Notes PDF

Notes on Randomized Algorithms
CPSC 469/569: Fall 2016
James Aspnes
2017-04-03 15:19
Contents
Table of contents i
List of figures xii
List of tables xiii
List of algorithms xiv
Preface xv
Syllabus xvi
Lecture schedule xix
1 Randomized algorithms 1
1.1 A trivial example . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Verifying polynomial identities . . . . . . . . . . . . . . . . . 3
1.3 Randomized QuickSort . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 Brute force method: solve the recurrence . . . . . . . 5
1.3.2 Clever method: use linearity of expectation . . . . . . 6
1.4 Classifying randomized algorithms by their goals . . . . . . . 8
1.4.1 Las Vegas vs Monte Carlo . . . . . . . . . . . . . . . . 8
1.4.2 Randomized complexity classes . . . . . . . . . . . . . 9
1.5 Classifying randomized algorithms by their methods . . . . . 10
2 Probability theory 12
2.1 Probability spaces and events . . . . . . . . . . . . . . . . . . 13
2.1.1 General probability spaces . . . . . . . . . . . . . . . . 13
2.2 Boolean combinations of events . . . . . . . . . . . . . . . . . 15
2.3 Conditional probability . . . . . . . . . . . . . . . . . . . . . 18
i
CONTENTS ii
2.3.1 Application to algorithm analysis . . . . . . . . . . . . 18

2.3.1.1 Racing coin-flips . . . . . . . . . . . . . . . . 19
2.3.1.2 Kargers min-cut algorithm . . . . . . . . . . 21
2.3.2 Conditional probability and independence . . . . . . . 22
3 Random variables 24
3.1 Operations on random variables . . . . . . . . . . . . . . . . . 25
3.2 Random variables and events . . . . . . . . . . . . . . . . . . 25
3.3 Measurability . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.1 Linearity of expectation . . . . . . . . . . . . . . . . . 28
3.4.1.1 Linearity of expectation for infinite sequences 29
3.4.2 Expectation and inequalities . . . . . . . . . . . . . . 30
3.4.3 Expectation of a product . . . . . . . . . . . . . . . . 30
3.4.3.1 Walds equation (simple version) . . . . . . . 31
3.5 Conditional expectation . . . . . . . . . . . . . . . . . . . . . 33
3.5.1 Expectation conditioned on an event . . . . . . . . . . 33
3.5.2 Expectation conditioned on a random variable . . . . 34
3.5.2.1 Calculating conditional expectations . . . . . 35
3.5.2.2 The law of iterated expectation . . . . . . . . 37
3.5.2.3 Conditional expectation as orthogonal pro-
jection (optional) . . . . . . . . . . . . . . . 37
3.5.3 Expectation conditioned on a -algebra . . . . . . . . 39
3.5.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.6 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.6.1 Geometric random variables . . . . . . . . . . . . . . . 41
3.6.2 Coupon collector . . . . . . . . . . . . . . . . . . . . . 43
3.6.3 Hoares FIND . . . . . . . . . . . . . . . . . . . . . . . 44
4 Basic probabilistic inequalities 46

4.1 Union bound (Booles inequality) . . . . . . . . . . . . . . . . 46
4.1.1 Applications . . . . . . . . . . . . . . . . . . . . . . . 47
4.1.1.1 Balls in bins . . . . . . . . . . . . . . . . . . 47
4.1.1.2 Independent sets . . . . . . . . . . . . . . . . 47
4.2 Markovs inequality . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.1 Applications . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.1.1 The union bound . . . . . . . . . . . . . . . 49
4.2.1.2 Fair coins . . . . . . . . . . . . . . . . . . . . 49
4.2.1.3 Randomized QuickSort . . . . . . . . . . . . 49
4.2.1.4 Balls in bins . . . . . . . . . . . . . . . . . . 49
CONTENTS iii
4.3 Jensens inequality . . . . . . . . . . . . . . . . . . . . . . . . 50

4.3.0.1 Proof . . . . . . . . . . . . . . . . . . . . . . 50
4.3.1 Applications . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1.1 Fair coins: lower bound . . . . . . . . . . . . 51
4.3.1.2 Fair coins: upper bound . . . . . . . . . . . . 52
4.3.1.3 Sifters . . . . . . . . . . . . . . . . . . . . . . 52
5 Concentration bounds 54
5.1 Chebyshevs inequality . . . . . . . . . . . . . . . . . . . . . . 54
5.1.1 Computing variance . . . . . . . . . . . . . . . . . . . 55
5.1.1.1 Alternative formula . . . . . . . . . . . . . . 55
5.1.1.2 Variance of a Bernoulli random variable . . . 56
5.1.1.3 Variance of a sum . . . . . . . . . . . . . . . 56
5.1.1.4 Variance of a geometric random variable . . 57
5.1.2 More examples . . . . . . . . . . . . . . . . . . . . . . 61
5.1.2.1 Flipping coins . . . . . . . . . . . . . . . . . 61
5.1.2.2 Balls in bins . . . . . . . . . . . . . . . . . . 61
5.1.2.3 Lazy select . . . . . . . . . . . . . . . . . . . 62
5.2 Chernoff bounds . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.1 The classic Chernoff bound . . . . . . . . . . . . . . . 64
5.2.2 Easier variants . . . . . . . . . . . . . . . . . . . . . . 66
5.2.3 Lower bound version . . . . . . . . . . . . . . . . . . . 67
5.2.4 Two-sided version . . . . . . . . . . . . . . . . . . . . 67
5.2.5 Almost-independent variables . . . . . . . . . . . . . . 68
5.2.6 Other tail bounds for the binomial distribution . . . . 69
5.2.7 Applications . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.7.1 Flipping coins . . . . . . . . . . . . . . . . . 69
5.2.7.2 Balls in bins again . . . . . . . . . . . . . . . 70
5.2.7.3 Flipping coins, central behavior . . . . . . . 70
5.2.7.4 Permutation routing on a hypercube . . . . . 71
5.3 The Azuma-Hoeffding inequality . . . . . . . . . . . . . . . . 74
5.3.1 Hoeffdings inequality . . . . . . . . . . . . . . . . . . 74
5.3.2 Azumas inequality . . . . . . . . . . . . . . . . . . . . 77
5.3.3 The method of bounded differences . . . . . . . . . . . 82
5.3.4 Applications . . . . . . . . . . . . . . . . . . . . . . . 83
5.3.4.1 Sprinkling points on a hypercube . . . . . . . 83
5.3.4.2 Chromatic number of a random graph . . . . 84
5.3.4.3 Balls in bins . . . . . . . . . . . . . . . . . . 84
5.3.4.4 Probabilistic recurrence relations . . . . . . . 85
5.3.4.5 Multi-armed bandits . . . . . . . . . . . . . . 86
CONTENTS iv
The UCB1 algorithm . . . . . . . . . . . . . . . 87

Analysis of UCB1 . . . . . . . . . . . . . . . . . 88
5.4 Anti-concentration bounds . . . . . . . . . . . . . . . . . . . . 91
5.4.1 The Berry-Esseen theorem . . . . . . . . . . . . . . . . 91
5.4.2 The Littlewood-Offord problem . . . . . . . . . . . . . 92
6 Randomized search trees 93

6.1 Binary search trees . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1.1 Rebalancing and rotations . . . . . . . . . . . . . . . . 94
6.2 Random insertions . . . . . . . . . . . . . . . . . . . . . . . . 94
6.3 Treaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.3.1 Assumption of an oblivious adversary . . . . . . . . . 98
6.3.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.3.2.1 Searches . . . . . . . . . . . . . . . . . . . . 99
6.3.2.2 Insertions and deletions . . . . . . . . . . . . 100
6.3.2.3 Other operations . . . . . . . . . . . . . . . . 101
6.4 Skip lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7 Hashing 104
7.1 Hash tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.2 Universal hash families . . . . . . . . . . . . . . . . . . . . . . 105
7.2.1 Linear congruential hashing . . . . . . . . . . . . . . . 108
7.2.2 Tabulation hashing . . . . . . . . . . . . . . . . . . . . 109
7.3 FKS hashing . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.4 Cuckoo hashing . . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.4.1 Structure . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.4.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.5 Practical issues . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.6 Bloom filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.6.1 Construction . . . . . . . . . . . . . . . . . . . . . . . 116
7.6.2 False positives . . . . . . . . . . . . . . . . . . . . . . 117
7.6.3 Comparison to optimal space . . . . . . . . . . . . . . 119
7.6.4 Applications . . . . . . . . . . . . . . . . . . . . . . . 120
7.6.5 Counting Bloom filters . . . . . . . . . . . . . . . . . . 121
7.6.6 Count-min sketches . . . . . . . . . . . . . . . . . . . 122
7.6.6.1 Initialization and updates . . . . . . . . . . . 122
7.6.6.2 Queries . . . . . . . . . . . . . . . . . . . . . 123
7.6.6.3 Finding heavy hitters . . . . . . . . . . . . . 125
7.7 Locality-sensitive hashing . . . . . . . . . . . . . . . . . . . . 126
7.7.1 Approximate nearest neighbor search . . . . . . . . . . 126
CONTENTS v
7.7.1.1 Locality-sensitive hash functions . . . . . . . 127

7.7.1.2 Constructing an (r1 , r2 )-PLEB . . . . . . . . 128
7.7.1.3 Hash functions for Hamming distance . . . . 129
7.7.1.4 Hash functions for `1 distance . . . . . . . . 132
8 Martingales and stopping times 133

8.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.2 Submartingales and supermartingales . . . . . . . . . . . . . 134
8.3 The optional stopping theorem . . . . . . . . . . . . . . . . . 135
8.4 Proof of the optional stopping theorem (optional) . . . . . . . 136
8.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.5.1 Random walks . . . . . . . . . . . . . . . . . . . . . . 138
8.5.2 Walds equation . . . . . . . . . . . . . . . . . . . . . 140
8.5.3 Waiting times for patterns . . . . . . . . . . . . . . . . 141
9 Markov chains 143

9.1 Basic definitions and properties . . . . . . . . . . . . . . . . . 144
9.1.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.2 Convergence of Markov chains . . . . . . . . . . . . . . . . . . 147
9.2.1 Stationary distributions . . . . . . . . . . . . . . . . . 148
9.2.1.1 Stationary distributions of reversible chains . 149
9.2.2 Total variation distance . . . . . . . . . . . . . . . . . 150
9.2.2.1 Total variation distance and expectation . . 152
9.2.3 Mixing time . . . . . . . . . . . . . . . . . . . . . . . . 152
9.2.4 Coupling of Markov chains . . . . . . . . . . . . . . . 153
9.2.5 Irreducible and aperiodic chains . . . . . . . . . . . . 154
9.2.6 Convergence of finite irreducible aperiodic Markov chains155
9.2.7 The Metropolis-Hastings algorithm . . . . . . . . . . . 156
9.3 The coupling method . . . . . . . . . . . . . . . . . . . . . . . 158
9.3.1 Random walk on a cycle . . . . . . . . . . . . . . . . . 158
9.3.2 Random walk on a hypercube . . . . . . . . . . . . . . 159
9.3.3 Various shuffling algorithms . . . . . . . . . . . . . . . 160
9.3.3.1 Move-to-top . . . . . . . . . . . . . . . . . . 161
9.3.3.2 Random exchange of arbitrary cards . . . . . 161
9.3.3.3 Random exchange of adjacent cards . . . . . 162
9.3.3.4 Real-world shuffling . . . . . . . . . . . . . . 163
9.3.4 Path coupling . . . . . . . . . . . . . . . . . . . . . . . 164
9.3.4.1 Sampling graph colorings . . . . . . . . . . . 165
9.3.4.2 Metropolis-Hastings and simulated annealing 168
Single peak . . . . . . . . . . . . . . . . . . . . 168
CONTENTS vi
Single peak with very small amounts of noise . 169

9.4 Spectral methods for reversible chains . . . . . . . . . . . . . 170
9.4.1 Spectral properties of a reversible chain . . . . . . . . 170
9.4.2 Analysis of symmetric chains . . . . . . . . . . . . . . 171
9.4.3 Analysis of asymmetric chains . . . . . . . . . . . . . . 173
9.4.4 Conductance . . . . . . . . . . . . . . . . . . . . . . . 174
9.4.5 Easy cases for conductance . . . . . . . . . . . . . . . 175
9.4.6 Edge expansion using canonical paths . . . . . . . . . 176
9.4.7 Congestion . . . . . . . . . . . . . . . . . . . . . . . . 178
9.4.8 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 179
9.4.8.1 Lazy random walk on a line . . . . . . . . . . 179
9.4.8.2 Random walk on a hypercube . . . . . . . . 179
9.4.8.3 Matchings in a graph . . . . . . . . . . . . . 180
9.4.8.4 Perfect matchings in dense bipartite graphs . 183
10 Approximate counting 185

10.1 Exact counting . . . . . . . . . . . . . . . . . . . . . . . . . . 185
10.2 Counting by sampling . . . . . . . . . . . . . . . . . . . . . . 186
10.2.1 Generating samples . . . . . . . . . . . . . . . . . . . 187
10.3 Approximating #DNF . . . . . . . . . . . . . . . . . . . . . . 188
10.4 Approximating #KNAPSACK . . . . . . . . . . . . . . . . . 190
10.5 Approximating exponentially improbable events . . . . . . . . 192
10.5.1 Matchings . . . . . . . . . . . . . . . . . . . . . . . . . 193
10.5.2 Other problems . . . . . . . . . . . . . . . . . . . . . . 194
11 The probabilistic method 195

11.1 Randomized constructions and existence proofs . . . . . . . . 195
11.1.1 Unique hats . . . . . . . . . . . . . . . . . . . . . . . . 196
11.1.2 Ramsey numbers . . . . . . . . . . . . . . . . . . . . . 197
11.1.3 Directed cycles in tournaments . . . . . . . . . . . . . 199
11.2 Approximation algorithms . . . . . . . . . . . . . . . . . . . . 200
11.2.1 MAX CUT . . . . . . . . . . . . . . . . . . . . . . . . 200
11.2.2 MAX SAT . . . . . . . . . . . . . . . . . . . . . . . . 201
11.3 The Lovsz Local Lemma . . . . . . . . . . . . . . . . . . . . 204
11.3.1 General version . . . . . . . . . . . . . . . . . . . . . . 205
11.3.2 Symmetric version . . . . . . . . . . . . . . . . . . . . 206
11.3.3 Applications . . . . . . . . . . . . . . . . . . . . . . . 206
11.3.3.1 Graph coloring . . . . . . . . . . . . . . . . . 206
11.3.3.2 Satisfiability of k-CNF formulas . . . . . . . 207
11.3.4 Non-constructive proof . . . . . . . . . . . . . . . . . . 207
CONTENTS vii
11.3.5 Constructive proof . . . . . . . . . . . . . . . . . . . . 210
12 Derandomization 214
12.1 Deterministic vs. randomized algorithms . . . . . . . . . . . . 215
12.2 Adlemans theorem . . . . . . . . . . . . . . . . . . . . . . . . 216
12.3 Limited independence . . . . . . . . . . . . . . . . . . . . . . 217
12.3.1 MAX CUT . . . . . . . . . . . . . . . . . . . . . . . . 217
12.4 The method of conditional probabilities . . . . . . . . . . . . 218
12.4.1 A trivial example . . . . . . . . . . . . . . . . . . . . . 219
12.4.2 Deterministic construction of Ramsey graphs . . . . . 219
12.4.3 MAX CUT using conditional probabilities . . . . . . . 220
12.4.4 Set balancing . . . . . . . . . . . . . . . . . . . . . . . 221
13 Quantum computing 222

13.1 Random circuits . . . . . . . . . . . . . . . . . . . . . . . . . 222
13.2 Bra-ket notation . . . . . . . . . . . . . . . . . . . . . . . . . 225
13.2.1 States as kets . . . . . . . . . . . . . . . . . . . . . . . 225
13.2.2 Operators as sums of kets times bras . . . . . . . . . . 226
13.3 Quantum circuits . . . . . . . . . . . . . . . . . . . . . . . . . 226
13.3.1 Quantum operations . . . . . . . . . . . . . . . . . . . 228
13.3.2 Quantum implementations of classical operations . . . 229
13.3.3 Representing Boolean functions . . . . . . . . . . . . . 230
13.3.4 Practical issues (which we will ignore) . . . . . . . . . 231
13.3.5 Quantum computations . . . . . . . . . . . . . . . . . 231
13.4 Deutschs algorithm . . . . . . . . . . . . . . . . . . . . . . . 232
13.5 Grovers algorithm . . . . . . . . . . . . . . . . . . . . . . . . 233
13.5.1 Initial superposition . . . . . . . . . . . . . . . . . . . 233
13.5.2 The Grover diffusion operator . . . . . . . . . . . . . . 233
13.5.3 Effect of the iteration . . . . . . . . . . . . . . . . . . 234
A Assignments 237
A.1 Assignment 1: due Sunday, 2016-09-18, at 17:00 . . . . . . . . 237
A.1.1 Bubble sort . . . . . . . . . . . . . . . . . . . . . . . . 238
A.1.2 Finding seats . . . . . . . . . . . . . . . . . . . . . . . 240
A.2 Assignment 2: due Thursday, 2016-09-29, at 23:00 . . . . . . 242
A.2.1 Technical analysis . . . . . . . . . . . . . . . . . . . . 242
A.2.2 Faulty comparisons . . . . . . . . . . . . . . . . . . . . 245
A.3.1 Painting with sprites . . . . . . . . . . . . . . . . . . . 246
A.3.2 Dynamic load balancing . . . . . . . . . . . . . . . . . 248
CONTENTS viii

A.4.1 Re-rolling a random treap . . . . . . . . . . . . . . . . 249
A.4.2 A defective hash table . . . . . . . . . . . . . . . . . . 252
A.5.1 A spectre is haunting Halloween . . . . . . . . . . . . 254
A.5.2 Colliding robots on a line . . . . . . . . . . . . . . . . 254
A.6.1 Another colliding robot . . . . . . . . . . . . . . . . . 258
A.6.2 Return of the sprites . . . . . . . . . . . . . . . . . . . 261
A.7 Final exam . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
A.7.1 Virus eradication (20 points) . . . . . . . . . . . . . . 263
A.7.2 Parallel bubblesort (20 points) . . . . . . . . . . . . . 264
A.7.3 Rolling a die (20 points) . . . . . . . . . . . . . . . . . 265
B Sample assignments from Spring 2014 267

B.1 Assignment 1: due Wednesday, 2014-09-10, at 17:00 . . . . . 267
B.1.1 Bureaucratic part . . . . . . . . . . . . . . . . . . . . . 267
B.1.2 Two terrible data structures . . . . . . . . . . . . . . . 267
B.1.3 Parallel graph coloring . . . . . . . . . . . . . . . . . . 271
B.2.1 Load balancing . . . . . . . . . . . . . . . . . . . . . . 273
B.2.2 A missing hash function . . . . . . . . . . . . . . . . . 273
B.3.1 Tree contraction . . . . . . . . . . . . . . . . . . . . . 274
B.3.2 Part testing . . . . . . . . . . . . . . . . . . . . . . . . 277
Using McDiarmids inequality and some cleverness278
B.4.1 A doubling strategy . . . . . . . . . . . . . . . . . . . 279
B.4.2 Hash treaps . . . . . . . . . . . . . . . . . . . . . . . . 280
B.5.1 Agreement on a ring . . . . . . . . . . . . . . . . . . . 281
B.5.2 Shuffling a two-dimensional array . . . . . . . . . . . . 284
B.6.1 Sampling colorings on a cycle . . . . . . . . . . . . . . 285
B.6.2 A hedging problem . . . . . . . . . . . . . . . . . . . . 286
B.7 Final exam . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
B.7.1 Double records (20 points) . . . . . . . . . . . . . . . 288
B.7.2 Hipster graphs (20 points) . . . . . . . . . . . . . . . . 289
Using the method of conditional probabilities . 289
Using hill climbing . . . . . . . . . . . . . . . . 290
CONTENTS ix
B.7.3 Storage allocation (20 points) . . . . . . . . . . . . . . 290

B.7.4 Fault detectors in a grid (20 points) . . . . . . . . . . 291
C Sample assignments from Spring 2013 294

C.1 Assignment 1: due Wednesday, 2013-01-30, at 17:00 . . . . . 294
C.1.1 Bureaucratic part . . . . . . . . . . . . . . . . . . . . . 294
C.1.2 Balls in bins . . . . . . . . . . . . . . . . . . . . . . . . 294
C.1.3 A labeled graph . . . . . . . . . . . . . . . . . . . . . 295
C.1.4 Negative progress . . . . . . . . . . . . . . . . . . . . . 295
C.2 Assignment 2: due Thursday, 2013-02-14, at 17:00 . . . . . . 297
C.2.1 A local load-balancing algorithm . . . . . . . . . . . . 297
C.2.2 An assignment problem . . . . . . . . . . . . . . . . . 299
C.2.3 Detecting excessive collusion . . . . . . . . . . . . . . 299
C.3.1 Going bowling . . . . . . . . . . . . . . . . . . . . . . 301
C.3.2 Unbalanced treaps . . . . . . . . . . . . . . . . . . . . 302
C.3.3 Random radix trees . . . . . . . . . . . . . . . . . . . 304
C.4.1 Flajolet-Martin sketches with deletion . . . . . . . . . 305
C.4.2 An adaptive hash table . . . . . . . . . . . . . . . . . 307
C.4.3 An odd locality-sensitive hash function . . . . . . . . . 309
C.5 Assignment 5: due Friday, 2013-04-12, at 17:00 . . . . . . . . 310
C.5.1 Choosing a random direction . . . . . . . . . . . . . . 310
C.5.2 Random walk on a tree . . . . . . . . . . . . . . . . . 311
C.5.3 Sampling from a tree . . . . . . . . . . . . . . . . . . . 312
C.6 Assignment 6: due Friday, 2013-04-26, at 17:00 . . . . . . . . 313
C.6.1 Increasing subsequences . . . . . . . . . . . . . . . . . 313
C.6.2 Futile word searches . . . . . . . . . . . . . . . . . . . 314
C.6.3 Balance of power . . . . . . . . . . . . . . . . . . . . . 316
C.7 Final exam . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
C.7.1 Dominating sets . . . . . . . . . . . . . . . . . . . . . 317
C.7.2 Tricolor triangles . . . . . . . . . . . . . . . . . . . . . 318
C.7.3 The n rooks problem . . . . . . . . . . . . . . . . . . . 319
C.7.4 Pursuing an invisible target on a ring . . . . . . . . . 319
D Sample assignments from Spring 2011 321

D.1 Assignment 1: due Wednesday, 2011-01-26, at 17:00 . . . . . 321
D.1.1 Bureaucratic part . . . . . . . . . . . . . . . . . . . . . 321
D.1.2 Rolling a die . . . . . . . . . . . . . . . . . . . . . . . 321
D.1.3 Rolling many dice . . . . . . . . . . . . . . . . . . . . 323
CONTENTS x
D.1.4 All must have candy . . . . . . . . . . . . . . . . . . . 323

D.2.1 Randomized dominating set . . . . . . . . . . . . . . . 324
D.2.2 Chernoff bounds with variable probabilities . . . . . . 326
D.2.3 Long runs . . . . . . . . . . . . . . . . . . . . . . . . . 327
D.3.1 Longest common subsequence . . . . . . . . . . . . . . 329
D.3.2 A strange error-correcting code . . . . . . . . . . . . . 331
D.3.3 A multiway cut . . . . . . . . . . . . . . . . . . . . . . 332
D.4.1 Sometimes successful betting strategies are possible . 333
D.4.2 Random walk with reset . . . . . . . . . . . . . . . . . 335
D.4.3 Yet another shuffling algorithm . . . . . . . . . . . . . 337
D.5 Assignment 5: due Thursday, 2011-04-07, at 23:59 . . . . . . 338
D.5.1 A reversible chain . . . . . . . . . . . . . . . . . . . . 338
D.5.2 Toggling bits . . . . . . . . . . . . . . . . . . . . . . . 339
D.5.3 Spanning trees . . . . . . . . . . . . . . . . . . . . . . 341
D.6 Assignment 6: due Monday, 2011-04-25, at 17:00 . . . . . . . 342
D.6.1 Sparse satisfying assignments to DNFs . . . . . . . . . 342
D.6.2 Detecting duplicates . . . . . . . . . . . . . . . . . . . 343
D.6.3 Balanced Bloom filters . . . . . . . . . . . . . . . . . . 344
D.7 Final exam . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
D.7.1 Leader election . . . . . . . . . . . . . . . . . . . . . . 347
D.7.2 Two-coloring an even cycle . . . . . . . . . . . . . . . 348
D.7.3 Finding the maximum . . . . . . . . . . . . . . . . . . 349
D.7.4 Random graph coloring . . . . . . . . . . . . . . . . . 350
E Sample assignments from Spring 2009 351

E.1 Final exam, Spring 2009 . . . . . . . . . . . . . . . . . . . . . 351
E.1.1 Randomized mergesort (20 points) . . . . . . . . . . . 351
E.1.2 A search problem (20 points) . . . . . . . . . . . . . . 352
E.1.3 Support your local police (20 points) . . . . . . . . . . 353
E.1.4 Overloaded machines (20 points) . . . . . . . . . . . . 354
F Probabilistic recurrences 355

F.1 Recurrences with constant cost functions . . . . . . . . . . . . 355
F.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
F.3 The Karp-Upfal-Wigderson bound . . . . . . . . . . . . . . . 356
F.3.1 Waiting for heads . . . . . . . . . . . . . . . . . . . . . 358
F.3.2 Quickselect . . . . . . . . . . . . . . . . . . . . . . . . 358
CONTENTS xi
F.3.3 Tossing coins . . . . . . . . . . . . . . . . . . . . . . . 358

F.3.4 Coupon collector . . . . . . . . . . . . . . . . . . . . . 359
F.3.5 Chutes and ladders . . . . . . . . . . . . . . . . . . . . 359
F.4 High-probability bounds . . . . . . . . . . . . . . . . . . . . . 360
F.4.1 High-probability bounds from expectation bounds . . 361
F.4.2 Detailed analysis of the recurrence . . . . . . . . . . . 361
F.5 More general recurrences . . . . . . . . . . . . . . . . . . . . . 362
Bibliography 363
Index 377
List of Figures
4.1 Proof of Jensens inequality . . . . . . . . . . . . . . . . . . . 51
5.1 Comparison of Chernoff bound variants . . . . . . . . . . . . 67

5.2 Hypercube network with n = 3 . . . . . . . . . . . . . . . . . 71
6.1 Tree rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

6.2 Balanced and unbalanced binary search trees . . . . . . . . . 95
6.3 Binary search tree after inserting 5 1 7 3 4 6 2 . . . . . . . . 95
6.4 Inserting values into a treap . . . . . . . . . . . . . . . . . . . 97
6.5 Tree rotation shortens spines . . . . . . . . . . . . . . . . . . 100
6.6 Skip list . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
11.1 Tricky step in MAX SAT argument . . . . . . . . . . . . . . . 204
A.1 Filling a screen with Space Invaders . . . . . . . . . . . . . . 247

A.2 Hidden Space Invaders . . . . . . . . . . . . . . . . . . . . . . 262
B.1 Example of tree contraction for Problem B.3.1 . . . . . . . . 275
C.1 Radix tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304

C.2 Word searches . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
xii
List of Tables
3.1 Sum of two dice . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.2 Various conditional expectations on two independent dice . . 42
5.1 Concentration bounds . . . . . . . . . . . . . . . . . . . . . . 55
7.1 Hash table parameters . . . . . . . . . . . . . . . . . . . . . . 106
xiii
List of Algorithms
7.1 Insertion procedure for cuckoo hashing . . . . . . . . . . . . . 113
A.1 One pass of bubble sort . . . . . . . . . . . . . . . . . . . . . . 238
C.1 Adaptive hash table insertion . . . . . . . . . . . . . . . . . . . 308
D.1 Dubious duplicate detector . . . . . . . . . . . . . . . . . . . . 343

D.2 Randomized max-finding algorithm . . . . . . . . . . . . . . . 349
xiv
Preface
These are notes for the Fall 2016 semester version of the Yale course CPSC
469/569 Randomized Algorithms. This document also incorporates the
lecture schedule and assignments, as well as some sample assignments from
previous semesters. Because this is a work in progress, it will be updated
frequently over the course of the semester.
Notes from previous versions of the course can be found at http://www.
cs.yale.edu/homes/aspnes/classes/469/notes-2014.pdf, http://www.
cs.yale.edu/homes/aspnes/classes/469/notes-2013.pdf, http://www.
cs.yale.edu/homes/aspnes/classes/469/notes-2011.pdf, and http://
www.cs.yale.edu/homes/aspnes/pinewiki/CS469(2f)2009.html. Some
sample assignments from these semesters can also be found in the appendix.
Much of the structure of the course follows the textbook, Mitzenmacher
and Upfalss Probability and Computing: Randomized Algorithms and Proba-
bilistic Analysis [MU05], with some material from Motwani and Raghavans
Randomized Algorithms [MR95]. In most cases youll find these textbooks
contain much more detail than what is presented here, so it is probably
better to consider this document a supplement to them than to treat it as
your primary source of information.
I would like to thank my many students and teaching fellows over the
years for their help in pointing out errors and omissions in earlier drafts of
these notes.
xv
Syllabus
Description
A study of randomized algorithms from several areas: graph algorithms,
algorithms in algebra, approximate counting, probabilistically checkable
proofs, and matrix algorithms. Topics include an introduction to tools from
probability theory, including some inequalities such as Chernoff bounds.
Meeting times
Monday and Wednesday 11:3512:50 in WTS A53.
On-line course information

The lecture schedule, course notes, and all assignments can be found in a sin-
gle gigantic PDF file at http://www.cs.yale.edu/homes/aspnes/classes/
469/notes.pdf. You should probably bookmark this file, as it will be up-
dated frequently.
For office hours, see http://www.cs.yale.edu/homes/aspnes#calendar.
Staff
The instructor for the course is James Aspnes. Office: AKW 401. Email:
[email protected]. URL: http://www.cs.yale.edu/homes/aspnes/.
The teaching fellow for the course is Lin Chen. Office: 17 Hillhouse,
Room 333. Email: [email protected]. URL: http://campuspress.yale.
edu/lchen/.
xvi
SYLLABUS xvii
Textbook
The textbook for the class is: Michael Mitzenmacher and Eli Upfal. Prob-
ability and Computing: Randomized Algorithms and Probabilistic Analysis.
Cambridge University Press, 2005. ISBN 0521835402. QA274 M574X 2005.
Reserved books at Bass library

These are other textbooks on randomized algorithms:
Rajeev Motwani and Prabhakar Raghavan, Randomized Algorithms.
Cambridge University Press, 1995. ISBN 0521474655. QA274 M68X
1995. Also available at http://www.books24x7.com/marc.asp?isbn=
0521474655 from Yale campus IP addresses.
A classic textbook in the field.
Juraj Hromkovi, Design and Analysis of Randomized Algorithms: In-
troduction to Design Paradigms. Springer, 2005. ISBN 9783540239499.
QA274 .H76X 2005 (LC). Also available at http://dx.doi.org/10.
1007/3-540-27903-2 from Yale campus IP addresses.
Intended to be a gentler introduction to randomized algorithms than
Motwani and Raghavan, but not as comprehensive.
These are general references on probability theory:
William Feller, An Introduction to Probability Theory and Its Appli-
cations, volumes 1 and 2. Wiley, 1968 (volume 1, 3rd edition); Wiley
1971 (volume 2, 2nd edition). QA273 F43 1968.
The probability theory analog of Knuths Art of Computer Program-
ming: comprehensive, multiple volumes, every theoretical computer
scientist of the right generation owns a copy. Volume 1, which covers
discrete probability, is the most useful for computer science.
Geoffrey R. Grimmett and David R. Stirzaker, Probability and Random
Processes. Oxford University Press, 2001. ISBN 0198572220. QA273
G74X 2001.
Similar in scope to Feller. A good alternative if you are on a budget.
Course requirements
Six homework assignments (60% of the semester grade) plus a final exam
(40%).
SYLLABUS xviii
Use of outside help

Students are free to discuss homework problems and course material with
each other, and to consult with the instructor or a TA. Solutions handed in,
however, should be the students own work. If a student benefits substantially
from hints or solutions received from fellow students or from outside sources,
then the student should hand in their solution but acknowledge the outside
sources, and we will apportion credit accordingly. Using outside resources in
solving a problem is acceptable but plagiarism is not.
Clarifications for homework assignments

From time to time, ambiguities and errors may creep into homework assign-
ments. Questions about the interpretation of homework assignments can
be sent to the instructor at [email protected]. Clarifications will
appear in an updated version of the assignment.
In some circumstances, you may be able to get a faster response using
Piazza, at http://piazza.com/yale/fall2016/cpsc469. Note that ques-
tions you ask there are visible to other students if not specifically marked
private, so be careful about broadcasting your draft solutions.
Late assignments
Late assignments will not be accepted without a Deans Excuse.
Lecture schedule
As always, the future is uncertain, so you should take parts of the schedule
that havent happened yet with a grain of salt. Readings refer to chapters or
sections in the course notes, except for those specified as in MU or MR, which
refer to the course textbook [MU05] or to the supplemental text [MR95].
Office hours, lecture times, and assignment due dates can be found at
http://www.cs.yale.edu/homes/aspnes#calendar.
2016-08-31 Randomized algorithms. What they are and what we will do

with them. Worst-case analysis, average-case analysis, and worst-case
randomized analysis. Las Vegas vs. Monte Carlo algorithms. Example
algorithms: Find the Lady, polynomial equality testing, randomized
QuickSort (analyzed by solving the recurrence). Readings: Chapter 1;
MU 1.1.
2016-09-02 Randomized Quicksort analyzed using linearity of expectation.

Basics of probability theory as used in analyzing randomized algorithms:
probability spaces, independence, the inclusion-exclusion formula, and
conditional probability. Kargers min-cut algorithm. Readings: Chap-
ter 2; MU 1.2, 1.4, and 2.5.
2016-09-07 Law of total probability. Random variables and expectations.

Expectation conditioned on an event. Expectation of a product of
independent random variables. Finite version of Walds equation.
Readings: Chapter 3, excluding 3.5.2 and 3.5.3; MU 2.1-2.4.
2016-09-12 Expectation conditioned on a random variable. The union

bound. Readings: 3.5.2 and 4.1; MU 2.3.
2016-09-14 More probabilistic inequalities: Markovs inequality, Jensens

inequality. Chebyshevs inequality and variance. Readings: 4.25.1;
MU 3.13.4, 2.1.2.
xix
LECTURE SCHEDULE xx
2016-09-19 Chernoff bounds and applications. Valiants randomized rout-

ing algorithm for the hypercube. Readings: section-chernoff-bounds;
MU 4.14.2, 4.5.1.
2016-09-21 Concentration bounds for sums of bounded independent random

variables and martingales with bounded increments: Azuma-Hoeffding
inequality, The method of bounded differences and McDiarmids in-
equality. Readings: 5.3 through 5.3.3. MU 12.1, 12.4.
2016-09-26 Applications of the method of bounded differences and Azuma-

Hoeffding. High-probability bound on QuickSort. Readings: 5.3.4,
except 5.3.4.5. MU 12.5.
2016-09-28 More concentration bounds: the UCB1 algorithm for multi-

armed bandits. Start of randomized search trees and treaps. Readings:
5.3.4.5, Chapter 6 through 6.3.1; [ACBF02], MR 8.18.2, [SA96].
2016-10-03 Analyis of treaps. Skip lists. Readings: Rest of Chapter 6; MR

8.3.
2016-10-05 Start of hash tables. Readings: Chapter 7 through 7.2; MU

5.5.1, 13.3.1, and 13.3.2.
2016-10-10 Perfect hashing, FKS hashing, and cuckoo hashing. Readings:

7.37.4; MU 13.3.3, [PR04].
2016-10-12 Bloom filters and variants. Readings: 7.6 through 7.6.5; MU

5.5.3.
2016-10-17 Count-min sketches and locality-sensitive hashing. Readings:

7.6.6, 7.7.1, 7.7.1.1, and 7.7.1.3; MU 13.4, [IM98].
2016-10-24 More locality-sensitive hashing: constructing an -PLEB. Mar-

tingales, stopping times, and the optional stopping theorem. Readings:
Rest of 7.7, 8.18.4; MU 12.2.
2016-10-26 Applications of the optional stopping theorem: random walks

with absorbing barriers, Walds equation, waiting times for patterns.
Readings: 8.5; MU 12.3.
2016-10-31 Start of Markov chains: Definition of a Markov chain. Tran-

sition probabilities and transition matrices. Existence of a stationary
distribution for finite chains. Irreducible and aperiodic chains. Markov
LECTURE SCHEDULE xxi
Chain Monte Carlo sampling. Total variation distance and the cou-
pling lemma. Readings: 9.1, 9.2 up through 9.2.4; MU 7.17.2,
11.111.4.
2016-11-02 Markov chain convergence; showing bounds on mixing time

using couplings. Readings: rest of 9.2.
2016-11-07 Path coupling. Sampling graph colorings when k 2 + 1. A

bit about adjusting reversible Markov chains using Metropolis-Hastings.
Readings: 9.2.7 and 9.3.4; MU 11.511.6.
2016-11-09 Markov chain convergence via spectral methods. Conductance

and Cheegers inequality. Readings: 9.49.4.5; MR 11.3.
2016-11-14 Bounding conductance using canonical paths. Readings: 9.4.69.4.8,

rest of MR 11.3. Note: The bound claiming at most N m canonical
paths across any edge in the random walk on matching, which I tried to
do in class, and which appeared in earlier versions of these notes, doesnt
work. The issue is that the auxiliary information ((X Y ) \ (S T ))
may have two extra edges on top of some underlying matching, not
just one. Accounting for the second edge gives a bound of N m2 that
you can read about in the current, corrected version of 9.4.8.3 if you
care.
2016-11-16 Approximate counting: sampling lemma, Karp-Luby algorithm

for #DNF, Dyers algorithm for #KNAPSACK. Readings: Chapter 10
except 10.5; MU 10.110.3, [Dye03].
2016-11-28 The probabilistic method: randomized constructions and ex-

istence proofs, approximate MAX CUT, randomized rounding for
approximate MAX SAT (but not the proof that it works). Readings:
Chapter 11 through 11.2.2; MU 6.16.2.
2016-11-30 More probabilistic method: proof for Goemans-Williamson

MAX SAT algorithm, Lovsz local lemma. Readings: Rest of Chap-
ter 11 (you can skip 11.3.4); [MT10].
2016-12-05 Derandomization: Adlemans theorem, pairwise independence,

method of conditional probabilities. Readings: Chapter 12; MR 2.3,
MU 6.3 and 13.1.2.
2016-12-07 Quantum computing. Readings: Chapter 13; [Deu89, Gro96].

LECTURE SCHEDULE xxii
2016-12-18 The final exam was given Sunday, December 18th, 2016, start-
ing at 14:00, in HLH17 101. It was be a closed-book test covering
all material discussed during the semester. Sample solutions are in
Appendix A.7.
Chapter 1
Randomized algorithms
A randomized algorithm flips coins during its execution to determine what

to do next. When considering a randomized algorithm, we usually care
about its expected worst-case performance, which is the average amount
of time it takes on the worst input of a given size. This average is computed
over all the possible outcomes of the coin flips during the execution of the
algorithm. We may also ask for a high-probability bound, showing that
the algorithm doesnt consume too much resources most of the time.
In studying randomized algorithms, we consider pretty much the same
issues as for deterministic algorithms: how to design a good randomized
algorithm, and how to prove that it works within given time or error bounds.
The main difference is that it is often easier to design a randomized algo-
rithmrandomness turns out to be a good substitute for cleverness more
often than one might expectbut harder to analyze it. So much of what
one does is develop good techniques for analyzing the often very complex
random processes that arise in the execution of an algorithm. Fortunately,
in doing so we can often use techniques already developed by probabilists
and statisticians for analyzing less overtly algorithmic processes.
Formally, we think of a randomized algorithm as a machine M that
computes M (x, r), where x is the problem input and r is the sequence of
random bits. Our machine model is the usual random-access machine or
RAM model, where we have a memory space that is typically polynomial in
the size of the input n, and in constant time we can read a memory location,
write a memory location, or perform arithmetic operations on integers of
up to O(log n) bits.1 In this model, we may find it easier to think of the
1
This model is unrealistic in several ways: the assumption that we can perform arithmetic
on O(log n)-bit quantities in constant omits at least a factor of (log log n) for addition
1
CHAPTER 1. RANDOMIZED ALGORITHMS 2
random bits as supplied as needed by some subroutine, where generating

a random integer of size O(log n) takes constant time; the justification for
this assumption is that it takes constant time to read the next O(log n)-sized
value from the random input.
Because the number of these various constant-time operations, and thus
the running time for the algorithm as a whole, may depend on the random
bits, it is now a random variablea function on points in some probability
space. The probability space consists of all possible sequences r, each
of which is assigned a probability Pr [r] (typically 2|r| ), and the running
time for M on some input x is generally given as an expected value2
Er [time(M (x, r))], where for any X,
X
Er [X] = X(r) Pr [r] . (1.0.1)
r
We can now quote the performance of M in terms of this expected value:

where we would say that a deterministic algorithms runs in time O(f (n)),
where n = |x| is the size of the input, we instead say that our randomized algo-
rithm runs in expected time O(f (n)), which means that Er [time(M (x, r))] =
O(f (|x|)) for all inputs x.
This is distinct from traditional worst-case analysis, where there is no
r and no expectation, and average-case analysis, where there is again no
r and the value reported is not a maximum but an expectation over some
distribution on x. The following trivial example shows the distinction.
1.1 A trivial example

Let us consider a variant of the classic card game Find the Lady3 Here a
dealer puts down three cards and we want to find a specific card among the
three (say, the Queen of Spades). In this version, the dealer will let us turn
over as many cards as we want, but each card we turn over will cost us a
dollar. If we find the queen, we get two dollars.
and probably more for multiplication in any realistic implementation; while the assumption
that we can address nc distinct locations in memory in anything less than nc/3 time in the
worst case requires exceeding the speed of light. But for reasonably small n, this gives a
pretty good approximation of the performance of real computers, which do in fact perform
arithmetic and access memory in a fixed amount of time, although with fixed bounds on
the size of both arithmetic operands and memory.
2
Well see more details of these and other concepts from probability theory in Chapters 2
and 3.
3
Often called Three-card Monte, but that will get confusing when we talk about
Monte Carlo algorithms later.
Because this is a toy example, we assume that the dealer is not cheating.
The author is not responsible for the results of applying the analysis below
to real-world games where this assumption does not hold.
A deterministic algorithm tries the cards in some fixed order. A clever
dealer will place the Queen in the last place we look: so the worst-case payoff
is a loss of a dollar.
In the average case, if we assume that all three positions are equally likely
to hold the target card, then we turn over one card a third of the time, two
cards a third of the time, and all three cards a third of the time; this gives
an expected payoff of
1
(1 + 2 + 3) 2 = 0.
3
But this requires making assumptions about the distribution of the cards,
and we are no longer doing worst-case analysis.
The trick to randomized algorithms is that we can can obtain the same
expected payoff even in the worst case by supplying the randomness ourselves.
If we turn over cards in a random order, then the same analysis for the
average case tells us we get the same expected payoff of 0but unlike the
average case, we get this expected performance no matter where the dealer
places the cards.
1.2 Verifying polynomial identities

A less trivial example is described in [MU05, 1.1]. Here we are given two
products of polynomials and we want to determine if they compute the same
function. For example, we might have
p(x) = (x 7)(x 3)(x 1)(x + 2)(2 x + 5)

q(x) = 2 x5 13 x4 21 x3 + 127 x2 + 121 x 210
These expressions both represent degree-5 polynomials, and it is not

obvious without multiplying out the factors of p whether they are equal or
not. Multiplying out all the factors of p may take as much as O(d2 ) time if we
assume integer multiplication takes unit time and do it the straightforward
way.4 We can do better than this using randomization.
The trick is that evaluating p(x) and q(x) takes only O(d) integer opera-
tions, and we will find p(x) = q(x) only if either (a) p(x) and q(x) are really
the same polynomial, or (b) x is a root of p(x) q(x). Since p(x) q(x)
4
It can be faster if we do something sneaky like use fast Fourier transforms [SS71].
has degree at most d, it cant have more than d roots. So if we choose x

uniformly at random from some much larger space, its likely that we will not
get a root. Indeed, evaluating p(11) = 112320 and q(11) = 120306 quickly
shows that p and q are not in fact the same.
This is an example of a Monte Carlo algorithm, which is an algorithm
that runs in a fixed amount of time but only gives the right answer some of
the time. (In this case, with probability 1 d/r, where r is the size of the
range of random integers we choose x from.) Monte Carlo algorithms have
the unnerving property of not indicating when their results are incorrect,
but we can make the probability of error as small as we like by running
the algorithm repeatedly. For this particular algorithm, the probability of
error after k trials is only (d/r)k , which means that for fixed d/r we need
O(log(1/)) iterations to get the error bound down to any given . If we are
really paranoid, we could get the error down to 0 by testing d + 1 distinct
values, but now the cost is as high as multiplying out p again.
The error for this algorithm is one-sided: if we find a witness to the fact
that p 6= q, we are done, but if we dont, then all we know is that we havent
found a witness yet. We also have the property that if we check enough
possible witnesses, we are guaranteed to find one.
A similar property holds in the classic Miller-Rabin primality test,
a randomized algorithm for determining whether a large integer is prime
or not.5 The original version, due to Gary Miller [Mil76] showed that, as
in polynomial identity testing, it might be sufficient to pick a particular
set of deterministic candidate witnesses. Unfortunately, this result depends
on the truth of the extended Riemann hypothesis, a notoriously difficult
open problem in number theory. Michael Rabin [Rab80] demonstrated that
choosing random witnesses was enough, if we were willing to accept a small
probability of incorrectly identifying a composite number as prime.
For many years it was open whether it was possible to test primality
deterministically in polynomial time without unproven number-theoretic
assumptions, and the randomized Miller-Rabin algorithm was one of the
most widely-used randomized algorithms for which no good deterministic
alternative was known. Finally, Agrawal et al. [AKS04] demonstrated how to
test primality deterministically using a different technique, although the cost
of their algorithm is high enough that Miller-Rabin is still used in practice.
5
We will not describe this algorithm here.
1.3 Randomized QuickSort

The QuickSort algorithm [Hoa61a] works as follows. For simplicity, we
assume that no two elements of the array being sorted are equal.
If the array has > 1 elements,
Pick a pivot p uniformly at random from the elements of the

array.
Split the array into A1 and A2 , where A1 contains all elements
< p elements > p.
Sort A1 and A2 recursively and return the sequence A1 , p, A2 .
Otherwise return the array.
The splitting step takes exactly n 1 comparisons, since we have to check

each non-pivot against the pivot. We assume all other costs are dominated by
the cost of comparisons. How many comparisons does randomized QuickSort
do on average?
There are two ways to solve this problem: the dumb way and the smart
way. Well do it the dumb way now and save the smart way for 1.3.2.
1.3.1 Brute force method: solve the recurrence

Let T (n) be the expected number of comparisons done on an array of n
elements. We have T (0) = T (1) = 0 and for larger n,
1 n1
X
T (n) = (T (k) + T (n 1 k)) . (1.3.1)
n k=0
Why? Because there are n equally-likely choices for our pivot (hence the
1/n), and for each choice the expected cost is T (k) + T (n 1 k), where k
is the number of elements that land in A1 . Formally, we are using here the
law of total probability, which says that for any random variable X and
partition of the probability space into events B1 . . . Bn , then
X
E [X] = Bi E [X | Bi ] ,
where
1 X
E [X | Bi ] = X()
Pr [Bi ] B
i
is the conditional expectation of X conditioned on Bi , which we can

think of as just the average value of X if we know that Bi occurred. (See
2.3.1 for more details.)
So now we just have to solve this ugly recurrence. We can reasonably
guess that when n 1, T (n) an log n for some constant a. Clearly this
holds for n = 1. Now apply induction on larger n to get
1 n1
X
T (n) = (n 1) + (T (k) + T (n 1 k))
n k=0
2 n1
X
= (n 1) + T (k)
n k=0
2 n1
X
= (n 1) + T (k)
n k=1
2 n1
X
(n 1) + ak log k
n k=1
Z n
2
(n 1) + ak log k
n k=1
!
2a n2 log n n2 1
= (n 1) + +
n 2 4 4
!
2a n2 log n n2 1
= (n 1) + +
n 2 4 4
an a
= (n 1) + an log n + .
2 2n
If we squint carefully at this recurrence for a while we notice that setting
a = 2 makes this less than or equal to an log n, since the remaining terms
become (n 1) n + 1/n = 1/n 1, which is negative for n 1. We can
thus confidently conclude that T (n) 2n log n (for n 1).
1.3.2 Clever method: use linearity of expectation

Alternatively, we can use linearity of expectation (which well discuss
further in 3.4.1 to compute the expected number of comparisons used by
randomized QuickSort.
Imagine we use the following method for choosing pivots: we generate a
random permutation of all the elements in the array, and when asked to sort
some subarray A0 , we use as pivot the first element of A0 that appears in our
list. Since each element is equally likely to be first, this is equivalent to the
actual algorithm. Pretend that we are always sorting the numbers 1 . . . n and
define for each pair of elements i < j the indicator variable Xij to be 1 if
i is compared to j at some point during the execution of the algorithm and 0
otherwise. Amazingly, we can actually compute the probability of this event
(and thus E [Xij ]): the only time i and j are compared is if one of them is
chosen as a pivot before they are split up into different arrays. How do they
get split up into different arrays? This happens if some intermediate element
k is chosen as pivot first, that is, if some k with i < k < j appears in the
permutation before both i and j. Occurrences of other elements dont affect
the outcome, so we can concentrate on the restriction of the permutations
to just the numbers i through j, and we win if this restricted permutation
starts with either i or j. This event occurs with probability 2/(j i + 1), so
we have E [Xij ] = 2/(j i + 1). Summing over all pairs i < j gives:

X X
E Xij = E [Xij ]
i<j i<j
X 2
=
i<j
ji+1
X ni+1
n1 X 2
=
i=1 k=2
k
n Xi
X 2
=
i=2 k=2
k
n
X 2(n k + 1)
=
k=2
k
n
2(n + 1)
X
= 2
k=2
k
n
X 2(n + 1)
= 2(n 1)
k=2
k
= 2(n + 1)(Hn 1) 2(n 1)
= 2(n + 1)Hn 4n.
Here Hn = ni=1 1i is the n-th harmonic number, equal to ln n + +

P
O(n1 ), where 0.5772 is the Euler-Mascheroni constant (whose exact

value is unknown!). For asymptotic purposes we only need Hn = (log n).
For the first step we are taking advantage of the fact that linearity of
expectation doesnt care about the variables not being independent. The
rest is just algebra.

This is pretty close to the bound of 2n log n we computed using the
recurrence in 1.3.1. Given that we now know the exact answer, we could in
principle go back and use it to solve the recurrence exactly.6
Which way is better? Solving the recurrence requires less probabilistic
handwaving (a more polite term might be insight) but more grinding
out inequalities, which is a pretty common trade-off. Since I am personally
not very clever I would try the brute-force approach first. But its worth
knowing about better methods so you can try them in other situations.
1.4 Classifying randomized algorithms by their goals

1.4.1 Las Vegas vs Monte Carlo
One difference between QuickSort and polynomial equality testing that
QuickSort always succeeds, but may run for longer than you expect; while
the polynomial equality tester always runs in a fixed amount of time, but may
produce the wrong answer. These are examples of two classes of randomized
algorithms, which were originally named by Lszl Babai [Bab79]:7
A Las Vegas algorithm fails with some probability, but we can tell
when it fails. In particular, we can run it again until it succeeds, which
means that we can eventually succeed with probability 1 (but with a
potentially unbounded running time). Alternatively, we can think of a
Las Vegas algorithm as an algorithm that runs for an unpredictable
amount of time but always succeeds (we can convert such an algorithm
back into one that runs in bounded time by declaring that it fails if it
runs too longa condition we can detect). QuickSort is an example of
a Las Vegas algorithm.
A Monte Carlo algorithm fails with some probability, but we cant

tell when it fails. If the algorithm produces a yes/no answer and the
failure probability is significantly less than 1/2, we can reduce the
probability of failure by running it many times and taking a majority of
the answers. The polynomial equality-testing algorithm in an example
of a Monte Carlo algorithm.
6
We wont.
7
To be more precise, Babai defined Monte Carlo algorithms based on the properties of
Monte Carlo simulation, a technique dating back to the Manhattan project. The term
Las Vegas algorithm was new.
The heuristic for remembering which class is which is that the names
were chosen to appeal to English speakers: in Las Vegas, the dealer can tell
you whether youve won or lost, but in Monte Carlo, le croupier ne parle que
Franais, so you have no idea what hes saying.
Generally, we prefer Las Vegas algorithms, because we like knowing
when we have succeeded. But sometimes we have to settle for Monte Carlo
algorithms, which can still be useful if we can get the probability of failure
small enough. For example, any time we try to estimate an average by
sampling (say, inputs to a function we are trying to integrate or political
views of voters we are trying to win over) we are running a Monte Carlo
algorithm: there is always some possibility that our sample is badly non-
representative, but we cant tell if we got a bad sample unless we already
know the answer we are looking for.
1.4.2 Randomized complexity classes

Las Vegas vs Monte Carlo is the typical distinction made by algorithm
designers, but complexity theorists have developed more elaborate classifi-
cations. These include algorithms with one-sided failure properties. For
these algorithms, we never get a bogus yes answer but may get a bogus
no answer (or vice versa). This gives us several complexity classes that act
like randomized versions of NP, co-NP, etc.:
The class R or RP (randomized P) consists of all languages L for which

a polynomial-time Turing machine M exists such that if x L, then
Pr [M (x, r) = 1] 1/2 and if x 6 L, then Pr [M (x, r) = 1] = 0. In
other words, we can find a witness that x L with constant probability.
This is the randomized analog of NP (but its much more practical,
since with NP the probability of finding a winning witness may be
exponentially small).
The class co-R consists of all languages L for which a poly-time Turing
machine M exists such that if x 6 L, then Pr [M (x, r) = 1] 1/2 and
if x L, then Pr [M (x, r) = 1] = 0. This is the randomized analog of
co-NP.
The class ZPP (zero-error probabilistic P ) is defined as RP co-RP.

If we run both our RP and co-RP machines for polynomial time, we
learn the correct classification of x with probability at least 1/2. The
rest of the time we learn only that weve failed (because both machines
return 0, telling us nothing). This is the class of (polynomial-time)
Las Vegas algorithms. The reason it is called zero-error is that we

can equivalently define it as the problems solvable by machines that
always output the correct answer eventually, but only run in expected
polynomial time.
The class BPP (bounded-error probabilistic P) consists of all languages

L for which a poly-time Turing machine exists such that if x 6 L,
then Pr [M (x, r) = 1] 1/3, and if x L, then Pr [M (x, r) = 1]
2/3. These are the (polynomial-time) Monte Carlo algorithms: if our
machine answers 0 or 1, we can guess whether x L or not, but we
cant be sure.
The class PP (probabilistic P) consists of all languages L for which a

poly-time Turing machine exists such that if x 6 L, then Pr [M (x, r) = 1]
1/2, and if x L, then Pr [M (x, r) = 1] < 1/2. Since there is only an
exponentially small gap between the two probabilities, such algorithms
are not really useful in practice; PP is mostly of interest to complexity
theorists.
Assuming we have a source of random bits, any algorithm in RP, co-RP,

ZPP, or BPP is good enough for practical use. We can usually even get
away with using a pseudorandom number generator, and there are good
reasons to suspect that in fact every one of these classes is equal to P.
1.5 Classifying randomized algorithms by their meth-

ods
We can also classify randomized algorithms by how they use their randomness
to solve a problem. Some very broad categories:8
Avoiding worst-case inputs, by hiding the details of the algorithm

from the adversary. Typically we assume that an adversary supplies
our input. If the adversary can see what our algorithm is going to
do (for example, he knows which door we will open first), he can use
this information against us. By using randomness, we can replace our
predictable deterministic algorithm by what is effectively a random
choice of many different deterministic algorithms. Since the adversary
doesnt know which algorithm we are using, he cant (we hope) pick
an input that is bad for all of them.
8
These are largely adapted from the introduction to [MR95].
Sampling. Here we use randomness to find an example or examples

of objects that are likely to be typical of the population they are drawn
from, either to estimate some average value (pretty much the basis of
all of statistics) or because a typical element is useful in our algorithm
(for example, when picking the pivot in QuickSort). Randomization
means that the adversary cant direct us to non-representative samples.
Hashing. Hashing is the process of assigning a large object x a small

name h(x) by feeding it to a hash function h. Because the names are
small, the Pigeonhole Principle implies that many large objects hash to
the same name (a collision). If we have few objects that we actually
care about, we can avoid collisions by choosing a hash function that
happens to map them to different places. Randomization helps here
by keeping the adversary from choosing the objects after seeing what
our hash function is.
Hashing techniques are used both in load balancing (e.g., insuring
that most cells in a hash table hold only a few objects) and in finger-
printing (e.g., using a cryptographic hash function to record a
fingerprint of a file, so that we can detect when it has been modified).
Building random structures. The probabilistic method shows

the existence of structures with some desired property (often graphs
with interesting properties, but there are other places where it can be
used) by showing that a randomly-generated structure in some class
has a nonzero probability of having the property we want. If we can
beef the probability up to something substantial, we get a randomized
algorithm for generating these structures.
Symmetry breaking. In distributed algorithms involving multi-

ple processes, progress may be stymied by all the processes trying to
do the same thing at the same time (this is an obstacle, for example,
in leader election, where we want only one process to declare itself
the leader). Randomization can break these deadlocks.
Chapter 2
Probability theory
In this chapter, we summarize the parts of probability theory that we

need for the course. This is not really a substitute for reading an actual
probability theory book like Feller [Fel68] or Grimmett and Stirzaker [GS01].
The basic idea of probability theory is that we want to model all possible
outcomes of whatever process we are studying simultaneously. This gives
the notion of a probability space, which is the set of all possible outcomes;
for example, if we roll two dice, the probability space would consist of all 36
possible combinations of values. Subsets of this space are called events; an
example in the two-dice space would be the event that the sum of the two
dice is 11, given by the set A = {h5, 6i , h6, 5i}. The probability of an event
A is given by a probability measure Pr [A]; for simple probability spaces,
this is just the sum of the probabilities of the individual outcomes contains
in A, while for more general spaces, we define the measure on events first
and the probabilities of individual outcomes are derived from this. Formal
definitions of all of these concepts are given later in this chapter.
When analyzing a randomized algorithm, the probability space describes
all choices of the random bits used by the algorithm, and we can think of the
possible executions of an algorithm as living within this probability space.
More formally, the sequence of operations carried out by an algorithm and the
output it ultimately produces are examples of random variablesfunctions
from a probability space to some other setwhich we will discuss in detail
in Chapter 3.
12
CHAPTER 2. PROBABILITY THEORY 13
2.1 Probability spaces and events

A discrete probability space is a countable set of points or outcomes
. Each in has a probability Pr [], which is a real value with 0
Pr [] 1. It is required that = 1.
P
P
An event A is a subset of ; its probability is Pr [A] = A Pr [].
We require that Pr [] = 1, and it is immediate from the definition that
Pr [] = 0.
The complement A or A of an event A is the event A. It is always
the case that Pr [A] = 1 Pr [A].
This fact is a special case of the general principle that if A1 , A2 , . . . forms
a partition of that is, if Ai Aj = when i 6= j and Ai = then
S
Pr [Ai ] = 1. It happens to be the case that A and A form a partition of

P
consisting of exactly two elements.

Whenever A1 , A2 , . . . are disjoint events (i.e., when Ai Aj = for all
i 6= j), it is the case that Pr [ Ai ] = Pr [Ai ]. This fact does not hold in
S P
general for events that are not disjoint.

For discrete probability spaces, all of these facts can be proven directly
from the definition of probabilities for events. For more general probability
spaces, its no longer possible to express the probability of an event as the
sum of the probabilities of its elements, and we adopt an axiomatic approach
instead.
2.1.1 General probability spaces

More general probability spaces consist of a triple (, F, Pr) where is
a set of points, F is a -algebra (a family of subsets of that contains
and is closed under complement and countable unions) of measurable
sets, and Pr is a function from F to [0, 1] that gives Pr [] = 1 and satisfies
S P
countable additivity: when A1 , . . . are disjoint, Pr [ Ai ] = Pr [Ai ].
This definition is needed for uncountable spaces, because (under certain set-
theoretic assumptions) we may not be able to assign a meaningful probability
to all subsets of .
Formally, this definition is often presented as three axioms of proba-
bility, due to Kolmogorov [Kol33]:
1. Pr [A] 0 for all A F.
2. Pr [] = 1.
3. For any countable collection of disjoint events A1 , A2 , . . . ,

" #
[ X
Pr Ai = Pr [Ai ] .
i i
Its not hard to see that the discrete probability spaces defined in the
preceding section satisfy these axioms.
General probability spaces arise in randomized algorithms when we have
an algorithm that might consume an unbounded number of random bits.
The problem now is that an outcome consists of countable sequence of bits,
and there are uncountably many such outcomes. The solution is to consider
as measurable events only those sets with the property that membership
in them can be determined after a finite amount of time. Formally, the
probability space the set {0, 1}N of all sequences of 0 and 1 values indexed
by the natural numbers, and the measurable sets F are all sets that can be
generated by countable unions1 cylinder sets, where a cylinder set consists
of all extensions xy of some finite prefix x. The probability measure itself is
obtained by assigning the set of all points that start with x the probability
2|x| , and computing the probabilities of other sets from the axioms. 2
An oddity that arises in general probability spaces is it may be that every
particular outcome has probability zero but their union has probability 1.
For example, the probability of any particular infinite string of bits is 0, but
the set containing all such strings is the entire space and has probability
1. This is where the fact that probabilities only add over countable unions
comes in.
Most randomized algorithms books gloss over general probability spaces,
with three good reasons. The first is that if we truncate an algorithm after
a finite number of steps, we are usually get back to a discrete probability
space, which avoids a lot of worrying about measurability and convergence.
The second is that we are often implicitly working in a probability space that
is either discrete or well-understood (like the space of bit-vectors described
above). The last is that the Kolmogorov extension theorem says that
if we specify Pr [A1 A2 Ak ] consistently for all finite sets of events
1
As well as complements and countable intersections. However, it is not hard to show
that that sets defined using these operations can be reduced to mere countable unions of
cylinder sets.
2
This turns out to give the same probabilities as if we consider each outcome as a
real number in the interval [0, 1] and use Lebesgue measure to compute the probability of
events. For some applications, thinking of our random values as real numbers (or even
sequences of real numbers) can make things easier: consider for example what happens
when we want to choose one of three outcomes with equal probability.
{A1 . . . Ak }, then there exists some probability space that makes these prob-
abilities work, even if we have uncountably many such events. So its usually
enough to specify how the events we care about interact, without worrying
about the details of the underlying space.
2.2 Boolean combinations of events

Even though events are defined as sets, we often think of them as representing
propositions that we can combine using the usual Boolean operations of NOT
(), AND (), and OR (). In terms of sets, these correspond to taking a
complement A = \ A, an intersection A B, or a union A B.
We can use the axioms of probability to calculate the probability of A:
Lemma 2.2.1.
h i
Pr A = 1 Pr [A] .
Proof. First note that A A = , so A A =

h i is a disjoint union of
3
countably many events. This gives Pr [A] + Pr A = Pr [] = 1.
For example, if our probability space consists of the six outcomes of a fair
i and A = [outcome is 3] with Pr [A] = 5/6, then Pr [outcome is not 3] =
diehroll,

Pr A = 1 1/6 = 5/6. Though this example is trivial, using the formula
does save us from having to add up the five cases where we dont get 3.
If we want to know the probability of A B, we need to know more
about the relationship between A and B. For example, it could be that
A and B are both events representing a fair coin coming up heads, with
Pr [A] = Pr [B] = 1/2. The probability of A B could be anywhere between
1/2 and 0:
For ordinary fair coins, wed expect that half the time that A happens,
B also happens. This gives Pr [A B] = (1/2) (1/2) = 1/4. To
make this formal, we might define our probability space as having
four outcomes HH, HT, TH, and TT, each of which occurs with equal
probability.
But maybe A and B represent the same fair coin: then A B = A and
Pr [A B] = Pr [A] = 1/2.
3
Countable need not be infinite, so 2 is countable.
At the other extreme, maybe A and B represent two fair coins welded
together so that if one comes up heads the other comes up tails. Now
Pr [A B] = 0.
With a little bit of tinkering, we could also find probabilities for

the outcomes in our four-outcome space to make Pr [A] = Pr [HH] +
Pr [HT] = 1/2 and Pr [B] = Pr [HH] + Pr [TH] = 1/2 while setting
Pr [A B] = Pr [HH] to any value between 0 and 1/2.
The difference between the nice case where Pr [A B] = 1/4 and the
other, more annoying cases where it doesnt is that in the first case we have
assumed that A and B are independent, which is defined to mean that
Pr [A B] = Pr [A] Pr [B].
In the real world, we expect events to be independent if they refer to
parts of the universe that are not causally related: if we flip two coins that
arent glued together somehow, then we assume that the outcomes of the
coins are independent. But we can also get independence from events that
are not causally disconnected in this way. An example would be if we rolled
a fair four-sided die labeled HH, HT, TH, TT, where we take the first letter
as representing A and the second as B.
Theres no simple formula for Pr [A B] when A and B are not disjoint,
even for independent events, but we can compute the probability by splitting
up into smaller, disjoint events and using countable additivity:
h i
(A B)
Pr [A B] = Pr (A B) (A B)
h i h i
+ Pr A B
= Pr [A B] + Pr A B
h i h i

= Pr [A B] + Pr A B + Pr A B + Pr [A B] Pr [A B]
= Pr [A] + Pr [B] Pr [A B] .
The idea is that we can compute Pr [A B] by adding up the individual

probabilities and then subtracting off the part where the counted the event
twice.
This is a special case of the general inclusion-exclusion formula, which
says:
Lemma 2.2.2. For any finite sequence of events A1 . . . An ,

" n #
[ X X X
Pr Ai = Pr [Ai ] Pr [Ai Aj ] + Pr [Ai Aj Ak ] . . .
i=1 i i<j i<j<k
" #
|S|+1
X \
= (1) Pr Ai . (2.2.1)
S{1...n},S6= iS
into 2n disjoint events BT , where BT = ( iT Ai )

T
Proof.
T Partition

/ Ai is the event that all Ai occur for i in T and no Ai occurs for i not
iT
in T . Then Ai is the union of all BT with T 3 i and Ai is the union of all
S
BT with T 6= .
That the right-hand side gives the probability of this event is a sneaky con-
sequence of the binomial theorem, and in particular the fact that ni=1 (1)i =
P
Pn i n
i=0 (1) 1 = (1 1) 1 is zero if n > 0 and 1 if n = 0. Using this
fact after rewriting the right-hand side using the BT events gives
" #
|S|+1
(1)|S|+1
X \ X X
(1) Pr Ai = Pr [BT ]
S{1...n},S6= iS S{1...n},S6= T S]

(1)|S|+1
X X
= Pr [BT ]
T {1...n} ST,S6=
n
!!
X X
i |T |
= Pr [BT ] (1)
T {1...n} i=1
i

Pr [BT ] ((1 1)|T | 1)
X
=
T {1...n}

Pr [BT ] (1 0|T | )
X
=
T {1...n}
X
= Pr [BT ]
T {1...n},T 6=
" n #
\
= Pr Ai .
i=1
2.3 Conditional probability

The probability of A conditioned on B or probability of A given B,
written Pr [A | B], is defined by
Pr [A B]
Pr [A | B] = , (2.3.1)
Pr [B]
provided Pr [B 6= 0]. If Pr [B] = 0, we generally cant condition on B.
Such conditional probabilities represent the effect of restricting our prob-
ability space to just B, which can think of as computing the probability of
each event if we know that B occurs. The intersection in the numerator
limits A to circumstances where B occurs, while the denominator normalizes
the probabilities so that, for example, Pr [ | B] = Pr [B | B] = 1.
2.3.1 Application to algorithm analysis

The reason we like conditional probability in algorithm analysis is that it
gives us a natural way to model the kind of case analysis that we are used
to applying to deterministic algorithms. Suppose we are trying to prove that
a randomized algorithm works (event A) with a certain probability. Most
likely, the first random thing the algorithm does is flip a coin, giving two

possible outcomes h B and
i B. Countable additivity tells us that Pr [A] =
Pr [A B] + Pr A B , which we can rewrite using conditional probability
as
h i h i
,
Pr [A] = Pr [A | B] Pr [B] + Pr A B Pr B (2.3.2)
a special case of the law of total probability.

h nicei about this expression is that we can often compute Pr [A | B]
Whats

and Pr A B by looking at what the algorithm does starting from the
and use the formula to
point where it has just gotten heads (B) or tails (B),
combine these values to get the overall probability of success.
For example, if
Pr [class occurs | snow] = 3/5,

Pr [class occurs | no snow] = 99/100, and
Pr [snow] = 1/10,
then
Pr [class occurs] = (3/5) (1/10) + (99/100) (1 1/10) = 0.951.

More generally, we can do the same computation for any partition of

into countably many disjoint events Bi :
" #
[
Pr [A] = Pr (A Bi )
i
X
= Pr [A Bi ]
i
X
= Pr [A | Bi ] Pr [Bi ] , (2.3.3)
i,Pr[Bi ]6=0
which is the law of total probability. Note that the last step works for
each term only if Pr [A | Bi ] is well-defined, meaning that Pr [Bi ] 6= 0. But
any such case contributes nothing to the previous sum, so we get the correct
answer if we simply omit any terms h from
ithe sum for which Pr [Bi ] = 0.

A special case arises when Pr A B = 0, which occurs, for example,
if A B. Then we just have Pr [A] = Pr [A | B] Pr [B]. If we consider an
event A = A1 A2 Ak , then we can iterate this expansion to get
Pr [A1 A2 Ak ] = Pr [A1 Ak1 ] Pr [Ak | A1 , . . . , Ak1 ]

= Pr [A1 Ak2 ] Pr [Ak1 | A1 , . . . , Ak2 ] Pr [Ak | A1 , . . . , Ak1 ]
= ...
k
Y
= Pr [Ai | A1 , . . . , Ai ] . (2.3.4)
i=1
Here Pr [A | B, C, . . .] is short-hand for Pr [B C . . .], the probability

that A occurs given that all of B, C, etc., occur.
2.3.1.1 Racing coin-flips

Suppose that I flip coins and allocate a space for each heads that I get before
the coin comes up tails. Suppose that you then supply me with objects (each
of which takes up one unit of space), one for each heads that you get before
you get tails. What are my chances of allocating enough space?
Lets start by solving this directly using the law of total probability. Let
Ai be the event that I allocate i spaces. The event Ai is the intersection of
i independent events that I get heads in the first i positions and the event
that I get tails in position i + 1; this multiplies out to (1/2)i+1 . Let Bi be
the similar event that you supply i objects. Let W be the event that I win.
To make the Ai partition the space, we must also add an extra event A
equal to the singleton set {HHHHHHH . . .} consisting of the all-H sequence;

this has probability 0 (so it wont have much of an effect), but we need to
include it since HHHHHHH . . . is not contained in any of the other Ai .
We can compute
Pr [W | Ai ] = Pr [B0 B1 Bi | Ai ]
= Pr [B0 B1 Bi ]
= Pr [B0 ] + Pr [B1 ] + + Pr [Bi ]
i
X
= (1/2)i
j=0
1 (1/2)i+1
= (1/2)
1 1/2
= 1 (1/2)i+1 . (2.3.5)
The clean form of this expression suggests strongly that there is a better
way to get it, and that this way involves taking the negation of the intersection
of i + 1 independent events that occur with probability 1/2 each. With a
little reflection, we can see that the probability that your objects dont fit in
my buffer is exactly (1/2)i+1
From the law of total probability (2.3.3),

X
Pr [W ] = (1 (1/2)i+1 )(1/2)i+1
i=0

X
=1 (1/4)i+1
i=0
1
=1 11 1/4
4
= 2/3.
This gives us our answer. However, we again see an answer that is

suspiciously simple, which suggests looking for another way to find it. We
can do this using conditional probability by defining new events Ci , where
Ci contains all sequences of coin-flips for both players where get i heads in a
row but at least one gets tails on the (i + 1)-th coin. These events plus the
probability-zero event C = {HHHHHHH . . . , TTTTTTT . . .} partition the
space, so Pr [W ] = i=0 Pr [W | Ci ] Pr [Ci ].
P
Now we ask, what is Pr [W | Ci ]? Here we only need to consider three

cases, depending on the outcomes of our (i + 1)-th coin-flips. The cases
hH, Ti and hT, Ti cause me to win, while the case hT, Hi causes me to
lose, and each case occurs with equal probability conditioned on Ci (which
excludes hH, Hi). So I win 2/3 of the time conditioned on Ci , and summing
Pr [W ] =
P
i=0 (2/3) Pr [Ci ] = 2/3 since Pr [Ci ] sums to 1, because the union
of these events includes the entire space except for the probability-zero event
C .
Still another approach is to compute the probability that our runs have
exactly the same length ( i 2i = 1/3), and argue by symmetry that
P
i=1 2
the remaining 2/3 probability is equally split between my run being longer
(1/3) and your run being longer (1/3). Since W occurs if my run is just as
long or longer, Pr[W ] = 1/3 + 1/3 = 2/3. A nice property of this approach is
that the only summation involved is over disjoint events, so we get to avoid
using conditional probability entirely.
2.3.1.2 Kargers min-cut algorithm

Here well give a simple algorithm for finding a global min-cut in a multi-
graph,4 due to David Karger [Kar93].
The idea is that we are given a multigraph G, and we want to partition
the vertices into nonempty sets S and T such that the number of edges with
one endpoint in S and one endpoint in T is as small as possible. There are
many efficient ways to do this, most of which are quite sophisticated. There
is also the algorithm we will now present, which solves the problem with
reasonable efficiency using almost no sophistication at all (at least in the
algorithm itself).
The main idea is that given an edge uv, we can construct a new multigraph
G1 by contracting the edge: in G1 , u and v are replaced by a single vertex,
and any edge that used to have either vertex as an endpoint now goes to the
combined vertex (edges with both endpoints in {u, v} are deleted). Kargers
algorithm is to contract edges chosen uniformly at random until only two
vertices remain. All the vertices that got packed into one of these become
S, the others become T . It turns out that this finds a minimum cut with
probability at least 1/ n2 .
Theorem 2.3.1. Given any min cut (S, T ) of a graph G on n vertices,

Kargers algorithm outputs (S, T ) with probability at least 1/ n2 .

Proof. Let (S, T ) be a min cut of size k. Then the degree of each vertex v
is at least k (otherwise (v, G v) would be a smaller cut), and G contains
4
Unlike ordinary graphs, multigraphs can have more than one edge between two vertices.
at least kn/2 edges. The probability that we contract an ST edge is thus

at most k/(kn/2) = 2/n, and the probability that we dont contract one is
1 2/n = (n 2)/n. Assuming we missed collapsing (S, T ) the first time,
we now have a new graph G1 with n 1 vertices in which the min cut is still
of size k. So now the chance that we miss (S, T ) is (n 3)/(n 1). We stop
when we have two vertices left, so the last step succeeds with probability
1/3.
We can compute the probability that the ST cut is never contracted by
applying (2.3.4), which just tells us to multiply all the conditional probabili-
ties together:
n
Y i2 2
= .
i=3
i n(n 1)
This tells us what happens when we are considering a particular min cut.
If the graph has more than one min cut, this only makes our life easier. Note
that since each min cut turns up with probability at least 1/ n2 , there cant
n
be more than 2 of them.5 But even if there is only one, we have a good
chance of finding it if we simply re-run the algorithm substantially more
than n2 times.
2.3.2 Conditional probability and independence

Rearranging (2.3.1) gives Pr [A B] = Pr [B] Pr [A | B] = Pr [A] Pr [B | A].
In many cases, knowing that B occurs tells us nothing about whether A
occurs; if so, we have Pr [A | B] = Pr [B], which implies that Pr [A B] =
Pr [A | B] Pr [B] = Pr [A] Pr [B]events A and B are independent. So
Pr [A | B] = Pr [A] gives an alternative criterion for independence when
Pr [B] is nonzero.6
A set of events A1 , A2 , . . . is independent if Ai is independent of B when
B is any Boolean formula of the Aj for j 6= i. The idea is that you cant
predict Ai by knowing anything about the rest of the events.
A set of events A1 , A2 , . . . is pairwise independent if each Ai and
Aj , i 6= j are independent. It is possible for a set of events to be pairwise

5
The suspiciously combinatorial appearance of the 1/ n2 suggests that there should
be some way of associating minimum cuts with particular pairs of vertices, but Im not
aware of any natural way to do this. It may be that sometimes the appearance of a simple
expression in a surprising context may just stem from the fact that there arent very many
distinct simple expressions.
6
If Pr [B] is zero, then A and B are always independent.
independent but not independent; a simple example is when A1 and A2 are

the events that two independent coins come up heads and A3 is the event
that both coins come up with the same value. The general version of pairwise
independence is k-wise independence, which means that any subset of k
(or fewer) events are independent.
Chapter 3
Random variables
A random variable on a probability space is just a function with do-

main .1 Rather than writing a random variable as f () everywhere, the
convention is to write a random variable as a capital letter (X, Y , S, etc.)
and make the argument implicit: X is really X(). Variables that arent
random (or arent variable) are written in lowercase.
For example, consider the probability space corresponding to rolling two
independent fair six-sided dice. There are 36 possible outcomes in this space,
corresponding to the 6 6 pairs of values hx, yi we might see on the two dice.
We could represent the value of each die as a random variable X or Y given
by X(hx, yi) = x or Y (hx, yi) = y, but for many applications, we dont care
so much about the specific values on each die. Instead, we want to know
the sum S = X + Y of the dice. This value S is also random variable; as a
function on , its defined by S(hx, yi) = x + y.
Random variables need not be real-valued. Theres no reason why we
cant think of the pair hx, yi itself a random variable, whose range is the
set [1 . . . 6] [1 . . . 6]. Similarly, if we imagine choosing a point uniformly at
random in the unit square [0, 1]2 , its coordinates are a random variable. For
a more exotic example, the random graph Gn,p obtained by starting with
n vertices and including each possible edge with independent probability p
is a random variable whose range is the set of all graphs on n vertices.
1
Technically, this only works for discrete spaces. In general, a random variable is a
measurable function from a probability space (, F) to some other set S equipped with
its own -algebra F 0 . What makes a function measurable in this sense is that that for any
set A in F 0 , the inverse image f 1 (A) must be in F. See 3.3 for more details.
24
CHAPTER 3. RANDOM VARIABLES 25
3.1 Operations on random variables

Random variables may be combined using standard arithmetic operators,
have functions applied to them, etc., to get new random variables. For
example, the random variable X/Y is a function from that takes on the
value X()/Y () on each point .
3.2 Random variables and events

Any random variable X allows us to define events based on its possible values.
Typically these are expressed by writing a predicate involving the random
variable in square brackets. An example would be the probability that the
sum of two dice is exactly 11: [S = 11]; or that the sum of the dice is less than
5: [S < 5]. These are both sets of outcomes; we could expand [S = 11] =
{h5, 6i , h6, 5i} or [S < 5] = {h1, 1i , h1, 2i , h1, 3i , h2, 1i , h2, 2i , h3, 1i}. This
allows us to calculate the probability that a random variable has particular
2 1 6
properties: Pr [S = 11] = 36 = 18 and Pr [S < 5] = 36 = 61 .
Conversely, given any event A, we can define an indicator random
variable 1A that is 1 when A occurs and 0 when it doesnt.2 Formally,
1A () = 1 for in A and 1A () = 0 for not in A.
Indicator variables are mostly useful when combined with other random
variables. For example, if you roll two dice and normally collect the sum of
the values but get nothing if it is 7, we could write your payoff as S 1[S6=7] .
The probability mass function of a random variable gives Pr [X = x]
for each possible value x. For example, our random variable S has the
probability mass function show in Table 3.1. For a discrete random variable
X, the probability mass function gives enough information to calculate the
probability of any event involving X, since we can just sum up cases using
countable additivity. This gives us another way to compute Pr [S < 5] =
Pr [S = 2] + Pr [S = 3] + Pr [S = 4] = 1+2+3 36 = 16 .
For two random variables, the joint probability mass function gives
Pr [X = x Y = y] for each pair of values x and y (this generalizes in the
obvious way for more than two variables).
2
Some people like writing A for these.
You may also see [P ] where P is some predicate, a convention known as Iverson
notation or the Iverson bracket that was invented by Iverson for the programming
language APL, appears in later languages like C where the convention is that true predicates
evaluate to 1 and false ones to 0, and ultimately popularized for use in mathematicswith
the specific choice of square brackets to set off the predicateby Graham et al. [GKP88].
Out of these alternatives, I personally find 1A to be the least confusing.
S Probability
2 1/36
3 2/36
4 3/36
5 4/36
6 5/36
7 6/36
8 5/36
9 4/36
10 3/36
11 2/36
12 1/36
Table 3.1: Probability mass function for the sum of two independent fair
six-sided dice
We will often refer to the probability mass function as giving the distri-
bution or joint distribution of a random variable or collection of random
variables, even though distribution (for real-valued variables) technically
refers to the cumulative distribution function F (x) = Pr [X x], which
is generally not directly computable from the probability mass function for
continuous random variables that take on uncountably many values. To
the extent that we can, we will try to avoid continuous random variables,
and the rather messy integration theory needed to handle them.
Two or more random variables are independent if all sets of events
involving different random variables are independent. In terms of proba-
bility mass functions, X and Y are independent if Pr [X = x Y = y] =
Pr [X = x] Pr [Y = y]. In terms of cumulative distribution functions, X and
Y are independent if Pr [X = x Y = y] = Pr [X = x] Pr [Y = y]. As with
events, we generally assume that random variables associated with causally
disconnected processes are independent, but this is not the only way we
might have independence.
Its not hard to see that the individual die values X and Y in our
two-dice example are independent, because every possible combination of
values x and y has the same probability 1/36 = Pr [X = x] Pr [Y = y]. If we
chose a different probability distribution on the space, we might not have
independence.
3.3 Measurability
For discrete probability spaces, any function on outcomes can be a random
variable. The reason is that any event in a discrete probability space has
a well-defined probability. For more general spaces, in order to be useful,
events involving a random variable should have well-defined probabilities.
For discrete random variables that take on only countably many values
(e.g., integers or rationals), its enough for the event [X = x] (that is, the
set { | X() = x}) to be in F for all x. For real-valued random variables,
we ask that the event [X x] be in F. In these cases, we say that X is
measurable with respect to F, or just measurable F. More exotic random
variables use a definition of measurability that generalizes the real-valued
version, which we probably wont need.3 Since we usually just assume that
3
The general version is that if X takes on values on another measure space (0 , F 0 ),
then the inverse image X 1 (A) = { | X() A} of any set A in F 0 is in F. This
means in particular that Pr maps through X to give a probability measure on 0 by
Pr0 [A] = Pr [X 1 (A)], and the condition on X 1 (A) being in F makes this work.
all of our random variables are measurable unless we are doing something
funny with F to represent ignorance, this issue wont come up much.
3.4 Expectation
The expectation or expected value of a random variable X is given by
P
E [X] = x x Pr [X = x]. This is essentially an average value of X weighted
by probability, and it only makes sense if X takes on values that can be
summed in this way (e.g., real or complex values, or vectors in a real- or
complex-valued vector space). Even if the expectation makes sense, it may
be that a particular random variable X doesnt have an expectation, because
the sum fails to converge.
For example,
P6 ifX and Y are independent fair six-sided dice, then E [X] =
E [Y ] = i=1 i 16 = 21 7
6 = 2 , while E [X + Y ] is the rather horrific
12
X 2 1 + 3 2 + 4 3 + 5 4 + 6 5 + 7 6 + 8 5 + 9 4 + 10 3 + 11 2 + 12 1
i Pr [X + Y = i] =
i=2
36
252
= = 7.
36
7 7
The fact that 7 = 2 + 2 here is not a coincidence.
3.4.1 Linearity of expectation

The main reason we like expressing the run times of algorithms in terms of
expectation is linearity of expectation: E [aX + bY ] = E [aX] + E [bY ]
for all random variables X and Y for which E [X] and E [Y ] are defined, and
all constants a and b. This means that we can compute the running time for
different parts of our algorithm separately and then add them together, even
if the costs of different parts of the algorithm are not independent.
P P
The general version is E [ ai Xi ] = ai E [Xi ] for any finite collection of
random variables Xi and constants ai , which follows by applying induction to
the two-variable case. A special case is E [cX] = c E [X] when c is constant.
For discrete random variables, linearity of expectation follows immediately
from the definition of expectation and the fact that the event [X = x] is the
disjoint union of the events [X = x, Y = y] for all y:

X
E [aX + bY ] = (ax + by) Pr [X = x Y = y]
x,y
X X
=a x Pr [X = x, Y = y] + b y Pr [X = x, Y = y]
x,y x,y
X X X X
=a x P r[X = x, Y = y] + b y Pr [X = x, Y = y]
x y y x
X X
=a xP r[X = x] + b y Pr [Y = y]
x y
= a E [X] + b E [Y ] .
Note that this proof does not require that X and Y be independent. The
sum of two fair six-sided dice always has expectation 72 + 72 = 7, whether
they are independent dice, the same die counted twice, or one die X and its
complement 7 X.
Linearity of expectation makes it easy to compute the expectations of
random variables that are expressed as sums of other random variables.
One example that will come up a lot is a binomial random variable,
which is the sum S = ni=1 Xi , where each Xi is an independent, identically
P
distributed random variable that is 1 with probability p and 0 with probability

q = 1p. These are called binomial random variables because the probability
that S is equal to k is given by
!
n k nk
Pr [S = k] = p q , (3.4.1)
k
which is the k-th term in the binomial expansion of (p + q)n . (In addition,
each Xi individually is called a Bernoulli random variable.4 ) In this case
each Xi has E [Xi ] = p, so E [S] is just np. It is possible to calculate this
fact directly from (3.4.1), but its a lot more work.
3.4.1.1 Linearity of expectation for infinite sequences

For infinite sequences of random variables, linearity of expectation may
break down. This is true even if the sequence is countable. An example
is the St. Petersburg paradox, in which a gambler bets $1 on a double-
or-nothing game, then bets $2 if she loses, then $4, and so on, until she
eventually wins and stops, up $1. If we represent the gamblers gain or loss
4
Named for Jacob Bernoulli, the 18th-century Swiss mathematician.
at stage i as a random variable Xi , its easy to show that E [Xi ] = 0, because

the gambler either wins 2i with equal probability, or doesnt play at all.
So
P P
i=0 E [Xi ] = 0. But E [ i=0 Xi ] = 1, because the probability that the
gambler doesnt eventually win is zero.5
Fortunately, these pathological cases dont come up often in algorithm
analysis, and with some additional side constraints we can apply linearity of
expectation even to infinite sums of random variables. The simplest is when
Xi 0 for all i; then E [
P P
i=0 Xi ] exists and is equal to i=0 E [Xi ] whenever
the sum of the expectations converges (this is a consequence of the monotone
convergence theorem). Another condition that works is if | ni=0 Xi | Y for
P
all n, where Y is a random variable with finite expectation; the simplest

version of this is when Y is constant. See [GS92, 5.6.12] or [Fel71, IV.2]for
more details.
3.4.2 Expectation and inequalities

If X Y (that is, if the event [X Y ] holds with probability 1), then
E [X] E [Y ]. For discrete spaces in which all outcomes occur with
non-zero probability, the proof is trivial:
X
E [X] = Pr [] X()

X
Pr [] Y ()

= E [Y ] .
(The claim continues to hold even in more general cases, but the proof is
more work.)
One special case of this that comes up a lot is that X 0 implies
E [X] 0.
3.4.3 Expectation of a product

When two random variables X and Y are independent, it also holds that
E [XY ] = E [X] E [Y ]. The proof (at least for discrete random variables) is
5
The trick here is that we are trading aPprobability-1 gain of 1 against a probability-0

loss of . So we could declare that E i=0
Xi involves 0 () and is undefined.
But this would lose the useful property that expectation isnt affected by probability-0
outcomes. As often happens in mathematics, we are forced to choose between candidate
definitions based on which bad consequences we most want to avoid, with no way to avoid
all of them. So the standard definition of expectation allows the St. Petersburg paradox
because the alternatives are worse.
straightforward:
XX
E [XY ] = xy Pr [X = x, Y = y]
x y
XX
= xy Pr [X = x] Pr [Y = y]
x y
! !
X X
= x Pr [X = x] Pr [Y = y]
x y
= E [X] E [Y ] .
For example, the expectation of the product of two independent fair
2
six-sided dice is 27 = 49 4 .
This is not true for arbitrary random variables. If we compute the
expectation of the product of a single fair six-sided die with itself, we get
11+22+33+44+55+66
6 = 91
6 which is much larger.
The difference E [XY ] E [X] E [Y ] is called the covariance of X and
Y , written Cov [X, Y ]. It will come back later when we look at concentration
bounds in Chapter 5.
3.4.3.1 Walds equation (simple version)

Computing the expectation of a product does not often come up directly in
the analysis of a randomized algorithm. Where we might expect to do it is
when we have a loop: one random variable N tells us the number of times
we execute the loop, while another random variable X tells us the cost of
each iteration. The problem is that if each iteration is randomized, then we
really have a sequence of random variables X1 , X2 , . . . , and what we want
to calculate is
N
" #
X
E Xi , (3.4.2)
i=1
where we cant use the sum formula directly because N is a random variable
and we cant use the product formula because the Xi are all different random
variables.
If N and the Xi are all independent,6 and N is bounded by some fixed
maximum n, then we can apply the product rule to get the value of (3.4.2)
6
When I presented this in lecture, an objection was raised for the randomized loop
example that N is most likely not independent of the Xi , because a successful iteration
of the loop that causes the algorithm to terminate may have a different distribution
on cost than a failed iteration. This turns out not to be a problem, because this
difference ends up being counted in the expectations of the Xi . For the specific case of
by throwing in a few indicator variables. The idea is that the contribution

of Xi to the sum is given by Xi 1[N i] , and because we assume that N is
h i
independent of the Xi , if we need to compute E Xi 1[N i] , we can do so by
h i
computing E [Xi ] E 1[N i] .
So we get
"N # " n #
X X
E Xi = E Xi 1[N i]
i=1 i=1
n
X h i
= E Xi 1[N i]
i=1
n
X h i
= E [Xi ] E 1[N i] .
i=1
For general Xi we have to stop here. But if we also know that the Xi all
have the same expectation , then E [Xi ] doesnt depend on i and we can
bring it out of the sum. This gives
n
X h i n
X h i
E [Xi ] E 1[N i] = E 1[N i]
i=1 i=1
= E [N ] . (3.4.3)
This equation is a special case of Walds equation, which we will see

again in 8.5.2. The main difference between this version and the general
version is that here we had to assume that N was independent of the Xi ,
which may not be true if our loop is a while loop, and termination is
correlated with the time taken by the last iteration.
But for simple cases, (3.4.3) can still be useful. For example, if we throw
one six-sided die to get N , and then throw N six-sided dice and add them
up, we get the same expected total 72 72 = 49 4 as if we just multiply two
six-sided dice. This is true even though the actual distribution of values is
very different in the two cases.
the loop, we can argue that if each attempts succeeds with probability p after s steps on
average, and fails with probability q = 1 p after f steps on average, then the expected
cost is s + f (1/p 1), since on average we get 1/p 1 failures before our first success.
Alternatively, we can compute = E [Xi ] = ps + (1 p)f and E [N ] = 1/p, giving
E [N ] = (ps + (1 p)f )/p = s + (1/p 1)f , which happens to be the same quantity. This
coincidence turns out to be a consequence of the full version of Walds equation, described
in 8.5.2.
3.5 Conditional expectation

We can also define a notion of conditional expectation, analogous to
conditional probability. There are three versions of this, depending on how
fancy we want to get about specifying what information we are conditioning
on.
3.5.1 Expectation conditioned on an event

The conditional expectation of X conditioned on an event A is written
E [X | A] and defined by
X X Pr [X = x A]
E [X | A] = x Pr [X = x | A] = . (3.5.1)
x x Pr [A]
This is essentially the weighted average value of X if we know that A occurs.
Most of the properties that we see with ordinary expectations continue
to hold for conditional expectation. For example, linearity of expectation
E [aX + bY | A] = a E [X | A] + b E [Y | A] (3.5.2)
holds whenever a and b are constant on A.
Similarly if Pr [X Y ] A = 1, E [X | A] E [Y | A].
Conditional expectation is handy because we can use it to compute
expectations by case analysis the same way we use conditional probabilities
using the law of total probability (see 2.3.1). If A1 , A2 , . . . are a countable
partition of , then
!
X X X
Pr [Ai ] E [X | Ai ] = Pr [Ai ] x Pr [X = x | Ai ]
i i x
!
X X Pr [X = x Ai ]
= Pr [Ai ] x
i x Pr [A]i
!
X X Pr [X = x Ai ]
= Pr [Ai ] x
i x Pr [A]i
XX
= x Pr [X = x Ai ]
i x
!
X X
= x Pr [X = x Ai ]
x i
X
= x Pr [X = x]
x
= E [X] . (3.5.3)
This is actually a special case of the law of iterated expectation, which

we will see in the next section.
3.5.2 Expectation conditioned on a random variable

In the previous section, we considered computing E [X] by breaking it up
into disjoint cases E [X | A1 ], E [X | A2 ], etc. But keeping track of all the
events in our partition of is a lot of book-keeping. Conditioning on a
random variable lets us combine all these conditional probabilities into a
single expression
E [X | Y ] ,
the expectation of X conditioned on Y which is defined to have the
value E [X | Y = y] whenever Y = y.7
Note that E [X | Y ] is generally a function of Y , unlike E [X] which is a
constant. This also means that E [X | Y ] is a random variable, and its value
can depend on which outcome we picked from our probability space .
The intuition behind the definition is that E [X | Y ] is the best estimate we
can make of X given that we know the value of Y but nothing else.
If we want to be formal about the definition, we can specific the value of
E [X | Y ] explicitly for each point :
E [X | Y ] () = E [X | Y = Y ()] . (3.5.4)
This is just another way of saying what we said already: if you want to know
what the expectation of X is conditioned on Y when you get outcome ,
find the value of Y at and condition on seeing that.
Here is a simple example. Suppose that X and Y are independent fair
coin-flips that take on the values 0 and 1 with equal probability. Then our
probability space has four elements, and looks like this:
h0, 0i h0, 1i
h1, 0i h1, 1i
where each tuple hx, yi gives the values of X and Y .

We can also define the total number of heads as Z = X + Y . If we label
all the points in our probability space with Z(), we get a picture that
looks like this:
0 1
1 2
7
If Y is not discrete, the situation is more complicated. See [Fel71, III.2 and V.9V.11]
or [GS92, 7.9].
This is what we see if we know the exact value of both coin-flips (or at
least the exact value of Z).
But now suppose we only know X, and want to compute E [Z] X. When
X = 0, E [Z] X = 0 = 12 0 + 12 1 = 12 ; and when X = 1, E [Z] X = 1 =
1 1 3
2 1 + 2 2 = 2 . So drawing E [Z] X over our probability space gives
1 1
2 2
3 3
2 2
Weve averaged the value of Z across each row, since each row corresponds
to one of the possible values of X.
If instead we compute E [Z] Y , we get this picture instead:
1 3
2 2
1 3
2 2
Now instead of averaging across rows (values of X) we average across

columns (values of Y ). So the left column shows E [Z] Y = 0 = 12 and the
right columns shows E [Z] Y = 1 = 32 , which is pretty much what wed expect.
Nothing says that we can only condition on X and Y . What happens if
we condition on Z?
Now we are going to get fixed values for each possible value of Z. If we
compute E [X] Z, then when Z = 0 this will be 0 (because Z = 0 implies
X = 0), and when Z = 2 this will be 1 (because Z = 2 implies X = 1). The
middle case is E [X | Z = 1] = 12 0 + 12 1 = 12 , because the two outcomes
h0, 1i and h1, 0i that give Z = 1 are equally likely. The picture is
1
0 2
1
2 1
3.5.2.1 Calculating conditional expectations

Usually we will not try to E [X | Y ] individually for each possible or even
each possible value y of Y . Instead, we can use various basic facts to compute
E [X | Y ] by applying arithmetic to random variables.
The two basic facts to start with are:
1. If X is a function of Y , then E [X | Y ] = X. Proof: Suppose X =
f (Y ). From (3.5.4), for each outcome , we have E [X | Y ] () =
E [X | Y = Y ()] = E [f (Y ()) | Y = Y ()] = f (Y ()) = X().
2. If X is independent of Y , then E [X | Y ] = E [X]. Proof: Now for each

, we have E [X | Y ] () = E [X | Y = Y ()] = E [X].
We also have a rather strong version of linearity of expectation. If A and

B are both functions of Z, then
E [AX + BY | Z] = A E [X | Z] + B E [Y | Z] . (3.5.5)
Here is a proof for discrete probability spaces. For each value z of Z, we

have
X Pr []
E [A(Z)X + B(Z)Y | Z = z] = (A(z)X() + B(z)Y ()
Pr [Z = z]
Z 1 (z)
X Pr [] X Pr []
= A(z) X() + B(z) Y ()
Pr [Z = z] Pr [Z = z]
Z 1 (z) Z 1 (z)
= A(z) E [X | Z = z] + B(z) E [Y | Z = z] ,
which is the value when Z = z of A E [X | Z] + B E [Y | Z].

This means that we can quickly simplify many conditional expectations.
If we go back to the example of the previous section, where Z = X + Y is
the sum of two independent fair coin-flips X and Y , then we can compute
E [Z | X] = E [X + Y | X]
= E [X | X] + E [Y | X]
= X + E [Y ]
1
=X+ .
2
Similarly, E [Z | Y ] = E [X + Y | Y ] = 12 + Y .
In some cases we have enough additional information to run this in
reverse. If we know Z and want to estimate X, we can use the fact that
X and Y are symmetric to argue that E [X | Z] = E [Y | Z]. But then
Z = E [X + Y | Z] = 2 E [X | Z], so E [X | Z] = Z/2. Note that this works
in general only if the events [X = a, Y = b] and [X = b, Y = a] have the
same probabilities for all a and b even if we condition on Z, which in this
case follows from the fact that X and Y are independent and identically
distributed and that addition is commutative. Other cases may be messier.8
Other facts, like X Y implies E [X | Z] E [Y | Z], can be proved
using similar techniques.
8
An example with X and Y i.d. but not independent is to imagine that we roll a
six-sided die to get X, and let Y = X + 1 if X < 6 and Y = 1 if X = 6. Now knowing
Z = X + Y = 3 tells me that X = 1 and Y = 2 exactly, neither of which is Z/2.
3.5.2.2 The law of iterated expectation

The law of iterated expectation says that
E [X] = E [E [X | Y ]] . (3.5.6)
When Y is discrete, this is just (3.5.3) in disguise. For each of the

(countably many) values of Y that occurs with nonzero probability, let Ay
be the event [Y = y]. Then these events are a countable partition of , and
X
E [E [X | Y ]] = Pr [Y = y] E [E [X | Y ] | Y = y]
y
X
= Pr [Y = y] E [X | Y = y]
y
= E [X] .
The trick here is that we use (3.5.3) to expand out the original expression in
terms of the events Ay , then notice that E [X | Y ] is equal to E [X | Y = y]
whenever Y = y.
So as claimed, conditioning on a variable gives a way to write averaging
over cases very compactly.
Its also not too hard to show that iterated expectation works with partial
conditioning:
E [E [X | Y, Z] | Y ] = E [X | Y ] . (3.5.7)
3.5.2.3 Conditional expectation as orthogonal projection (optional)

If you are comfortable with linear algebra, it may be helpful to think about
expectation conditioned on a random variable as a form of projection onto
a subspace. In this section, well give a brief description of how this works
for a finite, discrete probability space. For a more general version, see for
example [GS92, 7.9].
Consider the set of all real-valued random variables on the probability
space {TT, TH, HT, HH} corresponding to flipping two independent fair coins.
We can think of each such random variable as a vector in R4 , where the
four coordinates give the value of the variable on the four possible outcomes.
For example, the indicator variable for the event that the first coin is heads
would look like X = h0, 0, 1, 1i and the indicator variable for the event that
the second coin is heads would look like Y = h0, 1, 0, 1i.
When we add two random variables together, we get a new random vari-
able. This corresponds to vector addition: X + Y = h0, 0, 1, 1i + h0, 1, 0, 1i =
h0, 1, 1, 2i. Multiplying a random variable by a constant looks like scalar

multiplication 2X = 2 h0, 0, 1, 1i = h0, 0, 2, 2i. Because random variables
support both addition and scalar multiplication, and because these operation
obey the axioms of a vector space, we can treat the set of all real-valued
random variables defined on a given probability space as a vector space, and
apply all the usual tools from linear algebra to this vector space.
One thing in particular we can look at is subspaces of this vector space.
Consider the set of all random variables that are functions of X. These are
vectors of the form ha, a, b, bi, and adding any two such vectors or multiplying
any such vector by a scalar yields another vector of this form. So the functions
of X form a two-dimensional subspace of the four-dimensional space of all
random variables. An even lower-dimensional subspace is the one-dimensional
subspace of constants: vectors of the form ha, a, a, ai. As with functions of
X, this set is closed under addition and multiplication by a constant.
When we take the expectation of X, we are looking for a constant
that gives us the average value of X. In vector terms, this means that
E [h0, 0, 1, 1i] = h1/2, 1/2, 1/2, 1/2i. This expectation vector is in fact the
orthogonal projection of h0, 0, 1, 1i onto the subspace generated by 1 =
h1, 1, 1, 1i; we can tell this because the dot-product of X E [X] with 1
is h1/2, 1/2, 1/2, 1/2i h1, 1, 1, 1i = 0. If instead we take a conditional
expectation, we are again doing an orthogonal projection, but now onto a
higher-dimensional subspace. So E [X + Y | X] = h1/2, 1/2, 3/2, 3/2i is the
orthogonal projection of X + Y = h0, 1, 1, 2i onto the space of all functions
of X, which is generated by the vectors h0, 0, 1, 1i and h1, 1, 0, 0i. As in the
simple expectation, the dot-product of (X + Y ) E [X + Y | X] with either
of these basis vectors is 0.
Many facts about conditional expectation translate in a straightforward
way to facts about projections. Linearity of expectation is equivalent to
linearity of projection onto a subspace. The law of iterated expectation
E [E [X | Y ]] = E [X] says that projecting onto the subspace of function of Y
and then onto the subspace of constants is equivalent to projection directly
down to the subspace of constants; this is true in general for projection
operations. Its also possible to represent other features of probability spaces
in terms of expectations; for example, E [XY ] acts like an inner product
for random variables, E X 2 acts like the square of the Euclidean distance,

and the fact that E [X] is an orthogonal projection of X means that E [X]
is precisely the constant value that minimizes the distance E (X )2 .
We wont actually use any of these facts in the following, but having another
way to look at conditional expectation may be helpful in understanding how
it works.
3.5.3 Expectation conditioned on a -algebra

Expectation conditioned on a random variable is actually a special case of
the expectation of X conditioned on a -algebra F. Recall that a -algebra is
a family of subsets of that includes and is closed under complement and
countable union; for discrete probability spaces, this turns out to be the set
of all unions of equivalence classes for some equivalence relation on ,9 and
we think of F as representing knowledge of which equivalence class we are in,
but not which point in the equivalence class we land on. An example would
be if consists of all values (X1 , X2 ) obtained from two die rolls, and F
consists of all sets A such that whenever one point with X1 () + X2 () = s
is in A, so is every other point 0 with X1 ( 0 ) + X2 ( 0 ) = s. (This is the
-algebra generated by the random variable X1 + X2 .)
A discrete random variable X is measurable with respect to F, or
F-measurable, if every event [X = x] is contained in F; in other words,
knowing only where we are in F, we can compute exactly the value of X.
This gives a formal way to define (X): it is the smallest -algebra F such
that X is F-measurable.
If X is not F-measurable, the best approximation we can make to it given
that we only know where we are in F is E [X | F], which is defined as a random
variable Q that is (a) F-measurable; and (b) satisfies E [Q | A] = E [X | A]
for any event A F with Pr [A] 6= 0.
For discrete probability spaces, this just means that we replace X with
its average value across each equivalence class: property (a) is satisfied
because E [X | F] is constant across each equivalence class, meaning that
[E [X | F] = x] is a union of equivalence classes, and property (b) is satisfied
because we define E [E [X | F] | A] = E [X | A] for each equivalence class A,
and the same holds for unions of equivalence classes by a simple calculation.
This gives the same result as E [X | Y ] if F is generated by Y , or more
generally as E [X | Y1 , Y2 , . . .] if F is generated by Y1 , Y2 , . . . . In each case
the intuition is that we are getting the best estimate we can for X given the
information we have. It is also possible to define E [X | F] as a projection
onto the subspace of all random variables that are F-measurable, analogously
to the special case for E [X | Y ] described in 3.5.2.3.
9
Proof: Let F be a -algebra over a countable set . Let 0 if, for all A in F,
A if and only if 0 A; this is an equivalence relation on . To show that the
equivalence classes of are elementsTof F, for each 00 6 , let A00 be some element of F
that contains but not 00 . Then 00 A00 (a countable intersection of elements of F)
contains and all points 0 but no points 00 6 ; in other words, its the equivalence
class of . Since there are only countably many such equivalence classes, we can construct
all the elements of F by taking all possible unions of them.
Sometimes it is convenient to use more than one -algebra to represent

increasing knowledge over time. A filtration is a sequence of -algebras
F0 F1 F2 . . . , where each Ft represents the information we have
available at time t. That each Ft is a subset of Ft+1 means that any event
we can determine at time t we can also determine at all future times t0 > t:
though we may learn more information over time, we never forget what
we already know. A common example of a filtration is when we have a
sequence of random variables X1 , X2 , . . . , and define Ft as the -algebra
hX1 , X2 , . . . , Xt i generated by X1 , X2 , . . . , Xt .
When one -algebra is a subset of another, a version of the law of iterated
expectation applies: F F 0 implies E [E [X | F 0 ] | F] = E [X | F]. One way
to think about this is that if we forget everything about X we cant predict
from F 0 and then forget everything thats left that we cant predict from
F, we get to the same place as if we just forget everything except F to
begin with. The simplest version E [X | F 0 ] = E [X] is just what happens
when F is the trivial -algebra {, }, where all we know is that something
happened, but we dont know what.
3.5.4 Examples
Let X be the value of a six-sided die. Let A be the event that X is
even. Then
X
E [X | A] = x Pr [X = x | A]
x
1
= (2 + 4 + 6)
3
= 4.
Let X and Y be independent six-sided dice, and let Z = X + Y . Then

E [Z | X] is a random variable whose value is 1 + 7/2 when X = 1,
2 + 7/2 when X = 2, etc. We can write this succinctly by writing
E [Z | X] = X + 7/2.
Conversely, if X, Y , and Z are as above, we can also compute E [X] Z.

Here we are told what Z is and must make an estimate of X.
For some values of Z, this nails down X completely: E [X | Z = 2] = 1
because X you can only make 2 in this model as 1 + 1. For other values,
we dont know much about X, but can still compute the expectation.
For example, to compute E [X | Z = 5], we have to average X over all
the pairs (X, Y ) that sum to 5. This gives E [X | Z = 5] = 14 (1 + 2 + 3 +
4) = 52 . (This is not terribly surprising, since by symmetry E [Y | Z = 5]

should equal E [X | Z = 5], and since conditional expectations add
just like regular expectations, wed expect that the sum of these two
expectations would be 5.)
The actual random variable E [X | Z] summarizes these conditional
expectations for all events of the form [Z = z]. Because of the symmetry
argument above, we can write it succinctly as E [X | Z] = Z2 . Or we
could list its value for every in our underlying probability space, as
in done in Table 3.2 for this and various other conditional expectations
on the two-independent-dice space.
3.6 Applications
3.6.1 Geometric random variables
Suppose that we are running a Las Vegas algorithm that takes a fixed
amount of time T , but succeeds only with probability p (which we take to
be independent of the outcome of any other run of the algorithm). If the
algorithm fails, we run it again. How long does it take on average to get the
algorithm to work?
We can reduce the problem to computing E [T X] = T E [X], where X
is the number of times the algorithm runs. The probability that X = n is
exactly (1p)n1 p, because we need to get n1 failures with probability 1p
each followed by a single success with probability p, and by assumption all of
these probabilities are independent. A variable with this kind of distribution
is called a geometric random variable. We saw a special case of this
distribution earlier (2.3.1.1) when we were looking at how long it took to
get a tails out of a fair coin (in that case, p was 1/2).
Using conditional expectation, its straightforward to compute E [X]. Let
A be the event that the algorithm succeeds on the first run, i.e., then event
[X = 1]. Then
h i h i
E [X] = E [X | A] Pr [A] + E X A Pr A

h i
= 1 p + E X A (1 p).

h i
The tricky part here is to evaluate E X A . Intuitively, if we dont succeed

the first time, weve wasted one step and are back where we started, so it
X Y Z =X +Y E [X] E [X | Y = 3] E [X | Y ] E [Z | X] E [X | Z] E [X | X]
(1, 1) 1 1 2 7/2 7/2 7/2 1 + 7/2 2/2 1
(1, 2) 1 2 3 7/2 7/2 7/2 1 + 7/2 3/2 1
(1, 3) 1 3 4 7/2 7/2 7/2 1 + 7/2 4/2 1
(1, 4) 1 4 5 7/2 7/2 7/2 1 + 7/2 5/2 1
(1, 5) 1 5 6 7/2 7/2 7/2 1 + 7/2 6/2 1
(1, 6) 1 6 7 7/2 7/2 7/2 1 + 7/2 7/2 1
(2, 1) 2 1 3 7/2 7/2 7/2 2 + 7/2 3/2 2
(2, 2) 2 2 4 7/2 7/2 7/2 2 + 7/2 4/2 2
(2, 3) 2 3 5 7/2 7/2 7/2 2 + 7/2 5/2 2
(2, 4) 2 4 6 7/2 7/2 7/2 2 + 7/2 6/2 2
(2, 5) 2 5 7 7/2 7/2 7/2 2 + 7/2 7/2 2
(2, 6) 2 6 8 7/2 7/2 7/2 2 + 7/2 8/2 2
(3, 1) 3 1 4 7/2 7/2 7/2 3 + 7/2 4/2 3
(3, 2) 3 2 5 7/2 7/2 7/2 3 + 7/2 5/2 3
(3, 3) 3 3 6 7/2 7/2 7/2 3 + 7/2 6/2 3
(3, 4) 3 4 7 7/2 7/2 7/2 3 + 7/2 7/2 3
(3, 5) 3 5 8 7/2 7/2 7/2 3 + 7/2 8/2 3
(3, 6) 3 6 9 7/2 7/2 7/2 3 + 7/2 9/2 3
(4, 1) 4 1 5 7/2 7/2 7/2 4 + 7/2 5/2 4
(4, 2) 4 2 6 7/2 7/2 7/2 4 + 7/2 6/2 4
(4, 3) 4 3 7 7/2 7/2 7/2 4 + 7/2 7/2 4
(4, 4) 4 4 8 7/2 7/2 7/2 4 + 7/2 8/2 4
(4, 5) 4 5 9 7/2 7/2 7/2 4 + 7/2 9/2 4
(4, 6) 4 6 10 7/2 7/2 7/2 4 + 7/2 10/2 4
(5, 1) 5 1 6 7/2 7/2 7/2 5 + 7/2 6/2 5
(5, 2) 5 2 7 7/2 7/2 7/2 5 + 7/2 7/2 5
(5, 3) 5 3 8 7/2 7/2 7/2 5 + 7/2 8/2 5
(5, 4) 5 4 9 7/2 7/2 7/2 5 + 7/2 9/2 5
(5, 5) 5 5 10 7/2 7/2 7/2 5 + 7/2 10/2 5
(5, 6) 5 6 11 7/2 7/2 7/2 5 + 7/2 11/2 5
(6, 1) 6 1 7 7/2 7/2 7/2 6 + 7/2 7/2 6
(6, 2) 6 2 8 7/2 7/2 7/2 6 + 7/2 8/2 6
(6, 3) 6 3 9 7/2 7/2 7/2 6 + 7/2 9/2 6
(6, 4) 6 4 10 7/2 7/2 7/2 6 + 7/2 10/2 6
(6, 5) 6 5 11 7/2 7/2 7/2 6 + 7/2 11/2 6
(6, 6) 6 6 12 7/2 7/2 7/2 6 + 7/2 12/2 6
Table 3.2: Various conditional expectations on two independent dice

h i
should be the case that E X A = 1 + E [X]. If we want to be really careful,

we can calculate this out formally (no sensible person would ever do this):
h i
X
E XA = n Pr [X = n | X 6= 1]
n=1

X Pr [X = n]
= n
n=2
Pr [X 6= 1]

X (1 p)n1 p
= n
n=2
1p
X
= n(1 p)n2 p
n=2
X
= (n + 1)(1 p)n1 p
n=1

X
=1+ n(1 p)n1 p
n=1
= 1 + E [X] .
Since we know that E [X] = p + (1 + E [X])(1 p), a bit of algebra gives

E [X] = 1/p, which is about what wed expect.
There are more direct ways to get the same result. If we dont have
conditional expectation to work with, we can try computing the sum E [X] =
P n1 p directly. The easiest way to do this is probably to use
n=1 n(1 p)
generating functions (see, for example, [GKP88, Chapter 7] or [Wil06]).
An alternative argument is given in [MU05, 2.4]; this uses the fact that
E [X] = n=1 Pr [X n], which holds when X takes on only non-negative
P
integer values.
3.6.2 Coupon collector

In the coupon collector problem, we throw balls uniformly and indepen-
dently into n bins until every bin has at least one ball. When this happens,
how many balls have we used on average?10
Let Xi be the number of balls needed to go from i 1 nonempty bins
to i nonempty bins. Its easy to see that X1 = 1 always. For larger i, each
10
The name comes from the problem of collecting coupons at random until you have all
of them. A typical algorithmic application is having a cluster of machines choose jobs to
finish at random from some list until all are done. The expected number of job executions
to complete n jobs is given exactly by the solution to the coupon collector problem.
time we throw a ball, it lands in an empty bin with probability ni+1

n . This
ni+1
means that Xi has a geometric distribution with probability n , giving
n
E [Xi ] = ni+1 from the analysis in 3.6.1.
To get the total expected number of balls, take the sum
" n # n
X X
E Xi = E [Xi ]
i=1 i=1
n
X n
=
i=1
ni+1
n
X 1
=n
i=1
i
= nHn .
In asymptotic terms, this is (n log n).
3.6.3 Hoares FIND

Hoares FIND [Hoa61b], often called QuickSelect, is an algorithm for
finding the k-th smallest element of an unsorted array that works like
QuickSort, only after partitioning the array around a random pivot we
throw away the part that doesnt contain our target and recurse only on
the surviving piece. As with QuickSort, wed like to compute the expected
number of comparisons used by this algorithm, on the assumption that the
cost of the comparisons dominates the rest of the costs of the algorithm.
Here the indicator-variable trick gets painful fast. It turns out to be
easier to get an upper bound by computing the expected number of elements
that are left after each split.
First, lets analyze the pivot step. If the pivot is chosen uniformly, the
number of elements X smaller than the pivot is uniformly distributed in the
range 0 to n1. The number of elements larger than the pivot will be nX1.
In the worst case, we find ourselves recursing on the large pile always, giving
a bound on the number of survivors Y of Y max(X, n X + 1).
What is the expected value of Y ? By considering both ways the max can
go, we get E [Y ] = E [X | X > n X + 1] Pr [X > n X + 1]+E [n X + 1 | n X + 1 X].
For both conditional expectations we are choosing a value uniformly in either
the range d n1 n1
2 e to n 1 or d 2 e + 1 to n 1, and in either case the
expectation will be equal to the average of the two endpoints by symmetry.
So we get
n/2 + n 1 n/2 + n
E [Y ] Pr [X > n X + 1] + Pr [n X + 1 X]
2 2
3 1 3

= n Pr [X > n X + 1] + n Pr [n X + 1 X]
4 2 4
3
n.
4
Now let Xi be the number of survivors after i pivot steps. Note that
max(0, Xi 1) gives the number of comparisons at the following pivot step,
so that
P
i=0 Xi is an upper bound on the number of comparisons.
We have X0 = n, and from the preceding argument E [X1 ] (3/4)n. But
more generally, we can use the same argument to show that E [Xi+1 | Xi ]
(3/4)Xi , and by induction E [Xi ] (3/4)i n. We also have that Xj = 0 for
all j n, because we lose at least one element (the pivot) at each pivoting
step. This saves us from having to deal with an infinite sum.
Using linearity of expectation,
" # " n #
X X
E Xi = E Xi
i=0 i=0
n
X
= E [Xi ]
i=0
Xn
(3/4)i n
i=0
4n.
Chapter 4
Basic probabilistic
inequalities
Here were going to look at some inequalities useful for proving properties
of randomized algorithms. These come in two flavors: inequalities involving
probabilities, which are useful for bounding the probability that something
bad happens, and inequalities involving expectations, which are used to
bound expected running times. Later, in Chapter 5, well be doing both,
by looking at inequalities that show that a random variable is close to its
expectation with high probability.
4.1 Union bound (Booles inequality)

For any countable collection of events {Ai },
h[ i X
Pr Ai Pr [Ai ] . (4.1.1)
The direct way to prove this is to replace Ai with Bi = Ai \ i1

S
S S j=1 Ai .
Then Ai = Bi , but since the Bi are disjoint and each Bi is a subset of
the corresponding Ai , we get Pr [ Ai ] = Pr [ Bi ] = Pr [Bi ] Pr [Ai ].
S S P P
The typical use of the union bound is to show that if an algorithm can
fail only if various improbable events occur, then the probability of failure is
no greater than the sum of the probabilities of these events. This reduces
the problem of showing that an algorithm works with probability 1 to
constructing an error budget that divides the probability of failure among
all the bad outcomes.
46
CHAPTER 4. BASIC PROBABILISTIC INEQUALITIES 47
4.1.1 Applications
(See also the proof of Adlemans Theorem in Chapter 12.)
4.1.1.1 Balls in bins

Suppose we toss n balls into n bins. What is the likely value of the maximum
number of balls in any one bin?1
Consider all nk sets S of k balls. If we get at least k balls in some bin,
then one of these sets must all land in the same bin. Call the event that all
balls in S choose the same bin AS . The probability that AS occurs is exactly
nk+1 .
So the union bound says
" #
[ X
Pr AS Pr [AS ]
S S
!
n k+1
= n
k
nk k+1

n
k!
n
= .
k!
If we want this probability to be low, we should choose k so that k! n.
Stirlings formula says that k! 2k(k/e)k (k/e)k , which gives ln(k!)
k(ln k 1). If we set k = c ln n/ ln ln n, we get
c ln n
ln(k!) (ln c + ln ln n ln ln ln n 1)
ln ln n
c ln n

2
when n is sufficiently large.
It follows that the bound n/k! in this case is less than n/ exp(c ln n/2) =
n nc/2 = n1c/2 . For suitable choice of c we get a high probability that
every bin gets at most O(log n/ log log n) balls.
4.1.1.2 Independent sets

Given a graph G = (V, E) where |V | = n and |E| = m, mark a subset of

n/2 m vertices uniformly at random. The probability that any particular
1
Algorithmic version: we insert n elements into a hash table with n positions using a
random hash function. What is the maximum number of elements in any one position?

vertex is marked is then (n/2/ m)/n = 1/2 m, and the probability that

both endpoints of an edge are marked is (1/2 m) ((n/2/ m 1)/n <

1/(2 m))2 = 1/4m. So the probability that at least one of the edges has
two marked endpoints is at most m/4m = 1/4. We thus have a randomized
algorithm that outputs a set of size 2nm on average that is an independent
set with probability 3/4, without looking at any of the edges.
To see how this compares with more sensible algorithms, note that for
a graph with maximum degree , we can construct and independent set
n
of size +1 deterministically by marking some vertex, deleting all of its
neighbors, and repeating until we run out of vertices. qThe randomized
algorithm produces an independent set of size 2nnd = 2 n
in this case,
which is much worse.
4.2 Markovs inequality

This is the key tool for turning expectations of non-negative random variables
into (upper) bounds on probabilities. Used directly, it generally doesnt give
very good bounds, but it can work well if we apply it to E [f (X)] for a fast-
growing function f ; see Chebyshevs inequality (5.1) or Chernoff bounds
(5.2).
If X 0, then
E [X]
Pr [X ] . (4.2.1)

The proof is immediate from the law of total probability (2.3.3). We have
E [X] = E [X | X ] Pr [X ] + E [X | x < ] Pr [X < ]

Pr [X ] + 0 Pr [X < ]
= Pr [X ] ;
now solve for Pr [X ].

Markovs inequality doesnt work in reverse. For example, consider the
following game: for each integer k > 0, with probability 2k , I hgive youi 2k
dollars. Letting X be your payoff from the game, we have Pr X 2k =
P
2k = 2k ` = 2 . The right-hand side here is exactly what we
P
j=k `=0 2 2k
would get from Markovs inequality if E [X] = 2. But in this case, E [X] 6= 2;
in fact, the expectation of X is given by k k
P
k=1 2 2 , which diverges.
4.2.1 Applications
4.2.1.1 The union bound
Combining Markovs inequality with linearity of expectation and indicator
variables gives a succinct proof of the union bound:
h[ i hX i
Pr Ai = Pr 1Ai 1
hX i
E 1Ai
X
= E [1Ai ]
X
= Pr [Ai ] .
Note that for this to work for infinitely many events we need to use the fact
that 1Ai is always non-negative.
4.2.1.2 Fair coins

Flip n independent fair coins, and let S be the number of heads we get. Since
E [S] = n/2, we get Pr [S = n] = 1/2. This is much larger than the actual
value 2n , but its the best we can hope for if we only know E [S]: if we let
S be 0 or n with equal probability, we also get E [S] = n/2.
4.2.1.3 Randomized QuickSort

The expected running time for randomized QuickSort is O(n log n). It follows
that the probability that randomized QuickSort takes more than f (n) time is
O(n log n/f (n)). For example, the probability that it performs the maximum
n 2
2 = O(n ) comparisons is O(log n/n). (Its possible to do much better
than this.)

Suppose we toss m balls in n bins, uniformly and independently. What
is the probability that some particular bin contains at least k balls? The
probability that a particular ball lands in a particular bin is 1/n, so the
expected number of balls in the bin is m/n. This gives a bound of m/nk
that a bin contains k or more balls. We can combine this with the union
bound to show that the probability that any bin contains k or more balls is
at most m/k. Unfortunately this is not a very good bound.
4.3 Jensens inequality

This is mostly useful if we can calculate E [X] easily for some X, but what
we really care about is some other random variable Y = f (X).
Jensens inequality applies when f is a convex function, which means
that for any x, y, and 0 1, f (x + (1 )y) f (x) + (1 )f (y).
Geometrically, this means that the line segment between any two points on the
graph of f never goes below f ; i.e., that the set of points {(x, y) | y f (x)}
If we want to show that f is convex, its enough to show that that
is convex.
f x+y 2 f (x)+f
2
(y)
for all x and y (in effect, we only need to prove it for
the case = 1/2). If f is twice-differentiable, an even easier way is to show
that f 00 (x) 0 for all x.
The inequality says that if X is a random variable and f is convex then
f (E [X]) E [f (X)] . (4.3.1)
Alternatively, if f is concave (which means that f (x + (1 )y)

f (x) + (1 )f (y), or equivalently that f is convex), the reverse inequality
holds:
f (E [X]) E [f (X)] . (4.3.2)
In both cases, the direction of Jensens inequality matches the direction

of the inequality in the definition of convexity or concavity. This is not
surprising because convexity or or concavity is just Jensens inequality for
the random variable X that equals x with probability and y with probability
1. Jensens inequality just says that this continues to work for any random
variable for which the expectations exist.
4.3.0.1 Proof
Here is a proof for the case that f is convex and differentiable. The idea is
that if f is convex, then it lies above the tangent line at E [X]. So we can
define a linear function g that represents this tangent line, and get, for all x:
f (x) g(x) = f (E [X]) + (x E [X])f 0 (E [X]). (4.3.3)

f (x)
E [f (X)]
g(x)
f (E [X])
E [X]
Figure 4.1: Proof of Jensens inequality
But then
E [f (X)] E [g(X)]
= E f (E [X]) + (X E [X])f 0 (E [X])

= f (E [X]) + E [(X E [X])] f 0 (E [X])

= f (E [X]) + (E [X] E [X])f 0 (E [X])
= f (E [X]).
Figure 4.1 shows what this looks like for a particular convex f .
This is pretty much all linearity of expectation in action: E [X], f (E [X]),
and f 0 (E [X]) are all constants, so we can pull them out of expectations
whenever we see them.
The proof of the general case is similar, but for a non-differentiable convex
function it takes a bit more work to show that the bounding linear function
g exists.
4.3.1 Applications
4.3.1.1 Fair coins: lower bound
Suppose we flip n fair coins, and we want to get a lower bound on E X 2 ,

where X is the number of heads. The function f : x 7 x2 is convex (take

2
its second derivative), so (4.3.1) gives E X 2 (E [X])2 = n4 . (The actual

2
value for E X 2 is n4 + n4 , which can be found using generating functions.2 )

2
Heres how: The probability generating function for X is F (z) = E z k =
z k Pr [X = k] = 2n (1 + z)n . Then zF 0 (z) = 2n nz(1 + z)n1 = kz k Pr [X = k].
P P
k
4.3.1.2 Fair coins: upper bound

For an upper bound, we can choose a concave f . For example, if X is as in
the previous example, E [lg X] lg E [X] = lg n2 = lg n 1. This is probably
pretty close to the exact value, as we will see later that X will almost always
be within a factor of 1 + o(1) of n/2. Its not a terribly useful upper bound,
because if we use it with (say) Markovs inequality, the most we can prove
is that Pr [X = n] = Pr [lg X = lg n] lglgn1 1
n = 1 lg n , which is an even
worse bound than the 1/2 we can get from applying Markovs inequality to
X directly.
4.3.1.3 Sifters
Heres an example of Jensens inequality in action in the analysis of an actual
distributed algorithm. For some problems in distributed computing, its
useful to reduce coordinating a large number of processes to coordinating
a smaller number. A sifter [AA11] is a randomized mechanism for an
asynchronous shared-memory system that sorts the processes into winners
and losers, guaranteeing that there is at least one winner. The goal is to
make the expected number of winners as small as possible. The problem
is tricky, because processes can only communicate by reading and writing
shared variables, and an adversary gets to choose which processes participate
and fix the schedule of when each of these processes perform their operations.
The current best known sifter is due to Giakkoupis and Woelfel [GW12].
For n processes, it uses an array A of dlg ne bits, each of which can be read
or written by any of the processes. When a process executes the sifter, it
chooses a random index r 1 . . . dlg ne with probability 2r1 (this doesnt
exactly sum to 1, so the excess probability gets added to r = dlg ne). The
process then writes a 1 to A[r] and reads A[r + 1]. If it sees a 0 in its read
(or chooses r = dlg ne), it wins; otherwise it loses.
This works as a sifter, because no matter how many processes participate,
some process chooses a value of r at least as large as any other processs
value, and this process wins. To bound the expected number of winners, take
the sum over all r over the random variable Wr representing the winners
who chose this particular value r. A process that chooses r wins if it carries
Taking the
Pderivative of this a second time gives 2n n(1 + z)n1 + 2n n(n 1)z(1
+
n2 2 k1 2
z) = k
k z Pr [X = k]. Evaluate this monstrosity at z = 1 to get E X =
2
k2 Pr [X = k] = 2n n2n1 + 2n n(n 1)2n2 = n/2 + n(n 1)/4 = 2n+n4 n =
P
k
2
n /4 + n/4. This is pretty close to the lower bound we got out of Jensens inequality, but
we cant count on this happening in general.
out its read operation before any process writes r + 1. If the adversary wants
to maximize the number of winners, it should let each process read as soon
as possible; this effectively means that a process that choose r wins if no
process previously chooses r + 1. Since r is twice as likely to be chosen as
r + 1, conditioning on a process picking r or r + 1, there is only a 1/3 chance
that it chooses r + 1. So at most 1/(1/3) 1 = 2 = O(1) process on average
choose r before some process chooses r + 1. (A simpler argument shows that
the expected number of processes that win because they choose r = dlg ne is
at most 2 as well.)
Summing Wr 2 over all r gives at most 2dlg ne winners on average.
Furthermore, if k < n processes participate, essentially the same analysis
shows that only 2dlg ke processes win on average. So this is a pretty effective
tool for getting rid of excess processes.
But it gets better. Suppose that we take the winners of one sifter and
feed them into a second sifter. Let Xk be the number of processes left after
k sifters. We have that X0 = n and E [X1 ] 2dlg ne, but what can we
say about E [X2 ]? We can calculate E [X2 ] = E [E [X2 | X1 ]] 2dlg X1 e.
Unfortunately, the ceiling means that 2dlg xe is not a concave function,
but f (x) = 2(lg x + 1) 2dlg xe is. So E [X2 ] f (f (n)), and in general
E [Xi ] f (i) (n), where f (i) is the i-fold composition of f . All the extra
constants obscure what is going on a bit, but with a little bit of algebra it is
not too hard to show that f (i) (n) = O(1) for i = O(log n).3 So this gets rid
of all but a constant number of processes very quickly.
3
The log function counts how many times you need to hit n with lg to reduce it to
one or less. So log 1 = 0, log 2 = 1, log 4 = 2, log 16 = 3, log 65536 = 4, log 265536 = 5,
and after that it starts getting silly.
Chapter 5
Concentration bounds
If we really want to get tight bounds on a random variable X, the trick will
turn out to be picking some non-negative function f (X) where (a) we can
calculate E [f (X)], and (b) f grows fast enough that merely large values of
X produce huge values of f (X), allowing us to get small probability bounds
by applying Markovs inequality to f (X). This approach is often used to
show that X lies close to E [X] with reasonably high probability, what is
known as a concentration bound.
Typically concentration bounds are applied to sums of random variables,
which may or may not be fully independent. Which bound you may want to
use often depends on the structure of your sum. A quick summary of the
bounds in this chapter is given in Table 5.1. The rule of thumb is to use
Chernoff bounds (5.2) if you have a sum of independent 01 random variables;
the Azuma-Hoeffding inequality (5.3) if you have bounded variables with a
more complicated distribution that may be less independent; and Chebyshevs
inequality (5.1) if nothing else works but you can somehow compute the
variance of your sum (e.g., if the Xi are independent or have easily computed
covariance). In the case of Chernoff bounds, you will almost always end up
using one of the weaker but cleaner versions in 5.2.2 rather than the general
version in 5.2.1.
5.1 Chebyshevs inequality

Chebyshevs inequality allows us to show that a random variable is
close to its mean, by applying Markovs inequality to the variance of
2 = E X 2 2X E [X] + (E [X])2 =

X, given by Var [X] = E (X E [X])
E X 2 (E [X])2 . Its a fairly weak concentration bound, that is most useful

54
CHAPTER 5. CONCENTRATION BOUNDS 55
E[S]
e
Chernoff Xi {0, 1}, independent Pr [S (1 + ) E [S]] (1+)1+

t2
Azuma-Hoeffding |Xi | ci , martingale Pr [S t] exp 2 P c2i
Var[S]
Chebyshev Pr [|S E [S]| ] 2
P
Table 5.1: Concentration bounds for S = Xi (strongest to weakest)
when X is a sum of random variables with limited independence.

Using Markovs inequality, calculate
h i
Pr [|X E [X]| ] = Pr (X E [X])2 2
E (X E [X])2

2
Var [X]
= . (5.1.1)
2
5.1.1 Computing variance

At this point it would be reasonable to ask why we are going through
Var [X] = E (X E [X])2 rather than just using E [|X E [X]|]. The
reason is that Var [X] is usually easier to compute, especially if X is a sum.
In this section, well give some examples of computing variance, including
for various standard random variables that come up a lot in randomized
algorithms.
5.1.1.1 Alternative formula

The first step is to give an alternative formula for the variance that is more
convenient in some cases.
Expand
h i h i
E (X E [X])2 = E X 2 2X E [X] + (E [X])2
h i
= E X 2 2 E [X] E [X] + (E [X])2
h i
= E X 2 (E [X])2 .
This formula is easier to use if you are estimating the variance from a
P 2 P 2
sequence of samples; by tracking xi and xi , you can estimate E X
and E [X] in a single pass, without having to estimate E [X] first and then go
back for a second pass to calculate (xi E [X])2 for each sample. We wont
use this particular application much, but this explains why the formula is
popular with statisticians.
5.1.1.2 Variance of a Bernoulli random variable

Recall that a Bernoulli random variable is 1 with probability p and 0 with
probability q = 1 p; in particular, any indicator variable is a Bernoulli
random variable.
The variance of a Bernoulli random variable is easily calculated from the
alternative formula:
h i
Var [X] = E X 2 (E [X])2
= p p2
= p(1 p)
= pq.
5.1.1.3 Variance of a sum

P
If S = i Xi , then we can calculate
!2 " #!2
X X
Var [S] = E Xi E Xi
i i

XX XX
= E Xi Xj E [Xi ] E [Xj ]
i j i j
XX
= (E [Xi Xj ] E [Xi ] E [Xj ]) .
i j
For any two random variables X and Y , the quantity E [XY ]E [X] E [Y ]
is called the covariance of X and Y , written Cov [X, Y ]. If we take the
covariance of a variable and itself, covariance becomes variance: Cov [X, X] =
Var [X].
We can use Cov [X, Y ] to rewrite the above expansion as

" #
X X
Var Xi = Cov [Xi , Xj ] (5.1.2)
i i,j
X X
= Var [Xi ] + Cov [Xi , Xj ] (5.1.3)
i i6=j
X X
= Var [Xi ] + 2 Cov [Xi , Xj ] (5.1.4)
i i<j
Note that Cov [X, Y ] = 0 when X and Y are independent; this makes
Chebyshevs inequality particularly useful for pairwise-independent ran-
dom variables, because then we can just sum up the variances of the individual
variables.
P
A typical application is when we have a sum S = Xi of non-negative
random variables with small covariance; here applying Chebyshevs inequality
to S can often be used to show that S is not likely to be much smaller than
E [S], which can be handy if we want to show that some lower bound holds
on S with some probability. This complements Markovs inequality, which
can only be used to get upper bounds.
Pn
For example, suppose S = i=1 Xi , where the Xi are independent
Bernoulli random variables with E [Xi ] = p for all i. Then E [S] = np, and
P
Var [S] = i Var [Xi ] = npq (because the Xi are independent). Chebyshevs
inequality then says
npq
Pr [|S E [S]| ] 2 .

The highest variance is when p = 1/2. In this case, he probability that

S is more than n away from its expected value n/2 is bounded by 41 2 .
Well see better bounds on this problem later, but this may already be good
enough for many purposes.
More generally, the approach of bounding S from below by estimating
E [S] and either E S 2 or Var [S] is known as the second-moment method.

In some cases, tighter bounds can be obtained by more careful analysis.
5.1.1.4 Variance of a geometric random variable

Let X be a geometric random variable with parameter p as defined in
3.6.1, so that X takes on the values 1, 2, . . . and Pr [X = n] = q n1 p, where
q = 1 p as usual. What is Var [X]?
We know that E [X] = 1/p, so (E [X])2 = 1/p2 . Computing E X 2 is

trickier. Rather than do this directly from the definition of expectation, we

can exploit the memorylessness of geometric random variables to get it using

conditional expectations, just like we did for E [X] in 3.6.1.
Conditioning on the two possible outcomes of the first trial, we have
h i h i
E X2 = p + q E X2 X > 1 . (5.1.5)

We now argue that E X 2 X > 1 = E (X + 1)2 . The intuition is

that once we have flipped one coin the wrong way, we are back where
we started, except that now we have to add that extra coin to X. More
Pr[X 2 =n] n1
formally, we have, for n > 1, Pr X 2 = n X > 1 = Pr[X>1] = q q p =

q n2 p = Pr [X = n 1] = Pr [X + 1 = n]. So we get the same probability

mass function for X conditioned on X > 1 as for X + 1 with no conditioning.
Applying this observation to the right-hand side of (5.1.5) gives
h i h i
E X 2 = p + q E (X + 1)2
h i
= p + q E X 2 + 2 E [X] + 1
2q h i
= p + q E X2 + +q
p
h i 2q
= 1 + q E X2 + .
p
A bit of algebra turns this into
h i 1 + 2q/p
E X2 =
1q
1 + 2q/p
=
p
p + 2q
=
p2
2p
= .
p2
Now subtract (E [X])2 = 1

p2
to get
1p q
Var [X] = 2
= 2. (5.1.6)
p p
By itself, this doesnt give very good bounds on X. For example, if we
want to bound the probability that X = 1, we get

1

Pr [X = 1] = Pr X E [X] = 1
p
1

Pr |X E [X]| 1
p
Var [X]
2
1
p 1
q/p2
= 2
1
1
p
q
=
(1 p)2
1
= .
q
Since 1q 1, we could have gotten this bound with a lot less work.
The other direction is not much better. We can easily calculate that
Pr [X n] is exactly q n1 (because this corresponds to flipping n coins
the wrong way, no matter what happens with subsequent coins). Using
Chebyshevs inequality gives

1 1

Pr [X n] Pr X n

p p
q/p2
2
n p1
q
= .
(np 1)2
This at least has the advantage of dropping below 1 when n gets large
enough, but its only polynomial in n while the true value is exponential.
Where this might be useful is in analyzing the sum of a bunch of geometric
random variables, as occurs in the Coupon Collector problem discussed in
3.6.2.1 Letting Xi be the number of balls to take us from i 1 to i empty
bins, we have previously argued that Xi has a geometric distribution with
1
We are following here a similar analysis in [MU05, 3.3.1].
ni1
p= n , so
, 2
i1 ni1
Var [Xi ] =
n n
i1
=n ,
(n i 1)2
and
" n # n
X X
Var Xi = Var [Xi ]
i=1 i=1
n
X i1
= n .
i=1 (n i 1)2
Having the numerator go up while the denominator goes down makes
this a rather unpleasant sum to try to solve directly. So we will follow the
lead of Mitzenmacher and Upfal and bound the numerator by n, giving
" n # n
X X n
Var Xi n
i=1 i=1 (n i 1)2
n
X 1
= n2
i=1
i2

X 1
n2
i=1
i2
2
= n2 .
6
2
The fact that 1
P
i=1 i2 converges to 6 is not trivial to prove, and was first
shown by Leonhard Euler in 1735 some ninety years after the question was
first proposed.2 But its easy to show that the series converges to something,
so even if we didnt have Eulers help, wed know that the variance is O(n2 ).
Since the expected value of the sum is (n log n), this tells us that we
are likely to see a total waiting time reasonably close to this; with at least
constant probability, it will be within (n) of the expectation. In fact, the
distribution is much more sharply concentrated (see [MU05, 5.4.1] or [MR95,
3.6.3]), but this bound at least gives us something.
2
See http://en.wikipedia.org/wiki/Basel_Problem for a history, or Eulers original
paper [Eul68], available at http://eulerarchive.maa.org/docs/originals/E352.pdf,
for the actual proof in the full glory of its original 18th-century typesetting. Curiously,
though Euler announced his result in 1735, he didnt submit the journal version until 1749,
and it didnt see print until 1768. Things moved more slowly in those days.
5.1.2 More examples

Here are some more examples of Chebyshevs inequality in action. Some
of these repeat examples for which we previously got crummier bounds in
4.2.1.
5.1.2.1 Flipping coins

Let X be the sum of n independent fair coins. Let Xi be the indicator
variable for the event that the i-th coin comes up heads. Then Var [Xi ] = 1/4
and Var [X] = Var [Xi ] = n/4. Chebyshevs inequality gives Pr [X = n]
P
n/4 1
Pr [|X n/2| n/2] (n/2) 2 = n . This is still not very good, but its
getting better. Its also about the best we can do given only the assumption
of pairwise independence.
To see this, let n = 2m 1 for some m, and let Y1 . . . Ym be independent,
fair 01 random variables. For each non-empty subset S of {1 . . . m}, let
XS be the exclusive OR of all Yi for i S. Then (a) the Xi are pairwise
independent; (b) each Xi has variance 1/4; and thus (c) the same Chebyshevs
P
inequality analysis for independent coin flips above applies to X = S XS ,
giving Pr [|X n/2| = n/2] n1 . In this case it is not actually possible for
X to equal n, but we can have X = 0 if all the Yi are 0, which occurs with
probability 2m = n+1 1
. So the Chebyshevs inequality is almost tight in this
case.

Let Xi be the indicator that the i-th of m balls lands in a particular bin.
Then E [Xi ] = 1/n, giving E [ Xi ] = m/n, and Var [Xi ] = 1/n 1/n2 ,
P
giving Var [ Xi ] = m/n m/n2 . So the probability that we get k + m/n

P
or more balls in a particular bin is at most (m/n m/n2 )/k 2 < m/nk 2 , and
applying the union bound, the probability that we get k + m/n or more balls
in any of the n bins is less than m/k 2 . Setting this equal to andpsolving for
k gives a probability of at most of getting more than m/n + m/ balls
in any of the bins. This is not as good a bound as we will be able to prove
later, but its at least non-trivial.
5.1.2.3 Lazy select

This is the example from [MR95, 3.3]; essentially the same example, spe-
cialized to finding the median, also appears in [MU05, 3.4].3
We want to find the k-th smallest element S(k) of a set S of size n. (The
parentheses around the index indicate that we are considering the sorted
version of the set S(1) < S(2) < S(n) .) The idea is to:
1. Sample a multiset R of n3/4 elements of S with replacement and sort

them. This takes O(n3/4 log n3/4 ) = o(n) comparisons so far.
2. Use our sample to find an interval that is likely to contain S(k) . The
idea is to pick indices ` = (k n3/4 )n1/4 and r = (k + n3/4 )n1/4 and
use R(`) and R(r) as endpoints (we are omitting some floors and maxes
here to simplify the notation; for a more rigorous presentation see
[MR95]). The hope is that the interval P = [R(`) , R(r)] in S will both
contain S(k) , and be small, with |P | 4n3/4 + 2. We can compute the
elements of P in 2n comparisons exactly by comparing every element
with both R(`) and R(r) .
3. If both these conditions hold, sort P (o(n) comparisons) and return

S(k) . If not, try again.
We want to get a bound on how likely it is that P either misses S(k) or
is too big.
For any fixed k, the probability that one sample in R is less than or
equal to S(k) is exactly k/n, so the expected number X of samples S(k)
is exactly kn1/4 . The variance on X can be computed by summing the
variances of the indicator variables that each sample is S(k) , which gives
a bound Var [X] =hn3/4 ((k/n)(1 k/n)) 3/4
i n /4. Applying Chebyshevs

inequality gives Pr X kn1/4 n n3/4 /4n = n1/4 /4.

Now lets look at the probability that P misses S(k) because R(`) is too

big, where ` = kn1/4 n. This is
h i h i
Pr R(`) > S(k) = Pr X < kn1/4 n
n1/4 /4.
3
The history of this example is that Motwani and Raghavan adapted this algorithm
from a similar algorithm by Floyd and Rivest [FR75]. Mitzenmacher and Upfal give a
version that also includes the adaptations appearing Motwani and Raghavan, although
they dont say where they got it from, and it may be that both textbook versions come
from a common folklore source.
(with the caveat that we are being sloppy about round-off errors).
Similarly,
h i h i
Pr R(h) < S(k) = Pr X > kn1/4 + n
n1/4 /4.
So the total probability that P misses S(k) is at most n1/4 /2.

Now we want to show that |P | is small. We will do so by showing that
it is likely that R(`) S(k2n3/4 ) and R(h) S(k+2n3/4 ) . Let X` be the
number of samples in R that are S(k2n3/4 ) and Xr be the number of

samples in R that are S(k+2n3/4 ) . Then we have E [X` ] = kn1/4 2 n

and E [Xr ] = kn1/4 + 2 n, and Var [Xl ] and Var [Xr ] are both bounded by
n3/4 /4.
We can now compute
h i
Pr R(l) < S(k2n3/4 ) = P r[Xl > kn1/4 n] < n1/4 /4
by the same Chebyshevs inequality argument

h as before,
i and get the sym-
metric bound on the other side for Pr R(r) > S(k+2n3/4 ) . This gives a total
bound of n1/4 /2 that P is too big, for a bound of n1/4 = o(n) that the
algorithm fails on its first attempt.
The total expected number of comparisons is thus given by T (n) =
2n + o(n) + O(n1/4 T (n)) = 2n + o(n).
5.2 Chernoff bounds

To get really tight bounds, we apply Markovs inequality to exp(S), where
P
S = i Xi . This works best when the Xi are independent: if this is
the case, so are the variables exp(Xi ), and so we can easily calculate
Q Q
E [exp(S)] = E [ i exp(Xi )] = i E [exp(Xi )].
The quantity E [exp(S)], treated as a function of , is called the
moment h generating
i function of S, because it expands formally into
P
E X k k , the exponential generating function for the series of
k=0 k! h i
k-th moments E X k . Note that it may not converge for all S and ;4 we
will be careful to choose for which it does converge and for which Markovs
inequality gives us good bounds.

4
For example, the moment generating function for our earlier bad X with Pr X = 2k =
k
P k k
2 is equal to k
2 e , which diverges unless e /2 < 1.
5.2.1 The classic Chernoff bound

The basic Chernoff bound applies to sums of independent 01 random vari-
ables, which need not be identically distributed. For identically distributed
random variables, the sum has a binomial distribution, which we can either
compute exactly or bound more tightly using approximations specific to
binomial tails; for sums of bounded random variables that arent necessarily
01, we can use Hoeffdings inequality instead (see 5.3).
Let each Xi for i = 1 . . . n be a 01 random variable with expectation pi ,
so that E [S] = = i pi . Recall that the plan is to show Pr [S (1 +
P
h )]
i
is small when and are large, by applying Markovs inequality to E eS ,
where will be chosen to make the bound as i as possible for given .
h tight
The first step is to get an upper bound on E eS .
Compute
h i h P i
E eS = E e Xi
Y h i
= E eXi
i
Y
= pi e + (1 pi )e0
i
Y
= (pi e + 1 pi )
i
Y
= (1 + (e 1) pi )
i
Y 1)p
e(e i
i
P
(e 1) i pi
=e
1)
= e(e .
The sneaky inequality step in the middle uses the fact that (1 + x) ex
for all x, which itself is one of the most useful inequalities you can memorize.5
Whats nice about this derivation is that at the end, the pi have vanished.
We dont care what random variables we started with or how many of them
there were, but only about their expected sumh. i
Now that we have an upper bound on E eS , we can throw it into
5
For a proof of this inequality, observe that the function f (x) = ex (1 + x) has the
derivative ex 1, which is positive for x > 0 and negative for x < 0. It follows that x = 1
is the unique minimum of f , at which f (1) = 0.
Markovs inequality to get the bound we really want:

h i
Pr [S (1 + )] = Pr eS e(1+)
h i
E eS

e(1+)

e(e 1)
(1+)
e !

ee 1
= .
e(1+)
1(1+)

= ee .
We now choose to minimize the base in the last expression, by minimiz-

ing its exponent e 1 (1 + ). Setting the derivative of this expression
with respect to to zero gives e = (1 + ) or = ln(1 + ); luckily, this
value of is indeed greater than 0 as we have been assuming. Plugging this
value in gives

Pr [S (1 + )] e(1+)1(1+) ln(1+)
!
e
= . (5.2.1)
(1 + )1+
The base of this rather atrocious quantity is e0 /11 = 1 at = 0, and its

derivative is negative for 0 (the easiest way to show this is to substitute
= x 1 first). So the bound is never greater than 1 and is both decreasing
and less than 1 as soon as > 0. We also have that the bound is exponential
in for any fixed .
If we look at the shape of the base as a function of , we can observe that
when is very large, we can replace (1 + )1+ with without changing the
bound much (and to the extent that we change it, its an increase, so it still

works as a bound). This turns the base into e = (e/) = 1/(/e) . This is

pretty close to Stirlings formula for 1/! (there is a 2 factor missing).
For very small , we have that 1 + e , so the base becomes approxi-
e 2
mately e(1+) = e . This approximation goes in the wrong direction (its
smaller than the actual value) but with some fudging we can show bounds
2
of the form e /c for various constants c, as long as is not too big.
5.2.2 Easier variants

The full Chernoff bound can be difficult to work with, especially since its
hard to invert (5.2.1) to find a good that gives a particular bound.
Fortunately, there are approximate variants that substitute a weaker but less
intimidating bound. Some of the more useful are:
For 0 1.81,
2 /3
Pr [X (1 + )] e . (5.2.2)
(The actual upper limit is slightly higher.) Useful for small val-
ues of , especially because the bound can be inverted: ifpwe want
Pr [X (1 + )] exp( 2 /3) , we can use any with 3 ln(1/)/
1.81. The essential idea to the proof is to show that, in the given
range, e /(1 + )1+ exp( 2 /3). This is easiest to do numerically;
a somewhat more formal argument that the bound holds in the range
0 1 can be found in [MU05, Theorem 4.4].
For 0 4.11,
2 /4
Pr [X (1 + )] e . (5.2.3)
This is a slightly weaker bound than the previous

p that holds over a
larger range. It gives Pr [X (1 + )] if 4 ln(1/)/ 4.11.
Note that the version given on page 72 of [MR95] is not correct; it
claims that the bound holds up to = 2e 1 4.44, but it fails
somewhat short of this value.
For R 2e,
Pr [X R] 2R . (5.2.4)
Sometimes the assumption is replaced with the stronger R 6 (this

is the version given in [MU05, Theorem 4.4], for example); one can also
verify numerically that R 5 (i.e., 4) is enough. The proof of the
2e bound is that e /(1+)(1+) < e1+ /(1+)(1+) = (e/(1+))1+
2(1+) when e/(1 + ) 1/2 or 2e 1. Raising this to gives
Pr [X (1 + )] 2(1+) for 2e 1. Now substitute R for
(1 + ) (giving R 2e) to get the full result. Inverting this one gives
Pr [X R] when R min(2e, lg(1/)).
Figure 5.1 shows the relation between the various bounds, in the region
where they cross each other.
2R/
4.11+
e
(1+)(1+)
1.81+ R/ 4.32
2 /4
e
2 /3 e
Figure 5.1: Comparison of Chernoff bound variants with exponent omitted,

plotted in logarithmic scale relative to the standard bound. Each bound is
valid only in in the region where it exceeds e /(1 + )1+ .
5.2.3 Lower bound version

We can also use Chernoff bounds to show that a sum of independent 01
random variables isnt too small. The essential idea is to repeat the up-
per bound argument with a negative value of , which makes e(1) an
increasing function in . The resulting bound is:
!
e
Pr [S (1 )] . (5.2.5)
(1 )1
A simpler but weaker version of this bound is

2 /2
Pr [S (1 )] e . (5.2.6)
Both bounds hold for all with 0 1.
5.2.4 Two-sided version

If we combine (5.2.2) with (5.2.6), we get
2 /3
Pr [|S | ] 2e , (5.2.7)
for 0 1.81.
Suppose that weqwant this bound to be less than . Then we need
2
2e /3 or 3 ln(2/)

. Setting to exactly this quantity, (5.2.7)
becomes q
Pr |S | 3 ln(2/) , (5.2.8)
provided 2e/3 .
For asymptotic purposes, we can omit the constants, giving
Lemma 5.2.1. Let S be a sum of independent 01 E [S] = .

variables with
p
Then for any 0 < 2e /3 , S lies within O log(1/) of , with
probability at least 1 .
5.2.5 Almost-independent variables

Chernoff bounds generally dont work very well for variables that are not
independent, and in most such cases we must use Chebyshevs inequality
(5.1) or the Azuma-Hoeffding inequality (5.3) instead. But there is one
special case that comes up occasionally where it helps to be able to apply
the Chernoff bound to variables that are almost independent in a particular
technical sense.
Lemma 5.2.2. Let X1 , . . . , Xn be 01 random variables with the property

that E [Xi | X1 , . . . , Xi1 ] pi 1 for all i. Let = ni=1 pi and S =
P
Pn
i=1 Xi . Then (5.2.1) holds.
Alternatively, let E [Xi | X1 , . . . , Xi1 ] pi 0 for all i, and let =
Pn Pn
i=1 pi and S = i=1 Xi as before. Then (5.2.5) holds.
Proof. Rather than repeat the argument for independent variables, we will
employ a coupling, where we replace the Xi with independent Yi so that
Pn Pn
i=1 Yi gives a bound an i=1 Xi .
For the upper bound, let each Yi = 1 with independent probability
pi . Use the following process to generate a new Xi0 in increasing order
of i: if Yi = 0, set Xi0 = 0. Otherwise set Xi0 = 1 with probability
Pr Xi = 1 X1 = X10 , . . . Xi1
0 /pi , Then Xi0 Yi , and
Pr Xi0 = 1 X10 , . . . , Xi0 = Pr Xi = 1 X1 = X10 , . . . , Xi1 = Xi1

0

/pi Pr [Yi = 1]
= Pr Xi = 1 X1 = X10 , . . . , Xi1 = Xi1
0

.
It follows that the Xi0 have the same joint distribution as the Xi , and so
" n # " n #
Xi0
X X
Pr (1 + ) = Pr Xi (1 + )
i=1 i=1
" n #
X
Pr Yi (1 + )
i=1
!
e
.
(1 + )1+
For the other direction, generate the Xi first and generate the Yi using
the same rejection sampling trick. Now the Yi are independent (because
their joint distribution is) and each Yi is a lower bound on the corresponding
Xi .
The lemma is stated for the general Chernoff bounds (5.2.1) and (5.2.5),
but the easier versions follow from these, so they hold as well, as long as we
are careful to remember that the in the upper bound is not necessarily the
same as in the lower bound.
5.2.6 Other tail bounds for the binomial distribution

The random graph literature can be a good source for bounds on the binomial
distribution. See for example [Bol01, 1.3], which uses normal approximation
to get bounds that are slightly tighter than Chernoff bounds in some cases,
and [JLR00, Chapter 2], which describes several variants of Chernoff bounds
as well as tools for dealing with sums of random variables that arent fully
independent.
5.2.7 Applications
5.2.7.1 Flipping coins
Suppose S is the sum of n independent fair coin-flips. Then E [S] = n/2 and
Pr [S = n] = Pr [S 2 E [S]] is bounded using (5.2.1) by setting = n/2,

= 1 to get Pr [S = n] (e/4)n/2 = (2/ e)n . This is not quite as good as

the real answer 2n (the quantity 2/ e is about 1.213. . . ), but its at least
exponentially small.
5.2.7.2 Balls in bins again

Lets try applying the Chernoff bound to the balls-in-bins problem. Here
we let S = m
P
i=1 Xi be the number of balls in a particular bin, with Xi
the indicator that the i-th ball lands in the bin, E [Xi ] = pi = 1/n, and
E [S] = = m/n. To get a bound on Pr [S m/n + k], apply the Chernoff
bound with = kn/m to get
Pr [S m/n + k] = Pr [S (m/n)(1 + kn/m)]

ek
.
(1 + kn/m)1+kn/m
For m = n, this collapses to the somewhat nicer but still pretty horrifying
ek /(k + 1)k+1 .
Staying with m = n, if we are bounding the probability of having large
bins, we can use the 2R variant to show that the probability that any
particular bin has more than 2 lg n balls (for example), is at most n2 , giving
the probability that there exists such a bin of at most 1/n. This is not as
strong as what we can get out of the full Chernoff bound. If we take the
logarithm of ek /(k + 1)k+1 , we get k (k + 1) ln(k + 1); if we then substitute
k = lnc lnlnnn 1, we get
c ln n c ln n c ln n
1 ln
ln ln n ln ln n ln ln n
c 1 c

= (ln n) (ln c + ln ln n ln ln ln n)
ln ln n ln n ln ln n
c 1 c ln c c ln ln ln n

= (ln n) c+
ln ln n ln n ln ln n ln ln n
= (ln n)(c + o(1)).
So the probability of getting more than c ln n/ ln ln n balls in any one bin

is bounded by exp((ln n)(c + o(1))) = nc+o(1) . This gives a maximum bin
size of O(log n/ log log n) with any fixed probability bound na for sufficiently
large n.
5.2.7.3 Flipping coins, central behavior

Suppose we flip n fair coins, and let S be the number that come up heads.
We expect = n/2 heads on average. How many extra heads can we get, if
we want to stay within a probability bound of nc ?
p Pr [S (1 +p)(n/2)]
Here we use the small- approximation, which gives
exp( 2 n/6). Setting exp( 2 n/6) = nc gives = 6 ln nc /n = 6c ln n/n.
110 111
010 011
100 101
000 001
Figure 5.2: Hypercube network with n = 3
q
The actual excess over the mean is (n/2) = (n/2) 6c ln n/n = 32 cn ln n.
p
By symmetry, the same bound applies to extra tails. So if we flip 1000

coins and see more than 676 heads (roughly the bound when c=3), we can
reasonably conclude that either (a) our coin is biased, or (b) we just hit a
rare one-in-a-billion jackpot. p
In algorithm analysis, the (3/2)c part usually gets absorbed into the
c
with probability at least 1 n ,
asymptotic notation, and we just say that
the sum of n random bits is within O( n log n) of n/2.
5.2.7.4 Permutation routing on a hypercube

Here we use Chernoff bounds to show bounds on a classic permutation-
routing algorithm for hypercube networks due to Valiant [Val82]. The
presentation here is based on 4.2 of [MR95], which in turn is based on an
improved version of Valiants original analysis appearing in a follow-up paper
with Brebner [VB81]. Theres also a write-up of this in [MU05, 4.5.1].
The basic idea of a hypercube architecture is that we have a collection of
N = 2n processors, each with an n-bit address. Two nodes are adjacent if
their addresses differ by one bit (see Figure 5.2 for an example). Though
now mostly of theoretical interest, these things were the cats pajamas back
in the 1980s: see http://en.wikipedia.org/wiki/Connection_Machine.
Suppose that at some point in a computation, each processor i wants to
send a packet of data to some processor (i), where is a permutation of
the addresses. But we can only send one packet per time unit along each of
the n edges leaving a processor.6 How do we route the packets so that all of
them arrive in the minimum amount of time?
We could try to be smart about this, or we could use randomization.
Valiants idea is to first route each process is packet to some random
intermediate destination (i), then in the second phase, we route it from (i)
to its ultimate destination (i). Unlike , is not necessarily a permutation;
instead, (i) is chosen uniformly at random independently of all the other
(j). This makes the choice of paths for different packets independent of
each other, which we will need later to apply Chernoff bounds..
Routing is done by a bit-fixing: if a packet is currently at node x and
heading for node y, find the leftmost bit j where xj = 6 yj and fix it, by
sending the packet on to x[xj /yj ]. In the absence of contention, bit-fixing
routes a packet to its destination in at most n steps. The hope is that the
randomization will tend to spread the packets evenly across the network,
reducing the contention for edges enough that the actual time will not be
much more than this.
The first step is to argue that, during the first phase, any particular
packet is delayed at most one time unit by any other packet whose path
overlaps with it. Suppose packet i is delayed by contention on some edge uv.
Then there must be some other packet j that crosses uv during this phase.
From this point on, j remains one step ahead of i (until its path diverges),
so it cant block i again unless both are blocked by some third packet k (in
which case we charge is further delay to k). This means that we can bound
the delays for packet i by counting how many other packets cross its path.7
So now we just need a high-probability bound on the number of packets that
get in a particular packets way.
Following the presentation in [MR95], define Hij to be the indicator
variable for the event that packets i and j cross paths during the first phase.
Because each j chooses its destination independently, once we fix is path,
P
the Hij are all independent. So we can bound S = j6=i Hij using Chernoff
bounds. To do so, we must first calculate an upper bound on = E [S].
The trick here is to observe that any path that crosses is path must cross
one of its edges, and we can bound the number of such paths by bounding
how many paths cross each edge. For each edge e, let Te be the number
of paths that cross edge e, and for each j, let Xj be the number of edges
P P
that path j crosses. Counting two ways, we have e Te = j Xj , and so
6
Formally, we have a synchronous routing model with unbounded buffers at each node,
with a maximum capacity of one packet per edge per round.
7
A much more formal version of this argument is given as [MR95, Lemma 4.5].
hP i
Xj N (n/2). By symmetry, all the Te have the same
P
E[ e Te ] =E j
expectation, so we get E [Te ] N N
(n/2)
n = 1/2.
Now fix (i). This determines some path e1 e2 . . . ek for packet i. In
general we do not expect E [Te` | (i)] to equal E [Te` ], because conditioning
on is path crossing e` guarantees that at least one path crosses this edge that
might not have. However, if we let Te0 be the number of packets j 6= i that
cross e, then we have Te0 Te always, giving E [Te0 ] E [Te ], and because Te0
does not depend on is path, E [Te0 | (i)] = E [Te0 ] E [Te ] 1/2. Summing
this bound over all k n edges on is path gives E [Hij | (i)] n/2, which
implies E [Hij ] n/2 after removing the conditioning on (i).
Inequality (5.2.4) says that Pr [X R] 2R when R 2e. Letting
X = j6=i Hij and setting R = 3n gives R = 6(n/2) 6 > 2e, so
P
hP i
3n = N 3 . This says that any one packet reaches its
Pr j Hij 3n 2
random destination with at most 3n added delay (thus, in at most 4n time
units) with probability at least 1 N 3 . If we consider all N packets, the
total probability that any of them fail to reach their random destinations in
4n time units is at most N N 3 = N 2 . Note that because we are using
the union bound, we dont need independence for this stepwhich is good,
because we dont have it.
What about the second phase? Here, routing the packets from the random
destinations back to the real destinations is just the reverse of routing them
from the real destinations to the random destinations. So the same bound
applies, and with probability at most N 2 some packet takes more than
4n time units to get back (this assumes that we hold all the packets before
sending them back out, so there are no collisions between packets from
different phases).
Adding up the failure probabilities and costs for both stages gives a
probability of at most 2/N 2 that any packet takes more than 8n time units
to reach its destination.
The structure of this argument is pretty typical for applications of Cher-
noff bounds: we get a very small bound on the probability that something
bad happens by applying Chernoff bounds to a part of the problem where
we have independence, then use the union bound to extend this to the full
problem where we dont.
5.3 The Azuma-Hoeffding inequality

The problem with Chernoff bounds is that they only work for Bernoulli
random variables. Hoeffdings inequality is another concentration bound
based on the moment generating function that applies to any sum of bounded
independent random variables with mean 0.8 It has the additional useful
feature that it generalizes nicely to some collections of random variables
that are not mutually independent, as we will see in 5.3.2. This more
general version is known as Azumas inequality or the Azuma-Hoeffding
inequality.9
5.3.1 Hoeffdings inequality

This is the version for sums of bounded independent random variables.
Theorem 5.3.1. Let X1 . . . Xn be independent random variables with E [Xi ] =

0 and |Xi | ci for all i. Then for all t,
" n # !
X t2
Pr Xi t exp Pn 2 . (5.3.1)
i=1
2 i=1 ci
Proof. Let S = ni=1 Xi . As with Chernoff hbounds,

P
i well first calculate a
bound on the moment generating function E e S and then apply Markovs
inequality with a carefully-chosen . h i
Before jumping into this, it helps to get a bound on E eX when
E [X] = 0 and |X| c. The trick is that, for any , ex is a convex function.
Since we want an upper bound, we cant use Jensens inequality (4.3.1), but
we can use the fact that X is bounded and we know its expectation. Convexity
of ex means that, for any x with c x c, ex ec + (1 )ec,
where x = (c) + (1 )c. Solving for in terms of x gives = 12 1 xc
8
Note that the requirement that E [Xi ] = 0 can always be satisfied by considering
instead Yi = Xi E [Xi ].
9
The history of this is that Hoeffding [Hoe63] proved it for independent random variables,
and observed that the proof was easily extended to martingales, while Azuma [Azu67]
actually went and did the work of proving it for martingales.
1 x

and 1 = 2 1+ c . So
1 X c 1 X c
h i
X
E e E 1 e + 1+ e
2 c 2 c
ec + ec ec ec
= E [X] + E [X]
2 2c 2c
ec + ec
=
2
= cosh(c).
In other words, the worst possible X is a fair choice between c, and

in this case we get the hyperbolic cosine of c as its moment generating
function.
We dont like hyperbolic cosines much, because we are going to want
to take products of our bounds, and hyperbolic cosines dont multiply very
nicely. As before with 1 + x, wed be much happier if we could replace the
cosh with a nice exponential. The Taylor series expansion of cosh x starts
with 1 + x2 /2 + . . . , suggesting that we should approximate it with exp(x2 /2),
2
and indeed it is the case that for all x, cosh x ex /2 . This can be shown by
comparing the rest of the Taylor series expansions:
ex + ex
cosh x =
2
1 xn X
!
X (x)n
= +
2 n=0 n! n=0 n!

X x2n
=
n=0
(2n)!

X x2n

n=0
2n n!

X (x2 /2)n
=
n=0
n!
x2 /2
=e .
h i 2 /2
This gives E eX e(c) .
Using this bound and the independence of the Xi , we compute

n
" !#
h i X
S
E e = E exp Xi
i=1
" n #
Y
=E eXi
i=1
n
Y h i
= E eXi .
i=1
n
Y 2 /2
e(ci )
i=1
n
!
X 2 c2i
= exp
i=1
2
n
!
2 X
= exp c2i .
2 i=1
Applying Markovs inequality then gives (when > 0):

h i
Pr [S t] = Pr eS et
n
!
2 X
exp c2 t . (5.3.2)
2 i=1 i
Now we do the same trick as in Chernoff bounds and choose to minimize

the bound. If we write C for ni=1 c2i , this is done by minimizing the exponent
P
2
2 C t, which we do by taking the derivative with respect to and setting
it to zero: C t = 0, or = t/C. At this point, the exponent becomes
(t/C)2 t2
2 C (t/C)t = 2C .
Plugging this into (5.3.2) gives the bound (5.3.1) claimed in the theorem.
Lets see how good a bound this gets us for our usual test problem of
bounding Pr [S = n] where S = ni=1 Xi is the sum of n independent fair
P
coin-flips. To make the problem fit the theorem, we replace each Xi by

a rescaled version Yi = 2Xi 1 = 1 with equal probability; this makes
E [Yi ] = 0 as needed, with |Yi | ci = 1. Hoeffdings inequality (5.3.1) then
gives
" n # !
X n2
Pr Yi n exp
i=1
2n

= en/2 = ( e)n .

Since e 1.649 . . . , this is actually slightly better than the (2/ e)n
bound we get using Chernoff bounds.
On the other hand, Chernoff bounds work better if we have a more
skewed distribution on the Xi ; for example, in the balls-in-bins case, each
Xi is a Bernoulli random variable with E [Xi ] = 1/n. Using Hoeffdings
inequality, we get a bound ci on |Xi E [Xi ]| of only 1 1/n, which puts
Pn 2
i=1 ci very close to n, requiring t = ( n) before we get any non-trivial
bound out of (5.3.1), pretty much the same as in the fair-coin case (which is
not surprising, since Hoeffdings inequality doesnt know anything about the
distribution of the Xi ). But weve already seen that Chernoff gives us that
P
Xi = O(log n/ log log n) with high probability in this case.
Note: There is an asymmetrical version of Hoeffdings inequality in which
ai Xi bi , but E [Xi ] is still zero for all Xi . In this version, the bound is
" n # !
X 2t2
Pr Xi t exp Pn 2
. (5.3.3)
i=1 i=1 (bi ai )
This reduces to (5.3.1) when ai = ci and bi = ci . The proof is essentially

the
h same,
i but a little more analytic sneakery is required to show that
X 2 (b a )2 /8
E e i e i i ; see [McD89] for details. For most applications, the
only difference between the symmetric version (5.3.1) and the asymmetric
version (5.3.3) is a small constant factor on the resulting bound on t.
Hoeffdings inequality as stated above is not the tightest possible, but is
relatively simple and easy to work with. For a particularly strong version of
Hoeffdings inequality and a discussion of some variants, see [FGQ12].
5.3.2 Azumas inequality

A general rule of thumb is that most things that work for sums of independent
random variables also work for martingales, which are similar collections of
random variables that allow for more dependence.
Formally, a martingale is a sequence of random variables S0 , S1 , S2 , . . . ,
where E [St | S1 , . . . , St1 ] = St1 . In other words, given everything you
know up until time t 1, your best guess of the expected value at time t is
just wherever you are now.
Another way to describe a martingale is to take the partial sums St =

Pt
i=1 Xt of a martingale difference sequence, which is a sequence of
random variables X1 , X2 , . . . where E [Xt | X1 . . . Xt1 ] = 0. So in this
version, your expected change from time t 1 to t averages out to zero, even
if you try to predict it using all the information you have at time t 1.
In some cases it makes sense to allow extra information to sneak in. We
can represent this using -algebras, in particular by using a filtration of the
form F0 F1 F2 . . . , where each Ft is a -algebra (see 3.5.3). A
sequence S0 , S1 , S2 , . . . is adapted to a filtration F0 F1 F2 . . . if each
St is Ft -measurable. This means that at time t the sum of our knowledge
(Ft ) is enough to predict exactly the value of St . The subset relations also
mean that we remember everything there is to know about St0 for t0 < t.
The general definition of a martingale is a collection {(St , Ft ) | t N}
where
1. Each St is Ft -measurable; and
2. E [St+1 | Ft ] = St .
This means that even if we include any extra information we might have
at time t, we still cant predict St+1 any better than by guessing the current
value St . This alternative definition will be important in some special cases,
as when St is a function of some other collection of random variables that
we use to define the Ft . Because Ft includes at least as much information as
S0 , . . . , St , it will always be the case that any sequence {(St , Ft )} that is a
martingale in the general sense gives a sequence {St } that is a martingale in
the more specialized E [St+1 | S0 , . . . , St ] = St sense.
Martingales were invented to analyze fair gambling games, where your
return over some time interval is not independent of previous outcomes (for
example, you may change your bet or what game you are playing depending
on how things have been going for you), but it is always zero on average
given previous information.10 The nice thing about martingales is they allow
for a bit of dependence while still acting very much like sums of independent
random variables.
Where this comes up with Hoeffdings inequality is that we might have a
process that is reasonably well-behaved, but its increments are not technically
independent. For example, suppose that a gambler plays a game where
10
Real casinos give negative expected return, so your winnings in a real casino form
a supermartingale with St E [St+1 | S0 . . . St ]. On the other hand, the casinos take,
in a well-run casino, is a submartingale, a process with St E [St+1 | S0 . . . St ]. These
definitions also generalize in the obvious way to the {(St , Ft )} case.
she bets x units 0 x 1 at each round, and receives x with equal

probability. Suppose also that her bet at each round may depend on the
outcome of previous rounds (for example, she might stop betting entirely
if she loses too much money). If Xi is her take at round i, we have that
E [Xi | X1 . . . Xi1 ] = 0 and that |Xi | 1. This is enough to apply the
martingale version of Hoeffdings inequality, often called Azumas inequality.
Pk
Theorem 5.3.2. Let {Sk } be a martingale with Sk = i=1 Xi and |Xi | ci
for all i. Then for all n and all t 0:
!
t2
Pr [Sn t] exp Pn 2 . (5.3.4)
2 i=1 ci
h i 2

Proof. Basically, we just show that E eSn exp 2 ni=1 c2i just like
P
in the proof of Theorem 5.3.1and the rest follows using the same argument.
The
hQonly trickyi part is weh cani no longer use independence to transform
n X into ni=1 E eXi .
Q
E i=1 e i
Instead, we use the martingale property.

h For i have E [Xi | X1 . . . Xi1 ] =
each Xi , we
0 and |Xi | ci always. Recall that E eXi X1 . . . Xi1 is a random vari-

able that takes on the average value of eXi for each setting of X1 . . . Xi1 .
We can applyhthe same analysisias in the proof of 5.3.1 to show that this
X 2
means that E e i
X1 . . . Xi1 e(ci ) /2 always.

The trick is to use the fact that, for any random variables X and Y ,
E [XY ] = E [E [XY | X]] = E [X E [Y | X]]. hQ i 2
We argue by induction on n that E ni=1 eXi ni=1 e(c) /2 . The
Q
base case is when n = 0. For the induction step, compute

" n # " " n ##
Y Y
E eXi =E E eXi X1 . . . Xn1

i=1 i=1
" n ! #
Y h i
Xi Xn
=E e E e X1 . . . Xn1
i=1
" n1 ! #
(cn )2 /2
Y
Xi
E e e
i=1
"n1 #
Y
Xi 2 /2
=E e e(cn )
i=1
n1
!
(ci )2 /2 2 /2
Y
e e(cn )
i=1
n
Y 2 /2
= e(ci ) .
i=1
n
!
2 X
= exp c2 .
2 i=1 i
The rest of the proof goes through as before.
Some extensions:
The asymmetric version of Hoeffdings inequality (5.3.3) also holds for
martingales. So if each increment Xi satisfies ai Xi bi always,
" n # !
X 2t2
Pr Xi t exp Pn 2
. (5.3.5)
i=1 i=1 (bi ai )
The same bound works for bounded-difference supermartingales. A

supermartingale is a process where E [Xi | X1 . . . Xi1 ] 0; the idea is
that my expected gain at any step is non-positive, so my present wealth
is always superior to my future wealth.11 If E [Xi | X1 . . . Xi1 ] 0 and
|Xi | ci , then we can write Xi = Yi +Zi where Yi = E [Xi | X1 . . . Xi1 ]
0 is predictable from X1 . . . Xi1 and E [Zi | X1 . . . Xi1 ] = 0.12 Then
we can bound ni=1 Xi by observing that it is no greater than ni=1 Zi .
P P
11
The corresponding notion in the other direction is a submartingale. See 8.2.
12
This is known as a Doob decomposition and can be used to extract a martingale
{Zi } from any stochastic process {Xi }. For general processes, Yi = Xi Zi will still be
predictable, but may not satisfy Yi 0.
A complication is that we no longer have |Zi | ci ; instead, |Zi | 2ci

(since leaving out Yi may shift Zi up). But with this revised bound,
(5.3.4) gives
" n # " n #
X X
Pr Xi t Pr Zi t
i=1 i=1
!
t2
exp Pn 2 . (5.3.6)
8 i=1 ci
Suppose that we stop the process after the first time with S =
P
i=1 Xi t. This is equivalent to making a new variable Yi that is
zero whenever Si1 t and equal to Xi otherwise. This doesnt affect
the conditions E [Yi | Y1 . . . Yi1 ] = 0 or |Yi | ci , but it makes it so
Pn
Yi t if and only if maxkn ki=1 Xi t. Applying (5.3.4) to
P
P i=1
Yi then gives
k
" # !
X t2
Pr max Xi t exp Pn 2 . (5.3.7)
kn
i=1
2 i=1 ci
Since the conditions on Xi in Theorem 5.3.2 apply equally well to Xi ,

we have
" n # !
X t2
Pr Xi t exp Pn 2 . (5.3.8)
i=1
2 i=1 ci
which we can combine with (5.3.4) to get the two-sided bound

" n # !
X t2
Pr Xi t 2 exp Pn 2 . (5.3.9)

i=1
2 i=1 ci
The extension of Hoeffdings inequality to the case ai Xi bi works

equally well for Azumas inequality, giving the same bound as in (5.3.3).
Finally, one can replace the requirement that each ci be a constant

with a requirement that ci be predictable from X1 . . . Xi1 and that
Pn 2 Pn t2 /2C . This generally
i=1 ci C always and get Pr [ i=1 Xi t] e
doesnt come up unless you have an algorithm that explicitly cuts off
the process if c2i gets too big, but there is at least one example of
P
this in the literature [AW96].

5.3.3 The method of bounded differences

To use Azumas inequality, we need a bounded-difference martingale. The
easiest way to get such martingales is through the method of bounded dif-
ferences, which was popularized by a survey paper by McDiarmid [McD89].
For this reason the key result is often referred to as McDiarmids inequal-
ity.
The basic idea of the method is to structure a problem so that we are
computing a function f (X1 , . . . , Xn ) of a sequence of independent random
variables X1 , . . . , Xn . To get our martingale, well imagine we reveal the Xi
one at a time, and compute at each step the expectation of the final value of
f based on just the inputs weve seen so far.
Formally, let Ft = hX1 , . . . , Xt i, the -algebra generated by X1 through
Xt . This represents all the information we have at time t. Let Yt = E [f | Ft ],
the expected value of f given the values of the first t variables. Then
{hYt , FT i} forms a martingale, with Y0 = E [f ] and Yn = E [f | X1 , . . . , Xt ] =
f .13 So if we can find a bound ct on Yt Yt1 , we can apply Azumas
inequality to get bounds on Yn Y0 = f E [f ].
We do this by assuming that f has the bounded difference property,
which says that there are bounds ct such that for any x1 . . . xn and any x0t ,
we have
f (x1 . . . xt . . . xn ) f (x1 . . . x0 . . . xn ) ct .

t (5.3.10)
Now suppose that f has the bounded difference property, and con-
sider the possible values for Yt+1 Yt , which is the change in the ex-
pected value of f when we reveal xt+1 . Fix some sequence of values
x1, . . . , xt for X1 , . . . , Xt , and let x+t+1 be the value of Xt+1 that maximizes
E f (x1 , . . . , xt , x+ t + 1, Xt+2 , . . . , Xn and x t+1 be the value of Xt+1 that
minimizes E [f (x1 , . . . , xt , x t + 1, Xt+2 , . . . , Xn ]. Let us suppose also that
each Xi is discrete; the general case will require a better proof, but this will
give us some idea of what is going on. Then bounded difference means that
for any choice of xt+2 , . . . , xn , we have

|f (x1 , . . . , xt , x+
t+1 , xt+2 , . . . , xn ) f (x1 , . . . , xt , xt+1 , xt+2 , . . . , xn )| ct+1 ,
which implies

f (x1 , . . . , xt , x+
t+1 , xt+2 , . . . , xn ) f (x1 , . . . , xt , xt+1 , xt+2 , . . . , xn ) ct+1 ,
13
A sequence of random variables of the form Yt = E [Z | Ft ], where F0 F1 F2 . . .
is a called a Doob martingale. The proof that it is a martingale is immediate from
the general version of the law of iterated expectation E [Yt+1 ] Ft = E [E [Z] Ft+1 Ft ] =
E [Z] Ft = Yt .
the same thing without the absolute value. But then when we take expecta-
tions, we get
h i h i

E f (x1 , . . . , xt , x+
t+1 , Xt+2 , . . . , Xn E f (x1 , . . . , xt , xt+1 , Xt+2 , . . . , Xn +ct+1 ,
since we are just averaging the difference over all choices of xt+2 , . . . , xn .
This means that, for any fixed value of x1 through xt , the maximum and
minimum values of Yt+1 differ by at most ct+1 .
Since Yt must lie somewhere in between these values, we get that |Yt+1
Yt | so we can apply Azuma-Hoeffding to get Pr [Yn Y0 t]
ct+1 always,
2
exp 2 Ptn c2
. But this overestimates the possible range of Yt+1 ; we have
i=1 i
in fact shown that for any fixed x1 , . . . , xt , there exists bounds at+1
Yt+1 Yt bt+1 such that bt+1 at+1 = ct+1 . We would like to use this
to apply the asymmetric version of Azuma-Hoeffding given in (5.3.5). A
complication is that the specific values of at+1 and bt+1 depend on the
previous values x1 , . . . , xt , even though their difference bt+1 at+1 does not.
Fortunately, McDiarmid shows that the inequality works anyway, giving:
Theorem 5.3.3 (McDiarmids inequality [McD89]). Let X1 , . . . , Xn be

independent random variables and let f (X1 , . . . , Xn ) have the bounded differ-
ence property with bounds ci . Then
!
2t2
Pr [f (X1 , . . . , Xn ) E [f (X1 , . . . , Xn )] t] exp Pn 2 . (5.3.11)
i=1 ci
Since f satisfies the same constraints as f , the same bound holds for
Pr [f (X1 , . . . , Xn ) E [f (X1 , . . . , Xn )] t].
5.3.4 Applications
Here are some applications of the preceding inequalities. Most of these are
examples of the method of bounded differences.
5.3.4.1 Sprinkling points on a hypercube

Suppose you live in a hypercube, and the local government has conveniently
placed mailboxes on some subset A of the nodes. If you start at a random
location, how likely is it that your distance to the nearest mailbox deviates
substantially from the average distance?
We can describe your position as a bit vector X1 , . . . , Xn , where each
Xi is an independent random bit. Let f (X1 , . . . , Xn ) be the distance from
X1 , . . . , Xn to the nearest element of A. Then changing one of the bits

changes this function by at most 1. So we have Pr [|f E [f ]| 2t2 /n
t] 2e
by (5.3.11), giving a range of possible distances that is O( n log n) with
probability at least 1 nc for any fixed c > 0.14 Of course, without knowing
what A is, we dont know what E[f ] is; but at least we can be assured that
(unless A is very big) the distance we have to walk to send our mail will be
pretty much the same pretty much wherever we start.
5.3.4.2 Chromatic number of a random graph

Consider a random graph G(n, p) consisting of n vertices, where each pos-
sible edge appears with independent probability p. Let be the chromatic
number of this graph, the minimum number of colors necessary if we want
to assign a color to each vertex that is distinct for the colors of all of its
neighbors. The vertex exposure martingale shows us the vertices of the
graph one at a time, along with all the edges between vertices that have
been exposed so far. We define Xt to be the expected value of given this
information for vertices 1 . . . t.
If Zi is a random variable describing which edges are present between i
and vertices less than i, then the Zi are all independent, and we can write
= f (Z1 , . . . , Zn ) for some function f (this function may not be very easy
to compute, but it exists). Then Xt as defined above is just E [f | Z1 , . . . , Zt ].
Now observe that f has the bounded difference property with ct = 1: if I
change the edges for some vertex vt , I cant increase the number of colors
I need by more than 1, since in the worst case I can always take whatever
coloring I previously had for all the other vertices and add a new color for
vt . (This implies I cant decrease by more than one either.) McDiarmids
2
inequality (5.3.11) then says that Pr [| E []| t] 2e2t /n ; in other
words, the chromatic number of a random graph is tightly concentrated
around its mean, even if we dont know what that mean is. (This proof is
due to Shamir and Spencer [SS87].)

Suppose we toss m balls into n bins. How many empty bins do we get?
The probability that each bin individually is empty is exactly (1 1/n)m ,
which is approximately em/n when n is large. So the expected number of
empty bins is exactly n(1 1/n)m . If we let Xi be the bin that ball i gets
p1 2
14
Proof: Let t = 2
(c + 1)n ln n = O( n log n). Then 2e2t /n = 2e(c+1) ln n =
c1 c
2n < n when n is sufficiently large.
tossed into, and let Y = f (X1 , . . . , Xm ) be the number of empty bins, then
changing a single Xi can change f by at most 1. So from (5.3.11) we have
2
Pr [Y n(1 1/n)m + t] e2t /m .
5.3.4.4 Probabilistic recurrence relations

Most probabilistic recurrence arguments (as in Appendix F) can be inter-
preted as supermartingales: the current estimate of T (n) is always at least
the expected estimate after doing one stage of the recurrence. This fact can
be used to get concentration bounds using (5.3.6).
For example, lets take the recurrence (1.3.1) for the expected number of
comparisons for QuickSort:
1 n1
X
T (n) = (T (k) + T (n 1 k)) .
n k=0
We showed in 1.3.1 that the solution to this recurrence satisfies T (n)
2n ln n.
To turn this into a supermartingale, imagine that we carry out a pro-
cess where we keep around at each step t a set of unsorted blocks of size
nt1 , nt2 , . . . , ntkt for some kt (note that the superscripts on nti are not expo-
nents). One step of the process involves choosing one of the blocks (we can
do this arbitrarily without affecting the argument) and then splitting that
block around a uniformly-chosen pivot. We will track a random variable Xt
equal to Ct + ki=1 2nti ln nti , where Ct is the number of comparisons done
P t
so far and the summation gives an upper bound on the expected number of
comparisons remaining.
To show that this is in fact a supermartingale, observe that if we partition
a block of size n we add n to Ct but replace the cost bound 2n ln n by an
expected
1 n1
Z n
X 4
2 2k ln k n ln n
n k=0 n 2
!
4 n2 ln n n2
= ln 2 + 1
n 2 4
= 2n ln n n ln 2 + 1
< 2n ln n n.
The net change is less than ln 2. The fact that its not zero suggests
that we could improve the 2n ln n bound slightly, but since its going down,
we have a supermartingale.
Lets try to get a bound on how much Xt changes at each step. The Ct
part goes up by at most n 1. The summation can only go down; if we split
a block of size ni , the biggest drop we get is if we split it evenly,15 This gives
a drop of
n1 n1 n1

2n ln n 2 2 ln = 2n ln n 2(n 1) ln n
2 2 2n
2n
= 2n ln n 2(n 1)(ln n ln )
n1
2n 2n
= 2n ln n 2n ln n + 2n ln + 2 ln n 2 ln
n1 n1
= 2n O(1) + O(log n)
= O(n).
(with a constant tending to 2 in the limit).

So we can apply (5.3.6) with ct = O(n) to the at most n steps of the
algorithm, and get
2 /O(n3 )
Pr [Cn 2n ln n t] et .

This gives Cn = O(n3/2 ) with constant probability or O(n3/2 log n) with
all but polynomial probability. This is a rather terrible bound, but its a lot
better than O(n2 ).
Much tighter bounds are known: QuickSort in fact uses (n log n) com-
parisons with high probability [MH92]. If we arent too worried about
constants, an easy way to see the upper bound side of this is to adapt the
analysis of Hoares FIND (3.6.3). For each element, the number of other
elements in the same block is multiplied by a factor of at most 3/4 on average
each time the element is compared, so the chance that the element is not
by itself is at most (3/4)k n after k comparisons. Setting k = log4/3 (n2 /)
gives that any particular element is compared k or times with probability at
most /n. The union bound then gives a probability of at most that the
most-compared element is compared k or more times. So the total number
of comparisons is O(log(n/)) with probability 1 , which becomes O(log n)
with probability 1 nc if we set = nc for a fixed c.
5.3.4.5 Multi-armed bandits

In the multi-armed bandit problem, we must choose at each time step
one of a fixed set of k arms to pull. Pulling arm i at time t yields a return
15
This can be proven most easily using convexity of n ln n.
of Xit , a random payoff typically assumed to be between 0 and 1. Suppose

that all the Xit are independent, and that for each fixed i, all Xit have the
same distribution, and thus the same expected payoff. Suppose also that we
initially know nothing about these distributions. What strategy can we use
to maximize our expected payoff over a large number of pulls?
More specifically, how can we minimize our regret, defined as
k
T
X
E [Xi ] Ti , (5.3.12)
i=1
where is the expected payoff of the best arm, and Ti counts the number
of times we pull arm i?
The tricky part here is that when we pull an arm and get a bad return,
we dont know if we were just unlucky this time or its actually a bad arm.
So we have an incentive to try lots of different arms. On the other hand, the
more we pull a genuinely inferior arm, the worse our overall return. Wed like
to adopt a strategy that trades off between exploration (trying new arms)
and exploitation (collecting on the best arm so far) to do as best we can in
comparison to a strategy that always pulls the best arm.
The UCB1 algorithm Fortunately, there is a simple algorithm due to

Auer et al. [ACBF02] that solves this problem for us.16 To start with, we
pull each arm once. For any subsequent pull, suppose that for each i, we
P
have pulled the i-th arm ni times so far. Let n = ni , and let xi be the
average payoff from arm i so far. Then the UCB1 algorithm pulls the arm
that maximizes s
2 ln n
xi + . (5.3.13)
ni
UCB stands for upper confidence bound. (The 1 is because it is
the first, and simplest, of several algorithms of this general structure given
in the paper.) The idea is that we give arms that we havent tried a lot the
benefit of the doubt, and assume that their actual average payout lies at the
upper end of some plausible range of values.17
16
This is not the only algorithm for solving multi-armed bandit problems, and its not
even the only algorithm in the Auer et al. paper. But it has the advantage of being
relatively straightforward to analyze. For a more general survey of multi-armed bandit
algorithms for the regret model, see [BCB12].
17
The bound part is because we dont attempt to compute the exact upper end of
this confidence interval, which may be difficult, but instead use an upper bound derived
from Hoeffdings inequality. This distinguishes the UCB1 algorithm of [ACBF02] from the
upper confidence interval approach of Lai and Robbins [LR85] that it builds on.
q
The quantity 2 nlni n is a bit mysterious, but it arises in a fairly natural
way from the asymmetric version of Hoeffdings inequality. With a small
adjustment to deal with non-zero-mean variables, (5.3.3) says that, if S is a
sum of n random variables bounded between ai and bi , then
" n #
2/
Pn
(bi ai )2
(E [Xi ] Xi ) t e2t
X
Pr i=1 . (5.3.14)
i=1
ni
Now consider xi = n1i j=1
P
X where each Xj lies between 0 and 1. This
Pnij
is equivalent to having xi = j=1 Yj where Yj = Xj /ni lies between 0 and
1/ni . So (5.3.14) says that

" s #
2 ln n 2 2
Pr E [xi ] xi e2( (2 ln n)/ni ) /(ni (1/ni ) )
ni
= e4 ln n
= n4 . (5.3.15)
The same bound also applies to xi E [xi ].

So the bonus applied to xi is really a high-probability bound on how big
the difference
between the observed payoff and expected payoff might be.
The ln n part is there to make the error probability be small as a function
of n, since we will be summing over a number of bad cases polynomial in n
and not a particular ni . Applying the bonus to all arms make it likely that
the observed payoff of the best arm stays above its actual payoff, so we wont
forget to pull it. The hope is that over time the same bonus applied to other
arms will not boost them up so much that we pull them any more than we
have to.
However, in an infinite execution of the algorithm, even a bad arm will
be pulled infinitely often, as ln n rises enough to compensate for the last
increase in ni . This accounts for an O(log n) term in the regret, as we will
see below. It also prevents us from getting stuck refusing to give a good arm
a second chance just because we had an initial run of bad luck.
Analysis of UCB1 The following theorem is a direct quote of [ACBF02,

Theorem 1]:
Theorem 5.3.4 ([ACBF02]). For all K > 1, if policy UCB1 is run on K

machines having arbitrary reward distributions P1 , . . . , PK with support in
[0, 1], then its expected regret after any number n of plays is at most
! K
X ln n 2 X
8 + 1+ j . (5.3.16)
i:i <
i 3 j=1
Here i = E [Pi ] is the expected payoff for arm i, as before is maxi i ,

and i = i is the regret for pulling arm i. The theorem states that
our expected regret is bounded by a term for each arm worse that that
grows logarithmically in n and is inversely proportional to how close the arm
is to optimal, plus a constant additional loss corresponding to pulling every
arm a little more than 4 times on average. The logarithmic regret in n is
a bit of a nuisance, but an earlier lower bound of Lai and Robbins [LR85]
shows that something like this is necessary in the limit.
To prove Theorem 5.3.4, we need to get an upper bound on the number
of times each suboptimal arm is pulled during the first n pulls. Define
s
2 ln t
ct,s = , (5.3.17)
s
the bonus given to an arm that has been pulled s times in the first t pulls.
Fix some optimal arm. Let X i,s be the average return on arm i after s

pulls and X s be the average return on this optimal arm after s pulls.
If we pull arm i after t total pulls, when arm i has previously been pulled
si times and our optimal arm has been pulled s times, then we must have

X i,si + ct,si X s + Ct,s . (5.3.18)
This just says that arm i with its bonus looks better than the optimal
arm with its bonus.
To show that this bad event doesnt happen, we need three things:

1. The value of X s +ct,s should be at least . Were it to be smaller, the

observed value X s would be more than ct,s away from its expectation.
Hoeffdings inequality implies this doesnt happen too often.
2. The value of X i,si + ct,si should not be too much bigger than i . Well
again use Hoeffdings inequality to show that X i,si is likely to be at
most i + ct,si , making X i,si + ct,si at most i + 2ct,si .
3. The bonus ct,si should be small enough that even adding 2ct,si to i
is not enough to beat . This means that we need to pick si large
enough that 2ct,si i . For smaller values of si , we will just accept
that we need to pull arm i a few more times before we wise up.
More formally, if none of the following events hold:

X s + ct,s . (5.3.19)
X i,si i + ct,si (5.3.20)

i < 2ct,si , (5.3.21)

then X s + ct,s > > i + 2ct,si > X i,si , and we dont pull arm i because
the optimal arm is better. (We dont necessarily pull the optimal arm, but if
we dont, its because we pull some other arm that still isnt arm i.)
For (5.3.19) and (5.3.20), we repeat the argument in (5.3.15), plugging
in t for n and si or s for ni . This gives a probability of at most 2t4 that
either or both of these bad events occur.
For (5.3.21), we need to do something a little sneaky, because the state-
ment is not actually true when si is small. So we will give ` free pulls to arm
i, and only start comparing arm i to the optimal arm after we have done
at least this many pulls. The value of ` is chosen so that, when t n and
si > `,
2ct,si i ,
which expands to,

s
2 ln t
2 i ,
si
giving
8 ln t
si .
2i
So we must set ` to be at least

8 ln n 8 ln t
.
2i 2i
Because ` must be an integer, we actually get

8 ln n 8 ln n
`=d 2 e1+ .
i 2i
This explains (after multiplying by the regret i ) the first term in (5.3.16).
For the other sources of bad pulls, apply the union bound to the 2t4
error probabilities we previously computed for all choices of t n, s 1,
and si > `. This gives
n X
t1 t1
2t4 < 2 t2 t4
X X X
t=1 s =1 si =`+1 t=1

2
=2
6
2
= .
3
Again we have to multiply by the regret i for pulling the i-th arm, which
gives the second term in (5.3.16).
5.4 Anti-concentration bounds

It may be that for some problem you want to show that a sum of random
variables is far from its mean at least some of the time: this would be
an anti-concentration bound. Anti-concentration bounds are much less
well-understood than concentration bounds, but there are known results that
can help in some cases.
For variables where we know the distribution of the sum exactly (e.g.,
sums with binomial distributions, or sums we can attack with generating
functions), we dont need these. But they may be useful if computing the
distribution of the sum directly is hard.
5.4.1 The Berry-Esseen theorem

The Berry-Esseen theorem18 characterizes how quickly a sum of inde-
pendent identically-distributed random variables converges to a normal
distribution, as a function of the third moment of the random variables.
Its simplest version says that if we have n independent, identically-distributed
2 2
random
h ivariables X1 . . . Xn , with E [Xi ] = 0, Var [Xi ] = E Xi = , and
E |Xi |3 , then
n
" #
1 X C
sup Pr Xi x (x) 3 , (5.4.1)

<x< n i=1 n
18
Sometimes written Berry-Essen theorem to help with the pronunciation of Esseens
last name.
were C is an absolute constant and is the normal distribution function. Note

that the 3 in the denominator is really Var [Xi ]3/2 . Since the probability
bound doesnt depend on x, its more useful toward the middle of the
distribution than in the tails.
A classic proof of this result with C = 3 can be found in [Fel71, XVI.5].
More recent work has tightened the bound on C: the best currently known
constant is C < 0.4784 < 1/2, due to Shevtsova [She11].
As with most results involving sums of random variables, there are
generalizations to martingales. These are too involved to describe here, but
see [HH80, 3.6].
5.4.2 The Littlewood-Offord problem

The Littlewood-Offord problem asks, given a set of n complex numbers
x1 . . . xn with |xi | 1, for how many assignments of 1 to coefficients
1 . . . n it holds that | ni=1 i xi | r. Paul Erds showed [Erd45] that this
P

quantity was at mostcr2n / n, where c is a constant. The quantity c2n / n
n
here is really 12 bn/2c : Erdss proof shows that for each interval of length
2r, the number of assignments that give a sum in the interior of the interval
is bounded by at most the sum of the r largest binomial coefficients.
In random-variable terms, this means that if we are looking at ni=1 xi ,
P
where the xi are constants with |xi | 1 and the i are independent 1 fair
coin-flips, then Pr [| ni=1 i xi | r] is maximized by making all the xi equal
P
to 1. This shows that any distribution where the xi are all reasonably large
will not be any more concentrated than a binomial distribution.
There has been a lot of more recent work on variants of the Littlewood-
Offord problem, much of it by Terry Tao and Van Vu. See http://terrytao.
wordpress.com/2009/02/16/a-sharp-inverse-littlewood-offord-theorem/
for a summary of much of this work.
Chapter 6
Randomized search trees
These are data structures that are either trees or equivalent to trees, and use
randomization to maintain balance. Well start by reviewing deterministic
binary search trees and then add in the randomization.
6.1 Binary search trees

A binary search tree is a standard data structure for holding sorted data.
A binary tree is either empty, or it consists of a root node containing a
key and pointers to left and right subtrees. What makes a binary tree a
binary search tree is the invariant that, both for the tree as a whole and
any subtree, all keys in the left subtree are less than the key in the root,
while all keys in the right subtree are greater than the key in the root. This
ordering property means that we can search for a particular key by doing
binary search: if the key is not at the root, we can recurse into the left or
right subtree depending on whether it is smaller or bigger than the key at
the root.
The efficiency of this operation depends on the tree being balanced. If
each subtree always holds a constant fraction of the nodes in the tree, then
each recursive step throws away a constant fraction of the remaining nodes.
So after O(log n) steps, we find the key we are looking for (or find that
the key is not in the tree). But the definition of a binary search tree does
not by itself guarantee balance, and in the worst case a binary search tree
degenerates into a linked list with O(n) cost for all operations (see Figure 6.2.
93
CHAPTER 6. RANDOMIZED SEARCH TREES 94
y x
x C A y
A B B C
Figure 6.1: Tree rotations
6.1.1 Rebalancing and rotations

Deterministic binary search tree implementations include sophisticated re-
balancing mechanisms to adjust the structure of the tree to preserve balance
as nodes are inserted or delete. Typically this is done using rotations,
which are operations that change the position of a parent and a child while
preserving the left-to-right ordering of keys (see Figure 6.1).
Examples include AVL trees [AVL62], where the left and right subtrees
of any node have heights that differ by at most 1; red-black trees [GS78],
where a coloring scheme is used to maintain balance; and scapegoat
trees [GR93], where no information is stored at a node but part of the
tree is rebuilt from scratch whenever an operation takes too long. These all
give O(log n) cost per operation (amortized in the case of scapegoat trees),
and vary in how much work is needed in rebalancing. Both AVL trees and
red-black trees perform more rotations than randomized rebalancing does on
average.
6.2 Random insertions

Suppose we insert n keys into an initially-empty binary search tree in random
order with no rebalancing. This means that for each insertion, we follow the
same path that we would when searching for the key, and when we reach an
empty tree, we replace it with a tree consisting solely of the key at the root.
Since we chose a random order, each element is equally likely to be the
root, and all the elements less than the root end up in the left subtree, while
all the elements greater than the root end up in the right subtree, where
they are further partitioned recursively. This is exactly what happens in
randomized QuickSort (see 1.3.1), so the structure of the tree will exactly
mirror the structure of an execution of QuickSort. So, for example, we
can immediately observe from our previous analysis of QuickSort that the
total path lengththe sum of the depths of the nodesis (n log n),
4 1
2 6 2
1 3 5 7 3
Figure 6.2: Balanced and unbalanced binary search trees
1 7
3 6
2 4
Figure 6.3: Binary search tree after inserting 5 1 7 3 4 6 2
since the depth of each node is equal to 1 plus the number of comparisons it
participates in as a non-pivot, and (using the same argument as for Hoares
FIND in 3.6.3) that the height of the tree is O(log n) with high probability.1
When n is small, randomized binary search trees can look pretty scraggly.
Figure 6.3 shows a typical example.
1
The argument for Hoares FIND is that any node has at most 3/4 of the descendants
of its parent on average; this gives for any node x that Pr [depth(x) > d] (3/4)d1 n, or
a probability of at most nc that depth(x) > 1 + (c + 1) log(n)/ log(4/3) 1 + 6.952 ln n
for c = 1, which we need to apply the union bound. The right answer for the actual height
of a randomly-generated search tree in the limit is 4.31107 ln n [Dev88] so this bound
is actually pretty close. The real height is still nearly a factor of three worse than for a
completely balanced tree, which has max depth bounded by 1 + lg n 1 + 1.44269 ln n.
The problem with this approach in general is that we dont have any
guarantees that the input will be supplied in random order, and in the
worst case we end up with a linked list, giving O(n) worst-case cost for all
operations.
6.3 Treaps
The solution to bad inputs is the same as for QuickSort: instead of assuming
that the input is permuted randomly, we assign random priorities to each
element and organize the tree so that elements with higher priorities rise
to the top. The resulting structure is known as a treap [SA96], because it
satisfies the binary search tree property with respect to keys and the heap
property with respect to priorities.2
Theres an extensive page of information on treaps at http://faculty.
washington.edu/aragon/treaps.html, maintained by Cecilia Aragon, the
co-inventor of treaps; they are also discussed at length in [MR95, 8.2]. Well
give a brief description here.
To insert a new node in a treap, first walk down the tree according to
the key and insert the node as a new leaf. Then go back up fixing the heap
property by rotating the new element up until it reaches an ancestor with
the same or higher priority. (See Figure 6.4 for an example.) Deletion is
the reverse of insertion: rotate a node down to a leaf (by swapping with its
higher-priority child at each step), and then prune it off.
Because of the heap property, the root of each subtree is always the
element in that subtree with the highest priority. This means that the
structure of a treap is completely determined by the priorities and the keys,
no matter what order the elements arrive in. We can imagine in retrospect
that the treap is constructed recursively by choosing the highest-priority
element as the root, then organizing all smaller-index and all larger-index
nodes into the left and right subtrees by the same rule.
If we assign the priorities independently and uniformly at random from a
sufficiently large set ((n2 ) is enough in the limit), then we get no duplicates,
and by symmetry all n! orderings are equally likely. So the analysis of the
depth of a treap with random priorities is identical to the analysis of a binary
2
The name treap for this data structure is now standard but the history is a little
tricky. According to Seidel and Aragon, essentially the same data structure (though with
non-random priorities) was previously called a cartesian tree by Vuillemin [Vui80], and the
word treap was initially applied by McCreight to a different data structuredesigned
for storing two-dimensional datathat was called a priority search tree in its published
form. [McC85].
1,60 1,60
\ \
2,3 --> 3,26
\ /
(3,26) 2,3
1,60 1,60 1,60 5,78

\ \ \ /
3,26 --> 3,26 --> (5,78) --> 1,60
/ \ / \ / \
2,3 4,24 2,3 (5,78) 3,26 3,26
\ / / \ / \
(5,78) 4,24 2,3 4,24 2,3 4,24
5,78 5,78
/ \ / \
1,60 6,18 --> 1,60 7,41
\ \ \ /
3,26 (7,41) 3,26 6,18
/ \ / \
2,3 4,24 2,3 4,24
Figure 6.4: Inserting values into a treap. Each node is labeled with k, p where
k is the key and p the priority. Insertion of values not requiring rotations
are not shown.
search tree with random insertion order. Its not hard to see that the costs
of search insertion, and deletion operations are all linear in the depth of the
tree, so the expected cost of each of these operations is O(log n).
6.3.1 Assumption of an oblivious adversary

One caveat is that this only works if the priorities of the elements of the tree
are in fact independent. If operations on the tree are chosen by an adaptive
adversary, this assumption may not work. An adaptive adversary is one
that can observe the choice made by the algorithm and react to them: in this
case, a simple strategy would be to insert elements 1, 2, 3, 4, etc., in order,
deleting each one and reinserting it until it has a lower priority value than all
the smaller elements. This might take a while for the later elements, but the
end result is the linked list again. For this reason it is standard to assume in
randomized data structures that the adversary is oblivious, meaning that
it has to specify a sequence of operations without knowing what choices are
made by the algorithm. Under this assumption, whatever insert or delete
operations the adversary chooses, at the end of any particular sequence of
operations we still have independent priorities on all the remaining elements,
and the O(log n) analysis goes through.
6.3.2 Analysis
The analysis of treaps as carried out by Seidel and Aragon [SA96] is a nice
example of how to decompose a messy process into simple variables, much
like the linearity-of-expectation argument for QuickSort (1.3.2). The key
observation is that its possible to bound both the expected depth of any node
and the number of rotations needed for an insert or delete operation directly
from information about the ancestor-descendant relationship between nodes.
Define two classes of indicator variables. For simplicity, we assume that
the elements have keys 1 through n, which we also use as indices.
1. Ai,j indicates the event that i is an ancestor of j, where i is an ancestor

of j if it appears on the path from the root to j. Note that every node
is an ancestor of itself.
2. Ci;`,m indicates the event that i is a common ancestor of both ` and

m; formally, Ci;`,m = Ai,` Ai,m .
The nice thing about these indicator variables is that its easy to compute
their expectations.
For Ai,j , i will be the ancestor of j if and only if i has a higher priority
than j and there is no k between i and j that has an even higher prior-
ity: in other words, if i has the highest priority of all keys in the interval
[min(i, j), max(i, j)]. To see this, imagine that we are constructing the treap
recursively, by starting with all elements in a single interval and partitioning
each interval by its highest-priority element. Consider the last interval in
this process that contains both i and j, and suppose i < j (the j > i case is
symmetric). If the highest-priority element is some k with i < k < j, then i
and j are separated into distinct intervals and neither is the ancestor of the
other. If the highest-priority element is j, then j becomes the ancestor of i.
The highest-priority element cant be less than i or greater than j, because
then we get a smaller interval that contains both i and j. So the only case
where i becomes an ancestor of j is when i has the highest priority.
1
It follows that E [Ai,j ] = |ij|+1 , where the denominator is just the
number of elements in the range [min(i, j), max(i, j)].
For Ci;`,m , i is the common ancestor of both ` and m if and only if it is has
the highest priority in both [min(i, `), max(i, `)] and [min(i, m), max(i, m)].
It turns out that no matter what order i, `, and m come in, these intervals
overlap so that i must have the highest priority in [min(i, `, m), max(i, `, m)].
1
This gives E [Ci;`,m ] = max(i,`,m)min(i,`,m)+1 .
6.3.2.1 Searches
1.3 So
P
From the Ai,j variables we can compute depth(j) = i Ai,j
n
!
X 1
E [depth(j)] = 1
i=1
|i j| + 1

j n
X 1 X 1 1
= +
i=1
ji+1 i=j+1
i j + 1

j nj+1
X 1 X 1
= + 1
k=1
k k=2
k
= Hj + Hnj+1 2.
This is maximized at j = (n + 1)/2, giving 2H(n+1)/2 2 = 2 ln n + O(1).

So we get the same 2 ln n + O(1) bound on the expected depth of any one
3
We need the 1 because of the convention that the root has depth 0, making the
depth of a node one less than the number of its ancestors. Equivalently, we could exclude
j from the sum and count only proper ancestors.
4 2 6
/ \ / \ / \
*2 6* => 1 4 or 4 7
/ \ / \ / \ / \
1 *3 5* 7 *3 6* *2 5*
/ \ \
5* 7 *3
Figure 6.5: Rotating 4 right shortens the right spine of its left subtree by
removing 2; rotating left shortens the left spine of the right subtree by
removing 6.
node that we got for QuickSort. We can also sum over all j to get the exact
value of the expected total path length (but we wont). These quantities
bound the expected cost of searches.
6.3.2.2 Insertions and deletions

For insertions and deletions, the question is how many rotations we have to
perform to float a new leaf up to its proper location (after an insertion) or
to float a deleted node down to a leaf (before a deletion). Since insertion is
just the reverse of deletion, we can get a bound on both by concentrating on
deletion. The trick is to find some metric for each node that (a) bounds the
number of rotations needed to move a node to the bottom of the tree and
(b) is easy to compute based on the A and C variables
The left spine of a subtree is the set of all nodes obtained by starting
at the root and following left pointers; similarly the right spine is what we
get if we follow the right pointers instead.
When we rotate an element down, we are rotating either its left or right
child up. This removes one element from either the right spine of the left
subtree or the left spine of the right subtree, but the rest of the spines are
left intact (see Figure 6.5). Subsequent rotations will eventually remove all
these elements by rotating them above the target, while other elements in
the subtree will be carried out from under the target without ever appearing
as a child or parent of the target. Because each rotation removes exactly one
element from one or the other of the two spines, and we finish when both
are empty, the sum of the length of the spines gives the number of rotations.
To calculate the length of the right spine of the left subtree of some
element `, start with the predecessor ` 1 of `. Because there is no element
between them, either ` 1 is a descendant of ` or an ancestor of `. In the

former case (for example, when ` is 4 in Figure 6.5), we want to include all
ancestors of ` 1 up to ` itself. Starting with i Ai,`1 gets all the ancestors
P
of ` 1, and subtracting off i Ci;`1,` removes any common ancestors of

P
` 1 and `. Alternatively, if ` 1 is an ancestor of `, every ancestor of

` 1 is also an ancestor of `, so the same expression i Ai,`1 i Ci;`1,`
P P
evaluates to zero.
It follows that the expected length of the right spine of the left subtree is
exactly
" n n
# n n
X X X 1 X 1
E Ai,`1 Ci;`1,` =
i=1 i=1 i=1
|i (` 1)| + 1 i=1 max(i, `) min(i, ` 1) + 1
`1 n `1 n
X 1 X 1 X 1 X 1
= +
i=1
` i i=` i ` + 2 i=1 ` i + 1 i=` i (` 1) + 1
`1
X 1 n`+2
X 1 X `
1 n`+2
X 1
= +
j=1
j j=2
j j=2 j j=2
j
1
=1 .
`
By symmetry, the expected length of the left spine of the right subtree is
1
1 n`+1 . So the total expected number of rotations needed to delete the
`-th element is
1 1
2 2.
` n`+1
6.3.2.3 Other operations

Treaps support some other useful operations: for example, we can split a
treap into two treaps consisting of all elements less than and all elements
greater than a chosen pivot by rotating the pivot to the root (O(log n)
rotations on average, equal to the pivots expected depth as calculated in
6.3.2.1) and splitting off the left and right subtrees. The reverse merging
operation has the same expected complexity.
6.4 Skip lists

A skip list [Pug90] is a randomized tree-like data structure based on linked
lists. It consists of a level 0 list that is an ordinary sorted linked list, together
with higher-level lists that contain a random sampling of the elements at
HEAD TAIL
33 LEVEL 2
13 33 48 LEVEL 1
13 21 33 48 75 99 LEVEL 0
Figure 6.6: A skip list. The blue search path for 99 is superimposed on an
original image from [AS07].
lower levels. When inserted into the level i list, an element flips a coin that
tells it with probability p to insert itself in the level i + 1 list as well. The
result is that the element is represented by a tower of nodes, one in each
of the bottom 1 + X many layers, where X is a geometrically-distributed
random variable. An example of a small skip list is shown in Figure 6.6.
Searches in a skip list are done by starting in the highest-level list and
searching forward for the last node whose key is smaller than the target; the
search then continues in the same way on the next level down. To bound the
expected running time of a search, it helps to look at this process backwards;
the reversed search path starts at level 0 and continues going backwards
until it reaches the first element that is also in a higher level; it then jumps
to the next level up and repeats the process. The nice thing about this
reversed process is that it has a simple recursive structure: if we restrict a
skip list to only those nodes to the left of and at the same level or higher
of a particular node, we again get a skip list. Furthermore, the structure of
this restricted skip list depends only on coin-flips taken at nodes within it,
so its independent of anything that happens elsewhere in the full skip list.
We can analyze this process by tracking the number of nodes in the
restricted skip list described above, which is just the number of nodes in
the current level that are earlier than the current node. If we move left,
this drops by 1; if up, this drops to p times its previous value on average.
So the number of such nodes Xk after k steps satisfies E [Xk+1 | Xk ] =
(1 p)(Xk 1) + p(p E [Xk ]) = (1 p + p2 )Xk (1 p)2 (1 p + p2 )Xk ,
and in general we get E [Xk ] (1 p + p2 )k X0 = (1 p + p2 )n. (We can
substitute the rank of the starting node for n if we want a slightly tighter
bound.) This is minimized at p = 1/2, giving E [Xk ] (3/4)k n, suspiciously
similar to the bound we computed before for random binary search trees.
When Xk = 0, our search is done, so if T is the time to search, we have
Pr [T k] = Pr [Xk 1] (3/4)k n, by Markovs inequality. In particular, if
we want to guarantee that we finish with probability 1 , we need to run

for log4/3 (n/) steps. This translates into an O(log n) bound on the search
time, and the constant is even the same as our (somewhat loose) bound for
treaps.
The space per element of a skip list also depends on p. Every element
needs one pointer for each level it appears in. This gives one pointer for level
0, plus 1/(1 p) pointers on average for the higher levels. For constant p
this is O(1). However, the space cost can reduced (at the cost of increasing
search time) by adjusting p. For example, if space is at a premium, setting
p = 1/10 produces 10/9 pointers per node on averagenot much more than
in a linked listbut
gives O(log n) search time. In general the trade-off
still
1
is between n 1 + 1p total expected space and log1/(1p+p2 ) (n/) search
1
time. The constant factor in the search time is log(1p+p 2 ) , which for small
p is approximately 1/p.
Like treaps, skip lists can be split an merged in a straightforward way.
The difference is that in a skip list, its enough to cut (or recreate) all the
pointers crossing the boundary, without changing the structure of the rest of
the list.
Chapter 7
Hashing
The basic idea of hashing is that we have keys from a large set U , and wed
like to pack them in a small set M by passing them through some function
h : U M , without getting too many collisions, pairs of distinct keys x
and y with h(x) = h(y). Where randomization comes in is that we want this
to be true even if the adversary picks the keys to hash. At one extreme, we
could use a random function, but this will take a lot of space to store.1 So
our goal will be to find functions with succinct descriptions that are still
random enough to do what we want.
The presentation in this chapter is based largely on [MR95, 8.4-8.5]
(which is in turn based on work of Carter and Wegman [CW77] on universal
hashing and Fredman, Komls, and Szemerdi [FKS84] on O(1) worst-case
hashing); on [PR04] and [Pag06] for cuckoo hashing; and [MU05, 5.5.3] for
Bloom filters.
7.1 Hash tables

Here we review the basic idea of hash tables, which are implementations of
the dictionary data type mapping keys to values. The basic idea of hash
tables is usually attributed to Dumey [Dum56].2
1
An easy counting argument shows that almost all functions from U to M take |U | log|M |
bits to represent, no matter how clever you are about choosing your representation. This
forms the basis for algorithmic information theory, which defines an object as random
if there is no way to reduce the number of bits used to express it.
2
Caveat: This article is pretty hard to find, so I am basing this citation on its frequent
appearance in later sources. This is generally a bad idea that would not really be acceptable
in an actual scholarly publication.
104
CHAPTER 7. HASHING 105
Suppose we want to store n elements from a universe U of in a table

with keys or indices drawn from an index space M of size m. Typically we
assume U = [|U |] = {0 . . . |U | 1} and M = [m] = {0 . . . m 1}.
If |U | m, we can just use an array. Otherwise, we can map keys to
positions in the array using a hash function h : U M . This necessarily
produces collisions: pairs (x, y) with h(x) = h(y), and any design of a
hash table must include some mechanism for handling keys that hash to
the same place. Typically this is a secondary data structure in each bin,
but we may also place excess values in some other place. Typical choices
for data structures are linked lists (separate chaining or just chaining)
or secondary hash tables (see 7.3 below). Alternatively, we can push
excess values into other positions in the same hash table according to some
deterministic rule (open addressing or probing) or a second hash function
(see 7.4).
For most of these techniques, the costs of insertions and searches will
depend on how likely it is that we get collisions. An adversary that knows
our hash function can always choose keys with the same hash value, but we
can avoid that by choosing our hash function randomly. Our ultimate goal
is to do each search in O(1 + n/m) expected time, which for n m will be
much better than the (log n) time for pointer-based data structures like
balanced trees or skip lists. The quantity n/m is called the load factor of
the hash table and is often abbreviated as .
All of this only works if we are working in a RAM (random-access machine
model), where we can access arbitrary memory locations in time O(1) and
similarly compute arithmetic operations on O(log|U |)-bit values in time O(1).
There is an argument that in reality any actual RAM machine requires either
(log m) time to read one of m memory locations (routing costs) or, if one is
particularly pedantic, (m1/3 ) time (speed of light + finite volume for each
location). We will ignore this argument.
We will try to be consistent in our use of variables to refer to the different
parameters of a hash table. Table 7.1 summarizes the meaning of these
variable names.
7.2 Universal hash families

A family of hash functions H is 2-universal if for any x 6= y, Pr [h(x) = h(y)]
1/m for a uniform random h H. Its strongly 2-universal if for any
x1 6= x2 U , y1 , y2 M , Pr [h(x1 ) = y1 h(x2 ) = y2 ] = 1/m2 for a uniform
random h H. Another way to describe strong 2-universality is that the val-
U Universe of all keys

SU Set of keys stored in the table
n = |S| Number of keys stored in the table
M Set of table positions
m = |M | Number of table positions
= n/m Load factor
Table 7.1: Hash table parameters
ues of the hash function are uniformly distributed and pairwise-independent.

For k > 2, k-universal usually means strongly k-universal: Given
distinct x1 . . . xk , and any y1 . . . yk , Pr [h(xi ) = yi )i] = mk . This is equiv-
alent to the h(xi ) values for distinct xi and randomly-chosen h having a
uniform distribution and k-wise independence. It is possible to generalize
the weak version of 2-universality to get a weak version of k-universality
(Pr [h(xi ) are all equal] m(k1) ), but this generalization is not as useful
as strong k-universality.
To analyze universal hash families, it is helpful to have some notation for
counting collisions. Well mostly be doing counting rather than probabilities
because it saves carrying around a lot of denominators. Since we are assuming
uniform choices of h we can always get back probabilities by dividing by |H|.
Let (x, y, h) = 1 if x =6 y and h(x) = h(y), 0 otherwise. Abusing notation,
P
we also define, for sets X, Y , and H, (X, Y, H) = xX,yY,hH (x, y, h);
and allow lowercase variables to stand in for singleton sets, as in (x, Y, h) =
({x} , Y, {h}). Now the statement that H is 2-universal becomes x, y :
(x, y, H) |H|/m; this says that only 1/m of the functions in H cause any
particular distinct x and y to collide.
If H includes all functions U M , we get equality: a random function
gives h(x) = h(y) with probability exactly 1/m. But we might do better if
each h tends to map distinct values to distinct places. The following lemma
shows we cant do too much better:
Lemma
For any family H, there exist x, y such that (x, y, H)
7.2.1.
|H| m1
m 1 |U |1 .
Proof. Well count collisions in the inverse image of each element z. Since
all distinct pairs of elements of h1 (z) collide with each other, we have

(h1 (z), h1 (z), h) = h1 (z) h1 (z) 1 .

Summing over all z M gets all collisions, giving

X
(U, U, h) = h1 (z) h1 (z) 1 .

zM
Use convexity or Lagrange multipliers to argue that the right-hand side is

minimized subject to z h1 (z) = |U | when all pre-images are the same
P
size |U |/m. It follows that

X |U | |U |
(U, U, h) 1
zM
m m
|U | |U |

=m 1
m m
|U |
= (|U | m).
m
If we now sum over all h, we get
|H|
(U, U, H) |U |(|U | m).
m
There are exactly |U |(|U | 1) ordered pairs x, y for which (x, y, H) might
not be zero; so the Pigeonhole principle says some pair x, y has
|H| |U |(|U | m)

(x, y, H)
m |U |(|U | 1)
|H| m1

= 1 .
m |U | 1
m1
Since 1 |U |1 is likely to be very close to 1, we are happy if we get the
2-universal upper bound of |H|/m.
Why we care about this: With a 2-universal hash family, chaining using
linked lists costs O(1 + s/n) expected time per operation. The reason is that
the expected cost of an operation on some key x is proportional to the size
of the linked list at h(x) (plus O(1) for the cost of hashing itself). But the
expected size of this linked list is just the expected number of keys y in the
dictionary that collide with x, which is exactly s(x, y, H) s/n.
7.2.1 Linear congruential hashing

Universal hash families often look suspiciously like classic pseudorandom
number generators. Here is a 2-universal hash family based on taking
remainders. It is assumed that the universe U is a subset of Zp , the integers
mod p; effectively, this just means that every element x of U satisfies 0
x p 1.
Lemma 7.2.2. Let hab (x) = (ax + b mod p) mod m, where a Zp \ {0} , b
Zp , and p is a prime m. Then {hab } is 2-universal.
Proof. Again, we count collisions. Split hab (x) as g(fab (x)) where fab (x) =
ax + b mod p and g(x) = x mod m.
The intuition is that after feeding any distinct x and y through fab , all
distinct pairs of values r and s are equally likely. We then show that feeding
these values to g produces no more collisions than expected.
The formal statement of the intuition is that for any 0 x, y p 1
with x 6= y, (x, y, H) = (Zp , Zp , g).
To prove this, fix x and y, and consider some pair r 6= s Zp . Then the
equations ax + b = t and ay + b = u have a unique solution for a and b mod
p (because Zp is a finite field). So (fab (x), fab (y)) is a bijection between pairs
P P
a, b and pairs u, u and thus (x, y, H) = a,b (x, y, fab ) = t6=u (t, u, g) =
(Zp , Zp , g).
Now we just need to count how many distinct t and u collide. There are p
choices for t. For each t, there are at most dp/me values that map to the same
remainder mod m, but one of those values is equal to t, so this leaves only
dp/me 1 choices for u. But dp/me 1 (p + (m 1))/m 1 = (p 1)/m.
So (Zp , Zp , g) p(p 1)/m.
1
Since each choice of t and u occurs with probability p(p1) , this gives a
probability of a collision of at most 1/m.
A difficulty with this hash family is that it requires doing modular

arithmetic. A faster hash is given by Dietzfelbinger et al. [DHKP97], although
it requires a slight weakening of the notion of 2-universality. For each k and
` they define a class Hk,` of functions from [2k ] to [2` ] by defining
ha (x) = (ax mod 2k ) div 2k` ,
where x div y = bx/yc. They prove [DHKP97, Lemma 2.4] that if a is a

random odd integer with 0 < a < 2` , and x 6= y, Pr [ha (x) = ha (y)] 2`+1 .
This increases by a factor of 2 the likelihood of a collision, but any extra
costs from this can often be justified in practice by the reduction in costs
from working with powers of 2.
If we are willing to use more randomness (and more space), a method
called tabulation hashing (7.2.2) gives a simpler alternative that is 3-
universal.
7.2.2 Tabulation hashing

Tabulation hashing [CW77] is a method for hashing fixed-length strings
(or things that can be represented as fixed-length strings) into bit-vectors.
The description here follows Patrascu and Thorup [PT12].
Let c be the length of each string in characters, and let s be the size of
the alphabet. Initialize the hash function by constructing tables T1 . . . Tc
mapping characters to independent random bit-vectors of size lg m. Define
h(x) = T1 [x1 ] T2 [x2 ] . . . Tc [xc ],
where represents bitwise exclusive OR (what ^ does in C-like languages).3

This gives a family of hash functions that is 3-wise independent but not
4-wise independent.
The intuition for why the hash values might be independent is that if we
have a collection of strings, and each string brings in an element of T that
doesnt appear in the other strings, then that element is independent of the
hash values for the other strings and XORing it with the rest of the hash
value gives a random bit string that is independent of the hash values of the
other strings. In fact, we dont even need each string to include a unique
value; its enough if we can order the strings so that each string gets a value
that isnt represented among its predecessors.
More formally, suppose we can order the strings x1 , x2 , . . . , xn that we
0
are hashing so that each has a position ij such that xjij 6= xjij for any j 0 < j,
h 0
i
then we have, for each value v, Pr h(xj ) = v h(xj ) = vj 0 , j 0 < j = 1/m.

3
Letting m be a power of 2 and using exclusive OR is convenient on real computers. If
for some reason we dont like this approach, the same technique, with essentially the same
analysis, works for arbitrary m if we replace bitwise XOR with addition mod m.
It follows that the hash values are independent:

h i n
Y h i
Pr h(x1 ) = v1 , h(x2 ) = v2 , . . . , h(xn ) = vn ) = Pr h(xj ) = vj h(x1 ) = v1 . . . h(xj1 ) = vj1

j=1
1
=
mn
n
Y h i
= Pr h(xj ) = vj .
j=1
Now we want to show that when n = 3, this actually works for all possible
distinct strings x, y, and z. Let S be the set of indices i such that yi 6= xi ,
and similarly let T be the set of indices i such that zi 6= xi ; note that both
sets must be non-empty, since y 6= x and z 6= x. If S \ T is nonempty, then
(a) there is some index i in T where zi =6 xi , and (b) there is some index j in
S \ T where yi 6= xi = zi ; in this case, ordering the strings as x, z, y gives
the independence property above. If T \ S is nonempty, order them as x, z,
y instead. Alternatively, if S = T , then yi = 6 zi for some i in S (otherwise
y = z, since they both equal x on all positions outside S). In this case, xi ,
yi , and zi are all distinct.
For n = 4, we can have strings aa, ab, ba, and bb. If we take the
bitwise exclusive OR of all four hash values, we get zero, because each
character is included exactly twice in each position. So the hash values are
not independent, and we do not get 4-independence in general.
However, even though tabulation hashing is not 4-independent, most
reasonably small sets of inputs do give independence. This can be used to
show various miraculous properties like working well for cuckoo hashing.
7.3 FKS hashing

The FKS hash table, named for Fredman, Komls, and Szemerdi [FKS84],
is a method for storing a static set S so that we never pay more than
constant time for search (not just in expectation), while at the same time
not consuming too much space. The assumption that S is static is critical,
because FKS chooses hash functions based on the elements of S.
If we were lucky in our choice of S, we might be able to do this with
standard hashing. A perfect hash function for a set S U is a hash
function h : U M that is injective on S (that is, x 6= y h(x) 6= h(y)
when x, y S). Unfortunately, we can only count on finding a perfect hash
function if m is large:
n
Lemma 7.3.1. If H is 2-universal and |S| = n with 2 m, then there is
a perfect h H for S.
Proof. Let h be chosen uniformly at random from H. For each unordered pair
x 6= y in S, let Xxy be the indicator variable for the even that h(x) = h(y),
P
and let C = x6=y Xxy be the total number of collisions in S. Each Xxy
has expectation at most 1/m, so E [C] n2 /m < 1. But we we can write
E [C] as E [C | C = 0] Pr [C = 0] + E [C | C 1] Pr [C 6= 0] Pr [C 6= 0]. So
Pr [C 6= 0] n2 /m < 1, giving Pr [C = 0] > 0. But if C is zero with nonzero

probability, there must be some h that makes it 0. That h is perfect for

S.
Essentially the same argument shows that if n2 m, then

Pr [h is perfect for S] 1 . This can be handy if we want to find a

perfect hash function and not just demonstrate that it exists.
Using a perfect hash function, we get O(1) search time using O(n2 ) space.
But we can do better by using perfect hash functions only at the second
level of our data structure, which at top level will just be an ordinary hash
table. This is the idea behind the Fredman-Komls-Szemerdi (FKS) hash
table [FKS84].
The short version is that we hash to n = |S| bins, then rehash perfectly
within each bin. The top-level hash table stores a pointer to a header for
each bin, which gives the size of the bin and the hash function used within
it. The i-th bin, containing ni elements, uses O(n2i ) space to allow perfect
hashing. The total size is O(n) as long as we can show that ni=1 n2i = O(n).
P
The time to do a search is O(1) in the worst case: O(1) for the outer hash
plus O(1) for the inner hash.
Theorem 7.3.2. The FKS hash table uses O(n) space.
Proof. Suppose we choose h H as the outer hash function, where H is

some 2-universal family of hash functions. Compute:
n
X n
X
n2i = (ni + ni (ni 1))
i=1 i=1
= n + (S, S, h).
The last equality holds because each ordered pair of distinct values in S that
map to the same bucket i corresponds to exactly one collision in (S, S, h).
Since H is 2-universal, we have (S, S, H) |H|s(s 1)/n. But then
the Pigeonhole principle says there exists some h H with (S, S, h)
1 Pn 2
n(n 1)/n = n 1. This gives
|H| (S, S, H) i=1 ni n + (n 1) =
2n 1 = O(n).
If we want to find a good h quickly, increasing the size of the outer table
to n/ gives us a probability of at least 1 of getting a good one, using
essentially the same argument as for perfect hash functions.
7.4 Cuckoo hashing

The FKS hash guarantees that any search takes only two probes, but it only
works for static data. Cuckoo hashing [PR04] gives the same guarantee
for dynamic data.
The main idea is to use two hash functions h1 and h2 , and store each
element x in either h1 (x) or h2 (x). This may require moving other elements
to their alternate location to make room. But the payoff is that each search
takes only two reads, which can even be done in parallel. This is optimal by
a lower bound of Pagh [Pag01], which shows a matching upper bound for
static dictionaries.
Cuckoo hashing was invented by Pagh and Rodler [PR04]. The version
described here is based on a simplified version from notes of Pagh [Pag06] (the
main difference is that it uses just one table instead of the two tablesone
for each hash functionin [PR04]).
7.4.1 Structure
We have a table T of size n, with two separate, independent hash functions h1
and h2 . These functions are assumed to be k-universal for some sufficiently
large value k; as long as we never look at more than k values at once, this
means we can treat them effectively as random functions. In practice, using
crummy hash functions seems to work just fine, a common property of hash
tables. There are also specific hash functions that have been shown to work
with particular variants of cuckoo hashing [PR04, PT12]. We will avoid these
issues by assuming that our hash functions are actually random.
Every key x is stored either in T [h1 (x)] or T [h2 (x)]. So the search
procedure just looks at both of these locations and returns whichever one
contains x (or fails if neither contains x).
To insert a value x1 = x, we must put it in T [h1 (x1 )] or T [h2 (x1 )]. If one
or both of these locations is empty, we put it there. Otherwise we have to
kick out some value that is in the way (this is the cuckoo part of cuckoo
hashing, named after the bird that leaves its eggs in other birds nests). We
do this by letting x2 = T [h1 (x1 )] and writing x1 to T [h1 (x1 )]. We now have
a new nestless value x2 , which we swap with whatever is in T [h2 (x2 )]. If
that location was empty, we are done; otherwise, we get a new value x3 that
we have to put in T [h1 (x3 )] and so on. The procedure terminates when we
find an empty spot or if enough iterations have passed that we dont expect
to find an empty spot, in which case we rehash the entire table. This process
can be implemented succinctly as shown in Algorithm 7.1.
1 procedure insert(x)
2 if T (h1 (x) = x or T (h2 (x)) = x then
3 return
4 pos h1 (x)
5 for i 1 . . . n do
6 if T [pos] = then
7 T [pos] x
8 return
9 x T [pos]
10 if pos = h1 (x) then
11 pos h2 (x)
12 else
13 pos h1 (x)
14 If we got here, rehash the table and reinsert x.

Algorithm 7.1: Insertion procedure for cuckoo hashing. Adapted
from [Pag06]
A detail not included in the above code is that we always rehash (in
theory) after m2 insertions; this avoids potential problems with the hash
functions used in the paper not being universal enough. We will avoid this
issue by assuming that our hash functions are actually random (instead of
being approximately n-universal with reasonably high probability). For a
more principled analysis of where the hash functions come from, see [PR04].
An alternative hash family that is known to work for a slightly different
variant of cuckoo hashing is tabulation hashing, as described in 7.2.2; the
proof that this works is found in [PT12].
7.4.2 Analysis
The main question is how long it takes the insertion procedure to terminate,
assuming the table is not too full.
First lets look at what happens during an insert if we have a lot of
nestless values. We have a sequence of values x1 , x2 , . . . , where each pair of
values xi , xi+1 collides in h1 or h2 . Assuming we dont reach the loop limit,
there are three main possibilities (the leaves of the tree of cases below):
1. Eventually we reach an empty position without seeing the same key
twice.
2. Eventually we see the same key twice; there is some i and j > i such
that xj = xi . Since xi was already moved once, when we reach it
the second time we will try to move it back, displacing xi1 . This
process continues until we have restored x2 to T [h1 (x1 )], displacing x1
to T [h2 (x1 )] and possibly creating a new sequence of nestless values.
Two outcomes are now possible:
(a) Some x` is moved to an empty location. We win!
(b) Some x` is moved to a location weve already looked at. We lose!
We find we are playing musical chairs with more players than
chairs, and have to rehash.
Lets look at the probability that we get the last, closed loop case. Follow-
ing Pagh-Rodler, we let v be the number of distinct nestless keys in the loop.
Since v includes x1 , v is at least 1. We can now count how many different
ways such a loop can form.
There are at most v 3 choices for i, j, and `, mv1 choices of cells for the
loop, and nv1 choices for the non-x1 elements of the loop. For each non-xi
element, its initial placement may determined by either h1 or h2 ; this gives
another 2v1 choices.4 This gives a total of v 3 (2nm)v1 possible closed loops
starting with x1 that have v distinct nodes.
Since each particular loop allows us to determine both h1 and h2 for all v
of its elements, the probability that we get exactly these hash values (so that
the loop occurs) is m2v . Summing over all closed loops with v elements
gives a total probability of
v 3 (2nm)v1 m2v = v 3 (2n)v1 mv1
= v 3 (2n/m)v1 m2 .
4
The original analysis in [PR04] avoids this by alternating between two tables, so that
we can determine which of h1 or h2 is used at each step by parity.
Now sum over all v 1. We get

n
m2 v 3 (2n/m)v1 < m2
X X
v 3 (2n/m)v1 .
v=1 v=1
The series converges if 2n/m < 1, so for any fixed < 1/2, the probability
of any closed loop forming is O(m2 ).
If we do hit a closed loop, then we pay O(m) time to scan the existing
table and create a new empty table, and O(n) = O(m) time on average to
reinsert all the elements into the new table, assuming that this reinsertion
process doesnt generate any more closed loops and that the average cost of an
insertion that doesnt produce a closed loop is O(1), which we will show below.
But the rehashing step only fails with probability O(nm2 ) = O(m1 ), so if
it does fail we can just try again until it works, and the expected total cost
is still O(m). Since we pay this O(m) for each insertion with probability
O(m2 ), this adds only O(m1 ) to the expected cost of a single insertion.
Now we look at what happens if we dont get a closed loop. This doesnt
force us to rehash, but if the path is long enough, we may still pay a lot to
do an insertion.
Its a little messy to analyze the behavior of keys that appear more than
once in the sequence, so the trick used in the paper is to observe that for any
sequence of nestless keys x1 . . . xp , there is a subsequence of size p/3 with no
repetitions that starts with x1 . This will be either the sequence S1 given by
x1 . . . xj1 the sequence starting with the first place we try to insert x1 or
S2 given by x1 = xi+j1 . . . xp , the sequence starting with the second place
we try to insert x1 . Between these we have a third sequence S3 where we
undo some of the moves made in S1 . Because |S1 | + |S3 | + |S2 | p, at least
one of thee subsequences has size p/3. But |S3 | |S1 |, so it must be either
S1 or S2 .
We can then argue that the probability that we get a sequence of v
distinct keys in either S1 or S2 most 2(n/m)v1 . The (n/m)v1 is because
we need to hit a nonempty spot (which happens with probability at most
n/m) for the first v 1 elements in the path, and since we assume that our
hash functions are random, the choices of these v 1 spots are all independent.
The 2 is from the union bound over S1 and S2 . If T is the insertion time,
we get E [T ] =
P v1 = O(1), assuming n/m
v=1 Pr [T v]
P
v=1 2(n/m)
is bounded by a constant less than 1. Since we already need n/m 1/2 to
avoid the bad closed-loop case, we can use this here as well.
An annoyance with cuckoo hashing is that it has high space overhead
compared to more traditional hash tables: in order for the first part of the
analysis above to work, the table must be at least half empty. This can be
avoided at the cost of increasing the time complexity by choosing between
d locations instead of 2. This technique, due to Fotakis et al. [FPSS03], is
known as d-ary cuckoo hashing; for suitable choice of d it uses (1 + )n
space and guarantees that a lookup takes O(1/) probes while insertion
takes (1/)O(log log(1/)) steps in theory and appears to take O(1/) steps in
experiments done by the authors.
7.5 Practical issues

For large hash tables, local probing schemes are faster, because it is likely that
all of the locations probed to find a particular value will be on the same virtual
memory page. This means that a search for a new value usually requires one
cache miss instead of two. Hopscotch hashing [HST08] combines ideas
from linear probing and cuckoo hashing to get better performance than both
in practice.
Hash tables that depend on strong properties of the hash function may
behave badly if the user supplies a crummy hash function. For this reason,
many library implementations of hash tables are written defensively, using
algorithms that respond better in bad cases. See http://svn.python.org/
view/python/trunk/Objects/dictobject.c for an example of a widely-
used hash table implementation chosen specifically because of its poor theo-
retical characteristics.
7.6 Bloom filters

See [MU05, 5.5.3] for basics and a formal analysis or http://en.wikipedia.
org/wiki/Bloom_filter for many variations and the collective wisdom of
the unwashed masses. The presentation here mostly follows [MU05].
7.6.1 Construction
Bloom filters are a highly space-efficient randomized data structure invented
by Burton H. Bloom [Blo70] that store sets of data, with a small probability
that elements not in the set will be erroneously reported as being in the set.
Suppose we have k independent hash functions h1 , h2 , . . . , hk . Our mem-
ory store A is a vector of m bits, all initially zero. To store a key x, set
A[hi (x)] = 1 for all i. To test membership for x, see if A[hi (x)] = 1 for all
i. The membership test always gives the right answer if x is in fact in the
Bloom filter. If not, we might decide that x is in the Bloom filter anyway.
7.6.2 False positives

The probability of such false positives can be computed in two steps: first,
we estimate how many of the bits in the Bloom filter are set after inserting
n values, and then we use this estimate to compute a probability that any
fixed x shows up when it shouldnt.
If the hi are close to being independent random functions,5 then with
n entries in the filter we have Pr [A[i] = 1] = 1 (1 1/m)kn , since each of
the kn bits that we set while inserting the n values has one chance in m of
hitting position i.
Wed like to simplify this using the inequality 1 + x ex , but it goes
2
in the wrong direction; instead, well use 1 x exx , which holds for
0 x 0.683803 and in our application holds for m 2. This gives
Pr [A[i] = 1] 1 (1 1/m)kn
1 ek(n/m)(1+1/m)
= 1 ek(1+1/m)
0
= 1 ek
where = n/m is the load factor and 0 = (1 + 1/m) is the load factor
fudged upward by a factor of 1 + 1/m to make the inequality work.
Suppose now that we check to see if some value x that we never inserted
in the Bloom filter appears to be present anyway. This occurs if A[hi (x)] = 1
for all i. Since each hi (x) is assumed to be an independent uniform probe of
A, the probability that they all come up 1 conditioned on A is
P k
A[i]
. (7.6.1)
m
0

We have an upper bound E [ A[i]] m 1 ek , and if we were born
P
luckier might be able to get an upper bound on the expectation of (7.6.1)

by applying Jensens inequality to the function f (x) = xk . But sadly this
5
We are going sidestep the rather deep swamp of how plausible this assumption
is and what assumption we should be making instead; however, it is known [KM08]
that starting with two sufficiently random-looking hash functions h and h0 and setting
hi (x) = h(x) + ih0 (x) works.
inequality also goes in the wrong direction, because f is convex for k > 1.
P
So instead we will prove a concentration bound on S = A[i].
Because the A[i] are not independent, we cant use off-the-shelf Chernoff
bounds. Instead, we rely on McDiarmids inequality. Our assumption is
that the locations of the kn ones that get written to A are independent.
Furthermore, changing the location of one of these writes changes S by at most
2
1. So McDiarmids inequality (5.3.11)
q gives Pr [S E [S] + t] e2t /kn ,
which is bounded by nc for t 12 ckn log n. So as long as a reasonably
large fraction of the array is likely to be full, the relative error from assuming
S = E [S] is likely to be small. Alternatively, if the array is mostly empty,
then we dont care about the relative error so much because the probability
of getting a false positive will already be exponentially small as a function of
k.
So lets assume for simplicity that our false positive probability is exactly
0
(1 ek )k . We can choose k to minimize this quantity for fixed 0 by
doing the usual trick of taking a derivative and setting it to zero; to avoid
weirdness with the k in the exponent, it helps to take the logarithm first
(which doesnt affect the location of the minimum), and it further helps to
0
take the derivative with respect to x = e k instead of k itself. Note that
when we do this, k = 10 ln x still depends on x, and we will deal with this
by applying this substitution at an appropriate point.
Compute
d d
ln (1 x)k = k ln(1 x)
dx dx
d 1

= 0 ln x ln(1 x)
dx
1 ln(1 x) ln x

= 0 .
x 1x
Setting this to zero gives (1 x) ln(1 x) = x ln x, which by symmetry

has the unique solution x = 1/2, giving k = 10 ln 2.
In other words, to minimize the false positive rate for a known load factor
, we want to choose k = 10 ln 2 = (1+1/m) 1
ln 2, which makes each bit one
with probability approximately 1 e ln 2 = 12 . This makes intuitive sense,
since having each bit be one or zero with equal probability maximizes the
entropy of the data.
0
The probability of a false positive is then 2k = 2 ln 2/ . For a given
maximum false positive rate , and assuming optimal choice of k, we need to
ln2 2 ln2 2
keep 0 ln(1/) or (1+1/m) ln(1/) .
Alternatively, if we fix and n, we need m/(1 + 1/m) n ln(1/)

ln2 2

1.442n lg(1/), which works out to m 1.442n lg(1/) + O(1). This is very
good for constant .
Note that for this choice of m, we have = O(1/ ln(1/)), giving k =
O(log log(1/)). So for polynomial , we get k = O(log log n). This means
that not only do we use little space, but we also have very fast lookups
(although not as fast as the O(1) cost of a real hash table).
7.6.3 Comparison to optimal space

If we wanted to design a Bloom-filter-like data structure from scratch and
had no constraints on processing power, wed be looking for something that
stored an index of size lg M into a family of subsets S1 , S2 , . . . SM of our
universe of keys U , where |Si | |U | for each i (giving the upper bound on
the false positive rate)6 and for any set A U of size n, A Si for at least
one Si (allowing us to store A).
Let N = |U |. Then each set Si covers N N

n of the n subsets of size n. If
we could. get them to overlap optimally (we cant), wed still need a minimum
N N
of n n = (N )n /(N )n (1/)n sets to cover everybody, where the
approximation assumes N n. Taking the log gives lg M n lg(1/),
meaning we need about lg(1/) bits per key for the data structure. Bloom
filters use 1/ ln 2 times this.
There are known data structures that approach this bound asymptotically.
The first of these, due to Pagh et al. [PPR05] also has other desirable
properties, like supporting deletions and faster lookups if we cant look up
bits in parallel.
More recently, Fan et al. [FAKM14] have described a variant of cuckoo
hashing (see 7.4) called a cuckoo filter. This is a cuckoo hash table that,
instead of storing full keys x, stores fingerprints f (x), where f is a hash
function with `-bit outputs. False positives now arise if we happen to hash a
value x0 with f (x0 ) = f (x) to the same location as x. If f is drawn from a
6
Technically, this gives a weaker bound on false positives. For standard Bloom filters,
assuming random hash functions, each key individually has at most an probability
of appearing as a false positive. The hypothetical data structure we are considering
herewhich is effectively deterministicallows the set of false positives to depend directly
on the set of keys actually inserted in the data structure, so in principle the adversary
could arrange for a specific key to appear as a false positive with probability 1 by choosing
appropriate keys to insert. So this argument may underestimate the space needed to get
make the false positives less predictable. On the other hand, we arent charging the Bloom
filter for the space needed to store the hash functions, which could be quite a bit if they
are genuine random functions.
2-universal family, this occurs with probability at most 2` . So the idea is

that by accepting an small rate of false positives, we can shrink the space
needed to store each key from the full key length to lg(1/) = ln(1/)/ ln 2,
the asymptotic minimum.
One complication is that, since we are throwing away the original key x,
when we displace a key from h1 (x) to h2 (x) or vice versa, we cant recompute
h1 (x) and h2 (x) for arbitrary h1 and h2 . The solution proposed by Fan et al.
is to let h2 (x) = h1 (x) g(f (x)), where g is a hash function that depends
only on the fingerprint. This means that when looking at a fingerprint f (x)
stored in position i, we dont need to know whether i is h1 (x) or h2 (x), since
whichever it is, the other location will be i g(f (x)). Unfortunately, this
technique and some other techniques used in the paper to crunch out excess
empty space break the standard analysis of cuckoo hashing, so the authors
can only point to experimental evidence that their data structure actually
works. However, a variant of this data structure has been shown to work by
Eppstein [Epp16].
7.6.4 Applications
Historically, Bloom filters were invented to act as a way of filtering queries
to a database table through fast but expensive7 RAM before looking up the
actual values on a slow but cheap tape drive. Nowadays the cost of RAM is
low enough that this is less of an issue in most cases, but Bloom filters are
still popular in networking and in distributed databases.
In networking, Bloom filters are useful in building network switches, where
incoming packets need to be matched against routing tables in fractions of a
nanosecond. Bloom filters work particularly well for this when implemented
in hardware, since the k hash functions can be computed in parallel. False
positives, if infrequent enough, can be handled by some slower backup
mechanism.
In distributed databases, Bloom filters are used in the Bloomjoin algo-
rithm [ML86]. Here we want to do a join on two tables stored on different
machines (a join is an operation where we find all pairs of rows, one in each
table, that match on some common key). A straightforward but expensive
way to do this is to send the list of keys from the smaller table across the
network, then match them against the corresponding keys from the larger
table. If there are ns rows in the smaller table, nb rows in the larger table,
and j matching rows in the larger table, this requires sending ns keys plus j
7
As much as $0.10/bit in 1970.
rows. If instead we send a Bloom filter representing the set of keys in the
smaller table, we only need to send lg(1/)/ ln 2 bits for the Bloom filter plus
an extra nb rows on average for the false positives. This can be cheaper
than sending full keys across if the number of false positives is reasonably
small.
7.6.5 Counting Bloom filters

Its not hard to modify a Bloom filter to support deletion. The basic trick is
to replace each bit with a counter, so that whenever a value x is inserted, we
increment A[hi (x)] for all i and when it is deleted, we decrement the same
locations. The search procedure now returns mini A[hi (x)] (which means
that it principle it can even report back multiplicities, though with some
probability of reporting a value that is too high). To avoid too much space
overhead, each array location is capped at some small maximum value c;
once it reaches this value, further increments have no effect. The resulting
structure is called a counting Bloom filter, due to Fan et al. [FCAB00].
Wed only expect this to work if our chances of hitting the cap is small.
Fan et al. observe that the probability that the m table entries include one
that is at least c after n insertions is bounded by
! c
nk 1 enk 1

m c
m
c m c mc
enk c

=m
cm
= m(ek/c)c .
k
(This uses the bound nk en

k , which follows from Stirlings formula.)
For k = 1 ln 2, this is m(e ln 2/c)c . For the specific value of c = 16
(corresponding to 4 bits per entry), they compute a bound of 1.37 1015 m,
which they argue is minuscule for all reasonable values of m (its a systems
paper).
The possibility that a long chain of alternating insertions and deletions
might produce a false negative due to overflow is considered in the paper,
but the authors state that the probability of such a chain of events is so
low that it is much more likely that the proxy server would be rebooted in
the meantime and the entire structure reconstructed. An alternative way of
dealing with this problem is to never decrement a maxed-out register; this
never produces a false negative, but may cause the filter to slowly fill up
with maxed-out registers, producing a higher false-positive rate.
A fancier variant of this idea is the spectral Bloom filter [CM03],

which uses larger counters to track multiplicities of items. The essential idea
here is that we can guess that the number of times a particular value x was
inserted is equal to minmi=1 A[hi (x)]), with some extra tinkering to detect
errors based on deviations from the typical joint distribution of the A[hi (x)]
values. An even more sophisticated approach gives the count-min sketches
of the next section.
7.6.6 Count-min sketches

Count-min sketches are designed for use in data stream computation.
In this model, we are given a huge flood of datafar too big to storein a
single pass, and want to incrementally build a small data structure, called
a sketch, that will allow us to answer statistical questions about the data
after weve processed it all. The motivation is the existence of data sets that
are too large to store at all (network traffic statistics), or too large to store
in fast memory (very large database tables). By building a sketch we can
make one pass through the data set but answer queries after the fact, with
some loss of accuracy.
An example of a problem in this model is that we are presented with a
sequence of pairs (it , ct ) where 1 it n is an index and ct is a count, and
we want to construct a sketch that will allows us to approximately answer
P
statistical queries about the vector a given by ai = t,i[t]=i ct . The size of
the sketch should be polylogarithmic in the size of a and the length of the
stream, and polynomial in the error bounds. Updating the sketch given a
new data point should be cheap.
A solution to this problem is given by the count-min sketch of Cormode
and Muthukrishnan [CM05] (see also [MU05, 13.4]). This gives approxi-
mations of ai , ri=` ai , and a b (for any fixed b), and can be used for more
P
complex tasks like finding heavy hittersindices with high weight. The
easiest case is approximating ai when all the ct are non-negative, so well
start with that.
7.6.6.1 Initialization and updates

To construct a count-min sketch, build a two-dimensional array c with depth
d = dln(1/)e and width w = de/e, where is the error bound and is
the probability of exceeding the error bound. Choose d independent hash
functions from some 2-universal hash family; well use one of these hash
function for each row of the array. Initialize c to all zeros.
The update rule: Given an update (it , ct ), increment c[j, hj (it )] by ct for
j = 1 . . . d. (This is the count part of count-min.)
7.6.6.2 Queries
Lets start with point queries. Here we want to estimate ai for some
fixed i. There are two cases; the first handles non-negative increments only,
while the second handles arbitrary increments. In both cases we will get an
estimate whose error is linear in both the error parameter and the `1 -norm
kak1 = i |ai | of a. It follows that the relative error will be low for heavy
P
points, but we may get a large relative error for light points (and especially
large for points that dont appear in the data set at all).
For the non-negative case, to estimate ai , compute a i = minj c[j, hj (i)].
(This is the min part of coin-min.) Then:
Lemma 7.6.1. When all ct are non-negative, for a

i as defined above:
i ai ,
a (7.6.2)
and
ai ai + kak1 ] 1 .
Pr [ (7.6.3)
Proof. The lower bound is easy. Since for each pair (i, ct ) we increment each
c[j, hj (i)] by ct , we have an invariant that ai c[j, hj (i)] for all j throughout
the computation, which gives ai a i = minj c[j, hj (i)].
For the upper bound, let Iijk be the indicator for the event that (i 6=
k) (hj (i) = hj (k)), i.e., that we get a collision between i and k using hj .
The 2-universality property of the hj gives E [Iijk ] 1/w /e.
Now let Xij = nk=1 Iijk ak . Then c[j, hj (i)] = ai + Xij . (The fact that
P
Xij 0 gives an alternate proof of the lower bound.) Now use linearity of
expectation to get
" n #
X
E [Xij ] = E Iijk ak
k=1
n
X
= ak E [Iijk ]
k=1
Xn
ak (/e)
k=1
= (/e)kak1 .
So Pr [c[j, hj (i)] > ai + kak1 ] = Pr [Xij > e E [Xij ]] < 1/e, by Markovs in-
equality. With d choices for j, and each hj chosen independently, the proba-
bility that every count is too big is at most (1/e)d = ed exp( ln(1/)) =
.
Now lets consider the general case, where the increments ct might be
negative. We still initialize and update the data structure as described in
7.6.6.1, but now when computing a i , we use the median count instead of
the minimum count: a i = median {c[j, hj (i)] | j = 1 . . . n}. Now we get:
Lemma 7.6.2. For a

i as defined above,
i ai + 3kak1 ] > 1 1/4 .

Pr [ai 3kak1 a (7.6.4)
Proof. The basic idea is that for the median to be off by t, at least d/2
rows must give values that are off by t. Well show that for t = 3kak1 , the
expected number of rows that are off by t is at most d/8. Since the hash
functions for the rows are chosen independently, we can use Chernoff bounds
to show that with a mean of d/8, the chances of getting all the way to d/2
are small.
In detail, we again define the error term Xij as above, and observe that
" #
X
E [|Xij |] = E Iijk ak

k
n
X
|ak E [Iijk ]|
k=1
Xn
|ak |(/e)
k=1
= (/e)kak1 .
Using Markovs inequality, we get Pr [|Xij |] > 3kak1 ] = Pr [|Xij | > 3e E [Xij ]] <
1/3e < 1/8. In order for the median to be off by more than 3kak1 , we need
d/2 of these low-probability events to occur. The expected number that
occur is = d/8, so applying the standard Chernoff bound (5.2.1) with = 3
we are looking at
Pr [S d/2] = Pr [S (1 + 3)]
(e3 /44 )d/8
(e3/8 /2)ln(1/)
= ln 23/8
< 1/4
(the actual exponent is about 0.31, but 1/4 is easier to deal with). This
immediately gives (7.6.4).
One way to think about this is that getting an estimate within kak1 of
the right value with probability at least 1 requires 3 times the width and
4 times the depthor 12 times the space and 4 times the timewhen we
arent assuming increments are non-negative.
Next, we consider inner products. Here we want to estimate a b, where
a and b are both stored as count-min sketches using the same hash functions.
The paper concentrates on the case where a and b are both non-negative,
which has applications in estimating the size of a join in a database. The
method is to estimate a b as minj w k=1 ca [j, k] cb [j, k].
P
For a single j, the sum consists of both good values and bad collisions; we
have w
Pn
k=1 ca [j, k] cb [j, k] =
P P
k=1 ai bi + p6=q,hj (p)=hj (q) ap bq . The second
term has expectation
X X
Pr [hj (p) = hj (q)] ap bq (/e)ap bq
p6=q p6=q
X
(/e)ap bq
p,q
(/e)kak1 kbk1 .
As in the point-query case, we get probability at most 1/e that a single j

gives a value that is too high by more than kak1 kbk1 , so the probability
that the minimum value is too high is at most ed .
7.6.6.3 Finding heavy hitters

Here we want to find the heaviest elements in the set: those indices i for
which ai exceeds kak1 for some constant threshold 0 < 1.
The easy case is when increments are non-negative (for the general case,
see the paper), and uses a method from a previous paper by Charikar et
al. [CCFC04]. Because kak1 = i ai , we know that there will be at most 1/

P
heavy hitters. But the tricky part is figuring out which elements they are.
Instead of trying to find the elements after the fact, we extend the data
structure and update procedure to track all the heavy elements found so
far (stored in a heap), as well as kak1 = ct . When a new increment (i, c)
P
comes in, we first update the count-min structure and then do a point query
i kak1 , we insert i into the heap, and if not, we delete i along
on ai ; if a
with any other value whose stored point-query estimate has dropped below
threshold.
The trick here is that the threshold kak1 only increases over time
(remember that we are assuming non-negative increments). So if some
element i is below threshold at time t, it can only go above threshold if it
shows up again, and we have a probability of at least 1 of including it then.
This means that every heavy hitter appears in the heap with probability at
least 1 .
The total space cost for this data structure is the cost of the count-min
structure plus the cost of the heap; this last part will be O((1 + )/) with
reasonably high probability, since this is the maximum number of elements
that have weight at least kak1 /(1 + ), the minimum needed to get an
apparent weight of kak1 even after taking into account the error in the
count-min structure.
7.7 Locality-sensitive hashing

Locality-sensitive hashing was invented by Indyk and Motwani [IM98] to
solve the problem of designing a data structure that finds approximate nearest
neighbors to query points in high dimension. Well mostly be following this
paper in this section, concentrating on the hashing parts.
7.7.1 Approximate nearest neighbor search

In the nearest neighbor search problem (NNS for short), we are given a
set of n points P in a metric space with distance function d, and we want
to construct a data structure that allows us to quickly find the closet point
p in P to any given query point q. We could always compute the distance
between q and each possible p, but this takes time O(n), and wed like to
get lookups to be sublinear in n.
Indyk and Motwani were particularly interested in what happens in Rd
for high dimension d under various natural metrics. Because the volume of
a ball in a high-dimensional space grows exponentially with the dimension,
this problem suffers from the curse of dimensionality [Bel57]: simple

techniques based on, for example, assigning points in P to nearby locations
in a grid may require searching exponentially many grid locations. Indyk and
Motwani deal with this through a combination of randomization and solving
the weaker problem of -nearest neighbor search (-NNS), where its
OK to return a different point p0 as long as d(q, p0 ) (1 + ) minpP d(q, p).
This problem can be solved by reduction to a simpler problem called
-point location in equal balls or -PLEB. In this problem, we are given
n radius-r balls centered on points c in a set C, and we want a data structure
that returns a point c0 C with d(q, c0 ) (1 + )r if there is at least one
point c with d(q, c) r. If there is no such point, the data structure may or
may not return a point (its allowed to say no, or alternatively it might return
a point that is too far away, which we can simply discard). The difference
between an -PLEB and NNS is that an -PLEB isnt picky about returning
the closest point to q if there are multiple points that are all good enough.
Still, we can reduce NNS to -PLEB.
maxx,yP d(x,y)
The easy reduction is to use binary search. Let R = minx,yP,x6 =y d(x,y)
.
Given a point q, look for the minimum ` (1 + )0 , (1 + )1 , . . . , R for

which an -PLEB data structure with radius ` and centers P returns a point
p with d(q, p) (1 + )`; then return this point as the approximate nearest
neighbor.
This requires O(log1+ R) instances of the -PLEB data structure and
O(log log1+ R) queries. The blowup as a function of R can be avoided using
a more sophisticated data structure called a ring-cover tree, defined in the
paper. We wont talk about ring-cover trees because they are (a) complicated
and (b) not randomized. Instead, well move directly to the question of how
we solve -PLEB.
7.7.1.1 Locality-sensitive hash functions

Definition 7.7.1 ([IM98]). A family of hash functions H is (r1 , r2 , p1 , p2 )-
sensitive for d if, for any points p and q, if h is chosen uniformly from
H,
1. If d(p, q) r1 , then Pr [h(p) = h(q)] p1 , and
2. If d(p, q) > r2 , then Pr [h(p) = h(q)] p2 .
These are useful if p1 > p2 and r1 < r2 ; that is, we are more likely to hash
inputs together if they are closer. Ideally, we can choose r1 and r2 to build
-PLEB data structures for a range of radii sufficient to do binary search as
described above (or build a ring-cover tree if we are doing it right). For the
moment, we will aim for an (r1 , r2 )-PLEB data structure, which returns a
point within r1 with high probability if one exists, and never returns a point
farther away than r2 .
There is some similarity between locality-sensitive hashing and a more gen-
eral dimension-reduction technique known as the Johnson-Lindenstrauss
lemma [JL84]; this says that projecting n points in a high-dimensional space
to O(2 log n) dimensions using an appropriate random matrix preserves
`2 distances between the points to within relative error (in fact, even a
random matrix with 1 entries is enough [Ach03]). Unfortunately, dimension
reduction by itself is not enough to solve approximate nearest neighbors
in sublinear time, because we may still need to search a number of boxes
exponential in O(2 log n), which will be polynomial in n.
7.7.1.2 Constructing an (r1 , r2 )-PLEB

The first trick is to amplify the difference between p1 and p2 so that we can
find a point within r1 of our query point q if one exists. This is done in three
stages: First, we concatenate multiple hash functions to drive the probability
that distant points hash together down until we get few collisions: the idea
here is that we are taking the AND of the events that we get collisions in the
original hash function. Second, we hash our query point and target points
multiple times to bring the probability that nearby points hash together up:
this is an OR. Finally, we iterate the procedure to drive down any remaining
probability of failure below a target probability : another AND.
For the first stage, let k = log1/p2 n and define a composite hash function
log1/p n
g(p) = (h1 (p) . . . hk (p)). If d(p, q) > r2 , Pr [g(p) = g(q)] pk2 = p2 2
=
1/n. Adding this up over all n points in our data structure gives us one false
match for q on average.
However, we may not be able to find the correct match for q, since p1
may not be all that much larger than p2 . For this, we do a second round of
amplification, where now we are taking the OR of events we want instead of
the AND of events we dont want.
Let ` = n , where = log(1/p 1) log p1
log(1/p2 ) = log p2 < 1, and choose hash functions
g1 . . . g` independently as above. To store a point p, put it in a bucket for
gj (p) for each j; these buckets are themselves stored in a hash table (by
hashing the value of gj (p) down further) so that they fit in O(n) space.
Suppose now that d(p, q) r1 for some p. Then
Pr [gj (p) = gj (q)] pk1

log1/p2 n
= p1
log 1/p
log 1/p1
=n 2

=n
= 1/`.
So by searching through ` independent buckets we find p with probability

at least 1 (1 1/`)` = 1 1/e + o(1). Wed like to guarantee that we
only have to look at O(n ) points (most of which we may reject) during this
process; but we can do this by stopping if we see more than 2` points. Since
we only expect to see ` bad points in all ` buckets, this event only happens
with probability 1/2. So even adding it to the probability of failure from the
hash functions not working right we still have only a constant probability of
failure 1/e + 1/2 + o(1).
Iterating the entire process O(log(1/)) times then gives the desired
bound on the probability that this process fails to find a good point if one
exists.
the costs gives a cost of a query of O(k` log(1/)) =
Multiplying out all
O n log1/p2 n log(1/) hash function evaluations and O(n log(1/)) dis-

tance
computations. The cost to insert a point is just O(k` log(1/)) =

O n log1/p2 n log(1/) hash function evaluations, the same number as for
a query.
7.7.1.3 Hash functions for Hamming distance

Suppose that our points are d-bit vectors and that we use Hamming dis-
tance for our metric. In this case, using the family of one-bit projections
{hi | hi (x) = xi } gives a locality-sensitive hash family [ABMRT96].
Specifically, we can show this family is (r, r(1 + ), 1 dr , 1 r(1+)
d )-
sensitive. The argument is trivial: if two points p and q are at distance r
or less, they differ in at most r places, so the probability that they hash
together is just the probability that we dont pick one of these places, which
is at least 1 dr . Essentially the same argument works when p and q are far
away.
These are not particularly clever hash functions, so the heavy lifting will
be done by the (r1 , r2 )-PLEB construction. Our goal is to build an -PLEB
for any fixed r, which will correspond to an (r, r(1 + ))-PLEB. The main
thing we need to do, following [IM98] as always, is compute a reasonable
p1 ln(1r/d)
bound on = log log p2 = ln(1(1+)r/d) . This is essentially just a matter of
hitting it with enough inequalities, although there are a couple of tricks in
the middle.
Compute
ln(1 r/d)
=
ln(1 (1 + )r/d)
(d/r) ln(1 r/d)
=
(d/r) ln(1 (1 + )r/d)
ln((1 r/d)d/r )
=
ln((1 (1 + )r/d)d/r )
ln(e1 (1 r/d))

ln e(1+)
1 + ln(1 r/d)
=
(1 + )
1 ln(1 r/d)
= . (7.7.1)
1+ 1+
Note that we used the fact that 1 + x ex for all x in the denominator
and (1 x)1/x e1 (1 x) for x [0, 1] in the numerator. The first fact is
our usual favorite inequality.
The second can be proved in a number of ways. The most visually
intuitive is that (1 x)1/x and e1 (1 x) are equal at x = 1 and equal
in the limit as x goes to 0, while (1 x)1/x is concave in between 0 and
1 and e1 (1 x) is linear. Unfortunately it is rather painful to show that
(1 x)1/x is in fact concave. An alternative is to rewrite the inequality
(1 x)1x e1 (1 x) as (1 x)1/x1 e1 , apply a change of variables
y = 1/x to get (1 1/y)y1 e1 for y [1, ), and then argue that (a)
equality holds in the limit as y goes to infinity, and (b) the left-hand-side is
a nonincreasing function, since

d d
ln (1 1/y)y1 = [(y 1) (ln(y 1) ln y)]
dy dy
1 1

= ln(1 1/y) + (y 1)
y1 y
= ln(1 1/y) + 1 (1 1/y)
= ln(1 1/y) + 1/y
1/y + 1/y
= 0.
We now return to (7.7.1). Wed really like the second term to be small
enough that we can just write n as n1/(1+) . (Note that even though it looks
negative, it isnt, because ln(1 r/d) is negative.) So we pull a rabbit out of
a hat by assuming that r/d < 1/ ln n.8 This assumption can be justified by
modifying the algorithm so that d is padded out with up to d ln n unused
junk bits if necessary. Using this assumption, we get
n < n1/(1+) n ln(11/ ln n)/(1+)

= n1/(1+) (1 1/ ln n) ln n
en1/(1+) .
Plugging into the formula for (r1 , r2 )-PLEB gives O(n1/(1+) log n log(1/))
hash function evaluations per query, each of which costs O(1) time, plus
O(n1/(1+) log(1/)) distance computations, which will take O(d) time each.
If we add in the cost of the binary search, we have to multiply this by
O(log log1+ R log log log1+ R), where the log-log-log comes from having to
adjust so that the error doesnt accumulate too much over all O(log log R)
steps. The end result is that we can do approximate nearest-neighbor queries
in

O n1/(1+) log(1/)(log n + d) log log1+ R log log log1+ R
time. For reasonably large, this is much better than naively testing against
all points in our database, which takes O(nd) time (although it does produce
an exact result).
8
Indyk and Motwani pull this rabbit out of a hat a few steps earlier, but its pretty
much the same rabbit either way.
7.7.1.4 Hash functions for `1 distance

Essentially the same approach works for (bounded) `1 distance, using dis-
cretization, where we replace a continuous variable over some range with
a discrete variable. Suppose we are working in [0, 1]d with the `1 metric.
Represent each coordinate xi as a sequence of d/ values xij in unary, for
j = 1 . . . d, with xij = 1 if j/d < xi . Then the Hamming distance between
the bit-vectors representing x and y is proportional to the `1 distance between
the original vectors, plus an error term that is bounded by . We can then
use the hash functions for Hamming distance to get a locality-sensitive hash
family.
A nice bit about this construction is that we dont actually have to build
the bit-vectors; instead, we can specify a coordinate xi and a threshold c
and get the same effect by recording whether xi > c or not.
Note that this does increase the cost slightly: we are converting d-
dimensional vectors into (d/)-long bit vectors, so the log(n+d) term becomes
log(n + d/). When n is small, this effectively multiples the cost of a query by
an extra log(1/). More significant is that we have to cut in half to obtain
the same error bounds, because we now pay error for the data structure
itself and an additional error for the discretization. So our revised cost for
the `1 case is

O n1/(1+/2) log(1/)(log n + d/) log log1+/2 R log log log1+/2 R .
Chapter 8
Martingales and stopping

times
In 5.3.2, we used martingales to show that the outcome of some process

was tightly concentrated. Here we will show how martingales interact with
stopping times, which are random variables that control when we stop
carrying out some task. This will require a few new definitions.
8.1 Definitions
The general form of a martingale {Xt , Ft } consists of:
A sequence of random variables X0 , X1 , X2 , . . . ; and
A filtration F0 F1 F2 . . . , where each -algebra Ft represents

our knowledge at time t;
subject to the requirements that:
1. The sequence of random variables is adapted to the filtration, which

just means that each Xt is Ft -measurable or equivalently that Ft (and
thus all subsequent Ft0 for t0 t) includes all knowledge of Xt ; and
2. The martingale property
E [Xt+1 | Ft ] = Xt (8.1.1)
holds for all t.
133
CHAPTER 8. MARTINGALES AND STOPPING TIMES 134
Together with this more general definition of a martingale, we will also

use the following definition of a stopping time. Given a filtration {Ft }, a
random variable is a stopping time for {Ft } if N {} and the event
[ t] is Ft -measurable for all t.1 In simple terms, is a stopping time if
you know at time t whether to stop there or not.
What we like about martingales is that iterating the martingale property
shows that E [Xt ] = E [X0 ] for all fixed t. We will show that, under reasonable
conditions, the same holds for X when is a stopping time. (The random
variable X is defined in the obvious way, as a random variable that takes
on the value of Xt when = t.)
8.2 Submartingales and supermartingales

In some cases we have a process where instead of getting equality in (8.1.1),
we get an inequality instead. A submartingale replaces (8.1.1) with
Xt E [Xt+1 | Ft ] (8.2.1)
while a supermartingale satisfies
Xt E [Xt+1 | Ft ] . (8.2.2)
In each case, what is sub or super is the value at the current time
compared to the expected value at the next time. Intuitively, a submartingale
corresponds to a process where you win on average, while a supermartingale
is a process where you lose on average. Casino games (in profitable casinos)
are submartingales for the house and supermartingales for the player.
Sub- and supermartingales can be reduced to martingales by subtracting
off the expected change at each step. For example, if {Xt } is a submartingale
with respect to {Ft }, then the process {Yt } defined recursively by
Y0 = X0
Yt+1 = Yt + Xt+1 E [Xt+1 | Ft ]
1
Different authors impose different conditions on the range of ; for example, Mitzen-
macher and Upfal [MU05] exclude the case = . We allow = to represent the
outcome where we never stop. This can be handy for modeling processes where this
outcome is possible, although in practice we will typically insist that it occurs only with
probability zero.
is a martingale, since
E [Yt+1 | Ft ] = E [Yt + Xt+1 E [Xt+1 | Ft ] | Ft ]

= Yt + E [Xt+1 | Ft ] E [Xt+1 | Ft ]
= Yt .
One way to think of this is that Yt = Xt + t , where t is a predictable,

non-decreasing drift process that starts at 0. For supermartingales, the
same result holds, but now t is non-increasing. This ability to decompose
an adapted stochastic process into the sum of a martingale and a predictable
drift process is known as the Doob decomposition theorem.
8.3 The optional stopping theorem

If (Xt , Ft ) is a martingale, then applying induction to the martingale property
shows that E [Xt ] = E [X0 ] for any fixed time t. The optional stopping
theorem shows that this also happens for X when is a stopping time,
under various choices of additional conditions:
Theorem 8.3.1. Let (Xt , Ft ) be a martingale and a stopping time for

{Ft }. Then E [X ] = E[X0 ] if at least one of the following conditions holds:
1. Bounded time. There is a fixed n such that n always.
2. Bounded range and finite time. There is a fixed M such that

|Xt | M always, and Pr [ < ] = 1.
3. Bounded increments and finite expected time. There is a fixed

c such that |Xt+1 Xt | c always, and E [ ] < .
4. General case. All three of the following conditions hold:
(a) Pr [ < ] = 1,
(b) E [|X |] < , and
h i
(c) limt E Xt 1[ >t] = 0.
It would be nice if we could show E [X ] = E[X0 ] without the side

conditions, but in general this isnt true. For example, the double-after-losing
martingale strategy in the St. Petersburg paradox (see 3.4.1.1) eventually
yields +1 with probability 1, so if is the time we stop playing, we have
Pr [ < ] = 1, E [|X |] < , but E [X ] = 1 6= E [X0 ] = 0. But in order for
this to happen, we have to violate all of bounded time ( is not bounded),

bounded range (|Xt | roughly doubles every step until we stop), bounded
increments (|Xt+1 Xt | doubles every step as well), or one of the three
conditions of the general case (the last one: limt E [Xt ] 1[ >t] = 1 6= 0).
The next section gives a proof of the optional stopping theorem. The
intuition is that for any fixed n, we can truncate X to Xmin(,n) and show
h i
that E Xmin(,n) = E [X0 ]. This immediately gives the bounded time case.
h i
For the other cases, the argument is that limn E Xmin(,n) converges to
h i
E [X ] provided the missing part E X Xmin(,n) converges to zero. How
we do this depends on which assumptions we are making.
A proof of the optional stopping theorem is given in the next section.
You probably dont need to understand this proof to apply the theorem, and
may wish to skip directly to applications in 8.5.
8.4 Proof of the optional stopping theorem (op-

tional)
We start with the finite-time case, which we can enforce by truncating the
actual process.
Lemma 8.4.1. Let (Xht , Ft ) be aimartingale and a stopping time for {Ft }.
Then for any n N, E Xmin(,n) = E[X0 ].
Pt
Proof. Define Yt = X0 + i=1 (Xt Xt1 )1[ >t1]
h
. Then (Yt , Ft ) is a martin-
i
gale, because we can calculate E [Yt+1 | Ft ] = E Yt + (Xt+1 Xt )1[ >t] Ft =

Yt + 1[ >t] E [Xt+1 Xt | Ft ] = Yt ; effectively, we are treating 1[ t1] as a

sequence of bets, and we know that h adjusting i our bets doesnt change the
martingale property. But then E Xmin(,n) = E [Yn ] = E [Y0 ] = E [X0 ].
This gives us the bounded-time

h i for free: If n always, then
variant
X = Xmin(,n) , and E [X ] = E Xmin(,n) = E [X0 ].
For the unbounded-time variants, we will apply some version of the
following strategy:
h i
1. Observe that since E Xmin(,n) = E [X0 ] is a constant for any fixed n,
h i
limn E Xmin(,n) converges to E [X0 ].
h i
2. Argue using whatever assumptions we are making that limn E Xmin(,n)
also converges to E [X ].
3. Conclude that E [X0 ] = E [X ], since they are both limits of the same
sequence.
For the middle step, start with
X = Xmin(,n) + 1[ >n] (X Xn ).
This holds because either n, and we just get X , or > n, and we get
Xn + (X Xn ) = X .
Taking the expectation of both sides gives
h i h i
E [X ] = E Xmin(,n) + E 1[ >n] (X Xn ) .
h i
= E [X0 ] + E 1[ >n] (X Xn ) .
So if we can show that the right-hand term goes to zero in the limit, we are
done. h i
For the bounded-range case, we have |X Xn | 2M , so E 1[ >n] (X Xn )

2M Pr [ > n]. Since in this case we assume Pr [ < ] = 1, limn Pr [ > n] =

0, and the theorem holds.
For bounded increments, we have

h i X

E (X Xn )1[ >n] = E (Xt+1 Xt )1[ >t]

tn

X
E |(Xt+1 Xt )|1[ >t]
tn

X
E c1[ >t]
tn

X
cE 1[ >t] .
tn
P
But E [ ] = t=0 1[ >t] . Under the assumption that this sequence converges,
its tail goes to zero, and again the theorem holds.
For the general case, we can expand
h i h i h i
E [X ] = E Xmin(,n) + E 1[ >n] X E 1[ >n] Xn
which implies
h i h i h i
lim E [X ] = lim E Xmin(,n) + lim E 1[ >n] X lim E 1[ >n] Xn ,
n n n n
assuming all these limits exist and are finite. Weve already established that
the first limit is E [X0 ], which is exactly what we want. So we just need to
show that the other two limits both converge tohzero. Forithe last limit, we
just use condition (4c), which gives limn E 1[ >n] Xn = 0; no further
argument is needed. But we still need to show that the middle limit also
vanishes. h i P h i
Here we use condition (4b). Observe that E 1[ >n] X = t=n+1 E 1 X
[ =t] t .
h i
Compare this with E [X ] =
P
t=0 E 1[ =t] Xt ; this is an absolutely conver-
gent series (this is why we need condition (4b)), so in the limit the sum of
the terms for i = 0 . . . n converges to E [X ]. But this means that the sum
of the remaining terms for i = n + 1 . . . converges to zero. So the middle
term goes to zero as n goes to infinity. This completes the proof.
8.5 Applications
Here we give some example of the Optional Stopping Theorem in action. In
each case, the trick is to find an appropriate martingale and stopping time,
and let the theorem do all the work.
8.5.1 Random walks

Let Xt be an unbiased 1 random walk that starts at 0, adds 1 to its
current position with equal probability at each step, and stops if it reaches
+a or b.2 Wed like to calculate the probability of reaching +a before b.
Let be the time at which the process stops.
We can easily show that Pr [ < ] = 1 and E [ ] < by observing that
from any state of the random walk, there is a probability of at least 2(a+b)
that it stops within a + b steps (by flipping heads a + b times in a row),
so that if we consider a sequence of intervals of length a + b, the expected
number of such intervals we can have before we stop is at most 2a+b , giving
E [ ] (a + b)2a+b (we can do better than this).
We also have bounded increments by the definition of the process
(bounded range also works, at least up until time ). So E [X ] = E[X0 ] = 0
2
This is called a random walk with two absorbing barriers.
and the probability p of landing on +a instead of b must satisfy pa (1

b
p)b = 0, giving p = a+b .
Now suppose we want to find E [ ]. Let Yt = Xt2 t. Then Yt+1 =
(Xt 1)2 (t + 1) = Xt2 2Xt + 1 (t + 1) = (Xt2 t) 2Xt = Yt 2Xt . Since
the plus and minus cases are equally likely, they cancel out in expectation
and E [Yt+1 | Ft ] = Yt : we just showed Yt is a martingale.3 We can also show
it has bounded increments (at least up until time ), because |Yt+1 Yt | =
2|Xt | max(a, b). 2
From Theorem 8.3.1, E [Y ] = 0, which gives E [ ] = E X . But we
can calculate E X2 : it is a2 Pr [X = a] + b2 Pr [Xt = b] = a2 (b/(a + b)) +

b2 (a/(a + b)) = (a2 b + b2 a)/(a + b) = ab.

If we have a random walk that only stops at +a,4 then if is the
first time at which X = a, is a stopping time. However, in this case
3
This construction generalizes in a nice way to arbitrary martingales. Suppose {Xt } is
a martingale with respect to {Ft }. Let t = Xt Xt1 , and let Vt = Var [t | Ft1 ] be
the conditional variance of the t-th increment (note that this is a random variable
Pt that
may depend on previous outcomes). We can easily show that Yt = Xt2 i=1 Vi is a
martingale. The proof is that
" t
#
X
E [Yt | Ft1 ] = E Xt2 Vt Ft1

i=1
t
X
= E (Xt1 + t )2 Ft1

Vi
i=1
t
X
2

+ 2Xt1 t + 2t Ft1

= E Xt1 Vi
i=1
t
2
X
+ 2Xt1 E [t | Ft1 ] + E 2t Ft1

= Xt1 Vi
i=1
t
2
X
= Xt1 + 0 + Vt Vi
i=1
t1
2
X
= Xt1 Vi
i=1
= Yt1 .
Pt
For the 1 random walk case, we have Vt = 1 always, giving V = t and E X2 =
i=1 i
2
E X0 + E [ ] when is a stopping time satisfying the conditions of the Optional Stopping
P
Theorem. For the general case, the same argument gives E X2 = E X02 + E V
t=1 t
instead: the expected square position of Xt is incremented by the conditional variance at
each step.
4
This would be a random walk with one absorbing barrier.
E [X ] = a = 6 E [X0 ] = 0. So the optional stopping theorem doesnt apply in

this case. But we have bounded increments, so Theorem 8.3.1 would apply if
E [ ] < . It follows that the expected time until we reach a is unbounded,
either because sometimes we never reach a, or because we always reach a
but sometimes it takes a very long time. 5
We can also consider a biased random walk where +1 occurs with
probability p > 1/2 and 1 with probability q = 1 p. If Xt is the position
of the random walk at time t, it isnt a martingale. But Yt = Xt (p q)t is,
and it even has bounded increments. So if is the time at which Xt = a and
E [ ] is finite,6 the optional stopping theorem gives E [Y ] = E[Y0 ] = 0, which
a
gives E [a (p q)t] = 0 or E [t] = pq , pretty much what wed expect.
8.5.2 Walds equation

Suppose we run a Las Vegas algorithm until it succeeds, and the i-th attempt
costs Xi , where all the Xi are independent, satisfy 0 Xi c for some c,
and have a common mean E [Xi ] = .
Let N be the number of times we run the algorithm. Since we can tell
when we are done, N is a stopping time with respect to some filtration {Fi }
7
i Xi are adapted. Suppose also that E [N ] exists. What is
toh which the
PN
E i=1 Xi ?
If N were not a stopping time, this might be a very messy problem
indeed. But when N is a stopping time, we can apply it to the martingale
Yt = ti=1 (Xi ). This has bounded increments (0 Xi c, so c
P
Xi E [Xi ] c), and weve already said E [N ] is finite (which implies

5
In fact, we always reach a. An easy way to see this is to imagine a sequence of intervals
Pi 2
of length n1 , n2 , . . . , where ni+1 = a + n
j=1 j
. At the end of the i-th interval, we
Pi
are no lower than j=0 nj , so we only need to go up ni+1 positions to reach a by the
end of the (i + 1)-th interval. Since this is just one standard deviation, it occurs with
constant probability, so after a finite expected number of intervals, we will reach +a. Since
there are infinitely many intervals, we reach +a with probability 1.
6
Exercise: Show E [ ] is finite.
7
A stochastic process {Xt } is adapted to a filtration {Ft } if each Xt is Ft -measurable.
Pr [N < ] = 1), so Theorem 8.3.1 applies. We thus have
0 = E [YN ]
"N #
X
=E (Xi )
i=1
"N # "N #
X X
=E Xi E
i=1 i=1
"N #
X
=E Xi E [N ] .
i=1
Rearranging this gives Walds equation:

"N #
X
E Xi = E [N ] . (8.5.1)
i=1
This is the same formula as in 3.4.3.1, but weve eliminated the bound
on N and allowed for much more dependence between N and the Xi .8
8.5.3 Waiting times for patterns

Lets suppose we flip coins until we see some pattern appear: for example,
we might flip coins until we see HTHH. What is the expected number of
coin-flips until this happens?
A very clever trick due to Li [Li80] solves this problem exactly using the
Optional Stopping Theorem. Suppose our pattern is x1 x2 . . . xk . We imagine
an army of gamblers, one of which shows up before each coin-flip. Each
gambler starts by betting $1 that next coin-flip will be x1 . If she wins, she
bets $2 that the next coin-flip will be x2 , continuing to play double-or-nothing
until either she loses (and is $1) or wins her last bet on xk (and is up 2k 1).
Because each gamblers winnings form a martingale, so does their sum, and
so the expected total return of all gamblers up to the stopping time at
which our pattern first occurs is 0.
We can now use this fact to compute E [ ]. When we stop at time , we
have one gambler who has won 2k 1. We may also have other gamblers
8
In fact, looking closely at the proof reveals that we dont even need the Xi to be
independent of each other. We just need that E [Xi+1 | Fi ] = for all i to make (Yt , Ft ) a
martingale. But if we dont carry any information from one iteration of our Las Vegas
algorithm to the next, well get independence anyway. So the big payoff is not having to
worry about whether N has some devious dependence on the Xi .
who are still in play. For each i with x1 . . . xi = xki+1 . . . xk , there will be a
gambler with net winnings ij=1 2j1 = 2i 1. The remaining gamblers will
P
all be at 1.
Let i = 1 if x1 . . . xi = xki+1 . . . xk , and 0 otherwise. Then the number
of losers is given by ki=1 i and the total expected payoff is
P
k k
" #
X X
i
E [X ] = E ( i ) + i 2 1
i=1 i=1
k
" #
X
i
= E + i 2
i=1
= 0.
It follows that E [ ] = ki=1 i 2i .

P
As a quick test, the pattern H has E [ ] = 21 = 2. This is consistent with

what we know about geometric distributions.
For a longer example, the pattern HTHH only overlaps with its prefix H
so in this case we have E [ ] = i 2i = 16 + 2 = 18. But HHHH overlaps
P
with all of its prefixes, giving E [ ] = 16 + 8 + 4 + 2 = 30. At the other

extreme, THHH has no overlap at all and gives E [ ] = 16.
In general, for a pattern of length k, we expect a waiting time somewhere
between 2k and 2k+1 2almost a factor of 2 difference depending on how
much overlap we get.
This analysis generalizes in the obvious way to biased coins and larger
alphabets; see the paper [Li80] for details.
Chapter 9
Markov chains
A (discrete time) Markov chain is a sequence of random variables X0 , X1 , X2 , . . . ,

which we think of as the position of some particle at increasing times in N,
where the distribution of Xt+1 depends only on the value of Xt . A typical
example of a Markov chain is a random walk on a graph: each Xt is a node
in the graph, and a step moves to one of the neighbors of the current node
chosen at random, each with equal probability.
Markov chains come up in randomized algorithms both because the
execution of any randomized algorithm is, in effect, a Markov chain (the
random variables are the states of the algorithm); and because we can
often sample from distributions that are difficult to sample from directly
by designing a Markov chain that converges to the distribution we want.
Algorithms that use this latter technique are known as Markov chain
Monte Carlo algorithms, and rely on the fundamental fact that a Markov
chain that satisfies a few straightforward conditions will always converge in
the limit to a stationary distribution, no matter what state it starts in.
An example of this technique that predates randomized algorithms is
card shuffling: each permutation of the deck is a state, and the shuffling
operation sends the deck to a new state each time it is applied. Assuming
the shuffling operation is not too deterministic, it is possible to show that
enough shuffling will eventually produce a state that is close to being a
uniform random permutation. The big algorithmic question for this and
similar Markov chains is how quickly this happens: what is the mixing time
of the Markov chain, a measure of how long we have to run it to get close
to its limit distribution (this notion is defined formally in 9.2.3). Many of
the techniques in this chapter will be aimed at finding bounds on the mixing
time for particular Markov processes.
143
CHAPTER 9. MARKOV CHAINS 144
If you want to learn more about Markov chains than presented here or
in the textbook, they are usually covered in general probability textbooks
(for example, in [Fel68] or [GS01]), mentioned in many linear algebra text-
books [Str03], covered in some detail in stochastic processes textbooks [KT75],
and covered in exquisite detail in many books dedicated specifically to the
subject [KS76, KSK76]. Good sources for mixing times for Markov chains
are the textbook of Levin, Peres, and Wilmer [LPW09] and the survey pa-
per by Montenegro and Tetali [MT05]. An early reference on the mixing
times for random walks on graphs that helped inspire much subsequent
work is the Aldous-Fill manuscript [AF01], which can be found on-line at
http://www.stat.berkeley.edu/~aldous/RWG/book.html.
9.1 Basic definitions and properties

A Markov chain or Markov process is a stochastic process1 where the
distribution of Xt+1 depends only on the value of Xt and not any previous
history. Formally, this means that
Pr [Xt+1 = j | Xt = it , Xt1 = it1 , . . . , X0 = i0 ] = Pr [Xt+1 = j | Xt = it ] .

(9.1.1)
A stochastic process with this property is called memoryless: at any time,
you know where you are, and you can figure out where you are going, but
you dont know where you were before.
The state space of the chain is just the set of all values that each Xt can
have. A Markov chain is finite or countable if it has a finite or countable
state space, respectively. Well mostly be interested in finite Markov chains
(since we have to be able to fit them inside our computer), but countable
Markov chains will come up in some contexts.2
Well also assume that our Markov chains are homogeneous, which
means that Pr [Xt+1 = j | Xt = i] doesnt depend on t.
1
A stochastic process is just a sequence of random variables {St }, where we usually
think of t as representing time and the sequence as representing the evolution of some
system over time. Here we are considering discrete-time processes, where t will typically
be a non-negative integer.
2
If the state space is not countable, we run into the same measure-theoretic issues
as with continuous random variables, and have to replace (9.1.1) with the more general
condition that
E 1[Xt A] Xt , Xt1 , . . . , X0 = E 1[Xt A] Xt ,
provided A is measurable with respect to some appropriate -algebra. We dont really
want to deal with this, and for the most part we dont have to, so we wont.
For a homogeneous countable Markov chain, we can describe its behavior

completely by giving the state space and the one-step transition proba-
bilities pij = Pr [Xt+1 = j | Xt = i]. Given pij , we can calculate two-step
transition probabilities
(2)
pij = Pr [Xt+2 = j | Xt = i]
X
= Pr [Xt+2 = j | Xt+1 = k] Pr [Xt+1 = k | Xt = i]
k
X
= pik pkj .
k
This is identical to the formula for matrix multiplication. For a Markov

chain with n states, we can specify the transition probabilities pij using an
n n transition matrix P with Pij = pij , and the two-step transition
probabilities are given by p(2) )ij = Pij2 . More generally, the t-step transition
(t)
probability pij = (P t )ij .
Conversely, given any matrix with non-negative entries where the rows
P
sum to 1 ( j Pij = 1, or P 1 = 1, where 1 in the second equation stands for
the all-ones vector), there is a corresponding Markov chain given by pij = Pij .
Such a matrix is called a stochastic matrix; and for every stochastic matrix
there is a corresponding finite Markov chain and vice versa.
(s+t)
The general formula for (s+t)-step transition probabilities is that pij =
P (s) (t)
k pik pkj . This is known as the Chapman-Kolmogorov equation and
is equivalent to the matrix identity P s+t = P s P t .
A distribution over states of a finite Markov chain at some time t can
be given by a row vector x, where xi = P r[Xt = i]. To compute the
distribution at time t + 1, we use the law of total probability: Pr [Xt+1 = j] =
i Pr [Xt = i] Pr [Xt+1 = j | Xt = i] =
P P
i xi pij . Again we have the formula
for matrix multiplication (where we treat x as a 1 n matrix); so the
distribution vector at time t + 1 is just xP , and at time t + n is xP n .
We like Markov chains for two reasons:
1. They describe what happens in a randomized algorithm; the state
space is just the set of all states of the algorithm, and the Markov
property holds because the algorithm cant remember anything that
isnt part of its state. So if we want to analyze randomized algorithms,
we will need to get good at analyzing Markov chains.
2. They can be used to do sampling over interesting distributions. Under
appropriate conditions (see below), the state of a Markov chain con-
verges to a stationary distribution. If we build the right Markov
chain, we can control what this stationary distribution looks like, run
the chain for a while, and get a sample close to the stationary distribu-
tion.
In both cases we want to have a bound on how long it takes the Markov
chain to converge, either because it tells us when our algorithm terminates,
or because it tells us how long to mix it up before looking at the current
state.
9.1.1 Examples
A fair 1 random walk. The state space is Z, the transition probabilities
are pij = 1/2 if |i j| = 1, 0 otherwise. This is an example of a Markov
chain that is also a martingale.
A fair 1 random walk on a cycle. As above, but now the state space
is Z/m, the integers mod m. An example of a finite Markov chain.
Random walks with absorbing/reflecting barriers.
Random walk on a graph G = (V, E). The state space is V , the

transition probabilities are puv = 1/d(u) if uv E.
One can also have more general transition probabilities, where the
probability of traversing a particular edge is a property of the edge and
not the degree of its source. In principle we can represent any Markov
chain as a random walk on graph in this way: the states become vertices,
and the transitions become edges, each labeled with its transition
probability. Its conventional in this representation to exclude edges
with probability 0 and include self-loops for any transitions i i.
If the resulting graph is small enough or has a nice structure, this can
be a convenient way to draw a Markov chain.
The Markov chain given by Xt+1 = Xt + 1 with probability 1/2, and 0

with probability 1/2. The state space is N.
A finite-state machine running on a random input. The sequence

of states acts as a Markov chain, assuming each input symbol is
independent of the rest.
A classic randomized algorithm for 2-SAT, due to Papadimitriou [Pap91].

State is a truth-assignment. The transitional probabilities are messy
but arise from the following process: pick an unsatisfied clause, pick
one of its two variables uniformly at random, and invert it. Then
there is an absorbing state at any satisfying assignment. With a bit
of work, it can be shown that the Hamming distance between the
current assignment and some satisfying assignment follows a random
walk biased toward 0, giving a satisfying assignment after O(n2 ) steps
on average.
A similar process works for 2-colorability, 3-SAT, 3-colorability, etc.,

although for NP-hard problems, it may take a while to reach an
absorbing state. The constructive Lovsz Local Lemma proof from
11.3.5 also follows this pattern.
9.2 Convergence of Markov chains

We want to use Markov chains for sampling. Typically this means that we
have some subset S of the state space, and we want to know what proportion
of the states land in S. If we cant sample states from from the state space a
priori, we may be able to get a good approximation by running the Markov
chain for a while and hoping that it converges to something predictable.
To show this, we will proceed though several steps:
1. We will define a class of distributions, known as stationary distribu-

tions, that we hope to converge to (9.2.1
2. We will define a distance between distributions, the total variation

distance (9.2.2).
3. We will define the mixing time of a Markov chain as the minimum

time for the distribution of the position of a particle to get within a
given total variation distance of the stationary distribution (9.2.3).
4. We will describe a technique called coupling that can be used to

bound total variation distance (9.2.4), in terms of the probability that
two dependent variables X and Y with the distributions we are looking
at are or are not equal to each other.
5. We will define reducible and periodic Markov chains, which have

structural properties that prevent convergence to a unique stationary
distribution (9.2.5).
6. We will use a coupling between two copies of a Markov chain to

show that any Markov chain that does not have these properties does
converge in total variation distance to a unique stationary distribution

(9.2.6).
7. Finally, we will show that if we can construct a coupling between two

copies of a particular chain that causes both copies to reach the same
state quickly on average, then we can use this expected coupling time
to bound the mixing time of the chain that we defined previously. This
will give us a practical tool for showing that many Markov chains not
only converge eventually but converge at a predictable rate, so we can
tell when it is safe to stop running the chain and take a sample (9.3).
Much of this section follows the approach of Chapter 4 of Levin et
al. [LPW09].
9.2.1 Stationary distributions

For a finite Markov chain, there is a transition matrix P for which each row
sums to 1. We can write this fact compactly as P 1 = 1, where 1 is the
all-ones column vector. This means that 1 is a right eigenvector of P with
eigenvalue 1.3 Because the left eigenvalues of a matrix are equal to the right
eigenvalues, this means that there will be at least one left eigenvector such
that P = , and in fact it is possible to show that there is at least one
such that represents a probability distribution in that each i 0 and
i = 1. Such a distribution is called a stationary distribution4 of the
P
Markov chain, and if is unique, the probability i is called the stationary

probability for i.
Every finite Markov chain has at least one stationary distribution, but
it may not be unique. For example, if the transition matrix is the identity
matrix (meaning that the particle never moves), then all distributions are
stationary.
If a Markov chain does have a unique stationary distribution, we can
calculate it from the transition matrix, by observing that the equations
P = P
and
1 = 1
3
Given a square matrix A, a vector x is a right eigenvector of A with eigenvalue is
Ax = x. Similarly, a vector y is a left eigenvalue of A with eigenvalue if yA = A.
4
Spelling is important here. A stationery distribution would mean handing out office
supplies.
together give n + 1 equations in n unknowns, which we can solve for .

Often this will be impractical, especially if our state space is large enough
or messy enough that we cant write down the entire matrix. In these cases
we may be able to take advantage of a special property of some Markov
chains, called reversibility.
9.2.1.1 Stationary distributions of reversible chains

A Markov chain with transition matrix P is reversible if there is a distribu-
tion such that, for all i and j,
i pij = j pji . (9.2.1)
This is called the detailed balance equationsit says that in the station-
ary distribution, the probability of seeing a transition from i to j is equal to
the probability of seeing a transition from j to i). If this is the case, then
P P
i i pij = i j pji = j , which means that is stationary.
This often gives a very quick way to compute the stationary distribution,
since if we know i , and pij 6= 0, then j = i pij /pji . If the transition
probabilities are reasonably well-behaved (for example, if pij = pij for all
i, j), we may even be able to characterize the stationary distribution up to
a constant multiple even if we have no way to efficiently enumerate all the
states of the process.
What (9.2.1) says is that if we start in the stationary distribution and ob-
serve either a forward transition hX0 , X1 i or a backward transition hX1 , X0 i,
we cant tell which is which; Pr [X0 = i X1 = j] = Pr [X1 = i X0 = j].
This extends to longer sequences. The probability that X0 = i, X1 = j, and
X2 = k is given by i pij pjk = pji j pjk = pji pkj k , which is the probability
that X0 = k, X1 = j, and X0 = i. (A similar argument works for finite
sequences of any length.) So a reversible Markov chain is one with no arrow
of time in the stationary distribution.
A typical reversible Markov chain is a random walk on a graph, where
a step starting from a vertex u goes to one of its neighbors v, which each
1
neighbor chosen with probability d(u) . This has a stationary distribution
d(u)
u = P
u d(u)
d(u)
= ,
2|E|
which satisfies
d(u 1 d(v 1
u puv = = = v pvu .
2|E| d(u) 2|E| d(v)
If we dont know in advance, we can often guess it by observing that
i pij = j pji
implies
pij
j = i , (9.2.2)
pji
provided pij 6= 0. This gives us the ability to calculate k starting from any
initial state i as long as there is some chain of transitions i = i0 i1 i2
. . . i` = k where each step im im+1 has pim ,im+1 6= 0. For a random walk
on a graph, this implies that is unique as long as the graph is connected.
This of course only works for reversible chains; if we try to do this with a
non-reversible chain, we are likely to get a contradiction.
For example, if we consider a biased random walk on the n-cycle, which
moves from i to (i + 1) mod n with probability p and in the other direction
with probability q = 1 p, then applying (9.2.2) repeatedly would give
i
i = 0 pq . This is not a problem when p = q = 1/2, since we get i = 0
for all i and can deduce that i = 1/n is the unique stationary distribution.
But if we try it for p = 2/3, then we get i = 0 2i , which is fine up until we
hit 0 = n = 0 2n . So for p 6= q, this process is not reversible, which is not
surprising if we realize that the n = 60, p = 1 case describes precisely the
movement of the second hand on a clock.5
Reversible chains have a special role in Markov chain theory, because
some techniques for proving convergence only apply to reversible chains (see
9.4.4). They are also handy for sampling from large, irregular state spaces,
because by tuning the transition probabilities locally we can often adjust
the relative likelihoods of states in the stationary distribution to be closer to
what we want (see 9.2.7).
9.2.2 Total variation distance

Definition 9.2.1. Let X and Y be random variables defined on the same
probability space. Then the total variation distance between X and Y ,
5
It happens to be the case that i = 1/n is a stationary distribution for any value of p,
we just cant prove this using (9.2.1).
written dT V (X, Y ) is given by

dT V (X, Y ) = sup (Pr [X A] Pr [Y A]) , (9.2.3)
A
where the supremum is taken over all sets A for which Pr [X A] and
Pr [Y A] are both defined.6
An equivalent definition is
dT V (X, Y ) = sup|Pr [X A] Pr [Y A]|.
A
The reason this is equivalent is that if Pr [X A] Pr [Y A] is negative,

we can replaced A by its complement.
Less formally, given any test set A, X and Y land in A with probabilities
that differ by at most dT V (X, Y ). This is usually what we want for sampling,
since this says that if we are testing some property (represented by A) of
the states we are sampling, the answer we get for how likely this property is
to occur is close to the correct answer.
Total variation distance is a property of distributions, and is not affected
by dependence between X and Y . For finite Markov chains, we can define
the total variation distance between two distributions x and y as
X X
dT V (x, y) = max (xi yi ) = max| (xi yi )|.
A A
iA iA
A useful fact is that dT V (x, y) is directly connected to the `1 distance between

x and y. If we let B = {i} xi yi , then
X
dT V (x, y) = max (xi yi )
A
iA
X
(xi yi ),
iB
because if A leaves out an element of B, or includes and element of B, this

can only reduce the sum. But if we consider B instead, we get
X
dT V (y, x) = max (yi xi )
A
iA
X
(yi xi ).
iB
6
P For discrete random variables, this just means all A, since we can write Pr [X A] as
xA
Pr [X = x]; we can also replace sup with max for this case. For continuous random
variables, we want that X 1 (A) and Y 1 (A) are both measurable. If our X and Y range
over the states of a countable Markov chain, we will be working with discrete random
variables, so we can just consider all A.
Now observe that

X
kx yk1 = |xi yi |
i
X X
= (xi yi ) + (yi , xi )
iB iB
= dT V (x, y) + dT V (y, x)
= 2dT V (x, y).
So dT V (x, y) = 12 kx yk1 .
9.2.2.1 Total variation distance and expectation

Sometimes its useful to translate a bound on total variation to a bound on
the error when getting the expectation of a random variable. The following
lemma may be handy:
Lemma 9.2.2. Let x and y be two distributions of some discrete random
variable Z. Let Ex (Z) and Ey (Z) be the expectations of Z with respect to
each of these distributions. Suppose that |Z| M always. Then
1
|Ex (Z) Ey (Z)| M dT V (x, y). (9.2.4)
2
Proof. Compute

X
|Ex (Z) Ey (Z)| = z Pr(Z = z) Pr(Z = z)

x y
z

X

|z| Pr(Z = z) Pr(Z = z)
x y
z

X
M Pr
(Z = z) Pr(Z = z)
x y
z
M kx yk1
1
= M dT V (x, y).
2
9.2.3 Mixing time

We are going to show that well-behaved finite Markov chains eventually
converge to some stationary distribution in total variation distance. This
means that for any > 0, there is a mixing time tmix () such that for any
initial distribution x and any t tmix (),
dT V (xP t , ) .
It is common to standardize as 1/4: if we write just tmix , this means

tmix (1/4). The choice of 1/4 is somewhat arbitrary, but has some nice
technical properties that we will see below.
9.2.4 Coupling of Markov chains

A coupling of two random variables X and Y is a joint distribution on
hX, Y i that gives the correct marginal distribution for each of X and Y
while creating a dependence between them with some desirable property (for
example, minimizing total variation distance or maximize Pr [X = Y ]).
We will use couplings between Markov chains to prove convergence.
Here we take two copies of the chain, once of which starts in an arbitrary
distribution, and one of which starts in the stationary distribution, and show
that we can force them to converge to each other by carefully correlating their
transitions. Since the second chain is always in the stationary distribution,
this will show that the first chain converges to the stationary distribution as
well.
The tool that makes this work is the Coupling Lemma:
Lemma 9.2.3. For any discrete random variables X and Y ,
dT V (X, Y ) Pr [X 6= Y ] .
Proof. Let A be any set for which Pr [X A] and Pr [Y A] are defined.

Then
Pr [X A] = Pr [X A Y A] + Pr [X A Y 6 A] ,
Pr [Y A] = Pr [X A Y A] + Pr [X 6 A Y A] ,
and thus
Pr [X A] Pr [Y A] = Pr [X A Y 6 A] Pr [X 6 A Y A]
Pr [X A Y 6 A]
Pr [X 6= Y ] .
Since this holds for any particular set A, it also holds when we take the
maximum over all A to get dT V (X, Y ).
For Markov chains, our goal will be to find a useful coupling between a se-
quence of random variables X0 , X1 , X2 , . . . corresponding to the Markov chain
starting in an arbitrary distribution with a second sequence Y0 , Y1 , Y2 , . . .
corresponding to the same chain starting in a stationary distribution. What
will make a coupling useful is if Pr [Xt 6= Yt ] is small for reasonably large t:
since Yt has the stationary distribution, this will show that dT V (xP t , ) is
also small.
Our first use of this technique will be to show, using a rather generic
coupling, that Markov chains with certain nice properties converge to their
stationary distribution in the limit. Later we will construct specialized
couplings for particular Markov chains to show that they converge quickly.
9.2.5 Irreducible and aperiodic chains

Not all chains are guaranteed to converge to their stationary distribution.
If some states are not reachable from other states, it may be that starting
in one part of the chain will keep us from ever reaching another part of the
chain. Even if the chain is not disconnected in this way, we might still not
converge if the distribution oscillates back and forth due to some periodicity
in the structure of the chain. But if these conditions do not occur, then we
will be able to show convergence.
Let ptij be Pr [Xt = j | X0 = i].
A Markov chain is irreducible if, for all i and j, there exists some t
such that ptij 6= 0. This says that we can reach any state from any other
state if we wait long enough. If we think of a directed graph of the Markov
chain where the states are vertices and each edge represents a transition that
occurs with nonzero probability, the Markov chain is irreducible if its graph
is strongly connected.
The period of a state i of a Markov chain is gcd t > 0 ptii 6= 0 . If

the period of i is m, then we can only return to i at times that are multiples
of m. If m = 1, state i is said to be aperiodic. A Markov chain as a whole
is aperiodic if all of its states are aperiodic. In graph-theoretic terms, this
means that the graph of the chain is not k-partite for any k > 1. Reversible
chains are also an interesting special case: if a chain is reversible, it cant
have a period greater than 2, since we can always step off a node and step
back.
If our Markov chain is not aperiodic, we can make it aperiodic by flipping
a coin at each step to decide whether to move or not. This gives a lazy
Markov chain whose transition probabilities are given by 12 pij when i 6= j
and 12 + 12 pij when i = j. This doesnt affect the stationary distribution: if
we replace our transition

matrix P with a new transition matrix P +I
2 , and
P +I 1 1 1 1
P = , then 2 = 2 P + 2 I = 2 + 2 = .
Unfortunately there is no quick fix for reducible Markov chains. But
since we will often be designing the Markov chains we will be working with,
we can just take care to make sure they are not reducible to begin with.
We will later need the following lemma about aperiodic Markov chains,
which is related to the Frobenius problem of finding the minimum value
that cannot be constructed using coins of given denominations:
Lemma 9.2.4. Let i be an aperiodic state of some Markov chain. Then

there is a time t0 such that ptii 6= 0 for all t t0 .
Proof. Let S = {t | pii (t) 6= 0}. Since gcd(S) = 1, there is a finite subset S 0
of S such that gcd S 0 . Write the elements of S 0 as m1 , m2 , . . . , mk and let
M = kj=1 mj . From the extended Euclidean algorithm, there exist integer
Q
coefficients aj with |aj | M/mj such that kj=1 aj mj = 1. We would like

P
to use each aj as the number of times to go around the length mj loop from
i to i. Unfortunately many of these aj will be negative.
To solve this problem, we replace aj with bj = aj + M/mj . This makes all
the coefficients non-negative, and gives kj=1 bj mj = kM + 1. This implies
P
that there is a sequence of loops that gets us from i back to i in kM + 1

steps, or in other words that pkM ii
+1
6= 0. By repeating this sequence ` times,
`kM +`
we can similarly show that pii 6= 0 for any `.
We can also pad any of these sequences out by as many copies of M as
we like. In particular, given `kM + `, where ` {0, . . . , M 1}, we can add
(kM `)M to it to get (kM )2 + `. This means that we can express any
t (kM )2 , . . . , (kM )2 + M 1 as a sum of elements of S, or equivalently

that ptii =
6 0 for any such t. But for larger t, we can just add in more copies
of M . So in fact ptii 6= 0 for all t t0 = (kM )2 .
9.2.6 Convergence of finite irreducible aperiodic Markov chains

We can now show:
Theorem 9.2.5. Any finite irreducible aperiodic Markov chain converges to

a unique stationary distribution in the limit.
Proof. Consider two copies of the chain {Xt } and {Yt }, where X0 starts in
some arbitrary distribution x and Y0 starts in a stationary distribution .
Define a coupling between {Xt } and {Yt } by the rule: (a) if Xt = 6 Yt , then
Pr [Xt+1 = j Yt+1 = j 0 | Xt = i Yt = i0 ] = pij pi0 j 0 ; and (b) if Xt = Yt ,
then Pr [Xt+1 = Yt+1 = j | Xt = Yt = i] = pij . Intuitively, we let both chains

run independently until they collide, after which we run them together. Since
each chain individually moves from state i to state j with probability pij
in either case, we have that Xt evolves normally and Yt remains in the
stationary distribution.
Now let us show that dT V (xP t , ) Pr [Xt 6= Yt ] goes to zero in the
limit. Pick some state i. Let r be the maximum over all states j of the first
passage time fji where fji is the minimum time t such that ptji 6= 0. Let s
be a time such that ptii 6= 0 for all t s (the existence of such an s is given
by Lemma 9.2.4).
Suppose that at time `(r + s), where ` N, X`(r+s) = j = 6 j 0 = Y`(r+s) .
0 0
Then there are times `(r + s) + u and `(r + s) + u , where u, u r, such that
X reaches i at time `(r + s) + u and Y reaches i at time `(r + s) + u0 with
nonzero probability. Since (r + s u) s, then having reached i at these
times, X and Y both return to i at time `(r + s) + (r + s) = (` + 1)(r + s) with
nonzero hprobability. Let > 0 be ithe product of hthese nonzero probabilities;
i
then Pr X(`+1)(r+s) 6= Y(`+1)(r+s) (1 ) Pr X`(r+s) = Y`(r+s) , and in
general we have Pr [Xt 6= Yt ] (1 )bt/(r+s)c , which goes to zero in the
limit. This implies that dT V (xP t , ) also goes to zero in the limit (using the
Coupling Lemma), and since any initial distribution (including a stationary
distribution) converges to , is the unique stationary distribution as
claimed.
This argument requires that the chain be finite, because otherwise we

cannot take the maximum over all the first passage times. For infinite Markov
chains, it is not always enough to be irreducible and aperiodic to converge
to a stationary distribution (or even to have a stationary distribution at all).
However, with some additional conditions a similar result can be shown: see
for example [GS01, 6.4].
9.2.7 The Metropolis-Hastings algorithm

The basic idea of the Metropolis-Hastings algorithm[MRR+ 53, Has70]
(sometimes just called Metropolis) is that we start with a reversible Markov
chain P with a known stationary distribution , but wed rather get a chain
Q on the same states with a different stationary distribution , where
i = f (i)/ j f (j) is proportional to some function f 0 on states that we
P
can compute easily.

A typical application is that we want to sample according to Pr [i | A],
but A is highly improbable (so we cant just use rejection sampling, where
we sample random points from the original distribution until we find one for
which A holds), and Pr [i | A] is easy to compute for any fixed i but tricky
to compute for arbitrary events (so we cant use divide-and-conquer). If we
let f (i) Pr [i | A], then Metropolis-Hastings will do exactly what we want,
assuming it converges in a reasonable amount of time.
Let q be the transition probability for Q. Define, for i 6= j,
!
i f (j)
qij = pij min 1,
j f (i)
!
i j
= pij min 1,
j i
and let qii be whatever probability is left over. Now consider two states i
and j, and suppose that i f (j) j f (i). Then
qij = pij
which gives
i qij = i pij ,
while
j qji = j pji (j i /i j )
= pji (j i /i )
= i (pji j )/i
= i (pij i )/i
= i pij
(note the use of reversibility of P in the second-to-last step). So we have

j qji = i pij = i qij and Q is a reversible Markov chain with stationary
distribution .
We can simplify this when our underlying chain P has a uniform stationary
distribution (for example, when its the random walk on a graph with
maximum degree , where we traverse each edge with probability 1/).
Then we have i = j for all i, j, so the new transition probabilities qij are
1
just min(1, f (j)/f (i)). Most of our examples of reversible chains will be
instances of this case (see also [MU05, 10.4.1]).
9.3 The coupling method

In order to use the stationary distribution of a Markov chain to do sampling,
we need to have a bound on the rate of convergence to tell us when it is safe
to take a sample. There are two standard techniques for doing this: coupling,
where we show that a copy of the process starting in an arbitrary state can
be made to converge to a copy starting in the stationary distribution; and
spectral methods, where we bound the rate of convergence by looking at the
second-largest eigenvalue of the transition matrix. Well start with coupling
because it requires less development.
(See also [Gur00] for a survey of the relationship between the various
methods.)
Note: these notes will be somewhat sketchy. If you want to read more
about coupling, a good place to start might be Chapter 11 of [MU05]; Chapter
4-3 (http://www.stat.berkeley.edu/~aldous/RWG/Chap4-3.pdf) of the
unpublished but nonetheless famous Aldous-Fill manuscript (http://www.
stat.berkeley.edu/~aldous/RWG/book.html, [AF01]), which is a good
place to learn about Markov chains and Markov chain Monte Carlo methods
in general; or even an entire book [Lin92]. Well mostly be using examples
from the Aldous-Fill text.
9.3.1 Random walk on a cycle

Lets suppose we do a random walk on Zm , where to avoid periodicity at
each step we stay put with probability 1/2, move counterclockwise with
probability 1/4, and move clockwise with probability 1/4: in other words, we
are doing a lazy unbiased random walk. Whats a good choice for a coupling
to show this process converges quickly?
Specifically, we need to create a joint process (Xt , Yt ), where each of the
marginal processes X and Y looks like a lazy random walk on Zm , X0 has
whatever distribution our real process starts with, and Y0 has the stationary
distribution. Our goal is to structure the combined process so that Xt = Yt
as soon as possible.
Let Zt = Xt Yt (mod m). If Zt = 0, then Xt and Yt have collided and
we will move both together. If Zt 6= 0, then flip a coin to decided whether
to move Xt or Yt ; whichever one moves then moves up or down with equal
probability. Its not hard to see that this gives a probability of exactly 1/2
that Xt+1 = Xt , 1/4 that Xt+1 = Xt + 1, and 1/4 that Xt+1 = Xt 1, and
similarly for Yt . So the transition functions for X and Y individually are the
same as for the original process.
Whichever way the first flip goes, we get Zt+1 = Zt 1 with equal
probability. So Z acts as an unbiased random walk on Zm with an absorbing
barriers at 0; this is equivalent to a random walk on 0 . . . m with absorbing
barriers at both endpoints. The expected time for this random walk to reach
a barrier starting from an arbitrary initial state is at most m2 /4, so if is
the first time at which X = Y , we have E [ ] m2 /4.7
Using Markovs inequality, after t = 2(m2 /4) = m2 /2 steps we have
Pr [Xt 6= Yt ] = Pr > m2 /2 mE[ ]

2 /2 1/2. We can also iterate the whole
argument, starting over in whatever state we are in at time t if we dont
converge. This gives at most a 1/2 chance of not converging for each interval
of m2 /2 steps. So after m2 /2 steps we will have Pr [Xt 6= Yt ] 2 . This
gives tmix () 12 m2 lg(2/), where as before tmix () is the time needed to
make dT V (Xt , ) (see 9.2.3).
The choice of 2 for the constant in Markovs inequality could be improved.
The following lemma gives an optimized version of this argument:
Lemma 9.3.1. Let the expected coupling time, at which two coupled pro-
cesses {Xt } and {Yt } starting from an arbitrary state are first equal, be T .
Then dT V (XT , YT ) for T T e ln(2/).
Proof. Essentially the same argument as above, but replacing 2 with a

constant c to be determined. Suppose we restart the process every cT
steps. Then at time t we have a total variation bounded by 2cbt/cT c . The
expression ct/cT is minimized by minimizing c1/c or equivalently ln c/c,
which occurs at c = e. This gives a time to reach an /2 probability of of
T e ln(2/).
Its worth noting that the random walk example was very carefully rigged
to make the coupling argument clean. A similar argument still works (perhaps
with a change in the bound) for other ergodic walks on the ring, but the
details are messier.
9.3.2 Random walk on a hypercube

Start with a bit-vector of length n. At each step, choose an index uniformly
at random, and set the value of the bit-vector at that index to 0 or 1 with
equal probability. How long until we get a nearly-uniform distribution over
all 2n possible bit-vectors?
7
If we know that Y0 is uniform, then Z0 is also uniform, and we can use this fact to get
a slightly smaller bound on E [ ], around m2 /6. But this will cause problems if we want to
re-run the coupling starting from a state where Xt and Yt have not yet converged.
Here we apply the same transformation to both the X and Y vec-

tors. Its easy to see that the two vectors will be equal once every in-
dex has been selected once. The waiting time for this to occur is just
the waiting time nHn for the coupon collector problem. We can either
use this expected time directly to show that the process mixes in time
O(n log n log(1/)) as above, or apply a sharp concentration bound to the
coupon collector process. It is known that (see [MU05, 5.4.1] or [MR95,
3.6.3]), limn Pr [T n(ln n + c)] = 1 exp( exp(c))), so in the limit
n ln n + n ln ln(1/(1 )) = n ln n + O(n log(1/)) would seem to be enough.
But this is a little tricky: we dont know from this bound alone how fast the
probability converges as a function of n, so to do this right we need to look
into the bound in more detail.
Fortunately, we are looking at large enough values of c that we can get a
bound that is just as good using a simple argument. We have
t
1

Pr [there is an empty bin after t insertions] n 1
n
t/n
ne ,
and setting t = n ln n + cn gives a bound of e c. We can then set c = ln(1/)
to get a Pr [Xt 6= Yt ] at time n ln n + n ln(1/).
We can improve the bound slightly by observing that, on average, half
the bits in X0 and Y0 are already equal; doing this right involves summing
over a lot of cases, so we wont do it.
This is an example of a Markov chain with the rapid mixing property:
the mixing time is polylogarithmic in the number of states (2n in this case)
and 1/. For comparison, the random walk on the ring is not rapid mixing,
because the coupling time is polynomial in n = m rather than log n.
9.3.3 Various shuffling algorithms

Here we have a deck of n cards, and we repeatedly apply some random
transformation to the deck to converge to a stationary distribution that
is uniform over all permutations of the cards (usually this is obvious by
symmetry, so we wont bother proving it). Our goal is to show that the
expected coupling time at which our deck ends up in the same permutation
as an initially-stationary deck is small. We do this by counting how many
cards St are in the same position in both decks, and showing that, for a
suitable coupling, (a) this quantity never decreases, and (b) it increases with
some nonzero probability at each step. The expected coupling time is then
k 1/ Pr [St+1 = k + 1 | St = k].
P
9.3.3.1 Move-to-top
This is a variant of card shuffling that is interesting mostly because it gives
about the easiest possible coupling argument. At each step, we choose one
of the cards uniformly at random (including the top card) and move it to
the top of the deck. How long until the deck is fully shuffled, i.e., until the
total variation distance between the actual distribution and the stationary
distribution is bounded by ?
Here the trick is that when we choose a card to move to the top in the
X process, we choose the same card in the Y process. Its not hard to see
that this links the two cards together so that they are always in the same
position in the deck in all future states. So to keep track of how well the
coupling is working, we just keep track of how many cards are linked in this
way, and observe that as soon as n 1 are, the two decks are identical.
Note: Unlike some of the examples below, we dont consider two cards to
be linked just because they are in the same position. We are only considering
cards that have gone through the top position in the deck (which corresponds
to some initial segment of the deck, viewed from above). The reason is that
these cards never become unlinked: if we pick two cards from the initial
segment, the cards above them move down together. But deeper cards that
happen to match might become separated if we pull a card from one deck
that is above the matched pair while its counterpart in the other deck is
below the matched pair.
Having carefully processed the above note, given k linked cards the
probability that the next step links another pair of cards is exactly (n k)/n.
So the expected time until we get k + 1 cards is n/(n k), and if we sum
these waiting times for k = 0 . . . n 1, we get nHn , the waiting time for the
coupon collector problem. So the bound on the mixing time is the same as
for the random walk on a hypercube.
9.3.3.2 Random exchange of arbitrary cards

Here we pick two cards uniformly and independently at random and swap
them. (Note there is a 1/n chance they are the same card; if we exclude this
case, the Markov chain has period 2.) To get a coupling, we reformulate this
process as picking a random card and a random location, and swapping the
chosen card with whatever is in the chosen location in both the X and Y
processes.
First lets observe that the number of linked cards never decreases. Let
xi , yi be the position of card i in each process, and suppose xi = yi . If
neither card i nor position xi is picked, i doesnt move, and so it stays linked.
If card i is picked, then both copies are moved to the same location; it stays
linked. If position xi is picked, then it may be that i becomes unlinked; but
this only happens if the card j that is picked has xj 6= yj . In this case j
becomes linked, and the number of linked cards doesnt drop.
Now we need to know how likely it is that we go from k to k + 1 linked
cards. Weve already seen a case where the number of linked cards increases;
we pick two cards that arent linked and a location that contains cards that
arent linked. The probability of doing this is ((n k)/n)2 , so our total
expected waiting time is n2 (n k)2 = n2 k 2 n 2 /6 (see http:
P P
//en.wikipedia.org/wiki/Basel_problem for an extensive discussion of

the useful but non-obvious fact that kN+ k 2 = (2) = 2 /6.) The final
P
bound is O(n2 log(1/)).

This bound is much worse that the bound for move-to-top, which is sur-
prising. In fact, the real bound is O(n log n) with high probability, although
the proof uses very different methods (see http://www.stat.berkeley.edu/
~aldous/RWG/Chap7.pdf). This shows that the coupling method doesnt
always give tight bounds (or perhaps we need a better coupling?).
9.3.3.3 Random exchange of adjacent cards

Suppose now that we only swap adjacent cards. Specifically, we choose one
of the n positions i in the deck uniformly at random, and then swap the
cards a positions i and i + 1 (mod n) with probability 1/2. (The 1/2 is there
for the usual reason of avoiding periodicity.)
So now we want a coupling between the X and Y processes where each
1
possible swap occurs with probability 2n on both sides, but somehow we
correlate things so that like cards are pushed together but never pulled apart.
The trick is that we will use the same position i on both sides, but be sneaky
about when we swap. In particular, we will aim to arrange things so that
once some card is in the same position in both decks, both copies move
together, but otherwise one copy changes its position by 1 relative to the
1
other with a fixed probability 2n .
The coupled process works like this. Let D be the set of indices i where
the same card appears in both decks at position i or at position i + 1. Then
we do:
1
1. For i D, swap (i, i + 1) in both decks with probability 2n .
1
2. For i 6 D, swap (i, i + 1) in the X deck only with probability 2n .
1
3. For i 6 D, swap (i, i + 1) in the Y deck only with probability 2n .
|D|
4. Do nothing with probability 2n .
Its worth checking that the total probability of all these events is |D|/2n+
2(n |D|)/2n + |D|/2n = 1. More important is that if we consider only one
1
of the decks, the probability of doing a swap at (i, i + 1) is exactly 2n (since
we catch either case 1 or 2 for the X deck or 1 or 3 for the Y deck).
Now suppose that some card c is at position x in X and y in Y . If
x = y, then both x and x 1 are in D, so the only way the card can move
is if it moves in both decks: linked cards stay linked. If x = 6 y, then c
moves in deck X or deck Y , but not both. (The only way it can move
in both is in case 1, where i = x and i + 1 = y or vice verse; but in
this case i cant be in D since the copy of c at position x doesnt match
whatever is in deck Y , and the copy at position y doesnt match whats
in deck X.) In this case the distance x y goes up or down by 1 with
1
equal probability 2n . Considering x y (mod n), we have a lazy random
walk that moves with probability 1/n, with absorbing barriers at 0 and
n. The worst-case expected time to converge is n(n/2)2 = n3 /4, giving
Pr time for c to become linked n 3 /8 2 using the usual argument.

Now apply the union bound to get Pr time for every c to become linked n3 /8

n2 to get an expected coupling time of O(n3 log n).

To simplify the argument, we assumed that the deck wraps around, so
that we can swap the first and last card in addition to swapping physically
adjacent cards. If we restrict to swapping i and i + 1 for i {0, . . . , n 2},
we get a slightly different process that converges in essentially the same time
O(n3 log n). A result of David Bruce Wilson [Wil04]) shows both this upper
bound holds and that the bound is optimal up to a constant factor.
9.3.3.4 Real-world shuffling

In real life, the handful of people who still use physical playing cards tend
to use a dovetail shuffle, which is closely approximated by the reverse
of a process where each card in a deck is independently assigned to a left
or right pile and the left pile is place on top of the right pile. Coupling
doesnt really help much here. Instead, the process can be analyzed using
more sophisticated techniques due to Bayer and Diaconis [BD92]. The short
version of the result is that (log n) shuffles are needed to randomize a deck
of size n.
9.3.4 Path coupling

If the states of our Markov process are the vertices of a graph, we may be
able to construct a coupling by considering a path between two vertices and
showing how to shorten this path on average at each step. This technique is
known as path coupling. Typically, the graph we use will be the graph of
possible transitions of the underlying Markov chain (possibly after making
all edges undirected).
There are two ideas at work here. The first is that the expected distance
E [d(Xt , Yt )] between Xt and Yt in the graph gives an upper bound on
Pr [Xt 6= Yt ] (by Markovs inequality, since if Xt 6= Yt then d(Xt , Yt ) 1).
The second is that to show that E [d(Xt+1 , Yt+1 ) | Ft ] d(Xt , Yt ) for some
< 1, it is enough to show how to contract a single edge, that is, to show
that E [d(Xt+1 , Yt+1 | d(Xt , Yt ) = 1] . The reason is that if we have a
coupling that contracts one edge, we can apply this inductively along each
edge in the path to get a coupling between all the vertices in the path that
still leaves each pair with expected distance at most . The result for the
whole path then follows from linearity of expectation.
Formally, instead of just looking at Xt and Yt , consider a path of inter-
mediate states Xt = Z0,t Z1,t Z2,t . . . Zm,t = Yt , where d(Zi,t , Zi+1,t ) = 1 for
each i (the vertices are adjacent in the graph). We now construct a coupling
only for adjacent nodes that reduces their distance on average. The idea is
that d(Xt , Yt ) d(Zi,t , Zi+1,t ), so if the distance between each adjacent
P
pair shrinks on average, so does the total length of the path.

The coupling on each edge gives a joint conditional probability
Pr Zi,t+1 = zi0 , Zi+1,t+1 = zi+1

0

Zi,t = zi , Zi+1,t = zi+1 .
We can extract from this a conditional distribution on Zi+1,t+1 given the

other three variables:
0 Zi,t+1 = z 0 , Zi,t = zi , Zi+1,t = zi+1 .

Pr Zi+1,t+1 = zi+1 i
Multiplying these conditional probabilities together lets us compute a joint

distribution on Xt+1 , Yt+1 conditioned on Xt , Yt , which is the ordinary cou-
pling we really want.
Its worth noting that the path is entirely notional, and we dont actually
use it to construct an explicit coupling or even keep it around between steps
of the coupled Markov chains. The only purpose of Z0 , Z1 , . . . , Zm is to show
that X and Y move closer together. Even though we could imagine that we
are coalescing these nodes together to create a new path at each step (or
throwing in a few extra nodes if some Zi , Zi+1 move away from each other),
we could also imagine that we start with a fresh path at each step and throw
it away as soon as it has done its job.
9.3.4.1 Sampling graph colorings

For example, lets look at sampling k-colorings of a graph with maximum
degree . We will assume that k 2 + 1.8 For smaller values of k, its
possible that the chain still converges, but our analysis will not show this.
Consider the following chain on proper k-colorings of a graph with
maximum degree . At each step, we choose one of the n nodes v, compute
the set S of colors not found on any of vs neighbors, and recolor v with a
color chosen uniformly from S (which may be its original color).
Suppose pij 6= 0. Then i and j differ in at most one place v, and so the set
S of permitted colors for each processthose not found on vs neighborsare
1
the same. This gives pij = pji = n|S| , and the detailed balance equations
(9.2.1) hold when i is constant for all i. So we have a reversible Markov
chain with a uniform stationary distribution. Now we will apply a path
coupling to show that we converge to this stationary distribution reasonably
quickly when k is large enough.
Well think of colorings as vectors. Given two colorings x and y, let d(x, y)
be the Hamming distance between them, which is the number of nodes u
for which xu 6= yu . To show convergence, we will construct a coupling that
shows that d(X t , Y t ) converges to 0 over time starting from arbitrary initial
points X 0 and Y 0 .
A complication is that its not immediately evident that the length of
the shortest path from X t to Y t in the transition graph of our Markov chain
is d(X t , Y t ). The problem is that it may not be possible to transform X t
into Y t one node at a time without producing improper colorings. With
enough colors, we can explicitly construct a short path between X t and Y t
that uses only proper colorings; but for this particular process it is easier to
simply extend the Markov chain to allow improper colorings, and show that
our coupling works anyway. This also allows us to start with an improper
coloring for X 0 if we are particularly lazy. The stationary distribution is
not affected, because if i is a proper coloring and j is an improper coloring
that differs from i in exactly one place, we have pij = 0 and pji 6= 0, so the
detailed balance equations hold with j = 0.
8
An analysis based on a standard coupling for a different version of the Markov chain
that works for the same bound on k is given in [MU05, 11.5]. Even more sophisticated
results and a history of the problem can be found in [DGM02].
The natural coupling to consider given adjacent X t and Y t is to pick the

same node and the same new color for both, provided we can do so. If we
pick the one node v on which they differ, and choose a color that is not used
by any neighbor (which will be the same for both copies of the process, since
all the neighbors have the same colors), then we get X t+1 = Y t+1 ; this event
occurs with probability at least 1/n. If we pick a node that is neither v nor
adjacent to it, then the distance between X and Y doesnt change; either
both get a new identical color or both dont.
Things get a little messier when we pick some node u adjacent to v, an
event that occurs with probability at most /n. Let c be the color of v in
X t , c0 the color of u in Y t , and T the set of colors that do not appear among
the other neighbors of v. Let ` = |T | k ( 1).
Conditioning on choosing u to recolor, X t+1 picks a color uniformly from
T \ {c} and Y t+1 picks a color uniformly from T \ {c0 }. Wed like these colors
to be the same if possible, but these are not the same sets, and they arent
even necessarily the same size.
There are three cases:
1. Neither c nor c0 are in T . Then X t+1 and Y t+1 are choosing a new
color from the same set, and we can make both choose the same color:
the distance between X and Y is unchanged.
2. Exactly one of c and c0 is in T . Suppose that its c. Then |T \ {c}| = `1

and |T \ {c0 }| = |T | = `. Let X t+1 choose a new color c00 first. Then
let Yut+1 = c00 with probability `1 ` (this gives a probability of ` of
1
picking each color in T \ {c}, which is what we want), and let Yut+1 = c
with probability 1` . Now the distance between X and Y increases with
probability 1` .
3. Both c and c0 are in T . For each c00 in T \ {c, c0 }, let Xut+1 = Yut+1 = c00
1
with probability `1 ; since there are ` 2 such c00 , this accounts for `2
`1
of the probability. Assign the remaining `1 1
to Xut+1 = c0 , Yut+1 = c.
In this case the distance between X and Y increases with probability
1
`1 , making this the worst case.
Putting everything together, we have a 1/n chance of picking a node

that guarantees to reduce d(X, Y ) by 1, and at most a /n chance of
1
picking a node that may increase d(X, Y ) by at most `1 on average, where
1
` k + 1, giving a maximum expected increase of n k . So
h i 1 1
E d(X t+1 , Y t+1 ) d(X t , Y t ) d(X t , Y t ) = 1 +

n n k
1

= 1 +
n k
1 (k ) +

=
n k
1 k 2

= .
n k
So we get
h i
dT V (X t , Y t ) Pr X t 6= Y t
1 k 2 t
h i
1 E d(X 0 , Y 0 )
n k
t k 2

exp n.
n k
For fixed k and with k > 2, this is e(t/n) n, which will be less than
for t = (n(log n + log(1/))).
When k = 2, it may be possible to use a variant of this argument to
prove convergence, since in this case the expected change in d(X, Y ) at each
step is at most 0, and it changes by 1 or more with probability (1/n). So we
might be able to argue that we have a random walk that is at worst unbiased
and at best biased in the direction we like. The convergence bound we would
get from this is not so goodO(n3 log(1/))but it is still polylogarithmic
in the size of the state space. However, there is an important step that needs
to be filled in to make this argument work: we need to show that a typical
expectation-0 random change in the length of one part of the path is not
exactly compensated for by perfectly anticorrelated expectation-0 change in
the length of another part of the path, and we havent done this. This issue
doesnt come up in standard path coupling because we dont mind if different
parts of the path are anticorrelated as long as they all shrink on average,
and indeed any anticorrelation just makes the shrinkage more predictable.
For k < 2, the argument collapses completely: its more likely than
not that X and Y move away from each other, so even if Xs distribution
converges to the stationary distribution, we wont see X and Y coalesce any
time soon using this particular coupling.
9.3.4.2 Metropolis-Hastings and simulated annealing

Recall that the Metropolis-Hastings algorithm constructs a reversible Markov
chain with a desired stationary distribution from any reversible Markov chain
on the same states (see 9.2.7 for details.)
A variant, which generally involves tinkering with the chain while its
running, is the global optimization heuristic known as simulated annealing.
Here we have some function g that we are trying to minimize. So we set
f (i) = exp(g(i)) for some > 0. Running Metropolis-Hastings gives
a stationary distribution that is exponentially weighted to small values of
g; if i is the global minimum and j is some state with high g(j), then
(i) = (j) exp((g(j) g(i)), which for large enough goes a long way
towards compensating for the fact that in most problems there are likely to
be exponentially more bad js than good is. The problem is that the same
analysis applies if i is a local minimum and j is on the boundary of some
depression around i; large means that it is exponentially unlikely that we
escape this depression and find the global minimum.
The simulated annealing hack is to vary over time; initially, we set
small, so that we get conductance close to that of the original Markov chain.
This gives us a sample that is roughly uniform, with a small bias towards
states with smaller g(i). After some time we increase to force the process
into better states. The hope is that by increasing slowly, by the time we
are stuck in some depression, its a deep oneoptimal or close to it. If it
doesnt work, we can randomly restart and/or decrease repeatedly to jog
the chain out of whatever depression it is stuck in. How to do this effectively
is deep voodoo that depends on the structure of the underlying chain and the
shape of g(i), so most of the time people who use simulated annealing just
try it out with some generic annealing schedule and hope it gives some
useful result. (This is what makes it a heuristic rather than an algorithm.
Its continued survival is a sign that it does work at least sometimes.)
Here are toy examples of simulated annealing with provable convergence
times.
Single peak Lets suppose x is a random walk on an n-dimensional hy-

percube (i.e., n-bit vectors where we set 1 bit at a time), g(x) = |x|, and we
want to maximize g. Now a transition that increase |x| is accepted always
and a transition that decreases |x| is accepted only with probability e .
For large enough , this puts a constant fraction of on the single peak at
x = 1; the observation is that that there are only nk nk points with k
zeros, so the total weight of all points is at most (1) k0 nk exp(k) =
P
(1) exp(ln n)k = (1)/(1exp(ln n)) = (1)O(1) when > ln n,

P
giving (1) = (1) in this case.

So what happens with convergence? Let p = exp(). Lets try doing
a path coupling between two adjacent copies x and y of the Metropolis-
Hastings process, where we first pick a bit to change, then pick a value to
assign to it, accepting the change in both processes if we can. The expected
change in |x y| is then (1/2n)(1 p), since if we pick the bit where x
and y differ, we have probability 1/2n of setting both to 1 and probability
p/2n of setting both to 0, and if we pick any other bit, we get the same
distribution of outcomes in both processes. This gives a general bound of
E [|Xt+1 Yt+1 | | |Xt Yt |] (1 (1 + p)/2n)|Xt Yt |, from which we have
E [|Xt Yt |] exp(t(1 + p)/2n) E [|X0 Y0 |] n exp(t(1 + p)/2n). So
after t = 2n/(1 + p) ln(n/) steps, we expect to converge to within of the
stationary distribution in total variation distance. This gives an O(n log n)
algorithm for finding the peak.
This is kind of a silly example, but if we suppose that g is better disguised
(for example, g(x) could be |x r| where r is a random bit vector), then we
wouldnt really expect to do much better than O(n). So O(n log n) is not
bad for an algorithm with no brains at all.
Single peak with very small amounts of noise Now well let g :
2n N be some arbitrary Lipschitz function, that is, a function with the
property that |g(x) g(y)| |x y|, and ask for what values of p = e
the Metropolis-Hastings walk with f (i) = eg(i) can be shown to converge
quickly. Given adjacent states x and y, with xi 6= yi but xj = yj for all
j 6= i, we still have a probability of at least (1 + p)/2n of coalescing the
states by setting xi = yi . But now there is a possibility that if we try to
move to (x[j/b], y[j/b]) for some j and b, that x rejects while y does not
or vice versa (note if xj = yj = b, we dont move in either copy of the
process). Conditioning on j and b, this occurs with probability 1 p precisely
when x[j/b] < x and y[j/b] y or vice versa, giving an expected increase in
|x y| of (1 p)/2n. We still get an expected net change of 2p/2n = p/n
provided there is only one choice of j and b for which this occurs. So we
converge in time () (n/p) log(n/) in this case.9
One way to think of this is that the shape of the neighborhoods of nearby
points is similar. If I go up in a particular direction from point x, its very
likely that I go up in the same direction from some neighbor y of x.
9
P You might reasonably ask if such functions g exist. One example is g(x) = (x1 x2 ) +
i>2
xi .
If there are more bad choices for j and b, then we need a much larger value
of p: the expected net change is now (k(1p)1p)/2n = (k1(k+1)p)/2n,
which is only negative if p > (k 1)/(k + 1). This gives much weaker pressure
towards large values of g, which still tends to put us in high neighborhoods
but creates the temptation to fiddle with to try to push us even higher
once we think we are close to a peak.
9.4 Spectral methods for reversible chains

(See also http://www.stat.berkeley.edu/~aldous/RWG/Chap3.pdf, from
which many of the details in the notes below are taken.)
The problem with coupling is that (a) it requires cleverness to come up
with a good coupling; and (b) in many cases, even that doesnt workthere
are problems for which no coupling that only depends on current and past
transitions coalesces in a reasonable amount of time.10 When we run into
these problems, we may be able to show convergence instead using a linear-
algebraic approach, where we look at the eigenvalues of the transition matrix
of our Markov chain. This approach works best for reversible Markov chains,
where i pij = j pij for all states i and j and some distribution .
9.4.1 Spectral properties of a reversible chain

Suppose that P is the transition matrix of an irreducible, reversible Markov
chain. Then it has a unique stationary distribution that is a left eigen-
vector corresponding to the eigenvalue 1, which just means that P = 1.
For irreducible aperiodic chains, P will have a total of n real eigenvalues
1 > 2 3 . . . n , where 1 = 1 and n > 1. (This follows for
symmetric chains from the Perron-Frobenius theorem, which we will not
attempt to prove here; for general irreducible aperiodic reversible chains, we
will show below that we can reduce to the symmetric case.) Each eigenvalue i
has a corresponding eigenvector ui , a nonzero vector that satisfies ui P = i ui .
In Markov chain terms, these eigenvectors correspond to deviations from the
stationary distribution that shrink over time.
For example, the transition matrix
" #
p q
S=
q p
10
Such a coupling is called a causal coupling; an example of a Markov chain for which
causal couplings are known not to work is the one used for sampling perfect matchings in
bipartite graphs as described in 9.4.8.4 [KR99].
corresponding to a Markov chain on two states that stays in the same

state with probability p and switches
h i to the other
h state
i with probability
q = 1 p has eigenvectors u1 = 1 1 and u2 = 1 1 with corresponding
eigenvalues 1 = 1 and 2 = p q, as shown by computing
" #
h i p q h i h i
1 1 = p+q q+p =1 1 1
q p
and
" #
h i p q h i h i
1 1 = p q q p = (p q) 1 1 .
q p
9.4.2 Analysis of symmetric chains

To make our life easier, we will assume that in addition to being reversible,
our Markov chain has a uniform stationary distribution. Then i = j for all
i and j, and so reversibility implies pij = pji for all i and j as well, meaning
that the transition matrix P is symmetric. Symmetric matrices have the
very nice property that their eigenvectors are all orthogonal (this is the
spectral theorem), which allows for a very straightforward decomposition
of the distribution on the initial state (represented as a vector x0 ) as a linear
combination of the eigenvectors. h For i example, if initially we put all our
weight on state 1, we get x0 = 1 0 = 21 u1 + 12 u2 .
If we now take a step in the chain, we multiply x by P . We can express
this as
1 1 1 2

xP = u + u P
2 2
1 1 1
= u P + u2 P
2 2
1 1
= 1 u1 + 2 u2 .
2 2
This uses the defining property of eigenvectors, that ui P = i ui .
In general, if x = i ai ui , then xP = i ai i ui and xP t = ai ti ui .
P P P
t
For any eigenvalue i with |i | < 1, i goes to zero in the limit. So only
those eigenvectors with i = 1 survive. For irreducible aperiodic chains,
i i

these consist only of the stationary distribution = u / u 1 . For periodic

chains, there may be some additional eigenvalues with |i | = 1, but the only
possibility that arises for a reversible chain is n = 1, corresponding to a

chain with period 2.11
Assuming that |2 | |n |, as t grows large t2 will dominate the other
smaller eigenvalues, and so the size of 2 will control the rate of convergence
of the underlying Markov process.
This assumption is always true for lazy walks that stay put with proba-
bility 1/2, because we can think of such a walk as having transition matrix
1
2 (P + I), where I is the identity matrix and P is the transition matrix
of the unlazy
version
of the walk. If xP = x for some x and , then
1 1
x 2 (P + I) = 2 ( + 1) x. This means that x is still an eigenvector of the
lazy walk, and its corresponding eigenvalue is 12 ( + 1) 12 ((1) + 1) 0.
But what does having small 2 mean for total variation distance? If
xt is a vector representing the distribution of our position at time t, then
dT V = 12 ni=1 xti i = 12 xt 1 . But we know that x0 = + ni=2 ci ui
P P
ci , and xt = + ni=2 ti ci ui . So we are looking for a

P
for some coefficients
Pn
bound on i=2 ti ci ui 1 .
It turns out that it is easier to get a bound on the `2 norm ni=2 ti ci ui 2 .
P
Here we use the fact that the eigenvectors P are orthogonal.

2 This means that
the Pythagorean theorem holds and ni=2 ti ci ui 2 = ni=2 2t 2 i 2
i ci ku k2 =
P
Pn 2t 2 i i 2
i=2 i ci if we normalize each u so that ku k2 = 1. But then
2 n
X
t
x = 2t 2
i ci

2
i=2
n
X
2t 2
2 ci
i=2
n
X
= 2t
2 c2i
i=2
= 2 kx0
2t
k22 .
Now take the square root of both sides to get

t
x =< t2 x0 .

2
11
Chains with periods greater than 2 (which are never reversible) have pairs of complex-
valued eigenvalues that are roots of unity, which happen to cancel out to only produce
real probabilities in vP t . Chains that arent irreducible will have one eigenvector with
eigenvalue 1 for each final component; the stationary distributions of these chains are linear
combinations of these eigenvectors (which are just the stationary distributions on each
component).
To translate this back into dT V , we use the inequality

kxk2 kxk1 n kxk2 ,
which holds for any n-dimensional vector x. Because kx0 k1 = kk1 = 1,

kx0 k1 2 by the triangle inequality, which also gives kx0 k2 2. So
1
dT V (xt , ) = kxt k1
2

n t 0
kx k2
2 2
t2 n.
t
to get dT V (x , n) , we will need t ln(2 ) + ln
If we want n ln or
1 1
t ln(1/2) 2
ln n + ln(1/) .
1
The factor ln(1/ 2)
= ln1 2 can be approximated by 2 = 1
1
2
, which is
called the mixing rate or relaxation time of the Markov chain. Indeed,
our old friend 1 + x ex implies ln(1 + x) x, which
gives ln 2 2 1
1 1 1
and thus ln 2 12 = 2 . So 2 2 ln n + ln(1/) gives a conservative
estimate on the time needed to achieve dT V (xt , ) starting from an
arbitrary distribution.
Its worth comparing this to the mixing time tmix = arg mint dT V (xt , )
1/4 that we used with coupling. Here we have to throw in an extra factor of
1
2 ln n to get the bound down to 1, but after that the bound continues to drop
by a factor of e every 2 steps. In constrast, we have to go through another
tmix steps for each factor of 2 improvement in total variation distance, even
if the initial drop avoids the log factor. Since we cant always tell whether
coupling arguments or spectral arguments will do better for a particular
chain, and since convergence bounds are often hard to obtain no matter how
many techniques we throw at them, we will generally be happy with any
reasonable bound on either 2 or tmix .
9.4.3 Analysis of asymmetric chains

If the stationary distribution is not uniform, then in general the transition
matrix will not be symmetric. We can make it symmetric by scaling the
1/2
probability of being in the i-th state by i . The idea is to decompose
our transition matrix P as A1 , where is a diagonal matrix with
1/2
q
i
ii = i , and Aij = j Pij . Then A is symmetric, because
s
i
Aij = Pij
j
s
i j
= Pji
j i
j
r
= Pji
i
= Aji .
This means in particular that A has orthogonal eigenvectors u1 , u2 , . . . , un

with corresponding eigenvalues 1 2 . . . n . These eigenvalues carry
over to P = A1 , since (ui 1 )P = ui 1 A1 = ui A1 = i ui .
So now given an initial distribution x0 , we can first pass it through 1
and then apply the same reasoning as before to show that k1 xi1 k2
shrinks by 2 with every step of the Markov chain. The difference now is
that the initial distance is affected by the scaling we did in 1 ; so instead

of getting dT V (xt , ) t2 n, we get dT V (xt , ) t2 (min )1/2 , where min
is the smallest probability of any single node in the stationary distribution
. The bad initial x0 in this case is the one that puts all its weight on this
node, since this maximizes its distance from after scaling.
(A uniform is a special case of this, since when i = 1/n for all i,
1/2
min = (1/n)1/2 = n.)
For more details on this, see [AF01, 3.4].
So now we just need a tool for bounding 2 . For a small chain with a
known transition matrix, we can just feed it to our favorite linear algebra
library, but most of the time we will not be able to construct the matrix
explicitly. So we need a way to bound 2 indirectly, in terms of other
structural properties of our Markov chain.
9.4.4 Conductance
The conductance or Cheeger constant to (S) of a set S of states in a
Markov chain is
P
iS,j6Si pij
(S) = . (9.4.1)
(S)
This is the probability of leaving S on the next step starting from the
stationary distribution conditioned on being in S. The conductance is a
measure of how easy it is to escape from a set; it can also be thought of as a

weighted version of edge expansion.
The conductance of a Markov chain as a whole is obtained by taking the
minimum of (S) over all S that occur with probability at most 1/2:
= min (S). (9.4.2)

0<(S)1/2
The usefulness of conductance is that it bounds 2 :
Theorem 9.4.1. In a reversible Markov chain,
1 2 2 1 2 /2. (9.4.3)
The bound (9.4.3) is known as the Cheeger inequality. We wont

attempt to prove it here, but the intuition is that in order for a reversible
chain not to mix, it has to get stuck in some subset of the states. Having
high conductance prevents this disaster, while having low conductance causes
it.
For lazy walks we always have 2 = max , and so we can convert (9.4.3)
to a bound on the relaxation time:
Corollary 9.4.2.
1 2
2 2 . (9.4.4)
2
In other words, high conductance implies low relaxation time and vice
versa, up to squaring.
9.4.5 Easy cases for conductance

For very simple Markov chains we can compute the conductance directly.
Consider a lazy random walk on a cycle. Any proper subset S has at least two
outgoing edges, each of which carries a flow of 1/4n, giving S (1/2n)/(S).
If we now take the minimum of S over all S with (S) 1/2, we get 1/n,
which gives 2 2n2 . This is essentially the same bound as we got from
coupling.
Heres a slightly more complicated chain. Take two copies of Kn , where
n is odd, and join them by a path with n edges. Now consider S for a
lazy random walk on this graph where S consists of half the graph, split
at the edge in the middle of the path. There is a single outgoing edge uv,
with (u) = d(u)/2|E| = 2/(2n(n 1)n/2 + n) = 2n2 and puv = 1/4, for
(u)puv = n2 /2. By symmetry, we have (S) 1/2 as n , giving

S n2 (1/2)/(1/2) = n2 . So we have n2 /2 2 2n4 .
How does this compare to the actual mixing time? In the stationary
distribution, we have a constant probability of being in each of the copies of
Kn . Suppose we start in the left copy. At each step there is a 1/n chance
that we are sitting on the path endpoint. We then step onto the path with
probability 1/n, and reach the other end before coming back with probability
1/n. So (assuming we can make this extremely sloppy handwaving argument
rigorous) it takes at least n3 steps on average before we reach the other copy
of Kn , which gives us a rough estimate of the mixing time of (n3 ). In
this case the exponent is exactly in the middle of the bounds derived from
conductance.
9.4.6 Edge expansion using canonical paths

Here and below we are mostly following the presentation in [Gur00], but
with slightly different examples (and probably more errors).
For more complicated Markov chains, it is helpful to have a tool for
bounding conductance that doesnt depend on intuiting what sets have the
smallest boundary. The canonical paths [JS89] technique does this by
assigning a unique path xy from each state x to each state y in a way that
doesnt send too many paths across any one edge. So if we have a partition
of the state space into sets S and T , then there are |S| |T | paths from states
in S to states in T , and since (a) every one of these paths crosses an ST
edge, and (b) each ST edge carries at most paths, there must be at least
|S| |T |/ edges from S to T . Note that because there is no guarantee we
chose good canonical paths, this is only useful for getting lower bounds on
conductanceand thus upper bounds on mixing timebut this is usually
what we want.
Lets start with a small example. Let G = Cn Cm , the n m torus.
A lazy random walk on this graph moves north, east, south, or west with
probability 1/8 each, wrapping around when the coordinates reach n or m.
Since this is a random walk on a regular graph, the stationary distribution
is uniform. What is the relaxation time?
Intuitively, we expect it to be O(max(n, m)2 ), because we can think of this
two-dimensional random walk as consisting of two one-dimensional random
walks, one in the horizontal direction and one in the vertical direction, and
we know that a random walk on a cycle mixes in O(n2 ) time. Unfortunately,
the two random walks are not independent: every time I take a horizontal
step is a time I am not taking a vertical step. We can show that the expected
coupling time is O(n2 + m2 ) by running two sequential instances of the

coupling argument for the cycle, where we first link the two copies in the
horizontal direction and then in the vertical direction. So this gives us one
bound on the mixing time. But what happens if we try to use conductance?
Here it is probably easiest to start with just a cycle. Given points x
and y on Cn , let the canonical path xy be a shortest path between them,
breaking ties so that half the paths go one way and half the other. Then each
each is crossed by exactly k paths of length k for each k = 1 . . . (n/2 1),
and at most n/4 paths of length n/2 (0 if n is odd), giving a total of
(n/2 1)(n/2)/2 + n/4 = n2 /8 paths across the edge.
If we now take an ST partition where |S| = m, we get at least m(n
m)/ = 8m(n m)/n2 ST edges. This peaks at m = n/2, where we
get 2 edgesexactly the right numberand in general when m n/2 we
get at least 8m(n/2)/n2 = 4m/n outgoing edges, giving a conductance
S (1/4n)(4m/n)/(m/n) = 1/n. (This is essentially what we got before,
except we have to divide by 2 because we are doing a lazy walk. Note that
for small m, the bound is a gross underestimate, since we know that every
nonempty proper subset has at least 2 outgoing edges.)
Now lets go back to the torus Cn Cm . Given x and y, define xy to be
the L-shaped path that first changes x1 to y1 by moving the shortest distance
vertically, then changes x2 to y2 by moving the shortest distance horizontally.
For a vertical edge, the number of such paths that cross it is bounded by
n2 m/8, since we get at most n2 /8 possible vertical path segments and for
each such vertical path segment there are m possible horizontal destinations
to attach to it. For a horizontal edge, n and m are reversed, giving m2 n/8
paths. To make our life easier, lets assume n m, giving a maximum of
= n2 m/8 paths.
For |S| nm/2, this gives at least
|S| |G \ S| |S|(nm/2) 4|S|

2
=
n m/8 n
outgoing edges. We thus have
1 4|S|/n 1
(S) = .
8nm |S|/nm 2n
This gives 2 2/2 8n2 . Given that we assumed n m, this is essentially

the same O(n2 + m2 ) bound that we would get from coupling.
9.4.7 Congestion
For less symmetric chains, we weight paths by the probabilities of their
endpoints when counting how many cross each edge, and treat the flow
across the edge as a capacity. This gives the congestion of a collection of
canonical paths = {xy }, which is computed as
1 X
() = max x y ,
uvE u puv xy 3uv
and we define = min ().

The intuition here is that the congestion bounds the ratio between the
canonical path flow across an edge and the Markov chain flow across the
edge. If the congestion is not too big, this means that any cut that has a lot
of canonical path flow across it also has a lot of Markov chain flow. When
(T ) 1/2, the total canonical path flow (S)(T ) is at least (1/2)(S); so
this means that when (S) is large but less than 1/2, the Markov chain flow
leaving S is also large. This gives a lower bound on conductance.
Formally, we have the following lemma:
Lemma 9.4.3. For any reversible aperiodic irreducible Markov chain,
1
(9.4.5)
2
from which it follows that
2 82 . (9.4.6)
Proof. Pick some set of canonical paths with () = . Now pick some
ST partition with (S) = . Consider a flow where we route (x)(y) units
of flow along each path xy with x S and y T . This gives a total flow of
(S)(T ) (S)/2. We are going to show that we need a lot of capacity
across the ST cut to carry this flow, which will give the lower bound on
conductance.
For any ST edge uv, we have
1 X
x y
u puv xy 3uv
or
X
x y u puv .
xy 3uv
Since each ST path crosses at least one ST edge, we have

X
(S)(T ) = x y
xS,yT
X X
x y
uS,vT,uvE xy 3uv
X
u puv .
uS,vT,uvE
X
= u puv .
uS,vT,uvE
But then
P
uS,tT,uvE u puv
(S) =
S
(S)(T )/

(S)
(T )
=

1
.
2
To get the bound on 2 , use (9.4.4) to compute 2 2/2 82 .
9.4.8 Examples
Here are some more examples of applying the spectral method.
9.4.8.1 Lazy random walk on a line

Consider a lazy random walk on a line with reflecting barriers at 0 and n 1.
Here x = n1 for all x. There arent a whole lot of choices for canonical paths;
the obvious choice for xy with x < y is x, x+1, x+2, . . . , y. This puts (n/2)2
paths across the middle edge (which is the most heavily loaded), each of
which has weight x y = n2 . So the congestion is (1/n)(1/4)
1
(n/2)2 n2 = n,
giving a mixing time of at most 8n2 . This is a pretty good estimate.
9.4.8.2 Random walk on a hypercube

Lets try a more complicated example: the random walk on a hypercube
from 9.3.2. Here at each step we pick some coordinate uniformly at random
and set it to 0 or 1 with equal probability; this gives a transition probability

1
puv = 2n whenever i and j differ by exactly one bit. A plausible choice for
canonical paths is to let xy use bit-fixing routing, where we change bits
in x to the corresponding bits in y from left to right. To compute congestion,
pick some edge uv, and let k be the bit position in which u and v differ.
A path xy will cross uv if uk . . . un = xk . . . xn (because when we are at u
we havent fixed those bits yet) and v1 . . . vk = y1 . . . yk (because at v we
have fixed all of those bits). There are 2k1 choices of x1 . . . xk1 consistent
with the first condition and 2nk choices of yk+1 . . . yn consistent with the
second, giving exactly 2n1 total paths xy crossing uv. Since each path
occurs with weight x y = 22n , and the flow across uv is u puv = 2n 2n 1
,
we can calculate the congestion
1 X
() = max x y
uvE u puv xy 3uv
1
= 2n1 22n
2n /2n
= n.
This gives a relaxation time 2 82 = 8n2 . In this case the bound is

substantially worse than what we previously proved using coupling.
The fact that the number of canonical paths that cross a particular edge
is exactly one half the number of nodes in the hypercube is not an accident:
if we look at what information we need to fill in to compute x and y from u
and v, we need (a) the part of x weve already gotten rid of, plus (b) the
part of y we havent filled in yet. If we stitch these two pieces together, we
get all but one of the n bits we need to specify a node in the hypercube, the
missing bit being the bit we flip to get from u to v. This sort of thing shows
up a lot in conductance arguments where we build our canonical paths by
fixing a structure one piece at a time.
9.4.8.3 Matchings in a graph

A matching in a graph G = (V, E) is a subset of the edges with no two
elements adjacent to each other; equivalently, its a subgraph that includes
all the vertices in which each vertex has degree at most 1. We can use a
random walk to sample matchings from an arbitrary graph uniformly.
Here is the random walk. Let St be the matching at time t. At each
step, we choose an edge e E uniformly at random, and flip a coin to decide
whether to include it or not. If the coin comes up tails, set St+1 = St \{e} (this
may leave St+1 = St if St already omitted e); otherwise, set St+1 = St {e}
unless one of the endpoints of e is already incident to an edge in St , in which
case set St+1 = St .
Because this chain contains many self-loops, its aperiodic. Its also
straightforward to show that any transition between two adjacent matchings
1
occurs with probability exactly 2m , where m = |E|, and thus that the chain
is reversible with a uniform stationary distribution. Wed like to bound the
congestion of the chain to show that it converges in a reasonable amount of
time.
Let N be the number of matchings in G, and let n = |V | and m = |E| as
usual. Then S = 1/N for all S and S pST = 2N1m . Our congestion for any
transition ST will then be 2N mN 2 = 2m/N times the number of paths
that cross ST ; ideally this number of paths will be at most N times some
small polynomial in n and/or m.
Suppose we are trying to get from some matching X to another matching
Y . The graph X Y has maximum degree 2, so each of its connected
components is either a path or a cycle; in addition, we know that the edges
in each of these paths or cycles alternate whether they come from X or Y ,
which among other things implies that the cycles are all even cycles.
We can transform X to Y by processing these components in a methodical
way: first order the components (say, by the increasing identity of the smallest
vertex in each), then for each component replace the X edges with Y edges.
If we do this cleverly enough, we can guarantee that for any transition ST ,
the set of edges (X Y ) \ (S T ) always consists of a matching plus at most
two extra edges, and that S, T , (X Y ) \ (S T ) are enough to reconstruct
X and Y . Since there are at most N m2 choices of (X Y ) \ (S T ), this
will give at most N m2 canonical paths across each transition.12
Here is how we do the replacement within a component. If C is a
cycle with k vertices, order the vertices v0 , . . . , vk1 such that v0 has the
smallest index in C and v0 v1 X. Then X C consists of the even-
numbered edges e2i = v2i v2+1 and Y C consists of the odd-numbered edges
e2i+1 = v2i+1 v2i+2 , where we take vk = v0 . We now delete e0 , then alternate
between deleting e2i and adding e2i1 for i = 1 . . . k/2 1. Finally we add
ek1 .
The number of edges in C at each step in this process is always one of
k/2, k/2 1, or k/2 2, and since we always add or delete an edge, the
number of edges in C (S T ) for any transition will be at least k/2 1,
which means that the total degree of the vertices in C (S T ) is at least
12 m

We could improve the constant a bit by using N 2
, but we wont.
k 2. We also know that SU T is a matching, since its either equal to S or

equal to T . So there are at least k 2 vertices in C (S T ) with degree
1, and none with degree 2, leaving at most 2 vertices with degree 0. In the
complement C \ (S T ), this becomes at most two vertices with degree 2,
with the rest having degree 1. If these vertices are adjacent, removing the
edge between them leaves a matching; if not, removing one edge adjacent to
each does so. So two extra edges are enough.
A similar argument works for paths, which we will leave as an exercise
for the reader.
Now suppose we know S, T , and (XU Y ) \ (S T ). We can compute
X Y = ((X Y ) \ (S T )) (S T ); this means we can reconstruct the
graph X Y , identify its components, and so on. We know which component
C we are working on because we know which edge changes between S and T .
For any earlier component C 0 , we have C 0 S = C 0 Y (since we finished
already), and similarly C 0 \ S = C 0 X. For components C 0 we havent
reached yet, the reverse holds, C 0 S = C X and C 0 \ S = C Y . This
determines the edges in both X and Y for all edges not in C.
For C, we know from C (X Y ) which vertex is v0 . In a cycle, whenever
we remove an edge, its lower-numbered endpoint is always an even distance
from v0 , and similarly when we add an edge, its lower-numbered endpoint
is always an odd distance from v0 . So we can orient the cycle and tell
which edges in S T have already been added (and are thus part of Y ) and
which have been removed (and are thus part of X). Since we know that
C Y = C \ X, this is enough to reconstruct both C X and C Y . (The
same general idea works for paths, although for paths we do not need to be as
careful to figure out orientation since we can only leave v0 in one direction.)
The result is that given S and T , there are at most N m2 choices of X and
2 2
Y such that the canonical path XY crosses ST . This gives NNm1N1 =
2m
2m3 , which gives 2 8m6 . This is not great but it is at least polynomial in
the size of the graph.13
13
However, it might not give rapid mixing if N is small enough, which can occur in
dense graphs. For example, on a clique we have N = m + 1, since no matching can contain
more than one edge, making the bound 8m6 = (N 6 ) which is polynomial in N but not
polynomial in log N . Polynomial in log N is what we want. The actual mixing time in this
case is (m) (for the upper bound, wait to delete the only edge, and do so on both sides
of the coupling; for the lower bound, observe that until we delete the only edge we are still
in our initial state). This is much better than the upper bound from canonical paths, but
it still doesnt give rapid maxing.
9.4.8.4 Perfect matchings in dense bipartite graphs

(Basically doing [MR95, 11.3].)
A similar chain can be used to sample perfect matchings in dense bipartite
graphs, as shown by Jerrum and Sinclair [JS89] based on an algorithm by
Broder [Bro86] that turned out to have a bug in the analysis [Bro88].
A perfect matching of a graph G is a subgraph that includes all the
vertices and gives each of them exactly one incident edge. Well be looking
for perfect matchings in a bipartite graph consisting of n left vertices
u1 , . . . , un and n right vertices v1 , . . . , vn , where every edge goes between a
left vertex and a right vertex. Well also assume that the graph is dense,
which in this case means that every vertex has at least n/2 neighbors.
This density assumption was used in the original Broder and Jerrum-Sinclair
papers but removed in a later paper by Jerrum, Sinclair, and Vigoda [JSV04].
The random walk is similar to the random walk from the previous section
restricted to the set of matchings with either n 1 or n edges. At each
step, we first flip a coin (the usual lazy walk trick); if it comes up heads,
we choose an edge uv uniformly at random and apply one of the following
transformations depending on how uv fits in the current matching mt :
1. If uv mt , and |mt | = n, set mt+1 = mt \ {uv}.
2. If u and v are both unmatched in mt , and |mt | = n 1, set mt+1 =

mt {uv}.
3. If exactly one of u and v is matched to some other node w, and

|mt | = n 1, perform a rotation that deletes the w edge and adds uv.
4. If none of these conditions hold, set mt+1 = mt .
The walk can be started from any perfect matching, which can be found
in O(n5/2 ) time using a classic algorithm of Hopcroft and Karp [HK73] that
repeatedly searches for an augmenting path, which is a path in G between
two unmatched vertices that alternates between edges not in the matching
and edges in the matching. (Well see augmenting paths again below when we
show that any near-perfect matching can be turned into a perfect matching
using at most two transitions.)
We can show that this walk converges in polynomial time using a canonical
path argument. This is done in two stages: first, we define canonical paths
between all perfect matchings. Next, we define a short path from any
matching of size n 1 to some nearby perfect matching, and build paths
between arbitrary matchings by pasting one of these short paths on one or
both ends of a long path between perfect matchings. This gives a collection
of canonical paths that we can show to have low congestion.
To go between perfect matchings X and Y , consider X Y as a collection
of paths and even cycles as in 9.4.8.3. In some standard order, fix each path
cycle by first deleting one edge to make room, then using rotate operations
to move the rest of the edges, then putting the last edge back in. For any
transition ST along the way, we can use the same argument that we can
compute X Y from S T by supplying the missing edges, which will consist
of a matching of size n plus at most one extra edge. So if N is the number
of matchings of size n or n 1, the same argument used previously shows
that at most N m of these long paths cross any ST transition.
For the short paths, we must use the density property. The idea is that for
any matching that is not perfect, we can find an augmenting path of length at
most 3 between two unmatched nodes on either side. Pick some unmatched
nodes u and v. Each of these nodes is adjacent to at least n/2 neighbors; if
any of these neighbors are unmatched, we just found an augmenting path
of length 1. Otherwise the n/2 neighbors of u and the n/2 nodes matched
to neighbors of v overlap (because v is unmatched, leaving at most n 1
matched nodes and thus at most n/21 nodes that are matched to something
thats not a neighbor of v). So for each matching of size n 1, in at most
two steps (rotate and then add an edge) we can reach some specific perfect
matching. There are at most m2 ways that we can undo this, so each perfect
matching is associated with at most m2 smaller matchings. This blows up
the number of canonical paths crossing any transition by roughly m4 ; by
counting carefully we can thus show congestion that is O(m6 ) (O(m4 ) from
the blow-up, m from the m in N m, and m from 1/pST ).
It follows that for this process, 2 = O(m12 ). (I think a better analysis is
possible.)
As noted earlier, this is an example of a process for which causal coupling
doesnt work in less than exponential time [KR99], a common problem with
Markov chains that dont have much symmetry. So its not surprising that
stronger techniques were developed specifically to attack this problem.
Chapter 10
Approximate counting
(See [MR95, Chapter 11] for details.)

Basic idea: we have some class of objects, we want to know how many of
them there are. Ideally we can build an algorithm that just prints out the
exact number, but for many problems this is hard.
A fully polynomial-time randomized approximation scheme or
FPRAS for a numerical problem outputs a number that is between (1 )
and (1 + ) times the correct answer, with probability at least 3/4 (or some
constant bounded away from 1/2we can amplify to improve it), in time
polynomial in the input size n and (1/). In this chapter, well be hunting
for FPRASs. But first we will discuss briefly why we cant just count directly
in many cases.
10.1 Exact counting

A typical application is to a problem in the complexity class #P, problems
that involve counting the number of accepting computation paths in a
nondeterministic polynomial-time Turing machine. Equivalently, these are
problems that involve counting for some input x the number of values r such
that M (x, r) = 1, where M is some machine in P. An example would be the
problem #SAT of counting the number of satisfying assignments of some
CNF formula.
The class #P (which is usually pronounced sharp P or number P)
was defined by Leslie Valiant in a classic paper [Val79]. The central result
in this paper is Valiants theorem. This shows that any problem in #P
can be reduced (by Cook reductions, meaning that we are allowed to
use the target problem as a subroutine instead of just calling it once) to the
185
CHAPTER 10. APPROXIMATE COUNTING 186
problem of computing the permanent of a square 01 matrix A, where the

P Q
permanent is given by the formula i Ai,(i) , where the sum ranges over
all n! permutations of the indices of the matrix. An equivalent problem
is counting the number of perfect matchings (subgraphs including all
vertices in which every vertex has degree exactly 1) of a bipartite graph.
Other examples of #P-complete problems are #SAT (defined above) and
#DNF (like #SAT, but the input is in DNF form; #SAT reduces to #DNF
by negating the formula and then subtracting the result from 2n ).
Exact counting of #P-hard problems is likely to be very difficult: Todas
theorem [Tod91] says that being able to make even a single query to a #P-
oracle is enough to solve any problem in the polynomial-time hierarchy,
which contains most of the complexity classes you have probably heard
of. Nonetheless, it is often possible to obtain good approximations to such
problems.
10.2 Counting by sampling

If many of the things we are looking for are in the target set, we can count
by sampling; this is what poll-takers do for a living. Let U be the universe
we are sampling from and G be the good set of points we want to count.
Let = |G|/|U |. If we can sample uniformly from U , then we can estimate
by taking N independent samples and dividing the number of samples in G
by N . This will give us an answer whose expectation is , but the accuracy
may be off. Since the variance of each sample is (1 ) (when is

small), we get Var [ Xi ] m, giving a standard deviation of m. For
P

this to be less than our allowable error m, we need 1 m or N 21 .
This gets bad if is exponentially small. So if we are to use sampling, we
need to make sure that we only use it when is large.
On the positive side, if is large enough we can easily compute how
many samples we need using Chernoff bounds. The following lemma gives
a convenient estimate; it is based on [[MR95, Theorem 11.1]] with a slight
improvement on the constant:
Lemma 10.2.1. Sampling N times gives relative error with probability at

least 1 provided 1.81 and
3 2
N ln . (10.2.1)
2
Proof. Suppose we take N samples, and let X be the total count for these
samples. Then E [X] = N , and (5.2.7) gives (for 1.81):

2 /3
Pr [|X N | N ] 2eN .
2 /3
Now set 2eN and solve for N to get (10.2.1).
10.2.1 Generating samples

An issue that sometimes comes up in this process is that it is not always
obvious how to generate a sample from a given universe U . What we want
is a function that takes random bits as input and produces an element of U
as output.
In the simplest case, we know |U | and have a function f from {0, . . . , |U | 1}
to U (the application of such a function is called unranking). Here we can
do rejection sampling: we choose a bit-vector of length dlg|U |e, interpret
it as an integer x expressed in binary, and try again until x < |U |. This
requires less than 2dlg|U |e bits on average in the worst case, since we get a
good value at least half the time.
Unranking (and the converse operation of ranking) can be a non-trivial
problem in combinatorics, and there are entire textbooks on the subject [SW13].
But some general principles apply. If U is a union of disjoint sets U1 U2
. . . U Un , we can compute the sizes of each |Ui |, and we can unrank each
Ui , then we can map some x {0 . . . |U | 1} to a specific element of U
by choosing the maximum i such that i1 j=1 |Uj | x and choosing the k-th
P
Pi1
element of Ui , where k = x j=1 |Uj |. This is essentially what happens if
we want to compute the x-th day of the year. The first step of this process
requires computing the sizes of O(n) sets if we generate the sum one element
at a time, although if we have a fast way of computing ji=1 |Ui | for each
P
i, we can cut this time to O(log n) such computations down using binary
search. The second step depends on whatever mechanism we use to unrank
within Ui .
An example would be generating one of the nk subsets of a set of size

k uniformly at random. If we let S be the set of all such k-subsets, we can

express S as i = 1n Si where i is the set of all k-subsets that have the i-th
S
element as their smallest element in some fixed ordering. But then we can
ni
easily compute |Si | = k=1 for each i, and apply the above technique to pick
a particular Si to start with and then recurse within Si to get the rest of the
elements.
A more general case is when we cant easily sample from U but we can
sample from some T U . Here rejection sampling comes to our rescue
as long as |U |/|T | is large enough. For example, if we want to generate a

k-subset of n elements when k n, we can choose a list of k elements
with replacement and discard any lists that contain duplicates. But this
doesnt work so well for larger k. In this particular case, there is an easy
out: we can sample k elements without replacement and forget their order.
This corresponds to sampling from a universe T of ordered k-subsets, then
mapping down to U . As long as the inverse image of each element of U has
the same size in T , this will give us a uniform sample from U .
When rejection sampling fails, we may need to come up with a more clever
approach to concentrate on the particular values we want. One approach that
can work well if we can express U as a union of sets that are not necessarily
disjoint is the Karp-Luby technique [KL85], discussed in 10.3 below.
All of this assumes that we are interested in getting uniform samples.
For non-uniform samples, we replace the assumption that we are trying to
generate a uniform X {0 . . . m 1} with some X {0 . . . m 1} for which
we can efficiently calculate Pr [X i] for each i. In this case, we can (in
principle) generate a continuous random variable Y [0, 1] and choose the
maximum i such that Pr [X i] Y , which we can find using binary search.
The complication is that we cant actually generate Y in finite time.
Instead, we generate Y one bit at a time; this gives a sequence of rational
values Yt where each Yt is of the form k/2t . We can stop when every
value in interval [Yt , Yt + 2t ) lies with the range corresponding to some
particular value of X. This technique is closely related to arithmetic
coding [WNC87], and requires generating H(X) + O(1) bits on average,
where H(X) = m1
P
i=0 Pr [X = i] lg Pr [X = i] is the (base 2) entropy of
X.
10.3 Approximating #DNF

A classical algorithm of Karp and Luby [KL85] gives a FPRAS for #DNF.
Well describe how this works, mostly following the presentation in [MR95,
11.2]. The key idea of the Karp-Luby technique is to express the set S
whose size we want to know, as a union of a polynomial number of simpler
sets S1 , . . . , Sk . If for each i we can sample uniformly from Si , compute
|Si |, and determine membership in Si , then some clever sampling will let us
approximate |S| by sampling from a disjoint union of the Si to determine
how much |Si | overestimates |S|.
P
To make this concrete, lets look at the specific problem studied by Karp
and Luby of approximating the number of satisfying assignment to a DNF
formula. A DNF formula is a formula that is in disjunctive normal form:

it is an OR of zero or more clauses, each of which is an AND of variables
or their negations. An example would be (x1 x2 x3 ) (x1 x4 ) x2 ).
The #DNF problem is to count the number of satisfying assignments of a
formula presented in disjunctive normal form.
Solving #DNF exactly is #P-complete, so we dont expect to be able to
do it. Instead, well get a FPRAS by cleverly sampling solutions. The need
for cleverness arises because just sampling solutions directly by generating
one of the 2n possible assignments to the n variables may find no satisfying
assignments at all, since the size of any individual clause might be big enough
that getting a satisfying assignment for that clause at random is exponentially
unlikely.
So instead we will sample pairs (x, i), where x is an assignment that
satisfies clause Ci ; these are easier to find, because if we know which clause
Ci we are trying to satisfy, we can read off the satisfying assignment from
its variables. Let S be the set of such pairs. For each pair (x, i), define
P
f (x, i) = 1 if and only if Cj (x) = 0 for all j < i. Then (x,i)S f (x, i) counts
every satisfying assignment x, because (a) there exists some i such that
x satisfies Ci , and (b) only the smallest such i will have f (x, i) = 1. In
effect, f is picking out a single canonical satisfied clause from each satisfying
assignment. Note that we can compute f efficiently by testing x against all
clauses Cj with j < i.
Our goal is to estimate the proportion of good pairs with f (x, i) = 1
P
out of all pairs in S, and then use this to estimate (x,i)S f (x, i) = |S|.
If we can sample from S uniformly, the proportion of good pairs with
f (x, i) = 1 is at least 1/m, because every satisfying assignment x contributes
at most m pairs total, and one of them is good.
The only tricky part is figuring out how to sample pairs (x, i) with
Ci (x) = 1 so that all pairs occur with the same probability. Let Ui =
{(x, i) | Ci (x) = 1}. Then we can compute |Ui | = 2nki where ki is the
number of literals in Ci . Using this information, we can sample i first with
probability |Ui |/ j |Uj |, then sample x from Ui just by picking independent
P
uniform random values for the n k variables not fixed by Ci .

4
With N 2 (1/m) ln 2 = 4m
2
ln 2 , we obtain an estimate for the propor-
tion of pairs (x, i) with f (x, i) = 1 that is within relative error of with
probability at least 1 . Multiplying this by |Ui | then gives the desired
P
count of satisfying assignments.

Its worth noting that theres nothing special about DNF formulas in
this method. Essentially the same trick will work for estimating the size of
the union of any collection of sets Ui where we can (a) compute the size of
each Ui ; (b) sample from each Ui individually; and (c) test membership of
our sample x in Uj for j < i.
10.4 Approximating #KNAPSACK

Here is an algorithm for approximating the number of solutions to a KNAP-
SACK problem, due to Dyer [Dye03]. Well concentrate on the simplest
version, 01 KNAPSACK, following the analysis in Section 2.1 of [Dye03].
For the 01 KNAPSACK problem, we are given a set of n objects of
weight 0 a1 a2 . . . an b, and we want to find a 01 assignment
x1 , x2 , . . . , xn such that ni=1 ai xi b (usually while optimizing some prop-
P
erty that prevents us from setting all the xi to zero). Well assume that the
ai and b are all integers.
For #KNAPSACK, we want to compute |S|, where S is the set of all
assignments to the xi that make ni=1 ai xi b.
P
There is a well-known fully polynomial-time approximation scheme

for optimizing KNAPSACK, based on dynamic programming. The idea is
that a maximum-weight solution can be found exactly in time polynomial in
b, and if b is too large, we can reduce it by rescaling all the ai and b at the
cost of a small amount of error. A similar idea is used in Dyers algorithm:
the KNAPSACK problem is rescaled so that size of the solution set S 0 of the
rescaled version can be computed in polynomial time. Sampling is then used
to determine what proportion of the solutions in S 0 correspond to solutions
of the original problem.
Scaling step: Let a0i = bn2 ai /bc. Then 0 a0i n2 for all i. Taking the
floor creates some error: if we try to reconstruct ai from a0i , the best we can do
is argue that a0i n2 ai /b < a0i + 1 implies (b/n2 )a0i ai < (b/n2 )a0i + (b/n2 ).
The reason for using n2 as our rescaled bound is that the total error in the
upper bound on ai , summed over all i, is bounded by n(b/n2 ) = b/n, a fact
that will become important soon.
Let S 0 = ~x ni=1 a0i xi n2 be the set of solutions to the rescaled
P
knapsack problem, where we substitute a0i for ai and n2 = (n2 /b)b for b.
Pn
Claim: S S 0 . Proof: ~x S if and only if i=1 ai xi b. But then
n n
a0i xi =
X X
bn2 ai /bcxi
i=1 i=1
n
X
(n2 /b)ai xi
i=1
n
X
= (n2 /b) ai xi
i=1
n2 ,
which shows ~x S 0 .
The converse does not hold. However, we can argue that any ~x S 0 can
be shoehorned into S by setting at most one of the xi to 0. Consider the
set of all positions i such that xi = 1 and ai > b/n. If this set is empty,
then ni=1 ai xi ni=1 b/n = b, and ~x is already in S. Otherwise, pick any
P P
position i with xi = 1 and ai > b/n, and let yj = 0 when j = i and yj = xj

otherwise. Then
n
X n
X
a j yj = aj xj ai
j=1 j=1
n
((b/n2 )a0j + b/n2 )xj b/n
X
<
j=1
n
a0j xj + b/n b/n
X
(b/n2 )
j=1
2 2
(b/n )n
= b.
Applying this mapping to all elements ~x of S maps at most n + 1 of them to

each ~y in S 0 ; it follows that |S 0 | (n + 1)|S|, which means that if we can
sample elements of S 0 uniformly, each sample will hit S with probability at
least 1/(n + 1). n P o
To compute |S 0 |, let C(k, m) = ~x ki=1 a0i xi m be the number of

subsets of {a01 , . . . , a0k } that sum to m or less. Then C(k, m) satisfies the
recurrence
C(k, m) = C(k 1, m a0k ) + C(k 1, m)

C(0, m) = 1
where k ranges from 0 to n and m ranges from 0 to n2 , and we treat

C(k 1, m a0k ) = 0 if m a0k < 0. The idea is that C(k 1, m a0k ) counts
all the ways to make m if we include a0k , and C(k 1, m) counts all the ways
to make m if we exclude it. The base case corresponds to the empty set
(which sums to m no matter what m is).
We can compute a table of all values of C(k, m) by iterating through m
in increasing order; this takes O(n3 ) time. At the end of this process, we can
read off |S 0 | = C(n, n2 ).
But we can do more than this: we can also use the table of counts to
sample uniformly from S 0 . The probability that x0n = 1 for a uniform random
element of S 0 is exactly C(n1, n2 a0n )/C(n, n2 ); having chosen x0n = 1 (say),
the probability that x0n1 = 1 is then C(n2, n2 a0n a0n1 )/C(n2, n2 a0n ),
and so on. So after making O(n) random choices (with O(1) arithmetic
operations for each choice to compute the probabilities) we get a uniform
element of S 0 , which we can test for membership in S in an additional O(n)
operations.
Weve already established that |S|/|S 0 | 1/(n + 1), so we can apply
Lemma 10.2.1 to get relative error with probability at least 1 using
3(n+1)
2
ln 2 samples. This gives a cost of O(n2 log(1/)/2 ) for the sampling
step, or a total cost of O(n3 + n2 log(1/)/2 ) after including the cost of
building the table (in practice, the second term will dominate unless we are

willing to accept = (1/ n)).
It is possible to improve this bound. Dyer [Dye03] shows that using
randomized rounding on p the a0i instead of just truncating them gives a
FPRAS that runs in O(n 5/2 log(1/) + n2 /2 ) time.
10.5 Approximating exponentially improbable events

For #DNF and #KNAPSACK, we saw how restricting our sampling to
a cleverly chosen sample space could boost the hit ratio to something
that gave a reasonable number of samples using Lemma 10.2.1. For other
problems, it is often not clear how to do this directly, and the best sample
spaces we can come up with make our target points an exponentially small
fraction of the whole.
In these cases, it is sometimes possible to approximate this exponentially
small fraction as a product of many more reasonable ratios. The idea is to
express our target set as the last of a sequence of sets S0 , S1 , . . . , Sk , where
we can compute the size of S0 and can estimate |Si+1 |/|Si | accurately for
each i. This gives |Sk | = |S0 | ki=1 |S|Si+1 |
Q
i|
, with a relative error that grows
roughly linearly with k. Specifically, if we can approximate each |Si+1 |/|Si |

ratio to between 1 and 1 + of the correct value, then the product of
these ratios will be between (1 )k and (1 + )k of the correct value; these
bounds approach 1 k and 1 + k in the limit as k goes to zero, using
the binomial theorem, although to get a real bound we will need to do more
careful error analysis.
10.5.1 Matchings
We saw in 9.4.8.3 that a random walk on matchings on a graph with m
edges has mixing time 2 8m6 , where the walk is defined by selecting an
edge uniformly at random and flipping whether it is in the matching or not,
while rejecting any steps that produce a non-matching. This allows us to
sample matchingsof a graph with total variation distance from the uniform
distribution in O m6 log m log 1 time.
Suppose now that we want to count matchings instead of sampling them.
Its easy to show that for any particular edge uv G, at least half of all
matchings in G dont include uv: the reason is that if M is a matching in G,
then M 0 = M \ {uv} is also a matching, and at most two matchings M 0 and
M 0 {uv} are mapped to any one M 0 by this mapping.
Order the edges of G arbitrarily as e1 , e2 , . . . , em . Let Si be the set of
matchings in G \ {e1 . . . ei }. Then S0 is the set of all matchings, and weve
just argued that i+1 = |Si+1 |/|Si | 1/2. We also know that |Sm | counts
the number of matchings in a graph with no edges, so its exactly one. So
|Si |
we can use the product-of-ratios trick to compute S0 = m
Q
i=1 |Si+1 | .

A random walk of length O 8m6 log m log 1 can sample matchings from
Si with a probability 0 of getting a matching in Si+1 that is between (1
)i+1 and (1+)i+1 . From Lemma 10.2.1, we canestimate 0 within relative

error with probability at least 1 using O 2 0 log = O 1 log 1
1 1
samples. Combined with the error on 0 ,this gives relative error at most
+ + in O m log m log 1 log 1 log 1 operations.1 If we then multiply
6
out all the estimates for |Si |/|Si+1 |, we get an estimate of S0 that is at most
(1+++)m times the correct value with probability at least 1m (with a
similar bound on the other side), in total time O m7 log m log 1 log 1 log 1 .
To turn this into a fully polynomial-time approximation scheme, given
1
This is the point where sensible people start hauling out the O notation, where a
(n)) if it O(f (n)g) where g is polylogarithmic in n and any other parameters
function is O(f
that may be running around ( 1 , 1 , etc.).
, , and m, we need to select , , and to get relative error with

probability at least 1 . Letting = /m gets the part. For , we need
(1 + + + )m 1 + . Suppose that < 1 and let = = /6m. Then
m

(1 + + + )m 1 +
2m
e/2
1 + .

Plugging these values into our cost formula gives O m5 log m log2 6m
log m
=

O m7 log m log m + log 1 + log 1 = O(m 7 ) time.

10.5.2 Other problems

Similar methods work for other problems that self-reduce by restricting
particular features of the solution. Examples include colorings (fix the color
of some vertex), independent sets (remove a vertex), and approximating the
volume of a convex body (take the intersection with a sphere of appropriate
radius; see [MR95, 11.4] for a slightly less sketchy description). We will not
go into details on any of these applications here.
Chapter 11
The probabilistic method
The probabilistic method is a tool for proving the existence of objects with
particular combinatorial properties, by showing that some process generates
these objects with nonzero probability.
The relevance of this to randomized algorithms is that in some cases
we can make the probability large enough that we can actual produce such
objects.
Well mostly be following Chapter 5 of [MR95] with some updates for
more recent results. If youd like to read more about these technique, a
classic reference on the probabilistic method in combinatorics is the text of
Alon and Spencer [AS92].
11.1 Randomized constructions and existence proofs

Suppose we want to show that some object exists, but we dont know how to
construct it explicitly. One way to do this is to devise some random process for
generating objects, and show that the probability that it generates the object
we want is greater than zero. This implies that something we want exists,
because otherwise it would be impossible to generate; and it works even if the
nonzero probability is very, very small. The systematic development of the
method is generally attributed to the notoriously productive mathematician
Paul Erds and his frequent collaborator Alfrd Rnyi.
From an algorithmic perspective, the probabilistic method is useful
mainly when we can make the nonzero probability substantially larger than
zeroand especially if we can recognize when weve won. But sometimes
just demonstrating the existence of an object is a start.
We give a couple of example of the probabilistic method in action below.
195
CHAPTER 11. THE PROBABILISTIC METHOD 196
In each case, the probability that we get a good outcome is actually pretty
high, so we could in principle generate a good outcome by retrying our
random process until it works. There are some more complicated examples
of the method for which this doesnt work, either because the probability of
success is vanishingly small, or because we cant efficiently test whether what
we did succeeded (the last example below may fall into this category). This
means that we often end up with objects whose existence we can demonstrate
even though we cant actually point to any examples of them. For example,
it is known that there exist sorting networks (a special class of circuits
for sorting numbers in parallel) that sort in time O(log n), where n is the
number of values being sorted [AKS83]; these can be generated randomly
with nonzero probability. But the best explicit constructions of such networks
take time (log2 n), and the question of how to find an explicit network that
achieves O(log n) time has been open for over 25 years now despite many
efforts to solve it.
11.1.1 Unique hats

A collection of n robots each wishes to own a unique hat. Unfortunately, the
State Ministry for Hat Production only supplies one kind of hat. A robot
will only be happy if (a) it has a hat, and (b) no robot it sees has a hat.
Fortunately each robot can only see k other robots. How many robots can
be happy?
We could try to be clever about answering this question, or we could apply
the probabilistic method. Suppose we give each robot a hat with independent
probability p. Then the probability that any particular robot r is happy is
pq k , where q = 1 p is the probability that a robot doesnt have a hat and q k
gives the probability that none of the robots that r sees has a hat. If we let
Xr be the indicator for the event that r is happy, we have E [Xr ] = pq k and
E [Xr ] = npq k . Since
P P
the expected number of happy robots is E [ Xr ] =
we can achieve this value on average, there must be some specific assignment
that achieves it as well.
To choose a good p, we apply the usual calculus trick of looking for a
maximum by looking for the place where the derivative is zero. We have
d k k k1 (since dq = 1), so we get nq k nkpq k1 = 0.
dp npq = nq nkpq dp
Factoring out n and q k1 gives q pk = 0 or 1 p pk = 0 or p =
1/(k + 1). For this value of p, the expected number of happy robots is exactly
k
1 1
n k+1 1 k+1 . For large k the last factor approaches 1/e, giving us
n
approximately (1/e) k+1 happy robots.
Up to constant factors this is about as good an outcome as we can hope

for: we can set things up so that our robots are arranged in groups of k + 1,
where each robot sees all the other robots in its group, and here we can
clearly only have 1 happy robot per group, for a maximum of n/(k + 1)
happy robots.
Note that we can improve the constant 1/e slightly by being smarter
than just handing out hats at random. Starting with our original n robots,
look for a robot that is observed by the smallest number of other robots.
Since there are only nk pairs of robots (r1 , r2 ) where r1 sees r2 , one of the n
robots is only seen by at most k other robots. Now give this robot a hat, and
give no hat to any robot that sees it or that it sees. This produces 1 happy
robot while knocking out at most 2k + 1 robots total. Now remove these
robots from consideration and repeat the analysis on the remaining robots,
to get another happy robot while again knocking out at most 2k + 1 robots.
Iterating this procedure until no robots are left gives at least n/(2k + 1)
n
happy robots; this is close to (1/2) k+1 for large k, which is a bit better
n
than (1/e) k+1 . (It may be possible to improve the constant further with
more analysis.) This shows that the probabilistic method doesnt necessarily
produce better results than we can get by being smart. But the hope is
that with the probabilistic method we dont have to be smart, and for some
problems it seems that solving them without using the probabilistic method
requires being exceptionally smart.
11.1.2 Ramsey numbers

Consider a collection of n schoolchildren, and imagine that each pair of
schoolchildren either like each other or hate each other. We assume that
these preferences are symmetric: if x likes y, then y likes x, and similarly if x
hates y, y hates x. Let R(k, h) be the smallest value for n that ensures that
among any group of n schoolchildren, there is either a subset of k children
that all like each other or a subset of h children that all hate each other.1
It is not hard to show that R(k, h) is finite for all k and h.2 The exact
1
In terms of graphs, any graph G with at least R(k, h) nodes contains either a clique
of size k or an independent set of size h.
2
A simple proof, due to Erds and Szekeres [ES35], proceeds by showing that R(k, h)
R(k 1, h) + R(k, h 1). Given a graph G of size at least R(k 1, h) + R(k, h 1), choose
a vertex v and partition the graph into two induced subgraphs G1 and G2 , where G1
contains all the vertices adjacent to v and G2 contains all the vertices not adjacent to v.
Either |G1 | R(k 1, h) or |G2 | R(k, h 1). If |G1 | R(k 1, h), then G1 contains
either a clique of size k 1 (which makes a clique of size k in G when we add v to it)
or an independent set of size h (which is also in G). Alternatively, if |G2 | R(k, h 1),
value of R(k, h) is known only for small values of k and h.3 But we can use
the probabilistic method to show that for k = h, it is reasonably large. The
following theorem is due to Erds, and was the first known lower bound on
R(k, k).
Theorem 11.1.1 ([Erd47]). If k 3, then R(k, k) > 2k/2 .
Proof. Suppose each pair of schoolchildren flip a fair coin to decide whether
they like each other or not. Then the probability that any particular set of
k
k schoolchildren all like each other is 2(2) and the probability that they all
hate each other is the same. Summing over both possibilities and all subsets
k
gives a bound of nk 21(2) on the probability that there is at least one subset

in which everybody likes or everybody hates everybody. For n = 2k/2 , we

have
!
n 1(k) nk 1(k)
2 2 2 2
k k!
2 /2+1k(k1)/2
2k
=
k!
2 2
2k /2+1k /2+k/2
=
k!
21+k/2
=
k!
< 1.
Because the probability that there is an all-liking or all-hating subset

is less than 1, there must be some chance that we get a collection that
doesnt have one. So such a collection exists. It follows that R(k, k) > 2k/2 ,
because we have shown that not all collections at n = 2k/2 have the Ramsey
property.
then G2 either gives us a clique of size k by itself or an independent set of size h after
adding v. Together with the fact that R(1, h) = R(k, 1) = 1, this recurrence gives
R(k, h) (k1)+(h1)
k1
.
3
There is a fairly current table at http://en.wikipedia.org/wiki/Ramseys_Theorem.
Some noteworthy values are R(3, 3) = 6, R(4, 4) = 18, and 43 R(5, 5) 49. One problem
with computing exact values is that as R(k, h) grows, the number of graphs one needs
n
to consider gets very big. There are 2( 2 ) graphs with n vertices, and even detecting the
presence or absence of a moderately-large clique or independent set in such a graph can be
expensive. This pretty much rules out any sort of brute-force approach based on simply
enumerating candidate graphs.
The last step in the proof uses the fact that 21+k/2 < k! for k 3, which
can be tested explicitly for k = 3 and proved by induction for larger k. The
resulting bound is a little bit weaker than just saying that n must be large
k
enough that nk 21(2) 1, but its easier to use.

The proof can be generalized to the case where k 6= h by tweaking the

bounds and probabilities appropriately. Note that even though this process
generates a graph with no large cliques or independent sets with reasonably
high probability, we dont have any good way of testing the result, since
testing for the existence of a clique is NP-hard.
11.1.3 Directed cycles in tournaments

In the previous example we used the probabilistic method to show that there
existed a structure that didnt contain some particular substructure. For
this example we will do the reverse: show that there exists a structure that
contains many instances of a particular substructure.
Imagine that n wrestlers wrestle in a round-robin tournament, so that
ever wrestler wrestles every other wrestler exactly once, and so that for each
pair of wrestlers one wrestler beats the other. Consider a cycle of wrestlers
x1 , x2 , . . . , xn such that each wrestler xi beats xi+1 and xn beats x1 . (Note
that we consider two cycles equivalent if one is a cyclic shift of the other.)
Theorem 11.1.2. For n 3, there exists a tournament with at least 2n (n

1)! directed cycles.
Proof. Technical detail: We need n 3 because the edges in a 2-cycle always

go in opposite directions.
Flip a coin to decide the outcome of each pairing. For any particular
sequence x1 , . . . , xn , the probability that it is a directed cycle is then 2n
(since there are n edges in the sequence and each has independent probability
21 of going the right way). Now let range over all (n 1)! cycles and
define X as the indicator variable for the event that each edge is directed
the right way. Weve just shown that E [X ] = 2n ; so the expected total
number of cycles is E [ X ] = E [X ] = 2n (n 1)!. But if the average
P P
value is at least 2n (n 1)!, there must be some specific outcome that gives
a value at least this high.
As in the Ramsey graph example, Theorem 11.1.2 says that generating a

random tournament is likely to give us many directed cycles, but whether a
particular tournament has this property may be hard to test. This is not
unusual for probabilistic-method arguments.
11.2 Approximation algorithms

One use of a randomized construction is to approximate the solution to
an otherwise difficult problem. In this section, we start with a trivial
approximation algorithm for the largest cut in a graph, and then show a
more powerful randomized approximation algorithm, due to Goemans and
Williamson [GW94], that gets a better approximation ratio for a much larger
class of problems.
11.2.1 MAX CUT

Weve previously seen (2.3.1.2) a randomized algorithm for finding small
cuts in a graph. What if we want to find a large cut?
Here is a particularly brainless algorithm that finds a large cut. For each
vertex, flip a coin: if the coin comes up heads, put the vertex in S, otherwise,
put in it T . For each edge, there is a probability of exactly 1/2 that it is
included in the ST cut. It follows that the expected size of the cut is exactly
m/2.
One consequence of this is that every graph has a global cut that includes
at least half the edges. Another is that this algorithm finds such a cut,
1
with probability at least m+1 . To prove this, let X be the random variable
representing the number of edges in the cut, and let p be the probability
that X m/2. Then
m/2 = E [X]
= (1 p) E [X | X < m/2] + p E [X | X m/2]
m1
(1 p) + pm.
2
Solving this for p gives the claimed bound.4
By running this enough times to get a good cut, we get a polynomial-time
randomized algorithm for approximating the maximum cut within a factor
of 2, which is pretty good considering that MAX CUT is NP-hard.
There exist better approximation algorithms. Goemans and Williamson [GW95]
give a 0.87856-approximation algorithm for MAX CUT based on randomized
rounding of a semidefinite program. We wont attempt to present this here,
but we will describe (in 11.2.2) an earlier result, also due to Goemans and
4
This is tight for m = 1, but I suspect its an underestimate for larger m. The
main source of slop in the analysis seems to be the step E [X | X m/2] m; using a
concentration bound, we should be able to show a much stronger inequality here and thus
a much larger lower bound on p.
3
Williamson [GW94], that gives a 4 -approximation to MAX SAT using a
similar technique.
11.2.2 MAX SAT

Like MAX CUT, MAX SAT is an NP-hard optimization problem that sounds
like a very short story about Max. We are given a satisfiability problem
in conjunctive normal form: as a set of m clauses, each of which is the
OR of a bunch of variables or their negations. We want to choose values
for the n variables that satisfy as many of the clauses as possible, where
a clause is satisfied if it contains a true variable or the negation of a false
variable.5
We can instantly satisfy at least m/2 clauses on average by assigning
values to the variables independently and uniformly at random; the analysis
is the same as in 11.2.1 for large cuts, since a random assignment makes
the first variable in a clause true with probability 1/2. Since this approach
doesnt require thinking and doesnt use the fact that many of our clauses
may have more than one variable, we can probably do better. Except we
cant do better in the worst case, because it might be that our clauses consist
entirely of x and x for some variable x; clearly, we can only satisfy half of
these. We could do better if we knew all of our clauses consisted of at least
k distinct literals (satisfied with probability 1 2k ), but we cant count on
this. We also cant count on clauses not being duplicated, so it may turn
out that skipping a few hard-to-satisfy small clauses hurts us if they occur
many times.
Our goal will be to get a good approximation ratio, defined as the
ratio between the number of clauses we manage to satisfy and the actual
maximum that can be satisfied. The tricky part in designing a approxima-
tion algorithms is showing that the denominator here wont be too big.
We can do this using a standard trick, of expressing our original problem as
an integer program and then relaxing it to a linear program6 whose
5
The presentation here follows [MR95, 5.2], which in turn is mostly based on a classic
paper of Goemans and Williamson [GW94].
6
A linear program is an optimization problem where we want to maximize (or
minimize) some linear objective function of the variables subject to linear-inequality
constraints. A simple example would be to maximize x + y subject to 2x + y 1 and
x + 3y 1; here the optimal solution is the assignment x = 25 , y = 15 , which sets the
objective function to its maximum possible value 35 . An integer program is a linear
program where some of the variables are restricted to be integers. Determining if an integer
program even has a solution is NP-complete; in contrast, linear programs can be solved in
polynomial time. We can relax an integer program to a linear program by dropping the
solution doesnt have to consist of integers. We then convert the fractional

solution back to an integer solution by rounding some of the variables to inte-
ger values randomly in a way that preserves their expectations, a technique
known as randomized rounding.7
Here is the integer program (taken from [MR95, 5.2]). We let zj {0, 1}
represent whether clause Cj is satisfied, and let yi {0, 1} be the value of
variable xi . We also let Cj+ and Cj be the set of variables that appear in
Cj with and without negation. The problem is to maximize
m
X
zj
j=1
subject to X X
yi + (1 yi ) zj ,
iCj+ iCj
for all j.
The main trick here is to encode OR in the constraints; there is no
requirement that zj is the OR of the yi and (1 yi ) values, but we maximize
the objective function by setting it that way.
Sadly, solving integer programs like the above is NP-hard (which is not
surprising, since if we could solve this particular one, we could solve SAT).
But if we drop the requirements that yi , zj {0, 1} and replace them with
0 yi 1 and 0 zj 1, we get a linear programsolvable in polynomial
timewith an optimal value at least as good as the value for the integer
program, for the simple reason that any solution to the integer program is
also a solution to the linear program.
requirements that variables be integers; this will let us find a fractional solution that is
at least as good as the best integer solution, but might be undesirable because it tells us
to do something that is ludicrous in the context of our original problem, like only putting
half a passenger on the next plane.
Linear programming has an interesting history. The basic ideas were developed indepen-
dently by Leonid Kantorovich in the Soviet Union and George Dantzig in the United States
around the start of the Second World War. Kantorovichs work had direct relevance to
Soviet planning problems, but wasnt pursued seriously because it threatened the political
status of the planners, required computational resources that werent available at the
time, and looked suspiciously like trying to sneak a capitalist-style price system into the
planning process; for a fictionalized account of this tragedy, see [Spu12]. Dantzigs work,
which included the development of the simplex method for solving linear programs, had
a higher impact, although its publication was delayed until 1947 by wartime secrecy.
7
Randomized rounding was invented by Raghavan and Thompson [RT87]; the particular
application here is due to Goemans and Williamson [GW94].
The problem now is that the solution to the linear program is likely to
be fractional: instead of getting useful 01 values, we might find out we are
supposed to make xi only 2/3 true. So we need one more trick to turn the
fractional values back into integers. This is the randomized rounding step:
given a fractional assignment yi , we set xi to true with probability yi .
So what does randomized rounding do to clauses? In our fractional
solution, a clause might have value zj , obtained by summing up bits and
pieces of partially-true variables. Wed like to argue that the rounded version
gives a similar probability that Cj is satisfied.
Suppose Cj has k variables; to make things simpler, well pretend that
Cj is exactly x1 x2 . . . xk . Then the probability that Cj is satisfied is
exactly 1 ki=1 (1 yi ). This quantity is minimized subject to ki=1 yi zj
Q P
by setting all yi equal to zj /k (easy application of Lagrange multipliers, or

can be shown using a convexity argument). Writing z for zj , this gives
k
Y
Pr [Cj is satisfied] = 1 (1 yi )
i=1
Yk
1 (1 z/k)
i=1
= 1 (1 z/k)k
z(1 (1 1/k)k ).
z(1 1/e).
The second-to-last step looks like a typo, but it actually works. The idea
is to observe that the function f (z) = 1 (1 z/k)k is concave (Proof:
d2
dz 2
f (z) = k1
k (1 z/k)
k2 < 0), while g(z) = z(1 (1 1/k)k ) is linear,
so since f (0) = 0 = g(0) and f (1) = 1 (1 1/k)k = g(1), any point in

between must have f (z) g(z). An example of this argument for k = 3 is
depicted in Figure 11.1.
Since each clause is satisfied with probability at least zj (1 1/e), the
expected number of satisfied clauses is at least (1 1/e) j zj , which is at
P
least (1 1/e) times the optimum. This gives an approximation ratio of

slightly more than 0.632, which is better than 1/2, but still kind of weak.
So now we apply the second trick from [GW94]: well observe that, on a
per-clause basis, we have a randomized rounding algorithm that is good at
satisfying small clauses (the coefficient (1 (1 1/k)k ) goes all the way up
to 1 when k = 1), and our earlier dumb algorithm that is good at satisfying
big clauses. We cant combine these directly (the two algorithms demand
Figure 11.1: Tricky step in MAX SAT argument, showing (1 (1 z/k)k )

z(1 (1 1/k)k ) when 0 z 1. The case k = 3 is depicted.
different assignments), but we can run both in parallel and take whichever
gives a better result.
To show that this works, let Xj be indicator for the event that clause j is
satisfied by the randomized-rounding algorithm and Yj the indicator for the
even that it is satisfied by the simple algorithm. Then if Cj has k literals,
E [Xj ] + E [Yj ] (1 2k ) + (1 (1 1/k)k )

zj
((1 2k ) + (1 (1 1/k)k ))
zj
= (2 2k (1 1/k)k )
zj .
The coefficient here is exactly 3/2 when k = 1 or k = 2, and rises thereafter,

so for integer
hP ki we have
hP E i[Xj ] + E [Yj ] (3/2)zj . Summing over all j then
j Yj (3/2)
P
gives E j Xj +E jzj . But then one of the two expected
P
sums must beat (3/4) j zj , giving us a (3/4)-approximation algorithm.
11.3 The Lovsz Local Lemma

Suppose we have a finite set of bad events A, and we want to show that with
nonzero
hT probability,
i none of these events occur. Formally, we want to show

Pr AA A > 0.
Our usual trick so far has been to use the union bound (4.1.1) to show that
P
AA Pr [A] < 1. But this only works if the events are actually improbable.
If the union bound doesnt work, we might be lucky enough to have the
hT i h i
events be independent; in this case, Pr AA A = AA Pr A > 0, as
Q
long as each event A occurs with positive probability. But most of the time,
well find that the events we care about arent independent, so this wont
work either.
The Lovsz Local Lemma [EL75] handles a situation intermediate
between these two extremes, where events are generally not independent of
each other, but each collection of events that are not independent of some
particular event A has low total probability. In the original version; its
non-constructive: the lemma shows a nonzero probability that none of the
events occur, but this probability may be very small if we try to sample
the events at random and there is no guidance for how to find a particular
outcome that makes all the events false.
Subsequent work [Bec91, Alo91, MR98, CS00, Sri08, Mos09, MT10]
showed how, when the events A are determined by some underlying set
of independent variables and independence between two events is detected
by having non-overlapping sets of underlying variables, an actual solution
could be found in polynomial expected time. The final result in this series,
due to Moser and Tardos [MT10], gives the same bounds as in the original
non-constructive lemma, using the simplest algorithm imaginable: whenever
some bad event A occurs, squish it by resampling all of its variables, and
continue until no bad events are left.
11.3.1 General version

A formal statement of the general lemma is:8
Lemma 11.3.1. Let A = A1 , . . . , Am be a finite collection of events on some

probability space, and for each A A, let (A) be a set of events such that
A is independent of all events not in + A = {A} (A). If there exist real
numbers 0 < xA < 1 such that, for all events A A
Y
Pr [A] xA (1 xB ), (11.3.1)
B(A)
then
" #
A
\ Y
Pr (1 xA ). (11.3.2)
AA AA
8
This version is adapted from [MT10].
In particular, this means that the probability that none of the Ai occur
is not zero, since we have assumed xAi < 1 holds for all i.
The role of xA in the original proof is to act as an upper bound on the
probability that A occurs given that some collection of other events doesnt
occur. For the constructive proof, the xA are used to show a bound on the
number of resampling steps needed until none of the A occur.
11.3.2 Symmetric version

For many applications, having to come up with the xA values can be awkward.
The following symmetric version is often used instead:
Corollary 11.3.2. Let A and be as in Lemma 11.3.1. Suppose that there

are constants p and d, such that for hall A iA, we have Pr [A] p and
|(A)| d. Then if ep(d + 1) < 1, Pr AA A 6= 0.
T
Proof. Basically, we are going to pick a single value x such that xA = x for
all A in A, and (11.3.1) is satisfied. This works as long as p x(1 x)d ,
d |(A)| =
as in for all A, Pr [A] p x(1 x) x(1 x)
Qthis case we have,
xA B(A) (1 xB ) .
d
For fixed d, x(1 x)d is maximized using the usual trick: dx x(1 x)d =
d d1 1
(1 x) xd(1 x) = 0 gives (1 x) xd = 0 or x = d+1 . So now we
d d
1 1 1
need p d+1 1 d+1 . It is possible to show that 1/e < 1 d+1 for
all d 0 (see the solution to Problem D.1.4). So ep(d + 1) 1 implies p
d
1
e(d+1) 1 1
d+1
1
d+1 x(1 x)|(A)| as required by Lemma 11.3.1.
11.3.3 Applications
11.3.3.1 Graph coloring
Lets start with a simple application of the local lemma where we know what
the right answer should be. Suppose we want to color the vertices of a cycle
with c colors, so that no edge has two endpoints with the same color. How
many colors do we need?
Using brains, we can quickly figure out that c = 3 is enough. Without
brains, we could try coloring the vertices randomly: but in a cycle with n
vertices and n edges, on average n/c of the edges will be monochromatic,
since each edge is monochromatic with probability 1/c. If these bad events
were independent, we could argue that there was a (1 1/c)n > 0 probability
that none of them occurred, but they arent, so we cant. Instead, well use
the local lemma.
The set of bad events A is just the set of events Ai = [edge i is monochromatic].
Weve already computed p = 1/c. To get d, notice that each edge only shares
a vertex with two other edges, so |(Ai )| 2. Corollary 11.3.2 then says that
there is a good coloring as long as ep(d + 1) = 3e/c 1, which holds as long
as c 9. Weve just shown we can 9-color a cycle. If we look more closely
2
at the proof of Corollary 11.3.2, we can see that p 31 1 31 = 27 4
would
also work; this says we can 7-color a cycle. Still not as good as what we can
do if we are paying attention, but not bad for a procedure that doesnt use
the structure of the problem much.
11.3.3.2 Satisfiability of k-CNF formulas

A more sophisticated application is demonstrating satisfiability for k-CNF
formulas where each variable appears in a bounded number of clauses. Recall
that a k-CNF formula consists of m clauses, each of which consists of
exactly k variables or their negations (collectively called literals. It is
satisfied if at least one literal in every clause is assigned the value true.
Suppose that each variable appears in at most ` clauses. Let A consist of
all the events Ai = [clause i is not satisfied]. Then, for all i, Pr [Ai ] = 2k
exactly, and |(Ai )| d = k(` 1) since each variable in Ai is shared with
at most ` 1 other clauses, and Ai will be independent of all events Aj with
which it doesnt share a variable. So if ep(d + 1) = e2k (k(` 1) + 1) 1,
k
which holds if ` 2ek +1, Corollary 11.3.2 tells us that a satisfying assignment
exists.9
Corollary 11.3.2 doesnt let us actually find a satisfying assignment, but
it turns out we can do that too. Well return to this when we talk about the
constructive version of the local lemma in 11.3.5
11.3.4 Non-constructive proof

This is essentially the same argument presented in [MR95, 5.5], but we
adapt the notation to match the statement in terms of neighborhoods (A)
instead of edges in a dependency graph.10
9
To avoid over-selling this claim, its worth noting that the bound on ` only reaches 2
at k = 4, although it starts growing pretty fast after that.
10
The two formulations are identical, since we can always represent the neighborhoods
(A) by creating an edge from A to B when B is in (A); and conversely we can convert
a dependency graph into a neighborhood structure by making (A) the set of successors of
We show by induction on |S| that for any A and any S A with A 6 S,

" #
\
Pr A
xA .
B (11.3.3)

BS
When |S| = 0, this just says Pr [A] xA , which follows immediately from
(11.3.1).
For larger S, split S into S1 = S (A), the events in S that might
not be independent of A; and S2 = S \ (A), the events in S that we know
to be independent of A. If S2 = S, then A is independent
h T iof all events

in S, and (11.3.3) follows immediately from Pr A BS B = Pr [A]

xA B(A) (1 xB ) xA . Otherwise |S2 | < |S|, which means that we can

Q
assume the induction hypothesis holds for S2 .

T
Write C1 for the event BS1 B and C2 for the event T
BS2 B. Then
" #
= Pr [A C1 C2 ]
\
Pr A B

BS
Pr [C1 C2 ]
Pr [A C1 | C2 ] Pr [C2 ]
=
Pr [C1 | C2 ] Pr [C2 ]
Pr [A C1 | C2 ]
= . (11.3.4)
Pr [C1 | C2 ]
We dont need to do anything particularly clever with the numerator:
Pr [A C1 | C2 ] Pr [A | C2 ]
= Pr [A]

Y
xA (1 xB ) , (11.3.5)
B(A)
from (11.3.1) and the fact that A is independent of all B in S2 and thus of
C2 .
T
For the denominator, we expand C1 backh out to BSi1 B and break
out the induction hypothesis. To bound Pr
T
B C2 , we order S1
BS1
arbitrarily as {B1 , . . . , Br }, for some r, and show by induction on ` as ` goes
from 1 to r that
" ` # `

i C2
\ Y
Pr B (1 xBi ). (11.3.6)

i=1 i=1
A in the dependency graph.

The proof is that, for ` = 1,

h i
1 C2 = 1 Pr [B1 | C2 ]
Pr B
1 xB1
using the outer induction hypothesis (11.3.3), and for larger `, we can compute
" ` # " `1
! # "`1 #

i C2 = Pr B` i C2 Pr i C2
\ \ \
Pr B B B

i=1 i=1 i=1
`1
Y
(1 xB` ) (1 xBi )
i=1
`
Y
= (1 xBi ),
i=1
where the second-to-last step uses the outer induction hypothesis (11.3.3)
for the first term and the inner induction hypothesis (11.3.6) for the rest.
This completes the proof of the inner induction.
When ` = r, we get
" r #

i C2
\
Pr [C1 | C2 ] = Pr B

i=1
Y
(1 xB ). (11.3.7)
BS1
Substituting (11.3.5) and (11.3.7) into (11.3.4) gives

Q

x (1 x ) ,
" #
\ A B(A) B
Pr A

B
BS1 (1 xB )
Q

BS

Y
= xA (1 xB )
B(A)\S1
xA .
This completes the proof of the outer induction.
To get the bound (11.3.2), we reach back inside the proof and repeat
the argument for (11.3.7) with AA A in place of C1 and without the
T
conditioning on C2 . We order A arbitrarily as {A1 , A2 , . . . , Am } and show

by induction on k that
" k # k
Ai
\ Y
Pr (1 xAk ). (11.3.8)
i=1 i=1
For the base case we have k = 0 and Pr [] 1, using the usual conventions
on empty products. For larger k, we have
" k
i k1
#
h
Ai = Pr Ak Ai
\ \
Pr
i=1 i=1
k1
Y
(1 xAk ) (1 xAi )
i=1
k
Y
(1 xAk ),
i=1
where in the second-to-last step we use (11.3.3) for the first term and (11.3.8)
for the big product.
Setting k = n finishes the proof.
11.3.5 Constructive proof

We now describe he constructive proof of the Lovsz local lemma due to
Moser and Tardos [MT10], which is based on a slightly more specialized
construction of Moser alone [Mos09]. This version applies when our set of
bad events A is defined in terms of a set of independent variables P, where
each A A is determined by some set of variables vbl(A) P, and (A) is
defined to be the set of all events B 6= A that share at least one variable
with A; i.e., (A) = {B A {A} | vbl(B) vbl(A) 6= }.
In this case, we can attempt to find an assignment to the variables that
makes none of the A occur using the obvious algorithm of sampling an initial
state randomly, then resampling all variables in vbl(A) whenever we see some
bad A occur. Astonishingly, this actually works in a reasonable amount of
time, without even any cleverness in deciding which A to resample at each
step, if the conditions for Lemma 11.3.1 hold for x that are not too large.
In particular, we will show that a good assignment is found after each A is
xA
resampled at most 1x A
times on average.
Lemma 11.3.3. Under the conditions of Lemma 11.3.1, the Moser-Tardos

procedure does at most
X xA
A
1 xA
resampling steps on average.

We defer the proof of Lemma 11.3.3 for the moment. For most applica-
tions, the following symmetric version is easier to work with:
Corollary 11.3.4. Under the conditions of Corollary 11.3.2, the Moser-

Tardos procedure does at most m/d resampling steps on average.
1
Proof. Follows from Lemma 11.3.3 and the choice of xA = d+1 in the proof
of Corollary 11.3.2.
How this expected m/d bound translates into actual time depends on
the cost of each resampling step. The expensive part at each step is likely to
be the cost of finding an A that occurs and thus needs to be resampled.
Intuitively, we might expect the resampling to work because if each A A
has a small enough neighborhood (A) and a low enough probability, then
whenever we resample As variables, its likely that we fix A and unlikely
that we break too many B events in As neighborhood. It turns out to be
tricky to quantify how this process propagates outward, so the actual proof
uses a different idea that essentially looks at this process in reverse, looking
for each resampled event A at a set of previous events whose resampling we
can blame for making A occur, and then showing that this tree (which will
include every resampling operation as one of its vertices) cant be too big.
The first step is to fix some strategy for choosing which event A to
resample at each step. We dont care what this strategy is; we just want
it to be the case that the sequence of events depends only on the random
choices made by the algorithm in its initial sampling and each resampling.
We can then define an execution log C that lists the sequence of events
C1 , C2 , C3 , . . . that are resampled at each step of the execution.
From C we now construct a witness tree Tt for each resampling step
t whose nodes are labeled with events, with the property that the children
of a node v labeled with event Av are labeled with events in + (Av ) =
{Av } (Av ). The root of Tt is labeled with Ct ; to construct the rest of
the tree, we work backwards through Ct1 , Ct2 , . . . , C1 , and for each event
Ci we encounter we attach Ci as a child of the deepest v we can find with
Ci + (Av ), choosing arbitrarily if there is more than one such v, and
discarding Ci if there is no such v.
Now we can ask what the probability is that we see some particular
witness tree T in the execution log. Each vertex of T corresponds to some
event Av that we resample because it occurs; in order for it to occur, the
previous assignments of each variable in vbl(Av ) must have made Av true,
which occurs with probability Pr [Av ]. But since we resample all the variables
in Av , any subsequent assignments to these variables are independent of the
ones that contributed to v; with sufficient handwaving (or a rather detailed

coupling argument as found in [MT10]) this gives that each event Av occurs
Q
with independent probability Pr [Av ], giving Pr [T ] = vT Pr [Av ].
Why do we care about this? Because every event we resample is the root
of some witness tree, and we can argue that every event we resample is the
root of a distinct witness tree. The proof is that since we only discard events
B that have vbl(B) disjoint from all nodes already in the tree, once we put
A at the root, any other instance of A gets included. So the witness tree
rooted at the i-th occurrence of A in C will include exactly i copies of A,
unlike the witness tree rooted at the j-th copy for j 6= i.
Now comes the sneaky trick: well count how many distinct witness trees
T we can possibly have rooted at A, given that each occurs with probability
Q
vT Pr [Av ]. This is done by constructing a branching process using the
xB values from Lemma 11.3.1 as probabilities of a node with label A having
a kid labeled B for each B in + (A), and doing algebraic manipulations on
Q
the resulting probabilities until vT Pr [Av ] shows up.
Formally, consider the process where we construct a tree by starting
with a root labeled A, and for each vertex v with label Av , giving it a child
labeled B for each B + (Av ) with independent probability xB . Well now
calculate the probability pT that this process generates a particular tree T .
Let x0B = xB C(B) (1 xC ). Note that (11.3.1) says precisely that
Q
Pr [B] x0B .
For each vertex v in T , let Wv + (Av ) be the set of events B + (Av )
that dont occur as labels of children of v. The probability of getting T is
equal to the product of the probabilities at each v of getting all of its children
and none of its non-children. The non-children of v collectively contribute
BWv (1 xB ) to the product, and v itself contributes xAv (via the product
Q
for its parent), unless v is the root node. So we can express the giant product
as

1 Y Y
pT = xAv (1 xB ) .
xA vT BW v
We dont like the Wv very much, so we get rid of them by pushing a (1 xAv )
up from each vertex v, which we compensate for by dividing by 1 xAv at v
itself, with an extra 1 xA showing up at the root. This gives

1 xA Y xAv Y
pT = (1 xB ) .
xA vT 1 xAv
B+ (Av )

1 xA Y Y
= xAv (1 xB ) .
xA vT B(A ) v
1 xA Y 0
= xAv .
xA vT
Now we can bound the expected number of trees rooted at A that appear
in C, assuming (11.3.1) holds. Letting TA be the set of all such trees and
NA the number that appear in C, we have
X
E [NA ] = Pr [T appears in C]
T TA
X Y
Pr [A(v)]
T TA vT
x0Av
X Y

T TA vT
X xA
= pT
T TA
1 xA
xA X
= pT
1 xA T T
A
xA
.
1 xA
And were done.
Chapter 12
Derandomization
Derandomization is the process of taking a randomized algorithm and

turning it into a deterministic algorithm. This is useful both for practical
reasons (deterministic algorithms are more predictable, which makes them
easier to debug and gives hard guarantees on running time) and theoretical
reasons (if we can derandomize any randomized algorithm we could show
results like P = RP,1 which would reduce the number of complexity classes
that complexity theorists otherwise have to deal with). It may also be the
case that derandomizing a randomized algorithm can be used for probability
amplification, where we replace a low probability of success with a higher
probability, in this case 1.
There are basically two approaches to derandomization:
1. Reduce the number of random bits used down to O(log n), and then
search through all choices of random bits exhaustively. For example, if
we only need pairwise independence, we could use the XOR technique
from 5.1.2.1 to replace a large collection of variables with a small
collection of random bits.
Except for the exhaustive search part, this is how randomized algo-
rithms are implemented in practice: rather than burning random bits
continuously, a pseudorandom generator is initialized from a seed
consisting of a small number of random bits. For pretty much all of
the randomized algorithms we know about, we dont even need to use
a particularly strong pseudorandom generator. This is largely because
1
The class RP consists of all languages L for which there is a polynomial-time random-
ized algorithm that correctly outputs yes given an input x in L with probability at least
1/2, and never answers yes given an input x not in L. See 1.4.2 for a more extended
description of RP and other randomized complexity classes.
214
CHAPTER 12. DERANDOMIZATION 215
current popular generators are the products of a process of evolution:

pseudorandom generators that cause wonky behavior or fail to pass
tests that approximate the assumptions made about them by typical
randomized algorithms are abandoned in favor of better generators.2
From a theoretical perspective, pseudorandom generators offer the
possibility of eliminating randomization from all randomized algorithms,
except there is a complication. While (under reasonable cryptographic
assumptions) there exist cryptographically secure pseudorandom
generators whose output is indistinguishable from a genuinely random
source by polynomial-time algorithms (including algorithms originally
intended for other purposes), such generators are inherently incapable
of reducing the number of random bits down to the O(log n) needed
for exhaustive search. The reason is that any pseudorandom generator
with only polynomially-many seeds cant be cryptographically secure,
because we can distinguish it from a random source by just checking
its output against the output for all possible seeds. Whether there
is some other method for transforming an arbitrary algorithm in RP
or BPP into a deterministic algorithm remains an open problem in
complexity theory (and beyond the scope of this course).
2. Start with a specific randomized protocol and analyze its behavior

enough that we can replace the random bits it uses with specific,
deterministically-chosen bits we can compute. This is the main ap-
proach we will describe below. A non-constructive variant of this shows
that we can always replace the random bits used by all inputs of a given
size with a few carefully-selected fixed sequences (Adlemans Theorem,
described in 12.2). More practical is the method of conditional
probabilities, which chooses random bits sequentially based on which
value is more likely to give a good outcome (see 12.4).
12.1 Deterministic vs. randomized algorithms

In thinking about derandomization, it can be helpful to have more than
one way to look at a randomized algorithm. So far, weve describe random-
ized algorithms as random choices of deterministic algorithms (Mr (x)) or,
2
Having cheaper computers helps here as well. Nobody would have been willing to
spend 2496 bytes on the state vector for Mersenne Twister [MN98] back in 1975, but by
1998 this amount of memory was trivial for pretty much any computing device except the
tiniest microcontrollers.
equivalently, as deterministic algorithms that happen to have random inputs

(M (r, x)). This gives a very static view of how randomness is used in the
algorithm. A more dynamic view is obtained by thinking of the computation
of a randomized algorithm as a computation tree, where each path through
the tree corresponds to a computation with a fixed set of random bits and a
branch in the tree corresponds to a random decision. In either case we want
an execution to give us the right answer with reasonably high probability,
whether that probability measures our chance of getting a good deterministic
machine for our particular input or landing on a good computation path.
12.2 Adlemans theorem

The idea of picking a good deterministic machine is the basis for Adlemans
theorem[Adl78], a classic result in complexity theory. Adlemans theorem
says that we can always replace randomness by an oracle that presents
us with a fixed string of advice pn that depends only on the size of the
input n and has size polynomial in n. The formal statement of the theorem
relates the class RP, which is the class of problems for which there exists
a polynomial-time Turing machine M (x, r) that outputs 1 at least half the
time when x L and never when x 6 L; and the class P/poly, which is the
class of problems for which there is a polynomial-sized string pn for each
input size n and a polynomial-time Turing machine M 0 such that M 0 (x, p|x| )
outputs 1 if and only if x L.
Theorem 12.2.1. RP P/poly.
Proof. The intuition is that if any one random string has a constant proba-
bility of making M happy, then by choosing enough random strings we can
make the probability that M fails using on every random string for any given
input so small that even after we sum over all inputs of a particular size,
the probability of failure is still small using the union bound (4.1.1). This is
an example of probability amplification, where we repeat a randomized
algorithm many times to reduce its failure probability.
Formally, consider any fixed input x of size n, and imagine running M
repeatedly on this input with n + 1 independent sequences of random bits
r1 , r2 , . . . , rn+1 . If x 6 L, then M (x, ri ) never outputs 1. If x L, then for
each ri , there is an independent probability of at least 1/2 that M (x, ri ) = 1.
So Pr [M (x, ri ) = 0] 1/2, and Pr [iM (x, ri ) = 0] 2(n+1) . If we sum
this probability of failure for each individual x L of length n over the at
most 2n such elements, we get a probability that any of them fail of at most
2n 2(n+1) = 1/2. Turning this upside down, any sequence of n + 1 random

inputs includes a witness that x L for all inputs x with probability at
least 1/2. It follows that a good sequence r1 , . . . , rn+1 , exists.
Our advice pn is now some good sequence pn = hr1 . . . rn+1 i, and the
deterministic advice-taking algorithm that uses it runs M (x, ri ) for each
ri and returns true if and only if at least one of these executions returns
true.
The classic version of this theorem shows that anything you can do with
a polynomial-size randomized circuit (a circuit made up of AND, OR, and
NOT gates where some of the inputs are random bits, corresponding to the r
input to M ) can be done with a polynomial-size deterministic circuit (where
now the pn input is baked into the circuit, since we need a different circuit
for each size n anyway).
A limitation of this result is that ordinary algorithms seem to be better
described by uniform families of circuits, where there exists a polynomial-
time algorithm that, given input n, outputs the circuit Cn for processing size-n
inputs. In contrast, the class of circuits generated by Adlemans theorem
is most likely non-uniform: the process of finding the good witnesses ri is
not something we know how to do in polynomial time (with the usual caveat
that we cant prove much about what we cant do in polynomial time).
12.3 Limited independence

For some algorithms, it may be that full independence is not needed for all
of the random bits. If the amount of independence needed is small enough,
this may allow us to reduce the actual number of random bits consumed
down to O(log n), at which point we can try all possible sequences of random
bits in polynomial time.
Variants of this technique have been used heavily in the cryptography
and complexity; see [LW05] for a survey of some early work in this area.
Well do a quick example of the method before before moving onto more
direct approaches.
12.3.1 MAX CUT

Lets look at the randomized MAX CUT algorithm from 11.2.1. In the
original algorithm, we use n independent random bits to assign the n vertices
to S or T . The idea is that by assigning each vertex independently at
random to one side of the cut or the other, each edge appears in the cut with
probability 1/2, giving a total of m/2 edges in the cut in expectation.
Suppose that we replace these n independent random bits with n pairwise-
independent bits generated by taking XORs of subsets of dlg(n + 1)e inde-
pendent random bits as described in 5.1.2.1. Because the bits are pairwise-
independent, the probability that the two endpoints of an edge are assigned
to different sides of the cut is still exactly 1/2. So on average we get m/2
edges in the cut as before, and there is at least one sequence of random bits
that guarantees a cut at least this big.
But with only dlg(n+1)e random bits, there are only 2dlog(n+1)e < 2(n+1)
possible sequences of random bits. If we try all of them, then we find a cut
of size m/2 always. The total cost is O(n(n + m)) if we include the O(n + m)
cost of testing each cut. Note that this algorithm does not generate all 2n
possible cuts, but among those it does generate, there must be a large one.
In this particular case, well see below how to get the same result at a
much lower cost, using more knowledge of the problem. So we have the
typical trade-off between algorithm performance and algorithm designer
effort.
12.4 The method of conditional probabilities

The method of conditional probabilities [Rag88] follows an execution
of the randomized algorithm, but at each point where we would otherwise
make a random decision, it makes a decision that minimizes the conditional
probability of losing.
Structurally, this is similar to the method of bounded differences (see
5.3.3). Suppose our randomized algorithm generates m random values
X1 , X2 , . . . , Xm . Let f (X1 , . . . , Xm ) be the indicator variable for our ran-
domized algorithm failing (more generally, we can make it an expected cost
or some other performance measure). Extend f to shorter sequences of val-
ues by defining f (x1 , . . . , xk ) = E [f (x1 , . . . , xk , Xk+1 , . . . , Xm )]. Then Yk =
f (X1 , . . . , Xk ) is a Doob martingale, just as in the method of bounded differ-
ences. This implies that, for any partial sequence of values x1 , . . . , xk , there
exists some next value xk+1 such that f (x1 , . . . , xk ) f (x1 , . . . , xk , xk+1 ). If
we can find this value, we can follow a path on which f always decreases, and
obtain an outcome of the algorithm f (x1 , . . . , xm ) less than or equal to the
initial value f (hi). If our outcomes are 01 (as in failure probabilities), and
our initial value for f is less than 1, this means that we reach an outcome
with f = 0.
The tricky part here is that it may be hard to compute f (x1 , . . . , xk ).

(Its always possible to do so in principle by enumerating all assignments of
the remaining variables, but if we have time to do this, we can just search
for a winning assignment directly.) What makes the method of conditional
probabilities practical in many cases is that its not necessary for f to compute
the actual probability of failure, as long as (a) f gives an upper bound on the
real probability of failure, at least in terminal configurations, and (b) f has the
property used in the argument that for any partial sequence x1 , . . . , xk there
exists an extension x1 , . . . , xk , xk+1 with f (x1 , . . . , xk ) f (x1 , . . . , xk , xk+1 ).
Such an f is called a pessimistic estimator. If we can find a pessimistic
estimator that is easy to compute and starts out less than 1, then we can
just follow it down the tree to a leaf that doesnt fail.
12.4.1 A trivial example

Here is a very bad randomized algorithm for generating a string of n zeros:
flip n coins, and output the results. This has a 1 2n probability of failure,
which is not very good.
We can derandomize this algorithm and get rid of all probability of failure
at the same time. Let f (x1 , . . . , xk ) be the probability of failure after the
first k bits. Then we can easily compute f (its 1 if any of the bits are 1
and 1 2(nk) otherwise). We can use this to find a bit xk+1 at each stage
that reduces f . After noticing that all of these bits are zero, we improve the
deterministic algorithm even further by leaving out the computation of f .
12.4.2 Deterministic construction of Ramsey graphs

Here is an example that is slightly less trivial. For this example we let f be
a count of bad events rather than a failure probability, but the same method
applies.
Recall from 11.1.2 that if k 3, for n 2k/2 there exists a graph with
n nodes and no cliques or independent sets of size k. The proof of this fact
is to observe that each subset of k vertices is bad with probability 2k+1 ,
and when nk 2k+1 < 1, the expected number of bad subsets in Gn,1/2 is less
than 1, showing that some good graph exists.
We can turn this into a deterministic nO(log n) algorithm for finding a
Ramsey graph in the worst case when n = 2k/2 . The trick is to set the
edges to be present or absent one at a time, and for each edge, take the
value that minimizes the expected number of bad subsets conditioned on
the choices so far. We can easily calculate the conditional probability that
a subset is bad in O(k) time: if it already has both a present and missing
edge, its not bad. Otherwise, if weve already set ` of its edges to the
same value, its bad with probability exactly 2k+` . Summing this over all
O(nk ) subsets takes
O(nk k) time 2
per edge, and we have O(n ) edges, giving
O(nk+2 k) = O n(2 lg n+2+lg lg n) = nO(log n) time total.
Its worth mentioning that there are better deterministic constructions
in the literature. The best construction that I am aware of is given by an
c
algorithm of Barak et al. [BRSW06], which constructs a graph with k log k
vertices with no clique or independent set of size k for any fixed c.
12.4.3 MAX CUT using conditional probabilities

Again we consider the algorithm from 11.2.1 that assigns vertices to S
and T at random. To derandomize this algorithm, at each step we pick a
vertex and assign it to the side of the cut that maximizes the conditional
expectation of the number of edges that cross the cut. We can compute this
conditional expectation as follows:
1. For any edge that already has both endpoints assigned to S or T , its
either in the cut or it isnt: add 0 or 1 as appropriate to the conditional
expectation.
2. For any edge with only one endpoint assigned, theres a 1/2 probability
that the other endpoint gets assigned to the other side (in the original
randomized algorithm). Add 1/2 to the conditional expectation for
these edges.
3. For any edge with neither endpoint assigned, we again have a 1/2
probability that it crosses the cut. Add 1/2 for these as well.
So now let us ask how assigning a particular previously unassigned vertex

v to S or T affects the conditional probability. For any neighbor w of v that
is not already assigned, adding v to S or T doesnt affect the 1/2 contribution
of vw. So we can ignore these. The only effects we see are that if some
neighbor w is in S, assigning v to S decreases the conditional expectation by
1/2 and assigning v to T increases the expectation by 1/2. So to maximize
the conditional expectation, we should assign v to whichever side currently
holds fewer of vs neighborsthe obvious greedy algorithm, which runs in
O(n + m) time if we are reasonably clever about keeping track of how many
neighbors each unassigned node has in S and T . The advantage of the
method of conditional probabilities here is that we immediately get that
the greedy algorithm achieves a cut of size m/2, which might require actual
intelligence to prove otherwise.
12.4.4 Set balancing

Here we have a collection of vectors v1 , v2 , . . . , vn in {0, 1}m . Wed like to find
1 coefficients 1 , 2 , . . . n that minimize maxj |Xj | where Xj = ni=1 i vij .
P
If we choose the i randomly, Hoeffdings inequality (5.3.1) says for each

fixed j that Pr [|Xj | > t] < 2 exp(t2 /2n) (since there are at most n non-zero
values vij ). Setting 2 exp(t 2 /2n) 1/m gives ti 2n ln 2m. So by the
h
union bound, we have Pr maxj |Xj | > 2n ln 2m < 1: a solution exists.
Because there may be very complicated dependence between the Xj , it
is difficult to calculate the probability of the event j [|Xj | t], whether
S
conditioned on some of the i or not. However, we can calculate the proba-

bility of the individual events [|Xj | t] exactly. Conditioning on 1 , . . . , k ,
the expected value of Xj is just ` = ki=1 i vij , and the distribution of
P
Y = Xj E [Xj ] is the sum r n k independent 1 random variables. So
Pr [|Xj | > t | 1 , . . . , k ] = 1 Pr [t ` Y t `]
t`
!
X r r
= 2 .
i=t`
i
This last expression involves a linear number of terms, each of which we can
calculate using a linear number of operations on rational numbers that fit
in a linear number of bits, so we can calculate the probability exactly in
polynomial time by just adding them up.
For our pessimistic estimator, we take
n
X h i
U (1 , . . . , k ) = Pr |Xj | > 2n ln 2m 1 , . . . , k .

i=j
Note that we can calculate each term in the sum by adding up a big pile of
binomial coefficients. Furthermore, since each term in the sum is a Doob mar-
tingale, the sum is a martingale as well, so E [U (1 , . . . , k+1 ) | 1 , . . . , k ] =
U (1 , . . . , k ). It follows that for any choice of 1 , . . . , k there exists some
k+1 such that U (1 , . . . , k ) U (1 , . . . , k+1 ), and we can determine this
winning k+1 explicitly. Our previous argument shows that U (hi) < 1, which
implies that our final value U (1 , . . . , n ) will also be less than 1. Since U
is an integer,
this means it must be 0, and we find an assignment in which
|Xj | < 2n ln 2m for all j.
Chapter 13
Quantum computing
Quantum computing is a currently almost-entirely-theoretical branch of

randomized algorithms that attempts to exploit the fact that probabilities
at a microscopic scale arise in a mysterious way from more fundamental
probability amplitudes, which are complex-valued and can cancel each
other out where probabilities can only add. In a quantum computation, we
replace random bits with quantum bitsqubits for shortand replace
random updates to the bits with quantum operations.
To explain how this works, well start by re-casting our usual model of a
randomized computation to make it look more like the standard quantum
circuits of Deutsch [Deu89]. Well then get quantum computation by
replacing all the real-valued probabilities with complex-valued amplitudes.
13.1 Random circuits

Lets consider a very simple randomized computer whose memory consists of
two bits. We can describe our knowledge of the state of this machine using a
vector of length 4, with the coordinates in the vector giving the probability
of states 00, 01, 10, and 11. For example, if we know that the initial state is
always 00, we would have the (column) vector

1
0
.

0
0
Any such state vector x for the system must consist of non-negative
real values that sum to 1; this is just the usual requirement for a discrete
222
CHAPTER 13. QUANTUM COMPUTING 223
probability space. Operations on the state consist of taking the old values of
one or both bits and replacing them with new values, possibly involving both
randomness and dependence on the old values. The law of total probability
applies here, so we can calculate the new state vector x0 from the old state
vector x by the rule
x0b0 b0 = xb1 b2 Pr Xt+1 = b01 b02 Xt = b1 b2 .

X
1 2
b1 b2
These are linear functions of the previous state vector, so we can summarize
the effect of our operation using a transition matrix A, where x0 = Ax.1
For example, if we negate the second bit 2/3 of the time while leaving
the first bit alone, we get the matrix

1/3 2/3 0 0
2/3 1/3 0 0
A= .

0 0 1/3 2/3
0 0 2/3 1/3
One way to derive this matrix other than computing each entry directly is
that it is the tensor product of the matrices that represent the operations
on each individual bit. The idea here is that the tensor product of A and B,
written A B, is the matrix C with Cij,k` = Aik Bj` . Were cheating a little
bit by allowing the C matrix to have indices consisting of pairs of indices,
one for each of A and B; there are more formal definitions that justify this
at the cost of being harder to understand.
In this particular case, we have

1/3 2/3 0 0 " # " #
2/3 1/3 0 0 1 0 1/3 2/3
= .

0 0 1/3 2/3 0 1 2/3 1/3

0 0 2/3 1/3
The first matrix in the tensor product gives the update rule for the first bit
(the identity matrixdo nothing), while the second gives the update rule for
the second.
1
Note that this is the reverse of the convention we adopted for Markov chains in
Chapter 9. There it was convenient to have Pij = pij = Pr [Xt+1 = j | Xt = i]. Here we
defer to the physicists and make the update operator come in front of its argument, like
any other function.
Some operations are not decomposable in this way. If we swap the values
of the two bits, we get the matrix

1 0 0 0
0 0 1 0
S= ,

0 1 0 0
0 0 0 1
which maps 00 and 11 to themselves but maps 01 to 10 and vice versa.

The requirement for all of these matrices is that they be stochastic. This
means that each column has to sum to 1, or equivalently that 1A = 1, where
1 is the all-ones vector. This just says that our operations map probability
distributions to probability distributions; we dont apply an operation and
find that the sum of our probabilities is now 3/2 or something. (Proof: If
1A = 1, then |Ax|1 = 1(Ax) = (1A)x = 1x = |x|1 .)
A randomized computation in this model now consists of a sequence
of these stochastic updates to our random bits, and at the end performing a
measurement by looking at what the values of the bits actually are. If we
want to be mystical about it, we could claim that this measurement collapses
the probability distribution over states representing our previous ignorance
into a single unique state, but really we are opening the box to see what we
got.
For example, we could generate two bias-2/3 coin-flips by starting with
00 and using the algorithm flip second, swap bits, flip second, or in matrix
terms:
xout = ASAxin

1/3 2/3 0 0 1 0 0 0 1/3 2/3 0 0 1
2/3 1/3 0 0 0
0 1 0 2/3 1/3 0
0 0

=

0 0 1/3 2/3 0 1 0 0 0 0 1/3 2/3 0

0 0 2/3 1/3 0 0 0 1 0 0 2/3 1/3 0

1/9
2/9
= .

2/9
4/9
When we look at the output, we find 11 with probability 4/9, as wed

expect, and similarly for the other values.
13.2 Bra-ket notation

A notational oddity that scares many people away from quantum mechanics in
general and quantum computing in particular is the habit of practitioners in
these fields of writing basis vectors and their duals using bra-ket notation,
a kind of typographical pun invented by the physicist Paul Dirac [Dir39].
This is based on a traditional way of writing an inner product of two
vectors x and y in bracket form as hx|yi. The interpretation of this is
hx|yi = x y, where x is the conjugate transpose of x.
For real-valued x the conjugate transpose is the same as the transpose.
For complex-valued x, each coordinate xi = a + bi is replaced by its complex
conjugate x i = a bi.) Using the conjugate transpose makes hx|xi equal
|x|22 when x is complex-valued.
For example, for our vector xin above that puts all of its probability on
00, we have

1
h i 0
hxin |xin i = 1 0 0 0 = 1. (13.2.1)

0
0
The typographic trick is to split in half both hx|yi and its expansion. For
example, we could split (13.2.1) as

1
h i 0
hxin | = 1 0 0 0 |xin i = .

0
0
In general, wherever we used to have a bracket hx|yi, we now have a bra
hx| and a ket |yi. These are just row vectors and column vectors, and hx| is
always the conjugate transpose of |xi.
The second trick in bra-ket notation is to make the contents of the bra or
ket an arbitrary name. For kets, this will usually describe some state. As an
example, we might write xin as |00i to indicate that its the basis vector that
puts all of its weight on the state 00. For bras, this is the linear operator
that returns 1 when applied to the given state and 0 when applied to any
orthogonal state. So h00| |00i = h00|00i = 1 but h00| |01i = h00|01i = 0.
13.2.1 States as kets

This notation is useful for the same reason that variable names are useful.
Its a lot easier to remember that |01i refers to the distribution assigning
h i>
probability 1 to state 01 than that 0 1 0 0 is.
Non-basis vectors we can express using a linear combination of kets. For
example, we can write
1 2 2 4
xout = |00i + |01i + |10i + |11i .
9 9 9 9
This is not as compact as the vector, but has the advantage of clearly labeling
what states the probabilities apply to.
13.2.2 Operators as sums of kets times bras

A similar trick can be used to express operators, like the swap operator S.
We can represent S as a combination of maps from specific states to specific
other states. For example, the operator

0 0 0 0 0
1 h i 0 0 1 0
|01i h10| = 0 0 1 0 =

0 0 0 0 0

0 0 0 0 0
maps |10i to |01i (Proof: |01i h10| |10i = |01i h10|10i = |01i) and sends all
other states to 0. Add up four of these mappings to get

1 0 0 0
0 0 1 0
S = |00i h00| + |10i h01| + |01i h10| + |11i h11| = .

0 1 0 0
0 0 0 1
Here the bra-ket notation both labels what we are doing and saves writing a
lot of zeros.
13.3 Quantum circuits

So how do we turn our random circuits into quantum circuits?
The first step is to replace our random bits with quantum bits (qubits).
For a single random bit, the state vector represents a probability distri-
bution
p0 |0i + p1 |1i ,
where p0 and p1 are non-negative real numbers with p0 + p1 = 1.

For a single qubit, the state vector represents amplitudes
a0 |0i + a1 |1i ,
where a0 and a1 are complex numbers with |a0 |2 + |a1 |2 = 1.2 The reason for
this restriction on amplitudes is that if we measure a qubit, we will see state 0
with probability |a0 |2 and state 1 with probability |a1 |2 . Unlike with random
bits, these probabilities are not mere expressions of our ignorance but arise
through a still somewhat mysterious process from the more fundamental
amplitudes.3
With multiple bits, we get amplitudes for all combinations of the bits,
e.g.
1
(|00i + |01i + |10i + |11i)
2
gives a state vector in which each possible measurement will be observed
2
1
with equal probability 2 = 14 . We could also write this state vector as

1/2
1/2
.

1/2

1/2
2
The absolute value, norm, or magnitude |a + bi| of a complex number is given by
a2 + b2 . When b = 0, this is the same as the absolute value for the corresponding
real
number. For any complex number x, the norm p can also be written
p as x x, where
x
is
the complex conjugate of x. This is because (a + bi)(a bi) = a2 (bi)2 = a2 + b2 .
The appearance of the complex conjugate here explains why we define hx|yi = x y; the
conjugate transpose means that for hx|xi, when we multiply xi by xi we are computing a
squared norm.
3
In the old days of shut up and calculate, this process was thought to involve the
unexplained power of a conscious observer to collapse a superposition into a classical
state. Nowadays the most favored explanation involves decoherence, the difficulty of
maintaining superpositions in systems that are interacting with large, warm objects with lots
of thermodynamic degrees of freedom (measuring instruments, brains). The decoherence
explanation is particularly useful for explaining why real-world quantum computers have a
hard time keeping their qubits mixed even when nobody is looking at them. Decoherence by
itself does not explain which basis states a system collapses to. Since bases in linear algebra
are pretty much arbitrary, it would seem that we could end up running into a physics
version of Goodmans grue-bleen paradox [Goo83], but there are ways of dealing with this
too (einselection). Since all of this is (a) well beyond my own limited comprehension of
quantum mechanics and (b) irrelevant to the theoretical model we are using, these issues
will not be discussed further.
13.3.1 Quantum operations

In the random circuit model, at each step we pick a small number of random
bits, apply a stochastic transformation to them, and replace their values
with the results. In the quantum circuit model, we do the same thing, but
now our transformations must have the property of being unitary. Just as
a stochastic matrix preserves the property that the probabilities in a state
vector sum to 1, a unitary matrix preserves the property that the squared
norms of the amplitudes in a state vector sum to 1.
Formally, a square, complex matrix A is unitary if it preserves inner
products: hAx|Ayi = hx|yi for all x and y. Alternatively, A is unitary if
A A = AA = I, where A is the conjugate transpose of A, because if this
holds, hAx|Ayi = (Ax) (Ay) = x A Ay = x Iy = x y = hx|yi. Yet another
way to state this is that the columns of A form an orthonormal basis: this
means that hAi |Aj i = 0 if i 6= j and 1 if i = j, which is just a more verbose
way to say A A = I. The same thing also works if we consider rows instead
of columns.
The rule then is: at each step, we can operate on some constant number
of qubits by applying a unitary transformation to them. In principle, this
could be any unitary transformation, but some particular transformations
show up often in actual quantum algorithms.4
The simplest unitary transformations are permutations on states (like
the operator that swaps two qubits), and rotations of a single state. One
particularly important rotation is the Hadamard operator
r " #
1 1 1
H= .
2 1 1
q q
This maps |0i to the superposition 12 |0i + 12 |1i; since this superposition
collapses to either |0i or |1i with probability 1/2, the state resulting from
H |0i is q
the quantum-computing
q equivalent of a fair coin-flip. Note that
H |1i = 12 |0i 12 |1i 6= H |0i. Even though both yield the same prob-
abilities; these two superpositions have different phases and may behave
differently when operated on further. That H |0i and H |1i are different is
necessary, and indeed a similar outcome occurs for any quantum operation:
all quantum operations are reversible, because any unitary matrix U has
an inverse U .
4
Deutschs original paper [Deu89] shows that repeated applications of single-qubit
rotations and the CNOT operation (described in 13.3.2) are enough to approximate any
unitary transformation.
If we apply H in parallel to all the qubits in our system, we get the n-fold
tensor product H n , which (if we take our bit-vector indices
q as integers
1
0 . . . N 1 = 2n 1 represented in binary) maps |0i to N1 N i=0 |ii. So
P
n applications of H effectively scramble a deterministic initial state into a

uniform distribution across all states. Well see this scrambling operation
again when we look at Grovers algorithm in 13.5.
13.3.2 Quantum implementations of classical operations

One issue that comes up with trying to implement classical algorithms in the
quantum-circuit model is that classical operation are generally not reversible:
if I execution x x y, it may not be possible to reconstruct the old state
of x. So I cant implement AND directly as a quantum operation.
The solution is to use more sophisticated reversible operations from which
standard classical operations can be extracted as a special case. A simple
example is the controlled NOT or CNOT operator, which computes the
mapping (x, y) 7 (x, x y). This corresponds to the matrix (over the basis
|00i , |01i , |10i , |11i)

1 0 0 0
0 1 0 0
,

0 0 0 1

0 0 1 0
which is clearly unitary (the rows are just the standard basis vectors). We
could also write this more compactly as |00i h00| + |01i h01| + |11i h10| +
|10i h11|.
The CNOT operator gives us XOR, but for more destructive operations
we need to use more qubits, possibly including junk qubits that we wont look
at again but that are necessary to preserve reversibility. The Toffoli gate or
controlled controlled NOT gate (CCNOT) is a 3-qubit gate that was
originally designed to show that classical computation could be performed
reversibly [Tof80]. It implements the mapping (x, y, z) 7 (x, y, (x y) z),
which corresponds to the 8 8 matrix

1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0

0 0 1 0 0 0 0 0

0 0 0 1 0 0 0 0
.
0 0 0 0 1 0 0 0

0 0 0 0 0 1 0 0

0 0 0 0 0 0 0 1

0 0 0 0 0 0 1 0
By throwing in some extra qubits we dont care about, Toffoli gates

can implement basic operations like NAND ((x, y, 1) 7 (x, y, (x y))),
NOT ((x, 1, 1) 7 (x, 1, x)), and fan-out ((x, 1, 0) 7 (x, 1, x)).5 This gives
a sufficient basis for implementing all classical circuits.
13.3.3 Representing Boolean functions

Suppose we have a quantum circuit that computes a Boolean function f .
There are two conventions for representing the output of f :
1. We can represent f (x) by XORing it with an extra qubit y: |x, yi 7

|x, y f (x)i. This is the approach taken by the CNOT (f (x) = x) and
CCNOT (f (x1 x2 ) = x1 x2 ) gates.
2. We can represent f (x) by changing the phase of |xi: |xi 7 (1)f (x) |xi.
The actual unitary operator corresponding to this is x (1)f (x) |xi hx|,
P
which in matrix form looks like a truth table for f expressed as 1

values along the diagonal, e.g.:

1 0 0 0
0 1 0 0

0 0 1 0

0 0 0 1
5
In the case of fan-out, this only works with perfect accuracy for classical bits and not
superpositions, which run into something called the no-cloning theorem. For example,
applying CCNOT to 12 |010i + 12 |110i yields 12 |010i + 12 |111i. This works, sort of,
but the problem is that the first and last bits are still entangled, meaning we cant operate
on them independently. This is actually not all that different from what happens in the
probabilistic case (if I make a copy of a random variable X, its correlated with the original
X), but it has good or bad consequences depending on whether you want to prevent people
from stealing your information undetected or run two copies of a quantum circuit on
independent replicas of a single superposition.
for the XOR function.

This has the advantage of requiring one fewer qubit, but we cant
observe the value of f (x) directly, because the amplitude of |xi is
unchanged. Where this comes in handy is when we can use the change in
phase in the context of a larger quantum algorithm to get cancellations
for some values of x.
The first representation makes more sense for modeling classical circuits.
The second turns out to be more useful when f is a subroutine in a larger
quantum algorithm. Fortunately, the operator with matrix
" #
1 0
0 1
converts the |0i|1i representation into the 1 representation, so we can

build any desired classical f using 01 logic and convert it to 1 as needed.
13.3.4 Practical issues (which we will ignore)

The big practical question is whether any of these operationsor even non-
trivial numbers of independently manipulable qubitscan be implemented
in a real, physical system. As theorists, we can ignore these issues, but in
real life they are what would make quantum computing science instead of
science fiction.6
13.3.5 Quantum computations

Putting everything together, a quantum computation consists of three stages:
1. We start with a collection of qubits in some known state x0 (e.g.,

|000 . . . 0i).
2. We apply a sequence of unitary operators A1 , A2 , . . . Am to our qubits.
3. We take a measurement of the final superposition Am Am1 . . . A1 x0

that collapses it into a single state, with probability equal to the square
of the amplitude of that state.
Our goal is for this final state to tell us what we want to know, with
reasonably high probability.
6
Current technology, sadly, still puts quantum computing mostly in the science fiction
category.
13.4 Deutschs algorithm

We now have enough machinery to describe a real quantum algorithm.
Known as Deutschs algorithm, this computes f (0) f (1) while evaluating
f once [Deu89]. The trick, of course, is that f is applied to a superposition.
Assumption: f is implemented reversibly, as a quantum computation
that maps |xi to (1)f (x) |xi. To compute f (0) f (1), evaluate
r
1
Hf H |0i = Hf (|0i + |1i)
r2
1
= H (1)f (0) |0i + (1)f (1) |1i
2
1
= (1)f (0) + (1)f (1) |0i + (1)f (0) (1)f (1) |1i .
2
Suppose now that f (0) = f (1) = b. Then the |1i terms cancel out and
we are left with
1
2 (1)b |0i = (1)b |0i .
2
This puts all the weight on |0i, so when we take our measurement at the
end, well see 0.
Alternatively, if f (0) = b 6= f (1), its the |0i terms that cancel out,
leaving (1)b |1i. Again the phase depends on b, but we dont care about the
phase: the important thing is that if we measure the qubit, we always see 1.
The result in either case is that with probability 1, we determine the
value of f (0) f (1), after evaluating f once (albeit on a superposition of
quantum states).
This is kind of a silly example, because the huge costs involved in building
our quantum computer probably swamp the factor-of-2 improvement we got
in the number of calls to f . But a generalization of this trick, known as the
Deutsch-Josza algorithm [DJ92], solves the much harder (although still a bit
contrived-looking) problem of distinguishing a constant Boolean function on
n bits from a function that outputs one for exactly half of its inputs. No
deterministic algorithm can solve this problem without computing at least
2n /2 + 1 values of f , giving an exponential speed-up.7
7
The speed-up compared to a randomized algorithm that works with probability 1
is less impressive. With randomization, we only need to look at O(log 1/) values of f to
see both a 0 and a 1 in the non-constant case. But even here, the Deutsch-Josza algorithm
does have the advantage of giving the correct answer always.
13.5 Grovers algorithm

Grovers algorithm [Gro96] is one of two main exemplars for the astonishing
power of quantum computers.8 The idea of Grovers algorithm is that if we
have a function f on N = 2n possible inputs whose value is 1 for exactly

one possible input w, we can find this w with high probability using O( N )
quantum evaluations of f . As with Deutschs algorithm, we assume that f
is encoded as an operator (conventionally written Uw ) that maps each |xi to
(1)f (x) |xi.
The basic outline of the algorithm:
q
1. Start in the superposition |si = 1
x |xi = H n |0i.
P
N
2. Alternate between applying the Grover diffusion operator D =

2 |si hs| I and the f operator Uw = 2 |wi hw| I. Do this O( n)
times (the exact number of iterations is important and will be calculated
below).
3. Take a measurement of the state. It will be w with high probability.
Making this work requires showing that (a) we can generate the original
superposition |si, (b) we can implement D efficiently using unitary operations
on a constant number of qubits each, and (c) we actually get w at the end
of this process.
13.5.1 Initial superposition

To get the initial superposition, start with |0qn i and apply the Hadamard
transform to each bit individually; this gives N1 x |xi as claimed.
P
13.5.2 The Grover diffusion operator

We have the definition D = 2 |si hs| I.
Before we try to implement this, lets start by checking that it is in fact
8
The other is Shors algorithm [Sho97], which allows a quantum computer to factor
n-bit integers in time polynomial in n. Sadly, Shors algorithm is a bit too complicated to
talk about here.
unitary. Compute
DD = (2 |si hs| I)2

= 4 |si hs| |si hs| 4 |si hs| + I 2
= 4 |si hs| 4 |si hs| + I
= I.
Here we use the fact that |si hs| |si hs| = |si hs|si hs| = |si (1) hs| = |si hs|.
Recall that |si = H n |0n i, where H n is the result of applying H to
each of the n bits individually. We also have that H = H and HH = I,
from which H n H n = I as well.
So we can expand
D = 2 |si hs| I
= 2H n |0n i (H n |0n i) I
= 2H n |0n i h0n | H n I
= H n (2 |0n i h0n | I) H n .
The two copies of H n involve applying H to each of the n bits, which

we can do. The operator in the middle, 2 |0n i h0n | I, maps |0n i to |0n i and
maps all other basis vectors |xi to |xi. This can be implemented as a NOR
of all the qubits, which we can do using our tools for classical computations.
So the entire operator D can be implemented using O(n) qubit operations,
most of which can be done in parallel.
13.5.3 Effect of the iteration

To see what happens when we apply Uw D, it helps to represent the state in
terms of a particular two-dimensional basis. The idea here is that the initial
state |si and the operation Uw D are symmetric with respect to any basis
vectors |xi that arent |wi, so instead of tracking all of these non-|wi vectors
separately, we will represent all of them by a single composite vector
s
1 X
|ui = |xi .
N 1 x6=w
q
The coefficient N 11 is chosen to make hu|ui = 1. As always, we like
our vectors to have length 1.
Using |ui, we can represent

r s
1 N 1
|si = |wi + |ui . (13.5.1)
N N
q
1
A straightforward calculation shows that this indeed puts N amplitude
on each |xi. q
Now were going to bring in some trigonometry. Let = sin1 N1 , so
q q
that sin = N1 and cos = 1 sin2 = NN1 . We can then rewrite
(13.5.1) as
|si = (sin ) |wi + (cos ) |ui . (13.5.2)
Lets look at what happens if we expand D using (13.5.2):
D = 2 |si hs| I
= 2 ((sin ) |wi + (cos ) |ui) ((sin ) hw| + (cos ) hu|) I
= (2 sin2 1) |wi hw| + (2 sin cos ) |wi hu| + (2 sin cos ) |ui hw| + (2 cos2 1) |ui hu|
= ( cos 2) |wi hw| + (sin 2) |wi hu| + (sin 2) |ui hw| + (cos 2) |ui hu|
" #
cos 2 sin 2
= ,
sin 2 cos 2
where the matrix is over the basis (|wi , |ui).

Multiplying by Uw negates all the |wi coefficients. So we get
" #" #
1 0 cos 2 sin 2
Uw D =
0 1 sin 2 cos 2
" #
cos 2 sin 2
= . (13.5.3)
sin 2 cos 2
Aficionados of computer graphics, robotics, or just matrix algebra in

general may recognize (13.5.3) as the matrix that rotates two-dimensional
vectors by 2. Since we started with |si at an angle of , after t applications
of this matrix we will be at an angle of (2t + 1), or in state
(sin(2t + 1)) |wi + (cos(2t + 1)) |ui .
Ideally, we pick t so that (2t + 1) = /2, which would put all of the
amplitude on |wi. Because t is an integer, we cant do this exactly, but
setting t = b /21
2 c will get us somewhere between /2 2 and /2. Since
q
1
N,this gives us a probability of seeing |wi in our final measurement
p
of 1 O( 1/N ) after O( N ) iterations of Uw D.
Sadly, this is as good as it gets. A lower bound of Bennet et al. [BBBV97]
shows that any quantum algorithm using Uw as the representation for f must
apply Uw at least ( N ) times to find w. So we get a quadratic speedup
but not the exponential speedup wed need to solve NP-complete problems
directly.
Appendix A
Assignments
Assignments are typically due Thursdays at 23:00.

Assignments should be submitted using the drop box mechanism in
classesv2. To use this, you will need to log into https://classesv2.yale.
edu/ using your NetId, then click on the Drop Box tool in the list of tools
on the left. You should then be shown a folder with your name on it, to
which you can upload files using the Add pulldown menu.
Assignments should be submitted in PDF format. How you get your
assignment into this format is up to you. I personally recommend learning
LATEX if you havent already, but exporting to PDF from your favorite word
processor or scanning in legibly hand-written solutions are also sensible
options. Assignments that are not submitted in PDF format may be treated
unjustly. I would also recommend giving your assignment an informative
filename to avoid confusion.
Questions about assignments can be sent to the instructor directly at
[email protected], or submitted to the course Piazza site at http:
//piazza.com/yale/fall2016/cpsc469.
A.1 Assignment 1: due Sunday, 2016-09-18, at

17:00
Bureaucratic part
Send me email! My address is [email protected].
In your message, include:
1. Your name.
237
APPENDIX A. ASSIGNMENTS 238
2. Your status: whether you are an undergraduate, grad student, auditor,

etc.
3. Whether you are taking the course as CPSC 469 or CPSC 569.
4. Anything else youd like to say.
(You will not be graded on the bureaucratic part, but you should do it
anyway.)
A.1.1 Bubble sort

Algorithm A.1 implements one pass of the (generally pretty slow) bubble
sort algorithm. This involves comparing each array location with its succes-
sor, and swapping the elements if they are out of order. The full bubble sort
algorithm repeats this loop until the array is sorted, but here we just do it
once.
1 procedure BubbleSortOnePass(A, n)
2 for i 1 to n 1 do
3 if A[i] > A[i + 1] then
4 Swap A[i] and A[i + 1]
Algorithm A.1: One pass of bubble sort
Suppose that A starts out as a uniform random permutation of distinct

elements. As a function of n, what is the exact expected number of swaps
performed by Algorithm A.1?
Solution
The answer is n Hn , where Hn = ni=1 1i is the n-th harmonic number.
P
There are a couple of ways to prove this. Below, we let Ai represent the
original contents of A[i], before doing any swaps. In each case, we will use the
fact that after i iterations of the loop, A[i] contains the largest of Ai , . . . , Ai ;
this is easily proved by induction on i.
Let Xij be the indicator variable for the event that Ai is eventually
swapped with Aj . For this to occur, Ai must be bigger than Aj , and
must be present in A[j 1] after j 1 passes through the loop. This hap-
pens if and only if Ai is the largest value in A1 , . . . , Aj . Because these
values are drawn from a uniform random permutation, by symmetry

Ai is largest with probability exactly 1/j. So E [Xij ] = 1/j.
Now sum Xij over all pairs i < j. It is easiest to do this by summing
over j first:

X X
E Xij = E [Xij ]
i<j i<j
n j1
X X
= E [Xij ]
j=2 i=1
n j1
X X1
=
j=2 i=1
j
n
X j1
=
j=2
j
n
X j1
=
j=1
j
n
X 1
= (1 )
j=1
j
n
X 1
=n
j=1
j
= n Hn .
Alternatively, lets count how many values are not swapped from A[i]
to A[i 1]. We can then subtract from n to get the number that are.
Let Yi be the indicator variable for the event that Ai is not swapped
into A[i 1]. This occurs if, when testing A[i 1] against A[i], A[i] is
larger. Since we know that at this point A[i 1] is the largest value
among A1 , . . . , Ai1 , Yi = 1 if and only if Ai is greater than all of
A1 , . . . , Ai1 , or equivalently if Ai is the largest value in A1 , . . . , Ai .
Again by symmetry we have E [Yi ] = 1i , and summing over all i gives
an expected Hn values that are not swapped down. So there are n Hn
values on average that are swapped down, which also gives the expected
number of swaps.
A.1.2 Finding seats

A huge lecture class with n students meets in a room with m n seats.
Because the room is dark, loud, and confusing, the students adopt the
following seat-finding algorithm:
1. The students enter the room one at a time. No student enters until
the previous student has found a seat.
2. To find a seat, a student chooses one of the m seats uniformly at random.

If the seat is occupied, the student again picks a seat uniformly at
random, continuing until they either find a seat or have made k attempts
that led them to an occupied seat.
3. If a student fails to find a seat in k attempts, they contact an Under-

graduate Seating Assistant, who leads them to an unoccupied seat.
Give an asymptotic (big-) expression for the number of students who

are helped by Undergraduate Seating Assistants, as a function of n, m, and
k, where k is a constant independent of n and m. As with any asymptotic
formula, this expression should work for large n and m; and should be as
simple as possible.1
Solution
Define the i-th student to be the student who takes their seat after i students
have already sat (note that we are counting from zero here, which makes
things a little easier). Let Xi be the indicator variable for the even that the
i-th student does not find a seat on their own and needs help from a USA.
Each attempt to find a seat fails with probability i/m. Since each attempt is
independent of the others, the probability that all k attempts fail is (i/m)k .
The number of students who need help is n1
P
i=0 Xi , so the expected
1
An earlier version of this problem allowed k to grow with n, which causes some trouble
with the asymptotics.
number is
"n1 # n1
X X
E Xi = E [Xi ]
i=0 i=0
n1
X
= (i/m)k
i=0
n1
k
X
=m ik .
i=0
Unfortunately there is no simple expression for ik that works for all

P
k. However, we can approximate it from above and below using integrals.

From below, we have
n1
X n1
X
k
i = ik
i=0 i=1
Z n1
xk dx
0
n1
1

= xk+1
k+1 0
1
= (n 1)k+1
k
+1
1

k+1
= n .
k+1
From above, an almost-identical computation gives
n1
X Z n
k
i xk dx
i=0 0
n
1

= xk+1
k+1 0
(
= f rac1k + 1n k + 1)
1
= O( nk+1 ).
k+1

1 k+1
Combining the two cases gives that the sum is k+1 n . We must
also include mk , giving the full answer
1

nk+1 mk .
k+1
k

I like writing this as n (n/m)
k+1 to emphasize that the cost per student
(n/m)k

is k+1 , but there are a lot of different ways to write the same thing.
Note that since we are dealing with asymptotics, we could also replace the
1 1
k+1 with just k and still be correct, but that seems less informative somehow.
A.2 Assignment 2: due Thursday, 2016-09-29, at

23:00
A.2.1 Technical analysis
It is well known that many human endeavorsincluding sports reporting,
financial analysis, and Dungeons and Dragonsinvolve building narra-
tives on top of the output of a weighted random number generator [Mun11],
a process that is now often automated [Hol16]. For this problem, we are
going to consider a simplified model of automated financial analysis.
Let us model the changes in a stock price over n days using a sequence of
independent 1 random variables X1 , X2 , . . . , Xn , where each Xi is 1 with
probability 1/2. Let Si = ij=1 Xj , where i {0, . . . , n}, be the price after i
P
days.
Our automated financial reporter will declare a dead cat bounce if a
stock falls in price for two days in a row, followed by rising in price, followed
by falling in price again. Formally, a dead cat bounce occurs on day i if
i 4 and Si4 > Si3 > Si2 < Si1 > Si . Let D be the number of dead
cat bounces over the n days.
1. What is E [D]?
2. What is Var [D]?
3. Our investors will shut down our system if it doesnt declare at least
one dead cat bounce during the first n days. What upper bound you
can get on Pr [D = 0] using Chebyshevs inequality?
4. Show Pr [D = 0] is in fact exponentially small in n.
Note added 2016-09-28: Its OK if your solutions only work for sufficiently
large n. This should save some time dealing with weird corner cases when n
is small.
Solution
1. Let Di be the indicator variable for the event that a dead cat bounce
occurs on day i. Let p = E [Di ] = Pr [Di = 1]. Then
1
p = Pr [Xi3 = 1 Xi2 = 1 Xi1=+1 Xi = 1] = ,
16
since the Xi are independent.
Then
" n #
X
E [D] = E Di
i=4
n
X
= E [Di ]
i=4
n
X 1
=
i=4
16
n3
= ,
16
assuming n is at least 3.
15
2. For variance, we cant just sum up Var [Di ] = p(1 p) = 256 , because
the Di are not independent. Instead, we have to look at covariance.
Each Di depends on getting a particular sequence of four values for
Xi3 , Xi2 , Xi1 , and Xi . If we consider how Di overlaps with Dj for
j > i, we get these cases:
Variable Pattern Correlation
Di --+-
Di+1 --+- inconsistent
Di+2 --+- inconsistent
Di+3 --+- overlap in one place
For larger j, there is no overlap, so Di and Dj are independent for
j 4, giving Cov [Di , Dj ] = 0 for these cases.
Because Di cant occur with Di+1 or Di+2 , when j {i + 1, i + 2} we
have
Cov [Di , Dj ] = E [Di Dj ] E [Di ] E [Dj ] = 0 (1/16)2 = 1/256.

When j = i + 3, the probability that both Di and Dj occur is 1/27 ,

since we only need to get 7 coin-flips right. This gives
Cov [Di , Di+3 ] = E [Di Di+3 ]E [Di ] E [Di+3 ] = 1/128(1/16)2 = 1/256.
Now apply (5.1.4):

" n #
X
Var [D] = Var Di
i=4
n
X n1
X n2
X n3
X
= Var [Di ] + 2 Cov [Di , Di+1 ] + 2 Cov [Di , Di+2 ] + 2 Cov [Di , Di+3 ]
i=4 i=4 i=4 i=4
15 1 1 1
= (n 3) + 2(n 4) + 2(n 5) + 2(n 6)
256 256 256 256
13n 39
= ,
256
where to avoid special cases, we assume n is at least 6.
3. At this point we are just doing algebra. For n 6, we have:
Pr [D = 0] Pr [|D E [D]| E [D]]

n3

= Pr |D E [D]|
16
(13n 39)/256

((n 3)/16)2
13n 39
=
(n 3)2
13
= .
n3
This is (1/n), which is not an especially strong bound.
4. Here is an upper bound that avoids the mess of computing the exact
probability, but is still exponentially small in n. Consider the events
[D4i = 1] for i {1, 2, . . . , bn/4c}. These events are independent
since they are functions of non-overlapping sequences of increments.
So we can compute Pr [D = 0] Pr [D4i = 0, i {1, . . . , bn/4c}] =
(15/16)bn/4c .
This expression is a little awkward, so if we want to get an asymptotic
estimate we can simplify it using 1 + x ex to get (15/16)bn/4c
exp((1/16)bn/4c) = O(exp(n/64)).
With a better analysis, we should be able to improve the constant in

the exponent; or, if we are real fanatics, calculate Pr [D = 0] exactly.
But this answer is good enough given what the problems asks for.
A.2.2 Faulty comparisons

Suppose we want to do a comparison-based sort of an array of n distinct
elements, but an adversary has sabotaged our comparison subroutine. Specif-
ically, the adversary chooses a sequence of times t1 , t2 , . . . , tn , such that for
each i, the ti -th comparison we do returns the wrong answer. Fortunately,
this can only happen n times. In addition, the adversary is oblivious: thought
it can examine the algorithm before choosing the input and the times ti , it
cannot predict any random choices made by the algorithm.
A simple deterministic work-around is to replace each comparison in our
favorite (n log n)-comparison sorting algorithm with 2n + 1 comparisons,
and take the majority value. But this gives us a (n2 log n)-comparison
algorithm. We would like to do better using randomization.
Show that in this model, for every fixed c > 0 and sufficiently large n, it
is possible to sort correctly with probability at least 1 nc using O(n log2 n)
comparisons.
Solution
Well adapt QuickSort. If the bad comparisons occurred randomly, we could
just take the majority of (log n) faulty comparisons to simulate a non-faulty
comparison with high probability. But the adversary is watching, so if we do
these comparisons at predictable times, it could hit all (log n) of them and
stay well within its fault budget. So we are going to have to be more sneaky.
Heres this idea: Suppose we have a list of n comparisons hx1 , y1 i , hx2 , y2 i , . . . , hxn , yn i
that we want to perform. We will use a subroutine that carries out these
comparisons with high probability by doing kn ln n+n possibly-faulty compar-
isons, with k a constant to be chosen below, where each of the possibly-faulty
comparisons looks at xr and yr where each r is chosen independently and uni-
formly at random. The subroutine collects the results of these comparisons
for each pair hxi , yi i and takes the majority value.
At most n of these comparisons are faulty, and we get an error only if
some pair hxi , yi i gets more faulty comparisons than non-faulty ones. Let Bi
be the number of bad comparisons of the pair hxi , yi i and Gi the number of
good comparisons. We want to bound the probability of the event Bi > Gi .
The probability that any particular comparison lands on a particular
pair is exactly 1/n; so E [Bi ] = 1 and E [Gi ] = k ln n. Now apply Cher-

noff bounds. From (5.2.4), Pr [Bi (k/2) ln n] 2(k/2) ln n = n(k ln 2)/2)
provided (k/2) ln n is at least 6. In the other direction, (5.2.6) says that
2
Pr [Gi (k/2) log n] e(1/2) k log n/2 = nk/8 . So we have a probability of
at most nk/2 + nk/8 that we get the wrong result for this particular pair,
and from the union bound we have a probability of at most n1k/2 + n1k/8
that we get the wrong result for any pair. We can simplify this a bit by
observe that n must be at least 2 (or we have nothing to sort), so we can put
a bound of n2k/8 on the probability of any error among our n simulated
comparisons, provided k 12/ ln 2. Well choose k = max(12/ ln 2, 8(c + 3))
to get a probability of error of at most nc1 .
So now lets implement QuickSort. The first round of QuickSort picks a
pivot and performs n 1 comparisons. We perform these comparisons using
our subroutine (note we can always throw in extra comparisons to bring
the total up to n). Now we have two piles on which to run the algorithm
recursively. Comparing all nodes in each pile to the pivot for that pile
requires n 3 comparisons, which can again be done by our subroutine.
At the next state, we have four piles, and we can again perform the n 7
comparisons we need using the subroutine. Continue until all piles have size
1 or less; this takes O(log n) rounds with high probability. Since each round
does O(n log n) comparisons and fails with probability at most nc1 , the
entire process takes O(n log2 n) comparisons and fails with probability at
most nc1 O(log n) nc when n is sufficiently large.

23:00
A.3.1 Painting with sprites
In the early days of computing, memory was scarce, and primitive home
game consoles like the Atari 2600 did not possess enough RAM to store
the image they displayed on your television. Instead, these machines used a
graphics chip that would composite together fixed sprites, bitmaps stored
in ROM that could be displayed at locations specified by the CPU.2 One
2
In the actual Atari 2600 graphics hardware, the CPU did this not by writing down the
locations in RAM somewhere (which would take memory), but by counting instruction
cycles since the graphics chip started refreshing the screen and screaming display this
sprite NOW! when the CRT beam got to the right position. For this problem we will
assume a more modern approach, where we can just give a list of locations to the graphics
hardware.
Figure A.1: Filling a screen with Space Invaders. The left-hand image places
four copies of a 46-pixel sprite in random positions on a 40 40 screen. The
right-hand image does the same thing with 24 copies.
common programming trick of this era was to repurpose existing data in

memory as sprites, for example by searching through a programs machine
code instructions to find sequences of bytes that looked like explosions when
displayed.
For this problem, we will imagine that we have a graphics device that can
only display one sprite. This is a bitmap consisting of m ones, at distinct rel-
ative positions hy1 , x1 i , hy2 , x2 i , . . . hym , xm i. Displaying a sprite at position
y, x on our nn screen sets the pixels at positions h(y + yi ) mod n, (x + xi ) mod ni
for all i {1, . . . , m}. The screen is initially blank (all pixels 0) and setting
a pixel at some position hy, xi changes it to 1. Setting the same pixel more
than once has no effect.
We would like to use these sprites to simulate white noise on the screen,
by placing them at independent uniform random locations with the goal
of setting roughly half of the pixels. Unfortunately, because the contents
of the screen are not actually stored anywhere, we cant detect when this
event occurs. Instead, we want to fix the number of sprites ` so that we
get 1/2 o(1) of the total number of pixels set to 1 with high probability,
by which we mean that (1/2 o(1))n2 total pixels are set with probability
1 nc for any fixed c > 0 and sufficiently large n, assuming m is fixed.
An example of this process, using Taito Corporations classic Space
Invader bitmap, is shown in Figure A.1.
Compute the value of ` that we need as a function of n and m, and show
that this choice of ` does in fact get 1/2 o(1) of the pixels set with high
probability.
Solution
We will apply the following strategy. First, well choose ` so that each
individual pixel is set with probability close to 1/2. Linearity of expectation
then gives roughly n2 /2 total pixels set on average. To show that we get
close to this number with high probability, well use the method of bounded
differences.
The probability that pasting in a single copy of the sprite sets a particular
pixel is exactly m/n2 . So the probability that the pixel is not set after ` sprites
ln(1/2)
is (1 m/n2 )` . This will be exactly 1/2 if ` = log1m/n2 1/2 = ln(1m/n 2) .
Since ` must be an integer, we can just round this quantity to the nearest
integer to pick our actual `. For example, when n = 40 and m = 46,
ln(1/2)
ln(146/402 )
= 23.7612... which rounds to 24.
For large n, we can use the approximation ln(1 + x) x (which holds for
small x) to approximate ` as (n2 ln 2/m) + o(1). We will use this to get a
concentration bound using (5.3.11).
We have ` independent random variables corresponding to the positions
of the ` sprites. Moving a single sprite changes the number of set pixels by
at most m. So the probability that the total number of set pixels deviates
from its expectation by more than an ln n is bounded by
2
an ln n
!
a2 n2 ln n
2 exp = 2 exp 2

`m2 (n ln 2/m + o(1))m2
!
a2 ln n
= 2 exp
m ln 2 + o(m2 /n2 )

exp (a2 /m) ln n
2 /m
= na ,

where the inequality holds when n is sufficiently large.
Now choose a = cm
to get aprobability of deviating by more than n cm ln n of at most nc .
Since n cm ln n = o(1)n2 , this give us our desired bound.
A.3.2 Dynamic load balancing

A low-cost competitor to better-managed data center companies offers a
scalable dynamic load balancing system that supplies a new machine whenever
a new job shows up, so that there are always n machines available to run
n jobs. Unfortunately, one of the ways the company cuts cost is by not
hiring programmers to replace the code from their previous randomized
load balancing mechanism, so when the i-th job arrives, it is still assigned
uniformly at random to one of the i available machines. This means that job
1 is always assigned to machine 1; job 2 is assigned with equal probability to
machine 1 or 2; job 3 is assigned with equal probability to machine 1, 2, or
3; and so on. These choices are all independent. The company claims that
the maximum load is still not too bad, and that the reduced programming
costs that they pass on to their customers make up for their poor quality
control and terrible service.
Justify the first part of this claim by showing that, with high probability,
the most loaded machine after n jobs has arrived has load O(log n).
Solution
Let Xij be the indicator variable for the event that machine i gets job j.
Then E [Xij ] = 1/j for all i j, and E [Xij ] = 0 when i > j.
Let Yi = nj=1 Xij be the load on machine i.
P
hP i
n Pn Pn
Then E [Yi ] = E j=1 Xij j=i 1/j
= j=1 1/j = Hn ln n + 1.
From (5.2.4), we have Pr [Yi R] 2 R as long as R > 2e E [Yi ]. So if
we let R = (c + 1) lg n, we get a probability of at most nc1 that we get
more than R jobs on machine i. Taking the union bound over all i gives
a probability of at most nc that any machine gets a load greater than
(c + 1) lg n. This works as long as (c + 1) lg n 2e(ln n + 1), which holds for
sufficiently large c. For smaller c, we can just choose a larger value c0 that
0
does work, and get that Pr [max Yi (c0 + 1) lg n] nc nc .
So for any fixed c, we get that with probability at least 1 nc the
maximum load is O(log n).

23:00
A.4.1 Re-rolling a random treap
Suppose that we add to a random treap (see 6.3) an operation that re-rolls
the priority of the node with a given key x. This replaces the old priority for
x with a new priority generated uniformly at random, independently of the
old priority any other priorities in the treap. We assume that the range of
priorities is large enough that the probability that this produces a duplicate
is negligible, and that the choice of which node to re-roll is done obliviously,
without regard to the current priorities of nodes or the resulting shape of
the tree.
Changing the priority of x may break the heap property. To fix the
heap property, we either rotate x up (if its priority now exceeds its parents
priority) or down (if its priority is now less than that of one or both of its
children). In the latter case, we rotate up the child with the higher priority.
We repeat this process until the heap property is restored.
Compute the best constant upper bound you can on the expected number
of rotations resulting from executing a re-roll operation.
Solution
The best possible bound is at most 2 rotations on average in the worst case.
There is an easy but incorrect argument that 2 is an upper bound, which
says that if we rotate up, we do at most the same number of rotations as
when we insert a new element, and if we rotate down, we do at most the same
number of rotations as when we delete an existing element. This gives the
right answer, but for the wrong reasons: the cost of deleting x, conditioned
on the event that re-rolling its priority gives a lower priority, is likely to be
greater than 2, since the conditioning means that x is likely to be higher up
in the tree than average; the same thing happens in the other direction when
x moves up. Fortunately, it turns out that the fact that we dont necessarily
rotate x all the way to or from the bottom compensates for this issue.
This can be formalized using the following argument, for which I am
indebted to Adit Singha and Stanislaw Swidinski. Fix some element x, and
suppose its old and new priorities are p and p0 . If p < p0 , we rotate up,
and the sequence of rotations is exactly the same as we get if we remove
all elements of the original treap with priority p or less and then insert a
new element x with priority p0 . But now if we condition on the number k of
elements with priority greater than p, their priorities together with p0 are all
independent and identically distributed, since they are all obtained by taking
their original distribution and conditioning on being greater than p. So all
(k + 1)! orderings of these priorities are equally likely, and this means that
we have the same expected cost as an insertion into a treap with k elements,
which is at most 2. Averaging over all k shows that the expected cost of
rotating up is at most 2, and, since rotating down is just the reverse of this
process with a reversed distribution on priorities (since we get it by choosing
p0 as our old priority and p as the new one), the expected cost of rotating
down is also at most 2. Finally, averaging the up and down cases gives that
the expected number of rotations without conditioning on anything is at
most 2.
We now give an exact analysis of the expected number of rotations, which
will show that 2 is in fact the best bound we can hope for.
The idea is to notice that whenever we do a rotation involving x, we
change the number of ancestors of x by exactly one. This will always be
a decrease in the number of ancestors if the priority of x went up, or an
increase if the priority of x went down, so the total number of rotations will
be equal to the change in the number of ancestors.
Letting Ai be the indicator for the event that i is an ancestor of x before
the re-roll, and A0i the indicator for the even that i is an ancestor of x after the
re-roll, then the number of rotations is just | i Ai i A0i |, which is equal
P P
to i |Ai A0i | since we know that all changes Ai A0i have the same sign. So
P
the expected number of rotations is just E [ i |Ai A0i |] = i E [|Ai A0i |],
P P
by linearity of expectation.
So we have to compute E [|Ai A0i |]. Using the same argument as in
6.3.2, we have that Ai = 1 if and only if i has the highest initial priority
of all elements in the range [min(i, x), max(i, x)], and the same holds for A0i
if we consider updated priorities. So we want to know the probability that
changing only the priority of x to a new random value changes whether i
has the highest priority.
Let k = max(i, x) min(i, x) + 1 be the number of elements in the
range under consideration. To avoid writing a lot of mins and maxes, lets
renumber these elements as 1 through k, with i = 1 and x = k (this may
involve flipping the sequence if i > x). Let X1 , . . . , Xk be the priorities of
these elements, and let Xk0 be the new priority of x. These k + 1 random
variables are independent and identically distributed, so conditioning on the
even that no two are equal, all (k + 1)! orderings of their values are equally
likely.
So now let us consider how many of these orderings result in |Ai A0i | = 1.
For Ai to be 1, X1 must exceed all of X2 , . . . , Xk . For A0i to be 0, X1 must
not exceed all of X2 , . . . , Xk1 , Xk0 . The intersection of these events is when
Xk0 > X1 > max(X2 , . . . , Xk ). Since X2 , . . . , Xk can be ordered in any of
(k 1)! ways, this gives
(k 1)! 1
Pr Ai = 1 A0i = 0 =

= .
(k + 1)! k(k + 1)
In the other direction, for Ai to be 0 and A0i to be 1, we must have Xk >

X1 > max(X1 , . . . , Xk1 , Xk ). This again gives

1
Pr Ai = 0 A0i = 1 =

k(k + 1)
and combining the two cases gives
2
E Ai A0i =

.
k(k + 1)
Now sum over all i:
n
E A0i Ai
X
E [number of rotations] =
i=1
x1 n
X 2 X 2
= +
i=1
(x i + 1)(x i + 2) i=x+1 (i x + 1)(i x + 2
x nx+1
X 2 X 2
= +
k=2
k(k + 1) k=2
k(k + 1)
x nx+1
2 2 2 2
X X
= +
k=2
k k+1 k=2
k k+1
2 2
=2 .
x+1 nx+2
In the limit as n goes to infinity, choosing x = bn/2c makes both fractional
terms converge to 0, making the expected number of rotations arbitrarily
close to 2. So 2 is the best possible bound that doesnt depend on n.
A.4.2 A defective hash table

A foolish programmer implements a hash table with m buckets, numbered
0, . . . , m 1, with the property that any value hashed to an even-numbered
bucket is stored in a linked list as usual, but any value hashed to an odd-
numbered bucket is lost.
1. Suppose we insert a set S of n = |S| items into this hash table, using
a hash function h chosen at random from a strongly 2-universal hash
family H. Show that there is a constant c such that, for any t > 0, the
probability that at least n/2 + t items are lost is at most cn/t2 .
2. Suppose instead that we insert a set S of n = |S| items into this hash
table, using a hash function h chosen at random from a hash family
H 0 that is 2-universal, but that may not be strongly 2-universal. Show

that if n m/2, it is possible for an adversary that knows S to design
a 2-universal hash family H 0 that ensures that all n items are lost with
probability 1.
Solution
1. Let Xi be the indicator variable for the event that the i-th element
of S is hashed to an odd-numbered bucket. Since H is strongly 2-
universal, E [Xi ] 1/2 (with equality when m is even), from which
it follows that Var [Xi ] = E [Xi ] (1 E [Xi ]) 1/4; and the Xi are
pairwise independent. Letting Y = ni=1 Xi be the total number of
P
items lost, we get E [Y ] n/2 and Var [Y ] n/4. But then we can
apply Chebyshevs inequality to show
Pr [Y n/2 + t] Pr [Y E [Y ] + t]
Var [Y ]

t2
n/4
2
t
1
= (n/t2 ).
4
So the desired bound holds with c = 1/4.
2. Write [m] for the set {0, . . . , m 1}. Let S = {x1 , x2 , . . . , xn }.

Consider the set H 0 of all functions h : U [m] with the property
that h(xi ) = 2i + 1 for each i. Choosing a function h uniformly from
this set corresponds to choosing a random function conditioned on this
constraint; and this conditioning guarantees that every element of S is
sent to an odd-numbered bucket and lost. This gives us half of what
we want.
The other half is that H 0 is 2-universal. Observe that for any y 6 S,
the conditioning does not constrain h(y), and so h(y) is equally likely
to be any element of [m]; in addition, h(y) is independent of h(z) for
any z 6= y. So for any y 6= z, Pr [h(y) = h(z)] = 1/m if one or both of
y and z is not an element of S, and Pr [h(y) = h(z)] = 0 if both are
elements of S. In either case, Pr [h(y) = h(z)] 1/m, and so H 0 is
2-universal.

23:00
A.5.1 A spectre is haunting Halloween
An adult in possession of n pieces of leftover Halloween candy is distributing
them to n children. The candies have known integer values v1 , v2 , . . . , vn ,
and the adult starts by permuting the candies according to a uniform random
permutation . They then give candy (1) to child 1, candy (2) to child 2,
and so on. Or at least, that is their plan.
Unbeknownst to the adult, child n is a very young Karl Marx, and after
observing the first k candies (1), . . . , (k), he can launch a communist
revolution and seize the means of distribution. This causes the remaining
n k candies (k + 1) through (n) to be distributed evenly among the
remaining n k children, so that each receives a value equal to exactly
1 Pn
nk i=k+1 v(i) . Karl may declare the revolution at any time before the last
candy is distributed (even at time 0, when no candies have been distributed).
However, his youthful understanding of the mechanisms of historical de-
terminism are not detailed enough to predict the future, so his decision to
declare a revolution after seeing k candies can only depend on those k candies
and his knowledge of the values of all of the candies but not their order.
Help young Karl optimize his expected take by devising an algorithm that
takes as input v1 , . . . , vn and v(1) , . . . , v(k) , where 0 k < n, and outputs
whether or not to launch a revolution at this time. Compute Karls exact
expected return as a function of v1 , . . . , vn when running your algorithm,
and show that no algorithm can do better.
A.5.2 Colliding robots on a line

Suppose we have k robots on a line of n positions, numbered 1 through
n. No two robots can pass each other or occupy the same position, so we
can specify the positions of all the robots at time t as an increasing vector
X t = X1t , X2t , . . . , Xkt . At each time t, we pick one of the robots i uniformly

at random, and also choose a direction d = 1 uniformly and independently

at random. Robot i moves to position X t+1 = Xit + d if (a) 0 Xit + d < n,
and (b) no other robot already occupies position Xit + d. If these conditions
do not hold, robot i stays put, and Xit+1 = Xit .
1. Given k and n, determine the stationary distribution of this process.
2. Show that the mixing time tmix for this process is polynomial in n.
Solution
1. The stationary distribution is a uniform distribution on all nk place-

ments of the robots. To prove this, observe that two vectors increasing
vectors x and y are adjacent if and only if there is some i such that
xj = yj for all j 6= i and xj + d = yj for some d {1, +1}. In this
1
case, the transition probability pxy is 2n , since there is a n1 chance that
1
we choose i and a 2 chance that we choose d. But this is the same
as the probability that starting from y we choose i and d. So we
have pxy = pyx for all adjacent x and y, which means that a uniform
distribution satisfies x pxy = y pyx for all x and y.
To show that this stationary distribution is unique, we must show that
there is at least one path between any two states x and y. One way to
do this is to show that there is a path from any state x to the state
h1, . . . , ki, where at each step we move the lowest-index robot i that
is not already at position i. Since we can reverse this process to get
to y, this gives a path between any x and y that occurs with nonzero
probability.
2. This one could be done in a lot of ways. Below Ill give sketches of
three possible approaches, ordered by increasing difficulty. The first
reduces to card shuffling by adjacent swaps, the second uses an explicit
coupling, and the third uses conductance.
Of these approaches, I am personally only confident of the coupling
argument, since its the one I did before handing out the problem, and
indeed this is the only one I have written up in enough detail below to
be even remotely convincing. But the reduction to card shuffling is also
pretty straightforward and was used in several student solutions, so I
am convinced that it can be made to work as well. The conductance
idea I am not sure works at all, but it seems like it could be made to
work with enough effort.
(a) Lets start with the easy method. Suppose that instead of colliding
robots, we have a deck of n cards, of which k are specially marked.
Now run a shuffling algorithm that swaps adjacent cards at each
step. If we place a robot at the position of each marked card,
the trajectories of the robots follow the pretty much the same
distribution as in the colliding-robots process. This is trivially
the case when we swap a marked card with an unmarked card (a
robot moves), but it also works when we swap two marked cards
(no robot moves, since the positions of the set of marked cards
stays the same; this corresponds to a robot being stuck).
Unfortunately we cant use exactly the same process we used in
9.3.3.3, because this (a) allows swapping the cards in the first
and last positions of the deck, and (b) doesnt include any moves
corresponding to a robot at position 1 or n trying to move off the
end of the line.
The first objection is easily dealt with, and indeed the cited result
of Wilson [Wil04] doesnt allow such swaps either. The second can
be dealt with by adding extra no-op moves to the card-shuffling
process that occur with probability 1/n, scaling the probabilities
of the other operations to keep the sum to 1. This doesnt affect
the card-shuffling convergence argument much, but it is probably
a good idea to check that everything still works.
Finally, even after fixing the card-shuffling argument, we still have
to argue that convergence in the card-shuffling process implies
convergence in the corresponding colliding-robot process. Here is
where the definition of total variation distance helps. Let C t be the
permutation of the cards after t steps, and let f : C t 7 X t map
permutations of cards to positions of robots. Let and 0 be the
stationary distributions of the card and robot processes, respec-
t 0
t
tively. Then d (f (C ), f ()) = max A Pr X A (A)
=
t T V1

1 t

maxA Pr C f (A) (f (A)) maxB Pr C B (B) =

dT V (C t , ). So convergence of C implies convergence of X = f (C).
(b) Alternatively, we can construct a coupling between two copies of
the process X t and Y t , where as usual we start X 0 in our initial
distribution (whatever it is) and Y 0 in the uniform stationary
distribution. At each step we generate a pair (i, d) uniformly at
random from {1, . . . , n} {1, +1}. Having generated this pair,
we move robot i in direction d in both processes if possible. It is
0 0
clear that once X t = Y t , it will continue to hold that X t = Y t
for all t0 > t.
To show convergence we must show that X t will in fact eventually
hit Y t . To do so, let us track Z t = kX t Y t k1 = ki=1 |Xit Yit |.
P
We will show that Zt is a supermartingale with respect to the

history Ft generated by the random variables hX s , Y s i for s t.
Specifically, we will show that at each step, the expected change
in Z t conditioning on X t and Y t is non-positive.
Fix X t , and consider all 2n possible moves (i, d). Since (i, d)
doesnt change Xjt for any j 6= i, the only change in Z t will occur
if one of Xit and Yit can move and the other cant. There are
several cases:
i. If i = k, d = +1, and exactly one of Xkt and Ykt is n, then the
copy of the robot not at n moves toward the copy at n, giving
Zt+1 Zt = 1. The same thing occurs if i = 1, d = 1, and
exactly one of X1t and Y1t is 1.
It is interesting to note that these two cases will account for
the entire nonzero part of E [Zt+1 Zt | Ft ], although we will
not use this fact.
ii. If i < k, d = +1, Xit + 1 = Xi+1 t , but Y t + 1 < X t , and
i i+1
Xi Yi , then robot i can move right in Y t but not in X t .
t t
This gives Z t+1 Z t = +1.

However, in this case the move (i+1, 1) moves robot i+1 left
t > Y t + 1 > X t = X t+1 ,
in Y t but not in X t , and since Yi+1 i i i
this move gives Z t+1 t
Z = 1. Since the moves (i, +1) and
(i + 1, 1) occur with equal probability, these two changes in
Z t cancel out on average.
Essentially the same analysis applies if we swap X t with Y t ,
and if we consider moves (i, 1) where i > 1 in which one copy
of the robot is blocked but not the other. If we enumerate all
such cases, we find that for every move of these types that
increase Z t , there is a corresponding move that decreases Z t ;
it follows that these changes sum to zero in expectation.
Because each Xit and Yit lies in the range [1, n], it holds trivially
that Z t k(n 1) < n2 . Consider an unbiased 1 random walk
W t with a reflecting barrier at n2 (meaning that at n2 we always
move down) and an absorbing barrier at 0. Then the expected
time to reach 0 starting from W t = x is given by x(2n2 x) n4 .
Since this unbiased random walk dominates the biased random
walk corresponding to changes in Z t , this implies that Z t can
change at most n4 times on average before reaching 0.
Now let us ask how often Z t changes. The analysis above implies
that there is a chance of at least 1/n that Z t changes in any state
where some robot i with Xit 6= Yit is blocked from moving freely
in one or both directions in either the X t or Y t process. To show
that this occurs after polynomially many steps, choose some i with
Xit 6= Yit that is not blocked in either process. Then i is selected
every n1 steps on average, and it moves according to a 1 random
walk as long as it is not blocked. But a 1 random walk will

reach position 1 or n in O(n2 ) steps on average (corresponding to
O(n3 ) steps of the original Markov chain), and i will be blocked
by the left or right end of the line if it is not blocked before
then. It follows that we reach a state in which Z t changes with
probability at least n1 after O(n3 ) expected steps, which means
that Z t changes every O(n4 ) on average. Since we have previously
established that Z t reaches 0 after O(n4 ) expected changes, this
gives a total of O(n8 ) expected steps of the Markov process before
Z t = 0, giving X t = Y t . Markovs inequality and the Coupling
Lemma then give tmix = O(n8 ).
(c) I believe it may also be possible to show convergence using a con-
ductance argument. The idea is to represent a state hx1 , . . . , xk i
as the differences h1 , . . . , n i, where 1 = x1 and i+1 =
ni+1 xi . This
x representation
o makes the set of possible states
i 1, ki=1 i n look like a simplex, a pretty well-
P
behaved geometric object. In principle it should be possible

to show an isoperimetric inequality that the lowest-conductance
sets in this simplex are the N points closest to any given corner, or
possibly get a lower bound on conductance using canonical paths.
But the version of the process is messy (two coordinates change
every time a robot moves), and there are some complication with
this not being a lazy random walk, so I didnt pursue this myself.

23:00
A.6.1 Another colliding robot
A warehouse consists of an n n grid of locations, which we can think of
as indexed by pairs (i, j) where 1 i, j n. At night, the warehouse is
patrolled by a robot executing a lazy random walk. Unfortunately for the
robot, there are also m crates scattered about the warehouse, and if the
robot attempts to walk onto one of the m grid locations occupied by a crate,
it fails to do so, and instead emits a loud screeching noise. We would like to
use the noises coming from inside the warehouse to estimate m.
Formally, if the robot is at position (i, j) at time t, then at time t + 1
it either (a) stays at (i, j) with probability 1/2; or (b) chooses a position
(i0 , j 0 ) uniformly at random from {(i 1, j), (i + 1, j), (i, j 1), (i, j + 1)}
and attempts to move to it. If the robot chooses not to move, it makes no
noise. If it chooses to move, but (i0 , j 0 ) is off the grid or occupied by one of
the m crates, then the robot stays at (i, j) and emits the noise. Otherwise it
moves to (i0 , j 0 ) and remains silent. The robots position at time 0 can be
any unoccupied location on the grid.
To keep the robot from getting walled in somewhere, whatever adversary
placed the crates was kind enough to ensure that if a crate was placed at
position (i, j), then all of the eight positions
(i 1, j + 1) (i, j + 1) (i + 1, j + 1)
(i 1, j) (i + 1, j)
(i 1, j 1) (i, j 1) (i + 1, j 1)
reachable by a kings move from (i, j) are unoccupied grid locations. This
means that they are not off the grid and not occupied by another crate, so
that the robot can move to any of these eight positions.
Your job is to devise an algorithm for estimating m to within relative
error with probability at least 1 , based on the noises emitted by the robot.
The input to your algorithm is the sequence of bits x1 , x2 , x3 , . . . , where xi
is 1 if the robot makes a noise on its i-th step. Your algorithm should run in
time polynomial in n, 1/, and log(1/).
Solution
It turns out that is a bit of a red herring: we can in fact compute the exact
number of crates with probability 1 in time polynomial in n and log(1/).
The placement restrictions and laziness make this an irreducible aperiodic
chain, so it has a unique stationary distribution . It is easy to argue from
reversibility that this is uniform, so each of the N = n2 m unoccupied
positions occurs with probability exactly 1/N .
It will be useful to observe that we can assign three unique unoccupied
positions to each crate, to the east, south, and southeast, and this implies
m n2 /4.
The idea now is to run the robot until the distribution on its position is
close to , and then see if it hits an obstacle on the next step. We can easily
count the number of possible transitions that hit an obstacle, since there are
4m incoming edges to the crates, plus 4n incoming edges to the walls. Since
1
each edge uv has probability u puv = 8N of being selected in the stationary
distribution, the probability q that we hit an obstacle starting from is
exactly n+m n+m
2N = n2 m . This function is not trivial to invert, but we dont
have to invert it: if we can compute its value (to some reasonable precision),
we can compare it to all n2 /4 + 1 possible values of m, and see which it

matches. But we still have to deal with both getting enough samples and
the difference between our actual distribution when we take a sample and
the stationary distribution. For the latter, we need a convergence bound.
We will get this bound by constructing a family of canonical paths
between each distinct pair of locations (i, j) and (i0 , j 0 ). In the absence of
crates, we can use the L-shaped paths that first change i to i0 and then j to
j 0 , one step at a time in the obvious way. If we encounter a crate at some
position along this path, we replace that position with a detour that uses
up to three locations off the original path to get around the obstacle. The
general rule is that if we are traveling left-to-right or right-to-left, we shift
up one row to move past the obstacle, and if we are moving up-to-down or
down-to-up, we shift left one column. An annoying special case is when the
obstacle appears exactly at the corner (i0 , j); here we replace the blocked
position with the unique diagonally adjacent position that still gives us a
path.
Consider the number of paths (i, j) (i0 , j 0 ) crossing a particular left-
right edge (x, y) (x + 1, y). There are two cases we have to consider:
1. We cross the edge as part of the left-to-right portion of the path (this
includes left-to-right moves that are part of detours). In this case we
have |j y| 1. This gives at most 3 choices for j, giving at most 3n3
possible paths. (The constants can be improved here.)
2. We cross the edge as part of a detour on the vertical portion of the

path. Then |i x| 1, and so we again have at most 3n3 possible
paths.
This gives at most 6n3 possible paths across each edge, giving a congestion
1
(6n3 )ij i0 j 0
ij pij,i0 j 0
1
= 1 (6n3 )N 2
N (1/8)
= 24n3 N 1
3
24n3 ( n2 )1
4
= 32n,
since we can argue from the placement restrictions that N n2 /4. This
immediately gives a bound 2 82 O(n2 ). using Lemma 9.4.3.
So lets run for d512 ln ne = O(n2 ln n) steps for each sample. Starting
from any initial location, we will reach some distribution with dT V , =
O(n50 ). Let X be the number of obstacles (walls or crates) adjacent to the
current position, then we can apply Lemma 9.2.2 to get |E (X) E (X)|
4dT V (, ) = O(n50 ). The same bound (up to constants) also applies to
the probability = X/8 of hitting an obstacle, giving = 4(n + m)/(n2
m) O(n50 ). Note that is (n1 for all values of m.
Now take n10 ln(1/) samples, with a gap of d512 ln ne steps between
p
The expected number of positive samples is = (+O(n50 ))n10 ln(1/) =

p
each. p
(n9 ln(1/)), and from Lemma 5.2.2, the probability that the num-
ber of positive samples exceeds (1 + n4 ) is at most exp(n8 /3) =
exp((n ln(1/))) = (n) /2 for sufficiently large n. A similar bound
holds on the other side. So with probability at least 1 we get an estimate
of = 4(n + m)/(n2 m) that is accurate to within a relative error of n4 .
d
Since dm = O(n2 ) throughout the interval spanned by m, and since itself
is (n ), any relative error that is o(n1 ) will give us an exact answer for
1
m when we round to the nearest value of . We are done.

Weve chosen big exponents because all we care about is getting a polyno-
mial bound. With a cleaner analysis we could probably get better exponents.
It is also worth observing that the problem is ambiguous about whether we
have random access to the data stream, and whether the time taken by the
robot to move around counts toward our time bound. If we are processing
the stream after the fact, and are allow to skip to specific places in the stream
at no cost, we can get each sample in O(1) time. This may further reduce
the exponents, and demonstrates why it is important to nail down all the
details of a model.
A.6.2 Return of the sprites

The Space Invaders from Problem A.3.1 are back, and now they are trying
to hide in random n n bitmaps, where each pixel is set with probability
1/2. Figure A.2 shows an example of this, with two sprites embedded in an
otherwise random bitmap.
The sprites would like your assistance in making their hiding places
convincing. Obviously its impossible to disguise their presence completely,
but to make the camouflage as realistic as possible, they would like you to
provide them with an algorithm that will generate an n n bitmap uniformly
at random, conditioned on containing at least two copies of a given m-pixel
sprite at distinct offsets. (It is OK for the sprites to overlap, but they cannot
be directly on top of each other, and sprites that extend beyond the edges of
Figure A.2: Two hidden Space Invaders. On the left, the Space Invaders hide
behind random pixels. On the right, their positions are revealed by turning
the other pixels gray.
the bitmap wrap around as in Problem A.3.1.) Your algorithm should make
the probability of each bitmap that contains at least two copies of the sprite
be exactly equally likely, and should run in expected time polynomial in n
and m.
Solution
Rejection sampling doesnt work here because if the sprites are large, the
chances of getting two sprites out of a random bitmap are exponentially
small. Generating a random bitmap and slapping two sprites on top of it
also doesnt work, because it gives a non-uniform distribution: if the sprites
2
overlap in k places, there are 2n 2m+k choices for the remaining bits, which
means that we would have to adjust the probabilities to account for the effect
of the overlap. But even if we do this, we still have issues with bitmaps that
contain more than two sprites: a bitmap with three sprites can be generated
in three different ways, and it gets worse quickly as we generate more. It
may be possible to work around these issues, but a simpler approach is to
use the sampling mechanism from Karp-Luby [KL85] (see also 10.3).
2
Order all n2 positions u < v for the two planted sprites in lexicographic
order. For each pair of positions u < v, let Auv be the set of all bitmaps with
2
sprites at u and v. Then we can easily calculate |Auv | = 2n 2m+k where k
is the number of positions where sprites at u and v overlap, and so we can
sample a particular Auv with probability |Auv |/ st |Ast | and then choose an
P
element of Auv uniformly at random by filling in the positions not covered

by the sprites with independent random bits. To avoid overcounting, we
discard any bitmap that contains a sprite at a position less than max(u, v);
this corresponds to discarding any bitmap that appears in Ast for some
st < uv. By the same argument as in Karp-Luby, we expect to discard at
2
most n2 bitmaps before we get one that works. This gives an expected O(n4 )
attempts, each of which requires about O(n2 ) work to generate a bitmap
and O(n2 m) work to test it, for a nicely polynomial O(n6 m) expected time
in total.
If we dont want to go through all this analysis, we could also reduce
to Karp-Luby directly by encoding the existence of two sprites as a large
2
DNF formula consisting of n2 clauses, and then argue that the sampling
mechanism inside Karp-Luby gives us what we want.
A.7 Final exam

Write your answers in the blue book(s). Justify your answers. Work alone.
Do not use any notes or books.
There are three problems on this exam, each worth 20 points, for a total
of 60 points. You have approximately three hours to complete this exam.
A.7.1 Virus eradication (20 points)

Viruses have infected a network structured as an undirected graph G = (V, E).
At time 0, all n = |V | nodes are infected. At each step, as long as there is at
least one infected node, we can choose a node v to disinfect; this causes v
1
to become uninfected, but there is an independent 2d(v) chance that each of
vs neighbors becomes infected if it is not infected already, where d(v) is the
degree of v.
Let Xt be the number of nodes that are infected after t disinfection steps.
Let be the first time at which X = 0. Show that, no matter what strategy
is used to select nodes to disinfect, E [ ] = O(n).
Clarification added during exam: When choosing a node to disinfect, the
algorithm can only choose nodes that are currently infected.
Solution
Let Yt = Xt + 12 min(t, ). We will show that {Yt } is a supermartingale.
Suppose we disinfect node v at time t, and that v has d0 d(v) uninfected
d0
neighbors. Then E [Xt+1 | Xt ] = Xt 1 + 2d(v) Xt 12 , because we always
disinfect v and each of its d0 uninfected neighbors contributes 2d(v) 1

new
infections on average. We also have min(, t) increasing by 1 in this case.
This gives E [Yt+1 | Yt ] Yt 21 + 12 = Yt when t < ; it also holds trivially
that Yt+1 = Yt when t . So E [Yt+1 | Yt ] Yt always.
Now apply the optional stopping theorem. We have bounded incre-
ments, because |Yt+1 Yt | n + 12 . We also have bounded expected
time, because there is a nonzero probability that each sequence of n
consecutive disinfections produces no new infections, giving E [ ] 1/. So
n = Y0 E [Y ] = 12 E [ ], giving E [ ] 2n = O(n).
A.7.2 Parallel bubblesort (20 points)

Suppose we are given a random bit-vector and choose to sort it using parallel
bubblesort. The initial state consists of n independent fair random bits
X10 , X20 , . . . , Xn0 , and in each round, for each pair of positions Xit and Xi+1
t ,
t+1 t+1 t t
we set new values Xi = 0 and Xi+1 = 1 if Xi = 1 and Xi+1 = 0. When it
holds that Xit = 1 and Xi+1 t = 0, we say that there is a swap at position i in
round t. Denote the total number of swaps in round t by Yt .
1. What is E [Y0 ]?
2. What is E [Y1 ]?
3. Show that both Y0 and Y1 are concentrated around their expectations,

by showing, for any fixed c > 0 and each t {0, 1}, that there exists a
function ft (n) = o(n) such that Pr [|Yt E [Yt ]| ft (n)] nc .
Solution
1. This is a straightforward application of linearity of expectation. Let Zit ,
for each i {1, . . . , n 1}, be the indicator for the event that Xit = 1
t
and Xi+1 = 0. For t = 0, Xi0 and Xi+1 0 are independent, so this event
h Pn1 i Pn1
occurs with probability 14 . So E [Y0 ] = E Zi0 = E Zi0 =

i=1 i=1
n1
(n 1) 14 = 4 .
2. Again we want to use linearity of expectation, but we have to do a little

more work to calculate E Zi1 . For Zi1 to be 1, we need Xi1 = 1 and

1
Xi+1 = 0. If we ask where that 1 in Xi1 came from, either it started in
0
Xi1 and moved to position i because Xi0 = 0, or it started in Xi0 and
didnt move because Xi+10 1
= 1. Similarly the 0 in Xi+1 either started
0 0
in Xi+2 and moved down to replace Xi+1 = 1 or started in Xi+1 1 and
didnt move because Xi0 = 0. Enumerating all the possible assignments

0 , X 0 , X 0 , and X 0
of Xi1 i i+1 i+2 consistent with these conditions gives
0110, 1000, 1001, 1010, and 1110, making E Zi1 = 5 24 = 16 5

when
2 i n 2. This accounts for n 3 of the positions.
There are two annoying special cases: when i = 1 and when i = n 1.
We can handle these with the above analysis by pretending that X0t = 0
t
and Xn+1 = 1 for all t; this leaves the initial patterns 0110 for i = 1
and 1001 for i = n 1, each of which occurs with probability 81 .
1 Pn3 5 1 5n11
Summing over all cases then gives E [Y1 ] = 8 + i=2 16 + 8 = 16 .
3. The last part is just McDiarmids inequality. Because the algorithm

is deterministic aside from the initial random choice of input, We
can express each Yt as a function gt (X10 , . . . , Xn0 ) of the independent
random variables X10 , . . . , Xn0 . Changing some Xi0 can change Zi1 0 and
0
Zi , for a total change to Y0 of at most 2; similarly, changing Xi can 0
change at most Zi21 , Z 1 , Z 1 , and Z i , for a total change of at most

i1 i i+1
4. So McDiarmid says that, for each t {0, 1}, Pr [|Yt E [Yt ]| s]
2
2 exp (s /n). For any fixed c > 0, there is some s = ( n log n) =
o(n) such that the right-hand side is less than nc .
A.7.3 Rolling a die (20 points)

Suppose we have a device that generates a sequence of independent fair
random coin-flips, but what we want is a six-sided die that generates the
values 1, 2, 3, 4, 5, 6 with equal probability.
1. Give an algorithm that does so using O(1) coin-flips on average.
2. Show that any correct algorithm for this problem will use more than n
coin-flips with probability at least 2O(n) .
Solution
1. Use rejection sampling: generate 3 bits, and if the resulting binary
number is not in the range 1, . . . , 6, try again. Each attempt consumes
3 bits and succeeds with probability 3/4, so we need to generate 43 3 = 4
bits on average.
It is possible to improve on this by reusing the last bit of a discarded
triple of bits as the first bit in the next triple. This requires a more
complicated argument to show uniformity, but requires only two bits
for each attempt after the first, for a total of 3 + 14 43 2 = 11

3 bits on
1
average. This is still O(1), so unless that 3 bit improvement is really
important, its probably easiest just to do rejection sampling.
2. Suppose that we have generated n bits so far. Since 6 does not evenly
divide 2n for any n, we cannot assign an output from 1, . . . , 6 to all
possible 2n sequences of bits without giving two outputs different
probabilities. So we must keep going in at least one case, giving a
probability of at least 2n = 2O(n) that we continue.
Appendix B
Sample assignments from

Spring 2014
Assignments are typically due Wednesdays at 17:00. Assignments should be

submitted in PDF format via the classesv2 Drop Box.
B.1 Assignment 1: due Wednesday, 2014-09-10, at

17:00
B.1.1 Bureaucratic part
1. Your name.

etc.
anyway.)
B.1.2 Two terrible data structures

Consider the following data structures for maintaining a sorted list:
267
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 268
1. A sorted array A. Inserting a new element at position A[i] into an

array that current contains k elements requires moving the previous
values in A[i] . . . A[k] to A[i + 1] . . . A[k + 1]. Each element moved costs
one unit.
For example, if the array currently contains 1 3 5, and we insert 2,
the resulting array will contain 1 2 3 5 and the cost of the operation
will be 2, because we had to move 3 and 5.
2. A sorted doubly-linked list. Inserting a new element at a new position

requires moving the head pointer from its previous position to a neigh-
bor of the new node, and then to the new node; each pointer move
costs one unit.
For example, if the linked list currently contains 1 3 5, with the head
pointer pointing to 5, and we insert 2, the resulting linked list will
contain 1 2 3 5, with the head pointing to 2, and the cost will be 2,
because we had to move the pointer from 5 to 3 and then from 3 to 2.
Note that we do not charge for updating the pointers between the new
element and its neighbors. We will also assume that inserting the first
element is free.
Suppose that we insert the elements 1 through n, in random order, into

both data structures. Give an exact closed-form expression for the expected
cost of each sequence of insertions.
Solution
1. Suppose we have already inserted k elements. Then the next element
is equally likely to land in any of the positions A[1] through A[k + 1].
The number of displaced elements is then uniformly distributed in 0
through k, giving an expected cost for this insertion of k2 .
Summing over all insertions gives
n1
X k 1 n1
X
=
k=0
2 2 k=0
n(n 1)
= .
4
An alternative proof, which also uses linearity of expectation, is to

define Xij as the indicator variable for the event that element j moves
when element i < j is inserted. This is 1 if an only if j is inserted
before i, which by symmetry occurs with probability exactly 1/2. So

the expected total number of moves is
X X 1
E [Xij ] =
1i<jn 1i<jn
2
!
1 n
=
2 2
n(n 1)
= .
4
Its always nice when we get the same answer in situations like this.
2. Now we need to count how far the pointer moves between any two
consecutive elements. Suppose that we have already inserted k 1 > 0
elements, and let Xk be the cost of inserting the k-th element. Let
i and j be the indexes in the sorted list of the new and old pointer
positions after the k-th insertion. By symmetry, all pairs of distinct
positions i 6= j are equally likely. So we have

1 X
E [Xk ] = |i j|
k(k 1) i6=j

1 X X
= (j i) + (i j)
k(k 1) 1i<jk 1j<ik
2 X
= (j i)
k(k 1) 1i<jk
k X j1
2 X
= `
k(k 1) j=1 `=1
k
2 X j(j 1)
=
k(k 1) j=1 2
k
1 X
= j(j 1).
k(k 1) j=1
k
1 X
= j2 j
k(k 1) j=1
1 (2k + 1)k(k + 1) k

= 2
k(k 1) 6 k+1
1 (2k 2)k(k + 1)
=
k(k 1) 6
k+1
= .
3
This is such a simple result that we might reasonably expect that there
is a faster way to get it, and wed be right. A standard trick is to
observe that we can simulate choosing k points uniformly at random
from a line of n points by instead choosing k + 1 points uniformly at
random from a cycle of n + 1 points, and deleting the first point chosen
to turn the cycle back into a line. In the cycle, symmetry implies that
the expected distance between each point and its successor is the same
as for any other point; there are k + 1 such distances, and they add up
to n + 1, so each expected distance is exactly n+1
k+1 .
In our particular case, n (in the formula) is k and k (in the formula) is
2, so we get k+1
3 . Note we are sweeping the whole absolute value thing
under the carpet here, so maybe the more explicit derivation is safer.
However we arrive at E [Xk ] = k+13 (for k > 1), we can sum these
expectations to get our total expected cost:
" n # n
X X
E Xk = E [Xk ]
k=2 k=2
n
X k+1
=
k=2
3
1 n+1
X
= `
3 `=3
1 (n + 1)(n + 2)

= 3
3 2
(n + 1)(n + 2)
= 1.
6
Its probably worth checking a few small cases to see that this answer
actually makes sense.
For large n, this shows that the doubly-linked list wins, but not by much:
we get roughly n2 /6 instead of n2 /4. This is a small enough difference that
in practice it is probably dominated by other constant-factor differences that
we have neglected.
B.1.3 Parallel graph coloring

Consider the following algorithm for assigning one of k colors to each node
in a graph with m edges:
1. Assign each vertex u a color cu , chosen uniformly at random from all
k possible colors.
2. For each vertex u, if u has a neighbor v with cu = cv , assign u a new
color c0u , again chosen uniformly at random from all k possible colors.
Otherwise, let c0u = cu .
Note that any new colors c0 do not affect the test cu = cv . A node changes
its color only if it has the same original color as the original color of one or
more of its neighbors.
Suppose that we run this algorithm on an r-regular1 triangle-free2 graph.
As a function of k, r, and m, give an exact closed-form expression for the
1
Every vertex has exactly r neighbors.
2
There are no vertices u, v, and w such that all are neighbors of each other.
expected number of monochromatic3 edges after running both steps of the

algorithm.
Solution
We can use linearity of expectation to compute the probability that any
particular edge is monochromatic, and then multiply by m to get the total.
Fix some edge uv. If either of u or v is recolored in step 2, then the
probability that c0u = c0v is exactly 1/k. If neither is recolored, the probability
that c0u = c0v is zero (otherwise cu = cv , forcing both to be recolored). So we
can calculate the probability that c0u = c0v by conditioning on the event A
that neither vertex is recolored.
This event occurs if both u and v have no neighbors with the same color.
The probability that cu = cv is 1/k. The probability that any particular
neighbor w of u has cw = cu is also 1/k; similarly for any neighbor w of
v. These events are all independent on the assumption that the graph is
triangle-free (which implies that no neighbor of u is also a neighbor of v).
So the probability that none of these 2r 1 events occur is (1 1/k)2r1 .
We then have
h i h i
Pr [cu = cv ] = Pr cu = cv A Pr A + Pr [cu = cv | A] Pr [A]

2r1 !
1 1

= 1 1 .
k k
Multiply by m to get
2r1 !
m 1

1 1 .
k k

For large k, this is approximately m (2r1)/k , which is a little
k 1e
bit better than than the m k expected monochrome edges from just running
step 1.
Repeated application of step 2 may give better results, particular if k is
large relative to r. We will see this technique applied to a more general class
of problems in 11.3.5.
3
Both endpoints have the same color.

17:00
B.2.1 Load balancing
Suppose we distribute n processes independently and uniformly at random
among m machines, and pay a communication cost of 1 for each pair of
processes assigned to different machines. Let C be the total communication
cost, that is, the number of pairs of processes assigned to different machines.
What is the expectation and variance of C?
Solution
Let Cij be the communication cost between machines i and j. This is just
an indicator variable for the event that i and j are assigned to different
1
machines, which occurs with probability 1 m
P
. We have C = 1i<jn Cij .
1. Expectation is a straightforward application of linearity of expectation.
There are n2 pairs of processes, and E [Cij ] = 1 m 1
for each pair, so
!
n 1

E [C] = 1 .
2 m
2. Variance is a little trickier because the Cij are not independent. But
they are pairwise independent: even if we fix the location of i and j,
1
the expectation of Cjk is still 1 m , so Cov [Cij , Cjk ] = E [Cij Cjk ]
E [Cij ] E [Cjk ] = 0. So we can compute
!
n 1 1
X
Var [C] = Var [Cij ] = 1 .
1i<j<n
2 m m
B.2.2 A missing hash function

A clever programmer inserts Xi elements in each of m buckets in a hash table,
where each bucket i is implemented as a balanced binary search tree with
search cost at most dlg(Xi + 1)e. We are interested in finding a particular
target element x, which is equally likely to be in any of the buckets, but we
dont know what the hash function is.
Suppose that we know that the Xi are independent and identically
distributed with E [Xi ] = k, and that the location of x is independent of the
values of the Xi . What is best upper bound we can put on the expected cost
of finding x?
Solution
The expected cost of searching bucket i is E [dlg(Xi + 1)e]. This is the
expectation of a function of Xi , so we would like to bound it using Jensens
inequality (4.3).
Unfortunately the function f (n) = dlg(n + 1)e is not concave (because
of the ceiling), but 1 + lg(n + 1) > dlg(n + 1)e is. So the expected cost of
searching bucket i is bounded by 1 + lg(E [Xi ] + 1) = 1 + lg(k + 1).
Assuming we search the buckets in some fixed order until we find x, we
will search Y buckets where E [Y ] = n+12 . Because Y is determined by the
position of x, which is independent of the Xi , Y is also independent of the
Xi . So Walds equation (3.4.3) applies, and the total cost is bounded by
n+1
(1 + lg(k + 1)) .
2

17:00
B.3.1 Tree contraction
Suppose that you have a tree data structure with n nodes, in which each
node i has a pointer parent(u) to its parent (or itself, in the case of the root
node root).
Consider the following randomized algorithm for shrinking paths in the
tree: in the first phase, each node u first determines its parent v = parent(u)
and its grandparent w = parent(parent(u)). In the second phase, it sets
parent0 (u) to v or w according to a fair coin-flip independent of the coin-flips
of all the other nodes.
Let T be the tree generated by the parent pointers and T 0 the tree
generated by the parent0 pointers. An example of two such trees is given in
Figure B.1.
Recall that the depth of a node is defined by depth(root) = 0 and
depth(u) = 1 + depth(parent(u)) when u 6= root. The depth of a tree is equal
to the maximum depth of any node.
Let D be depth of T , and D0 be the depth of T 0 . Show that there is a
constant a, such that for any fixed c > 0, with probability at least 1 nc it
holds that
p p
aDO D log n D0 a D + O D log n . (B.3.1)
0 0
1 1 5
2 5 2 3 6
3 4 6 4
Figure B.1: Example of tree contraction for Problem B.3.1. Tree T is on

the left, T 0 on the right. Nodes 1, 3, 5, and 6 (in boldface) switch to their
grandparents. The other nodes retain their original parents.
Solution
For any node u, let depth(u) be the depth of u in T and depth0 (u) be the
depth of u in T 0 . Note that depth0 (u) is a random variable. We will start by
0
computing E depth (u) as a function of depth(u), by solving an appropriate
recurrence.
Let S(k) = depth0 (u) when depth(u) = k. The base cases are S(0) = 0
(the depth of the root never changes) and S(1) = 1 (same for the roots
children). For larger k, we have
1 1
E depth0 (u) = E 1 + depth0 (parent(u)) + E 1 + depth0 (parent(parent(u)))

2 2
or
1 1
S(k) = 1 + S(k 1) + S(k 2).
2 2
There are various ways to solve this recurrence. The most direct may be
to define a generating function F (z) = k
P
k=0 S(k)z . Then the recurrence
becomes
z 1 1
F = + zF + z 2 F.
1z 2 2
Solving for F gives

z
1z
F = 1 1 2
1 2z 2z
2z
= .
(1 z)(2 z z 2 )
2z
=
(1 z)2 (2 + z)
!
1/3 1/9 1/18
= 2z + +
(1 z) 2 (1 z) 1 + 12 z
2 z 2 z 1 z
= + + ,
3 (1 z) 2 9 1 z 9 1 + 12 z
from which we can read off the exact solution

k1
2 2 1 1

S(k) = k+ +
3 9 9 2
when k 1.4
We can easily show that depth0 (u) is tightly concentrated around depth(u)
using McDiarmids inequality (5.3.11). Let Xi , for i = 2 . . . depth u, be the
choice made by us depth-i ancestor. Then changing one Xi changes depth0 (u)
by at most 1. So we get
2 /(depth(u)1)
Pr depth0 (u) E depth0 (u) + t e2t

(B.3.2)
and similarly
2 /(depth(u)1)
Pr depth0 (u) E depth0 (u) t e2t

. (B.3.3)
q
1
Let t = 2 D ln (1/). Then the right-hand side of (B.3.2) and (B.3.3)
becomes e D ln(1/)/(depth(u)1) < e ln(1/) = . For = 12 nc1 , we get t =
q q
1
= c+1

c+1 )
2 D ln (2n 2 D(ln n + ln 2) = O D log n when c is constant.
0
For the lower bound on D , when can apply (B.3.3) to some u
single node
with depth (u) = D; this node by itself will give D0 23 DO D log n with

4
A less direct but still effective approach is to guess that S(k) grows linearly, and find a
and b such that S(k) ak + b. For this we need ak + b 1 + 12 (a(k 1)+ b)+ 12 (a(k 2) + b).
The bs cancel, leaving ak 1 + ak 32 a. Now the aks cancel, leaving us with 0 1 32 a
or a 2/3. We then go back and make b = 1/3 to get the right bound on S(1), giving the
bound S(k) 32 k + 13 . We can then repeat the argument for S(k) a0 k + b0 to get a full
bound 23 k S(k) 32 k + 13 .
probability at least 1 12 nc1 . For the upper bound, we need to take the max-
imum over all nodes. In general, an upper bound on the maximum of a bunch
of random variables is likely to be larger than an upper bound on any one of
the random variables individually, because there is a lot of room for one of
the variables to get unlucky, but we can happly the union bound to get around
i
this. For each individual u, we have Pr depth0 (u) 23 D + O D log n
1 c1
h
0 2D + O
i
u 12 nc1 = 12 nc . This com-
P
2 n , so Pr D 3 D log n
pletes the proof.
B.3.2 Part testing

You are running a factory that produces parts for some important piece of
machinery. Many of these parts are defective, and must be discarded. There
are two levels of tests that can be performed:
A normal test, which passes a part with probability 2/3.
A rigorous test, which passes a part with probability 1/3.
At each step, the part inspectors apply the following rules to decide which
test to apply:
For the first part, a fair coin-flip decides between the tests.
For subsequent parts, if the previous part passed, the inspectors become
suspicious and apply the rigorous test; if it failed, they relax and apply
the normal test.
For example, writing N+ for a part that passes the normal test, N- for
one that fails the normal test, R+ for a part that passes the rigorous test,
and R- for one that fails the rigorous test, a typical execution of the testing
procedure might look like N- N+ R- N- N+ R- N+ R- N- N- N- N+ R+ R-
N+ R+. This execution tests 16 parts and passes 7 of them.
Suppose that we test n parts. Let S be the number that pass.
1. Compute E [S].
2. Show that there a constant c > 0 such that, for any t > 0,
2 /n
Pr [|S E [S]| t] 2ect . (B.3.4)
Solution
Using McDiarmids inequality and some cleverness Let Xi be the
indicator variable for the event that part i passes, so that S = ni=1 Xi .
P
1. We can show by induction that E [Xi ] = 1/2 for all i. The base case
is X1 , where Pr [part 1 passes] = 12Pr [part

1 passes rigorous test] +
1
2 Pr [part 1 passes normal test] = 12 13 + 23 = 12 . For i > 1, E [Xi1 ] =
1/2 implies that part i is tested with the normal and rigorous tests
with equal probability, so the analysis for X1 carries through and gives
E [Xi ] = 1/2 as well. Summing over all Xi gives E [S] = n/2.
2. We cant use Chernoff, Hoeffding, or Azuma here, because the Xi are
not independent, and do not form a martingale difference sequence even
after centralizing them by subtracting off their expectations. So we
are left with McDiarmids inequality unless we want to do something
clever and new (we dont). Applying McDiarmid to the Xi directly
doesnt work so well, but we can make it work with a different set of
variables that generate the same outcomes.
Let Yi {A, B, C} be the grade of part i, where A means that it passes
both the rigorous and the normal test, B means that fails the rigorous
test but passes the normal test, and C means that it fails both tests.
In terms of the Xi , Yi = A means Xi = 1, Yi = C means Xi = 0,
and Yi = B means Xi = 1 Xi1 (when i > 1). We get the right
probabilities for passing each test by assigning equal probabilities.
We can either handle the coin-flip at the beginning by including an
extra variable Y0 , or we can combine the coin-flip with Y1 by assuming
that Y1 is either A or C with equal probability. The latter approach
improves our bound a little bit since then we only have n variables and
not n + 1.
6 i and ask what happens if Yi
Now suppose that we fix all Yj for j =
changes.
(a) If j < i, then Xj is not affected by Yi .

(b) Let k > i be such that Yk {A, C}. Then Xk is not affected by
Yi , and neither is Xj for any j > k.
It follows that changing Yi can only change Xi , . . . , Xi+` , where ` is

the number of B grades that follow position i.
There are two cases for the sequence X0 . . . X` :
(a) If Xi = 0, then X1 = 1, X2 = 0, etc.

(b) If Xi = 1, then X1 = 0, X2 = 1, etc.
If ` is odd, changing Yi thus has no effect on i+`

P
j=i Xi , while if ` is
even, changing Yi changes the sum by 1. In either case, the effect of
changing Yi is bounded by 1, and McDiarmids inequality applies with
ci = 1, giving
2
Pr [|S E [S]| t] 2e2t /n .

17:00
B.4.1 A doubling strategy
A common strategy for keeping the load factor of a hash table down is to
double its size whenever it gets too full. Suppose that we start with a hash
table of size 1 and double its size whenever two items in a row hash to the
same location.
Effectively, this means that we attempt to insert all elements into a table
of size 1; if two consecutive items hash to the same location, we start over
and try to do the same to a table of size 2, and in general move to a table of
size 2k+1 whenever any two consecutive elements hash to the same location
in our table of size 2k .
Assume that we are using an independently chosen 2-universal hash
function for each table size. Show that the expected final table size is O(n).
Solution
Let X be the random variable representing the final table size. Our goal is
to bound E [X].
First lets look at the probability that we get at least one collision between
consecutive elements when inserting n elements into a table with m locations.
Because the pairs of consecutive elements overlap, computing the exact
probability that we get a collision is complicated, but we only need an upper
bound.
We have n 1 consecutive pairs, and each produces a collision with
probability at most 1/m. This gives a total probability of a collision of at
most (n 1)/m.
Let k = dlg ne, so that 2k n. Then the probability of a consecutive
collision in a table with 2k+` locations is at most (n1)/2k+` < 2k /2k+` = 2` .
Since the events that collisions occur at each table size are independent, we
can compute, for ` > 0,
h i h i
Pr X = 2k+` Pr X 2k+`
`1
2i
Y

i=0
`(`1)/2
=2 .
From this it follows that

X h i
E [X] = 2i Pr X = 2i
i=0
h i
X h i
< 2k Pr X 2k + 2k+` Pr X = 2k+`
l=1

2k+` 2`(`1)/2
X
2k +
l=1

X
2k + 2k 2``(`1)/2
l=1

2 2`)/2
2(`
X
2k + 2k
l=1
k
= O(2 ),
since the series converges to some constant that does not depend on k. But
we chose k so that 2k = O(n), so this gives us our desired bound.
B.4.2 Hash treaps

Consider the following variant on a treap: instead of choosing the heap
keys randomly, we choose a hash function h : U [1 . . . m] from some
strongly 2-universal hash family, and use h(x) as the heap key for tree key x.
Otherwise the treap operates as usual.5
Suppose that |U | n. Show that there is a sequence of n distinct tree
keys such that the total expected time to insert them into an initially empty
hash treap is (n2 /m).
5
If m is small, we may get collisions in the heap key values. Assume in this case that a
node will stop rising if its parent has the same key.
Solution
Insert the sequence 1 . . . n.
Let us first argue by induction on i that all elements x with h(x) = m
appear as the uppermost elements of the right spine of the treap. Suppose
that this holds for i 1. If h(i) = m, then after insertion i is rotated up
until it has a parent that also has heap key m; this extends the sequence
of elements with heap key m in the spine by 1. Alternatively, if h(i) < m,
then i is never rotated above an element x with h(x) = m, so the sequence
of elements with heap key m is unaffected.
Because each new element has a larger tree key than all previous elements,
inserting a new element i requires moving past any elements in the right
spine, and in particular requires moving past any elements j < i with
h(j) = m. So the expected cost of inserting i is at least the expected
number of such elements j. Because h is chosen from a strongly 2-universal
hash family, Pr [h(j) = m] = 1/m for any j, and by linearity of expectation,
E [|{j < i | h(j) = m}|] = (i 1)/m. Summing this quantity over all i gives
a total expected insertion cost of at least n(n 1)/2m = (n2 /m).

17:00
B.5.1 Agreement on a ring
Consider a ring of n processes at positions 0, 1, . . . n 1. Each process i
has a value A[i] that is initially 0 or 1. At each step, we choose a process
r uniformly at random, and copy A[r] to A[(r + 1) mod n]. Eventually, all
A[i] will have the same value.
Let At [i] be the value of A[i] at time t, and let Xt = n1
P
i=0 At [i] be the
total number of ones in At . Let be the first time at which X {0, n}.
1. Suppose that we start with X0 = k. What is the probability that we

eventually reach a state that is all ones?
2. What is E [ ], assuming we start from the worst possible initial state?
Solution
This is a job for the optional stopping theorem. Essentially we are going to
follow the same analysis from 8.5.1 for a random walk with two absorbing
barriers, applied to Xt .
Let Ft be the -algebra generated by A0 , . . . , At . Then {Ft } forms a

filtration, and each Xt is Ft -measurable.
1. For the first part, we show that (Xt , Ft ) is a martingale. The intuition
is that however the bits in At are arranged, there are always exactly
the same number of positions where a zero can be replaced by a one as
there are where a one can be replaced by a zero.
Let Rt be the random location chosen in state At . Observe that
E [Xt+1 | Ft , Rt = i] = Xt + A[i] A[(i + 1) mod n].
But then
n1
X 1
E [Xt+1 | Ft ] = (Xt + A[i] A[(i + 1) mod n])
i=0
n
1 1
= Xt + Xt Xt
n n
= Xt ,
which establishes the martingale property.

We also have that (a) is a stopping time with respect to the Fi ; (b)
Pr [ < ] = 1, because from any state there is a nonzero chance that
all A[i] are equal n steps later; and (c) Xt is bounded between 0 and
n. So the optional stopping theorem applies, giving
E [X ] = E [X0 ] = k.
But then
E [X ] = n Pr [X = n] + 0 Pr [X = 0] = k,
so E [X ] = k/n.
2. For the second part, we use a variant on the Xt2 t martingale.

Let Yt count the number of positions i for which At [i] = 1 and At [(i +
1) mod n] = 0. Then, conditioning on Ft , we have

2
Xt + 2Xt + 1

with probability Yt /n,
2
Xt+1 = Xt2 2Xt + 1 with probability Yt /n, and

X 2

otherwise.
t
The conditional expectation sums to

h i
2
E Xt+1 Ft = Xt2 + 2Yt /n.

Let Zt = Xt2 2t/n when t and X2 2 /n otherwise. Then, for

t < , we have
h i
2
E [Zt+1 | Ft ] = E Xt+1 2(t + 1)/n Ft

= Xt2 + 2Yt /n 2(t + 1)/n

= Xt2 2t/n + 2Yt /n 2/n
= Zt + 2Yt /n 2/n
Zt .
For t , Zt+1 = Zt , so in either case Zt has the submartingale

property.
We have previously established that has bounded expectation, and
its easy to see that Zt has bounded step size. So the optional stopping
theorem applies to Zt , and E [Z0 ] E [Z ].
Let X0 = k. Then E [Z0 ] = k 2 , and E [Z ] = (k/n) n2 2 E [ ] /n =
kn 2 E [ ] /n. But then
k 2 kn 2 E [ ] /n
which gives
kn k 2
E [ ]
2/n
kn2 k 2 n
= .
2
This is maximized at k = bn/2c, giving
bn/2c n2 (bn/2c)2 n
E [ ] .
2
For even n, this is just n3 /4 n3 /8 = n3 /8.

For odd n, this is (n 1)n2 /4 (n 1)2 n/8 = n3 /8 n/8. So there
is a slight win for odd n from not being able to start with an exact
half-and-half split.
To show that this bound applies in the worst case, observe that if we
start with have contiguous regions of k ones and n k zeros in At , then
(a) Yt = 1, and (b) the two-region property is preserved in At+1 . In
this case, for t < it holds that E [Zt+1 | Ft ] = Zt + 2Yt /n 2/n = Zt ,
2 2n
so Zt is a martingale, and thus E [ ] = kn k 2 . This shows that the
initial state with bn/2c consecutive zeros and dn/2e consecutive ones
(or vice versa) gives the claimed worst-case time.
B.5.2 Shuffling a two-dimensional array

A programmer working under tight time constraints needs to write a proce-
dure to shuffle the elements of an n n array, so that all (n2 )! permutations
are equally likely. Some quick Googling suggests that this can be reduced
to shuffling a one-dimensional array, for which the programmers favorite
language provides a convenient library routine that runs in time linear in
the size of the array. Unfortunately, the programmer doesnt read the next
paragraph about converting the 2-d array to a 1-d array first, and instead
decides to pick one of the 2n rows or columns uniformly at random at each
step, and call the 1-d shuffler on this row or column.
Let At be the state of the array after t steps, where each step shuffles
one row or column, and let B be a random variable over permutations of
the original array state that has a uniform distribution. Show that the 2-d
shuffling procedure above is asympotically worse than the direct approach,
by showing that there is some f (n) = (n) such that after f (n) steps,
dT V (At , B) = 1 o(1).6
Solution
Consider the n diagonal elements in positions Aii . For each such element,
there is a 1/n chance at each step that its row or column is chosen. The
6
Ive been getting some questions about what this means, so here is an attempt to
translate it into English.
Recall that f (n) is (g(n)) if limn fg(n)
(n)
goes to infinity, and f (n) is o(g(n)) if
limn fg(n)
(n)
goes to zero.
The problem is asking you to show that there is some f (n) that is more than a constant
times n, such that the total variation distance between Af (n) becomes arbitrarily close to
1 for sufficiently large n.
So for example, if you showed that at t = n4 , dT V (At , B) 1 log12 n , that would
2 n
demonstrate the claim, because limn nn goes to infinity and limn 1/ log
1
= 0. These
functions are, of course, for illustration only. The actual process might or might not
converge by time n4 .)
time until every diagonal node is picked at least once maps to the coupon
collector problem, which means that it is (n log n) with high probability
using standard concentration bounds.
Let C be the event that there is at least one diagonal element that is
in its original position. If there is some diagonal node that has not moved,
C holds; so with probability 1 o(1), At holds at some time t that is
(n log n) = (n). But by the union bound, C holds in B with probability
at most n n2 = 1/n. So the difference between the probability of C in At
and B is at least 1 o(1) 1/n = 1 o(1).

17:00
B.6.1 Sampling colorings on a cycle
Devise a Las Vegas algorithm for sampling 3-colorings of a cycle.
Given n > 1 nodes in a cycle, your algorithm should return a random
coloring X of the nodes with the usual constraint that no edge should have
the same color on both endpoints. Your algorithm should generate all possible
colorings with equal probability, and run in time polynomial in n on average.
Solution
Since its a Las Vegas algorithm, Markov chain Monte Carlo is not going to
help us here. So we can set aside couplings and conductance and just go
straight for generating a solution.
First, lets show that we can generate uniform colorings of an n-node
line in linear time. Let Xi be the color of node i, where 0 i < n. Choose
X0 uniformly at random from all three colors; then for each i > 0, choose
Xi uniformly at random from the two colors not chosen for Xi1 . Given a
coloring, we can show by induction on i that it can be generated by this
process, and because each choice is uniform and each coloring is generated
only once, we get all 3 2n1 colorings of the line with equal probability.
Now we try hooking Xn1 to X0 . If Xn1 = X0 , then we dont have a
cycle coloring, and have to start over. The probability that this event occurs
is at most 1/2, because for every path coloring with Xn1 = X0 , there is
another coloring where we replace Xn1 with a color not equal to Xn2 or
X0 . So after at most 2 attempts on average we get a good cycle coloring.
This gives a total expected cost of O(n).
B.6.2 A hedging problem

Suppose that you own a portfolio of n investment vehicles numbered 1, . . . n,
where investment i pays out aij {1, +1} at time j, where 0 < j m. You
have carefully chosen these investments so that your total payout ni=1 aij
P
for any time j is zero, eliminating all risk.

Unfortunately, in doing so you have run afoul of securities regulators, who
demand that you sell off half of your holdingsa demand that, fortunately,
you anticipated by making n even.
This will leave you with a subset S consisting of n/2 of the ai , and your
net worth at time t will now be wt = tj=1 iS aij . If your net worth drops
P P
too low at any time t, all your worldly goods will be repossessed, and you
will have nothing but your talent for randomized algorithms to fall back on.
Note: an earlier version of this problem demanded a tighter
bound.
Show that when n and m are sufficiently large,
it is always possible to
choose a subset S of size n/2 so that wt m n ln nm for all 0 < t m,
and give an algorithm that finds such a subset in time polynomial in n and
m on average.
Solution
Suppose that we flip a coin independently to choose whether to include each
investment i. There are two bad things that can happen:
1. We lose too much at some time t from the investments the coin chooses.
2. We dont get exactly n/2 heads.
If we can show that the sum of the probabilities of these bad events is
less than 1, we get the existence proof we need. If we can show that it is
enough less than 1, we also get an algorithm, because we can test in time
O(nm) if a particular choice works.
Let Xi be the indicator variable for the event that we include investment
i. Then

n
X t
X
wt = Xi aij
i=1 j=1

n t
1 1
X X
= + Xi aij
i=1
2 2 j=1
t X n n t
!
1X 1
X X
= aij + Xi aij
2 j=1 i=1 i=1
2 i=1
n t
!
1
X X
= Xi aij .
i=1
2 i=1
h i
Because Xi 12 is always 12 , E Xi 12 = 0, and each aij is 1, each

term
in the
P outermost
sum is a zero-mean random variable that satisfies
t t m
Xi 12 a 2 2 . So Hoeffdings inequality says

j=1 ij
h i 2 2
Pr wt E [wt ] < m n ln nm em n ln nm/(2n(m/2) )
= e2 ln nm
= (nm)2 .
Summing over all t, the probability that this bound is violated for any t
is at most n21m .
For the second source of error, we have Pr [ ni=1 Xi 6= n/2] = 1 n2 /2n =
P

1 (1/ n). So the total probability that the random assignment fails is

bounded by 1 (1/ n) + 1/n, giving a probability that it succeeds of at

least (1/ n) 1/(n2 m) = ( n). It follows that generating and testing
random assignments gives an assignment with the desired characteristics

after ( n) trials on average, giving a total expected cost of (n3/2 m).
B.7 Final exam

There are four problems on this exam, each worth 20 points, for a total
B.7.1 Double records (20 points)

Let A[1 . . . n] be an array holding the integers 1 . . . n in random order, with
all n! permutations of the elements of A equally likely.
Call A[k], where 1 k n, a record if A[j] < A[k] for all j < k.
Call A[k], where 2 k n, a double record if both A[k 1] and A[k]
are records.
Give an asymptotic (big-) bound on the expected number of double
records in A.
Solution
Suppose we condition on a particular set of k elements appearing in positions
A[1] through A[k]. By symmetry, all k! permutations of these elements are
equally likely. Putting the largest two elements in A[k 1] and A[k] leaves
(k 2)! choices for the remaining elements, giving a probability of a double
record at k of exactly (k2)!
k!
1
= k(k1) .
Applying linearity of expectation gives a total expected number of double
records of
n n
X 1 X 1

k=2
k(k 1) k=2 (k 1)2

X 1

k=2
(k 1)2

X 1
=
k=1
k2
2
=
6
= O(1).
Since the expected number of double records is at least 1/2 = (1) for
n 2, this gives a tight asymptotic bound of (1).
I liked seeing our old friend 2 /6 so much that I didnt notice an easier
exact bound, which several people supplied in their solutions:

n n
1 1 1
X X
=
k=2
k(k 1) k=2 k 1 k
n1 n
X 1 X 1
=
k=1
k k=2 k
1
=1 .
n
= O(1).
B.7.2 Hipster graphs (20 points)

Consider the problem of labeling the vertices of a 3-regular graph with labels
from {0, 1} to maximize the number of happy nodes, where a node is happy
if its label is the opposite of the majority of its three neighbors.
Give a deterministic algorithm that takes a 3-regular graph as input and
computes, in time polynomial in the size of the graph, a labeling that makes
at least half of the nodes happy.
Solution
This problem produced the widest range of solutions, including several very
clever deterministic algorithms. Here are some examples.
Using the method of conditional probabilities If we are allowed a

randomized algorithm, its easy to get at exactly half of the nodes on average:
simply label each node independently and uniformly at random, observe that
each node individually has a 1/2 chance at happiness, and apply linearity of
expectation.
To turn this into a deterministic algorithm, well apply the method of
conditional expectations. Start with an unlabeled graph. At each step, pick
a node and assign it a label that maximizes the expected number of happy
nodes conditioned on the labeling so far, and assuming that all later nodes
are labeled independently and uniformly at random. We can compute this
conditional expectation in linear time by computing the value for each node
(there are only 34 possible partial labelings of the node and its immediate
neighbors, so computing the expected happiness of any particular node can
be done in constant time by table lookup; to compute the table, we just
enumerate all possible assignments to the unlabeled nodes). So in O(n2 )
time we get a labeling in which the number of happy nodes is at least the
n/2 expected happy nodes we started with.
With a bit of care, the cost can be reduced to linear: because each new
labeled node only affects its own probability of happiness and those of its
three neighbors, we can update the conditional expectation by just updating
the values for those four nodes. This gives O(1) cost per step or O(n) total.
Using hill climbing The following algorithm is an adaptation of the

solution of Rose Sloan, and demonstrates that it is in fact possible to make
all of the nodes happy in linear time.
Start with an arbitrary labeling (say, all 0). At each step, choose an
unhappy node and flip its label. This reduces the number of monochromatic
edges by at least 1. Because we have only 3n/2 edges, we can repeat this
process at most 3n/2 times before it terminates. But it only terminates
when there are no unhappy nodes.
To implement this in linear time, maintain a queue of all unhappy nodes.
B.7.3 Storage allocation (20 points)

Suppose that you are implementing a stack in an array, and you need to
figure out how much space to allocate for the array. At time 0, the size of
the stack is X0 = 0. At each subsequent time t, the user flips an independent
fair coin to choose whether to push onto the stack (Xt = Xt1 + 1) or pop
from the stack (Xt = Xt1 1). The exception is that when the stack is
empty, the user always pushes.
You are told in advance that the stack will only be used for n time units.
Your job is to choose a size s for the array so that it will overflow at most
half the time: Pr [maxt Xt > s] < 1/2. As an asymptotic (big-O) function of
n, what is the smallest value of s you can choose?
Solution
We need two ideas here. First, well show that Xt t2 is a martingale, despite
the fact that Xt by itself isnt. Second, well use the idea of stopping Xt
when it hits s + 1, creating a new martingale Yt = Xt 2 (t ) where is
the first time where Xt = s (and t is shorthand for min(t, )). We can
then apply Markovs inequality to Xn .
To save time, well skip directly to showing that Yt is a martingale. There
are two cases:
1. If t < , then
1 1
E [Yt+1 | Yt , t < ] = (Xt + 1)2 (t + 1) + (Xt 1)2 (t + 1)
2 2
2
= Xt + 1 (t + 1)
= Xt2 t
= Yt .
2. If t , then Yt+1 = Yt , so E [Yt+1 | Yt , t ] = Yt .
In either case the martingale property holds. 2 2

It follows that E [Yn ] = E [Y0 ] = 0, or E Xn n = 0, giving E Xn =
n. Now apply Markovs inequality:
Pr [max Xt > s] = Pr [Xn s + 1]

h i
2
= Pr Xn (s + 1)2
2

E Xn

(s + 1)2
n
=
( 2n + 1)2
n
<
2n
= 1/2.

So s = 2n = O( n) is enough.
B.7.4 Fault detectors in a grid (20 points)

A processing plant for rendering discarded final exam problems harmless
consists of n2 processing units arranged in an nn grid with coordinates (i, j)
each in the range 1 through n. We would like to monitor these processing
units to make sure that they are working correctly, and have access to a
supply of monitors that will detect failures. Each monitor is placed at some
position (i, j), and will detect failures at any of the five positions (i, j),
(i 1, j), (i + 1, j), (i, j 1), and (i, j + 1) that are within the bounds of
the grid. This plus-shaped detection range is awkward enough that the
engineer designing the system has given up on figuring out how to pack the
detectors properly, and instead places a detector at each grid location with
independent probability 1/4.
The engineer reasons that since a typical monitor covers 5 grid locations,
using n2 /4 monitors on average should cover (5/4)n2 locations, with the
extra monitors adding a little bit of safety to deal with bad random choices.
So few if any processing units should escape.
1. Compute the exact expected number of processing units that are not
within range of any monitor, as a function of n. You may assume
n > 0.
c, the actual number of unmonitored processing

2. Show that for any fixed
units is within O(n log n) of the expected number with probability at
least 1 nc .
Solution
1. This part is actually more annoying, because we have to deal with
nodes on the edges. There are three classes of nodes:
(a) The four corner nodes. To be unmonitored, each corner node

needs to have no monitor in its own location or either of the two
adjacent locations. This event occurs with probability (3/4)3 .
(b) The 4(n 2) edge nodes. These have three neighbors in the grid,
so they are unmonitored with probability (3/4)4 .
(c) The (n 2)2 interior nodes. These are each unmonitored with
probability (3/4)5 .
Adding these cases up gives a total expected number of unmonitored

nodes of
243 2 81 189
4 (3/4)3 + 4(n 2) (3/4)4 + (n 2)2 (3/4)5 = n n+ .
1024 32 64
(B.7.1)
For n = 1, this analysis breaks down; instead, we can calculate directly
that the sole node in the grid has a 3/4 chance of being unmonitored.
Leaving the left-hand side of (B.7.1) in the original form is probably a
good idea for understanding how the result works, but the right-hand
side demonstrates that this strategy leaves slightly less than a quarter
of the processing units uncovered on average.
2. Let Xij be the indicator for a monitor at position (i, j). Recall that
we have assumed that these variables are independent.
Let S = f (X11 , . . . , Xnn ) compute the number of uncovered processing

units. Then changing a single Xij changes the value of f by at most 5.
So we can apply McDiarmids inequality to get
2
P
c2ij )
Pr [|S E [S]| t] 2e2t /(
2 2 2 2 2
= 2e2t /(n 5 ) = 2e2t /(25n ) .
(The actual bound is slightly better, because we are overestimating cij

for boundary nodes.)

Set this to nc and solve for t = n (25/2)(c ln n + ln 2) = O(n log n).
p
Then the bound becomes 2ec ln nln 2 = nc , as desired.

Appendix C

Spring 2013
C.1 Assignment 1: due Wednesday, 2013-01-30, at

17:00
C.1.1 Bureaucratic part
1. Your name.

etc.
anyway.)
C.1.2 Balls in bins

Throw m balls independently and uniformly at random into n bins labeled
1 . . . n. What is the expected number of positions i < n such that bin i and
bin i + 1 are both empty?
294
APPENDIX C. SAMPLE ASSIGNMENTS FROM SPRING 2013 295
Solution
If we can figure out the probability that bins i and i + 1 are both empty for
some particular i, then by symmetry and linearity of expectation we can just
multiply by n 1 to get the full answer.
For bins i and i + 1 to be empty, every ball must choose another bin.
This occurs with probability (1 2/n)m . The full answer is thus n(1 2/n)m ,
or approximately ne2m/n when n is large.
C.1.3 A labeled graph

Suppose you are given a graph G = (V, E), where |V | = n, and you want to
assign labels to each vertex such that the sum of the labels of each vertex
and its neighbors modulo n + 1 is nonzero. Consider the nave randomized
algorithm that assigns a label in the range 0 . . . n to each vertex independently
and uniformly at random and then tries again if the resulting labeling doesnt
work. Show that this algorithm will find a correct labeling in time polynomial
in n on average.
Solution
Well use the law of total probability. First observe that the probability that
a random labeling yields a zero sum for any single vertex and its neighbors is
exactly 1/(n + 1); the easiest way to see this is that after conditioning on the
values of the neighbors, there is only one value in n + 1 that can be assigned
to the vertex itself to cause a failure. Now sum this probability over all n
vertices to get a probability of failure of at most n/(n + 1). It follows that
after n + 1 attempts on average (each of which takes O(n2 ) time to check all
the neighborhood sums), the algorithm will find a good labeling, giving a
total expected time of O(n3 ).
C.1.4 Negative progress

An algorithm has the property that if it has already run for n steps, it runs
for an additional n + 1 steps on average. Formally, let T 0 be the random
variable representing the running time of the algorithm, then
E [T | T n] = 2n + 1. (C.1.1)
For each n 0, what is the conditional probability Pr [T = n | T n]

that the algorithm stops just after its n-th step?
Solution
Expand (C.1.1) using the definition of conditional expectation to get

X
2n + 1 = x Pr [T = x | T n]
x=0

X Pr [T = x T n]
= x
x=0
Pr [T n]

1 X
= x Pr [T = x] ,
Pr [T n] x=n
which we can rearrange to get

X
x Pr [T = x] = (2n + 1) Pr [T n] , (C.1.2)
x=n
provided Pr [T n] is nonzero. We can justify this assumption by observing

that (a) it holds for n = 0, because T 0 always; and (b) if there is some
n > 0 such that Pr [T n] = 0, then E [T | T n 1] = n 1, contradicting
(C.1.1).
Substituting n + 1 into (C.1.2) and subtracting from the original gives
the equation

X
X
n Pr [T = n] = x Pr [T = x] x Pr [T = x]
x=n x=n+1
= (2n + 1) Pr [T n] (2n + 3) Pr [T n + 1]
= (2n + 1) Pr [T = n] + (2n + 1) Pr [T n + 1] (2n + 3) Pr [T n + 1]
= (2n + 1) Pr [T = n] 2 Pr [T n + 1] .
Since we are looking for Pr [T = n | T n] = Pr [T = n] / Pr [T n],
having an equation involving Pr [T n + 1] is a bit annoying. But we can
borrow a bit of Pr [T = n] from the other terms to make it work:
n Pr [T = n] = (2n + 1) Pr [T = n] 2 Pr [T n + 1]
= (2n + 3) Pr [T = n] 2 Pr [T = n] 2 Pr [T n + 1]
= (2n + 3) Pr [T = n] 2 Pr [T n] .
A little bit of algebra turns this into
Pr [T = n] 2
Pr [T = n | T n] = = .
Pr [T n] n+3
C.2 Assignment 2: due Thursday, 2013-02-14, at

17:00
C.2.1 A local load-balancing algorithm
Suppose that we are trying to balance n jobs evenly between two machines.
Job 1 chooses the left or right machine with equal probability. For i > 1, job
i chooses the same machine as job i 1 with probability p, and chooses the
other machine with probability 1 p. This process continues until every job
chooses a machine. We want to estimate how imbalanced the jobs are at the
end of this process.
Let Xi be +1 if the i-th job chooses the left machine and 1 if it chooses
the right. Then S = ni=1 Xi is the difference between the number of jobs
P
that choose the left machine and the number that choose the right. By
symmetry, the expectation of S is zero. What is the variance of S as a
function of p and n?
Solution
P
To compute the variance, well use (5.1.4), which says that Var [ i Xi ] =
P P
i Var [Xi ] + 2 i<j Cov [Xi , Xj ].
Recall that Cov [Xi , Xj ] = E [Xi Xj ] E [Xi ] E [Xj ]. Since the last term
is 0 (symmetry again), we just need to figure out E [Xi Xj ] for all i j (the
i = j case gets us Var [Xi ]).
First, lets compute E [Xj = 1 | Xi = 1]. Its easiest to do this starting
with the j = i case: E [Xi | Xi = 1] = 1. For larger j, compute
E [Xj | Xj1 ] = pXj1 + (1 p)(Xj1 )

= (2p 1)Xj1 .
It follows that
E [Xj | Xi = 1] = E [(2p 1)Xj1 | Xi = 1]

= (2p 1) E [Xj1 | Xi = 1] . (C.2.1)
The solution to this recurrence is E [Xj | Xi = 1] = (2p 1)ji .

We next have
Cov [Xi , Xj ] = E [Xi Xj ]

= E [Xi Xj | Xi = 1] Pr [Xi = 1] + E [Xi Xj | Xi = 1] Pr [Xi = 1]
1 1
= E [Xj | Xi = 1] + E [Xj | Xi = 1]
2 2
1 1
= E [Xj | Xi = 1] + E [Xj | Xi = 1]
2 2
= E [Xj | Xi = 1]
= (2p 1)ji
as calculated in (C.2.1).
So now we just need to evaluate the horrible sum.
X X X n
n X
Var [Xi ] + 2 Cov [Xi , Xj ] = n + 2 (2p 1)ji
i i<j i=1 j=i+1
n ni
X X
=n+2 (2p 1)k
i=1 k=1
n
X (2p 1) (2p 1)ni+1
=n+2
i=1
1 (2p 1)
n
n(2p 1) 1 X
=n+ (2p 1)m
1p 1 p m=1
n(2p 1) (2p 1) (2p 1)n+1
=n+ .
1p 2(1 p)2
(C.2.2)
This covers all but the p = 1 case, for which the geometric series formula
fails. Here we can compute directly that Var [S] = n2 , since S will be n
with equal probability.
For smaller values of p, plotting (C.2.2) shows the variance increasing
smoothly starting at 0 (for even n) or 1 (for odd n) at p = 0 to n2 in the limit
as p goes to 1, with an interesting intermediate case of n at p = 1/2, where
all terms but the first vanish. This makes a certain intuitive sense: when
p = 0, the processes alternative which machine they take, which gives an
even split for even n and a discrepancy of 1 for odd n; when p = 1/2, the
processes choose machines independently, giving variance n; and for p = 1,
the processes all choose the same machine, giving n2 .
C.2.2 An assignment problem

Here is an algorithm for placing n balls into n bins. For each ball, we first
select two bins uniformly at random without replacement. If at least one
of the chosen bins is unoccupied, the ball is placed in the empty bin at no
cost. If both chosen bins are occupied, we execute an expensive parallel scan
operation to find an empty bin and place the ball there.
1. Compute the exact value of the expected number of scan operations.
2. Let c > 0. Show that the absolute value of the difference between the
actual
number of scan operations and the expected number is at most
O( cn log n) with probability at least 1 nc .
Solution
1. Number the balls 1 to n. Forball
i,there are i1 bins already occupied,
giving a probability of i1
n
i2
n1 that we choose an occupied bin on
both attempts and incur a scan. Summing over all i gives us that the
expected number of scans is:
n n1
i1 i2 1
X X
= (i2 i)
i=1
n n1 n(n 1) i=1
1 (n 1)n(2n 1) (n 1)n)

=
n(n 1) 6 2
(2n 1) 1
=
6 2
n2
,
3
provided n 2. For n < 2, we incur no scans.
2. Its tempting to go after this using Chernoffs inequality, but in this case
Hoeffding gives a better bound. Let S be the number of scans. Then S is
the sum of n independent Bernoulli random variables,
so 5.3.3says that
2
Pr [|S E [S]| t] 2e2t /n . Now let t = cn ln n = O( cn log n)
to make the right-hand side 2n2c nc for sufficiently large n.
C.2.3 Detecting excessive collusion

Suppose you have n students in some unnamed Ivy League universitys
Introduction to Congress class, and each generates a random 1 vote, with
both outcomes having equal probability. It is expected that members of the

same final club will vote together, so it may be that many groups of up to k
students each will all vote the same way (flipping a single coin to determine
the vote of all students in the group, with the coins for different groups
being independent). However, there may also be a much larger conspiracy
of exactly m students who all vote the same way (again based on a single
independent fair coin), in violation of academic honesty regulations.
Let c > 0. How large must m be in asymptotic terms as a function of
n, k, and c so that the existence of a conspiracy can be detected solely by
looking at the total vote, where the probability of error (either incorrectly
claiming a conspiracy when none exists or incorrectly claiming no conspiracy
when one exists) is at most nc ?
Solution
Let S be the total vote. The intuition here is that if there is no conspiracy,
S is concentrated around 0, and if there is a conspiracy, S is concentrated
around m. So if m is sufficiently large and |S| m/2, we can reasonably
guess that there is a conspiracy.
We need to prove two bounds: first, that the probability that we see
|S| m/2 when there is no conspiracy is small, and second, that the
probability that we see |S| < m/2 when there is a conspiracy is large.
For the first case, let Xi be the total vote cast by the i-th group. This
will be ni with equal probability, where ni k is the size of the group.
P
This gives E [Xi ] = 0. We also have that ni = n.
Because the Xi are all bounded, we can use Hoeffdings inequality (5.3.1),
so long as we can compute an upper bound on n2i . Here we use the fact
P
P 2
that ni is maximized subject to 0 ni k and ni = 0 by setting as
P
many ni as possible to k; this follows from convexity of x 7 x2 .1 We thus

have
X
n2i bn/kck 2 + (n mod k)2
bn/kck 2 + (n mod k)k
(n/k)k 2
= nk.
1
The easy way to see this is that if f is strictly convex, then P f 0 is increasing. So if
0 < ni njP < k, increasing nj by while decreasing ni by leaves ni unchanged while
increasing f (ni ) by (f 0 (nj ) f 0 (ni )) + O(2 ), which will be positive when is small
enough. So at any global maximum, we must have that at least one of ni or nj equals 0 or
k for any i 6= j.
Now apply Hoeffdings inequality to get

2 /2nk
Pr [|S| m/2] 2e(m/2) .
We want to set m so that the right-hand side is less than nc . Taking

logs as usual gives
ln 2 m2 /8nk c ln n,
so the desired bound holds when

q
m 8nk(c ln n + ln 2)
p
= ( ckn log n).
For the second case, repeat the above analysis on the n m votes except
the m from the conspiracy. Again we get that if m = ( ckn log n), the
c
that these votes exceed m/2 is bounded by n . So in both cases
probability
m = ( ckn log n) is enough.

17:00
C.3.1 Going bowling
For your crimes, you are sentenced to play n frames of bowling, a game
that involves knocking down pins with a heavy ball, which we are mostly
interested in because of its complicated scoring system.
In each frame, your result may be any of the integers 0 through 9, a
spare (marked as /), or a strike (marked as X). We can think of your result
on the i-th frame as a random variable Xi . Each result gives a base score
Bi , which is equal to Xi when Xi {0 . . . 9} and 10 when Xi {/, X}. The
actual score Yi for frame i is the sum of the base scores for the frame and up
to two subsequent frames, according to the rule:

Bi

when Xi {0 . . . 9},
Yi = B +B
i i+1 when Xi = /, and

B + B +B when Xi = X.
i i+1 i+2
To ensure that Bi+2 makes sense even for i = n, assume that there exist
random variables Xn+1 and Xn+2 and the corresponding Bn+1 and Bn+2 .
Suppose that the Xi are independent (but not necessarily identically

distributed). Show that your final score S = ni=1 Yi is exponentially con-
P
centrated2 around its expected value.
Solution
This is a job for McDiarmids inequality (5.3.11). Observe that S is a function
of X1 . . . Xn+2 . We need to show that changing any one of the Xi wont
change this function by too much.
From the description of the Yi , we have that Xi can affect any of Yi2 (if
Xi2 = X), Yi1 (if Xi1 = /) and Yi . We can get a crude bound by observing
P
that each Yi ranges from 0 to 30, so changing Xi can change Yi by at most
90, giving ci 90. A better bound can be obtained by observing that Xi
contributes only Bi to each of Yi2 and Yi1 , so changing Xi can only change
these values by up to 10; this gives ci 50. An even more pedantic bound
can be obtained by observing that X1 , X2 , Xn+1 , and Xn+2 are all special
cases, with c1 = 30, c2 = 40, cn+1 = 20, and cn+2 = 10, respectively; these
values can be obtained by detailed meditation on the rules above.
We thus have n+2 2 2 2 2 2 2
i=1 ci = (n 2)50 + 30 + 40 + 20 + 10 = 2500(n
P
2) + 3000 = 2500n 2000, assuming n 2. This gives Pr [|S E [S]| t]

exp(2t2 /(2500n 2000)), with the symmetric bound holding on the other
side as well.
Forthe standard game of bowling, with n = 10, this bound starts to bite
at t = 11500 107, which is more than a third of the range between the
minimum and maximum possible scores. Theres a lot of variance in bowling,
but this looks like a pretty loose bound for players who dont throw a lot
of strikes. For large n, we get the usual bound of O n log n with high
probability: the averaging power of endless repetition eventually overcomes
any slop in the constants.
C.3.2 Unbalanced treaps

Recall that a treap (6.3) is only likely to be balanced if the sequence of
insert and delete operations applied to it is independent of the priorities
chosen by the algorithm.
Suppose that we insert the keys 1 through n into a treap with random
priorities as usual, but then allow the adversary to selectively delete whichever
keys it wants to after observing the priorities assigned to each key.
2
This means more concentrated than you can show using Chebyshevs inequality.
Show that there is an adversary strategy that produces a path in the

treap after deletions that has expected length ( n).
Solution
An easy way to do this is to produce a tree that consists of a single path,
which we can do by arranging that the remaining keys have priorities that
are ordered the same as their key values.

Heres a simple strategy that works. Divide the keys into n ranges of

n keys each (1 . . . n, n + 1 . . . 2 n, etc.).3 Rank the priorities from 1

to n. From each range (i 1) n . . . i n, choose a key to keep whose priority

is also ranked in the range (i 1) n . . . i n (if there is one), or choose no
key (if there isnt). Delete all the other keys.

For a particular range, we are drawing n samples without replacement

from the n priorities, and there are n possible choices that cause us
tokeep a key in that range. The probability that every draw misses is
Q n n
i=1 (1 n/(n i + 1)) (1 1/ n) e1 . So each range contributes
1
at least 1 e keys on average. Summing over all n ranges gives a
sequence of keys with increasing priorities with expected length at least

(1 e1 ) n = ( n).
An alternative solution is to apply the Erds-Szekeres theorem [ES35],
which says that every sequence of length k 2 + 1 has either an increasing
subsequence of length k + 1 or a decreasing sequence of k + 1. Consider the
sequence of priorities corresponding to the keys 1 . . . n; letting k = b n 1c
gives a subsequence of length at least n 1 that is either increasing or
decreasing. If we delete all other elements of the treap, the elements cor-
responding to this subsequence will form a path, giving the desired bound.
Note that this does not require any probabilistic reasoning at all.

Though not required for the problem, its possible to show that ( n)
is the best possible bound here. The idea is that the number of possible
sequences of keys that correspond to a path of length k in a binary search
tree is exactly nk 2k1 ; the nk corresponds to choosing the keys in the path,

and the 2k1 is because for each node except the last, it must contain either
the smallest or the largest of the remaining keys because of the binary search
tree property.
Since each such sequence will be a treap path only if the priorities are
decreasing (with probability 1/k!), the union bound says that the probability
3
To make our life easer, well assume that n is a square. This doesnt affect the
asymptotic result.
*
/ \
a / \ b
/ \
* *
/ \ \
a / \ b \ b
/ \ \
aa ab bb
Figure C.1: A radix tree, storing the strings aa, ab, and ba.
n k1
of having any length-k paths is at most k 2 /k!. But
!
n k1 (2n)k
2 /k!
k 2(k!)2
(2n)k

2(k/e)2k
1
= (2e2 n/k 2 )k .
2

This is exponentially small for k 2e2 n, showing that with high probability

all possible paths have length O( n).
C.3.3 Random radix trees

A radix tree over an alphabet of size m is a tree data structure where each
node has up to m children, each corresponding to one of the letters in the
alphabet. A string is represented by a node at the end of a path whose edges
are labeled with the letters in the string in order. For example, in Figure C.1,
the string ab is stored at the node reached by following the a edge out of
the root, then the b edge out of this child.
The only nodes created in the radix tree are those corresponding to stored
keys or ancestors of stored keys.
Suppose you have a radix tree into which you have already inserted
n strings of length k from an alphabet of size m, generated uniformly at
random with replacement. What is the expected number of new nodes you
need to create to insert a new string of length k?
Solution
We need to create a new node for each prefix of the new string that is not
already represented in the tree.
For a prefix of length `, the chance that none of the n strings have this
n
prefix is exactly 1 m` . Summing over all ` gives that the expected
n
number of new nodes is k`=0 1 m` .
P
There is no particularly clean expression for this, but we can observe

that (1 m` )n exp(nm` ) is close to zero for ` < logm n and close to 1
for ` > logm n. This suggests that the expected value is k logm n + O(1).

17:00
C.4.1 Flajolet-Martin sketches with deletion
A Flajolet-Martin sketch [FNM85] is a streaming data structure for
approximately counting the number of distinct items n in a large data
stream using only m = O(log n) bits of storage.4 The idea is to use a hash
function h that generates each value i {1, . . . m} with probability 2i ,
and for each element x that arrives in the data stream, we write a 1 to
A[h(x)]. (With probability 2m we get a value outside this range and write

nowhere.) After inserting n distinct elements, we estimate n as n = 2k ,
where k = max {k | A[k] = 1}, and argue that this is likely to be reasonably
close to n.
Suppose that we modify the Flajolet-Martin sketch to allow an element
x to be deleted by writing 0 to A[h(x)]. After n insertions and d deletions
(of distinct elements in both cases), we estimate the number of remaining

elements n d as before by n \ d = 2k , where k = max {k | A[k] = 1}.
Assume that we never delete an element that has not previously been
inserted, and that the values of h are for different inputs are independent of
each other and of the sequence of insertions and deletions.
Show that there exist constants c > 1 and > 0, such that for n
i elements then deleting d n of
h large, after inserting n distinct
sufficiently
them, Pr (n d)/c n d (n d)c 2/3.
\
4
If you actually need to do this, there exist better data structures for this problem. See
[KNW10].
Solution
Well apply the usual error budget approach and show that the probability
\
that n d is too big and the probability that n \ d is too small are both
small. For the moment, we will leave c and as variables, and find values
that work at the end.
Lets start with the too-big side. To get A[k] = 1, we need h(xi ) = k for
some xi that is inserted but not subsequently deleted. There are nd such xi ,
and each gives h(xi ) = k with probability 2k . So Pr [A[k] = 1] (n d)2k .
This gives
h i h i
\
Pr n d (n d)c = Pr kb dlg ((n d)c)e

(n d)2k
X

k=dlg((nd)c)e
= 2(n d)2dlg((nd)c)e
2
.
c
On the too-small side, fix k =hdlg ((n d)/c))e. iSince A[k]

h = 1 gives k i
k dlg ((n d)/c))e, we have Pr n \ d < (n d)/c = Pr k < lg(n d)/c
Pr [A[k] = 0]. (We might be able to get a better bound by looking at larger
indices, but to solve the problem this one k will turn out to be enough.)
Let x1 . . . xnd be the values that are inserted and not later deleted, and
xnd+1 . . . xn the values that are inserted and then deleted. For A[k] to be
zero, either (a) no xi for i in 1 . . . xnd has h(xi ) = k; or (b) some xi for
i in n d + 1 . . . xn has h(xi ) = k. The probability of the first event is
nd d
1 2k ; the probability of the second is 1 1 2k . So we have
nd d
Pr [A[k] = 0] 1 2k + 1 1 2k

exp 2k (n d) + 1 exp 2k + 22k d

exp 2dlg((nd)/c))e (n d) + 1 exp 2 2dlg((nd)/c))e d

exp 2 lg((nd)/c)) (n d) + 1 exp 2 2 lg((nd)/c))+1 d
4cd

c
= e + 1 exp
nd
4cd
ec +
nd
c 4c
e + .
1
So our total probability of error is bounded by 2c + ec + 1
4c
. Let c = 8
1 8 128 1
and = 1/128 to make this less than 4 + e + 127 16 0.313328 < 1/3,
giving the desired bound.
C.4.2 An adaptive hash table

Suppose we want to build a hash table, but we dont know how many
elements we are going to put in it, and because we allow undisciplined C
programmers to obtain pointers directly to our hash table entries, we cant
move an element once we assign it a position to it. Here we will consider a
data structure that attempts to solve this problem.
i
Construct a sequence of tables T0 , T1 , . . . , where each Ti has mi = 22
slots. For each table Ti , choose k independent strongly 2-universal hash
functions hi1 , hi2 , . . . hik .
The insertion procedure is given in Algorithm C.1. The essentially idea
is that we make k attempts (each with a different hash function) to fit x into
T0 , then k attempts to fit it in T1 , and so on.
If the tables Ti are allocated only when needed, the space complexity of
this data structure is given by the sum of mi for all tables that have at least
one element.
Show that for any fixed > 0, there is a constant k such that after
inserting n elements:
1. The expected cost of an additional insertion is O(log log n), and
2. The expected space complexity is O n2+ .

1 procedure insert(x)
2 for i 0 to do
3 for j 1 to k do
4 if Ti [hij (x)] = then
5 Ti [hij (x)] x
6 return
Algorithm C.1: Adaptive hash table insertion
Solution
The idea is that we use Ti+1 only if we get a collision in Ti . Let Xi be the
indicator for the event that there is a collision in Ti . Then

X
E [steps] 1 + E [Xi ] (C.4.1)
i=0
and

X
E [space] m0 + E [Xi ] mi+1 . (C.4.2)
i=0
To bound E [Xi ], lets calculate an upper bound on the probability that

a newly-inserted element xn+1 collides with any of the previous n elements
x1 . . . xn in table Ti . This occurs if, for every location hij (xn+1 ), there is some
xr and some j 0 such that hij 0 (xr ) = hij (xn+1 ). The chance that this occurs
for any particular j, j 0 , and r is at most 1/mi (if j = j 0 , use 2-universality
of hij , and if j 6= j 0 , use independence and uniformity), giving a chance
that it occurs for fixed j that is at most n/mi . The chance that it occurs
for all j is at most (n/mi )k , and the expected number of such collisions
summed over any n + 1 elements that we insert is bounded by n(n/mi )k
(the first element
cant collide
with any previous elements). So we have
E [Xi ] min 1, n(n/mi ) . k
Let ` be the largest value such that m` n2+ . We will show that, for
an appropriate choice of k, we are sufficiently unlikely to get a collision in
round ` that the right-hand sides of (C.4.2) and (C.4.2) end up being not
much more than the corresponding sums up to ` 1.
From our choice of `, it follows that (a) ` lg lg n2+ = lg lg n+lg(2+) =

O(log log n); and (b) m`+1 > n2+ , giving m` = m`+1 > n1+/2 . From this
we get E [X` ] n(n/m` )k < n1k/2 .
By choosing k large enough, we can make this an arbitrarily small

polynomial in n. Our goal is to wipe out the E [X` ] m`+1 and subsequent
terms in (C.4.2).
Observe that E [Xi ] mi+1 n(n/mi )k m2i = nk+1 m2k
i . Lets choose k so
that this is at most 1/mi , when i `, so we get a nice convergent series.5
This requires nk+1 m3k
i 1 or k + 1 + (3 k) logn mi 0. If i `, we
have logn mi > 1 + /2, so we win if k + 1 + (3 k)(1 + /2) 0. Setting
k 8/ + 3 works. (Since we want k to be an integer, we probably want
k = d8/e + 3 instead.)
So now we have

X
E [space] m0 + E [Xi ] mi+1
i=0
`
X X 1
mi +
i=0 i=`
mi
2
2m` +
m`
2+
= O(n ).
For E [steps], compute the same sum without all the mi+1 factors. This
makes the tail terms even smaller, so they is still bounded by a constant,
and the head becomes just `i=0 1 = O(log log n).
P
C.4.3 An odd locality-sensitive hash function

A deranged computer scientist decides that if taking one bit from a random
index in a bit vector is a good way to do locality-sensitive hashing (see
7.7.1.3), then taking the exclusive OR of k independently chosen indices
must be even better.
Formally, given a bit-vector x1 x2 . . . xn , and a sequence of indices i1 i2 . . . ik ,
Lk
define hi (x) = j=1 xij . For example, if x = 00101 and i = 3, 5, 2,
hi (x) = 1 1 0 = 0.
Suppose x and y are bit-vectors of length n that differ in m places.
1. Give a closed-form expression for the probability that hi (x) 6= hi (y),

assuming i consists of k indices chosen uniformly and independently at
random from 1 . . . n
5
We could pick a smaller k, but why not make things easier for ourselves?
2. Use this to compute the exact probability that hi (x) 6= hi (y) when
m = 0, m = n/2, and m = n.
Hint: You may find it helpful to use the identity (a mod 2) = 12 (1(1)a ).
Solution
1. Observe that hi (x) 6= hi (y) if an only if i chooses an odd number of
indices where x and y differ. Let p = m/n be the probability that each
index in i hits a position where x and y differ, and let q = 1 p. Then
the event that we get an odd number of differences is
k k
! !
X k j kj X 1 k
(j mod 2) p q = 1 (1)j pj q kj
j=0
j j=0
2 j
k k
! !
1X k j kj 1 X k
= p q (p)j q kj
2 j=0 j 2 j=0 j
1 1
= (p + q)k (p + q)k
2 2
1 (1 2(m/n))k
= .
2
11k
2. For m = 0, this is 2 = 0.
k
For m = n/2, its 102 = 1
2 (assuming k > 0).
k
For m = n, its 1(1)
2 = (k mod 2).
In fact, the chances of not colliding as a function of m are symmetric
around m = n/2 if k is even and increasing if k is odd. So we can only hope
to use this as locality-sensitive hash function in the odd case.
C.5 Assignment 5: due Friday, 2013-04-12, at 17:00

C.5.1 Choosing a random direction
Consider the following algorithm for choosing a random direction in three
dimensions. Start at the point (X0 , Y0 , Z0 ) = (0, 0, 0). At each step, pick one
of the three coordinates uniformly at random and add 1 to it with equal
probability. Continue until the resulting vector has length at least k, i.e.,
until Xt2 + Yt2 + Zt2 k 2 . Return this vector.
What is the expected running time of this algorithm, as an asymptotic
function of k?
Solution
The trick is to observe that Xt2 + Yt2 + Zt2 t is a martingale, essentially
following the same analysis as for Xt2 t for a one-dimensional random walk.
Suppose we pick Xt to change. Then
h
2
i 1
E Xt+1 Xt = (Xt + 1)2 + (Xt 1)2

2
= Xt2 + 1.
So
h i
2 2 2
E Xt+1 + Yt+1 + Zt+1 (t + 1) Xt , Yt , Zt , X changes = Xt2 + Yt2 + Zt2 t.

But by symmetry, the same equation holds if we condition on Y or Z changing.

It follows that E Xt+1 + Yt+1 + Zt+1 (t + 1) Xt , Yt , Zt = Xt2 +Yt2 +Zt2
2 2 2

t, and that we have a martingale as claimed.
Let be the first time at which Xt2 +Yt2 +Zt2 k 2 . From the optional stop-
ping theorem (specifically, the bounded-increments case of Theorem 8.3.1),
E X2 + Y2 + Z2 = 0, or equivalently E [ ] = E X2 + Y2 + Z2 . This

immediately gives E [ ] k 2 .
To get an upper bound, observe that X21 + Y21 + Z21 < k 2 , and
that exactly one of these three term increases between 1 and . Suppose
its X (the other cases are symmetric). Increasing X by 1 sets X2 =
X21 + 2X 1 + 1. So we get

X2 + Y2 + Z2 = X21 + Y21 + Z21 + 2X 1 + 1
< k 2 + 2k + 1.
So we have k 2 E [ ] < k 2 + 2k + 1, giving E [ ] = (k 2 ). (Or k 2 + O(k)
if we are feeling really precise.)
C.5.2 Random walk on a tree

Consider the following random walk on a (possibly unbalanced) binary search
tree: At each step, with probability 1/3 each, move to the current nodes
parent, left child, or right child. If the target node does not exist, stay put.
Suppose we adapt this random walk using Metropolis-Hastings (see 9.2.7)
so that the probability of each node at depth d in the stationary distribution
is proportional to d .
Use a coupling argument to show that, for any constant > 2, this
adapted random walk converges in O(D) steps, where D is the depth of the
tree.
Solution
As usual, let Xt be a copy of the chain starting in an arbitrary initial state
and Yt be a copy starting in the stationary distribution.
From the Metropolis-Hastings algorithm, the probability that the walk
moves to a particular child is 1/3, so the probability that the depth increases
after one step is at most 2/3. The probability that the walk moves to the
parent (if we are not already at the root) is 1/3.
Well use the same choice (left, right, or parent) in both the X and
Y processes, but it may be that only one of the particles moves (because
the target node doesnt exist). To show convergence, well track Zt =
max(depth(Xt ), depth(Yt )). When Zt = 0, both Xt and Yt are the root
node.
There are two ways that Zt can change:
1. Both processes choose parent; if Zt is not already 0, Zt+1 = Zt 1.
This case occurs with probability 1/3.
2. Both processes choose one of the child directions. If the appropriate

child exists for the deeper process (or for either process if they are
at the same depth), we get Zt+1 = Zt + 1. This even occurs with
probability at most 2/3 < 1/3.
So the expected change in Zt+1 conditioned on Zt > 0 is at most 1/3 +
2/3 = (1/3)(2/ 1). Let be the first time at which Zt = 0. Then
the process Zt0 = Zt (1/3)(2/ 1)t for t and 0 for t > is a
supermartingale, so E [Z ] = E [Z0 ] = E [max(depth(X0 ), depth(Y0 ))] D.
3D
This gives E [ ] 2/1 .
C.5.3 Sampling from a tree

Suppose we want to sample from the stationary distribution of the Metropolis-
Hastings walk in the previous problem, but we dont want to do an actual
random walk. Assuming > 2 is a constant, give an algorithm for sampling
exactly from the stationary distribution that runs in constant expected time.
Your algorithm should not require knowledge of the structure of the tree.
Its only input should be a pointer to the root node.
Clarification added 2013-04-09: Your algorithm can determine the chil-
dren of any node that it has already found. The idea of not knowing the
structure of the tree is that it cant, for example, assume a known bound on
the depth, counts of the number of nodes in subtrees, etc., without searching
through the tree to find this information directly.
Solution
Well use rejection sampling. The idea is to choose a node in the infinite
binary tree with probability proportional to d , and then repeat the process
if we picked a node that doesnt actually exist. Conditioned on finding a
depth(x)
node i that exists, its probability will be P depth(j) .
j
If we think of a node in the infinite tree as indexed by a binary strength
of length equal to its depth, we can generate it by first choosing the length
X and then choosing the bits in the string. We want Pr [X = n] to be
proportional to 2n n = (2/)n . Summing the geometric series gives
(2/)n
Pr [X = n] = .
1 (2/)
This is a geometric distribution, so we can sample it by repeatedly flipping

a biased coin. The number of such coin-flips is O(X), as is the number of
random bits we need to generate and the number of tree nodes we need to
check. The expected value of X is given by the infinite series

X (2/)n n
;
n=0
1 (2/)
this series converges to some constant by the ratio test.

So each probe costs O(1) time on average, and has at least a constant
1
probability of success, since we choose the root with Pr [X = 0] = 12/ .
Using Walds equation (8.5.1), the total expected time to run the algorithm
is O(1).
This is a little surprising, since the output of the algorithm may have
more than constant length. But we are interested in expectation, and when
> 2 most of the weight lands near the top of the tree.
C.6 Assignment 6: due Friday, 2013-04-26, at 17:00

C.6.1 Increasing subsequences
Let S1 , . . . , Sm be sets of indices in the range 1 . . . n. Say that a permutation
of 1 . . . n is increasing on Sj if (i1 ) < (i2 ) < < (ikj ) where i1 <
i2 < < ikj are the elements of Sj .
Given a fully polynomial-time randomized approximation scheme that
takes as input n and a sequence of sets S1 , . . . , Sm , and approximates the
number of permutations that are increasing on at least one of the Sj .
Solution
This can be solved using a fairly straightforward application of Karp-
Luby [KL85] (see 10.3). Recall that for Karp-Luby we need to be able to
express our target set U as the union of a polynomial number of covering sets
Uj , where we can both compute the size of each Uj and Psample uniformly from
it. We can then estimate |U | = j,xUj f (j, x) = j |Uj | Pr [f (j, x) = 1]
P
where f (j, x) is the indicator for the event that x 6 Uj 0 for any j 0 < j
and in the probability, the pair (j, x) is chosen uniformly at random from
{(j, x) | x Uj }.
In this case, let Uj be the set of all permutations that are increasing on
Sj . We can specify each such permutation by specifying the choice of which
kj = |Sj | elements are in positions i1 . . . ikj (the order of these elements
is determined by the requirement that the permutation be increasing on
Sj ) and specifying the order of the remaining n kj elements. This gives
n
kj (n kj )! = (n)nkj such permutations. Begin by computing these counts
for all Sj , as well as their sum.
We now wish to sample uniformly from pairs (j, ) where each is an
element of Sj . First sample each j with probability |Sj |/ ` |S` |, using the
P
counts weve already computed. Sampling a permutation uniformly from Sj

mirrors the counting argument: choose a kj -subset for the positions in Sj ,
then order the remaining elements randomly. The entire sampling step can
easily be done in O(n + m) time.
Computing f (j, ) requires testing to see if it is increasing for any Sj 0 for
j < j 0 ; without doing anything particularly intelligent, this takes O(nm) time.
So we can construct and test one sample in O(nm) time. Since each sample
has at least a = 1/m chance
f (j, ) = 1, from Lemma 10.2.1
of having
1 1 2 1
we need O 2 log = O m log samples to get relative error with

probability at least 1 , for a total cost of O m2 n2 log 1 .
C.6.2 Futile word searches

A word search puzzle consists of an nn grid of letters from some alphabet
, where the goal is to find contiguous sequences of letters in one of the eight
orthogonal or diagonal directions that form words from some lexicon. For
example, in Figure C.2, the left grid contains an instance of aabc (running
up and left from the rightmost a character on the last line), while the right
grid contains no instances of this word.
For this problem, you are asked to build an algorithm for constructing
bacab bacab
ccaac ccaac
bbbac babac
bbaaa bbaaa
acbab acbab
Figure C.2: Non-futile (left) and futile (right) word search grids for the
lexicon {aabc, ccca}
word search puzzles with no solution for a given lexicon. That is, given a set
of words S over some alphabet and a grid size n, the output should be an
n n grid of letters such that no word in S appears as a contiguous sequence
of letters in one of the eight directions anywhere in the grid. We will refer to
such puzzles as futile word search puzzles.
1. Suppose the maximum length of any word in S is k. Let pS be the

probability that some word in S is a prefix of an infinite string generated
by picking letters uniformly and independently from . Show that
there is a constant c > 0 such that for any k, , and S, pS < ck 2
implies that there exists, for all n, an n n futile word search puzzle
for S using only letters from .
2. Give an algorithm that constructs a futile word search puzzle given S

and n in expected time polynomial in |S|, k, and n, provided pS < ck 2
as above.
Solution
1. Well apply the symmetric version of the Lovsz local lemma. Suppose
the grid is filled in independently and uniformly at random with
characters from . Given a position ij in the grid, let Aij be the
event that there exists a word in S whose first character is at position
ij; observe that Pr [Aij ] 8pS by the union bound (this may be an
overestimate, both because we might run off the grid in some directions
and because the choice of initial character is not independent). Observe
also that Aij is independent of any event Ai0 j 0 where |i i0 | 2k 1 or
|j j 0 | 2k 1, because no two words starting at these positions can
overlap. So we can build a dependency graph with p 8pS and d
(4k3)2 . The Lovsz local lemma shows that there exists an assignment
where no Aij occurs provided ep(d + 1) < 1 or 8epS ((4k 3)2 + 1) < 1.
1
This easily holds if pS < 8e(4k 1 2 .
2 ) = 128e k
2. For this part, we can just use Moser-Tardos [MT10], particularly the
symmetric version described in Corollary 11.3.4. We have a collection
of m = O(n2 ) bad events, with d = (k 2 ), so the expected number of
resamplings is bounded by m/d = O(n2 /k 2 ). Each resampling requires
checking every position in the new grid for an occurrence of some string
in S; this takes O(n2 k |S|) time per resampling even if we are not
very clever about it. So the total expected cost is O(n4 |S|/k).
With some more intelligence, this can be improved. We dont need to
recheck any position at distance greater than k from any of the at most
k letters we resample, and if we are sensible, we can store S using a
radix tree or some similar data structure that allows us to look up all
words that appear as a prefix of a given length-k string in time O(k).
This reduces the cost of each resampling to O(k 3 ), with an additive
cost of O(k |S|) to initialize the data structure. So the total expected
cost is now O(n2 k + |S|).
C.6.3 Balance of power

Suppose you are given a set of n MMORPG players, and a sequence of
subsets S1 , S2 , . . . , Sm of this set, where each subset Si gives the players who
will participate in some raid. Before any of the raids take place, you are to
assign each player permanently to one of three factions. If for any i, |Si |/2
or more of the players are in the same faction, then instead of carrying out
the raid they will overwhelm and rob the other participants.
Give a randomized algorithm for computing a faction assignment that
prevents this tragedy from occurring (for all i) and thus allows all m raids to
be completed without incident, assuming that m > 1 and mini |Si | c ln m
for some constant c > 0 that does not depend on n or m. Your algorithm
should run in expected time polynomial in n and m.
Solution
Assign each player randomly to a faction. Let Xij be the number of players
in Si that are assigned to faction j. Then E [Xij ] = |Si |/3. Applying the
Chernoff bound (5.2.2), we have
1

Pr [Xij |Si |/2] = Pr Xij 1 + E [Xij ]
2
2 !
1
exp (|Si |/3) /3
2
= e|Si |/36 .
Let c = 3 36 = 108. Then if mini |Si | c ln m, for each i, j, it holds

that Pr [Xij |Si |/2] e3 ln m = m3 . So the probability that this bound
is exceeded for any i and j is at most (3m)m2 = 3/m2 . So a random
assignment works with at least 1/4 probability for m > 1.
We can generate and test each assignment in O(nm) time. So our
expected time is O(nm).
C.7 Final exam

C.7.1 Dominating sets

A dominating set in a graph is a subset S of the vertices for which every
vertex v is either in S or adjacent to a vertex in S.
Show that for any graph G, there is an aperiodic, irreducible Markov
chain on the dominating sets of G, such that (a) the transition rule for
the chain can be implemented in polynomial time; and (b) the stationary
distribution of the chain is uniform. (You do not need to say anything about
the convergence rate of this chain.)
Solution
Suppose we are have state St at time t. We will use a random walk where
we choose a vertex uniformly at random to add to or remove from St , and
carry out the action only if the resulting set is still a dominating set.
In more detail: For each vertex v, with probability 1/n, St+1 = St {v}
if v 6 St , St+1 = St \ {v} if v St and St \ {v} is a dominating set, and
St+1 = St otherwise. To implement this transition rule, we need to be able
to choose a vertex v uniformly at random (easy) and test in the case where
v St if St \ {v} is a dominating set (also polynomial: for each vertex, check
if it or one of its neighbors is in St \ {v}, which takes time O(|V | + |E|)).
Note that we do not need to check if St {v} is a dominating set.
For any pair of adjacent states S and S 0 = S \ {v} the probability of
moving from S to S 0 and the probability of moving from S 0 to S are both 1/n.
So the Markov chain is reversible with a uniform stationary distribution.
This is an aperiodic chain, because there exist minimal dominating sets
for which there is a nonzero chance that St+1 = St .
It is irreducible, because for any dominating set S, there is a path to
the complete set of vertices V by adding each vertex in V \ S one at a time.
Conversely, removing these vertices from V gives a path to S. This gives a
path S V T between any two dominating sets S and T .
C.7.2 Tricolor triangles

Suppose you wish to count the number of assignments of colors {r, g, b} to
nodes of a graph G that have the property that some triangle in G contains
all three colors. For example, the four-vertex graph shown below is labeled
with one of 18 such colorings (6 permutations of the colors of the triangle
nodes times 3 unconstrained choices for the degree-1 node).
r
g r
b
Give a fully polynomial-time randomized approximation scheme for this

problem.
Solution
Though it is possible, and tempting, to go after this using Karp-Luby (see
10.3), naive sampling is enough.
If a graph has at least one triangle (which can be checked in O(n3 )
time just by enumerating all possible triangles), then the probability that
that particular triangle is tricolored when colors are chosen uniformly and
independently at random is 6/27 = 2/9. This gives a constant hit rate , so
by Lemma 10.2.1, we can get relative error with 1 probability using

1
O 2
log 1 samples. Each sample costs O(n3 ) time to evaluate (again, brute-

force checking of all possible triangles), for a total cost of O n3 2 log 1 .
C.7.3 The n rooks problem

The n rooks problem requires marking as large a subset as possible of the
squares in an n n grid, so that no two squares in the same row or column
are marked.6
Consider the following randomized algorithm that attempts to solve this
problem:
1. Give each of the n2 squares a distinct label using a uniformly chosen

random permutation of the integers 1 . . . n2 .
2. Mark any square whose label is larger than any other label in its row
and column.
What is the expected number of marked squares?
Solution
Each square is marked if it is the largest of the 2n 1 total squares in its
row and column. By symmetry, each of these 2n 1 squares is equally likely
to be the largest, so the probability that a particular square is marked is
1
exactly 2n1 . By linearity of expectation, the total expected number of
n2
marked squares is then 2n1 .
C.7.4 Pursuing an invisible target on a ring

Suppose that you start at position 0 on a ring of size 2n, while a target
particle starts at position n. At each step, starting at position i, you can
choose whether to move to any of positions i 1, i, or i + 1. At the same
time, the target moves from its position j to either j 1 or j + 1 with equal
probability, independent of its previous moves or your moves. Aside from
knowing that the target starts at n at time 0, you cannot tell where the
target moves.
Your goal is to end up on the same node as the target after some step.7
Give an algorithm for choosing your moves such that, for any c > 0, you
6
This is not actually a hard problem.
7
Note that you must be in the same place at the end of the step: if you move from 1 to
2 while the target moves from 2 to 1, that doesnt count.
encounter the target in at most 2n steps with probability at least 1 nc

for sufficiently large n.
Solution
The basic idea is to just go through positions 0, 1, 2, . . . until we encounter
the target, but we have to be a little careful about parity to make sure we
dont pass it by accident.8
Let Xi = 1 be the increment of the targets i-th move, and let Si =
Pi
j=1 Xj , so that its position after i steps is n + Si mod 2n.
Let Yi be the position of the pursuer after i steps.
First move: stay at 0 if n is odd, move to 1 if n is even. The purpose of
this is to establish the invariant that n + Si Yi is even starting at i = 1.
For subsequent moves, let Yi+1 = Yi + 1. Observe that this maintains the
invariant.
We assume that n is at least 2. This is necessary to ensure that at time
1, Y1 n + S1 .
Claim: if at time 2n, Y2n n+S2n , then at some time i 2n, Yi = n+Si .
Proof: Let i be the first time at which Yi n + Si ; under the assumption
that n 2, i 1. So from the invariant, we cant have Yi = n + Si + 1,
and if Yi n + Si + 2, we have Yi1 Yi 1 n + Si + 1 n + Si1 ,
contradicting our assumption that i is minimal. The remaining alternative
is that Yi = n + Si , giving a collision at time i.
We now argue that Y2n n 1 is very likely to be at least n + S2n . Since
S2n is a sum of 2n independent 1 variables, from Hoeffdings inequality we
2
have Pr [Yn < n + S2n ] Pr [S2n n] en /4n = en/4 . For sufficiently
large n, this is much smaller than nc for any fixed c.
8
This is not the only possible algorithm, but there are a lot of plausible-looking
algorithms that turn out not to work. One particularly tempting approach is to run to
position n using the first n steps and then spend the next n steps trying to hit the target in
the immediate neighborhood of n, either by staying put (a sensible strategy when lost in the
woods in real life, assuming somebody is looking for you), or moving in a random walk of
some sort starting at n. This doesnt work if we want a high-probability bound. To see this,
observe that the target has a small but nonzero constant probability in the limit of begin

at some position greater than or equal to n + 4 n after exactly n/2 steps. Conditioned on

starting at n + 4 n or above, its chances of moving below n + 4 n 2 n = n + 2 n
4n/2(3n/2) 4/3
at any time in the next 3n/2 steps is bounded by e =e (Azuma), and a

similar bound holds independently for our chances of getting up to n + 2 n or above.
Multiplying out all these constants gives a constant probability of failure. A similar but
bigger disaster occurs if we dont rush to n first.
Appendix D

Spring 2011
D.1 Assignment 1: due Wednesday, 2011-01-26, at

17:00
D.1.1 Bureaucratic part
1. Your name.

etc.
anyway.)
D.1.2 Rolling a die

The usual model of a randomized algorithm assumes a source of fair, inde-
pendent random bits. This makes it easy to generate uniform numbers in the
range 0 . . . 2n 1, but not so easy for other ranges. Here are two algorithms
for generating a uniform random integer 0 s < n:
321
APPENDIX D. SAMPLE ASSIGNMENTS FROM SPRING 2011 322
Rejection sampling generates a uniform random integer 0 s <

2dlg ne . If s < n, return s; otherwise keep trying until you get some
s < n.
Arithmetic coding or range coding generates a sequence of bits

r1 , r2 , . . . rk until the half-open interval [ ki=1 2i ri , ki=1 2i ri +2k1 )
P P
is a subset of [s/n, (s + 1)/n) for some s; it then returns s.
1. Show that both rejection sampling and range coding produce a uniform
value 0 s < n using an expected O(log n) random bits.
2. Which algorithm has a better constant?
3. Does there exist a function f and an algorithm that produces a uniform

value 0 s < n for any n using f (n) random bits with probability 1?
Solution
1. For rejection sampling, each sample requires dlg ne bits and is accepted
with probability n/2dlg ne 1/2. So rejection sampling returns a value
after at most 2 samples on average, using no more than an expected
2dlg ne < 2(lg n + 1) expected bits for the worst n.
For range coding, we keep going as long as one of the n 1 nonzero end-
points s/n lies inside the current interval. After k bits, the probability
that one of the 2k intervals contains an endpoint is at most (n 1)2k ;
in particular, it drops below 1 as soon as k = 2dlg ne and continues to
drop by 1/2 for each additional bit, requiring 2 more bits on average.
So the expected cost of range coding is at most dlg ne + 2 < lg n + 3
bits.
2. Weve just shown that range coding beats rejection sampling by a

factor of 2 in the limit, for worst-case n. Its worth noting that other
factors might be more important if random bits are cheap: rejection
sampling is much easier to code and avoids the need for division.
3. There is no algorithm that produces a uniform value 0 s < n for all

n using any fixed number of bits. Suppose such an algorithm existed.
Fix some n. For all n values s to be equally likely, the sets of random
bits M 1 (s) = {r | M (r) = s} must have the same size. But this can
only happen if n divides 2f (n) , which works only for n a power of 2.
D.1.3 Rolling many dice

Suppose you repeatedly roll an n-sided die. Give an asymptotic (big-)
bound on the expected number of rolls until you roll some number you have
already rolled before.
Solution
In principle, it is possible to compute this value exactly, but we are lazy.
m
For a lower bound, observe that after m rolls, each of the 2 pairs of rolls
has probability 1/n of being equal, for an expected total of m 2 /n duplicates.

For m = n/2, this is less than 1/8, which shows that the expected number

of rolls is ( n).

For the upper bound, suppose we have already rolled the die n times.
If we havent gotten a duplicate already, each new roll has probability at

least n/n = 1/ n of matching a previous roll. So after an additional n
rolls on average, we get a repeat. This shows that the expected number of

rolls is O( n).

Combining these bounds shows that we need ( n) rolls on average.
D.1.4 All must have candy

A set of n0 children each reach for one of n0 candies, with each child choosing
a candy independently and uniformly at random. If a candy is chosen by
exactly one child, the candy and child drop out. The remaining n1 children
and candies then repeat the process for another round, leaving n2 remaining
children and candies, etc. The process continues until ever child has a candy.
Give the best bound you can on the expected number of rounds until
every child has a candy.
Solution
Let T (n) be the expected number of rounds remaining given we are starting
with n candies. We can set up a probabilistic recurrence relation T (n) =
1 + T (n Xn ) where Xn is the number of candies chosen by eactly one
child. It is easy to compute E [Xn ], since the probability that any candy gets
chosen exactly once is n(1/n)(1 1/n)n1 = (1 1/n)n1 . Summing over
all candies gives E [Xn ] = n(1 1/n)n1 .
The term (1 1/n)n1 approaches e1 in the limit, so for any fixed > 0,
we have n(1 1/n)n1 n(e1 ) for sufficiently large n. We can get a
quick bound by choosing so that e1 1/4 (for example) and then
applying the Karp-Upfal-Wigderson inequality (F.3.1) with (n) = n/4 to

get
Z n
1
E [T (n)] dt
1 t/4
= 4 ln n.
There is a sneaky trick here, which is that we stop if we get down to 1 candy
instead of 0. This avoids the usual problem with KUW and ln 0, by observing
that we cant ever get down to exactly one candy: if there were exactly one
candy that gets grabbed twice or not at all, then there must be some other
candy that also gets grabbed twice or not at all.
This analysis is sufficient for an asymptotic estimate: the last candy gets
grabbed in O(log n) rounds on average. For most computer-sciency purposes,
wed be done here.
We can improve the constant slightly by observing that (1 1/n)n1 is
in fact always greater than or equal to e1 . The easiest way to see this is
to plot the function, but if we want to prove it formally we can show that
(1 1/n)n1 is a decreasing function by taking the derivative of its logarithm:
d d
ln(1 1/n)n1 = (n 1) ln(1 1/n)
dn dn
n 1 1
= ln(1 1/n) + .
1 1/n n2
and observing that it is negative for n > 1 (we could also take the derivative
of the original function, but its less obvious that its negative). So if it
approaches e1 in the limit, it must do so from above, implying (11/n)n1
e1 .
This lets us apply (F.3.1) with (n) = n/e, giving E [T (n)] e ln n.
If we skip the KUW bound and use the analysis in F.4.2 instead, we get
that Pr [T (n) ln n + ln(1/)] . This suggests that the actual expected
value should be (1 + o(1)) ln n.

17:00
D.2.1 Randomized dominating set
A dominating set in a graph G = (V, E) is a set of vertices D such that
each of the n vertices in V is either in D or adjacent to a vertex in D.
Suppose we have a d-regular graph, in which every vertex has exactly d

neighbors. Let D1 be a random subset of V in which each vertex appears
with independent probability p. Let D be the union of D1 and the set of
all vertices that are not adjacent to any vertex in D1 . (This construction is
related to a classic maximal independent set algorithm of Luby [Lub85], and
has the desirable property in a distributed system of finding a dominating
set in only one round of communication.)
1. What would be a good value of p if our goal is to minimize E [|D|], and

what bound on E [|D|] does this value give?
2. For your choice of p above, what bound can you get on Pr [|D| E [|D|] t]?
Solution
1. First lets compute E [|D|]. Let Xv be the indicator for the event
that v D. Then Xv = 1 if either (a) v is in D1 , which occurs
with probability p; or (b) v and all d of its neighbors are not in D1 ,
which occurs with probability (1 p)d+1 . Adding these two cases gives
E [Xv ] = p + (1 p)d+1 and thus
X
E [|D|] = E [Xv ] = n p + (1 p)d+1 . (D.2.1)
v
We optimize E [|D|] in the usual way, by seeking a minimum for E [Xv ].

Differentiating with respect to p and setting to 0 gives 1 (d + 1)(1
p)d = 0, which we can solve to get p = 1 (d + 1)1/d . (We can easily
observe that this must be a minimum because setting p to either 0 or 1
gives E [Xv ] = 1.)
The value of E [|D|] for this value of p is the rather nasty expression
n(1 (d + 1)1/d + (d + 1)11/d ).
Plotting the d factor up suggests that it goes to ln d/d in the limit, and
both Maxima and www.wolframalpha.com agree with this. Knowing
the answer, we can prove it by showing
1 (d + 1)1/d + (d + 1)11/d 1 d1/d + d11/d

lim = lim
d ln d/d d ln d/d
1 d1/d
= lim
d ln d/d
1 e ln d/d
= lim
d ln d/d

1 1 ln d/d + O(ln2 d/d2 )
= lim
d ln d/d
ln d/d + O(ln2 d/d2 )
= lim
d ln d/d
= lim (1 + O(ln d/d))
d
= 1.
This lets us write E [|D|] = (1 + o(1))n ln d/d, where we bear in mind

that the o(1) term depends on d but not n.
2. Suppose we fix Xv for all but one vertex u. Changing Xu from 0 to

1 can increase |D| by at most one (if u wasnt already in D) and can
decrease it by at most d 1 (if u wasnt already in D and adding u to
D1 lets all d of us neighbors drop out). So we can apply the method
of bounded differences with ci = d 1 to get
!
t2
Pr [|D| E [|D|] t] exp .
2n(d 1)2
A curious feature of this bound is that it doesnt depend on p at all. It

may be possible to get a tighter bound using a better analysis, which

might pay off for very large d (say, d n).
D.2.2 Chernoff bounds with variable probabilities

Let X1 . . . Xn be a sequence of Bernoulli random variables, where for all i,
E [Xi | X1 . . . Xi1 ] pi . Let S = ni=1 Xi and = ni=1 pi . Show that, for
P P
all 0,
!
e
Pr [S (1 + )] .
(1 + )1+
Solution
Let St = ti=1 Xi , sohthati S = Sn , and let t = ti=1 pi . Well show by
P P
induction on t that E eSt exp(e1 t ), when > 0.

Compute
h i h i
E eS = E eSn1 eXn
h h ii
= E eSn1 E eXn X1 , . . . , Xn1

h i
= E eSn1 (Pr [Xn = 0 | X1 , . . . , Xn1 ] + e Pr [Xn = 1 | X1 , . . . , Xn1 ])
h i
= E eSn1 (1 + (e 1) Pr [Xn = 1 | X1 , . . . , Xn1 ])
h i
E eSn1 (1 + (e 1)pn )
h i
E eSn1 exp ((e 1)pn )
h i
E eSn1 exp ((e 1)pn )
exp (e 1)n1 ) exp ((e 1)pn )
= exp (e 1)n ) .
Now apply the rest of the proof of (5.2.1) to get the full result.
D.2.3 Long runs

Let W be a binary string of length n, generated uniformly at random. Define
a run of ones as a maximal sequence of contiguous ones; for example, the
string 11100110011111101011 contains 5 runs of ones, of length 3, 2, 6, 1,
and 2.
Let Xk be the number of runs in W of length k or more.
1. Compute the exact value of E [Xk ] as a function of n and k.

2. Give the best concentration bound you can for |Xk E [Xk ]|.
Solution
1. Well compute the probability that any particular position i = 1 . . . n is
the start of a run of length k or more, then sum over all i. For a run of
length k to start at position i, either (a) i = 1 and Wi . . . Wi+k1 are
all 1, or (b) i > 1, Wi1 = 0, and Wi . . . Wi+k1 are all 1. Assuming
n k, case (a) adds 2k to E [Xk ] and case (b) adds (n k)2k1 , for
a total of 2k + (n k)2k1 = (n k + 2)2k1 .
2. We can get an easy bound without too much cleverness using McDi-
armids inequality (5.3.11). Observe that Xk is a function of the inde-
pendent random variables W1 . . . Wn and that changing one of these bits
changes Xk by at most 1 (this can happen in several ways: a previous
run of length k1 can become a run of length k or vice versa, or two runs
of length k or more separated by a single zero may become a single
2 run,
t
or vice versa). So (5.3.11) gives Pr [|X E [X]| t] 2 exp 2n .
We can improve on this a bit by grouping the Wi together into blocks
of length `. If we are given control over a block of ` consecutive bits
and want to minimize the number of runs, we can either (a) make
all the bits zero, causing no runs to appear within the block and
preventing adjacent runs from extending to length k using bits from
the block, or (b) make all the bits one, possibly creating a new run but
possibly also causing two existing runs on either side of the block to
merge into one. In the first case, changing all the bits to one except
for a zero after every k consecutive ones creates at most b `+2k1k+1 c
new runs. Treating each of the dne` blocks as a2 single variable
then
t
gives Pr [|X E [X]| t] 2 exp 2dn/`e(b(`+2k1)/(k+1)c)2 . Staring
at plots of the denominator for a while suggests that it is minimized at
` = k + 3, the largest value with
b(` + 2k 1)/(k + 1)c 2. This gives
t2
Pr [|X E [X]| t] 2 exp 8dn/(k+3)e , improving the bound on t
p p
from ( n log(1/)) to ( (n/k) log(1/)).
For large k, the expectation of any individual Xk becomes small, so wed
expect that Chernoff bounds would work better on the upper bound
side than the method of bounded differences. Unfortunately, we dont
have independence. But from Problem D.2.2, we know that the usual
Chernoff bound works as long as we can show E [Xi | X1 , . . . , Xi1 ] pi
for some sequence of fixed bounds pi .
For X1 , there are no previous Xi , and we have E [X1 ] = 2k exactly.
For Xi with i > 1, fix X1 , . . . , Xi1 ; that is, condition on the event
Xj = xj for all j < i with some fixed sequence x1 , . . . , xi1 . Lets call
this event A. Depending on the particular values of the xj , its not
clear how conditioning on A will affect Xi ; but we can split on the
value of Wi1 to show that either it has no effect or Xi = 0:
E [Xi | A] = E [Xi | A, Wi1 = 0] Pr [Wi1 = 0 | A] + E [Xi | A, Wi1 = 1] Pr [Wi1 = 1 | A]
2k Pr [Wi1 = 0 | A] + 0
2k .
So we have pi 2k for all 1 i n k + 1.This gives =

k
2 (n k + 1), and Pr [X (1 + )] (1+)e (1+) .
h can set = 1 (since X cant

If we want a two-sided bound, we i
drop below 0 anyway, and get Pr |X E [X]| > 2k (n k + 1)
k
e 2 (nk+1)
4 . This is exponentially small for k = o(lg n). If k is much
bigger than lg n, then we have E [X] 1, so Markovs inequality alone
gives us a strong concentration bound.
However, in both cases, the bounds are competitive withp
the previous
bounds from McDiarmids inequality only if E [X] = O( n log(1/)).
So McDiarmids inequality wins for k = o(log n), Markovs inequality
wins for k = (log n), and Chernoff bounds may be useful for a small
interval in the middle.

17:00
D.3.1 Longest common subsequence
A common subsequence of two sequences v and w is a sequence u of length
k such that there exist indices i1 < i2 < < ik and j1 < j2 < < jk with
u` = vi` = wj` for all `. For example, ardab is a common subsequence of
abracadabra and cardtable.
Let v and w be words of length n over an alphabet of size n drawn
independently and uniformly at random. Give the best upper bound you
can on the expected length of the longest common subsequence of v and w.
Solution
Lets count the expectation of the number Xk of common subsequences of
length k. We have nk choices of positions in v, and nk choices of positions

in w; for each such choices, there is a probability of exactly nk that the

corresponding positions match. This gives

!2
n
E [Xk ] = nk
k
nk
< .
(k!)2
nk
<
(k/e)2k
!k
ne2
= .
k2
Wed like this bound to be substantially less than 1. We cant reasonably

expect this to happen unless the base of the exponent is less than 1, so we

need k > e n.

If k = (1 + )e n for any > 0, then E [Xk ] < (1 + )2e n < n1 for
sufficiently large n. It follows that the expected length of the longest common

subsequent is at most (1 + )e n for sufficiently large n (because if there
are no length-k subsequences, the longest subsequence has length at most
k 1, and if there is at least one, the longest has length at most n; this gives
a bound of at most (1 1/n)(k 1) + (1/n)n < k). So in general we have

the length of the longest common subsequence is at most (1 + o(1))e n.
Though it is not required by the problem, here is a quick argument that

the expected length of the longest common subsequence is ( n), based
on the Erds-Szekeres theorem [ES35].1 The Erds-Szekeres theorem
says that any permutation of n2 + 1 elements contains either an increasing
sequence of n + 1 elements or a decreasing sequence of n + 1 elements. Given
two random sequences of length n, let S be the set of all elements that appear
in both, and consider two permutations and of S corresponding to the
order in which the elements appear in v and w, respectively (if an element
appears multiple times, pick one of the occurrences at random). Then the
Erds-Szekeres
p theorem says that contains a sequence of length at least
b || 1c that is either increasing or decreasing with respect to the order
given by ; by symmetry, the probability that it is increasing is at least 1/2.
This gives
hp an expected
i value for the longest common subsequence that is at
least E || 1 /2.
Let X = ||. We can compute a lower bound E [X] easily; each possible el-
ement fails to occur in v with probability (11/n)n e1 , and similarly for w.
1
As suggested by Benjamin Kunsberg.
So the chance that an element appears in both sequences is at 1 2

h i least (1e ) ,
and thus E [X] n(1 e1 )2 . What we want is E X 1 /2; but here the
h i p
fact that x is concave means that E X 1 E [X 1] by Jensens
hp i p
inequality (4.3.1). So we have E || 1 /2 n(1 e1 )2 1/2
1e 1
2 n.
This is not a very good bound (empirically, the real bound seems to be

in the neighborhood of 1.9 n when n = 10000), but it shows that the upper

bound of (1 + o(1))e n is tight up to constant factors.
D.3.2 A strange error-correcting code

Let be an alphabet of size m + 1 that includes m non-blank symbols
and a special blank symbol. Let S be a set of nk strings of length n with

non-blank symbols in exactly k positions each, such that no two strings in S

have non-blank symbols in the same k positions.
For what value of m can you show S exists such that no two strings in S
have the same non-blank symbols in k 1 positions?
Solution
This is a job for the Lovsz Local Lemma. And its even symmetric, so we
can use the symmetric version (Corollary 11.3.2).
Suppose we assign the non-blank symbols to each string uniformly and
independently at random. For each A S with |A| = k, let XA be the string
that has non-blank symbols in all positions in A. For each pair of subsets
A, B with |A| = |B| = k and |A B| = k 1, let CA,B be the event that XA
and XB are identical on all positions in A B. Then Pr [CA,B ] = mk+1 .
We now now need to figure out how many events are in each neighborhood
(CA,B ). Since CA,B depends only on the choices of values for A and B, it
is independent of any events CA0 ,B 0 where neither of A0 or B 0 is equal to A
or B. So we can make (CA,B ) consist of all events CA,B 0 and CA0 ,B where
B 0 6= B and A0 6= A.
For each fixed A, there are exactly (n k)k events B that overlap it in
k 1 places, because we can specify B by choosing the elements in B \ A
and A \ B. This gives (n k)k 1 events CA,B 0 where B 0 6= B. Applying
the same argument for A0 gives a total of d = 2(n k)k 2 events in
(CA,B ). Corollary 11.3.2 applies if ep(d + 1) 1, which in this case means
em(k1) (2(n k)k 1) 1. Solving for m gives
m (2e(n k)k 1)1/(k1) . (D.3.1)
For k n, the (n k)1/(k1) n1/(k1) term dominates the shape of

the right-hand side asymptotically as k gets large, since everything else goes
to 1. This suggests we need k = (log n) to get m down to a constant.
Note that (D.3.1) doesnt work very well when k = 1.2 For the k = 1
case, there is no overlap in non-blank positions between different strings, so
m = 1 is enough.
D.3.3 A multiway cut

Given a graph G = (V, E), a 3-way cut is a set of edges whose endpoints
lie in different parts of a partition of the vertices V into three disjoint parts
ST U =V.
1. Show that any graph with m edges has a 3-way cut with at least 2m/3
edges.
2. Give an efficient deterministic algorithm for finding such a cut.
Solution
1. Assign each vertex independently to S, T , or U with probability 1/3
each. Then the probability that any edge uv is contained in the cut is
exactly 2/3. Summing over all edges gives an expected 2m/3 edges.
2. Well derandomize the random vertex assignment using the method of

conditional probabilities. Given a partial assignment of the vertices,
we can compute the conditional expectation of the size of the cut
assuming all other vertices are assigned randomly: each edge with
matching assigned endpoints contributes 0 to the total, each edge
with non-matching assigned endpoints contributes 1, and each edge
with zero or one assigned endpoints contributes 2/3. Well pick values
for the vertices in some arbitrary order to maximize this conditional
expectation (since our goal is to get a large cut). At each step, we need
only consider the effect on edges incident to the vertex we are assigning
whose other endpoints are already assigned, because the contribution
of any other edge is not changed by the assignment. Then maximizing
2
Thanks to Brad Hayes for pointing this out.
the conditional probability is done by choosing an assignment that

matches the assignment of the fewest previously-assigned neighbors:
in other words, the natural greedy algorithm works. The cost of this
algorithm is O(n + m), since we loop over all vertices and have to check
each edge at most once for each of its endpoints.

17:00
D.4.1 Sometimes successful betting strategies are possible
You enter a casino with X0 = a dollars, and leave if you reach 0 dollars or
b or more dollars, where a, b N. The casino is unusual in that it offers
arbitrary fair games subject to the requirements that:
Any payoff resulting from a bet must be a nonzero integer in the range
Xt to Xt , inclusive, where Xt is your current wealth.
The expected payoff must be exactly 0. (In other words, your assets
Xt should form a martingale sequence.)
For example, if you have 2 dollars, you may make a bet that pays off 2
with probability 2/5, +1 with probability 2/5 and +2 with probability 1/5;
but you may not make a bet that pays off 3, +3/2, or +4 under any
circumstances, or a bet that pays off 1 with probability 2/3 and +1 with
probability 1/3.
1. What strategy should you use to maximize your chances of leaving

with at least b dollars?
2. What strategy should you use to maximize your changes of leaving

with nothing?
3. What strategy should you use to maximize the number of bets you
make before leaving?
Solution
1. Let Xt be your wealth at time t, and let be the stopping time when
you leave. Because {Xt } is a martingale, E [X0 ] = a = E [X ] =
Pr [X b] E [X | X b]. So Pr [X b] is maximized by making
E [X | X b] as small as possible. It cant be any smaller than b,
which can be obtained exactly by making only 1 bets. The gives a

probability of leaving with b of exactly a/b.
2. Here our goal is to minimize Pr [X b], so we want to make E [X | X b]

as large as possible. The largest value of X we can possibly reach
is 2(b 1); we can obtain this value by betting 1 until we reach
b 1, then making any fair bet with positive payoff b 1 (for example,
(b 1) with equal probability works, as does a bet that pays off b 1
with probability 1/b and 1 with probability (b 1)/b). In this case
a
we get a probability of leaving with 0 of 1 2(b1) .
3. For each t, let t = Xt Xt1 and Vt = Var[t | Ft ]. We have previously

shown (see the footnote to 8.5.1) that E X = E X0 + E [ t=1 Vt ]
2
2 P
where is an appropriate stopping time. When we stop, we know

that X2 (2(b 1))2 , which puts an upper bound on E [ i=1 Vi ]. We
P
can spend this bound most parsimoniously by minimizing Vi as much

as possible. If we make each t = 1, we get the smallest possible
value for Vt (since any change contributes at least 1 to the variance).
However, in this case we dont get all the way out to 2(b 1) at the
high end; instead, we stop at b, giving an expected number of steps
equal to a(b a).
We can do a bit better than this by changing our strategy at b 1.
Instead of betting 1, lets pick some x and place a bet that pays off
b1 with probability 1b and 1 with probability b1 1
b = 1 b . (The idea
here is to minimize the conditional variance while still allowing ourselves
to reach 2(b 1).) Each ordinary random walk step has Vt = 1; a big
2 2
bet starting at b 1 has Vt = 1 1b + (b1)
b = b1+b b2b+1 = b 1b .
To analyze this process, observe that starting from a, we first spending
a
a(ba1) steps on average to reach either 0 (with probability 1 b1 or
a
b1 (with probability b1 . In the first case, we are done. Otherwise, we
take one more step, then with probability 1b we lose and with probability
b1
b we continue starting from b 2. We can write a recurrence for our
expected number of steps T (a) starting from a, as:
a b1

T (a) = a(b a 1) + 1+ T (b 2) . (D.4.1)
b1 b
When a = b 2, we get
b2 b1

T (b 2) = (b 2) + 1+ T (b 2)
b1 b
1 b2

= (b 2) 1 + + T (b 2),
b1 b
which gives
(b 2) 2b1
b1
T (b 2) =
2/b
b(b 2)(2b 1)
= .
2(b 1)
Plugging this back into (D.4.1) gives
a b 1 b(b 2)(2b 1)

T (a) = a(b a 1) + 1+
b1 b 2(b 1)
a a(b 2)(2b 1)
= ab a2 + a + +
b1 2(b 1)
3
= ab + O(b). (D.4.2)
2
This is much better than the a(b a) value for the straight 1 strategy,
especially when a is also large.
I dont know if this particular strategy is in fact optimal, but thats
what Id be tempted to bet.
D.4.2 Random walk with reset

Consider a random walk on N that goes up with probability 1/2, down with
probability 3/8, and resets to 0 with probability 1/8. When Xt > 0, this
gives:

Xt + 1with probability 1/2,

Xt+1 = Xt 1 with probability 3/8, and

0 with probability 1/8.
When Xt = 0, we let Xt+1 = 1 with probability 1/2 and 0 with probability

1/2.
1. What is the stationary distribution of this process?

2. What is the mean recurrence time n for some state n?
3. Use n to get a tight asymptotic (i.e., big-) bound on 0,n , the

expected time to reach n starting from 0.
Solution
1
1. For n > 0, we have n = 2 n1 + 38 n+1 , with a base case 0 =
1 3 3
8 + 8 0 + 8 1 .
The n expression is a linear homogeneous recurrence, so its solution
consists of linear combinations of terms bn , where b satisfies 1 =
1 1
2b + 38 b. The solutions to this equation are b = 2/3 and b = 2; we
can exclude the b = 2 case because it would make our probabilities
blow up for large n. So we can reasonably guess n = a(2/3)n when
n > 0.
1
For n = 0, substitute 0 = 8 + 38 0 + 38 a(2/3) to get 0 = 1
5 + 25 a. Now
substitute
1 = (2/3)a
1 3
= 0 + a(2/3)2
2 8
1 1 2 3
= + a + a(2/3)2
2 5 5 8
1 11
= + a,
10 30
which we can solve to get a = 1/3.
So our candidate is 0 = 1/3, n = (1/3)(2/3)n , and in fact we can
drop the special case for 0 .
Pn Pn n 1/3
As a check, i=0 n = (1/3) i=0 (2/3) = 12/3 = 1.
2. Since n = 1/n , we have n = 3(3/2)n .
3. In general, let k,n be the expected time to reach n starting at k. Then

n = n,n = 1 + 18 0,n + 12 n+1,n + 38 n1,n 1 + 0,n /8. It follows
that 0,n 8n + 1 = 24(3/2)n + 1 = O((3/2)n ).
For the lower bound, observe that n n,0 + 0,n . Since there is a
1/8 chance of reaching 0 from any state, we have n,0 8. It follows
that n 8 + 0,n or 0,n n 8 = ((3/2)n ).
D.4.3 Yet another shuffling algorithm

Suppose we attempt to shuffle a deck of n cards by picking a card uniformly
at random, and swapping it with the top card. Give the best bound you can
on the mixing time for this process to reach a total variation distance of
from the uniform distribution.
Solution
Its tempting to use the same coupling as for move-to-top (see 9.3.3). This
would be that at each step we choose the same card to swap to the top
position, which increases by at least one the number of cards that are in the
same position in both decks. The problem is that at the next step, these two
cards are most likely separated again, by being swapped with other cards in
two different positions.
Instead, we will do something slightly more clever. Let Zt be the number
of cards in the same position at time t. If the top cards of both decks are
equal, we swap both to the same position chosen uniformly at random. This
has no effect on Zt . If the top cards of both decks are not equal, we pick
a card uniformly at random and swap it to the top in both decks. This
increases Zt by at least 1, unless we happen to pick cards that are already in
the same position; so Zt increases by at least 1 with probability 1 Zt /n.
Lets summarize a state by an ordered pair (k, b) where k = Zt and b is
0 if the top cards are equal and 1 if they are not equal. Then we have a
Markov chain where (k, 0) goes to (k, 1) with probability nk n (and otherwise
stays put); and (k, 1) goes to (k + 1, 0) (or higher) with probability nk
n and
k
to (k, 0) with probability n .
n
Starting from (k, 0), we expect to wait nk steps on average to reach
(k, 1), at which point we move to (k + 1, 0) or back to (k, 0) in one more step;
n
we iterate through this process nk times on average before we are successful.
This gives an expected number of steps to get from (k, 0) to (k + 1, 0) (or
n n
possibly a higher value) of nk nk + 1) . Summing over k up to n 2
(since once k > n 2, we will in fact have k = n, since k cant be n 1), we
get
n2
n n
X
E [ ] +1
k=0
nk nk
n
!
X n2 n
= +
m=2
m2 m
!
2 2
n 1 + n ln n.
6
= O(n2 ).
So we expect the deck to mix in O(n2 log(1/)) steps. (I dont know if
this is the real bound; my guess is that it should be closer to O(n log n) as
in all the other shuffling procedures.)
D.5 Assignment 5: due Thursday, 2011-04-07, at

23:59
D.5.1 A reversible chain
Consider a random walk on Zm , where pi,i+1 = 2/3 for all i and pi,i1 = 1/3
for all i except i = 0. Is it possible to assign values to p0,m1 and p0,0 to
make this chain reversible, and if so, what stationary distribution do you
get?
Solution
Suppose we can make this chain reversible, and let be the resulting
stationary distribution. From the detailed balance equations, we have
(2/3)i = (1/3)i+1 or i+1 = 2i for i = 0 . . . m 2. The solution to
i
this recurrence is i = 2i 0 , which gives i = 2m21 when we set 0 to get
P
i i = 1.
Now solve 0 p0,m1 = m1 pm1,0 to get
m1 pm1,0
p0,m1 =
0
= 2m1 (2/3)
= 2m /3.
This is greater than 1 for m > 1, so except for the degenerate cases of
m = 1 and m = 2, its not possible to make the chain reversible.
D.5.2 Toggling bits

Consider the following Markov chain on an array of n bits a[1], a[2], . . . a[n].
At each step, we choose a position i uniformly at random. We then change
A[i] to A[i] with probability 1/2, provided i = 1 or A[i 1] = 1 (if neither
condition holds hold, do nothing).3
1. What is the stationary distribution?
2. How quickly does it converge?
Solution
1. First lets show irreducibility. Starting from an arbitrary configuration,
repeatedly switch the leftmost 0 to a 1 (this is always permitted
by the transition rules); after at most n steps, we reach the all-1
configuration. Since we can repeat this process in reverse to get to any
other configuration, we get that every configuration is reachable from
every other configuration in at most 2n steps (2n 1 if we are careful
about handling the all-0 configuration separately).
We also have that for any two adjacent configurations x and y, pxy =
1
pyx = 2n . So we have a reversible, irreducible, aperiodic (because
there exists at least one self-loop) chain with a uniform stationary
distribution x = 2n .
2. Here is a bound using the obvious coupling, where we choose the same
position in X and Y and attempt to set it to the same value. To
show this coalesces, given Xt and Yt define Zt to be the position of the
rightmost 1 in the common prefix of Xt and Yt , or 0 if there is no 1 in
the common prefix of Xt and Yt . Then Zt increases by at least 1 if we
1
attempt to set position Zt + 1 to 1, which occurs with probability 2n ,
and decreases by at most 1 if we attempt to set Zt to 0, again with
1
probability 2n .
It follows that Zt reaches n no later than a 1 random walk on 0 . . . n
with reflecting barriers that takes a step every 1/n time units on
average. The expected number of steps to reach n from the worst-case
3
Motivation: Imagine each bit represents whether a node in some distributed system
is inactive (0) or active (1), and you can only change your state if you have an active
left neighbor to notify. Also imagine that there is an always-active base station at 1
(alternatively, imagine that this assumption makes the problem easier than the other
natural arrangement where we put all the nodes in a ring).
starting position of 0 is exactly n2 . (Proof: model the random walk

with a reflecting barrier at 0 by folding a random walk with absorbing
barriers at n in half, then use the bound from 8.5.1.) We must then
multiply this by n to get an expected n3 steps in the original process.
So the two copies coalesce in at most n3 expected steps. My suspicion
is one could improve this bound with a better analysis by using the
bias toward increasing Zt to get the expected time to coalesce down to
O(n2 ), but I dont know any clean way to do this.
The path coupling version of this is that we look at two adjacent
configurations Xt and Yt , use the obvious coupling again, and see
what happens to E [d(Xt+1 , Yt+1 ) | Xt , Yt ], where the distance is the
number of transitions needed to convert Xt to Yt or vice versa. If
we pick the position i where Xt and Yt differ, then we coalesce; this
occurs with probability 1/n. If we change the 1 to the left of i to
a 0, then d(Xt+1 , Yt+1 ) rises to 3 (because to get from Xt+1 to Yt+1 ,
we have to change position i 1 to 1, change position i, and then
change position i 1 back to 0); this occurs with probability 1/2n if
i > 1. But we can also get into trouble if we try to change position
i + 1; we can only make the change in one of Xt and Yt , so we get
d(Xt+1 , Yt+1 ) = 2 in this case, which occurs with probability 1/2n
when i < n. Adding up all three cases gives a worst-case expected
change of 1/n + 2/2n + 1/2n = 1/2n > 0. So unless we can do
something more clever, path coupling wont help us here.
However, it is possible to get a bound using canonical paths, but the
best bound I could get was not as good as the coupling bound. The
basic idea is that we will change x into y one bit at a time (from left
to right, say), so that we will go through a sequence of intermediate
states of the form y[1]y[2] . . . y[i]x[i + 1]x[i + 2] . . . x[n]. But to change
x[i + 1] to y[i + 1], we may also need to reach out with a tentacle of 1
bits from from the last 1 in the current prefix of y (and then retract it
afterwards). Given a particular transition where we change a 0 to a 1,
we can reconstruct the original x and y by specifying (a) which bit i at
or after our current position we are trying to change; (b) which 1 bit
before our current position is the last real 1 bit in y as opposed to
something we are creating to reach out to position i; and (c) the values
of x[1] . . . x[i 1] and y[i + 1] . . . y[i]. A similar argument applies to
1 0 transitions. So we are routing at most n2 2n1 paths across each
transition, giving a bound on the congestion

1

n2 2n1 22n
2n /2n
= n3 .
The bound on 2 that follows from this is 8n6 , which is pretty bad
(although the constant could be improved by counting the (a) and
(b) bits more carefully). As with the coupling argument, it may be
that there is a less congested set of canonical paths that gives a better
bound.
D.5.3 Spanning trees

Suppose you have a connected graph G = (V, E) with n nodes and m edges.
Consider the following Markov process. Each state Ht is a subgraph of G
that is either a spanning tree or a spanning tree plus an additional edge. At
each step, flip a fair coin. If it comes up heads, choose an edge e uniformly
at random from E and let Ht+1 = Ht {e} if Ht is a spanning tree and let
Ht+1 = Ht \ {e} if Ht is not a spanning tree and Ht \ {e} is connected. If it
comes up tails and Ht is a spanning tree, let Ht+1 be some other spanning
tree, sampled uniformly at random. In all other cases, let Ht+1 = Ht .
Let N be the number of states in this Markov chain.
1. What is the stationary distribution?
2. How quickly does it converge?
Solution
1. Since every transition has a matching reverse transition with the same
transition probability, the chain is reversible with a uniform stationary
distribution H = 1/N .
2. Heres a coupling that coalesces in at most 4m/3 + 2 expected steps:
(a) If Xt and Yt are both trees, then send them to the same tree
with probability 1/2; else let them both add edges independently
(or we could have them add the same edgeit doesnt make any
difference to the final result).
(b) If only one of Xt and Yt is a tree, with probability 1/2 scramble
the tree while attempting to remove an edge from the non-tree,
and the rest of the time scramble the non-tree (which has no
effect) while attempting to add an edge to the tree. Since the
non-tree has at least three edges that can be removed, this puts
(Xt+1 , Yt+1 in the two-tree case with probability at least 3/2m.
(c) If neither Xt nor Yt is a tree, attempt to remove an edge from
both. Let S and T be the sets of edges that we can remove from
Xt and Yt , respectively, and let k = min(|S|, |T |) 3. Choose k
edges from each of S and T and match them, so that if we remove
one edge from each pair, we also remove the other edge. As in
the previous case, this puts (Xt+1 , Yt+1 in the two-tree case with
probability at least 3/2m.
To show this coalesces, starting from an arbitrary state, we reach

a two-tree state in at most 2m/3 expected steps. After one more
step, we either coalesce (with probability 1/2) or restart from a new
arbitrary state. This gives an expected coupling time of at most
2(2m/3 + 1) = 4m/3 + 2 as claimed.
D.6 Assignment 6: due Monday, 2011-04-25, at

17:00
D.6.1 Sparse satisfying assignments to DNFs
Given a formula in disjunctive normal form, wed like to estimate the number
of satisfying assignments in which exactly w of the variables are true. Give
a fully polynomial-time randomized approximation scheme for this problem.
Solution
Essentially, were going to do the Karp-Luby covering trick [KL85] described
in 10.3, but will tweak the probability distribution when we generate our
samples so that we only get samples with weight w.
n
Let U be the set of assignment with weight w (there are exactly w such
assignments, where n is the number of variables). For each clause Ci , let
Ui = {x U | Ci (x) = 1}. Now observe that:
|Ui |. Let ki = |Ci | be the number of variables in Ci

1. We can compute

+ +
and ki = Ci the number of variables that appear in positive form
nk
in Ci . Then |Ui | = wk+i is the number of ways to make a total of w
i
variables true using the remaining n ki variables.
2. We can sample uniformly from Ui , by sampling a set of w ki+ true

variables not in Ci uniformly from all variables not in Ci .
3. We can use the values computed for |Ui | to sample i proportionally to
the size of |Ui |.
So now we sample pairs (i, x) with x Ui uniformly at random by
sampling i first, then sampling x Ui . As in the original algorithm, we then
count (i, x) if and only if Ci is the leftmost clause for which Ci (x) = 1. The
same argument that at least 1/m of the (i, x) pairs count applies, and so we
get the same bounds as in the original algorithm.
D.6.2 Detecting duplicates

Algorithm D.1 attempts to detect duplicate values in an input array S of
length n.
1 Initialize A[1 . . . n] to
2 Choose a hash function h
3 for i 1 . . . n do
4 x S[i]
5 if A[h(x)] = x then
6 return true
7 else
8 A[h(x)] x
9 return false
Algorithm D.1: Dubious duplicate detector
Its easy to see that Algorithm D.1 never returns true unless some value
appears twice in S. But maybe it misses some duplicates it should find.
1. Suppose h is a random function. What is the worst-case probability
that Algorithm D.1 returns false if S contains two copies of some
value?
2. Is this worst-case probability affected if h is drawn instead from a
2-universal family of hash functions?
Solution
1. Suppose that S[i] = S[j] = x for i < j. Then the algorithm will see x in
A[h(x)] on iteration j and return true, unless it is overwritten by some
value S[k] with i < k < j. This occurs if h(S[k]) = h(x), which occurs
with probability exactly 1 (1 1/n)ji1 if we consider all possible k.
This quantity is maximized at 1 (1 1/n)n2 1 (1 1/n)2 /e
1 (1 1/2n)/e when i = 1 and j = n.
2. As it happens, the algorithm can fail pretty badly if all we know is that
h is 2-universal. What we can show is that the probability that some
S[k] with i < j < k gets hashed to the same place as x = S[i] = S[j]
in the analysis above is at most (j i 1)/n (n 2)/n = (1 2/n),
since each S[k] has at most a 1/n chance of colliding with x and the
union bound applies. But it is possible to construct a 2-universal family
for which we get exactly this probability in the worst case.
Let U = {0 . . . n}, and define for each a in {0 . . . n 1} ha : U n by
ha (n) = 0 and ha (x) = (x + a) mod n for x 6= n. Then H = {ha } is 2-
universal, since if x 6= y and neither x nor y is n, Pr [ha (x) = ha (y)] = 0,
and if one of x or y is n, Pr [ha (x) = ha (y)] = 1/n. But if we use this
family in Algorithm D.1 with S[1] = S[n] = n and S[k] = k for
1 < k < n, then there are n 2 choices of a that put one of the middle
values in A[0].
D.6.3 Balanced Bloom filters

A clever algorithmist decides to solve the problem of Bloom filters filling up
with ones by capping the number of ones at m/2. As in a standard Bloom fil-
ter, an element is inserted by writing ones to A[h1 (x)], A[h2 (x)], . . . , A[hk (x)];
but after writing each one, if the number of one bits in the bit-vector is more
than m/2, one of the ones in the vector (chosen uniformly at random) is
changed back to a zero.
Because some of the ones associated with a particular element might
be deleted, the membership test answers yes if at least 3/4 of the bits
A[h1 (x)] . . . A[hk (x)] are ones.
To simplify the analysis, you may assume that the hi are independent
random functions. You may also assume that (3/4)k is an integer.
1. Give an upper bound on the probability of a false positive when testing

for a value x that has never been inserted.
2. Suppose that we insert x at some point, and then follow this insertion
with a sequence of insertions of new, distinct values y1 , y2 , . . . . Assum-
ing a worst-case state before inserting x, give asymptotic upper and
lower bounds on the expected number of insertions until a test for x

fails.
Solution
1. The probability of a false positive is maximized when exactly half the
bits in A are one. If x has never been inserted, each A[hi (x)] is equally
likely to be zero or one. So Pr [false positive for x] = Pr [Sk (3/4)k]
when Sk is a binomial random variable with parameters 1/2 and k.
Chernoff bounds give
Pr [Sk (3/4)k] = Pr [Sk (3/2) E [Sk ]]

!k/2
e1/2

(3/2)3/2
(0.94734)k .
We can make this less than any fixed by setting k 20 ln(1/) or

thereabouts.
2. For false negatives, we need to look at how quickly the bits for x are
eroded away. A minor complication is that the erosion may start even
as we are setting A[h1 (x)] . . . A[hk (x)].
Lets consider a single bit A[i] and look at how it changes after (a)
setting A[i] = 1, and (b) setting some random A[r] = 1.
In the first case, A[i] will be 1 after the assignment unless it is set back
1
to zero, which occurs with probability m/2+1 . This distribution does
not depend on the prior value of A[i].
In the second case, if A[i] was previously 0, it becomes 1 with probability
1 1 1 m/2

1 =
m m/2 + 1 m m/2 + 1
1
= .
m+2
If it was previously 1, it becomes 0 with probability

1 1 1
= .
2 m/2 + 1 m+2
So after the initial assignment, A[i] just flips its value with probability
1
m+2 .
It is convenient to represent A[i] as 1; let Xit = 1 if A[i] = 0 at time
t, and 1 otherwise. Then Xit satisfies the recurrence
h i m+1 t 1
E Xit+1 Xit = X Xt

m+2 i m+2 i
m
= Xit .
m + 2
2

= 1 Xit .
m+2
t
We can extend this to E Xit Xi0 = 1 2
Xi0 e2t/(m+2) Xi0 .

m+2
1
Similarly, after setting A[i] = 1, we get E [Xi ] = 1 2 m/2+1 = 1
4
2m+1 = 1 o(1).
Let S t = ki=1 Xht i (x) . Let 0 be the time at which we finish inserting x.
P
Then each for each i we have

h i
1 o(1)e2k/(m+2) E Xh0i (x) 1 o(1),
from which it follows that

h i
k(1 o(1))e2k/(m+2) E S 0 1 o(1)
and in general that

h i
k(1 o(1))e2(k+t)/(m+2) E S t 1 o(1)e2t/(m+2) .
So for any fixed 0 < < 1 and sufficiently large m, we will have
E S = k for some t0 where t t0 k + t and t = (m ln(1/)).
t

We are now looking for the time at which S t drops below k/2 (the k/2
is because we are working with 1 variables). We will bound when
this time occurs using Markovs inequality.
t with E S t (3/4)k. Then E k S t

Lets look for the largest time
k/4 and Pr k S t k/2 1/2. It follows that after (m) k op-

erations, x is still visible with probability 1/2, which implies that the
expected time at which it stops being visible is at least ((m) k)/2.
To get the expected number of insert operations, we divide by k, to
get (m/k).
For the upper bound, apply the same reasoning to the first time at which
t
E S k/4. This occurs at time O(m) at the latest (with a different
constant), so after O(m) steps there is at most a 1/2 probability that
S t k/2. If S t is still greater than k/2 at this point, try again using
the same analysis; this gives us the usual geometric series argument
that E [t] = O(m). Again, we have to divide by k to get the number of
insert operations, so we get O(m/k) in this case.
Combining these bounds, we have that x disappears after (m/k)
insertions on average. This seems like about what we would expect.
D.7 Final exam

D.7.1 Leader election

Suppose we have n processes and we want to elect a leader. At each round,
each process flips a coin, and drops out if the coin comes up tails. We win if
in some round there is exactly one process left.
Let T (n) be the probability that this event eventually occurs starting
with n processes. For small n, we have T (0) = 0 and T (1) = 1. Show that
there is a constant c > 0 such that T (n) c for all n > 1.
Solution
Lets suppose that there is some such c. We will necessarily have c 1 = T (1),
so the induction hypothesis will hold in the base case n = 1.
For n 2, compute
n
!
n
2n
X
T (n) = T (k)
k=0
k
n1
!
n
= 2n T (n) + 2n nT (1) + 2n
X
T (k)
k=2
k
2n T (n) + 2n n + 2n (2n n 2)c.
Solve for T (n) to get
n + (2n n 2)c
T (n)
2n 1
n
2 n 2 + n/c

=c .
2n 1
This will be greater than or equal to c if 2n n 2 + n/c 2n 1 or

n
n/c n + 1, which holds if c n+1 . The worst case is n = 2 giving c = 2/3.
Valiant and Vazirani [VV86] used this approach to reduce solving general
instances of SAT to solving instances of SAT with unique solutions; they
prove essentially the result given above (which shows that fixing variables in
a SAT formula is likely to produce a SAT formula with a unique solution at
some point) with a slightly worse constant.
D.7.2 Two-coloring an even cycle

Here is a not-very-efficient algorithm for 2-coloring an even cycle. Every node
starts out red or blue. At each step, we pick one of the n nodes uniformly
at random, and change its color if it has the same color as at least one of
its neighbors. We continue until no node has the same color as either of its
neighbors.
Suppose that in the initial state there are exactly two monochromatic
edges. What is the worst-case expected number of steps until there are no
monochromatic edges?
Solution
Suppose we have a monochromatic edge surrounded by non-monochrome
edges, e.g. RBRRBR. If we pick one of the endpoints of the edge (say
the left endpoint in this case), then the monochromatic edge shifts in the
direction of that endpoint: RBBRBRB. Picking any node not incident to a
monochromatic edge has no effect, so in this case there is no way to increase
the number of monochromatic edges.
It may also be that we have two adjacent monochromatic edges: BRRRB.
Now if we happen to pick the node in the middle, we end up with no
monochromatic edges (BRBRB) and the process terminates. If on the other
hand we pick one of the nodes on the outside, then the monochromatic edges
move away from each other.
We can thus model the process with 2 monochromatic edges as a random
walk, where the difference between the leftmost nodes of the edges (mod
n) increases or decreases with equal probability 2/n except if the distance

is 1 or 1; in this last case, the distance increases (going to 2 or 2) with
probability 2/n, but decreases only with probability 1/n. We want to know
when this process hits 0 (or n).
Imagine a random walk starting from position k with absorbing barriers
at 1 and n 1. This reaches 1 or n 1 after (k 1)(n 1 k) steps on
average, which translates into (n/4)(k 1)(n 1 k) steps of our original
process if we take into account that we only move with probability 4/n
per time unit. This time is maximized by setting k = n/2, which gives
(n/4)(n/2 1)2 = n3 /16 n2 /4 + n/4 expected time units to reach 1 or n 1
for the first time.
At 1 or n 1, we wait an addition n/3 steps on average; then with
probability 1/3 the process finishes and with probability 2/3 we start over
from position 2 or n 2; in the latter case, we run (n/4)(n 3) + n/3 time
units on average before we may finish again. On average, it takes 3 attempts
to finish. Each attempt incurs the expected n/3 cost before taking a step,
and all but the last attempt incur the expected (n/4)(n 3) additional steps
for the random walk. So the last phase of the process the process adds
(1/2)n(n 3) + n = (1/2)n2 (5/4)n steps.
Adding up all of the costs gives n3 /16 n2 /4 + n/4 + n/3 + (1/2)n2
1 3
(5/4)n = 16 n + 41 n2 23 n steps.
D.7.3 Finding the maximum
1 Randomly permute A
2 m
3 for i 1 . . . n do
4 if A[i] > m then
5 m A[i]
6 return m
Algorithm D.2: Randomized max-finding algorithm
Suppose that we run Algorithm D.2 on an array with n elements, all of

which are distinct. What is the expected number of times Line 5 is executed
as a function of n?
Solution
Let Xi be the indicator variable for the event that Line 5 is executed on
the i-th pass through the loop. This will occur if A[i] is greater than A[j]
for all j < i, which occurs with probability exactly 1/i (given that A has
been permuted randomly). So the expected number of calls to Line 5 is
i = 1n E [Xi ] = ni=1 1i = Hn .
P P
D.7.4 Random graph coloring

Let G be a random d-regular graph on n vertices, that is, a graph drawn
uniformly from the family of all n-vertex graph in which each vertex has
exactly d neighbors. Color the vertices of G red or blue independently at
random.
1. What is the expected number of monochromatic edges in G?
2. Show that the actual number of monochromatic edges is tightly con-

centrated around its expectation.
Solution
The fact that G is itself a random graph is a red herring; all we really need
to know is that its d-regular.
1. Because G has exactly dn/2 edges, and each edge has probability 1/2
of being monochromatic, the expected number of monochromatic edges
is dn/4.
2. This is a job for Azumas inequality. Consider the vertex exposure

martingale. Changing the color of any one vertex changes the number of
monochromatic edges by at most d. So we have Pr [|X E [X]| t]
2 2
2 exp t2 /2 ni=1 d2 = 2et /2nd , which tells us that the deviation is
P

likely to be not much more than O(d n).
Appendix E

Spring 2009
E.1 Final exam, Spring 2009

E.1.1 Randomized mergesort (20 points)

Consider the following randomized version of the mergesort algorithm. We
take an unsorted list of n elements and split it into two lists by flipping an
independent fair coin for each element to decide which list to put it in. We
then recursively sort the two lists, and merge the resulting sorted lists. The
merge procedure involves repeatedly comparing the smallest element in each
of the two lists and removing the smaller element found, until one of the lists
is empty.
Compute the expected number of comparisons needed to perform this
final merge. (You do not need to consider the cost of performing the recursive
sorts.)
Solution
Color the elements in the final merged list red or blue based on which sublist
they came from. The only elements that do not require a comparison to
insert into the main list are those that are followed only by elements of the
351
APPENDIX E. SAMPLE ASSIGNMENTS FROM SPRING 2009 352
same color; the expected number of such elements is equal to the expected
length of the longest monochromatic suffix. By symmetry, this is the same as
the expected longest monochromatic prefix, which is equal to the expected
length of the longest sequence of identical coin-flips.
The probability of getting k identical coin-flips in a row followed by a
different coin-flip is exactly 2k ; the first coin-flip sets the color, the next k 1
must follow it (giving a factor of 2k+1 , and the last must be the opposite
color (giving an additional factor of 21 ). For n identical coin-flips, there is
a probability of 2n+1 , since we dont need an extra coin-flip of the opposite
color. So the expected length is n1 k + n2n+1 = Pn k + n2n .
P
k=1 k2 k=0 k2Pn
We can simplify the sum using generating functions. The sum k=0 2k z k
n+1
is given by 1(z/2)
Pn k k1 =
1z/2 . Taking the derivative with respect to z gives k=0 2 kz
n+1 2 n
(1/2) 1(z/2)
1z/2 + (1/2) (n+1)(z/2)
1z/2 . At z = 1 this is 2(1 2n1 ) 2(n +
1)2n = 2 (n + 2)2n . Adding the second term gives E [X] = 2 2 2n =
2 2n+1 .
Note that this counts the expected number of elements for which we do
not have to do a comparison; with n elements total, this leaves n 2 + 2n+1
comparisons on average.
E.1.2 A search problem (20 points)

Suppose you are searching a space by generating new instances of some
problem from old ones. Each instance is either good or bad; if you generate a
new instance from a good instance, the new instance is also good, and if you
generate a new instance from a bad instance, the new instance is also bad.
Suppose that your start with X0 good instances and Y0 bad instances,
and that at each step you choose one of the instances you already have
uniformly at random to generate a new instance. What is the expected
number of good instances you have after n steps?
Hint: Consider the sequence of values {Xt /(Xt + Yt )}.
Solution
We can show that the suggested sequence is a martingale, by computing

Xt+1 Xt Xt + 1 Yt Xt

E Xt , Yt = +
Xt+1 + Yt+1 Xt + Yt Xt + Yt + 1 Xt + Yt Xt + Yt + 1
Xt (Xt + 1) + Yt Xt
=
(Xt + Yt )(Xt + Yt + 1)
Xt (Xt + Yt + 1)
=
(Xt + Yt )(Xt + Yt + 1)
Xt
= .
Xt + Yt
h i
From the martingale property we have E XnX+Y n
n
= X0X+Y
0
0
. But Xn +
Yn = X0 + Y0 + n,
a constant,
so we can multiply both sides by this value to
X0 +Y0 +n
get E [Xn ] = X0 X0 +Y0 .
E.1.3 Support your local police (20 points)

At one point I lived in a city whose local police department supported
themselves in part by collecting fines for speeding tickets. A speeding ticket
would cost 1 unit (approximately $100), and it was unpredictable how often
one would get a speeding ticket. For a price of 2 units, it was possible
to purchase a metal placard to go on your vehicle identifying yourself as
a supporter of the police union, which (at least according to local legend)
would eliminate any fines for subsequent speeding tickets, but which would
not eliminate the cost of any previous speeding tickets.
Let us consider the question of when to purchase a placard as a problem
in on-line algorithms. It is possible to achieve a strict1 competitive ratio of 2
by purchasing a placard after the second ticket. If one receives fewer than 2
tickets, both the on-line and off-line algorithms pay the same amount, and at
2 or more tickets the on-line algorithm pays 4 while the off-line pays 2 (the
off-line algorithm purchased the placard before receiving any tickets at all).
1. Show that no deterministic algorithm can achieve a lower (strict)

competitive ratio.
2. Show that a randomized algorithm can do so, against an oblivious

adversary.
1
I.e., with no additive constant.
Solution
1. Any deterministic algorithm essentially just chooses some fixed number
m of tickets to collect before buying the placard. Let n be the actual
number of tickets issued. For m = 0, the competitive ratio is infinite
when n = 0. For m = 1, the competitive ratio is 3 when n = 1. For
m > 2, the competitive ratio is (m + 2)/2 > 2 when n = m. So m = 2
is the optimal choice.
2. Consider the following algorithm: with probability p, we purchase a

placard after 1 ticket, and with probability q = 1 p, we purchase a
placard after 2 tickets. This gives a competitive ratio of 1 for n = 0,
1 + 2p for n = 1, and (3p + 4q)/2 = (4 p)/2 = 2 p/2 for n 2.
There is a clearly a trade-off between the two ratios 1 + 2p and 2 p/2.
The break-even point is when they are equal, at p = 2/5. This gives a
competitive ratio of 1 + 2p = 9/5, which is less than 2.
E.1.4 Overloaded machines (20 points)

Suppose n2 jobs are assigned to n machines with each job choosing a machine
independently and uniformly at random. Let the load on a machine be the
number of jobs assigned to it. Show that for any fixed > 0 and sufficiently
large n, there is a constant c < 1 such that the maximum load exceeds
(1 + )n with probability at most ncn .
Solution
This is a job for Chernoff bounds. For any particular machine, the load S is
a sum of independent indicator variables and the mean load is = n. So we
have
!n
e
Pr [S (1 + )] .
(1 + )1+
Observe that e /(1 + )1+ < 1 for > 0. One proof of this fact is to take
the log to get (1+) log(1+), which equals 0 at = 0, and then show that
d
the logarithm is decreasing by showing that d = 1 1+
1+ log(1 + ) =
log(1 + ) < 0 for all > 0.
So we can let c = e /(1 + )1+ to get a bound of cn on the probability
that any particular machine is overloaded and a bound of ncn (from the
union bound) on the probability that any of the machines is overloaded.
Appendix F
Probabilistic recurrences
Randomized algorithms often produce recurrences with ugly sums embedded

inside them (see, for example, (1.3.1)). Wed like to have tools for pulling
at least asymptotic bounds out of these recurrences without having to deal
with the sums. This isnt always possible, but for certain simple recurrences
we can make it work.
F.1 Recurrences with constant cost functions

Let us consider probabilistic recurrences of the form T (n) = 1 + T (n Xn ),
where Xn is a random variable with 0 < Xn n and T (0) = 0. We assume
we can compute a lower bound on E [Xn ] for each n, and we want to translate
this lower bound into an upper bound on E [T (n)].
F.2 Examples
How long does it take to get our first heads if we repeatedly flip
a coin that comes up heads with probability p? Even though we
probably already know the answer to this, We can solve it by solving
the recurrence T (1) = 1 + T (1 X1 ), T (0) = 0, where E [X1 ] = p.
Hoares FIND [Hoa61b], often called QuickSelect, is an algorithm

for finding the k-th smallest element of an unsorted array. It works like
QuickSort, only after partitioning the array around a random pivot
we throw away the part that doesnt contain our target and recurse
only on the surviving piece. How many rounds of this must we do?
Here E [Xn ] is more complicated, since after splitting our array of size
355
APPENDIX F. PROBABILISTIC RECURRENCES 356
n into piles of size n0 and n n0 1, we have to pick one or the other

(or possibly just the pivot alone) based on the value of k.
Suppose we start with n biased coins that each come up heads with
probability p. In each round, we flip all the coins and throw away the
ones that come up tails. How many rounds does it take to get rid of
all of the coins? (This essentially tells us how tall a skip list [Pug90]
can get.) Here we have E [Xn ] = (1 p)n.
In the coupon collector problem, we sample from 1 . . . n with re-

placement until we see ever value at least once. We can model this
by a recurrence in which T (k) is the time to get all the coupons given
there are k left that we havent seen. Here Xn is 1 with probability
k/n and 0 with probability (n k)/n, giving E [Xn ] = k/n.
Lets play Chutes and Ladders without the chutes and ladders. We
start at location n, and whenever its our turn, we roll a fair six-sided
die X and move to n X unless this value is negative, in which case
we stay put until the next turn. How many turns does it take to get to
0?
F.3 The Karp-Upfal-Wigderson bound

This is a bound on the expected number of rounds to finish a process where
we start with a problem instance of size n, and after one round of work we get
a new problem instance of size n Xn , where Xn is a random variable whose
distribution depends on n. It was original described in a paper by Karp,
Upfal, and Wigderson on analyzing parallel search algorithms [KUW88]. The
bound applies when E [Xn ] is bounded below by a non-decreasing function
(n).
Lemma F.3.1. Let a be a constant, let T (n) = 1 + T (n Xn ), where for
each n, Xn is an integer-valued random variable satisfying 0 Xn n a
and let T (a) = 0. Let E [Xn ] (n) for all n > a, where is a positive
non-decreasing function of n. Then
Z n
1
E [T (n)] dt. (F.3.1)
a (t)
To get an intuition for why this works, imagine that Xn is the speed
at which we drop from n, expressed in units per round. Traveling at this
speed, it takes 1/Xn rounds to cross from k + 1 to k for any such interval
we pass. From the point of view of the interval [k, k + 1], we dont know
which n we are going to start from before we cross it, but we do know that
for any n k + 1 we start from, our speed will Rbe at least (n) (k + 1)
on average. So the time it takes will be at most kk+1 (t)
1
dt on average, and
the total time is obtained by summing all of these intervals.
Of course, this intuition is not even close to a real proof (among other
things, there may be a very dangerous confusion in there between 1/ E [Xn ]
and E [1/Xn ]), so we will give a real proof as well.
Proof of Lemma F.3.1. This is essentially the same proof as in Motwani and
Raghavan [MR95], but we add some extra detail to allow for the possibility
that Xn = 0.
Let p = Pr [Xn = 0], q = 1 p = Pr [Xn 6= 0]. Note we have q > 0
because otherwise E [Xn ] = 0 < (n). Then we have
E [T (n)] = 1 + E [T (n Xn )]
= 1 + p E [T (n Xn ) | Xn = 0] + q E [T (n Xn ) | Xn 6= 0]
= 1 + p E [T (n)] + q E [T (n Xn ) | Xn 6= 0] .
Now we have E [T (n)] on both sides, which we dont like very much. So
we collect it on the left-hand side:
(1 p) E [T (n)] = 1 + q E [T (n Xn ) | Xn 6= 0] ,
divide both sides by q = 1 p, and apply the induction hypothesis:
E [T (n)] = 1/q + E [T (n Xn ) | Xn 6= 0]
= 1/q + E [E [T (n Xn ) | Xn ] | Xn 6= 0]
"Z #
nXn1
1/q + E dt Xn 6= 0

a (t)
Z n Z n
1 1

= 1/q + E dt dt Xn 6= 0
a (t) nXn (t)
Z n
1 Xn

1/q + dt E Xn 6= 0
a (t) (n)
Z n
1 E [Xn | Xn 6= 0]
1/q + dt .
a (t) (n)
The second-to-last step uses the fact that (t) (n)f ort n.
It may seem like we dont know what E [Xn | Xn 6= 0] is. But we know
that Xn 0, so we have E [Xn ] = p E [Xn | Xn = 0] + q E [Xn | Xn 6= 0] =
q E [Xn | Xn 6= 0]. So we can solve for E [Xn | Xn 6= 0] = E[Xn ]/q. So lets

plug this in:
Z n
1 E [Xn ] /q
E [T (n)] 1/q + dt
a (t) (n)
Z n
1
1/q + dt 1/q
a (t)
Z n
1
= dt.
a (t)
This concludes the proof.
Now we just need to find some applications.
F.3.1 Waiting for heads

For the recurrence T (1) = 1+T (1X1 ) with E [X1 ] = p, we set (n) = p and
get E [T (1)] 01 p1 dt = p1 , which happens to be exactly the right answer.
R
F.3.2 Quickselect
In Quickselect, we pick a random pivot and split the original array of size
n into three piles of size m (less than the pivot), 1 (the pivot itself), and
nm1 (greater than the pivot). We then figure out which of the three piles
contains the k-th smallest element (depend on how k compares to m 1) and
recurse, stopping when we hit a pile with 1 element. Its easiest to analyze
this by assuming that we recurse in the largest of the three piles, i.e., that
our recurrence is T (n) = 1 + max(T (m), T (n m 1)), where m is uniform
in 0 . . . n 1. The exact value of E [max(m, n m 1)] is a little messy to
compute (among other things, it depends on whether n is odd or even), but
its not hard to see that its always less than (3/4)n. So letting (n) = n/4,
we get
Z n
1
E [T (n)] dt = 4 ln n.
1 t/4
F.3.3 Tossing coins

Here we have E [Xn ] = (1 p)n. If we let (n) = (1 p)n and plug into the
formula without thinking about it too much, we get
Z n
1 1
E [T (n)] dt = (ln n ln 0).
0 (1 p)t 1p
That ln 0 is trouble. We can fix it by making (n) = (1 p)dne, to get

Z n
1
E [T (n)] dt
0+ (1 p)dte
n
1 X 1
=
1 p k=1 k
Hn
= .
1p
F.3.4 Coupon collector

Now that we know how to avoid dividing by zero, this is easy and fun. Let
(x) = dxe/n, then we have
Z n
n
E [T (n)] dt
0+ dte
n
X n
=
k=1
k
= nHn .
As it happens, this is the exact answer for this case. This will happen
whenever X is always a 01 variable1 and we define (x) = E [X | n = dxe],
which can be seen by spending far too much time thinking about the precise
sources of error in the inequalities in the proof.
F.3.5 Chutes and ladders

Let (n) be the expected drop from position n. We have to be a little
bit careful about small n, but we can compute that in general (n) =
1 Pmin(n,6)
6 i=0 i. For fractional values x we will set (x) = (dxe) as before.
1
A 01 random variable is also called a Bernoulli random variable, but 01 is
shorter to type and more informative. Even more confusing, the underlying experiment
that gives rise to a Bernoulli random variable goes by the entirely different name of a
Poisson trial. Jacob Bernoulli and Simon-Denis Poisson were great mathematicians,
but there are probably better ways to remember them.
Then we have
Z n
1
E [T (n)] dt
0+ (t)
n
X 1
=
k=1
(k)
We can summarize the values in the following table:
P
n (n) 1/(n) 1/(k)
1 1/6 6 6
2 1/2 2 8
3 1 1 9
4 5/3 3/5 48/5
5 5/2 2/5 10
6 7/2 2/7 10 + (2/7)(n 5) = (2/7)n + 65/7
This is a slight overestimate; for example, we can calculate by hand

that the expected waiting time for n = 2 is 6 and for n = 3 that it is
20/3 = 6 + 2/3.
We can also consider the generalized version of the game where we start
at n and drop by 1 . . . n each turn as long as the drop wouldnt take us below
0. Now the expected drop from position k is k(k + 1)/2n, and so we can
apply the formula to get
n
X 2n
E [T (n)] .
k=1
k(k + 1)
1
The sum of k(k+1) when k goes from 1 to n happens to have a very nice
n 1 2
value; its exactly n+1 = 1 + n+1 . So in this case we can rewrite the bound
n 2n2
as 2n n+1 = n+1 .
F.4 High-probability bounds

So far we have only considered bounds on the expected value of T (n).
Suppose we want to show that T (n) is in fact small with high probability,
2
Pn 1
Pn1 1
Proof: Trivially true for n = 0; for larger n, compute k=1 k(k+1) k=1 k(k+1)
+
1 n1 1 (n1)(n+1)1 n2
n(n+1)
= n
+ n(n+1)
= n(n+1)
= n(n+1)
= n/(n + 1).
i.e., a statement of the form Pr [T (n) t] . There are two natural ways
to do this: we can repeatedly apply Markovs inequality to the expectation
bound, or we can attempt to analyze the recurrence in more detail. The first
method tends to give weaker bounds but its easier.
F.4.1 High-probability bounds from expectation bounds

Given E [T (n)] m, we have Pr [T (n) m] 1 . This does not give
a very good bound on probability; if we want to show Pr [T (n) t] nc
for some constant c (a typical high-probability bound), we need t nc m.
But we can get a better bound if m bounds the expected time starting from
any reachable state, as is the case for the class of problems we have been
considering.
The idea is that if T (n) exceeds m, we restart the analysis and ar-
gue that Pr [T (n) 2m | T (n) m] 1 , from which it follows that
Pr [T (n) 2m] 2 . In general, for any non-negative integer k, we have
Pr [T (n) km] k . Now we just need to figure out how to pick to
minimize this quantity for fixed t.
Let t = km. Then k = t/m and we seek to minimize t/m . Taking
the logarithm gives (t/m)(ln )/. The t/m factor is irrelevant to the
minimization problem, so we are left with minimizing (ln )/. Taking
the derivative gives 2 + 2 ln ; this is zero when ln = 1 or = e.
(This popular constant shows up often in problems like this.) So we get
Pr [T (n) kem] ek , or, letting k = ln(1/), Pr [T (n) em ln(1/)] .
So, for example, we can get an nc bound on the probability of running
too long by setting our time bound to em ln(nc ) = cem ln n = O(m log n).
We cant hope to do better than O(m), so this bound is tight up to a log
factor.
F.4.2 Detailed analysis of the recurrence

As Lance Fortnow has explained,3 getting rid of log factors is what theoretical
computer science is all about. So wed like to do better than an O(m log n)
bound if we can. In some cases this is not too hard.
Suppose for each n, T (n) = 1 + T (X), where E [X] n for a fixed
constant . Let X0 = n, and let X1 , X2 , etc., be the sequence of sizes of
the remaining problem at time 1, 2, etc. Then we have E [X1 ] n from
our assumption. But we also have E [X2 ] = E [E [X2 | X1 ]] E [X1 ] =
E [X1 ] 2 n, and by induction we can show that E [Xk ] k n for all k.
3
http://weblog.fortnow.com/2009/01/soda-and-me.html
Since Xk is integer-valued, E [Xk ] is an upper bound on Pr [Xk > 0]; we thus

get Pr [T (n) k] = Pr [Xk > 0] k n. We can solve for the value of k that
makes this less than : k = log(n/)/ log = log1/ n + log1/ (1/).
For comparison, the bound on the expectation of T (n) from Lemma F.3.1
is H(n)/(1 ). This is actually pretty close to log1/ n when is close
to 1, and is not too bad even for smaller . But the difference is that the
dependence on log(1/) is additive with the tighter analysis, so for fixed c
we get Pr [T (n) t] nc at t = O(log n + log nc ) = O(log n) instead of
O(log n log nc ) = O(log2 n).
F.5 More general recurrences

We didnt do these, but if we had to, it might be worth looking at Rouras
Improved Master Theorems [Rou01].
Bibliography
[AA11] Dan Alistarh and James Aspnes. Sub-logarithmic test-and-set

against a weak adversary. In Distributed Computing: 25th
International Symposium, DISC 2011, volume 6950 of Lecture
Notes in Computer Science, pages 97109. Springer-Verlag,
September 2011.
[ABMRT96] Arne Andersson, Peter Bro Miltersen, Sren Riis, and Mikkel
Thorup.
p Static dictionaries on AC0 RAMs: Query time
( log n/ log log n) is necessary and sufficient. In FOCS, pages
441450, 1996.
[ACBF02] Peter Auer, Nicol Cesa-Bianchi, and Paul Fischer. Finite-time

analysis of the multiarmed bandit problem. Machine Learning,
47:235256, 2002.
[Ach03] Dimitris Achlioptas. Database-friendly random projections:

Johnson-Lindenstrauss with binary coins. J. Comput. Syst.
Sci., 66(4):671687, June 2003.
[Adl78] Leonard Adleman. Two theorems on random polynomial time.

In Proceedings of the 19th Annual Symposium on Foundations
of Computer Science, pages 7583, Washington, DC, USA, 1978.
IEEE Computer Society.
[AF01] David Aldous and James Allen Fill. Reversible Markov chains
and random walks on graphs. Unpublished manuscript, avail-
able at http://www.stat.berkeley.edu/~aldous/RWG/book.
html, 2001.
[AKS83] M. Ajtai, J. Komls, and E. Szemerdi. An O(n log n) sorting

network. In STOC 83: Proceedings of the fifteenth annual ACM
363
BIBLIOGRAPHY 364
symposium on Theory of computing, pages 19, New York, NY,

USA, 1983. ACM.
[AKS04] Manindra Agrawal, Neeraj Kayal, and Nitin Saxena. PRIMES

is in P. Annals of Mathematics, 160:781793, 2004.
[Alo91] Noga Alon. A parallel algorithmic version of the local lemma.

Random Structures & Algorithms, 2(4):367378, 1991.
[AS92] Noga Alon and Joel H. Spencer. The Probabilistic Method. John
Wiley & Sons, 1992.
[AS07] James Aspnes and Gauri Shah. Skip graphs. ACM Transactions
on Algorithms, 3(4):37, November 2007.
[AVL62] G. M. Adelson-Velskii and E. M. Landis. An information

organization algorithm. Doklady Akademia Nauk SSSR, 146:263
266, 1962.
[AW96] James Aspnes and Orli Waarts. Randomized consensus in

O(n log2 n) operations per processor. SIAM Journal on Com-
puting, 25(5):10241044, October 1996.
[Azu67] Kazuoki Azuma. Weighted sums of certain dependent random

variables. Thoku Mathematical Journal, 19(3):357367, 1967.
[Bab79] Lszl Babai. Monte-carlo algorithms in graph isomorphism

testing. Technical Report D.M.S. 79-10, Universit de Montral,
1979.
[BBBV97] Charles H. Bennett, Ethan Bernstein, Gilles Brassard, and

Umesh Vazirani. Strengths and weaknesses of quantum com-
puting. SIAM J. Comput., 26(5):15101523, October 1997.
[BCB12] Sbastian Bubeck and Nicol Cesa-Bianchi. Regret analysis

of stochastic and nonstochastic multi-armed bandit problems.
Foundations and Trends in Machine Learning, 5(1):1122, 2012.
[BD92] Dave Bayer and Persi Diaconis. Trailing the dovetail shuffle to
its lair. Annals of Applied Probability, 2(2):294313, 1992.
[Bec91] Jzsef Beck. An algorithmic approach to the lovsz local lemma.

i. Random Structures & Algorithms, 2(4):343365, 1991.
BIBLIOGRAPHY 365
[Bel57] Richard Earnest Bellman. Dynamic Programming. Princeton

University Press, 1957.
[Blo70] Burton H. Bloom. Space/time trade-offs in hash coding with

allowable errors. Commun. ACM, 13(7):422426, 1970.
[Bol01] Bla Bollobs. Random Graphs. Cambridge University Press,

second edition, 2001.
[Bro86] Andrei Z. Broder. How hard is to marry at random? (on

the approximation of the permanent). In Proceedings of the
Eighteenth Annual ACM Symposium on Theory of Computing,
28-30 May 1986, Berkeley, California, USA, pages 5058, 1986.
[Bro88] Andrei Z. Broder. Errata to how hard is to marry at random?

(on the approximation of the permanent). In Proceedings of the
Twentieth Annual ACM Symposium on Theory of Computing,
2-4 May 1988, Chicago, Illinois, USA, page 551, 1988.
[BRSW06] Boaz Barak, Anup Rao, Ronen Shaltiel, and Avi Wigderson.
2-source dispersers for sub-polynomial entropy and Ramsey
graphs beating the Frankl-Wilson construction. In Proceedings
of the thirty-eighth annual ACM symposium on Theory of com-
puting, STOC 06, pages 671680, New York, NY, USA, 2006.
ACM.
[CCFC04] Moses Charikar, Kevin Chen, and Martin Farach-Colton. Find-

ing frequent items in data streams. Theor. Comput. Sci.,
312(1):315, 2004.
[CM03] Saar Cohen and Yossi Matias. Spectral bloom filters. In Alon Y.
Halevy, Zachary G. Ives, and AnHai Doan, editors, Proceed-
ings of the 2003 ACM SIGMOD International Conference on
Management of Data, San Diego, California, USA, June 9-12,
2003, pages 241252, 2003.
[CM05] Graham Cormode and S. Muthukrishnan. An improved data

stream summary: the count-min sketch and its applications. J.
Algorithms, 55(1):5875, 2005.
[CS00] Artur Czumaj and Christian Scheideler. Coloring non-uniform

hypergraphs: a new algorithmic approach to the general Lovsz
BIBLIOGRAPHY 366
local lemma. In Proceedings of the eleventh annual ACM-

SIAM symposium on Discrete algorithms, SODA 00, pages
3039, Philadelphia, PA, USA, 2000. Society for Industrial and
Applied Mathematics.
[CW77] J. Lawrence Carter and Mark N. Wegman. Universal classes of

hash functions (extended abstract). In Proceedings of the ninth
annual ACM symposium on Theory of computing, STOC 77,
pages 106112, New York, NY, USA, 1977. ACM.
[Deu89] David Deutsch. Quantum computational networks. Proceedings

of the Royal Society of London. A. Mathematical and Physical
Sciences, 425(1868):7390, 1989.
[Dev88] Luc Devroye. Applications of the theory of records in the study

of random trees. Acta Informatica, 26(1-2):123130, October
1988.
[DGM02] Martin Dyer, Catherine Greenhill, and Mike Molloy. Very

rapid mixing of the Glauber dynamics for proper colorings on
bounded-degree graphs. Random Struct. Algorithms, 20:98114,
January 2002.
[DHKP97] Martin Dietzfelbinger, Torben Hagerup, Jyrki Katajainen, and

Martti Penttonen. A reliable randomized algorithm for the
closest-pair problem. J. Algorithms, 25(1):1951, 1997.
[Dir39] P. A. M. Dirac. A new notation for quantum mechanics. Math-

ematical Proceedings of the Cambridge Philosophical Society,
35:416418, 6 1939.
[DJ92] David Deutsch and Richard Jozsa. Rapid solution of prob-

lems by quantum computation. Proceedings of the Royal Soci-
ety of London. Series A: Mathematical and Physical Sciences,
439(1907):553558, 1992.
[Dum56] A. I. Dumey. Indexing for rapid random-access memory. Com-

puters and Automation, 5(12):69, 1956.
[Dye03] Martin Dyer. Approximate counting by dynamic programming.

In Proceedings of the thirty-fifth annual ACM symposium on
Theory of computing, STOC 03, pages 693699, New York, NY,
USA, 2003. ACM.
BIBLIOGRAPHY 367
[EL75] Paul Erds and Laszlo Lovsz. Problems and results on 3-

chromatic hypergraphs and some related questions. In A. Haj-
nal, R. Rado, and V. T. Ss, editors, Infinite and Finite Sets (to
Paul Erds on his 60th birthday), pages 609627. North-Holland,
1975.
[Epp16] David Eppstein. Cuckoo filter: Simplification and analysis.

In Rasmus Pagh, editor, 15th Scandinavian Symposium and
Workshops on Algorithm Theory, SWAT 2016, June 22-24,
2016, Reykjavik, Iceland, volume 53 of LIPIcs, pages 8:18:12.
Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2016.
[Erd45] P. Erds. On a lemma of Littlewood and Offord. Bulletin of

the American Mathematical Society, 51(12):898902, 1945.
[Erd47] P. Erds. Some remarks on the theory of graphs. Bulletin of

the American Mathematical Society, 53:292294, 1947.
[ES35] P. Erds and G. Szekeres. A combinatorial problem in geometry.

Compositio Mathematica, 2:463470, 1935.
[Eul68] M. L. Euler. Remarques sur un beau rapport entre les sries

des pussances tant directes que rciproques. Mmoires de
lAcadmie des Sciences de Berlin, 17:83106, 1768.
[FAKM14] Bin Fan, David G. Andersen, Michael Kaminsky, and Michael

Mitzenmacher. Cuckoo filter: Practically better than bloom.
In Aruna Seneviratne, Christophe Diot, Jim Kurose, Augustin
Chaintreau, and Luigi Rizzo, editors, Proceedings of the 10th
ACM International on Conference on emerging Networking Ex-
periments and Technologies, CoNEXT 2014, Sydney, Australia,
December 2-5, 2014, pages 7588. ACM, 2014.
[FCAB00] Li Fan, Pei Cao, Jussara M. Almeida, and Andrei Z. Broder.

Summary cache: a scalable wide-area web cache sharing proto-
col. IEEE/ACM Trans. Netw., 8(3):281293, 2000.
[Fel68] William Feller. An Introduction to Probability Theory and Its

Applications, volume 1. Wiley, third edition, 1968.
[Fel71] William Feller. An Introduction to Probability Theory and Its

Applications, volume 2. Wiley, second edition, 1971.
BIBLIOGRAPHY 368
[FGQ12] Xiequan Fan, Ion Grama, and Liu Quansheng. Hoeffdings

inequality for supermartingales. Available as http://arxiv.
org/abs/1109.4359, July 2012.
[FKS84] M.L. Fredman, J. Komls, and E. Szemerdi. Storing a Sparse

Table with O(1) Worst Case Access Time. Journal of the ACM
(JACM), 31(3):538544, 1984.
[FNM85] Philippe Flajolet and G Nigel Martin. Probabilistic counting

algorithms for data base applications. Journal of computer and
system sciences, 31(2):182209, 1985.
[FPSS03] Dimitris Fotakis, Rasmus Pagh, Peter Sanders, and Paul Spi-
rakis. Space efficient hash tables with worst case constant
access time. In Helmut Alt and Michel Habib, editors, STACS
2003, volume 2607 of Lecture Notes in Computer Science, pages
271282. Springer Berlin Heidelberg, 2003.
[FR75] Robert W. Floyd and Ronald L. Rivest. Expected time bounds

for selection. Commun. ACM, 18(3):165172, 1975.
[GKP88] Ronald L. Graham, Donald E. Knuth, and Oren Patashnik.

Concrete Mathematics: A Foundation for Computer Science.
Addison-Wesley, 1988.
[Goo83] Nelson Goodman. Fact, Fiction, and Forecast. Harvard Uni-

versity Press, 1983.
[GR93] Igal Galperin and Ronald L. Rivest. Scapegoat trees. In

Proceedings of the Fourth Annual ACM-SIAM Symposium on
Discrete Algorithms, pages 165174. Society for Industrial and
Applied Mathematics, 1993.
[Gro96] Lov K Grover. A fast quantum mechanical algorithm for

database search. In Proceedings of the twenty-eighth annual
ACM symposium on Theory of computing, pages 212219.
ACM, 1996. Available as http://arxiv.org/abs/quant-ph/
9605043.
[GS78] Leo J. Guibas and Robert Sedgewick. A dichromatic framework

for balanced trees. In Foundations of Computer Science, 1978.,
19th Annual Symposium on, pages 821. IEEE, 1978.
BIBLIOGRAPHY 369
[GS92] G. R. Grimmet and D. R. Stirzaker. Probability and Random

Processes. Oxford Science Publications, second edition, 1992.
[GS01] Geoffrey R. Grimmett and David R. Stirzaker. Probability

and Random Processes. Oxford University Press, third edition,
2001.
[Gur00] Venkatesen Guruswami. Rapidly mixing Markov chains: A

comparison of techniques. Available at ftp://theory.lcs.
mit.edu/pub/people/venkat/markov-survey.ps, 2000.
[GW94] Michel X. Goemans and David P. Williamson. New 3/4-

approximation algorithms for the maximum satisfiability prob-
lem. SIAM J. Discret. Math., 7:656666, November 1994.
[GW95] Michel X. Goemans and David P. Williamson. Improved ap-

proximation algorithms for maximum cut and satisfiability
problems using semidefinite programming. Journal of the ACM,
42(6):11151145, 1995.
[GW12] George Giakkoupis and Philipp Woelfel. On the time and space
complexity of randomized test-and-set. In Darek Kowalski and
Alessandro Panconesi, editors, ACM Symposium on Princi-
ples of Distributed Computing, PODC 12, Funchal, Madeira,
Portugal, July 16-18, 2012, pages 1928. ACM, 2012.
[Has70] W. K. Hastings. Monte carlo sampling methods using Markov

chains and their applications. Biometrika, 57(1):97109, 1970.
[HH80] P. Hall and C.C. Heyde. Martingale Limit Theory and Its
Application. Academic Press, 1980.
[HK73] John E. Hopcroft and Richard M. Karp. An n5/2 algorithm

for maximum matchings in bipartite graphs. SIAM J. Comput.,
2(4):225231, 1973.
[Hoa61a] C. A. R. Hoare. Algorithm 64: Quicksort. Commun. ACM,

4:321, July 1961.
[Hoa61b] C. A. R. Hoare. Algorithm 65: find. Commun. ACM, 4:321322,

July 1961.
BIBLIOGRAPHY 370
[Hoe63] Wassily Hoeffding. Probability inequalities for sums of bounded

random variables. Journal of the American Statistical Associa-
tion, 58(301):1330, March 1963.
[Hol16] Jonathan Holmes. AI is already making in-
roads into journalism but could it win a Pulitzer?
https://www.theguardian.com/media/2016/apr/03/
artificla-intelligence-robot-reporter-pulitzer-prize,
April 2016. Accessed September 18th, 2016.
[HST08] Maurice Herlihy, Nir Shavit, and Moran Tzafrir. Hopscotch
hashing. In Proceedings of the 22nd international symposium
on Distributed Computing, DISC 08, pages 350364, Berlin,
Heidelberg, 2008. Springer-Verlag.
[IM98] Piotr Indyk and Rajeev Motwani. Approximate nearest neigh-
bors: Towards removing the curse of dimensionality. In STOC,
pages 604613, 1998.
[JL84] William B. Johnson and Joram Lindenstrauss. Extensions of
Lipschitz mappings into a Hilbert space. In Conference in
Modern Analysis and Probability (New Haven, Connecticut,
1982), number 26 in Contemporary Mathematics, pages 189
206. American Mathematical Society, 1984.
[JLR00] Svante Janson, Tomasz uczak, and Andrezej Ruciski. Ran-
dom Graphs. John Wiley & Sons, 2000.
[JS89] Mark Jerrum and Alistair Sinclair. Approximating the perma-
nent. SIAM J. Comput., 18(6):11491178, 1989.
[JSV04] Mark Jerrum, Alistair Sinclair, and Eric Vigoda. A polynomial-
time approximation algorithm for the permanent of a matrix
with nonnegative entries. J. ACM, 51(4):671697, 2004.
[Kar93] David R. Karger. Global min-cuts in rnc, and other ramifi-
cations of a simple min-out algorithm. In Proceedings of the
fourth annual ACM-SIAM Symposium on Discrete algorithms,
SODA 93, pages 2130, Philadelphia, PA, USA, 1993. Society
for Industrial and Applied Mathematics.
[KL85] Richard M. Karp and Michael Luby. Monte-carlo algorithms
for the planar multiterminal network reliability problem. J.
Complexity, 1(1):4564, 1985.
BIBLIOGRAPHY 371
[KM08] Adam Kirsch and Michael Mitzenmacher. Less hashing, same

performance: Building a better bloom filter. Random Struct.
Algorithms, 33(2):187218, 2008.
[KNW10] Daniel M. Kane, Jelani Nelson, and David P. Woodruff. An opti-
mal algorithm for the distinct elements problem. In Proceedings
of the Twenty-Ninth ACM SIGMOD-SIGACT-SIGART Sym-
posium on Principles of Database Systems, PODS 10, pages
4152, New York, NY, USA, 2010. ACM.
[Kol33] A.N. Kolmogorov. Grundbegriffe der Wahrscheinlichkeitsrech-
nung. Springer, 1933.
[KR99] V. S. Anil Kumar and H. Ramesh. Markovian coupling vs.
conductance for the Jerrum-Sinclair chain. In FOCS, pages
241252, 1999.
[KS76] J.G. Kemeny and J.L. Snell. Finite Markov Chains: With
a New Appendix Generalization of a Fundamental Matrix.
Undergraduate Texts in Mathematics. Springer, 1976.
[KSK76] John G. Kemeny, J. Laurie. Snell, and Anthony W. Knapp.
Denumerable Markov Chains, volume 40 of Graduate Texts in
Mathematics. Springer, 1976.
[KT75] Samuel Karlin and Howard M. Taylor. A First Course in
Stochastic Processes. Academic Press, second edition, 1975.
[KUW88] Richard M. Karp, Eli Upfal, and Avi Wigderson. The complexity
of parallel search. Journal of Computer and System Sciences,
36(2):225253, 1988.
[Li80] Shuo-Yen Robert Li. A martingale approach to the study
of occurrence of sequence patterns in repeated experiments.
Annals of Probability, 8(6):11711176, 1980.
[Lin92] Torgny Lindvall. Lectures on the Coupling Method. Wiley, 1992.
[LPW09] David A. Levin, Yuval Peres, and Elizabeth L. Wimmer. Markov
Chains and Mixing Times. American Mathematical Society,
2009.
[LR85] T. L. Lai and Herbert Robbins. Asymptotically efficient adap-
tive allocation rules. Advances in Applied Mathematics, 6(1):4
22, March 1985.
BIBLIOGRAPHY 372
[Lub85] Michael Luby. A simple parallel algorithm for the maximal

independent set problem. In Proceedings of the seventeenth
annual ACM symposium on Theory of computing, pages 110,
New York, NY, USA, 1985. ACM.
[LW05] Michael Luby and Avi Wigderson. Pairwise independence and
derandomization. Foundations and Trends in Theoretical Com-
puter Science, 1(4):237301, 2005.
[McC85] Edward M. McCreight. Priority search trees. SIAM J. Comput.,
14(2):257276, 1985.
[McD89] Colin McDiarmid. On the method of bounded differences. In
Surveys in Combinatorics, 1989: Invited Papers at the Twelfth
British Combinatorial Conference, pages 148188, 1989.
[MH92] Colin McDiarmid and Ryan Hayward. Strong concentration
for Quicksort. In Proceedings of the Third Annual ACM-SIAM
Symposium on Discrete algorithms, SODA 92, pages 414421,
Philadelphia, PA, USA, 1992. Society for Industrial and Applied
Mathematics.
[Mil76] Gary L. Miller. Riemanns hypothesis and tests for primality.
Journal of Computer and System Sciences, 13(3):300317, 1976.
[ML86] Lothar F. Mackert and Guy M. Lohman. R* optimizer vali-
dation and performance evaluation for distributed queries. In
Wesley W. Chu, Georges Gardarin, Setsuo Ohsuga, and Yahiko
Kambayashi, editors, VLDB86 Twelfth International Confer-
ence on Very Large Data Bases, August 25-28, 1986, Kyoto,
Japan, Proceedings, pages 149159. Morgan Kaufmann, 1986.
[MN98] Makoto Matsumoto and Takuji Nishimura. Mersenne twister:
a 623-dimensionally equidistributed uniform pseudo-random
number generator. ACM Trans. Model. Comput. Simul., 8:330,
January 1998.
[Mos09] Robin A. Moser. A constructive proof of the Lovsz local lemma.
In Proceedings of the 41st annual ACM Symposium on Theory
of Computing, STOC 09, pages 343350, New York, NY, USA,
2009. ACM.
[MR95] Rajeev Motwani and Prabhakar Raghavan. Randomized Algo-
rithms. Cambridge University Press, 1995.
BIBLIOGRAPHY 373
[MR98] Michael Molloy and Bruce Reed. Further algorithmic aspects

of the local lemma. In Proceedings of the thirtieth annual ACM
symposium on Theory of computing, STOC 98, pages 524529,
New York, NY, USA, 1998. ACM.
[MRR+ 53] Nicholas Metropolis, Arianna W. Rosenbluth, Marshall N.

Rosenbluth, Augusta H. Teller, and Edward Teller. Equa-
tion of state calculations by fast computing machines. J. Chem.
Phys., 21(6):10871092, 1953.
[MT05] Ravi Montenegro and Prasad Tetali. Mathematical aspects of

mixing times in markov chains. Foundations and Trends in
Theoretical Computer Science, 1(3), 2005.
[MT10] Robin A. Moser and Gbor Tardos. A constructive proof of the

general Lovsz local lemma. J. ACM, 57:11:111:15, February
2010.
[MU05] Michael Mitzenmacher and Eli Upfal. Probability and Comput-

ing: Randomized Algorithms and Probabilistic Analysis. Cam-
bridge University Press, 2005.
[Mun11] Randall Munroe. Sports. http://xkcd.com/904/, 2011. Ac-

cessed September 18th, 2016.
[Pag01] Rasmus Pagh. On the cell probe complexity of membership

and perfect hashing. In STOC, pages 425432, 2001.
[Pag06] Rasmus Pagh. Cuckoo hashing for undergraduates.

Available at http://www.it-c.dk/people/pagh/papers/
cuckoo-undergrad.pdf, 2006.
[Pap91] Christos H. Papadimitriou. On selecting a satisfying truth

assignment (extended abstract). In 32nd Annual Symposium on
Foundations of Computer Science, San Juan, Puerto Rico, 1-4
October 1991, pages 163169. IEEE Computer Society, 1991.
[PPR05] Anna Pagh, Rasmus Pagh, and S. Srinivasa Rao. An optimal

bloom filter replacement. In Proceedings of the Sixteenth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2005,
Vancouver, British Columbia, Canada, January 23-25, 2005,
pages 823829. SIAM, 2005.
BIBLIOGRAPHY 374
[PR04] Rasmus Pagh and Flemming Friche Rodler. Cuckoo hashing.

J. Algorithms, 51(2):122144, 2004.
[PT12] Mihai Patrascu and Mikkel Thorup. The power of simple
tabulation hashing. J. ACM, 59(3):14, 2012.
[Pug90] William Pugh. Skip Lists: A Probabilistic Alternative to Bal-
anced Trees. Communications of the ACM, 33(6):668676, June
1990.
[Rab80] Michael O Rabin. Probabilistic algorithm for testing primality.
Journal of Number Theory, 12(1):128138, 1980.
[Rag88] Prabhakar Raghavan. Probabilistic construction of determin-
istic algorithms: approximating packing integer programs. J.
Comput. Syst. Sci., 37:130143, October 1988.
[Rou01] Salvador Roura. Improved master theorems for divide-and-
conquer recurrences. J. ACM, 48:170205, March 2001.
[RT87] Prabhakar Raghavan and Clark D. Tompson. Randomized
rounding: a technique for provably good algorithms and algo-
rithmic proofs. Combinatorica, 7:365374, December 1987.
[SA96] Raimund Seidel and Cecilia R. Aragon. Randomized
search trees. Algorithmica, 16(4/5):464497, 1996. Avail-
able at http://people.ischool.berkeley.edu/~aragon/
pubs/rst96.pdf.
[She11] Irina Shevtsova. On the asymptotically exact constants in
the Berry-Esseen-Katz inequality. Theory of Probability & Its
Applications, 55(2):225252, 2011. A preprint is available at
http://arxiv.org/pdf/1111.6554.pdf.
[Sho97] Peter W Shor. Polynomial-time algorithms for prime factoriza-
tion and discrete logarithms on a quantum computer. SIAM
journal on computing, 26(5):14841509, 1997.
[Spu12] Francis Spufford. Red Plenty. Graywolf Press, 2012.
[Sri08] Aravind Srinivasan. Improved algorithmic versions of the Lovsz
local lemma. In Proceedings of the nineteenth annual ACM-
SIAM symposium on Discrete algorithms, SODA 08, pages
611620, Philadelphia, PA, USA, 2008. Society for Industrial
and Applied Mathematics.
BIBLIOGRAPHY 375
[SS71] A. Schnhage and V. Strassen. Schnelle multiplikation groer

zahlen. Computing, 7(34):281292, 1971.
[SS87] E. Shamir and J. Spencer. Sharp concentration of the chromatic

number on random graphs Gn,p . Combinatorica, 7:121129,
January 1987.
[Str03] Gilbert Strang. Introduction to linear algebra. Wellesley-

Cambridge Press, 2003.
[SW13] Dennis Stanton and Dennis White. Constructive Combinatorics.

Undergraduate Texts in Mathematics. Springer, 2013.
[Tod91] Seinosuke Toda. Pp is as hard as the polynomial-time hierarchy.

SIAM J. Comput., 20(5):865877, 1991.
[Tof80] Tommaso Toffoli. Reversible computing. Technical Report

MIT/LCS/TM-151, MIT Laboratory for Computer Science,
1980.
[Val79] Leslie G. Valiant. The complexity of computing the permanent.

Theor. Comput. Sci., 8:189201, 1979.
[Val82] Leslie G. Valiant. A scheme for fast parallel communication.

SIAM J. Comput., 11(2):350361, 1982.
[VB81] Leslie G. Valiant and Gordon J. Brebner. Universal schemes

for parallel communication. In STOC, pages 263277. ACM,
1981.
[Vui80] Jean Vuillemin. A unifying look at data structures. Communi-

cations of the ACM, 23(4):229239, 1980.
[VV86] Leslie G. Valiant and Vijay V. Vazirani. Np is as easy as

detecting unique solutions. Theor. Comput. Sci., 47(3):8593,
1986.
[Wil04] David Bruce Wilson. Mixing times of lozenge tiling and

card shuffling Markov chains. Annals of Applied Probability,
14(1):274325, 2004.
[Wil06] Herbert S. Wilf. generatingfunctionology. A. K. Peters, third

edition, 2006. Second edition is available on-line at http:
//www.math.penn.edu/~wilf/DownldGF.html.
BIBLIOGRAPHY 376
[WNC87] Ian H. Witten, Radford M. Neal, and John G. Cleary. Arith-

metic coding for data compression. Communications of the
ACM, 30(6):520540, 1987.
Index
(r1 , r2 , p1 , p2 )-sensitive, 127 oblivious, 98

3-way cut, 332 advice, 216
G(n, p), 84 Aldous-Fill manuscript, 158
, 174 algorithm
-NNS, 127 Deutschs, 232
-PLEB, 127 UCB1, 87
-nearest neighbor search, 127 algorithmic information theory, 104
-point location in equal balls, 127 amplitude
F-measurable, 39 probability, 222
-algebra, 13, 39 ancestor, 98
generated by a random variable, common, 98
39 annealing schedule, 168
2 , 173 anti-concentration bound, 91
d-ary cuckoo hashing, 116 aperiodic, 154
k-CNF formula, 207 approximation algorithm, 201
k-universal, 106 approximation ratio, 201
k-wise independence, 23 approximation scheme
tmix , 153 fully polynomial-time, 190
#DNF, 186, 189 fully polynomial-time randomized,
#P, 185 185
#SAT, 185 arithmetic coding, 188, 322
01 random variable, 359 arm, 86
2-universal, 105 Atari 2600, 246
augmenting path, 183
absolute value, 227 average-case analysis, 2
adapted, 78, 140 AVL tree, 94
adapted sequence of random variables, avoiding worst-case inputs, 10
133 axioms of probability, 13
adaptive adversary, 98 Azumas inequality, 74, 79
Adlemans theorem, 216 Azuma-Hoeffding inequality, 74
adversary
adaptive, 98 balanced, 93
377
INDEX 378
bandit common ancestor, 98

multi-armed, 86 common subsequence, 329
Bernoulli complement, 13
Jacob, 29 complex conjugate, 225
computation
Bernoulli random variable, 29, 359
Berry-Esseen theorem, 91 randomized, 224
biased random walk, 140 computation path, 216
binary search tree, 93 computation tree, 216
balanced, 93 concave, 50
binary tree, 93 concentration bound, 54
binomial random variable, 29 coupon collector problem, 160
bipartite graph, 183 conditional expectation, 6, 33
bit-fixing routing, 72, 180 conditional probability, 18
Bloom filter, 116, 344 conductance, 174
counting, 121 congestion, 178
spectral, 122 conjugate
Bloomjoin, 120 complex, 225
bounded difference property, 82 conjugate transpose, 225
bowling, 301 conjunctive normal form, 201
bra, 225 continuous random variable, 27
bra-ket notation, 225 contraction, 21
branching process, 212 controlled controlled NOT, 229
bubble sort, 238 controlled NOT, 229
convex, 50
canonical paths, 176 convex function, 50
causal coupling, 170 Cook reduction, 185
CCNOT, 229 count-min sketch, 122
chaining, 105 countable additivity, 13
Chapman-Kolmogorov equation, 145 countable Markov chain, 144
Chebyshevs inequality, 54 counting Bloom filter, 121
Cheeger constant, 174 coupling, 153
Cheeger inequality, 175 causal, 170
chromatic number, 84 path, 164
Chutes and Ladders, 356 Coupling Lemma, 153
circuit coupling time, 159, 160
quantum, 222 coupon collector problem, 43, 356
clause, 189, 201, 207 concentration bounds, 160
clique, 197 course staff, xvi
CNOT, 229 covariance, 31, 56
collision, 11, 104, 105 cryptographic hash function, 11
INDEX 379
cryptographically secure pseudoran- dovetail shuffle, 163

dom generator, 215 drift process, 135
cuckoo filter, 119 Dungeons and Dragons, 242
cuckoo hash table, 112
cuckoo hashing, 112 eigenvalue, 170
d-ary, 116 eigenvector, 170
cumulative distribution function, 27 einselection, 227
curse of dimensionality, 127 entropy, 188
cut equation
minimum, 21 Walds, 32
cylinder set, 14 Erds-Szekeres theorem, 303, 330
error budget, 46
data stream computation, 122 Euler-Mascheroni constant, 7
dead cat bounce, 242 event, 12, 13
decoherence, 227 events
decomposition independent, 16
Doob, 80 execution log, 211
dense, 183 expectation, 28
dependency graph, 207 conditional, 33
depth conditioned on a -algebra, 39
tree, 274 conditioned on a random variable,
derandomization, 214 34
detailed balance equations, 149 conditioned on an event, 33
dictionary, 104 iterated
discrete probability space, 13 law of, 37
discrete random variable, 27 linearity of, 28
discretization, 132 expected time, 2
disjunctive normal form, 189 expected value, 2, 28
distance expected worst-case, 1
total variation, 150 exponential generating function, 63
distributed algorithms, 11
distribution, 27 false positive, 117
geometric, 41 filter
stationary, 148 cuckoo, 119
DNF formula, 189 filtration, 40, 133
dominating set, 317, 324 financial analysis, 242
Doob decomposition, 80 Find the Lady, 2
Doob decomposition theorem, 135 fingerprint, 11, 119
Doob martingale, 82, 218 fingerprinting, 11
double record, 288 finite Markov chain, 144
INDEX 380
first passage time, 156 heavy hitter, 122

Flajolet-Martin sketch, 305 high-probability bound, 1
FPRAS, 185 Hoares FIND, 44, 355
fractional solution, 202, 203 Hoeffdings inequality, 74
Frobenius problem, 155 homogeneous, 144
fully polynomial-time approximation hopscotch hashing, 116
scheme, 190 hypercube network, 71, 83
fully polynomial-time randomized ap-
proximation scheme, 185 inclusion-exclusion formula, 16
function independence, 16
measurable, 24 of events, 16
futile word search puzzle, 315 of sets of events, 22
pairwise, 22
gate independent, 16, 22
Toffoli, 229 independent events, 16
generated by, 39 independent random variables, 27
generating function independent set, 197
exponential, 63 index, 105
probability, 51 indicator random variable, 25
geometric distribution, 41 indicator variable, 7
geometric random variable, 41 inequality
graph Azumas, 74, 79
bipartite, 183 Azuma-Hoeffding, 74
random, 24, 84 Chebyshevs, 54
Grover diffusion operator, 233 Cheeger, 175
Hoeffdings, 74
Hadamard operator, 228 Markovs, 48
handwaving, 8, 212 McDiarmids, 82, 83
harmonic number, 7 inner product, 225
hash function, 11, 105 instructor, xvi
perfect, 110 integer program, 201
tabulation, 109 irreducible, 154
universal, 105 iterated expectation
hash table, 11, 104 law of, 37
cuckoo, 112 Iverson bracket, 25
hashing, 11 Iverson notation, 25
cuckoo, 112
hopscotch, 116 Johnson-Lindenstrauss lemma, 128
locality-sensitive, 126 joint distribution, 27
heap, 96 joint probability mass function, 25
INDEX 381
Kargers min-cut algorithm, 21 stochastic, 224

Karp-Upfal-Wigderson bound, 356 transition, 145
ket, 225 unitary, 228
key, 93, 105 maximum cut, 200
KNAPSACK, 190 McDiarmids inequality, 82, 83
Kolmogorov extension theorem, 14 measurable, 27, 39
measurable function, 24
Las Vegas algorithm, 8 measurable sets, 13
law of iterated expectation, 37 measure
law of total probability, 5, 18, 19 probability, 12
lazy Markov chain, 154 measurement, 224
leader election, 11 memoryless, 144
left spine, 100 method of bounded differences, 82
lemma method of conditional probabilities,
coupling, 153 218
Johnson-Lindenstrauss, 128 Metropolis, 156
Lovsz local, 205 Metropolis-Hastings
linear program, 201 convergence, 168
linearity of expectation, 6, 28 Metropolis-Hastings algorithm, 156
Lipschitz function, 169 Miller-Rabin primality test, 4
literal, 207 min-cut, 21
Littlewood-Offord problem, 92 minimum cut, 21
load balancing, 11 mixing
load factor, 105 rapid, 160
locality-sensitive hashing, 126 mixing rate, 173
Lovsz Local Lemma, 205 mixing time, 143, 153
moment, 63
magnitude, 227
moment generating function, 63
Markov chain, 143, 144
Monte Carlo algorithm, 4, 8
lazy, 154
Monte Carlo simulation, 8
Markov chain Monte Carlo, 143
multi-armed bandit, 86
Markov process, 144
multigraph, 21
Markovs inequality, 48
martingale, 77, 133 nearest neighbor search, 126
Doob, 218 network
vertex exposure, 84 sorting, 196
martingale difference sequence, 78 NNS, 126
martingale property, 133 no-cloning theorem, 230
matching, 180 non-uniform, 217
matrix norm, 227
INDEX 382
normal form Miller-Rabin, 4

conjunctive, 201 probabilistic method, 11, 195
disjunctive, 189 probabilistic recurrence, 85
notation probability, 13
bra-ket, 225 conditional, 18
number P, 185 stationary, 148
probability amplification, 214, 216
objective function, 201 probability amplitude, 222
oblivious adversary, 98 probability generating function, 51
open addressing, 105 probability mass function, 25
operator probability measure, 12
Grover diffusion, 233 probability of A conditioned on B, 18
Hadamard, 228 probability of A given B, 18
optimal solution, 201 probability space, 12
optional stopping theorem, 135 discrete, 13
outcomes, 13 probability theory, 12
probing, 105
pairwise independence, 57
product
pairwise independent, 22
inner, 225
paradox
tensor, 223
St. Petersburg, 29
pseudorandom generator, 214
partition, 13
cryptographically secure, 215
perfect hash function, 110
puzzle
perfect matching, 183
word search, 314
perfect matchings, 186
futile, 315
period, 154
periodic, 147 quantum bit, 222
permanent, 186 quantum circuit, 222
permutation routing, 71 quantum computing, 222
Perron-Frobenius theorem, 170 qubit, 222
pessimistic estimator, 219 query
phase, 228 point, 123
pivot, 5 QuickSelect, 44, 355
PLEB, 127 QuickSort, 5, 85
point location in equal balls, 127
point query, 123 radix tree, 304
points, 13 RAM, 1
Poisson trial, 359 random graph, 24, 84
polynomial-time hierarchy, 186 random structure, 11
primality test random variable, 2, 12, 24
INDEX 383
Bernoulli, 29 satisfiability, 207

binomial, 29 satisfiability problem, 201
discrete, 27 satisfy, 201
geometric, 41 scapegoat tree, 94
indicator, 25 search tree
random walk, 138 balanced, 93
biased, 140 binary, 93
unbiased, 138 second-moment method, 57
with one absorbing barrier, 139 seed, 214
with two absorbing barriers, 138 separate chaining, 105
random-access machine, 1 sharp P, 185
randomized computation, 224 sifter, 52
randomized rounding, 192, 202 simplex method, 202
range coding, 322 simulated annealing, 168
rapid mixing, 160 simulation
record, 288 Monte Carlo, 8
double, 288 sketch, 122
red-black tree, 94 count-min, 122
reducible, 147 Flajolet-Martin, 305
reduction, 185 sort
regret, 87 bubble, 238
rejection sampling, 156, 187, 322 sorting network, 196
relax, 201 Space Invaders, 247
relaxation, 201 spanning tree, 341
relaxation time, 173 spare, 301
reversibility, 149 spectral Bloom filter, 122
reversible, 149, 228 spectral theorem, 171
right spine, 100 spine
ring-cover tree, 127 left, 100
root, 3, 93 right, 100
rotation, 94 sports reporting, 242
tree, 94 sprite, 246
routing staff, xvi
bit-fixing, 72, 180 state space, 144
permutation, 71 state vector, 222
run, 327 stationary distribution, 143, 145, 148
stationary probability, 148
sampling, 9, 11 stochastic matrix, 145, 224
rejection, 156, 187, 322 stochastic process, 144
SAT, 348 stopping time, 133, 134
INDEX 384
strike, 301 conjugate, 225

strongly k-universal, 106 treap, 96, 302
strongly 2-universal, 105 tree
submartingale, 78, 80, 134 AVL, 94
subtree, 93 binary, 93
supermartingale, 78, 80, 85, 134 binary search, 93
symmetry breaking, 11 cartesian, 96
radix, 304
tabulation hashing, 109 red-black, 94
teaching fellow, xvi ring-cover, 127
tensor product, 223 rotation, 94
theorem scapegoat, 94
Adlemans, 216 witness, 211
Berry-Esseen, 91 tree rotation, 94
Doob decomposition, 135 truncation, 136
Erds-Szekeres, 303
Kolmogorov extension, 14 UCB1 algorithm, 87
no-cloning, 230 unary, 132
optional stopping, 135 unbiased random walk, 138
Perron-Frobenius, 170 uniform, 217
spectral, 171 unitary matrix, 228
Todas, 186 unitary transformation, 228
Valiants, 185 universal hash family, 105
third moment, 91 unranking, 187
Three-card Monte, 2 upper confidence bound, 87
time
coupling, 159, 160 Valiants theorem, 185
expected, 2 variance, 54
mixing, 153 vector
relaxation, 173 state, 222
stopping, 133, 134 vector space, 38
Todas theorem, 186 vertex exposure martingale, 84
Toffoli gate, 229
Walds equation, 32, 141
total path length, 94
witness, 4, 217
total variation distance, 150
witness tree, 211
transformation
word search puzzle, 314
unitary, 228
futile, 315
transition matrix, 145
worst-case analysis, 2
transition probabilities, 145
transpose

Notes PDF

Uploaded by

Copyright:

Available Formats

Notes PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Notes PDF

Uploaded by

Copyright:

Available Formats

What are some different types of randomized algorithms classified by their goals?

What are some different types of randomized algorithms classified by their goals?

How can randomized algorithms be classified based on their methods?

How can randomized algorithms be classified based on their methods?

Notes on Randomized Algorithms

CPSC 469/569: Fall 2016

List of figures xii

List of tables xiii

List of algorithms xiv

Lecture schedule xix

2.3.1 Application to algorithm analysis . . . . . . . . . . . . 18

4 Basic probabilistic inequalities 46

4.3 Jensens inequality . . . . . . . . . . . . . . . . . . . . . . . . 50

The UCB1 algorithm . . . . . . . . . . . . . . . 87

6 Randomized search trees 93

7.7.1.1 Locality-sensitive hash functions . . . . . . . 127

8 Martingales and stopping times 133

9 Markov chains 143

Single peak with very small amounts of noise . 169

10 Approximate counting 185

11 The probabilistic method 195

11.3.5 Constructive proof . . . . . . . . . . . . . . . . . . . . 210

13 Quantum computing 222

A.4 Assignment 4: due Thursday, 2016-11-03, at 23:00 . . . . . . 249

B Sample assignments from Spring 2014 267

B.7.3 Storage allocation (20 points) . . . . . . . . . . . . . . 290

C Sample assignments from Spring 2013 294

D Sample assignments from Spring 2011 321

D.1.4 All must have candy . . . . . . . . . . . . . . . . . . . 323

E Sample assignments from Spring 2009 351

F Probabilistic recurrences 355

F.3.3 Tossing coins . . . . . . . . . . . . . . . . . . . . . . . 358

4.1 Proof of Jensens inequality . . . . . . . . . . . . . . . . . . . 51

5.1 Comparison of Chernoff bound variants . . . . . . . . . . . . 67

6.1 Tree rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

11.1 Tricky step in MAX SAT argument . . . . . . . . . . . . . . . 204

A.1 Filling a screen with Space Invaders . . . . . . . . . . . . . . 247

B.1 Example of tree contraction for Problem B.3.1 . . . . . . . . 275

C.1 Radix tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304

3.1 Sum of two dice . . . . . . . . . . . . . . . . . . . . . . . . . . 26

5.1 Concentration bounds . . . . . . . . . . . . . . . . . . . . . . 55

7.1 Hash table parameters . . . . . . . . . . . . . . . . . . . . . . 106

7.1 Insertion procedure for cuckoo hashing . . . . . . . . . . . . . 113

A.1 One pass of bubble sort . . . . . . . . . . . . . . . . . . . . . . 238

C.1 Adaptive hash table insertion . . . . . . . . . . . . . . . . . . . 308

D.1 Dubious duplicate detector . . . . . . . . . . . . . . . . . . . . 343

On-line course information

Reserved books at Bass library

Use of outside help

Clarifications for homework assignments

2016-08-31 Randomized algorithms. What they are and what we will do

2016-09-02 Randomized Quicksort analyzed using linearity of expectation.

2016-09-07 Law of total probability. Random variables and expectations.

2016-09-12 Expectation conditioned on a random variable. The union

2016-09-14 More probabilistic inequalities: Markovs inequality, Jensens

2016-09-19 Chernoff bounds and applications. Valiants randomized rout-

2016-09-21 Concentration bounds for sums of bounded independent random

2016-09-26 Applications of the method of bounded differences and Azuma-

2016-09-28 More concentration bounds: the UCB1 algorithm for multi-

2016-10-24 More locality-sensitive hashing: constructing an -PLEB. Mar-