Recurrence vs Transience: An introduction to

random walks
Pablo Lessa∗
May 24, 2015

Preface
These notes are aimed at advanced undergraduate students of Mathematics.
Their purpose is to provide a motivation for the study of random walks in a
wide variety of contexts. I have chosen to focus on the problem of determining
recurrence or transience of the simple random walk on infinite graphs, especially
trees and Cayley graphs (associated to a symmetric finite generator of a discrete
group).
None of the results here are new and even less of them are due to me.
Except for the proofs the style is informal. I’ve used contractions and many
times avoided the use of the editorial we. The bibliographical references are
historically incomplete. I give some subjective opinions, unreliable anecdotes,
and spotty historical treatment. This is not a survey nor a text for experts.
This work is dedicated to François Ledrappier.

Contents
1 Introduction 2
1.1 Pólya’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 The theory of random walks . . . . . . . . . . . . . . . . . . . . . 3
1.3 Recurrence vs Transience: What these notes are about . . . . . . 7
1.4 Notes for further reading . . . . . . . . . . . . . . . . . . . . . . . 11

2 The entry fee: A crash course in probability theory 13

2.1 Probability spaces, random elements . . . . . . . . . . . . . . . . 13
2.2 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Markov chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 The strong Markov property . . . . . . . . . . . . . . . . . . . . . 18
∗ Partially supported by /CNPq/-Brazil project number 407129/2013-8.

1
 2.6 Simple random walks on graphs . . . . . . . . . . . . . . . . . . . 19
2.7 A counting proof of Pólya’s theorem . . . . . . . . . . . . . . . . 21
2.8 The classification of recurrent radially symmetric trees . . . . . . 23
2.9 Notes for further reading . . . . . . . . . . . . . . . . . . . . . . . 25

3 The Flow Theorem 25

3.1 The finite flow theorem . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 A proof of the flow theorem . . . . . . . . . . . . . . . . . . . . . 27
3.3 Adding edges to a transient graph cannot make it recurrent . . . 28
3.4 Recurrence of groups is independent of finite symmetric generat-
ing set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5 Boundary and Core of a tree . . . . . . . . . . . . . . . . . . . . 29
3.6 Logarithmic capacity and recurrence . . . . . . . . . . . . . . . . 30
3.7 Average meeting height and recurrence . . . . . . . . . . . . . . . 31
3.8 Positive Hausdorff dimension implies transience . . . . . . . . . . 32
3.9 Perturbation of trees . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.10 Notes for further reading . . . . . . . . . . . . . . . . . . . . . . . 34

4 The classification of recurrent groups 35

4.1 Sub-quadratic growth implies recurrence . . . . . . . . . . . . . . 35
4.2 Growth of groups . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.4 Isoperimetric inequalities . . . . . . . . . . . . . . . . . . . . . . 37
4.5 Growth of order 3 or more implies transience . . . . . . . . . . . 39
4.6 Notes for further reading . . . . . . . . . . . . . . . . . . . . . . . 40

1 Introduction
1.1 Pólya’s Theorem
A simple random walk, or drunkard’s walk, on a graph, is a random path ob-
tained by starting at a vertex of the graph and choosing a random neighbor
at each step. The theory of random walks on finite graphs is rich and inter-
esting, having to do with diversions such as card games and magic tricks, and
also being involved in the construction and analysis of algorithms such as the
PageRank algorithm which at one point in the late 90s was an important part
of the Google search engines.
In these notes we will be interested in walks on infinite graphs. I believe
it’s only a mild exaggeration to claim that the theory was born in 1920s with a
single theorem due to George Pólya.
Theorem 1 (Pólya’s Theorem). The simple random walk on the two dimen-
sional grid is recurrent but on the three dimensional grid it is transient.

Here the two dimensional grid Z2 is the graph whose vertices are pairs of
integers and where undirected edges are added horizontally and vertically so

2
that each vertex has 4 neighbors. The three dimensional grid Z3 is defined
similarly, each vertex has 6 neighbors.
A simple random walk is called recurrent if almost surely (i.e. with proba-
bility 1) it visits every vertex of the graph infinitely many times. The reader
can probably easily convince him or herself that such paths exist. The content
of the first part of Pólya’s theorem is that a path chosen at random on Z2 will
almost surely have this property (in other words the drunkard always makes it
home eventually).
Transience is the opposite of recurrence. It is equivalent to the property
that given enough time the walk will eventually escape any finite set of vertices
never to return. Hence it is somewhat counterintuitive that the simple random
walk on Z3 is transient but its shadow or projection onto Z2 is recurrent.

1.2 The theory of random walks

Starting with Pólya’s theorem one can say perhaps that the theory of random
walks is concerned with formalizing and answering the following question: What
is the relationship between the behavior of a random walk and the geometry of
the underlying space?
Since it is possible for a drunkard to walk on almost any mathematical struc-
ture the theory has rich interactions with various parts of math. Meaningful and
interesting answers to this question have been obtained in a wide variety of con-
texts ranging from random matrix products to Brownian motion on Riemannian
manifolds.
We can illustrate this by looking at the most obvious generalization of the
context of Pólya’s theorem, which is simply to consider in place of Z2 and Z3
the simple random walk on the d-dimensional grid Zd for d = 1, 2, 3, . . ..
Pólya’s theorem can be rephrased as follows: Suppose two random walkers
(let’s say lovers) begin at different vertices of Zd . If d ≤ 2 then they will almost
surely eventually meet. However, if d ≥ 3 then there is a positive probability
that they will never do so (how sad).
In spite of the fact that the two lovers might never meet on Z3 , it is known
that almost surely each lover will at some point be at the same vertex as the
other lover was some time before (and hence be able to smell their perfume, there
is hope!). Let’s summarize the situation by saying that Z3 has the “Perfume
property”. A gem from the theory of random walks on Zd (due to Erdös and
Taylor) is the following.
Theorem 2 (The perfume theorem). The d-dimensional grid has the perfume
property if and only if d ≤ 4.
Brownian motion on Rd is a scaling limit of the simple random walk on Zd
(in a precise sense given by Donsker’s theorem, which is a generalization of the
central limit theorem). Hence one expects Brownian paths to have analogous
behaviour to those of the simple random walk in the same dimension.
A theorem of Dvoretsky, Erdös, and Kakutani implies that two Brownian
paths in R4 almost surely do not intersect (so the strict analog of the perfume

3
 30

24 25 26
28
27,29

23

22

18,20 9 8
21 7

0,2,10,12,16 6
11,17,19

1,3,13,15 4, 14 5

Figure 1: Bêbado
I generated this by flipping a brazilian “1 real” coin twice for each step. The
result was almost to good of an illustration of Pólya’s theorem, returning to the
origin 4 times before leaving the box from (−5, −5) to (5, 5) after exactly 30
steps.

4
 Figure 2: If you drink don’t space walk.
More coin flipping. I kept the previous 2d walk and added a 3rd coordinate,
flipping coins so that this third coordinate would change one third of the time
on average (if you have to know, if both coins landed heads I would flip one
again to see whether to increment or decrement the new coordinate, if both
landed tails I would just ignore that toss, in any other case I would keep the
same third coordinate and advance to the next step in the original 2d walk).
The result is a 3d simple random walk. Clearly it returns to the origin a lot less
than its 2d shadow.

5
property does not hold in R4 ). However, the two paths will pass arbitrarily close
to one another so the perfume property does hold in a slightly weaker sense.
Brownian motion makes sense on Riemannian manifolds (basically because a
list of instructions of the type “forward 1 meter, turn right 90 degrees, forward 1
meter, etc” can be followed on any Riemannian surface, this idea is formalized
by the so-called Eells-Ellsworthy-Malliavin construction of Brownian motion)
so a natural object to study is Brownian motion on the homogeneous surface
geometries (spherical, Euclidean, and hyperbolic). A beautiful result (due to
Jean-Jaques Prat) in this context is the following:
Theorem 3 (Hyperbolic Brownian motion is transient). Brownian motion on
the hyperbolic plane escapes to infinity at unit speed and has an asymptotic
direction.
The hyperbolic plane admits polar coordinates (r, θ) with respect to any
chosen base point. Hence Brownian motion can be described as a random curve
(rt , θt ) indexed on t ≥ 0. Prat’s result is that rt /t → 1 and the limit θ∞ = lim θt
exists almost surely (and in fact eiθ∞ is necessarily uniform on the unit circle by
symmetry). This result is very interesting because it shows that the behaviour of
Brownian motion can change drastically even if the dimension of the underlying
space stays the same (i.e. curvature affects the behavior of Brownian motion).
Another type of random walk is obtained by taking two 2 × 2 invertible
matrices A and B and letting g1 , . . . , gn , . . . be independent and equal to either A
or B with probability 1/2 in each case. It was shown by Furstenberg and Kesten
that the exponential growth of the norm of the product An = g1 · · · gn exists,
i.e. χ = lim n1 log(|An |) (this number is called the first Lyapunov exponent of
the sequence, incredibly this result is a trivial corollary of a general theorem of
Kingman proved only a few years later using an almost completely disjoint set
of ideas).
The sequence An can be seen as a random walk on the group of invertible
matrices. There are three easy examples where χ = 0. First, one can take both
A and B to be rotation matrices (in this case one will even have recurrence in
the following weak sense: An willpass arbitrarily close  to theidentity matrix).
1 1 −1 1 −1
Second, one can take A = and B = A = . And third, one
  0 1
  0 1
0 1 2 0
can take A = and B = . It is also clear that conjugation by a
1 0 0 1/2
matrix C (i.e. changing A and B to C AC and C −1 BC respectively) doesn’t
−1

change the behavior of the walk. A beautiful result of Furstenberg (which has
many generalizations) is the following.
Theorem 4 (Furstenberg’s exponent theorem). If a sequence of independent
random matrix products An as above has χ = 0 then either both matrices A and
B are conjugate to rotations with respect to the same conjugacy matrix, or they
both fix a one-dimensional subspace of R2 , or they both leave invariant a union
of two one-dimensional subspaces.

6
1.3 Recurrence vs Transience: What these notes are about
From the previous subsection the reader might imagine that the theory of ran-
dom walks is already too vast to be covered in three lectures. Hence, we con-
centrate on a single question: Recurrence vs Transience. That is we strive to
answer the following:
Question 1. On which infinite graphs is the simple random walk recurrent?
Even this is too general (though we will obtain a sharp criterion, the Flow
Theorem, which is useful in many concrete instances). So we restrict even fur-
ther to just two classes of graphs: Trees and Cayley graphs (these two families,
plus the family of planar graphs which we will not discuss, have recieved special
attention in the literature because special types of arguments are available for
them, see the excellent recent book by Russell Lyons and Yuval Perez “Proba-
bility on trees and networks” for more information).
A tree is simply a connected graph which has no non-trivial closed paths
(a trivial closed path is the concatenation of a path and its reversal). Here are
some examples.
Example 1 (A regular tree). Consider the regular tree of degree three T3 (i.e.
the only connected tree for which every vertex has exactly 3 neighbors). The
random walk on T3 is clearly transient. Can you give a proof ?
Example 2 (The Canopy tree). The Canopy tree is an infinite tree seen from
the canopy (It’s branches all the way down!). It can be constructed as follows:
1. There is a “leaf vertex” for each natural number n = 1, 2, 3, . . ..
2. The leaf vertices are split into consecutive pairs (1, 2), (3, 4), . . . and each
pair is joined to a new “level two” vertex v1,2 , v3,4 , . . ..
3. The level two vertices are split into consecutive pairs and joined to level
three vertices and so on and so forth.
Is the random walk on the Canopy tree recurrent?
Example 3 (Radially symmetric trees). Any sequence of natural numbers a1 , a2 , . . .
defines a radially symmetric tree, which is simply a tree with a root vertex having
a1 children, each of which have a2 children, each of which have a3 children, etc.
Two simple examples are obtained by taking an constant equal to 1 (in which
case we get half of Z on which the simple walk is recurrent) or 2 (in which case
we get an infinite binary tree where the simple random walk is transient). More
interesting examples are obtained using sequences where all terms are either 1
or 2 but where both numbers appear infinitely many times. It turns out that such
sequences can define both recurrent and transient trees (see Corollary 1).
Example 4 (Self-similar trees). Take a finite tree with a distinguished root
vertex. At each leaf attach another copy of the tree (the root vertex replacing
the leaf ). Repeat ad infinitum. That’s a self-similar tree. A trivial example

7
is obtained when the finite tree used has only one branch (the resulting tree is
half of Z and therefore is recurrent). Are any other self-similar trees recurrent?
A generalization of this construction (introduced by Woess and Nagnibeda) is
to have n rooted finite trees whose leaves are labeled with the numbers 1 to n.
Starting with one of them one attaches a copy of the k-th tree to each leaf labeled
k, and so on ad infinitum.
Given any tree one can always obtain another by subdividing a few edges.
That is replacing an edge by a chain of a finite number of vertices (this concept
of subdivision appears for example in the characterization of planar graphs, a
graph is planar if and only if it doesn’t contain a subdivision of the complete
graph in 5 verteces or the 3 houses connected to electricity, water and telephone
graph1 ). For example a radially symmetric tree defined by a sequence of ones
and twos (both numbers appearing infinitely many times) is a subdivision of
the infinite binary tree.
Question 2. Can one make a transient tree recurrent by subdividing its edges?
Besides trees we will be considering the class of Cayley graphs which is
obtained by replacing addition on Zd with a non-commutative operation. In
general, given a finite symmetric generator F of a group G (i.e. g ∈ F implies
g −1 ∈ F , an example is F = {(±1, 0), (0, ±1)} and G = Z2 ) the Cayley graph
associated to (G, F ) has vertex set G and an undirected edge is added between
x and y if and only if x = yg for some g ∈ F (notice that this relationship is
symmetric).
Let’s see some examples.
Example 5 (The free group in two generators). The free group F2 in two
generators is the set of finite words in the letters N, W, E, W (north, south,
east, and west) considered up to equivalence with respect to the relations N S =
SN = EW = W E = e (where e is the empty word). It’s important to note
that these are the only relations (for example N E 6= EN ). The Cayley graph of
F2 (we will always consider the generating set {N, W, E, W }) is a regular tree
where each vertex has 4 neighbors.
Example 6 (The modular group). The modular group Γ is the group of frac-
tional linear transformations of the form
az + b
g(z) =
cz + d
where a, b, c, d ∈ Z and ad − bc = 1. We will always consider the generating
set F = {z 7→ z + 1, z 7→ z − 1, z 7→ −1/z}. Is the simple random walk on the
corresponding Cayley graph recurrent or transient?

Example 7 (A wallpaper group). The wallpaper group ∗632 is a group of isome-

tries of the Euclidean plane. To construct it consider a regular hexagon in the
1A more standard name might be the (3, 3) bipartite graph.

8
 Figure 3: Random walk on the modular group
Starting at the triangle to the right of the center and choosing to go through
either the red, green or blue side of the triangle one is currently at, one obtains
a random walk on the modular group. Here the sequence red-blue-red-blue-
green-red-blue-blue-red-red is illustrated.

9
 Figure 4: Wallpaper
I made this wallpaper with the program Morenaments by Martin von Gagern.
It allows you to draw while applying a group of Euclidean transformations to
anything you do. For this picture I chose the group ∗632.

plane and let F be the set of 12 axial symmetries with respect to the 6 sides, 3
diagonals (lines joining opposite vertices) and 3 lines joining the midpoints of
opposite sides. The group ∗632 is generated by F and each element of it pre-
serves a tiling of the plane by regular hexagons. The strange name is a case of
Conway notation and refers to the fact that 6 axii of symmetry pass through the
center of each hexagon in this tiling, 3 pass through each vertex, and 2 through
the midpoint of each side (the lack of asterisk would indicate rotational instead
of axial symmetry). Is the simple random walk on the Cayley graph of ∗632
recurrent?

Example 8 (The Heisenberg group). The Heisenberg group or Nilpotent group

N il is the group of 3 × 3 matrices of the form
 
1 x z
g = 0 1 y 
0 0 1

where x, y, z ∈ Z. We consider the generating set F with 6 elements defined one

of x, y or b being ±1 and the other two being 0. The vertex set of the Cayley
graph can be identified with Z3 . Can you picture the edges? The late great Bill
Thurston once gave a talk in Paris (it was the conference organized with the
money that Perelman refused, but that’s another story) where he said that N il
was a design for an efficient highway system . Several years later I understood
what he meant (I think). How many different vertices can you get to from a
given vertex using only n steps?

10
 Figure 5: Heisenberg
A portion of the Cayley graph of the Heisenberg group.

Example 9 (The Lamplighter group). Imagine Z with a lamp at each vertex.

Only a finite number of lamps are on and there’s a single inhabitant of this
world (the lamplighter) standing at a particular vertex. At any step he may
either move to a neighboring vertex or change the state (on to off or off to on)
of the lamp at his current vertex. This situation can be modeled as a group
Lamplighter(Z). The vertex set is the set of pairs (x, A) where x ∈ Z and A is
a finite subset of Z (the set of lamps which are on) the group operation is

(x, A)(y, B) = (x + y, A4(B + x)).

where the triangle denotes symmetric difference. The “elementary moves” corre-
spond to multiplication by elements of the generating set F = {(±1, {}), (0, {0})}.
Is the random walk on Lamplighter(Z) recurrent?

Question 3. Can it be the case that the Cayley graph associated to one gen-
erator for a group G is recurrent while for some other generator it’s transient?
i.e. Can we speak of recurrent vs transient groups or must one always include
the generator?

1.4 Notes for further reading

There are at least two very good books covering the subject matter of these notes
and more: The book by Woess “Random walks on infinite graphs and groups”
[Woe00] and “Probability on trees and infinite networks” by Russell Lyons and
Yuval Peres [LP15]. The lecture notes by Benjamini [Ben13] can give the reader
a good idea of where some of the more current research is heading.
The reader interested in the highly related area of random walks on finite
graphs might begin by looking at the survey article by Laslo Lovasz [Lov96]
(with the nice dedication to the random walks of Paul Erdös).

11
 Figure 6: Drunken Lamplighter
I’ve simulated a random walk on the lamplighter group. After a thousand
steps the lamplighter’s position is shown as a red sphere and the white spheres
represent lamps that are on. One knows that the lamplighter will visit every
lamp infinitely many times, but will he ever turn them all off again?

For the very rich theory of the simple random walk on Zd a good starting
point is the classical book by Spitzer [Spi76] and the recent one by Lawler
[Law13].
For properties of classical Brownian motion on Rd it’s relatively painless and
highly motivating to read the article by Einstein [Ein56] (this article actually
motivated the experimental confirmation of the existence of atoms via observa-
tion of Brownian motion, for which Jean Perrin recieved the Nobel prize later
on!). Mathematical treatment of the subject can be found in several places such
as Mörters and Peres’ [MP10].
Anyone interested in random matrix products should start by looking at the
original article by Furstenberg [Fur63], as well as the excellent book by Bougerol
and Lacroix [BL85].
Assuming basic knowledge of stochastic calculus and Riemannian geometry
I can recommend Hsu’s book [Hsu02] for Brownian motion on Riemannian man-
ifolds, or Stroock’s [Str00] for those preferring to go through more complicated
calculations in order to avoid the use of stochastic calculus . Purely analytical
treatment of the heat kernel (including upper and lower bounds) is well given in
Grigoryan’s [Gri09]. The necessary background material in Riemannian geom-
etry is not very advanced into the subject and is well treated in several places
(e.g. Petersen’s book [Pet06]), similarly, for the Stochastic calculus background
there are several good references such as Le Gall’s [LG13].

12
2 The entry fee: A crash course in probability
theory
Let me quote Kingman who said it better than I can.

The writer on probability always has a dilemma, since the mathe-

matical basis of the theory is the arcane subject of abstract measure
theory, which is unattractive to many potential readers...
That’s the short version. A longer version is that there was a long conceptual
struggle between discrete probability (probability = favorable cases over total
cases; the theory of dice games such as Backgammon) and geometric proba-
bility (probability = area of the good region divided by total area; think of
Buffon’s needle or Google it). After the creation of measure theory at the turn
of the 20th century people started to discover that geometric probability based
on measure theory was powerful enough to formulate all concepts usually stud-
ied in probability. This point of view was clearly stated for the first time in
Kolmogorov’s seminal 1933 article (the highlights being, measures on spaces of
infinite sequences and spaces of continuous functions, and conditional expecta-
tion formalized as a Radon-Nikodym derivative). Since then probabilistic con-
cepts are formalized mathematically as statements about measurable functions
whose domains are probability spaces. Some people were (and some still are)
reluctant about the predominant role played by measure theory in probability
theory (after all, why should a random variable be a measurable function?) but
this formalization has been tremendously successful and has not only permitted
the understanding of somewhat paradoxical concepts such as Brownian motion
and Poisson point processes but has also allowed for deep interactions between
probability theory and other areas of mathematics such as potential theory and
dynamical systems.
We can get away with a minimal amount of measure theory here thanks to
the fact that both time and space are discrete in our case of interest. But some
is necessary (without definitions there can be no proofs). So here it goes.

2.1 Probability spaces, random elements

A probability space is a triplet consisting of a set Ω (the points of which are
sometimes called “elementary outcomes”), a family F of subsets of Ω (the el-
ements of which are called “events”) which is closed under complementation
and countable union (i.e. F is a σ-algebra), and a function P : F → [0, 1] (the
probability) satisfying P [Ω] = 1 and more importantly
" #
[ X
P An = P [An ]
n n

for any countable (or finite) sequence A1 , . . . , An , . . . of pairwise disjoint ele-

ments of F.

13
 A random element is a function x from a probability space Ω to some com-
plete separable metric space X (i.e. a Polish space) with the property that the
preimage of any open set belongs to the σ-algebra F (i.e. x is measurable).
Usually if the function takes values in R it’s called a random variable, and as
Kingman put it, a random elephant is a measurable function into a suitable
space of elephants.
The point is that given a suitable (which means Borel, i.e. belonging to
the smallest σ-algebra generated by the open sets) subset A of the Polish space
X defined by some property P one can give meaning to “the probability that
the
 random  element x satisfies property P” simply by assigning it the number
P x−1 (A) which sometimes will be denoted by P [x satisfies property P ] or
P [x ∈ A].
Some people don’t like the fact that P is assumed to be countably additive
(instead of just finitely additive). But this is crucial for the theory and is
necessary in order to make sense out of things such as P [lim xn exists ] where
xn is a sequence of random variables (results about the asymptotic behaviour of
sequences of random elements abound in probability theory, just think of or look
up the Law of Large Numbers, Birkhoff’s ergodic theorem, Doob’s martingale
convergence theorem, and of course Pólya’s theorem).

2.2 Distributions
The distribution of a random element x is the Borel (meaning defined on the
σ-algebra of Borel sets) probability measure µ defined by µ(A) = P [x ∈ A].
Similarly the joint distribution of a pair of random elements x and y of spaces
X and Y is a probability on the product space X × Y , and the joint distribution
of a sequence of random variables is a probability on a sequence space (just
group all the elements together as a single random element and consider its
distribution).
Two events A and B of a probability space Ω are said to be independent if
P [A ∩ B] = P [A] P [B]. Similarly, two random elements x and y are said to be
indpendent if P [x ∈ A, y ∈ B] = P [x ∈ A] P [y ∈ B] for all Borel sets A and B
in the corresponding ranges of x and y. In other words they are independent
if their joint distribution is the product of their individual distributions. This
definitions generalizes to sequences. The elements of a (finite or countable)
sequence are said to be independent if their joint distribution is the product of
their individual distributions.
A somewhat counterintuitive fact is that independence of the pairs (x, y), (x, z)
and (y, z) does not imply indepedence of the sequence x, y, z. An example is
obtained by letting x and y be independent with P [x = ±1] = P [y = ±1] = 1/2
and z = xy (the product of x and y). A random element is independent from
itself if and only if it is almost surely constant (this strange legalistic loophole
is actually used in the proof of some results, such as Kolmogorov’s zero-one law
or the ergodicity of the Gauss continued fraction map; one shows that an event
has probability either zero or one by showing that it is independent from itself).
The philosophy in probability theory is that the hypothesis and statements

14
of the theorems should depend only on (joint) distributions of the variables
involved (which are usually assumed to satisfy some weakened form of indepen-
dence) and not on the underlying probability space (of course there are some
exceptions, notably Skorohod’s representation theorem on weak convergence
where the results is that a probability space with a certain sequence of variables
defined on it exists). The point of using a general space Ω instead of some fixed
Polish space X with a Borel probability µ is that, in Kolmogorov’s framework,
one may consider simultaneously random objects on several different spaces and
combine them to form new ones. In fact one could base the whole theory (pretty
much) on Ω = [0, 1] endowed with Lebesgue measure on the Borel sets and a few
things would be easier (notably conditional probabilities), but not many peo-
ple like this approach nowadays (the few who do study “standard probability
spaces”).

2.3 Markov chains

Consider a countable set X. The space of sequences X N = {ω = (ω1 , ω2 , . . .) :
ωi ∈ X} of elements of X is a complete separable metric space when endowed
with the distance X
d(ω, ω 0 ) = 2−n
0 }
{n:ωn 6=ωn

and the topology is that of coordinate-wise convergence.

For each finite sequence x1 , . . . , xn ∈ X we denote the subset of X N consist-
ing of infinite sequences begining with x1 , . . . , xn by [x1 , . . . , xn ].
P A probability on X can be identified with a function p : X → [0, 1] such that
p(x) = 1. By a transition matrix we mean a function q : X × X → [0, 1] such
x P
that q(x, y) = 1 for all x.
y∈X
The point of the following theorem is that, interpreting p(x) to be the prob-
ability that a random walk will start at x and q(x, y) as the probability that
the next step will be at y given that it is currently at x, the pair p, q defines a
unique probability on X N .
Theorem 5 (Kolmogorov extension theorem). For each probability p on X and
transition matrix q there exists a unique probability µp,q on X N such that
µp,q ([x1 , . . . , xn ]) = p(x1 )q(x1 , x2 ) · · · q(xn−1 , xn ).
A Markov chain with initial distribution p and transition matrix q is a se-
quence of random elements x1 , . . . , xn , . . . (defined on some probability space)
whose joint distribution is µp,q .
Markov chains are supposed to model “memoryless processes”. That is to
say that what’s going to happen next only depends on where we are now (plus
randomness) and not on the entire history of what happened before. To formal-
ize this we need the notion of conditional probability with respect to an event
which in our case can be simply defined as
P [B|A] = P [B] /P [A]

15
where P [A] is assumed to be positive (making sense out of conditional prob-
abilities when the events with respect to which one wishes to condition have
probability zero was one of the problems which pushed towards the develop-
ment of the Kolmogorov framework... but we wont need that).
We can now formalize the “memoryless” property of Markov chains (called
the Markov property) in its simplest form.
Theorem 6 (Weak Markov property). Let x0 , x1 , . . . be a Markov chain on a
countable set X with initial distribution p and transition matrix q. Then for
each fixed n the sequence y0 = xn , y1 = xn+1 , . . . is also a Markov chain with
transition matrix q (the initial distribution is simply the distribution of xn ).
Furthermore the sequence y = (y0 , y1 , . . .) is conditionally independent from
x0 , . . . , xn−1 given xn by which we mean

P [y ∈ A|x0 = a0 , . . . , xn = an ] = P [y ∈ A|xn = an ]

for all a0 , . . . , an ∈ X.

Proof. When the set A yields an event of {y1 = b1 , . . . , ym = bm } the proof is

by direct calculation. The general result follows by approximation. One uses
the fact from measure theory (which we will not prove) that given a probability
measure µ on X N any Borel set A can be approximated by a countable union of
sets defined by fixing the value of a finite number of coordinates (approximation
means that the probability symmetric difference can be made arbitrarily small).
This level of generality is needed in applications (e.g. a case we will use is when
A is the event that y eventually hits a certain point x ∈ X).
A typical example of a Markov chain is the simple random walk on Z (i.e. a
Markov chain on Z whose transition matrix satisfies q(n, n ± 1) = 1/2 for all n).
For example, let pk be the probability that a simple random walk x0 , x1 , . . . on
Z starting at k eventually hits 0 (i.e. pk = P [xn = 0 for some n]). Then from
the weak Markov property (with n = 1) one may deduce pk = 21 pk−1 + 12 pk+1
for all k 6= 0 (do this now!). Since p0 = 1 it follows from this that in fact pk = 1
for all k. Hence, the simple random walk on Z is recurrent!
A typical example of a random sequence which is not a Markov chain is
obtained by sampling without replacement. For example, suppose you are turn-
ing over the cards of a deck (after shuffling) one by one obtaining a sequence
x1 , . . . , x52 (if you want the sequence to be infinite, rinse and repeat). The
more cards of the sequence one knows the more one can bound the conditional
probability of the last card being the ace of spades for example (this fact is used
by people who count cards in casinos like in the 2008 movie “21”). Hence this
sequence is not reasonably modeled by a Markov chain.

2.4 Expectation
Given a real random variable x defined on some probability space its expectation
is defined in the following three cases:

16
 n
P
1. If x assumes only finitely many values x1 , . . . , xn then E [x] = P [x = xk ] xk .
k=1

2. If x assumes only non-negative values then E [x] = sup{E [y]} where the
supremum is taken over all random variables with 0 ≤ y ≤ x which assume
only finitely many values. In this case one may have E [x] = +∞.

3. If E [|x|] < +∞ then one defines E [x] = E [x+ ] − E [x− ] where x+ =

max(x, 0) and x− = max(−x, 0).
Variables with E [|x|] = +∞ are said to be non-integrable. The contrary to
this is to be integrable or to have finite expectation.
We say a sequence of random variables xn converges almost surely to a ran-
dom variable x if P [lim xn = x] = 1. The bread and butter theorems regarding
expectation are the following:
Theorem 7 (Monotone convergence). Let xn be a monotone increasing (in the
weak sense) sequence of non-negative random variables which converges almost
surely to a random variable x. Then E [x] = lim E [xn ] (even in the case where
the limit is +∞).
Theorem 8 (Dominated convergence). Let xn be a sequence of random vari-
ables converging almost surely to a limit x, and suppose there exists an inte-
grable random variable y such that |xn | ≤ y almost surely for all n. Then
E [x] = lim E [xn ] (here the limit must be finite and less than E [y] in absolute
value) .
In probability theory the dominated convergence theorem can be used when
one knows that all variables of the sequence xn are uniformly bounded by
some constant C. This is because the probability of the entire space is finite
(counterexamples exist on infinite measure spaces such as R endowed with the
Lebesgue measure).
As a student I was sometimes bothered by having to think about non-
integrable functions to prove results about integrable ones (e.g. most theo-
rems in probability, such as the Law of Large Numbers or the ergodic theorem,
require some sort of integrability hypothesis). Wouldn’t life be easier if we
just did the theory for bounded functions for example? The thing is there are
very natural examples of non-bounded and even non-integrable random vari-
ables. For example, consider a simple random walk x0 , x1 , . . . on Z starting at
1. Let τ be the first n ≥ 1 such that xτ = 0. Notice that τ is random but
we have shown using the weak Markov property that it is almost surely finite.
However, it’s possible to show that E [τ ] = +∞! This sounds paradoxical: τ
is finite but its expected value is infinite. But, in fact, I’m pretty sure that
with some thought a determined and tenacious reader can probably calculate
P[τ = n] = P [x1 6= 0,P. . . , xn−1 6= 0, xn = 0] for each n. The statement is that
P
P [τ = n] = 1 but nP [τ = n] = +∞. That doesn’t sound so crazy does it?
In fact there’s an indirect proof that E [τ ] = +∞ using the Markov property.
Let e1 = E [τ ], e2 be the expected value of the first hitting time of 0 for the

17
random walk starting at 2, etc (so en is the expected hitting time of 0 for a
simple random walk starting at n). Using the weak Markov property one can
get
en−1 + en+1
en = 1 +
2
for all n ≥ 1 (where e0 = 0) which gives us en+1 − en = (en − en−1 ) − 2 so
any solution to this recursion is eventually negative. The expected values of the
hittings times we’re considering, if finite, would all be positive, so they’re not
finite.

2.5 The strong Markov property

A stopping time for a Markov chain x0 , x1 , . . . is a random non-negative integer
τ (we allow τ = +∞ with positive probability) such that for all n = 0, 1, . . .
the set {τ = n} can be written as a countable union of sets of the form {x0 =
a0 , . . . , xn = an }. Informally you can figure out if τ = n by looking at x0 , . . . , xn ,
you don’t need to look at “the future”.
The concept was originally created to formalize betting strategies (the de-
cision to stop betting shouldn’t depend on the result of future bets). And is
central to Doob’s Martingale convergence theorem which roughly states that
all betting strategies are doomed to failure if the game is rigged in favor of the
casino.
Stopping times are also natural and important for Markov chains. An ex-
ample, which we have already encountered is the first hitting time of a subset
A ⊂ X which is defined by
τ = min{n ≥ 0 : xn ∈ A}.
A non-example is the last exit time τ = max{n : xn ∈ A}. It’s not a stopping
time because one cannot determine if the chain is going to return or not to A
after time n only by looking at x0 , . . . , xn .
On the set on which a stopping time is τ finite for a chain x0 , x1 , . . . one can
define xτ as the random variable which equals xk exactly when τ = k for all
k = 0, 1, . . ..
If τ is a stopping time so are τ + 1, τ + 2, . . .. But the same is not true in
general for τ − 1 (even if one knows that τ > 1).
An event A is said to occur “before the stopping time τ ” if for each finite
n one can write A ∩ {τ = n} as a countable union of events of the form {x0 =
a0 , . . . , xn = an }. A typical example is the event that the chain hits a point x
before time τ (it sounds tautological but that’s just because the name for this
type of events is well chosen). The point is on each set {τ = n} you can tell if
A happened by looking at x0 , . . . , xn .
The following is a generalization of the Weak Markov Property (Theorem
6) to stopping times. To state it, given a stopping time τ for a Markov chain
x0 , . . . , xn , . . . defined on some probability space (Ω, F, P), we will need to con-
sider the modified probability defined by Pτ = P1τ <+∞ /P [τ < +∞] (i.e. the
measure of an event A is P [A ∩ τ < +∞] /P [τ < +∞]). Here it goes.

18
Theorem 9 (Strong Markov property). Let x0 , x1 , . . . be a Markov chain with
transition matrix q and τ be a finite stopping. Then with respect to the prob-
ability Pτ the sequence y = (y0 , y1 , . . .) where yn = xτ +n is a Markov chain
with transition matrix q. Furthermore y is independent from event prior to τ
conditionally on on xτ by which we mean that
Pτ [y ∈ B|A, xτ = x] = Pτ [y ∈ B|xτ = x]
for all A occurring prior to τ , all x ∈ X, and all Borel B ⊂ X N .
The reader should try to work out the proof in enough detail to understand
where the fact that τ is a stopping time is used.
To see why the generality of the strong Markov property is useful let’s con-
sider the simple random walk on Z starting at 0. We have shown that the
probability of returning to 0 is 1. The strong Markov property allows us to con-
clude the intuitive fact that almost surely the random walk will hit 0 infinitely
many times. That is one can go from
P [xn = 0 for some n > 0] = 1
to
P [xn = 0 for infinitely many n] = 1
.
To see this just consider τ = min{n > 0 : xn = 0} the first return time to
0. We have already seen that the xn returns to 0 almost surely. The strong
Markov property allows one to calculate the probability that xn will visit 0
at least twice as P [τ < +∞] Pτ [xτ +n = 0 for some n > 0] and tells us that the
second factor is exactly the probability that xn returns to zero at least once
(which is 1). Hence the probability of returning twice is also equal to 1 and so
on and so forth. Finally, after we’ve shown that event {xn = 0 at least k times}
has full probability their intersection (being countable) also does.
This application may not seem very impressive. It gets more interesting
when the probability of return is less than 1. We will use this in the next
section.

2.6 Simple random walks on graphs

We will now define precisely, for the first time, our object of study for these
notes.
Given a graph X (we will always assume graphs are connected, undirected,
and locally finite) a simple random walk on X is a Markov chain x0 , . . . , xn , . . .
whose transition probabilities are defined by
number of edges joining x to y
q(x, y) = .
total number of edges with x as an endpoint
The distribution of the initial vertex x0 is unspecified. The most common
choice is just a deterministic vertex x and in this case we speak of a simple
random starting at x.

19
Definition 1 (Recurrence and Transience). A simple random walk {xn : n ≥ 0}
on a graph X is said to be recurrent if almost surely it visits every vertex in X
infinitely many times i.e.

P [xn = x for infinitely many n] = 1.

If this is not the case then we say the walk is transient.

Since there are countably many vertices in X recurrence is equivalent to
the walk visiting each fixed vertex infinitely many times almost surely (that
is P [xn = x for infinitely many n] = 1). In fact we will now show something
stronger and simultaneously answer the question of whether or not recurrence
depends on the initial distribution of the random walk (it doesn’t).
Lemma 1 (Recurrence vs Transience Dichotomy). Let x be a vertex of a graph
X and p be the probability that a simple random walk starting at x will return
to x at some positive time. Then p = 1 if and only if all simple random walks
on X are recurrent. Furthermore for any simple random walk x0 , x1 , . . . the
expected number of visits to x is given by
+∞
X
E [|{n > 0 : xn = x}|] = P [xn = x] = P [xn = x for some n] /(1 − p).
n=0

Proof. Define f : X → R by setting f (y) to be the probability that the simple

random walk starting at y will visit x (in particular f (x) = 1). By the weak
Markov property one has
X
f (y) = q(y, z)f (z)
z∈X

for all y 6= x (functions satisfying this equation are said to be harmonic) and
X
p= q(x, z)f (z).
z∈X

If p = 1 then using the above equation one concludes that f (y) = 1 at

all neighbors of x. Iterating the same argument one sees that f (y) = 1 at all
vertices at distance 2 from x and etc. This shows that the probability that a
simple random walk starting at any vertex will eventually visit x is 1, and hence
all simple random walks on X visit the vertex x almost surely at least once.
Given any simple random walk xn and letting τ be the first hitting time of x,
one knows that τ is almost surely finite. Also, by the strong Markov property,
one has that xτ , xτ +1 , . . . is a simple random walk starting at x. Hence τ2 the
second hitting time of x is also almost surely finite, and so on so forth. This
shows that p = 1 implies that every simple random walk hits x infinitely many
times almost surely. Also, given any vertex y the probability q that a simple
random walk starting at x will return to x before hitting y is strictly less than 1.
The strong Markov property implies that the probability of returning to x for

20
the n-th time without hitting y is q n , since this goes to 0 one obtains that any
simple random walk will almost surely hit y eventually, and therefore infinitely
many times. In short all simple random walks on X are recurrent.
We will now establish the formula for the expected number of visits.
Notice that the number of visits to x of a simple random walk xn is simply
+∞
X
1{xn =x}
n=0

where 1A is the indicator of an event A (i.e. the function taking the value 1
exactly on A and 0 everywhere else in the domain probability space Ω). Hence
the first equality in the formula for the expected number of visits follows simply
by the monotone convergence theorem.
The equality of this to the third term in the case p = 1 is trivial (all terms
being infinite). Hence we may assume from now on that p < 1.
In order to establish the second equality we use the sequence of stopping
times τ, τ1 , τ2 , . . . defined above. The probability that the number of visits to x is
exactly n is shown by strong Markov property to be P [τ < +∞] pn−1 (1−p).
Pthe n−1
Using the fact that np = 1/(1 − p)2 one obtains the result.

2.7 A counting proof of Pólya’s theorem

For such a nice result Pólya’s theorem admits a decidedly boring proof. Let
pn = P [xn = 0] for each n where xn is a simple random walk starting at 0 on Zd
(d = 1 or 2). In viewP of the previous subsection itP
suffices to estimate pn well
enough to show that pn = +∞ when d = 2 and pn < +∞ when d = 3.
Noticing that the total number of paths of length n in Zd starting at 0 is
(2d)n one obtains

pn = (2d)−n |closed paths of length of length n starting at 0|

so that all one really needs to estimate is the number of closed paths of length
n starting and ending at 0 (in particular notice that pn = 0 if n is odd, which
is quite reasonable).
For the case d = 2 the number of closed paths of length 2n can be seen
 2
2n
to be by bijection as follows: Consider the 2n increments of a closed
n
path. Associate to each increment one of the four letters aAbB according to the
following table
(1, 0) a
(−1, 0) B
.
(0, 1) A
(0, −1) b
Notice that the total number of ‘a’s (whether upper or lowercased) is n and
that the total number of uppercase letters is also n. Reciprocally if one is given
two subsets A, U of {1, . . . , 2n} with n elements each one can obtain a closed

21
path by interpreting the set A as the set of increments labeled with ‘a’ or ‘A’
and the set U as those labelled with an upper case letter.
Hence, one obtains for d = 2 that
 2
2n
p2n = 4−2n .
n
√ n n


Stirling’s formula n! ∼ 2πn e implies that

1
p2n ∼ ,
πn
P
which clearly implies that the series pn diverges. This shows that the simple
random walk on Z2 is recurrent.
For d = 3 the formulaP gets a bit tougher to analyze
P but we can grind it out.
First notice that if p6n converges then so does pn (because the number
of closed paths of length 2n is an increasing function of n). Hence we will bound
only the number of closed paths of length 6n.
The number of closed paths of length 6n is (splitting the 6n increments into
6 groups according to their value and noticing that there must be the same
number of (1, 0, 0) as (−1, 0, 0),etc)
X (6n)! (6n)! X 1 (6n)! X (3n)!
≤ ≤
(a!)2 (b!)2 (c!)2 (n!)3 a!b!c! (3n)!(n!)3 a!b!c!
a+b+c=3n a+b+c=3n a+b+c=3n
(6n)! 1
= 33n ∼ 66n √ ,
(3n)!(n!)3 4π 3 n3
where the first inequality is because a!b!c! ≥ (n!)3 when a + b + c = 3n and we
use that 33n is the sum of (3n)!/(a!b!c!) over a + b + c = 3n. √
Hence we have shown thatPfor all  > 0 one has p6n ≤ (1 + )/ 4π 3 n3 for
n large enough. This implies pn < +∞ and hence the random walk on Z3 is
transient.
This proof is unsatisfactory (to me at least) for several reasons. First, you
don’t learn a lot about why the result is true from this proof. Second, it’s very
inflexible, remove or add a single edge from the Zd graph and the calculations
become a lot more complicated, not to mention if you remove or add edges at
random or just consider other types of graphs. The ideas we will discuss in the
following section will allow for a more satisfactory proof, and in particular will
allow one to answer questions such as the following:
Question 4. Can a graph with a transient random walk become recurrent after
one adds edges to it?
At one point while learning some of these results I imagined the following:
Take Z3 and add an edge connecting each vertex (a, b, c) directly to 0. The
simple random walk on the resulting graph is clearly recurrent since at each step
it has probability 1/7 of going directly to 0. The problem with this example

22
is that it’s not a graph in our sense and hence does not have a simple random
walk (how many neighbors does 0 have? how would one define the transition
probabilities for 0?). In fact, we will show that the answer to the above question
is negative.

2.8 The classification of recurrent radially symmetric trees

David Blackwell is a one of my favorite mathematicians because of his great life
story combined with some really nice mathematical results. He was a black man,
born in Centralia, Illinois in 1919 (a town which today has some 13 thousand in-
habitants, and which is mostly notable for a mining accident which killed around
100 people in the late 1940s). He went to the university at Urbana-Champagne
starting at the age of 16 and ended up doing his Phd. under the tutelage of
Joseph Doob (who was one of the pioneering adopters of Kolmogorov’s formal-
ism for probability). He did a postdoc at the Institute of Advanced Study (where
he recalled in an interview having fruitful interactions for example with Von
Neumann). Apparently during that time he wasn’t allowed to attend lectures
at Princeton University because of his color. Later on he was also considered
favorably for a position at Berkeley which it seems also didn’t work out because
of racism (though Berkeley did hire him some 20 years later). He was the first
black member of the National Academy of Sciences.
Anyway, I’d like to use this section to talk about a very elegant result from
a paper he wrote in 1955. The result will allows us to immediately classify the
radially symmetric trees whose simple random walk is recurrent. The idea is to
study the Markov chain on the non-negative integers whose transition matrix
satisfies q(0, 0) = 1, q(n, n + 1) = pn and qn = 1 − pn = q(n, n − 1) for all n ≥ 1.
The question of interest is: What is the probability that such a Markov chain
started at a vertex n will eventually hit zero? Let’s look at two examples.
When pn = 1/2 for all n one gets the simple random walk “stopped (or
absorbed or killed, terminologies vary according to the different author’s moods
apparently) at 0”. We have already shown that the probability of hitting zero
for the walk started at any n is equal to 1. We did this by letting f (n) be
the corresponding hitting probability and noting that (by the weak Markov
property) f (0) = 1 and f satisfies
1 1
f (n) = f (n − 1) + f (n + 1)
2 2
for all n ≥ 1. Since there are no non-constant bounded solutions to the above
equation we get that f (n) = 1 for all n.
When pn = 2/3 (a chain which goes right with probability 2/3 and left with
probability 1/3) we can set f (n) once again to be the probability of hitting 0 if
the chain starts at n and notice that f (0) = 1 and f satisfies
1 2
f (n) = f (n − 1) + f (n + 1)
3 3
for all n ≥ 1. It turns out that there are non-constant bounded solutions to this
equation, in fact g(n) = 2−n is one such solution and all other solutions are of

23
the form α + βg for some constants α and β. Does this mean that there’s a
positive probability of never hitting 0? The answer is yes and that’s the content
of Blackwell’s theorem (this is a special case of a general result, the original
paper is very readable and highly recommended).
Theorem 10 (Blackwell’s theorem). Consider the transition matrix on Z+
defined by q(0, 0) = 1, q(n, n + 1) = pn and qn = 1 − pn = q(n, n − 1) for all
n ≥ 1. Then any Markov chain with this transition matrix will eventually hit 0
almost surely if and only if the equation

f (n) = qn f (n − 1) + pn f (n + 1)

has no non-constant bounded solution.

Proof. One direction is easy. Setting f (n) to be the probability of hitting 0 for
the chain starting at n one has that f is a solution and f (0) = 1. Therefore if
the probability doesn’t equal 1 one has a non-constant bounded solution.
The other direction is a direct corollary of Doob’s Martingale convergence
theorem (which regrettably we will not do justice to in these notes). The proof
goes as follows. Suppose f is a non-constant bounded solution. Then the
Martingale convergence theorem yields that the random limit L = lim f (xn )
n→+∞
exists almost surely and is non-constant. If the probability of hitting 0 were 1
then one would have L = f (0) almost surely contradicting the fact that L is
non-constant.
Blackwell’s result allows the complete classification of radially symmetric
trees whose random walk is recurrent.
Corollary 1 (Classification of recurrent radially symmetric trees). Let T be
a radially symmetric tree where the n-th generation vertices have an children.
Then the simple random walk on T (with any starting distribution) is recurrent
if and only if
+∞
X 1
= +∞.
n=1
a 1 a2 · · · an

Proof. The distance to the root of a random walk on T is simply a Markov chain
on the non-negative integers with transition probability q(0, 1) = 1, q(n, n−1) =
1/(1 + an ) = qn , and q(n, n + 1) = an /(1 + an ) = pn . Clearly such a chain
will eventually almost surely hit 0 if and only if the corresponding chain with
modified transition matrix q(0, 0) = 1 does.
Hence by Blackwell’s result the probability of hitting 0 is 1 for all simple
random walks on T if and only if there are no non-constant bounded solutions
to the equation
f (n) = qn f (n − 1) + pn f (n + 1)
on the non-negative integers.
It’s a linear recurrence of order two so one can show the space of solutions
is two dimensional. Hence all solutions are of the form α + βg where g is the

24
solution with g(0) = 0 and g(1) = 1. Manipulating the equation one gets the
equation g(n + 1) − g(n) = (g(n) − g(n − 1))/an for the increments of g. From
this it follows that
1 1 1
g(n) = 1 + + + ··· +
a1 a1 a2 a1 a2 · · · an−1

for all n. Hence the random walk is recurrent if and only if the g(n) → +∞
when n → +∞, as claimed.

2.9 Notes for further reading

For results from basic measure theory and introductory measure-theoretic prob-
ability I can recommend Ash’s book [Ash00]. Varadhan’s books [Var01] and
[Var07] are also excellent and cover basic Markov chains (including the Strong
Markov property which is [Var01, Lemma 4.10]) and more.
I quoted Kingman’s book [Kin93] on Poisson point processes a couple of
times in the section and, even though it is in principle unrelated to the material
actually covered in these notes, any time spent reading it will not be wasted.
The history of measure-theoretic probability is very interesting in its own
right. The reader might want to take a look at the articles by Doob (e.g. [Doo96]
or [Doo89]) about resistance to Kolmogorov’s formalization. The original article
by Kolmogorov [Kol50] is still readable and one might also have a look at the
discussion given in [SV06].
The counting proof of Polya’s theorem is essentially borrowed from [Woe00].
Blackwell’s original article is [Bla55] and the reader might want to look at
Kaimanovich’s [Kai92] as well.

3 The Flow Theorem

Think of the edges of a graph as tubes of exactly the same size which are
completely filled with water. A flow on the graph is a way of specifying a
direction and speed for the water in each tube in such a way that the amount
of water which enters each vertex per unit time equals the amount exiting it.
We allow a finite number of vertices where there is a net flux of water outward
or inwards (such vertices are called sources and sinks respectively). A flow with
a single source and no sink (such a thing can only exist on an infinite graph) is
called a source flow. The energy of a flow is the total kinetic energy of the water.
A flow on an infinite graph may have either finite or infinite energy. Here’s a
beautiful theorem due to Terry Lyons (famous in households world wide as the
creator of the theory of Rough paths, and deservedly so).

Theorem 11 (The flow theorem). The simple random walk on a graph is tran-
sient if and only if the graph admits a finite energy source flow.

25
3.1 The finite flow theorem
The flow theorem can be reduced to a statement about finite graphs but first
we need some notation.
So far, whenever X was a graph, we’ve been abusing notation by using X
to denote the set of vertices as well (i.e. x ∈ X means x is a vertex of X).
Let’s complement this notation by setting E(X) to be the set of edges of X.
We consider all edges to be directed so that each edge e ∈ E(X) has a starting
vertex e− and an end e+ . The fact that our graphs are undirected means there
is a bijective involution e 7→ e−1 sending each edge to another one with the start
and end vertex exchanged (loops can be their own inverses).
A field on X is just a function ϕ from E(X) to R satisfying ϕ(e−1 ) = −ϕ(e)
for all edges e ∈ E(X). Any function f : X → R has an asociated gradient field
∇f defined by ∇f (e) P = f (e+ ) − f (e− ). The energy of a field ϕ : E(X) → R
1 2
is defined by E(ϕ) = Pa field ϕ : E(X) → R is a
2 ϕ(e) . The divergence of
function div(ϕ) : X → R defined by div(ϕ)(x) = ϕ(e).
e− =x
Obviously all the definitions were chosen by analogy with vector calculus in
Rn . Here’s the analogous result to integration by parts.
Lemma 2 (Summation by parts). Let f be a function and ϕ a field on a finite
graph X. Then one has
X X
∇f (e)ϕ(e) = −2 f (x)div(ϕ)(x).
e∈E(X) x∈X

A flow between two vertices a and b of a graph X is a field with divergence

0 at all vertices except a and b, with positive divergence at a and negative
divergence at b. The point of the following theorem is that if X is finite there is
a unique (up to constant multiples) energy-minimizing flow from a to b and it’s
directly related to the behaviour of the simple random walk on the graph X.
Theorem 12 (Finite flow theorem). Let a and b be vertices of a finite graph X
and define f : X → R by

f (x) = P [xn hits b before a]

where xn is a simple random walk starting at x. Then the following properties

hold:
1. The gradient of f is a flow from a to b and satisfies div(∇f )(a) = E(∇f ) =
P [xn = b before returning to a] where xn is a simple random walk starting
at a.
E(∇f ) 2
2. Any flow ϕ from a to b satisfies E(ϕ) ≥ div(∇(f ))(a)2 div(ϕ)(a) with equal-
ity if and only if ϕ = λ∇f for some constant λ.
Proof. Notice that f (a) = 0, f (b) = 1 and f (x) ∈ (0, 1) for all other vertices x.
Hence the divergence at a is positive and the divergence at b is negative. The

26
fact that the divergence of the gradient is zero at the rest of the vertices follows
from the weak Markov property (f is harmonic except at a and b). Hence f is
a flow from a to b as claimed. The formula for divergence at a follows directly
from the definition, and for the energy one uses summation by parts.
The main point of the theorem is that ∇f is the unique energy minimizing
flow (for fixed divergence at a). To see this consider any flow ϕ from a to b with
the same divergence as ∇f at a.
We will first show that unless ϕ is the gradient of a function it cannot
minimize energy. For this purpose assume that e1 , .P . . , en is a closed path in the
graph (i.e. e1+ = e2− , . . . , en+ = e1− ) such that ϕ(ek ) 6= 0. Let ψ be the
flow such that ψ(ek ) = 1 for all ek and ψ(e) = 0 unless e = ek or e = e−1 k for
some k (so ψ is the unit flow around the closed path). One may calculate to
obtain
∂t E (ϕ + tψ)|t=0 6= 0
so one obtains for some small t (either positive or negative) a flow with less
energy than ϕ and the same divergence at a. This shows that any energy
minimizing flow with the same divergence as ∇f at a must be the gradient of a
function.
Hence it suffices to show that if g : X → R is a function which is harmonic ex-
cept at a and b and such that div(∇g)(a) = div(∇f )(a) then ∇g = ∇f . For this
purpose notice that because ∇g is a flow one has div(∇g)(b) = −div(∇g)(a) =
−div(∇f )(a) = div(∇f )(b). Hence f − g is harmonic on the entire graph and
therefore constant.
Exercise 1. Show that all harmonic functions on a finite graph are constant.

3.2 A proof of the flow theorem

Let X be an infinite graph now and xn be a simple random walk starting at a.
For each n let pn be the probability that d(xn , a) = n before xn returns to a
and notice that

lim pn = p = P [xn never returns to a] .

n→+∞

For each n consider the finite graph Xn obtained from X by replacing all
vertices with d(x, a) ≥ n by a single vertex bn (edges joining a vertex at distance
n − 1 to one at distance n now end in this new vertex, edges joining two vertices
at distance larger than n from a in X are erased).
The finite flow theorem implies that any flow from a to bn in Xn with
divergence 1 at a has energy greater than or equal to 1/pn . Notice that if X
is recurrent pn → 0 and this shows there is no finite energy source flow (with
soure a) on X.
On the other hand if X is transient then p > 0, so by the finite flow theorem
there are fields ϕn with divergence 1 at a, divergence 0 at vertices with d(x, a) 6=
n and with energy less than 1/p. For each edge e ∈ E(X) the sequence ϕn (e) is
bounded, hence, using a diagonal argument, there is a subsequence ϕnk which

27
converges to a finite energy flow with divergence 1 at a and 0 at all other vertices
(i.e. a finite energy source flow). This completes the proof (in fact one could
show that the sequence ϕn converges directly if one looks at how it is defined
in the finite flow theorem).

3.3 Adding edges to a transient graph cannot make it re-

current
A first, very important, and somewhat surprising application of the flow theorem
is the following.
Corollary 2. If a graph X has a transient subgraph then X is transient.

3.4 Recurrence of groups is independent of finite symmet-

ric generating set
Suppose G is a group which admits both F and F 0 as finite symmetric generating
sets. The Cayley graphs associated to F and F 0 are different and in general one
will not be a subgraph of the other (unless F ⊂ F 0 or something like that, for
example the Cayley graph of Z2 with respect to {(±1, 0), (0, ±1), (±1, ±1)} is
also recurrent by the previous corollary). However the flow theorem implies the
following:

Corollary 3. Recurrence or transience of a Cayley graph depends only on the

group and not on the finite symmetric generator one chooses.
Proof. Suppose that the graph with respect to one generator admits a finite
energy source flow ϕ. Notice that ϕ is the gradient of a function f : G → R
with f = 0 at the source. The gradient of the same function with respect to the
other graph structure also has finite energy (prove this!).
This does not conclude the proof because the gradient of f on the second
graph is not necessarily a flow. We will sketch how to solve this issue.
Modifying f slightly one can show that for any n one can find a function
taking the value 0 at the identity element of the group and 1 outside of the
the ball of radius n and such that the energy of the flow is bounded by a
constant which does not depend on n. Mimicking the argument used to prove
the finite flow one can in fact find a a harmonic function for each n satisfying the
same restrictions and whose gradient has energy bounded by the same constant.
Taking a limit (as in the proof of the flow theorem) one obtains a function whose
gradient is a finite energy source flow on the second Cayley graph.

This is a special case of a very important result which we will not prove but
which can also be proved via the flow theorem. Two metric spaces (X, d) and
(X 0 , d0 ) are said to be quasi-isometric if they are large-scale bi-lipshitz i.e. there
exists a function f : X → X 0 (not necesarilly continuous) and a constant C
such that
d(x, y)/C − C ≤ d(f (x), f (y)) ≤ Cd(x, y) + C,

28
and f (X) is C-dense in X 0 .
The definition is important because it allows one to relate discrete and con-
tinuous objects. For example Z2 and R2 are quasi-isometric (projection to the
nearest integer point gives a quasi-isometry in one direction, and simple inclu-
sion gives one in the other).

Exercise 2. Prove that Cayley graphs of the same group with respect to different
generating sets are quasi-isometric.
An important fact is that recurrence is invariant under quasi-isometry.
Corollary 4 (Kanai’s lemma). If two graphs of bounded degree X and X 0 are
quasi-isometric then either they are both recurrent or they are both transient.
The bounded degree hypothesis can probably be replaced by something
sharper (I really don’t know how much is known in that direction). The problem
is that adding multiple edges between the same two vertices doesn’t change the
distance on a graph but it may change the behavior of the random walk.

Exercise 3. Prove that the random walk on the graph with vertex set {0, 1, 2, . . .}
where n is joined to n + 1 by 2n edges is transient.

3.5 Boundary and Core of a tree

Let T be a tree and fix a root vertex a ∈ T . We define the boundary ∂T of
T as the set of all infinite sequences of edges (e0 , e1 , . . .) ∈ E(T )N such that
en+ = e(n+1)− and d(a, en− ) = n for all n. We turn the boundary into a metric
space by using the following distance
0
d((e0 , e1 , . . .), (e00 , e01 , . . .)) = e− min{n:en 6=en } .

Exercise 4. Prove that ∂T is a compact metric space.

The core T̊ of T is the subtree consisting in all edges and which appear in
some path of ∂T and the vertices they go through.

Exercise 5. Prove that given the metric space ∂T it is possible to reconstruct

T̊ . That is, two rooted trees have isometric boundaries if and only if their cores
are isomorphic as graphs.
The flow theorem implies the following.
Corollary 5. The transience or recurrence of a tree depends only on its bound-
ary (or equivalently on its core). That is, if two trees have isometric boundaries
then they are either both recurrent or both transient.

29
3.6 Logarithmic capacity and recurrence
Short digresion: There’s a beautiful criterion due to Kakutani which answers
the question of whether or not a Brownian motion in R2 will hit a compact set
K with positive probability or not. Somewhat surprisingly it is not necesary
for K to have positive measure for this to be the case. The sharp criterion is
given by whether or not K can hold a distribution of electric charge in such a
way that the potential energy (created by the repulsion between charges of the
same sign) is finite. In this formulation charges are considered to be infinitely
divisible so that if one has a unit of charge at a single point then the potential
energy is infinite (because one can consider that one has two half charges at the
same spot, and they will repulse each other infinitely). Sets with no capacity
to hold charges (such as a single point, or a countable set) will not be hit by
Brownian motions, but sets with positive capacity will. End digresion.
Definition 2. Let (X, d) be a compact metric space. Then (X, d) is is said to
be Polar (or have zero capacity) if and only if for every probability measure µ
on X one has Z
− log(d(x, y))dµ(x)dµ(y) = ∞.

Otherwise the space is said to have positive capacity. In general the capacity is
defined as (I hope I got all the signs right, another solution would be to let the
potential energy above be negative and use a supremum in the formula below)
 Z 
C(X, d) = exp − inf − log(d(x, y))dµ(x)dµ(y)

where the infimum is over all probability measures on X.

There is a correspondence between finite energy source flows on a tree and

finite measures on the boundary. Furthermore the energy of a source flow has a
simple relationship to the above type of integral for the corresponding measure.
Theorem 13. Let T be a rooted tree with root a. Given a unit source flow ϕ
with source a there is a unique probability measure µ on ∂T such that for each
edge e ∈ E(T ) the measure of the set of paths in ∂T which contain e is ϕ(e).
This correspondence between unit flows and probability measures is bijective.
Furthermore, one has
Z
E(ϕ) = − log(d(x, y))dµϕ (x)dµϕ (y).

Proof. For each e ∈ E(T ) let [e] denote the subset of ∂T consisting of paths
Sn
containing the edge e. Notice that ∂T \ [e] = [ei ] where the ei are the
i=1
remaining edges at the same level (i.e. joining vertices at the same distance
from a) as e. Also the sets [e] are compact so that if we write [e] as a countable
union of disjoint sets [ei ] then in fact the union is finite. Since this implies

30
that [e] 7→ ϕ(e) is countably additive on the algebra of sets generated by the [e]
one obtains by Caratheodory’s extension theorem that it extends to a unique
probability measure.
The inverse direction is elementary. Given a measure µ on ∂T one defines
ϕ(e) = µ([e]) and verifies that it is a unit flow simply by additivity of µ.
The important part of the statement is the relationship between the energy
of the flow ϕ and the type of integral used to define the capacity of ∂T .
To prove this we use a well known little trick in probability. It’s simply
the observation that if X is a random variable taking value in the non-negative
integers then X
E [X] = P [X ≥ n] .
n

The random variable we will consider is X = − log(d(x, y)) where x and y

are independent elements of ∂T with distribution µ. Notice that E [X] is exactly
the integral we’re interested in. Our result follows from the observation that
X
P [X ≥ n] = ϕ(e)2
{e∈E(X):d(e− ,a)=n−1,d(e+ ,a)=n}

which is simply the sum of probabilities of all the different ways two infinite
paths can coincide up to distance n from a.
The folowing corollary is immediate (notice that team Lyons has the ball
and Terry has passed it over to Russell).
Corollary 6 (Russell Lyons (1992)). The simple random walk on a tree T is
transient if and only if ∂T has positive capacity. It is recurrent if and only if
∂T is polar.

3.7 Average meeting height and recurrence

One can also state the criterion
R for recurrence of trees is in more combinatorial
terms using the fact that − log(d(x, y))dµ(x)dµ(y) is the expected value of
the distance from a of the first vertex where the paths x and y separate. To
formalize this, given two vertices c, d ∈ T define their meeting height m(c, d) to
be the distance from a at which the paths joining c to a and d to a first meet.
The average meeting height of a set b1 , . . . , bn of vertices is defined as
1 X
  m(bi , bj ).
n i<j
2

Corollary 7 (Benjamini and Peres (1992)). The simple random walk on a tree
T is transient if and only if there is a finite constante C > 0 such that for any
finite n there are n vertices b1 , . . . , bn ∈ T whose average meeting height is less
than C.

31
Proof. Suppose there is a finite unit flow ϕ and let x1 , . . . , xn , . . . be random
independent paths in ∂T with distribution µϕ . Then for each n one has
 
 1 X 
E
 n  = E(ϕ).
m(xi , xj )
1≤i<j≤n
2

Therefore there is a positive probabibility that the average meeting distance

between the first n random paths (or suitably chosen vertices on them) will be
less than C = 2E(ϕ).
In the other direction suppose there exists C and for each n one has paths
xn,1 , . . . , xn,n in ∂T such that their average meeting distance is less than C.
Then any weak limit2 µ of the sequence of probabilities
1X
µn = δxn,i
n
R
will satisfy − log(d(x, y))dµ(x)dµ(y) ≤ C (here δx denotes the probability
giving total mass to x). Which implies that the boundary has positive capacity
and that the tree is transient.

3.8 Positive Hausdorff dimension implies transience

Some readers might be familiar with Hausdorff dimension, which is a notion
of size for a compact metric space for which one obtains for example that the
dimension of the unit interval [0, 1] is 1 while that of the middle thirds Cantor
set is log(2)/ log(3) ≈ 0.631. The boundary of a tree usually looks somewhat
like a Cantor set, so this gives a superficial reason why dimension might be
relevant in our context.
Definition 3. The d-dimensional Hausdorff content of a compact metric space
(X, d) is the infimum of X
rid
over all sequences of positive diameters ri such that X can be covered by balls
with those diameters.
The Hausdorff dimension of X is the infiumum over all d > 0 such that the
d-dimensional content of X is 0.

Consider the task of proving that the Hausdorff dimension of [0, 1] is 1.

First one notices that one can easily cover [0, 1] by n balls of length rn = 100/n.
Hence if d > 1 one gets
nrnd = n(100/n)d → 0.
2 We haven’t discussed weak convergence in these notes. The things every person on the

planet should know about this notion are: First, that it’s the notion that appears in the
statement of the Central Limit Theorem. And, second, that a good reference is Billinsgley’s
book.

32
 This shows that the dimension of [0, 1] is less than or equal to 1 (the easy
part). But how does one show that it is actually 1? This is harder because one
needs to control all possible covers not just construct a particular sequence of
them.
The trick is to use Lebesgue measure µ on [0, 1]. The existence of Lebesgue
measure (and the fact that it is countably
P additive) implies that if one covers
[0, 1] by intervals of length ri then ri ≥ 1. Hence the 1-dimensional Hausdorff
content of [0, 1] is positive (greater than or equal to 1) and one obtains that the
dimension must be 1.
For compact subsets of Rn the above trick was carried out to its maximal
potential in a Phd. thesis from the 1930s. The result works for general compact
metric spaces and is now known by name of the author of said thesis.
Lemma 3 (Frostman’s Lemma). The d-dimensional Hausdorff content of (X, d)
is positive if and only if there exists a probability measure µ on X satisfying

µ(Br ) ≤ Crd

for some constant C and all balls Br of radius r > 0.

An immediate corollary is the following (the original proof is more elemen-
tary, the paper is very readable and recommended).
Corollary 8 (Russel Lyons 1990). Let T be a tree whose boundary has positive
Hausdorff dimension. Then the simple random walk on T is transient.
Proof. By Frostman’s theorem there’s a probability µ on the ∂T such that
µ(Br ) ≤ Crd for every ball Br of radius r > 0. This allows one to estimate the
integral defining capacity.
To see this, fix x and notice that the set of y such that − log(d(x, y)) ≥ t is
exactly the ball Be−t (x) (radius e−t and centered at x). This implies that (by
a variant of the trick we used in the proof of Theorem 13)
Z Z +∞ Z +∞
− log(d(x, y))dµ(y) ≤ µ({y : − log(d(x, y)) ≥ t}dt ≤ Ce−dt dt < +∞.
0 0

Since this is valid for all x one may itegrate over x to obtain that ∂T has
positive capacity.

3.9 Perturbation of trees

Our final corollary, combined with the exercises below, gives a pretty good idea
of how much one has to modify a tree to change the behavior of the simple
random walk on it.
Corollary 9. If T and T 0 are trees whose boundaries are Hölder homeomor-
phic (i.e. there is a homeomorphism f : ∂T → ∂T 0 satisfying d(f (x), f (y)) ≤
Cd(x, y)α for some constants C, α > 0) then the random walk on T is transient
if and only if the one on T 0 is.

33
Exercise 6. Prove that if T is a tree and a, b ∈ T then the boundaries one
obtains by considering a and b as the root vertex are Lipschitz homeomorphic
and the Lipschitz constant is no larger than ed(a,b) .
Exercise 7. Prove that if one subdivides the edges of a tree into no more than
N smaller edges then the boundary of the resulting tree is homeomorphic to the
original via a Hölder homeomorphism.
Exercise 8. Show that if T and T 0 are quasi-isometric trees then their bound-
aries are homeomorphic via a Hölder homeomorphism.

3.10 Notes for further reading

The original article by Terry Lyons [Lyo83] gives a relatively short proof of the
flow theorem which is based on the Martingale convergence theorem. Reading
this article one also learns of the flow theorem’s origins. Apparently, the theorem
was discovered as the translation to the discrete setting of a result from the
theory of Riemann surfaces. In short, it’s not to wrong to say that the flow
theorem for Brownian motion preceeded and inspired the corresponding theorem
for discrete random walks.
Our treatment of the flow theorem beginnning with finite graphs is based on
the very clearly written article by Levin and Peres [LP10]. Like them, we have
avoided the language of electrical networks even though it is a very important
topic in the subject (introduced by Crispin Nash-Williams in one of the first
interesting works on random walks on general infinite graphs [NW59]).
The many interesting results about trees are due to Russell Lyons ([Lyo90]
and [Lyo92], and Yuval Peres and Itai Benjamini [BP92].
The focus on quasi-isometry comes mostly from the more geometric works on
Riemannian manifolds (here one should mention the Furstenberg-Lyons-Sullivan
proceedure which allows one to associated a discrete random walk to a Brownian
motion on a Riemannian manifold [LS84]3 ). The original paper by Kanai is
[Kan86].
For Frostman’s lemma and much more geometric measure theory the reader
can have a look at [Mat95].
3 There’s a story explaining why Furstenberg is usually included in the name of this dis-

cretization proceedure. It turns out Furstenberg once wanted to prove that SL(2, Z) was not
a lattice in SL(3, R), and there was a very simple proof available using Kazhdan’s property
T, which was well known at the time. But Furstenberg didn’t know about property T, so he
concocted the fantastic idea of trying to translate the clearly distinct behavior of Brownian
motion on SL(2, R) and SL(3, R) to the discrete random walks on the corresponding lattices.
This was done mid-proof, but has all the essential elements later refined in Sullivan and Lyons’
paper (see [Fur71]).

34
4 The classification of recurrent groups
4.1 Sub-quadratic growth implies recurrence
We’ve seen as a consequence of the flow theorem that the simple random walk
behaves the same way (with respect to recurrence and transience) on all Cayley
graphs of a given group G. Hence one can speak of recurrent or transient groups
(as opposed to pairs (G, F )).
In this section G will always denote a group and F a finite symmetric gen-
erator of G. The Cayley graph will be denoted by G abusing notation slightly
and we keep with the notation E(G) for the set of (directed) edges of the graph.
We introduce a new notation which will be very important in this section.
Given a set A ⊂ G (or in any graph) we denote by ∂A the edge boundary of A,
i.e. the set of outgoing edges (edges starting in A and ending outside of A).
Let Bn denote the ball of radius n centered at the identity element of G.
That is, it is the set of all elements of the group which can be written as a
product of n or less elements of F . A first consequence of the flow theorem
(which in particular implies the recurrence of Z2 and of all wallpaper groups
such as *632) is the following:
Corollary 10. If there exists c > 0 such that |Bn | ≤ cn2 for all n then the
group G is recurrent.
Proof. Suppose ϕ is a unit source flow with source the identity element of G.
Notice that for all n one has
X
ϕ(e) = 1.
e∈∂Bn

Using Jensen’s inequality (in the version “average of squares is greater than
square of the average”) one gets
!2
1 X
2 1 X
ϕ(e) ≥ ϕ(e) ,
|∂Bn | |∂Bn |
e∈∂Bn e∈∂Bn

so that X 1
ϕ(e)2 ≥
|∂Bn |
e∈∂Bn
P
for all n. Hence it suffices to show that 1/|∂Bn | = +∞ to conclude that the
flow has infinite energy (notice that for the standard Cayley graph associated
to Z2 one can calculate |∂Bn | = 4 + 8n). Here the growth hypothesis must
be used and we leave it to the reader (see the exercise below and notice that
|F |(|Bn+1 | − |Bn |) ≥ |∂Bn |).
Exercise 9. Let ak , k = 1, 2, . . . be a sequence of positive numbers such that for
some positive constant c the inequality
n
X
an ≤ cn2
k=1

35
holds for all n. Prove that
+∞
X 1
= +∞.
an
k=1

4.2 Growth of groups

The function n 7→ |Bn | is sometimes called the growth function of the group
G. It depends on the chosen set of generators F but the functions associated to
two distinct generators are comparable up to a multiplicative constant. Hence
we can roughly classify groups into the following categories:
1. Polynomial growth: A group is said to have polynomial growth of degree
d if cnd ≤ |Bn | ≤ Cnd for all n and some constants c, C > 0. It is known
that having only the upper bound is sufficient to have a lower bound with
the same exponent d. Also, d is always an integer. This is a consequence
of a famous theorem from the 1980s due to Gromov.
2. Exponential growth: A group is said to have exponential growth if cλn ≤
|Bn | for all n and some constants c > 0, λ > 1. Notice that |Bn | ≤
(1 + |F |)n so no group can grow faster than exponentially.
3. Intermediate growth: The first examples of a group with subexponen-
tial but super-polynomial growth where discovered in the early 1980s by
Grigorshuk. We will not discuss any of these examples in these notes.

We have seen that a group with polynomial growth of degree d less than or
equal to 2 is recurrent. Our objective in this section is to prove that a group
with |Bn | ≥ cn3 for some c > 0 and all n must be transitive. These two cases
cover all posibilities by the theorem of Gromov classifying groups of polynomial
growth. Recall that a subgroup H of a group G is said to have finite index k if
there exist g1 , . . . , gk ∈ G such that every element of G can be written as hgi
for some h ∈ H and some i = 1, . . . , k. The final result is the following:
Theorem 14 (Varopoulos+Gromov). A group G is recurrent if and only if it
has polynomial growth of degree less than or equal to 2. This can only happen if
the group is either finite, has a subgroup isomorphic to Z with finite index, or
has a subgroup isomorphic to Z2 with finite index.
This is an example of a result whose statement isn’t very interesting (it
basically says that one shouldn’t study recurrence of Cayley graphs since it’s
too strong of a property) but for which the ideas involved in the proof are very
interesting (the flow theorem, isoperimetric inequalities, and Gromov’s theorem
on groups of polynomial growth).
We will not prove the theorem of Gromov which leads to the final clasifica-
tion. Only the fact that growth larger than cn3 implies transience.

36
4.3 Examples
Recall our list of examples from the first section: The d-dimensional grid Zd ,
the free group in two generators F2 , the Modular group of fractional linear
transformations Γ, the wallpaper group ∗632, the Heisenberg group N il, and
the lamplighter group Lamplighter(Z).
It is relatively simple to establish that Zd has polynomial growth of degree d,
that the free group in two generators has exponential growth, that the wallpaper
group has polynomial growth of degree d = 2, and that the lamplighter group
has exponential growth (with 2n moves the lamplighter can light the first n-
lamps in any of the 2n possible on-off combinations).
It turns out that the modular group has exponential growth. To see this it
−2
suffices to establish that the subgroup generated by z 7→ z + 2 and z 7→ 2z+1
is free (this subgroup is an example of a “congruence subgroup” which are
important in number theory, or so I’ve been told... by wikipedia). We leave
it to the interested reader to figure out a proof (several are possible, either by
trying to find an explicit fundamental domain on the action of the upper half
plane of C, by combinatorial analysis of the coefficients of compositions, or by
a standard argument for establishing freeness of a group called the Ping-Pong
argument which the reader can Google and learn about quite easily, there’s even
a relevant post in Terry Tao’s blog).
In view of Gromov’s theorem (which again, we will neither state fully nor
prove), working out the growth of the Heisenberg group and variants of it (which
was first done by Bass and Guivarc’h) turns out to be a key point in the proof
of Theorem 14. Hence any time the reader spends thinking about this issue is
well spent.
Exercise 10. Show that there exists c > 0 such that cn4 ≤ |Bn | for all n on
the Heisenberg group N il.

4.4 Isoperimetric inequalities

There exist graphs with exponential growth (for the natural definition) which
are recurrent. The simplest example is perhaps the Canopy tree (notice that
the ball of radius 2n centered at a leaf of the Canopy tree contains 2n leaves).
So clearly if growth is to characterize recurrence of Cayley graphs there must
be some special property of these graphs to explain this which does not hold in
general. The relevant property is characterized by the so-called isoperimetric
inequalities.
A Cayley graph is said to satisfy the strong isoperimetric inequality if there
is a positive constant c such that
c|A| ≤ |∂A|
for all finite subsets A ⊂ G. Notice that the strong isoperimetric inequality
implies exponential growth. The free group satisfies the strong isoperimetric
inequality. The existence of a finite index free subgroup implies that the Modu-
lar group also satisfies the strong isoperimetric inequality (there are other ways

37
of showing this, in fact the strong isoperimetric inequality is equivalent to a
property called non-amenability which has a near-infinite list of equivalent def-
initions which are not trivially equivalent, some of these definitions are simple
to verify on the Modular group).
A Cayley graph is said to satisfy the d-dimensional isoperimetric inequality
if there is a positive constant c such that

c|A|d−1 ≤ |∂A|d

for all finite subsets A ⊂ G. An analogous relationship with d = 3 is satisfied

between surface area and volume of a three dimensional submanifold in R3 .
The main point we will use which distinguishes Cayley graphs from general
graphs is the following:
Lemma 4. If there exists c > 0 such that |Bn | ≥ cnd for all n then G satisfies
an isoperimetric inequality of dimension d.
Proof. Fix a finite set A ⊂ G and let f = 1A be the indicator function of A (i.e.
f = 1 on A and f = 0 on G \ A). Notice that
X
|∇f |1 = |∇f | = 2|∂A|,
e∈E(G)

so the point is to bound the L1 norm of the gradient of f (i.e. the above sum)
from below in terms of |A|.
On the other hand X
|f |1 = |f (x)| = |A|
x∈G

so the objective is to bound |∇f |1 from below in terms of |f |1 .

In order to use the growth hypothesis we will introduce an auxiliary function
f˜ constructed by averaging f over suitable balls and in fact bound |f˜ − f | from
above and below in terms of |f |1 and |∇f |1 respectively.
The function f˜ is defined by letting n be the smallest integer such |Bn | ≥ 2|A|
and setting
1 X
f˜(x) = f (xg).
|Bn |
g∈Bn

The point of the choice of n is that if x ∈ A then f (x) = 1 but f˜(x) ≤ 1/2.
Hence the L1 norm of f˜ − f is bounded from below as follows
1
|A| ≤ |f˜ − f |1 .
2

If we can bound |f˜ − f |1 from above using |∇f |1 then we’re done.
To acomplish this notice that if g ∈ F then
X
|f (xg) − f (x)| ≤ |∇f |1 .
x∈G

38
 A slight generalization (left to the reader; Hint: triangle inequality) is that
if g = g1 · · · gk for some sequence of gi ∈ F then
X
|f (xg) − f (x)| ≤ k|∇f |1 .
x∈G

Using this one can bound the L1 norm of f˜ − f from above in terms of |∂A|
as follows
X 1 X X ˜
|f˜(x) − f (x)| ≤ |f (xg) − f (x)| ≤ n|∇f | = 2n|∂A|.
|Bn |
x∈G g∈Bn x∈G

Hence we have obtained

1
|A| ≤ |f˜ − f |1 ≤ 2n|∇f |1 = 4n|∂A|.
2
The result follows by noticing that the growth hypothesis implies that n ≤
C|A|1/d where C = c−1/d .

4.5 Growth of order 3 or more implies transience

Lemma 5. Let G be a group satisfying a 3-dimensional isoperimetric inequality.
Then G is transient.
Proof. Let a be the identity element of the group G and define for each n a
function fn : G → [0, 1] by

f (x) = P [d(a, xn ) ≥ n before xn = a]

where xn is a simple random walk on the Cayley graph starting at x.

Recall that (from the finite flow theorem) one had that ∇fn is a flow with
source at 1 and sinks at the points with d(a, x) = n and that

0 < Dn = div(∇fn )(a) = E(∇fn ).

Notice that since fn takes values in [0, 1] one has Dn ≤ deg(a) = |F | (recall F
is a finite symmetric generating set which was used to define the Cayley graph).
If one could bound Dn from below by a positive constant then this would imply
that (taking a limit of a subsequence of ∇fn ) there is a finite energy source flow
on G and hence G is transient as claimed.
To acomplish this let n be fixed and define a finite sequence of subsets
beginning with A1 = {a} using the rule that if Ak ⊂ Bn−1 then
 
2Dn
Ak+1 = Ak ∪ e+ : e ∈ ∂Ak , ∇fn (e) ≤ .
|∂Ak |

The sequence stops the first time Ak contains a point at distance n from A, let
N be the number of sets in the thus obtained sequence.

39
 The fact that the sequence stops follows because if Ak ⊂ Bn−1 one has
X
∇f (e) = Dn ,
e∈∂Ak

so at least half of the edges in Ak must lead to a point in Ak+1 .

The point of the definition is that there is a point x ∈ AN with d(a, x) = n
so that choosing a path e1 , . . . , eN −1 with e1− = a, e(N −1)+ = x and ei ∈ ∂Ai
for i = 1, . . . , N − 1 one obtains
k
X 2Dn
1 = f (x) ≤ .
i=1
|∂Ak |

Hence it suffices to bound

N
X 1
|∂Ak |
k=1

from above (in a way which doesn’t depend on n) to obtain a uniform lower
bound for Dn and hence prove that G is transient.
Here we use the 3-dimensional isoperimetric inequality, the fact we had noted
before that if k < N then at least half the edges of ∂Ak lead to Ak+1 , and the
fact that at most |F | edges can lead to any given vertex. Combining these facts
we obtain
1 2
|Ak+1 | − |Ak | ≥ |∂Ak | ≥ c|Ak | 3
2|F |
where c > 0 is the constant in the isoperimetric inequality divided by 2|F |.
This implies (see the exercise below) that |Ak | ≥ c3 k 3 /343. Hence
N +∞
X 1 X 49 49π 2
≤ =
|∂Ak | c3 k 2 6c3
k=1 k=1

where the exact sum in the final equation was included on a whim since only
convergence is needed to obtain the desired result.
Exercise 11. Let 1 = x1 < x2 < · · · < xN be a sequence satisfying xk+1 − xk ≥
2
cxk3 for all k = 1, . . . , N − 1. Prove that xk ≥ c3 k 3 /343 for all k = 1, . . . , N , or
at least prove that there exists a constant λ > 0 depending only on c such that
xk ≥ λ 3 k 3 .

4.6 Notes for further reading

As far as I know the first serious study of random walks on discrete groups was
carried out by Kesten in his thesis [Kes59b]. He was also the first to prove that
the strong isoperimetric inequality implies exponential decay of return proba-
bilities and hence transience (see [Kes59a] though this article is hard to find
and even harder to read, at least for me). This idea was later on refined and

40
improved by several people, the reader might want to check out Gerl’s article
[Ger88] and the more recent work by Virág [Vir00].
Kesten was also the first to introduce the idea that growth might determine
recurrence or transience (see [Kes67]) and the idea that polynomial growth of
order 2 was equivalent to recurrence is sometimes known as Kesten’s conjecture.
I find it interesting that Kesten’s conjecture was first established for contin-
uous groups (see [Bal81] which is the culmination of a series of works by several
people including Baldi, Guivarc’h, Keane, Roynette, Peyrière, and Lohoué).
The idea of using d-dimensional isoperimetric inequalities for estimating re-
turn probabilities was introduced quite successfully into the area of random
walks on discrete groups by Varopoulos in the mid-80s. The main result is that
a d-dimensional isoperimetric inequality implies a decay of return probabilities
of the order of n−d/2 (in particular if d ≥ 3 the series converges and the walk is
transient) which was proved in [Var85].
Instead of proving Varopoulos’ decay estimates we borrowed the proof given
in Mann’s excellent book [Man12] that growth implies isoperimetric inequalities
on finitely generated groups, and then proved that an isoperimetric inequality
of degree 3 or more implies transience using an argument from a paper by
Benjamini and Kozma [BK05].
There are several good references for Gromov’s theorem including Gromov’s
original paper (where the stunning idea of looking at a discrete group from far
away to obtain a continuous one is introduced), Mann’s book [Man12], and even
Tao’s blog.

References
[Ash00] Robert B. Ash. Probability and measure theory. Harcourt/Academic
Press, Burlington, MA, second edition, 2000. With contributions by
Catherine Doléans-Dade.
[Bal81] P. Baldi. Caractérisation des groupes de Lie connexes récurrents. Ann.
Inst. H. Poincaré Sect. B (N.S.), 17(3):281–308, 1981.
[Ben13] Itai Benjamini. Coarse geometry and randomness, volume 2100 of
Lecture Notes in Mathematics. Springer, Cham, 2013. Lecture notes
from the 41st Probability Summer School held in Saint-Flour, 2011,
Chapter 5 is due to Nicolas Curien, Chapter 12 was written by Ariel
Yadin, and Chapter 13 is joint work with Gady Kozma, École d’Été de
Probabilités de Saint-Flour. [Saint-Flour Probability Summer School].
[BK05] Itai Benjamini and Gady Kozma. A resistance bound via an isoperi-
metric inequality. Combinatorica, 25(6):645–650, 2005.
[BL85] Philippe Bougerol and Jean Lacroix. Products of random matrices
with applications to Schrödinger operators, volume 8 of Progress in
Probability and Statistics. Birkhäuser Boston, Inc., Boston, MA, 1985.

41
[Bla55] David Blackwell. On transient Markov processes with a countable
number of states and stationary transition probabilities. Ann. Math.
Statist., 26:654–658, 1955.
[BP92] Itai Benjamini and Yuval Peres. Random walks on a tree and capacity
in the interval. Ann. Inst. H. Poincaré Probab. Statist., 28(4):557–592,
1992.
[Doo89] J. L. Doob. Kolmogorov’s early work on convergence theory and foun-
dations. Ann. Probab., 17(3):815–821, 1989.
[Doo96] Joseph L. Doob. The development of rigor in mathematical probability
(1900–1950) [in development of mathematics 1900–1950 (luxembourg,
1992), 157–170, Birkhäuser, Basel, 1994; MR1298633 (95i:60001)].
Amer. Math. Monthly, 103(7):586–595, 1996.
[Ein56] Albert Einstein. Investigations on the theory of the Brownian move-
ment. Dover Publications, Inc., New York, 1956. Edited with notes
by R. Fürth, Translated by A. D. Cowper.
[Fur63] Harry Furstenberg. Noncommuting random products. Trans. Amer.
Math. Soc., 108:377–428, 1963.
[Fur71] Harry Furstenberg. Random walks and discrete subgroups of Lie
groups. In Advances in Probability and Related Topics, Vol. 1, pages
1–63. Dekker, New York, 1971.
[Ger88] Peter Gerl. Random walks on graphs with a strong isoperimetric
property. J. Theoret. Probab., 1(2):171–187, 1988.
[Gri09] Alexander Grigor’yan. Heat kernel and analysis on manifolds, vol-
ume 47 of AMS/IP Studies in Advanced Mathematics. American
Mathematical Society, Providence, RI; International Press, Boston,
MA, 2009.
[Hsu02] Elton P. Hsu. Stochastic analysis on manifolds, volume 38 of Graduate
Studies in Mathematics. American Mathematical Society, Providence,
RI, 2002.

[Kai92] Vadim A. Kaimanovich. Measure-theoretic boundaries of Markov

chains, 0-2 laws and entropy. In Harmonic analysis and discrete po-
tential theory (Frascati, 1991), pages 145–180. Plenum, New York,
1992.

[Kan86] Masahiko Kanai. Rough isometries and the parabolicity of Rieman-

nian manifolds. J. Math. Soc. Japan, 38(2):227–238, 1986.
[Kes59a] Harry Kesten. Full Banach mean values on countable groups. Math.
Scand., 7:146–156, 1959.

42
[Kes59b] Harry Kesten. Symmetric random walks on groups. Trans. Amer.
Math. Soc., 92:336–354, 1959.
[Kes67] H. Kesten. The Martin boundary of recurrent random walks on count-
able groups. In Proc. Fifth Berkeley Sympos. Math. Statist. and Prob-
ability (Berkeley, Calif., 1965/66), pages Vol. II: Contributions to
Probability Theory, Part 2, pp. 51–74. Univ. California Press, Berke-
ley, Calif., 1967.
[Kin93] J. F. C. Kingman. Poisson processes, volume 3 of Oxford Studies in
Probability. The Clarendon Press, Oxford University Press, New York,
1993. Oxford Science Publications.

[Kol50] A. N. Kolmogorov. Foundations of the Theory of Probability. Chelsea

Publishing Company, New York, N. Y., 1950.
[Law13] Gregory F. Lawler. Intersections of random walks. Modern Birkhäuser
Classics. Birkhäuser/Springer, New York, 2013. Reprint of the 1996
edition.

[LG13] Jean-François Le Gall. Mouvement brownien, martingales et calcul

stochastique, volume 71 of Mathématiques & Applications (Berlin)
[Mathematics & Applications]. Springer, Heidelberg, 2013.
[Lov96] L. Lovász. Random walks on graphs: a survey. In Combinatorics,
Paul Erdős is eighty, Vol. 2 (Keszthely, 1993), volume 2 of Bolyai
Soc. Math. Stud., pages 353–397. János Bolyai Math. Soc., Budapest,
1996.
[LP10] David A. Levin and Yuval Peres. Pólya’s theorem on random walks
via Pólya’s urn. Amer. Math. Monthly, 117(3):220–231, 2010.

[LP15] Russel Lyons and Yuval Peres. Random walks on infinite graphs and
groups. 2015.
[LS84] Terry Lyons and Dennis Sullivan. Function theory, random paths and
covering spaces. J. Differential Geom., 19(2):299–323, 1984.

[Lyo83] Terry Lyons. A simple criterion for transience of a reversible Markov

chain. Ann. Probab., 11(2):393–402, 1983.
[Lyo90] Russell Lyons. Random walks and percolation on trees. Ann. Probab.,
18(3):931–958, 1990.
[Lyo92] Russell Lyons. Random walks, capacity and percolation on trees. Ann.
Probab., 20(4):2043–2088, 1992.
[Man12] Avinoam Mann. How groups grow, volume 395 of London Mathemat-
ical Society Lecture Note Series. Cambridge University Press, Cam-
bridge, 2012.

43
[Mat95] Pertti Mattila. Geometry of sets and measures in Euclidean spaces,
volume 44 of Cambridge Studies in Advanced Mathematics. Cambridge
University Press, Cambridge, 1995. Fractals and rectifiability.
[MP10] Peter Mörters and Yuval Peres. Brownian motion. Cambridge Series
in Statistical and Probabilistic Mathematics. Cambridge University
Press, Cambridge, 2010. With an appendix by Oded Schramm and
Wendelin Werner.
[NW59] C. St. J. A. Nash-Williams. Random walk and electric currents in
networks. Proc. Cambridge Philos. Soc., 55:181–194, 1959.

[Pet06] Peter Petersen. Riemannian geometry, volume 171 of Graduate Texts

in Mathematics. Springer, New York, second edition, 2006.
[Spi76] Frank Spitzer. Principles of random walk. Springer-Verlag, New York-
Heidelberg, second edition, 1976. Graduate Texts in Mathematics,
Vol. 34.

[Str00] Daniel W. Stroock. An introduction to the analysis of paths on a

Riemannian manifold, volume 74 of Mathematical Surveys and Mono-
graphs. American Mathematical Society, Providence, RI, 2000.
[SV06] Glenn Shafer and Vladimir Vovk. The sources of Kolmogorov’s grund-
begriffe. Statist. Sci., 21(1):70–98, 2006.

[Var85] N. Th. Varopoulos. Isoperimetric inequalities and Markov chains. J.

Funct. Anal., 63(2):215–239, 1985.
[Var01] S. R. S. Varadhan. Probability theory, volume 7 of Courant Lecture
Notes in Mathematics. New York University, Courant Institute of
Mathematical Sciences, New York; American Mathematical Society,
Providence, RI, 2001.
[Var07] S. R. S. Varadhan. Stochastic processes, volume 16 of Courant Lecture
Notes in Mathematics. Courant Institute of Mathematical Sciences,
New York; American Mathematical Society, Providence, RI, 2007.

[Vir00] B. Virág. Anchored expansion and random walk. Geom. Funct. Anal.,
10(6):1588–1605, 2000.
[Woe00] Wolfgang Woess. Random walks on infinite graphs and groups, volume
138 of Cambridge Tracts in Mathematics. Cambridge University Press,
Cambridge, 2000.

44
Index
Canopy Tree, 7, 37

Flow Theorem, 7, 25, 27–29, 34–36, 39

Free Group, 8, 37

Gromov’s Theorem, 36, 37, 41

Heisenberg Group, 10, 37

Lamplighter Group, 11, 37

Modular Group, 8, 37, 38

Pólya’s Theorem, 2, 21

Radially Symmetric Tree, 7, 23, 24

Recurrence, 3, 6, 7, 20–24, 27–31, 35–
37, 41
Regular Tree, 7

Self-similar Tree, 7
Simple Random Walk, 2, 19

Transience, see Recurrence

Wallpaper Group, 8, 35

45