are several reasons for this:

1 An upper bound does not tell you much about what you could be missing
out on. The only way to demonstrate that your algorithm really is (close to)
optimal is to prove a lower bound showing that no algorithm can do better.
2 The second reason is that lower bounds are often more informative in the
sense that it usually turns out to be easier to get the lower bound right than
the upper bound. History shows a list of algorithms with steadily improving
guarantees until eventually someone hits upon the idea for which the upper
bound matches some known lower bound.
3 Finally, thinking about lower bounds forces you to understand what is hard
about the problem. This is so useful that the best place to start when attacking
a new problem is usually to try and prove lower bounds. Too often we have
not heeded our own advice and started trying to design an algorithm, only
to discover later that had we tackled the lower bound first, then the right
algorithm would have fallen in our laps with almost no effort at all.

So what is the form of a typical lower bound? In the chapters that follow, we
will see roughly two flavours. The first is the worst-case lower bound, which
corresponds to a claim of the form
‘For any policy you give me, I will give you an instance of a bandit problem ν on which
the regret is at least L’ .

Results of this kind have an adversarial flavour, which makes them suitable for
understanding the robustness of a policy. The second type is a lower bound on
the regret of an algorithm for specific instances. These bounds have a different
form that usually reads like the following:
‘If you give me a reasonable policy, then its regret on any instance ν is at least L(ν)’ .

The statement only holds for some policies – the ‘reasonable’ ones, whatever that
means. But the guarantee is also more refined because bound controls the regret
for these policies on every instance by a function that depends on this instance.
This kind of bound will allow us to show that the instance-dependent bounds
P
for stochastic bandits of O( i:∆i >0 ∆i + log(n)/∆i ) are not improvable. The
inclusion of the word ‘reasonable’ is unfortunately necessary. For every bandit
instance ν there is a policy that just chooses the optimal action in ν. Such policies
are not reasonable because they have linear regret for bandits with a different
optimal arm. There are a number of ways to define ‘reasonable’ in a way that is
simultaneously rigorous and, well, reasonable.
The contents of this part is roughly as follows. First we introduce the definition
of worst-case regret and discuss the line of attack for proving lower bounds
(Chapter 13). The next chapter takes us on a brief excursion into information
theory, where we explain the necessary mathematical tools (Chapter 14). Readers
 178

familiar with information theory could skim this chapter. The final three chapters
are devoted to applying information theory to prove lower bounds on the regret
for both stochastic and adversarial bandits.
13 Lower Bounds: Basic Ideas

The worst-case regret of a policy π on a set of stochastic bandit environments

E is

Rn (π, E) = sup Rn (π, ν) .

ν∈E

Let Π be the set of all policies. The minimax regret is

Rn∗ (E) = inf Rn (π, E) = inf sup Rn (π, ν) .

π∈Π π∈Π ν∈E

A policy is called minimax optimal for E if Rn (π, E) = Rn∗ (E). The value Rn∗ (E)
is of interest by itself. A small value of Rn∗ (E) indicates that the underlying bandit
problem is less challenging in the worst-case sense. A core activity in bandit
theory is to understand what makes Rn∗ (E) large or small, often focusing on its
behaviour as a function of the number of rounds n.

Minimax optimality is not a property of a policy alone. It is a property of a

policy together with a set of environments and a horizon.

Finding a minimax policy is generally too computationally expensive to be

practical. For this reason, we almost always settle for a policy that is nearly
minimax optimal.
One of the main results of this part is a proof of the following theorem, which
together with Theorem 9.1 shows that Algorithm 7 from Chapter 9 is minimax
optimal up to constant factors for 1-subgaussian bandits with suboptimality gaps
in [0, 1].

Theorem 13.1. Let E k be the set of k-armed Gaussian bandits with unit variance
and means µ ∈ [0, 1]k . Then there exists a universal
√ constant c > 0 such that for
all k > 1 and n ≥ k, it holds that Rn∗ (E k ) ≥ c kn.

We will prove this theorem in Chapter 15, but first we give an informal
justification.
 13.1 Main Ideas Underlying Minimax Lower Bounds 180

13.1 Main Ideas Underlying Minimax Lower Bounds

Let X1 , . . . , Xn be a sequence of independent Gaussian random variables with

unknown mean µ and known variance 1. Assume you are told that µ takes on
one of two values: µ = 0 or µ = ∆ for some known ∆ > 0. Your task is to guess
Pn
the value of µ based on your observation of X1 , . . . , Xn . Let µ̂ = n1 i=1 Xi be
the sample mean, which is Gaussian with mean µ and variance 1/n. While it is
not immediately obvious how easy this task is, intuitively we expect the optimal
decision is to predict that µ = 0 if µ̂ is closer to 0 than to ∆, and otherwise to
predict µ = ∆. For large n we expect our prediction will probably be correct.
Supposing that µ = 0 (the other case is symmetric), then the prediction will be
wrong only if µ̂ ≥ ∆/2. Using the fact that µ̂ is Gaussian with mean µ = 0 and
variance 1/n, combined with known bounds on the Gaussian tail probabilities
(see Eq. (13.4)), leads to
r    
1 8 n∆2 ∆
√ √ exp − ≤ P µ̂ ≥
n∆2 + n∆2 + 16 π 8 2
r  
1 8 n∆2
≤√ p exp − .
n∆2 + n∆2 + 32/π π 8
(13.1)
The upper and lower bounds only differ in the constant in the square root of the
denominator. One might believe that the decision procedure could be improved,
but the symmetry of the problem makes this seem improbable. The formula
exhibits the expected behaviour, which is that once n is large relative to 8/∆2 ,
then the probability that this procedure fails drops exponentially with further
increases in n. But the lower bound also shows that if n is small relative to 8/∆2 ,
then the procedure fails with constant probability.
The problem described is called hypothesis testing, and the ideas underlying
the argument above are core to many impossibility results in statistics. The next
task is to reduce our bandit problem to hypothesis testing. The high-level idea
is to select two bandit problem instances in such a way that the following two
conditions hold simultaenously:

1 Competition: An action, or, more generally, a sequence of actions that is good

for one bandit is not good for the other.
2 Similarity: The instances are ‘close’ enough that the policy interacting with
either of the two instances cannot statistically identify the true bandit with
reasonable statistical accuracy.

The two requirements are clearly conflicting. The first makes us want to choose
instances with means µ, µ0 ∈ [0, 1]k that are far from each other, while the second
requirement makes us want to choose them to be close to each other. The lower
bound will follow by optimising this trade-off.
Let us start to make things concrete by choosing bandits ν = (Pi )ki=1 and
 13.1 Main Ideas Underlying Minimax Lower Bounds 181

ν 0 = (Pi0 )ki=1 , where Pi = N (µi , 1) and Pi0 = N (µ0i , 1) are Gaussian and
µ, µ0 ∈ [0, 1]k . We will also assume that n is larger than k by some suitably
large constant factor. In order to prove a lower bound, it suffices to show that for
every strategy π, there exists a choice of µ and µ0 such that
√
max {Rn (π, ν), Rn (π, ν 0 )} ≥ c kn ,

where c > 0 is a universal constant. Let ∆ ∈ (0, 1/2] be a constant to be tuned

subsequently and choose µ = (∆, 0, 0, . . . , 0), which means that the first arm is
optimal in instance ν and

Rn (π, ν) = (n − E[T1 (n)])∆ , (13.2)

where the expectation is taken with respect to the induced measure on the
sequence of outcomes when π interacts with ν. Now we need to choose µ0 to
satisfy the two requirements above. Since we want ν and ν 0 to be hard to
distinguish and yet have different optimal actions, we should make µ0 as close to
µ except in a coordinate where π expects to explore the least. To this end, let

i = argminj>1 E[Tj (n)]

be the suboptimal arm in ν that π expects to play least often. From n =

P
E[T1 (n)] + j>1 E[Tj (n)] ≥ (k − 1)E[Ti (n)] we see that
n
E[Ti (n)] ≤
k−1
must hold. Then, define µ0 ∈ Rk by
(
µj , if j 6= i ;
µj =
0
2∆ , otherwise .

The regret in this bandit is

X
Rn (π, ν 0 ) = ∆E0 [T1 (n)] + 2∆E0 [Tj (n)] ≥ ∆E0 [T1 (n)] , (13.3)
j ∈1,i
/

where E0 [·] is the expectation operator on the sequence of outcomes when π

interacts with ν 0 . So now we have the following situation: the strategy π interacts
with either ν or ν 0 , and when interacting with ν, it expects to play arm i at
most n/(k − 1) times. But the two instances only differ when playing arm i. The
time has come to tune ∆. Because the strategy expects to play arm i only about
n/(k − 1) times, taking inspiration from the previous discussion on distinguishing
samples from Gaussian distributions with different means, we will choose
s r
1 k−1
∆= ≥ .
E[Ti (n)] n

If we are prepared to ignore the fact that Ti (n) is a random variable and take for
granted the claims in the first part of the chapter, then with this choice of ∆,
 13.2 Notes 182

the strategy cannot distinguish between instances ν and ν 0 , and in particular we

expect that E[T1 (n)] ≈ E0 [T1 (n)]. If E[T1 (n)] < n/2, then by Eq. (13.2) we have
r
n k−1 1p
Rn (π, ν) ≥ = n(k − 1) .
2 n 2
On the other hand, if E[T1 (n)] ≥ n/2, then
1p
n(k − 1) ,
Rn (π, ν 0 ) ≥ ∆E0 [T1 (n)] ≈ ∆E[T1 (n)] ≥
2
which completes our heuristic argument that there exists a universal constant
c > 0 such that
√
Rn∗ (E k ) ≥ c nk .
We have been sloppy in many places. The claims in the first part of the chapter
have not been proven yet, and Ti (n) is a random variable. Before we can present the
rigourous argument, we need a chapter to introduce some ideas from information
theory. Readers already familiar with these concepts can skip to Chapter 15 for
the proof of Theorem 13.1.

13.2 Notes

1 The worst-case regret has a game-theoretic interpretation. Imagine a game

between a protagonist and an antagonist that works as follows: for k > 1 and
n ≥ k the protagonist proposes a bandit policy π. The antagonist looks at the
policy and chooses a bandit ν from the class of environments considered. The
utility for the antagonist is the expected regret, and for the protagonist it is the
negation of the expected regret, which makes this a zero-sum game. Both players
aim to maximise their pay-offs. The game is completely described by n and E.
One characteristic value in a game is its minimax value. As described above,
this is a sequential game (the protagonist moves first, then the antagonist). The
minimax value of this game from the perspective of the antagonist is exactly
Rn∗ (E), while for the protagonist, it is supπ inf ν (−Rn (π, ν)) = −Rn∗ (E).
2 We mentioned that finding the minimax optimal policy is usually
computationally infeasible. In fact it is not clear we should even try. In classical
statistics, it often turns out that minimising the worst case leads to a flat
risk profile. In the language of bandits, this would mean that the regret is
the same for every bandit (where possible). What we usually want in practice
is to have low regret against ‘easy’ bandits and larger regret against ‘hard’
bandits. The analysis in Part II suggests that easy bandits are those where the
suboptimality gaps are large or very small. There is evidence to suggest that
the exact minimax optimal strategy may not exploit these easy instances, so
in practice one might prefer to find a policy that is nearly minimax optimal
and has much smaller regret on easy bandits. We will tackle questions of this
nature in Chapter 16.
 13.3 Bibliographic Remarks 183

3 The regret on a class of bandits E is a multi-objective criterion. Some policies

will be good for some instances and bad on others, and there are clear trade-
offs. One way to analyse the performance in a multi-objective setting is called
Pareto optimality. A policy is Pareto optimal if there does not exist another
policy that is a strict improvement – more precisely, if there does not exist a
π 0 such that Rn (π 0 , ν) ≤ Rn (π, ν) for all ν ∈ E and Rn (π 0 , ν) < Rn (π, ν) for
at least one instance ν ∈ E.
4 When we say a policy is minimax optimal up to constant factors for finite-armed
1-subgaussian bandits with suboptimality gaps in [0, 1], we mean there exists a
C > 0 such that
Rn (π, E k )
≤ C for all k and n ,
Rn∗ (E k )
where E k is the set of k-armed 1-subgaussian bandits with suboptimality gaps
in [0, 1]. We often say a policy is minimax optimal up to logarithmic factors,
by which we mean that
Rn (π, E k )
≤ C(n, k) for all k and n ,
Rn∗ (E k )
where C(n, k) is logarithmic in n and k. We hope the reader will forgive us
for not always specifying in the text exactly what is meant and promise that
statements of theorems will always be precise.

13.3 Bibliographic Remarks

The bound on Gaussian tails used in Eq. (13.1) is derived from §7.1.13 of the
reference book by Abramowitz and Stegun [1964], which bounds
Z ∞
exp(−x2 ) exp(−x2 )
√ ≤ exp(−t2 )dt ≤ p for all x ≥ 0 . (13.4)
x + x2 + 2 x x + x2 + 4/π

13.4 Exercises

13.1 (Minimax risk for hypothesis testing) Let Pµ = N (µ, 1) be the

Gaussian measure on (R, B(R)) with mean µ ∈ {0, ∆} and unit variance. Let
X : R → R be the identity random variable (X(ω) = ω). For decision rule
d : R → {0, ∆}, define the risk
R(d) = max Pµ (d(X) 6= µ) ,
µ∈{0,∆}

Prove that R(d) is minimised by d(x) = argminµ̃∈{0,∆} |x − µ̃| .

13.2 (Pareto optimal policies) Let k > 1 and E = EN k

(1) be the set of
Gaussian bandits with unit variance. Find a Pareto optimal policy for this class.
 13.4 Exercises 184

Hint Think about simple policies (not necessarily good ones) and use the
definition.
14 Foundations of Information Theory
( )

To make the arguments in the previous chapter rigourous and generalisable to

other settings, we need some tools from information theory and statistics. The
most important of these is the relative entropy, also known as the Kullback–
Leibler divergence named for Solomon Kullback and Richard Leibler (KL
divergence, for short).

14.1 Entropy and Optimal Coding

Alice wants to communicate with Bob. She wants to tell Bob the outcome of a
sequence n of independent random variables sampled from known distribution Q.
Alice and Bob agree to communicate using a binary code that is fixed in advance
in such a way that the expected message length is minimised. The entropy of Q
is the expected number of bits necessary per random variable using the optimal
code as n tends to infinity. The relative entropy between distributions P and Q
is the price in terms of expected message length that Alice and Bob have to pay
if they believe the random variables are sampled from Q when in fact they are
sampled from P .
Let P be a measure on [N ] with σ-algebra 2[N ] and X : [N ] → [N ] be the
identity random variable, X(ω) = ω. Alice observes a realisation of X and wants
to communicate the result to Bob using a binary code that they agree upon
in advance. For example, when N = 4, they might agree on the following code:
1 → 00, 2 → 01, 3 → 10, 4 → 11. Then if Alice observes a 3, she sends Bob a
message containing 10. For our purposes, a code is a function c : [N ] → {0, 1} ,
∗

where {0, 1} is the set of finite sequences of zeros and ones.

∗

Of course c must be injective so that no two numbers (or symbols) have the
same code. We also require that c be prefix free, which means that no code is a
prefix of any other. This is justified by supposing that Alice would like to tell
Bob about multiple samples. Then Bob needs to know where the message for one
symbol starts and ends.

Using a prefix code is not the only way to enforce unique decodability, but
all uniquely decodable codes have equivalent prefix codes (see Note 1).
 14.1 Entropy and Optimal Coding 186

The easiest choice is to use dlog2 (N )e bits

111 L 10101 B 011000
no matter the value of X. This simple code
E 010 D 01101 V 1100000
is sometimes effective, but is not entirely T 1101 C 00001 K 11000011
satisfactory if X is far from uniform. To A 1011 U 00000 X 110000100
understand why, suppose that N is extremely O 1001 F 110011 J 1100001011
large and P (X = 1) = 0.99, and the I 1000 M 110010 Q 11000010101
remaining probability mass is uniform over N 0111 W 110001 Z 11000010100
[N ] \ {1}. Then it seems preferable to have a S 0011 Y 101001
short code for one and slightly longer codes for H 0010 P 101000
the alternatives. With this in mind, a natural R 0001 G 011001

objective is to find a code that minimises the

expected code length. That is, Figure 14.1 A Huffman code for the
English alphabet, including space.
N
X
c∗ = argminc pi `(c(i)) , (14.1)
i=1

where the argmin is taken over valid codes and `(·) is a function that returns
the length of a code. The optimisation problem in (14.1) can be solved using
Huffman coding, and the optimal value satisfies
N
X
H2 (P ) ≤ pi `(c∗ (i)) ≤ H2 (P ) + 1 , (14.2)
i=1

where H2 (P ) is the entropy of P ,

 
X 1
H2 (P ) = pi log2 .
pi
i∈[N ]:pi >0

When pi = 1/N is uniform, the naive idea of using a code of uniform length is
recovered, but for non-uniform distributions, the code adapts to assign shorter
codes to symbols with larger probability. It is worth pointing out that the sum
is only over outcomes that occur with non-zero probability, which is motivated
by observing that limx→0+ x log(1/x) = 0 or by thinking of the entropy as an
expectation of the log probability with respect to P , and expectations should not
change when the value of the random variable is perturbed on a measure zero set.
It turns out that H2 (P ) is not just an approximation on the expected length of
the Huffman code, but is itself a fundamental quantity. Imagine that Alice wants
to transmit a long string of symbols sampled from P . She could use a Huffman
code to send Bob each symbol one at a time, but this introduces rounding errors
that accumulate as the message length grows. There is another scheme called
arithmetic coding for which the average number of bits per symbol approaches
H2 (P ) and the source coding theorem says that this is unimprovable.
The definition of entropy using base 2 makes sense from the perspective of
sending binary message. Mathematically, however, it is more convenient to define
 14.2 Relative Entropy 187

the entropy using the natural logarithm:

 
X 1
H(P ) = pi log . (14.3)
pi
i∈[N ]:pi >0

This is nothing more than a scaling of the H2 . Measuring information using base
2 logarithms has a unit of bits, and for the natural logarithm the unit is nats.
By slightly abusing terminology, we will also call H(P ) the entropy of P .

14.2 Relative Entropy

Suppose that Alice and Bob agree to use a code that is optimal when X is sampled
from distribution Q. Unbeknownst to them, however, X is actually sampled from
distribution P . The relative entropy between P and Q measures how much longer
the messages are expected to be using the optimal code for Q than what would be
obtained using the optimal code for P . Letting pi = P (X = i) and qi = Q(X = i),
assuming Shannon coding, working out the math while dropping d·e leads to the
definition of relative entropy as
     
X 1 X 1 X pi
D(P, Q) = pi log − pi log = pi log
qi pi qi
i∈[N ]:pi >0 i∈[N ]:pi >0 i∈[N ]:pi >0
(14.4)
From the coding interpretation, one conjectures that D(P, Q) ≥ 0. Indeed, this
is easy to verify using Jensen’s inequality. Still poking around the definition,
what happens when qi = 0 and pi = 0? This means that symbol i is superfluous
and the value of D(P, Q) should not be impacted by introducing superfluous
symbols. And again, it is not by the definition of the expectations. We also see
that the sufficient and necessary condition for D(P, Q) < ∞ is that for each i
with qi = 0, we also have pi = 0. The condition we discovered is equivalent to
saying that P is absolutely continuous with respect to Q. Note that absolute
continuity only implies a finite relative entropy when X takes on finitely many
values (Exercise 14.2).
This brings us back to defining relative entropy between probability measures
P and Q on arbitrary measurable spaces (Ω, F). When the support of P is
uncountable, defining the entropy via communication is hard because infinitely
many symbols are needed to describe some outcomes. This seems to be a
fundamental difficulty. Luckily, the impasse gets resolved automatically if we
only consider relative entropy. While we cannot communicate the outcome, for
any finite discretisation of the possible outcomes, the discretised values can be
communicated finitely, and all our definitions will work. Formally, a discretisation
to [N ] is specified by a F/2[N ] -measurable map X : Ω → [N ]. Then the entropy
of P relative Q can be defined as
D(P, Q) = sup sup D(PX , QX ) , (14.5)
N ∈N+ X
 14.2 Relative Entropy 188

where PX is the push-forward of P on [N ] defined by PX (A) = P (X ∈ A). The

inner supremum is over all F/2N -measurable maps. Informally we take all possible
discretisations X (with no limit on the ‘fineness’ of the discretisation) and define
D(P, Q) as the excess information when expecting to see X with X ∼ QX , while
in reality X ∼ PX . As we shall see soon, this is indeed a reasonable definition.

Theorem 14.1. Let (Ω, F) be a measurable space, and let P and Q be measures
on this space. Then,
R  
 log dP (ω) dP (ω) , if P  Q ;
D(P, Q) = dQ
∞ , otherwise .

Note that the relative entropy between P and Q can still be infinite even when
P  Q. Note also that in the case of discrete measures, the above expression
reduces to (14.4). For calculating relative entropies densities one often uses
densities: If λ is a common dominating σ-finite measure for P and Q (that is,
P  λ and Q  λ both hold), then letting p = dP dλ and q = dλ , if also P  Q,
dQ

the chain rule gives dQ dλ = dλ , which lets us write

dP dQ dP

Z  
p
D(P, Q) = p log dλ . (14.6)
q
This is probably the best-known expression for relative entropy and is often used
as a definition. Note that for probability measures, a common dominating σ-finite
measure can always be bound. For example, λ = P + Q always dominates both
P and Q.
Relative entropy is a kind of ‘distance’ measure between distributions P and
Q. In particular, D(P, Q) = 0 whenever P = Q, and otherwise D(P, Q) > 0.
However, strictly speaking, the relative entropy is not a distance because it
satisfies neither the triangle inequality nor is it symmetric. Nevertheless, it serves
the same purpose.
The relative entropy between many standard distributions is often quite easy
to compute. For example, the relative entropy between two Gaussians with means
µ1 , µ2 ∈ R and common variance σ 2 is
(µ1 − µ2 )2
D(N (µ1 , σ 2 ), N (µ2 , σ 2 )) = .
2σ 2
The dependence on the difference in means and the variance is consistent with
our intuition. If µ1 is close to µ2 , then the ‘difference’ between the distributions
should be small, but if the variance is very small, then there is little overlap, and
the difference is large. The relative entropy between two Bernoulli distributions
with means p, q ∈ [0, 1] is
   
p 1−p
D(B(p), B(q)) = p log + (1 − p) log ,
q 1−q
where 0 log(·) = 0. Due to its frequent appearance at various places, D(B(p), B(q))
 14.2 Relative Entropy 189

gets the honour of being abbreviated to d(p, q), which we have met before in
Definition 10.1.
We are nearing the end of our whirlwind tour of relative entropy. It remains
to state the key lemma that connects the relative entropy to the hardness of
hypothesis testing.

Theorem 14.2 (Bretagnolle–Huber inequality). Let P and Q be probability

measures on the same measurable space (Ω, F), and let A ∈ F be an arbitrary
event. Then,
1
P (A) + Q(Ac ) ≥ exp (− D(P, Q)) , (14.7)
2
where Ac = Ω \ A is the complement of A.

The proof may be found at the end of the chapter, but first some interpretation
and a simple application. Suppose that D(P, Q) is small; then P is close to Q
in some sense. Since P is a probability measure, we have P (A) + P (Ac ) = 1.
If Q is close to P , then we might expect that P (A) + Q(Ac ) should be large.
The purpose of the theorem is to quantify just how large. Note that if P is not
absolutely continuous with respect to Q, then D(P, Q) = ∞, and the result is
vacuous. Also note that the result is symmetric. We could replace D(P, Q) with
D(Q, P ), which sometimes leads to a stronger result because the relative entropy
is not symmetric.
Returning to the hypothesis-testing problem described in the previous chapter,
let X be normally distributed with unknown mean µ ∈ {0, ∆} and variance
σ 2 > 0. We want to bound the quality of a rule for deciding what is the real mean
from a single observation. The decision rule is characterised by a measurable
set A ⊆ R on which the predictor guesses µ = ∆ (it predicts µ = 0 on the
complement of A). Let P = N (0, σ 2 ) and Q = N (∆, σ 2 ). Then the probability
of an error under P is P (A), and the probability of error under Q is Q(Ac ). The
reader surely knows what to do next. By Theorem 14.2, we have
 
1 1 ∆2
P (A) + Q(A ) ≥ exp (− D(P, Q)) = exp − 2 .
c
2 2 2σ
If we assume that the signal-to-noise ratio is small, ∆2 /σ 2 ≤ 1, then
 
1 1 3
P (A) + Q(A ) ≥ exp −
c
≥ ,
2 2 10
which implies max {P (A), Q(Ac )} ≥ 3/20. This means that no matter how we
chose our decision rule, we simply do not have enough data to make a decision
for which the probability of error on either P or Q is smaller than 3/20.

Proof of Theorem 14.2 For reals a, b, we abbreviate max {a, b} = a ∨ b and

min {a, b} = a ∧ b. The result is trivial if D(P, Q) = ∞. On the other hand, by
Theorem 14.1, D(P, Q) < ∞ implies that P  Q. Let ν = P + Q. Then P, Q  ν,
which by Theorem 2.13 ensures the existence of the Radon–Nikodym derivatives
 14.3 Notes 190

R  
p = dP
dν and q = dν . By Eq. (14.6), D(P, Q) =
dQ
p log pq dν. For brevity, when
writing integrals with
R respect to ν, in this proof, we will drop dν. Thus, we will
write, for example, p log(p/q) for the above integral.
Instead of (14.7), we prove the stronger result that
Z
1
p ∧ q ≥ exp(− D(P, Q)) . (14.8)
2
R R R R R
This indeed is sufficient since p ∧ q = A p ∧ q + Ac p ∧ q ≤ A p + Ac q =
P (A) + Q(Ac ). We start with an inequality attributed to French mathematician
Lucien Le Cam, which lower-bounds the left-hand side of Eq. (14.8). The inequality
states that
Z Z 2
1 √
p∧q ≥ pq . (14.9)
2
How do we get this inequality? Starting from the right-hand side above, using
pq = (p ∧ q)(p ∨ q) and Cauchy–Schwarz we get
Z 2 Z 2 Z  Z 
√ p
pq = (p ∧ q)(p ∨ q) ≤ p∧q p∨q .

Now,
R using Rp ∧ q + p ∨ q = p + q, the proof is finished by substituting
p ∨ q = 2 − p ∧ q ≤ 2 and dividing both sides by two. It remains to lower-bound
the right-hand side of (14.9). For this, we use Jensen’s inequality. First, we write
(·)2 as exp(2 log(·)) and then move the log inside the integral:
Z 2  Z   Z r 
√ √ q
pq = exp 2 log pq = exp 2 log p
p>0 p
 Z    Z  
1 q p
≥ exp 2 p log = exp − p log
p>0 2 p pq>0 q
 Z  
p
= exp − p log = exp (− D(P, Q)) .
q
In the fourth and the last step, we used that since P  Q, q = 0 implies p = 0,
and so p > 0 implies q > 0, and eventually pq > 0. The result is completed by
chaining the inequalities.

14.3 Notes

1 A code c : N+ → {0, 1}∗ is uniquely decodable if i1 , . . . , in 7→ c(i1 ) · · · c(in )

is injective, where on the right-hand side the codes are simply concatenated.
Kraft’s inequality states that for any uniquely decodable code c,
∞
X
2−`(c(i)) ≤ 1 . (14.10)
i=1
 14.3 Notes 191

P∞ −`
Furthermore, for any (`n )∞ n=1 satisfying i=1 2 ≤ 1, there exists a prefix
code c : N+ → {0, 1}∗ such that `(c(i)) = `i . The second part justifies our
restriction to prefix codes rather than uniquely decodable codes in the definition
of the entropy.
2 The supremum in the definition given in Eq. (14.5) may often be taken over
a smaller set. Precisely, let (X , G) be a measurable space and suppose that
G = σ(F) where F is a field. Note that a field is defined by the same axioms
as a σ-algebra except that being closed under countable unions is replaced by
the condition that it be closed under finite unions. Then, for measures P and
Q on (X , G), it holds that

D(P, Q) = sup D(Pf , Qf ) ,

f

where the supremum is over F/2[n] -measurable functions. This result is known
as Dobrushin’s theorem.
R
3 In the proof of Theorem 14.2 we used the inequality P (A) + Q(Ac ) ≥ p ∧ q.
Looking at the proof, it is not hard to see that the inequality becomes an equality
when A = {p ≤ q} = {q/p ≥ 1}. Reader’s familiar with statistical decision
theory may recognize that this is a special case of the Neyman-Pearson
lemma which states that the most powerful test among all statistical tests in a
simple hypothesis testing problem at a given significance level is the likelihood
ratio test. Exercise 14.13 explores this connection.
4 How tight is Theorem 14.2? We remarked already that D(P, Q) = 0 if and only
if P = Q. But in this case, Theorem 14.2 only gives
1 1
1 = P (A) + Q(Ac ) ≥ exp (− D(P, Q)) = ,
2 2
which does not seem so strong. From where does the weakness arise? The
answer is in Le Cam’s inequality, Eq. (14.9), which can be refined by
Z 2 Z  Z  Z  Z 
√
pq ≤ p∧q p∨q = p∧q 2− p∧q .

By solving the quadratic inequality, we have

s Z 2
Z
√
P (A) + Q(A ) ≥
c
p∧q ≥1− 1− pq
p
≥ 1 − 1 − exp (− D(P, Q)) , (14.11)

which gives a modest improvement on Theorem 14.2 that becomes more

pronounced when D(P, Q) is close to zero, as demonstrated by Fig. 14.2. This
stronger bound might be useful for fractionally improving constant factors in
lower bounds, but we do not know of any application for which it is really
crucial, and the more complicated form makes it cumbersome to use. Part of
the reason for this is that the situation where D(P, Q) is small is better dealt
with using a different inequality, as explained in the next note.
 14.3 Notes 192

1 x/2
√
1− 1−x
p
1− log(1/x)/2

0.5

−0.5

0 1
x

Figure 14.2 Tightening the inequality of Le Cam. Here, x = exp(− D(P, Q)). Higher
values are better as the figure shows lower bounds on P (A) + Q(Ac ).

5 Another inequality from information theory is Pinsker’s inequality, which

states for measures P and Q on the same probability space (Ω, F) that
r
1
δ(P, Q) = sup P (A) − Q(A) ≤ D(P, Q) . (14.12)
A∈F 2
The quantity on the left-hand side is called the total variation distance
between P and Q, which is a distance on the space of probability measures on
a probability space. From this we can derive for any measurable A ∈ F that
r s  
1 1 1
P (A) + Q(A ) ≥ 1 −
c
D(P, Q) = 1 − log .
2 2 exp(− D(P, Q))
(14.13)
Examining Fig. 14.2 shows that this is an improvement on Eq. (14.11) when
D(P, Q) is small. However, we also see that in the opposite case, when D(P, Q)
is large, Eq. (14.13) is worse than Eq. (14.11), or the inequality in Theorem 14.2.
6 We saw the total variation distance in Eq. (14.12). There are two other
‘distances’ that are occasionally useful. These are the Hellinger distance
and the χ-squared distance, which, using the notation in the proof of
Theorem 14.2, are defined by
sZ s  Z 
√ √ 2 √
h(P, Q) = ( p − q) = 2 1 − pq (14.14)
Z 2 Z
(p − q) p2
χ2 (P, Q) = = − 1. (14.15)
q q
The Hellinger distance is bounded and exists for all probability measures P
and Q. A necessary condition for the χ2 -distance to exist is that P  Q. Like
 14.4 Bibliographic Remarks 193

the total variation distance, the Hellinger distance is actually a distance (it is
symmetric and satisfies triangle inequality), but the χ2 -‘distance’ is not. It is
possible to show (Tsybakov [2008], chapter 2) that
δ(P, Q)2 ≤ h(P, Q)2 ≤ D(P, Q) ≤ χ2 (P, Q) . (14.16)
All the inequalities are tight for some choices of P and Q, but the examples
do not chain together, as evidenced by Pinsker’s inequality, which shows that
δ(P, Q)2 ≤ D(P, Q)/2 (which is also tight for some P and Q).
7 The entropy for distribution P was defined as H(P ) in Eq. (14.3). If X is a
random variable, then H(X) is defined to be the entropy of the law of X. This
is a convenient notation because it allows one to write H(f (X)) and H(XY )
and similar expressions.

14.4 Bibliographic Remarks

There are many references for information theory. Most well known (and
comprehensive) is the book by Cover and Thomas [2012]. Another famous book
is the elementary and enjoyable introduction by MacKay [2003]. The approach
we have taken for defining and understanding the relative entropy is inspired by
an excellent shorter book by Gray [2011]. Theorem 14.1 connects our definition of
relative entropies to densities (the ‘classic definition’). It can be found in §5.2 of
the aforementioned book. Dobrushin’s theorem is due to him [Dobrushin, 1959].
An alternative source is lemma 5.2.2 in the book of Gray [2011]. Theorem 14.2 is
due to Bretagnolle and Huber [1979]. We also recommend the book by Tsybakov
[2008] as a good source for learning about information theoretic lower bounds in
statistical settings.

14.5 Exercises

14.1 Let P be a probability distribution on N+ and pi = P ({i}). Show that for

any prefix code c : N+ → {0, 1}∗ , it holds that
∞
X
pi `(c(i)) ≥ H2 (P ) .
i=1

Hint Use Kraft’s inequality from Note 1.

14.2 Find probability measures P and Q on N+ with P  Q and D(P, Q) = ∞.

14.3 Prove the inequality in Eq. (14.10) for prefix free codes c.

Hint Consider an infinite sequence of independent Bernoulli random variables

(Xn )∞
n=1 where Xn ∼ B(1/2). Viewing X as an infinite binary string, what is the
probability that X has a prefix that is a code for some symbol?
 14.5 Exercises 194

14.4 Let (Ω, F) be a measurable space, and let P, Q : F → [0, 1] be probability

measures. Let a < b and X : Ω → [a, b] be a F-measurable random variable.
Prove that
Z Z 
 
 X(ω)dP (ω) − X(ω)dQ(ω) ≤ (b − a)δ(P, Q) .

Ω Ω

14.5 (Entropy inequalities) Prove that each of the inequalities in Eq. (14.16)
is tight.

14.6 (Counting measure absolute continuity and derivatives) Let Ω be

P
a countable set and p : Ω → [0, 1] be a distribution on Ω so that ω∈Ω p(ω) = 1.
P
Let P be the measure associated with p, which means that P (A) = ω∈A p(ω).
Recall that the counting measure µ is the measure on (Ω, 2Ω ) given by µ(A) = |A|
if A is finite and µ(A) = ∞ otherwise.

(a) Show that P is absolutely continuous with respect to µ.

(b) Show that the Radon–Nykodim dP/dµ exists and that dP/dµ(ω) = p(ω).

14.7 (Relative entropy for Gaussian distributions) For each i ∈ {1, 2},
let µi ∈ R, σi2 > 0 and Pi = N (µi , σi2 ). Show that
   
1 σ22 σ12 (µ1 − µ2 )2
D(P1 , P2 ) = log + − 1 + .
2 σ12 σ22 2σ22

14.8 Let λ be the Lebesgue measure on (R, B(R)). Find

(a) a probability measure (R, B(R)) that is not absolutely continuous with
respect to λ; and
(b) a probability measure P on (R, B(R)) that is absolutely continuous to λ
with D(P, Q) = ∞ where Q = N (0, 1) is the standard Gaussian measure.

14.9 (Data processing inequality) Let P and Q be measures on (Ω, F), and
let G be a sub-σ-algebra of F and PG and QG be the restrictions of P and Q to
(Ω, G). Show that D(PG , QG ) ≤ D(P, Q).

14.10 Let (Ω, F) be a measurable space and P, Q : B(R) × Ω → [0, 1] be a pair

of probability kernels from (Ω, F) to (R, B(R)). Prove that

V = {ω ∈ Ω : D(P (· | ω), Q(· | ω)) = ∞} ∈ F .

Hint Apply Dobrushin’s theorem to the field of finite unions of rational-valued

intervals in R.
14.11 (Chain rule) Let P and Q be measures on (Rn , B(Rn )), and for t ∈ [n],
let Xt (x) = xt be the coordinate project from Rn → R. Then let Pt and Qt be
regular versions of Xt given X1 , . . . , Xt−1 under P and Q, respectively. Show
 14.5 Exercises 195

that

n
X
D(P, Q) = EP [D(Pt (· | X1 , . . . , Xt−1 ), Qt (· | X1 , . . . , Xt−1 ))] . (14.17)
t=1

Hint This is a rather technical exercise. You will likely need to apply a
monotone class argument [Kallenberg, 2002, theorem 1.1]. For the definition
of a regular version, see [Kallenberg, 2002, theorem 5.3] or Theorem 3.11.
Briefly, Pt is a probability kernel from (Rt−1 , B(Rt−1 )) to (R, B(R)) such that
Pt (A | x1 , . . . , xt−1 ) = P (Xt ∈ A | X1 , . . . , Xt−1 ) with P -probability one for all
A ∈ B(R).
14.12 (Chain rule (cont.)) Let P and Q be measures on (Rn , B(Rn )), and
for t ∈ [n], let Xt (x) = xt be the coordinate project from Rn → R. Then let Pt
and Qt be regular versions of Xt given X1 , . . . , Xt−1 under P and Q, respectively.
Let τ be a stopping time adapted to the filtration generated by X1 , . . . , Xn with
τ ∈ [n] almost surely. Show that

" τ #
X
D(P|Fτ , Q|Fτ ) = EP D(Pt (· | X1 , . . . , Xt−1 ), Qt (· | X1 , . . . , Xt−1 )) .
t=1

14.13 (Neyman-Pearson lemma) The simple hypothesis testing problem is

specified by a measurable space (X , F) and two distinct probability distributions
over this space, P and Q. The problem is to decide based on observing a random
element X taking values in X whether its distribution follows P or Q. The decision
rule, or test, is a measurable map f from X to (say) {P, Q}: when f (x) = P , the
rule decides in favor of P given the observation x, otherwise it decides in favor of
Q. There are two ways a test can be wrong and a test is characterized by the
probability of an incorrect decision in the two cases. In particular,
R the probability
of the test making an error when X follows P is e(f, P ) := R I {f (x) = Q} P (dx)
and the probability of error when X follows Q is e(f, Q) := I {f (x) = P } Q(dx).
Show that the following hold:

(a) If a test f is such that f (x) = P if p(x) > ηq(x) and f (x) = Q if p(x) < ηq(x)
and α = e(f, P ) then e(f, Q) = min{e(f 0 , Q) : e(f 0 , P ) ≤ α}. Here,
p = dP/dν and q = dQ/dν where ν is a common dominating measure
of P and Q.

(b) If f , as postulated in the previous part exist, and f 0 is a test such that
e(f 0 , Q) = e(f, Q) and e(f 0 , P ) ≤ α then e(f 0 , P ) = α. Furthermore, f 0 = f
holds except perhaps over the union of the set {x : p(x) = ηq(x)} and a set
that has zero measure under both P and Q.
 14.5 Exercises 196

In statistics, the upper bound α on e(f, P ) is called the significance level of

the test, while 1 − e(f, Q) is called its power, and breaking the symmetry,
e(f, P ) is called the type-I error, while e(f, Q) is called the type-II error.
Since the decision rule in the first part of the lemma can be expressed as a
function of the ratio p/q of densities (or likelihoods), the test in this part
is called a likelihood ratio test. The first part of the exercise says that the
most powerful test for a given significance level are the likelihood ratio
tests and the second part says that these are the only most powerful tests
(uniqueness).
15 Minimax Lower Bounds

After the short excursion into information theory, let us return to the world
of k-armed stochastic bandits. In what follows, we fix the horizon n > 0 and
the number of actions k > 1. This chapter has two components. The first is
an exact calculation of the relative entropy between measures in the canonical
bandit model for a fixed policy and different bandits. In the second component,
we prove a minimax lower bound that formalises the intuitive arguments given in
Chapter 13.

15.1 Relative Entropy Between Bandits

The following result will be used repeatedly. Some generalisations are provided
in the exercises.
Lemma 15.1 (Divergence decomposition). Let ν = (P1 , . . . , Pk ) be the reward
distributions associated with one k-armed bandit, and let ν 0 = (P10 , . . . , Pk0 ) be the
reward distributions associated with another k-armed bandit. Fix some policy π
and let Pν = Pνπ and Pν 0 = Pν 0 π be the probability measures on the canonical
bandit model (Section 4.6) induced by the n-round interconnection of π and ν
(respectively, π and ν 0 ). Then,
k
X
D(Pν , Pν 0 ) = Eν [Ti (n)] D(Pi , Pi0 ) . (15.1)
i=1

Proof Assume that D(Pi , Pi0 ) < ∞ for all i ∈ [k]. It follows that Pi  Pi0 . Define
Pk Pk
λ = i=1 Pi + Pi0 , which is the measure defined by λ(A) = i=1 (Pi (A) + Pi0 (A))
dPν
for any measurable set A. Theorem 14.1 shows that, as long as dP < +∞,
ν0
  
dPν
D(Pν , Pν 0 ) = Eν log .
dPν 0
Recalling that ρ is the counting measure over [k], we find that the Radon–Nikodym
derivative of Pν with respect to the product measure (ρ × λ)n is given in Eq. (4.7)
as
n
Y
pνπ (a1 , x1 , . . . , an , xn ) = πt (at | a1 , x1 , . . . , at−1 , xt−1 )pat (xt ) .
t=1
 15.2 Minimax Lower Bounds 198

The density of Pν 0 is identical except that pat is replaced by p0at . Then

pa (xt )
n X
dPν
log (a1 , x1 , . . . , an , xn ) = log 0 t ,
dPν 0 t=1
pat (xt )

where we used the chain rule for Radon–Nikodym derivatives and the fact that
the terms involving the policy cancel. Taking expectations of both sides,
  X  
pA (Xt )
n
dPν
Eν log (A1 , X1 , . . . , An , Xn ) = Eν log 0 t ,
dPν 0 t=1
pAt (Xt )

and
  " "  ##
pAt (Xt ) pAt (Xt )   
Eν log 0 = Eν Eν log 0  At = Eν D(PAt , PA0 t ) ,
pAt (Xt ) pAt (Xt ) 

where in the second equality we used that under Pν (·|At ), the distribution of Xt
is dPAt = pAt dλ. Plugging back into the previous display,
  X  
pAt (Xt )
n
dPν
Eν log (A1 , X1 , . . . , An , Xn ) = Eν log 0
dPν 0 t=1
pAt (Xt )
n k
" n
#
X   X X
= Eν D(PAt , PAt ) =0
Eν I {At = i} D(PAt , PAt )
0

t=1 i=1 t=1

k
X
= Eν [Ti (n)] D(Pi , Pi0 ) .
i=1

When the right-hand side of (15.1) is infinite, by our previous calculation, it is

not hard to see that the left-hand side will also be infinite.

We note in passing that the divergence decomposition holds regardless of

whether the action set is discrete or not. In its more general form, the sum
over the actions must be replaced by an integral with respect to an appropriate
non-negative measure, which generalises the expected number of pulls of arms.
For details, see Exercise 15.8.

15.2 Minimax Lower Bounds

Recall that EN
k
(1) is the class of Gaussian bandits with unit variance, which can
be parameterised by their mean vector µ ∈ Rk . Given µ ∈ Rk , let νµ be the
Gaussian bandit for which the ith arm has reward distribution N (µi , 1).

Theorem 15.2. Let k > 1 and n ≥ k − 1. Then, for any policy π, there exists a
mean vector µ ∈ [0, 1]k such that
1p
Rn (π, νµ ) ≥ (k − 1)n .
27
 15.2 Minimax Lower Bounds 199

Figure 15.1 The idea of the minimax lower bound. Given a policy and one environment,
the evil antagonist picks another environment so that the policy will suffer a large regret
in at least one environment.

Since νµ ∈ ENk
(1), it follows that the minimax regret for EN
k
(1) is lower-bounded
by the right-hand side of the above display as soon as n ≥ k − 1:
1p
Rn∗ (EN
k
(1)) ≥ (k − 1)n .
27
The idea of the proof is illustrated in Fig. 15.1.

Proof Fix a policy π. Let ∆ ∈ [0, 1/2] be some constant to be chosen later. As
suggested in Chapter 13, we start with a Gaussian bandit with unit variance
and mean vector µ = (∆, 0, 0, . . . , 0). This environment and π give rise to the
distribution Pνµ ,π on the canonical bandit model (Hn , Fn ). For brevity we will
use Pµ in place of Pνµ ,π , and expectations under Pµ will be denoted by Eµ . To
choose the second environment, let

i = argminj>1 Eµ [Tj (n)] .

Pk
Since j=1 Eµ [Tj (n)] = n, it holds that Eµ [Ti (n)] ≤ n/(k −1). The second bandit
is also Gaussian with unit variance and means

µ0 = (∆, 0, 0, . . . , 0, 2∆, 0, . . . , 0) ,

where specifically µ0i = 2∆. Therefore, µj = µ0j except at index i and the optimal
arm in νµ is the first arm, while in νµ0 arm i is optimal. We abbreviate Pµ0 = Pνµ0 ,π .
Lemma 4.5 and a simple calculation lead to
n∆ n∆
Rn (π, νµ ) ≥ Pµ (T1 (n) ≤ n/2) and Rn (π, νµ0 ) > Pµ0 (T1 (n) > n/2) .
2 2
Then, applying the Bretagnolle–Huber inequality from the previous chapter
 15.3 Notes 200

(Theorem 14.2),

n∆
Rn (π, νµ ) + Rn (π, νµ0 ) > (Pµ (T1 (n) ≤ n/2) + Pµ0 (T1 (n) > n/2))
2 (15.2)
n∆
≥ exp(− D(Pµ , Pµ0 )) .
4

It remains to upper-bound D(Pµ , Pµ0 ). For this, we use Lemma 15.1 and the
definitions of µ and µ0 to get

(2∆)2 2n∆2
D(Pµ , Pµ0 ) = Eµ [Ti (n)] D(N (0, 1), N (2∆, 1)) = Eµ [Ti (n)] ≤ .
2 k−1

Plugging this into the previous display, we find that

 
n∆ 2n∆2
Rn (π, νµ ) + Rn (π, νµ0 ) ≥ exp − .
4 k−1

p
The result is completed by choosing ∆ = (k − 1)/4n ≤ 1/2, where the inequality
follows from the assumptions in the theorem statement. The final steps are lower
bounding exp(−1/2) and using 2 max(a, b) ≥ a + b.

We encourage readers to go through the alternative proof outlined in

Exercise 15.2, which takes a slightly different path.

15.3 Notes

1 We used the Gaussian noise model because the KL divergences are so easily
calculated in this case, but all that we actually used was that D(Pi , Pi0 ) =
O((µi − µ0i )2 ) when the gap between the means ∆ = µi − µ0i is small. While
this is certainly not true for all distributions, it very often is. Why is that? Let
{Pµ : µ ∈ R} be some parametric family of distributions on Ω and assume that
distribution Pµ has mean µ. Assuming the densities are twice differentiable
and that everything is sufficiently nice that integrals and derivatives can be
exchanged (as is almost always the case), we can use a Taylor expansion about
 15.3 Notes 201

µ to show that
 
∂  1 ∂2 
D(Pµ , Pµ+∆ ) ≈ D(Pµ , Pµ+∆ ) ∆+ D(P , P )  ∆2
2 2 µ µ+∆ 
∂∆ ∆=0 ∂∆ ∆=0
Z   
∂ dPµ  1
= log dPµ  ∆ + I(µ)∆2
∂∆ Ω dPµ+∆ ∆=0 2
Z  
∂ dPµ+∆  1
=− log dPµ ∆ + I(µ)∆2
Ω ∂∆ dP µ

∆=0 2
Z 
∂ dPµ+∆  1
=− dPµ ∆ + I(µ)∆2
Ω ∂∆ dPµ

∆=0 2
Z 
∂ dPµ+∆  1
=− dPµ  ∆ + I(µ)∆2
∂∆ Ω dPµ ∆=0 2
Z 
∂  1
=− dPµ+∆  ∆ + I(µ)∆2
∂∆ Ω ∆=0 2
1
= I(µ)∆2 ,
2
where I(µ), introduced in the second line, is called the Fisher information
of the family (Pµ )µ at µ. Note that if λ is a common dominating measure for
(Pµ+∆ ) for ∆ small, dPµ+∆ = pµ+∆ dλ and we can write
Z 
∂2 
I(µ) = − 2
log pµ+∆  pµ dλ ,
∂∆ ∆=0

which is the form that is usually given in elementary texts. The upshot of all
this is that D(Pµ , Pµ+∆ ) for ∆ small is indeed quadratic in ∆, with the scaling
√
provided by I(µ), and as a result the worst-case regret is always O( nk),
provided the class of distributions considered is sufficiently rich and not too
bizarre.
√
2 We have now shown a lower bound that is Ω( nk), while many of the upper
bounds were O(log(n)). There is no contradiction because the logarithmic
bounds depended on the inverse suboptimality gaps, which may be very large.
3 Our lower bound was only proven for n ≥ k − 1. In Exercise 15.3, we ask you
to show that when n < k − 1, there exists a bandit such that
n(2k − n − 1) n
Rn ≥ > .
2k 2
4 The method used to prove Theorem 15.2 can be viewed as a generalisation
and strengthening of Le Cam’s method in statistics. Recall that Eq. (15.2)
establishes that for any µ and µ0 ,
n∆
inf sup Rn (π, ν) ≥ exp(− D(Pµ , Pµ0 )) .
π ν 8
To explain Le Cam’s method, we need a little notation. Let X be an outcome
space, P a set of measures on X and θ : P → Θ, where (Θ, d) is a metric space.
 15.4 Bibliographic Remarks 202

An estimator is a function θ̂ : X n → Θ. Le Cam’s method is used for proving

minimax lower bounds on the expected error of the estimator, which is
h i
inf sup EX1 ,...,Xn ∼P n d(θ̂(X1 , . . . , Xn ), θ(P )) . (15.3)
θ̂ P ∈P

The idea is to choose P0 , P1 ∈ P to maximise d(θ(P0 ), θ(P1 )) exp(−n D(P0 , P1 )),

on the basis that for any P0 , P1 ∈ P,
∆
Eq. (15.3) ≥ exp (−n D(P0 , P1 )) , (15.4)
8
where ∆ = d(θ(P0 ), θ(P1 )). There are two differences compared to the bandit
lower bound: (i) we deal with the sequential setting, and (ii) having chosen P0
we choose P1 in a way that depends on the algorithm. This provides a much
needed extra boost, without which the method would be unable to capture
how the characteristics of P are reflected in the minimax risk (or regret, in our
case).

15.4 Bibliographic Remarks

The first work on lower bounds that we know of was the remarkably precise
minimax analysis of two-armed Bernoulli bandits by Vogel [1960]. The Bretagnolle–
Huber inequality (Theorem 14.2) was first used for bandits by Bubeck et al.
[2013b]. As mentioned in the notes, the use of this inequality for proving lower
bounds is known as Le Cam’s method in statistics [Le Cam, 1973]. The proof
of Theorem 15.2 uses the same ideas as Gerchinovitz and Lattimore [2016],
while the alternative proof in Exercise 15.2 is essentially due to Auer et al.
[1995], who analysed the more difficult case where the rewards are Bernoulli (see
Exercise 15.4). Yu [1997] describes some alternatives to Le Cam’s method for the
passive, statistical setting. These alternatives can be (and often are) adapted to
the sequential setting.

15.5 Exercises

15.1 (Le Cam’s method) Establish the claim in Eq. (15.4).

15.2 (Alternative proof of Theorem 15.2) Here p you will prove

Theorem 15.2 with a different method. Let c > 0 and ∆ = 2c k/n, and for each
(i)
i ∈ {0, 1, . . . , k}, let µ(i) ∈ Rk satisfy µj = I {i = j} ∆. Further abbreviate the
notation in the proof of Theorem 15.2 by letting Ei [·] = Eµ(i) [·].
(a) Use Pinsker’s inequality (Eq. 14.12) and Lemma 15.1 and the result of
Exercise 14.4 to show
r
1 2 p
Ei [Ti (n)] ≤ E0 [Ti (n)] + n ∆ E0 [Ti (n)] = E0 [Ti (n)] + c nkE0 [Ti (n)] .
4
 15.5 Exercises 203

Pk
(b) Using the previous part, Jensen’s inequality and the identity i=1 E0 [Ti (n)] =
n, show that
k
X k p
X
Ei [Ti (n)] ≤ n + c nkE0 [Ti (n)] ≤ n + ckn .
i=1 i=1

(c) Let Ri = Rn (π, Gµ(i) ). Find a choice of c > 0 for which

k
X k
X
Ri = ∆ (n − Ei [Ti (n)]) ≥ ∆ (nk − n − ckn)
i=1 i=1
r r
k nk k
= 2c (nk − n − ckn) ≥
n 8 n
(d) Conclude that there exists an i ∈ [k] such that
1√
Ri ≥ kn .
8

The method used in this exercise is borrowed from Auer et al. [2002b] and
is closely related to the lower-bound technique known as Assouad’s method
in statistics [Yu, 1997].

15.3 (Lower bound for small horizons) Let k > 1 and n < k. Prove that
for any policy π there exists a Gaussian bandit with unit variance and means
µ ∈ [0, 1]k such that Rn (π, νµ ) ≥ n(2k − n − 1)/(2k) > n/2.

15.4 (Lower bounds for Bernoulli bandits) Recall from Table 4.1 that
EBk is the set of k-armed Bernoulli bandits. Show that there exists a universal
constant c > 0 such that for any 2 ≤ k ≤ n, it holds that:
√
Rn∗ (EBk ) = inf sup Rn (π, ν) ≥ c nk .
π ν∈E k
B

Hint Use the fact that KL divergence is upper bounded by the χ-squared
distance (Eq. (14.16)).
15.5 In Chapter 9 we proved that if π is the MOSS policy and ν ∈ ESG
k
(1), then
 
√ X
Rn (π, ν) ≤ C  kn + ∆i  ,
i:∆i >0

where C > 0 is a universal constant. Prove that the dependence on the sum
cannot be eliminated.

Hint You will have to use that Ti (t) is an integer for all t.
15.6 (Lower bound for explore-then-commit) Let ETCnm be the explore-
then-commit policy with inputs n and m respectively (Algorithm 1). Prove that
 15.5 Exercises 204

for all m, there exists a µ ∈ [0, 1]k such that

n o
Rn (ETCnm , νµ ) ≥ c min n, n2/3 k 1/3 ,

where c > 0 is a universal constant.

15.7 (Stopping-time version of divergence decomposition) Consider

the setting of Lemma 15.1, and let Ft = σ(A1 , X1 , . . . , At , Xt ) and τ be an
(Ft )-measurable stopping time. Then, for any random element X that is Fτ -
measurable,
k
X
D(PνX , Pν 0 X ) ≤ Eν [Ti (τ )] D(Pi , Pi0 ) ,
i=1

where PνX and Pν 0 X are the laws X under ν and ν 0 respectively.

15.8 (Divergence decomposition for more general action spaces) The

purpose of this exercise is to show that the divergence decomposition lemma
(Lemma 15.1) continues to hold for more general action spaces (A, G). Starting
from the set-up of Section 4.7, let Pν = Pνπ and Pν 0 = Pν 0 π be the measures on the
canonical bandit model induced by the interconnection of π and ν (respectively,
π and ν 0 ).
(a) Prove that
Z
D(Pν , Pν 0 ) = D(Pa , Pa0 ) dGν (a) , (15.5)
A
Pn
where Gν is a measure on (A, G) defined by Gν (B) = Eν [ t=1 I {At ∈ B}].
(b) Prove that
" n #
X
D(Pν , Pν 0 ) = E D(PAt , PAt ) .
0

t=1

Hint Use an appropriately adjusted form of the chain rule for relative entropy
from Exercise 14.11.
16 Instance-Dependent Lower Bounds

In the last chapter, we proved a lower bound on the minimax regret for subgaussian
bandits with suboptimality gaps in [0, 1]. Such bounds serve as a useful measure
of the robustness of a policy, but are often excessively conservative. This chapter
is devoted to understanding instance-dependent lower bounds, which try to
capture the optimal performance of a policy on a specific bandit instance.
Because the regret is a multi-objective criteria, an algorithm designer might
try and design algorithms that perform well on one kind of instance or another.
An extreme example is the policy that chooses At = 1 for all t, which suffers
zero regret when the first arm is optimal and linear regret otherwise. This is a
harsh trade-off, with the price for reducing the regret from logarithmic to zero
on just a few instances being linear regret on the remainder. Surprisingly, this
is the nature of the game in bandits. One can assign a measure of difficulty to
each instance such that policies performing overly well relative to this measure
on some instances pay a steep price on others. The situation is illustrated in
Fig. 16.1.

n over-specialised

reasonable, not instance optimal

instance optimal
Regret

√
n minimax optimal

Instances

Figure 16.1 On the x-axis, the instances are ordered according to the measure of difficulty,
and the y-axis shows the regret (on some scale). In the previous chapter, we proved
that no policy can be entirely below the horizontal ‘minimax optimal’ line. The results
in this chapter show that if the regret of a policy is below the ‘instance optimal’ line at
any point, then it must have regret above the shaded region for other instances. For
example, the ‘overly specified’ policy.
 16.1 Asymptotic Bounds 206

In finite time, the situation is a little messy, but if one pushes these ideas to
the limit, then for many classes of bandits one can define a precise notion of
instance-dependent optimality.

16.1 Asymptotic Bounds

We need to define exactly what is meant by a reasonable policy. If one is only

concerned with asymptotics, then a rather conservative definition suffices.

Definition 16.1. A policy π is called consistent over a class of bandits E if for

all ν ∈ E and p > 0, it holds that
Rn (π, ν)
lim = 0. (16.1)
n→∞ np
The class of consistent policies over E is denoted by Πcons (E).

Theorem 7.1 shows that UCB is consistent over ESG k

(1). The strategy that
always chooses the first action is not consistent on any class E unless the first
arm is optimal for every ν ∈ E.

Consistency is an asymptotic notion. A policy could be consistent and yet

play At = 1 for all t ≤ 10100 . For this reason, an assumption of consistency
is insufficient to derive non-asymptotic lower bounds. In Section 16.2, we
introduce a finite-time version of consistency that allows us to prove finite-
time instance-dependent lower bounds.

Recall that a class E of stochastic bandits is unstructured if E = M1 × · · · × Mk

with M1 , . . . , Mk sets of distributions. The main theorem of this chapter is a
generic lower bound that applies to any unstructured class of stochastic bandits.
After the proof, we will see some applications to specific classes. Let M be a set
of distributions with finite means, and let µ : M → R be the function that maps
P ∈ M to its mean. Let µ∗ ∈ R and P ∈ M have µ(P ) < µ∗ and define

dinf (P, µ∗ , M) = inf

0
{D(P, P 0 ) : µ(P 0 ) > µ∗ } .
P ∈M

Theorem 16.2. Let E = M1 × · · · × Mk and π ∈ Πcons (E) be a consistent policy

over E. Then, for all ν = (Pi )ki=1 ∈ E, it holds that
Rn X ∆i
lim inf ≥ c∗ (ν, E) = , (16.2)
n→∞ log(n) dinf (Pi , µ∗ , Mi )
i:∆i >0

where ∆i is the suboptimality gap of the ith arm in ν and µ∗ is the mean of the
optimal arm.

Proof Let µi be the mean of the ith arm in ν and di = dinf (Pi , µ∗ , Mi ). The
 16.1 Asymptotic Bounds 207

result will follow from Lemma 4.5, and by showing that for any suboptimal arm
i it holds that
Eνπ [Ti (n)] 1
lim inf ≥ .
n→∞ log(n) di
Fix a suboptimal arm i, and let ε > 0 be arbitrary and ν 0 = (Pj0 )kj=1 ∈ E be a
bandit with Pj0 = Pj for j 6= i and Pi0 ∈ Mi be such that D(Pi , Pi0 ) ≤ di + ε and
µ(Pi0 ) > µ∗ , which exists by the definition of di . Let µ0 ∈ Rk be the vector of means
of distributions of ν 0 . By Lemma 15.1, we have D(Pνπ , Pν 0 π ) ≤ Eνπ [Ti (n)](di + ε),
and by Theorem 14.2, for any event A,
1 1
Pνπ (A) + Pν 0 π (Ac ) ≥ exp (− D(Pνπ , Pν 0 π )) ≥ exp (−Eνπ [Ti (n)](di + ε)) .
2 2
Now choose A = {Ti (n) > n/2}, and let Rn = Rn (π, ν) and Rn0 = Rn (π, ν 0 ).
Then,
n
Rn + Rn0 ≥ (Pνπ (A)∆i + Pν 0 π (Ac )(µ0i − µ∗ ))
2
n
≥ min {∆i , µ0i − µ∗ } (Pνπ (A) + Pν 0 π (Ac ))
2
n
≥ min {∆i , µ0i − µ∗ } exp (−Eνπ [Ti (n)](di + ε)) .
4
Rearranging and taking the limit inferior leads to
 
n min{∆i ,µ0i −µ∗ }
log
Eνπ [Ti (n)] 1 4(Rn +Rn 0 )

lim inf ≥ lim inf

n→∞ log(n) di + ε n→∞ log(n)
 
1 log (Rn + Rn0 ) 1
= 1 − lim sup = ,
di + ε n→∞ log(n) di + ε
where the last equality follows from the definition of consistency, which says
that for any p > 0, there exists a constant Cp such that for sufficiently large n,
Rn + Rn0 ≤ Cp np , which implies that
log (Rn + Rn0 ) p log(n) + log(Cp )
lim sup ≤ lim sup = p,
n→∞ log(n) n→∞ log(n)
which gives the result since p > 0 was arbitrary and by taking the limit as ε
tends to zero.
Table 16.1 provides explicit formulas for dinf (P, µ∗ , M) for common choices of
M. The calculation of these quantities are all straightforward (Exercise 16.1).
The lower bound and definition of c∗ (ν, E) are quite fundamental quantities in
the sense that for most classes E, there exists a policy π for which
Rn (π, ν)
lim = c∗ (ν, E) for all ν ∈ E . (16.3)
n→∞ log(n)
This justifies calling a policy asymptotically optimal on class E if Eq. (16.3)
holds. For example, UCB from Chapter 8 and KL-UCB from Chapter 10 are
asymptotically optimal for EN k
(1) and EBk , respectively.
 16.2 Finite-Time Bounds 208

M P dinf (P, µ∗ , M)
(µ − µ∗ )2
{N (µ, σ 2 ) : µ ∈ R} N (µ, σ 2 )
2σ 2
 
1 (µ − µ∗ )2
{N (µ, σ 2 ) : µ ∈ R, σ 2 ∈ (0, ∞)} N (µ, σ 2 ) log 1+
2 σ2
   
µ 1−µ
{B(µ) : µ ∈ [0, 1]} B(µ) µ log + (1 − µ) log
µ∗ 1 − µ∗
 
2((a + b)/2 − µ∗ )2
{U(a, b) : a, b ∈ R} U(a, b) log 1+
b−a

Table 16.1 Expressions for dinf for different parametric families when the mean of P is less
than µ∗ .

16.2 Finite-Time Bounds

By making a finite-time analogue of consistency, it is possible to prove a finite-

time instance-dependent bound. First, a lemma that summarises what can be
obtained by chaining the Bretagnolle–Huber inequality (Theorem 14.2) with the
divergence decomposition lemma (Lemma 15.1).

Lemma 16.3. Let ν = (Pi ) and ν 0 = (Pi0 ) be k-armed stochastic bandits that
differ only in the distribution of the reward for action i ∈ [k]. Assume that i is
suboptimal in ν and uniquely optimal in ν 0 . Let λ = µi (ν 0 ) − µi (ν). Then, for
any policy π,
 
min{λ−∆i (ν),∆i (ν)}
log 4 + log(n) − log(Rn (ν) + Rn (ν 0 ))
Eνπ [Ti (n)] ≥ . (16.4)
D(Pi , Pi0 )

The lemma holds for finite n and any ν and can be used to derive finite-
time instance-dependent lower bounds for any environment class E that is rich
enough. The following result provides a finite-time instance-dependence bound
for Gaussian bandits where the asymptotic notion of consistency is replaced by
an assumption that the minimax regret is not too large. This assumption alone
is enough to show that no policy that is remotely close to minimax optimal can
be much better than UCB on any instance.

Theorem 16.4. Let ν ∈ EN k

be a k-armed Gaussian bandit with mean vector
µ ∈ Rk and suboptimality gaps ∆ ∈ [0, ∞)k . Let

E(ν) = {ν 0 ∈ EN
k
: µi (ν 0 ) ∈ [µi , µi + 2∆i ]} .

Suppose C > 0 and p ∈ (0, 1) are constants and π is a policy such that
 16.3 Notes 209

Rn (π, ν 0 ) ≤ Cnp for all n and ν 0 ∈ E(ν). Then, for any ε ∈ (0, 1],

!+
2 X (1 − p) log(n) + log ε∆i
8C
Rn (π, ν) ≥ . (16.5)
(1 + ε)2 ∆i
i:∆i >0

Proof Let i be suboptimal in ν, and choose ν 0 ∈ E(ν) such that µj (ν 0 ) = µj (ν)

for j 6= i and µj (ν 0 ) = µi + ∆i (1 + ε). Then, by Lemma 16.3 with λ = ∆i (1 + ε),
   
2 n  min {λ − ∆i , ∆i }
Eνπ [Ti (n)] ≥ log + log
∆2i (1 + ε)2 2Cnp 4
  
2 ε∆i
= 2 (1 − p) log (n) + log .
∆i (1 + ε)2 8C

Plugging this into the basic regret decomposition identity (Lemma 4.5) gives the
result.

When p = 1/2, the leading term in this lower bound is approximately half that
of the asymptotic bound. This effect may be real. The class of policies considered
is larger than in the asymptotic lower bound, and so there is the possibility that
the policy that is best tuned for a given environment achieves a smaller regret.

16.3 Notes

1 We mentioned that for most classes E there is a policy satisfying Eq. (16.3).
Its form is derived from the lower bound, and by making some additional
assumptions on the underlying distributions. For details, see the article
by Burnetas and Katehakis [1996], which is also the original source of
Theorem 16.2.
2 The analysis in this chapter only works for unstructured classes. Without this
assumption a policy can potentially learn about the reward from one arm
by playing other arms and this greatly reduces the regret. Lower bounds for
structured bandits are more delicate and will be covered on a case-by-case
basis in subsequent chapters.
3 The classes analysed in Table 16.1 are all parametric, which makes the
calculation possible analytically. There has been relatively little analysis
in the non-parametric case, but we know of three exceptions for which we
simply refer the reader to the appropriate source. The first is the class of
distributions with bounded support: M = {P : Supp(P ) ⊆ [0, 1]}, which has
been analysed exactly [Honda and Takemura, 2010]. The second is the class
of distributions with semi-bounded support, M = {P : Supp(P ) ⊆ (−∞, 1]}
[Honda and Takemura, 2015]. The third is the class of distributions with
bounded kurtosis, M = {P : KurtX∼P [X] ≤ κ} [Lattimore, 2017].
 16.4 Bibliographic Remarks 210

16.4 Bibliographic Remarks

Asymptotic optimality via a consistency assumption first appeared in the seminal

paper by Lai and Robbins [1985], which was later generalised by Burnetas and
Katehakis [1996]. In terms of upper bounds, there now exist policies that are
asymptotic optimal for single-parameter exponential families [Cappé et al., 2013].
Until recently, there were no results on asymptotic optimality for multi-parameter
classes of reward distributions. There has been some progress on this issue recently
for the Gaussian distribution with unknown mean and variance [Cowan et al.,
2018] and for the uniform distribution [Cowan and Katehakis, 2015]. There
are plenty of open questions related to asymptotically optimal strategies for
non-parametric classes of reward distributions. When the reward distributions
are discrete and finitely supported, an asymptotically optimal policy is given by
Burnetas and Katehakis [1996], though the precise constant is hard to interpret. A
relatively complete solution is available for classes with bounded support [Honda
and Takemura, 2010]. Already for the semi-bounded case, things are getting
murky [Honda and Takemura, 2015]. One of the authors thinks that classes with
bounded kurtosis are quite interesting, but here things are only understood up
to constant factors [Lattimore, 2017]. An asymptotic variant of Theorem 16.4 is
by Salomon et al. [2013]. Finite-time instance-dependent lower bounds have been
proposed by several authors, including Kulkarni and Lugosi [2000], for two arms,
and Garivier et al. [2019] and Lattimore [2018], for the general case. As noted
earlier, neither ETC policies, nor elimination-based algorithms are able to achieve
asymptotic optimality: as shown by Garivier et al. [2016b], these algorithms (no
matter how they are tuned) must incur an additional multiplicative penalty of
a factor of two on the standard Gaussian bandit problems as compared to the
optimal asymptotic regret.

16.5 Exercises

16.1 (Relative entropy calculations) Verify the calculations in Table 16.1.

16.2 (Rademacher noise) Let R(µ) be the shifted Rademacher distribution,

which for µ ∈ R and X ∼ R(µ) is characterised by P (X = µ + 1) =
P (X = µ − 1) = 1/2.
(a) Show that dinf (R(µ), µ∗ , M) = ∞ for any µ < µ∗ .
(b) Design a policy π for bandits with shifted Rademacher rewards such that
the regret is bounded by
k
X
Rn (π, ν) ≤ 3 ∆i for all n and ν ∈ Mk .
i=1

(c) The results from parts (a) and (b) seem to contradict the heuristic analysis
in Note 1 at the end of Chapter 15. Explain.
 16.5 Exercises 211

16.3 (Asymptotic lower bound for exponential families) Let M = {Pθ :

θ ∈ Θ} be an exponential family with sufficient statistic equal to the identity and
E = Mk and π be a consistent policy for E. Prove that the asymptotic upper
bound on the regret proven in Exercise 10.4 is tight.

16.4 (Unknown subgaussian constant) Let


M = P : there exists a σ 2 ≥ 0 such that P is σ 2 -subgaussian .
(a) Find a distribution P such that P ∈ / M.
(b) Suppose that P ∈ M has mean µ ∈ R. Prove that dinf (P, µ∗ , M) = 0 for all
µ∗ > µ.
(c) Let E = {(Pi ) : Pi ∈ M for all 1 ≤ i ≤ k}. Prove that if k > 1, then for all
consistent policies π,
Rn (π, ν)
lim inf =∞ for all ν ∈ E .
n→∞ log(n)
(d) Let f : N → [0, ∞) be any increasing function with limn→∞ f (n)/ log(n) =
∞. Prove there exists a policy π such that
Rn (π, ν)
lim sup =0 for all ν ∈ E ,
n→∞ f (n)
where E is as in the previous part.
(e) Conclude that there exists a consistent policy for E.

16.5 (Minimax lower bound) Use Lemma 16.3 to prove Theorem 15.2, possibly
with different constants.
2
16.6 (Refining the lower-order terms) Let k = 2, and for ν ∈ EN let
2
∆(ν) = max{∆1 (ν), ∆2 (ν)}. Suppose that π is a policy such that for all ν ∈ EN
with ∆(ν) ≤ 1, it holds that
C log(n)
Rn (π, ν) ≤ . (16.6)
∆(ν)
(a) Give an example of a policy satisfying Eq. (16.6).
(b) Assume that i = 2 is suboptimal for ν and that α ∈ (0, 1) be such that
1
Eνπ [T2 (n)] = 2∆(ν) 2 log(α). Let ν
0
be the alternative environment where
µ1 (ν 0 ) = µ1 (ν) and µ2 (ν 0 ) = µ1 (ν) + 2∆(ν). Show that
1
exp(− D(Pνπ , Pν 0 π )) =
.
α
(c) Let A be the event that T2 (n) ≥ n/2. Show that
2C log(n) 1 2C log(n)
Pνπ (A) ≤ and Pν 0 π (A) ≥ − .
n∆(ν)2 2α n∆(ν)2
(d) Show that
 
n∆(ν) 1 2C log(n)
Rn (π, ν ) ≥
0
− .
2 2α n∆(ν)2
 16.5 Exercises 212

n∆(ν)2
(e) Show that α ≥ 8C log(n) and conclude that
 
1 n∆(ν)2
Rn (π, ν) ≥ log .
2∆(ν) 8C log(n)

(f) Generalise the argument to an arbitrary number of arms.

In Exercise 7.6 you showed that there exists a bandit policy π such that for
some universal constant C > 0 and for any ν ∈ E[0,b] k
k-armed bandit with
rewards taking values in [0, b], the regret Rn (π, ν) of π on ν after n rounds
satisfies
X  
σi2
 
Rn (π, ν) ≤ C ∆i + b + log(n) ,
∆i
i:∆i >0

where ∆i = ∆i (ν) is the action gap of action i and σi2 = σi2 (ν) is the
variance of the reward of arm i. In particular, this is the inequality shown in
Eq. (7.14). The next exercise asks you to show that the appearance of both
σ2
b and ∆ii is necessary in this bound.

16.7 (Sharpness of Eq. (7.14)) Let k > 1, b > 0 and c > 0 be arbitrary. Show
that there is no policy π for which either
Rn (π, ν)
lim sup ≤ cb, k
∀ν ∈ E[0,b] (16.7)
n→∞ log(n)
or
Rn (π, ν) X σ 2 (ν)
lim sup ≤c i
, k
∀ν ∈ E[0,b] (16.8)
n→∞ log(n) ∆i (ν)
i:∆i >0

would hold true.

The intuition underlying this result is the following: Eq. (16.7) cannot hold
because this would mean that for some policy, the regret is logarithmic
with a constant independent of the gaps, while intuitively, if the variance
is constant, the coefficient of the logarithmic regret must increase as the
gaps get close. Similarly, Eq. (16.8) cannot hold either because we expect a
logarithmic regret with a coefficient proportional to the inverse gap even as
the variance gets zero, as the case of Bernoulli bandits shows. This exercise
is due to Audibert et al. [2007].

16.8 (Lower bound on regret variance) Let k > 1 and E ⊂ EN k

be the set
of k-armed Gaussian bandits with mean rewards in [0, 1] for all arms. Suppose
 16.5 Exercises 213

that π is a policy such that for all ν ∈ E,

Rn (π, ν) X 2(1 + p)
lim sup ≤ .
n→∞ log(n) ∆i
i:∆i >0

Prove that
log(V[R̂n (π, ν)])
lim sup sup ≥ 1,
n→∞ ν∈E (1 − p) log(n)
Pn
where R̂n (π, ν) = nµ∗ (ν) − t=1 µAt (ν).
17 High-Probability Lower Bounds

The lower bounds proven in the last two chapters were for stochastic bandits.
In this chapter, we prove high probability lower bounds for both stochastic and
adversarial bandits. Recall that for adversarial bandit x ∈ [0, 1]n×k , the random
regret is
n
X
R̂n = max xti − xtAt
i∈[k]
t=1

and the (expected) regret is Rn = E[R̂n ]. To set expectations, remember that in

Chapter 12 we proved two high-probability upper bounds on the regret of Exp3-
IX. In the first, we showed there exists a policy π such that for all adversarial
bandits x ∈ [0, 1]n×k and δ ∈ (0, 1), it holds with probability at least 1 − δ that
s  !
p kn 1
R̂n = O kn log(k) + log . (17.1)
log(k) δ

We also gave a version of the algorithm that depended on δ ∈ (0, 1) for which
with probability at least 1 − δ,
s  !
k
R̂n = O kn log . (17.2)
δ

The important difference is the order of quantifiers. In the first, we have a

single algorithm and a high-probability guarantee that holds simultaneously for
any confidence level. The second algorithm needs the confidence level to be
specified
p in advance. The price for using the generic algorithm appears to be
log(1/δ)/ log(k), which is usually quite small but not totally insignificant. We
will see that both bounds are tight up to constant factors, which implies that
knowing the desired confidence level in advance really does help. One reason
why choosing the confidence level in advance is not ideal is that the resulting
high-probability bound cannot be integrated to prove a bound in expectation.
For algorithms satisfying (17.1), the expected regret can be bounded by
Z ∞   p
Rn ≤ P R̂n ≥ u du = O( kn log(k)) . (17.3)
0
 17.1 Stochastic Bandits 215

On the other hand, if the high-probability bound only holds for a single δ, as in
(17.2), then it seems hard to do much better than
s  !
k
Rn ≤ nδ + O kn log ,
δ
p
which with the best choice of δ leads to a bound of O( kn log(n)).

17.1 Stochastic Bandits

For simplicity, we start with the stochastic setting before explaining how to
convert the arguments to the adversarial model. There is no randomness in the
expected regret, so in order to derive a high-probability bound, we define the
random pseudo-regret by
k
X
R̄n = Ti (n)∆i ,
i=1

which is a random variable through the pull counts Ti (n).

For all results in this section, we let E k ⊂ EN

k
denote the set of k-armed
Gaussian bandits with suboptimality gaps bounded by one. For µ ∈ [0, 1]d
we let νµ ∈ E k be the Gaussian bandit with means µ.

Theorem 17.1. Let n ≥ 1 and k ≥ 2 and B > 0 and π be a policy such that for
any ν ∈ E k ,
p
Rn (π, ν) ≤ B (k − 1)n . (17.4)

Let δ ∈ (0, 1). Then there exists a bandit ν in E k such that

   
1 1p 1
P R̄n (π, ν) ≥ min n, (k − 1)n log ≥ δ.
4 B 4δ
Proof Let ∆ ∈ (0, 1/2] be a constant to be tuned subsequently and ν = νµ where
the mean vector µ ∈ Rd is defined by µ1 = ∆ and µi = 0 for i > 1. Abbreviate
Rn = Rn (π, ν) and P = Pνπ and E = Eνπ . Let i = argmini>1 E[Ti (n)]. Then, by
Lemma 4.5 and the assumption in Eq. (17.4),
r
Rn B n
E[Ti (n)] ≤ ≤ . (17.5)
∆(k − 1) ∆ k−1

Define alternative bandit ν 0 = νµ0 where µ0 ∈ Rd is equal to µ except µ0i = µi +2∆.

Abbreviate P0 = Pν 0 π and R̄n = R̄n (π, ν) and R̄n0 = R̄n (π, ν 0 ). By Lemma 4.5, the
 17.1 Stochastic Bandits 216

Bretagnolle–Huber inequality (Theorem 14.2) and the divergence decomposition

(Lemma 15.1), we have
    
∆n ∆n n  n
P R̄n ≥ + P R̄n ≥
0 0
≥ P Ti (n) ≥ + P0 Ti (n) <
2 2 2 2
 r 
1 1 n
≥ exp (− D(P, P0 )) ≥ exp −2B∆ ≥ 2δ ,
2 2 k−1
where the last line follows by choosing
( r  )
1 1 k−1 1
∆ = min , log .
2 2B n 4δ

The result follows since max{a, b} ≥ (a + b)/2.

Corollary 17.2. Let n ≥ 1 and k ≥ 2. Then, for any policy π and δ ∈ (0, 1)
such that
s  
1
nδ ≤ n(k − 1) log , (17.6)
4δ

there exists a bandit problem ν ∈ E k such that

( s  )!
1 n(k − 1) 1
P R̄n (π, ν) ≥ min n, log ≥ δ. (17.7)
4 2 4δ

Proof We prove the result by contradiction. Assume that the conclusion does
not hold for π and let δ ∈ (0, 1) satisfy (17.6). Then, for any bandit problem
ν ∈ E k , the expected regret of π is bounded by
s   s  
n(k − 1) 1 1
Rn (π, ν) ≤ nδ + log ≤ 2n(k − 1) log .
2 4δ 4δ
p
Therefore, π satisfies the conditions of Theorem 17.1 with B = 2 log(1/(4δ)),
which implies that there exists some bandit problem ν ∈ E k such that (17.7)
holds, contradicting our assumption.

Corollary 17.3. Let k ≥ 2 and p ∈ (0, 1) and B > 0. Then, there does not
exist a policy π such that for all n ≥ 1, δ ∈ (0, 1) and ν ∈ E k ,
  
p 1
P R̄n (π, ν) ≥ B (k − 1)n logp < δ.
δ
Proof We proceed by contradiction. Suppose that such a policy exists. Choosing
δ sufficiently small and n sufficiently large ensures that
     
1 1 1 1p 1
log ≥ B log p
and n(k − 1) log ≤ n.
B 4δ δ B 4δ
 17.2 Adversarial Bandits 217

Now, by assumption, for any ν ∈ E k we have

Z ∞


Rn (π, ν) ≤ P R̄n (π, ν) ≥ x dx
0
p Z ∞   p
≤ B n(k − 1) exp −x1/p dx ≤ B n(k − 1) .
0

Therefore, by the Theorem 17.1, there exists a bandit ν ∈ E k such that

  
p 1
P R̄n (π, ν) ≥ B n(k − 1) log
δ
   
1 1 p 1
≥ P R̄n (π, ν) ≥ min n, n(k − 1) log ≥ δ,
4 B 4δ
which contradicts our assumption and completes the proof.

We suspect there exists a policy π and universal constant B > 0 such that
for all ν ∈ E k ,
  
√ 1
P R̄n (π, ν) ≥ B kn log ≤ δ.
δ

17.2 Adversarial Bandits

We now explain how to translate the ideas in the previous section to the adversarial
model. Let π = (πt )nt=1 be a fixed policy, and recall that for x ∈ [0, 1]n×k , the
random regret is
n
X
R̂n = max (xti − xtAt ) .
i∈[k]
t=1

Let Fx be the cumulative distribution function of the law of R̂n when policy π
interacts with the adversarial bandit x ∈ [0, 1]n×k .

Theorem 17.4. Let c, C > 0 be sufficiently small/large universal constants and

k ≥ 2, n ≥ 1 and δ ∈ (0, 1) be such that n ≥ Ck log(1/(2δ)). Then there exists a
reward sequence x ∈ [0, 1]n×k such that
s  !
1
1 − Fx c nk log ≥ δ.
2δ

The proof is a bit messy, but is not completely without interest. For the sake of
brevity, we explain only the high-level ideas and refer you elsewhere for the gory
details. There are two difficulties in translating the arguments in the previous
section to the adversarial model. First, in the adversarial model, we need the
 17.2 Adversarial Bandits 218

rewards to be bounded in [0, 1]. The second difficulty is we now analyse the
adversarial regret rather than the random pseudo-regret. Given a measure Q, let
X ∈ [0, 1]n×k and (At )nt=1 be a collection of random variables on a probability
space (Ω, F, PQ ) such that
(a) PQ (X ∈ B) = Q(B) for all B ∈ B([0, 1]n×k ); and
(b) PQ (At | A1 , X1 , . . . , At−1 , Xt−1 ) = πt (At | A1 , X1 , . . . , At−1 , Xt−1 ) almost
surely, where Xs = XtAs .
Then the regret is a random variable R̂n : Ω → R defined by
n
X
R̂n = max (Xti − XtAt ) .
i∈[k]
t=1

Suppose we sample X ∈ [0, 1]n×k from distribution Q on ([0, 1]n×k , B([0, 1]k )).
Claim 17.5. Suppose that X ∼ Q, where Q is a measure on [0, 1]n×k with the
Borel σ-algebra and that EQ [1 − FX (u)] ≥ δ. Then there exists an x ∈ [0, 1]n×k
such that 1 − Fx (u) ≥ δ.
The next step is to choose Q and argue that EQ [1 − FX (u)] ≥ δ for sufficiently
large u. To do this, we need a truncated normal distribution. Defining clipping
function

1 if x > 1


clip[0,1] (x) = 0 if x < 0


x otherwise .

Let σ and ∆ be positive constants to be chosen later and (ηt )nt=1 a sequence of
independent random variables with ηt ∼ N (1/2, σ 2 ). For each i ∈ [k], let Qi be
the distribution of X ∈ [0, 1]n×k , where

clip[0,1] (ηt + ∆) if j = 1


Xtj = clip[0,1] (ηt + 2∆) if j = i and i 6= 1


clip (η ) otherwise .
[0,1] t

Notice that under any Qi for fixed t, the random variables Xt1 , . . . , Xtk are not
independent, but for fixed j, the random variables X1j , . . . , Xnj are independent
and identically distributed. Let PQi be the law of X1 , A1 , . . . , An , Xn when policy
π interacts with adversarial bandit sampled from X ∼ Qi .
q

1
Claim 17.6. If σ > 0 and ∆ = σ k−1 2n log 8δ , then there exists an arm i such
that
PQi (Ti (n) < n/2) ≥ 2δ .
The proof of this claim follows along the same lines as the theorems in the
previous section. All that changes is the calculation of the relative entropy. The
last step is to relate Ti (n) to the random regret. In the stochastic model, this was
 17.3 Notes 219

straightforward, but for adversarial bandits there is an additional step. Notice

that under Qi , it holds that Xti − XtAt ≥ 0 and that if Xti , XtAt ∈ (0, 1) and
At 6= i, then Xti − XtAt ≥ ∆. From this we conclude that
n
!
X
R̂n ≥ ∆ n − Ti (n) − I {exists j ∈ [k] : Xtj ∈ {0, 1}} . (17.8)
t=1

The following claim upper-bounds the number of rounds in which clipping occurs
with high probability.

Claim 17.7. If σ = 1/10 and ∆ < 1/8 and n ≥ 32 log(1/δ), then

n
!
X n
PQi I {exists j ∈ [k] : Xtj ∈ {0, 1}} ≥ ≤ δ.
t=1
4

Combining Claim 17.6 and Claim 17.7 with Eq. (17.8) shows there exists an
arm i such that
 
n∆
PQi R̂n ≥ ≥ δ,
4
which by the definition of ∆ and Claim 17.5 implies Theorem 17.4.

17.3 Notes

1 The adversarial bandits used in Section 17.2 had the interesting property that
the same arm has the best reward in every round (not just the best mean).
This cannot be exploited by an algorithm, however, because it only gets a
single observation in each round.
2 In Theorem 17.4, we did not make any assumptions on the algorithm. If √ we
had assumed the algorithm enjoyed an expected regret bound of Rn ≤ B kn,
then we could conclude that for each sufficiently small δ ∈ (0, 1) there exists
an adversarial bandit such that
  
c√ 1
P R̂n ≥ kn log ≥ δ,
B 2δ
which shows that our high-probability upper bounds for Exp3-IX are nearly
tight.

17.4 Bibliographic Remarks

The results in this chapter are by Gerchinovitz and Lattimore [2016], who also
provide lower bounds on what is achievable when the loss matrix exhibits nice
structure such as low variance or similarity between losses of the arms.
 17.5 Exercises 220

17.5 Exercises

17.1 Prove each of the claims in Section 17.2.

Part V
Contextual and Linear
Bandits
 222

The algorithms introduced so far work well in stationary environments with

only a few actions. Real-world problems are seldom this simple. For example,
a bandit algorithm designed for targeted advertising may have thousands of
actions. Even more troubling, the algorithm has access to contextual information
about the user and the advertisement. Ignoring this information would make the
problem highly non-stationary, but algorithms introduced in the previous chapter
cannot make use of side information.
Large action sets are usually dealt with by introducing structure that allows the
algorithm to generalise from one action to another. For example, advertisements
can usually be associated with features describing their topic. Then the reward
can be described as a function of the features, usually assumed to be nice in some
way (linear or smooth, for example). Contextual information is dealt with in a
similar fashion by assuming the mean reward of an action is a function of the
context features and action features. As we explain, this leads to a model where
the action set is essentially changing in each round.
Of course the world is messy in other ways. Rewards are often delayed, and
may be unattributed, or the world may be non-stationary. The first of these
issues is discussed briefly in the introduction to Part VII while non-stationarity
is the subjection of Chapter 31.
Except for the first chapter, which is generic, the focus of this part will be on
the special case that the expected reward of each arm is a linear function of some
feature vector in a way that will be made precise in Chapter 19. Along the way,
we will discuss many generalisations and give references to the literature. One
aspect that will play a far larger role is computation. While finite-armed bandits
with few arms present no computation difficulties, when the number of actions
is very large or the information structure of the feedback model is not so easily
separable, then computation can be a serious challenge.
18 Contextual Bandits

In many bandit problems, the learner has access to additional information

that may help predict the quality of the actions. Imagine designing a movie
recommendation system where users sequentially request recommendations for
which movie to watch next. It would be inadvisable to ignore demographic
information about the user making the request, or other contextual history such
as previously watched movies or ratings. None of the algorithms presented so
far make use of this kind of additional information. Indeed, they optimise a
benchmark (the regret) that also disregards such contextual data. Essentially
they would try to identify the best single movie in hindsight. In this chapter,
we present an augmented framework and regret definition that better models
real-world problems where contextual information is available.

Whenever you design a new benchmark, there are several factors to consider.
Competing with a poor benchmark does not make sense, since even an
algorithm that perfectly matches the benchmark will perform poorly. At
the same time, competing with a better benchmark can be harder from a
learning perspective, and this penalty must be offset against the benefits.

The trade-off just described is fundamental to all machine learning problems.

In statistical estimation, the analoguous trade-off is known as the bias-variance
trade-off. We will not attempt to answer the question of how to resolve this trade-
off in this chapter because first we need to see how to effectively compete with
improved benchmarks. The good news is that many of the techniques developed
earlier are easily generalised.

18.1 Contextual Bandits: One Bandit per Context

While contextual bandits can be studied in both the adversarial and stochastic
frameworks, in this chapter we focus on the k-armed adversarial model. As usual,
the adversary secretly chooses (xt )nt=1 , where xt ∈ [0, 1]k with xti the reward
associated with arm i in round t. The adversary also secretly chooses a sequence
of contexts (ct )nt=1 , where ct ∈ C with C a set of possible contexts. In each round,
 18.1 Contextual Bandits: One Bandit per Context 224

Adversary secretly chooses rewards (xt )n

t=1 with xt ∈ [0, 1]
k

Adversary secretly chooses contexts (ct )n

t=1 with ct ∈ C

For rounds t = 1, 2, . . . , n:
Learner observes context ct ∈ C where C is an arbitrary fixed set of contexts.
Learner selects distribution Pt ∈ Pk−1 and samples At from Pt .
Learner observes reward Xt = xtAt .

Figure 18.1 Interaction protocol for k-armed contextual bandits.

the learner observes ct , chooses an action At and receives reward xtAt . The
interaction protocol is shown in Fig. 18.1.
A natural way to define the regret is to compare the rewards collected by
the learner with the rewards collected by the best context-dependent policy in
hindsight:
 
X X
Rn = E  max (xti − Xt ) . (18.1)
i∈[k]
c∈C t∈[n]:ct =c

If the set of possible contexts is finite, then a simple approach is to use a separate
instance of Exp3 for each context. Let
 
X
Rnc = E max (xti − Xt )
i∈[k]
t∈[n]:ct =c

be the regret due to context c ∈ C. When using a separate instance of Exp3 for
each context, we can use the results of Chapter 11 to bound
v
u n
u X
Rnc ≤ 2tk I {ct = c} log(k) , (18.2)
t=1

where the sum inside the square root counts the number of times context c ∈ C is
observed. Because this is not known in advance, it is important to use an anytime
version of Exp3 for which the above regret bound holds without needing to tune
a learning rate that depends on the number of times the context is observed (see
Exercise 28.13). Substituting (18.2) into the regret leads to
v
u n
X Xu X
Rn = Rnc ≤ 2 tk log(k) I {ct = c} . (18.3)
c∈C c∈C t=1

The magnitude of the right-hand side depends on the distribution of observed

contexts. On one extreme, there is only one observed context, and the bound is
the same as the standard finite-armed bandit problem. The other extreme occurs
 18.2 Bandits with Expert Advice 225

when all contexts are observed equally often, in which case we have
p
Rn ≤ 2 nk|C| log(k) . (18.4)
Jensen’s inequality applied to Eq. (18.3) shows that this really is the worst case
(Exercise 18.1).

The regret in Eq. (18.4) is different than the regret studied in Chapter 11. If
we ignore the context and run the standard Exp3 algorithm, then we would
have
" n # n
X X p
E Xt ≥ max xti − 2 kn log(k) .
i∈[k]
t=1 t=1

Using one version of Exp3 per context leads to

" n #
X X X p
E Xt ≥ max xti − 2 kn|C| log(k) .
i∈[k]
t=1 c∈C t∈[n]:ct =c

Which of these bounds is preferable depends on the magnitude of n and

how useful the context is. When n is very large, the second bound is more
likely to be preferable. On the other hand, the second bound is completely
vacuous when n ≤ 4k|C| log(k).

18.2 Bandits with Expert Advice

When the context set C is large, using one bandit algorithm per context will
almost always be a poor choice because the additional precision is wasted unless
the amount of data is enormous. Fortunately, however, it is seldom the case that
the context set is both large and unstructured. To illustrate a common situation,
we return to the movie recommendation theme, where the actions are movies
and the context contains user information such as age, gender and recent movie
preferences. In this case, the context space is combinatorially large, but there
is a lot of structure inherited from the fact that the space of movies is highly
structured and users with similar demographics are more likely to have similar
preferences.
We start by rewriting Eq. (18.1) in an equivalent form. Let Φ be the set of all
functions from C → [k]. Then,
" n
#
X
Rn = E max (xtφ(ct ) − Xt ) . (18.5)
φ∈Φ
t=1

The discussion above suggests that a slightly smaller set Φ may lead to more
reward. In what follows, we describe some of the most common ideas of how to
do this.
 18.2 Bandits with Expert Advice 226

Figure 18.2 Prediction with expert advice. The experts, upon seeing a foot give expert
advice on what socks should fit it best. If the owner of the foot is happy, the
recommendation system earns a cookie!

Partitions
Let P ⊂ 2C be a partition of C, which means that sets (or parts) in P are disjoint
and ∪P ∈P P = C. Then define Φ to be the set of functions from C to [k] that are
constant on each part in P. In this case, we can run a version of Exp3 for each
part, which means the regret depends on the number of parts |P| rather than on
the number of contexts.

Similarity Functions
Let s : C × C → [0, 1] be a function measuring the similarity between pairs of
contexts on the [0, 1]-scale. Then let Φ be the set of functions φ : C → [k] such
that the average dissimilarity
1 X
(1 − s(c, d))I {φ(c) 6= φ(d)}
|C|2
c,d∈C

is below a user-tuned threshold θ ∈ (0, 1). It is not clear anymore that we can
control the regret (18.5) using some simple meta-algorithm on Exp3, but keeping
the regret small is still a meaningful objective.

From Supervised Learning to Bandits with Expert Advice

Yet another option is to run your favorite supervised learning method, training
on batch data to find a collection of predictors φ1 , . . . , φM : C → [k]. Then we
could use a bandit algorithm to compete with the best of these in an online
fashion. This has the advantage that the offline training procedure can use the
power of batch data and the whole army of supervised learning, without relying
on potentially inaccurate evaluation methods that aim to pick the best of the
pack. And why pick if one does not need to?
The possibilities are endless, but ultimately we always end up with a set of
functions Φ with the goal of competing with the best of them. This suggests we
 18.2 Bandits with Expert Advice 227

Adversary secretly chooses rewards x ∈ [0, 1]n×k

Experts secretly choose predictions E (1) , . . . , E (n)
For rounds t = 1, 2, . . . , n:
Learner observes predictions of all experts, E (t) ∈ [0, 1]M ×k .
Learner selects a distribution Pt ∈ Pk−1 .
Action At is sampled from Pt and the reward is Xt = xtAt .

Figure 18.3 Interaction protocol for bandits with expert advice.

should think more generally about some subset Φ of functions without considering
the internal structure of Φ. In fact, once Φ has been chosen, the contexts play
very little role. All we need in each round is the output of each function.

18.2.1 Bandits with Expert Advice Framework

The bandits with expert advice setting is a k-armed adversarial bandit, but
with M experts making recommendations to the learner. At the beginning of each
round, the experts announce their predictions about which actions are the most
promising. For the sake of generality, the experts report a probability distribution
over the actions. The interpretation is that the expert, if the decision were left
to them, would choose the action for the round at random from the probability
distribution it reported. As discussed before, in an adversarial setting it is natural
to consider randomised algorithms, hence one should not be too surprised that the
experts are also allowed to randomise. An application to an important practical
problem is illustrated in Fig. 18.2.
The predictions of the M experts in round t are represented by a matrix
(t)
E (t) ∈ [0, 1]M ×k , where the mth row Em is a probability vector over [m]
(t)
representing the recommendations of expert m in round t. Since Em is a row
(t)
vector, for a k-dimensional vector x, the expression Em xt is well defined. The
learner and the environment interact according to the protocol in Fig. 18.3.
The regret measures the cumulative rewards collected by the learner relative
to the best expert in hindsight:
" n n
#
X X
(t)
Rn = E max Em xt − Xt . (18.6)
m∈[M ]
t=1 t=1

This framework assumes the experts are oblivious in the sense that their
predictions do not depend on the actions of the learner.
 18.3 Exp4 228

18.3 Exp4

The number 4 in Exp4 is not just an increased version number, but indicates
the four e’s in the long name of the algorithm, which is exponential weighting
for exploration and exploitation with experts. The idea of the algorithm is
very simple. Since exponential weighting worked so well in the standard bandit
problem, we aim to adopt it to the problem at hand. However, since the goal is to
compete with the best expert in hindsight, it is not the actions that we will score,
but the experts. Exp4 thus maintains a probability distribution Qt over experts
and uses this to come up with the next action in the obvious way, by first choosing
an expert Mt at random from Qt and then following the chosen expert’s advice
(t)
to choose At ∼ EMt . The reader is invited to check for themself that this is the
same as sampling At from Pt = Qt E (t) where Qt is treated as a row vector. Once
the action is chosen, one can use their favorite reward estimation procedure to
estimate the rewards for all the actions, which is then used to estimate how much
total reward the individual experts would have made so far. The reward estimates
are then used to update Qt using exponential weighting. The pseudocode of Exp4
is given in Algorithm 11.

1: Input: n, k, M , η, γ
2: Set Q1 = (1/M, . . . , 1/M ) ∈ [0, 1]1×M (a row vector)
3: for t = 1, . . . , n do
4: Receive advice E (t)
5: Choose the action At ∼ Pt , where Pt = Qt E (t)
6: Receive the reward Xt = xtAt
t =i}
7: Estimate the action rewards: X̂ti = 1 − I{A Pti +γ (1 − Xt )
8: Propagate the rewards to the experts: X̃t = E (t) X̂t
9: Update the distribution Qt using exponential weighting:
exp(η X̃ti )Qti
Qt+1,i = P for all i ∈ [M ]
j exp(η X̃tj )Qtj

10: end for

Algorithm 11: Exp4.

The algorithm uses O(M ) memory and O(M + k) computation per round
(when sampling in two steps). Hence it is only practical when both M and k are
reasonably small.

18.4 Regret Analysis

We restrict our attention to the case when γ = 0, which is the original algorithm.
The version where γ > 0 is called Exp4-IX and its analysis is left for Exercise 18.3.
 18.4 Regret Analysis 229

p
Theorem 18.1. Let γ = 0 and η = 2 log(M )/(nk), and denote by Rn the
expected regret of Exp4 defined in Algorithm 11 after n rounds. Then,
p
Rn ≤ 2nk log(M ) . (18.7)
After translating the notation, the proof of the following lemma can be extracted
from the analysis of Exp3 in the proof of Theorem 11.2 (Exercise 18.2).
Lemma 18.2. For any m∗ ∈ [M ], it holds that

log(M ) η X X
n
X n X
X M n M
X̃ − Qtm X̃tm ≤ + Qtm (1 − X̂tm )2 .
2 t=1 m=1
tm∗
t=1 t=1 m=1
η

Proof of Theorem 18.1 Let Ft = σ(E (1) , A1 , E (2) , A2 , . . . , At−1 , E (t) ) and
abbreviate Et [·] = E[ · | Ft ]. Let m∗ be the index of the best-performing expert in
hindsight:
n
X
(t)
m = argmaxm∈[M ]
∗
Em xt , (18.8)
t=1

which is not random by the assumption that the experts are oblivious. Applying
Lemma 18.2 shows that
log(M ) η X X
n
X n X
X M M M
X̃tm∗ − Qtm X̃tm ≤ + Qtm (1 − X̃tm )2 . (18.9)
t=1 t=1 m=1
η 2 t=1 m=1

When γ = 0 the estimator X̂ti is unbiased so that Et [X̂t ] = xt and

Et [X̃t ] = Et [E (t) X̂t ] = E (t) Et [X̂t ] = E (t) xt . (18.10)
Taking the expectation of both sides of Eq. (18.9) and using the tower rule for
conditional expectation and the fact that Qt is Ft -measurable leads to

log(M ) η X X 
n M

Rn ≤ + E Qtm (1 − X̃tm )2 . (18.11)
η 2 t=1 m=1

Like in Chapter 11, it is more convenient to work with losses. Let Ŷti = 1 − X̂ti ,
yti = 1 − xti and Ỹtm = 1 − X̃tm . Note that Ỹt = E (t) Ŷt and recall the notation
Ati = I {At = i}, which means that Ŷti = APtitiyti and
 !2   2
(t) (t)
Xk Emi yti Xk (t)
2
E ytA Emi
Et [Ỹtm ] = Et  mAt = (18.12)
t
≤ .
PtAt i=1
Pti i=1
Pti

Therefore, using the definition of Pti ,

" M # " M #
X X Xk (t)
2 Emi
E Qtm (1 − X̃tm ) ≤ E Qtm
m=1 m=1 i=1
Pti
" k P #
X M Qtm E (t)
=E m=1 mi
= k.
i=1
Pti
 18.5 Notes 230

Substituting into Eq. (18.11) leads to

log(M ) ηnk p
Rn ≤ + = 2nk log(M ) .
η 2
Let us see how this theorem can be applied to the contextual bandit where C is a
finite set and Φ is the set of all functions from C → [k]. To each of these functions
(t)
φ ∈ Φ, we associate an expert m with Emi = I {φ(ct ) = i}. Then M = k C , and
Theorem 18.1 says that
p
Rn ≤ 2nk|C| log(k) ,
which is the same bound we derived using an independent copy of Exp3 for each
context. More generally, if C is arbitrary (possibly infinite) and Φ is a finite set
of functions from C to [k], then the theorem ensures that
p
Rn ≤ 2nk log(|Φ|) .
These results seem quite promising already, but in fact there is another
improvement possible. The intuition is that learning should be easier if the
experts have a high degree of agreement. One way to measure this is by
t X
X k
(s)
Et∗ = max Emi .
m∈[M ]
s=1 i=1

In Exercise 18.7, you will show that if all experts make identical recommendations,
then Et∗ = t and that no matter how the experts behave,
En∗ ≤ n min(k, M ) . (18.13)
In this sense En∗ /n can be viewed as the effective number of experts, which
depends on the degree of disagreement in the expert’s recommendations. By
modifying the algorithm to use a time varying learning rate, one can prove the
following theorem.
Theorem
p 18.3. Assume the same conditions as in Theorem 18.1, except let
ηt = log(M )/Et∗ . Then there exists a universal constant C > 0 such that
p
Rn ≤ C En∗ log(M ) . (18.14)
The proof of Theorem 18.3 is not hard and is left to Exercise 18.4. The bound
tells us that Exp4 with the suggested learning rate is able to adapt to degree of
disagreement between the experts, which seems like quite an encouraging result.
As a further benefit, the learning rate does not depend on the horizon so the
algorithm is anytime.

18.5 Notes

1 The most important concept in this chapter is that there are trade-offs when
choosing the competitor class. A large class leads to a more meaningful definition
 18.5 Notes 231

of the regret, but also increases the regret. This is similar to what we have
observed in stochastic bandits. Tuning an algorithm for a restricted environment
class usually allows faster learning, but the resulting algorithms can fail when
interacting with an environment that does not belong to the restricted class.
2 The Exp4 algorithm serves as a tremendous building block for other bandit
problems by defining your own experts. An example is the application of Exp4
to non-stationary bandits that we explore in Chapter 31, which is one of the
rare cases where Exp4 can be computed efficiently with a combinatorially large
number of experts. When Exp4 does not have an efficient implementation, it
often provides a good starting place to derive regret bounds without worrying
about computation (for an example, see Exercise 18.5).
3 The bandits with expert advice framework is clearly more general than
contextual bandits. With the terminology of the bandits with expert advice
framework, the contextual bandit problem arises when the experts are given
by static C → [k] maps.
4 A significant challenge is that a naive implementation of Exp4 has running
time O(M + k) per round, which can be enormous if either M or k is large. In
general there is no solution to this problem, but in some cases the computation
can be reduced significantly. One situation where this is possible is when the
learner has access to an optimisation oracle that for any context/reward
sequence returns the expert that would collect the most reward in this sequence
(this is equivalent to solving the offline problem Eq. (18.8)). In Chapter 30
we show how to use an offline optimisation oracle to learn efficiently in
combinatorial bandit problems. The idea is to solve a randomly perturbed
optimisation problem (leading to the so-called follow-the-perturbed-leader class
of algorithms) and then show that the randomness in the outputs provides
sufficient exploration. However, as we shall see there, these algorithms will
have some extra information, which makes estimating the rewards possible.
5 In the stochastic contextual bandit problem, it is assumed that the
context/reward pairs form a sequence of independent and identically distributed
random variables. Let Φ be a set of functions from C to [k] and suppose the
learner has access to an optimisation oracle capable of finding
t
X
argmaxφ∈Φ xsφ(cs )
s=1

for any sequence of reward vectors x1 , . . . , xt and contexts c1 , . . . , ct . A simple

and efficient algorithm that exploits such an oracle is based on explore-then-
commit, which has O(n2/3 ) regret (Exercise 18.8). There is a more sophisticated
algorithm that is still polynomial-time and for which the regret is about the
same as the result in Theorem 18.1 [Agarwal et al., 2014]. The algorithm
computes importance-weighted estimates of the rewards in each round. These
are used to estimate the regret of all the experts. Based on this, a distribution
over the experts (with a small support) is computed by solving a feasibility
 18.6 Bibliographic Remarks 232

problem. The distribution is constrained so that the importance weights will not
be too large, while the regret estimates averaged over the chosen distribution
will stay small. To reduce the computation cost, this distribution is updated
periodically with the length of the interval between the updates exponentially
growing. The significance of this result is that it reduces contextual bandits
to (cost-sensitive) empirical risk minimisation (ERM), which means that any
advance in solving cost-sensitive ERM problems automatically translates to
bandits.
6 The development of efficient algorithms for ERM is a major topic in supervised
learning. Note that ERM can be NP-hard even in simple cases like linear
classification [Shalev-Shwartz and Ben-David, 2014, §8.7].
7 The bound on the regret stated in Theorem 18.3 is data dependent. Note
that in adversarial bandits the data and instance are the same thing, while
in stochastic bandits the instance determines the probability distributions
associated with each arm and the data corresponds to samples from those
distributions. In any case a data/instance-dependent bound should usually be
preferred if it is tight enough to imply the worst-case optimal bounds.
8 There are many points we have not developed in detail. One is high-probability
bounds, which we saw in Chapter 12 and can also be derived here. We also
have not mentioned lower bounds. The degree to which the bounds are tight
depends on whether or not there is additional structure in the experts. In later
chapters we will see examples when the results are essentially tight, but there
are also cases when they are not.
9 Theorem 18.3 is the first result where we used a time-varying learning rate.
As we shall see in later chapters, time-varying learning rates are a powerful
way to make online algorithms adapt to specific characteristics of the problem
instance.

18.6 Bibliographic Remarks

For a good account on the history of contextual bandits, see the article by Tewari
and Murphy [2017]. The Exp4 algorithm was introduced by Auer et al. [2002b],
and Theorem 18.1 essentially matches theorem 7.1 of their paper (the constant
in Theorem 18.1 is slightly smaller). McMahan and Streeter [2009] noticed that
neither the number of experts nor the size of the action set are what really matters
for the regret, but rather the extent to which the experts tend to agree. McMahan
and Streeter [2009] also introduced the idea of finding the distribution to be played
to be maximally ‘similar’ to Pt (i) while ensuring sufficient exploration of each of
the experts. The idea of explicitly optimising a probability distribution with these
objectives in mind is at the heart of several subsequent works [e.g. Agarwal et al.,
2014]. While Theorem 18.3 is inspired by this work, the result appears to be new
and goes beyond the work of McMahan and Streeter [2009] because it shows
that all one needs is to adapt the learning rate based on the degree of agreement
 18.7 Exercises 233

amongst the experts. Neu [2015a] proves high-probability bounds for Exp4-IX.
You can follow in his footsteps by solving Exercise 18.3. Another way to get
high-probability bounds is to generalise Exp3.P, which was done by Beygelzimer
et al. [2011]. As we mentioned in Note 5, there exist efficient algorithms for
stochastic contextual bandit problems when a suitable optimisation oracle is
available [Agarwal et al., 2014]. An earlier attempt to address the problem of
reducing contextual bandits to cost-sensitive ERM is by Dudı́k et al. [2011]. The
adversarial case of static experts is considered by Syrgkanis et al. [2016], who
√
prove suboptimal (worse than n) regret bounds under various conditions for
follow-the-perturbed-leader for the transductive setting when the contexts are
available at the start. The case when the contexts are independent and identically
distributed, but the reward is adversarial is studied by Lazaric and Munos [2009]
for the finite expert case, while Rakhlin and Sridharan [2016] considers the case
when an ERM oracle is available. The paper of Rakhlin and Sridharan [2016] also
considers the more realistic case when only an approximation oracle is available
for the ERM problem. What is notable about this work is that they demonstrate
regret bounds with a moderate blow-up, but without changing the definition
of the regret. Kakade et al. [2008] consider contextual bandit problems with
adversarial context-loss sequences, where all but one action suffers a loss of one in
every round. This can also be seen as an instance of multi-class classification
with bandit feedback where labels to be predicted are identified with actions
and the only feedback received is whether the label predicted was correct, with
the goal of making as few mistakes as possible. Since minimising the regret is in
general hard in this non-convex setting, just like most of the machine learning
literature on classification, Kakade et al. [2008] provide results in the form of
mistake bounds for linear classifiers where the baseline is not the number of
mistakes of the best linear classifier, but is a convex upper bound on it. The
recent book by Shalev-Shwartz and Ben-David [2014] lists some hardness results
for ERM. For a more comprehensive treatment of computation in learning theory,
the reader can consult the book by Kearns and Vazirani [1994].

18.7 Exercises

18.1 Let C be a finite context set, and let c1 , . . . , cn ∈ C be an arbitrary sequence

of contexts.
v
u n
X uX p
(a) Show that t I {ct = c} ≤ n|C|.
c∈C t=1
(b) Assume that n is an integer multiple of |C|. Show that the choice that
maximises the right-hand side of the previous inequality is the one when
each context occurs n/|C| times.

18.2 Prove Lemma 18.2.

 18.7 Exercises 234

18.3 In this exercise you will prove an analogue of Theorem 12.1 for Exp4-IX.
In the contextual setting, the random regret is
n 
X 
(t)
R̂n = max Em xt − Xt .
m∈[M ]
t=1

Design an algorithm that accepts a parameter δ ∈ (0, 1) such that

s  !!
p nk 1
P R̂n ≥ C nk log(m) + log ≤ δ.
log(m) δ

18.4 Prove Theorem 18.3.

Hint The key idea is to modify the analysis of Exp3 to handle decreasing
learning rates. Of course you can do this directly yourself, or you can peek ahead
to Chapter 28, and specifically Exercises 28.12 and 28.13.
18.5 Let x1 , . . . , xn be a sequence of reward vectors chosen in advance by
an adversary with xt ∈ [0, 1]k . Furthermore, let o1 , . . . , on be a sequence of
observations, also chosen in advance by an adversary with ot ∈ [O] for some
fixed O ∈ N+ . Then let H be the set of functions φ : [O]m → [k] where m ∈ N+ .
In each round the learner observes ot and should choose an action At based on
o1 , A1 , X1 , . . . , ot−1 , At−1 , Xt−1 , ot , and the regret is
n
X
Rn = max xtφ(ot ,ot−1 ,...,ot−m ) − xtAt ,
φ∈H
t=1

where ot = 1 for t ≤ 0. This means the learner is competing with the best
predictor in hindsight that uses only the last m observations. Prove there exists
an algorithm such that
p
E[Rn ] ≤ 2knOm log(k) .

18.6 In this problem we consider non-oblivious experts. Consider the following

modified regret definition:
" n n
#
X X
(t)
Rn0 = max E Em xt − Xt .
m∈[M ]
t=1 t=1

Show the following:

(a) Rn0 ≤ Rn regardless of whether the experts are oblivious or not.
(b) Theorem 18.1 remains valid for non-oblivious experts if in Eq. (18.7) we
replace Rn with Rn0 . In particular, explain how to modify the proof.
(c) Research question: give a non-trivial bound on Rn .

18.7 Prove Eq. (18.13).

18.8 (Explore-then-commit) Consider a stochastic contextual bandit

environment where (Ct )nt=1 is a sequence of contexts sampled from distribution
 18.7 Exercises 235

ξ on C and the rewards are (Xt )nt=1 , where the conditional law of Xt given Ct
and At is PCt At . The mean
R reward when choosing action i ∈ [k] having observed
context c ∈ C is µ(c, i) = x dPci (x). Let Φ be a subset of functions from C to
[k]. The regret is
" n #
X
Rn = n sup µ(φ) − E Xt ,
φ∈Φ t=1
R
where µ(φ) = µ(c, φ(c))dξ(c). Consider a variation of explore-then-commit,
which explores uniformly at random for the first m rounds. Then define
m
k X
µ̂(φ) = I {At = φ(Ct )} Xt .
m t=1

For rounds t > m, the algorithm chooses At = φ̂∗ (Ct ), where

n
X
φ̂ = argmaxφ∈Φ µ̂(φ) = argmaxφ∈Φ
∗
X̂tφ(Ct ) ,
t=1

where X̂ti = kI {At = φ(Ct )} Xt . When no maximiser exists you may assume
that µ̂(φ̂∗ ) ≥ supφ∈Φ µ̂(φ) − ε for any ε > 0 of your choice. Show that when Φ
is finite, then for appropriately tuned m the expected regret of this algorithms
satisfies
 
Rn = O n2/3 (k log(|Φ|))1/3 .

This algorithm is the explore-then-commit version of the epoch-greedy

algorithm by Langford and Zhang [2008]. You should not worry too much
about these details, but of course C should be associated with a σ-algebra
and the family of distributions (Pca : c ∈ C, a ∈ [k]) should be a probability
kernel from C × [k] to R.

18.9 Consider a stochastic contextual bandit problem with the same set-up as
the previous exercise and k = 2 arms. As before, let Φ be a set of functions from
C to [k]. Design a policy such that
" n # r
X n
Rn = n max µ(φ) − E Xt ≤ C ndk log ,
φ∈Φ
t=1
d

where C > 0 is a universal constant and d = VC(Φ) is the VC dimension of Φ.

Hint Use an initial period of exploration to choose a finite ‘representative’

subset of Φ, and then run Exp4 on this subset. The result that you need to know in
connection to the VC dimension is known as Sauer’s lemma, which states that if Φ
has VC dimension d, then for any sequence c = (ci )m i=1 ⊂ C, the cardinality of the
set Φc = {(φ(c1 ), . . . , φ(cm )) : φ ∈ Φ} is at most (em/d)d . You may also find it
 18.7 Exercises 236

useful that an i.i.d. sequence C1 , . . . , Cn is ‘exchangeable’: for any set A ⊂ C n and

any π : [n] → [n] bijection, P ((C1 , . . . , Cn ) ∈ A) = P (Cπ(1) , . . . , Cπ(n) ) ∈ A .
This property helps you to argue that the finite subset of Φ obtained by choosing
functions from Φ that all disagree on the first m elements on the contexts will be
representative of the behaviour of functions in Φ on the rest of the contexts.

We did not talk about VC dimension in this book. An introduction is given

by Shalev-Shwartz and Ben-David [2014], or there is the classic text by
Vapnik [1998]. The application to bandits is due to Beygelzimer et al. [2011].
19 Stochastic Linear Bandits

Contextual bandits generalise the finite-armed setting by allowing the learner

to make use of side information. This chapter focusses on a specific type of
contextual bandit problem in the stochastic set-up where the reward is assumed
to have a linear structure that allows for learning to transfer from one context to
another. This leads to a useful and rich model that will be the topic of the next
few chapters. To begin, we describe the stochastic linear bandit problem and
start the process of generalising the upper confidence bound algorithm.

19.1 Stochastic Contextual Bandits

The stochastic contextual bandit problem mirrors the adversarial contextual

bandit set-up discussed in Chapter 18. At the beginning of round t, the learner
observes a context Ct ∈ C, which may be random or not. Having observed the
context, the learner chooses their action At ∈ [k] based on the information
available. So far everything is the same as the adversarial setting. The difference
comes from the assumption that the reward Xt satisfies

Xt = r(Ct , At ) + ηt ,

where r : C × [k] → R is called the reward function and ηt is the noise, which
we will assume is conditionally 1-subgaussian. Precisely, let

Ft = σ(C1 , A1 , X1 , . . . , Ct−1 , At−1 , Xt−1 , Ct , At )

be the σ-field summarising the information available just before Xt is observed.

Then, we assume that
 2
λ
E [exp(ληt ) | Ft ] ≤ exp almost surely .
2

The noise could have been chosen to be σ-subgaussian for any known σ 2 , but
like in earlier chapters, we save ourselves some ink by fixing its value to σ 2 = 1.
Remember from Chapter 5 that subgaussian random variables have zero mean,
so the assumption also implies that E [ηt | Ft ] = 0 and E [Xt | Ft ] = r(Ct , At ).
 19.1 Stochastic Contextual Bandits 238

If F ⊂ G are σ-algebras, and E[exp(λη) | G] ≤ exp(λ2 /2) almost surely, then

by the tower rule, E[exp(λη) | F] ≤ exp(λ2 /2) almost surely. Hence, the
condition that ηt is subgaussian with respect to Ft can be ‘relaxed’ to the
condition that it be subgaussian with respect to any σ-algebra containing
Ft .

If r was given, then the action in round t with the largest expected return
is A∗t ∈ argmaxa∈[k] r(Ct , a). Notice that this action is now a random variable
because it depends on the context Ct . The loss due to the lack of knowledge of r
makes the learner incur the (expected) regret
" n n
#
X X
Rn = E max r(Ct , a) − Xt .
a∈[k]
t=1 t=1

Like in the adversarial setting, there is one big caveat in this definition of the
regret. Since we did not make any restrictions on how the contexts are chosen, it
could be that choosing a low-rewarding action in the first round might change
the contexts observed in subsequent rounds. Then the learner could potentially
achieve an even higher cumulative reward by choosing a ‘suboptimal’ arm initially.
As a consequence, this definition of the regret is most meaningful when the actions
of the learner do not (greatly) affect subsequent contexts.
One way to eventually learn an optimal policy is to estimate r(c, a) for each
(c, a) ∈ C × [k] pair. As in the adversarial setting, this is ineffective when the
number of context-action pairs is large. In particular, the worst-case regret over
all possible√contextual problems with M contexts and mean reward in [0, 1] is
at least Ω( nM k). While this may not look bad, M is often astronomical (for
example, 2100 ). The argument that gives rise to the mentioned lower bound
relies on designing a problem where knowledge of r(c, ·) for context c provides
no useful information about r(c0 , ·) for some different context c0 . Fortunately,
in most interesting applications, the set of contexts is highly structured, which
is often captured by the fact that r(·, ·) changes ‘smoothly’ as a function of its
arguments.
A simple, yet interesting assumption to capture further information about the
dependence of rewards on context is to assume that the learner has access to a
map ψ : C × [k] → Rd , and for an unknown parameter vector θ∗ ∈ Rd , it holds
that

r(c, a) = hθ∗ , ψ(c, a)i , for all (c, a) ∈ C × [k] . (19.1)

The map ψ is called a feature map, which is the standard nomenclature in

machine learning. The idea of feature maps is best illustrated with an example.
Suppose the context denotes the visitor of a website selling books, the actions are
books to recommend and the reward is the revenue on a book sold. The features
could indicate the interests of the visitors as well as the domain and topic of the
 19.2 Stochastic Linear Bandits 239

book. If the visitors and books are assigned to finitely many categories, indicator
variables of all possible combinations of these categories could be used to create
the feature map. Of course, many other possibilities exist. For example, you can
train a neural network (deep or not) on historical data to predict the revenue
and use the nonlinear map that we obtained by removing the last layer of the
neural network. The subspace Ψ spanned by the feature vectors {ψ(c, a)}c,a in
Rd is called the feature space.
If k · k is a norm on Rd then, an assumption on kθ∗ k implies smoothness of r.
In particular, from Hölder’s inequality,
|r(c, a) − r(c0 , a0 )| ≤ kθ∗ kkψ(c, a) − ψ(c0 , a0 )k∗ ,
where k · k∗ denotes the dual of k · k. Restrictions on kθ∗ k have a similar effect
to assuming that the dimensionality d is finite. In fact, one may push this to
the extreme and allow d to be infinite, an approach that can buy tremendous
flexibility and makes the linearity assumption less limiting.

19.2 Stochastic Linear Bandits

Stochastic linear bandits arise from realising that under Eq. (19.1), all that
matters is the feature vector that results from choosing a given action and not
the ‘identity’ of the action itself. This justifies studying the following simplified
model: in round t, the learner is given the decision set At ⊂ Rd , from which it
chooses an action At ∈ At and receives reward
Xt = hθ∗ , At i + ηt ,
where ηt is 1-subgaussian given A1 , A1 , X1 , . . . , At−1 , At−1 , Xt−1 , At and At . The
random (pseudo-)regret and regret are defined by
n
X
R̂n = max hθ∗ , a − At i ,
a∈At
t=1
" n #
h i X n
X
Rn = E R̂n = E max hθ∗ , ai − Xt ,
a∈At
t=1 t=1

respectively. Different choices of At lead to different settings, some of which we

have seen before. For example, if (ei )i are the unit vectors and At = {e1 , . . . , ed },
then the resulting stochastic linear bandit problem reduces to the finite-
armed setting. On the other hand, if At = {ψ(Ct , i) : i ∈ [k]}, then we have a
contextual linear bandit. Yet another possibility is a combinatorial action
set At ⊆ {0, 1} . Many combinatorial problems (such as matching, least-cost
d

problems in directed graphs and choosing spanning trees) can be written as linear
optimisation problems over some combinatorial set A obtained from considering
incidence vectors often associated with some graph. Some of these topics will be
covered later in Chapter 30.
 19.2 Stochastic Linear Bandits 240

As we have seen in earlier chapters, the UCB algorithm is an attractive approach

for finite-action stochastic bandits. Its best variants are nearly minimax optimal,
instance optimal and exactly optimal asymptotically. With these merits in mind,
it seems quite natural to try and generalise the idea to the linear setting.
The generalisation is based on the view that UCB implements the ‘optimism
in the face of uncertainty’ principle, which is to act in each round as if the
environment is as nice as plausibly possible. In finite-action stochastic bandits,
this means choosing the action with the largest upper confidence bound. In the
case of linear bandits, the idea remains the same, but the form of the confidence
bound is more complicated because rewards received yield information about
more than just the arm played.
The first step is to construct a confidence set Ct ⊂ Rd based on
(A1 , X1 , . . . , At−1 , Xt−1 ) that contains the unknown parameter vector θ∗ with
high probability. Leaving the details of how the confidence set is constructed
aside for a moment, and assuming that the confidence set indeed contains θ∗ , for
any given action a ∈ Rd , let

UCBt (a) = maxhθ, ai (19.2)

θ∈Ct

be an upper bound on the mean pay-off hθ∗ , ai of a. The UCB algorithm that
uses the confidence set Ct at time t then selects

At = argmaxa∈At UCBt (a) . (19.3)

UCB applied to linear bandits is known by various names, including LinRel

(linear reinforcement learning), LinUCB and OFUL (optimism in the f ace of
uncertainty for linear bandits). We will not be very dogmatic of this name and
call algorithms with the above construct instances of LinUCB.
The main question is how to choose the confidence set Ct ⊂ Rd . As usual, there
are conflicting desirable properties:

(a) Ct should contain θ∗ with high probability.

(b) Ct should be as small as possible.

At first sight it is not at all obvious what Ct should look like. After all, it is a
subset of Rd , not just an interval like the confidence intervals about the empirical
estimate of the mean reward for a single action that we saw in the previous
chapters. While we specify the analytic form of a possible construction for Ct here,
there are some details in choosing some of the parameters in this construction. As
they are both delicate and important, we dedicate the next chapter to discussing
them.
Following the idea for UCB, we need an analogue for the empirical estimate
of the unknown quantity, which in this case is θ∗ . There are several principles
one might use for deriving such an estimate. For now we use the regularised
 19.2 Stochastic Linear Bandits 241

least-squares estimator, which is

t
!
X
2
θ̂t = argminθ∈Rd (Xs − hθ, As i) + λkθk22 , (19.4)
s=1

where λ ≥ 0 is called the penalty factor. Choosing λ > 0 helps because it

ensures that the loss function has a unique minimiser even when A1 , . . . , At do
not span Rd , which simplifies the math. The solution to Eq. (19.4) is obtained
easily by differentiation and is

t
X
θ̂t = Vt−1 As Xs , (19.5)
s=1

where (Vt )t are d × d matrices given by

t
X
V0 = λI and Vt = V0 + As A>
s . (19.6)
s=1

So θ̂t is an estimate of θ∗ , which makes it natural to choose Ct to be centered at

θ̂t−1 . For what follows, we will simply assume that the confidence set Ct is closed
and satisfies
n o
Ct ⊆ Et = θ ∈ Rd : kθ − θ̂t−1 k2Vt−1 ≤ βt , (19.7)

where (βt )t is an increasing sequence of constants with β1 ≥ 1. The set Et is an

ellipsoid centred at θ̂t−1 and with principle axis being the eigenvectors of Vt with
corresponding lengths being the reciprocal of the eigenvalues. Notice that as t
grows, the matrix Vt has increasing eigenvalues, which means the volume of the
ellipse is also shrinking (at least, provided βt does not grow too fast). As noted
beforehand, the next chapter will be devoted to show that Ct = Et is a natural
choice for carefully chosen βt . In the rest of this chapter, we simply examine the
consequence of using a confidence set satisfying Eq. (19.7) and assume all the
desirable properties.

The impatient reader who is puzzled of the form Et may briefly think of
the case when ηs ∼ N (0, σ 2 ), A1 , . . . , At−1 are deterministic and span Rd so
that we can take λ = 0. In this case, one easily computes that with V = Vt−1 ,
Z = V 1/2 (θ̂t−1 − θ∗ ) ∼ N (0, I), or that kZk2 is the sum of d, independent
standard normal random variables, and thus it follows the χ2 -distribution
(with d degrees of freedom), from which one can find the appropriate value
of βt−1 . As we shall see, the expression one can get from this calculation,
will, more or less, be still correct in the general case.
 19.3 Regret Analysis 242

19.3 Regret Analysis

We prove a regret bound for LinUCB under the assumption that the confidence
intervals indeed contain the true parameter with high probability and boundedness
conditions on the action set and rewards.

Assumption 19.1. The following hold:

(a) 1 ≤ β1 ≤ β2 ≤ · · · ≤ βn .
(b) maxt∈[n] supa,b∈At hθ∗ , a − bi ≤ 1.
Sn
(c) kak2 ≤ L for all a ∈ t=1 At .
(d) There exists a δ ∈ (0, 1) such that with probability 1 − δ, for all t ∈ [n],
θ∗ ∈ Ct where Ct satisfies Eq. (19.7).

Theorem 19.2. Under the conditions of Assumption 19.1 with probability 1 − δ,

the regret of LinUCB satisfies
s   s  
det Vn dλ + nL2
R̂n ≤ 8nβn log ≤ 8dnβn log .
det V0 dλ

Theorem 20.5 in the next chapter shows that βn may be chosen to be

s    
p √ 1 dλ + nL2
βn = λm2 + 2 log + d log , (19.8)
δ dλ

where m2 is an upper bound on kθ∗ k2 . By choosing δ = 1/n, we obtain the

following corollary bounding the expected regret.

Corollary 19.3. Under the conditions of Assumption 19.1, the expected regret
of LinUCB with δ = 1/n is bounded by
√
Rn ≤ Cd n log(nL) ,

where C > 0 is a suitably large universal constant.

The proof of Theorem 19.2 depends on the following lemma, often called the
elliptical potential lemma.

Lemma 19.4. Let V0 ∈ Rd×d be positive definite and a1 , . . . , an ∈ Rd be a sequence

P
of vectors with kat k2 ≤ L < ∞ for all t ∈ [n], Vt = V0 + s≤t as a> s . Then,

n      
X det Vn trace V0 + nL2
1∧ kat k2V −1 ≤ 2 log ≤ 2d log .
t=1
t−1 det V0 d det(V0 )1/d

Proof Using that for any u ≥ 0, u ∧ 1 ≤ 2 ln(1 + u), we get

n 
X  X  
1 ∧ kat k2V −1 ≤2 log 1 + kat k2V −1 .
t−1 t−1
t=1 t
 19.3 Regret Analysis 243

 
det Vn
We now argue that this last expression is log det V0 . For t ≥ 1, we have
1/2 1/2
Vt = Vt−1 + at a>
t = Vt−1 (I + Vt−1 at at Vt−1 )Vt−1 .
> −1/2 −1/2

Now, since the determinant is a multiplicative map,

   
det(Vt ) = det(Vt−1 ) det I + Vt−1 at a> = det(Vt−1 ) 1 + kat k2V −1 ,
−1/2 −1/2
t Vt−1
t−1

where the second equality follows because the matrix I + yy has eigenvalues >

1 + kyk22 and 1 as well as the fact that the determinant of a matrix is the product
of its eigenvalues. Putting things together, we see that
n 
Y 
det(Vn ) = det(V0 ) 1 + kat k2V −1 , (19.9)
t−1
t=1

which is equivalent to the first inequality that we wanted to prove. To get the
second inequality, note that by the inequality of arithmetic and geometric means,
 d  d
1 trace V0 + nL2
Y d
det(Vn ) = λi ≤ trace Vn ≤ ,
i=1
d d

where λ1 , . . . , λd are the eigenvalues of Vn .

Proof of Theorem 19.2 By part (d) of Assumption 19.1, it suffices to prove the
bound on the event that θ∗ ∈ Ct for all rounds t ∈ [n]. Let A∗t = argmaxa∈At hθ∗ , ai
be an optimal action for round t and rt be the instantaneous regret in round t
defined by

rt = hθ∗ , A∗t − At i .

Let θ̃t ∈ Ct be the parameter in the confidence set for which hθ̃t , At i = UCBt (At ).
Then, using the fact that θ∗ ∈ Ct and the definition of the algorithm leads to

hθ∗ , A∗t i ≤ UCBt (A∗t ) ≤ UCBt (At ) = hθ̃t , At i .

Using Cauchy–Schwarz inequality and the assumption that θ∗ ∈ Ct and facts that
θ̃t ∈ Ct and Ct ⊆ Et leads to

rt = hθ∗ , A∗t − At i ≤ hθ̃t − θ∗ , At i ≤ kAt kV −1 kθ̃t − θ∗ kVt−1

t−1
p
≤ 2kAt kV −1 βt . (19.10)
t−1

By part (b) we also have rt ≤ 2, which, combined with βn ≥ max{1, βt }, yields

p p  
rt ≤ 2 ∧ 2 βt kAt kV −1 ≤ 2 βn 1 ∧ kAt kV −1 .
t−1 t−1

Then, by Cauchy–Schwarz inequality,

v v
n u n u n  
X u X u X
R̂n = rt ≤ tn rt2 ≤ 2tnβn 1 ∧ kAt k2V −1 . (19.11)
t−1
t=1 t=1 t=1
 19.4 Notes 244

The result is completed using Lemma 19.4, which depends on part (c) of
Assumption 19.1.

19.3.1 Computation
An obvious question is whether or not the optimisation problem in Eq. (19.3) can
be solved efficiently. First note that the computation of At can also be written as

(At , θ̃t ) = argmax(a,θ)∈At ×Ct hθ, ai . (19.12)

This is a bilinear optimisation problem over the set At × Ct . In general, not much
can be said about the computational efficiency of solving this problem. There are
two notable special cases, however.

(a) Suppose that a(θ) = argmaxa∈At hθ, ai can be computed efficiently for any θ
and that Ct = co(φ1 , . . . , φm ) is the convex hull of a finite set. Then At can
be computed by finding a(φ1 ), . . . , a(φm ) and choosing At = a(φi ), where i
maximises hφi , a(φi )i.
(b) Assume that Ct = Et is the ellipsoid given in Eq. (19.7) and At is a small
finite set. Then the action At from Eq. (19.12) can be found using
p
At = argmaxa∈At hθ̂t−1 , ai + βt kakV −1 , (19.13)
t−1

which may be solved by simply iterating over the arms and calculating
the term inside the argmax. Further implementation issues are explored in
Exercise 19.8.

19.4 Notes

1 It was mentioned that ψ may map its arguments to an infinite dimensional

space. There are several issues that arise in this setting. The first is whether or
not the algorithm can be computed efficiently. This is usually tackled via the
kernel trick, which assumes the existence of an efficiently computable kernel
function κ : (C × [k]) × (C × [k]) → R such that

hψ(c, a), ψ(c0 , a0 )i = κ((c, a), (c0 , a0 )) .

The trick is to rewrite all computations in terms of the kernel function so that
ψ(c, a) is neither computed, nor stored. The second issue is that the claim
made in Theorem 19.2 depends on the dimension d and becomes vacuous when
d is large or infinite. This dependence arises from Lemma 19.4. It is possible to
modify this result by replacing d with a data-dependent quantity that measures
the ‘effective dimension’ of the image of the data under φ. The final challenge
is to define an appropriate confidence set. See the bibliographic remarks for
further details and references.
 19.4 Notes 245

2 The bound given in Theorem 19.2 is essentially a worst-case style of bound,

with little dependence on the parameter θ∗ or the geometry of the action set.
Instance-dependent bounds for linear bandits are still an open topic of research,
and the asymptotics are only understood in the special case where the action
set is finite and unchanging (Chapter 25).
3 Theorem 20.5 in the next chapter shows that (βt )nt=1 as defined in Eq. (19.8)
can be replaced with a data-dependent quantity that is strictly smaller:
s    
1/2
√ 1 det(Vt−1 )
βt = m2 λ + 2 log + log . (19.14)
δ λd

Empirically this choice is never worse than the value suggested in Eq. (19.8)
and sometimes better, typically by a modest amount.
4 The application of Cauchy–Schwarzpin Eq. (19.11) often loses a logarithm,
as it does, for example, when rt = 1/t. Recently, however, a lower bound
for contextual linear bandits has been derived by constructing a sequence for
which this Cauchy–Schwarz is tight, as well as Lemma 19.4 [Li et al., 2019b].
5 In the worst case, the bound in Theorem 19.2 is tight up to logarithmic factors.
More details are in Chapter 24, which is devoted to lower bounds for stochastic
linear bandits. The environments for which the lower bound nearly matches the
upper bound have action sets that are either infinite or exponentially large in
the dimension. When |At | ≤ k for all rounds t, there are algorithms for which
the regret is
q 
3
Rn = O dn log (nk) .

The special case where the action set does not change with time is treated in
Chapter 22, where references to the literature are also provided.
6 The calculation in Eq. (19.13) shows that LinUCB has more than just a passing
resemblance to the UCB algorithm introduced in Chapter 7. The term hθ̂t−1 , ai
may be interpreted as an empirical estimate of the reward from choosing action
√
a, and βt kakV −1 is a bonus term that ensures sufficient exploration. If the
t−1

penalty term vanishes (λ = 0) and At = {e1 , . . . , ed } for all t ∈ [n], then θ̂i
becomes the empirical mean of action ei , and the matrix Vt is diagonal, with
its ith diagonal entry being the number of times action ei is used up to and
including round t. Then the bonus term has order
s
p βt
βt kei kV −1 = ,
t−1 Ti (t − 1)

where Ti (t − 1) is the number of times action ei has been chosen before the tth
round. So UCB for finite-armed bandits is recovered by choosing βt = 2 log(·),
where the term inside the logarithm can be chosen in a variety of ways as
discussed in earlier chapters. Notice now
p that the simple analysis given in this
chapter leads to a regret bound of O( dn log(·)), which is quite close to the
 19.5 Bibliographic Remarks 246

highly specialised analysis given in Chapters 7 to 9. Note however that the

dimension-free choice of βt does not satisfy Eq. (19.7), but this happens to be
unnecessary for the proof of Theorem 19.2 to go through.
7 An extension of considerably interest of the linear model is the generalised
linear model where the reward is
Xt = µ(hθ∗ , At i) + ηt , (19.15)
where µ−1 : R → R is called the link function. A common choice is the sigmoid
function: µ(x) = 1/(1 + exp(−x)) Bandits with rewards from a generalised
linear model have been studied by Filippi et al. [2010], who prove a bound with
a similar form as Theorem 19.2. Unfortunately, however, the bound depends in
a slightly unpleasant manner on the form of the link function, and it seems
there may be significant room for improvement. You will analyse and algorithm
for generalised linear bandits in Exercise 19.6.
8 Beyond optimism, there are at least three other principles for constructing
algorithms for stochastic linear bandits. The first is Thompson sampling, which
is a randomised Bayesian algorithm discussed at length in Chapter 36. The
second is a class of algorithms designed to achieve asymptotic optimality
in the special cases where the action set is fixed, or sampled i.i.d. from a
fixed distribution with finite support. These algorithms fall into the class of
‘optimisation-based’ algorithms that estimate the unknown parameter and
then solve an optimisation problem to determine an optimal allocation over
the actions [Lattimore and Szepesvári, 2017, Ok et al., 2018, Combes et al.,
2017, Hao et al., 2020]. A downside of optimisation-based approaches is that
so far the results have a very asymptotic nature and the algorithms are not
very practical. These ideas are discussed a little more in Chapter 25, where we
prove asymptotic lower bounds for linear bandits. The third design principle is
called information-directed sampling, which has a Bayesian version [Russo and
Van Roy, 2014a] and frequentist analogue [Kirschner and Krause, 2018]. In
rough terms, these algorithms choose a distribution over actions that minimises
the ratio of a squared expected instantaneous regret and the information gain
about the optimal action, which in the frequentist version is replaced by a
potential function that mimics the information gain.

19.5 Bibliographic Remarks

Stochastic linear bandits were introduced by Abe and Long [1999]. The first paper
to consider algorithms based on the optimism principle for linear bandits is by
Auer [2002], who considered the case when the number of actions is finite. The
core ideas of the analysis of optimistic algorithms (and more) is already present
in this paper. An algorithm based on confidence ellipsoids is described in the
papers by Dani et al. [2008], Rusmevichientong and Tsitsiklis [2010] and Abbasi-
Yadkori et al. [2011]. The regret analysis presented here, and the discussion of the
 19.6 Exercises 247

computational questions, is largely based on the former of these works, which also
√
stresses that an expected regret of Õ(d n) can be achieved regardless of the shape
of the decision sets At as long as the means are guaranteed to lie in a bounded
interval. Rusmevichientong and Tsitsiklis [2010] consider both optimistic and
explore-then-commit strategies, which they call ‘phased exploration and greedy
exploitation’ (PEGE). They focus on the case where At is the unit ball or some
other compact set with a smooth boundary and show that PEGE is optimal up to
logarithmic factors. The observation that explore-then-commit works for the unit
ball (and other action sets with a smooth boundary) was independently made
by Abbasi-Yadkori et al. [2009], further expanded in [Abbasi-Yadkori, 2009a].
Generalised linear models are credited to Nelder and Wedderburn [1972]. We
mentioned already that LinUCB was generalised to this model by Filippi et al.
[2010]. A more computationally efficient algorithm has recently been proposed by
Jun et al. [2017]. Nonlinear structured bandits where the pay-off function belongs
to a known set have also been studied [Anantharam et al., 1987, Russo and Van
Roy, 2013, Lattimore and Munos, 2014]. Kernelised versions of UCB have been
given by Srinivas et al. [2010], Abbasi-Yadkori [2012] and Valko et al. [2013b].
We mentioned early in the chapter that making assumptions on the norm θ∗ is
related to smoothness of the reward function with smoother functions leading
to stronger guarantees. For an example of where this is done, see the paper on
‘spectral bandits’ by Valko et al. [2014] and Exercise 19.7.

19.6 Exercises

19.1 (Least-squares solution) Prove that the solution given in Eq. (19.5) is
indeed the minimiser of Eq. (19.4).

19.2 (Action selection with ellipsoidal confidence sets) Show that the
action selection in LinUCB can indeed be done as shown in Eq. (19.13) when
Ct = Et is an ellipsoid given in Eq. (19.7).

19.3 (Elliptical potentials: You cannot have more than O(d) big
intervals) Let V0 = λI and a1 , . . . , an ∈ Rd be a sequence of vectors with
Pt
kat k2 ≤ L for all t ∈ [n]. Then let Vt = V0 + s=1 as a>s and show that the
number of times kat kV −1 ≥ 1 is at most
t−1

 
3d L2
log 1 + .
log(2) λ log(2)
 19.6 Exercises 248

The proof of Theorem 19.2 depended on part (b) of Assumption 19.1, which
asserts that the mean rewards are bounded by one. Suppose we replace this
assumption with the relaxation that there exists a B > 0 such that
max sup hθ∗ , a − bi ≤ B .
t∈[n] a,b∈At

Then, Exercise 19.3 allows you to bound the number of rounds when
kxt kV −1 ≥ 1, and in these rounds the naive bound of rt ≤ B is used. For
t−1
the remaining rounds, the analysis of Theorem 19.2 goes through unaltered.
As a consequence we see that the dependence on B is an additive constant
term that does not grow with the horizon.

19.4 (Computation cost savings with fixed large action set) When
the action set At = A is fixed and |A| = k, the total computation cost of
LinUCB after n rounds is O(kd2 n) in n rounds if the advice on implementation
of Exercise 19.8 is used. This can be reduced to O(k log(n) + d2 n) with almost
no increase of the regret, a significant reduction when k  d2 . For this, LinUCB
should be modified to work in phases, where in a given fixed it uses the same
action computed in the usual way at the beginning of the phase. A phase ends
when log det Vt (λ) increases by log(1 + ε).

kxk2A det A
(a) Prove that if 0 ≺ B  A then supx6=0 kxk2B
≤ det B .

(b) Let Assumption 19.1 hold. Let R̂n (βn ) be the regret bound of LinUCB
stated in Theorem 19.2. Show that with probability 1 − δ the random pseudo-
regret R̂n of the phased version of LinUCB, as described above, satisfies
R̂n ≤ R̂n ((1 + ε)βn ).

19.5 (Lipschitz reward functions) Consider the k-armed stochastic

contextual setting of Section 19.1, and assume that C = [0, 1] and that the
reward functions r(·, i) : C → [0, 1] are L-Lipschitz:

|r(x, i) − r(y, i)| ≤ L|x − y| for all x, y ∈ [0, 1] , i ∈ [k] .

(a) Construct an algorithm whose regret Rn after n rounds is O((Lk log k)1/3 n2/3 ).
(b) Show that the minimax optimal regret is of the order Ω((Lk)1/3 n2/3 ).
(c) Generalise the result to the case when C = [0, 1]d and in the definition
of Lipschitzness we use the Euclidean norm. Show the dependence on the
dimension in the lower and upper bounds. Discuss the influence of the choice
of the norm.

Hint Consider discretising C. Alternatively, use Exp4.

 19.6 Exercises 249

This exercise is inspired by the work of Perchet and Rigollet [2013], who focus
on improving the regret bound by adaptive discretisation when a certain
margin condition holds. There are many variations of the problem of the
previous exercise. For starters, the domain of contexts could be more general:
one may consider higher-order smoothness, continuous action and context
spaces. What is the role of the context distribution? In some applications,
the context distribution can be estimated for free, in which case you might
assume the context distribution is known. How to take a known context
distribution into account? To whet your appetite, if the context distribution
is concentrated on a handful of contexts, the discretisation should respect
which contexts the distribution is concentrated on. Instead of discretisation,
one may also consider function approximation. An interesting approach that
goes beyond discretisation is by Combes et al. [2017] (see also Magureanu
et al. 2014). The approach in these papers is to derive an asymptotic,
instance-dependent lower bound, which is then used to guide the algorithm
(much like in the track-and-stop algorithm in Section 33.2). An open problem
is to design algorithms that are simultaneously near minimax optimal and
asymptotically optimal. As described in Part II, this problem is now settled
for finite-armed stochastic bandits, the only case where we can say this in
the whole literature of bandits.

19.6 (Generalised linear bandits) In this exercise you will design and
analyse an algorithm for the generalised linear bandit problem mentioned in
Note 7. Let Θ be a convex compact subset of Rd and assume that θ∗ ∈ Θ. The
only difference relative to the standard model is that the reward is

Xt = µ(hθ∗ , At i) + ηt ,

where µ : R → R is a continuously differentiable function such that

 
c1 = min 1, min min µ (hθ, ai) > 0 and
0
a∈∪n
t=1 At θ∈Θ
 
c2 = max 1, max n
max µ (hθ, ai) < ∞ .
0
a∈∪t=1 At θ∈Θ

That c1 > 0 is assumed implies that µ is increasing on the relevant area of its
domain. Like in the standard model, for each t ≥ 1, ηt is 1-subgaussian given
A1 , X1 , . . . , At−1 , Xt−1 , At , and you may as well assume that rewards and feature
vectors are bounded:

max µ(hθ∗ , ai) − µ(hθ∗ , bi) ≤ 1 and max kak2 ≤ L and kθ∗ k2 ≤ m2 .
a,b∈∪n
t=1 At a∈∪n
t=1 At

Recall that λ is the regularisation parameter in the definition of Vt (see Eq. (19.6))
 19.6 Exercises 250

and let

t
X t
X

gt (θ) = λθ + µ(hθ, As i)As , Lt (θ) = gt (θ) − Xs As .

s=1 s=1 Vt−1

(a) Let βt be as in Eq. (19.8) and define a confidence set Ct by

n o
1/2
Ct = θ ∈ Θ : Lt−1 (θ) ≤ βt−1 .

Show that θ∗ ∈ Ct for all t with probability at least 1 − δ.

(b) Prove that for all θ, θ0 ∈ Θ, c1 kθ − θ0 kVt ≤ kgt (θ) − gt (θ0 )kV −1 .
t
(c) Consider the algorithm that chooses

At = argmaxa∈At max µ(hθ, ai) .

θ∈Ct

Prove that on the event that θ∗ ∈ Ct , for A∗t = argmaxa∈At µ(hθ∗ , ai),
1/2
2c2 βt−1
rt = µ(hθ∗ , A∗t i) − µ(hθ∗ , At i) ≤ kAt kV −1 .
c1 t−1

Pn
(d) Prove that with probability at least 1 − δ, the random regret R̂n = t=1 rt
is bounded by
s  
c2 nL2
R̂n ≤ 8ndβn log 1 + .
c1 d

Hint For (a), you should peek into the future and use Theorem 20.4. The
mean value theorem will help with Part (b).
19.7 (Spectral bandits) The regret of LinUCB can be improved considerably
if an appropriate norm of θ∗ is known to be small. In this exercise you will
investigate this phenomenon. Suppose that V0 is positive definite with eigenvalues
Pt
λ1 , . . . , λd , respective eigenvectors v1 , . . . , vd , and Vt = V0 + s=1 As A> s . All
other quantities are left unchanged, but the alternative value of V0 means that θ̂t
is heavily regularised in the direction of each vi for which λi is large. Without
loss of generality, assume that (λi )di=1 is increasing and let λ = λ1 be the smallest
eigenvalue. Define the ‘effective dimension’ by
 
n
deff = max i ∈ [d] : (i − 1)λi ≤ ∈ [d] .
log(1 + nL2 /λ)
   
det(Vt ) nL2
(a) Prove that log det(V 0)
≤ 2deff log 1 + λ .
(b) Let m > 0 be a user-defined constant and
s    
1/2 1 det(Vt )
βt = m + 2 log + log
δ det(V0 )

and let Ct = {θ : kθ − θ̂t−1 k2Vt−1 ≤ βt−1 }. Assume that kθ∗ kV0 ≤ m and
prove that θ∗ ∈ Ct for all t with probability at least 1 − δ.
 19.6 Exercises 251

(c) Prove that if kθ∗ kV0 ≤ m, then with probability at least 1 − δ, the random
regret of LinUCB in this setting is bounded by
s  
nL2
R̂n ≤ 8βn deff log 1 + .
λ
(d) Show that with an appropriate choice of δ,
√

E[R̂n ] = O deff n log(nL2 ) , (19.16)
where the last equality suppresses dependence on m and λ = λ1 .
(e) The result of the previous display explains the definition of the ‘effective
dimension’ deff . When V0 = I, which gives rise to uniform regularisation,
√
Corollary 19.3 states that for kθ∗ k ≤ m, E[R̂n ] = O(d n log(nL2 )). Given
Eq. (19.16), for a fixed n, deff can be thought of as replacing the dimension
d when V0 is chosen as any positive definite matrix with V0  λI. It follows
then that when deff ≤ d, the upper bound for non-uniform regularisation will
be smaller. Given this, explain the potential pros and cons of non-uniform
regularisation. What happens when kθ∗ kV0 ≤ m fails to hold? Show a bound
on the degradation of the expected regret as a function of max(0, kθ∗ kV0 −m).

Hint For Part (b), you should peek into the next chapter and modify
Theorem 20.5.

Valko et al. [2014] had a particular application in mind when designing

the algorithm in Exercise 19.7. Consider a large graph with k vertices
and similarity matrix between the vertices W ∈ [0, ∞)k×k . The graph
Laplacian is the matrix L = D − W , where D is diagonal with Dii =
Pk
j=1 Wij . Let µ ∈ [0, 1] be an unknown reward function and consider a
k

bandit algorithm with k actions corresponding to the vertices of the graph.

Without further assumptions, this is a finite-armed bandit, which for large
k is hopeless without further assumptions. Valko et al. [2014] assume the
rewards for well-connected vertices are similar – a kind of smoothness. Let
L = Q> ΛQ be the spectral decomposition of L and θ = Q> µ. Then,
1 X
Wij (µi − µj )2 = kθk2Λ .
2
i,j∈[k]

The left-hand side measures the variability of µ, weighted by connectivity of

the graph. Hence, assuming that kθkΛ is small corresponds to assuming that
µ changes only a little between well-connected vertices. Valko et al. [2014]
then let V0 = Λ + λI and analyse/implement the algorithm described in
Exercise 19.7. A more detailed exposition is by Valko [2016].

19.8 (Implementation) If the action set is the same in every round, then the
assumptions are satisfied for the various versions of UCB discussed in Chapters 7
 19.6 Exercises 252

LinUCB 4,000 LinUCB

40 UCB UCB
Expected regret

30 2,000

20
0
0 0.2 0.4 0.6 0.8 1 0 200 400 600 800 1,000
∆ k

Figure 19.1 The plot on the left compares the regret of UCB (Algorithm 6) and LinUCB
on a Gaussian bandit with k = 2, n = 1000 and varying suboptimality gaps ∆. The plot
on the right compares the same algorithms on a linear bandit with actions uniformly
distributed on the sphere and with d = 5 and n = 5000. The parameter θ is also
uniformly generated on the sphere.

to 9. How does LinUCB compare to UCB? Implement the version of LinUCB

using the value of βt given in Eq. (19.8) and/or Eq. (19.14) and compare it the
version of UCB given in Algorithm 6. In particular:
(a) Compare LinUCB with UCB on the 2-armed bandit with n = 1000 where the
reward distributions are Gaussian with unit variance and mean µ = (0, −∆)
where ∆ ∈ [0, 1], which for LinUCB corresponds to using At = {e1 , e2 } with
d = 2.
(b) Now compare LinUCB with UCB on k-armed stochastic linear bandits
where the d = 5 and At = A is composed of k unit vectors sampled from
the uniform distribution on the sphere (sample these vectors once). The
unknown parameter θ should also lie on the unit sphere and the noise should
be standard Gaussian. Plot the expected regret as a function of k ranging
from 2 to 1000 with a horizon of n = 5000.
(c) What conclusions can you draw from the experimental results you obtained?
(d) Show how one can use the Sherman-Morrison formula to implement
LinUCB using O(kd2 ) computation cost per round.
For parts (a) and (b) you should produce something comparable to Fig. 19.1.
20 Confidence Bounds for Least
Squares Estimators

In the last chapter, we derived a regret bound for a version of the upper confidence
bound algorithm that depended on a particular kind of confidence set. The purpose
of this chapter is to justify these choices.
Suppose a bandit algorithm has chosen actions A1 , . . . , At ∈ Rd and received
the rewards X1 , . . . , Xt with Xs = hθ∗ , As i + ηs where ηs is zero-mean noise.
Recall from the previous chapter that the penalised least-squares estimate of θ∗
is the minimiser of
t
X
Lt (θ) = (Xs − hθ, As i)2 + λkθk22 ,
s=1

where λ ≥ 0 is the penalty factor. This is minimised by

t
X t
X
θ̂t = Vt (λ)−1 Xs As with Vt (λ) = λI + As A>
s . (20.1)
s=1 s=1

It is convenient for the remainder to abbreviate Vt = Vt (0). Designing confidence

sets for θ∗ when A1 , . . . , At have been chosen by a bandit algorithm is a
surprisingly delicate matter. The difficulty stems from the fact that the actions
are neither fixed nor independent but are intricately correlated via the rewards.
We spend the first section of this chapter building intuition by making some
simplifying assumptions. Eager readers may skip directly to Section 20.1. For the
rest of this section, we assume the following:

1 No regularisation: λ = 0 and Vt is invertible.

2 Independent subgaussian noise: (ηs )s are independent and 1-subgaussian.
3 Fixed design: A1 , . . . , At are deterministically chosen without the knowledge of
X1 , . . . , Xt .

None of these assumptions is plausible in the bandit setting, but the simplification
eases the analysis and provides insight.

The assumption that λ = 0 means that in this section, θ̂t is just the ordinary
least squares estimator of θ. The requirement that Vt be non-singular means
that (As )ts=1 must span Rd , and so t must be at least d.
 Confidence Bounds for Least Squares Estimators 254

Comparing θ∗ and θ̂t in the direction x ∈ Rd , we have

* t
+ * t
+
X X

hθ̂t − θ∗ , xi = x, Vt−1
As Xs − θ∗ = x, Vt −1
As As θ ∗ + η s − θ ∗
>

s=1 s=1
* t
+ t
X X

= x, Vt−1 As η s = x, Vt−1 As ηs .
s=1 s=1

Since (ηs )s are independent and 1-subgaussian, by Lemma 5.4 and Theorem 5.3,
 v 
D E uu X t

2
 
1
P  x, θ̂t − θ∗ ≥ t2 x, Vt−1 As log  ≤ δ.
s=1
δ

Pt
2
A little linear algebra shows that s=1 x, Vt−1 As = kxk2V −1 and so
t

s  !
2 1
P hθ̂t − θ∗ , xi ≥ 2kxkV −1 log ≤ δ. (20.2)
t δ

If we only care about confidence bounds for one or a few vectors x, we could stop
here. For large action sets (with more than Ω(2d ) actions), one approach is to
convert this bound to a bound on kθ̂t − θ∗ kVt . To begin this process, notice that
1/2
1/2 Vt (θ̂t − θ∗ )
kθ̂t − θ∗ kVt = hθ̂t − θ∗ , Vt Xi , where X = .
kθ̂t − θ∗ kVt
The problem is that X is random, while we have only proven (20.2) for
deterministic x. The standard way of addressing problems like this is to use
a covering argument. First we identify a finite set Cε ⊂ Rd such that whatever
value X takes, there exists some x ∈ Cε that is ε-close to X. Then a union
bound and a triangle inequality allows one to finish. By its definition, we have
kXk22 = X > X = 1, which means that X ∈ S d−1 = {x ∈ Rd : kxk2 = 1}. Using
that X ∈ S d−1 , we see it suffices to cover S d−1 . The following lemma provides
the necessary guarantees on the size of the covering set.

Lemma 20.1. There exists a set Cε ⊂ Rd with |Cε | ≤ (3/ε)d such that for all
x ∈ S d−1 there exists a y ∈ Cε with kx − yk2 ≤ ε.

The proof of this lemma requires a bit work, but nothing really deep is needed.
This work is deferred to Exercises 20.3 and 20.4. Let Cε be the covering set given
by the lemma, and define event
( s  )
D E |Cε |
1/2
E = exists x ∈ Cε : Vt x, θ̂t − θ∗ ≥ 2 log .
δ

1/2
Using the fact that kVt xkV −1 = kxk2 = 1, and a union bound combined with
t
Eq. (20.2) shows that P (E) ≤ δ. When E does not occur, Cauchy–Schwarz shows
 20.1 Martingales and the Method of Mixtures 255

that
D E
1/2
kθ̂t − θ∗ kVt = max Vt x, θ̂t − θ∗
x∈S d−1
hD E D Ei
1/2 1/2
= max min Vt (x − y), θ̂t − θ∗ + Vt y, θ̂t − θ∗
x∈S d−1 y∈Cε
" s  #
|Cε |
< max min kθ̂t − θ∗ kVt kx − yk2 + 2 log
x∈S d−1 y∈Cε δ
s  
|Cε |
≤ εkθ̂t − θ∗ kVt + 2 log .
δ
Rearranging yields
s  
1 |Cε |
kθ̂t − θ∗ kVt < 2 log .
1−ε δ
Now there is a tension in the choice of ε > 0. The term in the denominator
suggests that ε should be small, but by Lemma 20.1 the cardinality of Cε grows
rapidly as ε tends to zero. By lazily choosing ε = 1/2,
s   !
1
P kθ̂t − θ∗ kVt ≥ 2 2 d log(6) + log ≤ δ. (20.3)
δ

Except for constants and other minor differences, this turns out to be about as
good as you can get. Unfortunately, however, this analysis only works because
Vt was assumed to be deterministic. When the actions are chosen by a bandit
algorithm, this assumption does not hold, and the ideas need to be modified.

20.1 Martingales and the Method of Mixtures

We now remove the limiting assumptions in the previous section. Of course some
conditions are still required. For the remainder of this section the following is
assumed:
1 There exists a θ∗ ∈ Rd such that Xt = hθ∗ , At i + ηt for all t ≥ 1.
2 The noise is conditionally 1-subgaussian:
 2
α
for all α ∈ R and t ≥ 1, E [exp(αηt ) | Ft−1 ] ≤ exp a.s. , (20.4)
2
where Ft−1 is such that A1 , X1 , . . . , At−1 , Xt−1 , At are Ft−1 -measurable.
3 In addition, we assume that λ > 0.
The inclusion of At in the definition of Ft−1 allows the noise to depend on past
choices, including the most recent action. This is often essential, as the case of
Bernoulli rewards shows. We have now dropped the assumption that (At )∞ t=1 are
fixed in advance.
 20.1 Martingales and the Method of Mixtures 256

The assumption that λ > 0 ensures that Vt (λ) is invertible and allows us
to relax the requirement that the actions span Rd . Notice also that in this
section, we allow the interaction sequence to be infinitely long.

Since we want exponentially decaying tail probabilities, one is tempted to try

the Cramér–Chernoff method:
  h  i
P kθ̂t − θ∗ k2Vt (λ) ≥ u2 ≤ inf E exp αkθ̂t − θ∗ k2Vt (λ) − αu2 .
α>0

Sadly, we do not know how to bound this expectation. Can we still somehow use
the Cramér–Chernoff method? We take inspiration from looking at the special
Pt
case of λ = 0 one last time, assuming that Vt = s=1 As A>
s is invertible. Let

t
X
St = η s As .
s=1

Pt
Recall that θ̂t = Vt−1 s=1 Xs As = θ∗ + Vt−1 St . Hence,
 
1 2 1 2 1 2
kθ̂t − θ∗ kVt = kSt kV −1 = max hx, St i − kxkVt .
2 2 t x∈Rd 2

The point of the second equality is to separate the martingale (St )t from Vt at
the price of the introduction of a maximum. This second equality is a special
case of (Fenchel) duality. As we shall see later in Chapter 26, for sufficiently
nice convex functions f one can show that with an appropriate function f ∗ ,
for any x ∈ Rd from the domain of f , f (x) = supu∈Rd hu, xi − f ∗ (u). The
advantage of this is that for any fixed u, x appears in a linear fashion.

The next lemma shows that the exponential of the term inside the maximum
is a supermartingale even when λ > 0.

Lemma 20.2. For all x ∈ Rd the process Mt (x) = exp(hx, St i − 12 kxk2Vt (λ) ) is an
F-adapted non-negative supermartingale with M0 (x) ≤ 1.

Proof of Lemma 20.2 That Mt (x) is Ft -measurable for all t and that it
is nonnegative are immediate from the definition. We need to show that
E[Mt (x) | Ft−1 ] ≤ Mt−1 (x) almost surely. The fact that (ηt ) is conditionally
1-subgaussian means that
! !
hx, At i
2 kxk2At A>
E [exp (ηt hx, At i) | Ft−1 ] ≤ exp = exp t
a.s.
2 2
 20.1 Martingales and the Method of Mixtures 257

Hence
   
1 2

E[Mt (x) | Ft−1 ] = E exp hx, St i − kxkVt  Ft−1

2
   
1 2

= Mt−1 (x)E exp ηt hx, At i − kxkAt A>  Ft−1

2 t

≤ Mt−1 (x) a.s.

Finally, note that M0 (x) ≤ 1 is immediate.

For simplicity, consider now again the case when λ = 0. Combining the lemma
and the linearisation idea almost works. The Cramér–Chernoff method leads to
      
1 1
P kθ̂t − θ∗ k2Vt ≥ log(1/δ) = P exp max hx, St i − kxk2Vt ≥ 1/δ
2 x∈Rd 2
   
1
≤ δE exp max hx, St i − kxk2Vt
x∈Rd 2
 
= δE max Mt (x) . (20.5)
x∈Rd

Lemma 20.2 shows that E[Mt (x)] ≤ 1. This seems quite promising, but
the presence of the maximum is a setback because E[maxx∈Rd Mt (x)] ≥
maxx∈Rd E[Mt (x)], which is the wrong direction to be used above. This means
we cannot directly use the lemma to bound Eq. (20.5). There are two ways to
proceed. The first is to use a covering argument over possible near-maximisers
of x, which eventually works. A more elegant way is to take inspiration from
Eq. (20.5) and use Laplace’s method for approximating integrals of well-behaved
exponentials, as we now explain.

20.1.1 Laplace’s Method ( )

We briefly review Laplace’s method for one-dimensional functions. Assume that
f : [a, b] → R is twice differentiable and has a unique maximum at x0 ∈ (a, b)
with −q = f 00 (x0 ) < 0. Laplace’s method for approximating f (x0 ) is to compute
the integral
Z b
Is = exp(sf (x))dx
a

for some large value of s > 0. From a Taylor expansion, we may write
q
f (x) = f (x0 ) − (x − x0 )2 + R(x) ,
2
where R(x) = o((x − x0 )2 ). Under appropriate technical assumptions,
Z b  
sq(x − x0 )2
Is ∼ exp(sf (x0 )) exp − dx as s → ∞ .
a 2
 20.1 Martingales and the Method of Mixtures 258

1 1

s=1 s=3

0 0
−5 0 5 −5 0 5

Figure 20.1 The plots depict Laplace’s approximation with f (x) = cos(x) exp(−x2 /20),
which is maximised at x0 = 0 and has q = −f 00 (x0 ) = 11/10. The solid line is a plot of
exp(sf (x))/ exp(sf (x0 )), and the dotted line is exp(−sq(x − x0 )2 ).

Furthermore, as s gets large,

Z b   Z ∞   r
sq(x − x0 )2 sq(x − x0 )2 2π
exp − dx ∼ exp − dx =
a 2 −∞ 2 sq
and hence
r
2π
Is ∼ exp(sf (x0 )) .
sq
It should also be clear that the fact that we integrate with respect to the Lebesgue
measure does not matter much. We could have integrated with respect to any
other measure as long as that measure puts a positive mass on the neighbourhood
of the maximiser. The method is illustrated in Fig. 20.1. The take-home message
is that if we integrate the exponential of a function that has a pronounced
maximum, then we can expect that the integral will be close to the exponential
function of the maximum.

20.1.2 Method of Mixtures

Laplace’s approximation suggests that
Z
max Mt (x) ≈ Mt (x)dh(x) , (20.6)
x Rd

where h is some measure on Rd chosen so that the integral can be calculated

in closed form. This is not a requirement of the method, but it does make the
argument shorter. The main benefit of replacing the maximum with an integral
is that we obtain the following lemma, which you will prove in Exercise 20.5.
R
Lemma 20.3. Let h be a probability measure on Rd ; then, M̄t = Rd Mt (x)dh(x)
is an F-adapted non-negative supermartingale with M̄0 = 1.
 20.1 Martingales and the Method of Mixtures 259

The following theorem is the key result from which the confidence set will be
derived.

Theorem 20.4. For all λ > 0 and δ ∈ (0, 1),

    
2 1 det(Vt (λ))
P exists t ∈ N : kSt kVt (λ)−1 ≥ 2 log + log ≤ δ.
δ λd
The proof will be given momentarily. First, though, the implications.

Theorem 20.5. Let δ ∈ (0, 1). Then, with probability at least 1 − δ, it holds that
for all t ∈ N,
s    
√ 1 det Vt (λ)
kθ̂t − θ∗ kVt (λ) < λkθ∗ k2 + 2 log + log .
δ λd

Furthermore, if kθ∗ k2 ≤ m2 , then P (exists t ∈ N+ : θ∗ ∈

/ Ct ) ≤ δ with
( s    )
√ 1 det Vt−1 (λ)
Ct = θ ∈ R : kθ̂t−1 − θkVt−1 (λ) < m2 λ + 2 log
d
+ log .
δ λd

Proof We only have to compare kSt kVt (λ)−1 and kθ̂t − θ∗ kVt (λ) :

kθ̂t − θ∗ kVt (λ) = kVt (λ)−1 St + (Vt (λ)−1 Vt − I)θ∗ kVt (λ)
≤ kSt kVt (λ)−1 + (θ∗> (Vt (λ)−1 Vt − I)Vt (λ)(Vt (λ)−1 Vt − I)θ∗ )1/2
= kSt kVt (λ)−1 + λ1/2 (θ∗> (I − Vt (λ)−1 Vt )θ∗ )1/2
≤ kSt kVt (λ)−1 + λ1/2 kθ∗ k ,

and the result follows from Theorem 20.4.

Proof of Theorem 20.4 Let H = λI ∈ Rd×d and h = N (0, H −1 ) and

Z
M̄t = Mt (x)dh(x)
Rd
Z  
1 1 1
=p exp hx, St i − kxk2Vt − kxk2H dx .
(2π)d det(H −1 ) Rd 2 2

By Lemma 20.3, M̄t is a non-negative supermartingale, and thus the maximal

inequality (Theorem 3.9) shows that
    
1 1
P sup log(M̄t ) ≥ log = P sup M̄t ≥ ≤ δ. (20.7)
t∈N δ t∈N δ

Now we turn to studying M̄t . Completing the square in the definition of M̄t
we get
1 1 1 1
hx, St i − kxk2Vt − kxk2H = kSt k2(H+Vt )−1 − kx − (H + Vt )−1 St k2H+Vt .
2 2 2 2
The first term kSt k2(H+Vt )−1 does not depend on x and can be moved outside the
 20.2 Notes 260

integral, which leaves a quadratic ‘Gaussian’ term that may be integrated exactly
and results in
 1/2  
det(H) 1 2
M̄t = exp kSt k(H+Vt )−1 . (20.8)
det(H + Vt ) 2
The result follows by substituting this expression into Eq. (20.7) and
rearranging.

20.2 Notes

1 Recall from the previous chapter that when kAt k2 ≤ L is assumed, then
  d  d
det Vt (λ) Vt (λ) nL2
≤ trace ≤ 1+ . (20.9)
λd λd λd
In general, the log determinant form should be preferred when confidence
intervals are used as part of an algorithm, but the right-hand side has a
concrete form that can be useful when stating regret bounds.
2 Plugging the bounds of the previous note into Theorem 20.5 and choosing
λ = 1 gives the confidence set
( s    )
1 nL 2
Ct = θ ∈ Rd : kθ̂t−1 − θkVt−1 (1) < m2 + 2 log + d log 1 + .
δ d
p
The dependence of the radius on n, d and δ, up to constants and a log(n)
factor, is the same as what we got in the fixed design case (cf. Eq. (20.3)),
which suggests that Theorem 20.5 can be quite tight. By considering the case
when each basis vector {e1 , . . . , ed } is played m times, then D = kθ̂t − θk2Vt
is distributed like a chi-squared distribution with d degrees of freedom. From
this, we see that the first term under the square root with the coefficient two
is stemming from variance of the noise, while the term that involved d log(n) √
is the bias (the expected value of D). In particular, this shows that the d
factor cannot be avoided.
3 If either of the above confidence sets is used (either the one from the theorem,
or that from Eq. (20.3)) to derive confidence bounds for the prediction error
h√θ̂t − θ, xi at some fixed x ∈ Rd , we get a confidence width that scales with
d (e.g., Eq. (19.13)), unlike the confidence width in Eq. (20.2), which is
independent of d. It follows that if one is interested in high-probability bounds
for the mean at a fixed input x, one should avoid going through a confidence
set for the whole parameter vector. What this leaves open is whether a bound
like in Eq. (20.2) is possible at a fixed input x, but with a sequential design.
In Exercise 20.2 you will answer this question in the negative. First note that
when the actions are chosen using a fixed design, integrating Eq. (20.2) shows
 20.3 Bibliographic Remarks 261

that E[hθ̂t − θ∗ , xi2 /kxk2V −1 ] = O(1). In the exercise, you will show that there
t
exists a sequential design such that
h i
E hθ̂t − θ∗ , xi2 /kxk2V −1 = Ω(d) ,
t
√
showing that for some sequential designs the factor d is necessary. It remains
an interesting open question to design confidence bounds for sequential design
for fixed x that adapts to the amount of dependence in the design.
4 Supermartingales arise naturally in proofs relying on the Cramér–Chernoff
method. Just one example is the proof of Lemma 12.2. One could rewrite
most of the proofs involving sums of random variables relying on the Cramér–
Chernoff method in a way that it would become clear that the proof hinges on
the supermartingale property of an appropriate sequence.

20.3 Bibliographic Remarks

Bounds like those given in Theorem 20.5 are called self-normalised bounds [de la
Peña et al., 2008]. The method of mixtures goes back to the work by Robbins
and Siegmund [1970]. In practice, the improvement provided by the method of
mixtures relative to the covering arguments is quite large. A historical account
of martingale methods in sequential analysis is by Lai [2009]. A simple proof of
Lemma 20.1 appears as lemma 2.5 in the book by van de Geer [2000]. Calculating
covering numbers (or related packing numbers) is a whole field by itself, with
open questions even in the most obvious examples. The main reference is by
Rogers [1964], which by now is a little old, but still interesting.

20.4 Exercises

20.1 (Lower bounds for fixed design) Let n = md for integer m and
A1 , . . . , An be a fixed design where each basis vector in {e1 , . . . , ed } is played
exactly m times. Then let (ηt )nt=1 be a sequence of independent standard Gaussian
random variables and Xt = hθ∗ , At i + ηt . Finally, let θ̂n be the ordinary least
squares estimator of θ∗ ∈ Rd . Show that
h i
E kθ̂n − θ∗ k2Vn = d .

This exercise shows that the d-dependence in Eq. (20.3) is unavoidable in

general for a self-normalised bound, even in the fixed design setting.

20.2 (Lower bounds for sequential design) Let n ≥ 2d and (ηt )nt=1 be a
sequence of independent standard Gaussian random variables. Find a sequence of
 20.4 Exercises 262

Pn
random vectors (At )nt=1 , with At ∈ Rd such that Vn = t=1 At A> t is invertible
almost surely and At is σ(A1 , η1 , . . . , At−1 , ηt−1 )-measurable for all t and
h i
E hθ̂n , 1i2 /k1k2V −1 ≥ cd ,
n

Pn
where c > 0 is a universal constant and Sn = t=1 ηt At and θ̂n = Vn−1 Sn .

Hint Choose A1 , . . . , Ad to be the standard basis vectors. Subsequently choose

selected basis vectors adaptively to push the estimate of hθ̂n , 1i away from zero.
For Exercise 20.4, where we ask you to prove Lemma 20.1, a few standard
definitions will be useful.

Definition 20.6 (Covering and packing). Let A ⊂ Rd . A subset C ⊂ A is said to

be an ε-cover of A if A ⊂ ∪x∈C B(x, ε), where B(x, ε) = {y ∈ Rd : kx − yk ≤ ε}
is the ε ball centered at x. An ε-packing of A is a subset P ⊂ A such that for any
x, y ∈ P, kx − yk > ε (note the strict inequality). The ε-covering number of A
is N (A, ε) = min{|C| : C is an ε-covering of A}, while the ε-packing number
of A is M (A, ε) = max{|P| : P is an ε-packing of A}, where we allow for both
the covering and packing numbers to take on the value of +∞.

The definitions can be repeated for pseudo-metric spaces. Let X be a set and
d : X × X → [0, ∞) be a function that is symmetric, satisfies the triangle
inequality and for which d(x, x) = 0 for all x ∈ X. Note that d(x, y) = 0 is
allowed for distinct x and y, so d need not be a metric. The basic results
concerning covering and packing stated in the next exercise remain valid
with this more general definition. In applications we often need the logarithm
of the covering and packing numbers, which are called the metric entropy
of X at scale ε. As we shall see, these are often close no matter whether we
consider packing or covering.

20.3 (Coverings and packings) Let A ⊂ Rd , B be the unit ball of Rd and

vol(·) the usual volume (measure under the Lebesgue measure). For brevity let
N (ε) = N (A, ε) and M (ε) = M (A, ε). Show that the following hold:

(a) ε → N (ε) is increasing as ε ≥ 0 is decreasing.

(b) M (2ε) ≤ N (ε) ≤ M (ε).
(c) We have
 d  d
1 vol(A) vol(A + 2ε B) (∗) vol( 32 A) 3 vol(A)
≤ N (ε) ≤ M (ε) ≤ ≤ ≤ ,
ε vol(B) vol( 2 B)
ε
vol( 2 B)
ε
ε vol(B)

where (∗) holds under the assumption that εB ⊂ A and that A is convex
and for U, V ⊂ Rd , c ∈ R, U + V = {u + v : u ∈ U , v ∈ V } and
cU = {cu : u ∈ U };
 20.4 Exercises 263

(d) Fix ε > 0. Then N (ε) < +∞ if and only if A is bounded. The same holds
for M (ε).

20.4 Use the results of the previous exercise to prove Lemma 20.1.

20.5 Prove Lemma 20.3.

Hint Use the ‘sections’ lemma [Kallenberg, 2002, Lemma 1.26] to established
that M̄t is Ft -measurable.
20.6 (Hoeffding–Azuma) Let X1 , . . . , Xn be a sequence of random variables
adapted to a filtration F = (Ft )t . Suppose that |Xt | ∈ [at , bt ] almost surely for
arbitrary fixed sequences (at ) and (bt ) with at ≤ bt for all t ∈ [n]. Show that for
any ε > 0,
!  
2n2 ε2
Xn
P (Xt − E[Xt | Ft−1 ]) ≥ ε ≤ exp − Pn .
t=1 (bt − at )
2
t=1

Hint It may help to recall Hoeffding’s lemma from Note 4 in Chapter 5, which
states that for a random variable X ∈ [a, b], the moment-generating function
satisfies

MX (λ) ≤ exp(λ2 (b − a)2 /8) .

20.7 (Extension of Hoeffding–Azuma) The following simple extension

of Hoeffding–Azuma is often useful. Let n ∈ N+ and (at ) and (bt ) be fixed
sequences with at ≤ bt for all t ∈ [n]. Let X1 , . . . , Xn be a sequence of random
variables adapted to a filtration F = (Ft )t and A be an event. Assume that
P (exists t ∈ [n] : A and Xt ∈
/ [at , bt ]) = 0 and ε > 0, and show that
!  
2n2 ε2
Xn
(a) P A ∩ (Xt − E[Xt | Ft−1 ]) ≥ ε ≤ exp − Pn .
t=1 (bt − at )
2
t=1
!  
2n2 ε2
Xn
(b) P (Xt − E[Xt | Ft−1 ]) ≥ ε ≤ P (A ) + exp − Pn
c
.
t=1 (bt − at )
2
t=1

The utility of this result comes from the fact that very often the range of
some adapted sequence is itself random and could be arbitrarily large with
low probability (when A does not hold). A reference for the above result is
the survey by McDiarmid [1998].

20.8 Let δ ∈ (0, 1) and F = (Ft )∞

t=1 be a filtration and (Xt )t=1 be F-adapted
∞

such that

for all λ ∈ R, E[exp(λXt ) | Ft−1 ] ≤ exp(λ2 σ 2 /2) a.s.

 20.4 Exercises 264

Pn
Let Sn = t=1 Xt . Show that
 v !
u √
u tσ 2 + 1 
P exists t : |St | ≥ t2σ 2 (t + 1) log
 ≤ δ.
δ

20.9 (Law of the iterated logarithm and method of mixtures) This

exercise uses the same notation as Exercise 20.8. Let f be a probability density
function supported on [0, ∞) and
Z ∞  
λ2 n
Mn = f (λ) exp λSn − dλ .
0 2

(a) Show that argmaxλ∈R λSn − λ2 n/2 = Sn /n.

(b) Suppose that f (λ) is decreasing for λ > 0. Show that for any ε > 0 and
Λn = Sn /n, that,
 
(1 − ε2 )S 2
Mn ≥ εΛn f (Λn (1 + ε)) exp
2n

(c) Use the previous result to show that for any δ ∈ (0, 1),
s     !
2n 1 1
P exists n : Sn ≥ inf log + log ≤ δ.
ε>0 (1 − ε2 ) δ εΛn f (Λn (1 + ε))
R∞
(d) Find an f such that 0
f (λ)dλ = 1 and f (λ) ≥ 0 for all λ ∈ R and
   
1 1
log = (1 + o(1)) log log
λf (λ) λ
as λ → 0.
(e) Use the previous results to show that
!
Sn
P lim sup p ≤1 = 1.
n→∞ 2n log log(n)

The last part of the previous exercises is one-half of the statement of the
law of iterated logarithm, which states that
Sn
lim sup p =1 almost surely .
n→∞ 2n log log(n)
In other words, the magnitude of the largest fluctuations of the partial sum
√
(Sn )n is almost surely of the order 2n log log n as n → ∞.

20.10 Let F = (Ft )nt=0 be a filtration and X1 , X2 , . . . , Xn be a sequence of

F-adapted random variables with Xt ∈ {−1, 0, 1} and µt = E[Xt | Ft−1 , Xt 6= 0],
 20.4 Exercises 265

which we define to be zero whenever P (Xt 6= 0 | Ft−1 ) = 0. Then, with St =

Pt Pt
s=1 (Xs − µs |Xs |) and Nt = s=1 |Xs |,
 s 
 √ 
c N
P exists t ≤ n : |St | ≥ 2Nt log and Nt > 0 ≤ δ ,
t
δ

where c > 0 is a universal constant.

This result
p appeared
√ in a paper by the authors and others with the constant
c = 4 2/π/ erf( 2) ≈ 3.43 [Lattimore et al., 2018].

20.11 (Sequential likelihood ratios and confidence sets) Let (Θ, G)

be a measurable space and (Pθ : θ ∈ Θ) be a probability kernel from (Θ, G)
to (R, B(R)). Assume there exists a common measure µ such that Pθ  µ for
all θ ∈ Θ, and let {pθ : θ ∈ Θ} be a family of densities with pθ = dPθ /dµ.
You may assume that pθ (x) is jointly measurable in θ and x. Such a choice
is guaranteed to exist, as we explain in Note 8 at the end of Chapter 34.
Fix θ∗ ∈ Θ and let (Xt )∞ t=1 be a sequence of independent random variables
with law Pθ∗ . Let θ̂t ∈ Θ be σ(X1 , . . . , Xt )-measurable. For θ ∈ Θ, define
Pt
Lt (θ) = s=1 log(pθ̂s−1 (Xs )/pθ (Xs )).


(a) Show that P supt≥1 Lt (θ∗ ) ≥ log(1/δ) ≤ δ for any δ ∈ (0, 1).
(b) Show that Θt = {θ : Lt (θ) < log(1/δ)} is a sequence of confidence sets such
that P (exists t ∈ N such that θ∗ 6∈ Θt ) ≤ δ.
(c) Let Θ = R and G = B(R) and Pθ = N (0, θ) and µ be the Lebesgue measure.
Pt
Then let θ̂t = 0 for t = 0 and 1t s=1 Xs otherwise. Write an expression
for Ct , and investigate how it compares to the usual confidence intervals for
Gaussian random variables.

Hint Use Cramér–Chernoff method and observe that (exp(Lt (θ∗ )))∞
t=1 is a
martingale.

The quantities pθ0 (Xs )/pθ (Xs ) are called likelihood ratios. That the
product of likelihood ratios forms a martingale is a cornerstone result of
classical parametric statistics. The sequential form that appears in the above
exercise is based on Lemma 2 of Lai and Robbins [1985], who cite Robbins
and Siegmund [1972] as the original source.
21 Optimal Design for Least Squares
Estimators

In the preceding chapter, we showed how to construct confidence intervals for

least squares estimators when the design is chosen sequentially. We now study
the problem of choosing actions for which the resulting confidence sets are small.
This plays an important role in the analysis of stochastic linear bandits with
finitely many arms (Chapter 22) and adversarial linear bandits (Part VI).

21.1 The Kiefer–Wolfowitz Theorem

Let η1 , . . . , ηn be a sequence of independent 1-subgaussian random variables and

a1 , . . . , an ∈ Rd be a fixed sequence with span(a1 , . . . , an ) = Rd and X1 , . . . , Xn
be given by Xt = hθ∗ , at i + ηt for some θ∗ ∈ Rd . The least squares estimator of
Pn Pn
θ∗ is θ̂ = V −1 t=1 at Xt with V = t=1 at a> t .

The least squares estimator used here is not regularised. This eases the
calculations, and the lack of regularisation will not harm us in future
applications.

Eq. (20.2) from Chapter 20 shows that for any a ∈ Rd and δ ∈ (0, 1),
s  !
2 1
P hθ̂ − θ∗ , ai ≥ 2kakV −1 log ≤ δ. (21.1)
δ

For our purposes, both a1 , . . . , an and a will be actions from some (possibly
infinite) set A ⊂ Rd and the question of interest is finding the shortest sequence
of exploratory actions a1 , . . . , an such that the confidence bound in the previous
display is smaller than some threshold for all a ∈ A. To solve this exactly
is likely an intractable exercise in integer programming. Finding an accurate
approximation turns out to be efficient for a broad class of action sets, however.
P
Let π : A → [0, 1] be a distribution on A so that a∈A π(a) = 1 and V (π) ∈ Rd×d
and g(π) ∈ R be given by
X
V (π) = π(a)aa> , g(π) = max kak2V (π)−1 . (21.2)
a∈A
a∈A
 21.1 The Kiefer–Wolfowitz Theorem 267

In the subfield of statistics called optimal experimental design, the distribution

π is called a design, and the problem of finding a design that minimises g is
called the G-optimal design problem. So how to use this? Suppose that π is
a design and a ∈ Supp(π) and
  
π(a)g(π) 1
na = 2
log . (21.3)
ε δ
Then, choosing each action a ∈ Supp(π) exactly na times ensures that
 
X g(π) 1
V = na aa ≥ 2 log
>
V (π) ,
ε δ
a∈Supp(π)

which by Eq. (21.1) means that for any a ∈ A, with probability 1 − δ,

s  
2 1
hθ̂ − θ∗ , ai ≤ 2kakV −1 log ≤ ε.
δ

By Eq. (21.3), the total number of actions required to ensure a confidence width
of no more than ε is bounded by
X X  π(a)g(π)  
1 g(π)
 
1
n= na = 2
log ≤ | Supp(π)| + 2 log .
ε δ ε δ
a∈Supp(π) a∈Supp(π)

The set Supp(π) is sometimes called the core set. The following theorem
characterises the size of the core set and the minimum of g.

Theorem 21.1 (Kiefer–Wolfowitz). Assume that A ⊂ Rd is compact and

span(A) = Rd . The following are equivalent:

(a) π ∗ is a minimiser of g.
(b) π ∗ is a maximiser of f (π) = log det V (π).
(c) g(π ∗ ) = d.

Furthermore, there exists a minimiser π ∗ of g such that | Supp(π ∗ )| ≤ d(d + 1)/2.

A design that maximises f is known as a D-optimal design, and thus the

theorem establishes the equivalence of G-optimal and D-optimal designs.

Proof We give the proof for finite A. The general case follows by passing to the
limit (Exercise 21.3). When it is convenient, distributions π on A are treated as
vectors in R|A| . You will show in Exercises 21.1 and 21.2 that f is concave and
that

(∇f (π))a = kak2V (π)−1 . (21.4)

Also notice that

!
X X
π(a)kak2V (π)−1 = trace π(a)aa V (π)
> −1
= trace(I) = d . (21.5)
a∈A a
 21.1 The Kiefer–Wolfowitz Theorem 268

(b)⇒(a): Suppose that π ∗ is a maximiser of f . By the first-order optimality

criterion (see Section 26.5), for any π distribution on A,
0 ≥ h∇f (π ∗ ), π − π ∗ i
X X
= π(a)kak2V (π∗ )−1 − π ∗ (a)kak2V (π∗ )−1
a∈A a∈A
X
= π(a)kak2V (π∗ )−1 − d.
a∈A

For an arbitrary a ∈ A, choosing π to be the Dirac at a ∈ A proves that

kak2V (π∗ )−1 ≤ d. Hence g(π ∗ ) ≤ d. Since g(π) ≥ d for all π by Eq. (21.5), it follows
that π ∗ is a minimiser of g and that minπ g(π) = d. (c) =⇒ (b): Suppose that
g(π ∗ ) = d. Then, for any π,
X
h∇f (π ∗ ), π − π ∗ i = π(a)kakV (π∗ )−1 − d ≤ 0 .
a∈A

And it follows that π is a maximiser of f by the first-order optimality conditions

∗

and the concavity of f . That (a) =⇒ (c) is now trivial. To prove the second
part of the theorem, let π ∗ be a minimiser of g, which by the previous part is a
maximiser of f . Let S = Supp(π ∗ ), and suppose that |S| > d(d + 1)/2. Since the
dimension of the subspace of d × d symmetric matrices is d(d + 1)/2, there must
be a non-zero function v : A → R with Supp(v) ⊆ S such that
X
v(a)aa> = 0 . (21.6)
a∈S

Notice that for any a ∈ S, the first-order optimality conditions ensure that
kak2V (π∗ )−1 = d (Exercise 21.5). Hence
X X
d v(a) = v(a)kak2V (π∗ )−1 = 0 ,
a∈S a∈S

where the last equality follows from Eq. (21.6). Let π(t) = π ∗ + tv and let
P
τ = max{t > 0 : π(t) ∈ PA }, which exists since v 6= 0 and a∈S v(a) = 0
and Supp(v) ⊆ S. By Eq. (21.6), V (π(t)) = V (π ∗ ), and hence f (π(τ )) = f (π ∗ ),
which means that π(τ ) also maximises f . The claim follows by checking that
| Supp(π(T ))| < | Supp(π ∗ )| and then using induction.

Geometric Interpretation
There is a geometric interpretation of the D-optimal design problem. Let π be a
P
D-optimal design for A and V = a∈A π(a)aa> and

E = x ∈ Rd : kxk2V −1 ≤ d ,
which is a centered ellipsoid. By Theorem 21.1, it holds that A ⊂ E with the core
set lying on the boundary (see Fig. 21.1). As you might guess from the figure,
the ellipsoid E is the minimum volume centered ellipsoid containing A. This is
known to be unique and the optimisation problem that characterises it is in fact
the dual of the log determinant problem that determines the D-optimal design.
 21.2 Notes 269

Figure 21.1 The minimum volume centered ellipsoid containing a point cloud. The points
on the boundary are the core set. The ellipse is E = {x : kxk2V (π)−1 = d}, where π is an
optimal design.

21.2 Notes

1 The letter ‘d’ in D-optimal design comes from the determinant in the objective.
The ‘g’ in G-optimal design stands for ‘globally optimal’. The names were coined
by Kiefer and Wolfowitz, though both problems appeared in the literature
before them.
2 In applications we seldom need an exact solution to the design problem. Finding
a distribution π such that g(π) ≤ (1 + ε)g(π ∗ ) will increase the regret of our
algorithms by a factor of just (1 + ε)1/2 .
3 The computation of an optimal design for finite action sets is a convex problem
for which there are numerous efficient approximation algorithms. The Frank–
Wolfe algorithm is one such algorithm, which can be used to find a near-optimal
solution for modestly sized problems. The algorithm starts with an initial π0
and updates according to

πk+1 (a) = (1 − γk )πk (a) + γk I {ak = a} , (21.7)

where ak = argmaxa∈A kak2V (πk )−1 and the step size is chosen to optimise f
along the line connecting πk and δak .
1 2
d kak kV (πk )−1 − 1
γk = argmaxγ∈[0,1] f ((1 − γ)πk + γδak ) = . (21.8)
kak k2V (πk )−1 − 1

If π0 is chosen to be the uniform distribution over A, then the number of

iterations before g(πk ) ≤ (1 + ε)g(π ∗ ) is at most O(d log log |A| + d/ε). For a
slightly more sophisticated choice of initialisation the dependence on |A| can
be eliminated entirely. More importantly, this other initialisation has a core
set of size O(d) and running the algorithm in Eq. (21.7) for just O(d log log d)
iterations is guaranteed to produce a design with g(π) ≤ 2g(π ∗ ) and a core set
of size O(d log log d).
 21.3 Bibliographic Remarks 270

4 If the action set is infinite, then approximately optimal designs can sometimes
still be found efficiently. Unfortunately the algorithms in the infinite case tend
to be much ‘heavier’ and less practical.
5 The smallest ellipsoid containing some set K ⊂ Rd is called the minimum
volume enclosing ellipsoid (MVEE) of K. As remarked, the D-optimal
design problem is equivalent to finding an MVEE of A with the added constraint
that the ellipsoid must be centred – or equivalently, finding the MVEE of the
symmetrised set A ∪ {−a : a ∈ A}. The MVEE of a convex set is also called
John’s ellipsoid, which has many applications in optimisation and beyond.
6 In Exercise 21.6, you will generalise Kiefer–Wolfowitz theorem to sets that do
not span Rd . When A is compact and dim(span(A)) = m ∈ [d], then there
exists a distribution π ∗ supported on at most m(m + 1)/2 points of A and for
which g(π ∗ ) = m = inf π g(π).

21.3 Bibliographic Remarks

The Kiefer–Wolfowitz theorem is due to Kiefer and Wolfowitz [1960]. The

algorithm in Note 3 is due to Fedorov [1972]. A similar variant was also proposed
by Wynn [1970], and the name Wynn’s method is sometimes used. The algorithm
is a specialisation of Frank–Wolfe’s algorithm, which was originally intended for
quadratic programming [Frank and Wolfe, 1956]. Todd [2016] wrote a nice book
about minimum volume ellipsoids and related algorithms, where you can also find
many more references and improvements to the basic algorithms. Chapter 3 of his
book also includes discussion of alternative initialisations and convergence rates
for various algorithms. The duality between D-optimal design and the MVEE
problem was shown by Silvey and Sibson [1972]. Although the connection between
minimum volume ellipsoids and experimental design is well known, previous
applications of these results to bandits used John’s theorem without appropriate
symmetrisation, which made the resulting arguments more cumbersome. For
more details on finding approximately optimal designs for infinite sets, see the
article by Hazan et al. [2016], references there-in and the book by Grötschel et al.
[2012].

21.4 Exercises

21.1 (Derivative of log determinant) Prove the correctness of the derivative

in Eq. (21.4).

Hint For square matrix A let adj(A) be the transpose of the cofactor matrix
of A. Use the facts that the inverse of a matrix A is A−1 = adj(A)> / det(A) and
 21.4 Exercises 271

that if A : R → Rd×d , then

 
d d
det(A(t)) = trace adj(A) A(t) .
dt dt
21.2 (Concavity of log determinant) Prove that H 7→ log det(H) is concave
where H is a symmetric, positive definite matrix.

Hint Consider t 7→ log det(H + tZ) for Z symmetric, and show that this is a
concave function.
21.3 (Kiefer–Wolfowitz for compact sets) Generalise the proof of
Theorem 21.1 to compact action sets.

21.4 Prove the second inequality in Eq. (21.8).

21.5 Let π ∗ be a G-optimal design and a ∈ Supp(π ∗ ). Prove that kak2V (π∗ )−1 = d.

21.6 Prove that if A is compact and dim(span(A)) = m ∈ [d], then there exists
a distribution π ∗ over A supported on at most m(m + 1)/2 points and for which
g(π ∗ ) = m.

21.7 (Implementation) Write a program that accepts as parameters a finite

set A ⊂ Rd and returns a design π : A → [0, 1] such that g(π) ≤ d + ε for some
given ε > 0. How robust is your algorithm? Experiment with different choices of
A and d, and report your results.

Hint The easiest pure way to do this is to implement the Frank–Wolfe algorithm
described in Note 3. All quantities can be updated incrementally using rank-one
update formulas, and this will lead to a significant speedup. You might like to
read the third chapter of the book by Todd [2016] and experiment with the
proposed variants.
22 Stochastic Linear Bandits with
Finitely Many Arms

The optimal design problem from the previous chapter has immediate applications
to stochastic linear bandits. In Chapter 19, we developed a linear version of the
√
upper confidence bound algorithm that achieves a regret of Rn = O(d n log(n)).
The only required assumptions were that the sequence of available action sets
were bounded. In this short chapter, we consider a more restricted setting where:

1 the set of actions available in round t is A ⊂ Rd and |A| = k for some natural
number k;
2 the reward is Xt = hθ∗ , At i + ηt where ηt is conditionally 1-subgaussian:

E[exp(ληt )|A1 , η1 , . . . , At−1 ] ≤ exp(λ2 /2) almost surely for all λ ∈ R; and

3 the suboptimality gaps satisfy ∆a = maxb∈A hθ∗ , b − ai ≤ 1 for all a ∈ A.

The key difference relative to Chapter 19 is that now the set of actions is finite
and does not change with time. Under these conditions, it becomes possible to
design a policy such that
p 
Rn = O dn log(nk) .

For moderately sized k, this bound improves the regret by a factor of d1/2 , which
in some regimes is large enough to be worth the effort. The policy is an instance
of phase-based elimination algorithms. As usual, at the end of a phase, arms that
are likely to be suboptimal with a gap exceeding the current target are eliminated.
In fact, this elimination is the only way the data collected in a phase is being
used. In particular, the actions to be played during a phase are chosen based
entirely on the data from previous phases: the data collected in the present phase
do not influence which actions are played. This decoupling allows us to make use
of the tighter confidence bounds available in the fixed design setting, as discussed
in the previous chapter. The choice of policy within each phase uses the solution
to an optimal design problem to minimise the number of required samples to
eliminate arms that are far from optimal.

Theorem 22.1. With probability at least 1 − δ, the regret of Algorithm 12 satisfies

s  
k log(n)
Rn ≤ C nd log ,
δ
 22.1 Notes 273

Input A ⊂ Rd and δ
Step 0 Set ` = 1 and let A1 = A
Step 1 Let t` = t be the current timestep and find G-optimal design π` ∈ P(A` )
with Supp(π` ) ≤ d(d + 1)/2 that maximises
X
log det V (π` ) subject to π` (a) = 1
a∈A`

Step 2 Let ε` = 2−` and

  
2dπ` (a) k`(` + 1) X
T` (a) = 2 log and T` = T` (a)
ε` δ
a∈A`

Step 3 Choose each action a ∈ A` exactly T` (a) times

Step 4 Calculate the empirical estimate:
+T`
t`X X
θ̂` = V`−1 At Xt with V` = T` (a)aa>
t=t` a∈A`

Step 5 Eliminate low rewarding arms:

 
A`+1 = a ∈ A` : maxhθ̂` , b − ai ≤ 2ε` .
b∈A`

Step 6 ` ← ` + 1 and Goto Step 1

Algorithm 12: Phased elimination with G-optimal exploration.

p
where C > 0 is a universal constant. If δ = O(1/n), then E[Rn ] ≤ C nd log(kn)
for an appropriately chosen universal constant C > 0.

The proof of this theorem follows relatively directly from the high-probability
correctness of the confidence intervals used to eliminate low-rewarding arms. We
leave the details to the reader in Exercise 22.1.

22.1 Notes

1 The assumption that the action set does not change is crucial for Algorithm 12.
Several complicated algorithms have been proposed and analysed for the case
where At is allowed to change from round to round under the assumption that
|At | ≤ k for all rounds. For these algorithms, it has been proven that
q 
Rn = O nd log3 (nk) . (22.1)

When k is small, these

√ results improve on the bound for LinUCB in Chapter 19
by a factor of up to d.
2 Algorithm 12 can be adapted to the case where k is infinite by using confidence
 22.2 Bibliographic Remarks 274

intervals derived in Chapter 20. Once the dust has settled, you should find the
regret is
 p 
Rn = O d n log(n) .

3 One advantage of Algorithm 12 is that it behaves well even when the linear
model is misspecified. Suppose the reward is Xt = hθ, At i + ηt + f (At ), where
ηt is noise as usual and f : A → R is some function with kf k∞ ≤ ε. Then the
regret of Algorithm 12 can be shown to be
p √ 
Rn = O dn log(nk) + nε d log(n) .

The linear dependence

√ on the horizon should be expected when k is large. The
presence of d in the second term is unfortunate, but unavoidable in many
regimes as discussed by Lattimore and Szepesvári [2019b].

22.2 Bibliographic Remarks

The algorithms achieving Eq. (22.1) for changing action sets are SupLinRel [Auer,
2002] and SupLinUCB [Chu et al., 2011]. Both introduce phases to decouple
the dependence of the design on the outcomes. Unfortunately the analysis of
these algorithms is long and technical, which prohibited us from presenting
the ideas here. These algorithms are also not the most practical relative to
LinUCB (Chapter 19) or Thompson sampling (Chapter 36). Of course this does
not diminish the theoretical breakthrough. Phased elimination algorithms have
appeared in many places, but the most similar to the algorithm presented here
is the work on spectral bandits by Valko et al. [2014] (and we have also met
them briefly in earlier chapters on finite-armed bandits). None of the works just
mentioned used the Kiefer–Wolfowitz theorem. This idea is apparently new, but
it is based on the literature on adversarial linear bandits where John’s ellipsoid
has been used to define exploration policies [Bubeck et al., 2012]. For more details
on adversarial linear bandits, read on to Part VI.
Ghosh et al. [2017] address misspecified (stochastic) linear bandits with a fixed
action set. In misspecified linear bandits, the reward is nearly a linear function of
the feature vectors associated with the actions. Ghosh et al. [2017] demonstrate
that in the favourable case when one can cheaply test linearity, an algorithm
that first runs a test and then switches to either
√ a linear bandit or a finite-armed
√
bandit based on the outcome will achieve ( k ∧ d) n regret up to log factors.
We will return to misspecified linear bandits a few more times in the book.

22.3 Exercises

22.1 In this exercise, you will prove Theorem 22.1.

 22.3 Exercises 275

(a) Use Theorem 21.1 to show that the length of the `th phase is bounded by
 
2d k`(` + 1) d(d + 1)
T` ≤ 2 log + .
ε` δ 2
(b) Let a∗ ∈ argmaxa∈A hθ∗ , ai be the optimal arm and use Eq. (21.1) to show
that
δ
P (exists phase ` such that a∗ ∈
/ A` ) ≤ .
k
(c) For action a define `a = min{` : 2ε` < ∆a } to be the first phase where the
suboptimality gap of arm a is smaller than 2ε` . Show that
δ
P (a ∈ A`a ) ≤ .
k
(d) Show that with probability at least 1 − δ the regret is bounded by
s  
k log(n)
Rn ≤ C dn log ,
δ
where C > 0 is a universal constant.
(e) Show that this implies Theorem 22.1 for the given choice of δ.

22.2 (Misspecified linear bandits) Assume the reward satisfies Xt =

hθ, At i + ηt + f (At ), where ηt is 1-subgaussian noise as usual and f : A → R is
some function with kf k∞ ≤ ε, show that the expected regret of Algorithm 12
with the choice δ = 1/n is
p √ 
Rn = O dn log(nk) + nε d log(n) .
23 Stochastic Linear Bandits with
Sparsity

In Chapter 19 we showed the linear variant of UCB has regret bounded by

√
Rn = O(d n log(n)) ,
which for fixed finite action sets can be improved to
p
Rn = O( dn log(nk)) .
For moderately sized action sets, these approaches lead to a big improvement
over what could be obtained by using the policies that do not make use of the
linear structure.
The situation is still not perfect, though. In typical applications, the features
are chosen by the user of the system, and one can easily imagine there are many
candidate features and limited information about which will be most useful. This
presents the user with a challenging trade-off. If they include many features, then
d will be large, and the algorithm may be slow to learn. But if a useful feature is
omitted, then the linear model will almost certainly be quite wrong. Ideally, one
should be able to add features without suffering much additional regret if the
added feature does not contribute in a significant way. This can be captured by
the notion of sparsity, which is the central theme of this chapter.

23.1 Sparse Linear Stochastic Bandits

Like in the standard stochastic linear bandit setting, at the beginning of round t,
the learner receives a decision set At ⊂ Rd . They then choose an action At ∈ At
and receive a reward
Xt = hθ∗ , At i + ηt , (23.1)
where (ηt )t is zero-mean noise and θ∗ ∈ Rd is an unknown vector. The only
difference in the sparse setting is that the parameter vector θ∗ is assumed to have
many zero entries. For θ ∈ Rd let
d
X
kθk0 = I {θi 6= 0} ,
i=1

which is sometimes called the zero-‘norm’ (quotations because it is not really a

norm; see Exercise 23.1).
 23.2 Elimination on the Hypercube 277

Assumption 23.1. The following hold:

(a) (Sparse parameter) There exist known constants m0 and m2 such that
kθ∗ k0 ≤ m0 and kθ∗ k2 ≤ m2 .
(b) (Bounded mean rewards) hθ∗ , ai ≤ 1 for all a ∈ At and all rounds t.
(c) (Subgaussian noise) The noise is conditionally 1-subgaussian:
for all λ ∈ R, E[exp(ληt ) | Ft−1 ] ≤ exp(λ2 /2) a.s. ,
where Ft = σ(A1 , X1 , . . . , At , Xt , At+1 ).
Much ink has been spilled on what can be said about the speed of learning in
linear models like (23.1) when (At )t are passively generated and the parameter
vector is known to be sparse. Most results are phrased about recovering θ∗ , but
there also exist a few results that quantify the error when predicting Xt . The
ideal outcome would be that the learning speed depends mostly on m0 , with only
a mild dependence on d. Almost all the results come under the assumption that
the covariance matrix of the actions (At )t is well conditioned.

The condition number of a positive definite matrix A is the ratio of its

largest and smallest eigenvalues. A matrix is well conditioned if it has a
small condition number.

The details are a bit more complicated than just the conditioning, but the
main point is that the usual assumptions imposed on the covariance matrix of
the actions for passive learning are never satisfied when the actions are chosen
by a good bandit policy. The reason is simple. Bandit algorithms want to choose
the optimal action as often as possible, which means the covariance matrix will
have an eigenvector that points (approximately) towards the optimal action with
a large corresponding eigenvalue. We need some approach that does not rely on
such strong assumptions.

23.2 Elimination on the Hypercube

As a warm-up, consider the case where the action set is the d-dimensional
hypercube: At = A = [−1, 1]d . To reduce clutter, we denote the true parameter
vector by θ = θ∗ . The hypercube is notable as an action set because it enjoys
perfect separability. For each dimension i ∈ [d], the value of Ati ∈ [−1, 1] can
be chosen independently of Atj for j 6= i. Because of this, the optimal action is
a∗ = sign(θ), where

1 , if θi > 0 ;


sign(θ)i = sign(θi ) = 0 , if θi = 0 ;


−1 , if θ < 0 .
i
 23.2 Elimination on the Hypercube 278

So learning the optimal action amounts to learning the sign of θi for each
dimension. A disadvantage of this structure is that in the worst case the sign
of each θi must be learned independently, which in Chapter 24 we show leads
√
to a worst-case regret of Rn = Ω(d n). On the positive side, the separability
means that θi can be estimated in each dimension independently while paying
absolutely no price for this experimentation when θi = 0. It turns out that this
√
allows us to design a policy for which Rn = O(kθk0 n), even without knowing
the value of kθk0 .
Let Gt = σ(A1 , X1 , . . . , At , Xt ) be the σ-algebra containing information up to
time t − 1 (this differs from Ft , which also includes information about the action
chosen). Now suppose that (Ati )di=1 are chosen to be conditionally independent
given Gt−1 , and further assume for some specific i ∈ [d] that Ati is sampled from
a Rademacher distribution so that P (Ati = 1 | Gt−1 ) = P (Ati = −1 | Gt−1 ) = 1/2.
Then
   
d 
X 
E[Ati Xt | Gt−1 ] = E Ati  Atj θj + ηt   Gt−1 
j=1 
X
= θi E[A2ti | Gt−1 ] + θj E[Atj Ati | Gt−1 ] + E[Ati ηt | Gt−1 ]
j6=i

= θi ,

where the first equality is the definition of Xt = hθ, At i+ηt , the second by linearity
of expectation and the third by the conditional independence of (Ati )i and the
fact that E[Ati | Gt−1 ] = 0 and E[A2ti | Gt−1 ] = 1. This looks quite promising, but
we should also check the variance. Using our assumptions that (ηt ) is conditionally
1-subgaussian and that hθ, ai ≤ 1 for all actions a, we have
2
V[Ati Xt | Gt−1 ] = E[A2ti Xt2 | Gt−1 ] − θi2 = E[(hθ, At i + ηt ) | Gt−1 ] − θi2 ≤ 2 .
(23.2)
And now we have cause for celebration. The value of θi can be estimated by
choosing Ati to be a Rademacher random variable independent of the choices in
other dimensions. All the policy does is treat all dimensions independently. For a
particular dimension (say i), it explores by choosing Ati ∈ {−1, 1} uniformly at
random until its estimate is sufficiently accurate to commit to either Ati = 1 or
Ati = −1 for all future rounds. How long this takes depends on |θi |, but note that
if |θi | is small, then the price of exploring is also limited. The policy that results
from this idea is called selective explore-then-commit (Algorithm 13, SETC).

Theorem 23.2. There exists a universal constants C, C 0 > 0 such that the regret
of SETC satisfies
X log(n) p
Rn ≤ 3kθk1 + C and Rn ≤ 3kθk1 + C 0 kθk0 n log(n) .
|θi |
i:θi 6=0

By appealing to the central limit theorem and the variance calculation in

 23.2 Elimination on the Hypercube 279

1: Input n and d
2: Set E1i = 1 and C1i = R for all i ∈ [d]
3: for t = 1, . . . , n do
4: For each i ∈ [d] sample Bti ∼ Rademacher
5: Choose action:

Bti if 0 ∈ Cti


(∀i) Ati = 1 if Cti ⊂ (0, ∞]


−1 if C ⊂ [−∞, 0) .
ti

6: Play At and observe Xt

7: Construct empirical estimators:
t
X Pt
Esi Asi Xs
(∀i) Ti (t) = Esi θ̂ti = s=1

s=1
Ti (t)

8: Construct confidence intervals:

s 
1 1  p 
(∀i) Wti = 2 + log n 2T (t) + 1
Ti (t) Ti (t)2
i

h i
(∀i) Ct+1,i = θ̂ti − Wti , θ̂ti + Wti

9: Update exploration parameters:

(
0 if 0 ∈
/ Ct+1,i or Eti = 0
(∀i) Et+1,i =
1 otherwise .
10: end for
Algorithm 13: Selective explore-then-commit.

Eq. (23.2), we should be hopeful that the confidence intervals used by the
algorithm are sufficiently large to contain the true θi with high probability, but
this still needs to be proven.

Lemma 23.3. Define τi = n ∧ max{t : Eti = 1}, and let Fi = I {θi ∈ / Cτi +1,i } be
the event that θi is not in the confidence interval constructed at time τi . Then
P (Fi ) ≤ 1/n.

The proof of Lemma 23.3 is left until after the proof of Theorem 23.2.

Proof of Theorem 23.2 Recalling the definition of the regret and using the
fact that the optimal action is a∗ = sign(θ), we have the following regret
 23.2 Elimination on the Hypercube 280

decomposition:

n
" # d
" n #!
X X X
Rn = maxhθ, ai − E hθ, At i = n|θi | − E Ati θi . (23.3)
a∈A
t=1 i=1 t=1
| {z }
Rni

Clearly, if θi = 0, then Rni = 0. And so it suffices to bound Rni for each i

with |θi | > 0. Suppose that |θi | > 0 for some i and the failure event Fi given
in Lemma 23.3 does not occur. Then θi ∈ Cτi +1,t , and by the definition of the
algorithm, Ati = sign(θi ) for all t ≥ τi . Therefore,
" n
# " n
#
X X
Rni = n|θi | − E Ati θi = E |θi | (1 − Ati sign(θi ))
t=1 t=1

≤ 2n|θi |P (Fi ) + |θi |E [I {Fic } τi ] . (23.4)

Since τi is the first round t when 0 ∈ / Ct+1,i it follows that if Fi does not occur,
then θi ∈ Cτi ,i and 0 ∈ Cτi ,i . Thus the width of the confidence interval Cτi ,i must
be at least |θi |, and so

s 
1 1 √

2Wτi −1,i = 4 + log n 2τi − 1 ≥ |θi | ,
τi − 1 (τi − 1)2

which after rearranging shows for some universal constant C > 0 that

C log(n)
I {Fic } (τi − 1) ≤ 1 + .
θi2

Combining this result with Eq. (23.4) leads to

C log(n)
Rni ≤ 2n|θi |P (Fi ) + |θi | + .
|θi |

Using Lemma 23.3 to bound P (Fi ) and substituting into the decomposition
Eq. (23.3) completes the proof of the first part. The second part is left as a treat
for you (Exercise 23.2).

P
Proof of Lemma 23.3 Let Sti = j6=i Atj θj and Zti = Ati ηt +Ati Sti . For t ≤ τi ,

1X
t
θ̂ti − θi = Zsi .
t s=1
 23.3 Online to Confidence Set Conversion 281

√
The next step is to show that Zti is conditionally 2-subgaussian for t ≤ τi :

E [exp(λZti ) | Gt−1 ] = E [E [exp(λZti ) | Ft−1 ] | Gt−1 ]

= E [exp(λAti Sti )E [exp(λAti ηt ) | Ft−1 ] | Gt−1 ]
  2 
λ 
≤ E exp(λAti Sti ) exp Gt−1
2 
 2
λ
= exp E [E [exp(λAti Sti ) | Gt−1 , Sti ] | Gt−1 ]
2
 2   2 2  
λ λ Sti 
≤ exp E exp  Gt−1
2 2
≤ exp(λ2 ) .

The first inequality used the fact that ηt is conditionally 1-subgaussian. The second-
to-last inequality follows because Ati is conditionally Rademacher for t ≤ τi ,
which is 1-subgaussian by Hoeffding’s lemma (5.11). The final inequality follows
because Sti ≤ kAt k∞ kθk1 ≤ 1. The result follows by applying the concentration
bound from Exercise 20.8.

23.3 Online to Confidence Set Conversion

A new plan is needed to relax the assumption that the action set is a hypercube.
The idea is to modify the ellipsoidal confidence set used in Chapter 19 to have a
smaller radius. We will see that modifying the algorithm in Chapter 19 to use
√
the smaller confidence intervals improves the regret to Rn = O( dpn log(n)).

Without assumptions
√ on the action set, one cannot hope to have a regret
smaller than O( dn). To see this, recall that d-armed bandits can be
represented as linear bandits with At = {e1 , . . . , ed }. For these problems,
Theorem 15.2 shows
√ that for any policy there exists a d-armed bandit for
which Rn = Ω( dn). Checking the proof reveals that when adapted to the
linear setting the parameter vector is 2-sparse.

The construction that follows makes use of a kind of duality between online
prediction and confidence sets. While we will only apply the idea to the sparse
linear case, the approach is generic.
The prediction problem considered is online linear prediction under the
squared loss. This is also known as online linear regression. The learner
interacts with an environment in a sequential manner where in each round
t ∈ N+ :

1 The environment chooses Xt ∈ R and At ∈ Rd in an arbitrary fashion.

2 The value of At is revealed to the learner (but not Xt ).
 23.3 Online to Confidence Set Conversion 282

3 The learner produces a real-valued prediction X̂t ∈ R in some way.

4 The environment reveals Xt to the learner and the loss is (Xt − X̂t )2 .
The regret of the learner relative to a linear predictor that uses the weights
θ ∈ Rd is
n
X n
X
ρn (θ) = (Xt − X̂t )2 − (Xt − hθ, At i)2 . (23.5)
t=1 t=1

We say that the learner enjoys a regret guarantee Bn relative to Θ ⊆ Rd if for

any strategy of the environment,
sup ρn (θ) ≤ Bn . (23.6)
θ∈Θ

The online learning literature has a number of powerful techniques for this
learning problem. Later we will give a specific result for the sparse case when
Θ = {x : kxk0 ≤ m0 }, but first we show how to use such a learning algorithm
to construct a confidence set. Take any learner for online linear regression, and
assume the environment generates Xt in a stochastic manner like in linear bandits:
Xt = hθ∗ , At i + ηt . (23.7)
Combining Eqs. (23.5) to (23.7) with elementary algebra,
n
X n
X
Qt = (X̂t − hθ∗ , At i)2 = ρn (θ∗ ) + 2 ηt (X̂t − hθ∗ , At i)
t=1 t=1
n
X
≤ Bn + 2 ηt (X̂t − hθ∗ , At i) , (23.8)
t=1

where the first equality serves as the definition of Qt . Let us now take stock for a
moment. If we could somehow remove the dependence on the noise ηt in the right-
hand side, then we could define a confidence set consisting of all θ that satisfy
the equation. Of course the noise has zero mean and is conditionally independent
of its multiplier, so the expectation of this term is zero. The fluctuations can be
controlled with high probability using a little concentration analysis. Let
t
X
Zt = ηs (X̂s − hθ∗ , As i) .
s=1

Since X̂t is chosen based on information available at the beginning of the round,
X̂t is Ft−1 -measurable, and so
for all λ ∈ R, E[exp(λ(Zt − Zt−1 )) | Ft−1 ] ≤ exp(λ2 σt2 /2) ,

where σt2 = (X̂t − hθ∗ , At i)2 . The uniform self-normalised tail bound
(Theorem 20.4) with λ = 1 implies that,
s  !
1 + Qt
P exists t ≥ 0 such that |Zt | ≥ (1 + Qt ) log ≤ δ.
δ2
 23.3 Online to Confidence Set Conversion 283

Provided this low-probability event does not occur, then from Eq. (23.8) we have
s  
1 + Qt
Qt ≤ Bt + 2 (1 + Qt ) log . (23.9)
δ2
While both sides depend on Qt , the left-hand side grows linearly, while the
right-hand side grows sublinearly in Qt . This means that the largest value of Qt
that satisfies the above inequality is finite. A tedious calculation then shows this
value must be less than
√ √ 
8 + 1 + Bt
βt (δ) = 1 + 2Bt + 32 log . (23.10)
δ
By piecing together the parts, we conclude that with probability at least 1 − δ
the following holds for all t:
t
X
Qt = (X̂s − hθ∗ , As i)2 ≤ βt (δ) .
s=1

We could define Ct+1 to be the set of all θ such that the above holds with θ∗
replaced by θ, but there is one additionally subtlety, which is that the resulting
Pt
confidence interval may be unbounded (think about the case that s=1 As A> s is
not invertible). In Chapter 19 we overcame this problem by regularising the least
squares estimator. Since we have assumed that kθ∗ k2 ≤ m2 , the previous display
implies that
t
X
kθ∗ k22 + (X̂s − hθ∗ , As i)2 ≤ m22 + βt (δ) .
s=1

All together, we have the following theorem:

Theorem 23.4. Let δ ∈ (0, 1) and assume that θ∗ ∈ Θ and supθ∈Θ ρt (θ) ≤ Bt . If
( t
)
X
2 2 2
Ct+1 = θ ∈ R : kθk2 +
d
(X̂s − hθ, As i) ≤ m2 + βt (δ) ,
s=1

then P (exists t ∈ N such that θ∗ 6∈ Ct+1 ) ≤ δ .

The confidence set in Theorem 23.4 is not in the most convenient form. By
Pt Pt
defining Vt = I + s=1 As A> s and St = s=1 As X̂s and θ̂t = Vt St and
−1

performing an algebraic calculation that we leave to the reader (see Exercise 23.5),
one can see that
t
X t
X
kθk22 + (X̂s − hθ, As i)2 = kθ − θ̂t k2Vt + (X̂s − hθ̂t , As i)2 + kθ̂t k22 . (23.11)
s=1 s=1

Using this, the confidence set can be rewritten in the familiar form of an ellipsoid:
( t
)
X
2 2 2 2 2
Ct+1 = θ ∈ R : kθ − θ̂t kVt ≤ m2 + βt (δ) − kθ̂t k2 −
d
(X̂s − hθ̂t , As i) .
s=1
 23.4 Sparse Online Linear Prediction 284

1: Input Online linear predictor and regret bound Bt , confidence parameter

δ ∈ (0, 1)
2: for t = 1, . . . , n do
3: Receive action set At
4: Computer confidence set:
( t−1
)
X
2 2 2
Ct = θ ∈ R : kθk2 +
d
(X̂s − hθ, As i) ≤ m2 + βt (δ)
s=1

5: Calculate optimistic action

At = argmaxa∈At maxhθ, ai
θ∈Ct

6: Feed At to the online linear predictor and obtain prediction X̂t

7: Play At and receive reward Xt
8: Feed Xt to online linear predictor as feedback
9: end for
Algorithm 14: Online linear predictor UCB (OLR-UCB).

It is not obvious that Ct+1 is not empty because the radius could be negative.
Theorem 23.4 shows, however, that with high probability θ∗ ∈ Ct+1 . At last we
have established all the conditions required for Theorem 19.2, which implies the
following theorem bounding the regret of Algorithm 14:
Theorem 23.5. With probability at least 1 − δ the pseudo-regret of OLR-UCB
satisfies
r  n
R̂n ≤ 8dn (m22 + βn−1 (δ)) log 1 + .
d

23.4 Sparse Online Linear Prediction

Theorem 23.6. There exists a strategy π for the learner such that for any
θ ∈ Rd , the regret ρn (θ) of π against any strategic environment such that
maxt∈[n] kAt k2 ≤ L and maxt∈[n] |Xt | ≤ X satisfies
n  o
ρn (θ) ≤ cX 2 kθk0 log(e + n1/2 L) + Cn log 1 + kθk 1
kθk0 + (1 + X 2 )Cn ,

where c > 0 is some universal constant and Cn = 2 + log2 log(e + n1/2 L).
Note that Cn = O(log log(n)), so by dropping the dependence on X and L, we
have
sup ρn (θ) = O(m0 log(n)) .
θ:kθk0 ≤m0 ,kθk2 ≤L

As a final catch, the rewards (Xt ) in sparse linear bandits with subgaussian noise
 23.5 Notes 285

are not necessarily bounded. However, the subgaussian property implies that with
probability 1 − δ, |ηt | ≤ log(2/δ). By choosing δ = 1/n2 and Assumption 23.1,
we have
 
2

1
P max |Xt | ≥ 1 + log 2n ≤ .
t∈[n] n

Putting all the pieces together shows that the expected regret of OLR-UCB when
using the predictor provided by Theorem 23.6 and when kθk0 ≤ m0 satisfies
p 
Rn = O dnm0 log(n)2 .

23.5 Notes

1 The strategy achieving the bound in Theorem 23.6 is not computationally

efficient. In fact we do not know of any polynomial time algorithm with
logarithmic regret for this problem. The consequence is that Algorithm 14 does
not yet have an efficient implementation.
2 While we focused on the sparse case, the results and techniques apply to other
settings. For example, we can also get alternative confidence sets from results
in online learning even for the standard non-sparse case. Or one may consider
additional or different structural assumptions on θ.
3 When the online linear regression results are applied, it is important to use the
tightest possible, data-dependent regret bounds Bn . In online learning most
regret bounds start as tight, data-dependent bounds, which are then loosened
to get further insight into the structure of problems. For our application,
naturally one should use the tightest available regret bounds (or modify the
existing proofs to get tighter data-dependent bounds). The gains from using
data-dependent bounds can be significant.
4 The confidence set used by Algorithm 14 depends on the sparsity parameter
, which must be known in advance. No algorithm can enjoy a regret of
m0p
O( kθ∗ k0 dn) for all kθ∗ k0 simultaneously (see Chapter 24).
5 The bound in Theorem 23.5 still depends on the ambient dimension. In general
this is unavoidable, as we show in Theorem 24.3. For this reason it recently
became popular to study the contextual setting with changing actions and
make assumptions on the distribution of the contexts so that techniques from
high-dimensional statistics can be brought to bear. These approaches are still
in their infancy and deciding on the right assumptions is a challenge. The
reader is referred to the recent papers by Kim and Paik [2019] and Bastani
and Bayati [2020].
 23.6 Bibliographical Remarks 286

23.6 Bibliographical Remarks

The selective explore-then-commit algorithm is due to the authors [Lattimore et al.,

2015]. The construction for the sparse case is from another paper co-authored
by one of the authors [Abbasi-Yadkori et al., 2012]. The online linear predictor
that competes with sparse parameter vectors and its analysis summarised in
Theorem 23.6 is due to Gerchinovitz [2013, theorem 10]. A recent paper by
Rakhlin and Sridharan [2017] also discusses the relationship between online
learning regret bounds and self-normalised tail bounds of the type given here.
Interestingly, what they show is that the relationship goes in both directions:
tail inequalities imply regret bounds, and regret bounds imply tail inequalities.
We are told by Francesco Orabona that confidence set constructions similar to
those in Section 23.3 have been used earlier in a series of papers by Claudio
Gentile and friends [Dekel et al., 2010, 2012, Crammer and Gentile, 2013, Gentile
and Orabona, 2012, 2014]. Carpentier and Munos [2012] consider a special case
where the action set is the unit sphere and the noise is vector valued so that the
reward is Xt = hAt , θ + ηt i. They prove bounds that essentially depend on the
sparsity of θ and E[kηt k22 ]. Our setting is recovered by choosing ηt to be a vector
of independent standard Gaussian random variables, but in this case the bounds
recovered by the proposed algorithm are suboptimal.

23.7 Exercises

23.1 (The zero-‘norm’) A norm on Rd is a function k · k : Rd → R such that

for all a ∈ R and x, y ∈ Rd , it holds that: (a) kxk = 0 if and only if x = 0 and
(b) kaxk = |a|kxk and (c) kx + yk ≤ kxk + kyk and (d) kxk ≥ 0. Show that k · k0
Pd
given by kxk0 = i=1 I {xi 6= 0} is not a norm.

23.2 (Minimax bound for SETC) Prove the second part of Theorem 23.2.

23.3 (Anytime algorithm) Algorithm 13 is not anytime (it requires advance

knowledge of the horizon). Design a modified version that does not require
this knowledge and prove a comparable regret bound to what was given in
Theorem 23.2.

Hint One way is to use the doubling trick, but a more careful approach will
lead to a more practical algorithm.
23.4 Complete the calculation to derive Eq. (23.10) from Eq. (23.9).

23.5 Prove the equality in Eq. (23.11).

24 Minimax Lower Bounds for
Stochastic Linear Bandits

Lower bounds for linear bandits turn out to be more nuanced than those for the
classical finite-armed bandit. The difference is that for linear bandits the shape
of the action set plays a role in the form of the regret, not just the distribution
of the noise. This should not come as a big surprise because the stochastic
finite-armed bandit problem can be modeled as a linear bandit with actions
being the standard basis vectors, A = {e1 , . . . , ek }. In this case the actions are
orthogonal, which means that samples from one action do not give information
about the rewards for other actions. Other action sets such as the unit ball
(A = B2d = {x ∈ Rd : kxk2 ≤ 1}) do not share this property. For example, if
d = 2 and A = B2d and an algorithm chooses actions e1 = (1, 0) and e2 = (0, 1)
many times, then it can deduce the reward it would obtain from choosing any
other action.
All results of this chapter have a worst-case flavour showing what is (not)
achievable in general, or under a sparsity constraint, or if the realisable assumption
is not satisfied. The analysis uses the information-theoretic tools introduced in
Part IV combined with careful choices of action sets. The hard part is guessing
what is the worst case, which is followed by simply turning the crank on the
usual machinery.
In all lower bounds, we use a simple model with Gaussian noise. For action
At ∈ A ⊆ Rd the reward is Xt = µ(At ) + ηt where ηt ∼ N (0, 1) is a sequence of
independent standard Gaussian noise and µ : A → R is the mean reward. We
will usually assume there exists a θ ∈ Rd such that µ(a) = ha, θi. We write Pµ to
indicate the measure on outcomes induced by the interaction of the fixed policy
and the Gaussian bandit paramterised by µ. Because we are now proving lower
bounds, it becomes necessary to be explicit about the dependence of the regret
on A and µ or θ. The regret of a policy is:
" n
#
X
Rn (A, µ) = n max µ(a) − Eµ Xt ,
a∈A
t=1

where the expectation is taken with respect to Pµ . Except in Section 24.4, we

assume the reward function is linear, which means there exists a θ ∈ Rd such
that µ(a) = ha, θi. In these cases, we write Rn (A, θ) and Eθ and Pθ . Recall the
Pt
notation used for finite-armed bandits by defining Tx (t) = s=1 I {As = x}.
 24.1 Hypercube 288

24.1 Hypercube

The first lower bound is for the hypercube action set and shows that the upper
bounds in Chapter 19 cannot be improved in general.
Theorem 24.1. Let A = [−1, 1]d and Θ = {−n−1/2 , n−1/2 }d . Then, for any
policy, there exists a vector θ ∈ Θ such that:
exp(−2) √
Rn (A, θ) ≥ d n.
8
Proof By the relative entropy identities in Exercise 15.8.(b) and Exercise 14.7,
we have for θ, θ0 ∈ Θ that
" n #
X
D(Pθ , Pθ0 ) = Eθ D(N (hAt , θi, 1), N (hAt , θ i, 1))
0

t=1

1X 
n

= Eθ hAt , θ − θ0 i2 . (24.1)
2 t=1

For i ∈ [d] and θ ∈ Θ, define

n
!
X
pθi = Pθ I {sign(Ati ) 6= sign(θi )} ≥ n/2 .
t=1

Now let i ∈ [d] and θ ∈ Θ be fixed, and let θj0 = θj for j 6= i and θi0 = −θi . Then,
by the Bretagnolle–Huber inequality (Theorem 14.2) and Eq. (24.1),
!
1 1X 1
n
pθi + pθ0 i ≥ exp − Eθ [hAt , θ − θ0 i2 ] ≥ exp (−2) . (24.2)
2 2 t=1 2

Applying an ‘averaging hammer’ over all θ ∈ Θ, which satisfies |Θ| = 2d , we get

X 1 X d
1 XX
d
d
pθi = pθi ≥ exp (−2) .
|Θ| i=1 |Θ| i=1 4
θ∈Θ θ∈Θ
Pd
This implies that there exists a θ ∈ Θ such that i=1 pθi ≥ d exp (−2) /4. By
the definition of pθi , the regret for this choice of θ is at least
" n d #
XX
Rn (A, θ) = Eθ (sign(θi ) − Ati )θi
t=1 i=1
r " n #
1X
d X
≥ Eθ I {sign(Ati ) 6= sign(θi )}
n i=1 t=1
√ X d n
!
n X
≥ Pθ I {sign(Ati ) 6= sign(θi )} ≥ n/2
2 i=1 t=1
√ X
exp(−2) √
d
n
= pθi ≥ d n,
2 i=1 8
 24.2 Unit Ball 289

where the first line follows since the optimal action satisfies a∗i = sign(θi ) for
i ∈ [d], the first inequality follows from a simple case-based analysis showing that
(sign(θi ) − Ati )θi ≥ |θi |I {sign(Ati ) 6= sign(θi )}, the second inequality is Markov’s
inequality (see Lemma 5.1), and the last inequality follows from the choice of
θ.

Except for logarithmic factors, this shows that the algorithm of Chapter 19
is near optimal for this action set. The same proof works when A = {−1, 1}d
is restricted to the corners of the hypercube, which is a finite-armed
p linear
bandit. In Chapter 22, we gave a policy with regret Rn = O( nd log(nk)),
where k = |A|. There is no contradiction because the action set in the above
proof has k = |A| = 2d elements.

24.2 Unit Ball

Lower-bounding the minimax regret when the action set is the unit ball presents
an additional challenge relative to the hypercube. The product structure of
the hypercube means that the actions of the learner in one dimension do not
constraint their choices in other dimensions. For the unit ball, this is not true, and
this complicates the analysis. Nevertheless, a small modification of the technique
allows us to prove a similar bound.

Theorem 24.2. Assume d ≤ 2n and let A = {x ∈ Rd : kxk2 ≤ 1}. Then

there exists a parameter
√ vector θ ∈ Rd with kθk22 = d2 /(48n) such that
√
Rn (A, θ) ≥ d n/(16 3).
1
p
Proof Let ∆ = 4√ 3
d/n and θ ∈ {±∆}d and for i ∈ [d], define τi = n ∧ min{t :
Pt 2
s=1 Asi ≥ n/d}. Then,
" n d  #
XX 1
Rn (A, θ) = ∆Eθ √ − Ati sign(θi )
t=1 i=1 d
√ " n d  2 #
∆ d XX 1
≥ Eθ √ − Ati sign(θi )
2 t=1 i=1 d
√ d "τ  2 #
∆ dX X i
1
≥ Eθ √ − Ati sign(θi ) ,
2 i=1 t=1 d

where the first inequality

√ uses that kAt k22 ≤ 1. Fix i ∈ [d]. For x ∈ {±1}, define
P τi
Ui (x) = t=1 (1/ d − Ati x)2 and let θ0 ∈ {±∆}d be another parameter vector
such that θj = θj0 for j 6= i and θi0 = −θi . Assume without loss of generality that
θi > 0. Let P and P0 be the laws of Ui (1) with respect to the bandit/learner
 24.2 Unit Ball 290

interaction measure induced by θ and θ0 , respectively. Then,

 r
4n 1
Eθ [Ui (1)] ≥ Eθ0 [Ui (1)] − +2 D(P, P0 )
d 2
v " #
 u
∆ 4n u X τi
≥ Eθ0 [Ui (1)] − +2 t E 2
Ati (24.3)
2 d t=1
 r
∆ 4n n
≥ Eθ [Ui (1)] − +2 +1 (24.4)
2
0
d d
√ r
4 3∆n n
≥ Eθ0 [Ui (1)] − , (24.5)
d d
where in the first inequality we used Pinsker’s inequality (Eq. (14.12)), the result
in Exercise 14.4, the bound
1 4n
Xτi Xτi Xτi
√ 2
Ui (1) = (1/ d − Ati ) ≤ 2 +2 A2ti ≤ + 2,
t=1 t=1
d t=1
d

and the assumption that d ≤ 2n. The inequality in Eq. (24.3) follows from the
chain rule for the relative entropy up to a stopping time (Exercise 15.7). Eq. (24.4)
is true by the definition of τi and Eq. (24.5) by the assumption that d ≤ 2n.
Then,
√ r
4 3n∆ n
Eθ [Ui (1)] + Eθ0 [Ui (−1)] ≥ Eθ0 [Ui (1) + Ui (−1)] −
d d
" # √ r √ r
4 3n∆ n 2n 4 3n∆ n
τ
τi X 2 i
n
= 2Eθ0 + Ati − ≥ − = .
d t=1
d d d d d d

The proof is completed using the randomisation hammer:

√ d
X ∆ dX X
Rn (A, θ) ≥ Eθ [Ui (sign(θi ))]
d
2 i=1 d
θ∈{±∆} θ∈{±∆}
√
∆ d
d
X X X
= Eθ [Ui (sign(θi ))]
2 i=1 θ−i ∈{±∆}d−1 θi ∈{±∆}
√ d
∆ dX X n √
≥ = 2d−2 n∆ d .
2 i=1 d
θ−i ∈{±∆}d−1
√ √
n∆ d d n
Hence there exists a θ ∈ {±∆} such that Rn (A, θ) ≥
d
= √ .
4 16 3

The same proof works when A = {x ∈ Rd : kxk2 = 1} is the unit sphere. In

fact, given a set X ⊂ Rd , a minimax lower bound that holds for A = co(X)
continues to hold for A = X.
 24.3 Sparse Parameter Vectors 291

24.3 Sparse Parameter Vectors

√
In Chapter 23 we gave an algorithm with Rn = Õ( dpn) where p ≥ kθk0 is a
known bound on the sparsity of the unknown parameter. Except for logarithmic
terms this bound cannot be improved. An extreme case is when p = 1, which
essentially√reduces to the finite-armed bandit problem where the minimax regret
has order dn (see Chapter 15). For this reason we cannot expect too much from
sparsity and in particular the worst-case bound will depend polynomially on the
ambient dimension d.
Constructing a lower bound for p > 1 is relatively straightforward. For simplicity
we assume that d = pk for some integer k > 1. A sparse linear bandit can mimic
the learner playing p finite-armed bandits simultaneously, each with k arms.
Rather than observing the reward for each bandit, however, the learner only
observes the sum of the rewards and the noise is added at the end. This is
sometimes called the multi-task bandit problem.

Theorem 24.3. Assume pd ≤ n and that d = pk for some integer k ≥ 2. Let

A = {ei ∈ Rk : i ∈ [k]}p ⊂ Rd . Then, for
p any policy there exists a parameter
√
vector θ ∈ Rd with kθk0 = p and kθk∞ ≤ d/(pn) such that Rn (A, θ) ≥ 81 pdn.

Proof Let ∆ > 0 and Θ = {∆ei : i ∈ [k]} ⊂ Rk . Given θ ∈ Θp ⊂ Rd and i ∈ [p],

(i)
let θ(i) ∈ Rk be defined by θk = θ(i−1)p+k , which means that

θ> = [θ(1)> , θ(2)> , . . . , θ(p)> ] .

Next define matrix V ∈ Rp×d to be a block-diagonal matrix with 1 × k blocks,

each containing the row vector (1, 2, . . . , k). For example, when p = 3, we have
 
1 ··· k 0 ··· 0 0 ··· 0
 
V =
0 ··· 0 1 ··· k 0 ··· 0.
 
0 ··· 0 0 ··· 0 1 ··· k

Let Bt = V At ∈ [k]p represent the vector of ‘base’ actions chosen by the learner
in each of the p bandits in round t. The optimal action in the ith bandit is
(i)
b∗i (θ) = argmaxb∈[k] θb .

The regret can be decomposed into the regrets in the p ‘base bandit’ problems (a
form of separability, again):
p
" n #
X X
Rn (θ) = ∆ Eθ I {Bti 6= bi } .
∗

i=1 t=1
| {z }
Rni (θ)
 24.4 Misspecified Models 292

For i ∈ [p], we abbreviate θ(−i) = (θ(1) , . . . , θ(i−1) , θ(i+1) , . . . , θ(p) ). Then,

p
1 X 1 X X
R n (θ) = Rni (θ)
|Θ|p p
|Θ|p i=1 p
θ∈Θ θ∈Θ
p
X 1 X 1 X
= Rni (θ)
|Θ|p−1 |Θ| (i)
i=1 θ (−i) ∈Θp−1 θ ∈Θ
p
1 X 1 X √
≥ kn (24.6)
8 |Θ|p−1
i=1 θ (−i) ∈Θp−1
1 √ 1p
= p kn = dpn .
8 8
Here, in the second equality, we use the convention that θ denotes the vector
obtained by ‘inserting’ θ(i) into θ(−i) at the ith ‘block’. Other
pthan this, the only
tricky step is the inequality, which follows by choosing ∆ ≈ k/n and repeating
the argument outlined in Exercise 15.2. We leave it to the reader to check the
details (Exercise 24.1).

24.4 Misspecified Models

An important generalisation of the linear model is the misspecified case, where

the mean rewards are not assumed to follow a linear model exactly. Suppose
that A ⊂ Rd is a finite set with |A| = k and that Xt = ηt + µ(At ), where
µ : A → R is an unknown function. Let θ ∈ Rd be the parameter vector for which
supa∈A |hθ, ai − µ(a)| is as small as possible:

θ = argminα∈Rd sup |hα, ai − µ(a)| .

a∈A

Then let ε = supa∈A |hθ, ai − µ(a)| be the maximum error. It would be very
pleasant to have an algorithm such that
" n #
X √ √
Rn (A, µ) = n max µ(a) − E µ(At ) = Õ(min{d n + εn, kn}) . (24.7)
a∈A
t=1

Unfortunately, it turns out that results of this kind are not achievable. To show
this, we will prove a generic bound for the classical finite-armed bandit problem
and afterwards show how this implies the impossibility of an adaptive bound like
the above.

Theorem 24.4. Let A = [k], and for µ ∈ [0, 1]k the reward is Xt = µAt + ηt and
the regret is
" n #
X
Rn (µ) = n max µi − Eµ µAt .
i∈A
t=1
 24.4 Misspecified Models 293

Define Θ, Θ0 ⊂ Rk by
 
Θ = µ ∈ [0, 1]k : µi = 0 for i > 1 Θ0 = µ ∈ [0, 1]k .
p
If V ∈ R is such that 2(k − 1) ≤ V ≤ n(k − 1) exp(−2)/8 and supµ∈Θ Rn (µ) ≤
V , then
n(k − 1)
sup Rn (µ0 ) ≥ exp(−2) .
µ0 ∈Θ0 8V
Pn
Proof Recall that Ti (n) = t=1 I {At = i} is the number of times arm i is
played after all n rounds. Let µ ∈ Θ be given by µ1 = ∆ = (k − 1)/V ≤ 1/2. The
regret is then decomposed as:
k
X
Rn (µ) = ∆ Eµ [Ti (n)] ≤ V .
i=2
Pk
Rearranging shows that i=2 Eµ [Ti (n)] ≤ ∆,
V
and so by the pigeonhole principle
there exists an i > 1 such that
V 1
Eµ [Ti (n)] ≤ = 2.
(k − 1)∆ ∆
Then, define µ0 ∈ Θ0 by

∆

 if j = 1
µ0j = 2∆ if j = i


0 otherwise .

Next, by Theorem 14.2 and Lemma 15.1, for any event A, we have
1 1
1
Pµ (A) + Pµ0 (Ac ) ≥ exp (D(Pµ , Pµ0 )) = exp −2∆2 E[Ti (n)] ≥ exp (−2) .
2 2 2
By choosing A = {T1 (n) ≤ n/2} we have
n∆ n(k − 1)
Rn (µ) + Rn (µ0 ) ≥ exp(−2) = exp(−2) .
4 4V
p
Therefore, by the assumption that Rn (µ) ≤ V ≤ n(k − 1) exp(−2)/8 we have
n(k − 1)
Rn (µ0 ) ≥ exp(−2) .
8V
As promised, we now relate this to the misspecified linear bandits. Suppose
that d = 1 (an absurd case) and that there are k arms A = {a1 , a2 , . . . , ak } ⊂ R1 ,
where a1 = (1) and ai = (0) for i > 1. Clearly, if θ > 0 and µ(ai ) = hai , θi, then
the problem can be modelled as a finite-armed bandit with means µ ∈ Θ ⊂ [0, 1]k .
In the general case, we just have a finite-armed bandit with µ ∈ Θ0 . If in the first
√
case we have Rn (A, µ) = O( n), then the theorem shows for large enough n that
√
sup Rn (A, µ) = Ω(k n) .
µ∈Θ0
 24.5 Notes 294

It follows that Eq. (24.7) is a pipe dream. To our knowledge, it is still an open
question of what is possible on this front. We speculate that for k ≥ d2 , there is
a policy for which
  
√ √ k√
Rn (A, θ) = Õ min d n + εn d, n .
d

24.5 Notes

1 The worst-case bound demonstrates the near optimality of the OFUL algorithm
for a specific action set. It is an open question to characterise the optimal
regret for a wide range of action sets. We will return to these issues in the next
part of the book, where we discuss adversarial linear bandits.
2 We return to misspecified bandits in the notes and exercises of Chapter 29,
where algorithms from the adversarial linear bandit framework are applied to
this problem in special cases. √
In many applications, the number of actions is so
√
large that Rn = Õ(d n + εn d) should be considered acceptable. There exist
algorithms achieving this bound, which for large k is essentially not improvable
in the worst case [Lattimore and Szepesvári, 2019b]. For small k,√recent work√by
Foster and Rakhlin [2020] shows that one can achieve Rn = Õ( dkn + εn k).

24.6 Bibliographic Remarks

Worst-case lower bounds for stochastic bandits have appeared in a variety of

places, all with roughly the same bound, but for different action sets. Our very
simple proof for the hypercube is new, but takes inspiration from the paper by
√
Shamir [2015]. Rusmevichientong and Tsitsiklis [2010] proved that Rn = Ω(d n)
when A is the unit sphere. Our proof for the unit ball strengthens their result
marginally and is much simpler. As far as we know, the first lower bound of
√
Ω(d n) was given by Dani et al. [2008] for an action set equal to the product
of two-dimensional disks. The results for the misspecified case are inspired by
the work of one of the authors on the Pareto-regret frontier for bandits, which
characterises what trade-offs are available when it is desirable to have a regret
that is unusually small relative to some specific arms [Lattimore, 2015a].

24.7 Exercises

24.1 Complete the missing steps to prove the inequality in Eq. (24.6).
25 Asymptotic Lower Bounds for
Stochastic Linear Bandits

The lower bounds in the previous chapter were derived by analysing the worst
case for specific action sets and/or constraints on the unknown parameter. In this
chapter, we focus on the asymptotics and aim to understand the influence of the
action set on the regret. We start with a lower bound, and argue that the lower
bound can be achieved. We finish by arguing that the optimistic algorithms (and
Thompson sampling) will perform arbitrarily worse than what can be achieved
by non-optimistic algorithms.

25.1 An Asymptotic Lower Bound for Fixed Action Sets

We assume that A ⊂ Rd is finite with |A| = k and that the reward is

Xt = hAt , θi + ηt , where θ ∈ Rd and (ηt )∞
t=1 is a sequence of independent
standard Gaussian random variables. Of course the regret of a policy in this
setting is
" n #
X
Rn (A, θ) = Eθ ∆At , ∆a = max
0
ha0 − a, θi ,
a ∈A
t=1

where the dependence on the policy is omitted for readability and Eθ [·] is the
expectation with respect to the measure on outcomes induced by the interaction
of the policy and the linear bandit determined by θ. Like the asymptotic lower
bounds in the classical finite-armed case (Chapter 16), the results of this chapter
are proven only for consistent policies. Recall that a policy is consistent in some
class of bandits E if the regret is sub-polynomial for any bandit in that class.
Here this means that

Rn (A, θ) = o(np ) for all p > 0 and θ ∈ Rd . (25.1)

The main objective of the chapter is to prove the following theorem on the
behaviour of any consistent policy and discuss the implications.

Theorem 25.1. Assume that A ⊂ Rd is finite and spans Rd , and suppose a

policy is consistent (satisfies Eq. 25.1). Let θ ∈ Rd be any parameter
 such
Pn
that there is a unique optimal action, and let Ḡn = Eθ t=1 t t . Then
A A >
 25.1 An Asymptotic Lower Bound for Fixed Action Sets 296

lim inf n→∞ λmin (Ḡn )/ log(n) > 0. Furthermore, for any a ∈ A, it holds that

∆2a
lim sup log(n)kak2Ḡ−1 ≤ .
n→∞ n 2

The reader should recognise kak2Ḡ−1 as the key term in the width of the
n
confidence interval for the least squares estimator (Chapter 20). This is quite
intuitive. The theorem is saying that any consistent algorithm must prove
statistically that all suboptimal arms are indeed suboptimal by making the
size of the confidence interval smaller than the suboptimality gap. Before the
proof of this result, we give a corollary that characterises the asymptotic regret
that must be endured by any consistent policy.

Corollary 25.2. Let A ⊂ Rd be a finite set that spans Rd and θ ∈ Rd be such

that there is a unique optimal action. Then, for any consistent policy,
Rn (A, θ)
lim inf ≥ c(A, θ) ,
n→∞ log(n)

where c(A, θ) is defined as

X
c(A, θ) = inf α(a)∆a
α∈[0,∞)A
a∈A
∆2a
subject to kak2H −1 ≤ for all a ∈ A with ∆a > 0 ,
α 2
P
with Hα = a∈A α(a)aa> .

The lower bound is complemented by a matching upper bound that we will

not prove.

Theorem 25.3. Let A ⊂ Rd be a finite set that spans Rd . Then there exists a
policy such that
Rn (A, θ)
lim sup ≤ c(A, θ) ,
n→∞ log(n)

where c(A, θ) is defined as in Corollary 25.2.

Proof of Theorem 25.1 The proof of the first part is simply omitted (see the
reference below for details). It follows along similar lines to what follows, essentially
that if Gn is not sufficiently large in every direction, then some alternative
parameter is not sufficiently identifiable. Let a∗ = argmaxa∈A ha, θi be the optimal
action, which we assumed to be unique. Let θ0 ∈ Rd be an alternative parameter
to be chosen subsequently, and let P and P0 be the measures on the sequence
of outcomes A1 , X1 , . . . , An , Xn induced by the interaction between the policy
and the bandit determined by θ and θ0 respectively. Let E[·] and E0 [·] be the
expectation operators of P and P0 , respectively. By Theorem 14.2 and Lemma 15.1,
 25.1 An Asymptotic Lower Bound for Fixed Action Sets 297

for any event E,

1
P (E) + P0 (E c ) ≥ exp (− D(P, P0 ))
2 " n #!  
1 1 X 2 1 1 0 2
= exp − E 0
hAt , θ − θ i = exp − kθ − θ kḠn .
2 2 t=1
2 2
(25.2)
A simple re-arrangement shows that
 
1 1
kθ − θ0 k2Ḡn ≥ log .
2 2P (E) + 2P0 (E c )
Now we follow the usual plan of choosing θ0 to be close to θ, but so that the
optimal action in the bandit determined by θ0 is not a∗ . Let ∆min = min{∆a :
a ∈ A, ∆a > 0} and ε ∈ (0, ∆min ) and H be a positive definite matrix to be
chosen later such that ka − a∗ k2H > 0. Then define
∆a + ε
θ0 = θ + H(a − a∗ ) ,
ka − a∗ k2H
which is chosen so that
ha − a∗ , θ0 i = ha − a∗ , θi + ∆a + ε = ε .
This means that a∗ is ε-suboptimal for bandit θ0 . We abbreviate Rn = Rn (A, θ)
and Rn0 = Rn (A, θ0 ). Then
" #
X n∆min nε
Rn = E Ta (n)∆a ≥ P (Ta∗ (n) < n/2) ≥ P (Ta∗ (n) < n/2) ,
2 2
a∈A
Pn
where Ta (n) = t=1 I {At = a}. Similarly, a∗ is ε-suboptimal in bandit θ0 so that
nε 0
Rn0 ≥ P (Ta∗ (n) ≥ n/2) .
2
Therefore,
2
P (Ta∗ (n) < n/2) + P0 (Ta∗ (n) ≥ n/2) ≤ (Rn + Rn0 ) . (25.3)
nε
Note that this holds for any choice of H with ka − a∗ kH > 0. The logical next
step is to select H (which determines θ0 ) to make (25.2) as large as possible. The
main difficulty is that this depends on n, so instead we aim to choose an H so
the quantity is large enough infinitely often. We start by just re-arranging things:
∗ 2
1 (∆a + ε)2 ka − a kH Ḡn H (∆a + ε)2
kθ − θ0 k2Ḡn = · 4 = ρn (H) ,
2 2 ∗
ka − a kH 2ka − a∗ k2Ḡ−1
n

where we introduced
ka − a∗ k2Ḡ−1 ka − a∗ k2H Ḡ
ρn (H) = n nH
.
ka − a∗ k4H
 25.1 An Asymptotic Lower Bound for Fixed Action Sets 298

Therefore, by choosing E to be the event that Ta∗ (n) < n/2 and using (25.3) and
(25.2), we have
 
(∆a + ε)2 nε
ρ (H) ≥ log ,
2ka − a∗ k2Ḡ−1 4Rn + 4Rn0
n
n

which after re-arrangement leads to

(∆a + ε)2 log((4Rn + 4Rn0 )/ε)
ρ (H) ≥ 1 − .
2 log(n)ka − a∗ k2Ḡ−1 log(n)
n
n

The definition of consistency means that Rn and Rn0 are both sub-polynomial,
which implies that the second term in the previous expression tends to zero for
large n and so by sending ε to zero,
ρn (H) 2
lim inf 2 ≥ 2. (25.4)
n→∞ log(n)ka − a kḠ−1
∗ ∆a
n

We complete the result using proof by contradiction. Suppose that

∆2a
lim sup log(n)ka − a∗ k2Ḡ−1 > . (25.5)
n→∞ n 2
Then there exists an ε > 0 and infinite set S ⊆ N such that
(∆a + ε)2
log(n)ka − a∗ k2Ḡ−1 ≥ for all n ∈ S .
n 2
Hence, by (25.4), lim inf n∈S ρn (H) > 1. We now choose H to be a cluster point of
the sequence (Ḡ−1
n /kḠn k)n∈S where kḠn k is the spectral norm of the matrix
−1 −1

Ḡn . Such a point must exist, since matrices in this sequence have unit spectral
−1

norm by definition and the set of such matrices is compact. We let S 0 ⊆ S be

a subset so that Ḡ−1n /kḠn k converges to H on n ∈ S . We now check that
−1 0

ka − a kH > 0:
∗

ka − a∗ k2Ḡ−1
ka − a∗ k2H = lim0 n
> 0,
n∈S kḠ−1
n k

where the last inequality follows from the assumption in (25.5) and the first part
of the theorem. Therefore,
ka − a∗ k2Ḡ−1 ka − a∗ k2H Ḡ
1 < lim inf ρn (H) ≤ lim inf n nH
= 1,
n∈S 0 n∈S ka − a∗ k4H
which is a contradiction, and hence (25.5) does not hold. Thus,
∆2a
lim sup log(n)ka − a∗ k2Ḡ−1 ≤ .
n→∞ n 2
We leave the proof of the corollary as an exercise for the reader. Essentially,
though, any consistent algorithm must choose its actions so that in expectation
∆2a
ka − a∗ k2Ḡ−1 ≤ (1 + o(1)) .
n 2 log(n)
 25.2 Clouds Looming for Optimism 299

Now, since a∗ will be chosen linearly, often it is easily shown for suboptimal a
that limn→∞ ka − a∗ kḠ−1
n
/kakḠ−1
n
→ 1. This leads to the required constraint on
the actions of the algorithm, and the optimisation problem in the corollary is
derived by minimising the regret subject to this constraint.

25.2 Clouds Looming for Optimism

The theorem and its corollary have disturbing

implications for policies based on the principle
of optimism in the face of uncertainty, which
is that they can never be asymptotically
optimal. The reason is that these policies
do not choose actions for which they have
collected enough statistics to prove they are
suboptimal, but in the linear setting it can
be worth playing these actions when they
are very informative about other actions for
which the statistics are not yet so clear. As
we shall see, a problematic example appears
in the simplest case where there is information sharing between the arms. Namely,
when the dimension is d = 2, and there are k = 3 arms.
Let A = {a1 , a2 , a3 }, where a1 = e1 and a2 = e2 and a3 = (1 − ε, γε) with
γ ≥ 1 and ε > 0 is small. Let θ = (1, 0) so that the optimal action is a∗ = a1
and ∆a2 = 1 and ∆a3 = ε. If ε is very small, then a1 and a3 point in nearly
the same direction, and so choosing only these arms does not provide sufficient
information to quickly learn which of a1 or a3 is optimal. On the other hand,
a2 and a1 − a3 point in very different directions, which means that choosing a2
allows a learning agent to quickly identify that a1 is in fact optimal. We now
show how the theorem and corollary demonstrate this. First we calculate the
optimal solution to the optimisation problem in Corollary 25.2. Recall we are
trying to minimise

X ∆2a
α(a)∆a subject to kak2H(α)−1 ≤ for all a ∈ A with ∆a > 0 ,
2
a∈A

P
where H(α) = a∈A α(a)aa> . Clearly we should choose α(a1 ) arbitrarily large,
then a computation shows that
 
 0 0 
lim H(α)−1 = .
α(a1 )→∞ 1
0 α(a3 )ε2 γ 2 +α(a2 )
 25.2 Clouds Looming for Optimism 300

The constraints mean that

1 1
= lim ka2 k2H(α)−1 ≤
α(a3 )ε γ + α(a2 ) α(a1 )→∞
2 2 2
2 2
γ ε ε2
= lim ka3 k2H(α)−1 ≤ .
α(a3 )ε γ + α(a2 ) α(a1 )→∞
2 2 2
Provided that γ ≥ 1, this reduces to the constraint that
α(a3 )ε2 + α(a2 ) ≥ 2γ 2 .
Since we are minimising α(a2 ) + εα(a3 ) we can easily see that α(a2 ) = 2γ 2 and
α(a3 ) = 0 provided that 2γ 2 ≤ 2/ε. Therefore, if ε is chosen sufficiently small
relative to γ, then the optimal rate of the regret is c(A, θ) = 2γ 2 , and so by
Theorem 25.3 there exists a policy such that
Rn (A, θ)
lim sup = 2γ 2 .
n→∞ log(n)
Now we argue that for γ sufficiently large and ε arbitrarily small that the regret
for any consistent optimistic algorithm is at least
Rn (A, θ)
lim sup = Ω(1/ε) ,
n→∞ log(n)
which can be arbitrarily worse than the optimal rate! So why is this so? Recall
that optimistic algorithms choose


At = argmaxa∈A max a, θ̃ ,
θ̃∈Ct

where Ct ⊂ Rd is a confidence set that we assume contains the true θ with high
probability. So far this does not greatly restrict the class of algorithms that we
might call optimistic. We now assume that there exists a constant c > 0 such
that
n p o
Ct ⊆ θ̃ : kθ̂t − θ̃kVt ≤ c log(n) ,
Pt
where Vt = s=1 As A> s . So now we ask how often we can expect the optimistic
algorithm to choose action a2 = e2 in the example described above. Since we
have assumed θ ∈ Ct with high probability, we have that
maxha1 , θ̃i ≥ 1 .
θ̃∈Ct

On the other hand, if Ta2 (t − 1) > 4c2 log(n), then

s
q log(n)
maxha2 , θ̃i = maxha2 , θ̃ − θi ≤ 2c ka2 kV −1 log(n) ≤ 2c < 1,
θ̃∈Ct θ̃∈Ct t Ta2 (t − 1)

which means that a2 will not be chosen more than 1 + 4c2 log(n) times. So if
γ = Ω(c2 ), then the optimistic algorithm will not choose a2 sufficiently often
and a simple computation shows it must choose a3 at least Ω(log(n)/ε2 ) times
 25.3 Notes 301

and suffers regret of Ω(log(n)/ε). The key take away from this is that optimistic
algorithms do not choose actions that are statistically suboptimal, but for linear
bandits it can be optimal to choose these actions more often to gain information
about other actions.

This conclusion generalises to structured bandit problems where choosing

one action allows you to gain information about the rewards of other actions.
In such models the optimism principle often provides basic guarantees, but
may fail to optimally exploit the structure of the problem.

25.3 Notes

1 All algorithms known to match the lower bound in Theorem 25.3 are based
on (or inspired by) solving the optimisation problem that defines c(A, θ) with
estimated value θ. Unfortunately, these algorithms are not especially practical
in finite time. As far as we know, none are simultaneously near-optimal in a
minimax sense. Constructing a practical asymptotically optimal algorithm for
linear bandits is a fascinating open problem.
2 In Chapter 36 we will introduce the randomised Bayesian algorithm called
Thompson sampling algorithm for finite-armed and linear bandits. While
Thompson sampling is often empirically superior to UCB, it does not overcome
the issues described here.

25.4 Bibliographic Remarks

The theorems of this chapter are by the authors: Lattimore and Szepesvári [2017].
The example in Section 25.2 first appeared in a paper by Soare et al. [2014],
which deals with the problem of best-arm identification for linear bandits (for an
introduction to best-arm identification, see Chapter 33). The optimisation-based
algorithms that match the lower bound are by Lattimore and Szepesvári [2017],
Ok et al. [2018], Combes et al. [2017] and Hao et al. [2020], with the latter
handling also the contextual case with finitely many contexts.

25.5 Exercises

25.1 Prove Corollary 25.2.

25.2 Prove the first part of Theorem 25.1.

 25.5 Exercises 302

25.3 Give examples of action sets A, parameter vectors θ ∈ Rd and vectors

a ∈ Rd such that:
(a) c(A ∪ {a}, θ) > c(A, θ); and
(b) c(A ∪ {a}, θ) < c(A, θ).
Part VI
Adversarial Linear Bandits
 304

The adversarial linear bandit is superficially a generalisation of the stochastic

linear bandit where the unknown parameter vector is chosen by an adversary.
There are many similarities between the two topics. Indeed, the techniques in
this part combine the ideas of optimal design presented in Chapter 22 with
the exponential weighting algorithm of Chapter 11. The intuitions gained by
studying stochastic bandits should not be taken too seriously, however. There are
subtle differences between the model of adversarial bandits introduced here and
the stochastic linear bandits examined in previous chapters. These differences
will be discussed at length in Chapter 29. The adversarial version of the linear
bandits turns out to be remarkably rich, both because of the complex information
structure and because of the challenging computational issues.
The part is split into four chapters, the first of which is an introduction to the
necessary tools from convex analysis and optimisation. In the first chapter on
bandits, we show how to combine the core ideas of the Exp3 policy of Chapter 11
with the optimal experimental design for least-squares estimators in Chapter 21.
When the number of actions is large (or infinite), the approach based on Exp3
is hard to make efficient. These shortcomings are addressed in the next chapter,
where we introduce the mirror descent and follow-the-regularised leader algorithms
for bandits and show how they can be used to design efficient algorithms. We
conclude the part with a discussion on the relationship between adversarial and
stochastic linear bandits, which is more subtle than the situation with finite-armed
bandits.
26 Foundations of Convex Analysis ( )

Our coverage of convex analysis is necessarily extremely brief. We introduce only

what is necessary and refer the reader to standard texts for the proofs.

26.1 Convex Sets and Functions

A set A ⊆ Rd is convex if for any x, y ∈ A it holds that αx + (1 − α)y ∈ A for all

α ∈ (0, 1). The convex hull of a collection of points x1 , x2 , . . . , xn ∈ Rd is the
smallest convex set containing the points, which also happens to satisfy
( n
)
X
co(x1 , x2 , . . . , xn ) = x∈R :x=
d
pi xi for some p ∈ Pn−1 .
i=1

The convex hull co(A) is also defined for an arbitrary set A ⊂ Rd and is still the
smallest convex set that contains A (see (c) in Figure 26.1). For the rest of the
section, we let A ⊆ Rd be convex. Let R̄ = R ∪ {−∞, ∞} be the extended real
number system and define operations involving infinities in the natural way (see
notes).

Definition 26.1. An extended real-valued function f : Rd → R̄ is convex if its

epigraph Ef = {(x, y) ∈ Rd × R : y ≥ f (x)} ⊂ Rd+1 is a convex set.

The term ‘epi’ originates in greek and it means upon or over: The epigraph of
a function is the set of points that sit on the top of the function’s graph.
The domain of an extended real-valued function on Rd is dom(f ) = {x ∈ Rd :
f (x) < ∞}. For S ⊂ Rd , a function f : S → R̄ is identified with the function
f¯ : Rd → R̄, which coincides with f on S and is defined to take the value ∞
outside of S. It follows that if f : S → R, then dom(f ) = S. A convex function is
proper if its range does not include −∞ and its domain is nonempty.

For the rest of the chapter, we will write ‘let f be a convex’ to mean that
f : Rd → R̄ is a proper convex function.
 26.1 Convex Sets and Functions 306

(a) (b) (c)

(d) (e) (f)

Figure 26.1 (a) is a convex set. (b) is a non-convex set. (c) is the convex hull of a
non-convex set. (d) is a convex function. (e) is non-convex, but all local minimums are
global. (f) is not convex.

Permitting convex functions to take values of −∞ is a convenient standard

because certain operations on proper convex functions result in improper
ones (infimal convolution, for example). These technicalities will never bother
us in this book, however.

A consequence of the definition is that for convex f , we have

f (αx + (1 − α)y) ≤ αf (x) + (1 − α)f (y)
for all α ∈ (0, 1) and x, y ∈ dom(f ) . (26.1)
In fact, the inequality holds for all x, y ∈ Rd .

Some authors use Eq. (26.1) as the definition of a convex function along
with a specification that the domain is convex: If A ⊆ Rd is convex, then
f : A → R is convex if it satisfies Eq. (26.1), with f (x) = ∞ assumed for
x∈/ A.

The reader is invited to prove that all convex functions are continuous on the
interior of their domain (Exercise 26.1).
A function is strictly convex if the inequality in Eq. (26.1) is always strict.
The Fenchel dual of a function f is f ∗ (u) = supx hx, ui − f (x), which is convex
because the maximum of convex functions is convex. The Fenchel dual has many
nice properties. Most important for us is that for sufficiently nice functions, ∇f ∗
is the inverse of ∇f (Theorem 26.6). Another useful property is that when f is
a proper convex function and its epigraph is closed, then f = f ∗∗ , where f ∗∗
denotes the bidual of f : f ∗∗ = (f ∗ )∗ . The Fenchel dual is also called the convex
 26.2 Jensen’s Inequality 307

conjugate. If f : Rd → R̄ is twice differentiable on the interior of its domain, then

convexity of f is equivalent to its Hessian having non-negative eigenvalues for
all x ∈ int(dom(f )). The field of optimisation is obsessed with convex functions
because all local minimums are global (see Fig. 26.1). This means that minimising
a convex function is usually possible (efficiently) using some variation of gradient
descent. A function f : Rd → R̄ is concave if −f is convex.

26.2 Jensen’s Inequality

One of the most important results for convex functions is Jensen’s inequality:
Theorem 26.2 (Jensen’s inequality). Let f : Rd → R̄ be a measurable convex
function and X be an Rd -valued random element on some probability space such
that E[X] exists and X ∈ dom(f ) holds almost surely. Then E[f (X)] ≥ f (E[X]).
If we allowed Lebesgue integrals to take on the value of ∞, the condition that
X is almost surely an element of the domain of f could be removed and the
result would still be true. Indeed, in this case we would immediately conclude
that E[f (X)] = ∞ and Jensen’s inequality would trivially hold.
The basic inequality of (26.1) is trivially
a special case of Jensen’s inequality. Jensen’s f (x)
inequality is so central to convexity that it
can actually be used as the definition (a Pn
(x̄, pk f (xk ))
function is convex if and only if it satisfies k=1

Jensen’s inequality). The proof of Jensen’s using

Definition 26.1 in full generality is left to the
reader (Exercise 26.2). However, we cannot resist (x̄, f (x̄))

to include here a simple ‘graphical proof’ that x1 x2 x3 x4 x5

works in the simple case when X is supported
Pn
on x1 , . . . , xn and P(X = xk ) = pk . Then, letting x̄ = k=1 pk xk , one can notice
P
that the point (x̄, k pk xk ) lies in the convex hull of {(xk , f (xk ))k }, which is a
convex subset of the epigraph Ef ⊂ Rd+1 . The result follows because (x̄, f (x̄)) is
on the boundary of Ef as shown in the figure. The direction of Jensen’s inequality
is reversed if ‘convex’ is replaced by ‘concave’.

26.3 Bregman Divergence

Let f : Rd → R be convex and x, y ∈ Rd with y ∈ dom(f ). The Bregman

divergence at y induced by f is defined by
Df (x, y) = f (x) − f (y) − ∇x−y f (y) ,
where ∇v f (y) = limh→0+ (f (y + hv) − f (y))/h ∈ R ∪ {−∞, ∞} is the directional
derivative of f at y in direction v. The directional derivative is always well defined
 26.3 Bregman Divergence 308

f (x)
f (y) + hx − y, ∇f (y)i

Df (x, y)

y x

Figure 26.2 The Bregman divergence Df (x, y) is the difference between f (x) and the
Taylor series approximation of f at y. When f is convex the, linear approximation is a
lower bound on the function and the Bregman divergence is positive.

for convex functions, but can be positive/negative infinity. When f is differentiable

at y, then ∇v f (y) = hv, ∇f (y)i and thus Df (x, y) = f (x) − f (y) − hx − y, ∇f (y)i,
which is the more usual definition. For the geometric intuition see Fig. 26.2. Let
dom(∇f ) denote the set of points in the domain of f where f is differentiable.
Theorem 26.3. The following hold:
(a) Df (x, y) ≥ 0 for all y ∈ dom(f ).
(b) Df (x, x) = 0 for all x ∈ dom(f ).
(c) Df (x, y) is convex as a function of x for any y ∈ dom(∇f ).
Part (c) does not hold in general when f is not differentiable at y, as you will
show in Exercise 26.14. The square root of the Bregman divergence shares many
properties with a metric, and for some choices of f , it actually is a metric. In
general, however, it is not symmetric and does not satisfy the triangle inequality.
Example 26.4. Let f (x) = 21 kxk22 . Then ∇f (x) = x and
1 1 1
Df (x, y) = kxk22 − kyk22 − hx − y, yi = kx − yk22 .
2 2 2
Example 26.5. Let A = [0, ∞)d , dom(f ) = A and for x ∈ A, f (x) =
Pd
i=1 (xi log(xi ) − xi ), where 0 log(0) = 0. Then, for y ∈ (0, ∞) , ∇f (y) = log(y)
d

and
d
X d
X d
X
Df (x, y) = (xi log(xi ) − xi ) − (yi log yi − yi ) − log(yi )(xi − yi )
i=1 i=1 i=1
d
X   d
X
xi
= xi log + (yi − xi ) .
i=1
yi i=1

Notice that if x, y ∈ Pd−1 are in the unit simplex, then Df (x, y) is the relative
entropy between probability vectors x and y. The function f is called the
unnormalised negentropy, which will feature heavily in many of the chapters
 26.4 Legendre Functions 309

that follow. When y 6> 0, the Bregman divergence is infinite if there exists an
P
i such that yi = 0 and xi > 0. Otherwise, Df (x, y) = i:xi >0 xi log(xi /yi ) +
Pd
i=1 (yi − xi ).

26.4 Legendre Functions

In this section we use various topological

notions such as the interior, closed set and 0 √
− x
boundary. The definitions of these terms are
given in the notes. Let f be a convex function
and A = dom(f ) and C = int(A). Then f is −1
Legendre if

(a) C is non-empty; −2
(b) f is differentiable and strictly convex on 0 1 2 3 4
C; and √
(c) limn→∞ k∇f (xn )k2 = ∞ for any Figure 26.3 f (x) = − x: the
archetypical Legendre function
sequence (xn )n with xn ∈ C for all n
and limn→∞ xn = x and some x ∈ ∂C.

The intuition is that the set {(x, f (x)) : x ∈ dom(A)} is a ‘dish’ with ever-
steepening edges towards the boundary of the domain. Legendre functions have
some very convenient properties:

Theorem 26.6. Let f : Rd → R̄ be a Legendre function. Then,

(a) ∇f is a bijection between int(dom(f )) and int(dom(f ∗ )) with the inverse

(∇f )−1 = ∇f ∗ ;
(b) Df (x, y) = Df ∗ (∇f (y), ∇f (x)) for all x, y ∈ int(dom(f )); and
(c) the Fenchel conjugate f ∗ is Legendre.

The next result formalises the ‘dish’ intuition by showing the directional
derivative along any straight path from a point in the interior to the boundary
blows up. You should supply the proof of the following results in Exercise 26.6.

Proposition 26.7. Let f be Legendre and x ∈ int(dom(f )) and y ∈

∂ int(dom(f )), then limα→1 hy − x, ∇f ((1 − α)x + αy)i = ∞.

Corollary 26.8. If f is Legendre and x∗ ∈ argminx∈dom(f ) f (x), then x∗ ∈

int(dom(f )).

Example 26.9. Let f be the Legendre function given by f (x) = 12 kxk22 , which
has domain dom(f ) = Rd . Then, f ∗ (x) = f (x) and ∇f and ∇f ∗ are the identity
functions.
 26.4 Legendre Functions 310

Pd √
Example 26.10. Let f (x) = −2 i=1 xi when xi ≥ 0 for all i and ∞
otherwise, which has dom(f ) = [0, ∞)d and int(dom(f )) = (0, ∞)d . The gradient
√
is ∇f (x) = −1/ x, which blows up (in norm) on any sequence (xn ) approaching
√
∂ int(dom(f )) = {x ∈ [0, ∞)d : xi = 0 for some i ∈ [d]}. Here, x stands for
√
the vector ( xi )i . In what follows we will often use the underlying convention
of extending univariate functions to vector by applying them componentwise.
Note that k∇f (x)k → 0 as kxk → ∞: ‘∞’ is not part of the boundary of dom(f ).
Strict convexity is also obvious so f is Legendre. In Exercise 26.8, we ask you
to calculate the Bregman divergences with respect to f and f ∗ and verify the
results of Theorem 26.6.
P
Example 26.11. Let f (x) = i xi log(xi ) − xi be the unnormalised negentropy,
which we met in Example 26.5. Similarly to the previous example, dom(f ) =
[0, ∞)d , int(dom(f )) = (0, ∞)d and ∂ int(dom(f )) = {x ∈ [0, ∞)d : xi =
0 for some i ∈ [d]}. The gradient is ∇f (x) = log(x), and thus k∇f (x)k → ∞ as
x → ∂ int(dom(f )). Strict convexity also holds, hence f is Legendre. You already
met the Bregman divergence Df (x, y), which turned out to be the relative entropy
when x, y belong to the simplex. Exercise 26.9 asks you to calculate the dual of f
(can you guess what this function will be?) and the Bregman divergence induced
by f ∗ and to verify Theorem 26.6.

The Taylor series of the Bregman divergence is often a useful approximation.

Let g(y) = Df (x, y), which for y = x has ∇g(y) = 0 and ∇2 g(y) = ∇2 f (x). A
second-order Taylor expansion suggests that
1 1
Df (x, y) = g(y) ≈ g(x) + hy − x, ∇g(x)i + ky − xk2∇2 f (x) = ky − xk2∇2 f (x) .
2 2
This approximation can be very poor if x and y are far apart. Even when x and y
are close, the lower-order terms are occasionally problematic, but nevertheless the
approximation can guide intuition. The next theorem, which is based on Taylor’s
theorem and measurable selections, gives an exact result (Exercise 26.15).

Theorem 26.12. If f is convex and twice differentiable in A = int(dom(f )) and

x, y ∈ A, then there exists an α ∈ [0, 1] and z = αx + (1 − α)y such that
1
Df (x, y) = kx − yk2∇2 f (z) .
2
Suppose furthermore that ∇2 f is continuous on int(dom(f )); then there exists a
measurable function g : int(dom(f )) × int(dom(f )) → int(dom(f )) such that for
all x, y ∈ int(dom(f )),
1
Df (x, y) = kx − yk2∇2 f (g(x,y)) .
2
The next result will be useful.

Theorem 26.13. Let η > 0 and f be Legendre and twice differentiable in

 26.5 Optimisation 311

A = int(dom(f )), x, y ∈ A, and let z ∈ [x, y] be the point such that Df (x, y) =
1 2
2 kx − yk∇2 f (z) . Then, for all u ∈ R ,
d

Df (x, y) η
hx − y, ui − ≤ kuk2(∇2 f (z))−1 .
η 2

Although the Bregman divergence is not symmetric, the right-hand side does
not depend on the order of x and y in the Bregman divergence except that
z ∈ [x, y] may be different.

Proof Strict convexity of f ensures that H = ∇2 f (z) is invertible. Applying

Cauchy–Schwarz,
q
hx − y, ui ≤ kx − ykH kukH −1 = kukH −1 2Df (x, y) .
Therefore,
q
Df (x, y) Df (x, y) η
hx − y, ui − ≤ kukH −1 2Df (x, y) − ≤ kuk2H −1 ,
η η 2
where the last step follows from the ever useful maxx∈R ax − bx2 = a2 /(4b) which
holds for any b > 0 and a ∈ R.

26.5 Optimisation

The first-order optimality condition states that if x ∈ Rd is the minimiser

of a differentiable function f : Rd → R, then ∇f (x) = 0. One of the things we
like about convex functions is that when f is convex, the first-order optimality
condition is both necessary and sufficient. In particular, if ∇f (x) = 0 for some
x ∈ Rd then x is a minimiser of f . The first-order optimality condition can also
be generalised to constrained minima: if f : Rd → R is convex and differentiable
and A ⊆ Rd is a non-empty convex set, then
x∗ ∈ argminx∈A f (x) ⇔ ∀x ∈ A : hx − x∗ , ∇f (x∗ )i ≥ 0 . (26.2)
The necessity of the condition on the right-hand side is easy to understand by
a geometric reasoning. If ∇f (x∗ ) = 0, then the said condition trivially holds. If
∇f (x∗ ) 6= 0, the hyperplane Hx∗ whose normal is ∇f (x∗ ) and goes through x∗
must be a supporting hyperplane of A at x∗ , with −∇f (x∗ ) being the outer
normal of A at x∗ otherwise x∗ could be moved by a small amount while staying
inside A and improving the value of f . Since A is convex, it thus lies entirely
on the side of Hx∗ that ∇f (x∗ ) points into. This is clearly equivalent to (26.2).
The sufficiency of the condition also follows from this geometric viewpoint as the
reader may verify from the figure.
The above statement continues to hold with a small modification even when f
 26.6 Projections 312

−∇f (x)
better point

−∇f (x∗ )

Figure 26.4 Illustration of first-order optimality conditions. The point at the top is not a
minimiser because the hyperplane with normal as gradient does not support the convex
set. The point at the right is a minimiser.

is not differentiable everywhere. In particular, in this case the equivalence (26.2)

holds for any x∗ ∈ dom(∇f ) with the modification that on the right side of the
equivalence, A should be replaced by A ∩ dom(f ):

Proposition 26.14. Let f : Rd → R̄ be a convex function and A ⊂ Rd a

non-empty convex set. Then, for any x∗ ∈ dom(∇f ), it holds that

x∗ ∈ argminx∈A f (x) ⇐⇒
∀x ∈ A ∩ dom(f ) : hx − x∗ , ∇f (x∗ )i ≥ 0 . (26.3)

Further, if f is Legendre, then x∗ ∈ argminx∈A f (x) implies x∗ ∈ dom(∇f ) and

hence also (26.3).

The part that concerns the Legendre objective f follows by noting that by
Corollary 26.8, x∗ ∈ int(dom(f )) combined with that by Theorem 26.6(a),
int(dom(f )) = dom(∇f ).

26.6 Projections

If A ⊂ Rd and x ∈ Rd , then the Euclidean projection of x on A is ΠA (x) =

argminy∈A kx − yk22 . One can also project with respect to a Bregman divergence
induced by convex function f . Let ΠA,f be defined by

ΠA,f (x) = argminy∈A Df (y, x) .

An important property of the projection is that minimising a Legendre function

f on a convex set A is (usually) equivalent to finding the unconstrained minimum
on the domain of f and then projecting that point on to A.

Theorem 26.15. Let f : Rd → R̄ be Legendre, A ⊂ Rd a non-empty, closed

convex set with A ∩ dom(f ) non-empty and assume that ỹ = argminz∈Rd f (z)
exists. Then the following hold:
 26.7 Notes 313

(a) y = argminz∈A f (z) exists and is unique;

(b) y = argminz∈A Df (z, ỹ).
√
The assumption that ỹ exists is necessary. For example f (x) = − x for x ≥ 0
and f (x) = ∞ for x < 0 is Legendre with domain dom(f ) = [0, ∞), but f does
not have a minimum on its domain.

26.7 Notes

1 The ‘infinity arithmetic’ on the extended real line is as follows:

α+∞=∞ for α ∈ (−∞, ∞]
α − ∞ = −∞ for α ∈ [−∞, ∞)
α · ∞ = ∞ and α · (−∞) = −∞ for α > 0
α · ∞ = −∞ and α · (−∞) = ∞ for α < 0
0 · ∞ = 0 · (−∞) = 0 .
Like α/0, the value of ∞ − ∞ is not defined. We also have α ≤ ∞ for all α
and α ≥ −∞ for all α.
2 There are many ways to define the topological notions used in this chapter.
The most elegant is also the most abstract, but there is no space for that here.
Instead we give the classical definitions that are specific to Rd and subsets.
Let A be a subset of Rd . A point x ∈ A is an interior point if there exists
an ε > 0 such that Bε (x) = {y : kx − yk2 ≤ ε} ⊂ A. The interior of A is
int(A) = {x ∈ A : x is an interior point}. The set A is open if int(A) = A
and closed if its complement Ac = Rd \ A is open. The boundary of A is
denoted by ∂A and is the set of points in x ∈ Rd such that for all ε > 0 the
set Bε (x) contains points from A and Ac . Note that points in the boundary
need not be in A. Some examples: ∂[0, ∞) = {0} = ∂(0, ∞) and ∂Rn = ∅.

26.8 Bibliographic Remarks

The main source for these notes is the excellent book by Rockafellar [2015]. The
basic definitions are in part I. The Fenchel dual is analysed in part III while
Legendre functions are found in part V. Convex optimisation is a huge topic. The
standard text is by Boyd and Vandenberghe [2004].

26.9 Exercises

26.1 Let f : Rd → R be convex. Prove that f is continuous on int(dom(f )).

26.2 Prove Jensen’s inequality (Theorem 26.2). Precisely, let X ∈ Rd be a

 26.9 Exercises 314

random variable for which E[X] exists and f : Rd → R̄ a measurable convex

function. Prove that E[f (x)] ≥ f (E[X]).

Hint Let x0 = E[X] ∈ Rd and define a linear function g : Rd → R such that

g(x0 ) = f (x0 ) and g(x) ≤ f (x) for all x ∈ Rd . To guarantee the existence of
g, you may use the supporting hyperplane theorem, which states that if
S ⊂ Rm is a convex set and s ∈ ∂S, then there exists a supporting hyperplane
containing s.
26.3 Let f : Rd → R ∪ {−∞, ∞}.

(a) Prove that f ∗∗ (x) ≤ f (x).

(b) Assume that f is convex and differentiable on int(dom(f )). Show that
f ∗∗ (x) = f (x) for x ∈ int(dom(f )).

As mentioned in the text, the assumption that f is differentiable can be

relaxed to an assumption that the epigraph of f is closed, in which case the
result holds over the whole domain. The proof is not hard, but you will need
to use the sub-differential rather than the gradient, and the boundary must
be treated with care.

26.4 For each of the real-valued functions below, decide whether or not it is
Legendre on the given domain:

(a) f (x) = x2 on [−1, 1].

√
(b) f (x) = − x on [0, ∞).
(c) f (x) = log(1/x) on [0, ∞) with f (0) = ∞.
(d) f (x) = x log(x) on [0, ∞) with f (0) = 0.
(e) f (x) = |x| on R.
(f) f (x) = max{|x|, x2 } on R.

26.5 Prove Theorem 26.3.

26.6 Prove Proposition 26.7 and Corollary 26.8.

26.7 Prove Proposition 26.14.

26.8 Let f be the convex function given in Example 26.10.

(a) For x, y ∈ dom(f ), find Df (x, y).

(b) Compute f ∗ (u) and ∇f ∗ (u).
(c) Find dom(∇f ∗ ).
(d) Show that for u, v ∈ (−∞, 0]d ,

(ui − vi )2
d
X
Df ∗ (u, v) = − .
i=1
ui vi2
 26.9 Exercises 315

(e) Verify the claims in Theorem 26.6.

26.9 Let f : Rd → R̄ be the unnormalised negentropy function from

P
Example 26.11. We have seen in Example 26.5 that Df (x, y) = i (xi log(xi /yi ) +
yi − xi ).
(a) Compute f ∗ (u) and ∇f ∗ (u).
(b) Find dom(∇f ∗ ).
(c) Show that for u, v ∈ Rd ,
d
X
Df ∗ (u, v) = exp(vi )(vi − ui ) + exp(ui ) − exp(vi ) .
i=1

(d) Verify the claims in Theorem 26.6.

26.10 Let f be Legendre. Show that f˜ given by f˜(x) = f (x) + hx, ui is also
Legendre for any u ∈ Rd .

26.11 Let f be the unnormalised negentropy function from Example 26.5.

(a) Prove that f is Legendre.
(b) Given y ∈ (0, ∞)d , prove that argminx∈Pd−1 Df (x, y) = y/kyk1 .

26.12 Let α ∈ [0, 1/d] and A = Pd−1 ∩ [α, 1]d and f be the unnormalised
negentropy function. Let y ∈ [0, ∞)d and x = argminx∈A Df (x, y) and assume
that y1 ≤ y2 ≤ · · · ≤ yd . Let m be the smallest value such that
d
X
ym (1 − (m − 1)α) ≥ α yj .
j=m

Show that
(
α if i < m
xi = Pd
(1 − (m − 1)α)yi / j=m yj otherwise .

26.13 (Generalised Pythagorean identity) Let A ⊂ Rd be convex and

closed and f : Rd → R̄ be a convex function with A ∩ dom(f ) non-empty.
(a) Suppose that x ∈ A and y ∈ Rd and z = ΠA,f (y) and f is differentiable at
y. Prove that
Df (x, y) ≥ Df (x, z) + Df (z, y) .
(b) Prove that the condition that f be differentiable at y cannot be relaxed.

26.14 Prove Theorem 26.3 and show that Part (c) does not hold in general
when f is not differentiable at y.

26.15 Prove Theorem 26.12

Hint For the first part, simply apply Taylor’s theorem. For the second part,
 26.9 Exercises 316

use a measurable selection theorem. For example, the theorem by Kuratowski

and Ryll-Nardzewski, which appears as Theorem 6.9.4 in the second volume of
the book by Bogachev [2007].
27 Exp3 for Adversarial Linear Bandits

The model for adversarial linear bandits is as follows. The learner is given an
action set A ⊂ Rd and the number of rounds n. As usual in the adversarial setting,
it is convenient to switch to losses. An instance of the adversarial problem is a
sequence of loss vectors y1 , . . . , yn taking values in Rd . In each round t ∈ [n], the
learner selects a possibly random action At ∈ A and observes a loss Yt = hAt , yt i.
The learner does not observe the loss vector yt . The regret of the learner after n
rounds is
" n # n
X X
Rn = E Yt − min ha, yt i .
a∈A
t=1 t=1

Clearly, the finite-armed adversarial bandits discussed in Chapter 11 is a special

case of adversarial linear bandits corresponding to the choice A = {e1 , . . . , ed },
where e1 , . . . , ed are the unit vectors of the d-dimensional standard Euclidean
basis.

For this chapter, we assume that


(a) for all t ∈ [n] the loss satisfies yt ∈ L = x ∈ Rd : supa∈A |ha, xi| ≤ 1 ;
and
(b) the action set A spans Rd .
The latter assumption is for convenience only and may be relaxed with a
little care (Exercise 27.7).

27.1 Exponential Weights for Linear Bandits

We adapt the exponential-weighting algorithm of Chapter 11. Like in that setting,

we need a way to estimate the individual losses for each action, but now we make
use of the linear structure to share information between the arms and decrease
the variance of our estimators. For now we assume that A is finite, which we
relax in Section 27.3. Let t ∈ [n] be the index of the current round. Assuming
the loss estimate for action a ∈ A in round s ∈ [n] is Ŷs (a), then the probability
distribution proposed by exponential weights is given by the probability mass
 27.1 Exponential Weights for Linear Bandits 318

function P̃t : A → [0, 1] given by

t−1
!
X
P̃t (a) ∝ exp −η Ŷs (a) ,
s=1

where η > 0 is the learning rate. To control the variance of the loss estimates,
it will be useful to mix this distribution with an exploration distribution π
P
(π : A → [0, 1] and a∈A π(a) = 1). The mixture distribution is
Pt (a) = (1 − γ)P̃t (a) + γπ(a) ,
where γ is a constant mixing factor to be chosen later. The algorithm then simply
samples its action At from Pt :
A t ∼ Pt .
Recall that Yt = hAt , yt i is the observed loss after taking action At . We need a
.
way to estimate yt (a) = ha, yt i. The idea is to use least squares to estimate yt with
Ŷt = Rt At Yt , where Rt ∈ Rd×d is selected so that Ŷt is an unbiased estimate of yt
given the history. Then the loss for a given action is estimated by Ŷt (a) = ha, Ŷt i.
To find the choice of Rt that makes Ŷt unbiased, let Et [·] = E [·|Pt ] and calculate
!
X
Et [Ŷt ] = Rt Et [At At ]yt = Rt
>
Pt (a)aa>
yt .
a∈A
| {z }
Qt

Using Rt = Q−1 t leads to Et [Ŷt ] = yt as desired. Of course Qt should be non-

singular, which will follow by choosing π so that
X
Q(π) = π(a)aa>
a∈A

is non-singular. The complete algorithm is summarised in Algorithm 15.

1: Input Finite action set A ⊂ Rd , learning rate η, exploration distribution π,

exploration parameter γ
2: for t = 1, 2, . . . , n do
3: Compute sampling distribution:
 P 
exp −η s=1 Ŷs (a)
t−1

Pt (a) = γπ(a) + (1 − γ) P  P .
a0 ∈A exp −η s=1 Ŷs (a )
t−1 0

4: Sample action At ∼ Pt
5: Observe loss Yt = hAt , yt i and compute loss estimates:
Ŷt = Q−1
t At Yt and Ŷt (a) = ha, Ŷt i .
6: end for
Algorithm 15: Exp3 for linear bandits.
 27.2 Regret Analysis 319

27.2 Regret Analysis

Theorem 27.1. Assume that A is non-empty and let k = |A|. For any exploration
distribution π, for some parameters η and γ, for all (yt )t with yt ∈ L, the regret
of Algorithm 15 satisfies
p
Rn ≤ 2 (2g(π) + d)n log(k) , (27.1)

where g(π) = maxa∈A kak2Q−1 (π) . Furthermore, there exists an exploration

distribution
p π and parameters η and γ such that g(π) ≤ d, and hence Rn ≤
2 3dn log(k).

The utility of (27.1) is that at times, calculating the distribution that minimises
g(π) or sampling from it may be difficult, in which case, one may employ a
distribution that trades off computation with the regret.

Proof Assume that the learning rate η is chosen so that for each round t the
loss estimates satisfy

η Ŷt (a) ≥ −1, ∀a ∈ A . (27.2)

Then, by adopting the proof of Theorem 11.1 (see Exercise 27.1), the regret is
bounded by
" #
log k Xn X
2
Rn ≤ + 2γn + η E Pt (a)Ŷt (a) . (27.3)
η t=1 a∈A

Note that we cannot use the proof that leads to the tighter constant (η getting
replaced by η/2 in the second term above) because we would loose too much in
other parts of the proof by guaranteeing that the loss estimates are bounded
by one (see below). To get a regrethP bound, it remains
i to set γ and η so that
2
(27.2) is satisfied and to bound E a Pt (a)Ŷt (a) . We start with the latter. Let
P 2
Mt = a Pt (a)Ŷt (a). By the definition of the loss estimate,

Ŷt2 (a) = (a> Q−1 2 2 > −1

t At Yt ) = Yt At Qt aa Qt At ,
> −1

P 2 2 > −1
which means that Mt = a Pt (a)Ŷt (a) = Yt At Qt At ≤ At Qt At =
> −1

trace(At At Qt ), and by the linearity of trace,

> −1

!
X
E[Mt | Pt ] ≤ trace Pt (a)aa Qt
> −1
= d.
a∈A

It remains to choose γ and η. Strengthen (27.2) to |η Ŷt (a)| ≤ 1 and note that
since |Yt | ≤ 1,

|η Ŷt (a)| = |ηa> Q−1 > −1

t At Yt | ≤ η|a Qt At | .
P
Recall that Q(π) = a∈A π(a)aa . Clearly Qt  γQ(π), and hence Qt
> −1

 27.3 Continuous Exponential Weights 320

Q(π)−1 /γ by Exercise 27.4. Using this and the Cauchy–Schwarz inequality shows
that
1 g(π)
t At | ≤ kakQ−1 kAt kQ−1 ≤ max ν Qt ν ≤
|a> Q−1 max ν > Q−1 (π)ν =
> −1
,
t t ν∈A γ ν∈A γ

which implies that

η
|η Ŷt (a)| ≤ g(π) . (27.4)
γ

Choosing γ = ηg(π) guarantees |η Ŷt (a)| ≤ 1. Plugging this choice into (27.3), we
get

log k p
Rn ≤ + ηn(2g(π) + d) = 2 (2g(π) + d)n log(k) ,
η
q
log(k)
where the last equality is derived by choosing η = (2g(π)+d)n finishing the proof
of (27.1).
For the second half, recall that by the Kiefer–Wolfowitz theorem (Theorem 21.1
and Exercise 21.6), there exists a sampling distribution π such that g(π) ≤ d.
Plugging this value into (27.1), finishes the proof.

27.3 Continuous Exponential Weights

The dependence on log(k) in the regret guarantee provided by Theorem 27.1

is objectionable when the number of arms is extremely large or infinite. One
approach is to find a finite subset C ⊆ A for which

sup min sup |ha − b, yi| ≤ 1/n .

a∈A b∈C y∈L

A standard calculation shows (Exercise 27.6) shows that C can always be chosen
so that log |C| ≤ d log(6dn). Then it is easy to check that
p Exp3 on C suffers regret
relative to the best action in A of at most Rn = O(d n log(nd)). The problem
with this approach is that C is exponentially large in d, which makes this algorithm
intractable in most situations. When A is convex, a more computationally
tractable approach is to use the continuous exponential weights algorithm.

For this section, we assume that A is convex and has positive Lebesgue
measure. The latter condition can be relaxed with some care (Exercise 27.10).

Let π be a probability measure supported on A. The continuous exponential

weights policy samples At from Pt = (1 − γ)P̃t + γπ, where P̃t is a measure
 27.3 Continuous Exponential Weights 321

supported on A defined by
R  P 
exp (a)
t−1
B
−η Ŷ
s=1 s da
P̃t (B) = R  P  . (27.5)
exp −η s=1 Ŷs (a) da
t−1
A

We will shortly see that the analysis in the previous section can be copied almost
verbatim to prove a regret bound for this strategy. But what has been bought here?
Rather than sampling from a discrete distribution on a large number of arms, we
now have to sample from a probability measure on a convex set. Sampling from
arbitrary probability measures is itself a challenging problem, but under certain
conditions there are polynomial time algorithms for this problem. The factors
that play the biggest role in the feasibility of sampling from a measure are (a)
the form of the measure or its density and (b) how the convex set is represented.
As it happens, the measure defined in the last display is log-concave, which
means that the logarithm of the density, with respect to the Lebesgue measure
on A, is a concave function.
Theorem 27.2. Let p(a) ∝ IA (a) exp(−f (a)) be a density with respect to the
Lebesgue measure on A such that f : A → R is a convex function. Then there
exists a polynomial-time algorithm for sampling from p, provided one can compute
the following efficiently:

1 (First-order information): ∇f (a) where a ∈ A.

2 (Euclidean projections): argminx∈A kx − yk2 where y ∈ Rd .
The probability distribution defined by Eq. (27.5) satisfies the first condition.
Efficiently computing a projection on to a convex set is a more delicate issue.
A general criterion that makes this efficient is access to a separation oracle,
which is a computational procedure φ that accepts a point x ∈ Rd as input and
responds φ(x) = true if x ∈ A and otherwise φ(x) = u, where hy, ui > hx, ui for
all y ∈ A (see Fig. 27.1).

x
A

Figure 27.1 Separation oracle returns the normal of a hyperplane that separates x from
A whenever x ∈
/ A. When x ∈ A, the separation oracle returns true.

Define log+ (x) = max(0, log(x)).

Theorem
R 27.3. Assume that A is compact, convex and has volume vol(A) =
A
da > 0. Then an appropriately tuned instantiation of the continuous
 27.4 Notes 322

exponential weights algorithm with Kiefer–Wolfowitz exploration has regret bounded

by
q
Rn ≤ 2d 3n(1 + log+ (2n/d)) .
The proof of Theorem 27.3 relies on the following proposition, which we leave
as an exercise (Exercise 27.11).
Proposition 27.4. Let K ⊂ Rd be a compact convex set with vol(K) > 0, u ∈ Rd
and let x∗ = argminx∈K hx, ui. Then,
    
vol(K) supx,y∈K hx − y, ui
log R ≤ d 1 + log + .
K
exp (−hx − x∗ , ui) dx d

The left-hand side in the above display is the logarithmic Laplace transform
of the uniform measure on K − {x∗ } evaluated at u.

Proof of Theorem 27.3 As before, choosing γ = dη ensures that |ηha, Ŷt i| ≤ 1 for
all a ∈ A (see the proof of Theorem 27.1). The standard argument (Exercise 27.9)
shows that
  
1   vol(A)
Rn ≤ E log R  P   + 3ηdn . (27.6)
η exp −η (Ŷ (a) − Ŷ (a∗ )) da
n
A t=1 t t

Pn
Using again that η|ha, Ŷt i| ≤ 1 and Proposition 27.4 with u = η t=1 Ŷt shows
that
d(1 + log+ (n/d)) q
Rn ≤ + 3ηdn ≤ 2d 3n(1 + log+ (2n/d)) .
η

27.4 Notes

1 A naive implementation of Algorithm 15 has computation complexity O(kd+d3 )

per round. There is also the one-off cost of computing the exploration
distribution, the complexity of which was discussed in Chapter 21. The real
problem is that k can be extremely large. This is especially true when the
action set is combinatorial. For example, when A = {a ∈ Rd : ai = ±1} is
the corners of the hypercube |A| = 2d , which is much too large unless the
dimension is small. Such problems call for a different approach that we present
in the next chapter and in Chapter 30.
2 It is not important to find exactly the optimal exploration distribution. All
that is needed is a bound on Eq. (27.4), which for the exploration distribution
based on the Kiefer–Wolfowitz theorem is just d. However, unlike in the finite
case, exploration is crucial and cannot be removed (Exercise 27.8).
 27.5 Bibliographic Remarks 323

√
3 The O( n) dependence of the regret on the horizon is not improvable, but the
linear dependence on the dimension is suboptimal for certain action sets and
optimal for others. An example where improvement is possible occurs when A
is the unit ball, which is analysed in the next chapter.
4 A slight modification of the set-up allows the action set to change in each
round, but where actions have identities. Suppose that k ∈ {1, 2, . . .} and At =
{a1 (t), . . . , ak (t)} and the adversary chooses losses so that maxa∈At |ha, yt i| ≤ 1
for all t. Then a straightforward adaptation of Algorithm 15 and Theorem 27.1
leads to an algorithm for which
" n #
X p
Rn = max E hAt − ai (t), yt i ≤ 2 3dn log(k) .
i∈[k]
t=1

The definition of the regret still compares the learner to the best single action
in hindsight, which makes it less meaningful than the definition of the regret
in Chapter 19 for stochastic linear bandits with changing action sets. These
differences are discussed in more detail in Chapter 29. See also Exercise 27.5.

27.5 Bibliographic Remarks

The results in Sections 27.1 and 27.2 follow the article by Bubeck et al. [2012],
with minor modifications to make the argument more pedagogical. The main
difference is that they used John’s ellipsoid over the action set for exploration,
which is only the right thing when John’s ellipsoid is also a central ellipsoid. Here
we use Kiefer–Wolfowitz, which is equivalent to finding the minimum volume
central ellipsoid containing the action set. Theorem 27.2, which guarantees
the existence of a polynomial time sampling algorithm for convex sets with
gradient information and projections is by Bubeck et al. [2015b]. We warn the
reader that these algorithms are not very practical, especially if theoretically
justified parameters are used. The study of sampling from convex bodies is quite
fascinating. There is an overview by Lovász and Vempala [2007], though it is a
little old. The continuous exponential weights algorithm is perhaps attributable
to Cover [1991] in the special setting of online learning called universal portfolio
optimisation. The first application to linear bandits is by Hazan et al. [2016].
Their algorithm and analysis are more complicated because they seek to improve
the computation properties by replacing the exploration distribution based on
Kiefer–Wolfowitz with an adaptive randomised exploration basis that can be
computed in polynomial time under weaker assumptions. Continuous exponential
weights for linear bandits using the core set of John’s ellipsoid for exploration
(rather than Kiefer–Wolfowitz) was recently p analysed by van der Hoeven et al.
[2018]. Another path towards an efficient O(d n log(·)) policy for convex action
sets is to use the tools from online optimisation. We explain some of these ideas
in more detail in the next chapter, but the reader is referred to the paper by
Bubeck and Eldan [2015].
 27.6 Exercises 324

27.6 Exercises

27.1 (‘Mixed’ Exp3 analysis) Prove Eq. (27.3).

27.2 (Dependence on the range of losses) Suppose that instead of assuming

yt ∈ L, we assume that yt ∈ {y ∈ Rd : supa∈A |ha, yi| ≤ b} for some known b > 0.
Modify the algorithm to accommodate this change, and explain how the regret
guarantee changes.

27.3 (Dependence on the range of losses (ii)) Now suppose that a < b
are known and yt ∈ {y ∈ Rd : ha, yi ∈ [a, b] for all a ∈ A}. How can you adapt
the algorithm now, and what is its regret?

27.4 (Inversion reverses Loewner orders) Let A, B ∈ Rd×d and suppose

that A  B and B is invertible. Show that A−1  B −1 .

27.5 (Changing action sets) Provide the necessary corrections to

Algorithm 15 and its analysis to prove the result claimed in Note 4.

Hint You will need to choose a new exploration distribution in every round.
Otherwise everything is more or less the same.
27.6 (Covering numbers for convex sets) For K ⊂ Rd let kxkK =
supy∈K |hx, yi|. Let A ⊂ Rd and L = {y : kykA ≤ 1}. Let N (A, ε) be the size of
the smallest subset C ⊆ A such that minx0 ∈C kx − x0 kL ≤ ε for all x ∈ A. Show
the following:

(a) When A = x ∈ Rd : kxkV −1 ≤ 1 , we have N (A, ε) ≤ (3/ε)d .
(b) When A is convex, bounded and span(A) = Rd we have N (A, ε) ≤ (3d/ε)d .
(c) For any bounded A ⊂ Rd we have N (A, ε) ≤ (6d/ε)d .

Hint For the first part, find a linear map from A to the Euclidean ball and
use the fact that the Euclidean ball can be covered with a set of size (3/ε)d . For
the second part use the fact that for any symmetric, convex and compact set K
there exists an ellipsoid E = {x : kxkV ≤ 1} such that E ⊆ K ⊆ dE.
27.7 (Low rank action sets (i)) In the definition of the algorithm and the
proof of Theorem 27.1, we assumed that A spans Rd and that it has positive
Lebesgue measure. Show that this assumption may be relaxed by carefully
adapting the algorithm and analysis.

27.8 (Necessity of exploration) We saw in Chapter 11 that the exponential

weights algorithm achieved near-optimal regret without mixing additional
exploration. Show that exploration is crucial here. More precisely, construct a
finite action set A and reward sequence yt ∈ L such that the regret of Algorithm 15
with γ = 0 becomes very poor (even with η optimally tuned) relative to the
optimal choice.
 27.6 Exercises 325

27.9 (Continuous exponential weights) Complete the missing steps in the

proof of Theorem 27.3.

27.10 (Low rank action sets (ii)) In the definition of the algorithm and
the proof of Theorem 27.3, we assumed that A spans Rd and that it has positive
Lebesgue measure. Show that this assumption may be relaxed by carefully
adapting the algorithm and analysis.

27.11 (Volume bounds) Prove Proposition 27.4.

28 Follow-the-Regularised-Leader and
Mirror Descent

In the last chapter, we showed that if A ⊂ Rd has k elements, then the regret of
Exp3 with a careful exploration distribution has regret
p
Rn = O( dn log(k)) .
We also showed the continuous version of this algorithm has regret at most
p
Rn = O(d n log(n)) .
Although this algorithm can often be made to run in polynomial time, the degree
tends to be high and the implementation complicated, making the algorithm
impractical. In many cases this can be improved, both in terms of the regret and
computation. In this chapter we demonstrate this in the case when A is the unit
ball by showing that for this case pthere is an efficient, low-complexity algorithm
for which the regret is Rn = O( dn log(n)). More importantly, however, we
introduce a pair of related algorithms called follow-the-regularised-leader and
mirror descent, which are powerful tools for the design and analysis of bandit
algorithms. In fact, the exponential weights algorithm turns out to be a special case.

28.1 Online Linear Optimisation

Mirror descent originated in the convex optimisa-

tion literature. The idea has since been adapted
to online learning and specifically to online lin-
ear optimisation. Online linear optimisation
is the full information version of the adversarial
linear bandit, where at the end of each round the
learner observes the full vector yt . Let A ⊂ Rd Figure 28.1 Mirror descent is a
be a convex set and L ⊂ Rd be an arbitrary set modern art, as well as science
called the loss space. Let y1 , . . . , yn be a sequence of loss vectors with yt ∈ L
for t ∈ [n]. In each round the learner chooses at ∈ A and subsequently observes
the vector yt . The regret relative to a fixed comparator a ∈ A is
n
X
Rn (a) = hat − a, yt i ,
t=1
 28.1 Online Linear Optimisation 327

and the regret is Rn = maxa∈A Rn (a). We emphasise that the only difference
relative to the adversarial linear bandit is that now yt is observed rather than
hat , yt i. Actions are not capitalised in this section because the algorithms presented
here do not randomise.

Mirror descent
The basic version of mirror descent has two extra parameters beyond n and A. A
learning rate η > 0 and a convex function F : Rd → R̄ with domain D = dom(F ).
Usually F will be Legendre. The function F is called a potential function or
regulariser. In the first round, mirror descent predicts

a1 = argmina∈A F (a) . (28.1)

Subsequently it predicts

at+1 = argmina∈A (ηha, yt i + DF (a, at )) , (28.2)

where DF (a, at ) is the F -induced Bregman divergence between a and at . Implicit

in the definition is that a1 , a2 , . . . are well-defined. The reader is invited to
construct examples when this is not the case (Exercise 28.2). A simple case when
(at )nt=1 are well-defined is when A is compact and F is Legendre.

Follow-the-Regularised-Leader
Like mirror descent, follow-the-regularised-leader depends on a convex potential
F with domain D = dom(F ) and predicts a1 = argmina∈A F (a). In subsequent
rounds t ∈ [n], the predictions are
t
!
X
at+1 = argmina∈A η ha, ys i + F (a) . (28.3)
s=1

The intuition is that the algorithm chooses at+1 to be the action that performed
best in hindsight with respect to the regularised loss. Again, the definition of
follow-the-regularised-leader implicitly assumes that (at )nt=1 are well-defined. As
for mirror descent, the regularisation serves to stabilise the algorithm, which
turns out to be a key property of good algorithms for online linear prediction.

Follow-the-leader chooses the action that appears best in hindsight,

Pt
at+1 = argmina∈A s=1 ha, ys i. In general this algorithm is not well suited
for online linear optimisation because the absence of regularisation makes it
unstable (Exercise 28.4).

Equivalence of Mirror Descent and Follow-the-Regularised-Leader

At first sight these algorithms do not look that similar. To clarify matters, let us
suppose that F is Legendre with domain D ⊆ A. In this setting, mirror descent
 28.1 Online Linear Optimisation 328

and follow-the-regularised-leader are identical. To see this, let

Φt (a) = ηha, yt i + DF (a, at ) = ηha, yt i + F (a) − F (at ) − ha − at , ∇F (at )i .

Now mirror descent chooses at+1 to minimise Φt . The reader should check that
the assumption that F is Legendre on domain D ⊆ A implies that the minimiser
occurs in the interior of D ⊆ A and that ∇Φt (at+1 ) = 0 (see Exercise 28.1). This
means that ηyt = ∇F (at ) − ∇F (at+1 ), and so
t
X t
X
∇F (at+1 ) = −ηyt + ∇F (at ) = ∇F (a1 ) − η ys = −η ys ,
s=1 s=1

where the last equality is true because a1 is chosen as the minimiser of F in

A ∩ D = D, and again the fact that F is Legendre ensures this minimum occurs
at an interior point where the gradient vanishes. Follow the regularised leader
Pt
chooses at+1 to minimise Φ0t (a) = η s=1 ha, ys i + F (a). The same argument
shows that ∇Φ0t (at+1 ) = 0, which means that
t
X
∇F (at+1 ) = −η ys .
s=1

The last two displays and the fact that the gradient for Legendre functions is
invertible shows that mirror descent and follow-the-regularised-leader are the
same in this setting.

The equivalence between these algorithms is far from universal. First of

all, it does not generally hold when F is not Legendre or its domain
is larger than A. Second, in many applications of these algorithms, the
learning rate or potential change with time, and in either case the
algorithms will typically produce different action sequences. For example, if
a learning rate ηt is used rather than η in the definition of Φt , then mirror
Pt
descent chooses ∇F (at+1 ) = − s=1 ηs ys , while follow-the-regularised-
Pt
leader chooses ∇F (at+1 ) = −ηt s=1 ys . We return to this issue in the
notes and exercises.

Example 28.1. Let A = Rd and F (a) = 12 kak22 . Then ∇F (a) = a and

D(a, at ) = 12 ka − at k22 . Clearly F is Legendre and D = A, so mirror descent and
follow-the-regularised-leader are the same. By simple calculus we see that
1
at+1 = argmina∈Rd ηha, yt i + ka − at k22 = at − ηyt ,
2
which may be familiar as online gradient descent with linear losses. For the
extension to nonlinear convex loss functions, see Note 12.
1 2
Example 28.2. Let A be a compact convex subset of Rd and F (a) = 2 kak2 .
 28.1 Online Linear Optimisation 329

Then mirror descent chooses

1
at+1 = argmina∈A ηha, yt i + ka − at k22 = Π(at − ηyt ) , (28.4)
2
where Π(a) is the Euclidean projection of a on to A. This algorithm is usually
called online projected gradient descent. On the other hand, for follow-the-
regularised-leader we have
!
1
Xt Xt
at+1 = argmina∈A η ha, ys i + ka − at k22 = Π −η ys ,
s=1
2 s=1

which may be a different choice than that of mirror descent.

Example 28.3. The exponential weights algorithm that appeared in various

forms on numerous occasions in earlier chapters is a special case of mirror descent
corresponding to choosing the constraint set A as the simplex in Rd and choosing
F to be the unnormalised negentropy function of Example 26.5. In this case
follow-the-regularised-leader chooses
t
X d
X
at+1 = argmina∈A η ha, ys i + ai log(ai ) − ai .
s=1 i=1

You will show in Exercise 28.8 that

 P 
exp −η s=1 ysi
t

at+1,i =P  P . (28.5)
exp
d t
j=1 −η s=1 ysj

28.1.1 A Two-Step Process for Implementation

Solving the optimisation problem in Eq. (28.2) is often made easier by using
Theorem 26.15 from Chapter 26. Assume F is Legendre and A is compact and
non-empty, and suppose that

∇F (a) − ηy ∈ int(dom(F ∗ )) for all a ∈ A ∩ D and y ∈ L . (28.6)

Then the solution to Eq. (28.2) can be found using the following two-step
procedure:

ãt+1 = argmina∈D ηha, yt i + DF (a, at ) and (28.7)

at+1 = argmina∈A DF (a, ãt+1 ) . (28.8)

Eq. (28.6) means the first optimisation problem can be evaluated explicitly as
the solution to

ηyt + ∇F (ãt+1 ) − ∇F (at ) = 0 . (28.9)

Since F is Legendre, Theorem 26.6 shows that ∇F is a bijection between

int(dom(F )) and int(dom(F ∗ )), which means that ãt+1 = (∇F )−1 (∇F (at ) − ηyt ).
 28.2 Regret Analysis 330

The optimisation problem in Eq. (28.8) is usually harder to calculate analytically,

but there are important exceptions, as we shall see.

All potentials and losses that appear in positive results in this book guarantee
that mirror descent (and also follow-the-regularised-leader) are well defined,
and that condition in Eq. (28.6) holds.

The two-step implementation of mirror descent also explains its name. The
update in round t can be seen as transforming the action at ∈ A into the ‘mirror’
(dual) space using ∇F , where it is combined with the most recent (scaled) loss
ηyt . Then ∇F −1 is used to transform the updated vector back to the original
(primal) space. The function ∇F is called the mirror map.

The same idea works for follow-the-regularised-leader. Assuming F is

Pt
Legendre, A is compact and nonempty and −η s=1 ys ∈ int(dom(F ∗ )),
then for follow-the-regularised-leader
t
!!
X
at+1 = ΠA,F ∇F −1
−η ys ,
s=1

where ΠA,F is the projection on to A with respect to DF as described in

Section 26.6.

Some of the differences between follow-the-regularised-leader and mirror descent

are illustrated in Fig. 28.2, which shows how the algorithms differ once projections
start to occur.

28.2 Regret Analysis

Although mirror descent and follow-the-regularised-leader are not the same, the
bounds presented here are identical. The theorem for mirror descent has two
parts, the first of which is a little stronger than the second. To minimise clutter,
we abbreviate DF by D.

Theorem 28.4 (Mirror descent regret bound). Let η > 0 and F be Legendre with
domain D and A ⊂ Rd be a non-empty convex set with int(dom(F )) ∩ A = 6 ∅. Let
a1 , . . . , an+1 be the actions chosen by mirror descent, which are assumed to be
well-defined. Then, for any a ∈ A, the regret of mirror descent is bounded by

F (a) − F (a1 ) X 1X
n n
Rn (a) ≤ + hat − at+1 , yt i − D(at+1 , at ) .
η t=1
η t=1

Furthermore, suppose that Eq. (28.6) holds and ã2 , ã3 , . . . , ãn+1 are given by
 28.2 Regret Analysis 331

Constraint set Primal Dual

a1 ∇F

−ηy1
∇F −1
a2
∇F
−ηy2
ΠA,F
a3 ∇F −1

∇F −ηy3
−ηy3
afollow-the-regularised-leader
4

∆F −1

∆F −1
aMD
4
∆F

Figure 28.2 Illustration of follow-the-regularised-leader and mirror descent. The

constraint set is A, and the function ΠA,F is the projection on to A with respect
to the Bregman divergence induced by Legendre function F . Follow-the-regularised-
leader accumulates the scaled losses in the dual space, mapping back to the primal
using the inverse map ∇F −1 . Mirror descent computes the next iterate by at+1 =
ΠA,F (∇F −1 (∇F (at ) − ηyt ). The algorithms generally behave differently in the presence
of projections. In the figure, the algorithms behave the same until the fourth iterate,
after which the projection that appeared in the computation of the third iterate breaks
the equivalence.

Eq. (28.7). Then,

!
1
n
X
Rn (a) ≤ F (a) − F (a1 ) + D(at , ãt+1 ) .
η t=1

Proof Fix a ∈ A. The result trivially holds when a 6∈ D. Hence, we assume that
a ∈ D. For the first part of the claim, we split the inner product:

hat − a, yt i = hat − at+1 , yt i + hat+1 − a, yt i .

In Exercise 28.1, you will show that at ∈ int(dom(F )), and hence the Bregman
divergence D(b, at ) = F (b) − F (at ) − hb − at , ∇F (at )i for any b ∈ dom(F ). By
definition, at+1 = argminb∈A ηhb, yt i + D(b, at ). Hence, the first-order optimality
conditions for at+1 (Proposition 26.14) show that

ha − at+1 , ηyt + ∇F (at+1 ) − ∇F (at )i ≥ 0 .

 28.2 Regret Analysis 332

Reordering and using the definition of the Bregman divergence,

1
hat+1 − a, yt i ≤ ha − at+1 , ∇F (at+1 ) − ∇F (at )i
η
1
= (D(a, at ) − D(a, at+1 ) − D(at+1 , at )) .
η
Using this, along with the definition of the regret,
n
X
Rn = hat − a, yt i
t=1

1X
Xn n
≤ hat − at+1 , yt i + (D(a, at ) − D(a, at+1 ) − D(at+1 , at ))
t=1
η t=1
!
1
Xn n
X
= hat − at+1 , yt i + D(a, a1 ) − D(a, an+1 ) − D(at+1 , at )
t=1
η t=1

F (a) − F (a1 ) 1 X
Xn n
≤ hat − at+1 , yt i + − D(at+1 , at ) , (28.10)
t=1
η η t=1

where the final inequality follows from the fact that D(a, an+1 ) ≥ 0 and
D(a, a1 ) ≤ F (a) − F (a1 ), the latter of which is true by the first-order optimality
conditions for a1 = argminb∈A F (b). To see the second part, note that
1
hat − at+1 , yt i = hat − at+1 , ∇F (at ) − ∇F (ãt+1 )i
η
1
= (D(at+1 , at ) + D(at , ãt+1 ) − D(at+1 , ãt+1 ))
η
1
≤ (D(at+1 , at ) + D(at , ãt+1 )) .
η
The result follows by substituting this into Eq. (28.10).

The assumption that a1 minimises the potential was only used to bound
D(a, a1 ) ≤ F (a) − F (a1 ). For a different initialisation, the following bound
still holds:
!
1 X n
Rn (a) ≤ D(a, a1 ) + D(at , ãt+1 ) . (28.11)
η t=1

As we shall see in Chapter 31, this is useful when using mirror descent to
analyse non-stationary bandits.

The first part of Theorem 28.4 also holds for follow-the-regularised-leader as

stated in the next result, the proof of which is left for Exercise 28.5.

Theorem 28.5 (Follow-the-regularised-leader regret bound). Let η > 0, F

be convex with domain D, A ⊆ Rd be a non-empty convex set. Assume that
 28.2 Regret Analysis 333

a1 , . . . , an+1 chosen by follow-the-regularised-leader are well defined. Then, for

any a ∈ A, the regret of follow-the-regularised-leader is bounded by
F (a) − F (a1 ) X 1X
n n
Rn (a) ≤ + hat − at+1 , yt i − D(at+1 , at ) .
η t=1
η t=1

We now give two applications of the regret bound of Theorem 28.4 for mirror
descent. The same results would hold for the same problems for FTRL just in this
case we would need to use Theorem 28.5. Let diamF (A) = maxa,b∈A F (a) − F (b)
be the diameter of A with respect to F .
Proposition 28.6 (Regret on the unit ball). Let A = B2d = {a ∈ Rd : kak2 ≤ 1}
be the standard unit ball and assume
pyt ∈ B2 for all t. Then mirror descent with
d
1 2
potential F (a) = 2 kak2 and η = 1/n is well defined and its regret satisfies
√
Rn ≤ n.
Proof That mirror descent is well defined follows by a direct calculation (cf.
Example 28.1). By Eq. (28.9), we have ãt+1 = at − ηyt so
1 η2
D(at , ãt+1 ) = kãt+1 − at k22 = kyt k22 .
2 2
Therefore, since diamF (A) = 1/2 and kyt k2 ≤ 1 for all t,

diamF (A) η X 1
n
ηn √
Rn ≤ + kyt k22 ≤ + = n.
η 2 t=1 2η 2

Proposition 28.7 (Regret on the simplex). Let A = Pd−1 be the probability

simplex and yt ∈ L = [0, 1]d forpall t. Then mirror descent with the unnormalised
p potential and η = 2 log(d)/n is well defined and its regret satisfies
negentropy
Rn ≤ 2n log(d).
Proof That mirror descent is well-defined follows because the simplex is compact.
The Bregman divergence with respect to the unnormalised negentropy potential
Pd
for a, b ∈ A is D(a, b) = i=1 ai log(ai /bi ). Therefore,

F (a) − F (a1 ) X 1X
n n
Rn (a) ≤ + hat − at+1 , yt i − D(at+1 , at )
η t=1
η t=1
log(d) X 1X1
n n
≤ + kat − at+1 k1 kyt k∞ − kat − at+1 k21
η t=1
η t=1
2
log(d) η X log(d) ηn p
n
≤ + kyt k2∞ ≤ + = 2n log(d) ,
η 2 t=1 η 2

where the first inequality follows from Theorem 28.4, the second from Pinsker’s
inequality and the facts that diamF (A) = log(d). In the third inequality, we
used ‘optimise to bound’. In particular, we used that for any a ∈ R and b > 0,
maxx∈R ax − bx2 /2 = a2 /(2b). The last inequality follows from the assumption
that kyt k∞ ≤ 1.
 28.3 Application to Linear Bandits 334

The last few steps in the above proof are so routine that we summarise their
use in a corollary, the proof of which we leave to the reader (Exercise 28.6).

Corollary 28.8. Let F be a Legendre potential and k · kt be a norm on Rd for

each t ∈ [n] such that DF (at+1 , at ) ≥ 12 kat+1 − at k2t . Then the regret of mirror
descent or follow-the-regularised-leader satisfies

diamF (A) η X
n
Rn ≤ + kyt k2t∗ ,
η 2 t=1

where kykt∗ = maxx:kxkt ≤1 hx, yi is the dual norm of k · kt .

It often happens that the easiest way to bound the regret of mirror descent is to
find a norm that satisfies the conditions of Corollary 28.8. Often, Theorem 26.13
provides a good approach.

Example 28.9. To illustrate a suboptimal application of mirror descent,

suppose we had chosen F (a) = 12 kak22 in the setting of Proposition 28.7. Then
DF (at+1 , at ) = 12 kat+1 −at k22 suggests choosing k·kt to be the standard Euclidean
norm. Since diamF (A) = 1/2 and k · k2∗ = k · k2 , applying Corollary 28.8 shows
that

1
n
ηX
Rn ≤ + kyt k22 .
2η 2 t=1

2
But now we√ see that kyt k2 canpbe as large as d, and tuning η would lead to a
rate of O( nd) rather than O( n log(d)).

Both Theorems 28.4 and 28.5 were presented for the oblivious case where
(yt )nt=1 are chosen in advance. This assumption was not used, however, and in
fact the bounds continue to hold when yt is chosen strategically as a function
of a1 , y1 , . . . , yt−1 , at . This is analogous to how the basic regret bound for
exponential weights continues to hold in the face of strategic losses. But
be cautioned, this result does not carry immediately to the application of
mirror descent to bandits, as discussed at the end in Note 9.

28.3 Application to Linear Bandits

We now show how mirror descent and follow-the-regularised-leader can be used to

construct algorithms for adversarial linear bandit problems. Like in the previous
chapter, the adversary chooses a sequence of vectors y1 , . . . , yn with yt ∈ L ⊂ Rd .
In each round the learner chooses At ∈ A ⊂ Rd and observes hAt , yt i. The regret
 28.3 Application to Linear Bandits 335

relative to action a ∈ A is
" n #
X
Rn (a) = E hAt − a, yt i .
t=1

The regret is Rn = maxa∈A Rn (a). The application of mirror descent and follow-
the-regularised-leader to linear bandits is straightforward. The only difficulty
is that the learner does not observe yt but instead hAt , yt i. The solution is to
replace yt with an estimator, which is typically some kind of importance-weighted
estimator as in the previous chapter. Because estimation of yt is only possible
using randomisation, the algorithm cannot play the suggested action of mirror
descent, but instead plays a distribution over actions with the same mean as the
proposed action. This is often necessary anyway, when A is not convex. Since
the losses are linear, the expected additional regret by playing according to the
distribution vanishes. The algorithm is summarised in Algorithm 16. We have
switched to capital letters because the actions are now randomised.

Theorem 28.10 (Regret of Mirror-Descent and FTRL with bandit feedback).

Suppose that Algorithm 16 is run with Legendre potential F , convex action set
A ⊂ Rd and learning rate η > 0 such that the loss estimators are unbiased:
E[Ŷt | Āt ] = yt for all t ∈ [n]. Then the regret for either variant of Algorithm 16,
provided that they are well defined, is bounded by
" #
F (a) − F (Ā1 ) X 1X
n n
Rn (a) ≤ E + hĀt − Āt+1 , Ŷt i − D(Āt+1 , Āt ) .
η t=1
η t=1

Furthermore, letting
Ãt+1 = argmina∈dom(F ) ηha, Ŷt i + DF (a, Āt )

and assuming that −η Ŷt + ∇F (a) ∈ ∇F (dom(F )) for all a ∈ A almost surely,
the regret of the mirror descent variation satisfies
diamF (A) 1 X 
n

Rn ≤ + E D(Āt , Ãt+1 ) .
η η t=1

Proof Using the definition of the algorithm and the assumption that Ŷt is
unbiased given Āt and that Pt has mean Āt leads to
     h h  ii

E [hAt , yt i] = E hĀt , yt i = E E hĀt , yt i | Āt = E E hĀt , Ŷt i  Āt ,

where the last equality used the linearity of expectations. Hence,

" n # " n #
X X
Rn (a) = E hAt , yt i − ha, yt i = E hĀt − a, Ŷt i ,
t=1 t=1

which is the expected random regret of mirror descent or follow-the-regularised-

leader on the recursively constructed sequence Ŷt . The result follows from
Theorem 28.4 or Theorem 28.5 and the note at the end of the last section
 28.4 Linear Bandits on the Unit Ball 336

that says these theorems continue to hold even for recursively constructed loss
sequences.

1: Input Legendre potential F , action set A and learning rate η > 0

2: Choose Ā1 = argmina∈A∩dom(F ) F (a)
3: for t = 1, . . . , n do
4: Choose measure Pt on A with mean Āt
5: Sample action At from Pt and observe hAt , yt i
6: Compute estimate Ŷt of the loss vector yt
7: Update:
Āt+1 = argmina∈A∩dom(F ) ηha, Ŷt i + DF (a, Āt ) (Mirror descent)
t
X
Āt+1 = argmina∈A∩dom(F ) η ha, Ŷs i + F (a)
s=1
(follow-the-regularised-leader)
8: end for
Algorithm 16: Online stochastic mirror descent/follow-the-regularised-leader.

28.4 Linear Bandits on the Unit Ball

To illustrate the power of these methods, we return to adversarial linear bandits

and the special case where the action set is the unit ball. In the previous chapter,
we showed that continuous exponential weights on the unit ball with Kiefer-
Wolfowitz exploration has a regret of
p
Rn = O(d n log(n)) .
Surprisingly, follow-the-regularised-leader
√ with a carefully chosen potential
improves on this bound by a factor of d.
For the remainder of this section, let k · k = k · k2 be the standard Euclidean
norm and A = B2d be the standard unit ball. In order to instantiate follow-the-
regularised-leader we need a potential, a sampling rule, an unbiased estimator and
a learning rate. Note that the only source of randomness is the randomisation in
the algorithm. Hence, let Et [·] = E[· | A1 , . . . , At−1 ]. We start with the sampling
rule and estimator. Recall that in round t we need to choose a distribution on
A with mean Āt and sufficient variability that the variance of the estimator is
not too large. Given the past, let Et and Ut be independent, where Et ∈ {0, 1} is
such that Et [Et ] = 1 − kĀt k and Ut is uniformly distributed on {±e1 , . . . , ±ed }.
The algorithm chooses
(1 − Et )Āt
At = Et Ut + .
kĀt k
 28.4 Linear Bandits on the Unit Ball 337

In other words, Et = 1 indicates that the algorithm explores, which happens with
probability 1 − kĀt k. Clearly, Et [At ] = Āt . (The sampling distribution Pt is just
the law of At given the past, which remains implicit.) For the estimator we use a
variant of the importance-weighted estimator from the last chapter:

d Et At hAt , yt i
Ŷt = . (28.12)
1 − kĀt k

The reader can check for themself that this estimator is unbiased. Next, we
inspect the contents of our magician’s hat and select the potential

F (a) = − log (1 − kak) − kak .

There is one more modification. Rather than instantiating follow-the-regularised-

leader with action set A, we use à = {x ∈ Rd : kxk2 ≤ r}, where r < 1 is a
radius to be tuned subsequently. The reason for this modification is to control
the variance of the estimator in Eq. (28.12), which blows up as Āt gets close to
the boundary. You will show in Exercise 28.7 that
! t−1
−η L̂t−1 X
Āt = Π with L̂t−1 = Ŷs , (28.13)
1 + ηkL̂t−1 k s=1

where Π(x) is the projection operator on to à with respect to k · k2 .

1: Input Learning rate η > 0

2: for t = 1, . . . , n do
3: Compute
! t−1
−η L̂t−1 X
Āt = Π with L̂t−1 = Ŷs
1 + ηkL̂t−1 k s=1

4: Sample Et ∈ {0, 1} from Binomial with bias 1 − kĀt k and Ut uniformly

on {e1 , . . . , ed }
5: Play action At = Et Ut + (1−E t )Āt
kĀt k
d Et At hAt , yt i
6: Observe hAt , yt i and estimate loss vector Ŷt =
1 − kĀt k
7: end for
Algorithm 17: Follow-the-regularised-leader for linear bandits on the unit ball

Theorem 28.11. Assume that (yt )nt=1 are a sequence of losses such that kyt k2 ≤ 1
for all t. Suppose that Algorithm 16 is run using the sampling rule, estimator
and potential as describedp above, shrunken action set à with r = 1 − 2ηd where
the learning rate is η = p log(n)/(3dn). Then, the algorithm is well defined and
its regret satisfies Rn ≤ 2 3nd log(n).
 28.4 Linear Bandits on the Unit Ball 338

You might notice that in some regimes this is smaller than the lower bound
for stochastic linear bandits (Theorem 24.2). There is no contradiction
because the adversarial and stochastic linear bandit models are actually
quite different. More details are in Chapter 29.

Proof That the algorithm is well defined follows because à is compact. Let
Pn
a∗ = argmina∈A t=1 ha, yt i be the optimal action. Then
" n # n
X X
Rn = E hAt − ra , yt i +
∗
hra∗ − a∗ , yt i ≤ Rn (ra∗ ) + (1 − r)n ,
t=1 t=1

where the inequality follows from the definition of A and Cauchy–Schwarz. By

Theorems 26.13 and 28.10,
" n #
diamF (Ã) η X
2
Rn (ra ) ≤
∗
+ E kŶt k(∇2 F (Zt ))−1 ,
η 2 t=1

where Zt ∈ [Āt , Āt+1 ] lies on the chord connecting Āt and Āt+1 . The algorithm is
stable in the sense that no matter how the losses are chosen, Āt+1 cannot
be too far from Āt . This also means that Zt is close to Āt . By definition,
ηkŶt k ≤ ηd/(1 − r) = 1/2. Combining this with Eq. (28.13) shows that
 
1 − kZt k 1 − kαĀt + (1 − α)Āt+1 k 1 − kĀt+1 k
≤ sup = max 1,
1 − kĀt k α∈[0,1] 1 − kĀt k 1 − kĀt k
( ) ( )
1 + ηkL̂t−1 k 1 + ηkL̂t−1 k
≤ max 1, ≤ max 1, ≤ 2.
1 + ηkL̂t k 1/2 + ηkL̂t−1 k
Here, the second inequality is proved by noting that if the maximum is not one,
kĀt+1 k < kĀt k. The next step is to find the Hessian of F , which is
I aa> I
∇2 F (a) = +  .
1 − kak kak(1 − kak)2 1 − kak
Therefore, (∇2 F (a))−1 ≤ (1 − kak)I, and so
h i h i  
2 2 2 (1 − kZt k)Et hUt , yt i2
E kŶt k(∇2 F (Zt ))−1 ≤ E (1 − kZt k)kŶt k = d E ≤ 2d .
(1 − kĀt k)2
The diameter satisfies diamF (Ã) ≤ log(1/(1 − r)), and hence
 
1 1
Rn ≤ (1 − r)n + log + ηnd
η 1−r
 
1 1
= log + 3ηnd
η 2ηd
p
≤ 2 3nd log(n) ,
where the last two relations follow from the choices of r and η, respectively.
 28.5 Notes 339

We could have used mirror descent rather than follow-the-regularised-

leader with a slightly more complicated proof and the same bound except
for constants. Using continuous exponential weights √ and the analysis in
Section 27.3 would yield a bound that is a factor of d worse than the above,
and we believe that this cannot be improved.

28.5 Notes

1 Our assumptions on the potential and action set in the analysis of mirror
descent (Theorem 28.4) can be relaxed significantly. What is important is
that F is convex and the directional derivative v 7→ ∇v F (x) is linear for
all values for which it exists. Our assumptions are chosen to ensure that
at ∈ int(dom(F )), which for Legendre F means that ∇F (at ) exists, and hence
∇v F (at ) = hv, ∇F (at )i is linear. A comprehensive examination of various
generalisations is given by Joulani et al. [2017]. For follow-the-regularised-
leader, convexity of F suffices, as you will show using directional derivatives in
Exercise 28.5.
2 Finding at+1 for both mirror descent and follow-the-regularised-leader requires
solving a convex optimisation problem. Provided the dimension is not too
large and the action set and potential are reasonably nice, there exist practical
approximation algorithms for this problem. The two-step process described in
Eqs. (28.7) and (28.8) is sometimes an easier way to go. Usually (28.7) can
be solved analytically, while (28.8) can be quite expensive. In some important
special cases, however, the projection step can be written in closed form or
efficiently approximated.
3 We saw that follow-the-regularised-leader
p with a carefully chosen potential
function achieves O( dn log(n)) regret on the `2 -ball. On the `∞ ball
√
(hypercube), the optimal regret is O(d n). Interestingly, as n tends to infinity
the optimal dependence on the dimension for A = Bpd = {x ∈ Rd : kxkp ≤ 1}
√
with p ≥ 1 is either d or d with a complete classification given by Bubeck
et al. [2018].
4 Adversarial linear bandits with A = Pk−1 are essentially equivalent to k-armed
adversarial bandits.
√ There exists a potential such that the resulting algorithm
satisfies Rn = O( kn), which matches the lower bound up to constant factors
√
and shaves a factor of log k from the upper bounds presented in Chapters 11
and 12. For more details, see Exercise 28.15.
5 Most of the bounds proven for adversarial bandits have a worst-case flavour.
The tools in this chapter can often be applied to prove adaptive bounds. In
Exercise 28.14, you will analyse a simple algorithm for k-armed adversarial
 28.5 Notes 340

bandits for which

v
u ! 
u Xn n
Rn = O tk 1 + min yta log .
a∈[k]
t=1
k

Bounds of this kind are called first-order bounds [Allenberg et al., 2006,
Abernethy et al., 2012, Neu, 2015b, Wei and Luo, 2018]. The log(n/k) term
can be improved to log(k) using a more sophisticated algorithm/analysis.
6 Both mirror descent and follow-the-regularised-leader depend on the potential
function. Currently there is no characterisation of exactly what this potential
should be or how to find it. At least in the full information setting, there are
quite general universality results showing that if a certain regret is achievable
by some algorithm, then that same regret is nearly achievable by mirror descent
with some potential [Srebro et al., 2011]. In practice this result is not useful for
constructing new potential functions, however. There have been some attempts
to develop ‘universal’ potential functions that exhibit nice behaviour for any
action sets [Bubeck et al., 2015b, and others]. These can be useful, but as yet
we do not know precisely what properties are crucial, especially in the bandit
case.
7 When the horizon is unknown, the learning rate cannot be tuned ahead of time.
One option is to apply the doubling trick. A more elegant solution is to use a
decreasing schedule of learning rates. This requires an adaptation of the proofs
of Theorems 28.4 and 28.5, which we outline in Exercises 28.11 and 28.12. This
is one situation where mirror descent and follow-the-regularised-leader are not
the same and where the latter algorithm is usually to be preferred.
8 In much of the literature the potential is chosen in such a way that mirror descent
and follow-the-regularised-leader are the same algorithm. For historical reasons,
the name mirror descent is more commonly used in the bandit community.
Unfortunately ‘mirror descent’ is often used, sometimes with qualifiers, when
the algorithm being analysed is actually follow-the-regularised-leader. This is
confusing and makes it hard to identify for which algorithm the results actually
hold. Naming aside, we encourage the reader to keep both algorithms in mind,
since the analysis of one or the other can sometimes be slightly easier.
9 Mirror descent and follow-the-regularised-leader are used as modules for
converting loss sequences to distributions. Since these losses depend on past
actions, it is crucial that both algorithm are well-behaved in the full-information
setting when the losses are chosen non-obliviously. This does not translate to
Pn
the bandit setting for a subtle reason. Let R̂n (a) = t=1 hAt − a, yt i be the
random regret so that
  " n n
#
X X
Rn = E max R̂n (a) = E hAt , yt i − min ha, yt i .
a∈A a∈A
t=1 t=1

The second sum is constant when the losses are oblivious, which means the
maximum can be brought outside the expectation, which is not true if the loss
 28.5 Notes 341

vectors are non-oblivious. It is still possible to bound the expected loss relative
to a fixed comparator a so that
" n #
X
Rn (a) = E hAt − a, yt i ≤ B ,
t=1

where B is whatever bound obtained from the analysis presented above. Using
maxa R̂n (a) ≤ maxa R̂n (a) − Rn (a) + maxa Rn (a) shows that
   
Rn = E max R̂n (a) ≤ B + E max R̂n (a) − Rn (a) .
a∈A a∈A

The second term on the right-hand side can be bounded using tools from
√
empirical process theory, but the resulting bound is O( n) only if V[R̂n (a)] =
O(n). In general, however, the variance can be much larger (for an example, see
Exercise 11.6). We emphasise again that the non-oblivious regret is a strange
measure because it does not capture the reactive nature of the environment.
The details of the application of empirical process theory is beyond the scope
of this book. For an introduction to that topic, we recommend the books by
van der Vaart and Wellner [1996], van de Geer [2000], Boucheron et al. [2013]
and Dudley [2014]. p
10 The price of bandit information on the unit ball is an extra d log(n) (compare
Proposition 28.6 and Theorem 28.11). Except for log factors, this is also true for
the simplex (Proposition
√ 28.7 and Note 4). One might wonder if the difference
is always about d, but this is not true. The price of bandit information can
be as high as Θ(d). Overall the dimension dependence in the regret in terms of
the action set is still not well understood except for special cases.
11 The poor behaviour of follow-the-leader in the full information setting depends
on (a) the environment being adversarial rather than stochastic and (b) the
action set having sharp corners. When either of these factors is missing, follow-
the-leader is a reasonable choice [Huang et al., 2017b]. Note that with bandit
feedback, the failure is primarily due to a lack of exploration (Exercises 4.11
and 4.12).
12 A generalisation of online linear optimisation is online convex optimisation,
where the adversary secretly chooses a sequence of convex functions f1 , . . . , fn .
In each round the learner chooses at ∈ A and observes the entire function ft .
As usual, the regret is relative to a ∈ A is
n
X
Rn (a) = ft (at ) − ft (a) .
t=1

One way to tackle this problem is to linearise the loss functions. Let
yt = ∇ft (at ). Then, by convexity of the loss functions,
n
X
Rn (a) ≤ hat − a, yt i ,
t=1

which shows that an algorithm for online linear optimisation can be used to
 28.6 Bibliographic Remarks 342

analyse the more general case. Now look again at Example 28.1 and notice
that online mirror descent with a quadratic potential and linearised losses is
really the same gradient descent we know and love. Online convex optimisation
is a rich topic by itself. We refer the interested reader to the books by Shalev-
Shwartz [2012] and Hazan [2016].
13 There is a nice application of online linear optimisation to minimax theorems.
Let X and Y be arbitrary sets. For any function f : X × Y → R,

inf sup f (x, y) ≥ sup inf f (x, y) .

x∈X y∈Y y∈Y x∈X

Under certain conditions, the inequality becomes an equality. Theorems

guaranteeing this are called minimax theorems. The following result by Sion
[1958] is one of the more generic variants. The statement uses notions of
quasi-convexity and semi-continuity, which are defined in the next note.

Theorem 28.12 (Sion’s minimax theorem). Suppose that X and Y are convex
subsets of linear topological spaces with at least one of X or Y compact. Let
f : X × Y → R be a function such that f (·, y) is lower semi-continuous and
quasi-convex for all y ∈ Y and f (x, ·) is upper semi-continuous and quasi-
concave for all x ∈ X. Then

inf sup f (x, y) = sup inf f (x, y) .

x∈X y∈Y y∈Y x∈X

There is a short topological proof of this theorem [Komiya, 1988]. You will
use the tools of online linear optimisation to analyse two special cases in
Exercise 28.16. When X and Y are probability simplexes and f is linear, the
resulting theorem is von Neumann’s minimax theorem [von Neumann, 1928].
The minimax theorems form a bridge between minimax adversarial regret and
Bayesian regret, which we discuss in Chapters 34 and 36.
14 Let X be a subset of a linear topological space and f : X → R. The function f
is quasi-convex if f −1 ((−∞, a)) is convex for all a ∈ R and quasi-concave
if −f is quasi-convex. f is upper semi-continuous if for all x ∈ X and
ε > 0 there exists a neighborhood U of x such that f (y) ≤ f (x) + ε for all
y ∈ U . It is lower semi-continuous if for all x ∈ X and ε > 0 there exists a
neighborhood U of x such that f (y) ≥ f (x) − ε for all y ∈ U .

28.6 Bibliographic Remarks

The results in this chapter come from a wide variety of sources. The online convex
optimisation framework was popularised by Zinkevich [2003]. The framework
has been briefly considered by Warmuth and Jagota [1997], then reintroduced
by Gordon [1999] (without noticing the earlier work of Warmuth and Jagota).
While the framework was introduced relatively recently, the core ideas have
been worked out earlier in the special case of linear prediction with nonlinear
 28.7 Exercises 343

losses (the book of Cesa-Bianchi and Lugosi [2006] can be used as a reference
to this literature). Mirror descent was first developed by Nemirovsky [1979] and
Nemirovsky and Yudin [1983] for classical optimisation. In statistical learning,
follow-the-regularised-leader is known as regularised risk minimisation and
has a long history. In the context of online learning, Gordon [1999] considered
follow-the-regularised-leader and called it ‘generalised gradient descent’. The name
seems to originate from the work of Shalev-Shwartz [2007] and Shalev-Shwartz and
Singer [2007]. An implicit form of regularisation is to add a perturbation of the
losses, leading to the ‘follow-the-perturbed-leader’ algorithm [Hannan, 1957, Kalai
and Vempala, 2002], which is further explored in the context of combinatorial
bandit problems in Chapter 30 (and see also Exercise 11.7). Readers interested in
an overview of online learning will like the short books by Shalev-Shwartz [2012]
and Hazan [2016], while the book by Cesa-Bianchi and Lugosi [2006] has a little
more depth (but is also older). As far as we know, the first explicit application of
mirror descent to bandits was by Abernethy et al. [2008]. Since then the idea has
been used extensively, with some examples by Audibert et al. [2013], Abernethy
et al. [2015], Bubeck et al. [2018] and Wei and Luo [2018]. Mirror descent has
been adapted in a generic way to prove high-probability bounds by Abernethy
and Rakhlin [2009]. The reader can find (slightly) different proofs of some mirror
descent results in the book by Bubeck and Cesa-Bianchi [2012]. The results for
the unit ball are from a paper by Bubeck et al. [2012], but we have reworked
the proof to be more in line with the rest of the book. Mirror descent can be
generalised to Banach spaces. For details, see the article by Sridharan and Tewari
[2010].

28.7 Exercises

28.1 Let F : Rd → R ∪ {∞} be Legendre with domain D ⊆ Rd and A ⊆ Rd be

convex, and for b ∈ int(D) and y ∈ Rd let Φ(a) = ha, yi + DF (a, b). Suppose that
c ∈ argmina∈A Φ(a) exists and A ∩ int(D) 6= ∅. Show that c ∈ int(D)

28.2 (Ill-defined actions) Given an example of a non-empty bounded convex

action set A, convex potential F and sequences of losses (yt )nt=1 where the choices
of mirror descent and/or follow-the-regularised-leader either:
(a) exist but are not unique;
(b) do not exist at all.
Prove that if F is Legendre and A is non-empty and compact, then (at )nt=1 exist
and are unique for both mirror descent and follow-the-regularised-leader.

28.3 Prove the correctness of the two-step procedure described in Section 28.1.1.

28.4 (Linear regret for follow-the-leader) Let A = [−1, 1], and let
y1 = 1/2 and ys = 1 for odd s > 1 and ys = −1 for even s > 1.
 28.7 Exercises 344

(a) Recall that follow-the-leader (without regularisation) chooses at =

Pt−1
argmina s=1 ha, ys i. Show that this algorithm suffers linear regret on the
above sequence.
(b) Implement follow-the-regularised-leader or mirror descent on this problem
with quadratic potential F (a) = a2 and plot at as a function of time.

28.5 (Regret for follow-the-regularised-leader) Prove Theorem 28.5.

28.6 (Regret in terms of local dual norms) Prove Corollary 28.8.

28.7 (Follow-the-regularised-leader for the unit ball) Prove the

equality in Eq. (28.13).

28.8 (Exponential weights as mirror descent) Prove the equality in

Eq. (28.5).

28.9 (Exp3 as mirror descent) Let A = Pk−1 be the simplex, F the

unnormalised negentropy potential and η > 0. Let P1 = argminp∈A F (p), and for
t > 1,
Pt+1 = argminp∈A ηhp, Ŷt i + DF (p, Pt ) ,

where Ŷti = I {At = i} yti /Pti and At is sampled from Pt .

(a) Show that the resulting algorithm is exactly Exp3 from Chapter 11.
(b) What happens if you replace mirror descent by follow-the-regularised-leader,
t
X
Pt+1 = argminp∈A hp, Ŷs i + F (p) ?
s=1

28.10 (Exp3 as mirror descent (ii)) Here you will show that the tools in
this chapter not only lead to the same algorithm, but also the same bounds.
(a) Let P̃t+1 = argminp∈[0,∞)k ηhp, Ŷt i + DF (p, Pt ). Show both relations in the
following display:
k
X   η2 Xk
DF (Pt , P̃t+1 ) = Pti exp(−η Ŷti ) − 1 + η Ŷti ≤ Pti Ŷti2 .
i=1
2 i=1
" n #
1 X ηnk
(b) Show that E DF (Pt , P̃t+1 ) ≤ .
η t=1
2
(c) Show that diamF (Pk−1 ) = log(k).
(d) Conclude that for appropriately tuned η > 0, the regret of Exp3 satisfies,
p
Rn ≤ 2nk log(k) .

Hint Use Theorem 26.6(b).

28.11 (Mirror descent and changing learning rates) Let A be a
convex set and y1 , . . . , yn ∈ L ⊆ Rd . Let F be Legendre with domain D with
 28.7 Exercises 345

A ∩ int(D) non-empty and assume that Eq. (28.6) holds. Let η0 , η1 , . . . , ηn > 0,
a1 , a2 , . . . , an+1 ∈ A and ã2 , . . . , ãn+1 be sequences so that η0 = ∞, a1 =
argmina∈A F (a) and

ãt+1 = argmina∈D ηt ha, yt i + DF (a, at ) ,

at+1 = argmina∈A∩D DF (a, ãt+1 ) .

Show that for all a ∈ A,

DF (at , ãt+1 ) DF (a, at ) − DF (a, ãt+1 )

n
X n
X n
X
(a) Rn (a) = hat − a, yt i ≤ + ;
t=1 t=1
ηt t=1
ηt
and  
DF (at , ãt+1 ) 1 1
n
X n
X
(b) Rn (a) ≤ + DF (a, at ) − .
t=1
ηt t=1
ηt ηt−1

The statement allows the time-varying learning rate sequence (ηt )t to be

constructed in any way. This flexibility can be useful when designing adaptive
algorithms. A sequence of learning rates (ηt )t is said to be non-anticipating
if for each t, ηt depends on data available at the end of round t.

28.12 (follow-the-regularised-leader and changing potentials) Like

in the previous exercise, let A be non-empty and convex and y1 , . . . , yn ∈ L ⊆ Rd .
Let F1 , . . . , Fn , Fn+1 be a sequence of convex functions and Φt (a) = Ft (a) +
Pt−1
s=1 ha, ys i and at = argmina∈A Φt (a), which you may assume are well defined.

(a) Show that

n
X
Rn (a) ≤ (hat − at+1 , yt i − DFt (at+1 , at ))
t=1
n
X
+ Fn+1 (a) − F1 (a1 ) + (Ft (at+1 ) − Ft+1 (at+1 )) .
t=1

(b) Show that if Ft = F/ηt and (ηt )n+1

t=1 is decreasing with ηn = ηn+1 , then
n  
F (a) − minb∈A F (b) X DF (at+1 , at )
Rn (a) ≤ + hat − at+1 , yt i − .
ηn t=1
ηt

Again, the statement applies to any sequence of Legendre functions, including

those that are constructed based on the past.

28.13 (Anytime version of Exp3) Consider the k-armed adversarial bandit

problem described in Chapter 11, where the adversary chooses (yt )nt=1 with
 28.7 Exercises 346

yt ∈ [0, 1]k . Let Pt ∈ Pk−1 be defined by

 Pt−1 
exp −ηt s=1 Ŷsi
Pti = P  Pt−1 ,
j=1 exp −ηt
k
s=1 Ŷsj

where (ηt )∞
t=1 is an infinite sequence of learning rates and Ŷti = I {At = i} yti /Pti
and At is sampled from Pt .
(a) Let A = Pk−1 be the simplex, F be the unnormalised negentropy potential,
Pt−1
Ft (p) = F (p)/ηt and Φt (p) = F (p)/ηt + s=1 hp, Ŷs i. Show that Pt is the
choice of follow-the-regularised-leader with potentials (Ft )nt=1 and losses
(Ŷt )nt=1 .
(b) Assume that (ηt )nt=1 are decreasing and then use Exercise 28.12 to show that
" n #
log(k) X DF (Pt+1 , Pt )
Rn ≤ +E hPt − Pt+1 , Ŷt i − .
ηn t=1
ηt

(c) Use Theorem 26.13 in combination with the facts that Ŷti ≥ 0 for all i and
Ŷti = 0 unless At = i to show that
DF (Pt+1 , Pt ) ηt
hPt − Pt+1 , Ŷt i − ≤ .
ηt 2PtAt
log(k) k X
n
(d) Prove that Rn ≤ + ηt .
ηn 2 t=1
p
(e) Choose (ηt )∞
t=1 so that Rn ≤ 2 nk log(k) for all n ≥ 1.

28.14 (The log barrier and first-order bounds) Your mission in this
exercise is to prove first-order bounds for finite-armed bandits as studied in
Chapter 11. The notation is the same as the previous exercise. Let (yt )nt=1 be
Pk
a sequence of loss vectors with yt ∈ [0, 1]k for all t and F (a) = − i=1 log(ai ).
Consider the instance of follow-the-regularised-leader for bandits that samples
At from Pt defined by
t−1
X
Pt = argminp∈Pk−1 ηt hp, Ŷs i + F (p) .
s=1

(a) Show a particular, non-anticipating choice of the learning rates (ηt )nt=1 so
that
v "n−1 #!
u  
u X n∨k
t
Rn ≤ k + 2 k 1 + E 2
ytAt log . (28.14)
t=1
k

(b) Prove that any algorithm satisfying Eq. (28.14) also satisfies
v !
  u n  
n∨k u X n∨k
Rn ≤ k + k log + C tk 1 + min yta log ,
k a∈[k]
t=1
k
 28.7 Exercises 347

where C is a suitably large universal constant.

Hint For choosing the learning rate, you might take inspiration from
Theorem 18.3.

The algorithm in this exercise is a simplified variant of the algorithm analysed

by Wei and Luo [2018].

28.15 (Minimax regret for finite-armed adversarial bandits) Let

(yt )nt=1 be a sequence of loss vectors with yt ∈ [0, 1]k for all t and F (a) =
Pk √
−2 i=1 ai . Consider the instance of follow-the-regularised-leader for k-armed
adversarial bandits that samples At ∈ [k] from Pt defined by
t−1
X
Pt = argminp∈Pk−1 η hp, Ŷs i + F (p) ,
s=1

where Ŷsi = I {As = i} ysi /Psi is the importance-weighted estimator of ysi and
η > 0 is the learning rate.

(a) Show that

t−1
!−2
X
Pti = λ+ Ŷsi ,
s=1

where λ ∈ R is the largest value such that Pti ∈ Pk−1 .

(b) Show that Pt+1,At ≤ PtAt for all t ∈ [n − 1].
(c) Show that ∇2 F (x) = 12 diag(x−3/2 ).
√
(d) Show that diamF (A) ≤ 2 k.
√
(e) Prove that the regret of this algorithm is bounded by Rn ≤ 8kn.
(f) What happens if you use mirror descent instead of follow-the-regularised-
leader. Are the resulting algorithms the same? And if not, what can you
prove for mirror descent?
(g) Explain how you would implement this algorithm.
(h) Prove
√ that if the learning rate is chosen in a time-dependent way to be ηt =
1/ t, then
√ the resulting instantiation of follow-the-regularised-leader satisfies
P
Rn = O( nk) for adversarial bandits and Rn = O( i:∆i >0 log(n)/∆i ) for
stochastic bandits with losses in [0, 1].

The algorithm in the above exercise is called the implicitly normalised

forecaster (INF) and was introduced by Audibert and Bubeck [2009]. The
last part of the exercise is very difficult. For ‘hints’, see the articles by
Zimmert and Seldin [2019] and Zimmert et al. [2019].
 28.7 Exercises 348

28.16 (Minimax theorem) In this exercise you will prove simplified versions
of Sion’s minimax theorem.
(a) Use the tools from online linear optimisation to prove Sion’s minimax theorem
when X = Pk−1 and Y = Pj−1 and f (x, y) = x> Gy for some G ∈ Rk×j .
(b) Generalise your result to the case when X and Y are non-empty, convex,
compact subsets of Rd and f : X × Y → R is convex/concave and has
bounded gradients.

Hint Consider a repeated simultaneous game where the first player chooses
(xt )∞
t=1 and the second player chooses (yt )t=1 . The loss in round t to the first
∞

player is f (xt , yt ), and the loss to the second player is −f (xt , yt ). See what
Pn Pn
happens to the average iterates x̄n = n1 t=1 xt and ȳn = n1 t=1 yt when (xt )
and (yt ) are chosen by (appropriate) regret-minimising algorithms. For the second
part, see Note 12. Also observe that there is nothing fundamental about X and
Y both having dimension d.
28.17 (Counterexample to Sion without compactness) Find examples
of X, Y and f that satisfy the conditions of Sion’s theorem except that neither
X nor Y are compact and where the statement does not hold. Can you choose f
to be bounded?
29 The Relation between Adversarial
and Stochastic Linear Bandits

The purpose of this chapter is to highlight some of the differences and connections
between adversarial and stochastic linear bandits. As it turns out, the connection
between these are not as straightforward as for finite-armed bandits. We focus
on three topics:

(a) For fixed action sets, there is a reduction from stochastic linear bandits to
adversarial linear bandits. This does not come entirely for free. The action
set needs to be augmented for things to work (Section 29.2).
(b) The adversarial and stochastic settings make different assumptions about
the variability of the losses/rewards. This will explain the apparently
contradictory
p result that the upper bound for adversarial bandits on the unit
ball is O( dn log(n)) (Theorem 28.11), while the lower bound for stochastic
√
bandits also for the unit ball is Ω(d n) (Theorem 24.2).
(c) When the action set is changing, the notion of regret in the adversarial
setting must be carefully chosen, and for the ‘right’ choice, we do not yet
have effective algorithms (Section 29.4).

We start with a unified view of the two settings.

29.1 Unified View

To make the notation consistent, we present the stochastic

and adversarial linear bandit frameworks again using losses
for both. Let A ⊂ Rd be the action set. In each round, the
learner chooses At ∈ A and receives the loss Yt , where

Yt = hAt , θi + ηt , (Stochastic setting) (29.1)

Yt = hAt , θt i , (Adversarial setting) (29.2)
Figure 29.1 A tricky
and (ηt )nt=1is a sequence of independent and identically relationship
distributed 1-subgaussian random variables and (θt )nt=1 is a sequence of loss
vectors chosen by the adversary. As noted earlier, the assumptions on the noise
can be relaxed significantly. For example, if Ft = σ(A1 , Y1 , . . . , At , Yt , At+1 ), then
the results of the previous chapters hold as soon as ηt is 1-subgaussian conditioned
 29.2 Reducing Stochastic Linear Bandits to Adversarial Linear Bandits 350

on Ft−1 . The expected regret for the two cases are defined as follows:
n
X
Rn = E [hAt , θi] − n inf ha, θi , (Stochastic setting)
a∈A
t=1
Xn
Rn = E [hAt , θt i] − n inf ha, θ̄n i . (Adversarial setting)
a∈A
t=1

1
Pn
In the last display, θ̄n = n t=1 θt is the average of the loss vectors chosen by
the adversary.

29.2 Reducing Stochastic Linear Bandits to Adversarial Linear

Bandits

To formalise the intuition that adversarial environments are harder than stochastic
environments, one may try to find a reduction where learning in the stochastic
setting is reduced to learning in the adversarial setting. Here, reducing problem
E (‘easy’) to problem H (‘hard’) just means that we can use algorithms designed
for problem H to solve instances of problem E. In order to do this, we need to
transform instances of problem E into instances of problem H and translate back
the actions of algorithms designed for H to actions for problem E. To get a regret
bound for problem E from a regret bound for problem H, one needs to ensure
that the losses translate properly between the problem classes.
Of course, based on our previous discussion, we know that if there is a reduction
from stochastic linear bandits to adversarial linear bandits, then somehow the
adversarial problem must change so that no contradiction is created in the curious
case of the unit ball. To be able to use an adversarial algorithm in the stochastic
environment, we need to specify a sequence (θt )t so that the adversarial feedback
matches the stochastic one. Comparing Eq. (29.1) and Eq. (29.2), we can see
that the crux of the problem is incorporating the noise ηt into θt while satisfying
the other requirements. One simple way of doing this is by introducing an extra
dimension for the adversarial problem.
In particular, suppose that the stochastic problem is d-dimensional so that
A ⊂ Rd . For the sake of simplicity, assume furthermore that the noise and
parameter vector satisfy |ha, θi + ηt | ≤ 1 almost surely for all a ∈ A and that
a∗ = argmina∈A ha, θi exists. Then define Aaug = {(a, 1) : a ∈ A} ⊂ Rd+1 and let
the adversary choose θt = (θ, ηt ) ∈ Rd+1 . Here, we slightly abuse notation: for
x ∈ Rd and y ∈ R, we use (x, y) to denote the d + 1 dimensional vector whose
first d components are those of x and whose last component is y. The reduction
is now straightforward: for t = 1, 2, . . . , do the following:

1 Initialise adversarial bandit policy with action set Aaug .

2 Collect action A0t = (At , 1) from the policy.
3 Play At and observe loss Yt .
 29.3 Stochastic Linear Bandits with Parameter Noise 351

4 Feed Yt to the adversarial bandit policy, increment t and repeat from step 2.

Suppose the adversarial policy guarantees a bound Bn on the expected regret:

" n n
#
X X
Rn = E
0
hAt , θt i − 0 inf
0 0
ha , θt i ≤ Bn .
a ∈Aaug
t=1 t=1

Let a0∗ = (a∗ , 1). Note that for any a = (a, 1) ∈ Aaug , hAt , θi − ha, θi =
0

hAt , θt i − ha0 , θt i and thus adversarial regret, and eventually Bn , will upper
0

bound the stochastic regret:

" n # " n #
X X
E hAt , θi − nha∗ , θi = E hAt , θt i − nha∗ , θ̄n i ≤ Rn0 ≤ Bn .
0 0

t=1 t=1

Therefore, the expected regret in the stochastic bandit is also at most Bn . We

have to emphasise that this reduction changes the geometry of the decision sets
for both the learner and the adversary. For example, if A = B2d is the unit ball,
then neither Aaug nor
( )
y ∈ Rd : sup |ha, yi| ≤ 1
a∈Aaug

are unit balls. It does not seem like this should make much difference, but at
√
least in the case of the ball, from our Ω(d n) lower bound on the regret for the
stochastic case, we see that the changed geometry must make the adversary more
powerful. This reinforces the importance of the geometry of the action set, which
we have already seen in the previous chapter.
While the reduction shows one way to use adversarial algorithms in stochastic
environments, the story seems to be unfinished. When facing a linear bandit
problem with some action set A, the user is forced to decide whether or not
the environment is stochastic. Strangely enough, for stochastic environments the
recommendation is to run your favorite adversarial linear bandit algorithm on the
augmented action set. What if the environment may or may not be stochastic?
One can still run the adversarial linear bandit algorithm on the original action
set. This usually works, but the algorithm may need to be tuned differently
(Exercises 29.2 and 29.3).

29.3 Stochastic Linear Bandits with Parameter Noise

The real reason for all these discrepancies is that the adversarial linear bandit
model is better viewed as relaxation of another class of stochastic linear bandits.
Rather than assuming the noise is added after taking an inner product, assume that
(θt )nt=1 is a sequence of vectors sampled independently from a fixed distribution
ν on Rd . The resulting model is called a stochastic linear bandit with
parameter noise. This new problem can be trivially reduced to adversarial
 29.3 Stochastic Linear Bandits with Parameter Noise 352

bandits when Supp(ν) is bounded (Exercise 29.1). In particular, there is no need

to change the action set.

Combining the stochastic linear bandits with parameter noise model with the
techniques in Chapter 24 is the standard method for proving lower bounds
for adversarial linear bandits.

Parameter noise environments form R a subset of all possible stochastic

environments. To see this, let θ = xν(dx) be the mean parameter vector
under ν. Then the loss in round t is
hAt , θt i = hAt , θi + hAt , θt − θi .
Let Et [·] = E[· | Ft−1 ]. By our assumption that ν has mean θ, the second term
vanishes in expectation, Et [hAt , θt − θi] = 0. This implies that we can make a
connection to the ‘vanilla’ stochastic setting by letting η̃t = hAt , θt − θi. Now
consider the conditional variance of η̃t :
Vt [η̃t ] = Et [hAt , θt − θi2 ] = A>
t Et [(θt − θ)(θt − θ) ]At = At ΣAt ,
> >
(29.3)
where Σ is the covariance matrix of multivariate distribution ν. Eq. (29.3) implies
that the variance of the noise η̃t now depends on the choice of action and in
particular the noise variance scales with the length of At . This can make parameter
noise problems easier. For example, if ν is a Gaussian with identity covariance,
then Vt [η̃t ] = kAt k22 so that long actions have more noise than short actions.
By contrast, in the usual stochastic linear bandit, the variance of the noise is
unrelated to the length of the action. In particular, even the noise accompanying
short actions can be large. This makes quite a bit of difference in cases when
the action set has both short and long actions. In the standard stochastic model,
shorter actions have the disadvantage of having a worse signal-to-noise ratio,
which an adversary can exploit.
This calculation also provides the reason for the different guarantees for the
unit ball. For stochastic linear bandits with 1-subgaussian noise the regret is
√
Õ(d n), while in √ the last chapter we showed that for adversarial linear bandits,
the regret is Õ( dn). This discrepancy is explained by the variance of the noise.
Suppose that ν is supported on the unit sphere. Then the eigenvalues of its
covariance matrix sum to one and if the learner chooses At from the uniform
probability measure µ on the sphere, then
Z
E[Vt [η̃t ]] = a> Σa dµ(a) = 1/d .

By contrast, in the standard stochastic model with 1-subgaussian noise, the

predictable variation of the noise is just 1. If the
√ adversary were allowed to choose
its loss vectors from the sphere of radius d, then the expected predictable
variation would be 1, matching the standard stochastic case, and the regret would
scale linearly in d, which also matches the vanilla stochastic case. This example
 29.4 Contextual Linear Bandits 353

further emphasises the importance of the assumptions that restrict the choices of
the adversary.

The best way to think about the standard adversarial linear model is that it
generalises the stochastic linear bandit with parameter noise. Linear bandits
with parameter noise are sometimes easier than the standard model because
parameter noise limits the adversary’s control of the signal-to-noise ratio
experienced by the learner.

29.4 Contextual Linear Bandits

In practical applications the action set is usually changing from round to round.
Although it is possible to prove bounds for adversarial linear bandits with changing
action sets, the notion of regret makes the results less meaningful than what
one obtains in the stochastic setting. Suppose that (At )nt=1 are a sequence of
action sets. In the stochastic setting, the actions (At )t selected by the LinUCB
algorithm satisfy
" n #
X √
E hAt − a∗t , θi = Õ(d n) ,
t=1

where a∗t = argmaxa∈At ha, θi is the optimal action in round t. This definition of
the regret measures the right thing: the action a∗t really is the optimal action in
round t. The analogous result for adversarial bandits would be a bound on
" n #
X
Rn (Θ) = max E hAt − at (θ), yt i , (29.4)
θ∈Θ
t=1

where Θ is a subset of Rd and at (θ) = argmaxa∈At ha, θi. Unfortunately, however,

we do not currently know how to design algorithms for which this regret is small.
Forpfinite Θ, the techniques of Chapter 27 are easily adapted to prove a bound of
O( dn log |Θ|), but this algorithm is (a) not computationally efficient for large
|Θ|, and (b) choosing Θ as an ε-covering of a continuous set does not guarantee a
bound against the larger set. Providing a meaningful bound on Eq. (29.4) when
Θ is a continuous set like {θ : kθk2 ≤ 1} is a fascinating challenge. The reader
may recall that the result in Exercise 27.5 provides a bound for adversarial linear
bandits with changing action sets. However, in this problem the actions have
‘identities’, and the regret is measured with respect to the best action in hindsight,
which is a markedly different objective than the one in Eq. (29.4).
 29.5 Notes 354

29.5 Notes

1 For the reduction in Section 29.2, we assumed that |Yt | ≤ 1 almost surely.
This is not true for many classical noise models like the Gaussian. One way to
overcome this annoyance is to apply the adversarial analysis on the event that
|Yt | ≤ C for some constant C > 0 that is sufficiently large that the probability
that this event occurs is high. For example, if ηt is p a standard Gaussian and
supa∈A |ha, θi| ≤ 1, then C may be chosen to be 1 + 4 log(n), and the failure
event that there exists a t such that |hAt , θi + ηt | ≥ C has probability at most
1/n by Theorem 5.3 and a union bound.
2 The mirror descent analysis of adversarial linear bandits also works for
stochastic bandits. Recall that mirror descent samples At from a distribution
with a conditional mean of Āt , and suppose that θ̂t is a conditionally unbiased
estimator of θ. Then the regret for a stochastic linear bandit with optimal
action a∗ can be rewritten as
" n # " n # " n #
X X X
Rn = E hAt − a∗ , θi = E hĀt − a∗ , θi = E hĀt − a∗ , θ̂t i ,
t=1 t=1 t=1

which is in the standard format necessary for the analysis of mirror

descent/follow-the-regularised-leader. In the stochastic setting, the covariance
of the least squares estimator θ̂t will not be the same as in the adversarial
setting, however, which leads to different results. When θ̂t is biased, the bias
term can be incorporated into the above formula and then bounded separately.
3 Consider a stochastic bandit with A = B2d the unit ball and Yt = hAt , θi + ηt
where |Yt | ≤ 1 almost surely and kθk2 ≤ 1. Adapting p the analysis of the
algorithm in Section 28.4 leads to a bound of Rn = O(d n log(n)). Essentially
the only change is the variance calculation, which increases by roughly a factor
of d. The details of this calculation are left to you in Exercise 29.2. When A is
finite, the analysis of Exp3 with Kiefer–Wolfowitz
p exploration (Theorem 27.1)
leads to an algorithm for which Rn = O( dn log(k)). For convex A, you can
use continuous exponential weights (Section 27.3).
4 You might wonder whether or not an adversarial bandit algorithm is well
behaved for stochastic bandits where the model is almost linear (the misspecified
linear bandit). Suppose the loss is nearly linear in the sense that

Yt = `(At ) + ηt ,

where `(At ) = hAt , θi + ε(At ) and ε : A → R is some function with small

supremum norm. Because ε(At ) depends on the chosen action, it is not possible
to write Yt = hAt , θt i for θt independent of At . When A = B2d is the unit
ball, you will show in Exercise 29.4 that an appropriately
p tuned instantiation
√
of follow-the-regularised-leader satisfies Rn = O(d n log(n) + εn d), where
ε = supa∈A ε(a). This improves by logarithmic factors on the more generic
algorithm in Chapter 22.
 29.6 Bibliographic Remarks 355

29.6 Bibliographic Remarks

Linear bandits on the sphere with parameter noise have been studied by Carpentier
and Munos [2012]. However they consider the case where the action set is the
sphere and the components of the noise are independent so that the reward is
Xt = hAt , θ + ηt i where the coordinates of ηt ∈ Rd are independent with unit
Pd
variance. In this case, the predictable variation is V[Xt | At ] = i=1 A2ti = 1 for
all actions At and the parameter noise is equivalent to the standard model. We
are not aware of any systematic studies of parameter noise in the stochastic
setting. With only a few exceptions, the impact on the regret of the action set
and adversary’s choices is not well understood beyond the case where A is an
`p -ball, which has been mentioned in the previous section. A variety of lower
bounds illustrating the complications are given by Shamir
√ [2015]. Perhaps the
most informative is the observation that obtaining O( dn) regret is not possible
when A = {a + x : kxk2 ≤ 1} is a shifted unit ball with a = (2, 0, . . . , 0), which
also follows from our reduction in Section 29.2.

29.7 Exercises

29.1 (Reductions) Let A ⊂ Rd be an action set and L = {y ∈ Rd :

supa∈A |ha, yi| ≤ 1}. Take an adversarial linear bandit algorithm that enjoys
a worst-case guarantee Bn on its n-round expected regret Rn when the adversary
is restricted to playing θt ∈ L. Show that if this algorithm is used in a stochastic
linear bandit problem with parameter noise where θt ∼ ν and Supp(ν) ⊆ L, then
the expected regret Rn0 is still bounded by Bn .

29.2 (Follow-the-regularised-leader for stochastic bandits (i))

Consider a stochastic linear bandit with A = B2d and loss Yt = hAt , θi + ηt where
(ηt )nt=1 are independent with zero mean and Yt ∈ [−1, 1] almost surely. Adapt the
proof of Theorem 28.11 to show p that with appropriate tuning the algorithm in
Section 28.4 satisfies Rn ≤ Cd n log(n) for universal constant C > 0.

Hint Repeat the analysis in the proof of Theorem 28.11, update the learning
rate and check the bounds on the norm of the estimators.
29.3 (Follow-the-regularised-leader for stochastic bandits (ii))
Repeat the previous exercise using exponential weights or continuous exponential
weights with Kiefer–Wolfowitz exploration where

(a) A is finite; and

(b) A is convex.

29.4 (Misspecified linear bandits) Let A ⊂ Rd and (ηt )nt=1 be a sequence

 29.7 Exercises 356

of independent zero-mean random variables and assume the loss is

Yt = `(At ) + ηt ,
where `(At ) = hAt , θi + ε(At ) and ε = supa∈A ε(a) and |Yt | ≤ 1 almost surely.
(a) Suppose that A = B2d . Show that the expected regret of an appropriately
tuned version of the algorithm in Section 28.4 satisfies
p √
Rn ≤ C(d n log(n) + εn d) ,
where C > 0 is a universal constant.
(b) Do you think the result from Part (a) can be improved?
(c) Suppose that A is finite. What goes wrong in the analysis of exponential
weights with Kiefer–Wolfowitz exploration (Algorithm 15)?
Part VII
Other Topics
 358

In the penultimate part, we collect a few topics to which we could not dedicate
a whole part. When deciding what to include, we balanced our subjective views
on what is important, pedagogical and sufficiently well understood for a book.
Of course we have played favourites with our choices and hope the reader can
forgive us for the omissions. We spend the rest of this intro outlining some of the
omitted topics.

Continuous-Armed Bandits
There is a small literature on bandits where the number of actions is infinitely
large. We covered the linear case in earlier chapters, but the linear assumption
can be relaxed significantly. Let A be an arbitrary set and F a set of functions
from A → R. The learner is given access to the action set A and function class
F. In each round, the learner chooses an action At ∈ A and receives reward
Xt = f (At ) + ηt , where ηt is noise and f ∈ F is fixed, but unknown. Of course
this set-up is general enough to model all of the stochastic bandits so far, but is
perhaps too general to say much. One interesting relaxation is the case where A
is a metric space and F is the set of Lipschitz functions. We refer the reader to
papers by Kleinberg [2005], Auer et al. [2007], Kleinberg et al. [2008], Bubeck
et al. [2011], Slivkins [2014], Magureanu et al. [2014] and Combes et al. [2017], as
well as the book of Slivkins [2019].

Infinite-Armed Bandits
Consider a bandit problem where in each round the learner can choose to play
an arm from an existing pool of Bernoulli arms or to add another Bernoulli arm
to the pool with mean sampled from a uniform distribution. The regret in this
setting is defined as
" n #
X
Rn = n − E Xt .
t=1

This problem is studied by Berry et al. [1997], who show that Rn = Θ(n1/2 ) is the
optimal regret. There are now a number of strengthening and generalisations of
this work [Wang et al., 2009, Bonald and Proutiere, 2013, Carpentier and Valko,
2015, for example], which sadly must be omitted from this book. The notable
difficulty is generalising the algorithms and analysis to the case where reservoir
distribution from which the new arms are sampled is unknown and/or does not
exhibit a nice structure.

Duelling Bandits
In the duelling bandit problem, the learner chooses two arms in each round
At1 , At2 . Rather than observing a reward for each arm, the learner observes
the winner of a ‘duel’ between the two arms. Let k be the number of arms and
P ∈ [0, 1]k×k be a matrix where Pij is the probability that arm i beats arm j in
a duel. It is natural to assume that Pij = 1 − Pji . A common, but slightly less
justifiable, assumption is the existence of a total ordering on the arms such that
 359

if i  j, then Pij > 1/2. There are at least two notions of regret. Let i∗ be the
optimal arm so that i∗  j for all j 6= i∗ . Then the strong and weak regret are
defined by
" n #
X
Strong regret = E (Pi∗ ,At1 + Pi∗ ,At2 − 1) ,
t=1
" n #
X
Weak regret = E min {Pi∗ ,At1 − 1/2, Pi∗ ,At2 − 1/2} .
t=1

Both definitions measure the number of times arms with low probability of
winning a duel against the optimal arm is played. The former definition only
vanishes when At1 = At2 = i∗ , while the latter is zero as soon as i∗ ∈ {At1 , At2 }.
The duelling bandit problem was introduced by Yue et al. [2009] and has seen
quite a lot of interest since then [Yue and Joachims, 2009, 2011, Ailon et al.,
2014, Zoghi et al., 2014, Dudı́k et al., 2015, Jamieson et al., 2015, Komiyama
et al., 2015a, Zoghi et al., 2015, Wu and Liu, 2016, Zimmert and Seldin, 2019].

Convex Bandits
Let A ⊂ Rd be a convex set. The convex bandit problem comes in both stochastic
and adversarial varieties. In both cases, the learner chooses At from A. In the
stochastic case, the learner receives a reward Xt = f (At ) + ηt where f is an
unknown convex function and ηt is noise. In the adversarial setting, the adversary
chooses a sequence of convex functions f1 , . . . , fn and the learner receives reward
Xt = ft (At ). This turned out to be a major challenge over the last decade with
most approaches leading to suboptimal regret in terms of the horizon. The best
bounds in the stochastic case are by Agarwal et al. [2011], while in the adversarial
case there has been a lot of recent progress [Bubeck et al., 2015a, Bubeck and
Eldan, 2016, Bubeck et al., 2017]. In both cases the dependence of the regret on
√
the horizon is O( n), which is optimal in the worst case. Many open question
remain, such as the optimal dependence on the dimension, or the related problem
of designing practical low-regret algorithms. The interested reader may consult
Shamir [2013] and Hu et al. [2016] for some of the open problems.

Budgeted Bandits
In many problems, choosing an action costs some resources. In the bandits-with-
knapsacks problem, the learner starts with a fixed budget B ∈ [0, ∞)d over
d resource types. Like in the standard K-armed stochastic bandit, the learner
chooses At ∈ [K] and receives a reward Xt sampled from a distribution depending
on At . The twist is that the game does not end after a fixed number of rounds.
Instead, in each round, the environment samples a cost vector Ct ∈ [0, 1]d from a
distribution that depends on At . The game ends in the first round τ for which
Pτ
there exists an i ∈ [d] such that t=1 Cti > Bi . This line of work was started by
Badanidiyuru et al. [2013] and has been extended in many directions by Agrawal
and Devanur [2014], Tran-Thanh et al. [2012], Ashwinkumar et al. [2014], Xia
 360

et al. [2015], Agrawal and Devanur [2016], Tran-Thanh et al. [2010] and Hanawal
et al. [2015]. A somewhat related idea is the conservative bandit problem where
the goal is to minimise regret subject to the constraint that the learner must not
be much worse than some known baseline. The constraint limits the amount of
exploration and makes the regret guarantees slightly worse [Sui et al., 2015, Wu
et al., 2016, Kazerouni et al., 2017].

Learning with Delays

In many practical applications, the feedback to the learner is not immediate. The
time between clicking on a link and buying a product could be minutes, days,
weeks or longer. Similarly, the response to a drug does not come immediately.
In most cases, the learner does not have the choice to wait before making the
next decision. Buyers and patients just keep coming. Perhaps the first paper
for online learning with delays is by Weinberger and Ordentlich [2002], who
consider the full information setting. Recently this has become a hot topic, and
there has been a lot of follow-up work extending the results in various directions
[Joulani et al., 2013, Desautels et al., 2014, Cesa-Bianchi et al., 2016, Vernade
et al., 2017, 2018, Pike-Burke et al., 2018, and others]. Learning with delays is
an interesting example where the adversarial and stochastic models lead to quite
different outcomes. In general the increase in regret due to rewards being delayed
√
by at most τ rounds is a multiplicative τ factor for adversarial models and an
additive term only for stochastic models.

Graph Feedback
There is growing interest in feedback models that lie between the full information
and bandit settings. One way to do this is to let G be a directed graph with
K vertices. The adversary chooses a sequence of loss vectors in [0, 1]K as usual.
In each round, the learner chooses a vertex and observes the loss corresponding
to that vertex and its neighbours. The full information and bandit settings
are recovered by choosing the graph to be fully connected or have no edges
respectively, but of course there are many interesting regimes in between. There
are many variants on this basic problem. For example, G might change in each
round or be undirected. Or perhaps the graph is changing, and the learner only
observes it after choosing an action. The reader can explore this topic by reading
the articles by Mannor and Shamir [2011], Alon et al. [2013], Kocák et al. [2014]
and Alon et al. [2015] or the short book by Valko [2016].
30 Combinatorial Bandits

A combinatorial bandit is a linear bandit with an action set that is a subset

of the d-dimensional binary hypercube: A ⊂ {0, 1}d . Elements of A are thus
d-dimensional, binary-valued vectors. Each component may be on or off, but some
combinations are not allowed – hence the combinatorial structure. Combinatorial
bandit problems arise in many applications, some of which are detailed shortly.
The setting is studied in both the adversarial and stochastic models. We focus
on the former in this chapter and discuss the latter in the notes. In the adversarial
setting, as usual, the environment chooses a sequence of loss vectors y1 , . . . , yn
with yt ∈ Rd , and the regret of the learner is

" n #
X
Rn = max E hAt − a, yt i ,
a∈A
t=1

where as usual At is the action chosen by the learner in round t.

Unsurprisingly, the algorithms and analysis from Chapters 27 and 28 are
applicable in this setting. The main challenge is controlling the computation
complexity of the resulting algorithms. As we will soon argue, except in special
cases, it is natural to be hopeful when there exists an efficient optimisation
oracle that computes the map y 7→ argmina∈A ha, yi. The most important result
of this chapter gives a strategy based on follow-the-perturbed-leader that
makes a single call to such an optimisation oracle in every round for a suitable
chosen vector L̃t ∈ Rd (a perturbed estimate of the cumulative loss vector). This
is done in the semi-bandit setting, an in-between setting where the learner
receives semi-bandit feedback, which is the vector (At1 yt1 , . . . , Atd ytd ). Since
Ati ∈ {0, 1}, this is equivalent to observing yti for all i for which Ati = 1.
The rest of this chapter is organised as follows: the next section describes
some additional useful notation. There follows a section that describes selected
applications. In Section 30.3 we describe an application of Exp3 to the case
when the learner receives only bandit feedback and explain the computational
challenges that arise due to the combinatorial nature of the problem. Section 30.4
explains how online stochastic mirror descent can be applied to the semi-bandit
setting, which still fails to give an efficient algorithm. Finally, the follow-the-
perturbed-leader algorithm is introduced and analysed in Section 30.5.
 30.1 Notation and Assumptions 362

30.1 Notation and Assumptions

In the applications, a key quantity associated with combinatorial action sets is

the largest number of elements m that can be simultaneously ‘on’ in any given
action:

A ⊆ a ∈ {0, 1}d : kak1 ≤ m .

In Chapters 27 and 28, we assumed that yt ∈ {y : supa∈A |ha, yi| ≤ 1}. This
restriction is not consistent with the applications we have in mind, so instead we
assume that yt ∈ [0, 1]d , which by the definition of A ensures that |hAt , yt i| ≤ m
for all t. In the standard bandit model, the learner observes hAt , yt i in each round.

30.2 Applications

Shortest-Path Problems
Let G = (V, E) be a fixed graph with a finite set of vertices V and edges
E ⊆ V × V , with |E| = d. The online shortest-path problem is a game over n
rounds between an adversary and a learner. Given fixed vertices u, v ∈ V , the
learner’s objective in each round is to find the shortest path between u and v. At
the beginning of the game, the adversary chooses a sequence of vectors y1 , . . . , yn ,
with yt ∈ [0, 1]d and yti representing the length of the ith edge in E in round t.
In each round, the learner chooses a path between u and v. The regret of the
learner is the difference between the distance they travelled and the distance of
the optimal path in hindsight. A path is represented by a vector a ∈ {0, 1}d where
ai = 1 if the ith edge is part of the path. Let A be the set of paths connecting
vertices u and v, then the length of path a in round t is ha, yt i. In this problem,
m is the length of the longest path. Fig. 30.1 illustrates a typical example.

Ranking
Suppose a company has d ads and m locations in which to display them. In each
round t, the learner should choose the m ads to display, which is represented by a
vector At ∈ {0, 1}d with kAt k1 = m. As before, the adversary chooses yt ∈ [0, 1]d
that measures the quality of each placement and the learner suffers loss hAt , yt i.
This problem could also be called ‘selection’ because the order of the items play
no role. Problems where the order plays a direct role are analysed in Chapter 32.

Multitask Bandits
Consider playing m multi-armed bandits simultaneously, each with k arms. If
the losses for each bandit problem are observed, then it is easy to apply Exp3 or
Exp3-IX to each bandit independently. But now suppose the learner only observes
the sum of the losses. This problem is represented as a combinatorial bandit by
 30.3 Bandit Feedback 363

Beijing

12

13 7
Budapest
10 13
1
11 Abu Dhabi
Frankfurt
10
12 Singapore

8
Sydney

Figure 30.1 Shortest-path problem between Budapest and Sydney. The learner chooses
the path Budapest–Frankfurt–Singapore–Sydney. In the bandit setting, they observe
total travel time (21 hours), while in the semi-bandit they observe the length of each
flight on the route they took (1 hour, 12 hours, 8 hours).

letting d = mk and
( k
)
X
A= a ∈ {0, 1} : d
ai+kj = 1 for all 0 ≤ j < m .
i=1

In words, the d coordinates are partitioned into m parts and the learner needs to
select exactly one coordinate (“primitive action”) from each part. The resulting
problem is called the multi-task bandit problem: This problem is like making
m independent choices in parallel in m bandit problems blindly and then receiving
an aggregated feedback for all the m choices made. This scenario can arise in
practice when a company is making multiple independent interventions, but the
quality of the interventions are only observed via a single change in revenue.

30.3 Bandit Feedback

The easiest approach is to apply the version of Exp3 for linear bandits described
in Chapter 27. The only difference is that now |hAt , yt i| can be as large as m,
which increases the regret by a factor of m. We leave the proof of the following
theorem to the reader (Exercise 30.1).
Theorem 30.1. Consider the setting of Section 30.1. If Algorithm 15 is run on
action set A with appropriately chosen learning rate, then
s  
p 3/2 ed
Rn ≤ 2m 3dn log |A| ≤ m 12dn log .
m
There are two issues with this approach, both computational. First, the action
 30.4 Semi-bandit Feedback and Mirror Descent 364

set is typically so large that finding the core set of the central minimum volume
enclosing ellipsoid that determines the Kiefer–Wolfowitz exploration distribution
of Algorithm 15 is hopeless. Second, efficiently sampling from the resulting
exponential weights distribution may not be possible. There is no silver bullet
for these issues. The combinatorial bandit can model a repeated version of the
travelling salesman problem, which is hard even to approximate. Since an online
learning algorithm with O(np ) regret with p < 1 can be used to approximate
the optimal solution, it follows that no such algorithm can be computationally
efficient. There are, however, special cases where efficient algorithms exist, and
we give some pointers to the relevant literature on this at the end of the chapter.
One modification that greatly eases computation is to replace the optimal Kiefer–
Wolfowitz exploration distribution with a distribution that can be computed and
sampled from in an efficient manner, as noted after Theorem 27.1.

30.4 Semi-bandit Feedback and Mirror Descent

In the semi-bandit setting, the learner observes the loss associated with all non-
zero coordinates of the chosen action. The additional information is exploited by
noting that yt can now be estimated in each coordinate. Let
Ati yti
Ŷti = , (30.1)
Āti
where Āti = E[Ati | Ft−1 ] with Ft = σ(A1 , . . . , At ). An easy calculation shows
that E[Ŷt | Ft−1 ] = yt , so this estimate is still unbiased. Unsurprisingly we will
again use online stochastic mirror descent, which is summarised for this setting
in Algorithm 18.

1: Input A, η, F
2: Ā1 = argmina∈co(A) F (a)
3: for t = 1, . . . , n do
P
4: Choose distribution Pt on A such that a∈A Pt (a)a = Āt
5: Sample At ∼ Pt and observe At1 yt1 , . . . , Atd ytd
6: Compute Ŷti = Ati yti /Āti for all i ∈ [d]
7: Update Āt+1 = argmina∈co(A) ηha, Ŷt i + DF (a, Āt )
8: end for
Algorithm 18: Online stochastic mirror descent for semi-bandits.

Theorem 30.2. Consider the setting of Section 30.1. Let F : Rd → R be the

unnormalised negentropy potential:
d
X
F (a) = (ai log(ai ) − ai )
i=1
 30.4 Semi-bandit Feedback and Mirror Descent 365

for a ∈ [0, ∞)p

d
and F (a) = ∞ otherwise. Then, Algorithm 18 is well defined and,
provided η = 2m(1 + log(d/m))/(nd), its regret Rn satisfies
p
Rn ≤ 2nmd(1 + log(d/m)) .

Proof Since A is a finite set, the algorithm is well defined. In particular, Āt > 0
exists and is unique for all t ∈ [n]. By Theorem 28.10,
" n #
diamF (co(A)) X 1
Rn ≤ +E hĀt − Āt+1 , Ŷt i − DF (Āt+1 , Āt ) . (30.2)
η t=1
η

The diameter is easily bounded by noting that F is negative in co(A) and using
Jensen’s inequality:
d   
X 1
diamF (co(A)) ≤ sup ai + ai log ≤ m(1 + log(d/m)) .
a∈co(A) i=1 ai

For the second term in Eq. (30.2), let Ŷti0 = Ŷti I Āt+1,i ≤ Āti . Since Ŷt is
positive,
1 1
hĀt − Āt+1 , Ŷt i − DF (Āt+1 , Āt ) ≤ hĀt − Āt+1 , Ŷt0 i − DF (Āt+1 , Āt )
η η
d
η η X Ati
≤ kŶt0 k2∇2 F (Zt )−1 ≤ ,
2 2 i=1 Āti

where Zt is provided by Theorem 26.12 and lies on the chord [Āt , Āt+1 ]. The
final inequality follows because ∇2 F (z) = diag(1/z) and using the definition of
Ŷt0 , which ensures that the worst case occurs when Zt = Āt . Summing and taking
the expectation:
" n # " n d #
X 1 η X X Ati ηnd
E hĀt − Āt+1 , Ŷt i − DF (Āt+1 , Āt ) ≤ E = .
t=1
η 2 t=1 i=1
Āti 2

Putting together the pieces shows that

m(1 + log(d/m)) ηnd p
Rn ≤ + = 2nmd(1 + log (d/m)) .
η 2

Algorithm 18 plays mirror descent on the convex hull of the actions, which has
dimension d − 1. In principle it would be possible to do the same thing on the
set of distributions over actions, which has dimension
p |A| − 1. Repeating the
analysis leads to a suboptimal regret of O(m dn log(d/m)). We encourage
the reader to go through this calculation to see where things go wrong.

Like in Section 30.3, the main problem is computation. In each round the algorithm
P
needs to find a distribution Pt over A such that a∈A Pt (a) = Āt . Feasibility
follows from the definition of co(A), while Carathéodory’s theorem proves the
 30.5 Follow-the-Perturbed-Leader 366

support of Pt never needs to be larger than d+1. Since A is finite, we can write the
problem of finding Pt in terms of linear constraints, but naively the computation
complexity is polynomial in k = |A|, which is exponential in m. The algorithm also
needs to compute Āt+1 from Āt and Ŷt . This is a convex optimisation problem,
but the computation complexity depends on the representation of A and may be
intractable. See Note 6 for a few more details on this.

30.5 Follow-the-Perturbed-Leader

In this section, we help ourselves to find a computationally efficient algorithm

by adding the assumption that for all y ∈ [0, ∞)d , the optimisation problem of
finding

a∗ = argmina∈A ha, yi (30.3)

admits a computationally efficient algorithm. This assumption feels close to the

minimum one could get away with in the sense that if the offline problem in
Eq. (30.3) is hard to approximate, then any algorithm with low regret must also
be inefficient. A marginally more reasonable assumption is that Eq. (30.3) can be
approximated efficiently. For simplicity we assume exact solutions, however.
If the algorithm observed the losses in every round, adding a random vector to
the sum of previous losses and then finding the action that minimises the total
randomly perturbed loss leads to what is known as the follow-the-perturbed-
leader (FTPL) algorithm. As discussed before, the random perturbation is
necessary to achieve sublinear regret. In semi-bandit setting which is considered
here, the full loss vector is unobserved and hence needs to be estimated. Letting
Pt−1
L̂t−1 = s=1 Ŷs be the cumulative loss estimates before round t, FTPL chooses

At = argmina∈A ha, η L̂t−1 − Zt i , (30.4)

where η > 0 is the learning rate and Zt ∈ Rd is sampled from a carefully chosen
distribution Q. The random perturbations is chosen to both guard against worst-
case, and to induce necessary exploration. Notice that if η is small, then the
effect of Zt is larger and the algorithm can be expected to explore more, which is
consistent with the learning rate used in mirror descent or exponential weighting
studied in previous chapters.
Before defining the loss estimations and perturbation distribution, we make a
connection between FTPL and mirror descent. Given Legendre potential F with
dom(∇F ) = int(co(A)), online stochastic mirror descent chooses Āt so that

Āt = argmina∈co(A) ha, η Ŷt−1 i + DF (a, Āt−1 ) .

Taking derivatives and using the fact that dom(∇F ) = int(co(A)), we have

∇F (Āt ) = ∇F (Āt−1 ) − η Ŷt−1 = −η L̂t−1 .

 30.5 Follow-the-Perturbed-Leader 367

By duality (Theorem 26.6), this implies that Āt = ∇F ∗ (−η L̂t−1 ). On the other
hand, examining Eq. (30.4), we see that for FTPL,
h  i

Āt = E[At | Ft−1 ] = E argmina∈A ha, η L̂t−1 − Zt i  Ft−1 ,

where Ft = σ(Z1 , . . . , Zt ). Thus, in order to view FTPL as an instance of mirror

descent, it suffices to find a Legendre potential F with dom(∇F ) = int(co(A))
and
h  i

∇F ∗ (−η L̂t−1 ) = E argmina∈A ha, η L̂t−1 − Zt i  Ft−1
h  i

= E argmaxa∈A ha, Zt − η L̂t−1 i  Ft−1 .

Since L̂t−1 is more or less uncontrolled, the latter

R condition is most easily satisfied
by requiring that for any x ∈ R , ∇F (x) = Rd argmaxa∈co(A) ha, x + zi dQ(z).
d ∗

To remove clutter in the notation, define

a(x) = argmaxa∈A ha, xi ,

where a(x) is chosen to be an arbitrary maximiser if multiple maximisers exist.

Readers with some familiarity with convex analysis will remember that if a convex
set A has a smooth boundary, then the support function of A,

φ(x) = maxha, xi,

a∈A

satisfies ∇φ(x) = a(x). For combinatorial bandits, A is not smooth, but if Q is

absolutely continuous with respect to the Lebesgue measure, then you will show
in Exercise 30.5 that
Z Z
∇ φ(x + z) dQ(z) = a(x + z) dQ(z) for all x ∈ Rd .
Rd Rd

The key to this argument is that the derivative of φ exists almost everywhere
and is equal to a(x). All this shows is that FTPL can be interpreted as mirror
descent with potential F defined in terms of its Fenchel dual,
Z
F (x) =
∗
φ(x + z)dQ(z) . (30.5)
Rd

Of course we have not shown that F is Legendre or that int(dom(F )) = int(co(A)),

both of which you will do in Exercise 30.6 under appropriate conditions on Q.
There are more reasons for making this connection than mere curiosity. The
classical analysis of FTPL involves at least one ‘leap of faith’ in the analysis. In
contrast, the analysis via the mirror descent interpretation is more mechanical.
Recall that mirror descent depends on choosing a potential, an exploration
distribution and an estimator. We now make the choice of these explicit. The
exploration distribution is a distribution Pt on A such that
X
Āt = Pt (a)a ,
a∈A
 30.5 Follow-the-Perturbed-Leader 368

which in our case is implicitly defined by the distribution of Zt :

Pt (a) = P(a(Zt − η L̂t−1 ) = a | Ft−1 ) .

It remains to choose the loss estimator. A natural choice would be the same as
Eq. (30.1), which is Ŷti = Ati yti /Pti with Pti = P (Ati = 1 | Ft−1 ) = Āti . The
problem is that Pti does not generally have a closed-form solution. And while
Pti can be estimated by sampling, the number of samples required for sufficient
accuracy can be quite large. The next idea is to replace 1/Pti in the importance-
weighted estimator with a random variable with conditional expectation equal to
1/Pti . This is based on the following well-known result:

Lemma 30.3. Let U ∈ {1, 2, . . .} be geometrically distributed with parameter

θ ∈ [0, 1] so that P (U = j) = (1 − θ)j−1 θ. Then E[U ] = 1/θ.

You can sample from a geometric distribution with parameter θ by counting

the number of flips of a biased coin with bias θ until the first head. That is,
if (Xt )∞
t=1 is an independent sequence of Bernoulli random variables with
bias θ, then U = min{t ≥ 1 : Xt = 1} is geometrically distributed with
parameter θ.

Define a sequence of d-dimensional random vectors K1 , . . . , Kn , where (Kti )di=1

is a sequence of geometric random variables that are conditionally independent
given Ft so that the conditional law of Kti given Ft is Geometric(Pti ) and where
we now redefine Ft = σ(Z1 , K1 , . . . , Zt−1 , Kt−1 , Zt ). The estimator of yti can now
be defined by

Ŷti = min(β, Kti )Ati yti ,

where β is a positive integer to be chosen subsequently. Note that

E [Kti Ati yti | Ft−1 ] = yti .

The truncation parameter β is needed to ensure that Ŷti is never too large. We
have now provided all the pieces to define a version of FTPL that is a special
case of mirror descent. The algorithm is summarised in Algorithm 19.

Theorem 30.4. Consider the setting of Section 30.1. Let Q have density with
respect to the Lebesgue measure of q(z) = 2−d exp(−kzk1 ), and choose the
parameters η, β as follows:
s  
2(1 + log(d)) 1
η= , β= .
(1 + e2 )dnm ηm

p 19 is well defined and provided that ηm ≥ 1 its

Then the algorithm Algorithm
regret is bounded by Rn ≤ m 2(1 + e2 )nd(1 + log(d)).
 30.5 Follow-the-Perturbed-Leader 369

1: Input A, n, η, β, Q
2: L̂0 = 0 ∈ Rd
3: for t = 1, . . . , n do
4: Sample Zt ∼ Q
5: Compute At = argmaxa∈A ha, Zt − η L̂t−1 i
6: Observe At1 yt1 , . . . , Atd ytd
7: For each i ∈ [d] sample Kti ∼ Geometric(Pti )
8: For each i ∈ [d] compute Ŷti = min(β, Kti )Ati yti
9: L̂t = L̂t−1 + Ŷt
10: end for
Algorithm 19: Follow-the-perturbed-leader for semi-bandits.

Proof First, note that At is almost surely uniquely defined and so is Āt =
E [At | Ft−1 ]. Therefore, by isolating the bias in the loss estimators, and thanks
to Exercise 30.6, we can apply Theorem 28.4 to get that
" n # n
" #
X X
Rn (a) = E hAt − a, yt i = E hĀt − a, yt i
t=1 t=1
" n
# n
" #
X X
=E hĀt − a, Ŷt i + E hĀt − a, yt − Ŷt i
t=1 t=1
" # " n #
diamF (A) 1X
n X
≤ +E DF (Āt , Āt+1 ) + E hĀt − a, yt − Ŷt i .
η η t=1 t=1
(30.6)

Of the three terms, the diameter is most easily bounded. For Z ∼ Q,

F (a) = sup (ha, xi − F ∗ (x)) = sup (ha, xi − E[maxhb, x + Zi]) (30.7)

x∈Rd x∈Rd b∈A

1
d
X
≥ −E[maxhb, Zi] ≥ −mE[kZk∞ ] = −m ≥ −m(1 + log(d)) ,
b∈A
i=1
d

where the first inequality follows by choosing x = 0 and the second follows
from Hölder’s inequality and that kak1 ≤ m for any a ∈ A. The last equality is
non-trivial and is explained in Exercise 30.4. By the convexity of the maximum
function and the fact that Z is centered, we also have from Eq. (30.7) that
F (a) ≤ 0, which means that

diamF (A) = max F (a) − F (b) ≤ m(1 + log(d)) . (30.8)

a,b∈A

The next step is to bound the Bregman divergence induced by F . We will shortly
show that the Hessian ∇2 F ∗ (x) of F ∗ exists, so by Part (b) of Theorem 26.6
 30.5 Follow-the-Perturbed-Leader 370

and Taylor’s theorem, there exists an α ∈ [0, 1] and ξ = −η L̂t−1 − αη Ŷt such that

DF (Āt , Āt+1 ) = DF ∗ (∇F (Āt+1 ), ∇F (Āt ))

η2
= DF ∗ (−η L̂t−1 − η Ŷt , −η L̂t−1 ) = kŶt k2∇2 F ∗ (ξ) , (30.9)
2
where the last equality follows from Taylor’s theorem (see Theorem 26.12). To
calculate the Hessian, we use a change of variable to avoid applying the gradient
to the non-differentiable argmax:
Z
∇2 F ∗ (x) = ∇(∇F ∗ (x)) = ∇E [a(x + Z)] = ∇ a(x + z)q(z)dz
Rd
Z Z
=∇ a(u)q(u − x)du = a(u)(∇q(u − x))> du
Rd Rd
Z Z
= a(u) sign(u − x)> q(u − x)du = a(x + z) sign(z)> q(z)dz .
Rd Rd

Using the definition of ξ and the fact that a(x) is non-negative,

Z
∇2 F ∗ (ξ)ij = a(ξ + z)i sign(z)j q(z)dz (30.10)
Rd
Z
≤ a(ξ + z)i q(z)dz
d
ZR
= a(z − η L̂t−1 − αη Ŷt )i q(z)dz
d
ZR
= a(u − η L̂t−1 )i q(u + αη Ŷt )du
Rd
 Z
≤ exp kαη Ŷt k1 a(u − η L̂t−1 )i q(u)du
Rd
2
≤ e Pti , (30.11)

where the last inequality follows since α ∈ [0, 1] and Ŷti ≤ β = d1/(mη)e, ηm ≥ 1
and Ŷt has at most m non-zero entries. Continuing on from Eq. (30.9), we have

η2 e2 η 2 X e2 η 2 X X
d Xd d d
kŶt k2∇2 F ∗ (ξ) ≤ Pti Ŷti Ŷtj ≤ Pti Kti Ati Ktj Atj .
2 2 i=1 j=1
2 i=1 j=1

Chaining together the parts and taking the expectation shows that
 
e2 η X X
d d
E[DF (Āt , Āt+1 )] ≤ E Pti Kti Ati Ktj Atj 
2 i=1 j=1
 
e2 η 2 X X Ati Atj  e2 mdη 2
d d
= E ≤ .
2 i=1 j=1
Ptj 2

The last step is to control the bias term. For this, first note that since
 30.6 Notes 371

Ati yti ∈ {0, 1},

E[Ŷti | Ft ] = E[min(β, Kti )Ati yti | Ft ] = Ati yti E[min(β, Kti ) | Ft ]

Ati
= yti (1 − (1 − Pti )β ) ,
Pti
where the last equality follows from the definition of Kti using a direct calculation.
Thus, E[Ŷti | Ft−1 ] = (1 − (1 − Pti )β )yti and
" n # " n #
X X
E hĀt − a, yt − Ŷt i ≤ E hĀt , yt − Ŷt i
t=1 t=1
" n d #
XX dn dnmη
=E yti Pti (1 − Pti ) β
≤ = ,
t=1 i=1
2β 2

where the last inequality follows from using that for x ∈ [0, 1], s > 0,
x(1 − x)s ≤ xe−sx ≤ 1/s. Putting together all the pieces into Eq. (30.6) leads to
m(1 + log(d)) e2 dnmη dnmη p
Rn ≤ + + ≤ m 2(1 + e2 )nd(1 + log(d)) .
η 2 2

30.6 Notes

1 For a long time, it was speculated that the dependence of the regret on m3/2
in Theorem 30.1 (bandit feedback) might be improvable to m. Very recently,
however, the lower bound was increased to show the upper bound is √ tight
[Cohen et al., 2017]. For semi-bandits the worst-case lower bound is Ω( dnm)
(Exercise 30.8), which holds for large enough n and m ≤ d/2 and is matched
up to constant factors by online stochastic mirror descent with a different
potential (Exercise 30.7).
2 The implementation of FTPL shown in Algorithm 19 needs to sample Kti for
each i with Ati = 1. The conditional expected running time for this is Ati /Pti ,
which has expectation 1. It follows that the expected running time over the
whole n rounds is O(nd) calls to the oracle linear optimisation algorithm. It
can happen that the algorithm is unlucky and chooses Ati = 1 for some i
with Pti quite small and then sampling Kti could be time-consuming. Note,
however, that only min(Kti , β) is actually used by the algorithm, and hence
the sampling procedure can be truncated at β. This minor modification ensures
the algorithm needs at most O(βnd) calls to the oracle in the worst case.
3 While FTPL is excellent in the face of semi-bandit information, we do not know
of a general result for the bandit model. The main challenge is controlling the
variance of the least squares estimator without explicitly inducing exploration
using a sophisticated exploration distribution like what is provided by Kiefer–
Wolfowitz.
 30.6 Notes 372

4 Combinatorial bandits can also be studied in a stochastic setting. There are

several ways to do this. The first mirrors our assumptions for stochastic linear
bandits in Chapter 19, where the loss (more commonly reward) is defined by

Xt = hAt , θi + ηt , (30.12)

where θ ∈ Rd is fixed and unknown and ηt is the noise on which statistical

assumptions are made (for example, conditionally 1-subgaussian). There are
at least two alternatives. Suppose that θ1 , . . . , θn are sampled independently
from some multivariate distribution, and define the reward by

Xt = hAt , θt i . (30.13)

This latter version has ‘parameter noise’ (cf. Chapter 29) and is more closely
related to the adversarial set-up studied in this chapter. Finally, one can assume
additionally that the distribution of θt is a product distribution so that (θ1i )di=1
are also independent.
5 For some action sets, the off-diagonal elements of the Hessian in Eq. (30.10)
√
are negative, which improves the dependence on m to m. An example where
this occurs is when A = {a ∈ {0, 1}d : kak1 = m}. Let i 6= j, and suppose that
z, ξ ∈ Rd and zj ≥ 0. Then you can check that a(z + ξ)i ≤ a(z − 2zj ej + ξ)i ,
and so
Z
∇2 F ∗ (ξ)ij = a(z + ξ)i sign(z)j q(z)dz
Rd
Z Z ∞
= (a(z + ξ)i − a(z − 2zj ej + ξ)i )q(z)dzj dz−j
Rd−1 0
≤ 0,

where dz−j is shorthand for dz1 dz2 , . . . dzj−1 dzj+1 , . . . , dzd . You are asked to
complete all the details in Exercise 30.9. This result unfortunately does not
hold for every action set (Exercise 30.10).
6 In order to implement mirror descent or follow-the-regularised-leader with
bandit or semi-bandit information, one needs to solve two optimisation problems:
(a) a convex optimisation problem of the form argmina∈co(A) F (a) for some
convex F and (b) a linear optimisation problem to find a distribution P over A
with mean ā where ā ∈ co(A). More or less sufficient is an efficient membership
oracle for co(A) and evaluation oracle for F [Grötschel et al., 2012, Lee et al.,
2018]. Also necessary for bandits is to identify an exploration distribution,
which we discuss in the notes and bibliographic remarks of Chapter 27. This is
not required for semi-bandits, however, at least with the negentropy potential
used in by Algorithm 18.
 30.7 Bibliographic Remarks 373

30.7 Bibliographic Remarks

The online combinatorial bandit was introduced by Cesa-Bianchi and Lugosi

[2012], where you will also find the most comprehensive list of known applications
for which efficient algorithms exist. The regret bound for Exp3 given in
Theorem 30.1 for the bandit case is due to Bubeck and Cesa-Bianchi [2012]
(with a slightly different argument). While computational issues remain in the
bandit problem, there has been some progress in certain settings. Combes et al.
[2015b] propose playing mirror descent on the convex hull of the action set without
fancy exploration, which leads to near-optimal bounds for well-behaved action
sets. One could also use continuous exponential weights from Chapter 27. These
methods lead to computationally efficient algorithms for some action sets, but this
must be checked on a case-by-case basis. The full information setting has been
studied quite extensively [Koolen et al., 2010, and references from/to]. FTPL was
first proposed (in the full information context) by Hannan [1957], rediscovered
by Kalai and Vempala [2002, 2005] and generalised by Hutter and Poland [2005].
Poland [2005] and Kujala and Elomaa [2005] independently applied FTPL to finite-
armed adversarial bandits and showed near-optimal regret for this case. Poland
[2005] also proposed to use Monte Carlo simulation to estimate the probability
of choosing each arm needed in the construction of reward estimates. Kujala and
Elomaa [2007] extended the result to non-oblivious adversaries. For combinatorial
settings, suboptimal rates have been shown by Awerbuch and Kleinberg [2004],
McMahan and Blum [2004] and Dani and Hayes [2006]. Semi-bandits seem to have
been introduced in the context of shortest-path problems by György et al. [2007].
The general set-up and algorithmic analysis of FTPL presented follows the work
by Neu [2015a], who also introduced the idea to estimate the inverse probabilities
via a geometric random variable. Our analysis based on mirror descent is novel.
The analysis follows ideas of Abernethy et al. [2014], who present the core ideas in
the prediction with expert advice setting, Cohen and Hazan [2015], who consider
the combinatorial full information case, and Abernethy et al. [2015], who study
finite-armed bandits. The literature on stochastic combinatorial semibandits
is also quite large with algorithms and analysis in the frequentist [Gai et al.,
2012, Combes et al., 2015b, Kveton et al., 2015b] and Bayesian settings [Wen
et al., 2015, Russo and Van Roy, 2016]. These works focus on the case where the
reward is given by Eq. (30.13) and the components of θt are independent. When
the reward is given by Eq. (30.12), one can use the tools for stochastic linear
bandits developed in Part V. Some work also pushes beyond the assumption
that the rewards are linear [Chen et al., 2013, Lin et al., 2015, Chen et al.,
2016a,b, Wang and Chen, 2018]. The focus in these works is on understanding
what are the minimal structural assumptions on the reward function and action
spaces for which learning in combinatorially large action spaces is still feasible
statistically/computationally. Last of all, we mentioned that travelling salesman
is computationally hard to approximate, which you can read about in the paper
by Papadimitriou and Vempala [2006], and references there-in.
 30.8 Exercises 374

30.8 Exercises

30.1 (Mirror descent for combinatorial bandits) Prove Theorem 30.1.

Hint For the second inequality, you
may find it useful to know that for
Pm
0 ≤ m ≤ n, defining Φm (n) = i=0 ni , it holds that (m/n)m Φm (n) ≤ em .
30.2 (Efficient computation on m-sets) Provide an efficient implementation
of Algorithm 18 for the m-set: A = {a ∈ {0, 1}d : kak1 = m}.

30.3 (Efficient computation on shortest-path problems) Playing mirror

descent on co(A) leads to a good bound for bandit or semi-bandit problems, but
sometimes playing Exp3 over A is more efficient, even when A is exponentially
large. Design and analyse a variant of Exp3 for the online shortest-path problem
with semi-bandit feedback described in Section 30.2. Your challenge is to ensure
the following:
√
(a) a regret of Rn = O( n), with dependence on d and m omitted; and
(b) polynomial computation complexity in n and d.

Hint This is not the easiest exercise. Start by reading the paper by Takimoto
and Warmuth [2003], then follow up with that of György et al. [2007].
30.4 (Expected supremum norm of Laplace) Let Z be sampled from
measure on Rd with density f (z) = 2−d exp(−kzk1 ). The purpose of this exercise
is to show that
1
d
X
E[kZk∞ ] = . (30.14)
i=1
i

(a) Let X1 , . . . , Xd be independent standard exponentials. Show that kZk∞ and

max{X1 , . . . , Xd } have the same law.
(b) Let Mj = maxi≤j Xi . Prove for j ≥ 2 that
E[Mj ] = E[Mj−1 ] + E[exp(−Mj−1 )] .
(c) Prove by induction or otherwise that for all a, j ∈ {1, 2, . . .},
a!
E[exp(−aMj )] = Qa .
b=1 (j + b)

(d) Prove the claim in Eq. (30.14).

30.5 (Gradient of expected support function) Let A ⊂ Rd be a compact

set and φ(x) = maxa∈A ha, xi its support function. Let Q be a measure on Rd that
is absolutely continuous with respect to the Lebesgue measure, and let Z ∼ Q.
Show that
∇E[φ(x + Z)] = E [argmaxa∈A ha, x + Zi] .

Hint Recall that the support function φ of a non-empty compact set is a proper
 30.8 Exercises 375

convex function. Then, note that for any proper convex function f : Rd → R∪{∞},
the set Rd \ dom(∇f ) has Lebesgue measure zero [Rockafellar, 2015, Theorem
25.5]. Next, by Danskin’s theorem, the directional derivative of φ in the direction
v ∈ Rd is given by ∇v φ(x) = maxa∈A(x) ha, vi, where A(x) is the set of maximisers
of a 7→ ha, xi over A [Bertsekas, 2015, Proposition 5.4.8 in Appendix B]. Finally,
it is worth remembering the following result: let f be an extended real-valued
function with x ∈ Rd in the interior of its domain. Then, for some g ∈ Rd ,
∇v f (x) = hg, vi holds true for all v ∈ Rd if and only if ∇f (x) exists and is equal
to g.
30.6 A function f : Rd → R̄ is closed if its epigraph is a closed set. Let F ∗ be
the function defined in Section 30.5 and F be the proper convex closed function
and whose Fenchel dual is F ∗ .

(a) Show that the function F is well defined (F ∗ is the Fenchel dual of a proper
convex closed function, and there is only a single such function).
(b) For the remainder of the exercise, let Q be absolutely continuous with respect
to the Lebesgue measure with an everywhere positive density, and let A be
the convex hull of finitely many points in Rd whose span is Rd . Show that
the function F is Legendre.
(c) Show that int(dom(F )) = int(co(A)).

Hint For Part (a), it may be worth recalling that the bidual (the dual of the
dual) of a proper convex closed function f is itself: f = f ∗∗ (= (f ∗ )∗ ). Furthermore,
the Fenchel dual of a proper function is always a proper convex closed function.
30.7 (Minimax bound for combinatorial semi-bandits) Adapt the analysis
in Exercise 28.15 to derive an algorithm for combinatorial
√ bandits with semi-
bandit feedback for which the regret is Rn ≤ C mdn for universal constant
C > 0.

30.8 (Lower bound for combinatorial semi-bandits) Let m ≥ 1 and

d = km for some k > 1. Prove that for any √
algorithm there exists a combinatorial
semi-bandit such that Rn ≥ c min{nm, mdn} where c > 0 is a universal
constant.

Hint The most obvious choice is A = {a ∈ {0, 1}d : kak1 = m}, which are
sometimes called m-sets. A lower bound does hold for this action set [Lattimore
et al., 2018]. However, an easier path is to impose a little additional structure
such as multi-task bandits.
30.9 (Follow-the-perturbed-leader
√ for m-sets) Use the ideas in Note 5 to
prove that FTPL has Rn = Õ( mnd) regret when A = {a ∈ {0, 1}d : kak1 = m}.

Hint After proving the off-diagonal elements of the Hessian are negative, you
will also need to tune the learning rate. We do not know of a source for this
result, but the full information case was studied by Cohen and Hazan [2015].
 30.8 Exercises 376

30.10 Construct an action set and i 6= j and z ∈ Rd with zj > 0 such that
a(z)i ≥ a(z − 2zj ej )i .

Hint Consider the shortest-path problem defined by the graph below.

i j
start goal

Choose losses for the edges z, and think about what happens when the loss
associated with edge j decreases.
31 Non-stationary Bandits

The competitor class used in the standard definition of the

regret is not appropriate when the underlying environment
is changing. In this chapter we increase the power of
the competitor class to ‘track’ changing environments
and derive algorithms for which the regret relative to
this enlarged class is not too large. While the results
are specified to bandits with finitely many arms (both
stochastic and adversarial), many of the ideas generalise to
other models such as linear bandits. This chapter also
illustrates the flexibility of the tools presented in the Figure 31.1 This ban-
dit is definitely not sta-
earlier chapters, which are applied here almost without
tionary!
modification. We hope (and expect) that this will also be
true for other models you might study.

31.1 Adversarial Bandits

In contrast to stochastic bandits, the adversarial bandit model presented in

Chapter 11 does not prevent the environment from changing over time. The
problem is that bounds on the regret can become vacuous when the losses appear
non-stationary. To illustrate an extreme situation, suppose you face a two-armed
adversarial bandit with losses yt1 = I {t ≤ n/2} and yt2 = I {t > n/2}. If we run
Exp3 on this problem, then Theorem 11.2 guarantees that
" n # n
X X p
Rn = E ytAt − min yti ≤ 2nk log(k) .
i∈{1,2}
t=1 t=1
Pn
Since mini∈{1,2} t=1 yti = n/2, by rearranging we see that
" n #
X n p
E ytAt ≤ + 2nk log(k) .
t=1
2

To put this in perspective, a policy that plays each arm with probability half in
Pn
every round would have E[ t=1 ytAt ] = n/2. In other words, the regret guarantee
is practically meaningless.
What should we expect for this problem? The sequence of losses is so regular
 31.1 Adversarial Bandits 378

that we might hope that a clever policy will mostly play the second arm in the
first n/2 rounds and then switch to playing mostly the first arm in the second
n/2 rounds. Then the cumulative loss would be close to zero and the regret would
be negative. Rather than aiming to guarantee negative regret, we redefine the
regret by enlarging the competitor class as a way to ensure meaningful results.
Let Γnm ⊂ [k]n be the set of action sequences of length n with at most m − 1
changes:
( n−1
)
X
Γnm = (at ) ∈ [k] :
n
I {at 6= at+1 } ≤ m − 1 .
t=1

Then define the non-stationary regret with m − 1 change points by

" n # " n #
X X
Rnm = E ytAt − min E ytat .
a∈Γnm
t=1 t=1

The non-stationary regret is sometimes called the tracking regret because a

learner that makes it small must ‘track’ the best arm as it changes. Notice
that Rn1 coincides with the usual definition of the regret. Furthermore, on the
sequence described at the beginning of the section, we see that
" n #
X
Rn2 = E ytAt ,
t=1

which means a policy can only enjoy sublinear non-stationary regret if it detects
the change point quickly. The obvious question is whether or not such a policy
exists and how its regret depends on m.

Exp4 for Non-stationary Bandits

One idea is to use the Exp4 policy from Chapter 18 with a large set of experts,
one for each a ∈ Γnm . Theorem 18.1 shows that Exp4 with these experts suffers
regret of at most
p
Rnm ≤ 2nk log |Γnm | . (31.1)
Naively bounding log |Γnm | (Exercise 31.1) and ignoring constant factors shows
that
s  !
kn
Rnm = O nmk log . (31.2)
m

To see that you cannot do much better than this, imagine interacting with m
adversarial bandit environments sequentially, each with horizon n/m. No matter
what policy you propose, there exist choices of bandits
p such that the expected
regret suffered against each bandit is at least Ω( nk/m). After summing over
the m instances, we see that the worst-case regret is at least
√ 
Rnm = Ω nmk ,
 31.1 Adversarial Bandits 379

which matches the upper bound except for logarithmic factors. Notice how this
lower bound applies to policies that know the location of the changes, so it is
not true that things are significantly harder in the absence of this knowledge.
There is one big caveat with all these calculations. The running time of a naive
implementation of Exp4 is linear in the number of experts, which even for modestly
sized m is very large indeed.

Online Stochastic Mirror Descent

The computational issues faced by Exp4 are most easily overcome using the
tools from online convex optimisation developed in Chapter 28. The idea is to
use online stochastic mirror descent and the unnormalised negentropy potential.
Without further modification, this would be Exp3, which you will show does not
work for non-stationary bandits (Exercise 31.3). The trick is to restrict the action
set to the clipped simplex A = Pk−1 ∩ [α, 1]k where α ∈ [0, 1/k] is a constant to
be tuned subsequently. The clipping ensures the algorithm does not commit too
hard to any single arm. The rationale is that a strong commitment could prevent
the discovery of change points.
Let F : [0, ∞)k → R be the unnormalised negentropy potential and P1 ∈ A be
the uniform probability vector. In each round t, the learner samples At ∼ Pt and
updates its sampling distribution using

Pt+1 = argminp∈A ηhp, Ŷt i + DF (p, Pt ) , (31.3)

where η > 0 is the learning rate and Ŷti = I {At = i} yti /Pti is the importance-
weighted estimator of the loss of action i for round t. The solution to the
optimisation problem of Eq. (31.3) can be computed efficiently using the two-step
process:

P̃t+1 = argminp∈[0,∞)k ηhp, Ŷt i + DF (p, Pt ) ,

Pt+1 = argminp∈A DF (p, P̃t+1 ) .

The first of these sub-problems can be evaluated analytically, yielding P̃t+1,i =

Pti exp(−η Ŷti ). The second can be solved efficiently using the result in
Exercise 26.12. The algorithm enjoys the following guarantee on its regret:

Theorem 31.1. The expected regret of the policy sampling At ∼ Pt with Pt

defined in Eq. (31.3) is bounded by

m log(1/α) ηnk
Rnm ≤ αn(k − 1) + + .
η 2
Pn
Proof Let a∗ ∈ argmina∈Γnm t=1 ytat be an optimal sequence of actions in
hindsight constrained to Γnm . Then let 1 = t1 < t2 < · · · < tm < tm+1 = n + 1
so that a∗t is constant on each interval {ti , . . . , ti+1 − 1}. We abuse notation by
 31.2 Stochastic Bandits 380

writing a∗i = a∗ti . Then the regret decomposes into

" n # " m ti+1 −1 #
X X X
Rnm = E (ytAt − yta∗t ) = E (ytAt − yta∗t )
t=1 i=1 t=ti
" "ti+1 −1  ##
m
X X 

= E E (ytAt − yta∗i )  Pti .
t=t

i=1 i

The next step is to apply Eq. (28.11) and the solution to Exercise 28.10 to bound
the inner expectation, giving
"ti+1 −1  # "ti+1 −1  #
X  X 
 
E (ytAt − yta∗i )  Pti = E hPt − ea∗i , yt i  Pti
 
t=ti t=ti
"  #
ti+1 −1
X 

≤ α(ti+1 − ti )(k − 1) + E max hPt − p, yt i  Pti
p∈A
t=ti

"  #
ti+1 −1
X 

= α(ti+1 − ti )(k − 1) + E max hPt − p, Ŷt i  Pti
p∈A
t=ti

  
D(p, Pti ) ηk(ti+1 − ti ) 
≤ α(ti+1 − ti )(k − 1) + E max +  Pti .
p∈A η 2
By assumption, Pti ∈ A and so Pti j ≥ α for all j and D(p, Pti ) ≤ log(1/α).
Combining this observation with the previous two displays shows that
m log(1/α) ηnk
Rnm ≤ nα(k − 1) + + .
η 2
The learning rate and clipping parameters are approximately optimised by
p p
η = 2m log(1/α)/(nk) and α = m/(nk) ,
p √
which leads to a regret of Rnm ≤ mnk log(nk/m) + mnk. In typical
applications,
p the value p
of m is not known. In this case one can choose η =
√
log(1/α)/nk and α = 1/nk, and the regret increases by a factor of O( m).

31.2 Stochastic Bandits

To keep things simple, we will assume the rewards are Gaussian and that for
each arm i there is a function µi : [n] → R, and the reward is
Xt = µAt (t) + ηt ,
where (ηt )nt=1 is a sequence of independent standard Gaussian random variables.
The optimal arm in round t has mean µ∗ (t) = maxi∈[k] µi (t) and the regret is
n
" n #
X X
Rn (µ) = µ (t) − E
∗
µAt (t) .
t=1 t=1
 31.2 Stochastic Bandits 381

The amount of non-stationarity is modelled by placing restrictions on the functions

µi : [n] → R. To be consistent with the previous section, we assume the mean
vector changes at most m − 1 times, which amounts to saying that
n−1
X
max I {µi (t) 6= µi (t + 1)} ≤ m − 1 .
i∈[k]
t=1

If the locations of the change points were known then, thanks to the concavity of
log, running a new copy of UCB on each interval would lead to a bound of
  n 
mk
Rn (µ) = O m + log , (31.4)
∆min m
where ∆min is the smallest suboptimality gap over all m blocks and n ≥ m. This
is a non-vacuous bound for n large. Inspired by the results of the last section
that showed that the bound achieved by an omniscient policy that knows when
the changes occur can be achieved by a policy that does not, one then wonders
whether the same holds concerning the bound in Eq. (31.4). As it turns out, the
answer in this case is no.

Theorem 31.2. Let k = 2, and fix ∆ ∈ (0, 1) and a policy π. Let µ be so that
µi (t) = µi is constant for both arms and ∆ = µ1 − µ2 > 0. If the expected regret
Rn (µ) of policy π on bandit µ satisfies Rn (µ) = o(n), then for all sufficiently
large n, there exists a non-stationary bandit µ0 with at most two change points
and mint∈[n] |µ01 (t) − µ02 (t)| ≥ ∆ such that Rn (µ0 ) ≥ n/(22Rn (µ)).

The theorem implies that if a policy enjoys Rn (µ) = o(n1/2 ) for any non-trivial
(stationary) bandit, then its minimax regret is at least ω(n1/2 ) on some non-
stationary bandit. In particular, if Rn (µ) = O(log(n)), then its worst-case regret
against non-stationary bandits with at most two changes is at least Ω(n/ log(n)).
This dashes our hopes for a policy that outperforms Exp4 in a stochastic setting
with switches, even in an asymptotic sense. The reason for the negative result
is that any algorithm anticipating the possibility of an abrupt change must
frequently explore all suboptimal arms to check that no change has occurred.
There are algorithms designed for non-stationary bandits in the stochastic
setting with abrupt change points as described above. Those that come with
theoretical guarantees are based on forgetting or discounting data so that decisions
of the algorithm depend almost entirely on recent data. In the notes, we discuss
these approaches along with alternative models for non-stationarity. For now,
the advantage of the stochastic setting seems to be that in the stochastic setting
there are algorithms that do not need to know the number of changes, while, as
noted beforehand, such algorithms are not yet known (or maybe not possible) in
the nonstochastic setting.

Proof of Theorem 31.2 Let (Sj )L j=1 be a uniform partition of [n] into successive
intervals. Let P and E[·] denote the probabilities and expectations with respect
to the bandit determined by µ and P0 with respect to alternative non-stationary
 31.3 Notes 382

bandit µ0 to be defined shortly. By the pigeonhole principle, there exists a j ∈ [L]

such that
 
X E[T2 (n)]
E I {At = 2} ≤ . (31.5)
L
t∈Sj

Define an alternative non-stationary banditp with µ (t) = µ

0
except for t ∈ Sj
when we let µ2 (t) = µ2 + ε, where ε = 2L/E[T2 (n)] while
0
µ01 (t) = µ1 . Then,
by Theorem 14.2 and Lemma 15.1,
   
X |S | X |S | 1
I {At = 2} ≥ + P0  I {At = 2} < exp (− D(P, P0 ))
j  j 
P ≥
2 2 2
t∈Sj t∈Sj
 
1 E[T2 (n)]ε2 1
≥ exp − = .
2 2L 2e
By Markov’s inequality and Eq. (31.5),
   
X |S | 2 X 2E[T2 (n)] 1
I {At = 2} ≥ I {At = 2} ≤
j
P ≤ E ≤ 2 ,
2 |Sj | L|Sj | ∆ |Sj |
t∈Sj t∈Sj
 
where the last inequality follows by choosing L = 2∆2 E[T2 (n)] and assuming n
is large enough that L ≤ n. Then ε ≥ 2∆ so that µ0 satisfies the assumptions of
the theorem. Therefore,
   
1 1 ε|Sj | 1 1 |Sj |∆ j n k 1 1
Rn (µ0 ) ≥ − 2 ≥ − 2 ≥ − .
2e ∆ |Sj | 4 2e ∆ |Sj | 2 L 4e∆ 2∆
Then, using Rn (µ) = ∆E [T2 (n)], the definition of L and the assumption that
Rn (µ) = o(n), it follows that for sufficiently large n,
n
Rn (µ0 ) ≥ .
22Rn (µ)
where the constant is chosen so that 1/22 < 1/(8e).

31.3 Notes

1 Environments that appear non-stationary can often be made stationary by

adding context. For example, when bandit algorithms are used for on-line
advertising, gym membership advertisements are received more positively in
January than July. A bandit algorithm that is oblivious to the time of year
will perceive this environment as non-stationary. You could tackle this problem
by using one of the algorithms in this chapter. Or you could use a contextual
bandit algorithm and include the time of year in the context. The reader is
encouraged to consider whether or not adding contextual information might
be preferable to using an algorithm designed for non-stationary bandits.
 31.3 Notes 383

2 The negative results for stochastic non-stationary bandits do not mean that
trying to improve on the adversarial bandit algorithms is completely hopeless.
First of all, the adversarial bandit algorithms are not well suited for exploiting
distributional assumptions on the noise, which makes things irritating when
the losses/rewards are Gaussian (which are unbounded) or Bernoulli (which
have small variance near the boundaries). There have been several algorithms
designed specifically for stochastic non-stationary bandits. When the reward
distributions are permitted to change abruptly, as in the last section, then the
two main algorithms are based on the idea of ‘forgetting’ rewards observed in
the distant past. One way to do this is with discounting. Let γ ∈ (0, 1) be
the discount factor , and define
t
X t
X
µ̂γi (t) = γ t−s I {As = i} Xs Tiγ (t) = γ t−s I {As = i} .
s=1 s=1

Then, for appropriately tuned constant α, the discounted UCB policy chooses
each arm once and subsequently
 v !
u k
u α X
At = argmaxi∈[k] µ̂γi (t − 1) + t γ log Tiγ (t − 1)  .
Ti (t − 1) i=1

The idea is to ‘discount’ rewards that occurred far in the past, which makes
the algorithm most influenced by recent events. A similar algorithm called
sliding-window UCB uses a similar approach, but rather than discounting past
rewards with a geometric discount function, it simply discards them altogether.
Let τ ∈ N+ be a constant, and define
t
X t
X
µ̂τi (t) = I {As = i} Xs Tiτ (t) = I {As = i} .
s=t−τ +1 s=t−τ +1

Then sliding-window UCB chooses

 r 
α
At = argmaxi∈[k] µ̂τi (t − 1) + log(t ∧ τ ) .
Tiτ (t − 1)
Regrettably, however, these algorithms suffer from a tuning problem.
p There
is no choice of γ and τ for which the algorithms enjoy Rn = O( n log(n)) in
a minimax sense. On the positive side, there is empirical evidence to support
the use of these algorithms when the stochastic assumption holds. Recently,
more complicated algorithms were proposed that can adapt to the number
of switches in a stochastic environment and match the regret of an optimally
tuned adversarial algorithm [Auer et al., 2019, Chen et al., 2019].
3 An alternative way to model non-stationarity in stochastic bandits is to assume
the mean pay-offs of the arms are slowly drifting. One way to do this is to
assume that µi (t) follows a reflected Brownian motion in some interval. It is
not hard to see that the regret is necessary linear in this case because the best
arm changes in any round with constant probability. The objective in this case
 31.4 Bibliographic Remarks 384

is to understand the magnitude of the linear regret in terms of the size of the
interval or volatility of the Brownian motion.
4 Yet another idea is to allow the means to change in an arbitrary way, but
restrict the amount of total variation. Let µt = (µ1 (t), . . . , µk (t)) and
n−1
X
Vn = kµt − µt+1 k∞
t=1

be the cumulative change in mean rewards measured in terms of the supremum

norm. Then, for each V ∈ [1/k, n/k], there exists a policy such that for all
bandits with Vn ≤ V , it holds that
Rn ≤ C(V k log(k))1/3 n2/3 . (31.6)
This bound is nearly tight in a minimax sense. The lower bound is obtained by
partitioning [n] into m parts, where in each part all arms
p have equal means
except for the optimal arm, which is better by ∆ = c mk/n for universal
√
constant c ∈ R. Thep usual argument shows that the total regret is Ω( kmn),
while Vn ≤ 2cm3/2 k/n. Tuning m so that Vn ≤ V completes the proof.
Recent work shows that it is possible to achieve Eq. (31.6) without knowing
V . That is, there exists an algorithm that is able to adapt to V . In fact,
the algorithm mentioned in Note 2, which is able to adapt to the number of
switches, can accomplish this.

31.4 Bibliographic Remarks

Non-stationary bandits have quite a long history. The celebrated Gittins index is
based on a model where each arm is associated with a Markov chain that evolves
when played, the reward depends on the state, and the state of the chosen Markov
chain is observed after it evolves [Gittins, 1979, Gittins et al., 2011]. The classical
approaches, as discussed in Chapter 35, address this problem in the Bayesian
framework, and the objective is primarily to design efficient algorithms rather
than understanding the frequentist regret. Even more related is the restless
bandit, which is the same as Gittins’s set-up except the Markov chain for every
arm evolves in every round, while the learner still only observes the state and
reward for the action they chose. As a result, the learner needs to reason about the
evolution of all the Markov chains, which makes this problem rather challenging.
Restless bandits were introduced by Whittle [1988] in the Bayesian framework,
where most of the results are not especially positive. There has been some interest
in a frequentist analysis, but the challenging nature of the problem makes it
difficult to design efficient algorithms with meaningful regret guarantees [Ortner
et al., 2012]. Certainly there is potential for more work in this area.
The ideas in Section 31.1 are mostly generalisations of algorithms designed
for the full information setting, notably the fixed share algorithm [Herbster and
Warmuth, 1998]. The first algorithm designed for the adversarial non-stationary
 31.5 Exercises 385

bandit is Exp3.S by Auer et al. [2002b]. This algorithm can be interpreted as an

efficient version of Exp4, where experts correspond to sequences of actions that
have the permitted number of changes and where the initialisation is carefully
chosen so that the computation needed to run Exp4 is made tractable [György
et al., 2019]. See also the analysis of fixed share in the book by Cesa-Bianchi
and Lugosi [2006]. The Exp3.P policy was originally developed in order to prove
high-probability bounds for finite-armed adversarial bandits [Auer et al., 2002b],
but Audibert and Bubeck [2010b] proved that with appropriate tuning it also
enjoys the same bounds as Exp3.S. Presumably this also holds for Exp3-IX.
Mirror descent has been used to prove tracking bounds in the full information
setting by Herbster and Warmuth [2001]. A more recent reference is by György
and Szepesvári [2016], which makes the justification for clipping explicit. The
latter paper considers the linear prediction setting and provides bounds on the
regret that scale with the complexity of the sequence of losses as measured by
the cumulative change of consecutive loss vectors. The advantage of this is that
the complexity measure can distinguish between abrupt and gradual changes.
This is similar to the approach of Besbes et al. [2014]. The lower bound for
stochastic non-stationary bandits is by Garivier and Moulines [2011], though
our proof differs in minor ways. We mentioned that there is a line of work on
stochastic non-stationary bandits where the rewards are slowly drifting. The
approach based on Brownian motion is due to Slivkins and Upfal [2008], while
the variant described in Note 4 is by Besbes et al. [2014], who also gave the lower
bound described there. The idea of discounted UCB was introduced without
analysis by Kocsis and Szepesvári [2006]. The analysis of this algorithm and
also of sliding-window UCB algorithm is by Garivier and Moulines [2011]. The
sliding-window algorithm has been extended to linear bandits [Cheung et al.,
2019] and learning in Markov decision processes [Gajane et al., 2018]. Contextual
bandits have also been studied in the non-stationary setting [Luo et al., 2018,
Chen et al., 2019].
p We are not aware of an algorithm for the adversarial setting
with Rn = O( mkn log(n)) when the √ number of switches is unknown. Auer
et al. [2018] prove a bound of Rn = O( mkn log(n)) in the stochastic setting
when k = 2. The idea underlying this work has been extended to the k-armed
case [Auer et al., 2019], as well as to the contextual case [Chen et al., 2019],
the latter of which also shows that adapting to the total shift of distributions
described in Note 4 is possible. The key novelty in these algorithms is adding
explicit exploration whose durations are multi-scale, which is made possible by
extra randomisation.

31.5 Exercises

31.1 (Exp4 for non-stationary bandits) Let n, m, k ∈ N+ . Prove (31.2).

In particular, specify first what the experts predict in each round and how
Theorem 18.1 gives rise to (31.1) and how (31.2) follows from (31.1).
 31.5 Exercises 386

Hint For the second part, you may find it useful to show the
following well-
Pm
known inequality: for 0 ≤ m ≤ n, defining Φm (n) = i=0 ni , it holds that
(m/n)m Φm (n) ≤ em .
31.2 (Lower bound for adversarial non-stationary bandits) Let
n, m, k ∈ N+ be such that n ≥ mk. Prove that for any policy π there exists an
adversarial bandit (yti ) such that
√
Rnm ≥ c nmk ,
where c > 0 is a universal constant.

31.3 (Unsuitability of Exp3 for non-stationary bandits) Prove for all

sufficiently large n that Exp3 from Chapter 11 has Rn2 ≥ cn for some universal
constant c > 0.

31.4 (Empirical comparison) Let k = 2 and n = 1000, and define adversarial

bandit in terms of losses with yt1 = I {t < n/2} and yt2 = I {t ≥ n/2}. Plot the
expected regret of Exp3, Exp3-IX and the variant of online stochastic mirror
descent proposed in this chapter. Experiment with a number of learning rates for
each algorithm.
32 Ranking

Ranking is the process of producing

an ordered shortlist of m items from a
larger collection of ` items. These tasks
come in several flavours. Sometimes the
user supplies a query, and the system
responds with a shortlist of items.
In other applications the shortlist is
produced without an explicit query.
For example, a streaming service might
provide a list of recommended movies
when you sign in. Our focus here is on
the second type of problem.
We examine a sequential version of
the ranking problem where the learner Figure 32.1 A classic ranking problem:
selects a ranking, receives feedback which hats to put where on the stand? Higher
about its quality and repeats the and towards the front attracts more attention.
process over n rounds. The feedback
will be in the form of ‘clicks’ from the user, which comes from the view that
ranking is a common application in on-line recommendation systems and the user
selects the items they like by clicking on them. The objective of the learner is to
maximise the expected number of clicks.

Ranking is a huge topic, and our approach is necessarily quite narrow. In fact
there is still a long way to go before we have a genuinely practical algorithm
for large-scale online ranking problems. As usual, we summarise alternative
ideas in the notes.

Stochastic Ranking
A permutation on [`] is an invertible function σ : [`] → [`]. Let A be the set of
all permutations on [`]. In each round t the learner chooses an action At ∈ A,
which should be interpreted as meaning the learner places item At (k) in the kth
position. Equivalently, A−1
t (i) is the position of the ith item. Since the shortlist
has length m, the order of At (m + 1), . . . , At (`) is not important and is included
 32.1 Click Models 388

only for notational convenience. After choosing their action, the learner observes
Cti ∈ {0, 1} for each i ∈ [`], where Cti = 1 if the user clicked on the ith item.
Note that the user may click on multiple items. We will assume a stochastic
model where the probability that the user clicks on position k in round t only
depends on At and is given by v(At , k), with v : A × [`] → [0, 1] an unknown
function. The regret over n rounds is
" n #
X̀ X X̀
Rn = n max v(a, k) − E Cti .
a∈A
k=1 t=1 i=1

A naive way to minimise the regret would be to create a finite-armed bandit

where each arm corresponds to a ranking of the items and then apply your
favourite algorithm from Part II. The problem is that these algorithms treat the
arms as independent and cannot exploit any structure in the ranking. This is
almost always unacceptable because the number of ways to rank m items from a
collection of size ` is `!/(` − m)!. Ranking illustrates one of the most fundamental
dilemmas in machine learning: choosing a model. A rich model leads to low
misspecification error, but takes longer to fit. A coarse model can suffer from
large misspecification error. In the context of ranking, a model corresponds to
assumptions on the function v.

32.1 Click Models

The only way to avoid the curse of dimensionality is to make assumptions. A

natural way to do this for ranking is to assume that the probability of clicking on
an item depends on (a) the underlying quality of that item and (b) the location of
that item in the chosen ranking. A formal definition of how this is done is called
a click model. Deciding which model to use depends on the particulars of the
problem at hand, such as how the list is presented to the user and whether or not
clicking on an item diverts them to a different page. This issue has been studied
by the data retrieval community, and there is now a large literature devoted
to the pros and cons of different choices. We limit ourselves to describing the
popular choices and give pointers to the literature at the end of the chapter.

Document-Based Model
The document-based model is one of the simplest click models, which assumes
the probability of clicking on a shortlisted item is equal to its attractiveness.
Formally, for each item i ∈ [`], let α(i) ∈ [0, 1] be the attractiveness of item i.
The document-based model assumes that

v(a, k) = α(a(k))I {k ≤ m} .

The unknown quantity in this model is the attractiveness function, which has
just ` parameters.
 32.1 Click Models 389

Position-Based Model
The document-based model might occasionally be justified, but in most cases the
position of an item in the ranking also affects the likelihood of a click. A natural
extension that accounts for this behaviour is called the position-based model,
which assumes that
v(a, k) = α(a(k))χ(k) ,
where χ : [`] → [0, 1] is a function that measures the quality of position k. Since
the user cannot click on items that are not shown, we assume that χ(k) = 0 for
k > m. This model is richer than the document-based model, which is recovered
by choosing χ(k) = I {k ≤ m}. The number of parameters in the position-based
models is m + `.

Cascade Model
The position-based model is not suitable for applications where clicking on an
item takes the user to a different page. In the cascade model, it is assumed
that the learner scans the shortlisted items in order and only clicks on the first
item they find attractive. Define χ : A × [`] → [0, 1] by

1 if k = 1


χ(a, k) = 0 if k > m

Qk−1

(1 − α(a(k ))) otherwise ,
0
0 k =1

which is the probability that the user has not clicked on the first k − 1 items.
Then the cascade model assumes that
v(a, k) = α(a(k))χ(a, k) . (32.1)
The first term in the factorisation is the attractiveness function, which measures
the probability that the user is attracted to the ith item. The second term can be
interpreted as the probability that the user examines that item. This interpretation
is also valid in the position-based model. It is important to emphasise that
v(a, k) is the probability of clicking on the kth position when taking action
a ∈ A. This does not mean that Ct1 , . . . , Ct` are independent. The assumptions
only restricts the marginal distribution of each Cti , which is sufficient for our
purposes. Nevertheless, in the cascade model, it would be standard to assume
that CtAt (k) = 0 if there exists an k 0 < k such that CtAt (k0 ) = 1, and otherwise
P(CtAt (k) = 1 | At , CtAt (1) = 0, . . . , CtAt (k−1) = 0) = I {k ≤ m} α(At (k)) .
Like the document-based model, the cascade model has ` parameters.

Generic Model
We now introduce a model that generalises the last three. Previous models
essentially assumed that the probability of a click factorises into an attractiveness
probability and an examination probability. We deviate from this norm by making
 32.1 Click Models 390

a a0
1
2 i j
3
4 j i
5

Figure 32.2 Part (c) of Assumption 32.1 says that the probability of clicking in the
second position on the left list is larger than the probability of clicking on the second
position on the right list by a factor of α(i)/α(j). For the fourth position, the probability
is larger for the right list than the left by the same factor.

assumptions directly on the function v. Given α : [`] → [0, 1], an action a is

called α-optimal if the shortlisted items are the m most attractive sorted by
attractiveness: α(a(k)) = maxk0 ≥k α(a(k 0 )) for all k ∈ [m].

Assumption 32.1. There exists an attractiveness function α : [`] → [0, 1] such

that the following four conditions are satisfied. Let a ∈ A and i, j, k ∈ [`] be such
that α(i) ≥ α(j), and let σ be the permutation that exchanges i and j.

(a) v(a, k) = 0 for all k > m.

Pm Pm
k=1 v(a , k) = maxa∈A k=1 v(a, k) for all α-optimal actions a .
∗ ∗
(b)
(c) For all i and j with α(i) ≥ α(j),
α(i)
v(a, a−1 (i)) ≥ v(σ ◦ a, a−1 (i)) ,
α(j)
where σ is the permutation on [`] that exchanges i and j.
(d) If a is an action such that α(a(k)) = α(a∗ (k)) for some α-optimal action a∗ ,
then v(a, k) ≥ v(a∗ , k) .

These assumptions may appear quite mysterious. At some level they are
chosen to make the proof go through, while simultaneously generalising the
document-based, position-based and cascade models (32.1). The choices are
not entirely without basis or intuition, however. Part (a) asserts that the user
does not click on items that are not placed in the shortlist. Part (b) says that
α-optimal actions maximise the expected number of clicks. Note that there
are multiple optimal rankings if α is not injective. Part (c) is a little more
restrictive and is illustrated in Fig. 32.2. One way to justify this is to assume
that v(a, k) = α(a(k))χ(a, k), where χ(a, k) is viewed as the probability that the
user examines position k. It seems reasonable to assume that the probability
the user examines position k should only depend on the first k − 1 items. Hence
v(a, 2) = α(i)χ(a, 2) = α(i)χ(a0 , 2) = α(i)/α(j)v(a0 , 2). In order the make the
argument for the fourth position, we need to assume that placing less attractive
items in the early slots increases the probability that the user examines later
positions (searching for a good result). This is true for the position-based and
 32.2 Policy 391

cascade models, but is perhaps the most easily criticised assumption. Part (d)
says that the probability that a user clicks on a position with a correctly placed
item is at least as large as the probability that the user clicks on that position in
an optimal ranking. The justification is that the items a(1), . . . , a(k − 1) cannot
be more attractive than a∗ (1), . . . , a∗ (k − 1), which should increase the likelihood
that the user makes it the kth position.
The generic model has many parameters, but we will see that the learner does
not need to learn all of them in order to suffer small regret. The advantage of
this model relative to the previous ones is that it offers more flexibility, and yet
it is not so flexible that learning is impossible.

32.2 Policy

We now explain the policy for learning to rank when v is unknown, but satisfies
Assumption 32.1. After the description is an illustration that may prove helpful.

Step 0: Initialisation
The policy takes as input a confidence parameter δ ∈ (0, 1) and ` and m. The
policy maintains a binary relation Gt ⊆ [`] × [`]. In the first round t = 1 the
relation is empty: G1 = ∅. You should think of Gt as maintaining pairs (i, j)
for which the policy has proven with high probability that α(i) < α(j). Ideally,
Gt ⊆ {(i, j) ∈ [`] × [`] : α(i) < α(j)}.

Step 1: Defining a Partition

In each round t, the learner computes a partition of the actions based on a
topological sort according to relation Gt . Given A ⊂ [`], define minGt (A) to be
the set of minimum elements of A according to relation Gt :
minGt (A) = {i ∈ A : (i, j) ∈
/ Gt for all j ∈ Gt } .
Then let Pt1 , Pt2 , . . . be the partition of [`] defined inductively by
d−1
!
[
Ptd = minGt [`] \ Ptc .
c=1

Finally, let Mt = max{d : Ptd 6= ∅}. The reader should check that if Gt does not
have cycles, then Mt is well defined and finite and that Pt1 , . . . , PtMt is indeed a
partition of [`] (Exercise 32.5). The event that Gt contains cycles is a failure event.
In order for the policy to be well defined, we assume it chooses some arbitrary
fixed action in this case.

Step 2: Choosing an Action

Let It1 , . . . , ItMt be a partition of [`] defined inductively by
Itd = [|∪c≤d Ptc |] \ [|∪c<d Ptc |] .
 32.2 Policy 392

Next let Σt ⊆ A be the set of actions σ such that σ(Itd ) = Ptd for all d ∈ [Mt ].
The algorithm chooses At uniformly at random from Σt . Intuitively the policy
first shuffles the items in Pt1 and uses these as the first |Pt1 | entries in the ranking.
Then Pt2 is shuffled, and the items are appended to the ranking. This process is
repeated until the ranking is complete. For an item i ∈ [`], we denote by Dti the
unique index d such that i ∈ Ptd .

Step 3: Updating the Relation

Pt Pt
For any pair of items i, j ∈ [`], define Stij = s=1 Usij and Ntij = s=1 |Usij |,
where

Utij = I {Dti = Dtj } (Cti − Ctj ) .

All this means is that Stij tracks the difference between the number of clicks
on items i and j over rounds when they share a partition. As a final step, the
relation Gt+1 is given by
 v !
 u p
u c Ntij 
Gt+1 = Gt ∪ (j, i) : Stij ≥ t2Ntij log ,
 δ 

where c ≈ 3.43 is the universal constant given in Exercise 20.10. In the analysis we
will show that if α(i) ≥ α(j), then with high probability Stji is never large enough
for Gt+1 to include (i, j). In this sense, with high probability, Gt is consistent
with the order on [`] induced by sorting in decreasing order with respect to α(·).
Note that Gt is generally not a partial order because it need not be transitive.

Illustration
Suppose ` = 5 and m = 4, and in round t the relation is Gt = {(3, 1), (5, 2), (5, 3)},
which is represented in the graph below, where an arrow from j to i indicates
that (j, i) ∈ Gt .

Pt1 1 2 4 It1 = {1, 2, 3}

Pt2 3 It2 = {4}

Pt3 5 It3 = {5}

This means that in round t the first three positions in the ranking will contain
items from Pt1 = {1, 2, 4} but with random order. The fourth position will be
item 3, and item 5 is not shown to the user.
 32.3 Regret Analysis 393

Part (a) of Assumption 32.1 means that items in position k > m are never
clicked. As a consequence, the algorithm never needs to actually compute
the partitions Ptd for which min Itd > m because items in these partitions
are never shortlisted.

32.3 Regret Analysis

Theorem 32.2. Let v satisfy Assumption 32.1, and assume that α(1) > α(2) >
· · · > α(`). Let ∆ij = α(i) − α(j) and δ ∈ (0, 1). Then the regret of TopRank is
bounded by
  √ 
X̀ min{m,j−1}
X 6(α(i) + α(j)) log c δ n
Rn ≤ δnm`2 + 1 + .
j=1 i=1
∆ ij

s  √ 
2 c n
Furthermore, Rn ≤ δnm` + m` + 4m3 `n log .
δ

By choosing δ = n−1 the theorem shows that the expected regret is at most
 
X̀ min{m,j−1}
X α(i) log(n) p 
Rn = O   and Rn = O m3 `n log(n) .
j=1 i=1
∆ij

The algorithm does not make use of any assumed ordering on α(·), so the
assumption is only used to allow for a simple expression for the regret. The core
idea of the proof is to show that (a) if the algorithm is suffering regret as a
consequence of misplacing an item, then it is gaining information so that Gt will
get larger and, (b) once Gt is sufficiently rich, the algorithm is playing optimally.
Let Ft = σ(A1 , C1 , . . . , At , Ct ) and Pt (·) = P(· | Ft ) and Et [·] = E[· | Ft ]. For each
t ∈ [n], let Ft be the failure event that there exists i = 6 j ∈ [`] and s < t such that
Nsij > 0 and
  q
 Xs  p
 
Ssij − Eu−1 [Uuij | Uuij 6= 0] |Uuij | ≥ 2Nsij log(c Nsij /δ) .
 
u=1

Lemma 32.3. Let i and j satisfy α(i) ≥ α(j) and d ≥ 1. On the event that
i, j ∈ Psd and d ∈ [Ms ] and Usij 6= 0, the following hold almost surely:
∆ij
(a) Es−1 [Usij | Usij 6= 0] ≥ .
α(i) + α(j)
(b) Es−1 [Usji | Usji 6 0] ≤ 0 .
=

Proof For the remainder of the proof, we focus on the event that i, j ∈ Psd and
d ∈ [Ms ] and Usij =
6 0. We also discard the measure zero subset of this event where
 32.3 Regret Analysis 394

Ps−1 (Usij =
6 0) = 0. From now on, we omit the ‘almost surely’ qualification on
conditional expectations. Under these circumstances, the definition of conditional
expectation shows that
Ps−1 (Csi = 1, Csj = 0) − Ps−1 (Csi = 0, Csj = 1)
Es−1 [Usij | Usij 6= 0] =
Ps−1 (Csi 6= Csj )
Ps−1 (Csi = 1) − Ps−1 (Csj = 1)
=
Ps−1 (Csi 6= Csj )
Ps−1 (Csi = 1) − Ps−1 (Csj = 1)
≥
Ps−1 (Csi = 1) + Ps−1 (Csj = 1)
Es−1 [v(As , As−1 (i)) − v(As , A−1
s (j))]
= , (32.2)
Es−1 [v(As , As (i)) + v(As , As (j))]
−1 −1

where in the second equality we added and subtracted Ps−1 (Csi = 1, Csj = 1).
By the design of TopRank, the items in Ptd are placed into slots Itd uniformly at
random. Let σ be the permutation that exchanges the positions of items i and j.
Then using Part (c) of Assumption 32.1,
X
Es−1 [v(As , A−1
s (i))] = Ps−1 (As = a)v(a, a−1 (i))
a∈A
α(i) X
≥ Ps−1 (As = a)v(σ ◦ a, a−1 (i))
α(j)
a∈A
α(i) X
= Ps−1 (As = σ ◦ a)v(σ ◦ a, (σ ◦ a)−1 (j))
α(j)
a∈A
α(i)
= Es−1 [v(As , A−1
s (j))] ,
α(j)
where the second equality follows from the fact that a−1 (i) = (σ ◦ a)−1 (j) and the
definition of the algorithm ensuring that Ps−1 (As = a) = Ps−1 (As = σ ◦ a). The
last equality follows from the fact that σ is a bijection. Using this and continuing
the calculation in Eq. (32.2) shows that
 
s (i)) − v(As , As (j))
Es−1 v(As , A−1 −1
Eq. (32.2) =  
s (i)) + v(As , As (j))
Es−1 v(As , A−1 −1

2
=1−    
1 + Es−1 v(As , As (i)) /Es−1 v(As , A−1
−1
s (j))
2
≥1−
1 + α(i)/α(j)
α(i) − α(j) ∆ij
= = .
α(i) + α(j) α(i) + α(j)
The second part follows from the first since Usji = −Usij .

The next lemma shows that the failure event occurs with low probability.

Lemma 32.4. It holds that P(Fn ) ≤ δ`2 .

 32.3 Regret Analysis 395

Proof The proof follows immediately from Lemma 32.3, the definition of Fn , the
union bound over all pairs of actions, and a modification of the Azuma–Hoeffding
inequality in Exercise 20.10.

Lemma 32.5. On the event Ftc , it holds that (i, j) ∈

/ Gt for all i < j.

Proof Let i < j so that α(i) ≥ α(j). On the event Ftc , either Nsji = 0 or
s r cp 
X
Ssji − Eu−1 [Uuji | Uuji 6= 0]|Uuji | < 2Nsji log Nsji for all s < t .
u=1
δ

When i and j are in different blocks in round u < t, then Uuji = 0 by definition.
On the other hand, when i and j are in the same block, Eu−1 [Uuji | Uuji 6= 0] ≤ 0
almost surely by Lemma 32.3. Based on these observations,
r cp 
Ssji < 2Nsji log Nsji for all s < t ,
δ

which by the design of TopRank implies that (i, j) ∈

/ Gt .

Lemma 32.6. Let Itd ∗

= min Ptd be the most attractive item in Ptd . Then, on
P
event Ft , it holds that Itd
c ∗
≤ 1 + c<d |Ptd | for all d ∈ [Mt ].
P
Proof Let i∗ = min ∪c≥d Ptc . Then i∗ ≤ 1 + c<d |Ptd | holds trivially for any
Pt1 , . . . , PtMt and d ∈ [Mt ]. Now consider two cases. Suppose that i∗ ∈ Ptd . Then
it must be true that i∗ = Itd ∗
, and our claim holds. On the other hand, suppose
that i ∈ Ptc for some c > d. Then by Lemma 32.5 and the design of the partition,
∗

there must exist a sequence of items id , . . . , ic in blocks Ptd , . . . , Ptc such that
id < · · · < ic = i∗ . From the definition of Itd ∗
, Itd
∗
≤ id < i∗ . This concludes our
proof.

Lemma 32.7. On the event Fnc and for all i < j, it holds that
 √ 
6(α(i) + α(j)) c n
Snij ≤ 1 + log .
∆ij δ

Proof The result is trivial when Nnij = 0. Assume from now on that Nnij > 0.
By the definition of the algorithm, arms i and j are not in the same block once
Stij grows too large relative to Ntij , which means that
r cp 
Snij ≤ 1 + 2Nnij log Nnij .
δ

On the event Fnc and part (a) of Lemma 32.3, it also follows that
r cp 
∆ij Nnij
Snij ≥ − 2Nnij log Nnij .
α(i) + α(j) δ
 32.3 Regret Analysis 396

Combining the previous two displays shows that

r cp  r cp 
∆ij Nnij
− 2Nnij log Nnij ≤ Snij ≤ 1 + 2Nnij log Nnij
α(i) + α(j) δ δ
r cp 
√
≤ (1 + 2) Nnij log Nnij . (32.3)
δ

Using the fact that Nnij ≤ n and rearranging the terms in the previous display
shows that

√  √ 
(1 + 2 2)2 (α(i) + α(j))2 c n
Nnij ≤ log .
∆2ij δ

The result is completed by substituting this into Eq. (32.3).

Proof of Theorem 32.2 The first step in the proof is an upper bound on the
expected number of clicks in the optimal list a∗ . Fix time t, block Ptd and recall
that Itd∗
= min Ptd is the most attractive item in Ptd . Let k = A−1t (Itd ) be the
∗

position of item Itd

∗
and σ be the permutation that exchanges items k and Itd ∗
.
By Lemma 32.6, on the event Ft , we have Itd ≤ k. From Parts (c) and (d) of
c ∗

Assumption 32.1, we have v(At , k) ≥ v(σ ◦ At , k) ≥ v(a∗ , k). Hence, on the event
Ftc , the expected number of clicks on Itd
∗
is bounded from below by those on
items in a ,
∗

h i X
Et−1 CtItd
∗ = Pt−1 (A−1
t (Itd ) = k)Et−1 [v(At , k) | At (Itd ) = k]
∗ −1 ∗

k∈Itd
1 X 1 X
= Et−1 [v(At , k) | A−1
t (Itd ) = k] ≥
∗
v(a∗ , k) ,
|Itd | |Itd |
k∈Itd k∈Itd

where we also used the fact that TopRank randomises within each block to
guarantee that Pt−1 (A−1
t (Itd ) = k) = 1/|Itd | for any k ∈ Itd . Using this and the
∗

design of TopRank,

m
X Mt X
X Mt
X h i
v(a∗ , k) = v(a∗ , k) ≤ |Itd |Et−1 CtItd
∗ .
k=1 d=1 k∈Itd d=1

Therefore, under event Ftc , the conditional expected regret in round t is bounded
 32.4 Notes 397

by
   
m
X X̀ XMt X̀
v(a∗ , k) − Et−1  Ctj  ≤ Et−1  |Ptd |CtItd
∗ − Ctj 
k=1 j=1 d=1 j=1
 
XMt X
= Et−1  (CtItd
∗ − Ctj )

d=1 j∈Ptd
Mt
X X
= Et−1 [UtItd
∗ j]

d=1 j∈Ptd

X̀ min{m,j−1}
X
≤ Et−1 [Utij ] . (32.4)
j=1 i=1

Pmin{m,j−1}
The last inequality follows by noting that Et−1 [UtItd
∗ j] ≤
i=1 Et−1 [Utij ].
To see this, use part (a) of Lemma 32.3 to show that Et−1 [Utij ] ≥ 0 for i < j and
Lemma 32.6 to show that when Itd ∗
> m, then neither Itd∗
nor j are not shown to
the user in round t so that UtItd
∗ j = 0. Substituting the bound in Eq. (32.4) into

the regret leads to

X̀ min{m,j−1}
X
Rn ≤ nmP(Fn ) + E [I {Fnc } Snij ] , (32.5)
j=1 i=1

where we used the fact that the maximum number of clicks over n rounds is
nm. The proof of the first part is completed by using Lemma 32.4 to bound
the first term and Lemma 32.7 to bound the second. The problem-independent
bound follows from Eq. (32.5) and by stopping early in the proof of Lemma 32.7
(Exercise 32.6).

32.4 Notes

1 At no point in the analysis did we use the fact that v is fixed over time. Suppose
that v1 , . . . , vn are a sequence of click-probability functions that all satisfy
Assumption 32.1 with the same attractiveness function. The regret in this
setting is
n Xm
" n #
X X X̀
Rn = vt (a∗ , k) − E Cti .
t=1 k=1 t=1 i=1

Then the bounds in Theorem 32.2 still hold without changing the algorithm.
2 The cascade model is usually formalised in the following more restrictive fashion.
Let {Zti : i ∈ [`], t ∈ [n]} be a collection of independent Bernoulli random
variables with P (Zti = 1) = α(i). Then define Mt as the first item i in the
 32.4 Notes 398

shortlist with Zti = 1:


Mt = min k ∈ [m] : ZtAt (k) = 1 ,

where the minimum of an empty set is ∞. Finally let Cti = 1 if and only if
Mt ≤ m and At (Mt ) = i. This set-up satisfies Eq. (32.1), but the independence
assumption makes it possible to estimate α without randomisation. Notice that
in any round t with Mt ≤ m, all items i with A−1 t (i) < Mt must have been
unattractive (Zti = 0), while the clicked item must be attractive (Zti = 1).
This fact can be used in combination with standard concentration analysis to
estimate the attractiveness. The optimistic policy sorts the ` items in decreasing
order by their upper confidence bounds and shortlists the first m. When the
confidence bounds are derived from Hoeffding’s inequality , this policy is called
CascadeUCB, while the policy that uses Chernoff’s lemma is called CascadeKL-
UCB. The computational cost of the latter policy is marginally higher than
the former, but the improvement is also quite significant because in practice
most items have barely positive attractiveness.
3 The linear dependence of the regret on ` is unpleasant when the number of
items is large, which is the case in many practical problems. Like for finite-
armed bandits, one can introduce a linear structure on the items by assuming
that α(i) = hθ, φi i where θ ∈ Rd is an unknown parameter vector and (φi )`i=1
are known feature vectors. This has been investigated in the cascade model by
Zong et al. [2016] and with a model resembling that of this chapter by Li et al.
[2019a].
4 There is an adversarial variant of the cascade model. In the ranked bandit
model an adversary secretly chooses a sequence of sets S1 , . . . , Sn , with St ⊆ [`].
In each round t the learner chooses At ∈ A and receives a reward Xt (At ), where
Xt : A → [0, 1] is given by Xt (a) = I {St ∩ {a(1), . . . , a(k)} =
6 ∅}. The feedback
is the position of the clicked action, which is Mt = min{k ∈ [m] : At (k) ∈ St }.
The regret is
n
X
Rn = (Xt (a∗ ) − Xt (At )) ,
t=1

where a∗ is the optimal ranking in hindsight:

n
X
a∗ = argmina∈A Xt (a) . (32.6)
t=1

Notice that this is the same as the cascade model when St = {i : Zti = 1}.
5 A challenge in the ranked bandit model is that solving the offline problem (Eq.
32.6) for known S1 , . . . , Sn is NP-hard. How can one learn when finding an
optimal solution to the offline problem is hard? First, hardness only matters if
|A| is large. When ` and m are not too large, then exhaustive search is quite
feasible. If this is not an option, one may use an approximation algorithm.
It turns out that in a certain sense, the best one can do is to use a greedy
 32.5 Bibliographic Remarks 399

algorithm, We omit the details, but the highlight is that there exist efficient
algorithms such that
" n #  
X 1 Xn  p 
E Xt (At ) ≥ 1 − max Xt (a) − O m n` log(`) .
t=1
e a∈A t=1

See the article by Radlinski et al. [2008] for more details.

6 By modifying the reward function, one can also define an adversarial variant
of the document-based model. As in the previous note, the adversary secretly
chooses S1 , . . . , Sn as subsets of [`], but now the reward is

Xt (a) = |St ∩ {a(1), . . . , a(k)}| .

The feedback is the positions of the clicked items, St ∩ {a(1), . . . , a(k)}. For
this model, there are no computation issues. In fact, the problem can be
analysed using a reduction to combinatorial semi-bandits, which we ask you to
investigate in Exercise 32.3.
7 The position-based model can also be modelled in the adversarial setting by
letting Stk ⊂ [`] for each t ∈ [n] and k ∈ [m]. Then, defining the reward by
m
X
Xt (a) = I {At (k) ∈ Stk } .
k=1

Again, the feedback is the positions of the clicked items, {k ∈ [m] : At (k) ∈ Stk }.
This model can also be tackled using algorithms for combinatorial semi-bandits
(Exercise 32.4).

32.5 Bibliographic Remarks

The policy and analysis presented in this chapter is by the authors and others
[Lattimore et al., 2018]. The most related work is by Zoghi et al. [2017], who
assumed a factorisation of the click probabilities v(a, k) = α(a(k))χ(a, k) and then
made assumptions on χ. The assumptions made here are slightly less restrictive,
and the bounds are simultaneously stronger. Some experimental results comparing
these algorithms are given by Lattimore et al. [2018]. For more information on
click models, we recommend the survey paper by Chuklin et al. [2015] and the
article by Craswell et al. [2008]. Cascading bandits were first studied by Kveton
et al. [2015a], who proposed algorithms based on UCB and KL-UCB and prove
finite-time instance-dependence upper bounds and asymptotic lower bounds that
match in specific regimes. Around the same time, Combes et al. [2015a] proposed
a different algorithm for the same model that is also asymptotically optimal. The
optimal regret has a complicated form and is not given explicitly in all generality.
We remarked in the notes that the linear dependence on ` is problematic for
large `. To overcome this problem, Zong et al. [2016] introduce a linear variant
where the attractiveness of an item is assumed to be an inner product between an
 32.6 Exercises 400

unknown parameter and a known feature vector. A slightly generalised version of

this set-up was simultaneously studied by Li et al. [2016], who allowed the features
associated with each item to change from round to round. The position-based
model is studied by Lagree et al. [2016], who suggest several algorithms and
provide logarithmic regret analysis for some of them. Asymptotic lower bounds
are also given that match the upper bounds in some regimes. Katariya et al. [2016]
study the dependent click model introduced by Guo et al. [2009]. This differs
from the models proposed in this chapter because the reward is not assumed
to be the number of clicks and is actually unobserved. We leave the reader to
explore this interesting model on their own. The adversarial variant of the ranking
problem mentioned in the notes is due to Radlinski et al. [2008]. Another related
problem is the rank-1 bandit problem, where the learner chooses one of ` items
to place in one of m positions, with all other positions left empty. This model has
been investigated by Katariya et al. [2017a,b], who assume the position-based
model. The cascade feedback model is also used in a combinatorial setting by
Kveton et al. [2015c], but this paper does not have a direct application to ranking.
A more in-depth discussion on ranking can be found in the recent book on bandits
in information retrieval by Glowacka [2019], which discusses a number of practical
considerations, like the cold-start problem.

32.6 Exercises

32.1 (Click models and assumptions) Show that the document-based,

position-based and cascade models all satisfy Assumption 32.1.

32.2 (Diversity) Most ranking algorithms are based on assigning an

attractiveness value to each item and shortlisting the m most attractive items.
Radlinski et al. [2008] criticise this approach in their paper as follows:
“The theoretical model that justifies ranking documents in this way is the probabilistic
ranking principle [Robertson, 1977]. It suggests that documents should be ranked by their
probability of relevance to the query. However, the optimality of such a ranking relies
on the assumption that there are no statistical dependencies between the probabilities
of relevance among documents – an assumption that is clearly violated in practice. For
example, if one document about jaguar cars is not relevant to a user who issues the
query jaguar, other car pages become less likely to be relevant. Furthermore, empirical
studies have shown that given a fixed query, the same document can have different
relevance to different users [Teevan et al., 2007]. This undermines the assumption that
each document has a single relevance score that can be provided as training data to the
learning algorithm. Finally, as users are usually satisfied with finding a small number of,
or even just one, relevant document, the usefulness and relevance of a document does
depend on other documents ranked higher.”

The optimality criterion Radlinski et al. [2008] had in mind is to present at least
one item that the user is attracted to. Do you find this argument convincing?
Why or why not?
 32.6 Exercises 401

The probabilistic ranking principle was put forward by Maron and Kuhns
[1960]. The paper by Robertson [1977] identifies some sufficient conditions
under which the principle is valid and also discusses its limitations.

32.3 (Adversarial ranking as a semi-bandit (i)) Frame the adversarial

variant of the document-based model in Note 6 as a combinatorial semi-bandit
and use the results in Chapter 30 to prove a bound on the regret of
p
Rn ≤ 2m`n(1 + log(`)) .

32.4 (Adversarial ranking as a semi-bandit (ii)) Adapt your solution to

the previous exercise to the position-based model in Note 7, and prove a bound
on the regret of
p
Rn ≤ m 2`n(1 + log(`)) .

32.5 (Cycles in partial order) Prove that if Gt does not contain cycles, then
Mt defined in Section 32.2 is well defined and that Pt1 , . . . , PtMt is a partition of
[`].

32.6 (Worst-case bound for TopRank) Prove the second part of

Theorem 32.2.
33 Pure Exploration

All the policies proposed in this book so far were

designed to maximise the cumulative reward. As
a consequence, the policies must carefully balance
exploration against exploitation. But what happens
if there is no price to be paid for exploring? Imagine,
for example, that a researcher has k configurations
of a new drug and a budget to experiment on
n mice. The researcher wants to find the most
promising drug configuration for subsequent human Figure 33.1 The mouse never
trials, but is not concerned with the outcomes benefits from the experiment.
for the mice. Problems of this nature are called
pure exploration problems. Although there are
similarities to the cumulative regret setting, there are also differences. This
chapter outlines a variety of pure exploration problems and describes the basic
algorithmic ideas.

33.1 Simple Regret

Let ν be a k-armed stochastic bandit and π = (πt )n+1

t=1 be a policy. One way to
measure the performance of a policy in the pure exploration setting is the simple
regret,
 
Rnsimple (π, ν) = Eνπ ∆An+1 (ν) .
The action chosen in round n + 1 has a special role. In the example with the mice,
it represents the configuration recommended for further investigation at the end
of the trial. We start by analysing the uniform exploration (UE) policy, which
explores deterministically for the first n rounds and recommends the empirically
best arm in round n + 1. The pseudocode is provided in Algorithm 20.
Theorem 33.1. Let π be the policy of Algorithm 20 and ν ∈ ESG k
(1) be a 1-
subgaussian bandit. Then, for all n ≥ k,
 
 2

X bn/kc ∆i (ν) 
Rnsimple (π, ν) ≤ min ∆ + ∆i (ν) exp − .
∆≥0 4
i:∆i (ν)>∆
 33.1 Simple Regret 403

1: for t = 1, . . . , n do
2: Choose At = 1 + (t mod k)
3: end for
4: Choose An+1 = argmaxi∈[k] µ̂i (n)

Algorithm 20: Uniform exploration.

Proof Let ∆i = ∆i (ν) and P = Pνπ . Assume without loss of generality that
∆1 = 0, and let i be a suboptimal arm with ∆i > ∆. Observe that An+1 = i
implies that µ̂i (n) ≥ µ̂1 (n). Now Ti (n) ≥ bn/kc is not random, so by Theorem 5.3
and Lemma 5.4,
 
bn/kc ∆2i
P (µ̂i (n) ≥ µ̂1 (n)) = P (µ̂i (n) − µ̂1 (n) ≥ 0) ≤ exp − . (33.1)
4

The definition of the simple regret yields

k
X X
Rnsimple (π, ν) = ∆i P (An+1 = i) ≤ ∆ + ∆i P (An+1 = i) .
i=1 i:∆i >∆

The proof is completed by substituting Eq. (33.1) and taking the minimum over
all ∆ ≥ 0.

The theorem highlights some important differences between the simple regret
and the cumulative regret. If ν is fixed and n tends to infinity, then the simple
regret converges to zero exponentially fast. On the other hand, if n is fixed and ν
is allowed to vary, then we are in a worst-case regime.
p Theorem 33.1 can be used
to derive a bound in this case by choosing ∆ = 2 log(k)/ bn/kc, which after a
short algebraic calculation shows that for n ≥ k there exists a universal constant
C > 0 such that
r
k log(k)
Rn (UE, ν) ≤ C
simple
for all ν ∈ ESG
k
(1) . (33.2)
n
In Exercise 33.1 we ask you to use the techniques of Chapter 15 to prove pthat for
all policies there exists a bandit ν ∈ EN (1) such that Rn (π, ν) ≥ C k/n for
k simple

some universal constant C > 0. It turns out the logarithmic dependence on k in

Eq. (33.2) is tight for uniform exploration (Exercise 33.2), but there exists another
policy for which the simple regret matches the aforementioned lower bound up to
constant factors. There are several ways to do this, but the most straightforward
is via a reduction from algorithms designed for minimising cumulative regret.

Proposition 33.2. Let π = (πt )nt=1 be a policy, and define

1X
n
πn+1 (i | a1 , x1 , . . . , an , xn ) = I {at = i} .
n t=1
 33.2 Best-Arm Identification with a Fixed Confidence 404

Then the simple regret of (πt )n+1

t=1 satisfies

Rn (π, ν)
Rnsimple ((πt )n+1
t=1 , ν) = ,
n
where Rn (π, ν) is the cumulative regret of policy π = (πt )nt=1 on bandit ν.
Proof By the regret decomposition identity (4.5),
" k #
X Ti (n)  
Rn (π, ν) = nE ∆i = nE ∆An+1 = nRnsimple ((πt )n+1
t=1 , ν) ,
i=1
n

where the first equality follows from the definition of the cumulative regret, the
third from the definition of πn+1 and the last from the definition of the simple
regret.
Corollary 33.3. For all n there exists a policypπ such that for all ν ∈ ESG
k
(1)
with ∆(ν) ∈ [0, 1] it holds that Rn (π, ν) ≤ C k/n, where C is a universal
k simple

constant.
Proof Combine the previous result with Theorem 9.1.
Proposition 33.2 raises our hopes that policies designed for minimising the
cumulative regret might also have well-behaved simple regret. Unfortunately this
is only true in the intermediate regimes where the best arm is hard to identify.
Policies with small cumulative regret spend most of their time playing the optimal
arm and play suboptimal arms just barely enough to ensure they are not optimal.
In pure exploration this leads to a highly suboptimal policy for which the simple
regret is asymptotically polynomial, while we know from Theorem 33.1 that the
simple regret should decrease exponentially fast. More details and pointers to
the literature are given in Note 2 at the end of the chapter.

33.2 Best-Arm Identification with a Fixed Confidence

Best-arm identification is a variant of pure exploration where the learner is

rewarded only for identifying an exactly optimal arm. There are two variants of
best-arm identification. In this section we consider the fixed confidence setting
when the learner is given a confidence level δ ∈ (0, 1) and should use as few
samples as possible to output an arm that is optimal with probability at least
1 − δ. In the other variant the learner has to make a decision after n rounds and
the goal is to minimise the probability of selecting a suboptimal arm. We treat
this alternative in the next section.
In the fixed confidence setting, the learner chooses a policy π = (πt )∞ t=1
as normal. The number of rounds is not fixed in advance, however, the
learner chooses a stopping time τ adapted to filtration F = (Ft )∞ t=0 with
Ft = σ(A1 , X1 , . . . , At , Xt ). The learner also chooses a Fτ -measurable random
variable ψ taking values in [k]. The stopping time represents the time when the
 33.2 Best-Arm Identification with a Fixed Confidence 405

learner halts and ψ ∈ [k] is the recommended action, which by the measurability
assumption only depends on (A1 , X1 , . . . , Aτ , Xτ ). Note that in line with our
definition of stopping times (see Definition 3.6), it is possible that τ = ∞, which
just means the learner cannot ever make up their mind to stop. This behaviour
of a learner, of course, will not be encouraged! The function ψ is called the
selection rule.
Definition 33.4. A triple (π, τ, ψ) is sound at confidence level δ ∈ (0, 1) for
environment class E if for all ν ∈ E,
Pνπ (τ < ∞ and ∆ψ (ν) > 0) ≤ δ . (33.3)
The objective in fixed confidence best-arm identification is to find a sound
learner for which Eνπ [τ ] is minimised over environments ν ∈ E. Since this is a
multi-objective criteria, there is a priori no reason to believe that a single optimal
learner should exist. Conveniently, however, the condition that the learner must
satisfy Eq. (33.3) plays the role of the consistency assumption in the asymptotic
lower bounds in Chapter 16, which allows for a sense of instance-dependent
asymptotic optimality. The situation in finite time is more complicated, as we
discuss in Note 7.

If E is sufficiently rich and ν has multiple optimal arms, then no sound learner
can stop in finite time with positive probability. The reason is that there is
no way to reject the hypothesis that one optimal arm is fractionally better
than another. You will investigate this in Exercise 33.10. Also note that
in our definition, I {τ = t} is a deterministic function of A1 , X1 , . . . , At , Xt .
None of the results that follow would change if you allowed τ or ψ to also
depend on some exogenous source of randomness.

33.2.1 Lower Bound

We start with the lower bound, which serves as a target for the upper bound to
follow. Let E be an arbitrary set of k-armed stochastic bandit environments, and
for ν ∈ E define
i∗ (ν) = argmaxi∈[k] µi (ν) and Ealt (ν) = {ν 0 ∈ E : i∗ (ν 0 ) ∩ i∗ (ν) = ∅} ,
which is the set of bandits in E with different optimal arms than ν.
Theorem 33.5. Assume that (π, τ, ψ) is sound
for E at confidence level δ ∈ (0, 1),
1
and let ν ∈ E. Then Eνπ [τ ] ≥ c (ν) log 4δ , where
∗

k
!!
X
c (ν) = sup
∗ −1
inf αi D(νi , νi )
0
(33.4)
α∈Pk−1 0 ν ∈Ealt (ν)
i=1

with c (ν) = ∞ when c (ν)

∗ ∗ −1
= 0.
 33.2 Best-Arm Identification with a Fixed Confidence 406

Proof The result is trivial when Eνπ [τ ] = ∞. For the remainder, assume that
Eνπ [τ ] < ∞, which implies that Pνπ (τ = ∞) = 0. Next, let ν 0 ∈ Ealt (ν) and
define event E = {τ < ∞ and ψ ∈ / i∗ (ν 0 )} ∈ Fτ . Then,
2δ ≥ Pνπ (τ < ∞ and ψ ∈
/ i∗ (ν)) + Pν 0 π (τ < ∞ and ψ ∈
/ i∗ (ν 0 ))
≥ Pνπ (E c ) + Pν 0 π (E)
!
1 Xk
≥ exp − Eνπ [Ti (τ )] D(νi , νi ) ,
0
(33.5)
2 i=1

where the first inequality follows from the definition of soundness and the last
from the Bretagnolle–Huber inequality (Theorem 14.2) and the stopping time
version of Lemma 15.1 (see Exercise 15.7). The second inequality holds because
Pνπ (τ = ∞) = 0 and i∗ (ν) ∩ i∗ (ν 0 ) = ∅ and
E c = {τ = ∞} ∪ {τ < ∞ and ψ ∈ i∗ (ν 0 )}
⊆ {τ = ∞} ∪ {τ < ∞ and ψ ∈
/ i∗ (ν)} .
Rearranging Eq. (33.5) shows that
 
1
k
X
Eνπ [Ti (τ )] D(νi , νi0 ) ≥ log , (33.6)
i=1
4δ

which implies that Eνπ [τ ] > 0. Using this, the definition of c∗ (ν) and Eq. (33.6),

Eνπ [τ ] Xk
= Eνπ [τ ] sup 0 inf αi D(νi , νi0 )
c∗ (ν) α∈Pk−1 ν ∈Ealt (ν) i=1

Eνπ [Ti (τ )]
k
X
≥ Eνπ [τ ] inf D(νi , νi0 ) (33.7)
ν 0 ∈Ealt (ν)
i=1
Eνπ [τ ]
k
X
= inf Eνπ [Ti (τ )] D(νi , νi0 )
ν 0 ∈Ealt (ν)
i=1
 
1
≥ log ,
4δ
where the last inequality follows from Eq. (33.6). Rearranging completes the proof.
Note, in the special case that c∗ (ν)−1 = 0, the assumption that Eνπ [τ ] < ∞
would lead to a contradiction.

Theorem 33.5 does not depend on E being unstructured. The assumption

that the bandits are finite armed could also be relaxed with appropriate
measureability assumptions.

In a moment, we will prove that the bound in Theorem 33.5 is asymptotically

optimal as δ → 0 when E = EN k
(1), a result that holds more generally for
Bernoulli bandits or when the distributions come from an exponential family.
 33.2 Best-Arm Identification with a Fixed Confidence 407

Before this, we devote a little time to understanding the constant c∗ (ν). Suppose
that α∗ (ν) ∈ Pk−1 satisfies
k
X
c∗ (ν)−1 = inf αi∗ (ν) D(νi , νi0 ) .
ν 0 ∈Ealt (ν)
i=1

A few observations about this optimisation problem:

(a) The value of α∗ (ν) is unique when E = EN k

(1) and ν ∈ E has a unique
optimal arm. Uniqueness continues to hold when E is unstructured with
distributions from an exponential family.
(b) The inequality in Eq. (33.7) is tightest when Eνπ [Ti (τ )]/Eνπ [τ ] = αi∗ (ν),
which shows a policy can only match the lower bound by playing arm i
exactly in proportion to αi∗ (ν) in the limit as δ tends to zero.
2
(c) When E = EN (1) and ν ∈ E has a unique optimal arm, then
1 
c∗ (ν)−1 = sup 0 inf α(µ1 (ν) − µ1 (ν 0 ))2 + (1 − α)(µ2 (ν) − µ2 (ν 0 ))2
2 α∈[0,1] ν ∈Ealt (ν)
1 2 1 2
= sup α(1 − α) (µ1 (ν) − µ2 (ν)) = (µ1 (ν) − µ2 (ν)) .
2 α∈[0,1] 8

In this case we observe that α1∗ (ν) = α2∗ (ν) = 1/2.

(d) Suppose that σ 2 ∈ (0, ∞)k is fixed and E = {(N (µi , σi2 )ki=1 : µ ∈ Rk }. You
are asked in Exercise 33.4 to verify that when k = 2,
2(σ1 + σ2 )2
c∗ (ν) = . (33.8)
∆22
which unsurprisingly shows the problem becomes harder as the variance of
either of the arms increases. In Exercise 33.4, you will show when k ≥ 2, it
holds that
v
u 2 X 2
2σi2∗ X 2σ 2 2σ 2 X 2σ 2 u 2σi∗ 2σi
+ (ν) + + 2
i ∗ i ∗ i
2 2 ≤ c ≤ 2 2
t 2 ,
∆min ∆i ∆min ∆i ∆min ∆2i
i6=i
∗ ∗ i6=i ∗ i6=i

where ∆min = mini6=i∗ ∆i is the smallest suboptimality gap. This bound

faithfully captures the intuition that each suboptimal arm must be played
sufficiently often to be distinguished from the optimal arm, while the optimal
arm must be observed sufficiently many times so that it can be distinguished
from the second best arm. For k = 2, this bound is smaller than the value
of c∗ (ν), as shown in (33.8), showing that there is room for improvement in
this case.

33.2.2 Policy, Stopping/Selection Rule and Upper Bounds

The bound in Theorem 33.5 is asymptotically tight for many environment classes.
For simplicity, we focus on the Gaussian case.
 33.2 Best-Arm Identification with a Fixed Confidence 408

For this section, we assume that E = EN

k
(1) is the set of k-armed Gaussian
bandits with unit variance.

We need to construct a triple (π, τ, ψ) that is sound for E and for which Eνπ [τ ]
matches the lower bound in Theorem 33.5 as δ → 0. Both are derived using
the insights provided by the lower bound. The policy should choose action i in
proportion to αi∗ (ν), which must be estimated from data. The stopping rule is
motivated by noting that Eq. (33.6) implies that a sound stopping rule must
satisfy
 
1
X k
Eνπ [Ti (τ )] D(νi , νi0 ) ≥ log for all ν 0 ∈ Ealt (ν) .
i=1
4δ

If the inequality is tight, then we might guess that a reasonable stopping rule as
the first round t when
 
1
Xk
inf Ti (t) D(ν ,
i iν 0
) & log .
ν ∈Ealt (ν)
0 δ
i=1

There are two problems: (a) ν is unknown, so the expression cannot be evaluated,
and (b) we have replaced the expected number of pulls with the actual number of
pulls. Still, let us persevere. To deal with the first problem, we can try replacing
ν by the Gaussian bandit environment with mean vector µ̂(t), which we denote
by ν̂(t). Then let

1
k
X Xk
Zt = inf Ti (t) D(ν̂i (t), νi0 ) = inf Ti (t)(µ̂i (t) − µi (ν 0 ))2 .
ν 0 ∈Ealt (ν̂(t))
i=1
2 µ0 ∈Ealt (µ̂(t)) i=1

We will show there exists a choice of βt (δ) ≈ log(t/δ) such that if τ = min{t : Zt >
βt (δ)}, then the empirically optimal arm at time τ is the best arm with probability
at least 1 − δ. The next step is to craft a policy for which the expectation of τ
matches the lower bound asymptotically. As we remarked earlier, if the policy
is to match the lower bound, it should play arm i approximately in proportion
to αi∗ (ν). This suggests estimating α∗ (ν) by α̂(t) = α∗ (ν̂(t)) and then playing
the arm for which tα̂i (t) − Ti (t) is maximised. If α̂(t) is inaccurate, then perhaps
the samples collected will not allow the algorithm to improve its estimates. To
overcome this last challenge, the policy includes enough forced exploration to
ensure that eventually α̂(t) converges to α∗ (ν) with high probability. Combining
all these ideas leads to the track-and-stop policy (Algorithm 21).

Theorem 33.6. Let (π, τ, ψ) be the policy, stopping time and selection rule of
track-and-stop (Algorithm 21). There exists a choice of βt (δ) such that track-and-
stop is sound and for all ν ∈ E with |i∗ (ν)| = 1 it holds that
Eνπ [τ ]
lim = c∗ (ν) .
δ→0 log(1/δ)
 33.2 Best-Arm Identification with a Fixed Confidence 409

1: Input δ and βt (δ)

2: Choose each arm once and set t = k
3: while Zt < βt (δ) do √
4: if argmini∈[k] Ti (t) ≤ t then
5: Choose At+1 = argmini∈[k] Ti (t)
6: else
7: Choose At+1 = argmaxi∈[k] (tα̂i∗ (t) − Ti (t))
8: end if
9: Observe reward Xt+1 , update statistics and increment t
10: end while
11: return ψ = i∗ (ν̂(t)), τ = t
Algorithm 21: Track-and-stop.

Note that only π does not depend on δ inside the limit statement of the theorem,
but the stopping time does. The following lemma guarantees the soundness of
(π, τ, ψ).
Lemma 33.7. Let f : [k, ∞) → R be given by f (x) = exp(k − x)(x/k)k and
βt (δ) = k log(t2 + t) + f −1 (δ). Then, for τ = min{t : Zt ≥ βt (δ)}, it holds that
P (i∗ (ν̂(τ )) 6= i∗ (ν)) ≤ δ.

The inverse f −1 (δ) is well defined because f is strictly decreasing on [k, ∞)

with f (k) = 1 and limx→∞ f (x) = 0. In fact, the inverse has a closed-
form solution in terms of the Lambert W function. By staring at the form
of f one can check that limδ→0 f −1 (δ)/ log(1/δ) = 1 or equivalently that
f −1 (δ) = (1 + o(1)) log(1/δ).

Proof of Lemma 33.7 Notice that |i∗ (ν̂(t))| > 1 implies that Zt = 0. Hence
|i∗ (ν̂(τ ))| = 1 for τ < ∞, and the selection rule is well defined. Abbreviate
µ = µ(ν) and ∆ = ∆(ν), and assume without loss of generality that ∆1 = 0. By
the definition of τ and Zt ,
( k )
1X 2
{ν ∈ Ealt (ν̂(τ ))} ⊆ Ti (τ )(µ̂i (τ ) − µi ) ≥ βτ (δ) .
2 i=1

Using the definition of Ealt (ν̂(τ )) yields

!
1X
k
2
P (1 6∈ i (ν̂(τ ))) = P (ν ∈ Ealt (ν̂(τ ))) ≤ P
∗
Ti (τ )(µ̂i (τ ) − µi ) ≥ βτ (δ) .
2 i=1

Then apply Lemma 33.8 and Proposition 33.9 from Section 33.2.3.
A candidate for βt (δ) can be extracted from the proof and satisfies βt (δ) ≈
2k log t + log(1/δ). This can be improved to approximately k log log(t) + log(1/δ)
 33.2 Best-Arm Identification with a Fixed Confidence 410

by using a law of the iterated logarithm bound instead of Lemma 33.8. Below,
we sketch the proof of Theorem 33.6. A more complete outline is given in
Exercise 33.6.

Proof sketch of Theorem 33.6 Lemma 33.7 shows that (π, τ ) are sound. It
remains to control the expectation of the stopping time. The intuition is
straightforward. As more samples are collected, we expect that α̂(t) ≈ α∗ (ν) and
µ̂ ≈ µ and

Ti (t)(µ̂i (t) − µi (ν̃))2

k
X
Zt = inf
ν̃∈Ealt (ν̂(t))
i=1
2

α∗ (ν)(µi (ν) − µi (ν̃))2

k
X
≈t inf i
ν̃∈Ealt (ν)
i=1
2
t
= .
c∗ (ν)
Provided the approximation is reasonably accurate, the algorithm should halt
once
t
≥ βt (δ) = (1 + o(1)) log(1/δ) ,
c∗ (ν)

which occurs once t ≥ (1 + o(1))c∗ (ν) log(1/δ).

33.2.3 Concentration
The first concentration theorem follows from Corollary 5.5 and a union bound.

Lemma 33.8. Let (Xt )∞

t=1 be a sequence of independent Gaussian random variables
Pn
with mean µ and unit variance. Let µ̂n = n1 t=1 Xt . Then
 n 
P exists n ∈ N+ : (µ̂n − µ)2 ≥ log(1/δ) + log(n(n + 1)) ≤ δ .
2

As we remarked earlier, the log(n(n + 1)) term can be improved to

approximately log log(n). You can do this using peeling (Chapter 9) or the
method of mixtures (Exercise 20.9). Since (Xt ) are Gaussian, you can also use
the tangent approximation and the Bachélier–Levy formula (Exercise 9.4).

Proposition 33.9. Let g : N → R be increasing, and for each i ∈ [k], let

Si1 , Si2 , . . . be an infinite sequence of random variables such that for all δ ∈ (0, 1),

P (exists s ∈ N : Sis ≥ g(s) + log(1/δ)) ≤ δ .

 33.3 Best-Arm Identification with a Budget 411

Then, provided that (Si )ki=1 are independent and x ≥ 0,

! !
Xk Xk  x k
P exists s ∈ N :
k
Sisi ≥ kg si + x ≤ exp(k − x) .
i=1 i=1
k

Proof For i ∈ [k], let Wi = max{w ∈ [0, 1] : Sis < g(s)+log(1/w) for all s ∈ N},
where we define log(1/0) = ∞. Note that Wi are well defined. Then, for any
s ∈ Nk ,
k k k k
! k
X X X X X
Sisi ≤ g(si ) + log(1/Wi ) ≤ kg si + log(1/Wi ) .
i=1 i=1 i=1 i=1 i=1

By assumption, (Wi )ki=1 are independent and satisfy P (Wi ≤ x) ≤ x for all
x ∈ [0, 1]. The proof is completed using the result of Exercise 5.16.

33.3 Best-Arm Identification with a Budget

In the fixed-budget variant of best-arm identification, the learner is given the

horizon n and should choose a policy π = (πt )n+1
t=1 with the objective of minimising
the probability that An+1 is suboptimal. The constraint on the horizon rather
than the confidence level makes this setting a bit more nuanced than the fixed
confidence setting, and the results are not as clean.
A naive option is to use the uniform exploration policy, but as discussed in
Section 33.1, this approach leads to poor results when the suboptimality gaps
are not similar to each other. To overcome this problem, the sequential halving
algorithm divides the budget into L = dlog2 (k)e phases. In the first phase, the
algorithm chooses each arm equally often. The bottom half of the arms are then
eliminated, and the process is repeated.

1: Input n and k
2: Set L = dlog2 (k)e and A1 = [k].
3: for ` = 1, . . .j, L dok
4: Let T` = L|A n
`|
.
5: Choose each arm in A` exactly T` times
6: For each i ∈ A` compute µ̂`i as the empirical mean of arm i based on the
last T` samples
7: Let A`+1 contain the top d|A` |/2e arms in A`
8: end for
9: return An+1 as the arm in AL+1
Algorithm 22: Sequential halving.

Theorem 33.10. If ν ∈ ESG

k
(1) has mean vector µ = µ(ν) and µ1 ≥ · · · ≥ µk and
 33.4 Notes 412

π is sequential halving, then

 
n
Pνπ (∆An+1 > 0) ≤ 3 log2 (k) exp − ,
16H2 (µ) log2 (k)
where H2 (µ) = maxi:∆i >0 i
∆2i
.

The assumption on the ordering of the means is only needed for the clean
definition of H2 , which would otherwise be defined by permuting the arms. The
algorithm is completely symmetric. In Exercise 33.8 we guide you through the
proof of Theorem 33.10.
The quantity H2 (µ) looks a bit unusual, but arises naturally in the
analysis. It is related to a more familiar quantity as follows. Define H1 (µ) =
Pk 2 2
i=1 min{1/∆i , 1/∆min }. Then

H2 (µ) ≤ H1 (µ) ≤ (1 + log(k))H2 (µ) . (33.9)

Furthermore, both inequalities are essentially tight (Exercise 33.7). Let’s see how
the bound in Theorem 33.10 compares to uniform exploration, which is the same
as Algorithm 20. Like in the proof of Theorem 33.1, the probability that uniform
exploration selects a suboptimal arm is easily controlled using Theorem 5.3 and
Lemma 5.4:
 
X X bn/kc ∆2i
Pν,UE (∆An+1 > 0) ≤ P (µ̂i (n) ≥ µ̂1 (n)) ≤ exp − .
4
i:∆i >0 i:∆i >0

Suppose that ∆ = ∆2 = ∆k so that all suboptimal arms have the same

suboptimality gap. Then H2 = k/∆2 and terms in the exponent for sequential
halving and uniform exploration are Θ(n∆2 /(k log k)) and Θ(n∆2 /k), respectively,
which means that uniform exploration is actually moderately better than
sequential halving, at least if n is sufficiently large. On the other hand, if ∆2 = ∆
is small, but ∆i = 1 for all i > 2, then H2 = Θ(1/∆2 ) and the exponents are
Θ(n∆2 ) and Θ(n∆2 /k) respectively and sequential halving is significantly better.
The reason for the disparity is the non-adaptivity of uniform exploration, which
wastes many samples on arms i > 2. Although there are not asymptotically
matching upper and lower bounds in the fixed budget setting, the bound of
sequential halving is known to be roughly optimal.

33.4 Notes

1 The problems studied in this chapter belong to the literature on stochastic

optimisation, where the simple regret is called the expected suboptimality.
There are many variants of pure exploration. In the example at the start of the
chapter, a medical researcher may be interested in getting the most reliable
information about differences between treatments. This falls into the class of
pure information-seeking problems, the subject of optimal experimental design
from statistics, which we have met earlier.
 33.4 Notes 413

2 We mentioned that algorithms with logarithmic cumulative regret are not well
suited for pure exploration. Suppose π has asymptotically optimal cumulative
regret on E = ENk
, which means that limn→∞ Eνπ [Ti (n)]/ log(n) = 2/∆i (ν) for
all ν ∈ E. You will show in Exercise 33.5 that for any ε > 0, there exists a
ν ∈ E with a unique optimal arm such that
− log (Pνπ (An+1 6∈ i∗ (ν)))
lim inf ≤ 1 + ε.
n→∞ log(n)
This shows that using an asymptotically optimal policy for cumulative regret
minimisation leads to a best-arm identification policy for which the probability
of selecting a suboptimal arm decays only polynomially with n. This result
holds no matter how An+1 is selected.
3 A related observation is that the empirical estimates of the means after running
an algorithm designed for minimising the cumulative regret tend to be negatively
biased. This occurs because these algorithms play arms until their empirical
means are sufficiently small.
4 Although there is no exploration/exploitation dilemma in the pure exploration
setting, there is still an ‘exploration dilemma’ in the sense that the optimal
exploration policy depends on an unknown quantity. This means the policy
must balance (to some extent) the number of samples dedicated to learning
how to explore relative to those actually exploring.
5 Best-arm identification is a popular topic that lends itself to simple analysis
and algorithms. The focus on the correct identification of an optimal arm
makes us question the practicality of the setting, however. In reality, any
suboptimal arm is acceptable provided its suboptimality gap is small enough
relative to the budget, which is more faithfully captured by the simple
regret criterion. Of course the
simple regret may be bounded naively by
Rnsimple ≤ maxi ∆i P ∆An+1 > 0 , which is tight in some circumstances and
loose in others.
6 An equivalent form of the bound shown in Theorem 33.5 is
( k k
)
X X
Eνπ [τ ] ≥ min αi : α1 , . . . , αk ≥ 0, inf αi D(νi , νi ) ≥ log(4/δ) .
0
ν∈Ealt (ν)
i=1 i=1

This form follows immediately from Eq. (33.6) by noting that Eνπ [τ ] =
P
i Eνπ [Ti (τ )]. The version given in the theorem is preferred because it is
a closed form expression. Exercise 33.3 asks you to explore the relation between
the two forms.
7 The forced exploration in the track-and-stop algorithm is sufficient for
asymptotic optimality. We are uneasy about the fact that the proof would work√
for any threshold Ctp with p ∈ (0, 1). There is nothing fundamental about t.
We do not currently know of a principled way to tune the amount of forced
exploration or if there is better algorithm design for best-arm identification.
Ideally one should provided finite-time upper bounds that match the finite-time
 33.5 Bibliographical Remarks 414

lower bound provided by Theorem 33.5. The extent to which this is possible
appears to be an open question.
8 The choice of βt (δ) significantly influences the practical performance of track-
and-stop. We believe the analysis given here is mostly tight except that the
naive concentration bound given in Lemma 33.8 can be improved using a
finite-time version of the law of the iterated logarithm (see Exercise 20.9, for
example).
9 Perhaps the most practical set-up in pure exploration has not yet received any
attention, which is upper and lower instance-dependent bounds on the simple
regret. Even better would be to have an understanding of the distribution of
∆An+1 .

33.5 Bibliographical Remarks

In the machine learning literature, pure exploration for bandits seems to have
been first studied by Even-Dar et al. [2002], Mannor and Tsitsiklis [2004] and
Even-Dar et al. [2006] in the ‘Probability Approximately Correct’ setting, where
the objective is to find an ε-optimal arm with high probability with as few samples
as possible. After a dry spell, the field was restarted by Bubeck et al. [2009] and
Audibert and Bubeck [2010b]. The asymptotically optimal algorithm for the fixed
confidence setting of Section 33.2 was introduced by Garivier and Kaufmann
[2016], who also provide results for exponential families as well as in-depth
intuition and historical background. Degenne and Koolen [2019] and Degenne
et al. [2019] have injected some new ideas into the basic principles of track-and-
stop by incorporating a kind of optimism and solving the optimisation problem
incrementally using online learning, which leads to theoretical and practical
improvements. A similar problem is studied in a Bayesian setting by Russo
[2016], who focuses on designing algorithms for which the posterior probability of
choosing a suboptimal arm converges to zero exponentially fast with an optimal
rate. Even more recently, Qin et al. [2017] designed a policy that is optimal in both
the frequentist and Bayesian settings. The stopping rule used by Garivier and
Kaufmann [2016] is inspired by similar rules by Chernoff [1959]. The sequential
halving algorithm is by Karnin et al. [2013], and the best summary of lower
bounds is by Carpentier and Locatelli [2016]. Besides this there have been many
other approaches, with a summary by Jamieson and Nowak [2014]. The negative
result discussed in Note 2 is due to Bubeck et al. [2009]. Pure exploration has
recently become a hot topic and is expanding beyond the finite-armed case. For
example, to linear bandits [Soare et al., 2014] and continuous-armed bandits
[Valko et al., 2013a], tree search [Garivier et al., 2016a, Huang et al., 2017a] and
combinatorial bandits [Chen et al., 2014, Huang et al., 2018].
The continuous-armed case is also known as zeroth-order (or derivative-
free) stochastic optimisation and is studied under various assumptions on
the unknown reward function, usually assuming that A ⊂ Rd . Because of the
 33.5 Bibliographical Remarks 415

obvious connection to optimisation, this literature usually considers losses, or

cost, rather than reward, and the reward function is then called the objective
function. A big part of this literature poses only weak assumptions, such as
smoothness, on the objective function. Note that in the continuous-armed case,
regret minimisation may only be marginally more difficult than minimising the
simple regret because even the instance-dependent simple regret can decay at
a slow, polynomial rate. While the literature is vast, most of it is focused on
heuristic methods without rigorous finite-time analysis. Methods developed for
this case maintain an approximation to the unknown objective and often use
branch-and-bound techniques to focus the search for the optimal value. For a
taster of the algorithmic ideas, see [Conn et al., 2009, Rios and Sahinidis, 2013].
When the search for the optimum is organised cleverly, the methods can adapt to
‘local smoothness’ and enjoy various optimality guarantees [Valko et al., 2013a].
A huge portion of this literature considers the easier problem of finding a local
minimiser, or just a stationary point. Another large portion of this literature
is concerned with the case when the objective function is convex. Chapter 9 of
the classic book by Nemirovsky and Yudin [1983] describes two complementary
approaches (a geometric, and an analytic) and sketches their analysis. For the
class of strongly convex
p and smooth functions, it is known that the minimax
simple regret is Θ( d2 /n) [Shamir, 2013]. The main outstanding challenge is to
understand the dependence of simple regret on the dimension beyond the strongly
convex and smooth case. Hu et al. [2016] prove a lower bound of Ω(n−1/3 ) on
the simple-regret for algorithms that construct gradient estimates by injecting
random noise (as is done by Katkovnik and Kulchitsky [1972], Nemirovsky and
Yudin [1983] and others), which, together with the O(n−1/2 ) upper bound by
Nemirovsky and Yudin [1983] (see also Agarwal et al. 2013, Liang et al. 2014),
establishes the inferiority of this approach in the n  d regime. Interestingly,
empirical evidence favours these gradient-based techniques in comparison to the
‘optimal algorithms’. Thus, much room remains to improve our understanding of
this problem. This setting is to be contrasted to the one when unbiased noisy
estimates of the gradient are available where methods such as mirror descent
(see Chapter 28) give optimal rates. This is a much better understood problem
with matching lower and upper bounds available on the minimax simple regret
for various settings (for example, Chapter 5 of Nemirovsky and Yudin [1983], or
Rakhlin et al. [2012]).
Variants of the pure exploration problem are studied in a branch of statistics
called ranking and selection. The earliest literature on ranking and selection
goes back to at least the 1950s. A relatively recent paper that gives a glimpse
into a small corner of this literature is by Chan and Lai [2006]. The reason we
cite this paper is because it is particularly relevant for this chapter. Using our
terminology, Chan and Lai consider the PAC setting in the parametric setting
when the distributions underlying the arms belong to some known exponential
family of distributions. A procedure that is similar to the track-and-stop procedure
considered here is shown to be both sound and asymptotically optimal as the
 33.6 Exercises 416

confidence parameter approaches one. We also like the short and readable review
of the literature up to the 1980s from the perspective of simulation optimisation
by Goldsman [1983].
A related setting studied mostly in the operations research community is
ordinal optimisation. In its simplest form, ordinal optimisation is concerned
with finding an arm amongst the αk arms with the highest pay-offs. Ho et al. [1992],
who defined this problem in the stochastic simulation optimisation literature,
emphasised that the probability of failing to find one of the ‘good arms’ decays
exponentially with the number of observations n per arm, in contrast to the
slow n−1/2 decay of the error of estimating the value of the best arm, which
this literature calls the problem of cardinal optimisation. Given the results
in this chapter, this should not be too surprising. A nice twist in this literature
is that the error probability does not need to depend on k (see Exercise 33.9).
The price, of course, is that the simple regret is in general uncontrolled. In a
way, ordinal optimisation is a natural generalisation of best-arm identification.
As such, it also leads to algorithmic choices that are not the best fit when the
actual goal is to keep the simple regret small. Based on a Bayesian reasoning, a
heuristic expression for the asymptotically optimal allocation of samples for the
Gaussian best-arm identification problem is given by Chen et al. [2000]. They
call the problem of finding an optimal allocation the ‘optimal computing budget
allocation’ (OCBA) problem. Their work can be viewed as the precursor to the
results in Section 33.2. Glynn and Juneja [2015] gives further pointers to this
literature, while connecting it to the bandit literature.
Best-arm identification has also been considered in the adversarial setting
[Jamieson and Talwalkar, 2016, Li et al., 2018, Abbasi-Yadkori et al., 2018].
Another related setting is called the max-armed bandit problem, where the
objective is to obtain the largest possible single reward over n rounds [Cicirello
and Smith, 2005, Streeter and Smith, 2006a,b, Carpentier and Valko, 2014, Achab
et al., 2017].

33.6 Exercises

33.1 (Simple regret lower bound) Show there exists a universal constant
C > 0 such that for all
p n ≥ k > 1 and all policies π, there exists a ν ∈ EN such
k

that Rn (π, ν) ≥ C k/n.

simple

33.2 (Suboptimality of uniform exploration) Show there exists a universal

constant C > 0 such
p that for all n ≥ k > 1, there exists a ν ∈ EN such that
k

Rn (UE, ν) ≥ C k log(k)/n.
simple
 33.6 Exercises 417

33.3 Let L > 0 and D ⊂ [0, ∞)k \ {0} be non-empty. Show that
  !−1
inf kαk1 : α ∈ [0, ∞) , inf hα, di ≥ L =
k
sup inf hα, di L.
d∈D α∈Pk−1 d∈D

33.4 (Best-arm identification for Gaussian bandits) Let σ12 , . . . , σk2 be

fixed and E = {(N (µi , σi2 ))ki=1 : µ ∈ Rk } be the set of Gaussian bandits with given
variances. Let ν ∈ E be a bandit with µ1 (ν) > µi (ν) for all i > 1. Abbreviate
µ = µ(ν) and ∆ = ∆(ν).

(a) For any α ∈ [0, ∞)k show that

1 α1 αi ∆2i
k
X
inf αi D(νi , ν̃i ) = min .
ν̃∈Ealt (ν)
i=1
2 i>1 α1 σi2 + αi σ12

(b) Show that if k = 2, then c∗ (ν) = 2(σ1 + σ2 )2 /∆22 .

2σ12 2σi2
k
X
(c) Show that c (ν) ≥
∗
+ .
∆min i=2 ∆2i
(d) Show that
v
u
2 2 u 2σ 2 X
2σ 2σ 2σi2
Xk k
1 t 1
c (ν) ≤
∗
+ i
+ 2 . (33.10)
∆min i=2 ∆2i ∆2min i=2 ∆2i

(e) Show that if σi2 /∆2i = σ12 /∆2min for all i, then equality holds in Eq. (33.10).

33.5 (Suboptimality of cumulative regret algorithms for best-arm

identification) Suppose π is an asymptotically optimal bandit policy in E = EN
k

in the sense that

Rn (π, ν) X 2
lim = for all ν ∈ E .
n→∞ log(n) ∆i (ν)
i:∆i (ν)>0

(a) For any ε > 0, prove there exists a ν ∈ E with a unique optimal arm such
that
− log(Pνπ (∆An+1 > 0))
lim inf ≤ 1 + ε.
n→∞ log(n)
(b) Can you prove the same result with lim inf replaced by lim sup?
(c) What happens if the assumption that π is asymptotically optimal is replaced
with the assumption that there exists a universal constant C > 0 such that
X  log(n)

Rn (π, ν) ≤ C ∆i (ν) + .
∆i (ν)
i:∆i (ν)>0

33.6 (Analysis of track-and-stop) In this exercise, you will complete the

proof of Theorem 33.6. Assume that ν has a unique optimal arm. Make E a
 33.6 Exercises 418

metric space via the metric d(ν1 , ν2 ) = kµ(ν1 ) − µ(ν2 )k∞ . Let ε > 0 be a small
constant, and define random times
τν (ε) = 1 + max {t : d(ν̂t , ν) ≥ ε}
τα (ε) = 1 + max {t : kα∗ (ν) − α∗ (ν̂t )k∞ ≥ ε}
τT (ε) = 1 + max {t : kT (t)/t − αi∗ (ν)k∞ ≥ ε} .
Note, these are not stopping times. Do the following:
(a) Show that α∗ (ν) is unique.
(b) Show α∗ is continuous at ν.
(c) Prove that E[τν (ε)] < ∞ for all ε > 0.
(d) Prove that E[τα (ε)] < ∞ for all ε > 0.
(e) Prove that E[τT (ε)] < ∞ for all ε > 0.
(f) Prove that limδ→0 E[τ ]/ log(1/δ) ≤ c∗ (ν).

33.7 (Complexity measure comparison) Prove the following:

Pk
(a) Let L = 1 + i=3 1i and show that H2 (µ) ≤ H1 (µ) ≤ LH2 (µ). Combine this
with the fact that L ≤ 1 + log(k) to prove that Eq. (33.9) holds.
(b) Find µ and µ0 such that H2 (µ) = H1 (µ) and H1 (µ0 ) = LH2 (µ0 ). Conclude
the inequalities in Eq. (33.9) are tight.

33.8 (Analysis of sequential halving) The purpose of this exercise is to

prove Theorem 33.10. Assume without loss of generality that µ = µ(ν) satisfies
µ1 ≥ µ2 ≥ . . . ≥ µk . Given a set A ⊂ [k], let
 
 X 
TopM(A, m) = i ∈ [k] : I {j ∈ A} ≤ m
 
j≤i

be the top m arms in A. To make life easier, you may also assume that k is a
power of two so that |A` | = k21−` and T` = n2`−1 / log2 (k).
(a) Prove that |AL+1 | = 1.
(b) Let i be a suboptimal arm in A` , and suppose that 1 ∈ A` . Show that
 

T` ∆2i
P µ̂1 ≤ µ̂i i ∈ A` , 1 ∈ A` ≤ exp −
` `
.
4
(c) Let A0` = A` \ TopM(A` , d|A` |/4e) be the bottom three-quarters of the arms
in round `. Show that if the optimal arm is eliminated after the `th phase,
then
X  1
N` = I µ̂`i ≥ µ̂`1 ≥ |A0` | .
0
3
i∈A`

(d) Let i` = min A0` and show that

   
∆2i n2`−1 n∆2i`
E[N` | A` ] ≤ |A` | max0 exp −
0
≤ |A` | exp −
0
.
i∈A` 4 log2 (k) 16i` log2 (k)
 33.6 Exercises 419

(e) Combine the previous two parts with Markov’s inequality to show that
 
T ∆2i`
P (1 ∈
/ A`+1 | 1 ∈ A` ) ≤ 3 exp − .
16 log2 (k)i`
(f) Join the dots to prove Theorem 33.10.

33.9 Let P be a distribution over the measurable set X , µ : X → [0, 1] be

measurable, α, δ ∈ (0, 1), and define µ∗α = inf{y : P (µ(X) < y) ≥ 1 − α}. Show
that if n ≥ log(1/δ)/ log(1/(1 − α)), then for X1 , . . . , Xn ∼ P independent, with
probability 1 − δ, it holds that maxi∈[n] µ(Xi ) ≥ µ∗α .

33.10 (Multiple optimal arms and soundness) Throughout this exercise,

let k > 1.
(a) Let E = EN k
(1). Prove that for any sound pair (π, τ ) and ν ∈ E with
|i (ν)| > 1, it holds that Pνπ (τ = ∞) = 1.
∗

(b) Repeat the previous part with E = EBk .

(c) Describe an unstructured class of k-armed stochastic bandits E and ν ∈ E
with |i∗ (ν)| > 1 and sound pair (π, τ ) for which Pνπ (τ = ∞) = 0.

33.11 (Probably approximately correct algorithms) This exercise is

about designing (ε, δ)-PAC algorithms.
(a) For each ε > 0 and δ ∈ (0, 1) and number of arms k > 1, design a policy π
and stopping time τ such that for all ν ∈ E,
 
Ck k
Pνπ (∆Aτ ≥ ε) ≤ δ and Eνπ [τ ] ≤ 2 log ,
ε δ
for universal constant C > 0.
(b) It turns out the logarithmic dependence on k can be eliminated. Design a
policy π and stopping time τ such that for all ν ∈ E,
 
Ck 1
Pνπ (∆Aτ ≥ ε) ≤ δ and Eνπ [τ ] ≤ 2 log .
ε δ
(c) Prove a lower bound showing that the bound in part (b) is tight up to
constant factors in the worst case.

Hint Part (b) of the above exercise is a challenging problem. The simplest
approach is to use an elimination algorithm that operates in phases where at
the end of each phase, the bottom half of the arms (in terms of their empirical
estimates) are eliminated. For details, see the paper by Even-Dar et al. [2002].
34 Foundations of Bayesian Learning

e!TEX root = ../../book.tex

Bayesian methods have been used for bandits from the beginning of the field and
dominated research from 1950 until 1980. This chapter introduces the Bayesian
viewpoint and develops the technical tools necessary for applications in bandits.
Readers who are already familiar with the measure-theoretic Bayesian analysis
can skim Sections 34.4 and 34.6 for the notation used in subsequent chapters.

34.1 Statistical Decision Theory and Bayesian Learning

The fundamental challenge in learning problems is that the true environment

is unknown and policies that are optimal in one environment are usually not
optimal in another. This forces the user to make trade-offs, balancing performance
between environments. We have already discussed this in the context of finite-
armed bandits in Part IV. Here we take a step back and consider a more general
set-up.
Let E be a set of environments and Π a set of policies. These could be bandit
environments/policies, but for now an abstract view is sufficient. A loss function
is a mapping ` : E × Π → R with `(ν, π) representing the loss suffered by policy π
in environment ν. Of course you should choose a policy that makes the loss small,
but most choices are incomparable because the loss depends on the environment.
Fig. 34.1 illustrates a typical situation with four policies. Some policies can
be eliminated from consideration because they are dominated, which means
they suffer at least as much loss as some other policy on all environments and
more loss on at least one. A policy that is not dominated is called admissible
or Pareto optimal . Choosing between admissible policies is non-trivial. One
canonical choice of admissible policy (assuming it exists) is a minimax optimal
policy π ∈ argminπ0 supν `(ν, π 0 ). Minimax optimal policies enjoy robustness, but
the price may be quite large on average. Would you choose the minimax optimal
policy in the example in Fig. 34.1?
In the Bayesian viewpoint, the uncertainty in the environment is captured by
choosing a prior probability measure on E that reflects the user’s belief about the
environment the learner will face. Having committed to a prior, the Bayesian
optimal policy simply minimises the expected loss with respect to the prior.
 34.2 Bayesian Learning and the Posterior Distribution 421

1 π1 , admissible

Loss

π2 , dominated
π3 , minimax optimal

π4 , admissible
0
0 1
Environments

Figure 34.1 Loss as a function of the environment for four different polices π1 , . . . , π4 ,
when E = [0, 1]. Which policy would you choose?

When E is countable, a measure corresponds to a probability vector q ∈ P(E),

and the Bayesian optimal policy with respect to q is an element of
X
argmaxπ q(ν)`(ν, π) .
ν∈E

The Bayesian viewpoint is hard to criticise when the user really does know the
underlying likelihood of each environment and the user is risk-neutral. Even
when the distribution is not known exactly, however, sensible priors often yield
provably sensible outcomes, regardless of whether one is interested in the average
loss across the environments, or the worst-case loss, or some other metric.

A distinction is often made between the Bayesian and frequentist viewpoints,

which naturally leads to heated discussions on the merits of one viewpoint
relative to another. This debate does not interest us greatly. We prefer to
think about the pros and cons of problem definitions and solution methods,
regardless of the label on them. Bayesian approaches to bandits have their
strengths and weaknesses, and we hope to do them a modicum of justice
here.

34.2 Bayesian Learning and the Posterior Distribution

The last section explained the ‘forward view’, where a policy is chosen in advance
that minimises the expected loss. The Bayesian can also act sequentially by
updating their beliefs (the prior) as data is observed to obtain a new distribution
 34.2 Bayesian Learning and the Posterior Distribution 422

on the set of environments (more generally, the set of hypotheses). The new
distribution is called the posterior. This is simple and well defined when the
environment set is countable, but quickly gets technical for larger spaces. We
start gently with a finite case and then explain the measure-theoretic machinery
needed to rigourously treat the general case.
Suppose you are given a bag containing two marbles. A trustworthy source
tells you the bag contains either (a) two white marbles (ww) or (b) a white
marble and a black marble (wb). You are allowed to choose a marble from the
bag (without looking) and observe its colour, which we abbreviate by ‘observe
white’ (ow) or ‘observe black’ (ob). The question is how to update your ‘beliefs’
about the contents of the bag having observed one of the marbles. The Bayesian
way to tackle this problem starts by choosing a probability distribution on the
space of hypotheses, which, incidentally, is also called the prior. This distribution
usually reflects one’s beliefs about which hypotheses are more probable. In the
lack of extra knowledge, for the sake of symmetry, it seems reasonable to choose
P(ww) = 1/2 and P(wb) = 1/2. The next step is to think about the likelihood
of the possible outcomes under each hypothesis. Assuming that the marble is
selected blindly (without peeking into the bag) and the marbles in the bag are
well shuffled, these are
P(ow | ww) = 1 and P(ow | wb) = 1/2 .
The conditioning here indicates that we are including the hypotheses as part of
the probability space, which is a distinguishing feature of the Bayesian approach.
With this formulation we can apply Bayes’ law (Eq. (2.2)) to show that
P(ow | ww)P(ww) P(ow | ww)P(ww)
P(ww | ow) = =
P(ow) P(ow | ww)P(ww) + P(ow | wb)P(wb)
1 × 21 2
= = .
1 × 2 + 12 × 12
1 3
Of course P(wb | ow) = 1 − P(ww | ow) = 1/3. Thus, while in the lack of
observations, ‘a priori’, both hypotheses are equally likely, having observed a
white marble, the probability that the bag originally contained two white marbles
(and thus the bag has a white marble remaining in it) jumps to 2/3. An alternative
calculation shows that P(ww | ob) = 0, which makes sense because choosing a
black marble rules out the hypothesis that the bag contains two white marbles.
The conditional distribution P( · | ow) over the hypotheses is called the posterior
distribution and represents the Bayesian’s belief in each hypothesis after observing
a white marble.

34.2.1 A Rigorous Treatment of Posterior Distributions

A more sophisticated approach is necessary when the hypothesis and/or outcome
spaces are not discrete. In introductory texts, the underlying details are often
(quite reasonably) swept under the rug for the sake of clarity. Besides the desire
 34.2 Bayesian Learning and the Posterior Distribution 423

for generality, there are two reasons not to do this. First, having spent the effort
developing the necessary tools in Chapter 2, it would seem a waste not to use them
now. And second, the subtle issues that arise highlight some real consequences of
the differences between the Bayesian and frequentist viewpoints. As we shall see,
there is a real gap between these viewpoints.
Let Θ be a set called the hypothesis space and G be a σ-algebra on Θ. While
Θ is often a subset of a Euclidean space, we do not make this assumption. A prior
is a probability measure Q on (Θ, G). Next, let (U, H) be a measurable space and
P = (Pθ : θ ∈ Θ) be a probability kernel from (Θ, G) to (U, H). We call P the
model. Let Ω = Θ × U and F = G ⊗ H. The prior and the model combine to yield
a probability P = Q ⊗ P on (Ω, F). The prior is now the marginal distribution
of the joint probability measure: Q(A) = P(A × U). Suppose a random element
X on Ω describes what is observed. Then, generalizing the previous example
with the marbles, the posterior should somehow be the marginal of the joint
probability measure conditioned X. To make this more precise, let (X , J ) be a
measurable space and X : Ω → X a F/J -measurable map. The posterior having
observed that X = x should be a measure Q( · | x) on (Θ, G).

We abuse notation by letting θ : Ω → Θ denote the F/G-measurable random

element given by the projection: θ((φ, u)) = φ. This allows θ being used as
part of the probability expressions below.

Without much thought, we might try and apply Bayes’ law (Eq. (2.2)) to claim
that the posterior distribution having observed X(ω) = x should be a measure
on (Θ, G) given by
P (X = x | θ ∈ A) P (θ ∈ A)
Q(A | x) = P (θ ∈ A | X = x) = . (34.1)
P (X = x)
The problem with the ‘definition’ in (34.1) is that P (X = x) can have measure
zero, and then P (θ ∈ A | X = x) is not defined. This is not an esoteric problem.
Consider the problem when θ is randomly chosen from Θ = R and its distribution
is Q = N (0, 1), the parameter θ is observed in Gaussian noise with a variance of
one: U = R, Pθ = N (θ, 1) for all θ ∈ R and X(φ, u) = u for all (φ, u) ∈ Θ×U. Even
in this very simple example, we have P (X = x) = 0 for all x ∈ R. Having read
Chapter 2, the next attempt might be to define Q(A | X) as a σ(X)-measurable
random variable defined using conditional expectations: for A ∈ G,

Q(A | x) = E[I {θ ∈ A} | X](x) ,

where we remind the reader that E[I {θ ∈ A} | X] is a σ(X)-measurable random

variable that is uniquely defined except for a set of measure zero and also that
the notation on the right-hand side is explained in Fig. 2.4 in Chapter 2. For
most applications of probability theory, the choice of conditional expectation
does not matter. However, as we shortly illustrate with an example, this is not
 34.2 Bayesian Learning and the Posterior Distribution 424

true here. A related annoying issue is that Q( · | x) as defined above need not be
a measure. By assuming that (Θ, G) is a Borel space, this issue can be overcome
by using a regular version (Theorem 3.11), a result that we restate here using
the present notation.

Theorem 34.1. If (Θ, G) is a Borel space, then there exists a probability kernel
Q : X ×G → [0, 1] such that Q(A | X) = P (θ ∈ A | X) simultaneously for all A ∈ G
outside of some P-null set. Furthermore, for any two probability kernels Q, Q0
satisfying this condition, Q(· | x) = Q0 (· | x) for all x in some set of PX -probability
one.

The Posterior Density

Theorem 34.1 provides weak conditions under which a posterior exists but does
not suggest a useful way of finding it. In many practical situations, the posterior
can be calculated using densities. Given θ ∈ Θ let pθ be the Radon–Nikodym
derivative of Pθ with respect to some measure µ and let q(θ) be the Radon–
Nikodym derivative of Q with respect to another measure ν. Provided all terms
are appropriately measurable and non-zero, then
pθ (x)q(θ)
q(θ | x) = R (34.2)
Θ
p θ (x)q(θ)dν(θ)

is the Radon–Nikodym derivative of Q( · | x) with respect to ν, also known

as the posterior
R density of Q. In other words, for any A ∈ G, it holds that
Q(A | x) = A q(θ | x)dν(θ). This corresponds to the usual manipulation of
densities when µ and ν are the Lebesgue measures.
The reader may wonder about why all the fuss about the existence of Q( · | x)
in the previous section if we can get its density with a simple formula like (34.2).
In other words, why not flip around things and define Q( · | x) via (34.2)? The
crux of the problem is that oftentimes it is hard to come up with an appropriate
dominating measure µ, and in general the denominator in the right-hand side of
(34.2) could be zero from some particular value of x. But when we can identify
an appropriate measure µ and the denominators are non-zero, the above formula
can indeed be used as the definition of Q( · | x) (Exercise 34.4).

The Non-uniqueness Issue Frequentists Face

A minor annoyance when using Bayesian methods as part of a frequentist argument
is that the posterior need not be unique.

Example 34.2. Consider the situation when the hypothesis set is the [0, 1]
interval, the prior is the uniform distribution, and the observation is equal to the
hypothesis sampled. Formally, Θ = [0, 1] and the prior Q is the uniform measure
on (Θ, B(Θ)), Pθ = δθ is the Dirac measure on [0, 1] at θ, and X : [0, 1] → [0, 1] is
the identity: X(x) = x for all x ∈ [0, 1]. Let C ⊂ [0, 1] be an arbitrary countable
set and µ be an arbitrary probability measure on ([0, 1], B(R)). It is not hard to
 34.3 Conjugate Pairs, Conjugate Priors and the Exponential Family 425

see that the probability kernel

(
δx (A) , if x ∈
/ C;
Q(A | x) =
µ(A) , if x ∈ C

satisfies the conditions of Theorem 34.1 and is thus one of the many versions of
the posterior, regardless of the choice of C and µ!
A true Bayesian is unconcerned. If θ is sampled from the prior Q, then the
event {X ∈ C} has measure zero, and there is little cause to worry about events
that happen with probability zero. But for a frequentist using Bayesian techniques
for inference, this actually matters. If θ is not sampled from Q, then nothing
prevents the situation that θ ∈ C and the non-uniqueness of the posterior is an
issue (Exercise 34.12). Probability theory does not provide a way around this
issue.

It follows that one must be careful to specify the version of the posterior
being used when using Bayesian techniques for inference in a frequentist
setting because in the frequentist viewpoint, θ is not part of the probability
space and results are proven for Pθ for arbitrary fixed θ ∈ Θ. By contrast,
the all-in Bayesians include θ in the probability space and thus will not worry
about events with negligible prior probability, and for them any version of
the posterior will do.

Although it is important to be aware of the non-uniqueness of the posterior,

practically speaking it is hard to go wrong. In typical applications, there is a
‘canonical’ choice. For example, in the Gaussian prior and model case studied
below, it feels right to choose the posterior to be Gaussian. More generally,
preferring posteriors with continuous densities with respect to the Lebesgue
measure is generally a parsimonious choice.

34.3 Conjugate Pairs, Conjugate Priors and the Exponential Family

One of the strengths of the Bayesian approach is the ability to incorporate

explicitly specified prior beliefs. This is philosophically attractive and can be
enormously beneficial when the user has well-grounded prior knowledge about
the problem. When it comes to Bayesian algorithms, however, this advantage
is belied a little by the competing necessity of choosing a prior for which the
posterior can be efficiently computed or sampled from. The ease of computing
(or sampling from) the posterior depends on the interplay between the prior and
the model. Given the importance of computation, it is hardly surprising that
researchers have worked hard to find models and priors that behave well together.
A prior and model are called a conjugate pair if the posterior has the same
34.3 Conjugate Pairs, Conjugate Priors and the Exponential Family 426

parametric form as the prior. In this case, the prior is called a conjugate prior
to the model.

Gaussian Model/Gaussian Prior

Suppose that (Θ, G) = (Ω, F) = (R, B(R)) and X : Ω → Ω is the identity and
Pθ is Gaussian with mean θ and known signal variance σS2 . If the prior Q is
Gaussian with mean µP and prior variance σP2 , then the posterior distribution
having observed X = x can be chosen to be
 −1 !
µP /σP2 + x/σS2 1 1
Q( · | x) = N , + 2 .
1/σP2 + 1/σS2 σS2 σP

The proof is left to the reader in Exercise 34.1.

Following convention, from now on we sweep under the rug that this posterior
is one of many choices, which is justified because all posteriors must agree
almost everywhere.

The limiting regimes as the prior/signal variance tend to zero or infinity are
quite illuminating. For example, as σP2 → 0 the posterior tends to a Gaussian
N (µP , σP2 ), which is equal to the prior and indicates that no learning occurs.
This is consistent with intuition. If the prior variance is zero, then the statistician
is already certain of the mean, and no amount of data can change their belief.
On the other hand, as σP2 tends to infinity, we see the mean of the posterior
has no dependence on the prior mean, which means that all prior knowledge is
washed away with just one sample. You should think about what happens when
σS2 → {0, ∞}.
Notice how the model has fixed σS2 , suggesting that the model variance is
known. The Bayesian can also incorporate their uncertainty over the variance. In
this case, the model parameters are Θ = R × [0, ∞) and Pθ = N (θ1 , θ2 ). But is
there a conjugate prior in this case? Already things are getting complicated, so we
will simply let you know that the family of Gaussian-inverse-gamma distributions
is conjugate.

Bernoulli Model/Beta Prior

Suppose that Θ = [0, 1] and Pθ = B(θ) is Bernoulli with parameter θ. In this
case, it turns out that the family of beta distributions is conjugate, which for
parameters θ = (α, β) ∈ (0, ∞)2 is given in terms of its probability density
function with respect to the Lebesgue measure:
Γ(α + β)
pα,β (x) = xα−1 (1 − x)β−1 , (34.3)
Γ(α)Γ(β)
where Γ(x) is the Gamma function. Then the posterior having observed X = x ∈
{0, 1} is also a beta distribution with parameters (α + x, β + 1 − x). Unlike in
 34.3 Conjugate Pairs, Conjugate Priors and the Exponential Family 427

the Gaussian case, the posterior for the Bernoulli model and beta prior is unique
(Exercise 34.2).

34.3.1 Exponential Families

Both the Gaussian and Bernoulli families are examples of a more general family.
Let h be a measure on (R, B(R)) and S, η : R → R be two ‘suitable’ functions,
where S is called the sufficient statistic. Together, h, η and S define a measure
Pθ on (R, B(R)) for each θ ∈ Θ ⊆ R in terms of its density with respect to h:
dPθ
(x) = exp (η(θ)S(x) − A(θ)) ,
dh
R
where A(θ) = log R exp(η(θ)S(x))dh(x) is the log-partition function and
Θ = dom(A) = {θ : A(θ) < ∞} is the domain of A. Integrating the density
shows that for any B ∈ B(R) and θ ∈ Θ,
Z Z
dPθ
Pθ (B) = (x) dh(x) = exp (η(θ)S(x) − A(θ)) dh(x) .
B dh B

The collection (Pθ : θ ∈ Θ) is called a single-parameter exponential family.

An exponential family is regular if Θ is non-empty and open. It is non-singular
if A00 (θ) > 0 for all θ ∈ Θ.

Example 34.3. Let σ 2 > 0 and h = N (0, σ 2 ) and η(θ) = σθ and S(x) = σx . An
easy calculation shows that A(θ) = θ2 /(2σ 2 ), which has domain Θ = R and
Pθ = N (θ, σ 2 ).

Example 34.4. Let h = δ0 + δ1 be the sum of Dirac measures and S(x) = x

and η(θ) = θ. Then A(θ) = log(1 + exp(θ)) and Θ = R and Pθ = B(σ(θ)), where
σ(θ) = exp(θ)/(1 + exp(θ)) is the logistic function.

Example 34.5. The same family can be parameterised in many different ways.
Let h = δ0 + δ1 , S(x) = x and η(θ) = log(θ/(1 − θ)). Then A(θ) = − log(1 − θ)
and Θ = (0, 1) and Pθ = B(θ).

Exponential families have many nice properties, some of which you will prove
in Exercise 34.5. Of most interest to us here is the existence of conjugate priors.
Suppose that (Pθ : θ ∈ Θ) is a single-parameter exponential family determined by
h, η and S, where S(x) = x is the identity map. Let x0 , n0 ∈ R, and define prior
measure Q on (Θ, B(Θ)) in terms of its density q = dQ/dλ with λ the Lebesgue
measure:
exp (n0 x0 η(θ) − n0 A(θ))
q(θ) = R , (34.4)
Θ
exp (n0 x0 η(θ) − n0 A(θ)) dθ

where we assume that the integral in the denominator exists and is positive.
Suppose we observe X = x. Then a choice of posterior has density with respect
 34.3 Conjugate Pairs, Conjugate Priors and the Exponential Family 428

to the Lebesgue measure given by

exp (η(θ)(x + n0 x0 ) − (1 + n0 )A(θ))
q(θ | x) = R .
Θ
exp (η(θ)(x + n0 x0 ) − (1 + n0 )A(θ)) dλ(θ)
What this means is that after observing the value x, the posterior takes the form
of the prior except that the parameters (x0 , n0 ) associated with the prior get
updated to ((n0 x0 + x)/(n0 + 1), n0 + 1). The posterior is both easy to represent
and maintain. To see how exponential families recover previous examples, consider
the Bernoulli case of Example 34.5. Since
  n0 x 0
θ
exp(n0 x0 η(θ) − n0 A(θ)) = (1 − θ)n0 = θn0 x0 (1 − θ)n0 (1−x0 ) ,
1−θ
we see that the prior from (34.4) is a beta distribution with parameters
α = 1 + n0 x0 and β = 1 + n0 (1 − x0 ), as can be seen from (34.3). As expected,
the posterior update also works as described earlier.

There are important parametric families with conjugate priors that are not
exponential families. One example is the uniform family (U(a, b) : a < b),
which is conjugate to the Pareto family.

34.3.2 Sequences of Random Variables and the Markov Chain View

Let P n = (Pθn : θ ∈ Θ) be a probability kernel from (Θ, G) to (X n , H⊗n ) and
Q a prior on (Θ, G). Then, let θ and X1 , . . . , Xn be random elements on some
probability space (Ω, F, P), where θ ∈ Θ and Xt ∈ X such that
(a) the law of θ is Pθ = Q; and
(b) P(X1 , . . . , Xn ∈ B | θ) = Pθ (B) almost surely for all B ∈ H⊗n .
By definition, the posterior after observing (Xs )ts=1 is a probability kernel Qt
from (X t , H⊗t ) to (Θ, G) such that for any B ∈ G,
E[ I{θ ∈ B} | X1 , . . . , Xt ] = Qt ( B | X1 , . . . , Xt ) almost surely .
Then, by the tower rule, the conditional distribution of Xt+1 given (Xs )ts=1
almost surely satisfies
Z
P (Xt+1 ∈ B | X1 , . . . , Xt ) = Pθ (Xt+1 ∈ B | X1 , . . . , Xt )Qt (dθ | X1 , . . . , Xt ) .
Θ
(34.5)
This identity says that the conditional distribution of Xt+1 can be written in terms
of the model and posterior. In the fundamental setting where Pθn = Pθ ⊗ · · · ⊗ Pθ
is a product probability measure, then Eq. (34.5) reduces to
Z
P (Xt+1 ∈ B | X1 , . . . , Xt ) = Pθ (B)Qt (dθ | X1 , . . . , Xt ) ,
Θ
 34.4 The Bayesian Bandit Environment 429

which shows that in this case the posterior summarises all the useful information
in (Xs )ts=1 for predicting future data. By introducing a little measure-theoretic
machinery and making suitable regularity assumptions, it is possible to show that
the sequence Q1 , . . . , Qn is a time-inhomogeneous Markov chain. In many cases,
the posterior has a simple form, as you can see in the next two examples.
Example 34.6. Suppose Θ = [0, 1] and G = B([0, 1]) and Q = Beta(α, β)
and Pθ = B(θ) is Bernoulli. Then the posterior after t observations is Qt =
Pt
Beta(α+St , β+t−St ), where St = s=1 Xs . Furthermore, E[Xt+1 | X1 , . . . , Xt ] =
EQt [Xt+1 ] = (α + St )/(α + β + t), and hence
α + St
P (St+1 = St + 1 | St ) = ,
α+β+t
β + t − St
P (St+1 = St | St ) = .
α+β+t
So the posterior after t observations is a Beta distribution depending on St and
S1 , S2 , . . . , Sn follows a Markov chain evolving according to the above display.
Example 34.7. Let (Θ, G) = (R, B(R)) and Q = N (µ, σ 2 ) and Pθ = N (θ, 1).
Then, using the same notation as above the posterior is almost surely Qt =
N (µt , σt2 ), where
 −1
µ/σ 2 + St 2 1
µt = and σt = +t .
1/σ 2 + t σ2
Then S1 , S2 , . . . , Sn is a Markov chain with the conditional distribution of St+1
given St a Gaussian with mean St + µt and variance 1 + σt2 .

34.4 The Bayesian Bandit Environment

The Bayesian bandit model is the same as the frequentist version introduced
in Chapter 4, except that at the beginning of the game, an environment is
sampled from the prior. Of course, the chosen environment is not revealed to
the learner, but its presence forces us to change our conditions on the rewards
because the rewards are dependent on each other through the chosen environment.
For simplicity, we treat only the finite, k-armed case, but the more general set-up
is handled in the same was as in Chapter 4.
A k-armed Bayesian bandit environment is a tuple (E, G, Q, P ), where
(E, G) is a measurable space and Q is a probability measure on (E, G) called the
prior. The last element P = (Pνi : ν ∈ E, i ∈ [k]) is a probability kernel from
E × [k] to (R, B(R)), where Pνi is the reward distribution associated with the ith
arm in bandit ν. A Bayesian bandit environment and policy π = (πt )nt=1 interact
to produce a collection of random variables, ν ∈ E, (At )nt=1 and (Xt )nt=1 with
At ∈ [k] and Xt ∈ R that satisfy
(a) P (ν ∈ ·) = Q(·);
 34.5 Posterior Distributions in Bandits 430

(b) the conditional distribution of action At given ν, A1 , X1 , . . . , At−1 , Xt−1 is

πt (· | A1 , X1 , . . . , At−1 , Xt−1 ) almost surely; and
(c) the conditional distribution of the reward Xt given ν, A1 , X1 , . . . , At is PνAt
almost surely.

The existence of a probability space carrying random elements satisfying

these conditions is guaranteed by the Ionescu–Tulcea theorem (Theorem 3.3,
Exercise 34.9). The corresponding probability measure will be denoted by PQP π .

Most of the structure of a Bayesian bandit environment is in P , which

determines the reward distribution for each arm i in bandits ν ∈ E.

Example 34.8. A k-armed Bayesian Bernoulli bandit environment could be

defined by letting E = [0, 1]k , G = B(E) and Pνi = B(νi ). A natural prior in this
case would be a product of Beta(α, β) distributions:
Z Y
k
Q(A) = qi (xi )dx ,
A i=1

where qi (x) = xα−1 (1 − x)β−1 Γ(α + β)/(Γ(α)Γ(β)).

34.5 Posterior Distributions in Bandits

Let (E, G, Q, P ) be a k-armed Bayesian bandit environment. Assuming that (E, G)

is a Borel space, Theorem 34.1 guarantees the existence of the posterior: a
probability kernel Q( · | ·) from the space of histories to (E, G) so that

Q(A | a1 , x1 , . . . , at , xt )

is a regular version of E[IA (ν) | A1 , X1 , . . . , At , Xt ]. For explicit calculations, it is

worth adding some some extra structure: assume there exists a σ-finite measure λ
on (R, B(R)) such that Pνi  λ for all i ∈ [k] and ν ∈ E. Recall from Chapter 15
that the Radon–Nikodym derivative of Pνπ with respect to (ρ × λ)n is
n
Y
pνπ (a1 , x1 , . . . , an , xn ) = πt (at | a1 , x1 , . . . , at−1 xt−1 )pνat (xt ) , (34.6)
t=1

where pνa is the density of Pνa with respect to λ. Then the posterior after t
rounds is given by
R
pνπ (a1 , x1 , . . . , at , xt )dQ(ν)
Q(B | a1 , x1 , . . . , at , xt ) = RB
p (a , x1 , . . . , at , xt )dQ(ν)
E νπ 1
R Qt
pνas (xs )dQ(ν)
= RB Qts=1 , (34.7)
E s=1 pνas (xs )dQ(ν)
 34.6 Bayesian Regret 431

where the second equality follows from Eq. (34.6). The posterior is not
defined when the denominator is zero, which only occurs with probability zero
(Exercise 34.11). Note that the Radon–Nikodym derivatives pνa (x) are only
unique up to sets of Pνa -measure zero, and so the ‘choice’ of posterior has been
converted to a choice of the Radon–Nikodym derivatives, which, in all practical
situations is straightforward. Observe also that Eq. (34.7) is only well defined
if pνas (·) is G-measurable as a function of ν. Fortunately this is always possible
(see Note 8).

Example 34.9. The posterior for the Bayesian bandit in Example 34.8 in terms
of its density with respect to the Lebesgue measure is
k
Y α+si (ht )−1
q(θ | a1 , x1 , . . . , at , xt ) ∝ θi (1 − θi )β+ti (ht )−si (ht )−1 ,
ht i=1

Pt Pt
where si (ht ) = u=1 xu I {au = i} and ti (ht ) = u=1 I {au = i}. This means the
posterior is also the product of Beta distributions, each updated according to the
observations from the relevant arm.

34.6 Bayesian Regret

Recall that the regret of policy π in k-armed bandit environment ν over n rounds
is
" n #
X
Rn (π, ν) = nµ∗ − E Xt , (34.8)
t=1

where µ = maxi∈[k] µi and µi is the mean of Pνi . Given a k-armed Bayesian

∗

bandit environment (E, G, Q, P ) and a policy π, the Bayesian regret is

Z
BRn (π, Q) = Rn (π, ν)dQ(ν) .
E

The dependence on E, G and P is omitted on the grounds that these are

always self-evident from the context. The Bayesian optimal regret is BR∗n (Q) =
inf π BRn (π, Q), and the optimal (regret-minimizing) policy is

π ∗ = argminπ BRn (π, Q) . (34.9)

Note that the regret-minimising policy is the same as the reward-maximising

Pn
policy π ∗ = argmaxπ EPQP π [ t=1 Xt ], which is known as the Bayesian optimal
policy under prior Q. In all generality, there is no guarantee that the (Bayes)
optimal policy exists, but the non-negativity of the Bayesian regret ensures that
for any ε > 0, there exists a policy π with BRn (π, Q) ≤ BR∗n (Q) + ε.
 34.7 Notes 432

The fact that the expected regret Rn (π, ν) is non-negative for all ν and π
means that the Bayesian regret is always non-negative. Perhaps less obviously,
the Bayesian regret of the Bayesian optimal policy can be strictly greater
than zero (Exercise 34.8).

34.7 Notes

1 In Chapter 4, we defined the environment class E as a set of tuples of probability

distributions over the reals. In a Bayesian bandit environment (E, G, Q, P ) the
set E is arbitrary and the reward distributions are given by the probability
kernel P . The probability kernel and the change of notation is needed because
we are now integrating the regret over E, which may not be measurable without
additional conditions.
2 The Bayesian regret of an algorithm is less informative than the frequentist
regret. By this we mean that a bound on BRn (π, Q) does not generally imply
a meaningful bound on Rn (π, ν), while if Rn (π, ν) ≤ f (ν) for a measurable
function f , then BRn (π, Q) ≤ E[f (ν)]. This is not an argument against using
a Bayesian algorithm but rather an argument for the need to analyse the
frequentist regret of Bayesian algorithms.
3 The relationship between admissibility, Bayesian optimality and minimax
optimality is one of the main topics of statistical decision theory, which intersects
heavily with game theory. In many classical statistical settings, all Bayesian
optimal policies are admissible (Exercise 34.14), and all admissible policies
are either Bayesian optimal for some prior or the limit point of a sequence of
Bayesian optimal policies (Exercise 34.13). Be warned, however, that there
are counterexamples. A nice book with many examples is by Berger [1985].
In Exercise 34.15, you will prove that all admissible policies for stochastic
Bernoulli bandits are Bayesian optimal for some prior.
4 While admissibility and related notions of optimality are helpful in being clear
about the goals of algorithm design, we must recognise that these concepts are
too binary for most purposes. One problem with the classic decision theory
literature is that it puts too much emphasis on these narrow concepts. Who
would argue that a policy that is dominated, but just barely, is worth nothing?
Especially since the optimal policy is often intractable. Meaningful ways of
defining slightly worse usually consider a bigger picture when a policy design
approach (policy schema) is evaluated across many problem classes. In the
Bayesian setting, one may for example consider all k-armed stochastic Bayesian
bandits with (say) bounded rewards and consider policy schema that work no
matter what the environment is. An example of a policy schema is Thompson
sampling, since it can be instantiated for any of the environments. One may
ask whether such a policy schema is near Bayesian (or minimax) optimal across
 34.7 Notes 433

all of the considered environments. In fact, most bandit algorithm design is

better viewed as designing policy schema.
5 Many algorithms/statistical methods have Bayesian interpretations. One
example is ridge regression, which we saw in Chapter 20. Using the notation
of that chapter, the estimator given in Eq. (20.1) is the mean of the Bayesian
posterior when the model is Gaussian with known variance and the prior on the
unknown parameter is a Gaussian with zero mean and covariance I/λ. Another
example is the exponential weighting algorithm for prediction with expert
advice. Consider a sequence y1 , . . . , yn ∈ [d] and suppose there are set M of k
experts making predictions about yt . We write µ(· | y1 , . . . , yt−1 ) ∈ Pd−1 for the
distribution of yt predicted by expert µ ∈ M. In each round the learner observes
the predictions µ(· | y1 , . . . , yt−1 ) for all experts µ ∈ M and should make
a prediction ξ(· | y1 , . . . , yt−1 ) ∈ Pd−1 . Notice that defining µ(y1 , . . . , yn ) =
Qn
t=1 µ(yt | y1 , . . . , yt−1 ) makes µ(·) into a probability distribution on [d] . The
n

regret compares the learner’s performance relative to the best expert in M

under the logarithmic loss:
   
µ(yt | y1 , . . . , yt−1 ) µ(y1 , . . . , yn )
X n
Rn = max log = max log .
µ∈M
t=1
ξ(yt | y1 , . . . , yt−1 ) µ∈M ξ(y1 , . . . , yn )
A Bayesian approach to this problem is to assume that y1 , . . . , yn is sampled
from some unknown µ ∈ M and choose a prior distribution Q ∈ P(M) over
the experts. Then predict to minimise the Bayesian expected loss, which you
will show in Exercise 34.6 leads to
predictive dist. of µ
X µ(y1 , . . . , yt−1 )Q(µ)
ξ( · | y1 , . . . , yt−1 ) = µ( · | y1 , . . . , yt−1 ) P .
µ∈M ν∈M ν(y1 , . . . , yt−1 )Q(ν)
posterior Qt−1 (µ)
(34.10)
You will also show that when Q is taken to be the uniform distribution, the
regret is bounded by Rn ≤ log(k) for all sequences y1 , . . . , yn . Simple algebraic
manipulations show that the posterior is
 P  
exp − s=1 log µ(ys | y11,...,ys−1 )
t

Qt (µ) = P  P   ,
1
exp log
t
ν∈M − s=1 ν(ys | y1 ,...,ys−1 )

which is precisely the exponential weights distribution with learning rate

η = 1. The analogy should not be taken too seriously, however. That this
algorithm controls the regret for all sequences y1 , . . . , yn does not hold for more
general loss functions. For this, the learning rate must be chosen much more
conservatively. For more on the online learning approach to learning under the
logarithmic loss, see chapter 9 of the book by Cesa-Bianchi and Lugosi [2006].
The Bayesian approach is covered in the book by Hutter [2004].
6 Sion’s minimax theorem provides a connection between minimax optimal regret
and the maximum Bayesian optimal regret over all priors. Let Π be the space
 34.7 Notes 434

of all policies and P be a convex space of probability measures over policies and
Q be a convex space of probability measures on (E, G). Define L : P × Q → R
by
Z Z
L(S, Q) = Rn (π, ν)Q(dν)S(dπ) ,
Π E

which is linear in both arguments because integrals are linear as a function of

the measure. Suppose that L is continuous in both arguments and at least one
of P and Q is compact. Then, by Sion’s minimax theorem (Theorem 28.12),
sup inf L(S, Q) = inf sup L(S, Q) . (34.11)
Q∈Q S∈P S∈P Q∈Q

Usually P and Q include all Dirac measures:

{δπ : π ∈ ΠD } ⊆ P and {δν : ν ∈ E} ⊆ Q ,
where ΠD is the space of deterministic policies. Then the left-hand side of
Eq. (34.11) is supQ∈Q BR∗n (Q), and the right-hand side is the minimax regret
Rn∗ (E). Choosing P, Q and a measurable structure on Π is not always easy.
Examples may be found in Exercises 34.16 and 36.11.
7 The issue of conditioning on measure zero sets has been described in many
places. We do not know of a practical situation where things go awry. Sensible
choices yield sensible posteriors. The curious reader could probably burn a few
weeks reading through the literature on the Borel–Kolmogorov paradox
[Jaynes, 2003, §15.7].
8 Suppose that (Pθ : θ ∈ Θ) is a probability kernel from (Θ, G) to (R, B(R)) for
which there exists measure λ on (R, B(R)) such that Pθ  λ for all θ ∈ Θ.
Then there exists a family of densities pθ : R → [0, ∞) such that pθ (x) is jointly
measurable as a (θ, x) 7→ pθ (x) map and pθ = dPθ /dλ for all θ ∈ Θ. See the
proof of Lemma 1.2 in Ghosal and van der Vaart [2017] or sections 1.3 and 1.4
of the book by Strasser [2011].
9 The notion of a sufficient statistic is more general than its role in exponential
families. Let X and Y be random elements on the same measurable space
taking values in X and Y respectively. The random element Y is a sufficient
statistic for X given a family of distributions (Pθ )θ∈Θ over the probability space
carrying both X and Y if (i) Y is σ(X)-measurable and (ii) for all θ ∈ Θ, the
conditional distribution Pθ (X ∈ · |Y ) is independent of the value of θ. Formally,
(ii) means there exists a probability kernel P from Y to X such that for any
θ ∈ Θ, Pθ (X ∈ · |Y ) = P (Y, · ) holds Pθ -almost surely. Informally, this means,
that given Y , there is no information left about θ in X. Denoting by PX,θ the
distribution of X under Pθ and without loss of generality letting Y = y(X)
for some y : X → Y measurable map (recall Lemma 2.5), and assuming that
X, Y take values in Borel spaces, with the help of the disintegration theorem
(Theorem 3.12), it is not hard to see that if (PX,θ )θ have a common dominating
σ-finite measure µ and for any θ ∈ Θ, dµ (x) = h(x)gθ (y(x)) holds µ-almost
dPX,θ

surely for all x ∈ X for some some h : X → [0, ∞) and gθ : Y → [0, ∞) Borel
 34.8 Bibliographic Remarks 435

measurable maps, then Y is a sufficient statistic for X. The Fisher–Neyman

factorisation theorem states that the converse also holds. With some creative
matching of concepts, we can see that in single-parameter exponential families,
what we called a sufficient statistic satisfies the more general definition.

34.8 Bibliographic Remarks

The original essay by Thomas Bayes is remarkably

readable [Bayes, 1763]. There are many texts on
Bayesian statistics. For an introduction to the applied
side, there is the book by Gelman et al. [2014]. This
book offers lots of discussions and examples. A more
philosophical book that takes a foundational look
at probability theory from a Bayesian perspective
is by Jaynes [2003]. The careful definition of the
posterior can be found in several places, but the
recent book by Ghosal and van der Vaart [2017]
does an impeccable job. A worthy mention goes Thomas Bayes
to the article by Chang and Pollard [1997], which
uses disintegration (Theorem 3.12) to formalise the ‘private calculations’ that
probabilists so frequently make before writing everything carefully using Radon–
Nikodym derivatives and regular versions. Theorem 34.1 is a specification of the
theorem guaranteeing the existence of regular conditional probability measures
(Theorem 3.11). For a detailed presentation of exponential families, see the book
by Lehmann and Casella [2006]. A compendium of conjugate priors is by Fink
[1997].

34.9 Exercises

34.1 (Posterior calculations) Evaluate the posteriors for each pair of

conjugate priors in Section 34.3.

34.2 (Uniqueness of beta/Bernoulli posterior) Explain why the posterior

for the Bernoulli model with a beta prior is unique.

34.3 Use the tower rule to prove the identity in Eq. (34.5).

34.4 (Posterior in terms of density) Let P = (Pθ : θ ∈ Θ) be a probability

kernel from (Θ, G) to (X , H) and Q be a probability measure on Θ and P = Q ⊗ P
on Θ × X . As usual, let θ and X be the coordinate projections on Θ × X . Let
ν and µ be probability measures on (Θ, G) and (X , H) such that Q  ν and
 34.9 Exercises 436

Pθ  µ for all θ ∈ Θ, and define

pθ (x)q(θ)
q(θ | x) = R ,
p (x)q(ψ)dν(ψ)
Θ ψ

where pθ (x) = dPθ /dµ and q(θ) = dQ/dν. You may assume that pθ (x) is jointly
measurable in θ and x (see Note 8).
R
(a) Let N = {x : Θ pψ (x)q(ψ)dν(ψ) = 0} and show that PX (N ) = 0.
R
(b) Define Q(A | x) = A q(θ | x)dν(θ) for x ∈/ N and Q(A | x) be an arbitrary
fixed probability measure for x ∈ N . Show that Q( · | X) is a regular version
of P (θ ∈ · | X).

Hint The ‘sections’ lemma may prove useful (Lemma 1.26 in Kallenberg 2002),
along with the properties of the Radon–Nikodym derivative.
34.5 (Exponential families) Let A, T , h, η and Θ be as in Section 34.3.1.
(a) Prove that Pθ is indeed a probability measure.
(b) Let Eθ denote expectations with respect to Pθ . Show that A0 (θ) = Eθ [T ].
(c) Let θ ∈ Θ and X ∼ Pθ . Show that for all λ with λ + θ ∈ θ,
Eθ [exp(λT (X))] = exp(A(λ + θ) − A(θ)) .
(d) Given θ, θ0 ∈ Θ, show that
  
pθ (X)
d(θ, θ ) = Eθ log
0
= A(θ0 ) − A(θ) − (θ0 − θ)A0 (θ) . (34.12)
pθ0 (X)
(e) Let θ, θ0 ∈ Θ be such that A0 (θ0 ) ≥ A0 (θ) and X1 , . . . , Xn be independent
Pn
and identically distributed and T̂ = n1 t=1 T (Xt ). Show that
 
P T̂ ≥ A0 (θ0 ) ≤ exp (−nd(θ0 , θ)) .

Curiously, the function d of Eq. (34.12) is both the relative entropy D(Pθ , Pθ0 )
and the Bregman divergence between θ0 and θ induced by the convex function
A. See Section 26.3 for the definition of Bregman divergence.

34.6 (Exponential weights algorithm) Consider the setting of Note 5.

(a) Prove the claim in Eq. (34.10).
(b) Prove that when Q(µ) = 1/k is the uniform prior, then the regret is bounded
by Rn ≤ log(k) for any y1 , . . . , yn .

34.7 (Measurability of the regret) Let (E, G, Q, P ) be a Bayesian bandit

environment and π a policy. Prove that Rn (π, ν), defined in Eq. (34.8), is G-
measurable as a function of ν.

34.8 (Bayesian optimal regret can be positive) Construct an example

 34.9 Exercises 437

demonstrating that for some priors over finite-armed stochastic bandits, the
Bayesian regret is strictly positive: inf π BRn (π, Q) > 0.

Hint The key is to observe that under appropriate conditions, BRn (π, Q) = 0
would mean that π needs to know the identity of the optimal action under ν from
round one, which is impossible when ν is random and the model is rich enough.
34.9 (Canonical model) Prove the existence of a probability space carrying
the random variables satisfying the conditions in Section 34.4.

34.10 (Sufficiency of deterministic policies) Let ΠD be the set of all

deterministic policies and Π the space of all policies. Prove that for any k-armed
Bayesian bandit environment (E, G, Q, P ),

inf BRn (π, Q) = inf BRn (π, Q) .

π∈Π π∈ΠD

34.11 Prove that the denominator in Eq. (34.7) is almost surely non-zero.

34.12 (Bayesian optimal policies can be dominated) Consider the set-up

in Example 34.2. A Bayesian learner observes X ∼ Pθ and should choose an
action At ∈ [0, 1] that is σ(X)-measurable. Their loss is I {At 6= θ}.

(a) Show that the optimal choice is At = Xt .

(b) Give a Bayesian optimal algorithm with At 6= Xt on some non-empty
(measure zero) event.
(c) Give a Bayesian optimal algorithm and θ such that the loss when θ is true
(and so X ∼ Pθ ) is not zero.

34.13 (Admissible policies are Bayesian for finite environments) Let

E = {ν1 , . . . , νN } and Π be sets. Call the elements of E environments, and the
elements of Π policies (this is just to help to make connection to the rest of the
material). Let ` : Π × E → [0, ∞) be a positive loss function. Given a policy π, let
`(π) = (`(π, ν1 ), . . . , `(π, νN )) be the loss vector resulting from policy π. Define
S = {`(π) : π ∈ Π} ⊂ RN and

λ(S) = {x ∈ cl(S) : y 6< x for all y ∈ S} ,

where y 6< x is defined to mean it is not true that yi ≤ xi for all i with strict
inequality for at least one i (λ(S) is the Pareto frontier of set S, and its elements
are the non-dominated loss-outcome vectors in cl(S)). Prove that if λ(S) ⊆ S
and S is convex, then for every π ∗ ∈ Π such that `(π ∗ ) ∈ λ(S), there exists a
prior q ∈ P(E) such that
X X
q(ν)`(π ∗ , ν) = min q(ν)`(π, ν) .
π∈Π
ν∈E ν∈E

Hint Use the supporting hyperplane theorem, stated in the hint after
Exercise 26.2.
 34.9 Exercises 438

By identifying elements of E as ‘criteria’, the interpretation of the result of the

exercise in multi-criteria optimisation is that for non-empty, convex, closed
loss sets, solutions on the Pareto frontier (policies π such that `(π) ∈ λ(S))
can be obtained by minimizing a convex combination of the individual
criteria. There is also a connection to constrained optimisation where the
constraints are expressed as a bounds on linear combinations of the losses.

34.14 (Uniquely Bayes optimal policies are admissible) Let (E, G) be a

measurable space and Π an arbitrary set of the elements that we call policies. Let
` : Π × E → R be a function with `(π, ·) being G-measurable. Given a probability
measure Q on (E, G), a policy is called Bayesian optimal with respect to Q if
Z Z
`(π, ν)dQ(ν) = inf0
`(π 0 , ν)dQ(ν) .
E π ∈Π E

Prove the following:

(a) If π is the unique Bayesian optimal policy given prior Q, then π is admissible.
(b) There is an example when π is a Bayesian optimal policy and π is inadmissible.
(c) If E is countable and Supp(Q) = E, then any Bayes optimal policy π is
admissible.
(d) If π is Bayesian optimal with respect to prior Q, then it is admissible on
Supp(Q) ⊆ E.

34.15 (Admissible policies are Bayesian for Bernoulli bandits) Let E

be the set of k-armed Bernoulli bandits. Prove that every admissible policy is
Bayesian optimal for some prior.

Hint Argue that all policies can be written as convex combinations of

deterministic policies using an appropriate linear structure. Then identify the
spaces of environments and policies with compact metric spaces. Let (νj )∞ j=1 be
a dense subset of E and repeat the argument in the previous exercise with each
finite subset {ν1 , . . . , νj }, and then take the limit as j → ∞. You will probably
find Theorem 2.14 useful.
34.16 Let E = EBk be the space of k-armed stochastic Bernoulli bandits. Endow
E with a topology via the natural bijection to [0, 1]k , and let Q be the space of
all probability measures on (E, B(E)) with the weak* topology. Prove that
max BR∗n (Q) = Rn∗ (E) .
Q∈Q

Hint Use Theorem 2.14 and Sion’s theorem (Theorem 28.12).

35 Bayesian Bandits

The first section of this chapter provides simple bounds on the Bayesian optimal
regret, which are obtained by integrating the regret guarantees for frequentist
algorithms studied in Part II. This is followed by a short interlude on the basic
theory of optimal stopping, which we will need later. The next few sections
are devoted to special cases where computing the Bayesian optimal policy is
tractable. We start with the finite horizon Bayesian one-armed bandit problem
where the existence of a tractable solution is reduced to the computation of a
sequence of functions on the sufficient statistics of the arm with the unknown
pay-off. Next, the k-armed setting is considered. The main question is whether
there exists a solution that avoids considering joint sufficient statistics over all
arms, which would be intractable in the lack of further structure (see Note 2).
Avoiding the joint sufficient in general is not possible, but in the remarkable
case of the problem of maximising the total expected discounted reward over an
infinite horizon, where John C. Gittins’s celebrated result shows that the Bayesian
optimal policy takes the form of an ‘index’ policy that keeps statistics for each
arm separately (updated based on the arm’s observations only) to compute a
value (‘index’) for each arm, in each round choosing the arm with the highest
index.

35.1 Bayesian Optimal Regret for k-Armed Stochastic Bandits

Even in relatively benign set-ups, the computation of the Bayesian optimal policy
appears hopelessly intractable. Nevertheless, one can investigate the value of the
Bayesian optimal regret by proving upper and lower bounds.
For simplicity, we restrict our attention to Bernoulli bandits, but the arguments
generalise to other models. Let (E, G) = ([0, 1]k , B([0, 1]k )), and for ν ∈ [0, 1]k
let Pνj = B(νj ). Choose some prior Q on (E, G). The Bayesian optimal regret is
necessarily smaller than the minimax regret, which by Theorem 9.1 means that
√
BR∗n (Q) ≤ C kn ,
where C > 0 is a universal constant. The proof of the lower bound in Exercise 15.2
shows that for each n, there exists a prior Q for which
√
BR∗n (Q) ≥ c kn ,
 35.2 Optimal Stopping ( ) 440

where c > 0 is a√universal constant. These two together show that the
supQ BR∗n (Q) = Θ( kn).
Turning to the asymptotics for a fixed distribution, recall that that for any fixed
Bernoulli bandit environment, the asymptotic growth rate of regret is Θ(log(n)).
In stark contrast to this, the best we can say in the Bayesian case is that the
√ √
asymptotic growth rate of BR∗n (Q) is slower than n, but for some priors, n is
almost a lower bound on the growth rate. In particular, we ask you to prove the
following theorem in Exercise 35.1:

Theorem 35.1. For any prior Q,

BR∗n (Q)
lim sup = 0.
n→∞ n1/2
Furthermore, there exists a prior Q such that for all ε > 0,
BR∗n (Q)
lim inf = ∞.
n→∞ n1/2−ε
The lower bound has a worst-case flavour in the sense it holds for a specific prior.
The prior that yields the lower bound is a little unnatural because it assigns the
overwhelming majority of its mass to bandits with small suboptimality gaps. In
particular, Q({ν ∈ E : ∆min (ν) ≤ ∆}) ≥ c/ log(1/∆) for some constant c > 0. For
more regular priors, the Bayesian optimal regret satisfies BR∗n (Q) = Θ(log2 (n)).
See the bibliographic remarks for pointers to the literature.

35.2 Optimal Stopping ( )

We now make a detour to show some results of optimal stopping, which will be
used in the next sections to find tractable solutions to certain Bayesian bandit
problems.
The first setting we consider will be useful for the one-armed bandit problem.
Let (Ut )nt=1 be a sequence of random variables adapted to filtration F = (Ft )nt=1 .
Optimal stopping is concerned with finding solutions to optimisation problems of
the following form:

sup E[Uτ ] , (35.1)

τ ∈Rn
1

where Rn1 is the set of F-stopping times τ with 1 ≤ τ ≤ n. When n is finite,

the situation is conceptually straightforward. The idea is to use backwards
induction to define the expected optimal utility conditioned on the information
in Ft starting from t = n and working backwards to t = 1. The Snell envelope
is a sequence of random variables (Et )nt=1 defined by
(
Un , if t = n ;
Et =
max {Ut , E[Et+1 | Ft ]} , otherwise .
 35.2 Optimal Stopping ( ) 441

Intuitively, Et is the optimal expected value one can guarantee provided that
stage t was reached.
Theorem 35.2. Assume that n is finite and Ut is integrable for all t ∈ [n]. Then
the stopping time τ = min{t ∈ [n] : Ut = Et } ∈ Rn1 achieves the supremum in
Eq. (35.1).
Backwards induction is not directly applicable when the horizon is infinite.
There are several standard ways around this problem. For our purposes, the most
convenient workaround is to introduce a Markov structure. The connection to
the Bayesian bandit setting is that in the Bayesian setting, posteriors follow a
Markov process. The connection will be made explicit in a few examples in later
sections.
Let (S, G) be a Borel space and (Px : x ∈ S) be a probability kernel from S to
itself and u : S → R be S/B(R)-measurable. A Markov reward process is a
Markov chain (St )∞t=1 evolving according to P and a sequence of random variables
(Ut )∞
t=1 with Ut = u(St ). Define the filtration F = (Ft )∞
t=1 with Ft = σ(S1 , . . . , St ).
The (Markov) optimal stopping problem is
sup E[Uτ ] ,
τ ∈R1

where R1 is the set of F-adapted stopping times, and the initial distribution of
S1 is arbitrary. Inspired by the solution of the finite horizon problem define the
value function v : S → R by
v(x) = sup Ex [Uτ ] , (35.2)
τ ∈R1

where Px is the probability measure on the space carrying (St )∞ t=1 for which
Px (S1 = x) = 1 and Ex be the expectation
R with respect to Px . As before, the
idea is to stop when Ut is above S v(y)PSt (dy), the predicted optimal value of
continuing. Note that ties can be resolved in any way (depending on St , one may
or may not stop when the predicted optimal value of continuation is equal to Ut ).
The next result gives sufficient conditions under which stopping rules of this form
are indeed optimal.
Theorem 35.3. Assume for all x ∈ S that U∞ = limn→∞ Un exists Px -a.s. and
supn≥1 |Un | is Px -integrable. Then v satisfies the Wald–Bellman equation,
Z
v(x) = max{u(x), v(y)Px (dy)} for all x ∈ S .
S

Furthermore, limn→∞ v(Sn ) = U∞ Px -a.s., and the supremum in Eq. (35.2) is

achieved by any stopping time τ such that for all t,
R
(a) τ ≤ t on the event that Ut > RS v(y)PSt (dy); and
(b) τ > t on the event that Ut < S v(y)PSt (dy) and τ ≥ t.
The conditions are satisfied in many practical applications, e.g. if the Markov
chain is ergodic and the utility function is bounded over the state space. In
 35.3 One-armed Bayesian Bandits 442

our application, Un will be an accumulation of discounted rewards, which in all

standard situations converges very fast.

A natural choice of stopping time satisfying conditions (a) and (b) in

Theorem 35.3 is τ = min{t ≥ 1 : v(St ) = Ut }.
R The conditions express that
in the indifference region {x ∈ S : u(x) = S v(y)Px (dy)}, both stopping
and continuing are acceptable.

The proof of Theorem 35.2 is straightforward (Exercise 35.2). Measurability

issues make the proof of Theorem 35.3 more technical (Exercise 35.3). Pointers
to the literature are given in the notes, and a solution to the exercise is available.

35.3 One-armed bandits

The one-armed Bayesian bandit problem is a special

case where the Bayesian optimal policy has a simple
form that can often be computed efficiently. Before
reading on, you might like to refresh your memory
by looking at Exercises 4.10 and 8.2. Let (E, G, Q, P )
be a two-armed Bayesian bandit environment, where
Pν2 = δµ2 is a Dirac at fixed constant µ2 ∈ R for
all ν ∈ E. Because the mean of the second arm is
known in advance, we call this a one-armed Bayesian £34.5

bandit problem. In Part (a) of Exercise 4.10, you

showed that when the horizon is known, retirement
policies that choose the first arm until some random Figure 35.1 When will you
time before switching to the second arm until the stop playing? A one-armed
end of the game (pointwise over ν) dominate all Bayesian bandit.
other policies in terms of regret. Since we care about
Bayesian optimal policies, the result of Exercise 34.10 allows us to restrict our
attention to deterministic retirement policies.
These facts allow us to frame the Bayesian one-armed bandit problem in
terms of optimal stopping. Define a probability space (Ω, F, P) carrying random
elements ν ∈ E and Z = (Zt )nt=1 where

(a) the law of ν is Pν = Q; and

(b) P(Z ∈ · | ν) = Pν1
n
(·), which means that after conditioning on ν, the sequence
(Zt )t=1 is independent and identically distributed according to Pν1 .
n

Given a deterministic retirement policy π = (πt )nt=1 , define the random variable

τ = min{t ≥ 1 : πt (2 | 1, Z1 , . . . , 1, Zt−1 ) = 1} ,

where the minimum of an empty set in this case is n+1. Clearly τ is an F-stopping
 35.3 One-armed Bayesian Bandits 443

time, where F = (Ft )n+1

t=1 with Ft = σ(Z1 , . . . , Zt−1 ). In fact, this correspondence
between deterministic retirement policies and F-stopping times is a bijection. The
Bayesian expected reward when following the policy associated with stopping
time τ is
"τ −1 n
#
X X
E Zt + µ2 = E [Uτ ] ,
t=1 t=τ
Pt−1
where Ut = s=1 Zs + (n − t + 1)µ2 . Since minimizing the Bayesian regret is
equivalent to maximising the Bayesian expected cumulative reward, the problem
of finding the Bayesian optimal policy has been reduced to an optimal stopping
problem.

Proposition 35.4. If Z1 is integrable, then the Bayesian regret is minimised by

the retirement policy associated with stopping time τ = min{t ≥ 1 : Ut = Et },
where
(
Ut , if t = n + 1 ;
Et =
max{Ut , E[Et+1 | Ft ]} , otherwise .

The interpretation of Et is that it is the total expected optimal value

conditioned on the information available at the start of round t. The proposition
is an immediate corollary of Theorem 35.2 and the fact that integrability of
Z1 is equivalent to integrability of (Ut )n+1
t=1 . The optimal stopping time in
Proposition 35.4 can be rewritten in a more convenient form. For 1 ≤ t ≤ n + 1,
Pt−1
define Wt = Et − s=1 Zs , which can be seen as the optimal value to go for the
last n − t + 1 rounds. The definition of Et shows that Wn+1 = 0 and for t ≤ n,
t−1
!
X
Wt = max (n − t + 1)µ2 , E[Et+1 | Ft ] − Zs
s=1

= max ((n − t + 1)µ2 , E [Zt + Wt+1 | Ft ]) . (35.3)

Hence the optimal stopping time can be rewritten as

τ ∗ = min{t : Ut = Et } = min {t : Wt = (n − t + 1)µ2 } .

This should make intuitive sense. It is optimal to continue only if the expected
future reward from doing so is at least as large as what can be obtained by
stopping immediately. The difficulty is that E[Zt + Wt+1 | Ft ] can be quite a
complicated object. We now give two examples where E[Zt + Wt+1 | Ft ] has a
simple representation and thus computing the optimal stopping rule becomes
practical. The idea is to find a sequence of sufficient statistics (St )nt=0 so that
St ∈ S is Ft -measurable and Pν1 (Z1 , . . . , Zt ∈ · | St ) is independent of ν. Then
Et is σ(St )-measurable, and by Lemma 2.5 it follows that Et = vt (St ) for an
appropriately measurable function vt : S → R. For more on this, read the next
two subsections, and then do Exercise 35.4.
 35.3 One-armed Bayesian Bandits 444

35.3.1 Bernoulli Rewards

Let E = [0, 1], G = B([0, 1]) and for ν ∈ E, let Pν1 = B(ν) and Pν2 = δµ2 : the first
arm is Bernoulli, and the second is a Dirac at some fixed value µ2 ∈ [0, 1]. For the
prior, choose Q = Beta(α, β), a Beta prior. By the argument in Example 34.6 the
posterior at the start of round t is a Beta distribution Beta(α + St , β + t − 1 − St )
Pt−1
where St = s=1 Zs . Letting pt (s) = (α + s)/(α + β + t − 1), it follows that
α + St
E[Zt | Ft ] = = pt (St ) ,
α+β+t−1
P (St+1 = St + 1 | St ) = pt (St ) ,
P (St+1 = St | St ) = 1 − pt (St ) .
Now let wn+1 (s) = 0 for all s and
wt (s) = max {(n − t + 1)µ2 , pt (s) + pt (s)wt+1 (s + 1) + (1 − pt (s))wt+1 (s)} .
Then Wt = wt (St ), and hence the optimal policy can be computed by evaluating
wt (s) for all s ∈ {0, . . . , t} starting with t = n, then n − 1 and so on until t = 1.
The total computation for this backwards induction is O(n2 ), and the output
is a policy that can be implemented over all n rounds. By contrast, the typical
frequentist stopping rule requires only O(n) computations, so the overhead is
quite severe. The improvement in terms of the Bayes regret is not insignificant,
however, as illustrated by the following experiment.
Experiment 35.1 The horizon is set to n = 500 and µ2 = 1/2. The stopping
rules we compare are the Bayesian optimal policy with a Beta(1, 1) prior and the
‘frequentist’ stopping rule given by
 
log(n/t)
τ = min t ≥ 2 : µ̂t−1 < µ2 and d(µ̂t−1 , µ2 ) ≥ , (35.4)
t−1
Pt
where d(p, q) is the binary relative entropy and µ̂t = s=1 Xs /t is the empirical
estimate of µ1 based on the first t observations. Fig. 35.2 shows the expected regret
for different values of µ, with horizontal dotted lines indicating the expected regret
averaged over the prior. Note that although the prior is symmetric, the one-armed
bandit problem is not, which explicates the asymmetric behaviour of the Bayesian
optimal algorithm. The frequentist algorithm is even more asymmetric with very
small regret for µ1 > 1/2, but large regret for µ1 < 1/2. This is caused by a
conservative confidence interval in Eq. (35.4), which makes it stop consistently
later than its Bayesian counterpart, which makes it ‘win’ for µ1 > 1/2, but it also
makes it ‘lose’ when µ1 < 1/2, with an overall loss (naturally) when considering
the average over all environments.

35.3.2 Gaussian Rewards

Let (E, G) = (R, B(R)), where for ν ∈ R, we let Pν1 = N (ν, 1) and Pν2 = δµ2
for µ2 ∈ R fixed. Choose a Gaussian prior Q = N (µP , σP2 ) with mean µP ∈ R
 35.3 One-armed Bayesian Bandits 445

Bayesian optimal
8 Frequentist

Expected regret 6

0
0.2 0.4 0.6 0.8
µ1

Figure 35.2 The plot shows the expected regret for the Bayesian optimal algorithm
compared to the ‘frequestist’ algorithm in Eq. (35.4) on the Bernoulli 1-armed bandit
where µ2 = 1/2 and µ1 varies on the x-axis. The horizontal lines show the average
regret for each algorithm with respect to the prior, which is uniform.

and variance σP2 > 0. By the results in Section 34.3, the posterior Q(· | x1 , . . . , xt )
after observing rewards x1 , . . . , xt from the first arm is almost surely Gaussian
with mean µt and variance σt2 given by
Pt  −1
2 +
µP
σP s=1 xs 2 1
µt = and σ = t + . (35.5)
1 + σP−2
t
σP2
The posterior variance is independent of the observations, so the posterior is
determined entirely by its mean. As in the Bernoulli case, there exist functions
(wt )n+1
t=1 such that Wt = wt (µt−1 ) almost surely for all t ∈ [n]. Precisely,
wn+1 (µ) = 0 and for t ≤ n,
 Z ∞   
1 x2
wt (µ) = max (n − t + 1)µ2 , µ + √ exp − 2 wt+1 (µ + x)dx .
2π −∞ 2σt−1
(35.6)
The integral on the right-hand side does not have a closed-form solution, which
forces the use of approximate methods. Fortunately wt is a well-behaved function
and can be efficiently approximated. The favourable properties are summarised
in the next lemma, the proof of which is left to Exercise 35.5.
Lemma 35.5. The following hold:
(a) The function wt is increasing.
(b) The function wt is convex.
(c) limµ→∞ wt (µ)/µ = n − t + 1 and limµ→−∞ wt (µ) = (n − t + 1)µ2 .
There are many ways to approximate a function, but in order to propagate
the approximation using Eq. (35.6), it is convenient to choose a form for which
the integral in Eq. (35.6) can be computed analytically. Given the properties in
 35.4 Gittins Index 446

Lemma 35.5, a natural choice is to approximate wt using piecewise quadratic

functions. Let w̃n+1 (µ) = 0 and
 Z ∞   
1 x2
w̄t (µ) = max (n − t + 1)µ2 , µ + √ exp − 2 w̃t+1 (µ + x)dx .
2π −∞ 2σt
Then let −∞ < x1 ≤ x2 ≤ . . . ≤ xN < ∞, and for µ ∈ [xi , xi+1 ], define
w̃t (µ) = ai µ2 + bi µ + ci to be the unique quadratic approximation of w̄t (µ) such
that
w̃t (xi ) = w̄t (xi ) ,
w̃t (xi+1 ) = w̄t (xi+1 ) ,
w̃t ((xi + xi+1 )/2) = w̄t ((xi + xi+1 )/2) .
For µ < x1 , we approximate wt (µ) = (n − t + 1)µ2 , and for µ > xN , the linear
approximation w̃t (µ) = (n−t+1)µ is reasonable by Lemma 35.5. The computation
time for calculating the coefficients ai , bi , ci for all t and i ∈ [N ] is O(N n). We
encourage the reader to implement this algorithm and compare it to its natural
frequentist competitors (Exercise 35.11).

35.4 Gittins Index

Generalising the analysis in the previous section to multiple actions

is mathematically straightforward, but computationally intractable. The
computational complexity of backwards induction increases exponentially with
the number of arms, which is impractical unless the number of arms and horizon
are both small.
An index policy is a policy that in each round computes a real-valued index
for each arm and plays the arm with the largest index, while the index of an
arm is restricted to depend on statistics collected for that arm only (the time
horizon can also be used). Many policies we met earlier are index policies. For
example, most variants of the upper confidence bound algorithm introduced in
Part II are index policies. Sadly, however, the Bayesian optimal policy for finite
horizon bandits is not usually an index policy (see Note 6). John C. Gittins
proved that if one is prepared to modify the objective to a special kind of infinite
horizon problem, then the Bayesian optimal policy becomes an index policy. In
the remainder of this chapter, we explore his ideas.

35.4.1 A Discounted Retirement Game

We start by describing the discounted setting with one action and then generalise
to multiple actions. Besides discounting, another change is that the reward-
generating process is made into a Markov reward process, a strict generalisation
of the previous case. The motivation is that, as hinted on before, the posterior of
the arm with the unknown payoff evolves as a Markov process.
 35.4 Gittins Index 447

Let (St )∞
t=1 be a Markov chain on Borel space (S, G) evolving according to
probability kernel (Px : x ∈ S). As in Section 35.2, let (Ω, F, Px ) be a probability
space carrying (Sn )∞
n=1 with Sn ∈ S such that

(a) Px (S1 = x) = 1; and

(b) Px (Sn+1 ∈ · | Sn ) = PSn (·) with Px -probability one.

Expectations with respect to Px are denoted by Ex . Next, let γ ∈ R and r : S → R

be a G/B(R)-measurable function, both of which are known to the learner. In
each round t = 1, 2, . . ., the learner observes the state St and chooses one of two
options: (a) to retire and end the game or (b) pay the fixed cost γ to receive
a reward of r(St ) and continue for another round. The policy of a learner in
this game corresponds to choosing a F-stopping time τ with F = (Ft )t and
Ft = σ(S1 , . . . , St ), where τ = t means that the learner retires after observing
St at the start of round t. The α-discounted value of the game when starting in
state S1 = x is
"τ −1 #
X
vγ (x) = sup Ex α (r(St ) − γ) ,
t−1
(35.7)
τ ≥1 t=1

where α ∈ (0, 1) is the discount factor. To ensure that this is well defined, we
need the following assumption:
"∞ #
X
Assumption 35.6. For all x ∈ S, it holds that Ex α |r(St )| < ∞.
t−1

t=1

If the rewards are bounded, the assumption will hold. When the rewards are
unbounded, the assumption restricts the rate of growth of rewards over time.
The presence of discounting encourages the learner to obtain large rewards
earlier rather than later and is one distinction between this model and the finite-
horizon model studied for most of this book. A brief discussion of discounting is
left for the notes.
Fix a state x ∈ S. The map γ 7→ vγ (x) is decreasing and is always non-negative.
In fact, if γ is large enough, it is easy to see that retiring immediately (τ = 1)
achieves the supremum in the definition of vγ (x), and thus vγ (x) = 0. The
Gittins index, or fair charge, of a state x is the smallest value of γ for which
the learner is indifferent between retiring immediately and playing for at least
one round:

g(x) = inf {γ ∈ R : vγ (x) = 0} . (35.8)

Straightforward manipulation (Exercise 35.6) shows that

hP i
r(St )
τ −1 t−1
Ex t=1 α
g(x) = sup hP i . (35.9)
τ −1 t−1
τ ≥2 Ex t=1 α

The form in (35.9) will be useful for computation. It is not immediately clear that
 35.4 Gittins Index 448

a stopping time attaining the supremum in (35.9) exists. The following lemma
shows that it does and gives an explicit form.
Lemma 35.7. Let x ∈ S be arbitrary. The following hold under Assumption 35.6:
R
(a) vγ (x) = max{0, r(x) − γ + α S vγ (y)PxR(dy)} for all γ ∈ R.
(b) If γ ≤ g(x), then vγ (x) = r(x) − γ + α S vγ (y)Px (dy).
(c) The stopping time τ = min{t ≥ 2 : g(St ) ≤ γ} attains the supremum in
Eq. (35.9).
The result is relatively intuitive. The Gittins index represents the price the
learner should be willing to pay for the privilege of continuing to play. The optimal
policy continues to play as long as the actual value of the game is not smaller
than this price was at the start. The proof of Lemma 35.7 uses Theorem 35.3
and is left for the reader in Exercise 35.7.

35.4.2 Discounted Bandits and the Index Theorem

The generalisation of the discounted retirement game to multiple arms is quite
straightforward. As we will see, this will lead to a solution to the infinite horizon
discounted Bayesian k-armed bandit problem where the prior factorises over the
arms.
There are now k independent Markov chains sharing the same state space S.
We are also given a reward function r : S → R. In each round t, the learner first
observes the state of all chains S1 (t), . . . , Sk (t) and then chooses an action At ∈ [k]
to receive a reward r(SAt (t)) and to make the state of the chain underlying arm
k move according to a fixed transition kernel that is common to all chains. The
states of the other chains do not move. The goal is still to maximise the total
expected discounted reward. The interaction protocol is illustrated on Fig. 35.3.

The assumption that the Markov chains evolve on the same state space with
the same transition kernel is non-restrictive since the state space can always
be taken to be the union of k state spaces and the transition kernel defined
with k disconnected components.

Because the learner observes the state of all chains in each round, a policy π
now is a collection (πt )∞
t=1 , where πt is a probability kernel from (S ×[k])
k t−1
×S k
(history, including past observed states and actions) to [k]. Given a discount
rate α ∈ (0, 1), the objective is to find the policy maximising the cumulative
discounted reward:
"∞ #
X
argmaxπ Eπ α r(SAt (t)) ,
t−1

t=1

where the expectation is taken with respect to the distribution on state/action

sequences induced by the interaction of π and the k Markov chains.
 35.4 Gittins Index 449

t = 1 and initialise S1 (1), . . . , Sk (1)

Observe states S1 (t), . . . , Sk (t)

Choose action At ∈ [k] Increment t

Update Si (t + 1) = Si (t) for

Receive reward r(SAt (t)) i 6= At , and SAt (t + 1) ∼ PSAt (t) (·)

Figure 35.3 Interaction protocol for discounted bandits with Markov pay-offs

Example 35.8 (Bayesian k-armed Bernoulli bandits in the Markov framework).

To see the relation to Bayesian bandits with discounted rewards, consider the
following set-up. Let S = [0, ∞) × [0, ∞) and G = B(S). Then let the initial state
of each Markov chain be Si (1) = (1, 1), and define probability kernel (Ps : s ∈ S)
from (S, G) to itself by
x y
P(x,y) (A) = IA ((x + 1, y)) + IA ((x, y + 1)) .
x+y x+y
The reward function is r(x, y) = x/(x + y). The reader should check that this
corresponds to a Bernoulli bandit with Beta(1, 1) prior on the mean reward of
each arm (Exercise 35.8). The role of the state space is to maintain a sufficient
statistic for the posterior while the reward function is the expected reward given
the posterior.
Returning to the general problem, let g be the Gittins index function as defined
in Eq. (35.8) associated with the probability kernel (Px : x ∈ S) and reward
function r. A policy π ∗ that chooses in round t the arm At ∈ argmaxi∈[k] g(Si (t))
is called a Gittins index policy. One of the most celebrated theorems in the
study of bandits is that these policies are Bayesian optimal.
Theorem 35.9. Let π ∗ be a policy choosing in round t At = argmaxi g(Si (t))
with ties broken arbitrarily. Then, provided Assumption 35.6 holds for all Markov
chains (Siu )∞
u=1 , then
"∞ # "∞ #
X X
Eπ∗ α r(SAt (t)) = sup Eπ
t−1
α r(SAt (t)) ,
t−1
π
t=1 t=1

where the supremum is taken over all policies.

The remainder of the section is devoted to proving Theorem 35.9. The choice
of actions produces an interleaving of the rewards generated by each Markov
chain, and it will be useful to have a notation for these interleavings. For each
 35.4 Gittins Index 450

i ∈ [k], let gi = (git )∞

t=1 be a real-valued sequence and g = (g1 , . . . , gk ) be the
tuple of these sequences.

While this notation breaks our convention of putting the time index first in
the reward sequences of a multi-armed bandit, we prefer this notation here
as we need to consider reward sequences underlying individual arms.

Given an infinite sequence (at )∞ t=1 , taking values in [k], define the interleaving
sequence I(g, a) = (It (g, a))∞
t=1 by
t−1
X
It (g, a) = gat ,1+nat (a,t−1) with ni (a, t − 1) = I {as = i} .
s=1

Note that this is the same as the ‘reward-stack model’ of bandits mentioned on
page 65 in Chapter 4 except that here we have fixed sequences. The next lemma
follows from the Hardy–Littlewood inequality, a generalisation of the trivial
observation that the identical ordering of two sequences of numbers maximises
their inner product. We leave the proof to Exercise 35.9.

Lemma 35.10. Suppose that gi is decreasing for all i ∈ [k] and (a∗t )∞ t=1 is defined
recursively by a∗t = argmaxi gi,1+ni (a∗ ,t−1) and I ∗ (g) = I(g, a∗ ). Then, for any
α ∈ (0, 1),
∞
X ∞
X
αt−1 It∗ (g) = sup αt−1 It (g, a) .
t=1 a∈[k]N t=1

Proof of Theorem 35.9 Given a policy π = (πt )∞ t=1 , let (Ω, F, Pπ ) be a

probability space carrying random elements S1 , . . . , Sk , where Si = (Siu )∞u=1
is a sequence of states and (At )∞
t=1 is a sequence of actions such that

(a) Pπ (Si,u+1 ∈ · | Si1 , Si2 , . . . , Siu ) = PSiu (·);

(b) The sequences (Siu )∞ u=1 and (Sju )u=1 are independent for all i 6= j; and
∞

(c) Pπ (At ∈ · | S(1), A1 , . . . , At−1 , S(t)) = πt (· | S(1), A1 , . . . , At−1 , S(t)), where

Si (t) = Si,(1+Ti (t−1)) is the state of machine i observed by the learner at the
Pt
start of round t with Ti (t) = s=1 I {As = i}.

Let Ft = σ(S(1), A1 , S(2), . . . , At−1 , S(t), At ) be the σ-algebra containing

information available to the learner after choosing their action in round t. As
usual, Eπ denotes the expectation with respect to Pπ .
Given an arm i and round t, the prevailing charge is a random variable
Gi (t) = mins≤t g(Si (s)). The name comes from one of the early proofs of Gittins
theorem that constructed a game in which the prevailing charge was the fee paid
by the learner to play arm i in round t. The proof is decomposed into two steps.
In the first step, we relate the prevailing charge to the discounted cumulative
reward. The second step completes the proof by combining the first with an
interleaving argument using Lemma 35.10.
 35.4 Gittins Index 451

Part 1: The Prevailing Charge

Fix an arm i. We claim that
"∞ # "∞ #
X X
Eπ αt−1 r(Si (t))I {At = i} ≤ Eπ αt−1 Gi (t)I {At = i} .
t=1 t=1

Furthermore, equality holds for π = π . To prove this claim, let τ1 , τ2 , . . . be a

∗

sequence of stopping times defined recursively by

τ1 = min{t ≥ 1 : At = i} and
τj+1 = min{t > τj : At = i and g(Si (t)) ≤ Gi (τj )} ,

where the minimum of the empty set is defined to be infinite. Next, let

Tj = {t : At = i and τj ≤ t < τj+1 } and γj = Gi (τj ) .

Note that on the event {τj < ∞}, Gi (t) = γj for all t ∈ Tj . Furthermore,
g(Si (τj )) = γj . By definition, we have
"∞ #  
X ∞
X X
Eπ αt−1 (r(Si (t)) − Gi (t)) I {At = i} = Eπ  αt−1 (r(Si (t)) − γj ) .
t=1 j=1 t∈Tj

The claim follows by showing the term inside the sum on the right-hand side
vanishes for the Gittins index policy and is not positive for any other policy.
Fix j ≥ 1. By definition, for t ∈ Tj it holds that gi (Si (t)) ≥ Gi (t) = γj .
Combining this with Part (b) of Lemma 35.7, on {t ∈ Tj }, thanks to {t ∈ Tj } ∈
Ft ,
Z
vγj (Si (t)) + γj − r(Si (t)) = α vγj (y)PSi (t) (dy) = αEπ [vγj (Si (t + 1)) | Ft ] .
S

From this it follows that

   
X X

Eπ  αt−1 (r(Si (t)) − γj ) = Eπ  αt−1 vγj (Si (t)) − αvγj (Si (t + 1)) 
t∈Tj t∈Tj

≤ 0,

where the final inequality holds since vγj is non-negative, vγj (Si (τj )) = 0 and by
telescoping the sum, which is possible because whenever t0 is the smallest element
larger than t in Tj , then Si (t0 ) = Si (t + 1). We now argue that the inequality
is replaced by an equality for the Gittins index policy. The key observation
is that having played Aτj = i, the Gittins index policy continues playing arm
i until g(Si (t)) ≤ γj , which means that Tj = {τj , τj + 1, . . . , κj − 1}, where
κj = min{t > τj : g(Si (t)) ≤ γj }, which by Part (c) of Lemma 35.7 means that
 
X
Eπ∗  αt−1 (r(Si (t)) − γj ) | Fτj  = vγj (Si (τj )) = 0 .
t∈Tj
 35.5 Computing the Gittins Index 452

Part 2: Interleaving Prevailing Charges

Let Hiu = minv≤u g(Siv ). The key point is that the distribution of H = (Hiu )
does not depend on the choice of policy, and clearly Hiu is decreasing in u for
each i. For the Gittins index policy π ∗ ,
"∞ # "∞ #
X X
Eπ∗ α r(SAt (t)) = Eπ∗
t−1
α GAt (t)
t−1

t=1 t=1
" ∞
#
X
= Eπ∗ α t−1
It (H, A)
t=1
"∞ #
X
= Eπ∗ αt−1 It∗ (H) ,
t=1

where the first equality follows from part 1, the second by the definition of It
and H and the third by the definitions of It∗ from Lemma 35.10 and that of
the Gittins index policy, which always chooses an action that maximises the
prevailing charge. On the other hand, for any policy π,
"∞ # "∞ #
X X
Eπ α r(SAt (t)) ≤ Eπ
t−1
α GAt (t)
t−1

t=1 t=1
" n
#
X
= Eπ α t−1
It (H, A)
t=1
" n #
X
≤ Eπ αt−1 It∗ (H) ,
t=1

where the last line follows from Lemma 35.10. Finally, note that the law of H
under Pπ does not depend on π, and hence
" n # " n #
X X
Eπ α It (H) = Eπ∗
t−1 ∗
α It (H) .
t−1 ∗

t=1 t=1

Therefore, for all π,

"∞ # "∞ #
X X
Eπ∗ α r(SAt (t)) ≥ Eπ
t−1
α r(SAt (t)) ,
t−1

t=1 t=1

which completes the proof.

35.5 Computing the Gittins Index

We describe a simple approach that depends on the state space being finite.
References to more general methods are given in the bibliographic remarks.
Assume without loss of generality that S = {1, 2, . . . , |S|} and G = 2S . The matrix
form of the transition kernel is P ∈ [0, 1]|S|×|S| and is defined by Pij = Pi ({j}).
We also let r ∈ [0, 1]|S| be the vector of rewards so that ri = r(i). The standard
 35.6 Notes 453

basis vector is ei ∈ R|S| , and 1 ∈ R|S| is the vector with one in every coordinate.
For C ⊂ S, let QC be the transition matrix with (QC )ij = Pij IC (j). For each
i ∈ S, the goal is to find
hP i
)
τ −1 t−1
Ei t=1 α r(S t
g(i) = sup hP i ,
τ −1 t−1
τ ≥2 Ei t=1 α

where Ei is the expectation with respect to the measure Pi for which the initial
state is S1 = i. Lemma 35.7 shows that the stopping time τ = min{t ≥ 2 : g(St ) ≤
g(i)} attains the supremum in the above display. The set Ci = {j : g(j) > g(i)}
is called the continuation region, and Si = S \ Ci is the stopping region. Then
the Gittins index can be calculated as
hP i
Ei
τ −1 t−1
r(St ) P∞ t−1 > t−1
t=1 α i (I − αQCi )
e> −1
t=1 α ei QCi r r
g(i) = hP i = P∞ = .
Ei
τ −1 t−1
t=1 α
t−1 > t−1
ei QCi 1 ei (I − αQCi ) 1
> −1
t=1 α

All this suggests an induction approach where the Gittins index is calculated
for each state in decreasing order of their indices. To get started, note that the
maximum possible Gittins index is maxi ri and that this is achievable for state
i = argmaxj rj with the deterministic stopping time τ = 2. For the induction
step, assume that g(i) is known for the j states C = {i1 , i2 , . . . , ij } with the
largest Gittins indices. Then ij+1 is given by

i (I − αQC )
e> −1
r
ij+1 = argmaxi∈C .
/
ei (I − αQC ) 1
> −1

If Gauss–Jordan elimination is used for matrix inversion, then the computational

complexity of this algorithm is O(|S|4 ). A more sophisticated inversion algorithm
would reduce the complexity to O(|S|3+ε ) for some ε ≤ 0.373, but these are
seldom practical. When α is relatively small, the inversion can be replaced by
directly calculating the sums to some truncated horizon with little loss in accuracy.

35.6 Notes

1 Bayesian methods automatically and optimally exploit the assumptions encoded

in their prior. If we think of the prior as a way of enriching and refining the
standard formulation of bandits, this is an advantage. However, this blessing
can also be a curse. A policy that exploits its assumptions too heavily can
be brittle when those assumptions turn out to be wrong. This can have a
devastating effect in bandits where the cost of overly aggressive confidence
intervals is large.
2 We claimed that computing the Bayesian optimal policy is generally intractable
without discounting. This is a widely held belief, but we are not aware of any
lower-bound on the computation complexity. A good place to start might be
 35.6 Notes 454

to lower bound the computation complexity of finding the optimal action for
k-armed Bayesian bandits when the prior is a product of Beta distributions,
but without discounting.
3 The solution to optimal stopping problems is essentially a form of dynamic
programming, which is a method that trades memory for computation by
introducing recursively defined value functions that suffice for reconstructing an
optimal policy. In the one-armed bandit optimal stopping problem, thanks to
the factorisation lemma (Lemma 2.5), for any 0 ≤ t ≤ n, there exists a function
wt : Rt → R such that Wt = wt (X1 , . . . , Xt ) almost surely. This function can
be seen as the value function that captures the optimal value-to-go from stage
t on, and (35.3) gives a recursive construction for it, wn (x1 , . . . , xn ) = 0, and
for t < n,
Z
wt (x1 , . . . , xt ) = max((n − t)µ2 , xt+1 + wt+1 (x1 , . . . , xt , xt+1 )dPt (xt+1 )) ,

where Pt is the distribution of Xt+1 given X1 , . . . , Xt . The problem with

this general recursion is that the computation is prohibitive. The example
with Bernoulli rewards shows that sometimes a similar recursion holds on a
reduced ‘state space’ that avoids the combinatorial explosion that typically
arises. For Gaussian rewards, even the reduced ‘state space’ was uncountably
large, and a piecewise quadratic approximation was suggested. When this
kind of approximation is used, we get an instance of approximate dynamic
programming.
4 Discounted bandits with Markov pay-offs (Fig. 35.3) are a special case of
discounted Markov decision processes on which there is a large literature. More
details are in the bibliographic remarks in Chapter 38.
5 Economists have long recognised the role of time in the utility people place on
rewards. Most people view a promise of pizza (freshly made) a year from today
as less valuable than the same pizza tomorrow. Discounting rewards is one way
to model this kind of preference. The formal model is credited to renowned
American economist Paul Samuelson [1937], who, according to Frederick et al.
[2002], had serious reservations about both the normative and descriptive value
of the model. While discounting is not very common in the frequentist bandit
literature, it appears often in reinforcement learning, where it offers certain
technical advantages [Sutton and Barto, 1998].
6 Theorem 35.9 only holds for geometric discounting. If αt−1 is replaced by α(t),
where α(·) is not an exponential, then one can construct Markov chains for
which the optimal policy is not an index policy. The intuition behind this result
is that when α(t) is not an exponential function, then the Gittins index of
an arm can change even in rounds you play a different arm, and this breaks
the interleaving argument [Berry and Fristedt, 1985, chapter 6]. The Gittins
index theorem is brittle in other ways. For example, it no longer holds in the
multiple-play setting, where the learner can choose multiple arms in each round
[Pandelis and Teneketzis, 1999].
 35.7 Bibliographical Remarks 455

7 The previous note does not apply to one-armed bandits for which the
interleaving argument is not required. Given a Markov chain (St )t and horizon
n, the undiscounted Gittins index of state s is
hP i
)
τ −1
Es t=1 r(St
gn (s) = sup .
2≤τ ≤n Es [τ − 1]
If the learner receives reward µ2 by retiring, then the Bayesian optimal policy
is to retire in the first round t when gn−t+1 (St ) ≤ µ2 . A reasonable strategy
for undiscounted k-armed bandits is to play the arm At that maximises
gn−t+1 (Si (t)). Although this strategy is not Bayesian optimal anymore, it
nevertheless performs well in practice. In the Gaussian case, it even enjoys
frequentist regret guarantees similar to UCB [Lattimore, 2016a].
8 The form of the undiscounted Gittins index was analysed asymptotically
by Burnetas and Katehakis [1997b], who showed the index behaves like the
upper confidence bound provided by KL-UCB. This should not be especially
surprising and explains the performance of the algorithm in the previous note.
The asymptotic nature of the result does not make it suitable for proving regret
guarantees, however.
9 We mentioned that computing the Bayesian optimal policy in finite horizon
bandits is computationally intractable. But this is not quite true if n is small.
For example, when n = 50 and k = 5, the dynamic program for computing
the exact Bayesian optimal policy for Bernoulli noise and Beta prior has
approximately 1011 states. A big number to be sure, but not so large that the
table cannot be stored on disk. And this is without any serious effort to exploit
symmetries. For mission-critical applications with small horizon, the benefits
of exact optimality might make the computation worth the hassle.
10 The algorithm in Section 35.5 for computing Gittins index is called Varaiya’s
algorithm. In the bibliographic remarks, we give some pointers on where to
look for more sophisticated methods. The assumption that |S| is finite is less
severe than it may appear. When the discount rate is not too close to one, then
for many problems the Gittins index can be approximated by removing states
that are not reachable from the start state before the discounting means they
becomes close to irrelevant. When the state space is infinite, there is often a
topological structure that makes a discretisation possible.

35.7 Bibliographical Remarks

The classic text on optimal stopping is by Robbins et al. [1971], while a

more modern text is by Peskir and Shiryaev [2006], which includes a proof
of Theorem 35.2 (see theorem 1.2). With a little extra work, you can also extract
the proof of Theorem 35.3 from section 1.2 of that book. We are not aware
of a reference for Theorem 35.1, but Lai [1987] has shown that for sufficiently
 35.8 Exercises 456

regular priors and noise models, the asymptotic Bayesian optimal regret is
BR∗n ∼ c log(n)2 for some constant c > 0 that depends on the prior/model
(see theorem 3 of Lai [1987]). The Bayesian approach dominated research on
bandits from 1960 to 1980, with Gittins’s result (Theorem 35.9) receiving the
most attention [Gittins, 1979]. Gittins et al. [2011] has written a whole book on
Bayesian bandits. Another book that focusses mostly on the Bayesian problem is
by Berry and Fristedt [1985]. Although it is now more than 30 years old, this book
is still a worthwhile read and presents many curious and unintuitive results about
exact Bayesian policies. The book by Presman and Sonin [1990] also considers
the Bayesian case. As compared to the other books, here the emphasis is on a
case that is more similar to partial monitoring, the subject of Chapter 37 (in the
adversarial setting). As far as we know, the earliest fully Bayesian analysis is by
Bradt et al. [1956], who studied the finite horizon Bayesian one-armed bandit
problem, essentially writing down the optimal policy using backwards induction,
as presented here in Section 35.3. More general ‘approximation results’ are shown
by Burnetas and Katehakis [2003], who show that under weak assumptions the
Bayesian optimal strategy for one-armed bandits is asymptotically approximated
by a retirement policy reminiscent of Eq. (35.4). The very specific approach
to approximating the Bayesian strategy for Gaussian one-armed bandits is by
one of the authors [Lattimore, 2016a], where a precise approximation for this
special case is also given. There are at least four proofs of Gittins’s theorem
[Gittins, 1979, Whittle, 1980, Weber, 1992, Tsitsiklis, 1994]. All are summarised
in the review by Frostig and Weiss [1999]. There is a line of work on computing
and/or approximating the Gittins index, which we cannot do justice to. The
approach presented here for finite state spaces is due to Varaiya et al. [1985], but
more sophisticated algorithms exist with better guarantees. A nice survey is by
Chakravorty and Mahajan [2014], but see also the articles by Chen and Katehakis
[1986], Kallenberg [1986], Sonin [2008], Niño-Mora [2011] and Chakravorty and
Mahajan [2013]. There is also a line of work on approximations of the Gittins
index, most of which are based on approximating the discrete time stopping
problem with continuous time and applying free boundary methods [Yao, 2006,
and references therein]. The Gittins index has been generalised to continuous
time, where the challenge is to ensure the existence of solutions to the resulting
stochastic differential equations [Karoui and Karatzas, 1994]. We mentioned
restless bandits in Chapter 31 on non-stationary bandits, but they are usually
studied in the Bayesian context [Whittle, 1988, Weber and Weiss, 1990]. The
difference is that now the Markov chains for all actions evolve regardless of the
action chosen, but the learner only gets to observe the new state for the action
they chose.

35.8 Exercises

35.1 (Bounding the Bayesian optimal regret) Prove Theorem 35.1.

 35.8 Exercises 457

Hint For the first part, you should use the existence of a policy for Bernoulli
bandits such that
 
√ k log(n)
Rn (π, ν) ≤ C min kn, ,
∆min (ν)

where C > 0 is a universal constant and ∆min (ν) is the smallest positive
suboptimality gap. Then let En be a set of bandits for which there exists a
small enough positive suboptimality gap and integrate the above bound on En
and Enc . The second part is left as a challenge, though the solution is available.
35.2 (Finite horizon optimal stopping) Prove Theorem 35.2.

Hint Prove that (Et )nt=1 is a F-adapted supermartingale and that for stopping
time τ satisfying the conditions of the theorem that (Mt )nt=1 defined by Mt = Et∧τ
is a martingale. Then apply the optional stopping theorem (Theorem 3.8).
35.3 (Infinite horizon optimal stopping) Prove Theorem 35.3.

Hint This is a technical exercise. Use theorem 1.7 of Peskir and Shiryaev [2006],
and pass to the limit using the almost-sure convergence of (Ut )t as t → ∞. You
may find the ideas in the proof of theorem 1.11 of the same book useful. Be
careful, Peskir and Shiryaev adopt the convention that stopping times are almost
surely finite, while here we permit infinite stopping times.
35.4 This exercise uses the notation and setting of Section 35.3. Suppose that
(St )nt=0 is a sequence of random elements taking values in measurable space (S, H)
and with St being Ft /H-measurable and Pν1 (Z1 , . . . , Zt ∈ · | St ) is independent
of ν. Show that Et is σ(St )-measurable, and there exists a H/B(R))-measurable
function vt : S → R such that Et = vt (St ). You may assume that (E, G) is Borel.

35.5 Prove Lemma 35.5.

35.6 (Equivalence of Gittins definitions) Prove that the definitions of

the Gittins index given in Eq. (35.8) and Eq. (35.9) are equivalent.

35.7 Prove Lemma 35.7.

Hint Find a way to apply Theorem 35.3.

35.8 Consider that the discounted bandit with Markov pay-offs described in
Example 35.8. Show that there is a one-to-one correspondence φ between the
policies for this problem and the discounted Bayesian bandit with Beta(1, 1) on
the mean reward of each arm such that the total expected discounted reward
(value) is invariant under φ.

35.9 Prove Lemma 35.10.

Hint Use the Hardy–Littlewood inequality, which for infinite sequences states
 35.8 Exercises 458

that for any real, increasing sequences (xn )∞ , (yn )∞

n=1 and any bijection
+ +
P∞ P∞ n=1
σ : N → N it holds that n=1 xn yn ≥ n=1 xn yσ(n) .
35.10 (Correctness of Varaiya’s algorithm) Prove the correctness of
Varaiya’s algorithm, as explained in Section 35.5.

35.11 In this exercise, you will implement some Bayesian (near-)optimal 1-armed
bandit algorithms.
(a) Reproduce the experimental results in Experiment 1.
(b) Implement an approximation of the optimal policy for one-armed Gaussian
bandits and compare its performance to the stopping rule τα defined below
for a variety of different choices of α > 0.
( r )
2 max{0, log(αn/t)}
τα = min t ≥ 2 : µ̂t−1 + ≤ µ2 .
t−1
36 Thompson Sampling

“As all things come to an end, even this story, a day came at last when they were in
sight of the country where Bilbo had been born and bred, where the shapes of the land
and of the trees were as well known to him as his hands and toes.” – Tolkien [1937].

Like Bilbo, as the end nears, we return to where it all began, to the first algorithm
for bandits proposed by Thompson [1933]. The idea is a simple one. Before the
game starts, the learner chooses a prior over a set of possible bandit environments.
In each round, the learner samples an environment from the posterior and
acts according to the optimal action in that environment. Thompson only gave
empirical evidence (calculated by hand) and focused on Bernoulli bandits with
two arms. Nowadays these limitations have been eliminated, and theoretical
guarantees have been proven demonstrating the approach is often close to optimal
in a wide range of settings. Perhaps more importantly, the resulting algorithms
are often quite practical both in terms of computation and empirical performance.
The idea of sampling from the posterior and playing the optimal action is called
Thompson sampling, or posterior sampling.
The exploration in Thompson sampling comes from the randomisation. If the
posterior is poorly concentrated, then the fluctuations in the samples are expected
to be large and the policy will likely explore. On the other hand, as more data
is collected, the posterior concentrates towards the true environment and the
rate of exploration decreases. We focus our attention on finite-armed stochastic
bandits and linear stochastic bandits, but Thompson sampling has been extended
to all kinds of models, as explained in the bibliographic remarks.

Randomisation is crucial for adversarial bandit algorithms and can be useful

in stochastic settings (see Chapters 23 and 32 for examples). We should
be wary, however, that injecting noise into our algorithms might come at a
cost in terms of variance. What is gained or lost by the randomisation in
Thompson sampling is still not clear, but we leave this cautionary note as a
suggestion to the reader to think about some of the costs and benefits.
 36.1 Finite-Armed Bandits 460

36.1 Finite-Armed Bandits

Recalling the notation from Section 34.5, let k > 1 and (E, B(E), Q, P ) be a
k-armed Bayesian bandit environment. The learner chooses actions (At )nt=1 and
receives rewards (Xt )nt=1 , and the posterior after t observations is a probability
kernel Q( · | ·) from ([k] ×RR)t to (E, B(E)). Denote the mean of the ith arm in
bandit ν ∈ E by µi (ν) = R xdPνi (x). In round t, Thompson sampling samples
a bandit environment νt from the posterior of Q given A1 , X1 , . . . , At−1 , Xt−1
and then chooses the arm with the largest mean (Algorithm 23). A more precise
definition is that Thompson sampling is the policy π = (πt )∞ t=1 with

πt (a | a1 , x1 , . . . , at−1 , xt−1 ) = Q(Ba | a1 , x1 , . . . , at−1 , xt−1 ) ,

where Ba = {ν ∈ E : µa (ν) = argmaxb µb (ν)} ∈ B(E), with ties in the argmax

are resolved in an arbitrary, but systematic fashion.

1: Input Bayesian bandit environment (E, B(E), Q, P )

2: for t = 1, 2, . . . , n do
3: Sample νt ∼ Q( · | A1 , X1 , . . . , At−1 , Xt−1 )
4: Choose At = argmaxi∈[k] µi (νt )
5: end for
Algorithm 23: Thompson sampling.

Thompson sampling has been analysed in both the frequentist and the Bayesian
settings. We start with the latter where the result requires almost no assumptions
on the prior. In fact, after one small observation about Thompson sampling, the
analysis is almost the same as that of UCB.

Theorem 36.1. Let (E, B(E), Q, P ) be a k-armed Bayesian bandit environment

such that for all ν ∈ E and i ∈ [k], the distribution Pνi is 1-subgaussian (after
centering) with mean in [0, 1]. Then the policy π of Thompson sampling satisfies
p
BRn (π, Q) ≤ C kn log(n) ,

where C > 0 is a universal constant.

Proof Abbreviate µi = µi (ν) and let A∗ = argmaxi∈[k] µi be the optimal arm,

which depends on ν and is a random variable. When there are ties, we use the
same tie-breaking rule as in the algorithm in the definition of A∗ . For each t ∈ [n]
and i ∈ [k], let
s !
2 log(1/δ)
Ut (i) = clip[0,1] µ̂i (t − 1) + ,
1 ∨ Ti (t − 1)

where µ̂i (t − 1) is the empirical estimate of the reward of arm i after t − 1 rounds
and we assume µ̂i (t − 1) = 0 if Ti (t − 1) = 0. Let E be the event that for all
 36.2 Frequentist Analysis 461

t ∈ [n] and i ∈ [k],

s
2 log(1/δ)
|µ̂i (t − 1) − µi | < .
1 ∨ Ti (t − 1)
In Exercise 36.2, we ask you to prove that P (E c ) ≤ 2nkδ. Let Ft =
σ(A1 , X1 , . . . , At , Xt ) be the σ-algebra generated by the interaction sequence
by the end of round t. Note that Ut (i) is Ft−1 -measurable. The Bayesian regret is
" n # " n #
X X
BRn = E (µA∗ − µAt ) = E E [µA∗ − µAt | Ft−1 ] .
t=1 t=1

The key insight (Exercise 36.3) is to notice that the definition of Thompson
sampling implies the conditional distributions of A∗ and At given Ft−1 are the
same:
P (A∗ = · | Ft−1 ) = P (At = · | Ft−1 ) a.s. (36.1)
Using the previous display,
E [µA∗ − µAt | Ft−1 ] = E [µA∗ − Ut (At ) + Ut (At ) − µAt | Ft−1 ]
= E [µA∗ − Ut (A∗ ) + Ut (At ) − µAt | Ft−1 ] (Eq. (36.1))
= E [µA∗ − Ut (A ) | Ft−1 ] + E [Ut (At ) − µAt | Ft−1 ] .
∗

Using the tower rule for expectation shows that

" n n
#
X X
BRn = E (µA∗ − Ut (A∗ )) + (Ut (At ) − µAt ) . (36.2)
t=1 t=1

On the event E c the terms inside the expectation are bounded by 2n, while on
the event E, the first sum is negative and the second is bounded by
n
X n X
X k
I {E} (Ut (At ) − µAt ) = I {E} I {At = i} (Ut (i) − µi )
t=1 t=1 i=1
s Z r
Ti (n)
8 log(1/δ) 8 log(1/δ)
k X
X n X k
≤ I {At = i} ≤ ds
i=1 t=1
1 ∨ Ti (t − 1) i=1 0 s
k p
X p
= 32Ti (n) log(1/δ) ≤ 32nk log(1/δ) .
i=1

The proof is completed by choosing δ = n−2 and the fact that P (E c ) ≤ 2nkδ.

36.2 Frequentist Analysis

Bounding the frequentist regret of Thompson sampling is more technical than the
Bayesian regret. The trouble is the frequentist regret does not have an expectation
with respect to the prior, which means that At is not conditionally distributed in
 36.2 Frequentist Analysis 462

the same way as the optimal action (which is not random). Thompson sampling
can be viewed as an instantiation of follow-the-perturbed-leader, which we already
saw in action for adversarial combinatorial semi-bandits in Chapter 30. Here we
work with the stochastic setting and consider the general form algorithm given
in Algorithm 24.

1: Input Cumulative distribution functions F1 (1), . . . , Fk (1)

2: for t = 1, . . . , n do
3: Sample θi (t) ∼ Fi (t) independently for each i
4: Choose At = argmaxi∈[k] θi (t)
5: Observe Xt and update:
Fi (t + 1) = Fi (t) for i 6= At and FAt (t + 1) = Update(FAt (t), At , Xt )
6: end for
Algorithm 24: Follow-the-perturbed-leader

Thompson sampling is recovered by choosing F1 (1), . . . , Fk (1) to be the

cumulative distribution functions of the mean reward of each arm for a prior
that is independent over the arms (a product prior). Then letting Update be
the function that updates the posterior for the played arm. There are, however,
many alternatives ways to configure this algorithm.

The core property that we use in the analysis of Algorithm 24 (to be

presented soon) is that Fi (t + 1) = Fi (t) whenever At 6= i. When Update(·)
is a Bayesian update this corresponds to choosing an independent prior on
the distribution of each arm.

Let Fis be the cumulative distribution function used for arm i in all rounds
t with Ti (t − 1) = s. This quantity is defined even if Ti (n) < s by using the
reward-stack model from Section 4.6.
Theorem 36.2. Assume that arm 1 is optimal. Let i > 1 be an action and ε ∈ R
be arbitrary. Then the expected number of times Algorithm 24 plays action i is
bounded by
"n−1  # "n−1 #
X 1 X
E[Ti (n)] ≤ 1 + E −1 +E I {Gis > 1/n} , (36.3)
s=0
G1s s=0

where Gis = 1 − Fis (µ1 − ε).

In applications, ε is normally chosen to be a small positive constant. In this case,
the first sum in Eq. (36.3) measures the probability that the sample corresponding
to the first arm is nearly optimistic and tends to be smaller when the variance
of the perturbation is larger. The second sum measures the likelihood that
the sample from arm i is close to µ1 and is small when the variance of the
 36.2 Frequentist Analysis 463

perturbation is small. Balancing these two terms corresponds to optimising the

exploration/exploitation trade-off.

Proof of Theorem 36.2 Let Ft = σ(A1 , X1 , . . . , At , Xt ) and Ei (t) = {θi (t) ≤

µ1 − ε}. By definition,

P (θ1 (t) ≥ µ1 − ε | Ft−1 ) = G1T1 (t−1) a.s.

We start with a straightforward decomposition:

" n #
X
E[Ti (n)] = E I {At = i}
t=1
" n # " n
#
X X
=E I {At = i, Ei (t)} + E I {At = i, Eic (t)} . (36.4)
t=1 t=1

In order to bound the first term, let A0t = argmaxi6=1 θi (t). Then

P (At = 1, Ei (t) | Ft−1 ) ≥ P (A0t = i, Ei (t), θ1 (t) ≥ µ1 − ε | Ft−1 )

= P (θ1 (t) ≥ µ1 − ε | Ft−1 ) P (A0t = i, Ei (t) | Ft−1 )
G1T1 (t−1)
≥ P (At = i, Ei (t) | Ft−1 ) , (36.5)
1 − G1T1 (t−1)
where in the first equality we used the fact that θ1 (t) is conditionally independent
of A0t and Ei (t) given Ft−1 . In the second inequality, we used the definition of
G1s and the fact that

P (At = i, Ei (t) | Ft−1 ) ≤ (1 − P (θ1 (t) > µ1 − ε | Ft−1 ))P (A0t = i, Ei (t) | Ft−1 ) ,

which is true since {At = i, Ei (t) occurs} ⊆ {A0t = i, Ei (t) occurs} ∩ {θ1 (t) ≤
µ1 − ε}, and the two intersected events are conditionally independent given Ft−1 .
Therefore using Eq. (36.5), we have
 
1
P (At = i, Ei (t) | Ft−1 ) ≤ − 1 P (At = 1, Ei (t) | Ft−1 )
G1T1 (t−1)
 
1
≤ − 1 P (At = 1 | Ft−1 ) .
G1T1 (t−1)
Substituting this into the first term in Eq. (36.4) leads to
" n # " n   #
X X 1
E I {At = i, Ei (t) occurs} ≤ E − 1 P (At = 1 | Ft−1 )
t=1 t=1
G1T1 (t−1)
" n   #
X 1
=E − 1 I {At = 1}
t=1
G1T1 (t−1)
"n−1  #
X 1
≤E −1 , (36.6)
s=0
G1s

where in the last step we used the fact that T1 (t − 1) = s is only possible for one
 36.2 Frequentist Analysis 464

round where At = 1. Let T = {t ∈ [n] : 1 − FiTi (t−1) (µ1 − ε) > 1/n}. After some
calculation (Exercise 36.5), we get
" n # " # " #
X X X
E I {At = i, Eic (t) occurs} ≤ E I {At = i} + E I {Eic (t)}
t=1 t∈T t∈T
/
"n−1 # " #
X X1
≤E I {1 − Fis (µ1 − ε) > 1/n} + E
s=0
n
t∈T
/
"n−1 #
X
≤E I {Gis > 1/n} + 1 .
s=0

Putting together the pieces completes the proof.

By instantiating Algorithm 24 with different choices of perturbations, one

can prove that Thompson sampling enjoys frequentist guarantees in a number
of settings. The following theorem shows that Thompson sampling with an
appropriate prior is asymptotically optimal for the set of Gaussian bandits.
The reader is invited to prove this result by following the steps suggested in
Exercise 36.6.

Theorem 36.3. Suppose that Fi (1) = δ∞ is the Dirac at infinity and let
Update(Fi (t), At , Xt ) be the cumulative distribution function of the Gaussian
N (µ̂i (t), 1/t). Then the regret of Algorithm 24 on Gaussian bandit ν ∈ EN k
(1)
satisfies
Rn X 2
lim = .
n→∞ log(n) ∆i
i:∆i >0

Furthermore,
p there exists a universal constant C > 0 such that Rn ≤
C nk log(n).

The choice of update and initial distributions in Theorem 36.3 correspond to

Thompson sampling when the prior mean and variance are sent to infinity at
appropriate rates. For this choice, a finite-time analysis is also possible (see the
exercise).
Experiment 36.1 Empirically the algorithm described in Theorem 36.3 has a
smaller expected regret than the version of UCB analysed in Chapter 7. Compared
to more sophisticated algorithms, however, it has larger regret and larger variance.
AdaUCB (which we briefly met in Section 9.3) and Thompson sampling were
simulated on a two-armed Gaussian bandit with mean vector µ = (1/5, 0) and
unit variance and a horizon of n = 2000. The expected regret as estimated
over 100,000 independent runs was 23.8 for AdaUCB and 29.9 for Thompson
sampling. The figure below shows that contribution of the second moment of
P
R̂n = i ∆i Ti (n) for each algorithm, which shows that Thompson sampling has
a much larger variance than AdaUCB, despite its inferior expected regret.
 36.3 Linear Bandits 465

Proportion × Regret2
300 Thompson sampling 300 AdaUCB

200 200

100 100

0 0
0 100 200 300 400 0 100 200 300 400
Regret Regret

36.3 Linear Bandits

While the advantages of Thompson sampling in finite-armed bandits are relatively

limited, in the linear setting there is much to be gained, both in terms of
computation and empirical performance. Let A ⊂ Rd and (E, B(E), Q, P ) be a
Bayesian bandit environment where E ⊂ Rd and for θ ∈ E and a ∈ A, Pθa is
1-subgaussian with mean hθ, ai. Let θ : E → Rd be the identity map, which is a
random vector on (E, B(E), Q).

1: Input Bayesian bandit environment (E, B(E), Q, P )

2: for t ∈ 1, . . . , n do
3: Sample θt from the posterior
4: Choose At = argmaxa∈A ha, θt i
5: Observe Xt
6: end for
Algorithm 25: Thompson sampling for linear bandits.

The Bayesian regret is controlled using the techniques from the previous section
in combination with the concentration analysis in Chapter 20. A frequentist
analysis is also possible under slightly unsatisfying assumptions, which we discuss
in the notes and bibliographic remarks.
Theorem 36.4. Assume that kθk2 ≤ S with Q-probability one and supa∈A kak2 ≤
L and supa∈A |ha, θi| ≤ 1 with Q-probability one. Then the Bayesian regret of
Algorithm 25 is bounded by
s  
nS 2 L2
BRn ≤ 2 + 2 2dnβ 2 log 1 + ,
d
s  
nS 2 L2
where β = 1 + 2 log(n) + d log 1 + .
d
For fixed S and L, the upper bound obtained here is of order
 36.3 Linear Bandits 466

p
O(d n log(n) log(n/d)), which matches the upper bound obtained for Lin-UCB
in Corollary 19.3.
Proof We apply the same technique as used in the proof of Theorem 36.1. Define
the upper confidence bound function Ut : A → R by
1 X t
Ut (a) = ha, θ̂t−1 i + βkakV −1 , where Vt = 2 I + As A>
s .
t−1 S s=1

By Theorem 20.5 and Eq. (20.9), P(exists t ≤ n : kθ̂t−1 − θkVt−1 > β) ≤ 1/n. Let
Tn
Et be the event that kθ̂t−1 −θkVt−1 ≤ β, E = t=1 Et and A∗ = argmaxa∈A ha, θi.
Note that A∗ is a random variable because θ is random. Then
" n #
X
BRn = E ∗
hA − At , θi
t=1
" n
# " n
#
X X
= E IE c hA∗ − At , θi + E IE hA∗ − At , θi
t=1 t=1
"n
#
X
≤ 2 + E IE ∗
hA − At , θi
t=1
" n #
X
≤2+E IEt hA∗ − At , θi . (36.7)
t=1

Let Ft = σ(A1 , X1 , . . . , At , Xt ) and let Et−1 [·] = E[· | Ft−1 ]. As before,

P (A∗ = · | Ft−1 ) = P (At = · | Ft−1 ), and Ut (a), for any fixed a ∈ A, is Ft−1 -
measurable, and so Et−1 (Ut (A∗ )) = Et−1 (Ut (At )). It follows that the second term
in the above display is bounded by
Et−1 [IEt hA∗ − At , θi] = IEt Et−1 [hA∗ , θi − Ut (A∗ ) + Ut (At ) − hAt , θi]
≤ IEt Et−1 [Ut (At ) − hAt , θi]
h i
≤ IEt Et−1 hAt , θ̂t−1 − θi + βkAt kV −1
t
h i
≤ IEt Et−1 kAt kV −1 kθ̂t−1 − θkVt + βkAt kV −1
t t

≤ 2βkAt kV −1 .
t

Substituting this combined with IEt hA∗ − At , θi ≤ 2 into the second term of
Eq. (36.7), we get
" n # " n #
X X
E IEt hA − At , θi ≤ 2βE
∗
(1 ∧ kAt kV −1 )
t
t=1 t=1
v " n #
u
u X
≤ 2 nβ E
t 2 2
(1 ∧ kAt kV −1 ) (Cauchy-Schwarz)
t
t=1
s   
nS 2 L2
≤2 2dnβ 2 E log 1 + . (Lemma 19.4)
d
 36.4 Information Theoretic Analysis 467

Putting together the pieces shows that

s  
nS 2 L2
BRn ≤ 2 + 2 2dnβ 2 log 1 + .
d

36.3.1 Computation
An implementation of Thompson sampling for linear bandits needs to sample θt
from the posterior and then find the optimal action for the sampled parameter:

At = argmaxa∈A ha, θt i .

For some priors and noise models, sampling from the posterior is straightforward.
The most notable case is when Q is a multivariate Gaussian and the noise is
Gaussian with a known variance. More generally, there is a large literature devoted
to numerical methods for sampling from posterior distributions. Having sampled
θt , finding At is a linear optimisation problem. By comparison, LinUCB needs to
solve

At = argmaxa∈A maxha, θ̃i ,

θ̃∈C

which for large or continuous action sets is often intractable.

36.4 Information Theoretic Analysis

We now examine a Bayesian version of the adversarial k-armed bandit. As we

will see, the natural generalisation of Thompson sampling is still a reasonable
algorithm. Recall that in the adversarial bandit model studied in Part III, the
adversary secretly chooses a matrix x ∈ [0, 1]n×k at the start of the game and
the reward in round t is xtAt . In the Bayesian set-up, there is a prior probability
measure Q on [0, 1]n×k with the Borel σ-algebra. At the start of the game, a
reward matrix X is sampled from Q, but not revealed. The learner then chooses
actions (At )nt=1 and the reward in round t is XtAt . Formally, let π = (πt )nt=1 be a
policy and X ∈ [0, 1]n×k and (At )nt=1 be random elements on some probability
space (Ω, F, P) such that:

(a) the law of X is PX = Q; and

(b) P(At ∈ · | X, Ht−1 ) = πt ( · | Ht−1 ) with Ht = (A1 , X1A1 , . . . , At , XtAt ).

The inclusion of X in the conditional expectation in Part (b) implies that

P(At ∈ · | X, Ht−1 ) = P(At ∈ · | Ht−1 ) ,

which means that At and X are conditionally independent given Ht−1 . This is
consistent with our definition of the model where X is sampled first from Q and
 36.4 Information Theoretic Analysis 468

then At depends on X only through the history Ht−1 . The optimal action is
Pn
A∗ = argmaxa∈[k] t=1 Xta with ties broken arbitrarily. The Bayesian regret is
" n #
X
BRn = E (XtA∗ − XtAt ) .
t=1

Like in the previous sections, Thompson sampling is a policy π = (πt )nt=1 that
plays each action according to the conditional probability that it is optimal, which
means the following holds almost surely:

πt ( · | A1 , X1A1 , . . . , At−1 , Xt−1At−1 ) = P(A∗ ∈ · | A1 , X1A1 , . . . , At−1 , Xt−1At−1 ) .

The main result of this section is the following theorem:

Theorem 36.5. The Bayesian regret of Thompson sampling for Bayesian k-armed
adversarial bandits satisfies
p
BRn ≤ kn log(k)/2 .

The proof is done through a generic theorem that is powerful enough to analyse
a wide range of settings. For stating this result, we need some preparation. Let
Ft = σ(A1 , X1A1 , . . . , At , XtAt ) and Et [·] = E[· | Ft ] and Pt (·) = P( · | Ft ). Let
∆t = XtA∗ − XtAt denote the immediate regret of round t.
The promised generic theorem bounds the regret in terms of an ‘information
ratio’ that depends on the ratio of the squared expected instantaneous regret
conditioned on the past and a Bregman divergence with respect to some convex
function F to be chosen later.

Theorem 36.6. Let F : Rk → R ∪ {∞} be convex, and suppose there exists a

constant β ≥ 0 such that
p
Et−1 [∆t ] ≤ βEt−1 [DF (Pt (A∗ = ·), Pt−1 (A∗ = ·))] a.s.
p
Then BRn ≤ nβdiamF (Pk−1 ).

Proof Let Mt = Pt (A∗ = ·) ∈ Pk−1 . Using the directional derivative definition

of the Bregman divergence combined with Fatou’s lemma and convexity of F ,
 
Et−1 [DF (Mt , Mt−1 )] = Et−1 F (Mt ) − F (Mt−1 ) − ∇Mt −Mt−1 F (Mt−1 )
  
F ((1 − h)Mt−1 + hMt ) − F (Mt−1 )
= Et−1 lim inf F (Mt ) − F (Mt−1 ) −
h→0+ h
  
F ((1 − h)Mt−1 + hMt ) − F (Mt−1 )
≤ lim inf Et−1 F (Mt ) − F (Mt−1 ) −
h→0+ h
F (Mt−1 ) − Et−1 [F ((1 − h)Mt−1 + hMt )]
= Et−1 [F (Mt )] − F (Mt−1 ) + lim inf
h→0+ h
F (Mt−1 ) − F (Et−1 [(1 − h)Mt−1 + hMt ])
≤ Et−1 [F (Mt )] − F (Mt−1 ) + lim inf
h→0+ h
= Et−1 [F (Mt )] − F (Mt−1 ) , (36.8)
 36.4 Information Theoretic Analysis 469

where the first inequality follows from Fatou’s lemma and the second from the
convexity of F . The last equality is because Et−1 [Mt ] = Mt−1 . Hence,
" n # " n #
X Xp
BRn = E ∆t ≤ E βEt−1 [DF (Mt , Mt−1 )]
t=1 t=1
v " n #
u
u X p
t
≤ βnE Et−1 [DF (Mt , Mt−1 )] ≤ βndiamF (Pk−1 ) ,
t=1

where the first inequality follows from the assumption in the theorem, the second
by Cauchy–Schwarz, while the third follows by Eq. (36.8), telescoping and the
definition of the diameter.

It remains to choose F and show that the condition of the previous result can
be met. As you might have guessed, a good choice is the unnormalised negentropy
Pk
potential F (p) = a=1 pa log(pa ) − pa . Remember that in this case the resulting
Bregman divergence DF (p, q) is the relative entropy, D(p, q), between categorical
distributions parameterised by p and q, respectively.

Lemma 36.7. If Xti ∈ [0, 1] almost surely for all t ∈ [n] and i ∈ [k] and At is
chosen by Thompson sampling using any prior, then
r
k
Et−1 [∆t ] ≤ Et−1 [D(Pt (A∗ ∈ ·), Pt−1 (A∗ ∈ ·))] .
2
Proof Given a measure P, we write PX|Y ( · ) for P(X ∈ · | Y ). In our application
below, X is a random variable, and hence P(X ∈ · | Y ) can be chosen to be a
probability measure by Theorem 3.11. When Y is discrete, we write PX|Y =y (·)
for P(X ∈ · | Y = y). The result follows by chaining Pinsker’s inequality and
Cauchy–Schwarz:
k
X
Et−1 [∆t ] = Pt−1 (At = a) (Et−1 [Xta | A∗ = a] − Et−1 [Xta ])
a=1
r
1
k
X
≤ Pt−1 (At = a) D(Pt−1,Xta |A∗ =a , Pt−1,Xta )
a=1
2
v
u k
uk X
≤t Pt−1 (At = a)2 D(Pt−1,Xta |A∗ =a , Pt−1,Xta )
2 a=1
v
u k k
uk X X
≤t Pt−1 (At = a) Pt−1 (A∗ = a) D(Pt−1,Xta |A∗ =b , Pt−1,Xta )
2 a=1
b=1
r
k
= Et−1 [D(Pt (A∗ ∈ ·), Pt−1 (A∗ ∈ ·))] ,
2
where the final equality follows from Bayes’ law and is left as an exercise.
 36.5 Notes 470

Proof of Theorem 36.5 The result follows by combining Lemma 36.7, Theo-
rem 36.6 and the fact that the diameter of the unnormalised negentropy potential
is diamF (Pk−1 ) = log(k).

The reason for the name ‘information-theoretic’ is that historically

Theorem 36.6 was specified to the unnormalised negentropy when the
expected Bregman divergence is called the information gain or mutual
information. In this, sense Theorem 36.6 shows that the Bayesian regret is
well controlled if Et−1 [∆t ] can be bounded in terms of the information gain
about the optimal action, which seems rather natural. Other potentials can
be useful, however, as you will show in Exercise 36.10.

36.5 Notes

1 There are several equivalent ways to view Thompson sampling in stationary

stochastic multi-armed bandits: (a) select an arm according to the posterior
probability that the arm is optimal, or (b) sample an environment from the
posterior and play the optimal action in that environment. When the mean
rewards for each arm are independent under the posterior, then also equivalent
is (c) sample the mean reward for each arm and choose the arm with the largest
mean (Exercise 36.1). The algorithms in this chapter are based on (b), but all
are equivalent and simply correspond to sampling from different push-forward
measures of the posterior. Historically it seems that Thompson [1933] had
the form in (a) in mind, but there are reasons to remember the alternative
views. Though we are not aware of an example, in some instances beyond
finite-armed bandits, it might be more computationally efficient to sample
from a push-forward of the posterior than the posterior itself. Furthermore, in
more complicated situations like reinforcement learning, it may be desirable
to ‘approximate’ Thompson sampling, and approximating a sample from each
of the above three choices may lead to different algorithms. It is also good to
keep in mind that in the non-Bayesian setting there can be cheaper ways of
inducing sufficient exploration than sampling from a posterior, especially in
the context of structured bandit problems.
2 Thompson sampling is known to be asymptotically optimal in a variety of
settings – most notably, when the noise model follows a single-parameter
exponential family and the prior is chosen appropriately [Kaufmann et al.,
2012b, Korda et al., 2013]. Unfortunately, Thompson sampling is not a silver
bullet. The linear variant in Section 36.3 is not asymptotically optimal by the
same argument we presented for optimism in Chapter 25. Characterising the
conditions under which Thompson sampling is close to optimal remains an
open challenge.
 36.5 Notes 471

3 For the Gaussian noise model, it is known that Thompsonp sampling is not
minimax optimal. Its worst-case regret is Rn = Θ( nk log(k)) [Agrawal and
Goyal, 2013a].
4 An alternative to sampling from the posterior is to choose in each round
the arm that maximises a Bayesian upper confidence bound, which is a
quantile of the posterior. The resulting algorithm is called BayesUCB and
has excellent empirical and theoretical guarantees [Kaufmann et al., 2012a,
Kaufmann, 2018].
5 The prior has a significant effect on the performance of Thompson sampling.
In classical Bayesian statistics, a poorly chosen prior is quickly washed away by
data. This is not true in (stochastic, non-Bayesian) bandits because if the prior
underestimates the quality of an arm, then Thompson sampling may never
play that arm with high probability and no data is ever observed. We ask you
to explore this situation in Exercise 36.16.
6 An instantiation of Thompson sampling for stochastic contextual linear bandits
is known to enjoy near-optimal frequentist regret. In each round the algorithm
samples θt ∼ N (θ̂t−1 , rVt−1
−1
), where r = Θ(d) is a constant and
t
X t
X
Vt = I + As A>
s and θ̂t = Vt−1 Xs As .
s=1 s=1

Then At = argmaxa∈At hθt , ai. This corresponds to assuming the noise is

Gaussian with variance r and choosing prior Q = N (0, I). Provided the
rewards are conditionally 1-subgaussian, the frequentist regret of this algorithm
√
√
is Rn = Õ(d3/2 n), which is worse than LinUCB by a factor of d. The
increased regret is caused by the choice of noise model, which assumes the
variance is r = Θ(d) rather than r = 1. The reason to do this comes from the
analysis, which works by showing the algorithm is ‘optimistic’ with reasonable
probability. Very recently, an example was constructed showing that the blowup
of the variance is necessary. The most recent version of this result establishes
this for an action set with three (fixed) actions [Hamidi and Bayati, 2020].
Empirically, r = 1 often leads strong performance on many instances, though
clearly, as shown by the results of Hamidi and Bayati [2020], this depends on
what instances are used.
7 The analysis in Section 36.4 can be generalised to structured settings such as
linear bandits [Russo and Van Roy, 2016]. For linear bandits with an infinite
action set, the entropy of the optimal action may be infinite. The analysis
can be corrected in this case by discretising the action set and comparing to
a near-optimal action. This leads to a trade-off between the fineness of the
discretisation and its size, and when the trade-off ispresolved in an optimal
fashion, one obtains an upper bound of order O(d n log(1 + n/d)) on the
Bayesian regret, slightly improving previous analysis. The reader is referred to
the recent article by Dong and Van Roy [2018] for this analysis.
8 The information-theoretic ideas in Section 36.4 suggest that rather than
 36.5 Notes 472

sampling At from the posterior on A∗ , one can sample At from the distribution
Pt given by
Pk
a=1 pa (Et−1 [Xta | A = a] − Et−1 [Xta ])
∗
Pt = argminp∈Pk−1 Pk .
a=1 pa Et−1 [DF (Pt−1 (A = · | Xta ), Pt−1 (A = ·))]
∗ ∗

When F is the unnormalised negentropy, the resulting policy is called

information-directed sampling. Bayesian regret analysis for this algorithm
follows along similar lines to what was presented in Section 36.4. See
Exercise 36.9 or the paper by Russo and Van Roy [2014a] for more details.
9 The proof of Theorem 36.6 only used the fact that Mt = Pt (A∗ = ·) is a
martingale. The posterior is just one possible choice, but in some cases an
alternative martingale leads to improved bounds.
Pk √
10 Replacing the unnormalised negentropy
√ potential with F (p) = −2 i=1 pi
leads to a bound of BRn ≤ 2nk for any prior for finite-armed bandits
[Lattimore and Szepesvári, 2019c]. You will prove this in Exercise 36.10.
The same potential also led to minimax bounds for adversarial bandits in
Exercise 28.15, which suggests there is some kind of connection. This was
explored by Zimmert and Lattimore [2019], who show that the same techniques
used to bound the dual norm ‘stability’ terms in the analysis of mirror descent
also control the information ratio for a version of Thompson sampling.
11 Let E = [0, 1]n×k be the set of all adversarial bandits and Π the set of all
randomised policies and Q be the set of all finitely supported distributions on
E, which means that Q ∈ Q is a function Q : E → [0, 1] with Supp(Q) = {x :
P
Q(x) > 0} a finite set and x∈Supp(Q) Q(x) = 1. Given x ∈ E and π ∈ Π, let
Exπ be the expectation with respect to the interaction between policy π and
environment x. Then,
" n
#
X
Rn (E) = min sup Exπ max
∗
(xti − xtAt )
π∈Π x∈E i∈[k]
t=1
| {z }
Adversarial regret
" n
#
X X
= sup min Q(x)Exπ max (xti − xtAt ) (36.9)
Q∈Q π∈Π i∈[k]
t=1
x∈Supp(Q)
| {z }
Bayesian optimal regret
p
≤ nk log(k)/2 ,

where the second equality follows from Sion’s minimax theorem (Exercise 36.11)
and the inequality follows from Theorem 36.5. This bound is a factor
√ of two
better than what we gave in Theorem 11.2 and can be improved to 2nk using
the argument from the previous note and Exercise 36.10. The approach has
been used in more sophisticated settings, like the first near-optimal analysis
for adversarial convex bandits [Bubeck et al., 2015a, Bubeck and Eldan, 2016]
or partial monitoring [Lattimore and Szepesvári, 2019c]. As noted earlier, the
 36.6 Bibliographic Remarks 473

main disadvantage is that the technique does not lead to algorithms for the
adversarial setting.

36.6 Bibliographic Remarks

Thompson sampling has the honor of being the first bandit algorithm and
is named after its inventor [Thompson, 1933], who considered the Bernoulli
case with two arms. Thompson provided no theoretical guarantees, but argued
intuitively and gave hand-calculated empirical analysis. It would be wrong to
say that Thompson sampling was entirely ignored for the next eight decades,
but it was definitely not popular until recently, when a large number of authors
independently rediscovered the article/algorithm [Graepel et al., 2010, Granmo,
2010, Ortega and Braun, 2010, Chapelle and Li, 2011, May et al., 2012]. The
surge in interest was mostly empirical, but theoreticians followed soon with regret
guarantees. For the frequentist analysis, we followed the proofs by Agrawal and
Goyal [2012, 2013a], but the setting is slightly different. We presented results for
the ‘realisable’ case where the pay-off distributions are actually Gaussian, while
Agrawal and Goyal use the same algorithm but prove bounds for rewards bounded
in [0, 1]. Agrawal and Goyal [2013a] also analyse the Beta/Bernoulli variant of
Thompson sampling, which for rewards in [0, 1] is asymptotically optimal in
the same way as KL-UCB (see Chapter 10). This result was simultaneously
obtained by Kaufmann et al. [2012b], who later showed that for appropriate
priors, asymptotic optimality also holds for single-parameter exponential families
[Korda et al., 2013]. For Gaussian bandits with unknown mean and variance,
Thompson sampling is asymptotically optimal for some priors, but not others –
even quite natural ones [Honda and Takemura, 2014]. The Bayesian analysis of
Thompson sampling based on confidence intervals is due to Russo and Van Roy
[2014b]. Recently the idea has been applied to a wide range of bandit settings
[Kawale et al., 2015, Agrawal et al., 2017] and reinforcement learning [Osband
et al., 2013, Gopalan and Mannor, 2015, Leike et al., 2016, Kim, 2017]. The
BayesUCB algorithm is due to Kaufmann et al. [2012a], with improved analysis
and results by Kaufmann [2018]. The frequentist analysis of Thompson sampling
for linear bandits is by Agrawal and Goyal [2013b], with refined analysis by
Abeille and Lazaric [2017a] and a spectral version by Kocák et al. [2014]. A recent
paper analyses the combinatorial semi-bandit setting [Wang and Chen, 2018].
The information-theoretic analysis is by Russo and Van Roy [2014a, 2016], while
the generalising beyond the negentropy potential is by Lattimore and Szepesvári
[2019c]. As we mentioned, these ideas have been applied to convex bandits [Bubeck
et al., 2015a, Bubeck and Eldan, 2016] and also to partial monitoring [Lattimore
and Szepesvári, 2019c]. There is a tutorial on Thompson sampling by Russo
et al. [2018] that focuses mostly on applications and computational issues. We
mentioned there are other ways to configure Algorithm 24, for example the recent
article by Kveton et al. [2019].
 36.7 Exercises 474

36.7 Exercises

36.1 (Equivalent views) Prove the claimed equivalences in Note 1.

36.2 (Filling in steps in the proof of Theorem 36.1 (i)) Consider the
event E defined in Theorem 36.1, and prove that P (E c ) ≤ 2nkδ.

36.3 (Filling in steps in the proof of Theorem 36.1 (ii)) Prove Eq. (36.1).

36.4 (Removing logarithmic

√ factors) Improve the bound in Theorem 36.1
to show that BRn ≤ C kn where C > 0 is a universal constant.

Hint Replace the naive confidence intervals used in the proof of Theorem 36.1
by the more refined confidence bounds used in Chapter 9. The source for this
result is the paper by Bubeck and Liu [2013].
36.5 (Filling in steps in the proof of Theorem 36.2) Let Gi (s) =
1 − Fis (µ1 − ε). Show that
X n
X
(a) I {At = i} ≤ I {Gi (s − 1) > 1/n}; and
t∈T s=1
" # " #
X X
(b) E I {Eic (t)} ≤E 1/n .
t∈T
/ t∈T
/

36.6 (Frequentist bound for Thompson sampling) In this exercise you

will prove Theorem 36.3.

(a) Show that there exists a universal constant c > 0 such that
"∞  #  
X 1 c 1
E −1 ≤ 2 log .
t=1
G 1s (ε) ε ε

(b) Show that

" #
2 log(n)
n
X
E I {Gis > 1/n} ≤ + o(log(n)) .
s=1
(∆i − ε)2

(c) Use Theorem 36.2 and the fundamental regret decomposition (Lemma 4.5)
to prove Theorem 36.3.

Hint For (a) you may find it useful to know that for y ≥ 0,
r
2 exp(−y 2 /2)
1 − Φ(y) ≥ p ,
π y + y2 + 4
Ry
where Φ(y) = √12π −∞ exp(−x2 /2)dx is the cumulative distribution function of
the standard Gaussian [Abramowitz and Stegun, 1964, §7.1.13].
36.7 Prove the final equality in the proof of Lemma 36.7.
 36.7 Exercises 475

36.8 (Prediction with expert advice) Consider the adversarial Bayesian

framework from Section 36.4, but assume the learner observes the whole vector
Xt rather than just XtAt , which corresponds to the prediction with expert advice
setting. Prove that Thompson sampling in this setting has a Bayesian regret of
at most
p
BRn ≤ n log(k)/2 .

36.9 (Information-directed sampling) Prove that for any prior such that
Xti ∈ [0, 1] almost surely, the Bayesian regret of information-directed sampling
(see Note 8) satisfies
p
BRn ≤ kn log(k)/2 .

36.10 (Minimax Bayesian regret for Thompson sampling) Prove that for
any prior over adversarial k-armed bandits such that Xti ∈√ [0, 1] almost surely,
the Bayesian regret of Thompson sampling satisfies BRn ≤ 2kn.
Pk √
Hint Use the potential F (p) = −2 i=1 pi and the fact that the total
variation distance is upper-bounded by the Hellinger distance.
36.11 (From Bayesian to adversarial regret) Let E = {0, 1}n×k and Q
be the space of probability measures on E. Prove that

Rn∗ (E) = sup BR∗n (Q) .

Q∈Q

Hint Repeat the argument in the solution to Exercise 34.16, noting that Q is
finite dimensional. Take care to adapt the result in Exercise 4.4 to the adversarial
setting.
36.12 (From Bayesian to adversarial regret) Let E = [0, 1]n×k . Prove
that

Rn∗ (E) = sup BR∗n (Q) ,

Q∈Q

where Q is the set of probability measures on (E, 2E ) with finite support.

Hint That E is uncountably large introduces some challenges. Like in the

previous exercise, the idea is to express the regret of a policy as an integral
over the regret of deterministic policies, which can be viewed as functions
π : ∪nt=1 [0, 1]t−1 → [k]. Use Tychonoff’s theorem to argue that the space of
all deterministic policies is compact with respect to the product topology. Then
the space of regular probability measures over deterministic policies is compact
with the weak* topology by Theorem 2.14. Then carefully check continuity
and linearity of the Bayesian regret, and apply Sion’s theorem. Details are by
Lattimore and Szepesvári [2019c].
36.13 (Binary is the worst case) Prove that Rn∗ ({0, 1}n×k ) = Rn∗ ([0, 1]n×k ).
 36.7 Exercises 476

Hint Think about how to use a minimax optimal policy for {0, 1}n×k for
bandits in [0, 1]n×k .
36.14 (Implementation (i)) In this exercise, you will reproduce the results in
Experiment 1.
(a) Implement Thompson sampling as described in Theorem 36.3 as well as
UCB and AdaUCB.
(b) Reproduce the figures in Experiment 1 as well as UCB.
(c) How consistent are these results across different bandits? Run a few
experiments and report the results.
(d) Explain your findings. Which algorithm do you prefer and why.

36.15 (Implementation (ii)) Implement linear Thompson sampling with a

Gaussian prior as defined in Note 6 as well as LinUCB from Chapter 19 and
Algorithm 12. Compare these algorithms in a variety of regimes, report your
results, and tell an interesting story. Discuss the pros and cons of different choices
of r.

36.16 (Misspecified prior) Fix a Gaussian bandit with unit variance and
mean vector µ = (0, 1/10) and horizon n = 1000. Now consider Thompson
sampling with a Gaussian model with known unit covariance and a prior on the
unknown mean of each arm given by a Gaussian distribution with mean µP and
covariance σP2 I.
(a) Let the prior mean be µP = (0, 0), and plot the regret of Thompson sampling
as a function of the prior variance σP2 .
(b) Repeat the above with µP = (0, 1/10) and (0, −1/10) and (2/10, 1/10).
(c) Explain your results.
Part VIII
Beyond Bandits
37 Partial Monitoring

While in a bandit problem, the feedback that the learner receives from the
environment is the loss of the chosen action, in partial monitoring the coupling
between the loss of the action and the feedback received by the learner is loosened.
Consider the problem of learning to match
pennies when feedback is costly. Let c > 0 be
a known constant. At the start of the game, the
adversary secretly chooses a sequence i1 , . . . , in ∈
{heads, tails}. In each round, the learner chooses
an action At ∈ {heads, tails, uncertain}. The
loss for choosing action a in round t is
 Figure 37.1 Spam filtering
0 , if a = it ;

 is a potential application of
yta = c , if a = uncertain ; partial monitoring. The turtle

 (called Spam) was inherited by
1 , otherwise .
one of the authors.
So far this looks like a bandit problem. The difference is that the learner
never directly observes ytAt . Instead, the learner observes nothing unless
At = uncertain, in which case they observe the value of it . As usual, the
goal is to minimise the (expected) regret, which is
" n #
X
Rn = max E (ytAt − yta ) .
a∈[k]
t=1

How should a learner act in problems like this, where the loss is not directly
observed? Can we find a policy with sublinear regret? In this chapter we give
a more or less complete answer to these questions for finite adversarial partial
monitoring games, which include the above problems as a special case.

Matching pennies with costly feedback seems like an esoteric problem. But
think about adding contextual information and replace the pennies with
emails to be classified as spam or otherwise. The true label is only accessible
by asking a human, which replaces the third action. While the chapter does
not cover the contextual version, some pointers to the literature are added
at the end.
 37.1 Finite Adversarial Partial Monitoring Problems 479

37.1 Finite Adversarial Partial Monitoring Problems

A finite, k-action, d-outcome adversarial partial monitoring problem is

specified by a loss matrix L ∈ Rk×d and a feedback matrix Φ ∈ Σk×d , where
Σ is called the set of signals. We let m be the maximum number of distinct
symbols in any row of Φ. At the beginning of the game, the learner is given L and
Φ, and the environment secretly chooses n outcomes i1 , . . . , in with it ∈ [d]. The
loss of action a ∈ [k] in round t is yta = Lait . In each round t, the learner chooses
At ∈ [k] and receives feedback σt = ΦAt it . Given a partial monitoring problem
G = (Φ, L), the regret of policy π when the adversary chooses i1:n = (it )nt=1 is
" n #
X
Rn (π, i1:n , G) = max E (ytAt − yta ) .
a∈[k]
t=1

We omit the arguments of Rn when they can be inferred from the context.

To reduce clutter, we slightly abuse notation by using (ei ) to denote the

standard basis vectors of Euclidean spaces of potentially different dimensions.

37.1.1 Examples
The partial monitoring framework is rich enough to model a wide variety of
problems, a few of which are illustrated in the examples that follow. Many of the
examples are quite artificial and are included only to highlight the flexibility of
the framework and challenges of making the regret small.

Example 37.1 (Hopeless problem). Some partial monitoring problems are

completely hopeless in the sense that one cannot expect to make the regret
small. A simple example occurs when k = d = 2, m = 1 and
   
0 1 ⊥ ⊥
L= , Φ= . (37.1)
1 0 ⊥ ⊥

Note that rows/columns correspond to choices of the learner/adversary,

respectively. In both rows, the feedback matrix has identical entries for both
columns. As the learner has no way of distinguishing between different sequences
of outcomes, there is no way to learn and avoid linear regret. The reader is
encouraged to think of generalisations of this example where the game is still
hopeless.
 37.1 Finite Adversarial Partial Monitoring Problems 480

Two feedback matrices Φ ∈ Σk×d and Φ̃ ∈ Σ̃k×d encode the same information
if the pattern of identical entries in each row match. For example,
   
 ⊥ 4 ⊥   ♦ ♠ ♦ 
   
Φ = 1 2 2
  and Φ̃ = ♣ ♥ ♥


   
3 1 1 ♣ ♥ ♥

both encode the same information. Note that for these matrices m = 2 since
in any row there are at most two distinct symbols.

Example 37.2 (Trivial problem). Just as there are hopeless problems, there are
also trivial problems. This happens when one action dominates all others as in
the following problem:
   
 0 0 ⊥ ⊥ 
L= , Φ= .
1 1 ⊥ ⊥

In this game the learner can safely ignore the second action and suffer zero regret,
regardless of the choices of the adversary.

Example 37.3 (Matching pennies). The penny-matching problem mentioned in

the introduction has k = 3 actions d = 2 outcomes and is described by
   
 0 1   ⊥ ⊥
   
L = 1 0 ,
  Φ = ⊥ ⊥

. (37.2)
   
c c H T

Matching pennies is a hard game for c > 1/2 in the sense that the adversary can
force the regret of any policy to be at least Ω(n2/3 ). To see this, consider the
randomised adversary that chooses the first outcome with probability p and the
second with probability 1 − p. Let ε > 0 be a small constant to be chosen later
and assume p is either 1/2 + ε or 1/2 − ε. The techniques in Chapter 13 show that
the learner can only distinguish between these environments by playing the third
action about 1/ε2 times. If the learner does not choose to do this, then the regret
is expected to be Ω(nε). Taking these together shows the regret is lower-bounded
by Rn = Ω(min(nε, (c − 1/2 + ε)/ε2 )). Choosing ε = n−1/3 leads to a bound
of Rn = Ω((c − 1/2)n2/3 ). Notice that the argument fails when c ≤ 1/2. We
encourage you to pause for a minute to convince yourself about the correctness of
the above argument and to consider what might be the situation when c ≤ 1/2.

Example 37.4 (Bandits). Finite-armed adversarial bandits with binary losses

 37.1 Finite Adversarial Partial Monitoring Problems 481

can be represented in the partial monitoring framework. When k = 2, this is

possible with the following matrices:
   
0 1 0 1 0 1 0 1
L= , Φ=L= .
0 0 1 1 0 0 1 1

The number of columns for this game is 2k . For non-binary rewards, you would
need even more columns. A partial monitoring problem where Φ = L can be
called a bandit problem because the learner observes the loss of the chosen
p action.
In bandit games, Exp3 from Chapter 11 guarantees a regret of O( kn log(k)),
and as noted there, a more sophisticated algorithm will also remove the log(k)
factor. If
√ you completed Exercise 15.4 then you will know that, up to a constant
factor, kn is also the best possible regret in adversarial bandits with binary
losses.
Example 37.5 (Full information problems). One can also represent problems
where the learner observes all the losses. With binary losses and two actions, we
have
   
0 1 0 1 1 2 3 4
L= , Φ= .
0 0 1 1 1 2 3 4

Like for bandits, the size of the game grows quickly as more actions/outcomes
are added. A partial monitoring game where Φai = i for all a ∈ [k] and i ∈ [d]
can be called full information because the signal reveals the losses for all actions.
Example 37.6 (Dynamic pricing). A charity worker is going door to door selling
calendars. The marginal cost of a calendar is close to zero, but the wages of the
door knocker represents a fixed cost of c > 0 per occupied house. The question
is how to price the calendar. Each round corresponds to an attempt to sell a
calendar, and the action is the seller’s asking price from one of d choices. The
potential buyer will purchase the calendar if the asking price is low enough. Below
we give the corresponding matrices for case where both the candidate asking prices
and the possible values for the buyer’s private valuations are {$1, $2, $3, $4}:
   
c − 1 c − 1 c − 1 c − 1 Y Y Y Y
   
   
 c c − 2 c − 2 c − 2 N Y Y Y
L= 
,
 Φ = 

.

 c
 c c − 3 c − 3

N N Y Y
 
   
c c c c−4 N N N Y

Notice that observing the feedback is sufficient to deduce the loss so the problem
could be tackled with a bandit algorithm. But there is additional structure in
the losses here because the learner knows that if a calendar did not sell for $3,
then it would not sell for $4.
 37.2 The Structure of Partial Monitoring 482

37.2 The Structure of Partial Monitoring

The minimax regret of partial monitoring problem G = (L, Φ) is

Rn∗ (G) = inf max Rn (π, i1:n , G) .
π i1:n

One of the core questions in partial monitoring is to understand the growth of

Rn∗ (G) as a function of n for different games. We have seen examples where
Rn∗ (G) = 0 (Example 37.2)
1/2
Rn∗ (G) = Θ(n ) (Example 37.4)
Rn∗ (G) = Θ(n2/3 ) (Example 37.3)
Rn∗ (G) = Ω(n) . (Example 37.1)
The main result of this chapter is that there are no other options. A partial
monitoring game is called trivial if Rn∗ (G) = 0, easy if Rn∗ (G) = Θ(n1/2 ), hard
if Rn∗ (G) = Θ(n2/3 ) and hopeless if Rn∗ (G) = Ω(n). Furthermore, we will show
that any game can be classified using elementary linear algebra.
What makes matching pennies hard and bandits easy? To get a handle on this,
we need a geometric representation of partial monitoring games. The next few
paragraphs introduce a lot of new terminology that can be hard to grasp all at
once. At the end of the section, there is an example illustrating the concepts
(Example 37.10).

37.2.1 The Geometry of Losses and Actions

The geometry underlying partial monitoring comes from viewing the problem as
a linear prediction problem, where the adversary plays on the (d − 1)-dimensional
probability simplex and the learner plays on the rows of L. Define a sequence of
vectors (ut )nt=1 by ut = eit and let `a ∈ Rd be the ath row of matrix L. The loss
suffered in round t when choosing action a is yta = h`a , ut i.
Pt
Let ūt = 1t s=1 us ∈ Pd−1 be the probability vector of proportions of
the adversary’s choices over t rounds. An action a is optimal in hindsight if
h`a , ūn i ≤ minb6=a h`b , ūn i. The cell of an action a is the subset of Pd−1 on which
it is optimal:
 
Ca = u ∈ Pd−1 : maxh`a − `b , ui ≤ 0 ,
b∈[k]

which is a convex polytope. The collection {Ca : a ∈ [k]} is called the cell
decomposition of Pd−1 . Actions with Ca = ∅ are called dominated because
they are never optimal, no matter how the adversary plays. For non-dominated
actions we define the dimension of an action to be the dimension of the affine
hull of Ca . Readers unfamiliar with the affine hull should read Note 4 at the
end of the chapter. A non-dominated action is called Pareto optimal if it has
dimension d − 1, and degenerate otherwise. Actions a and b are duplicates if
 37.2 The Structure of Partial Monitoring 483

`a = `b . The set of all Pareto optimal actions is denoted by Π ⊆ [k]. A partial

monitoring game is called degenerate if it has any degenerate or duplicate actions.

Dominated and degenerate actions can never be uniquely optimal in hindsight,

but their presence can make the difference between a hard game and a
hopeless one. Consider the matching pennies game (Example 37.3). When
c > 1/2, the third action is dominated, but without it the learner would
suffer linear regret. Duplicate actions are only duplicate in the sense that
they have the same loss. They may have different feedback structures and
so cannot be trivially combined.

Neighbourhood relation
Pareto optimal actions a and b are neighbours if Ca ∩ Cb has dimension d − 2.
Note that if a and b are Pareto optimal duplicates, then Ca ∩ Cb has dimension
d − 1, and the definition means that a and b are not neighbours. For Pareto
optimal action a we let Na be the set consisting of a and its neighbours. Given
a pair of neighbours e = (a, b), we let Ne = Nab = {c ∈ [k] : Ca ∩ Cb ⊆ Cc } to
be the set of actions that are incident to e. The neighbourhood relation defines
an undirected graph over [k] with edges E = {(a, b) : a and b are neighbours},
which is called the neighbourhood graph.
The next result, which shows the connectedness of the neighborhood graph
induced by a set of actions whose cells cover the whole simplex, will play an
important role in subsequent proofs:
Lemma 37.7. Suppose that S is any set of Pareto optimal actions such that
∪a∈S Ca = Pd−1 . Then the graph with vertices S and edges from E is connected.
Let e = (a, b) ∈ E. The next lemma characterises actions in Ne as either a, b,
duplicates of a, b or degenerate actions c for which `c is a convex combination of
`a and `b . The situation is illustrated when d = 2 in Fig. 37.2.
Lemma 37.8. Let e = (a, b) ∈ E be neighbouring actions and c ∈ Ne be an action
such that `c ∈
/ {`a , `b }. Then

(a) there exists an α ∈ (0, 1) such that `c = α`a + (1 − α)`b ;

(b) Cc = Ca ∩ Cb ; and
(c) c has dimension d − 2.

Proof We use the fact that if X ⊆ Y ⊆ Rd and dim(X ) = dim(Y), then aff(X ) =
aff(Y) (Exercise 37.2). Introduce ker0 (x) = {u ∈ Rd : u> x = 0, u> 1 = 1}. Clearly,
Ca ∩Cb ⊆ Ca ∩Cc and aff(Ca ∩Cb ) = ker0 (`a −`b ) and aff(Ca ∩Cc ) = ker0 (`a −`c ).
By assumption dim(Ca ∩ Cb ) = d − 2. Since Ca ∩ Cb ⊆ Ca ∩ Cc , it holds that
dim(Ca ∩ Cc ) ≥ d − 2. Furthermore, dim(Ca ∩ Cc ) ≤ d − 2, since otherwise `c = `a .
Hence dim(Ca ∩ Cc ) = d − 2 and thus by the fact mentioned and our earlier
 37.2 The Structure of Partial Monitoring 484

h`3 , (u, 1 − u)i

h`
losses

2
i

,(
u)

u,
−

1
1

−
,
(u

u)
1,

i
h`
0 1
u

Figure 37.2 The figure shows the situation when d = 2 and `1 = (1, 0) and `2 = (0, 1)
and `3 = (1/2, 1/2). The x axis corresponds to P1 = [0, 1], the y axis to the losses.
Then C1 = [0, 1/2] and C2 = [1/2, 1], which both have dimension 1 = d − 1. Then
C3 = {1/2} = C1 ∩ C2 , which has dimension 0.

findings, ker0 (`a − `b ) = ker0 (`a − `c ). This implies (Exercise 37.3) that `a − `b
is proportional to `a − `c so that (1 − α)(`a − `b ) = `a − `c for some α = 6 1.
Rearranging shows that
`c = α`a + (1 − α)`b .
Now we show that α ∈ (0, 1). First note that α ∈ / {0, 1} since otherwise
`c ∈ {`a , `b }. Let u ∈ Ca be such that h`a , ui < h`b , ui, which exists since
dim(Ca ) = d − 1 and dim(Ca ∩ Cb ) = d − 2. Then
h`a , ui ≤ h`c , ui = αh`a , ui + (1 − α)h`b , ui = h`a , ui + (α − 1)h`a − `b , ui ,
which by the negativity of h`a − `b , ui implies that α ≤ 1. A symmetric argument
shows that α > 0. For (b), it suffices to show that Cc ⊂ Ca ∩ Cb . By de Morgan’s
law, for this it suffices to show that Pd−1 \ (Ca ∩ Cb ) ⊂ Pd−1 \ Cc . Thus, pick
some u ∈ Pd−1 \ (Ca ∩ Cb ). The goal is to show that u 6∈ Cc . The choice of u
implies that there exists an action e such that h`a − `e , ui ≥ 0 and h`b − `e , ui ≥ 0
with a strict inequality for either a or b (or both). Therefore, using the fact that
α ∈ (0, 1), we have
h`c , ui = αh`a , ui + (1 − α)h`b , ui > h`e , ui ,
which by definition means that u ∈
/ Cc , completing the proof of (b). Finally, (c)
is immediate from (b) and the definition of neighbouring actions.

37.2.2 Estimating Loss Differences

In order to achieve small regret, the learner needs to identify an optimal action.
How efficiently this can be done depends on the loss and feedback matrices. An
initial observation is that since the loss matrix is known, the learner can restrict
the search for the optimal action to the Pareto optimal actions. Furthermore,
by Lemma 37.7, it suffices to estimate the loss differences between neighbours
and then chain the estimates together along a connecting path. The second
important point is that to minimise the regret the learner only needs to estimate
 37.2 The Structure of Partial Monitoring 485

the differences in losses between Pareto optimal actions and not the actual losses
themselves. In fact, there exist games for which estimating the actual losses is
impossible, but estimating the differences is straightforward:

Example 37.9. Consider the partial monitoring game with

   
0 1 10 11 ♥ ♣ ♥ ♣
L= , Φ= .
1 0 11 10 ♣ ♥ ♣ ♥

The learner can never tell if the environment is playing in the first two columns or
the last two, but the differences between the losses of actions are easily deduced
from the feedback no matter the outcome and the action.

Only the loss differences between Pareto optimal actions need to be estimated.
There are games that are easy, but where some loss differences cannot be
estimated. For example, there is never any need to estimate the losses of a
dominated action.

Having decided we need to estimate the loss differences between neighbouring

Pareto optimal actions, the next question is how the learner can do this. Focusing
our attention on a single round, suppose the adversary secretly chooses an outcome
i ∈ [d] and the learner samples an action A from distribution p ∈ ri(Pk−1 ) and
observes σ = ΦAi . We are interested in finding an unbiased estimator of Lai − Lbi
for neighbouring actions a and b. Without loss of generality, the estimator can
be in the form of f (A, σ)/pA with some function f : [k] × Σ → R. Then, the
unbiasedness requirement takes the convenient form
  X
f (A, σ)
k
E = f (c, Φci ) .
pA c=1

In other words, f (A, σ)/pA is an unbiased estimator of Lai − Lbi regardless of

the adversaries’ choice if and only if
k
X
f (c, Φci ) = Lai − Lbi for all i ∈ [d] . (37.3)
c=1

A pair of neighbours a and b are called globally observable if there exists a

function f satisfying Eq. (37.3). The set of all functions f : [k] × Σ → R satisfying
glo
Eq. (37.3) is denoted by Eab . A pair of neighbours a and b are locally observable
if f can be chosen satisfying Eq. (37.3) with f (c, σ) = 0 whenever c ∈ / Nab . The set
loc
of functions satisfying this additional requirement are Eab . A partial monitoring
problem is called globally/locally observable if all pairs of neighbouring actions are
globally/locally observable. The global/local observability conditions formalise
the idea introduced in Example 37.3. Games that are globally observable but not
locally observable are hard because the learner cannot identify the optimal action
 37.3 Classification of Finite Adversarial Partial Monitoring 486

by playing near-optimal actions only. Instead it has to play badly suboptimal

actions to gain information, and this increases the minimax regret.

Example 37.10. The partial monitoring problem illustrated in Fig. 37.3 has six
actions, three feedbacks and three outcomes. The cell decomposition is shown on
the right with the 2-simplex parameterised by its first two coordinates u1 and
u2 so that u3 = 1 − u2 − u1 . Actions 1, 2 and 3 are Pareto optimal. There are
no dominated actions while actions 4 and 5 are 1-dimensional and action 6 is 0-
dimensional. The neighbours are (1, 3) and (2, 3), which are both locally observable,
and so the game is locally observable. Note that (1, 2) are not neighbours because
the intersection of their cells is (d − 3)-dimensional. Finally, N3 = {1, 2, 3} and
N1 = {1, 3} and N23 = {2, 3, 4}. Think about how we decided on what losses to
use to get the cell decomposition shown in Fig. 37.3.

    u2
0 1 1 1 2 3
   
 1 0 1  ⊥ ⊥ ⊥
   
    C4
1/2 1/2 1/2 ⊥ ⊥ ⊥
    C2
L=  Φ=  C6
3/4 1/4 3/4 1 2 3
   
   
 1 1/2 1/2 ⊥ ⊥ ⊥ C5 C3
    C1
1 1/4 3/4 ⊥ ⊥ ⊥ u1

Figure 37.3 Partial monitoring game with k = 6 and d = 3 and m = 3.

37.3 Classification of Finite Adversarial Partial Monitoring

The terminology in the last section finally allows us to state the main theorem of
this chapter that classifies finite adversarial partial monitoring games.

Theorem 37.11. The minimax regret of partial monitoring problem G = (L, Φ)

falls into one of four categories:


 0, if G has no pairs of neighbouring actions ;


Θ(n1/2 ) , if G is locally observable and has neighbouring actions ;
Rn∗ (G) =


 Θ(n2/3 ) , if G is globally observable, but not locally observable ;


Ω(n) , otherwise .

The Landau notation is used in the traditional mathematical sense and

obscures dependence on k, d, m and the finer structure of G = (L, Φ).
 37.4 Lower Bounds 487

The proof is split into parts by proving upper and lower bounds for each part.
First up is the lower bounds. We then describe a policy and analyse its regret.

37.4 Lower Bounds

Like for bandits, the lower bounds are most easily proven using a stochastic
adversary. In stochastic partial monitoring, we assume that u1 , . . . , un are
chosen independently at random from the same distribution. To emphasise
the randomness, we switch to capital letters. Given a partial monitoring game
G = (L, Φ) and probability vector u ∈ Pd−1 , the stochastic partial monitoring
environment associated with u samples a sequence of independently and identically
distributed random variables I1 , . . . , In with P (It = i) = ui and Ut = eIt . In each
round t, a policy chooses action At and receives feedback σt = ΦAt It . The regret
is
" n # " n #
X X
Rn (π, u) = max E h`At − `a , Ut i = max E h`At − `a , ui .
a∈[k] a∈[k]
t=1 t=1

The reader should check that Rn∗ (G)≥ inf π maxu∈Pd−1 Rn (π, u), which allows
us to restrict our attention to stochastic partial monitoring problems. Given
u, q ∈ Pd−1 , let D(u, q) be the relative entropy between categorical distributions
with parameters u and q respectively:
  X
(ui − qi )2
X d d
ui
D(u, q) = ui log ≤ , (37.4)
i=1
qi i=1
qi

where the second inequality follows from the fact that for measures P, Q we have
D(P, Q) ≤ χ2 (P, Q) (see Note 6 in Chapter 13).
Theorem 37.12. Let G = (L, Φ) be a globally observable partial monitoring
problem that is not locally observable. Then there exists a constant cG > 0 such
that Rn∗ (G) ≥ cG n2/3 .

Proof The proof involves several steps. Roughly, we need to define two alternative
stochastic partial monitoring problems. We then show these environments are
hard to distinguish without playing an action associated with a large loss. Finally
we balance the cost of distinguishing the environments against the linear cost of
playing randomly. Without loss of generality assume that Σ = [m].

Step 1: Defining the Alternatives

Let a, b be a pair neighbouring actions that are not locally observable. Then, by
definition, Ca ∩ Cb is a polytope of dimension d − 2. Let u be the centroid of
Ca ∩ Cb and
ε = min h`c − `a , ui . (37.5)
c∈N
/ ab
 37.4 Lower Bounds 488

Ca

ua
u Cb
ub
Cc

Figure 37.4 Lower-bound construction for hard partial monitoring problems. Shown is
Pd−1 , the cells Ca and Cb of two Pareto optimal actions, and two alternatives ua ∈ Ca
and ub ∈ Cb that induce the same distributions on the outcomes under both a and b.

The value of ε is well defined, since by global observability of G, but nonlocal

observability of (a, b), there must exist some action c ∈ / Nab . Furthermore, since
c∈/ Nab , it follows that ε > 0. As in the lower-bound constructions for stochastic
bandits, we now define two stochastic partial monitoring problems ua , ub by
choosing a direction q ∈ Rd and a small value ∆ such that ua = u − ∆q ∈ Ca
and ub = u + ∆q ∈ Cb (see Fig. 37.4). This means that action a is optimal if the
environment plays ua on average and b is optimal if the environment plays ub
on average. The direction q will be chosen so that using a and b alone it is not
possible to distinguish between ua and ub .
loc
The vector q is chosen as follows: Since (a, b) are not locally observable, Eab = ∅.
Equivalently, there does not exist a function f : [k] × Σ → R such that for all
i ∈ [d],
X
f (c, Φci ) = `ai − `bi . (37.6)
c∈Nab

In this form, it does not seem obvious what the next step should be. To clear
things up, we introduce some linear algebra. Let Sc ∈ {0, 1}m×d be the matrix
with (Sc )σi = I {Φci = σ}, which is chosen so that Sc ei = eΦci . Define the linear
map S : Rd → R|Nab |m by
 
Sa 
 
 
 Sb 
S= 
 ..  ,
 . 
 
 
Sc

which is the matrix formed by stacking the matrices {Sc : c ∈ Nab }. Then, an
elementary argument shows that there exists a function f satisfying Eq. (37.6) if
 37.4 Lower Bounds 489

and only if there exists a w ∈ R|Nab |m such that

`a − `b = S > w .

In other words, actions (a, b) are locally observable if and only if `a − `b ∈ im(S > ).
Since we have assumed that (a, b) are not locally observable, we must have
/ im(S > ). Let z ∈ im(S > ) and w ∈ ker(S) be such that `a − `b = z + w,
`a − `b ∈
which is possible since im(S > )⊕ker(S) = Rd . Since `a −`b ∈ / im(S > ), it holds that
w= 6 0 and h`a − `b , wi = hz + w, wi = hw, wi =6 0. Note also that 1 ∈ im(S > ) and
hence h1, wi = 0. Finally, let q = w/h`a − `b , wi. By construction, q ∈ Rd , q 6= 0
while Sq = 0, h`a − `b , qi = 1 and h1, qi = 0. Let ∆ > 0 be some small constant
to be chosen subsequently. With this, we define ua = u − ∆q and ub = u + ∆q so
that

h`b − `a , ua i = ∆ and h`a − `b , ub i = ∆ . (37.7)

We note that if ∆ is sufficiently small, then ua ∈ Ca and ub ∈ Cb because a and

b are Pareto optimal.

Step 2: Calculating the Relative Entropy

Given action c and r ∈ Pd−1 , let Pcr be the distribution on the feedback observed
by the learner when playing action c in stochastic partial monitoring environment
determined by r. That is Pcr (σ) = (Sc r)σ . Further, let Pr be the distribution
on the histories Hn = (A1 , Φ1 , . . . , An , Φn ) arising from the interaction of the
learner’s policy with the stochastic environment determined by r. Expectations
with respect to Pr are denoted by Er . A modification of Lemma 15.1 shows that
X
D(Pua , Pub ) = Eua [Tc (n)] D(Pcua , Pcub ) . (37.8)
c∈[k]

By the definitions of ua and ub , we have Sc ua = Sc ub for all c ∈ Nab . Therefore,

Pcua = Pcub and D(Pcua , Pcub ) = 0 for all c ∈ Nab . On the other hand, if
c∈/ Nab , then by the data processing inequality (Exercise 14.9) and Eq. (37.4),
for ∆ ≤ mini:qi 6=0 ui /(2|qi |),

(uai − ubi )2 qi2

d
X k
X
D(Pcua , Pcub ) ≤ D(ua , ub ) ≤ ≤ 4∆2 ≤ C̃u ∆2 ,
i=1
ubi i=1
ui − ∆|qi |

where we used that u ∈ Ca ∩ Cb is not on the boundary of Pd−1 , so ui > 0 for all
i and we defined C̃u as a suitably large constant that depends on u (q is entirely
determined by a and b). Therefore,
X
D(Pua , Pub ) ≤ C̃u E[Tc (n)]∆2 . (37.9)
c∈N
/ ab

Step 3: Comparing the Regret

By Eq. (37.5) and Hölder’s inequality, for c ∈ / Nab we have h`c − `a , ua i ≥
ε − h`c − `a , ∆qi ≥ ε − ∆kqk1 and h`c − `b , ub i ≥ ε − ∆kqk1 , where, for simplicity
 37.4 Lower Bounds 490

and without the loss of generality, we assumed that the losses lie in [0, 1]. Define
T̃ (n) to be the number of times an arm not in Nab is played:
X
T̃ (n) = Tc (n) .
c∈N
/ ab

By Lemma 37.8, for each action c ∈ Nab , there exists an α ∈ [0, 1] such that
`c = α`a + (1 − α)`b . Therefore, by Eq. (37.7),

h`c − `a , ua i + h`c − `b , ub i = (1 − α)h`b − `a , ua i + αh`a − `b , ub i = ∆ ,

(37.10)

which means that max(h`c − `a , ua i, h`c − `b , ub i) ≥ ∆/2. Define T̄ (n) as the

number of times an arm in Nab is played that is at least ∆/2 suboptimal in ua :
X  ∆

T̄ (n) = I h`c − `a , ua i ≥ Tc (n) .
2
c∈Nab

It also follows from (37.10) that if c ∈ Nab and h`c − `a , ua i < ∆ 2 , then
∆
P
h`c −`b , ub i ≥ 2 . Hence, under ub , the random pseudo-regret, c Tc (n)h`c −`b , ub i,
is at least (n − T̄ (n))∆/2. Assume that ∆ is chosen sufficiently small so that
∆kqk1 ≤ ε/2. By the above,

Rn (π, ua ) + Rn (π, ub )
   
X X
= Eua  Tc (n)h`c − `a , ua i + Eub  Tc (n)h`c − `b , ub i
c∈[k] c∈[k]

ε   n∆

≥ Eu T̃ (n) + Pua (T̄ (n) ≥ n/2) + Pub (T̄ (n) < n/2)
2 a 4
ε   n∆
≥ Eua T̃ (n) + exp (− D(Pua , Pub ))
2 8
ε   n∆  

≥ Eua T̃ (n) + exp −C̃u ∆2 Eua T̃ (n) ,
2 8
where the second inequality follows from the Bretagnolle–Huber inequality
(Theorem 14.2) and the third from Eqs. (37.8) and (37.9). The bound is completed
by choosing
 
ui ε
∆ = min min , ,
i:qi 6=0 2|qi | 2kqk1 n1/3

which is finite since q 6= 0. Straightforward calculation concludes the result

(Exercise 37.7).

We leave the following theorems as exercises for the reader (Exercises 37.8
and 37.9):

Theorem 37.13. If G is not globally observable and has at least two non-
dominated actions, then there exists a constant cG > 0 such that Rn∗ (G) ≥ cG n.
 37.5 Policy and Upper Bounds 491

Proof sketch Since G is not globally observable, there exists a pair of

neighbouring actions (a, b) that are not globally observable. Let u be the centroid
of Ca ∩Cb . Let S ∈ Rkm×d be the stack of matrices from {Sc : c ∈ [k]}. Then, using
the same argument as the previous proof, we have `a − `b ∈ / im(S > ). Now define
q ∈ R such that h1, qi = 0, h`a − `b , qi = 1 and Sq = 0. Let ∆ > 0 be sufficiently
d

small and ua = u−∆q and ub = u+∆q. Show that D(Pua , Pub ) = 0 for all policies
and complete the proof in the same fashion as the proof of Theorem 37.12.

Theorem 37.14. Let G = (L, Φ) be locally observable and have at least one pair
of neighbours. Then there exists a constant cG > 0 such that for all large enough
√
n the minimax regret satisfies Rn∗ (G) ≥ cG n.

Proof sketch By assumption, there exists a pair of neighbouring actions (a, b).
Define u as the centroid of Ca ∩ Cb and ua and ub be the centroids of Ca and
Cb respectively. For sufficiently small ∆ > 0, let va = (1 − ∆)u − ∆ua and
vb = (1 − ∆)u + ∆ub . Then

(vai − vbi )2
d
X
D(Pva , Pvb ) ≤ n ≤ cG n∆2 ,
i=1
vbi
√
where cG > 0 is a game-dependent constant. Let ∆ = 1/ n and apply the ideas
in the proof of Theorem 37.12.

37.5 Policy and Upper Bounds

We now describe a policy for globally and locally observable games, and prove
its regret is O(n1/2 ) for locally observable games and O(n2/3 ) otherwise. For
the remainder of this section, fix a globally observable game G = (L, Φ). The
glo loc
estimation functions in Eab and Eab are designed to combine with importance-
weighting to estimate the loss differences between actions a and b. For this section,
it is more convenient to define estimation functions for the whole loss vector up
to constant shifts. Let E vec be the set of all functions f : [k] × Σ → Rk such that:

(a) f (a, σ)b = 0 for all b ∈

/ Π; and
(b) for each outcome i ∈ [d], there exists a constant c ∈ R with
k
X
f (a, Φai )b = Lbi + c for all b ∈ Π .
a=1

The intuition is that E vec is the set of functions that serve as unbiased loss
difference estimators in the sense that when A ∼ p ∈ ri(Pk−1 ), then
 
hea − eb , f (A, ΦAi )i
E = h`a − `b , ei i for all Pareto optimal a, b and i ∈ [d] .
pA
As we will see in the proof of Theorem 37.16, if G is globally observable, then
 37.5 Policy and Upper Bounds 492

E vec is non-empty. By identifying functions and vectors, E vec ⊂ Rk(km) , a view

that will be useful later.
The policy for partial monitoring combines exponential weights with a careful
exploration strategy. A little reminder about exponential weights and some new
notation will be useful. Given a probability vector q ∈ Pk−1 , define a function
Ψq : Rk → R by

Ψq (z) = hq, exp(−z) + z − 1i ,

where the exponential function is applied component-wise. You might recognise

Ψq as the Bregman divergence

Ψq (z) = DF ∗ (∇F (q) − z, ∇F (q)) ,

where F is the unnormalised negentropy potential. Suppose that (ŷt )nt=1 is an

arbitrary sequence of vectors with ŷt ∈ Rk , η > 0 and
 P 
exp −η s=1 ŷsa
t−1

Qta = P  P , t ∈ [n] .
b=1 exp −η
k t−1
s=1 ŷsb

Recall from Theorem 28.4 that for any a∗ ∈ [k],

log(k) 1 X
n X
X k n
Qta (ŷta − ŷta∗ ) ≤ + ΨQt (η ŷt ) . (37.11)
t=1 a=1
η η t=1

Exp3 is derived by defining ŷt as the importance-weighted loss estimator and

sampling At from Qt . We will do something similar in partial monitoring, but
with two significant differences: (a) the importance-weighted estimator must
depend on the feedback and loss matrices, and (b) the algorithm will sample
At from an alternative distribution Pt that is optimised to balance the regret
suffered relative to Qt and the information gained.

The definition of global observability does not imply that loss differences
between dominated and degenerate actions can be estimated. Consequentially,
the distribution Qt used by the new algorithm will be supported on Pareto
optimal actions only. The actual distribution Pt used when choosing an
action may also include degenerate actions, however.

The optimisation problem for balancing information and regret explicitly

optimises a worst case upper bound on the right-hand side of Eq. (37.11). For
η > 0 and q ∈ Pk−1 with Supp(q) ⊆ Π, let
"  #
1 1 X ηf (a, Φai )
k
optq (η) = infvec max (p − q) Lei + 2
>
p a Ψq .
f ∈E i∈[d] η η a=1 pa
p∈ri(Pk−1 )
(37.12)
 37.5 Policy and Upper Bounds 493

Of course, optq (η) depends on the game G, which is hidden from the notation
to reduce clutter. The first term in the right-hand side of Eq. (37.12) measures
the additional regret when playing p rather than q, while the second corresponds
to the expectation of the second term in Eq. (37.11) when the algorithm uses
importance-weighting using estimation function f . The optimisation problem is
convex and hence amenable to efficient computation (see Note 9 for some details).
The worst-case value over all q is

opt∗ (η) = sup{optq (η) : q ∈ Pk−1 , Supp(q) ⊆ Π} .

The function q 7→ optq (η) is generally not convex, so opt∗ (η) may be hard to
compute. This causes a minor problem when setting the learning rate, which can
be mitigated by adapting the learning rate online as discussed in Note 7.

We say that f ∈ E vec and p ∈ ri(Pk−1 ) solve Eq. (37.12) with precision
ε ≥ 0 if
"  #
1 1 X ηf (a, Φai )
k
max (p − q) Lei + 2
>
p a Ψq ≤ optq (η) + ε . (37.13)
i∈[d] η η a=1 pa

Such approximately optimal solutions exist for any ε > 0, but may not exist
for ε = 0 because the constraint on p is not compact.

The convexity of the inner maximum in Eq. (37.12) can be checked using the
following construction. The perspective of a convex function f : Rd → R
is a function g : Rd+1 → R given by
(
uf (x/u) , if u > 0 ;
g(x, u) = (37.14)
∞, otherwise .
The perspective is known to be convex (Exercise 37.1). Since Ψq is convex
and the max of convex functions is convex, it follows that the term inside of
the infimum of Eq. (37.12) is convex.

The full algorithm is given as Algorithm 26.

Theorem 37.15. For any η > 0 and ε > 0, the regret of Algorithm 26 is bounded
by

log(k)
Rn ≤ + nη(opt∗ (η) + ε) .
η

Proof The result follows from the definitions of E vec and the regret, and the
Pn
bound for exponential weights in Eq. (37.11). Let a∗ = argmina∈Π t=1 h`a , ut i.
 37.5 Policy and Upper Bounds 494

1: Input: η, ε, L and Φ
2: for t ∈ 1, . . . , n do
3: Compute exponential weights distribution Qt ∈ Pk−1 by:
 P 
IΠ (a) exp −η s=1 ŷsa
t−1

Qta = P  P .
exp
t−1
b∈Π −η s=1 ŷsb

4: Solve Eq. (37.12) with q = Qt and precision ε to find Pt ∈ Pk−1 and

ft ∈ E vec
5: Sample At ∼ Pt and observe σt
6: Set ŷt = ft (At , σt )/PtAt
7: end for
Algorithm 26: Exponential weights for partial monitoring. Recall that Π denotes the set of
Pareto optimal actions.

Then,
" n
#
X
Rn = E LAt it − La∗ it
t=1
" n X
k
#
X
=E Pta (Lait − La∗ it )
t=1 a=1
" n X
k
# " n k #
X XX
=E Qta (Lait − La∗ it ) + E (Pta − Qta )Lait
t=1 a=1 t=1 a=1
" n X
k
# " n k #
X XX
=E Qta (ŷta − ŷta∗ ) + E (Pta − Qta )Lait .
t=1 a=1 t=1 a=1

The first expectation is bounded using the definition of Qt and Eq. (37.11) by
" n k # " n #
XX log(k) 1 X
E Qta (ŷta − ŷta∗ ) ≤ + E ΨQt (η ŷt )
t=1 a=1
η η t=1
" n k  #
log(k) 1 XX ηft (a, Φait )
= + E Pta ΨQt .
η η t=1 a=1
Pta

Combining the two displays, using the definitions of Pt , ft and εt , and substituting
the definition of optQt (η) ≤ opt∗ (η) completes the proof.

The extent to which this result is useful depends on the behaviour of opt∗ (η)
for different classes of games. The following two theorems bound the value of
the optimisation problem for globally observable and locally observable games
respectively. An apparently important quantity in the regret upper bounds for
both globally and locally observable games is the minimum magnitude of the
 37.5 Policy and Upper Bounds 495

estimation functions. Let

vglo = max min kf k∞ and vloc = max min kf k∞ .

e∈E glo e∈E f ∈E loc
f ∈Ee e

In the remainder of this chapter we assume that the losses are between zero
and one L ∈ [0, 1]k×d .

√
Theorem 37.16. For all globally observable games, opt∗ (η) ≤ 2vglo k 2 / η for all
2
η ≤ 1/ max{1, vglo k 4 }.

Theorem 37.17. For all locally observable games, opt∗ (η) ≤ 9k 3 max(1, vloc
2
) for
2
all η ≤ 1/(2k max(1, vloc )).

The proofs follow in subsequent sections. Combining Theorem 37.15 with

Theorem 37.17 shows that for an appropriately tuned learning rate, the regret of
Algorithm 26 on locally observable games is bounded by
 p 
Rn = O vloc k 3/2 n log(k) .

By using Theorem 37.16, it follows that for globally observable games the regret
is bounded by
 
Rn = O (vglo kn)2/3 (log(k))1/3 .

These results establish the upper bounds in the classification theorem for locally
and globally observable games. The quantities vglo and vloc only depend on G but
may be exponentially large in d. We walk you through the proof of the following
proposition in Exercise 37.12.

Proposition 37.18. The following hold:

(a) If G is globally observable, then vglo ≤ d1/2 k d/2 .

(b) If G is locally observable, then vloc ≤ d1/2 k d/2 .
(c) If G is locally observable and non-degenerate, then vloc ≤ m.

The only property of non-degenerate games used in Part (c) is that |Ne | = 2
for all e ∈ E. It is illustrative to bound opt∗ (η) for well-known games. The next
proposition shows that Algorithm 26 recovers the usual bounds for bandits and
the full information setting.

Proposition 37.19. The following hold:

(a) For bandit games (Φ = L), opt∗ (η) ≤ k/2.

(b) For full information games (Φai = i for all a and i), opt∗ (η) ≤ 1/2.

You will prove this proposition in Exercise 37.14 by making explicit choices of
p ∈ ri(Pk−1 ) and f ∈ E vec .
 37.6 Proof of Theorem 37.16 496

37.6 Proof of Theorem 37.16

The definition of global (and local) observability is defined in terms of the existence
of functions serving as unbiased loss estimators between pairs of neighbouring
actions. To make a connection between E vec and E glo (and E loc ) we need the
concept of an in-tree on the neighbourhood graph. Let S be a subset of Pareto
optimal actions with no duplicate actions and ∪a∈S Ca = Pd−1 . An in-tree on the
graph (S, E) is a set of edges T ⊂ E such that (S, T ) is a directed tree with all
edges pointing towards a special vertex called the root, denoted by rootT and
such that V (T ), the set of vertices underlying T , is the same as S. Provided
the game is non-trivial, then such a tree exists by Lemma 37.7. Given a Pareto
optimal action b, let pathT (b) ⊆ T denote the path from b to the root. The path
is empty when b is the root. When b is not the root, we let parT (b) denote the
unique Pareto optimal action such that (b, parT (b)) ∈ T .
Abbreviate v = vglo and let T ⊂ E be an arbitrary in-tree over the Pareto
optimal actions. For each e ∈ E, let fe ∈ Eeglo be such that kfe k∞ ≤ v. Then
define f : [k] × Σ → Rk by
X
f (a, σ)b = fe (a, σ) .
e∈pathT (b)

By the triangle inequality, maxa∈[k],σ∈Σ kf (a, σ)k∞ ≤ kv. Furthermore, f ∈ E vec ,

since for any outcome i,
k
X k
X X
f (a, Φai )b = fe (a, Φai ) = Lbi − Lroot(T )i .
a=1 a=1 e∈pathT (b)

√
Let p = (1 − γ)q + γ1/k with γ = vk 2 η. By the condition in the theorem that
η ≤ 1/ max{1, v 2 k 4 }, it holds that γ ≤ 1 and hence p ∈ ri(Pk−1 ). The next step
is to bound the minimum possible value of the loss estimator. For actions a and
b and outcome i,
ηf (a, Φai )b ηvk 2 √
≥− = − η ≥ −1 ,
pa γ
where in the final inequality we used the fact that η ≤ 1. Next, using the fact
that exp(−x) ≤ x2 + 1 − x for x ≥ −1, it follows that for any z ≥ −1,
k
X
Ψq (z) ≤ qb zb2 , (37.15)
b=1

which is the inequality we have used long ago in Chapter 11. Using this,
  X
1 X ηf (a, Φai ) k4 v2 k2 v
k k X
qb 2
p Ψ ≤ f (a, Φ ) ≤ = √ ,
η 2 a=1
a q ai b
pa a=1
pa γ η
b∈Π

where we used that kf (a, σ)k∞ ≤ kv and p ≥ γ/k1 and that q ∈ Pk−1 . For the
 37.7 Proof of Theorem 37.17 497

other component of the objective,

1 γ vk 2 vk 2
(p − q)> Lei = (1/k − q)> Lei = √ (1/k − q)> Lei ≤ √ .
η η η η

Combining the previous two displays shows that for any i ∈ [d],
 
1 1 X ηf (a, Φai ) 2vk 2
k
(p − q) Lei + 2
>
p a Ψq ≤ √ ,
η η a=1 pa η

which is the desired result.

37.7 Proof of Theorem 37.17

Exploiting local observability is not straightforward. To gain some insight let us

consider the matching pennies game with c = 1/4:
   
 0 1  ⊥ ⊥
    1 3 2
L=
 1 0 
 Φ=
⊥ ⊥
   
1/4 1/4 H T

The figure on the right-hand side is the neighbourhood graph. Notice that the third
action is revealing and also separates the first two actions in the neighbourhood
graph. Clearly, loss differences can be estimated between all pairs of neighbours
in this graph, and hence the game is locally observable. Let’s suppose now that
q = (1/2 − ε/2, 1/2 − ε/2, ε) and p = q. The obvious estimation function f ∈ E vec
is given by

(0, 1, 1/4) , if a = 3 and σ = 1 ;

 >

f (a, σ) = (1, 0, 1/4)> , if a = 3 and σ = 2 ;



0 , otherwise .

Examining the second term in Eq. (37.12) and using a second order Taylor
approximation,
3   3
1 X ηf (a, Φai ) 1 X qb 2 1 1−ε
p Ψ ≈ Lbi = + ,
2 2 32 4ε
a q
η a=1 pa p3
b=1

which holds for both i = 1 and i = 2. This is bad news. The appearance of p3 = q3
in the denominator means the objective can be arbitrarily large when ε is small.
Taylor’s theorem shows that the approximation is not to blame, provided that
η is suitably small. The main issue is that q and p assign most of their mass to
two actions that are not neighbours and hence cannot be distinguished without
 37.7 Proof of Theorem 37.17 498

playing a third action. Now suppose that p is constructed by transferring mass

from the first two actions to the third by:
p = q − min(q1 , q2 )(e1 + e2 ) + 2 min(q1 , q2 )e3 .
The first observation is that this can only decrease the expected loss:
3
(p − q)> L = − min(q1 , q2 )1 ≤ 0 .
4
This takes care of the first term in the objective. Let us assume without loss of
generality that p1 = max(p1 , p2 , p3 ), and let

(0, 1, 1/4)> ,

 if a = 3 and σ = 1 ;
f (a, σ) = (0, −1, −3/4) , if a = 3 and σ = 2 ;
>


0 , otherwise .
Using again a Taylor approximation suggests the second term in the objective is
now well behaved:
 
1 X ηf (a, Φai ) 1 X qb
k k
p Ψ ≈ f (3, Φ3i )2b
η 2 a=1 2
a q
pa p3
b=1
   
1 q2 q3 1 1
≤ + ≤ +1 .
2 p3 p3 2 2
Things are starting to look more promising. By transferring the mass in q towards
the revealing action and shifting the loss estimators to be zero on the most played
action, we have gained control of the stability term and simultaneously decreased
the expected loss of p relative to q.

37.7.1 Duality and the Water Transfer Operator

The water transfer operator, which we will introduce momentarily, provides the
generalisation of the specific argument just given. The first step is an application
of Sion’s minimax theorem (Theorem 28.12) to Eq. (37.12), which shows that
"  #
(p − q)> Lλ 1 X X ηf (a, Φai )
d k
optq (η) = max infvec + 2 λi p a Ψq .
λ∈Pd−1 f ∈E η η i=1 a=1 pa
p∈Pk−1
(37.16)
By exchanging the max and the inf, we free ourselves from finding a distribution
p and estimation function f such that the objective is controlled for all choices
of the adversary. Now we only need to find a p and f for each distribution over
outcomes λ ∈ Pd−1 .
Fix therefore an arbitrary distribution λ ∈ Pd−1 . Let S be an arbitrary subset
of Pareto optimal actions containing no duplicates and for which ∪a∈S Ca = Pd−1
and let T ⊂ E be an in-tree over S. Given an edge e = (a, b) ∈ E, let
αe : Ne → [0, 1] be the mapping such that `c = (1 − αe (c))`a + αe (c)`b for
 37.7 Proof of Theorem 37.17 499

Root
7
5 8 9 12
6

Degenerate
4 11
actions

2 1 10
3

Figure 37.5 The large nodes are Pareto optimal actions in S. The smaller nodes inside
are their duplicates, which are not part of S. The remaining nodes are degenerate
actions that are linear combinations of Pareto optimal actions. The arrows indicate
the in-tree. A vector y ∈ Rk is T -increasing if it is constant on duplicate actions and
otherwise increasing in the direction of the arrows. In this case, the constraint is that
y1 = y2 = y3 ≤ y4 ≤ y5 = y6 = y7 ≤ y8 ≤ y9 ≤ y12 and y10 ≤ y11 ≤ y12 .

all c ∈ Ne , which exists by Lemma 37.8. Note that αe (c) = 0 when c is a

duplicate of a and αe (c) = 1 when it is a duplicate of b. A vector y ∈ Rk is called
T -increasing if for all e = (a, b) ∈ T and c, d ∈ Ne with αe (d) ≥ αe (c), it holds
that yd ≥ yc . A vector y is called T -decreasing if −y is T -increasing. This concept
is illustrated in Fig. 37.5.

Lemma 37.20. Given an in-tree T ⊂ E and distribution q ∈ Pk−1 there exists a

distribution r ∈ Pk−1 such that

(a) r ≥ q/k;
(b) r is T -increasing; and
(c) hr − q, yi ≤ 0 for all T -decreasing vectors y ∈ Rk .

Proof For simplicity, we give the proof for the special case that all actions
are Pareto optimal and there are no duplicates, in which case S = [k].
The proof is generalised in Exercise 37.11. Given an action a ∈ [k], let
ancT (a) = ∪e∈pathT (a) Ne ∪ {a} be the set of ancestors of a, including a and
descT (a) = {b : a ∈ ancT (b)} be the set of descendants of a. Define r by
X qb
ra = .
| ancT (b)|
b∈descT (a)

Let us first confirm that r ∈ Pk−1 . That r ≥ 0 is obvious and

k
X k
X X X k X
qb qb
ra = = = 1.
a=1 b∈descT (a)
| ancT (b)| | ancT (b)|
a=1 b=1 a∈ancT (b)
 37.7 Proof of Theorem 37.17 500

For Part (a), the definition means that ra ≥ qa /| ancT (a)| ≥ qa /k. That r is
T -increasing follows immediately from the definition. (c) follows because
k
X X Xk X
qb yb qb
hr, yi = ya ≥ = hq, yi .
| ancT (b)| a=1 | ancT (b)|
a=1 b∈descT (a) b∈descT (a)

The existence of the mapping q 7→ r given by Lemma 37.21 was originally proven
using a ‘water flowing’ argument and was called the water transfer operator.

Lemma 37.21. Let S as before. Then, for any λ ∈ Pd−1 , there exists an in-tree
T ⊂ E over vertices S such that Lλ is T -decreasing.

Proof Again, we outline the argument for games with no degenerate or duplicate
actions, leaving the complete proof for Exercise 37.11. Let a be an action such
that λ ∈ Ca . First, assume that λ ∈ ri(Ca ). The root of our tree will be
a (the reader may find helpful to check Fig. 37.6). Next, for b 6= a, define
par(b) = argminc∈Nb e> c Lλ and then let T = {(b, par(b)) : b 6= a}. Clearly,
V (T ) = [k]. Provided that T really is a tree, the fact that Lλ is T -decreasing is
obvious from the definition of the parent function. That T is a tree follows by
showing that for any (b, d) ∈ T , e>
d Lλ < eb Lλ, which we will prove now. For this,
>

let ω ∈ ri(Cb ) and c ∈ Nb such that Cc ∩ [ω, λ] 6= ∅. These exist by Exercise 37.10
(see also Fig. 37.6). We now show that e> c Lλ < eb Lλ from which the desired
>

result follows. To show this let

f (α) = (eb − ec )> L((1 − α)ω + αλ) .

It suffices to show that f (1) > 0. The following hold: (a) f is linear; (b) f (0) < 0,
since ω ∈ ri(Cb ); and (c) there exists an α ∈ (0, 1) such that f (α) = 0, which
holds because Cc ∩ [ω, λ] 6= ∅ and λ ∈ ri(Ca ). Thus f (1) > 0, establishing the
result for λ ∈ ri(Ca ). When λ is on the boundary of Ca , let (λ(i) )∞
i=1 be a sequence
in ri(Ca ) so that limi→∞ λ(i) → λ. For each i, let T (i) ⊂ E be an in-tree such
that Lλ(i) is T (i) -decreasing. Since there are only finitely many trees, by selecting
a subsequence we conclude that there exists an in-tree T ⊂ E such that Lλ(i) is
T -decreasing for all i. The result follows by taking the limit.

This concludes the building of the tools needed to control optq (η) for locally
observable games.

Proof of Theorem 37.17 Abbreviate v = vloc and let λ ∈ Pd−1 be arbitrary. By

Lemma 37.21, there exists an in-tree T ⊂ E over S such that Lλ is T -decreasing.
Hence, by Lemma 37.20, there exists a T -increasing r ∈ Pk−1 such that r ≥ q/k
and (r − q)> Lλ ≤ 0. Let p = (1 − γ)r + γ1/k with γ = ηvk 2 and
X
f (a, σ)b = fe (a, σ) ,
e∈pathT (b)

where fe ∈ Eeloc has kfe k∞ ≤ v. The same argument as in the proof of

 37.7 Proof of Theorem 37.17 501

Ca
λ

Cb
Cc ω

Figure 37.6 The core argument used in the proof of Lemma 37.21.

Theorem 37.17 shows that f ∈ E vec . Moving to the objective in Eq. (37.16), we
lower-bound the loss estimates:
ηf (a, σ)b η X ηvk 2
= fe (a, σ) ≥ − = −1 . (37.17)
pa pa γ
e∈pathT (b)

Fix i ∈ [k]. The stability term is bounded using the properties of p and f as
follows:
  X
1 X ηf (a, Φai )
k k X
qb
pa Ψq ≤ f (a, Φai )2b
η 2 a=1 pa a=1 b∈Π
p a
 2
Xk X
qb  X
= fe (a, Φai )
pa
a=1 b∈Π e∈pathT (b)
 2
Xk X X
q
≤ 2v 2
b 
I {a ∈ Ne }
ra
a=1 b∈Π e∈pathT (b)

Xk X
qb 
≤ 8v 2 I a ∈ ∪e∈pathT (b) Ne
a=1
ra
b∈Π
k X
X qb
≤ 8v 2
a=1 b∈Π
rb
3 2
≤ 8k v .
Here, in the first inequality we used Eq. (37.15) and Eq. (37.17). The second
inequality follows by the definition of v = vloc and the choice of fe ∈ Eeloc , and
also because pa ≥ ra /2 by the condition on η in the theorem statement. The
third since any action a is in Ne for at most two edges in e ∈ pathT (b) (because
V (T ) ⊂ Π and it has no duplicates). The fourth inequality is true since r is
T -increasing and the fifth because r ≥ q/k. Finally, by Part (c) of Lemma 37.20
 37.8 Proof of the Classification Theorem 502

and the fact that Lλ is T -decreasing,

1 1−γ γ
(p − q)> Lλ = (r − q)> Lλ + (1/k − q)> Lλ ≤ k 2 v ≤ k 3 max(1, v 2 ) .
η η η
Combining the previous two displays shows that
 
(p − q)> Lλ 1 X X ηf (a, Φai )
d k
+ 2 λi p a Ψq ≤ 9k 3 max(1, v 2 ) .
η η i=1 a=1 pa

Since the right-hand side is independent of λ, the result follows from

Eq. (37.16).

37.8 Proof of the Classification Theorem

Almost all the results are now available to prove Theorem 37.11. In Section 37.4,
we showed that if G is globally observable and not locally observable, then
Rn∗ (G) = Ω(n2/3 ). We also proved that if G is locally observable and has
√
neighbours, then Rn∗ (G) = Ω( n). This last result is complemented by the policy
and analysis in Sections 37.5 to 37.7, where we showed that for globally observable
√
games Rn∗ (G) = O(n2/3 ) and for locally observable games Rn∗ (G) = O( n).
Finally we proved that if G is not globally observable, then Rn∗ (G) = Ω(n). All
that remains is to prove that if G has no neighbouring actions, then Rn∗ (G) = 0.
Theorem 37.22. If G has no neighbouring actions, then Rn∗ (G) = 0.
Proof Since G has no neighbouring actions, there exists an action a such that
Ca = Pd−1 and the policy that chooses At = a for all rounds suffers no regret.

37.9 Notes

1 The next three notes are covering some basic definitions and facts in linear
algebra. There are probably hundreds of introductory texts on linear algebra.
A short and intuitive exposition is by Axler [1997].
2 A non-empty set L ⊆ Rn is a linear subspace of Rn if αv + βw ∈ L for
all α, β ∈ R and v, w ∈ L. If L and M are linear subspaces of Rn , then
L ⊕ M = {v + w : L ∈ L, w ∈ M }. The orthogonal complement of linear
subspace L is L⊥ = {v ∈ Rn : hu, vi = 0 for all u ∈ L}. The following
properties are easily checked: (i) L⊥ is a linear subspace, (ii) (L⊥ )⊥ = L and
(iii) (L ∩ M )⊥ = L⊥ ⊕ M ⊥ .
3 Let A ∈ Rm×n be a matrix and recall that matrices of this form correspond
to linear maps from Rn → Rm where the function A : Rn → Rm is given by
matrix multiplication, A(x) = Ax. The image of A is im(A) = {Ax : x ∈ Rn },
and the kernel is ker(A) = {x ∈ Rn : Ax = 0}. Notice that im(A) ⊆ Rm and
ker(A) ⊆ Rn . One can easily check that im(A) and ker(A> ) are linear subspaces,
 37.9 Notes 503

and an elementary theorem in linear algebra says that im(A) ⊕ ker(A> ) = Rm

for any matrix A ∈ Rm×n . Finally, if u ∈ im(A) and v ∈ ker(A> ), then
hu, vi = 0.
4 Given a set A ⊆ Rd , the affine hull is the set
( j j
)
X X
aff(A) = αi xj : j > 0, α ∈ Rj , xi ∈ A for all i ∈ [j] and αi = 1 .
i=1 i=1

Its dimension is the smallest m such that there exist vectors v1 , . . . , vm ∈ Rd

such that aff(A) = x◦ + span(v1 , . . . , vm ) for any x◦ ∈ A.
5 We introduced the stochastic variant of partial monitoring to prove our lower
bounds. Of course our upper bounds also apply to this setting, which means the
classification theorem also holds in the stochastic case. The interesting question
is to understand the problem-dependent regret, which for partial monitoring
problem G = (L, Φ) is
" n #
X
Rn (π, u) = max E h`At − `a , Ut i ,
a∈[k]
t=1

where U, U1 , . . . , Un is a sequence of independent and identically distributed

random vectors with Ut ∈ {e1 , . . . , ed } and E[U ] = u ∈ Pd−1 . Provided G is
not hopeless, one can derive an algorithm for which the regret is logarithmic,
and like in bandits there is a sense of asymptotic optimality. The open research
question is to understand the in-between regime where the horizon is not yet
large enough that the asymptotically optimal logarithmic regret guarantees
become meaningful, but not so small that minimax is acceptable.
6 More generally, a stochastic partial monitoring problem by a probability
kernel (Pθ,a : θ ∈ Θ, a ∈ A) from (Θ × A, F ⊗ G) to (Σ × R, H ⊗ B(R)). The
environment chooses θ ∈ Θ, and the learner chooses (At )nt=1 with At ∈ A and
observes (σt )nt=1 in a sequential manner, where (σt , Xt ) ∼ Pθ,At (·). The reward
Xt of round t is unobserved. As before, the learner’s goal is to maximise the
total expected reward or, equivalently, to minimise regret. The special case
of the previous note is has been studied under the name of finite stochastic
partial monitoring.
7 The optimal tuning of the bound for Algorithm 26 depends on opt∗ (η), which
may be hard to compute. A simple way to address this problem is to use an
adaptive learning rate:
( s )
1 log(k)
ηt = min , Pt−1 ,
B 1 + s=1 Vs
where Vt = max{0, optQt (ηt )} and B is chosen large enough that η1 is
sufficiently small to satisfy the conditions needed in either Theorem 37.16
or Theorem 37.17. An excessively large B only affects the regret in an additive
fashion. The adaptive algorithm only needs to solve the optimisation in
Eq. (37.12) and not opt∗ (η). Another benefit of the adaptive algorithm is
 37.9 Notes 504

that it only depends on the game through the constant B. Furthermore, the
bound depends on (Vt )nt=1 , rather than opt∗ (η), which may sometimes be
beneficial. The analysis of the algorithm uses the same techniques as developed
in Exercise 28.13 and is given by Lattimore and Szepesvári [2019d].
8 Algorithm 26 can be modified in several ways. One enhancement is to drop the
constraint that f ∈ E vec in the optimisation problem and introduce the worst
case bias of f as a penalty. Certainly this does not make the bounds worse. A
more significant change is to introduce a moment-generating function into the
optimisation problem, which leads to high-probability bounds [Lattimore and
Szepesvári, 2019d].
9 The optimisation problem in Algorithm 26 is convex and can be solved using
standard solvers when k and d are small and η is not too small. When η is small
and/or k or d is large, then numerical instability is a real challenge. One way to
address this issue is to approximate the exponential in the definition of Ψq with
a quadratic and add constraints on p and f that ensure the approximation is
reasonable. Since the analysis uses p and f satisfying these conditions, none of
the theory changes. What is bought by this approximation is that the resulting
optimisation problem becomes a second order cone program, rather than an
exponential cone program, and these are better behaved. More details are in
our paper: [Lattimore and Szepesvári, 2019d].
10 Partial monitoring has many potential applications. We already mentioned
dynamic pricing and spam filtering. In the latter case, acquiring the true label
comes at a price, which is a typical component of hard partial monitoring
problems. In general, there are many set-ups where the learner can pay extra
for high-quality information. For example, in medical diagnosis the doctor can
request additional tests before recommending a treatment plan, but these cost
time and money. Yet another potential application is quality testing in factory
production where the quality control team can choose which items to test (at
great cost).
11 There are many possible extensions to the partial monitoring framework. We
have only discussed problems where the number of actions/feedbacks/outcomes
is potentially infinite, but nothing prevents studying a more general setting.
Suppose the learner chooses a sequence of real-valued outcomes i1 , . . . , in with
it ∈ [0, 1]. In each round, the learner chooses At ∈ [k] and observes ΦAt (it ),
where Φa : [0, 1] → Σ is a known feedback function. The loss is determined
by a collection of known functions La : [0, 1] → [0, 1]. We do not know of any
systematic study of this setting. The reader can no doubt imagine generalising
this idea to infinite action sets or introducing a linear structure for the loss.
12 A pair of Pareto-optimal actions (a, b) are called weak neighbours if
Ca ∩ Cb = 6 ∅ and pairwise observable if there exists a function g satisfying
Eq. (37.3) and with g(c, f ) = 0 whenever c ∈ / {a, b}. A partial monitoring
problem is called a point-locally observable game if all weak neighbours are
pairwise observable. All point-locally observable games are locally observable,
but the converse is not true. Bartók [2013] designed a policy for this type of
 37.9 Notes 505

game for which

1 p
Rn ≤ kloc n log(n) ,
εG

where εG > 0 is a game-dependent constant and kloc is the size of the largest
A ⊆ [k] of Pareto optimal actions such that ∩a∈A Ca = 6 ∅. Using a different
policy, Lattimore and Szepesvári [2019a] have shown that as the horizon grows,
the game-dependence diminishes so that

Rn p
lim sup √ ≤ 8(2 + m) 2kloc log(k) .
n→∞ n

13 Linear regret is unavoidable in hopeless games, but that does not mean there
is nothing to play for. Rustichini [1999] considered a version of the regret that
captures the performance of policies in this harsh setting. Given p ∈ Pd−1
define set I(p) ⊆ Pd−1 by
( d
)
X
I(p) = q ∈ Pd−1 : (pi − qi )I {Φai = f } = 0 for all a ∈ [k] and f ∈ [m] .
i=1

This is the set of distributions over the outcomes that are indistinguishable
from p by the learner using any actions. Then define
d
X
f (p) = max min qi Lai .
q∈I(p) a∈[k]
i=1

Rustichini [1999] proved there exist policies such that

" #
1X
n
lim max E LAt it − f (ūn ) = 0 ,
n→∞ i1:n n t=1
Pn
where ūn = n1 t=1 eit ∈ Pd−1 is the average outcome chosen by the adversary.
Intuitively this means the learner does not compete with the best action in
hindsight with respect to the actual outcomes. Instead, the learner competes
with the best action in hindsight with respect to an outcome sequence that is
indistinguishable from the actual outcome sequence. Rustichini did not prove
rates on the convergence of the limit. This has been remedied recently, and we
give some references in the bibliographic remarks.
14 Partial monitoring is still quite poorly understood. With some exceptions, we
do not know how the regret should depend on d, k, m or the structure of
G. Lower bounds that depend on these quantities are also missing, and the
lower bounds proven in Section 37.4 are surely very conservative. We hope this
chapter inspires more activity in this area. The setting described in Note 13 is
even more wide open, where the dependence on n is still not nailed down.
 37.10 Bibliographical Remarks 506

37.10 Bibliographical Remarks

The first work on partial monitoring is by Rustichini [1999], who focussed on

finding Hannan consistent policies in the adversarial setting. Rustichini shows
how to reduce the problem to Blackwell approachability (see Cesa-Bianchi and
Lugosi [2006]) and uses this to deduce the existence of a Hannan consistent
strategy. Rustichini also used a refined notion of regret that allows one to
distinguish between learners even in the case of hopeless games (see Note 13).
The first non-asymptotic result in the setting of this chapter is due to Piccolboni
and Schindelhauer [2001], who derive a policy with regret O(n3/4 ) for globally
observable games. Cesa-Bianchi et al. [2006] reduced the dependence to O(n2/3 )
and proved a wide range of other results for specific classes of problems. The first
O(n1/2 ) bound for non-degenerate locally observable games is due to Foster and
Rakhlin [2012]. The classification theorem when d = 2 is due to Bartók et al.
[2010] (extended version: Antos et al. [2013]). With the exception of degenerate
games, the classification of adversarial partial monitoring games is by Bartók
et al. [2014]. The case of degenerate games was resolved by the present authors
[Lattimore and Szepesvári, 2019a]. The policies mentioned in Note 12 are due to
Bartók [2013] and Lattimore and Szepesvári [2019a]. We warn the reader that
neighbours are defined differently by Foster and Rakhlin [2012] and Bartók [2013],
which can lead to confusion. Additionally, although both papers are largely
correct, in both cases the core proofs contain errors that cannot be resolved
without changing the policies [Lattimore and Szepesvári, 2019a]. Algorithm 26
and its analysis is also by the present authors [Lattimore and Szepesvári, 2019d],
which is a followup on an earlier information-theoretic analysis [Lattimore and
Szepesvári, 2019c].
There is a growing literature on the stochastic setting where it is common to
study both minimax and asymptotic bounds. In the latter case, one can obtain
asymptotically optimal logarithmic regret for games that are not hopeless. We
refer the reader to papers by Bartók et al. [2012], Vanchinathan et al. [2014]
and Komiyama et al. [2015b] as a good starting place. As we mentioned, partial
monitoring can model problems that lie between bandits and full information.
There are now several papers on this topic, but in more restricted settings and
consequentially with more practical algorithms and bounds. One such model is
when the learner is playing actions corresponding to vertices on a graph and
observes the losses associated with the chosen vertex and its neighbours [Mannor
and Shamir, 2011, Alon et al., 2013]. A related result is in the finite-armed
Gaussian setting where the learner selects an action At ∈ [k] and observes a
Gaussian sample from each arm, but with variances depending on the chosen
action. Like partial monitoring, this problem exhibits many challenges and is
not yet well understood [Wu et al., 2015]. We mentioned in Note 13 that for
hopeless games, the definition of the regret can be refined. A number of authors
have studied this setting and proved sublinear regret guarantees. As usual, the
price of generality is that the bounds are correspondingly a bit worse [Mannor
 37.11 Exercises 507

and Shimkin, 2003, Perchet, 2011, Mannor et al., 2014]. There has been some
work on infinite partial monitoring games. Lin et al. [2014] study a stochastic
setting with finitely many actions, but infinitely many outcomes and a particular
linear structure for the feedback. Chaudhuri and Tewari [2016] also consider a
linear setting with global observability and prove O(n2/3 log(n)) regret using
an explore-then-commit algorithm. Kirschner et al. [2020] study a version of
information-directed sampling in partial monitoring setting with a linear feedback
structure and finitely or infinitely many actions.
One can also add context, as usual. The special case of stochastic finite
contextual partial monitoring has been considered by Bartók and Szepesvári
[2012]. In this version, the learner is still given the matrices (L, Φ), but also a set of
functions F that map a sequence (xt )t of contexts to outcome distributions, with
the assumption that the outcome in round t is generated from f (xt ) with f ∈ F
unknown to the learner. A special case, apple tasting with context (equivalently,
matching pennies with context) is the subject of the paper of Helmbold et al.
[2000]. The aforementioned paper by Kirschner et al. [2020] also studies the
contextual partial monitoring problem in a linear setting.

37.11 Exercises

37.1 (Perspective) Prove that the perspective as defined in Eq. (37.14) is

convex.

37.2 (Affine sets and dimension) Let X ⊆ Y ⊆ Rd and dim(X ) = dim(Y).

Prove that aff(X ) = aff(Y).

37.3 (Modified kernel) Recall that ker0 (x) = {u : u> x = 0 and u> 1 = 1}.
Show that if ker0 (x) = ker0 (y) 6= ∅ then x and y are proportional.

37.4 (Structure of examples) Calculate the neighbourhood structure, cell

decomposition and action classification for each of the examples in this chapter.

37.5 (Apple tasting) Apples arrive sequentially from the farm to a processing
facility. Most apples are fine, but occasionally there is a rotten one. The only way
to figure out whether an apple is good or rotten is to taste it. For some reason
customers do not like bite marks in the apples they buy, which means that tested
apples cannot be sold. Good apples yield a unit reward when sold, while the sale
of a bad apple costs the company c > 0.

(a) Formulate this problem as a partial monitoring problem: determine L and Φ.

(b) What is the minimax regret in this problem?
(c) What do you think about this problem? Will actual farmers be excited about
your analysis?

37.6 (Two-action partial monitoring games are trivial, hopeless or

 37.11 Exercises 508

easy) Let G = (L, Φ) be a partial monitoring game with k = 2 actions. Prove

that G is either trivial, hopeless or easy.

37.7 (Complete lower bound for hard games) Complete the last step in
the proof of Theorem 37.12.

37.8 (Lower bound for easy games) Prove Theorem 37.14.

37.9 (Lower bound for hopeless games) Prove Theorem 37.13.

37.10 Let a and b be non-duplicate Pareto optimal actions and λ ∈ ri(Ca ). Show
there exists an ω ∈ ri(Cb ) and neighbour c of b such that Cc ∩ [ω, λ] 6= ∅.

Hint It may be useful to look at Fig. 37.6 to get some tips. The figure depicts
a slightly different situation, but is still useful when it is changed a little.
37.11 Generalise the proofs of Lemma 37.20 and Lemma 37.21 to handle duplicate
and degenerate actions.

37.12 Prove Proposition 37.18.

Hint For Part (a), let S ∈ Rkm×d be obtained by stacking (Sc )kc=1 , defined√as
in the proof of Theorem 37.12. Then argue that for globally observable games, d
times the reciprocal of the smallest non-zero singular value of S is an upper bound
on vglo and then use the fact that S > S has integer-valued coefficients. Part (b)
follows in a similar fashion. For Part (c), use a graph-theoretic argument.
37.13 Let m = |Σ| = 2 and d = 2k − 1 and construct a globally observable game
for which there exists a pair of neighbouring actions a, b for which
min kf k ≥ C2d/2 ,
glo
f ∈Eab

where C > 0 is a universal constant.

37.14 Prove Proposition 37.19.

Hint Find choices of p and f that reduce the algorithm to Exp3 and exponential
weights respectively.
37.15 (Lower bound depending on the number of feedbacks) Consider
G = (L, Φ) given by
 
1 0 1 0 ··· 1 0
L=  and
0 1 0 1 ··· 0 1
 
1 2 2 3 3 4 ··· m−1 m−1 m 
Φ= .
1 1 2 2 3 3 ··· m−2 m−1 m−1
 37.11 Exercises 509

opt∗ (η)

0
0.2 0.4 0.6 0.8 1
c

Figure 37.7 The value of opt∗ (η) as a function of c in matching pennies (Example 37.3).

(a) Show this game is locally observable.

(b) Prove that for n ≥ m, there exists a universal constant c > 0 such that
√
Rn∗ (G) ≥ c(m − 1) n.

The source for previous exercise is the paper by the authors [Lattimore and
Szepesvári, 2019a].

37.16 (Divergence decomposition for partial monitoring) Complete the

necessary modification of Lemma 15.1 to show that Eq. (37.8) is true.

37.17 (Algorithm for classifying games) Write a program that accepts as

input matrices L and Φ and outputs the classification of the game.

37.18 (Implementation (i)) Implement a solver for the optimisation problem

in Eq. (37.12). Consider the matching pennies problem (Example 37.3). Let
η = 1/100 and plot opt∗ (η) as a function of the cost c. Explain your results.

Hint The convex optimisation problem in Eq. (37.12) seems to cause problems
for some solvers (see Note 9 for some mitigating strategies). We assume that
many libraries can be made to work. Our implementation used the splitting cone
solver by O’Donoghue et al. [2016, 2017]. Your plot should resemble Fig. 37.7.

37.19 (Implementation (ii)) In this exercise you will compare empirically or

otherwise Algorithm 26 to exponential weights and Exp3 on full information and
bandit games. Specifically:

(a) For full information games, exponential weights behaves like Algorithm 26
except that ŷt = yt and Pt = Qt . Does the solution to the optimisation
 37.11 Exercises 510

problem used by Algorithm 26 lead to the same loss estimators and

distribution Pt ?
(b) For bandits, Exp3 uses ŷta = yta I {At = a} /Pta . Does Algorithm 26 end up
using the same loss estimators? Does Pt = Qt ?

Hint You can approach this problem by using your solution to Exercise 37.18
and comparing values empirically. Alternatively, you can theoretically analyse
Eq. (37.12) in these special cases. Some of these questions are answered by
Lattimore and Szepesvári [2019d].
38 Markov Decision Processes

Bandit environments are a sensible model for many simple problems, but they do
not model more complex environments where actions have long-term consequences.
A brewing company needs to plan ahead when ordering ingredients, and the
decisions made today affect their position to brew the right amount of beer in
the future. A student learning mathematics benefits not only from the immediate
reward of learning an interesting topic but also from their improved job prospects.
A Markov decision process (MDP) is a simple way to incorporate long-term
planning into the bandit framework. Like in bandits, the learner chooses actions
and receives rewards. But they also observe a state, and the rewards for different
actions depend on the state. Furthermore, the actions chosen affect which state
will be observed next.

38.1 Problem Set-Up

An MDP is defined by a tuple M = (S, A, P, r, µ). The first two items S and A
are sets called the state space and action space, and S = |S| and A = |A| are
their sizes, which may be infinite. An MDP is finite if S, A < ∞. The quantity
P = (Pa : a ∈ A) is called the transition function with Pa : S × S → [0, 1]
so that Pa (s, s0 ) is the probability that the learner transitions from state s to
s0 when taking action a. The fourth element of the tuple is r = (ra : a ∈ A),
which is a collection of reward functions with ra : S → [0, 1]. When the learner
takes action a in state s, it receives a deterministic reward of ra (s). The last
element is µ ∈ P(S), which is a distribution over the states that determines
the starting state. The transition and reward functions are often represented by
vectors or matrices. When the state space is finite, we may assume without loss
of generality that S = [S]. We write Pa (s) ∈ [0, 1]S as the probability vector with
s0 th coordinate given by Pa (s, s0 ). In the same way, we let Pa ∈ [0, 1]S×S be the
right stochastic matrix with (Pa )s,s0 = Pa (s, s0 ). Finally, we view ra as a vector
in [0, 1]S in the natural way.
The interaction protocol is similar to bandits. Before the game starts, the initial
state S1 is sampled from µ. In each round t, the learner observes the state St ∈ S,
chooses an action At ∈ A and receives reward rAt (St ). The environment then
 38.1 Problem Set-Up 512

samples St+1 from the probability vector PAt (St ), and the next round begins
(Fig. 38.1).

t = 1 and sample S1 ∼ µ

Observe state St

Choose action At ∈ A Increment t

Receive reward rAt (St ) Update St+1 ∼ PAt (St )

Figure 38.1 Interaction protocol for Markov decision processes.

Although the action set is the same in all states, this does not mean that
Pa (s) or ra (s) has any relationship to Pa (s0 ) or ra (s0 ) for states s 6= s0 . In
this sense, it might be better to use an entirely different set of actions for
each state, which would not change the results we present. And while we
are at it, of course one could also allow the number of actions to vary over
the state space.

Histories and Policies

Before considering the learning problem, we explain how to act in a known MDP.
Because there is no learning going on, we call our protagonist the ‘agent’ rather
than ‘learner’. In a stochastic bandit, the optimal policy given knowledge of the
bandit is to choose the action with the largest expected reward in every round.
In an MDP, the definition of optimality is less clear.
The history Ht = (S1 , A1 , . . . , St−1 , At−1 , St ) in round t contains the
information available before the action for the round is to be chosen. Note
that state St is included in Ht . The actions are also included because the agent
may randomise. For simplicity the rewards are omitted because the all-knowing
agent can recompute them if needed from the state-action pairs.
A policy is a (possibly randomised) map from the set of possible histories
to actions. Simple policies include memoryless policies, which choose actions
based on only the current state, possibly in a randomised manner. The set
of such policies is denoted by ΠM , and its elements are identified with maps
P
π : A × S → [0, 1] with a∈A π(a | s) = 1 for any s ∈ S so that π(a | s) is
interpreted as the probability that policy π takes action a in state s.
 38.1 Problem Set-Up 513

A memoryless policy that does not randomise is called a memoryless

deterministic policy. To reduce clutter, such policies are written as S → A
maps, and the set of all such policies is denoted by ΠDM . A policy is called a
Markov policy if the actions are randomised and depend only on the round
index and the previous state. These policies are represented by fixed sequences of
memoryless policies. Under a Markov policy, the sequence of states (S1 , S2 , . . . )
evolve as a Markov chain (see Section 3.2). If the Markov policy is memoryless,
this chain is homogeneous.

1, 1
1, 1

1 trap state high reward state

2 2 2 2
2 5,0 5,0 5,0 5 , 10
5,0 3
5 , 10
3 3 3 3
5,0 5,0 5,0 5,0

2 3 4 5 6

1 1 1 1
5,1 5,1 5,1 5,1
1
4 5 , 12
5,1 4 4 4 4
5,1 5,1 5,1 5 , 12

Figure 38.2 A Markov decision process with six states and two actions represented by
solid and dashed arrows, respectively. The numbers next to each arrow represent the
probability of transition and reward for the action respectively. For example, taking the
solid action in state 3 results in a reward of 0, and the probability of moving to state
4 is 3/5, and the probability of moving to state 3 is 2/5. For human interpretability
only, the actions are given consistent meaning across the states (blue/solid actions
‘increment’ the state index, black/dashed actions decrement it). In reality there is no
sense of similarity between states or actions built into the MDP formalism.

Probability Spaces
It will be convenient to allow infinitely long interactions between the learner and
the environment. In line with Fig. 38.1, when the agent or learner follows a policy
π in MDP M = (S, A, P, r, µ), such a never-ending interaction gives rise to a
random process (S1 , A1 , S2 , A2 , . . . ) so that for any s, s0 ∈ S, a ∈ A and t ≥ 1,
(a) P(S1 = s) = µ(s);
(b) P(St+1 = s0 | Ht , At ) = PAt (St , s0 ); and
(c) P(At = a | Ht ) = π(a | Ht ).
Meticulous readers may wonder whether there exists a probability space (Ω, F, P)
holding the infinite sequence of random variables (S1 , A1 , S2 , A2 , . . . ) that satisfy
(a)–(c). The Ionescu–Tulcea theorem (Theorem 3.3) furnishes us with a positive
answer (Exercise 38.1). Item (b) above is known as the Markov property. Of
 38.1 Problem Set-Up 514

course the measure P depends on the policy, Markov decision process and the
initial distribution. For most of the chapter, these quantities will be fixed and the
dependence is omitted from the notation. In the few places where disambiguation
is necessary, we provide additional notation. In addition to this, to minimise
clutter, we allow ourselves to write P(· | S1 = s), which just means the probability
distribution that results from the interconnection of π and M , while replacing µ
with an alternative initial state distribution that is a Dirac at s.

Traps and the Diameter of a Markov Decision Process

A significant complication in MDPs is the potential for traps. A trap is a subset of
the state space from which there is no escape. For example, the MDP in Fig. 38.2
has a trap state. If being in the trap has a suboptimal yield in terms of the
reward, the learner should avoid the trap. But since the learner can only discover
that an action leads to a trap by trying that action, the problem of learning while
competing with a fully informed agent is hopeless (Exercise 38.28).
To avoid this complication, we restrict our attention to MDPs with no traps.
An MDP is called strongly connected or communicating if for any pair of
states s, s0 ∈ S, there exists a policy such that when starting from s there is a
positive probability of reaching s0 some time in the future while following the
policy. One can also define a real-valued measure of the connectedness of an MDP
called the diameter. MDPs with smaller diameter are usually easier to learn
because a policy can recover from mistakes more quickly.

Definition 38.1. The diameter of an MDP M is

D(M ) = max0 min Eπ [min{t ≥ 1 : St = s} | S1 = s0 ] − 1 ,

s6=s π∈ΠDM

where the expectation is taken with respect to the law of Markov chain (St )∞
t=1
induced by the interaction between π and M .

A number of observations are in order about this definition. First, the order
of the maximum and minimum means that for any pair of states a different
policy may be used. Second, travel times are always minimised by deterministic
memoryless policies, so the restriction to these policies in the minimum is
inessential (Exercise 38.3). Finally, the definition only considers distinct states.
We also note that when the number of states is finite, it holds that D(M ) < ∞ if
and only if M is strongly connected (Exercise 38.4). The diameter of an MDP
with S states and A actions cannot be smaller than logA (S) − 3 (Exercise 38.5).

For the remainder of this chapter, unless otherwise specified, all MDPs are
assumed to be strongly connected.
 38.2 Optimal Policies and the Bellman Optimality Equation 515

38.2 Optimal Policies and the Bellman Optimality Equation

We now define the notion of an optimal policy and outline the proof that there
exists a deterministic memoryless optimal policy. Along the way, we define what is
called the Bellman optimality equation. Methods that solve this equation are the
basis for finding optimal policies in an efficient manner and also play a significant
role in learning algorithms. Throughout, we fix a strongly connected MDP M .
The gain of a policy π is the long-term average reward expected from using
that policy when starting in state s:

1X π
n
ρπs = lim E [rAt (St ) | S1 = s] ,
n→∞ n
t=1

where Eπ denotes the expectation on the interaction sequence when policy π

interacts with MDP M . In general, the limit need not exist, so we also introduce

1X π
n
ρ̄πs = lim sup E [rAt (St ) | S1 = s] ,
n→∞ n t=1

which exists for any policy. Of course, whenever ρπs exists we have ρπs = ρ̄πs . The
optimal gain is a real value

ρ∗ = max sup ρ̄πs ,

s∈S π

where the supremum is taken over all policies. A π policy is an optimal policy
if ρπ = ρ∗ 1. For strongly connected MDPs, an optimal policy is guaranteed to
exist. This is far from trivial, however, and we will spend the next little while
outlining the proof.

MDPs that are not strongly connected may not have a constant optimal
gain. This makes everything more complicated, and we are lucky not to have
to deal with such MDPs here.

Before continuing, we need some new notation. For a memoryless policy π, define
X X
Pπ (s, s0 ) = π(a | s)Pa (s, s0 ) and rπ (s) = π(a | s)ra (s) . (38.1)
a∈A a∈A

We view Pπ as an S × S transition matrix and rπ as a vector in RS . With this

notation, Pπ is the transition matrix of the homogeneous Markov chain S1 , S2 , . . .
when At ∼ π(· | St ). The gain of a memoryless policy π satisfies

1 X t−1
n
ρπ = lim Pπ rπ = Pπ∗ rπ , (38.2)
n→∞ n
t=1

1
Pn
where Pπ∗ = limn→∞ n t=1 Pπt−1 is called the stationary transition matrix,
 38.2 Optimal Policies and the Bellman Optimality Equation 516

the existence of which you will prove in Exercise 38.7. For each k ∈ N, define
k
X
vπ(k) = Pπt−1 (rπ − ρπ ) .
t=1
(k)
For s ∈ S, vπ (s) gives the total expected excess reward collected by π when
the process starts at state s and lasts for k time steps. The (differential) value
function of a policy is a function vπ : S → R defined as the Cesàro sum of the
sequence (Pπt (rπ − ρπ ))t≥0 ,
1 X (k)
n
vπ = lim vπ = ((I − Pπ + Pπ∗ )−1 − Pπ∗ )rπ . (38.3)
n→∞ n
k=1

Note, the second equality above is non-trivial (Exercise 38.7). The definition
implies that vπ (s) − vπ (s0 ) is the ‘average’ long-term advantage of starting in
state s relative to starting in state s0 when following policy π. These quantities
are only defined for memoryless policies where they are also guaranteed to exist
(Exercise 38.7). The definition of Pπ∗ implies that Pπ∗ Pπ = Pπ∗ , which in turn
implies that Pπ∗ vπ = 0. Combining this with Eqs. (38.2) and (38.3) shows that
for any memoryless policy π,
ρπ + vπ = rπ + Pπ vπ . (38.4)
A value function is a function v : S → R, and its span is given by
span(v) = max v(s) − min v(s) .
s∈S s∈S

As with other quantities, value functions are associated with vectors in RS . A

greedy policy with respect to value function v is a deterministic memoryless
policy πv given by
πv (s) = argmaxa∈A ra (s) + hPa (s), vi .
There may be many policies that are greedy with respect to some value function v
due to ties in the maximum. Usually the ties do not matter, but for consistency
and for the sake of simplifying matters, we assume that ties are broken in a
systematic fashion. In particular, this makes πv well defined for any value function.
One way to find the optimal policy is as the greedy policy with respect to a
value function that satisfies the Bellman optimality equation, which is
ρ + v(s) = max (ra (s) + hPa (s), vi) for all s ∈ S . (38.5)
a∈A

This is a system of S nonlinear equations with unknowns ρ ∈ R and v ∈ RS .

The reader will notice that if v : S → R is a solution to Eq. (38.5), then so is
v + c1 for any constant c ∈ R, and hence the Bellman optimality equation lacks
unique solutions. It is not true that the optimal value function is unique up to
translation, even when M is strongly connected (Exercise 38.11). The v-part of
a solution pair (ρ, v) of Eq. (38.5) is called an optimal (differential) value
function.
 38.2 Optimal Policies and the Bellman Optimality Equation 517

Theorem 38.2. The following hold:

(a) There exists a pair (ρ, v) that satisfies the Bellman optimality equation.
(b) If (ρ, v) satisfies the Bellman optimality equation, then ρ = ρ∗ and πv is
optimal.
(c) There exists a deterministic memoryless optimal policy.

Proof sketch The proof of part (a) is too long to include here, but we guide you
through it in Exercise 38.10. For part (b), let (ρ, v) satisfy the Bellman equation
and π ∗ = πv be the greedy policy with respect to v. Then, by Eq. (38.2),

1 X t−1 1 X t−1
n n
ρπ = lim Pπ∗ rπ∗ = lim Pπ∗ (ρ1 + v − Pπ∗ v) = ρ1 .
∗

n→∞ n n→∞ n
t=1 t=1

Next, let π be an arbitrary Markov policy. We show that ρ̄π ≤ ρ1. The result is
then completed using the result of Exercise 38.2, where you will prove that for any
policy π, there exists a Markov policy with the same expected rewards. Denote
by πt the memoryless policy used at time t = 1, 2, . . . when following the Markov
(t) (0)
policy π, and for t ≥ 1, let Pπ = Pπ1 . . . Pπt , while for t = 0, let Pπ = I. Thus,
(t)
Pπ (s, s ) is the probability of ending up in state s while following π from state
0 0
Pn (t−1)
s for t time steps. It follows that ρ̄π = lim supn→∞ n1 t=1 Pπ rπt . Fix t ≥ 1.
Using the fact that π ∗ is the greedy policy with respect to v gives

Pπ(t−1) rπt = Pπ(t−1) (rπt + Pπt v − Pπt v)

≤ Pπ(t−1) (rπ∗ + Pπ∗ v − Pπt v)
= Pπ(t−1) (ρ1 + v − Pπt v)
= ρ1 + Pπ(t−1) v − Pπ(t) v .

Taking the average of both sides over t ∈ [n] and then taking the limit shows
that ρ̄π ≤ ρ1, finishing the proof. Part (c) follows immediately from the first
two parts.

The theorem shows that there exist solutions to the Bellman optimality equation
and that the greedy policy with respect to the resulting value function is an
optimal policy. We need one more result about solutions to the Bellman optimality
equation, the proof of which you will provide in Exercise 38.13.

Lemma 38.3. Suppose that (ρ, v) satisfies the Bellman optimality equation. Then
span(v) ≤ D(M ).

The map T : RS → RS defined by (T v)(s) = maxa∈A ra (s) + hPa (s), vi is

called the Bellman operator. The Bellman optimality equation can be
written as ρ1 + v = T v.
 38.3 Finding an Optimal Policy ( ) 518

38.3 Finding an Optimal Policy ( )

There are many ways to find an optimal policy, including value iteration, policy
iteration and enumeration. These ideas are briefly discussed in Note 12. Here
we describe a two-step approach based on linear programming. Consider the
following constrained linear optimisation problem:

minimise ρ (38.6)
ρ∈R,v∈RS

subject to ρ + v(s) ≥ ra (s) + hPa (s), vi for all s, a .

Recall that a constrained optimisation problem is said to be feasible if the set

of values that satisfy the constraints are non-empty.

Theorem 38.4. The optimisation problem in Eq. (38.6) is feasible, and if (ρ, v)
is a solution, then ρ = ρ∗ is the optimal gain.

Solutions (ρ, v) to the optimisation problem in Eq. (38.6) need not satisfy
the Bellman optimality equation (Exercise 38.12).

Proof of Theorem 38.4 Theorem 38.2 guarantees the existence of a pair (ρ∗ , v ∗ )
that satisfies the Bellman optimality equation:

ρ∗ + v ∗ (s) = max ra (s) + hPa (s), v ∗ i for all s, a .

a∈A

Hence the pair (ρ , v ) satisfies the constraints in Eq. (38.6) and witnesses
∗ ∗

feasibility. Next, let (ρ, v) be a solution of Eq. (38.6). Since (ρ∗ , v ∗ ) satisfies the
constraints, ρ ≤ ρ∗ is immediate. It remains to prove that ρ ≥ ρ∗ . Let π = πv
be the greedy policy with respect to v and π ∗ be greedy with respect to v ∗ . By
Theorem 38.2, ρ∗ = ρπ . Furthermore,
∗

Pπt ∗ rπ∗ ≤ Pπt ∗ (rπ + Pπ v − Pπ∗ v) ≤ Pπt ∗ (ρ1 + v − Pπ∗ v) = ρ1 + Pπt ∗ v − Pπt+1
∗ v .

Pn−1
Summing over t shows that ρ∗ 1 = limn→∞ n1 t=0 Pπt ∗ rπ∗ ≤ ρ1, which completes
the proof.

Having found the optimal gain, the next step is to find a value function that
satisfies the Bellman optimality equation. Let s̃ ∈ S, and consider the following
linear program:

minimise hv, 1i (38.7)

v∈RS

subject to ρ∗ + v(s) ≥ ra (s) + hPa (s), vi for all s, a

v(s̃) = 0 .

The second constraint is crucial in order for the minimum to exist, since otherwise
the value function can be arbitrarily small.
 38.3 Finding an Optimal Policy ( ) 519

Theorem 38.5. There exists a state s̃ ∈ S such that the solution v of Eq. (38.7)
satisfies the Bellman optimality equation.

Proof The result follows by showing that ε = v + ρ∗ 1 − T v = 0. The first

constraint in Eq. (38.7) ensures that ε ≥ 0. It remains to show that ε ≤ 0. Let
π ∗ be an optimal policy and π be the greedy policy with respect to v. Then

Pπt ∗ rπ∗ ≤ Pπt ∗ (rπ + Pπ v − Pπ∗ v) = Pπt ∗ (ρ∗ 1 + v − ε − Pπ∗ v) .

Hence ρ∗ 1 = ρπ 1 ≤ ρ∗ 1 − Pπ∗∗ ε and Pπ∗∗ ε ≤ 0. Since ε ≥ 0 and Pπ∗∗ is right

∗

stochastic, Pπ∗∗ ε = 0. Choose s̃ to be a state such that Pπ∗∗ (s, s̃) > 0 for some s ∈ S,
which exists because Pπ∗∗ is right stochastic. Then 0 = (Pπ∗∗ ε)(s) ≥ Pπ∗∗ (s, s̃)ε(s̃)
and hence ε(s̃) = 0. It follows that ṽ = v − ε also satisfies the constraints in
Eq. (38.7). Because v is a solution to Eq. (38.7), hṽ, 1i ≥ hv, 1i, implying that
hε, 1i ≤ 0. Since we already showed that ε ≥ 0, it follows that ε = 0.

The theorem only demonstrates the existence of a state s̃ for which the solution
of Eq. (38.7) satisfies the Bellman optimality equation. There is a relatively
simple procedure for finding such a state using the solution to Eq. (38.6), but its
analysis depends on the basic theory of duality from linear programming, which
is beyond the scope of this text. More details are in Note 11 at the end of the
chapter. Instead we observe that one can simply solve Eq. (38.7) for all choices
of s̃ and take the first solution that satisfies the Bellman optimality equation.

38.3.1 Efficient Computation

The linear programs in Eq. (38.6) and Eq. (38.7) can be solved efficiently under
assumptions that will be satisfied in subsequent applications.

The algorithm proposed in this subsection is guaranteed to run in polynomial

time, which is a standard objective in theoretical computer science. Its
practical performance, however, is usually much worse than alternatives that
suffer from exponential running time in the worst case. These issues are
discussed in Note 12 at the end of the chapter.

The general form of a linear program is an optimisation problem of the form

minimise
n
hc, xi
x∈R

subject to Ax ≥ b ,

where c ∈ Rn and A ∈ Rm×n and b ∈ Rm are parameters of the problem. This

general problem can be solved in time that depends polynomially on n and m.
When m is very large or infinite, these algorithms may become impractical, but
nevertheless one can often still solve the optimisation problem in time polynomial
 38.3 Finding an Optimal Policy ( ) 520

in n only, provided that the constraints satisfy certain structural properties. Let
K ⊂ Rn be convex, and consider the optimisation problem

minimise
n
hc, xi (38.8)
x∈R

subject to x ∈ K .

Algorithms for this problem generally have a slightly different flavour because K
may have no corners. Suppose the following holds:

(a) There exists a known R > 0 such that K ⊂ {x ∈ Rn : kxk2 ≤ R}.

(b) There exists a separation oracle, which we recall from Chapter 27, is a
computational procedure to evaluate some function φ on Rn with φ(x) =
true for x ∈ K, and otherwise φ(x) = u with hy, ui > hx, ui for all y ∈ K
(see Fig. 27.1).
(c) There exists a δ > 0 and x0 ∈ Rd such that {x ∈ Rn : kx − x0 k2 ≤ δ} ⊂ K.

Under these circumstances, the ellipsoid method accepts as input the size of the
bounding sphere R, the separation oracle and an accuracy parameter ε > 0. Its
output is a point x in time polynomial in n and log(R/(δε)) such that x ∈ K and
hc, xi ≤ hc, x∗ i + ε, where x∗ is the minimiser of Eq. (38.8). The reader can find
references to this method at the end of the chapter.
The linear programs in Eq. (38.6) and Eq. (38.7) do not have bounded feasible
regions because if v is feasible, then v + c1 is also feasible for any c ∈ R. For
strongly connected MDPs with diameter D, however, Lemma 38.3 allows us to
add the constraint that kvk∞ ≤ D. If the rewards are bounded in [0, 1], then we
may also add the constraint that 0 ≤ ρ ≤ 1. Together these imply that for (ν, ρ)
in the feasible region,

k(ρ, v)k22 = ρ2 + kvk22 ≤ 1 + Skvk2∞ ≤ 1 + SD2 .

√
Then set R = 1 + D2 S. When the diameter is unknown, one may use a doubling
procedure. In order to guarantee the feasible region contains a small ball, we
add some slack to the constraints. Let ε > 0, and consider the following linear
program:

minimise ρ (38.9)
ρ∈R,v∈RS

subject to ε + ρ + v(s) ≥ ra (s) + hPa (s), vi for all s, a .

v(s) ≥ −D for all s
v(s) ≤ D for all s
ρ ≤ 1 + ε for all s
ρ ≥ −ε for all s .

Note that for any x in the feasible region of Eq. (38.9), there exists a y that is
feasible for Eq. (38.6) with kx − yk∞ ≤ ε. Furthermore, the solution to the above
linear program is at most ε away from the solution to Eq. (38.6). What we have
 38.4 Learning in Markov Decision Processes 521

bought by adding this slack is that now the linear program in Eq. (38.9) satisfies
the conditions (a) and (c) above. The final step is to give a condition when a
separation oracle exists for the convex set determined by the constraints in the
above program. Define convex set K by

K = {(ρ, v) ∈ Rd+1 : ε + ρ + v(s) ≥ ra (s) + hPa (s), vi for all s, a} . (38.10)

Assuming that

argmaxa∈A (ra (s) + hPa (s), vi) (38.11)

can be solved efficiently, Algorithm 27 provides a separation oracle for K. For the
specialised case considered later, Eq. (38.11) is trivial to compute efficiently. The
feasible region defined by the constraints in Eq. (38.9) is the intersection of K with
a small number of half-spaces. In Exercise 38.15, you will show how to efficiently
Tn
extend a separation oracle for arbitrary convex set K to i=1 Hk ∩ K, where
(Hk )nk=1 are half-spaces. You will show in Exercise 38.14 that approximately
solving Eq. (38.7) works in the same way as above, as well as the correctness of
Algorithm 27.

In Theorem 38.2, we assumed an exact solution of the Bellman optimality

equation, which may not be possible in practice. Fortunately, approximate
solutions to the Bellman optimality equation with approximately greedy
policies yield approximately optimal policies. Details are deferred to
Exercise 38.16.

1: function SeparationOracle(ρ, v)
2: For each s ∈ S find a∗s ∈ argmaxa (ra (s) + hPa (s), vi)
3: if ε + ρ + v(s) ≥ ra∗s (s) + hPa∗s (s), vi for all s ∈ S then
4: return true
5: else
6: Find state s with ε + ρ + v(s) < ra∗s (s) + hPa∗s (s), vi
7: return (1, es − Pa∗s (s))
8: end if
9: end function
Algorithm 27: Separation oracle for Eq. (38.6).

38.4 Learning in Markov Decision Processes

The problem of finding an optimal policy in an unknown MDP is no longer just

an optimisation problem, and the notion of regret is introduced to measure the
price of the uncertainty. For simplicity we assume that only the transition matrix
 38.5 Upper Confidence Bounds for Reinforcement Learning 522

is unknown while the reward function is given. This assumption is not especially
restrictive as the case where the rewards are also unknown is easily covered using
either a reduction or a simple generalisation, as we explain in the notes. The
regret of a policy π is the deficit of rewards suffered relative to the expected
average reward of an optimal policy:
n
X
R̂n = nρ∗ − rAt (St ) .
t=1

The reader will notice we are comparing the non-random nρ∗ to the random
sum of rewards received by the learner, which was also true in the study of
stochastic bandits. The difference is that ρ∗ is an asymptotic quantity while for
stochastic bandits the analogous quantity was nµ∗ . The definition stills makes
sense, however, because for MDPs with finite diameter D the optimal expected
cumulative reward over n rounds is at least nρ∗ − D so the difference is negligible
(Exercise 38.17). The main result of this chapter is the following:
Theorem 38.6. Let S, A and n be natural numbers and δ ∈ (0, 1). There
exists an efficiently computable policy π that when interacting with any MDP
M = (S, A, P, r) with S states, A actions, rewards in [0, 1] and any initial state
distribution satisfies with probability at least 1 − δ,
p
R̂n < CD(M )S An log(nSA/δ) ,
where C is a universal constant.
In Exercise 38.18, we ask you to use the assumption that the rewards are
bounded to find a choice of δ ∈ (0, 1) such that
p
E[R̂n ] ≤ 1 + CD(M )S 2An log(n) . (38.12)
This result is complemented by the following lower bound:
Theorem 38.7. Let S ≥ 3, A ≥ 2, D ≥ 6 + 2 logA S and n ≥ DSA. Then for
any policy π there exists a Markov decision process with S states, A actions and
diameter at most D such that
√
E[R̂n ] ≥ C DSAn ,
where C > 0 is again a universal constant.
√
The upper and lower bounds are separated by a factor of at least DS, which
is a considerable gap. Recent work has made progress towards closing this gap as
we explain in the notes.

38.5 Upper Confidence Bounds for Reinforcement Learning

Reinforcement learning is the subfield of machine learning devoted to

designing and studying algorithms that learn to maximise long-term reward
 38.5 Upper Confidence Bounds for Reinforcement Learning 523

in sequential context. The algorithm that establishes Theorem 38.6 is called

UCRL2 because it is the second version of the ‘upper confidence bounds for
reinforcement learning’ algorithm. Its pseudocode is shown in Algorithm 28.
At the start of each phase, UCRL2 computes an optimal policy for the
statistically plausible MDP with the largest optimal gain. The details of this
computation are left to the next section. This policy is then implemented until
the number of visits to some state-action pair doubles when a new phase starts
and the process begins again. The use of phases is important, not just for
computational efficiency. Recalculating the optimistic policy in each round may
lead to a dithering behaviour in which the algorithm frequently changes its plan
and suffers linear regret (Exercise 38.19).
To complete the specification of the algorithm, we must define confidence
sets on the unknown quantity, which in this case is the transition matrix. The
confidence sets are centered at the empirical transition probabilities defined by
Pt
I {Su = s, Au = a, Su+1 = s0 }
P̂t,a (s, s ) = u=1
0
,
1 ∨ Tt (s, a)
Pt
where Tt (s, a) = u=1 I {Su = s, Au = a} is the number of times action a was
taken in state s. As before, we let P̂t,a (s) be the vector whose s0 th entry is
P̂t,a (s, s0 ). Given a state-action pair s, a, define
( s )
SLt−1 (s, a)
Ct (s, a) = P ∈ P(S) : kP − P̂t−1,a (s)k1 ≤ , (38.13)
1 ∨ Tt−1 (s, a)

where for Tt (s, a) > 0 we set

 
4SATt (s, a)(1 + Tt (s, a))
Lt (s, a) = 2 log ,
δ
and for Tt (s, a) = 0 we set Lt (s, a) = 1. Note that in this case Ct+1 (s, a) = P(S).
Then define

Ct = {P = (Pa (s))s,a : Pa (s) ∈ Ct (s, a) for all s, a ∈ S × A} . (38.14)

Clearly Tt (s, a) cannot be larger than the total number of rounds n, so

 
4SAn(n + 1)
Lt (s, a) ≤ L = 2 log . (38.15)
δ
The algorithm operates in phases k = 1, 2, 3, . . . with the first phase starting in
round τ1 = 1 and the (k + 1)th phase starting in round τk+1 defined inductively
by

τk+1 = 1 + min {t : Tt (St , At ) ≥ 2Tτk −1 (St , At )} ,

which means that the next phase starts once the number of visits to some
state-action pair at least doubles.
 38.5 Upper Confidence Bounds for Reinforcement Learning 524

1: Input S, A, r, δ ∈ (0, 1)
2: t=0
3: for k = 1, 2, . . . do
4: τk = t + 1
5: Find πk as the greedy policy with respect to vk satisfying Eq. (38.16)
6: do
7: t ← t + 1, observe St and take action At = πk (St )
8: while Tt (St , At ) < 2Tτk −1 (St , At )
9: end for
Algorithm 28: UCRL2.

38.5.1 The Extended Markov Decision Process

The confidence set Ct defines a set of plausible transition probability functions at
the start of round t. Since the reward function is known already, this corresponds
to a set of plausible MDPs. The algorithm plays according to the optimal policy
in the plausible MDP with the largest gain. There is some subtlety because
the optimal policy is not unique, and what is really needed is to find a policy
that is greedy with respect to a value function satisfying the Bellman optimality
equation in the plausible MDP with the largest gain. Precisely, at the start of
the kth phase, the algorithm must find a value function vk , gain ρk and MDP
Mk = (S, A, Pk , r) with Pk ∈ Cτk such that

ρk + vk (s) = max ra (s) + hPk,a (s), vk i for all s ∈ S and a ∈ A ,

a∈A
(38.16)
ρk = max max max ρπs (P ) ,
s∈S π∈ΠDM P ∈Cτk

where ρπs (P ) is the gain of deterministic memoryless policy π starting in state s

in the MDP with transition probability function P . The algorithm then plays
according to πk defined as the greedy policy with respect to vk . There is quite a
lot hidden in these equations. The gain is only guaranteed to be constant when
Mk has a finite diameter, but this may not hold for all plausible MDPs. As it
happens, however, solutions to Eq. (38.16) are guaranteed to exist and can be
found efficiently. To see why this is true we introduce the extended MDP M̃k ,
which has state space S and state-dependent action space Ãs given by

Ãs = {(a, P ) : a ∈ A, P ∈ Cτk (s, a)} .

The reward function of the extended MDP is r̃(a,P ) (s) = ra (s), and the transitions
are P̃a,P (s) = Pa (s). The action space in the extended MDP allows the agent to
choose both a ∈ A and a plausible transition vector Pa (s) ∈ Cτk (s, a). By the
definition of the confidence sets, for any pair of states s, s0 and action a ∈ A,
there always exists a transition vector Pa (s) ∈ Cτk (s, a) such that Pa (s, s0 ) > 0,
which means that M̃k is strongly connected. Hence solving the Bellman optimality
equation for M̃k yields a value function vk and constant gain ρk ∈ R that satisfy
 38.6 Proof of Upper Bound 525

Eq. (38.16). A minor detail is that the extended action sets are infinite, while
the analysis in previous sections only demonstrated existence of solutions to
the Bellman optimality equation for finite MDPs. You should convince yourself
that Ct (s, a) is convex and has finitely many extremal points. Restricting the
confidence sets to these points makes the extended MDP finite without changing
the optimal policy.

38.5.2 Computing the Optimistic Policy ( )

Here we explain how to efficiently solve the Bellman optimality equation for the
extended MDP. The results in Section 38.3 show that the Bellman optimality
equation for M̃k can be solved efficiently provided that for any value function
v ∈ RS computing
 
argmaxa∈A ra (s) + max hP, vi (38.17)
P ∈Cτk (s,a)

can be carried out in an efficient manner. The inner optimisation is another linear
program with S variables and O(S) constraints and can be solved in polynomial
time. This procedure is repeated for each a ∈ A to compute the outcome of
(38.17). In fact the inner optimisation can be solved more straightforwardly by
sorting the entries of v and then allocating P coordinate by coordinate to be as
large as allowed by the constraints in decreasing order of v. The total computation
cost of solving Eq. (38.17) in this way is O(S(A + log S)). Combining this with
Algorithm 27 gives the required separation oracle.
The next problem is to find an R such that the set of feasible solutions to the
linear programs in Eq. (38.6) and Eq. (38.7) are contained in√the set {x : kxk ≤ R}.
As discussed in Section 38.3.1, a suitable value is R = 1 + D2 S, where D is
√
an upper bound on the diameter of the MDP. It turns out that D = n works
because for each pair of states s, s0 , there exists an action a and P ∈ Cτk (s, a)
√ √
such that P (s, s0 ) ≥ 1 ∧ (1/ n) so D(M̃k ) ≤ n. Combining this with the tools
developed in Section 38.3 shows that the Bellman optimality equation for M̃k may
be solved using linear programming in polynomial time. Note that the additional
constraints require a minor adaptation of the separation oracle, which we leave
to the reader.

38.6 Proof of Upper Bound

The proof is developed in three steps. First we decompose the regret into phases
and define a failure event where the confidence intervals fail. In the second step,
we bound the regret in each phase, and in the third step we sum over the phases.
Recall that M = (S, A, P, r) is the true Markov decision process with diameter
D = D(M ). The initial state distribution is µ ∈ P(S), which is arbitrary.
 38.6 Proof of Upper Bound 526

Step 1: Failure Events and Decomposition

Let K be the (random) number of phases, and for k ∈ [K], let Ek =
{τk , τk + 1, . . . , τk+1 − 1} be the set of rounds in the kth phase, where τK+1
is defined to be n + 1. Let T(k) (s, a) be the number of times state-action pair s, a
is visited in the kth phase:
X
T(k) (s, a) = I {St = s, At = a} .
t∈Ek

Define F as the failure event that P ∈

/ Cτk for some k ∈ [K].

Lemma 38.8. P (F ) ≤ δ/2 .

The proof is based on a concentration inequality derived for categorical

distributions and is left for Exercise 38.21. When F does not hold, the true
transition kernel is in Cτk for all k, which means that ρ∗ ≤ ρk and

n
X K X
X
R̂n = (ρ∗ − rAt (St )) ≤ (ρk − rAt (St )) .
t=1 k=1 t∈Ek
| {z }
R̃k

In the next step, we bound R̃k under the assumption that F does not hold.

Step 2: Bounding the Regret in Each Phase

Assume that F does not occur and fix k ∈ [K]. Recall that vk is a value function
satisfying the Bellman optimality equation in the optimistic MDP Mk and ρk is
its gain. Hence

ρk = rπk (s) − vk (s) + hPk,πk (s), vk i for all s ∈ S . (38.18)

As noted earlier, solutions to the Bellman optimality equation remain solutions

when translated, so we may assume without loss of generality that vk is such that
kvk k∞ ≤ span(vk )/2, which means that

1 D
kvk k∞ ≤ span(vk ) ≤ , (38.19)
2 2

where the second inequality follows from Lemma 38.3 and the fact that when F
does not hold, the diameter of the extended MDP M̃k is at most D and vk also
satisfies the Bellman optimality equation in this MDP. By the definition of the
policy, we have At = πk (St ) for t ∈ Ek , which implies that

ρk = rAt (St ) − vk (St ) + hPk,At (St ), vk i for all t ∈ Ek .

 38.6 Proof of Upper Bound 527

Rearranging and substituting yields

X
R̃k = (−vk (St ) + hPk,At (St ), vk i)
t∈Ek
X X
= (−vk (St ) + hPAt (St ), vk i) + hPk,At (St ) − PAt (St ), vk i
t∈Ek t∈Ek
X D X
≤ (−vk (St ) + hPAt (St ), vk i) + kPk,At (St ) − PAt (St )k1 , (38.20)
2
t∈Ek t∈Ek
| {z } | {z }
(A) (B)

where the inequality follows from Hölder’s inequality and Eq. (38.19). Let
Et [·] denote the conditional expectation with respect to P conditioned on
σ(S1 , A1 , . . . , St−1 , At−1 , St ). To bound (A), we reorder the terms and use the
fact that span(vk ) ≤ D on the event F c . We get
X
(A) = (vk (St+1 ) − vk (St ) + hPAt (St ), vk i − vk (St+1 ))
t∈Ek
X
= vk (Sτk+1 ) − vk (Sτk ) + (hPAt (St ), vk i − vk (St+1 ))
t∈Ek
X
≤D+ (Et [vk (St+1 )] − vk (St+1 )) ,
t∈Ek

where the second equality used that max Ek = τk+1 − 1 and min Ek = τk . We
leave this here for now and move on to term (B) in Eq. (38.20). The definition of
the confidence intervals and the assumption that F does not occur shows that
√
D LS X T(k) (s, a)
(B) ≤ p .
2 1 ∨ Tτk −1 (s, a)
(s,a)∈S×A

Combining the bounds (A) and (B) yields

√
X D LS X T(k) (s, a)
R̃k ≤ D + (Et [vk (St+1 )] − vk (St+1 )) + p .
2 1 ∨ Tτk −1 (s, a)
t∈Ek (s,a)∈S×A

Step 3: Bounding the Number of Phases and Summing

Let Kt be the phase in round t so that t ∈ EKt . By the work in the previous two
steps, if F does not occur, then
K
X n
X
R̂n ≤ R̃k ≤ KD + (Et [vKt (St+1 )] − vKt (St+1 ))
k=1 t=1
√
T(k) (s, a)
X K
X
D LS
+ p .
2 1 ∨ Tτk −1 (s, a)
(s,a)∈S×A k=1
 38.7 Proof of Lower Bound 528

The first sum is bounded using a version of Hoeffding–Azuma (Exercise 20.7):

r !
n log(2/δ)
Xn
δ
P F and
c
(Et [vKt (St+1 )] − vKt (St+1 )) ≥ D ≤ .
t=1
2 2
p
For the second term, we note that T(k) (s, a)/ 1 ∨ Tτk −1 (s, a) cannot be large
too often. A continuous approximation often provides intuition for the correct
form. Recalling the thousands of integrals you did at school, for any differentiable
f : [0, ∞) → R,
Z K 0
f (k) p p
p dk = 2 f (K) − 2 f (0) . (38.21)
0 f (k)
Here we are thinking of f (k) as the continuous approximation of Tτk −1 (s, a) and
its derivative as T(k) (s, a). In Exercise 38.22, we ask you to make this argument
rigourous by showing that
K
X T(k) (s, a) √ p
p ≤ 2+1 Tn (s, a) .
k=1
1 ∨ Tτk −1 (s, a)
P
Then by Cauchy–Schwarz and the fact that s,a∈S×A Tn (s, a) = n,
XXp √
Tn (s, a) ≤ SAn .
s∈S a∈A

It remains to bound the number of phases. A new phase starts when the visit
count for some state-action pair doubles. Hence K cannot be more than the
number of times the counters double in total for each of the states. It is easy to
see that 1 + log2 Tn (s, a) gives an upper bound on how many times the counter
for this pair may double (the constant 1 is there to account for the counter
P
changing from zero to one). Thus K ≤ K 0 = s,a 1 + log2 Tn (s, a). Noting that
P
0 ≤ Tn (s, a) and s,a Tn (s, a) = n and relaxing Tn (s, a) to take real values, we
find that the value of K 0 is the largest when Tn (s, a) = n/(SA), which shows that
  n 
K ≤ SA 1 + log2 .
SA
Putting everything together gives the desired result.

38.7 Proof of Lower Bound

The lower bound is proven by crafting a difficult MDP that models a bandit
with approximately SA arms. This is a cumbersome endeavour, but intuitively
straightforward, and the explanations that follow should be made clear in Fig. 38.3.
Given S and A, the first step is to construct a tree of minimum depth with at
most A children for each node using exactly S − 2 states. The root of the tree is
denoted by s◦ and transitions within the tree are deterministic, so in any given
node, the learner can simply select which child to transition to. Let L be the
 38.7 Proof of Lower Bound 529

number of leaves, and label these states s1 , . . . , sL . The last two states are sg
and sb (‘good’ and ‘bad’ respectively). For each i ∈ [L], the learner can take any
action a ∈ A and transitions to either the good state or the bad state according
to
1 1
Pa (si , sg ) = + ε(a, i) and Pa (si , sb ) = − ε(a, i) .
2 2
The function ε will be chosen so that ε(a, i) = 0 for all (a, i) pairs except one. For
this special state-action pair, we let ε(a, i) = ∆ for appropriately tuned ∆ > 0.
The good state and the bad state have the same transitions for all actions:
Pa (sg , sg ) = 1 − δ , Pa (sg , s◦ ) = δ ,
Pa (sb , sb ) = 1 − δ , Pa (sb , s◦ ) = δ .
Choosing δ = 4/D, which under the assumptions of the theorem is guaranteed
to be in (0, 1], ensures that the diameter of the described MDP is at most D,
regardless of the value of ∆. The reward function is ra (s) = 1 if s = sg and
ra (s) = 0 otherwise.
The connection to finite-armed bandits is straightforward. Each time the learner
arrives in state s◦ , it selects which leaf to visit and then chooses an action from
that leaf. This corresponds to choosing one of k = LA = Ω(SA) meta actions.
The optimal policy is to select the meta action with the largest probability of
transitioning to the good state. The choice of δ means the learner expects to
stay in the good/bad state for approximately D rounds, which also makes the
diameter of this MDP about D. This means the learner expects to make about
n/D decisions and p the rewards√are roughly in [0, D], so we should expect the
regret to be Ω(D kn/D) = Ω( nDSA).

s◦

δ, 1 δ, 0
s1 s2 s3

sg sb
Good state Bad state
1 − δ, 1 1 − δ, 0

Figure 38.3 Lower-bound construction for A = 2 and S = 8. The resulting MDP is

roughly equivalent to a bandit with six actions.
 38.7 Proof of Lower Bound 530

One could almost claim victory here and not bother with the proof. As usual,
however, there are some technical difficulties, which in this case arise because the
number of visits to the decision state s◦ is a random quantity. For this reason we
give the proof, leaving as exercises the parts that are both obvious and annoying.

Proof of Theorem 38.7 The proof follows the path suggested in Exercise 15.2.
We break things up into two steps. Throughout we fix an arbitrary policy π.

Step 1: Notation and Facts about the MDP

Let d be the depth of the tree in the MDP construction, L the number of leaves
and k = LA. Define the set of state–action pairs for which the state is a leaf of
the tree by

L = {(s, a) : a ∈ A and s is a leaf of the tree} .

By definition, this has k elements. Let M0 be the MDP with ε(s, a) = 0 for all
(s, a) ∈ L. Then let Mj be the MDP with ε(s, a) = ∆ for the jth state-action
pair in the above set. Define stopping time τ by
( t
)
X n
τ = n ∧ min t : I {Su = s◦ } ≥ −1 ,
u=1
D

which is the first round when the number of visits to state s◦ is at least n/D − 1,
or n if s◦ is visited fewer times than n/D. Next, let Tj be the number of visits to
Pk
state-action pair j ∈ [k] until stopping time τ and Tσ = j=1 Tj . For 0 ≤ j ≤ k,
let Pj be the law of T1 , . . . , Tk induced by the interaction of π and Mj . And
let Ej [·] be the expectation with respect to Pj . None of the following claims is
surprising, but they are all tiresome to prove to some extent. The claims are
listed in increasing order of difficulty and left to the reader in Exercise 38.24.

Claim 38.9. For all j ∈ [k], the diameter is bounded by D(Mj ) ≤ D .

Claim 38.10. There exist universal constants 0 < c1 < c2 < ∞ such that

DE0 [Tσ ]/n ∈ [c1 , c2 ] .

Claim 38.11. Let Rnj be the expected regret of policy π in MDP Mj over n
rounds. There exists a universal constant c3 > 0 such that

Rnj ≥ c3 ∆D Ej [Tσ − Tj ] .

Step 2: Bounding the Regret

Notice that M0 and Mj only differ when state-action pair j is visited. In
Exercise 38.30, you are invited to use this fact and the chain rule for relative
entropy given in Exercise 14.12 to prove that

D(P0 , Pj ) = E0 [Tj ]d(1/2, 1/2 + ∆) , (38.22)

 38.8 Notes 531

where d(p, q) is the relative entropy between Bernoulli distributions with means
p and q, respectively. Now ∆ will be chosen to satisfy ∆ ≤ 1/4. It follows from
the entropy inequalities in Eq. (14.16) that

D(P0 , Pj ) ≤ 4∆2 E0 [Tj ] . (38.23)

Using the fact that 0 ≤ Tσ − Tj ≤ Tσ ≤ n/D, Exercise 14.4 and Pinsker’s

inequality (Eq. (14.12)) and (38.23),
r q
n D(P0 , Pj ) n∆
Ej [Tσ − Tj ] ≥ E0 [Tσ − Tj ] − ≥ E0 [Tσ − Tj ] − 2E0 [Tj ] .
D 2 D
Summing over j and applying Cauchy–Schwarz yields
k
X k
X k q
n∆ X
Ej [Tσ − Tj ] ≥ E0 [Tσ − Tj ] − 2E0 [Tj ]
j=1 j=1
D j=1
n∆ p
≥ (k − 1)E0 [Tσ ] − 2kE0 [Tσ ]
Dr
c1 n(k − 1) n∆ 2c2 nk
≥ −
D D D
c1 n(k − 1)
≥ , (38.24)
2D
where the last inequality follows by choosing
r
c1 (k − 1) D
∆= .
2 2c2 nk
By Eq. (38.24), there exists a j ∈ [k] such that

c1 n(k − 1)
Ej [Tσ − Tj ] ≥ .
2Dk
Then, for the last step, apply Claim 38.11 to show that
r
c2 c3 n(k − 1)2 D
Rnj ≥ c3 D∆Ej [Tσ − Tj ] ≥ 1 .
4k 2c2 nk
Naive bounding and simplification concludes the proof.

38.8 Notes

1 MDPs in applications can have millions (or ‘billions and billions’) of states,
which should make the reader worried that the bound in Theorem 38.6 could
be extremely large. The takeaway should be that learning in large MDPs
without additional assumptions is hard, as attested by the lower bound in
Theorem 38.7.
 38.8 Notes 532

2 The key to choosing the state space is that the state must be observable and
sufficiently informative that the Markov property is satisfied. Blowing up the
size of the state space may help to increase the fidelity of the approximation
(the entire history always works), but will almost always slow down learning.
3 We simplified the definition of MDPs by making the rewards a deterministic
function of the current state and the action chosen. A more general definition
allows the rewards to evolve in a random fashion, jointly with the next state.
In this definition, the mean reward functions are dropped and the transition
kernel Pa is replaced with an S → S × R stochastic kernel, call it, P̃a . Thus,
for every s ∈ S, P̃a (s) is a probability measure over S × R. The meaning of this
is that when action a is chosen in state s, a random transition, (S, R) ∼ P̃a (s)
happens to state S, while reward R R is received. Note that the mean reward
along this transition is ra (s) = xP̃a (s, ds0 , dx).
4 A state s ∈ S is absorbing if Pa (s, s) = 1 for all a ∈ A. An MDP is episodic if
there exists an absorbing state that is reached almost surely by any policy. The
average reward criterion is meaningless in episodic MDPs because all policies
are optimal. In this case the usual objective is to maximise the expected reward
until the absorbing state is reached without limits or normalisation, sometimes
with discounting. An MDP is finite-horizon if it is episodic and the absorbing
state is always reached after some fixed number of rounds. The simplification
of the setting eases the analysis and preserves most of the intuition from the
general setting.
5 A partially observable MDP (POMDP) is a generalisation where the learner
does not observe the underlying state. Instead they receive an observation
that is a (possibly random) function of the state. Given a fixed (known) initial
state distribution, any POMDP can be mapped to an MDP at the price of
enlarging the state space. A simple way to achieve this is to let the new state
space be the space of all histories. Alternatively you can use any sufficient
statistic for the hidden state as the state. A natural choice is the posterior
distribution over the hidden state given the interaction history, which is called
the belief space. While the value function over the belief space has some nice
structure, in general even computing the optimal policy is hard [Papadimitriou
and Tsitsiklis, 1987].
6 We called the all-knowing entity that interacts with the MDP an agent. In
operations research the term is decision maker and in control theory it is
controller. In control theory the environment would be called the controlled
system or the plant (for power-plant, not a biological plant). Acting in
an MDP is studied in control theory under stochastic optimal control,
while in operations research the area is called multistage decision making
under uncertainty or multistage stochastic programming. In the control
community the infinite horizon setting with the average cost criterion is perhaps
the most common, while in operations research the episodic setting is typical.
7 The definition of the optimal gain that is appropriate for MDPs that are not
strongly connected is a vector ρ∗ ∈ RS given by ρ∗s = supπ ρ̄πs . A policy is
 38.8 Notes 533

optimal if it achieves the supremum in this definition and such a policy always
exists as long as the MDP is finite. In strongly connected MDPs, the two
definitions coincide. For infinite MDPs, everything becomes more delicate and
a large portion of the literature on MDPs is devoted to this case.
8 In applications where the asymptotic nature of gain optimality is unacceptable,
there are criteria that make finer distinctions between the policies. A memoryless
policy π ∗ is bias optimal if it is gain optimal and vπ∗ ≥ vπ for all memoryless
policies π. Even more sensitive criteria exist. Some keywords to search for are
Blackwell optimality and n-discount optimality.
9 The Cesàro sum of a real-valued sequence (an )n is the asymptotic average of
its partial sums. Let sn = a0 + · · · + an−1 be the nth partial sum. The Cesàro
sum of this sequence is A = limn→∞ n1 (s1 + · · · + sn ) when this limit exists.
The idea is that Cesàro summation smoothes out periodicity, which means that
for certain sequences the Cesáro sum exists while sn does not converge. For
example, the alternating sequence (+1, −1, +1, −1, . . . ) is Cesàro summable,
and its Cesàro sum is easily seen to be 1/2, while it is not summable in the
normal sense. If a sequence is summable, then its sum and its Cesàro sum
coincide. The differential value of a policy is defined as a Cesàro sum so that it
is well defined even if the underlying Markov chain has periodic states.
10 For γ ∈ (0, 1), the γ-discounted average of sequence (an )n is Aγ = (1 −
P∞
γ) n=0 γ n an . An elementary argument shows that if Aγ is well defined, then
P∞ Pn
Aγ = (1 − γ)2 n=1 γ n−1 sn . Suppose the Cesàro sum A = limn→∞ n1 t=1 st
P
exists, then using the fact that 1 = (1 − γ)2 n=1 γ n−1 n, we have Aγ − A =
∞

2
P∞ n−1 P∞
(1−γ) n=1 γ (sn −nA). It is not hard to see that | n=1 γ n−1 (sn −nA)| =
O(1/(1 − γ)), and thus Aγ − A = O(1 − γ) as γ → 1, which means that
limγ→1 Aγ = A. The value limγ→1 Aγ is called the Abel sum of (an )n . Put
simply, the Abel sum of a sequence is equal to its Cesàro sum when the latter
exists. Abel summation is stronger in the sense that there are sequences that
are Abel summable but not Cesàro summable. The approach of approximating
Cesàro sums through γ-discounted averages, and taking the limit as γ → 1 is
called the vanishing discount approach and is one of the standard ways
to prove that the (average reward) Bellman equation has a solution (see
Exercises 38.9 and 38.10). As an aside, the systematic study of how to define
the ‘sum’ of a divergent series is a relatively modern endeavour. An enjoyable
historical account is given in the first chapter of the book on the topic by
Hardy [1973].
11 Given a solution (ρ, v) to Eq. (38.6), we mentioned a procedure for finding
a state s̃ ∈ S that is recurrent under some optimal policy. This works as
follows. Let C0 = {(s, a) : ρ + v(s) = ra (s) + hPa (s), vi} and I0 = {s :
(s, a) ∈ C0 for some a ∈ A}. Then define Ck+1 and Ik+1 inductively by the
following algorithm. First find an (s, a) ∈ Ck such that Pa (s, s0 ) > 0 for some
s0 6∈ Ik . If no such pair exists, then halt. Otherwise let Ck+1 = Ck \ {(s, a)}
and Ik+1 = {s : (s, a) ∈ Ck+1 for some a ∈ A}. Now use the complementary
slackness conditions of the dual program to Eq. (38.6) to prove that the
 38.8 Notes 534

algorithm halts with some non-empty Ik and that these states are recurrent
under some optimal policy. For more details, have a look at Exercise 4.15 of
the second volume of the book by Bertsekas [2012].
12 We mentioned enumeration, value iteration and policy iteration as other
methods for computing optimal policies. Enumeration just means enumerating
all deterministic memoryless policies and selecting the one with the highest
gain. This is obviously too expensive. Policy iteration is an iterative process
that starts with a policy π0 . In each round, the algorithm computes πk+1
from πk by computing vπk and then choosing πk+1 to be the greedy policy
with respect to vπk . This method may not converge to an optimal policy,
but by slightly modifying the update process, one can prove convergence.
For more details, see chapter 4 of volume 2 of the book by Bertsekas [2012].
Value iteration works by choosing an arbitrary value function v0 and then
inductively defining vk+1 = T vk , where (T v)(s) = maxa∈A ra (s) + hPa (s), vi
is the Bellman operator. Under certain technical conditions, one can prove
that the greedy policy with respect to vk converges to an optimal policy. Note
that vk+1 = Ω(k), which can be a problem numerically. A simple idea is to
let vk+1 = T vk − δk where δk = maxs∈S vk (s). Since the greedy policy is the
same for v and v + c1, this does not change the mathematics, but improves
the numerical situation. The aforementioned book by Bertsekas is again a
good source for more details. Unfortunately, none of these algorithms have
known polynomial time guarantees on the computation complexity of finding
an optimal policy without stronger assumptions than we would like. In practice,
however, both value and policy iteration work quite well, while the ellipsoid
method for solving linear programs should be avoided at all costs. Of course
there are other methods for solving linear programs, and these can be effective.
13 Theorem 38.6 is vacuous when the diameter is infinite, but you might wonder if
the bound continues to hold in certain ‘nice’ cases. Unfortunately, the algorithm
is rather brittle. UCRL2 suffers linear regret if there is a single unreachable
state with reward larger than the optimal gain (Exercise 38.27).
14 One can modify the concept of regret to allow for MDPs that have traps. We
restrict our attention to policies with sublinear regret in strongly connected
MDPs, which must try and explore the whole state space and hence almost
surely become trapped in a strongly communicating subset of the state space.
The regret is redefined by ‘restarting the clock’ at the time when the policy
gets trapped. For details, see Exercise 38.29.
15 The assumption that the reward function is known can be relaxed without
difficulty. It is left as an exercise to figure out how to modify algorithm and
analysis to the case when r is unknown and reward observed in round t is
bounded in [0, 1] and has conditional mean rAt (St ). See Exercise 38.23.
16 Although√it has not been done yet in this setting, the path to removing the
spurious S from the bound is to avoid the application of Cauchy–Schwarz
in Eq. (38.20). Instead one should define confidence intervals directly on
hP̂k − P, vk i, where the dependence on the state and action has been omitted.
 38.9 Bibliographical Remarks 535

Of course, the algorithm must be changed to use the improved confidence

intervals. At first sight, it seems that one could apply Hoeffding’s bound
directly to the inner product, but there is a subtle problem that has spoiled a
number of attempts: vk and P̂k are not independent. This non-independence
is unfortunately quite pernicious and appears from many angles. We advise
extreme caution. Some references for guidance are given in the bibliographic
remarks.

38.9 Bibliographical Remarks

The study of sequential decision-making has a long

history, and we recommend the introduction of the
book by Puterman [2009] as a good starting point.
One of the main architects in modern times is Richard
Bellman, who wrote an influential book [Bellman,
1954]. His autobiography is so entertaining that
reading it slowed the writing of this chapter: The
Eye of the Hurricane [Bellman, 1984]. As a curiosity,
Bellman knew about bandit problems after accidentally
encountering a paper by Thompson [1935]. For the
tidbit, see page 260 of the aforementioned biography.
MDPs are studied by multiple research communities, Richard Bellman
including control, operations research and artificial
intelligence. The two-volume book by Bertsekas [2012] provides a thorough and
formal introduction to the basics. The perspective is quite interdisciplinary,
but with a slight (good) bias towards the control literature. The perspective
of an operations researcher is most precisely conveyed in the comprehensive
book by Puterman [2009]. A very readable shorter introductory book is by Ross
[1983]. Arapostathis et al. [1993] surveyed existing analytical results (existence,
uniqueness of optimal policies, validity of the Bellman optimality equation) for
average-reward MDPs with an emphasis on continuous state and action space
models. The online lecture notes of Kallenberg [2016] are a recent comprehensive
alternative account for the theory of discrete MDPs. There are many texts on
linear/convex optimisation and the ellipsoid method. The introductory book on
linear optimisation by Bertsimas and Tsitsiklis [1997] is a pleasant read, while
the ellipsoid method is explained in detail by Grötschel et al. [2012].
The problem considered in this chapter is part of a broader field called
reinforcement learning (RL), which has recently seen a surge of interest. The
books by Sutton and Barto [2018] and Bertsekas and Tsitsiklis [1996] describe
the foundations. The first book provides an intuitive introduction aimed at
computer scientists, while the second book focuses on the theoretical results of
the fundamental algorithms. A book by one of the present authors focuses on
 38.9 Bibliographical Remarks 536

cataloguing the range of learning problems encountered in reinforcement learning

and summarising the basic ideas and algorithms [Szepesvári, 2010].
The UCRL algorithm and the upper and lower regret analysis is due to
Auer et al. [2009] and Jaksch et al. [2010]. Our proofs differ in minor ways. A
more significant difference is that these works used value iteration for finding
the optimistic policy and hence cannot provide polynomial time computation
guarantees. In practice this may be preferable to linear programming anyway.
The number of rigourous results for bounding the regret of various algorithms is
limited. One idea is to replace the optimistic approach with Thompson sampling,
which was first adapted to reinforcement learning by Strens [2000] under the
name PSRL (posterior sampling reinforcement learning). Agrawal and Jia [2017]
recently made an attempt to improve the dependence of the regret on the state
space. The proof is not quite correct, however, and at the time of writing the
holes have not yet been patched. Azar et al. [2017] also improve upon the UCRL2
bound, but for finite-horizon
√ episodic problems, where they derive an optimistic
algorithm with regret Õ( HSAn), which after √ adapting UCRL to the episodic
setting improves on its regret by a factor of SH. The main innovation is to
use Freedman’s Bernstein-style inequality for computing bonuses directly while
computing action values using backwards induction from the end of the episode
rather than keeping confidence estimates for the transition probabilities. An
issue with both of these improvements is that lower-order terms in the bounds
mean they only hold for large n. It remains to be seen if these terms arise from
the analysis or if the algorithms need modification. UCRL2 will fail in MDPs
with infinite diameter, even if the learner starts in a subset of the states that
is strongly connected from which it cannot escape. This limitation was recently
overcome by Fruit et al. [2018], who provide an algorithm with roughly the same
regret as UCRL2, but where the dependence on the diameter and state space are
replaced with those of the sub-MDP in which the learner starts and from which
it is assumed there is no escape.
Tewari and Bartlett [2008] use an optimistic version of linear programming
to obtain finite-time logarithmic bounds with suboptimal instance-dependent
constants. Note this paper mistakenly drops some constants from the confidence
intervals, which after fixing would make the constants even worse and seems to
have other problems, as well [Fruit et al., 2018]. Similar results are also available
for UCRL2 [Auer and Ortner, 2007]. Burnetas and Katehakis [1997a] prove
asymptotic guarantees with optimal constants, but with the crucial assumption
that the support of the next-state distributions Pa (s) are known. Lai and Graves
[1997] also consider asymptotic optimality. However, they consider general state
spaces where the set of transition probabilities is smoothly parameterised with a
known parameterisation but under the weakened goal of competing with the best
of finitely many memoryless policies given to the learner as black boxes.
Finite-time regret for large state and action space MDPs under additional
structural assumptions are also considered by Abbasi-Yadkori and Szepesvári
[2011], Abbasi-Yadkori [2012] and Ortner and Ryabko [2012]. Abbasi-Yadkori and
 38.10 Exercises 537

√
Szepesvári [2011] and Abbasi-Yadkori [2012] give algorithms with O( n) regret
for linearly parameterised MDP problems with quadratic cost (linear quadratic
regulation, or LQR), while Ortner and Ryabko [2012] give O(n(2d+1)/(2d+2) ) regret
bounds under a Lipschitz assumption, where d is the dimensionality of the state
space. The algorithms in these works are not guaranteed to be computationally
efficient because they rely on optimistic policies. In theory, this could be addressed
by Thompson sampling, which is considered by Abeille and Lazaric [2017b], who
obtain partial results for the LQR setting. Thompson sampling has also been
studied in the Bayesian framework by Osband et al. [2013], Abbasi-Yadkori
and Szepesvári [2015], Osband and Van Roy [2017] and Theocharous et al.
[2017], of which Abbasi-Yadkori and Szepesvári [2015] and Theocharous et al.
[2017] consider general parametrisations, while the other papers are concerned
with finite state-action MDPs. Learning in MDPs has also been studied in the
probability approximately correct (PAC) framework introduced by Kearns and
Singh [2002], where the objective is to design policies for which the number
of badly suboptimal actions is small with high probability. The focus of these
papers is on the discounted reward setting rather than average reward. The
algorithms are again built on the optimism principle. Algorithms that are known
to be PAC-MDP include R-max [Brafman and Tennenholtz, 2003, Kakade, 2003],
MBIE [Strehl and Littman, 2005, 2008], delayed Q-learning [Strehl et al., 2006],
the optimistic-initialisation-based algorithm of Szita and Lőrincz [2009], MorMax
by Szita and Szepesvári [2010], and an adaptation of UCRL by Lattimore and
Hutter [2012], which they call UCRLγ. The latter work presents optimal results
(matching upper and lower bounds) for the case when the transition structure
is sparse, while the optimal dependence on the number of state-action pairs
is achieved by delayed Q-learning and Mormax [Strehl et al., 2006, Szita and
Szepesvári, 2010], though the Mormax bound is better in its dependency on the
discount factor. The idea to incorporate the uncertainty in the transitions into
the action space to solve the optimistic optimisation problem appeared in the
analysis of MBIE [Strehl and Littman, 2008]. A hybrid between stochastic and
adversarial settings is when the reward sequence is chosen by an adversary, while
transitions are stochastic. This problem has been introduced by Even-Dar et al.
[2004]. State-of-the-art results for the bandit case are due to Neu et al. [2014],
where the reader can also find further pointers to the literature. The case when
the rewards and the transitions probability distributions are chosen adversarially
is studied by [Abbasi-Yadkori et al., 2013].

38.10 Exercises

38.1 (Existence of probability space) Let M = (S, A, P ) be a finite

controlled Markov environment, which is a finite MDP without the reward
function. A policy π = (πt )∞
t=1 is a sequence of probability kernels where πt
is from (S × A)t−1 × S to A. Given a policy π and initial state distribution
 38.10 Exercises 538

µ ∈ P(S), show there exists a probability space (Ω, F, P) and an infinite sequence
of random elements S1 , A1 , S2 , A2 , . . . such that for any s ∈ S, a ∈ A and t ∈ N,

(a) P(S1 = s) = µ(s);

(b) P(St+1 = s | S1 , A1 , . . . , St , At ) = PAt (St , s); and
(c) P(At = a | S1 , A1 , . . . , St ) = πt (a | S1 , A1 , . . . , St−1 , At−1 , St ).

Hint Use Theorem 3.3.

38.2 (Sufficiency of Markov policies) Let M = (S, A, P ) be a finite

controlled Markov environment, π be an arbitrary policy and µ ∈ P(S) an
arbitrary initial state distribution. Denote by Pπµ the probability distribution that
results from the interconnection of π and M , while the initial state distribution
is µ.

(a) Show there exists a Markov policy π 0 such that

Pπµ (St = s, At = a) = Pπµ (St = a, At = a)

0

holds for all t ≥ 1 and s, a ∈ S × A.

(b) Conclude that for any policy π there exists a Markov policy π 0 such that for
any s ∈ S, ρ̄πs = ρ̄πs .
0

Hint Define π 0 inductively starting at t = 1. Puterman [theorem 5.5.1 2009]

proves this result and credits Strauch [1966].

38.3 (Deterministic policies minimise travel time) Let P be some

transition structure over some finite state space S and some finite action space
A. Show that the expected travel time between two states s, s0 of S is minimised
by a deterministic policy.

Hint Let τ ∗ (s, s0 ) be the shortest expected travel time between some arbitrary
pairs of states, which for s = s0 is defined to be zero. Show that τ ∗ satisfies the
fixed point equation
(
0, if s = s0 ;
τ (s, s ) =
∗ 0
P
1 + mina s00 Pa (s, s ) τ (s , s ) , otherwise .
00 ∗ 00 0

38.4 (Strongly connected ⇔ finite diameter) Let M be a finite MDP.

Prove that D(M ) < ∞ is equivalent to M being strongly connected.

38.5 (Diameter lower bound) Let M = (S, A, P, r) be any MDP. Show that
D(M ) ≥ logA (S) − 3.
Hint Denote by d∗ (s, s0 ) the minimum expected time it takes to reach
state s0 when starting from state s. The definition of d∗ can be extended to
arbitrary initial distributions µ0 over states and sets U ⊂ S of target states:
 38.10 Exercises 539

P P
d∗ (µ0 , U ) = s µ0 (s)d∗ (s, s0 ). Prove by induction on the size of U that
s0 ∈U
  
X  X 

d∗ (µ0 , U ) ≥ min knk  0 ≤ nk ≤ Ak , k ≥ 0, nk = |U | (38.25)
  
k≥0 k≥0

and then conclude that the proposition holds by choosing U = S [Jaksch et al.,
2010, corollary 15].

38.6 (State visitation probabilities and cumulative reward) Let

M = (S, A, P, r) be an MDP and π a memoryless policy and i, j ∈ [S].
(a) Show that e>i Pπ ej is the probability of arriving in state j from state i in t
t

rounds using policy π.

Pn
(b) Show that e> i t=1 Pπ rπ is the expected cumulative reward collected by
t

policy π over n rounds when starting in state i.

38.7 (Stochastic matrices) Let P be any S × S right stochastic matrix. Show

that the following hold:
Pn−1
(a) An = n1 t=0 P t is right stochastic.
(b) An + n1 (P n − I) = An P = P An .
Pn−1
(c) P ∗ = limn→∞ n1 t=0 P t exists and is right stochastic.
(d) P ∗ P = P P ∗ = P ∗ P ∗ = P ∗ .
(e) The matrix H = (I − P + P ∗ )−1 is well defined.
Pn Pi−1
(f) Let U = H − P ∗ . Then U = limn→∞ n1 i=1 k=0 (P k − P ∗ ).
Pn Pi−1
(g) Let r ∈ RS and ρ = P ∗ r. Then v = limn→∞ n1 i=1 k=0 P k (r − ρ) is well
defined and satisfies (38.3).
(h) With the notation of the previous part, v + ρ = r + P v.

Hint Note that the first four parts of this exercise are the same as in Chapter 37.
For parts (c) and (d), you will likely find it useful that the space of right stochastic
matrices is compact. Then show that all cluster points of (An ) are the same. For
(g), show that v = U r.

The previous exercise shows that the gain and differential value function of
any memoryless policy in any MDP are well defined. The matrix H is called
the fundamental matrix, and U is called the deviation matrix.

38.8 (Discounted MDPs) Let γ ∈ (0, 1), and define the operator Tγ : RS → RS
by
(Tγ v)(s) = max ra (s) + γhPa (s), vi .
a∈A

(a) Prove that Tγ is a contraction with respect to the supremum norm:

kTγ v − Tγ wk∞ ≤ γkv − wk∞ for any v, w ∈ RS .
 38.10 Exercises 540

(b) Prove that there exists a v ∈ RS such that Tγ v = v.

(c) Let π be the greedy policy with respect to v. Show v = rπ + γPπ v.
(d) Prove that v = (I − γPπ )−1 rπ .
(e) Define the γ-discounted value function vγπ of a policy π as the function
that for any given state s ∈ S gives the total expected discounted reward
of the policy when it is started from state s. Let vγ∗ ∈ RS be defined by
vγ∗ (s) = maxπ vγπ (s), s ∈ S. We call π γ-discount optimal if vγ∗ = vγπ . Show
that if π is greedy with respect to v from part (b), then π is a γ-optimal
policy.

Hint For (b), you should use the contraction mapping theorem (or Banach
fixed point theorem), which says that if (X , d) is a complete metric space and
T : X → X satisfies d(T (x), T (y)) ≤ γd(x, y) for γ ∈ [0, 1), then there exists an
x ∈ X such that T (x) = x. For (e), use (d) and Exercise 38.2 to show that
it suffices to check that vγπ ≤ v for any Markov policy π. Verify this by using
the fact that Tγ is monotone (f ≤ g implies that Tγ f ≤ Tγ g) and showing that
π
vγ,n ≤ Tγn 0 holds for any n, where vγ,n
π
(s) is the total expected discounted reward
of the policy when it is started from state s and is followed for n steps.
38.9 (From discounting to average reward) Recall that H = (I − P +
P ∗ )−1 , U = H − P ∗ . For γ ∈ [0, 1), define Pγ∗ = (1 − γ)(I − γP )−1 . Show that

(a) limγ→1− Pγ∗ = P ∗ ;

Pγ∗ −P ∗
(b) limγ→1− 1−γ = U.

Hint For (a) start by manipulating the expressions Pγ∗ P and (Pγ∗ )−1 P ∗ . For
(b) consider H −1 (Pγ∗ − P ∗ ).
38.10 (Solution to Bellman optimality equation) In this exercise you
will prove part (a) of Theorem 38.2.

(a) Prove there exists a deterministic stationary policy π and increasing sequence
of discount rates (γn ) with γn < 1 and limn→∞ γn = 1 such that π is a
greedy policy with respect to the fixed point vn of Tγn for all n.
(b) For the remainder of the exercise, fix a policy π whose existence is guaranteed
by part (a). Show that ρπ = ρ1 is constant.
(c) Let v = vπ be the value function and ρ = ρπ the gain of policy π. Show that
(ρ, v) satisfies the Bellman optimality equation.

Hint For (a), use the fact that for finite MDPs there are only finitely many
memoryless deterministic policies. For (b) and (c), use Exercise 38.9.
38.11 (Counterintuitive solutions to the Bellman equation) Consider
the deterministic MDP shown below with two states and two actions. The first
action, stay, keeps the state the same and the second action, Go, moves the
learner to the other state while incurring a reward of −1. Show that in this
 38.10 Exercises 541

example, solutions (ρ, v) to the Bellman optimality equations (Eq. (38.5)) are
exactly the elements of the set

(ρ, v) ∈ R × R2 : ρ = 0, v(1) − 1 ≤ v(2) ≤ v(1) + 1 .
r=0 r=0

r = −1

1 2

r = −1

38.12 (Dangers of linear program relaxation) Give an example of an

MDP and a solution (ρ, v) to the linear program in Eq. (38.6) such that v does
not satisfy the Bellman optimality equation and the greedy policy with respect
to v is not optimal.

38.13 (Bound on span in terms of diameter) Let M be a strongly connected

MDP and (ρ, v) be a solution to the Bellman optimality equation. Show that
span(v) ≤ (ρ∗ − mins,a ra (s))D(M ).

Hint Note that by Theorem 38.2, ρ = ρ∗ . Fix some states s1 =

6 s2 and a
memoryless policy π. Show that
v(s2 ) − v(s1 ) ≤ (ρ∗ − min ra (s))Eπ [τs2 | S1 = s1 ] .
s,a

Note for the sake of curiosity that the above display continues to hold for weakly
communicating MDPs.

The proof of Theorem 4 in the paper by Bartlett and Tewari [2009] is

incorrect. The problem is that the statement needs to hold for any solution
v of the Bellman optimality equation. The proof uses an argument that
hinges on the fact that in an aperiodic strongly connected MDP, v is in the
set {c1 + limn→∞ T n 0 − nρ∗ : c ∈ R}. However, Exercise 38.11 shows that
there exist strongly connected MDPs where this does not hold.

38.14 (Separation oracles) Solve the following problems:

(a) Prove that Algorithm 27 provides a separation oracle for convex set K defined
in Eq. (38.10).
(b) Assuming that Algorithm 27 can be implemented efficiently, explain how to
find an approximate solution to Eq. (38.7).

38.15 (Combining separation oracles) Let K ⊂ Rd be a convex set and φ be

a separation oracle for K. Suppose that a1 , . . . , an is a collection of vectors with
ak ∈ Rd and b1 , . . . , bk be a collection of scalars. Let Hk = {x ∈ Rd : hak , xi ≥ bk }.
Tn
Devise an efficient separation oracle for k=1 K ∩ Hk .
 38.10 Exercises 542

38.16 (Approximate solutions to Bellman equation) Consider a strongly

connected MDP, and suppose that ρ and v approximately satisfy the Bellman
optimality equation in the sense that there exists an ε > 0 such that
 
 
ρ + v(s) − max ra (s) + hPa (s), vi ≤ ε for all state-action pairs s, a .
 a∈A 
(38.26)

(a) Show that ρ ≥ ρ∗ − ε.

(b) Let π̃ be the greedy policy with respect to v. Assume that π̃ is ε0 -
greedy with respect to v in the sense that rπ̃(s) (s) + hPπ̃(s) (s), vi ≥
maxa∈A ra (s) + hPa (s), vi − ε0 holds for all s ∈ S. Show that π̃ is 2ε + ε0
optimal: ρπ̃ ≥ ρ∗ − (2ε + ε0 ).
(c) Suppose that ρ∗ in Eq. (38.7) is replaced with ρ ∈ [ρ∗ , ρ∗ + δ]. Show that the
linear program remains feasible and the solution (ρ, v) satisfies Eq. (38.26)
with ε ≤ |S|2 δ.

38.17 (Average-optimal is nearly finite-time optimal) Let M be a

strongly connected MDP with rewards in [0, 1], diameter D < ∞ and optimal
gain ρ∗ . Let vn∗ (s) be the maximum total expected reward in n steps when the
process starts in state s. Prove that vn∗ (s) ≤ nρ∗ + D.

38.18 (High probability ⇒ expected regret) Prove that (38.12) follows

from Theorem 38.6.

38.19 (Necessity of phases) The purpose of this exercise is to show that

without phases, UCRL2 may suffer linear regret. For convenience, we consider the
modified version of UCRL2 in Exercise 38.23 that does not know the reward. Now
suppose we further modify this algorithm to re-solve the optimistic MDP in every
round (τk = k for all k). We make use of a two state deterministic MDP with
two actions A = {stay, go}, depicted in Fig. 38.4. The transitions underlying
the two actions are represented by dashed and solid arrows, respectively.

1/2 1/2

1 2

Figure 38.4 Transitions and rewards are deterministic. Numbers indicate the rewards.

(a) Find all memoryless optimal policies for the MDP in Fig. 38.4.
(b) Prove that the version of UCRL2 given in Exercise 38.23 modified to re-solve
the optimistic MDP in every round suffers linear regret on this MDP.
 38.10 Exercises 543

Hint Since UCRL2 and the environment are both deterministic you can
examine the behaviour of the algorithm on the MDP. You should aim to prove
that eventually the algorithm will alternate between actions stay and go.

[Long-term plans should have phases] The reason UCRL2, or more generally,
optimistic algorithms without an explicit introduction of phases fail is
because UCRL2 for creating a plan solves the infinite horizon problem where
a reward in a state other than the current one and that is larger by the
tiniest amount than the reward in the current state makes it worth to switch
to the other state. If we considered an finite horizon version of the problem
where experience is collected in episodes with some fixed start state or start
state distribution, an optimistic algorithm would eventually stop considering
switches because on a finite horizon, eventually the loss from using the
actions that switch would be assessed to be higher than the potential gain
from switching.

38.20 (Extended MDP is strongly connected) Let M̃k be the extended

MDP defined in Section 38.5.1 and Cτk be the confidence set defined in Eq. (38.13).
Prove that P ∈ Cτk implies that M̃k is strongly connected.

38.21 (Confidence sets) Prove Lemma 38.8.

Hint Use the result of Exercise 5.17 and apply a union bound over all state-
action pairs and the number of samples. Use the Markov property to argue that
the independence assumption in Exercise 5.17 is not problematic.
38.22 Let (ak ) and (Ak ) be non-negative numbers so that for any k ≥ 0,
ak+1 ≤ Ak = 1 ∨ (a1 + · · · + ak ). Prove that for any m ≥ 1,
m
X a √ p
p k ≤ 2+1 Am .
k=1
Ak−1
Pm−1
Hint The statement is trivial if k=1 ak ≤ 1. If this does not hold, use
induction based on m = n, n + 1, . . . , where n is the first integer such that
Pn−1
k=1 ak > 1.

38.23 (Unknown rewards) In this exercise, you will modify the algorithm
to handle the situation where r is unknown and rewards are stochastic. More
precisely, assume there exists a function ra (s) ∈ [0, 1] for all a ∈ A and s ∈ S.
Then, in each round, the learner observes St , chooses an action At and receives a
reward Xt ∈ [0, 1] with

E[Xt | At , St ] = rAt (St ) .

In order to accommodate the unknown reward function, we modify UCRL2 in

the following way. First, define the empirical reward at the start of the kth phase
 38.10 Exercises 544

by
k −1
I {Su = s, Au = a} Xu
τX
r̂k,a (s) = .
u=1
1 ∨ Tτk −1 (s, a)

Then, let r̃t,a (s) be an upper confidence bound given by

s
L
r̃k,a (s) = r̂t,a (s) + ,
2(1 ∨ Tτk −1 (s, a))
where L is as in the proof of Theorem 38.6. The modified algorithm operates
exactly like Algorithm 28, but replaces the unknown ra (s) with r̃k,a (s) when
solving the extended MDP. Prove that with probability at least 1 − 3δ/2, the
modified policy in the modified setting has regret at most
s  
nSA
R̂n ≤ CD(M )S nA log ,
δ
where C > 0 is a universal constant.

38.24 (Lower bound) In this exercise, you will prove the claims to complete
the proof of the lower bound.

(a) Prove Claim 38.9.

(b) Prove Claim 38.10.
(c) Prove Claim 38.11.

38.25 (Contextual bandits as MDPs) Consider the MDP M = (S, A, P, r),

where Pa (s) = p for some fixed categorical distribution p for any (s, a) ∈ S × A,
where mins∈S p(s) > 0. Assume that the rewards for action a in state s are
sampled from a distribution supported on [0, 1] (see Note 3). An MDP like this
defines nothing but a contextual bandit.

(a) Derive the optimal policy and the average optimal reward.
(b) Show an optimal value function that solves the Bellman optimality equation.
(c) Prove that the diameter of this MDP is D = maxs 1/p(s).
(d) Consider the algorithm that puts one instance of an appropriate version
of UCB into every state (the same idea was explored in the context of
adversarial bandits in Section√18.1). Prove that the expected regret of your
algorithm will be at most O( SAn).
(e) Does the scaling behaviour of the upper bound in Theorem 38.6 match the
actual scaling behaviour of the expected regret of UCRL2 in this example?
Why or why not?
(f) Design and run an experiment to confirm your claim.

38.26 (Implementation) This is a thinking and coding exercise to illustrate

the difficulty of learning in MDPs. The RiverSwim environment is originally due to
Strehl and Littman [2008]. The environment has two actions A = {left, right}
 38.10 Exercises 545

and S = [S] with S ≥ 2. In all states s > 1, action left deterministically leads
to state s − 1 and provides no reward. In state 1, action left leaves the state
unchanged and yields a reward of 0.05. The action right tends to make the agent
move right but not deterministically (the learner is swimming against a current).
With probability 0.3, the state is incremented, with a probability 0.6, the state
is left unchanged, while with probability of 0.1 the state is decremented. This
action incurs a reward of zero in all states except in state S, where it receives a
reward of 1. The situation when S = 5 is illustrated in Fig. 38.5.

1, 0.05 1, 0 1, 0 1, 0 1, 0
0.3, 0 0.3, 0 0.3, 0 0.3, 0

1 0.1, 0 2 0.1, 0 3 0.1, 0 4 0.1, 1 5

0.7, 0 0.6, 0 0.6, 0 0.6, 0 0.9, 1

current

Figure 38.5 The RiverSwim MDP when S = 5. Solid arrows correspond to action left
and dashed ones to action right. The right-hand bank is slippery, so the learner
sometimes falls back into the river.

(a) Show that the optimal policy always takes action right and calculate the
optimal average reward ρ∗ as a function of S.
(b) Implement the MDP and test the optimal policy when started from state 1.
Plot the total reward as a function of time and compare it with the plot of
t 7→ tρ∗ . Run multiple simulations to produce error bars. How fast do you
think the total reward concentrates around tρ∗ ? Experiment with different
values of S.
(c) The ε-greedy strategy can also be implemented in MDPs as follows: based
on the data previously collected, estimate the transition probabilities and
rewards using empirical means. Find the optimal policy π ∗ of the resulting
MDP, and if the current state is s, use the action π ∗ (s) with probability 1 − ε
and choose one of the two actions uniformly at random with the remaining
probability. To ensure the empirical MDP has a well-defined optimal policy,
mix the empirical estimate of the next state distributions Pa (s) with the
uniform distribution with a small mixture coefficient. Implement this strategy
and plot the trajectories it exhibits for various MDP sizes. Explain what you
see.
(d) Implement UCRL2 and produce the same plots. Can you explain what you
see?
(e) Run simulations in RiverSwim instances of various sizes to compare the
regret of UCRL2 and ε-greedy. What do you conclude?
 38.10 Exercises 546

38.27 (UCRL2 and unreachable states) Show that UCRL2 suffers linear
regret if there is a single unreachable state with reward larger than the optimal
gain.

Hint Think about the optimistic MDP and the optimistic transitions to the
unreachable state. The article by Fruit et al. [2018] provides a policy that mitigates
the problem.
38.28 (MDPs with traps (i)) Fix state space S, action space A and reward
function r. Let π be a policy with sublinear regret in all strongly connected
MDPs (S, A, r, P ). Now suppose that (S, A, r, P ) is an MDP that is not strongly
connected such that for all s ∈ S, there exists a state s0 that is reachable from s
under some policy and where ρ∗s0 < maxu ρ∗u . Finally, assume that ρ∗S1 = maxu ρ∗u
almost surely. Prove that π has linear regret on this MDP.

38.29 (MDPs with traps (ii)) This exercise develops the ideas mentioned in
Note 14. First, we need some definitions: fix S and A and define Π0 as the set
of policies (learner strategies) for MDPs with state space S and action space A
that achieve sublinear regret in any strongly connected MDP with state space S
and action space A. Now consider an arbitrary finite MDP M = (S, A, P, r). A
state s ∈ S is reachable from state s0 ∈ S if there is a policy that when started
in s0 reaches state s with positive probability after one or more steps. A set of
states C ⊂ S is a strongly connected component (SCC) if every state s ∈ U
is reachable from every other state s0 ∈ C, including s = s0 . A set C ⊆ S is
maximal if we cannot add more states to C and still maintain the SCC property.
A SCC C is called a maximal end component if there does not exist another
SCC C 0 with C ⊂ C 0 . Show the following:

(a) There exists at least one MEC and two MECs C1 and C2 , are either equal
or disjoint.
(b) Let C1 , . . . , Ck be all the distinct MECs of an MDP. The MDP structure
defines a connectivity over C1 , . . . , Ck as follows: for i 6= j, we say that Ci is
connected to Cj if from some state in Ci , it is possible to reach some state of
Cj with positive probability under some policy. Show that this connectivity
structure defines a directed graph, which must be acyclic.
(c) Let C1 , . . . , Cm with m ≤ k be the sinks (the nodes with no out edges) of
this graph. Show that if M is strongly connected, then m = 1 and C1 = S.
(d) Show that for any i ∈ [m] and for any policy π ∈ Π0 , it holds that π will
reach Ci in finite time with positive probability if the initial state distribution
assigns positive mass to the non-trap states S \ ∪i∈[m] Ci .
(e) Show that for i ≤ m, for any s ∈ Ci and any action a ∈ A, Pa (s, s0 ) = 0 for
any s0 ∈ S \ Ci , i.e., Ci is closed.
(f) Show that the restriction of M to Ci defined as

Mi = (Ci , A, (Pa (s))s∈Ci ,a∈A , (ra (s))s∈Ci ,a∈A )

 38.10 Exercises 547

is an MDP.
(g) Show that Mi is strongly connected.
(h) Let τ be the time when the learner enters one of C1 , . . . , Cm and let I ∈ [m]
be the index of the class that is entered at time τ . That is, Sτ ∈ CI . Show
that if M is strongly connected, then τ = 1 with probability one.
(i) We redefine the regret as follows:
"τ +n−1 #
X
Rn = E
0
rAt (St ) − nρ (MI ) .
∗

t=τ

Show that if M is strongly connected, then Rn = Rn0 . p

(j) Can you design a policy with Rn0 = O(E[D(MI )|CI |] An log(n))? Will
UCRL2 already satisfy this?

The logic of the regret definition in part (i) is that by part (d), reasonable
policies cannot control which trap they fall into in an MDP that has more
than one traps. As such, policies should not be penalised for what trap they
fall into. However, once a policy falls into some trap, we expect it to start
to behave near optimally. What this definition is still lacking is that it is
insensitive to how fast a policy gets trapped. The last part is quite subtle
[Fruit et al., 2018].

38.30 (Chain rule for relative entropy) Prove the claim in Eq. (38.22).

Hint Make use of the result in Exercise 14.12.

Bibliography

Y. Abbasi-Yadkori. Forced-exploration based algorithms for playing in bandits

with large action sets. PhD thesis, University of Alberta, 2009a. [247]
Y. Abbasi-Yadkori. Forced-exploration based algorithms for playing in bandits
with large action sets. Master’s thesis, University of Alberta, Department of
Computing Science, 2009b. [94]
Y. Abbasi-Yadkori. Online Learning for Linearly Parametrized Control Problems.
PhD thesis, University of Alberta, 2012. [247, 536, 537]
Y. Abbasi-Yadkori and Cs. Szepesvári. Regret bounds for the adaptive control of
linear quadratic systems. In Proceedings of the 24th Conference on Learning
Theory, pages 1–26, Budapest, Hungary, 2011. JMLR.org. [536]
Y. Abbasi-Yadkori and Cs. Szepesvári. Bayesian optimal control of smoothly
parameterized systems. In Proceedings of the 31st Conference on Uncertainty
in Artificial Intelligence, pages 2–11, Arlington, VA, United States, 2015. AUAI
Press. [537]
Y. Abbasi-Yadkori, A. Antos, and Cs. Szepesvári. Forced-exploration based
algorithms for playing in stochastic linear bandits. In COLT Workshop on
On-line Learning with Limited Feedback, 2009. [94, 247]
Y. Abbasi-Yadkori, D. Pál, and Cs. Szepesvári. Improved algorithms for linear
stochastic bandits. In Advances in Neural Information Processing Systems,
pages 2312–2320. Curran Associates, Inc., 2011. [246]
Y. Abbasi-Yadkori, D. Pál, and Cs. Szepesvári. Online-to-confidence-set
conversions and application to sparse stochastic bandits. In Proceedings of the
15th International Conference on Artificial Intelligence and Statistics, pages
1–9, La Palma, Canary Islands, 2012. JMLR.org. [286]
Y. Abbasi-Yadkori, P. L. Bartlett, V. Kanade, Y. Seldin, and Cs. Szepesvári.
Online learning in Markov decision processes with adversarially chosen
transition probability distributions. In Advances in Neural Information
Processing Systems, pages 2508–2516, USA, 2013. Curran Associates Inc. [537]
Y. Abbasi-Yadkori, P. Bartlett, V. Gabillon, A. Malek, and M. Valko. Best of
both worlds: Stochastic & adversarial best-arm identification. In Proceedings
of the 31st Conference on Learning Theory, 2018. [416]
N. Abe and P. M. Long. Associative reinforcement learning using linear
probabilistic concepts. In Proceedings of the 16th International Conference
 BIBLIOGRAPHY 549

on Machine Learning, pages 3–11, San Francisco, CA, USA, 1999. Morgan
Kaufmann Publishers Inc. [246]
M. Abeille and A. Lazaric. Linear Thompson sampling revisited. In Proceedings
of the 20th International Conference on Artificial Intelligence and Statistics,
pages 176–184, Fort Lauderdale, FL, USA, 2017a. JMLR.org. [473]
M. Abeille and A. Lazaric. Thompson sampling for linear-quadratic control
problems. In Proceedings of the 20th International Conference on Artificial
Intelligence and Statistics, pages 1246–1254, Fort Lauderdale, FL, USA, 2017b.
JMLR.org. [537]
J. D. Abernethy and A. Rakhlin. Beating the adaptive bandit with high probability.
In Proceedings of the 22nd Conference on Learning Theory, 2009. [171, 343]
J. D. Abernethy, E. Hazan, and A. Rakhlin. Competing in the dark: An efficient
algorithm for bandit linear optimization. In Proceedings of the 21st Conference
on Learning Theory, pages 263–274. Omnipress, 2008. [343]
J. D. Abernethy, E. Hazan, and A. Rakhlin. Interior-point methods for full-
information and bandit online learning. IEEE Transactions on Information
Theory, 58(7):4164–4175, 2012. [170, 340]
J. D. Abernethy, C. Lee, A. Sinha, and A. Tewari. Online linear optimization via
smoothing. In Proceedings of the 27th Conference on Learning Theory, pages
807–823, Barcelona, Spain, 2014. JMLR.org. [373]
J. D. Abernethy, C. Lee, and A. Tewari. Fighting bandits with a new kind of
smoothness. In Advances in Neural Information Processing Systems, pages
2197–2205. Curran Associates, Inc., 2015. [343, 373]
M. Abramowitz and I. A. Stegun. Handbook of mathematical functions: with
formulas, graphs, and mathematical tables, volume 55. Courier Corporation,
1964. [183, 474]
M. Achab, S. Clémençon, A. Garivier, A. Sabourin, and C. Vernade. Max k-
armed bandit: On the extremehunter algorithm and beyond. In Joint European
Conference on Machine Learning and Knowledge Discovery in Databases, pages
389–404. Springer, 2017. [416]
L. Adelman. Choice theory. In Saul I. Gass and Michael C. Fu, editors,
Encyclopedia of Operations Research and Management Science, pages 164–
168. Springer US, Boston, MA, 2013. [67]
A. Agarwal, D. P. Foster, D. J. Hsu, S. M. Kakade, and A. Rakhlin. Stochastic
convex optimization with bandit feedback. In Advances in Neural Information
Processing Systems, pages 1035–1043. Curran Associates, Inc., 2011. [359]
A. Agarwal, D. P. Foster, D. Hsu, S. M. Kakade, and A. Rakhlin. Stochastic
convex optimization with bandit feedback. SIAM Journal on Optimization, 23
(1):213–240, 2013. [415]
A. Agarwal, D. Hsu, S. Kale, J. Langford, L. Li, and R. Schapire. Taming the
monster: A fast and simple algorithm for contextual bandits. In Proceedings
of the 31st International Conference on Machine Learning, pages 1638–1646,
Bejing, China, 2014. JMLR.org. [231, 232, 233]
 BIBLIOGRAPHY 550

A. Agarwal, S. Bird, M. Cozowicz, L. Hoang, J. Langford, S. Lee, J. Li,

D. Melamed, G. Oshri, and O. Ribas. Making contextual decisions with
low technical debt. arXiv:1606.03966, 2016. [17]
R. Agrawal. Sample mean based index policies with O(log n) regret for the multi-
armed bandit problem. Advances in Applied Probability, pages 1054–1078, 1995.
[110, 120]
S. Agrawal and N. R. Devanur. Bandits with concave rewards and convex
knapsacks. In Proceedings of the 15th ACM conference on Economics and
computation, pages 989–1006. ACM, 2014. [359]
S. Agrawal and N. R. Devanur. Linear contextual bandits with knapsacks. In
Advances in Neural Information Processing Systems, pages 3458–3467. Curran
Associates Inc., 2016. [360]
S. Agrawal and N. Goyal. Analysis of Thompson sampling for the multi-armed
bandit problem. In Proceedings of the 25th Conference on Learning Theory,
2012. [473]
S. Agrawal and N. Goyal. Further optimal regret bounds for Thompson sampling.
In Proceedings of the 16th International Conference on Artificial Intelligence
and Statistics, pages 99–107, Scottsdale, Arizona, USA, 2013a. JMLR.org. [471,
473]
S. Agrawal and N. Goyal. Thompson sampling for contextual bandits with
linear payoffs. In Proceedings of the 30th International Conference on Machine
Learning, pages 127–135, Atlanta, GA, USA, 2013b. JMLR.org. [473]
S. Agrawal and R. Jia. Optimistic posterior sampling for reinforcement learning:
worst-case regret bounds. In Advances in Neural Information Processing
Systems, pages 1184–1194. Curran Associates, Inc., 2017. [536]
S. Agrawal, V. Avadhanula, V. Goyal, and A. Zeevi. Thompson sampling for
the MNL-bandit. In Proceedings of the 2017 Conference on Learning Theory,
pages 76–78, Amsterdam, Netherlands, 2017. JMLR.org. [473]
N. Ailon, Z. Karnin, and T. Joachims. Reducing dueling bandits to cardinal
bandits. In Proceedings of the 31st International Conference on Machine
Learning, pages II–856–II–864. JMLR.org, 2014. [359]
J. Aldrich. “but you have to remember P. J. Daniell of Sheffield”. Electronic
Journal for History of Probability and Statistics, 3(2), 2007. [53]
C. Allenberg, P. Auer, L. Györfi, and G. Ottucsák. Hannan consistency in on-line
learning in case of unbounded losses under partial monitoring. In Proceedings
of the 17th International Conference on Algorithmic Learning Theory, pages
229–243, Berlin, Heidelberg, 2006. Springer-Verlag. [156, 170, 340]
N. Alon, Y. Matias, and M. Szegedy. The space complexity of approximating the
frequency moments. In Proceedings of the 28th annual ACM symposium on
theory of computing, pages 20–29. ACM, 1996. [115]
N. Alon, N. Cesa-Bianchi, C. Gentile, and Y. Mansour. From bandits to experts:
A tale of domination and independence. In Advances in Neural Information
Processing Systems, pages 1610–1618. Curran Associates, Inc., 2013. [360, 506]
 BIBLIOGRAPHY 551

N. Alon, N. Cesa-Bianchi, O. Dekel, and T. Koren. Online learning with feedback

graphs: Beyond bandits. In Proceedings of the 28th Conference on Learning
Theory, pages 23–35, Paris, France, 2015. JMLR.org. [360]
V. Anantharam, P. Varaiya, and J. Walrand. Asymptotically efficient allocation
rules for the multiarmed bandit problem with multiple plays-part i: Iid rewards.
IEEE Transactions on Automatic Control, 32(11):968–976, 1987. [247]
J. R. Anderson, J. L. Dillon, and J. E. Hardaker. Agricultural decision analysis.
Monographs: Applied Economics. Iowa State University Press, 1977. [2]
F. J. Anscombe. Sequential medical trials. Journal of the American Statistical
Association, 58(302):365–383, 1963. [94]
A. Antos, G. Bartók, D. Pál, and Cs. Szepesvári. Toward a classification of
finite partial-monitoring games. Theoretical Computer Science, 473:77–99, 2013.
[506]
A. Arapostathis, V. S. Borkar, E. Fernandez-Gaucherand, M. K. Ghosh, and S. I.
Marcus. Discrete-time controlled Markov processes with average cost criterion:
a survey. SIAM Journal of Control and Optimization, 31(2):282–344, 1993.
[535]
R. Arora, O. Dekel, and A. Tewari. Online bandit learning against an adaptive
adversary: From regret to policy regret. In Proceedings of the 29th International
Conference on Machine Learning, Madison, WI, USA, 2012. Omnipress. [158]
B. Ashwinkumar, J. Langford, and A. Slivkins. Resourceful contextual bandits.
In Proceedings of the 27th Conference on Learning Theory, pages 1109–1134,
Barcelona, Spain, 2014. JMLR.org. [359]
J.-V. Audibert and S. Bubeck. Regret bounds and minimax policies under partial
monitoring. Journal of Machine Learning Research, 11:2785–2836, 2010a. [157]
J.-Y. Audibert and S. Bubeck. Minimax policies for adversarial and stochastic
bandits. In Proceedings of the 22nd Conference on Learning Theory, pages
217–226, 2009. [128, 157, 347]
J.-Y. Audibert and S. Bubeck. Best arm identification in multi-armed bandits.
In Proceedings of the 23rd Conference on Learning Theory, 2010b. [385, 414]
J.-Y. Audibert, R. Munos, and Cs. Szepesvári. Tuning bandit algorithms in
stochastic environments. In Proceedings of the 18th International Conference on
Algorithmic Learning Theory, pages 150–165, Berlin, Heidelberg, 2007. Springer
Berlin Heidelberg. [69, 86, 110, 114, 212]
J.-Y. Audibert, R. Munos, and Cs. Szepesvári. Exploration-exploitation tradeoff
using variance estimates in multi-armed bandits. Theoretical Computer Science,
410(19):1876–1902, 2009. [69]
J.-Y. Audibert, S. Bubeck, and G. Lugosi. Regret in online combinatorial
optimization. Mathematics of Operations Research, 39(1):31–45, 2013. [343]
P. Auer. Using confidence bounds for exploitation-exploration trade-offs. Journal
of Machine Learning Research, 3:397–422, 2002. [246, 274]
P. Auer and C. Chiang. An algorithm with nearly optimal pseudo-regret for
both stochastic and adversarial bandits. In Proceedings of the 29th Annual
 BIBLIOGRAPHY 552

Conference on Learning Theory, pages 116–120, New York, NY, USA, 2016.
JMLR.org. [157]
P. Auer and R. Ortner. Logarithmic online regret bounds for undiscounted
reinforcement learning. In Advances in Neural Information Processing Systems,
pages 49–56. MIT Press, 2007. [536]
P. Auer and R. Ortner. UCB revisited: Improved regret bounds for the stochastic
multi-armed bandit problem. Periodica Mathematica Hungarica, 61(1-2):55–65,
2010. [98, 128]
P. Auer, N. Cesa-Bianchi, Y. Freund, and R. E. Schapire. Gambling in a
rigged casino: The adversarial multi-armed bandit problem. In Foundations
of Computer Science, 1995. Proceedings., 36th Annual Symposium on, pages
322–331. IEEE, 1995. [97, 146, 159, 202]
P. Auer, N. Cesa-Bianchi, and P. Fischer. Finite-time analysis of the multiarmed
bandit problem. Machine Learning, 47:235–256, 2002a. [95, 110]
P. Auer, N. Cesa-Bianchi, Y. Freund, and R. E. Schapire. The nonstochastic
multiarmed bandit problem. SIAM Journal on Computing, 32(1):48–77, 2002b.
[171, 203, 232, 385]
P. Auer, R. Ortner, and Cs. Szepesvári. Improved rates for the stochastic
continuum-armed bandit problem. In International Conference on
Computational Learning Theory, pages 454–468. Springer, 2007. [358]
P. Auer, T. Jaksch, and R. Ortner. Near-optimal regret bounds for reinforcement
learning. In Advances in Neural Information Processing Systems, pages 89–96,
2009. [536]
P. Auer, P. Gajane, and R. Ortner. Adaptively tracking the best arm with
an unknown number of distribution changes. In European Workshop on
Reinforcement Learning 14, 2018. [385]
P. Auer, P. Gajane, and R. Ortner. Adaptively tracking the best bandit arm
with an unknown number of distribution changes. In Proceedings of the 32nd
Conference on Learning Theory, 2019. [383, 385]
B. Awerbuch and R. Kleinberg. Adaptive routing with end-to-end feedback:
Distributed learning and geometric approaches. In Proceedings of the 36th
annual ACM symposium on theory of computing, pages 45–53. ACM, 2004.
[373]
S. J. Axler. Linear algebra done right, volume 2. Springer, 1997. [502]
M. G. Azar, I. Osband, and R. Munos. Minimax regret bounds for reinforcement
learning. In Proceedings of the 34th International Conference on Machine
Learning, pages 263–272, Sydney, Australia, 06–11 Aug 2017. JMLR.org. [536]
A. Badanidiyuru, R. Kleinberg, and A. Slivkins. Bandits with knapsacks. In
Foundations of Computer Science (FOCS), 2013 IEEE 54th Annual Symposium
on, pages 207–216. IEEE, 2013. [359]
P. L. Bartlett and A. Tewari. Regal: A regularization based algorithm for
reinforcement learning in weakly communicating MDPs. In Proceedings of
the 25th Conference on Uncertainty in Artificial Intelligence, pages 35–42,
Arlington, VA, United States, 2009. AUAI Press. [541]
 BIBLIOGRAPHY 553

G. Bartók. A near-optimal algorithm for finite partial-monitoring games against

adversarial opponents. In Proceedings of the 26th Conference on Learning
Theory, pages 696–710. JMLR.org, 2013. [504, 506]
G. Bartók and Cs. Szepesvári. Partial monitoring with side information. In
Proceedings of the 23rd International Conference on Algorithmic Learning
Theory, pages 305–319, 2012. [507]
G. Bartók, D. Pál, and Cs. Szepesvári. Toward a classification of finite partial-
monitoring games. In Proceedings of the 21st International Conference on
Algorithmic Learning Theory, pages 224–238. Springer, 2010. [506]
G. Bartók, N. Zolghadr, and Cs. Szepesvári. An adaptive algorithm for
finite stochastic partial monitoring. In Proceedings of the 29th International
Conference on Machine Learning, pages 1779–1786, USA, 2012. Omnipress.
[506]
G. Bartók, D. P. Foster, D. Pál, A. Rakhlin, and Cs. Szepesvári. Partial
monitoring—classification, regret bounds, and algorithms. Mathematics of
Operations Research, 39(4):967–997, 2014. [506]
H. Bastani and M. Bayati. Online decision making with high-dimensional
covariates. Operations Research, 68(1):276–294, 2020. [285]
J. A. Bather and H. Chernoff. Sequential decisions in the control of a spaceship. In
Fifth Berkeley Symposium on Mathematical Statistics and Probability, volume 3,
pages 181–207, 1967. [17]
T. Bayes. LII. An essay towards solving a problem in the doctrine of chances. by
the late Rev. Mr. Bayes, FRS communicated by Mr. Price, in a letter to John
Canton, AMFR S. Philosophical transactions of the Royal Society of London,
53:370–418, 1763. [435]
R. Bellman. The theory of dynamic programming. Technical report, RAND
CORP SANTA MONICA CA, 1954. [535]
R. E. Bellman. Eye of the Hurricane. World Scientific, 1984. [535]
D. Berend and A. Kontorovich. On the concentration of the missing mass.
Electronic Communications in Probability, 18(3):1–7, 2013. [84]
J. O. Berger. Statistical Decision Theory and Bayesian Analysis. Springer Science
& Business Media, 1985. [432]
D. Bernoulli. Exposition of a new theory on the measurement of risk.
Econometrica: Journal of the Econometric Society, pages 23–36, 1954. [67]
A. C. Berry. The accuracy of the Gaussian approximation to the sum of
independent variates. Transactions of the American mathematical society,
49(1):122–136, 1941. [78]
D. Berry and B. Fristedt. Bandit problems : sequential allocation of experiments.
Chapman and Hall, London ; New York, 1985. [17, 454, 456]
D. A. Berry, R. W. Chen, A. Zame, D. C. Heath, and L. A. Shepp. Bandit
problems with infinitely many arms. The Annals of Statistics, 25(5):2103–2116,
1997. [358]
D. Bertsekas and J. N. Tsitsiklis. Neuro-Dynamic Programming. Athena Scientific,
1st edition, 1996. [535]
 BIBLIOGRAPHY 554

D. P. Bertsekas. Dynamic Programming and Optimal Control, volume 1-2. Athena

Scientific, Belmont, MA, 4 edition, 2012. [534, 535]
D. P. Bertsekas. Convex optimization algorithms. Athena Scientific Belmont,
2015. [375]
D. Bertsimas and J. N. Tsitsiklis. Introduction to linear optimization, volume 6.
Athena Scientific Belmont, MA, 1997. [535]
O. Besbes, Y. Gur, and A. Zeevi. Stochastic multi-armed-bandit problem with
non-stationary rewards. In Advances in Neural Information Processing Systems,
pages 199–207. Curran Associates, Inc., 2014. [385]
L. Besson and E. Kaufmann. What doubling tricks can and can’t do for multi-
armed bandits. arXiv:1803.06971, 2018. [97]
A. Beygelzimer, J. Langford, L. Li, L. Reyzin, and R. E. Schapire. An optimal
high probability algorithm for the contextual bandit problem. arXiv:1002.4058,
2010. [171]
A. Beygelzimer, J. Langford, L. Li, L. Reyzin, and R. Schapire. Contextual
bandit algorithms with supervised learning guarantees. In Proceedings of the
14th International Conference on Artificial Intelligence and Statistics, pages
19–26, Fort Lauderdale, FL, USA, 2011. JMLR.org. [233, 236]
P. Billingsley. Probability and measure. John Wiley & Sons, 2008. [41, 53]
D. Blackwell. Controlled random walks. In Proceedings of the International
Congress of Mathematicians, volume 3, pages 336–338, 1954. [146]
V. I. Bogachev. Measure theory, volume 2. Springer Science & Business Media,
2007. [42, 316]
T. Bonald and A. Proutiere. Two-target algorithms for infinite-armed bandits
with bernoulli rewards. In Advances in Neural Information Processing Systems,
pages 2184–2192, 2013. [358]
L. Bottou, J. Peters, J. Quiñonero-Candela, D. X. Charles, D. M. Chickering,
E. Portugaly, D. Ray, P. Simard, and E. Snelson. Counterfactual reasoning
and learning systems: The example of computational advertising. The Journal
of Machine Learning Research, 14(1):3207–3260, 2013. [171]
S. Boucheron, G. Lugosi, and P. Massart. Concentration inequalities: A
nonasymptotic theory of independence. OUP Oxford, 2013. [80, 341]
D. Bouneffouf and I. Rish. A survey on practical applications of multi-armed and
contextual bandits. arXiv:1904.10040, 2019. [17]
G. E. P. Box. Science and statistics. Journal of the American Statistical
Association, 71(356):791–799, 1976. [145]
G. E. P. Box. Robustness in the strategy of scientific model building. Robustness
in statistics, 1:201–236, 1979. [145]
S. Boyd and L. Vandenberghe. Convex optimization. Cambridge University Press,
2004. [313]
R. N. Bradt, S. M. Johnson, and S. Karlin. On sequential designs for maximizing
the sum of n observations. The Annals of Mathematical Statistics, pages
1060–1074, 1956. [456]
 BIBLIOGRAPHY 555

R. Brafman and M. Tennenholtz. R-MAX – a general polynomial time algorithm

for near-optimal reinforcement learning. Journal of Machine Learning Research,
3:213–231, 2003. [537]
J. Bretagnolle and C. Huber. Estimation des densités: risque minimax. Zeitschrift
für Wahrscheinlichkeitstheorie und verwandte Gebiete, 47(2):119–137, 1979.
[193]
S. Bubeck and N. Cesa-Bianchi. Regret analysis of stochastic and nonstochastic
multi-armed bandit problems. Foundations and Trends in Machine Learning,
5(1):1–122, 2012. [17, 110, 157, 343, 373]
S. Bubeck and R. Eldan. The entropic barrier: a simple and optimal universal
self-concordant barrier. In Proceedings of the 28th Conference on Learning
Theory, pages 279–279, Paris, France, 2015. JMLR.org. [323]
S. Bubeck and R. Eldan. Multi-scale exploration of convex functions and bandit
convex optimization. In Proceedings of the 29th Conference on Learning Theory,
pages 583–589, New York, NY, USA, 2016. JMLR.org. [359, 472, 473]
S. Bubeck and C. Liu. Prior-free and prior-dependent regret bounds for Thompson
sampling. In Advances in Neural Information Processing Systems, pages 638–
646. Curran Associates, Inc., 2013. [474]
S. Bubeck and A. Slivkins. The best of both worlds: Stochastic and adversarial
bandits. In Proceedings of the 25th Conference on Learning Theory, pages
42.1–42.23, 2012. [157]
S. Bubeck, R. Munos, and G. Stoltz. Pure exploration in multi-armed bandits
problems. In International conference on Algorithmic learning theory, pages
23–37. Springer, 2009. [414]
S. Bubeck, R. Munos, G. Stoltz, and Cs. Szepesvári. X-armed bandits. Journal
of Machine Learning Research, 12:1655–1695, 2011. [358]
S. Bubeck, N. Cesa-Bianchi, and S. Kakade. Towards minimax policies for online
linear optimization with bandit feedback. In Proceedings of the 25th Conference
on Learning Theory, pages 41–1. Microtome, 2012. [274, 323, 343]
S. Bubeck, N. Cesa-Bianchi, and G. Lugosi. Bandits with heavy tail. IEEE
Transactions on Information Theory, 59(11):7711–7717, 2013a. [115]
S. Bubeck, V. Perchet, and P. Rigollet. Bounded regret in stochastic multi-armed
bandits. In Proceedings of the 26th Annual Conference on Learning Theory,
pages 122–134, Princeton, NJ, USA, 2013b. JMLR.org. [202] √
S. Bubeck, O. Dekel, T. Koren, and Y. Peres. Bandit convex optimization: T
regret in one dimension. In Proceedings of the 28th Conference on Learning
Theory, pages 266–278, Paris, France, 2015a. JMLR.org. [359, 472, 473]
S. Bubeck, R. Eldan, and J. Lehec. Finite-time analysis of projected Langevin
Monte Carlo. In Advances in Neural Information Processing Systems, pages
1243–1251. Curran Associates, Inc., 2015b. [323, 340]
S. Bubeck, Y.T. Lee, and R. Eldan. Kernel-based methods for bandit convex
optimization. In Proceedings of the 49th Annual ACM SIGACT Symposium on
Theory of Computing, STOC 2017, pages 72–85, New York, NY, USA, 2017.
ACM. ISBN 978-1-4503-4528-6. [359]
 BIBLIOGRAPHY 556

S. Bubeck, M. Cohen, and Y. Li. Sparsity, variance and curvature in multi-armed

bandits. In Proceedings of the 29th International Conference on Algorithmic
Learning Theory, pages 111–127. JMLR.org, 07–09 Apr 2018. [170, 339, 343]
A. N. Burnetas and M. N. Katehakis. Optimal adaptive policies for sequential
allocation problems. Advances in Applied Mathematics, 17(2):122–142, 1996.
[120, 140, 209, 210]
A. N. Burnetas and M. N. Katehakis. Optimal adaptive policies for Markov
decision processes. Mathematics of Operations Research, 22(1):222–255, 1997a.
[536]
A. N. Burnetas and M. N. Katehakis. On the finite horizon one-armed bandit
problem. Stochastic Analysis and Applications, 16(1):845–859, 1997b. [455]
A. N. Burnetas and M. N. Katehakis. Asymptotic Bayes analysis for the
finite-horizon one-armed-bandit problem. Probability in the Engineering and
Informational Sciences, 17(1):53–82, 2003. [456]
R. R. Bush and F. Mosteller. A stochastic model with applications to learning.
The Annals of Mathematical Statistics, pages 559–585, 1953. [16]
O. Cappé, A. Garivier, O. Maillard, R. Munos, and G. Stoltz. Kullback–Leibler
upper confidence bounds for optimal sequential allocation. The Annals of
Statistics, 41(3):1516–1541, 2013. [120, 140, 141, 210]
A. Carpentier and A. Locatelli. Tight (lower) bounds for the fixed budget best
arm identification bandit problem. In Proceedings of the 29th Conference on
Learning Theory, pages 590–604, New York, NY, USA, 2016. JMLR.org. [414]
A. Carpentier and R. Munos. Bandit theory meets compressed sensing for high
dimensional stochastic linear bandit. In Proceedings of the 15th International
Conference on Artificial Intelligence and Statistics, pages 190–198, La Palma,
Canary Islands, 2012. JMLR.org. [286, 355]
A. Carpentier and M. Valko. Extreme bandits. In Advances in Neural Information
Processing Systems, pages 1089–1097. Curran Associates, Inc., 2014. [416]
A. Carpentier and M. Valko. Simple regret for infinitely many armed bandits. In
Proceedings of the 32nd International Conference on Machine Learning, pages
1133–1141, Lille, France, 2015. PMLR. [358]
O. Catoni. Challenging the empirical mean and empirical variance: a deviation
study. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 48(4):
1148–1185, 2012. [115]
N. Cesa-Bianchi and G. Lugosi. Prediction, learning, and games. Cambridge
University Press, 2006. [17, 157, 343, 385, 433, 506]
N. Cesa-Bianchi and G. Lugosi. Combinatorial bandits. Journal of Computer
and System Sciences, 78(5):1404–1422, 2012. [373]
N. Cesa-Bianchi, G. Lugosi, and G. Stoltz. Regret minimization under partial
monitoring. Mathematics of Operations Research, 31:562–580, 2006. [506]
N. Cesa-Bianchi, C. Gentile, Y. Mansour, and A. Minora. Delay and cooperation
in nonstochastic bandits. In Proceedings of the 29th Conference on Learning
Theory, pages 605–622, New York, NY, USA, 2016. JMLR.org. [360]
 BIBLIOGRAPHY 557

N. Cesa-Bianchi, C. Gentile, G. Lugosi, and G. Neu. Boltzmann exploration done

right. In Advances in Neural Information Processing Systems, pages 6284–6293.
Curran Associates, Inc., 2017. [95]
J. Chakravorty and A. Mahajan. Multi-armed bandits, Gittins index, and its
calculation. Methods and Applications of Statistics in Clinical Trials: Planning,
Analysis, and Inferential Methods, 2:416–435, 2013. [456]
J. Chakravorty and A. Mahajan. Multi-armed bandits, Gittins index, and its
calculation. Methods and Applications of Statistics in Clinical Trials: Planning,
Analysis, and Inferential Methods, 2:416–435, 2014. [456]
H. P. Chan and T. L. Lai. Sequential generalized likelihood ratios and adaptive
treatment allocation for optimal sequential selection. Sequential Analysis, 25:
179–201, 2006. [415]
J. T. Chang and D. Pollard. Conditioning as disintegration. Statistica Neerlandica,
51(3):287–317, 1997. [435]
O. Chapelle and L. Li. An empirical evaluation of Thompson sampling. In
Advances in Neural Information Processing Systems, pages 2249–2257. Curran
Associates, Inc., 2011. [473]
S. Chaudhuri and A. Tewari. Phased exploration with greedy exploitation in
stochastic combinatorial partial monitoring games. In Advances in Neural
Information Processing Systems, pages 2433–2441, 2016. [507]
C-H. Chen, J. Lin, E. Yücesan, and S. E. Chick. Simulation budget allocation
for further enhancing the efficiency of ordinal optimization. Discrete Event
Dynamic Systems, 10(3):251–270, 2000. [416]
S. Chen, T. Lin, I. King, M. R. Lyu, and W. Chen. Combinatorial pure exploration
of multi-armed bandits. In Advances in Neural Information Processing Systems,
pages 379–387. Curran Associates, Inc., 2014. [414]
W. Chen, Y. Wang, and Y. Yuan. Combinatorial multi-armed bandit: General
framework and applications. In Proceedings of the 30th International Conference
on Machine Learning, pages 151–159, Atlanta, Georgia, USA, 17–19 Jun 2013.
PMLR. [373]
W. Chen, W. Hu, F. Li, J. Li, Y. Liu, and P. Lu. Combinatorial multi-armed
bandit with general reward functions. In Advances in Neural Information
Processing Systems, pages 1659–1667. Curran Associates, Inc., 2016a. [373]
W. Chen, Y. Wang, Y. Yuan, and Q. Wang. Combinatorial multi-armed bandit
and its extension to probabilistically triggered arms. Journal of Machine
Learning Research, 17(50):1–33, 2016b. URL http://jmlr.org/papers/v17/
14-298.html. [373]
Y. Chen, C-W. Lee, H. Luo, and C-Y. Wei. A new algorithm for non-stationary
contextual bandits: Efficient, optimal, and parameter-free. arXiv:1902.00980,
2019. [383, 385]
Y. R. Chen and M. N. Katehakis. Linear programming for finite state multi-armed
bandit problems. Mathematics of Operations Research, 11(1):180–183, 1986.
[456]
 BIBLIOGRAPHY 558

H. Chernoff. Sequential design of experiments. The Annals of Mathematical

Statistics, 30(3):755–770, 1959. [17, 414]
H. Chernoff. A career in statistics. Past, Present, and Future of Statistical
Science, page 29, 2014. [140]
W. Cheung, D. Simchi-Levi, and R. Zhu. Learning to optimize under non-
stationarity. In Proceedings of the 22nd International Conference on Artificial
Intelligence and Statistics, pages 1079–1087. PMLR, 16–18 Apr 2019. [385]
W. Chu, L. Li, L. Reyzin, and R. Schapire. Contextual bandits with linear payoff
functions. In Proceedings of the 14th International Conference on Artificial
Intelligence and Statistics, pages 208–214, Fort Lauderdale, FL, USA, 2011.
JMLR.org. [274]
A. Chuklin, I. Markov, and M. de Rijke. Click Models for Web Search. Morgan
& Claypool Publishers, 2015. [399]
V. A. Cicirello and S. F. Smith. The max k-armed bandit: A new model of
exploration applied to search heuristic selection. In AAAI, pages 1355–1361,
2005. [416]
A. Cohen and T. Hazan. Following the perturbed leader for online structured
learning. In Proceedings of the 32nd International Conference on Machine
Learning, pages 1034–1042, Lille, France, 07–09 Jul 2015. JMLR.org. [373, 375]
A. Cohen, T. Hazan, and T. Koren. Tight bounds for bandit combinatorial
optimization. In Proceedings of the 2017 Conference on Learning Theory, pages
629–642, Amsterdam, Netherlands, 2017. JMLR.org. [371]
R. Combes, S. Magureanu, A. Proutiere, and C. Laroche. Learning to rank:
Regret lower bounds and efficient algorithms. In Proceedings of the 2015 ACM
SIGMETRICS International Conference on Measurement and Modeling of
Computer Systems, pages 231–244. ACM, 2015a. ISBN 978-1-4503-3486-0.
[399]
R. Combes, M. Shahi, A. Proutiere, and M. Lelarge. Combinatorial bandits
revisited. In Advances in Neural Information Processing Systems, pages 2116–
2124. Curran Associates, Inc., 2015b. [373]
R. Combes, S. Magureanu, and A. Proutière. Minimal exploration in structured
stochastic bandits. In Advances in Neural Information Processing Systems,
pages 1761–1769, 2017. [246, 249, 301, 358]
A. R. Conn, K. Scheinberg, and L. N. Vicente. Introduction to Derivative-Free
Optimization. SIAM, 2009. [415]
T. M. Cover. Universal portfolios. Mathematical Finance, 1(1):1–29, 1991. [323]
T. M. Cover and J. A. Thomas. Elements of information theory. John Wiley &
Sons, 2012. [193]
W. Cowan and M. N. Katehakis. An asymptotically optimal policy for uniform
bandits of unknown support. arXiv:1505.01918, 2015. [210]
W. Cowan, J. Honda, and M. N. Katehakis. Normal bandits of unknown means
and variances. Journal of Machine Learning Research, 18(154):1–28, 2018.
[210]
 BIBLIOGRAPHY 559

K. Crammer and C. Gentile. Multiclass classification with bandit feedback using

adaptive regularization. Machine learning, 90(3):347–383, 2013. [286]
N. Craswell, O. Zoeter, M. Taylor, and B. Ramsey. An experimental comparison of
click position-bias models. In Proceedings of the 2008 International Conference
on Web Search and Data Mining, pages 87–94. ACM, 2008. [399]
V. Dani and T. P. Hayes. Robbing the bandit: Less regret in online geometric
optimization against an adaptive adversary. In 17th Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 937–943, 2006. [373]
V. Dani, T. P. Hayes, and S. M. Kakade. Stochastic linear optimization under
bandit feedback. In Proceedings of the 21st Conference on Learning Theory,
pages 355–366, 2008. [246, 294]
V. H. de la Peña, T.L. Lai, and Q. Shao. Self-normalized processes: Limit theory
and Statistical Applications. Springer Science & Business Media, 2008. [80, 86,
261]
R. Degenne and W. M. Koolen. Pure exploration with multiple correct answers.
In Advances in Neural Information Processing Systems, pages 14591–14600.
Curran Associates, Inc., 2019. [414]
R. Degenne and V. Perchet. Anytime optimal algorithms in stochastic multi-
armed bandits. In Proceedings of the 33rd International Conference on Machine
Learning, pages 1587–1595, New York, NY, USA, 20–22 Jun 2016. JMLR.org.
[128]
R. Degenne, W. M. Koolen, and P. Ménard. Non-asymptotic pure exploration by
solving games. In Advances in Neural Information Processing Systems, pages
14492–14501. Curran Associates, Inc., 2019. [414]
O. Dekel, C. Gentile, and K. Sridharan. Robust selective sampling from single
and multiple teachers. In Proceedings of the 23rd Conference on Learning
Theory, pages 346–358, 2010. [286]
O. Dekel, C. Gentile, and K. Sridharan. Selective sampling and active learning
from single and multiple teachers. Journal of Machine Learning Research, 13:
2655–2697, 2012. [286]
A. Dembo and O. Zeitouni. Large deviations techniques and applications,
volume 38. Springer Science & Business Media, 2009. [83]
E. V. Denardo, H. Park, and U. G. Rothblum. Risk-sensitive and risk-neutral
multiarmed bandits. Mathematics of Operations Research, 32(2):374–394, 2007.
[68]
T. Desautels, A. Krause, and J. W. Burdick. Parallelizing exploration-exploitation
tradeoffs in gaussian process bandit optimization. Journal of Machine Learning
Research, 15:4053–4103, 2014. [360]
R. L. Dobrushin. Eine allgemeine formulierung des fundamentalsatzes von shannon
in der informationstheorie. Usp. Mat. Nauk, 14(6(90)):3–104, 1959. [193]
S. Dong and B. Van Roy. An information-theoretic analysis for Thompson
sampling with many actions. In Advances in Neural Information Processing
Systems, Red Hook, NY, USA, 2018. Curran Associates Inc. [471]
J. L. Doob. Stochastic processes. Wiley, 1953. [54]
 BIBLIOGRAPHY 560

M. Dudı́k, D. Hsu, S. Kale, N. Karampatziakis, J. Langford, L. Reyzin, and

T. Zhang. Efficient optimal learning for contextual bandits. In Proceedings of
the 27th Conference on Uncertainty in Artificial Intelligence, pages 169–178.
AUAI Press, 2011. [233]
M. Dudı́k, K. Hofmann, R. E. Schapire, A. Slivkins, and M. Zoghi. Contextual
dueling bandits. In Proceedings of the 28th Conference on Learning Theory,
pages 563–587, Paris, France, 2015. JMLR.org. [359]
R. M. Dudley. Uniform central limit theorems, volume 142. Cambridge University
Press, 2014. [80, 341]
C. G. Esseen. On the Liapounoff limit of error in the theory of probability.
Almqvist & Wiksell, 1942. [78]
E. Even-Dar, S. Mannor, and Y. Mansour. PAC bounds for multi-armed bandit
and Markov decision processes. In Computational Learning Theory, pages
255–270. Springer, 2002. [414, 419]
E. Even-Dar, S. M. Kakade, and Y. Mansour. Experts in a Markov decision
process. In Advances in Neural Information Processing Systems, pages 401–408,
Cambridge, MA, USA, 2004. MIT Press. [537]
E. Even-Dar, S. Mannor, and Y. Mansour. Action elimination and stopping
conditions for the multi-armed bandit and reinforcement learning problems.
Journal of Machine Learning Research, 7:1079–1105, 2006. [414]
V. V. Fedorov. Theory of optimal experiments. Academic Press, New York, 1972.
[270]
S. Filippi, O. Cappé, A. Garivier, and Cs. Szepesvári. Parametric bandits: The
generalized linear case. In Advances in Neural Information Processing Systems,
pages 586–594. Curran Associates, Inc., 2010. [246, 247]
D. Fink. A compendium of conjugate priors, 1997. [435]
D. Foster and A. Rakhlin. No internal regret via neighborhood watch. In
Proceedings of the 15th International Conference on Artificial Intelligence and
Statistics, pages 382–390, La Palma, Canary Islands, 2012. JMLR.org. [506]
D. J. Foster and A. Rakhlin. Beyond UCB: Optimal and efficient contextual
bandits with regression oracles. arXiv:2002.04926, 2020. [294]
M. Frank and P. Wolfe. An algorithm for quadratic programming. Naval Research
Logistics Quarterly, 3(1-2):95–110, 1956. [270]
S. Frederick, G. Loewenstein, and T. O’donoghue. Time discounting and time
preference: A critical review. Journal of Economic Literature, 40(2):351–401,
2002. [454]
E. Frostig and G. Weiss. Four proofs of Gittins’ multiarmed bandit theorem.
Applied Probability Trust, 70, 1999. [456]
R. Fruit, M. Pirotta, and A. Lazaric. Near optimal exploration-exploitation
in non-communicating markov decision processes. In Advances in Neural
Information Processing Systems, pages 2997–3007, 2018. [536, 546, 547]
Y. Gai, B. Krishnamachari, and R. Jain. Combinatorial network optimization with
unknown variables: Multi-armed bandits with linear rewards and individual
 BIBLIOGRAPHY 561

observations. IEEE/ACM Transactions on Networking, 20(5):1466–1478, 2012.

[373]
P. Gajane, R. Ortner, and P. Auer. A sliding-window algorithm for
Markov decision processes with arbitrarily changing rewards and transitions.
arXiv:1805.10066, 2018. [385]
A. Garivier. Informational confidence bounds for self-normalized averages and
applications. arXiv:1309.3376, 2013. [111, 127]
A. Garivier and O. Cappé. The KL-UCB algorithm for bounded stochastic
bandits and beyond. In Proceedings of the 24th Conference on Learning Theory,
2011. [139, 140]
A. Garivier and E. Kaufmann. Optimal best arm identification with fixed
confidence. In Proceedings of the 29th Conference on Learning Theory, pages
998–1027, New York, NY, USA, 2016. JMLR.org. [414]
A. Garivier and E. Moulines. On upper-confidence bound policies for switching
bandit problems. In Proceedings of the 22nd International Conference on
Algorithmic Learning Theory, pages 174–188, Berlin, Heidelberg, 2011. Springer
Berlin Heidelberg. [385]
A. Garivier, E. Kaufmann, and W. M. Koolen. Maximin action identification: A
new bandit framework for games. In Proceedings of the 29th Conference on
Learning Theory, pages 1028–1050, New York, NY, USA, 2016a. JMLR.org.
[414]
A. Garivier, T. Lattimore, and E. Kaufmann. On explore-then-commit strategies.
In Advances in Neural Information Processing Systems, pages 784–792. Curran
Associates, Inc., 2016b. [94, 119, 210]
A. Garivier, P. Ménard, and G. Stoltz. Explore first, exploit next: The true
shape of regret in bandit problems. Mathematics of Operations Research, 44
(2):377–399, 2019. [210]
A. Gelman, J. B. Carlin, H. S. Stern, D. B. Dunson, A. Vehtari, and D. B. Rubin.
Bayesian data analysis, volume 2. CRC Press Boca Raton, FL, 2014. [435]
C. Gentile and F. Orabona. On multilabel classification and ranking with partial
feedback. In Advances in Neural Information Processing Systems, pages 1151–
1159. Curran Associates, Inc., 2012. [286]
C. Gentile and F. Orabona. On multilabel classification and ranking with bandit
feedback. Journal of Machine Learning Research, 15(1):2451–2487, 2014. [286]
S. Gerchinovitz. Sparsity regret bounds for individual sequences in online linear
regression. Journal of Machine Learning Research, 14(Mar):729–769, 2013.
[286]
S. Gerchinovitz and T. Lattimore. Refined lower bounds for adversarial bandits. In
Advances in Neural Information Processing Systems, pages 1198–1206. Curran
Associates, Inc., 2016. [202, 219]
S. Ghosal and A. van der Vaart. Fundamentals of nonparametric Bayesian
inference, volume 44. Cambridge University Press, 2017. [434, 435]
A. Ghosh, S. R. Chowdhury, and A. Gopalan. Misspecified linear bandits. In
31st AAAI Conference on Artificial Intelligence, 2017. [274]
 BIBLIOGRAPHY 562

J. Gittins. Bandit processes and dynamic allocation indices. Journal of the Royal
Statistical Society. Series B (Methodological), 41(2):148–177, 1979. [384, 456]
J. Gittins, K. Glazebrook, and R. Weber. Multi-armed bandit allocation indices.
John Wiley & Sons, 2011. [17, 384, 456]
D. Glowacka. Bandit algorithms in information retrieval. Foundations and
Trends® in Information Retrieval, 13:299–424, 01 2019. [400]
P. Glynn and S. Juneja. Ordinal optimization – empirical large deviations rate
estimators, and stochastic multi-armed bandits. arXiv:1507.04564, 2015. [416]
D. Goldsman. Ranking and selection in simulation. In 15th conference on Winter
Simulation, pages 387–394, 1983. [416]
A. Gopalan and S. Mannor. Thompson sampling for learning parameterized
Markov decision processes. In Proceedings of the 28th Conference on Learning
Theory, pages 861–898, Paris, France, 2015. JMLR.org. [473]
G. J. Gordon. Regret bounds for prediction problems. In Proceedings of the 12th
Conference on Learning Theory, pages 29–40, 1999. [342, 343]
T. Graepel, J. Q. Candela, T. Borchert, and R. Herbrich. Web-scale Bayesian
click-through rate prediction for sponsored search advertising in microsoft’s
bing search engine. In Proceedings of the 27th International Conference on
Machine Learning, pages 13–20, USA, 2010. Omnipress. [473]
O. Granmo. Solving two-armed bernoulli bandit problems using a Bayesian
learning automaton. International Journal of Intelligent Computing and
Cybernetics, 3(2):207–234, 2010. [473]
R. M. Gray. Entropy and information theory. Springer Science & Business Media,
2011. [193]
K. Greenewald, A. Tewari, S. Murphy, and P. Klasnja. Action centered contextual
bandits. In Advances in Neural Information Processing Systems, pages 5977–
5985. Curran Associates, Inc., 2017. [17]
M. Grötschel, L. Lovász, and A. Schrijver. Geometric algorithms and combinatorial
optimization, volume 2. Springer Science & Business Media, 2012. [270, 372,
535]
F. Guo, C. Liu, and Y. M. Wang. Efficient multiple-click models in web search.
In Proceedings of the 2nd ACM International Conference on Web Search and
Data Mining, pages 124–131. ACM, 2009. [400]
A. György and Cs. Szepesvári. Shifting regret, mirror descent, and matrices. In
Proceedings of the 33rd International Conference on Machine Learning, pages
2943–2951, New York, NY, USA, 20–22 Jun 2016. JMLR.org. [385]
A. György, T. Linder, G. Lugosi, and G. Ottucsák. The on-line shortest path
problem under partial monitoring. Journal of Machine Learning Research, 8
(Oct):2369–2403, 2007. [373, 374]
A. György, D. Pál, and Cs. Szepesvári. Online learning: Algorithms for Big Data.
2019. [385]
P. R. Halmos. Measure Theory. Graduate Texts in Mathematics. Springer New
York, 1976. [53]
 BIBLIOGRAPHY 563

Nima Hamidi and Mohsen Bayati. On worst-case regret of linear thompson

sampling. arXiv preprint arXiv:2006.06790, 2020. [471]
M. Hanawal, V. Saligrama, M. Valko, and R. Munos. Cheap bandits. In
Proceedings of the 32nd International Conference on Machine Learning, pages
2133–2142, Lille, France, 07–09 Jul 2015. JMLR.org. [360]
J. Hannan. Approximation to Bayes risk in repeated play. Contributions to the
Theory of Games, 3:97–139, 1957. [146, 343, 373]
B. Hao, T. Lattimore, and Cs. Szepesvári. Adaptive exploration in linear
contextual bandit. In Proceedings of the 23rd International Conference on
Artificial Intelligence and Statistics, 2020. [246, 301]
G. H. Hardy. Divergent Series. Oxford University Press, 1973. [533]
E. Hazan. Introduction to online convex optimization. Foundations and Trends®
in Optimization, 2(3-4):157–325, 2016. [342, 343]
E. Hazan and S. Kale. A simple multi-armed bandit algorithm with optimal
variation-bounded regret. In Proceedings of the 24th Conference on Learning
Theory, pages 817–820. JMLR.org, 2011. [170]
E. Hazan, Z. Karnin, and R. Meka. Volumetric spanners: an efficient exploration
basis for learning. Journal of Machine Learning Research, 17(119):1–34, 2016.
[270, 323]
D. P. Helmbold, N. Littlestone, and P. M. Long. Apple tasting. Information and
Computation, 161(2):85–139, 2000. [507]
M. Herbster and M. K. Warmuth. Tracking the best expert. Machine Learning,
32(2):151–178, 1998. [384]
M. Herbster and M. K. Warmuth. Tracking the best linear predictor. Journal of
Machine Learning Research, 1(Sep):281–309, 2001. [385]
Y-C. Ho, R. S. Sreenivas, and P. Vakili. Ordinal optimization of DEDS. Discrete
Event Dynamic Systems, 1992. [416]
J. Honda and A. Takemura. An asymptotically optimal bandit algorithm for
bounded support models. In Proceedings of the 23rd Conference on Learning
Theory, pages 67–79, 2010. [120, 128, 139, 140, 209, 210]
J. Honda and A. Takemura. An asymptotically optimal policy for finite support
models in the multiarmed bandit problem. Machine Learning, 85(3):361–391,
2011. [120]
J. Honda and A. Takemura. Optimality of Thompson sampling for Gaussian
bandits depends on priors. In Proceedings of the 17th International Conference
on Artificial Intelligence and Statistics, pages 375–383, Reykjavik, Iceland,
2014. JMLR.org. [473]
J. Honda and A. Takemura. Non-asymptotic analysis of a new bandit algorithm for
semi-bounded rewards. Journal of Machine Learning Research, 16:3721–3756,
2015. [140, 209, 210]
X. Hu, Prashanth L.A., A. György, and Cs. Szepesvári. (Bandit) convex
optimization with biased noisy gradient oracles. In AISTATS, pages 819–
828, 2016. [359, 415]
 BIBLIOGRAPHY 564

R. Huang, M. M. Ajallooeian, Cs. Szepesvári, and M. Müller. Structured best arm

identification with fixed confidence. In Proceedings of the 28th International
Conference on Algorithmic Learning Theory, pages 593–616, Kyoto, Japan,
2017a. JMLR.org. [414]
R. Huang, T. Lattimore, A. György, and Cs. Szepesvári. Following the leader
and fast rates in online linear prediction: Curved constraint sets and other
regularities. Journal of Machine Learning Research, 18:1–31, 2017b. [341]
W. Huang, J. Ok, L. Li, and W. Chen. Combinatorial pure exploration with
continuous and separable reward functions and its applications. In IJCAI,
pages 2291–2297, 2018. [414]
M. Hutter. Universal artificial intelligence: Sequential decisions based on
algorithmic probability. Springer Science & Business Media, 2004. [433]
M. Hutter and J. Poland. Adaptive online prediction by following the perturbed
leader. Journal of Machine Learning Research, 6:639–660, 2005. [373]
C. Ionescu Tulcea. Mesures dans les espaces produits. Atti Accademia Nazionale
Lincei Rend, 7:208–211, 1949–50. [54]
E. L. Ionides. Truncated importance sampling. Journal of Computational and
Graphical Statistics, 17(2):295–311, 2008. [171]
V. I. Ivanenko and V. A. Labkovsky. On regularities of mass random phenomena.
arXiv:1204.4440, 2013. [146]
T. Jaksch, P. Auer, and R. Ortner. Near-optimal regret bounds for reinforcement
learning. Journal of Machine Learning Research, 99:1563–1600, August 2010.
ISSN 1532-4435. [536, 539]
K. Jamieson and R. Nowak. Best-arm identification algorithms for multi-armed
bandits in the fixed confidence setting. In Information Sciences and Systems
(CISS), 2014 48th Annual Conference on, pages 1–6. IEEE, 2014. [414]
K. Jamieson and A. Talwalkar. Non-stochastic best arm identification and
hyperparameter optimization. In Proceedings of the 19th International
Conference on Artificial Intelligence and Statistics, pages 240–248, 2016. [416]
K. Jamieson, S. Katariya, A. Deshpande, and R. Nowak. Sparse dueling bandits.
In Proceedings of the 18th International Conference on Artificial Intelligence
and Statistics, pages 416–424, San Diego, CA, USA, 2015. JMLR.org. [359]
E. T. Jaynes. Probability theory: the logic of science. Cambridge University Press,
2003. [434, 435]
A. Jefferson, L. Bortolotti, and B. Kuzmanovic. What is unrealistic optimism?
Consciousness and Cognition, 50:3–11, 2017. [110]
P. Joulani, A. György, and Cs. Szepesvári. Online learning under delayed feedback.
In Proceedings of the 30th International Conference on Machine Learning, pages
1453–1461, Atlanta, GA, USA, 2013. JMLR.org. [360]
P. Joulani, A. György, and Cs. Szepesvári. A modular analysis of adaptive (non-
)convex optimization: Optimism, composite objectives, and variational bounds.
In Proceedings of the 28th International Conference on Algorithmic Learning
Theory, pages 681–720, Kyoto University, Kyoto, Japan, 2017. JMLR.org. [339]
 BIBLIOGRAPHY 565

K. Jun, A. Bhargava, R. Nowak, and R. Willett. Scalable generalized linear

bandits: Online computation and hashing. In Advances in Neural Information
Processing Systems, pages 99–109. Curran Associates, Inc., 2017. [247]
L. P. Kaelbling. Learning in embedded systems. MIT Press, 1993. [110]
D. Kahneman and A. Tversky. Prospect theory: An analysis of decision under
risk. Econometrica, 47(2):263–91, 1979. [67]
S. Kakade. On The Sample Complexity Of Reinforcement Learning. PhD thesis,
University College London, 2003. [537]
S. M. Kakade, S. Shalev-Shwartz, and A. Tewari. Efficient bandit algorithms for
online multiclass prediction. In Proceedings of the 25th International Conference
on Machine Learning, pages 440–447, 2008. [233]
A. Kalai and S. Vempala. Geometric algorithms for online optimization. Technical
Report MIT-LCS-TR-861, MIT, 2002. [343, 373]
A. Kalai and S. Vempala. Efficient algorithms for online decision problems.
Journal of Computer and System Sciences, 71(3):291–307, 2005. [373]
L. Kallenberg. A note on M.N. Katehakis’ and Y.-R. Chen’s computation of the
Gittins index. Mathematics of Operations Research, 11(1):184–186, 1986. [456]
L. Kallenberg. Markov decision processes: Lecture notes, 2016. [535]
O. Kallenberg. Foundations of modern probability. Springer-Verlag, 2002. [41, 42,
52, 53, 54, 195, 263, 436]
Z. Karnin, T. Koren, and O. Somekh. Almost optimal exploration in multi-armed
bandits. In Proceedings of the 30th International Conference on Machine
Learning, pages 1238–1246, Atlanta, GA, USA, 2013. JMLR.org. [414]
N. El Karoui and I. Karatzas. Dynamic allocation problems in continuous time.
The Annals of Applied Probability, pages 255–286, 1994. [456]
S. Katariya, B. Kveton, Cs. Szepesvári, and Z. Wen. DCM bandits: Learning to
rank with multiple clicks. In Proceedings of the 33rd International Conference
on Machine Learning, pages 1215–1224, 2016. [400]
S. Katariya, B. Kveton, Cs. Szepesvári, C. Vernade, and Z. Wen. Bernoulli
rank-1 bandits for click feedback. In Proceedings of the 26th International Joint
Conference on Artificial Intelligence, 2017a. [400]
S. Katariya, B. Kveton, Cs. Szepesvári, C. Vernade, and Z. Wen. Stochastic rank-
1 bandits. In Proceedings of the 20th International Conference on Artificial
Intelligence and Statistics, 2017b. [400]
M. N. Katehakis and H. Robbins. Sequential choice from several populations.
Proceedings of the National Academy of Sciences of the United States of America,
92(19):8584, 1995. [110, 119, 120]
V. Ya Katkovnik and Yu Kulchitsky. Convergence of a class of random search
algorithms. Automation Remote Control, 8:1321–1326, 1972. [415]
E. Kaufmann. On Bayesian index policies for sequential resource allocation. The
Annals of Statistics, 46(2):842–865, 04 2018. [120, 128, 471, 473]
E. Kaufmann, O. Cappé, and A. Garivier. On Bayesian upper confidence bounds
for bandit problems. In Proceedings of the 15th International Conference on
 BIBLIOGRAPHY 566

Artificial Intelligence and Statistics, pages 592–600, La Palma, Canary Islands,

2012a. JMLR.org. [471, 473]
E. Kaufmann, N. Korda, and R. Munos. Thompson sampling: An asymptotically
optimal finite-time analysis. In Proceedings of the 23rd International Conference
on Algorithmic Learning Theory, volume 7568 of Lecture Notes in Computer
Science, pages 199–213. Springer Berlin Heidelberg, 2012b. ISBN 978-3-642-
34105-2. [120, 470, 473]
J. Kawale, H. H. Bui, B. Kveton, L. Tran-Thanh, and S. Chawla. Efficient
Thompson sampling for online matrix-factorization recommendation. In
Advances in Neural Information Processing Systems, pages 1297–1305. Curran
Associates, Inc., 2015. [473]
A. Kazerouni, M. Ghavamzadeh, Y. Abbasi, and B. Van Roy. Conservative
contextual linear bandits. In Advances in Neural Information Processing
Systems, pages 3910–3919. Curran Associates, Inc., 2017. [360]
M. Kearns and L. Saul. Large deviation methods for approximate probabilistic
inference. In Proceedings of the 14th Conference on Uncertainty in Artificial
Intelligence, page 311–319. Morgan Kaufmann Publishers Inc., 1998. [84]
M. Kearns and S. Singh. Near-optimal reinforcement learning in polynomial time.
Machine Learning, 49(2-3):209–232, 2002. [537]
M. J. Kearns and U. V. Vazirani. An introduction to computational learning
theory. MIT Press, 1994. [233]
J. Kiefer and J. Wolfowitz. The equivalence of two extremum problems. Canadian
Journal of Mathematics, 12(5):363–365, 1960. [270]
G-S. Kim and M. C. Paik. Doubly-robust lasso bandit. In Advances in Neural
Information Processing Systems, pages 5877–5887. Curran Associates, Inc.,
2019. [285]
M. J. Kim. Thompson sampling for stochastic control: The finite parameter case.
IEEE Transactions on Automatic Control, 62(12):6415–6422, 2017. [473]
J. Kirschner and A. Krause. Information directed sampling and bandits with
heteroscedastic noise. In Proceedings of the 31st Conference On Learning
Theory, pages 358–384. PMLR, 06–09 Jul 2018. [87, 246]
J. Kirschner, T. Lattimore, and A. Krause. Information directed sampling for
linear partial monitoring. arXiv:2002.11182, 2020. [507]
R. Kleinberg. Nearly tight bounds for the continuum-armed bandit problem.
In Advances in Neural Information Processing Systems, pages 697–704. MIT
Press, 2005. [358]
R. Kleinberg, A. Slivkins, and E. Upfal. Multi-armed bandits in metric spaces.
In Proceedings of the 40th Annual ACM Symposium on Theory of Computing,
pages 681–690. ACM, 2008. [358]
T. Kocák, G. Neu, M. Valko, and R. Munos. Efficient learning by implicit
exploration in bandit problems with side observations. In Advances in Neural
Information Processing Systems, pages 613–621. Curran Associates, Inc., 2014.
[170, 171, 360]
 BIBLIOGRAPHY 567

T. Kocák, M. Valko, R. Munos, and S. Agrawal. Spectral Thompson sampling.

In AAAI, pages 1911–1917, 2014. [473]
L. Kocsis and Cs. Szepesvári. Discounted UCB. In 2nd PASCAL Challenges
Workshop, pages 784–791, 2006. [17, 385]
H. Komiya. Elementary proof for Sion’s minimax theorem. Kodai Mathematical
Journal, 11(1):5–7, 1988. [342]
J. Komiyama, J. Honda, H. Kashima, and H. Nakagawa. Regret lower bound
and optimal algorithm in dueling bandit problem. In Proceedings of the
28th Conference on Learning Theory, pages 1141–1154, Paris, France, 2015a.
JMLR.org. [359]
J. Komiyama, J. Honda, and H. Nakagawa. Regret lower bound and optimal
algorithm in finite stochastic partial monitoring. In Advances in Neural
Information Processing Systems, pages 1792–1800. Curran Associates, Inc.,
2015b. [506]
W. M. Koolen, M. K. Warmuth, and J. Kivinen. Hedging structured concepts.
In Proceedings of the 23rd Conference on Learning Theory, pages 93–105.
Omnipress, 2010. [373]
N. Korda, E. Kaufmann, and R. Munos. Thompson sampling for 1-dimensional
exponential family bandits. In Advances in Neural Information Processing
Systems, pages 1448–1456. Curran Associates, Inc., 2013. [120, 141, 470, 473]
J. Kujala and T. Elomaa. On following the perturbed leader in the bandit setting.
In Proceedings of the 16th International Conference on Algorithmic Learning
Theory, pages 371–385, 2005. [373]
J. Kujala and T. Elomaa. Following the perturbed leader to gamble at multi-armed
bandits. In Proceedings of the 18th International Conference on Algorithmic
Learning Theory, pages 166–180. Springer, 2007. [373]
S. R. Kulkarni and G. Lugosi. Finite-time lower bounds for the two-armed bandit
problem. IEEE Transactions on Automatic Control, 45(4):711–714, 2000. [210]
B. Kveton, Cs. Szepesvári, Z. Wen, and A. Ashkan. Cascading bandits: Learning to
rank in the cascade model. In Proceedings of the 32nd International Conference
on Machine Learning, pages 767–776. JMLR.org, 2015a. [399]
B. Kveton, Z. Wen, A. Ashkan, and Cs. Szepesvári. Tight regret bounds for
stochastic combinatorial semi-bandits. In Proceedings of the 18th International
Conference on Artificial Intelligence and Statistics, pages 535–543, San Diego,
CA, USA, 2015b. JMLR.org. [373]
B. Kveton, Z. Wen, Z. Ashkan, and Cs. Szepesvári. Combinatorial cascading
bandits. In Advances in Neural Information Processing Systems, pages 1450–
1458. Curran Associates Inc., 2015c. [400]
B. Kveton, Cs. Szepesvári, S. Vaswani, Z. Wen, T. Lattimore, and
M. Ghavamzadeh. Garbage in, reward out: Bootstrapping exploration in
multi-armed bandits. In Proceedings of the 36th International Conference on
Machine Learning, pages 3601–3610, Long Beach, California, USA, 09–15 Jun
2019. PMLR. [473]
 BIBLIOGRAPHY 568

P. Lagree, C. Vernade, and O. Cappé. Multiple-play bandits in the position-

based model. In Advances in Neural Information Processing Systems, pages
1597–1605. Curran Associates Inc., 2016. [400]
T. L. Lai. Adaptive treatment allocation and the multi-armed bandit problem.
The Annals of Statistics, pages 1091–1114, 1987. [110, 119, 128, 140, 455, 456]
T. L. Lai. Martingales in sequential analysis and time series, 1945–1985. Electronic
Journal for history of probability and statistics, 5(1), 2009. [261]
T. L. Lai and T. Graves. Asymptotically efficient adaptive choice of control laws
in controlled Markov chains. SIAM Journal on Control and Optimization, 35
(3):715–743, 1997. [536]
T. L. Lai and H. Robbins. Asymptotically efficient adaptive allocation rules.
Advances in applied mathematics, 6(1):4–22, 1985. [68, 110, 119, 140, 210, 265]
J. Langford and T. Zhang. The epoch-greedy algorithm for multi-armed bandits
with side information. In Advances in Neural Information Processing Systems,
pages 817–824. Curran Associates, Inc., 2008. [94, 235]
P. Laplace. Pierre-Simon Laplace Philosophical Essay on Probabilities: Translated
from the fifth French edition of 1825 With Notes by the Translator, volume 13.
Springer Science & Business Media, 2012. [42]
T. Lattimore. The Pareto regret frontier for bandits. In Advances in Neural
Information Processing Systems, pages 208–216. Curran Associates, Inc., 2015a.
[157, 294]
T. Lattimore. Optimally confident UCB: Improved regret for finite-armed bandits.
arXiv:1507.07880, 2015b. [128]
T. Lattimore. Regret analysis of the finite-horizon Gittins index strategy for
multi-armed bandits. In Proceedings of the 29th Annual Conference on Learning
Theory, pages 1214–1245, New York, NY, USA, 2016a. JMLR.org. [119, 455,
456]
T. Lattimore. Regret analysis of the anytime optimally confident ucb algorithm.
arXiv:1603.08661, 2016b. [128]
T. Lattimore. A scale free algorithm for stochastic bandits with bounded kurtosis.
In Advances in Neural Information Processing Systems, pages 1584–1593.
Curran Associates, Inc., 2017. [115, 209, 210]
T. Lattimore. Refining the confidence level for optimistic bandit strategies.
Journal of Machine Learning Research, 2018. [98, 128, 130, 210]
T. Lattimore and M. Hutter. PAC bounds for discounted MDPs. In Proceedings
of the 23th International Conference on Algorithmic Learning Theory, volume
7568 of Lecture Notes in Computer Science, pages 320–334. Springer Berlin /
Heidelberg, 2012. [537]
T. Lattimore and R. Munos. Bounded regret for finite-armed structured bandits.
In Advances in Neural Information Processing Systems, pages 550–558. Curran
Associates, Inc., 2014. [247]
T. Lattimore and Cs. Szepesvári. The end of optimism? An asymptotic analysis of
finite-armed linear bandits. In Proceedings of the 20th International Conference
 BIBLIOGRAPHY 569

on Artificial Intelligence and Statistics, pages 728–737, Fort Lauderdale, FL,

USA, 2017. JMLR.org. [246, 301]
T. Lattimore and Cs. Szepesvári. Cleaning up the neighbourhood: A full
classification for adversarial partial monitoring. In Proceedings of the 30th
International Conference on Algorithmic Learning Theory, 2019a. [505, 506,
509]
T. Lattimore and Cs. Szepesvári. Learning with good feature representations in
bandits and in RL with a generative model. arXiv:1911.07676, 2019b. [274,
294]
T. Lattimore and Cs. Szepesvári. An information-theoretic approach to minimax
regret in partial monitoring. In Proceedings of the 32nd Conference on Learning
Theory, pages 2111–2139, Phoenix, USA, 2019c. PMLR. [472, 473, 475, 506]
T. Lattimore and Cs. Szepesvári. Exploration by optimisation in partial
monitoring. arXiv:1907.05772, 2019d. [504, 506, 510]
T. Lattimore, K. Crammer, and Cs. Szepesvári. Linear multi-resource allocation
with semi-bandit feedback. In Advances in Neural Information Processing
Systems, pages 964–972. Curran Associates, Inc., 2015. [286]
T. Lattimore, B. Kveton, S. Li, and Cs. Szepesvári. Toprank: A practical algorithm
for online stochastic ranking. In Advances in Neural Information Processing
Systems, pages 3949–3958. Curran Associates, Inc., 2018. [265, 375, 399]
B. Laurent and P. Massart. Adaptive estimation of a quadratic functional by
model selection. Annals of Statistics, pages 1302–1338, 2000. [80]
A. Lazaric and R. Munos. Hybrid stochastic-adversarial on-line learning. In
Proceedings of the 22nd Conference on Learning Theory, 2009. [233]
T. Le, Cs. Szepesvári, and R. Zheng. Sequential learning for multi-channel
wireless network monitoring with channel switching costs. IEEE Transactions
on Signal Processing, 62(22):5919–5929, 2014. [17]
L Le Cam. Convergence of estimates under dimensionality restrictions. The
Annals of Statistics, 1(1):38–53, 1973. [202]
Y. T. Lee, A. Sidford, and S. S. Vempala. Efficient convex optimization with
membership oracles. In Proceedings of the 31st Conference On Learning Theory,
pages 1292–1294. JMLR.org, 06–09 Jul 2018. [372]
E. L. Lehmann and G. Casella. Theory of point estimation. Springer Science &
Business Media, 2006. [435]
H. Lei, A. Tewari, and S. A. Murphy. An actor-critic contextual bandit algorithm
for personalized mobile health interventions. arXiv:1706.09090, 2017. [17]
J. Leike, T. Lattimore, L. Orseau, and M. Hutter. Thompson sampling is
asymptotically optimal in general environments. In Proceedings of the 32nd
Conference on Uncertainty in Artificial Intelligence, pages 417–426. AUAI
Press, 2016. [473]
H. R. Lerche. Boundary crossing of Brownian motion: Its relation to the law of
the iterated logarithm and to sequential analysis. Springer, 1986. [130]
D. A. Levin and Y. Peres. Markov chains and mixing times, volume 107. American
Mathematical Soc., 2017. [53]
 BIBLIOGRAPHY 570

L. A. Levin. On the notion of a random sequence. Soviet Mathematics Doklady,

14(5):1413–1416, 1973. [146]
L. Li, K. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar. Hyperband:
A novel bandit-based approach to hyperparameter optimization. Journal of
Machine Learning Research, 18(185):1–52, 2018. [416]
S. Li, B. Wang, S. Zhang, and W. Chen. Contextual combinatorial cascading
bandits. In Proceedings of the 33rd International Conference on Machine
Learning, pages 1245–1253, 2016. [400]
S. Li, T. Lattimore, and Cs. Szepesvári. Online learning to rank with features. In
Proceedings of the 36th International Conference on Machine Learning, pages
3856–3865, Long Beach, California, USA, 09–15 Jun 2019a. PMLR. [398]
Y. Li, Y. Wang, and Y. Zhou. Nearly minimax-optimal regret for linearly
parameterized bandits. In Proceedings of the 32nd Conference on Learning
Theory, pages 2173–2174, Phoenix, USA, 2019b. JMLR.org. [245]
T. Liang, H. Narayanan, and A. Rakhlin. On zeroth-order stochastic convex
optimization via random walks. arXiv:1402.2667, 2014. [415]
T. Lin, B. Abrahao, R. Kleinberg, J. Lui, and W. Chen. Combinatorial partial
monitoring game with linear feedback and its applications. In Proceedings of
the 31st International Conference on Machine Learning, pages 901–909, Bejing,
China, 22–24 Jun 2014. PMLR. [507]
T. Lin, J. Li, and W. Chen. Stochastic online greedy learning with semi-bandit
feedbacks. In Advances in Neural Information Processing Systems, pages
352–360. Curran Associates, Inc., 2015. [373]
N. Littlestone and M. K. Warmuth. The weighted majority algorithm. Information
and Computation, 108(2):212–261, 1994. [146, 159]
L. Lovász and S. Vempala. The geometry of logconcave functions and sampling
algorithms. Random Structures & Algorithms, 30(3):307–358, 2007. [323]
H. Luo, C-Y. Wei, A. Agarwal, and J. Langford. Efficient contextual bandits
in non-stationary worlds. In Proceedings of the 31st Conference On Learning
Theory, pages 1739–1776. JMLR.org, 06–09 Jul 2018. [385]
D. MacKay. Information theory, inference and learning algorithms. Cambridge
University Press, 2003. [193]
S. Magureanu, R. Combes, and A. Proutière. Lipschitz bandits: Regret lower
bound and optimal algorithms. In Proceedings of the 27th Conference on
Learning Theory, pages 975–999, 2014. [249, 358]
O. Maillard. Robust risk-averse stochastic multi-armed bandits. In Proceedings
of the 24th International Conference on Algorithmic Learning Theory, pages
218–233. Springer, Berlin, Heidelberg, 2013. [68]
O. Maillard, R. Munos, and G. Stoltz. Finite-time analysis of multi-armed
bandits problems with Kullback-Leibler divergences. In Proceedings of the 24th
Conference on Learning Theory, 2011. [140]
S. Mannor and O. Shamir. From bandits to experts: On the value of side-
observations. In Advances in Neural Information Processing Systems, pages
684–692. Curran Associates, Inc., 2011. [360, 506]
 BIBLIOGRAPHY 571

S. Mannor and N. Shimkin. On-line learning with imperfect monitoring. In

Learning Theory and Kernel Machines, pages 552–566. Springer, 2003. [506]
S. Mannor and J. N. Tsitsiklis. The sample complexity of exploration in the
multi-armed bandit problem. Journal of Machine Learning Research, 5:623–648,
December 2004. [414]
S. Mannor, V. Perchet, and G. Stoltz. Set-valued approachability and online
learning with partial monitoring. The Journal of Machine Learning Research,
15(1):3247–3295, 2014. [507]
H. Markowitz. Portfolio selection. The Journal of Finance, 7(1):77–91, 1952. [67]
M. E. Maron and J. L. Kuhns. On relevance, probabilistic indexing and
information retrieval. Journal of the ACM, 7(3):216–244, 1960. [401]
P. Martin-Löf. The definition of random sequences. Information and Control, 9
(6):602–619, 1966. [146]
A. Maurer and M. Pontil. Empirical Bernstein bounds and sample variance
penalization. arXiv:0907.3740, 2009. [86, 114]
B. C. May, N. Korda, A. Lee, and D. S. Leslie. Optimistic Bayesian sampling in
contextual-bandit problems. The Journal of Machine Learning Research, 13
(1):2069–2106, 2012. [473]
C. McDiarmid. Concentration. In Probabilistic methods for algorithmic discrete
mathematics, pages 195–248. Springer, 1998. [80, 86, 263]
H. B. McMahan and A. Blum. Online geometric optimization in the bandit
setting against an adaptive adversary. In Proceedings of the 17th Conference
on Learning Theory, volume 3120, pages 109–123. Springer, 2004. [373]
H. B. McMahan and M. J. Streeter. Tighter bounds for multi-armed bandits
with expert advice. In Proceedings of the 22nd Conference on Learning Theory,
2009. [232]
P. Ménard and A. Garivier. A minimax and asymptotically optimal algorithm
for stochastic bandits. In Proceedings of the 28th International Conference on
Algorithmic Learning Theory, pages 223–237, Kyoto University, Kyoto, Japan,
15–17 Oct 2017. JMLR.org. [120, 128, 140]
S. P. Meyn and R. L. Tweedie. Markov chains and stochastic stability. Springer
Science & Business Media, 2012. [52, 53]
V. Mnih, Cs. Szepesvári, and J.-Y. Audibert. Empirical Bernstein stopping. In
Proceedings of the 25th International Conference on Machine Learning, pages
672–679, New York, NY, USA, 2008. ACM. [86, 114]
S. Mukherjee, KP. Naveen, N. Sudarsanam, and B. Ravindran. Efficient-
UCBV: An almost optimal algorithm using variance estimates. In 32nd AAAI
Conference on Artificial Intelligence, 2018. [128]
J. A. Nelder and R. W. M. Wedderburn. Generalized linear models. Journal of
the Royal Statistical Society. Series A (General), 135(3):370–384, 1972. [247]
A. S. Nemirovsky. Efficient methods for large-scale convex optimization problems.
Ekonomika i Matematicheskie Metody, 15, 1979. [343]
A. S. Nemirovsky and D. B. Yudin. Problem Complexity and Method Efficiency
in Optimization. Wiley, 1983. [343, 415]
 BIBLIOGRAPHY 572

G. Neu. Explore no more: Improved high-probability regret bounds for non-

stochastic bandits. In Advances in Neural Information Processing Systems,
pages 3168–3176. Curran Associates, Inc., 2015a. [170, 171, 233, 373]
G. Neu. First-order regret bounds for combinatorial semi-bandits. In Proceedings
of the 28th Conference on Learning Theory, pages 1360–1375, Paris, France,
2015b. JMLR.org. [170, 340]
G. Neu, A. György, Cs. Szepesvári, and A. Antos. Online Markov decision
processes under bandit feedback. IEEE Transactions on Automatic Control,
59(3):676–691, December 2014. [537]
J. Von Neumann and O. Morgenstern. Theory of Games and Economic Behavior.
Princeton University Press, Princeton, 1944. [67]
J. Niño-Mora. Computing a classic index for finite-horizon bandits. INFORMS
Journal on Computing, 23(2):254–267, 2011. [456]
B. O’Donoghue, E. Chu, N. Parikh, and S. Boyd. Conic optimization via operator
splitting and homogeneous self-dual embedding. Journal of Optimization
Theory and Applications, 169(3):1042–1068, 2016. [509]
B. O’Donoghue, E. Chu, N. Parikh, and S. Boyd. SCS: Splitting conic solver,
version 2.1.1. https://github.com/cvxgrp/scs, November 2017. [509]
J. Ok, A. Proutiere, and D. Tranos. Exploration in structured reinforcement
learning. In Advances in Neural Information Processing Systems, Red Hook,
NY, USA, 2018. Curran Associates Inc. [246, 301]
P. A. Ortega and D. A. Braun. A minimum relative entropy principle for learning
and acting. Journal of Artificial Intelligence Research, pages 475–511, 2010.
[473]
R. Ortner and D. Ryabko. Online regret bounds for undiscounted continuous
reinforcement learning. In Advances in Neural Information Processing Systems,
pages 1763–1771, USA, 2012. Curran Associates Inc. [536, 537]
R. Ortner, D. Ryabko, P. Auer, and R. Munos. Regret bounds for restless Markov
bandits. In Proceedings of the 23rd International Conference on Algorithmic
Learning Theory, pages 214–228, Berlin, Heidelberg, 2012. Springer Berlin
Heidelberg. [384]
I. Osband and B. Van Roy. Why is posterior sampling better than optimism for
reinforcement learning? In Proceedings of the 34th International Conference
on Machine Learning, pages 2701–2710, Sydney, Australia, 06–11 Aug 2017.
JMLR.org. [537]
I. Osband, D. Russo, and B. Van Roy. (more) efficient reinforcement learning via
posterior sampling. In Advances in Neural Information Processing Systems,
pages 3003–3011. Curran Associates, Inc., 2013. [473, 537]
E. Ostrovsky and L. Sirota. Exact value for subgaussian norm of centered
indicator random variable. arXiv:1405.6749, 2014. [84]
D. G. Pandelis and D. Teneketzis. On the optimality of the gittins index rule for
multi-armed bandits with multiple plays. Mathematical Methods of Operations
Research, 50(3):449–461, 1999. [454]
 BIBLIOGRAPHY 573

C. H. Papadimitriou and J. N. Tsitsiklis. The complexity of Markov decision

processes. Mathematics of Operations Research, 12(3):441–450, 1987. [532]
C. H. Papadimitriou and S. Vempala. On the approximability of the traveling
salesman problem. Combinatorica, 26(1):101–120, 2006. [373]
V. Perchet. Approachability of convex sets in games with partial monitoring.
Journal of Optimization Theory and Applications, 149(3):665–677, 2011. [507]
V. Perchet and P. Rigollet. The multi-armed bandit problem with covariates.
The Annals of Statistics, 41(2):693–721, 04 2013. [249]
G. Peskir and A. Shiryaev. Optimal stopping and free-boundary problems. Springer,
2006. [54, 455, 457]
A. Piccolboni and C. Schindelhauer. Discrete prediction games with arbitrary
feedback and loss. In Computational Learning Theory, pages 208–223. Springer,
2001. [506]
C. Pike-Burke, S. Agrawal, Cs. Szepesvári, and S. Grünewälder. Bandits
with delayed, aggregated anonymous feedback. In Proceedings of the 35th
International Conference on Machine Learning, volume 80, pages 4102–4110.
JMLR.org, 10–15 Jul 2018. [360]
J. Poland. FPL analysis for adaptive bandits. In O. B. Lupanov, O. M.
Kasim-Zade, A. V. Chaskin, and K. Steinhöfel, editors, Stochastic Algorithms:
Foundations and Applications, pages 58–69, Berlin, Heidelberg, 2005. Springer
Berlin Heidelberg. [373]
D. Pollard. A user’s guide to measure theoretic probability, volume 8. Cambridge
University Press, 2002. [41]
E. L. Presman and I. N. Sonin. Sequential control with incomplete information.
The Bayesian approach to multi-armed bandit problems. Academic Press, 1990.
[17, 456]
M. Puterman. Markov decision processes: discrete stochastic dynamic
programming, volume 414. Wiley, 2009. [535, 538]
C. Qin, D. Klabjan, and D. Russo. Improving the expected improvement algorithm.
In Advances in Neural Information Processing Systems, pages 5381–5391.
Curran Associates, Inc., 2017. [414]
F. Radlinski, R. Kleinberg, and T. Joachims. Learning diverse rankings with
multi-armed bandits. In Proceedings of the 25th International Conference on
Machine Learning, pages 784–791. ACM, 2008. [399, 400]
A. N. Rafferty, H. Ying, and J. J. Williams. Bandit assignment for educational
experiments: Benefits to students versus statistical power. In Artificial
Intelligence in Education, pages 286–290. Springer, 2018. [17]
A. Rakhlin and K. Sridharan. BISTRO: An efficient relaxation-based method
for contextual bandits. In Proceedings of the 33rd International Conference on
Machine Learning, pages 1977–1985, 2016. [233]
A. Rakhlin and K. Sridharan. On equivalence of martingale tail bounds and
deterministic regret inequalities. In Proceedings of the 30th Conference on
Learning Theory, pages 1704–1722, Amsterdam, Netherlands, 2017. JMLR.org.
[286]
 BIBLIOGRAPHY 574

A. Rakhlin, O. Shamir, and K. Sridharan. Making gradient descent optimal for

strongly convex stochastic optimization. In Proceedings of the 29th International
Conference on Machine Learning, 2012. [415]
L. M. Rios and N. V. Sahinidis. Derivative-free optimization: a review of
algorithms and comparison of software implementations. Journal of Global
Optimization, 56(3):1247–1293, Jul 2013. [415]
H. Robbins. Some aspects of the sequential design of experiments. Bulletin of
the American Mathematical Society, 58(5):527–535, 1952. [17, 68, 94]
H. Robbins and D. Siegmund. Boundary crossing probabilities for the wiener
process and sample sums. The Annals of Mathematical Statistics, pages 1410–
1429, 1970. [261]
H. Robbins and D. Siegmund. A class of stopping rules for testing parametric
hypotheses. In Proceedings of the Sixth Berkeley Symposium on Mathematical
Statistics and Probability, pages 37–41. University of California Press, 1972.
[265]
H. Robbins, D. Sigmund, and Y. Chow. Great expectations: the theory of optimal
stopping. Houghton-Nifflin, 7:631–640, 1971. [455]
S. Robertson. The probability ranking principle in IR. Journal of Documentation,
33(4):294–304, 1977. [400, 401]
R. T. Rockafellar. Convex analysis. Princeton university press, 2015. [313, 375]
R. T. Rockafellar and S. Uryasev. Optimization of conditional value-at-risk.
Journal of Risk, 2:21–42, 2000. [67]
C. A. Rogers. Packing and covering. Cambridge University Press, 1964. [261]
S. M. Ross. Introduction to Stochastic Dynamic Programming. Academic Press,
New York, 1983. [535]
P. Rusmevichientong and J. N. Tsitsiklis. Linearly parameterized bandits.
Mathematics of Operations Research, 35(2):395–411, 2010. [94, 246, 247, 294]
D. Russo. Simple Bayesian algorithms for best arm identification. In Proceedings
of the 29th Annual Conference on Learning Theory, pages 1417–1418, New
York, NY, USA, 2016. JMLR.org. [414]
D. Russo and B. Van Roy. Eluder dimension and the sample complexity of
optimistic exploration. In Advances in Neural Information Processing Systems,
pages 2256–2264. Curran Associates, Inc., 2013. [247]
D. Russo and B. Van Roy. Learning to optimize via information-directed sampling.
In Advances in Neural Information Processing Systems, pages 1583–1591.
Curran Associates, Inc., 2014a. [246, 472, 473]
D. Russo and B. Van Roy. Learning to optimize via posterior sampling.
Mathematics of Operations Research, 39(4):1221–1243, 2014b. [473]
D. Russo and B. Van Roy. An information-theoretic analysis of Thompson
sampling. Journal of Machine Learning Research, 17(1):2442–2471, 2016. ISSN
1532-4435. [373, 471, 473]
D. J. Russo, B. Van Roy, A. Kazerouni, I. Osband, and Z. Wen. A tutorial on
Thompson sampling. Foundations and Trends in Machine Learning, 11(1):
1–96, 2018. [473]
 BIBLIOGRAPHY 575

A. Rustichini. Minimizing regret: The general case. Games and Economic

Behavior, 29(1):224–243, 1999. [505, 506]
A. Salomon, J. Audibert, and I. Alaoui. Lower bounds and selectivity of weak-
consistent policies in stochastic multi-armed bandit problem. Journal of
Machine Learning Research, 14(Jan):187–207, 2013. [210]
P. Samuelson. A note on measurement of utility. The Review of Economic Studies,
4(2):pp. 155–161, 1937. [454]
A. Sani, A. Lazaric, and R. Munos. Risk-aversion in multi-armed bandits. In
Advances in Neural Information Processing Systems, pages 3275–3283. Curran
Associates, Inc., 2012. [68]
Y. Seldin and G. Lugosi. An improved parametrization and analysis of the
EXP3++ algorithm for stochastic and adversarial bandits. In Proceedings
of the 2017 Conference on Learning Theory, pages 1743–1759, Amsterdam,
Netherlands, 2017. JMLR.org. [157]
Y. Seldin and A. Slivkins. One practical algorithm for both stochastic and
adversarial bandits. In Proceedings of the 31st International Conference on
Machine Learning, pages 1287–1295, Bejing, China, 2014. JMLR.org. [157]
S. Shalev-Shwartz. Online learning: Theory, algorithms, and applications. PhD
thesis, The Hebrew University of Jerusalem, 2007. [343]
S. Shalev-Shwartz. Online learning and online convex optimization. Foundations
and Trends in Machine Learning, 4(2):107–194, 2012. [342, 343]
S. Shalev-Shwartz and S. Ben-David. Understanding Machine Learning: From
Theory to Algorithms. Cambridge University Press, 2014. [232, 233, 236]
S. Shalev-Shwartz and Y. Singer. A primal-dual perspective of online learning
algorithms. Machine Learning, 69(2-3):115–142, 2007. [343]
O. Shamir. On the complexity of bandit and derivative-free stochastic convex
optimization. In Proceedings of the 26th Conference on Learning Theory, pages
3–24. JMLR.org, 2013. [359, 415]
O. Shamir. On the complexity of bandit linear optimization. In Proceedings of
the 28th Conference on Learning Theory, pages 1523–1551, Paris, France, 2015.
JMLR.org. [294, 355]
T. Sharot. The optimism bias. Current Biology, 21(23):R941–R945, 2011a. [109,
110]
T. Sharot. The optimism bias: A tour of the irrationally positive brain.
Pantheon/Random House, 2011b. [110]
D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche,
J. Schrittwieser, I. Antonoglou, V. Panneershelvam, and M. Lanctot. Mastering
the game of go with deep neural networks and tree search. Nature, 529(7587):
484–489, 2016. [17]
S. D. Silvey and B. Sibson. Discussion of Dr. Wynn’s and of Dr. Laycock’s papers.
Journal of Royal Statistical Society (B), 34:174–175, 1972. [270]
M. Sion. On general minimax theorems. Pacific Journal of mathematics, 8(1):
171–176, 1958. [342]
 BIBLIOGRAPHY 576

A. Slivkins. Contextual bandits with similarity information. Journal of Machine

Learning Research, 15(1):2533–2568, 2014. [358]
A. Slivkins. Introduction to multi-armed bandits. Foundations and Trends in
Machine Learning, 12(1-2):1–286, 2019. ISSN 1935-8237. [17, 358]
A. Slivkins and E. Upfal. Adapting to a changing environment: the Brownian
restless bandits. In Proceedings of the 21st Conference on Learning Theory,
pages 343–354, 2008. [385]
M. Soare, A. Lazaric, and R. Munos. Best-arm identification in linear bandits. In
Advances in Neural Information Processing Systems, pages 828–836. Curran
Associates, Inc., 2014. [301, 414]
I. M. Sonin. A generalized Gittins index for a Markov chain and its recursive
calculation. Statistics and Probability Letters, 78(12):1526–1533, 2008. [456]
N. Srebro, K. Sridharan, and A. Tewari. On the universality of online mirror
descent. In Advances in neural information processing systems, pages 2645–2653,
2011. [340]
K. Sridharan and A. Tewari. Convex games in banach spaces. In Proceedings of
the 23rd Conference on Learning Theory, pages 1–13. Omnipress, 2010. [343]
N. Srinivas, A. Krause, S. Kakade, and M. Seeger. Gaussian process optimization
in the bandit setting: No regret and experimental design. In Proceedings of the
27th International Conference on Machine Learning, page 1015–1022, Madison,
WI, USA, 2010. Omnipress. [247]
G. Stoltz. Incomplete information and internal regret in prediction of individual
sequences. PhD thesis, Université Paris Sud-Paris XI, 2005. [159]
H. Strasser. Mathematical theory of statistics: statistical experiments and
asymptotic decision theory, volume 7. Walter de Gruyter, 2011. [434]
R. E. Strauch. Negative dynamic programming. The Annals of Mathematical
Statistics, 37(4):871–890, 08 1966. [538]
M. J. Streeter and S. F. Smith. A simple distribution-free approach to the
max k-armed bandit problem. In International Conference on Principles and
Practice of Constraint Programming, pages 560–574. Springer, 2006a. [416]
M. J Streeter and S. F. Smith. An asymptotically optimal algorithm for the
max k-armed bandit problem. In Proceedings of the National Conference on
Artificial Intelligence, pages 135–142, 2006b. [416]
A. Strehl and M. Littman. A theoretical analysis of model-based interval
estimation. In Proceedings of the 22nd International Conference on Machine
learning, pages 856–863, New York, NY, USA, 2005. ACM. [537]
A. Strehl and M. Littman. An analysis of model-based interval estimation for
Markov decision processes. Journal of Computer and System Sciences, 74(8):
1309–1331, 2008. [537, 544]
A. Strehl, L. Li, E. Wiewiora, J. Langford, and M. Littman. PAC model-free
reinforcement learning. In Proceedings of the 23rd international conference on
Machine learning, pages 881–888, New York, NY, USA, 2006. ACM. [537]
 BIBLIOGRAPHY 577

M. J. A. Strens. A Bayesian framework for reinforcement learning. In Proceedings

of the 17th International Conference on Machine Learning, pages 943–950, San
Francisco, CA, USA, 2000. Morgan Kaufmann Publishers Inc. [536]
Y. Sui, A. Gotovos, J. Burdick, and A. Krause. Safe exploration for optimization
with gaussian processes. In Proceedings of the 32nd International Conference
on Machine Learning, pages 997–1005, Lille, France, 07–09 Jul 2015. JMLR.org.
[360]
Q. Sun, W. Zhou, and J. Fan. Adaptive huber regression: Optimality and phase
transition. arXiv:1706.06991, 2017. [115]
R. Sutton and A. Barto. Reinforcement Learning: An Introduction. MIT Press,
1998. [95, 454]
R. Sutton and A. Barto. Reinforcement Learning: An Introduction. MIT Press,
second edition, 2018. [535]
J.M. Swart. Large deviation theory, January 2017. URL http://staff.utia.
cas.cz/swart/lecture_notes/LDP8.pdf. [83]
V. Syrgkanis, A. Krishnamurthy, and R. Schapire. Efficient algorithms for
adversarial contextual learning. In Proceedings of the 33rd International
Conference on Machine Learning, pages 2159–2168, New York, NY, USA,
2016. JMLR.org. [233]
Cs. Szepesvári. Algorithms for Reinforcement Learning. Synthesis Lectures on
Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers,
2010. [536]
I. Szita and A. Lőrincz. Optimistic initialization and greediness lead to polynomial
time learning in factored MDPs. In Proceedings of the 26th International
Conference on Machine Learning, pages 1001–1008, New York, USA, 2009.
ACM. [537]
I. Szita and Cs. Szepesvári. Model-based reinforcement learning with nearly
tight exploration complexity bounds. In Proceedings of the 27th International
Conference on Machine Learning, pages 1031–1038, USA, 2010. Omnipress.
[537]
E. Takimoto and M. K. Warmuth. Path kernels and multiplicative updates.
Journal of Machine Learning Research, 4:773–818, 2003. [374]
M. Talagrand. The missing factor in Hoeffding’s inequalities. Annales de l’IHP
Probabilités et Statistiques, 31(4):689–702, 1995. [79]
G. Taraldsen. Optimal learning from the Doob-Dynkin lemma. arXiv:1801.00974,
2018. [38]
J. Teevan, S. T. Dumais, and E. Horvitz. Characterizing the value of personalizing
search. In Proceedings of the 30th Annual International ACM SIGIR Conference
on Research and Development in Information Retrieval, pages 757–758, New
York, NY, USA, 2007. ACM. [400]
A. Tewari and P. L. Bartlett. Optimistic linear programming gives logarithmic
regret for irreducible MDPs. In Advances in Neural Information Processing
Systems, pages 1505–1512. Curran Associates, Inc., 2008. [536]
 BIBLIOGRAPHY 578

A. Tewari and S. A. Murphy. From ads to interventions: Contextual bandits in

mobile health. In Mobile Health, pages 495–517. Springer, 2017. [232]
G. Theocharous, Z. Wen, Y. Abbasi-Yadkori, and N. Vlassis. Posterior sampling
for large scale reinforcement learning. arXiv:1711.07979, 2017. [537]
W. Thompson. On the likelihood that one unknown probability exceeds another
in view of the evidence of two samples. Biometrika, 25(3/4):285–294, 1933. [16,
68, 95, 459, 470, 473]
W. R. Thompson. On the theory of apportionment. American Journal of
Mathematics, 57(2):450–456, 1935. [535]
M. J. Todd. Minimum-volume ellipsoids: Theory and algorithms. SIAM, 2016.
[270, 271]
J. R. R. Tolkien. The Hobbit. Ballantine Books, 1937. [459]
L. Tran-Thanh, A. Chapman, E. Munoz de Cote, A. Rogers, and N. R. Jennings.
Epsilon–first policies for budget–limited multi-armed bandits. In Proceedings of
the 24th AAAI Conference on Artificial Intelligence, AAAI, pages 1211–1216,
2010. [360]
L. Tran-Thanh, A. Chapman, A. Rogers, and N. R. Jennings. Knapsack based
optimal policies for budget-limited multi-armed bandits. In Proceedings of the
26th AAAI Conference on Artificial Intelligence, AAAI’12, pages 1134–1140.
AAAI Press, 2012. [359]
J. A. Tropp. An introduction to matrix concentration inequalities. Foundations
and Trends® in Machine Learning, 8(1-2):1–230, 2015. [80]
J. N. Tsitsiklis. A short proof of the Gittins index theorem. The Annals of
Applied Probability, pages 194–199, 1994. [456]
A. B. Tsybakov. Introduction to nonparametric estimation. Springer Science &
Business Media, 2008. [193]
E. Uchibe and K. Doya. Competitive-cooperative-concurrent reinforcement
learning with importance sampling. In Proceedings of the International
Conference on Simulation of Adaptive Behavior: From Animals and Animats,
pages 287–296, 2004. [171]
M. Valko. Bandits on graphs and structures, 2016. [251, 360]
M. Valko, A. Carpentier, and R. Munos. Stochastic simultaneous optimistic
optimization. In Proceedings of the 30th International Conference on Machine
Learning, pages 19–27, Atlanta, GA, USA, 2013a. JMLR.org. [414, 415]
M. Valko, N. Korda, R. Munos, I. Flaounas, and N. Cristianini. Finite-time
analysis of kernelised contextual bandits. In Proceedings of the 29th Conference
on Uncertainty in Artificial Intelligence, pages 654–663, Arlington, VA, USA,
2013b. AUAI Press. [247]
M. Valko, R. Munos, B. Kveton, and T. Kocák. Spectral bandits for smooth graph
functions. In Proceedings of the 31st International Conference on Machine
Learning, pages 46–54, Bejing, China, 2014. JMLR.org. [247, 251, 274]
S. van de Geer. Empirical Processes in M-estimation, volume 6. Cambridge
University Press, 2000. [80, 261, 341]
 BIBLIOGRAPHY 579

D. van der Hoeven, T. van Erven, and W. Kotlowski. The many faces of
exponential weights in online learning. In Proceedings of the 31st Conference
on Learning Theory, pages 2067–2092, 2018. [323]
A. W. van der Vaart and J. A. Wellner. Weak Convergence and Empirical
Processes. Springer, New York, 1996. [341]
H. P. Vanchinathan, G. Bartók, and A. Krause. Efficient partial monitoring
with prior information. In Advances in Neural Information Processing Systems,
pages 1691–1699. Curran Associates, Inc., 2014. [506]
V. Vapnik. Statistical learning theory. 1998, volume 3. Wiley, New York, 1998.
[236]
P. Varaiya, J. Walrand, and C. Buyukkoc. Extensions of the multiarmed bandit
problem: The discounted case. IEEE Transactions on Automatic Control, 30
(5):426–439, 1985. [456]
C. Vernade, O. Cappé, and V. Perchet. Stochastic bandit models for delayed
conversions. In Proceedings of the 33rd Conference on Uncertainty in Artificial
Intelligence. AUAI Press, 2017. [360]
C. Vernade, A. Carpentier, G. Zappella, B. Ermis, and M. Brueckner. Contextual
bandits under delayed feedback. arXiv:1807.02089, 2018. [360]
S. Villar, J. Bowden, and J. Wason. Multi-armed bandit models for the optimal
design of clinical trials: benefits and challenges. Statistical science: a review
journal of the Institute of Mathematical Statistics, 30(2):199–215, 2015. [17]
W. Vogel. An asymptotic minimax theorem for the two armed bandit problem.
The Annals of Mathematical Statistics, 31(2):444–451, 1960. [202]
J. von Neumann. Zur theorie der gesellschaftsspiele. Mathematische annalen, 100
(1):295–320, 1928. [342]
V. G. Vovk. Aggregating strategies. Proceedings of Computational Learning
Theory, 1990. [146, 159]
S. Wang and W. Chen. Thompson sampling for combinatorial semi-bandits. In
Proceedings of the 35th International Conference on Machine Learning, pages
5114–5122, Stockholmsmässan, Stockholm Sweden, 10–15 Jul 2018. JMLR.org.
[373, 473]
Y. Wang, J-Y. Audibert, and R. Munos. Algorithms for infinitely many-armed
bandits. In Advances in Neural Information Processing Systems, pages 1729–
1736, 2009. [358]
M. K. Warmuth and A. Jagota. Continuous and discrete-time nonlinear gradient
descent: Relative loss bounds and convergence. In Electronic Proceedings of
the 5th International Symposium on Artificial Intelligence and Mathematics,
1997. [342]
P. L Wawrzynski and A. Pacut. Truncated importance sampling for reinforcement
learning with experience replay. In Proceedings of the International
Multiconference on Computer Science and Information Technology, pages 305–
315, 2007. [171]
R. Weber. On the Gittins index for multiarmed bandits. The Annals of Applied
Probability, 2(4):1024–1033, 1992. [456]
 BIBLIOGRAPHY 580

R. Weber and G. Weiss. On an index policy for restless bandits. Journal of

Applied Probability, 27(3):637–648, 1990. [456]
C-Y. Wei and H. Luo. More adaptive algorithms for adversarial bandits. In
Proceedings of the 31st Conference On Learning Theory, pages 1263–1291.
JMLR.org, 06–09 Jul 2018. [340, 343, 347]
M. J. Weinberger and E. Ordentlich. On delayed prediction of individual sequences.
In Information Theory, 2002. Proceedings. 2002 IEEE International Symposium
on, page 148. IEEE, 2002. [360]
Z. Wen, B. Kveton, and A. Ashkan. Efficient learning in large-scale combinatorial
semi-bandits. In Proceedings of the 32nd International Conference on Machine
Learning, volume 37, pages 1113–1122, Lille, France, 2015. JMLR.org. [373]
P. Whittle. Multi-armed bandits and the Gittins index. Journal of the Royal
Statistical Society (B), pages 143–149, 1980. [456]
P. Whittle. Restless bandits: Activity allocation in a changing world. Journal of
applied probability, 25(A):287–298, 1988. [384, 456]
D. Williams. Probability with martingales. Cambridge University Press, 1991.
[41]
H. Wu and X. Liu. Double Thompson sampling for dueling bandits. In Advances
in Neural Information Processing Systems, pages 649–657. Curran Associates,
Inc., 2016. [359]
Y. Wu, A. György, and Cs. Szepesvári. Online learning with gaussian payoffs
and side observations. In Advances in Neural Information Processing Systems,
pages 1360–1368. Curran Associates Inc., 2015. [506]
Y. Wu, R. Shariff, T. Lattimore, and Cs. Szepesvári. Conservative bandits. In
Proceedings of the 33rd International Conference on Machine Learning, pages
1254–1262, New York, NY, USA, 20–22 Jun 2016. JMLR.org. [360]
H. P. Wynn. The sequential generation of D-optimum experimental designs. The
Annals of Mathematical Statistics, pages 1655–1664, 1970. [270]
Y. Xia, H. Li, T. Qin, N. Yu, and T.-Y. Liu. Thompson sampling for budgeted
multi-armed bandits. In Proceedings of the 24th International Conference on
Artificial Intelligence, IJCAI, pages 3960–3966. AAAI Press, 2015. [359]
Y. Yao. Some results on the Gittins index for a normal reward process. In Time
Series and Related Topics, pages 284–294. Institute of Mathematical Statistics,
2006. [456]
B. Yu. Assouad, Fano, and Le Cam. In D. Pollard, E. Torgersen, and G. L. Yang,
editors, Festschrift for Lucien Le Cam: Research Papers in Probability and
Statistics, pages 423–435. Springer, 1997. [202, 203]
Y. Yue and T. Joachims. Interactively optimizing information retrieval systems as
a dueling bandits problem. In Proceedings of the 26th International Conference
on Machine Learning, pages 1201–1208. ACM, 2009. [359]
Y. Yue and T. Joachims. Beat the mean bandit. In Proceedings of the 28th
International Conference on Machine Learning, pages 241–248, New York, NY,
USA, June 2011. ACM. [359]
 BIBLIOGRAPHY 581

Y. Yue, J. Broder, R. Kleinberg, and T. Joachims. The k-armed dueling bandits

problem. In Proceedings of the 22nd Conference on Learning Theory, 2009.
[359]
J. Zimmert and T. Lattimore. Connections between mirror descent, thompson
sampling and the information ratio. In Advances in Neural Information
Processing Systems, pages 11973–11982. Curran Associates, Inc., 2019. [472]
J. Zimmert and Y. Seldin. An optimal algorithm for stochastic and adversarial
bandits. In AISTATS, pages 467–475, 2019. [157, 347, 359]
J. Zimmert, H. Luo, and C-Y. Wei. Beating stochastic and adversarial semi-
bandits optimally and simultaneously. In Proceedings of the 36th International
Conference on Machine Learning, pages 7683–7692, Long Beach, California,
USA, 09–15 Jun 2019. JMLR.org. [347]
M. Zinkevich. Online convex programming and generalized infinitesimal gradient
ascent. In Proceedings of the 20th International Conference on Machine
Learning, pages 928–935. AAAI Press, 2003. [342]
M. Zoghi, S. Whiteson, R. Munos, and M. Rijke. Relative upper confidence
bound for the k-armed dueling bandit problem. In Proceedings of the 31st
International Conference on Machine Learning, pages 10–18, Bejing, China,
2014. JMLR.org. [359]
M. Zoghi, Z. Karnin, S. Whiteson, and M. Rijke. Copeland dueling bandits. In
Advances in Neural Information Processing Systems, pages 307–315. Curran
Associates, Inc., 2015. [359]
M. Zoghi, T. Tunys, M. Ghavamzadeh, B. Kveton, Cs. Szepesvári, and Z. Wen.
Online learning to rank in stochastic click models. In Proceedings of the 34th
International Conference on Machine Learning, JMLR.org, pages 4199–4208,
2017. [399]
S. Zong, H. Ni, K. Sung, R. N. Ke, Z. Wen, and B. Kveton. Cascading bandits for
large-scale recommendation problems. In Proceedings of the 32nd Conference
on Uncertainty in Artificial Intelligence, 2016. [398, 399]
Index

σ-algebra, 21 Bayesian regret, 61, 431

restriction of, 42 Bayesian upper confidence bound algo-
a.s., 35 rithm, 471
Abel sum, 533 Bellman optimality equation, 516
absolutely continuous, 39 Bennett’s inequality, 86
action space, 511 Bernoulli
adapted, 27 bandit, 58, 70, 71, 120, 132–142, 158,
AdaUCB, 128, 464 162, 174, 202, 203, 426, 429, 430,
admissible, 420 435, 438, 439, 444, 449, 455, 457,
affine hull, 482 459, 473
almost surely, 35 distribution, 58, 75, 81, 84, 188
anytime, 94 random variable, 34, 54, 368
Exp3, 345 Bernstein’s inequality, 85, 86, 536
Exp3-IX, 170 empirical, 110, 113
Exp4, 230 beta distribution, 426
MOSS, 128 bias-variance trade-off, 223
UCB, 116 bidual, 306, 375
arithmetic coding, 186 bits, 187
Assouad’s method, 203 Borel σ-algebra, 22
asymptotic optimality Borel function, see Borel measurable
k-armed lower bounds, 206–207 function
k-armed upper bounds, 116, 128, 136 Borel measurable function, 23
best arm identification, 404–410 Borel space, 46
linear bandits, 295–299 boundary, 313
partial monitoring, 506 Bregman divergence, 307–309, 327,
ranking, 399 436
Thompson sampling, 470 Bretagnolle-Huber inequality, 189, 200,
Bachelier–Lévy formula, 130 208, 216, 288, 406
bandits with expert advice, 227 Brownian motion, 129, 385
Bayes rule, 28 canonical bandit model, 197
Bayesian bandit environment, 429 k-armed stochastic, 63
Bayesian optimal policy, 431 Bayesian, 437
1-armed bandit, 443 contextual, 70
discounted bandit, 449 infinite-armed stochastic, 66
 INDEX 583

Carathéodory’s extension theorem, 25 derivative-free stochastic optimisation,

cardinal optimisation, 416 414
cascade model, 389 descriptive theory, 67
categorical distribution, 87, 487, 526 deviation matrix, 539
Catoni’s estimator, 115 diameter
cell decomposition, 482 of convex set, 333
Cesàro sum, 516, 533 of MDP, 514
chain rule of probability measures, 53 differential value function, 516
change of measure, 39 discount factor, 383
Chernoff bound, 133, 398 discounting, 383, 447, 454
chi-squared distance, 192, 203 disintegration theorem, 53, 434, 435
chi-squared distribution, 80, 241 distribution, 21
click model, 388 Dobrushin’s theorem, 191, 194
closed function, 375 domain, 305
closed set, 313 dominated action, 482
complement, 21 dominating measure, 40
concave, 307 doubling trick, 94, 96, 158, 286, 340
conditional expectation, 33 dual norm, 334
conditional independence, 37 dynamic programming, 454
conditional probability, 28 easy partial monitoring game, 482
conjugate pair, 426 effective dimension, 250
conjugate prior, 425, 426 elimination algorithm, 94, 95, 97, 122,
consistent policy, 206, 295 210, 272, 277
contextual bandit, 70, 223, 382 elliptical potential lemma, 242
adversarial linear, 353 empirical risk minimisation, 232
stochastic, 231, 234, 237–239 entropy, 185, 186, 193
stochastic linear, 239–244, 353 events, 19
contextual partial monitoring, 507 exchangeable, 236
controlled Markov environment, 537 Exp3, 151, 224, 318, 320, 326, 344, 363,
convex hull, 305 377, 379, 481
convex optimisation, 326 Exp3-IX, 165, 214, 363, 385
core set, 267 Exp3.P, 172, 233, 385
counting measure, 40 Exp3.S, 385
covering, 254, 262 Exp4, 228, 248, 378
Cramér transform, 82, 83 expectation, 30
Cramér–Chernoff method, 76–78, 167, explore-then-commit, 90, 108, 231, 247
256, 257 exponential family, 120, 140, 210, 211,
cumulant generating function, 76, 141 407, 427, 427, 470
cumulative distribution function, 33 exponential weighting, 151, 366
D-optimal design, 267 algorithm, 157
data processing inequality, 194, 489 continuous, 320–322
degenerate action, 482 extended real line, 305
density, 39 feasible, 518
 INDEX 584

feature map, 238 hopeless partial monitoring game, 482

feature space, 239 Huffman coding, 186
feature vector, 239 hypercube, 277, 288, 322, 339
feedback matrix, 479 hypothesis space, 423
Fenchel dual, 83, 256, 306 hypothesis testing, 195
filtered probability space, 28 image, 502
filtration, 27 implicitly normalised forecaster, 347
finite additivity, 21 importance-weighted estimator, 149,
first-order bound, 170, 340, 346 150, 170, 231, 337
first-order optimality condition, 311 independent events, 29
Fisher information, 201 index, 446
fixed design, 253 index policy, 446
fixed share, 384 indicator function, 23
follow the leader, 327, 341 information-directed sampling, 246,
follow-the-perturbed-leader, 162, 231, 472, 475, 507
366, 462 instance-dependent bound, 232
follow-the-regularised-leader, 327, 354, integrable, 31
355 interior, 313
changing potentials, 345 Ionescu–Tulcea theorem, 48, 66, 430,
Frank–Wolfe algorithm, 269, 271 513
Fubini’s theorem, 40 isomorphic measurable spaces, 46
full information, 157, 384 Jensen’s inequality, 307
fundamental matrix, 539 John’s ellipsoid, 323
G-optimal design, 267 Kearns-Saul inequality, 84
gain, 515 kernel, 502
game theory, 182 kernel trick, 244
Gaussian tail lower bound, 474 Kiefer–Wolfowitz, 267, 274, 320, 323,
generalised linear bandit, 249 336, 354
generalised linear model, 246 Kraft’s inequality, 190
Gittins index, 384, 446 Kullback–Leibler divergence, 185
globally observable, 485 Laplace’s method, 257
gradient descent, 328 large deviation theory, 82
graph Laplacian, 251 law, 22
Hahn decomposition, 31 law of the iterated logarithm, 111, 123,
hard partial monitoring game, 482 264
Hardy–Littlewood, 450, 457, 458 law of total expectations, 36
heavy tailed, 76 Le Cam’s inequality, 190
Hedge, 156 Le Cam’s method, 201, 202
Hellinger distance, 192 learning rate, 151
Hoeffding’s inequality, 79, 135, 398 adaptive, 340, 346
Hoeffding’s lemma, 79, 83, 134, 157, time-varying, 230, 328, 345, 346
263 least-squares, 253
Hoeffding–Azuma, 263, 528 Lebesgue integral, 30
 INDEX 585

Lebesgue measure, 32 Neyman-Pearson lemma, 191, 195

Legendre function, 309–311, 327, 366 non-anticipating sequence, 345
light tailed, 76 non-oblivious, 157, 340
likelihood ratio, 265 non-parametric, 58
linear subspace, 502 non-singular exponential family, 427
link function, 246 non-stationary bandit, 68
Lipschitz bandit, 248 nonstationary, 158
locally observable, 485 null set, 35
log partition function, 427 oblivious, 334, 340
log-concave, 321 oblivious adversary, 157
logistic function, 427 one-armed bandit, 10, 71, 120
loss matrix, 479 Bayesian, 442–446
margin, 249 online gradient descent, 328
Markov chain, 48–49, 513 online learning, 17, 282, 326
Markov kernel, 48 online linear optimisation, 326
Markov policy, 513 online-to-confidence set conversion, 281
Markov process, 52 open set, 313
Markov property, 532 operator, 41
Markov reward process, 446 optimal experimental design, 267, 412
martingale, 49 optimal value function, 516
maximal end component, 546 optimisation oracle, 231, 371
maximal inequality, 51, 123, 259 optimism bias, 109
measurable map, 22 optional stopping theorem, 50
measurable set, 21 ordinal optimisation, 416
measurable space, 21 orthogonal complement, 502
measure, 21 outcome space, 19
median-of-means, 114 packing, 262
memoryless deterministic policy, 513 parameter noise, 351
memoryless policy, 513 parametric, 58
metric entropy, 262 Pareto optimal, 183, 420
minimax, 122 Pareto optimal action, 482
minimax optimal, 179 partial monitoring, 13
mirror descent, 157, 327, 366, 379 partially observable Markov decision
misspecified linear bandit, 274, 292, process, 532
354, 355 peeling device, 123
model, 423 permutation, 387
MOSS, 120, 122 Pinsker’s inequality, 133, 140, 192, 193,
multi-class classification with bandit 333
feedback, 233 point-locally observable, 504
multi-task bandit, 291, 363, 375 Poisson distribution, 81
nats, 187 policy, 64
negative correlation, 168 policy iteration, 534
neighbouring actions, 483 policy schema, 433
 INDEX 586

position-based model, 389 regular exponential family, 427

posterior, 422 regular version, 52, 195
potential function, 327 regularised risk minimisation, 343
predictable, 27 regulariser, 327
predictable variation, 85, 86 reinforcement learning, 13, 95, 454, 522
prediction with expert advice, 157 relative entropy, 187–190, 309, 436
preimage, 20 restless bandit, 384, 456
prescriptive theory, 67 retirement policy, 71
prior, 420, 423 reward-stack model, 65, 450
prior variance, 426 ridge regression, 253
probability distribution, 21 right stochastic matrix, 511
probability kernel, 48, 423 semi-bandit, 361, 364–371, 399, 401
probability measure, 21 semibandit, 473
probability space, 21 separation oracle, 321, 520
product σ-algebra, 25 sequential halving, 411
product kernel, 48 Sherman-Morrison formula, 252
product measure, 40, 65
signal variance, 426
projective, 47
signal-to-noise ratio, 352
push-forward, 22
signed measure, 21
quadratic variation, 170 similarity function, 226
Rademacher variable, 82 simple function, 31
Radon–Nikodym derivative, 39 Sion’s minimax theorem, 342, 348, 433,
random 472
element, 22 sliding window, 383
variable, 22 smoothness, 239
vector, 22 source coding theorem, 186
random table model, 65, 149 span, 516
random variable, 19 spectral bandit, 250
ranked bandit model, 398 state space, 511
ranking and selection, 415 static experts, 231
rate function, 83 stationary transition matrix, 516, 539
reactive adversary, 157 stochastic optimisation, 412
reduction, 350, 403 stochastic process, 47
regret, 10 stopping rule, 50
adversarial, 147 stopping time, 50
non-stationary, 378 strictly convex, 306
policy, 157 strongly connected component, 546
pseudo, 68 sub-σ-algebra, 21
pseudo, random, 215 submartingale, 49, 123
random, 68 suboptimality gap, 62
stochastic, 60 sufficient statistic, 434, 443, 449, 532
tracking, 378 of exponential family, 427
regret decomposition lemma, 62 supermartingale, 49, 261
 INDEX 587

supervised learning, 226

support, 41
support function, 367, 374
supporting hyperplane, 311
theorem, 314, 437
tail probability, 33
Thompson sampling, 68, 120, 459
for reinforcement learning, 536
total variation distance, 192, 192
tower rule, 36
track-and-stop algorithm, 409
transductive learning, 233
transition matrix, 515
trivial event, 43
trivial partial monitoring game, 482
UCB-V, 114
uniform exploration algorithm, 402
union bound, 77
unit ball, 4, 247, 262, 289, 333, 336, 341,
344, 349, 352, 354
unit sphere, 286, 290
universal constant, 4
unnormalised negentropy, 309, 310,
315, 329, 344, 364, 379, 469
unstructured bandits, 206
value function, 441
vanishing discount approach, 533
Varaiya’s algorithm, 455
VC dimension, 235
von Neumann-Morgenstern theorem, 67
Wald–Bellman equation, 441
weak neighbour, 504
weak* topology, 41
worst-case regret, 179
zeroth-order stochastic optimisation,
414