On random walks on The random graph

Laurent Miclo
(Toulouse School of Economics)

Joint work with Sourav Chatterjee and Persi Diaconis

(Stanford University)
Plan of the talk

1 Sphere random walks

2 On The random graph

3 Spectral gap

4 A Lyapounov argument

5 Hardy’s bounds

6 On a deterministic Rado graph

Plan

1 Sphere random walks

2 On The random graph

3 Spectral gap

4 A Lyapounov argument

5 Hardy’s bounds

6 On a deterministic Rado graph

The random graph

Consider The random graph G B pZ` , E q, where each undirected

edge tx, y u Ă Z` , with x ‰ y , belongs to E with probability 1/2,
independently for all of them.
When the probability 1{2 is replaced by ρ P p0, 1q, the
corresponding notions will receive ρ in index.
For any x P Z` , define

Npxq B ty P Z` : tx, y u P E u

the sphere of radius 1 around x for G.

An associated Markov kernel

Consider the probability measure Q given on Z` by

1
@ x P Z` , Qpxq B
21`x
and associate a Markov kernel K on Z` via

Qpy q
@ x, y P Z` , K px, y q B 1 py q
QpNpxqq Npxq

This kernel is reversible to the probability π given by

@ x P Z` , πpxq “ Z ´1 QpxqQpNpxqq

where Z ą 0 is the normalizing (random) constant.

Sphere "random walks"

For x P Z` , let pXtx qtě0 be a Markov process starting from x and

whose generator is K ´ Id.
We are interested in its speed of convergence to π. Define the
mixing time
" *
x x 1
τ B min t ě 0 : }LpXt q ´ π}tv ď
2

The main result of the talk is:

Theorem 1
There exist two (random) constants a, b ą 0 such that for any
x P Z` ,

τx ď bp1 ` log˚a pxqq

Iterated logarithms

Recall that for a ą 0, the iterated logarithm log˚a is defined as

follows.
Consider the smallest xa ě 0 such that for any x ě xa , we have

logpxq
loga pxq B ď x
logpaq

If x P r0, xa s, by definition, we take log˚a pxq “ 0. Otherwise log˚a pxq

is the minimal number of time one has to iterate loga , starting from
x, to get a number below xa . Namely

log a ˝ loga ˝ ¨ ¨ ¨ ˝ loga pxq ď xa

loooooooooooomoooooooooooon
log˚
a times

The function log˚a grows very slowly, as mentioned by Persi.

Plan

1 Sphere random walks

2 On The random graph

3 Spectral gap

4 A Lyapounov argument

5 Hardy’s bounds

6 On a deterministic Rado graph

The parent vertex

For any x P Z` , define the "parent" vertex of x by

ppxq B min Npxq “ arg max Q

Npxq

and consider the event

B B t@ x P N, ppxq ă xu

It happens with positive probability:

Lemma 2
We have PrBs ě 1{4. More generally, for any ρ P p0, 1q, we have
Pρ rBs ą 0.
Proof of Lemma 2 (1)
For the second statement, define Ax , for x P N, as the event that x
is not linked in Gρ to a smaller vertex. Namely,
č
Ax B tBty ,xu “ 0u
y PJ0,x´1K

These events are independent and PrAx s “ p1 ´ ρqx . We have

č
B “ Acx
xPN

Simple computations lead to

¨ ˛´1
ÿ
Pρ rBs “ ˝ ppnqp1 ´ ρqn ‚
nPZ`

where ppnq is the number of partitions of n. Since this quantity

?
behaves like an exponential of n for large n, we get Pρ rBs ą 0.
Proof of Lemma 2 (2)
For the first bound, we could try to use upper bound on the
partition numbers. It is simpler to use Kounias-Hunter-Worsley
bound for unions of pairwise independent events:
» fi ¨ ˛
ď ÿ ÿ
P– Ax fl ď 1 ^ ˝ PrAx s ´ PrA1 s PrAy s‚
xPJnK xPJnK y PJ2,nK

using that

PrA1 s ě PrA2 s ě ¨ ¨ ¨ ě PrAn s

We get
« ff
ď
c
P rB s “ P Ax
xPN
1 1 1 3
ď lim ` ´ n`1 “
nÑ8 2 4 2 4

A tree

Consider the set of edges

F B ttx, ppxqu : x P Nu

and the corresponding graph T B pZ` , F q.

Under B, it is clear that T is a tree. In fact this is always true:
Lemma 3
The graph T is a tree.

Proof
To show T does not contain cycles, note that when y “ ppxq and
z “ ppy q, then z ă x, because tx, zu Ă Npy q.
Furthermore T is connected, since following p, one ends up being
decreasing and attaining 0.

An estimate on p
Lemma 4
A.s. there exists only a finite number of x P N such that
ppxq ą 2 log2 p1 ` xq. In particular, a.s. there exists a finite
(random) K ě 2 such that

@ x P N, ppxq ď K logp1 ` xq

Proof
ÿ
Prppxq ą 2 logp1 ` xqs
xPN
ÿ
“ PrBt0,xu “ 0, Bt1,xu “ 0, ..., Btt2 logp1`xqu,xu “ 0s
xPN
ÿ 1
“
xPN
21`t2 logp1`xqu
ÿ 1
ď ă `8 
xPN
p1 ` xq2
Plan

1 Sphere random walks

2 On The random graph

3 Spectral gap

4 A Lyapounov argument

5 Hardy’s bounds

6 On a deterministic Rado graph

Spectral gap

Our goal is to obtain a positive spectral gap, first under B.

Proposition 5

On B, there exists a random constant Λ ą 0 such that

@ f P L2 pπq, Λπrpf ´ πrf sq2 s ď Epf q

where in the r.h.s. E is the Dirichlet form defined by

1 ÿ
@ f P L2 pπq, Epf q B pf py q ´ f pxqq2 πpxqK px, y q
2 x,y P Z
`
First Dirichlet eigenvalue

Since πrpf ´ πrf sq2 s ď πrpf ´ f p0qq2 s, the previous result is an

immediate consequence of the existence of a positive first Dirichlet
eigenvalue:
Proposition 6

On B, there exists a random constant Λ ą 0 such that

@ f P L2 pπq, Λπrpf ´ f r0sq2 s ď Epf q (1)

The proof of Proposition 6 is based on the pruning of G into T and

on the resort to Cheeger’s inequalities for trees.
Pruning (1)

Define the Markov kernel KT via

$
& K px,řy q , if tx, y u P F
@ x, y P Z` , KT px, y q B 1 ´ zPZ` ztxu KT px, zq , if x “ y
%
0 , otherwise

the Dirichlet form ET given by

ÿ
@ f P L2 pπq, ET pf q “ Z pf py q ´ f pxqq2 πpxqK px, y q
tx,y uPF

and the (non-negative) measure µ through

@ x P N, µpxq B QpxqQpppxqq (2)

Pruning (2)
Proposition 7

On B, there exists λ ą 0 such that

@ f P L2 pµq, λµrpf ´ f p0qq2 s ď ET pf q (3)

Proposition 6 follows with Λ B λ{2, due to the inclusion

Npxq Ă Jppxq, 8J and to the exponential feature of Q implying
@ x P N, Qpppxqq ď QpNpxqq ď 2Qpppxqq
Indeed,
λ ÿ
λπrpf ´ f p0qq2 s “ pf pxq ´ f p0qq2 QpxqQpNpxqq
Z xPN
2λ ÿ
ď pf pxq ´ f p0qq2 QpxqQpppxqq
Z xPN
2λ 2
“ µrpf ´ f p0qq2 s “ ET pf q ď 2Epf q
Z Z
Dirichlet-Cheeger
For any A Ă N, define BA B ttx, y u : x P A, y R Au. Endow the
set of edges with the measure ν induced by

νptx, y uq B Z πpxqKT px, y q

Define the Dirichlet-Cheeger constant

νpBAq
ι B inf ě 0
APA µpAq

where

A B tA Ă N : A ‰ Hu

The Dirichlet-Cheeger inequality states

ι2
λ ě
2
On B (1)

Proposition 8
On B, we have ι ě 1{2.

Proof
Decomposing an element of A into its T -connected components
and including each component into the substree Ta generated its
smallest element a, we get

νpBTa q
ι “ inf
aPN µpTa q
QpaqQpppaqq
“ inf
aPN µpTa q
On B (2)

Note that on B,

@ x P Ta , ppxq ě ppaq
Ta Ă Ja, 8J

We deduce
ÿ
µpTa q “ QpxqQpppxqq
xPTa
ÿ
ď Qpppaqq Qpxq
xPTa
ÿ
ď Qpppaqq Qpxq
xPJa,8J
“ 2QpppaqqQpaq


Ouside B
In the general case, note from Lemma 4 that there exists a
(random) vertex x0 P Z` such that

@ x ą x0 , ppxq ă x

We deduce there exists x1 ě x0 such that

@ a ą x1 , @ x P Ta , ppxq ă x

and from the above proof

νpBTa q 1
inf ě
aąx1 µpTa q 2

By finiteness of Jx1 K, we also have

νpBTa q
inf ą 0
aPJx1 K µpTa q

and thus ι ą 0.
Plan

1 Sphere random walks

2 On The random graph

3 Spectral gap

4 A Lyapounov argument

5 Hardy’s bounds

6 On a deterministic Rado graph

Coming back close to 0
The spectral gap and Lemma 4 imply the bound
1
}LpXtx q ´ π}tv ď a e ´Λt
πpxq
ď px ` 1qK 2x{2 e ´Λt

suggesting a mixing time of order logpx ` 2q starting from x. This

is ok for small x. Fixing x0 P Z` to be specified later on, it leads us
to consider

Sx B inftn P Z` : Xnx ď x0 u

Lemma 9
There exist (random) constants x0 P Z` and a ą 0 such that

@ x P Z` , ErS x s ď 2 log˚a pxq

Proof of Theorem 1

Let be given t ě 0 and A Ă Z` . Conditioning by the events before

S x , say FS x , we get for any x P Z` ,

|LpXtx qrAs ´ πpAq| “ |PrXtx P As ´ πpAq|

ď |PrXtx P A|FS x , S x ď t{2s ´ πpAq|
`PrS x ą t{2s

On one hand, using the strong Markov property at S x and the

bound deduced from the spectral gap, and on the other hand
resorting to Lemma 9, we get
n 4
|LpXtx qrAs ´ πpAq| ď px0 ` 1qK 2x0 {2 e ´Λt{2 ` log˚a pxq
2 t
and the r.h.s. can be made as small as we desire, and uniformly in
A, by the choice of t “ bp1 ` log˚a pxqq and b large enough. 
Proof of Lemma 9 (1)

It is sufficient to work with the imbedded Markov chain, abusively

denoted the same. Recalling Lemma 4, define

@ n P Z` , Znx B log˚a pXnx q

1
with a B 2 2K .
Lemma 9 will be proven if we can find x0 (independent from x)
such that
n ^ Sx
ˆ ˙
x
Zn^S x `
2 n

is a supermartingale. Indeed, letting n go to infinity in

n ^ Sx
„ 
x
E Zn^S x ` ď Z0x “ log˚a pxq
2

we get the desired bound.

Proof of Lemma 9 (2)

The previous supermartingale property amounts to see that for any

x ě x0 ,
1
ErZ1x ´ Z0x s ď (4)
2
Indeed, consider y B 2K logpxq “ loga pxq. For z ď y , we have

log˚a pzq ď log˚a py q “ log˚a ploga pxqq “ log˚a pxq ´ 1

thus

ErpZ1x ´ Z0x q1X1x ďy s ď p´1qPrX1x ď y s “ ´1 ` PrX1x ą y s

Proof of Lemma 9 (3)

We deduce, via Lemma 4,

ÿ Qpzq
ErZ1x ´ Z0x s ď ´1 ` p1 ` log˚a pzqq
ząy
QpNpxqq
1 ÿ
ď ´1 ` p1 ` log˚a pzqqQpzq
Qpppxqq ząy
1 ÿ
ď ´1 ` p1 ` log˚a pzqqQpzq
QptK logpxquq ząy
1 ÿ
“ ´1 ` p1 ` log˚a pzqqQpzq
Qpty {2uq ząy

Due to the exponential feature of Q, the last term of the r.h.s. is as

small we desire for y large enough. We can thus find x0 P Z` , such
that (4) holds for any x ě x0 .

Plan

1 Sphere random walks

2 On The random graph

3 Spectral gap

4 A Lyapounov argument

5 Hardy’s bounds

6 On a deterministic Rado graph

A family of subtrees
Since the above approach is based on the analysis of the smallest
Dirichlet eigenvalue on trees whose root is absorbing, we can also
resort to Hardy’s inequalities.
Let us recall the principle. See T as a tree rooted in 0 and for any
x P Z` , denote by hpxq the height of x in T . Consider T the set of
all subtrees T of T satisfying the conditions
T does not contain 0,
there exists M ě 1 such that hpxq ď M for all x P T ,
if x P T has a child in T , then all children of x belong to T .
Each T P T has a root that is denoted r pT q. When T is not
reduced to a singleton, the set of sons of r pT q will be denoted
SpT q (it is the same in T and in T ). For y P SpT q, write Ty for
the set of all offsprings of y in T , so we have the decomposition
ğ
T “ tr pT qu \ Ty
y PSpT q
Notions relative to T

Recall we have defined a functional on the edges:

νptx, y uq B Z πpxqKT px, y q

(where Z is the normalization in π). We now extend it on T (no

longer as a measure) via the iteration
when T is the singleton tr pT qu, we take

νpT q B νptr pT q, ppr pT qquq

when T is not a singleton, then ν satisfies

1 1 1
“ `ř
νpT q νptmpT quq y PSpT q νpTy q

Also for T P T, let T ˚ be the set of all offsprings in T of the leaves

of T (themselves included).
Hardy’s inequality for trees

Consider S Ă T the set of T P T which are such that r pT q is a son

of 0.
Finally define

µpT ˚ q
A B sup
T P S νpT q

The interest of this quantity is the Hardy inequality:

1
A ď ď 16A (5)
λ
where λ is the best constant in Proposition 7, namely the smallest
Dirichlet eigenvalue for the Markov process associated to KT and
absorbed at 0.
An alternative proof of the spectral gap

It is sufficient to show that A ă `8 (a.s.).

It is possible to do so, based on Lemmas 2, 3 and 4. This proof is
more involved than the use of the Cheeger’s inequality. But a priori
it provides a better estimate, since (5) gives almost matching upper
and lower bounds.

To illustrate the improvement, let us turn to a deterministic

situation.
Plan

1 Sphere random walks

2 On The random graph

3 Spectral gap

4 A Lyapounov argument

5 Hardy’s bounds

6 On a deterministic Rado graph

The Rado graph

On the set of vertices Z` , an edge is put between x ă y when in

the dyadic decomposition of y , the coefficient in front of 2x is 1.
Exemple: 9 “ 20 ` 23 , so the edges t0, 9u and t3, 9u belong to the
Rado graph, called G .
Thus for any given x P Z` , the neighbors y ą x are exactly the
n2x , where n is an odd number.
The graph G is isomorphic to G, a.s.
All the previous notions, such as the set of neighbors Npxq or the
parent ppxq of a vertex x P Z` , are defined as before.
Associated sphere random walks

For any δ P p0, 1q, let Q be the probability given by

@ x P Z` , Qpxq B p1 ´ δqδ x

and consider the associated sphere random walk.

The previous arguments can be extended (simplified in fact) to get
Theorem 10
There exists a constant b ą 0 depending on δ such that for any
x P Z` , the mixing time τ x satisfies

τx ď bp1 ` log˚2 pxqq

Cheeger’s vs Hardy’s bounds

Concerning the proof of Proposition 7, giving an estimate of a

smallest Dirichlet eigenvalue on a tree rooted in 0, both Cheeger or
Hardy methods are available. Cheeger’s inequality leads to the
bound
p1 ´ δq2
λ ě (6)
2
while Hardy’s inequality implies
1´δ
λ ě
16p2 _ rlog logp2{ logp1{δqqsq

which is better than (6) as δ goes to 1´ (δ “ 1 would correspond

to the problematic case “pick a neighbor at random”).
Questions

The previous models lead to a lot of unanswered questions. Here

are two examples for the determinist Rado graph.
It can be shown the smallest Dirichlet eigenvalue of
Proposition 7 is positive if and only if Qpxq is of the same
order as QpJx, 8Jq. Is this assertion also true for the spectral
gap of the corresponding sphere random walk? More generally
and in the spirit of [Benjamini and Schramm, Every graph with
a positive Cheeger constant contains a tree with a positive
Cheeger constant, 1997], does a Markov process with a
positive Cheeger constant “contain” a Markov process induced
on a tree with a positive Cheeger constant?
For the sphere Markov process on the Rado graph, do the
pruning procedure and the Hardy’s estimate give the right
order of the spectral gap?