Random walks and algorithms - handout exercises

1. (M) Prove that there exist constants d ∈ N and c > 0 such that the following holds: let
G(n, d) be a uniform random d-regular graph on n vertices. Let λ1 (G) ≥ λ2 (G) ≥ ... ≥
λn (G) be the eigenvalues of its adjacency matrix. Then
 
P λ1 (G(n, d)) − λ2 (G(n, d)) > cd → 1, as n → ∞

Remark: Picking a uniform graph out of all possible d-regular graphs is not a very
tractable construction. It turns out that there is a constructive way to do it, which is much
easier to work with. The idea is to consider the n vertices, each connected to d half-edges
and take a uniform perfect matching on these nd half-edges to determine how to connect
them. If we condition on the resulting graph to be simple (hence, to have no loops and
no double-edges), the graph we get turns out to be a uniform sample of a G(n, d) graph
(why?).

2. Let K be a convex body in Rd , suppose Rthat it is encapsulated between a ball of radius 1

xdx
and a ball of radius n10 . Define b(K) = RK dx to be the barycenter of K. Suppose you’re
K
given access to K via a membership oracle (= a black box which answers the question ”is
x ∈ K).

(a) (M) Find a polynomial time algorithm to approximate b(K) up to a small error.
Hence prove that for every ε > 0 the algorithm generates a point Y such that P(|Y −
b(K)| < ε) > 1 − o(1) using poly(n, ε) steps.
(b) (M) Do the same thing for estimating the covariance matrix of K, hence, if X is
a random vector uniformly distributed in K, find a polynomial time algorithm that
approximates its covariance matrix

Cov(X) := E[(X − E[X])(X − E[X])T ]

in the sense that the algorithm output a matrix M such that 2Cov(X)  M 
Cov(X) with high probability.

3. (M) Let f (x) be a density in Rn with bounded support. We define the ”hit-and-run”
Markov chain (whose state space is Rn ) as follows: suppose that we are at some position
Xt . Then Xt+1 is generated by: (i) Generating a uniform random direction θ ∈ Sn−1 .
(ii) Considering the one-dimensional affine subspace Xt + span(θ), and taking Xt to be
a random point in this subspace picked according to the density proportional to f (x).
Prove that the stationary distribution of this Markov chain is f (x). Bonus: suppose that
f (x) is the indicator of a convex body. Can you come up with an idea how to prove that
the chain mixes in polynomial time?

1
4. (M) Define the 1 − 3-tree as the following infinite tree. Fix a root with two children. Now
suppose that all nodes except for the root has either one or three children, so that the k-th
level has 2k nodes (hence half the nodes on each level have 3 descendants and the other
half have 1 descendant) and so that if the tree is drawn on a plane, the degree of the nodes
is monotone from left to right (thus, a node of degree 4 always has an ancestor of degree
4, unless it’s the root). Is the random walk on this tree transient or recurrent? How many
distinct infinite rays does this tree have?

5. (R) Consider the following self-interacting random walk on Z2 : whenever a vertex is

reached by the walker for the first time, the walker takes a step either right or left with
probabilities (1/2, 1/2) and whenever the walker reaches a vertex which has already been
visited, the walker goes either up or down with probabilities (1/2, 1/2). Is this random
walk transient or recurrent? Is there limiting shape for the range? (hence, is it true that
after rescaling the space properly the set of vertices visited by the random walk before
time n looks like an interval?).

6. (E) Let Z1 , ..., ZT be random variables. Let X1 , ..., XT be real valued random variables
such that Xi is a (deterministic) function of Zi and such that for every 1 ≤ t < T
and z1 , .., zt , one has that the increment Xt+1 − Xt , conditioned on the event {Z1 =
z1 , ..., Zt = zt }, is a Gaussian with mean zero and variance at most 1. Prove that
√
P(|XT | > λ T ) < 2 exp(−λ2 /2), ∀λ > 0.

7. Let X0 , X1 , X2 , ... be an irreducible and aperiodic Markov chain with stationary measure
π and let S0 , S1 , S2 , ... be the associated evolving set process with S0 = {X0 } = {x0 }
being some deterministic point. Denote by v(t) the total-variation distance between Xt
and the stationary measure.
π(y)
(a) (E) Prove that P(Xt = y) = π(x0 )
P(y ∈ St ) for all y, t.
(b) (M) Let T be the first time that St is either the empty-set or the whole state-space.
Prove that
1
v(t) ≤ P(T > t).
π(x0 )
(c) (H) Prove that
C p
v(t) ≤ E min(π(St ), 1 − π(St )).
π(x0 )
(d) (H) Can you think of a way to use this idea to show that a Cheeger inequality implies
mixing bounds?

8. (M) Suppose that you’re given N directions θi ∈ Sn−1 , for 1 ≤ i ≤ N . Suggest an

algorithm to approximate the spherical surface area of the set

S = Sn−1 ∩ {x; hx, θi i > 0, ∀1 ≤ i ≤ N }

up to a small multiplicative error, with complexity poly(nN ).

9. (H) The goal of this question is to find a variant of the constructive hypergraph discrep-
ancy algorithm for the case where, instead of ±1, one assigns to each point a value from
some given subset of Rd .

2
 Let L = {s1 , .., sk } ⊂ Rd such that ksi k2 ≤ 1 for all i and such that L contains the origin
in its convex hull. Let V = [n] be a vertex set, and let S ⊂ 2V be a family of m subsets.
The goal is to find a ”coloring” ξ : V → L such that for all S ∈ S,

X
ξ(v) < D



v∈S 2

for a number D as small as possible. Find a way to tweak the algorithm we saw in class
so that this can be done for
p
D = C(d) n log((m + n)/n)

where C(d) depends only on d. What is the best dependence on d that can be attained?

10. (E) Let S1 , S2 , ... be the evolving set process associated with a Markov chain with state
space S. Denote by K(S, S1 ) its transition kernel. Consider the Doob transform associ-
ated to the process (what we called in class ”modified process”) with transition kernel
µ(S1 )
K̃(S, S1 ) = K(S, S1 )
µ(S)
where µ is the stationary measure of the Markov chain. Prove that the modified process
is the same as the original process conditioned on the event {Si 6= ∅ for all i}.

11. (M) In [Andersen-Peres, Proposition 5] (the paper on the evolving sets process), it is
shown that for most points in a set A whose conductance is φ, the escape time from A
(hence the typical time it takes a random walk to exit the set A) is bounded by O(1/φ).
Note that this bound is not tight, for example, if A is an interval in Z (in this case, the
escape time is of the order 1/φ2 ). Can you find a family of bounded-degree graphs and
respective subsets A where this bound is actually tight? Bonus: (H?) can you think of
a natural sufficient condition on the graph and on the set under which this bound can be
improved to 1/φ2 ?

12. Consider the random walk X0 , X1 , .. on K = {−1, 1}d where X0 = (1, ..., 1) and at each
step, a coordinate is selected at random, and is then toggled with probability 21 . Let u be
the uniform measure on K and define vd (t) to be the total variation distance between the
law of Xt and u.

(a) (E) Prove that for a universal constant C > 0 large enough one has that

lim vd (Cd log d) = 0

d→∞

(hence, the random walk mixes in less than Cd log d steps).

(b) (H) Prove that for every δ > 0 one has

lim vd (1 − δ) 21 d log d = 1


d→∞

and on the other hand

lim vd (1 + δ) 12 d log d = 0


d→∞

This phenomenon is called ”cutoff”, which intuitively means that the mixing of the
random walk occurs in a small window around the time 12 log2 d.

3
13. (M) Let K ⊂ Rd be a convex body. Assume that D ⊂ K ⊂ RD, where D is the unit
Euclidean ball. Let f : K → [−M, M ] be a function with Lipschitz constant L. A.
Provided access to a membership oracle to K and aRfunction which evaluates f in time
O(1) for a point x ∈ K, find a way to approximate K f (x)dx up to an additive error of
ε (with high probability), with complexity poly(d, ε, L, M, R). B. What happens if one
omits the assumption that f is L-Lipschitz? How should the algorithm be modified to
overcome this?

14. (M) The Binary Branching Random walk is defined as follows: a single particle starts
from a prescribed vertex (say, the origin). At every time step, every particle duplicates
itself, dividing into two particles, which will now each do an independent random walk,
so that after t iterations one has 2t particles, which all do an independent step. Prove that
this random walk is recurrent in Z3 in the sense that there are infinite times in which the
origin will be occupied by some particle. Is it true that from some point in time and on
the origin will always be occupied?

15. Let G be a a connected infinite graph with degrees bounded by d, rooted at some vertex
v0 . Let v0 = X0 , X1 , X2 , ... be a random walk starting from v0 . Use evolving sets to
prove the following claims:

(a) (M) For every time t > 0 one has

c(d)
P(Xt = v0 ) < √
t
for a constant c(d) depending only on d.
(b) (M) Suppose now that G is an expander in the sense that for every A ⊂ G one has
|∂A|
|A|
> φ > 0. Prove that for all t > 0,

P(Xt = v0 ) < C(d, φ) exp(−tc(d, φ)).

with C(d, φ), c(d, φ) > 0.

(c) (H) Suppose that you don’t know what G is, you only know that it has a bounded
degree. You’re given the sequence t1 , t2 , .... of return times of Xt to v0 . Can you
determine whether or not the graph G is finite or infinite only by looking at this
sequence?