MSRI Sheet 1
MSRI Sheet 1
MSRI Sheet 1
Exercise Sheet 1
Convexity, Martingales and Concentration inequalities
We do not expect you to solve all the exercises in this sheet. You can use these
exercises as useful references and perhaps to practice in the future.
Easy
Exercise 1
Goal of this exercise is to become familiar with the concept of convexity and subdifferential.
Exercise 2
Let Mt be an (Ft )-martingale. Prove that:
1. E[Mt ] = E[M0 ], for all t ≥ 0.
2. max(Mt , 0) is an (Ft )-sub-martingale.
3. for any convex function φ such that φ(Mt ) is integrable for all t, φ(Mt ) is an (Ft )-sub-
martingale.
Exercise 3
Let (Xi )i∈N be a sequence of bounded, Pt independent random variables with mi := E[Xi ] and
σi2 := Var(Xi ) for all i ∈ N. Set St = i=1 Xi and Ft = σ(X1 , . . . , Xt ).
1. Find a sequence (at ) of real numbers such that St + at is an Ft -martingale.
2. Find two sequences (at ) and (bt ) of real numbers such that St2 + at St + bt is an Ft -martingale.
3. Fix λ ∈ R. Find a sequence (aλt ) such that (aλt )−1 exp(λSt ) is an Ft -martingale.
Intermediate
Exercise 4
The goal of this exercise is to derive Chernoff ’s inequality and apply it (for more examples of
concentration inequalities, see [3]).
Let us denote the moment generating function MX (t) = E[etX ].
1 The claim holds also for the Lebesgue measure.
1
MSRI Summer School July 3, 2023
1. Prove
P(X ≥ α) ≤ inf MX (t)e−tα .
t≥0
2. Compute the moment generating function of a Bernoulli random variable with parameter p.
3. Let XiPbe independent Bernoulli random variables with parameter pi . Consider their sum
N
SN = i=1 Xi , and denotes its mean by µ = E[SN ]. Prove, for any t > µ, we have
eµ t
P(SN ≥ t) ≤ e−µ .
t
Exercise 5
Let (Xn )n∈N be a stochastic process that starts at X0 = 1, such that Xn ∈ {0, 2n } for all n, and
is defined by the following operation:
(
Xn+1 = 0 if Xn = 0
n
Xn+1 ∼ Unif {0, 2 } if Xn ̸= 0
Exercise 6
The goal of
this exercise is to derive some concentration inequalities on the hypercube2 .
1 1 n
n
Let Q = − 2 , 2 be the unit cube.
√
1. Show that diam(Qn ) = n.
2. Let X = (X1 , . . . , Xn ) and Y = (Y1 , . . . , Yn ) be two independent random vectors that are
distributed uniformly on the cube. Show
n
E[||X − Y ||22 ] = .
6
What can you conclude?
√ √ √
3. Consider Q̃n = 12Qn = [− 3, P3]n . Let Y be a random vector distributed on Q̃n . Apply
the central limit theorem on √1n i Yi .
Exercise 7
Let KR ⊂ Rn be a convex set and suppose Voln (K) = 1. Let f : K → R be a C 1 -smooth function
with K f dλ = 0, where λ is the Lebesgue measure.
1. Derive Z Z Z
1
f 2 (x)dλ(x) = |f (x) − f (y)|2 dxdy .
K 2 K K
2 In [2] you can find a similar derivation for the Euclidean ball as well.
2
MSRI Summer School July 3, 2023
2. Show that
Z 1
f (x) − f (y) = ∇f (1 − t)x + ty (y − x)dt .
0
3. Show that
Z Z 1 Z Z
2 2
2
f (x)dλ(x) = diam (K) ∇f (1 − t)x + ty dxdydt ,
1
K 2 K K
4. Show that
Z Z
2
dy ≤ 2n |∇f (z)|2 dz.
∇f (1 − t)x + ty
K K
5. Conclude Z Z
f 2 dλ ≤ CP (K) |∇f |2 dλ,
K K
Exercise 8
2. Prove that if (Xt )t∈[0,∞) isa martingale and f : R → R is convex (but not necessarily
increasing), then, if f (Xt ) t∈[0,∞) ∈ L1 for all t ≥ 0, we have that f (Xt ) t∈[0,∞) is a
submartingale.
3. Give an example of a martingale (Xt )t∈[0,∞) and a convex f : R → R such that f (Xt ) t∈[0,∞)
is a submartingale but not a martingale.
Advanced
Exercise 9
The following theorem (see [3]) proves that dense graphs are almost regular.
Prove there is an absolute constant C such that the following holds: consider a random graph
G ∼ G(n, p) with expected degree satisfying d ≥ C log n. Then, with probability at least 0.9, the
following occurs: all vertices of G have degrees between 0.9d and 1.1d.
Exercise 10
The following inequality is known as Bennett’s inequality. (see [3] for a more general statement).
Let X1 , . . . , XN be i.i.d. random variables. Assume that E[Xi ] = 0, Var(Xi ) = σ 2 and |Xi |≤ M ∀i.
Prove, for any t > 0, that we have
N
!
σ2
X Kt
P Xi ≥ t ≤ exp − 2 h ,
i=1
K σ2
Exercise 11
Let f : [0, 1] → R be a Lipschitz function. Prove there ∃g measurable and bounded such that
∀x ∈ R it holds Z x
f (x) = f (0) + g(t)dt .
0
Hint [1]: let X ∼ Unif[0, 1]. Find two suitable sequences of r.v. (Xn ) and (Zn ) depending on X.
3
MSRI Summer School July 3, 2023
References
[1] Max Fathi. Probabilités M1. 2023.
[2] Bo’az Klartag. Regularity through convexity in high dimensions. 2014.
[3] Roman Vershynin. High-Dimensional Probability. An Introduction with Applications in Data
Science. Cambridge University Press, 2018.