Random walks and electric networks

Peter G. Doyle J. Laurie Snell

Version dated 5 July 2006

GNU FDL∗

Acknowledgement
This work is derived from the book Random Walks and Electric Net-
works, originally published in 1984 by the Mathematical Association of
America in their Carus Monographs series. We are grateful to the MAA
for permitting this work to be freely redistributed under the terms of
the GNU Free Documentation License.

∗
Copyright (C) 1999, 2000, 2006 Peter G. Doyle and J. Laurie Snell. Derived
from their Carus Monograph, Copyright (C) 1984 The Mathematical Association
of America. Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, as published by the Free
Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no
Back-Cover Texts.

1
Preface
Probability theory, like much of mathematics, is indebted to physics as
a source of problems and intuition for solving these problems. Unfor-
tunately, the level of abstraction of current mathematics often makes it
difficult for anyone but an expert to appreciate this fact. In this work
we will look at the interplay of physics and mathematics in terms of an
example where the mathematics involved is at the college level. The
example is the relation between elementary electric network theory and
random walks.
Central to the work will be Polya’s beautiful theorem that a random
walker on an infinite street network in d-dimensional space is bound to
return to the starting point when d = 2, but has a positive probability
of escaping to infinity without returning to the starting point when
d ≥ 3. Our goal will be to interpret this theorem as a statement about
electric networks, and then to prove the theorem using techniques from
classical electrical theory. The techniques referred to go back to Lord
Rayleigh, who introduced them in connection with an investigation of
musical instruments. The analog of Polya’s theorem in this connection
is that wind instruments are possible in our three-dimensional world,
but are not possible in Flatland (Abbott [1]).
The connection between random walks and electric networks has
been recognized for some time (see Kakutani [12], Kemeny, Snell, and
Knapp [14], and Kelly [13]). As for Rayleigh’s method, the authors
first learned it from Peter’s father Bill Doyle, who used it to explain a
mysterious comment in Feller ([5], p. 425, Problem 14). This comment
suggested that a random walk in two dimensions remains recurrent
when some of the streets are blocked, and while this is ticklish to prove
probabilistically, it is an easy consequence of Rayleigh’s method. The
first person to apply Rayleigh’s method to random walks seems to have
been Nash-Williams [24]. Earlier, Royden [30] had applied Rayleigh’s
method to an equivalent problem. However, the true importance of
Rayleigh’s method for probability theory is only now becoming appre-
ciated. See, for example, Griffeath and Liggett [9], Lyons [20], and
Kesten [16].
Here’s the plan of the work: In Section 1 we will restrict ourselves
to the study of random walks on finite networks. Here we will establish
the connection between the electrical concepts of current and voltage
and corresponding descriptive quantities of random walks regarded as
finite state Markov chains. In Section 2 we will consider random walks

2
on infinite networks. Polya’s theorem will be proved using Rayleigh’s
method, and the proof will be compared with the classical proof using
probabilistic methods. We will then discuss walks on more general
infinite graphs, and use Rayleigh’s method to derive certain extensions
of Polya’s theorem. Certain of the results in Section 2 were obtained
by Peter Doyle in work on his Ph.D. thesis.
To read this work, you should have a knowledge of the basic concepts
of probability theory as well as a little electric network theory and
linear algebra. An elementary introduction to finite Markov chains as
presented by Kemeny, Snell, and Thompson [15] would be helpful.
The work of Snell was carried out while enjoying the hospitality of
Churchill College and the Cambridge Statistical Laboratory supported
by an NSF Faculty Development Fellowship. He thanks Professors
Kendall and Whittle for making this such an enjoyable and rewarding
visit. Peter Doyle thanks his father for teaching him how to think like a
physicist. We both thank Peter Ney for assigning the problem in Feller
that started all this, David Griffeath for suggesting the example to be
used in our first proof that 3-dimensional random walk is recurrent
(Section 2.2.9), and Reese Prosser for keeping us going by his friendly
and helpful hectoring. Special thanks are due Marie Slack, our secre-
tary extraordinaire, for typing the original and the excessive number of
revisions one is led to by computer formatting.

1 Random walks on finite networks

1.1 Random walks in one dimension
1.1.1 A random walk along Madison Avenue
A random walk, or drunkard’s walk, was one of the first chance pro-
cesses studied in probability; this chance process continues to play an
important role in probability theory and its applications. An example
of a random walk may be described as follows:
A man walks along a 5-block stretch of Madison Avenue. He starts
at corner x and, with probability 1/2, walks one block to the right and,
with probability 1/2, walks one block to the left; when he comes to
the next corner he again randomly chooses his direction along Madison
Avenue. He continues until he reaches corner 5, which is home, or
corner 0, which is a bar. If he reaches either home or the bar, he stays
there. (See Figure 1.)

3
 Figure 1: ♣

The problem we pose is to find the probability p(x) that the man,
starting at corner x, will reach home before reaching the bar. In looking
at this problem, we will not be so much concerned with the particular
form of the solution, which turns out to be p(x) = x/5, as with its
general properties, which we will eventually describe by saying “p(x) is
the unique solution to a certain Dirichlet problem.”

1.1.2 The same problem as a penny matching game

In another form, the problem is posed in terms of the following game:
Peter and Paul match pennies; they have a total of 5 pennies; on each
match, Peter wins one penny from Paul with probability 1/2 and loses
one with probability 1/2; they play until Peter’s fortune reaches 0 (he
is ruined) or reaches 5 (he wins all Paul’s money). Now p(x) is the
probability that Peter wins if he starts with x pennies.

1.1.3 The probability of winning: basic properties

Consider a random walk on the integers 0, 1, 2, . . . , N . Let p(x) be the
probability, starting at x, of reaching N before 0. We regard p(x) as
a function defined on the points x = 0, 1, 2, . . . , N . The function p(x)
has the following properties:

(a) p(0) = 0.

(b) p(N ) = 1.

(c) p(x) = 21 p(x − 1) + 21 p(x + 1) for x = 1, 2, . . . , N − 1.

Properties (a) and (b) follow from our convention that 0 and N are
traps; if the walker reaches one of these positions, he stops there; in
the game interpretation, the game ends when one player has all of the

4
pennies. Property (c) states that, for an interior point, the probability
p(x) of reaching home from x is the average of the probabilities p(x − 1)
and p(x + 1) of reaching home from the points that the walker may
go to from x. We can derive (c) from the following basic fact about
probability:
Basic Fact. Let E be any event, and F and G be events such that
one and only one of the events F or G will occur. Then

P(E) = P(F ) · P(E given F ) + P(G) · P(E given G).

In this case, let E be the event “the walker ends at the bar”, F
the event “the first step is to the left”, and G the event “the first
step is to the right”. Then, if the walker starts at x, P(E) = p(x),
P(F ) = P(G) = 21 , P(E given F ) = p(x−1), P(E given G) = p(x+1),
and (c) follows.

1.1.4 An electric network problem: the same problem?

Let’s consider a second apparently very different problem. We connect
equal resistors in series and put a unit voltage across the ends as in
Figure 2.

Figure 2: ♣

Voltages v(x) will be established at the points x = 0, 1, 2, 3, 4, 5. We

have grounded the point x = 0 so that v(0) = 0. We ask for the voltage
v(x) at the points x between the resistors. If we have N resistors, we
make v(0) = 0 and v(N ) = 1, so v(x) satisfies properties (a) and (b) of
Section 1.1.3. We now show that v(x) also satisfies (c).
By Kirchhoff’s Laws, the current flowing into x must be equal to
the current flowing out. By Ohm’s Law, if points x and y are connected

5
by a resistance of magnitude R, then the current ixy that flows from x
to y is equal to
v(x) − v(y)
ixy = .
R
Thus for x = 1, 2, . . . , N − 1,

v(x − 1) − v(x) v(x + 1) − v(x)

+ = 0.
R R
Multiplying through by R and solving for v(x) gives

v(x + 1) + v(x − 1)
v(x) =
2
for x = 1, 2, . . . , N − 1. Therefore, v(x) also satisfies property (c).
We have seen that p(x) and v(x) both satisfy properties (a), (b),
and (c) of Section 1.1.3. This raises the question: are p(x) and v(x)
equal? For this simple example, we can easily find v(x) using Ohm’s
Law, find p(x) using elementary probability, and see that they are the
same. However, we want to illustrate a principle that will work for very
general circuits. So instead we shall prove that these two functions are
the same by showing that there is only one function that satisfies these
properties, and we shall prove this by a method that will apply to more
general situations than points connected together in a straight line.

Exercise 1.1.1 Referring to the random walk along Madison Avenue,

let X = p(1), Y = p(2), Z = p(3), and W = p(4). Show that properties
(a), (b), and (c) of Section 1.1.3 determine a set of four linear equations
with variables X, Y , Z and W . Show that these equations have a
unique solution. What does this say about p(x) and v(x) for this special
case?

Exercise 1.1.2 Assume that our walker has a tendency to drift in one
direction: more specifically, assume that each step is to the right with
probability p or to the left with probability q = 1 − p. Show that
properties (a), (b), and (c) of Section 1.1.3 should be replaced by

(a) p(0) = 0.

(b) p(N ) = 1.

(c) p(x) = q · p(x − 1) + p · p(x + 1).

6
Exercise 1.1.3 In our electric network problem, assume that the re-
sistors are not necessarily equal. Let Rx be the resistance between x
and x + 1. Show that
1 1
Rx−1 Rx
v(x) = 1 1 v(x − 1) + 1 1 v(x + 1).
Rx−1
+ Rx Rx−1
+ Rx

How should the resistors be chosen to correspond to the random walk

of Exercise 1.1.2?

1.1.5 Harmonic functions in one dimension; the Uniqueness

Principle
Let S be the set of points S = {0, 1, 2, . . . , N }. We call the points of
the set D = {1, 2, . . . , N − 1} the interior points of S and those of
B = {0, N } the boundary points of S. A function f (x) defined on S is
harmonic if, at points of D, it satisfies the averaging property
f (x − 1) + f (x + 1)
f (x) = .
2
As we have seen, p(x) and v(x) are harmonic functions on S having
the same values on the boundary: p(0) = v(0) = 0; p(N ) = v(N ) =
1. Thus both p(x) and v(x) solve the problem of finding a harmonic
function having these boundary values. Now the problem of finding
a harmonic function given its boundary values is called the Dirichlet
problem, and the Uniqueness Principle for the Dirichlet problem asserts
that there cannot be two different harmonic functions having the same
boundary values. In particular, it follows that p(x) and v(x) are really
the same function, and this is what we have been hoping to show. Thus
the fact that p(x) = v(x) is an aspect of a general fact about harmonic
functions.
We will approach the Uniqueness Principle by way of the Maxi-
mum Principle for harmonic functions, which bears the same relation
to the Uniqueness Principle as Rolle’s Theorem does to the Mean Value
Theorem of Calculus.
Maximum Principle . A harmonic function f (x) defined on S
takes on its maximum value M and its minimum value m on the bound-
ary.
Proof. Let M be the largest value of f . Then if f (x) = M for x
in D, the same must be true for f (x − 1) and f (x + 1) since f (x) is
the average of these two values. If x − 1 is still an interior point, the

7
same argument implies that f (x − 2) = M ; continuing in this way, we
eventually conclude that f (0) = M . That same argument works for
the minimum value m. ♦
Uniqueness Principle. If f (x) and g(x) are harmonic functions
on S such that f (x) = g(x) on B, then f (x) = g(x) for all x.
Proof. Let h(x) = f (x) − g(x). Then if x is any interior point,
h(x − 1) + h(x + 1) f (x − 1) + f (x + 1) g(x − 1) + g(x + 1)
= − ,
2 2 2
and h is harmonic. But h(x) = 0 for x in B, and hence, by the Max-
imum Principle, the maximum and mininium values of h are 0. Thus
h(x) = 0 for all x, and f (x) = g(x) for all x. ♦
Thus we finally prove that p(x) = v(x); but what does v(x) equal?
The Uniqueness Principle shows us a way to find a concrete answer:
just guess. For if we can find any harmonic function f (x) having the
right boundary values, the Uniqueness Principle guarantees that
p(x) = v(x) = f (x).
The simplest function to try for f (x) would be a linear function; this
leads to the solution f (x) = x/N . Note that f (0) = 0 and f (N ) = 1
and
f (x − 1) + f (x + 1) x−1+x+1 x
= = = f (x).
2 2N N
Therefore f (x) = p(x) = v(x) = x/N .
As another application of the Uniqueness Principle, we prove that
our walker will eventually reach 0 or N . Choose a starting point x with
0 < x < N . Let h(x) be the probability that the walker never reaches
the boundary B = {0, N }. Then
1 1
h(x) = h(x + 1) + h(x − 1)
2 2
and h is harmonic. Also h(0) = h(N ) = 0; thus, by the Maximum
Principle, h(x) = 0 for all x.

Exercise 1.1.4 Show that you can choose A and B so that the func-
tion f (x) = A(q/p)x + B satisfies the modified properties (a), (b) and
(c) of Exercise 1.1.2. Does this show that f (x) = p(x)?

Exercise 1.1.5 Let m(x) be the expected number of steps, starting at

x, required to reach 0 or N for the first time. It can be proven that
m(x) is finite. Show that m(x) satisfies the conditions

8
 (a) m(0) = 0.
(b) m(N ) = 0.
(c) m(x) = 21 m(x + 1) + 21 m(x − 1) + 1.

Exercise 1.1.6 Show that the conditions in Exercise 1.1.5 have a unique
solution. Hint: show that if m and m̄ are two solutions, then f = m− m̄
is harmonic with f (0) = f (N ) = 0 and hence f (x) = 0 for all x.

Exercise 1.1.7 Show that you can choose A, B, and C such that
f (x) = A + Bx + Cx2 satisfies all the conditions of Exercise 1.1.5. Does
this show that f (x) = m(x) for this choice of A, B, and C?

Exercise 1.1.8 Find the expected duration of the walk down Madison
Avenue as a function of the walker’s starting point (1, 2, 3, or 4).

1.1.6 The solution as a fair game (martingale)

Let us return to our interpretation of a random walk as Peter’s fortune
in a game of penny matching with Paul. On each match, Peter wins
one penny with probability 1/2 and loses one penny with probability
1/2. Thus, when Peter has k pennies his expected fortune after the
next play is
1 1
(k − 1) + (k + 1) = k,
2 2
so his expected fortune after the next play is equal to his present for-
tune. This says that he is playing a fair game; a chance process that can
be interpreted as a player’s fortune in a fair game is called a martingale.
Now assume that Peter and Paul have a total of N pennies. Let
p(x) be the probability that, when Peter has x pennies, he will end up
with all N pennies. Then Peter’s expected final fortune in this game is

(1 − p(x)) · 0 + p(x) · N = p(x) · N.

If we could be sure that a fair game remains fair to the end of the
game, then we could conclude that Peter’s expected final fortune is
equal to his starting fortune x, i.e., x = p(x) · N . This would give
p(x) = x/N and we would have found the probability that Peter wins
using the fact that a fair game remains fair to the end. Note that the
time the game ends is a random time, namely, the time that the walk
first reaches 0 or N for the first time. Thus the question is, is the
fairness of a game preserved when we stop at a random time?

9
 Unfortunately, this is not always the case. To begin with, if Peter
somehow has knowledge of what the future holds in store for him, he
can decide to quit when he gets to the end of a winning streak. But
even if we restrict ourselves to stopping rules where the decision to
stop or continue is independent of future events, fairness may not be
preserved. For example, assume that Peter is allowed to go into debt
and can play as long as he wants to. He starts with 0 pennies and
decides to play until his fortune is 1 and then quit. We shall see that
a random walk on the set of all integers, starting at 0, will reach the
point 1 if we wait long enough. Hence, Peter will end up one penny
ahead by this system of stopping.
However, there are certain conditions under which we can guarantee
that a fair game remains fair when stopped at a random time. For our
purposes, the following standard result of martingale theory will do:
Martingale Stopping Theorem. A fair game that is stopped at
a random time will remain fair to the end of the game if it is assumed
that there is a finite amount of money in the world and a player must
stop if he wins all this money or goes into debt by this amount.
This theorem would justify the above argument to obtain p(x) =
x/N .
Let’s step back and see how this martingale argument worked. We
began with a harmonic function, the function f (x) = x, and inter-
preted it as the player’s fortune in a fair game. We then considered the
player’s expected final fortune in this game. This was another harmonic
function having the same boundary values and we appealed to the Mar-
tingale Stopping Theorem to argue that this function must be the same
as the original function. This allowed us to write down an expression
for the probability of winning, which was what we were looking for.
Lurking behind this argument is a general principle: If we are given
boundary values of a function, we can come up with a harmonic function
having these boundary values by assigning to each point the player’s
expected final fortune in a game where the player starts from the given
point and carries out a random walk until he reaches a boundary point,
where he receives the specified payoff. Furthermore, the Martingale
Stopping Theorern allows us to conclude that there can be no other
harmonic function with these boundary values. Thus martingale theory
allows us to establish existence and uniqueness of solutions to a Dirich-
let problem. All this isn’t very exciting for the cases we’ve been con-
sidering, but the nice thing is that the same arguments carry through
to the more general situations that we will be considering later on.

10
 The study of martingales was originated by Levy [19] and Ville
[34]. Kakutani [12] showed the connection between random walks and
harmonic functions. Doob [4] developed martingale stopping theorems
and showed how to exploit the preservation of fairness to solve a wide
variety of problems in probability theory. An informal discussion of
martingales may be found in Snell [32].

Exercise 1.1.9 Consider a random walk with a drift; that is, there is
a probability p 6= 21 of going one step to the right and a probability
q = 1 − p of going one step to the left. (See Exercise 1.1.2.) Let
w(x) = (q/p)x ; show that, if you interpret w(x) as your fortune when
you are at x, the resulting game is fair. Then use the Martingale
Stopping Theorem to argue that

w(x) = p(x)w(N ) + (1 − p(x))w(0).

Solve for p(x) to obtain

 x
q
p
−1
p(x) =  N .
q
p
−1

Exercise 1.1.10 You are gambling against a professional gambler; you

start with A dollars and the gambler with B dollars; you play a game
in which you win one dollar with probability p < 21 and lose one dollar
with probability q = 1 − p; play continues until you or the gambler runs
out of money. Let RA be the probability that you are ruined. Use the
result of Exercise 1.1.9 to show that
 B
p
1− q
RA =  N
p
1− q

with N = A + B. If you start with 20 dollars and the gambler with 50

dollars and p = .45, find the probability of being ruined.

Exercise 1.1.11 The gambler realizes that the probability of ruining

you is at least 1 − (p/q)B (Why?). The gambler wants to make the
probability at least .999. For this, (p/q)B should be at most .001. If
the gambler offers you a game with p = .499, how large a stake should
she have?

11
1.2 Random walks in two dimensions
1.2.1 An example
We turn now to the more complicated problem of a random walk on a
two-dimensional array. In Figure 3 we illustrate such a walk. The large

Figure 3: ♣

dots represent boundary points; those marked E indicate escape routes

and those marked P are police. We wish to find the probability p(x)
that our walker, starting at an interior point x, will reach an escape
route before he reaches a policeman. The walker moves from x = (a, b)
to each of the four neighboring points (a + 1, b), (a − 1, b), (a, b + 1),
(a, b−1) with probability 41 . If he reaches a boundary point, he remains
at this point.
The corresponding voltage problem is shown in Figure 4. The
boundary points P are grounded and points E are connected and fixed
at one volt by a one-volt battery. We ask for the voltage v(x) at the
interior points.

1.2.2 Harmonic functions in two dimensions

We now define harmonic functions for sets of lattice points in the plane
(a lattice point is a point with integer coordinates). Let S = D ∪ B
be a finite set of lattice points such that (a) D and B have no points
in common, (b) every point of D has its four neighboring points in S,
and (c) every point of B has at least one of its four neighboring points

12
 Figure 4: ♣

in D. We assume further that S hangs together in a nice way, namely,

that for any two points P and Q in S, there is a sequence of points
Pj in D such that P, P1 , P2 , . . . , Pn , Q forms a path from P to A. We
call the points of D the interior points of S and the points of B the
boundary points of S.
A function f defined on S is harmonic if, for points (a, b) in D, it
has the averaging property

f (a + 1, b) + f (a − 1, b) + f (a, b + 1) + f (a, b − 1)
f (a, b) = .
4
Note that there is no restriction on the values of f at the boundary
points.
We would like to prove that p(x) = v(x) as we did in the one-
dimensional case. That p(x) is harmonic follows again by considering
all four possible first steps; that v(x) is harmonic follows again by
Kirchhoff’s Laws since the current coming into x = (a, b) is

v(a + 1, b) − v(a, b) v(a − 1, b) − v(a, b) v(a, b + 1) − v(a, b) v(a, b − 1) − v(a, b)

+ + + = 0.
R R R R
Multiplying through by R and solving for v(a, b) gives

v(a + 1, b) + v(a − 1, b) + v(a, b + 1) + v(a, b − 1)

v(a, b) = .
4

13
 Thus p(x) and v(x) are harmonic functions with the same boundary
values. To show from this that they are the same, we must extend the
Uniqueness Principle to two dimensions.
We first prove the Maximum Principle. If M is the maximum value
of f and if f (P ) = M for P an interior point, then since f (P ) is the
average of the values of f at its neighbors, these values must all equal
M also. By working our way due south, say, repeating this argument at
every step, we eventually reach a boundary point Q for which we can
conclude that f (Q) = M . Thus a harmonic function always attains
its maximum (or minimum) on the boundary; this is the Maximum
Principle. The proof of the Uniqueness Principle goes through as before
since again the difference of two harmonic functions is harmonic.
The fair game argument, using the Martingale Stopping Theorem,
holds equally well and again gives an alternative proof of the existence
and uniqueness to the solution of the Dirichlet problem.

Exercise 1.2.1 Show that if f and g are harmonic functions so is

h = a · f + b · g for constants a and b. This is called the superposition
principle.

Exercise 1.2.2 Let B1 , B2 , . . . , Bn be the boundary points for a region

S. Let ej (a, b) be a function that is harmonic in S and has boundary
value 1 at Bj and 0 at the other boundary points. Show that if arbitrary
boundary values v1 , v2 , . . . , vn are assigned, we can find the harmonic
function v with these values from the solutions e1 , e2 , . . . , en .

1.2.3 The Monte Carlo solution

Finding the exact solution to a Dirichlet problem in two dimensions is
not always a simple matter, so before taking on this problem, we will
consider two methods for generating approximate solutions. In this
section we will present a method using random walks. This method
is known as a Monte Carlo method, since random walks are random,
and gambling involves randomness, and there is a famous gambling
casino in Monte Carlo. In Section 1.2.4, we will describe a much more
effective method for finding approximate solutions, called the method
of relaxations.
We have seen that the solution to the Dirichlet problem can be
found by finding the value of a player’s final winning in the following
game: Starting at x the player carries out a random walk until reaching
a boundary point. He is then paid an amount f (y) if y is the boundary

14
point first reached. Thus to find f (x), we can start many random walks
at x and find the average final winnings for these walks. By the law of
averages (the law of large numbers in probability theory), the estimate
that we obtain this way will approach the true expected final winning
f (x).
Here are some estimates obtained this way by starting 10,000 ran-
dom walks from each of the interior points and, for each x, estimating
f (x) by the average winning of the random walkers who started at this
point.
1 1
1.824 .785 1
1 .876 .503 .317 0
1 0 0
This method is a colorful way to solve the problem, but quite inef-
ficient. We can use probability theory to estimate how inefficient it is.
We consider the case with boundary values I or 0 as in our example.
In this case, the expected final winning is just the probability that the
walk ends up at a boundary point with value 1. For each point x, as-
sume that we carry out n random walks; we regard each random walk
to be an experiment and interpret the outcome of the ith experiment
to be a “success” if the walker ends at a boundary point with a 1 and
a “failure” otherwise. Let p = p(x) be the unknown probability for
success for a walker starting at x and q = 1 − p. How many walks
should we carry out to get a reasonable estimate for p? We estimate p
to be the fraction p̄ of the walkers that end at a 1.
We are in the position of a pollster who wishes to estimate the
proportion p of people in the country who favor candidate A over B.
The pollster chooses a random sample of n people and estimates p
as the proportion p̄ of voters in his sample who favor A. (This is a
gross oversimplification of what a pollster does, of course.) To estimate
the number n required, we can use the central limit theorem. This
theorem states that, if Sn , is the number of successes in n independent
experiments, each having probability p for success, then for any k > 0
!
Sn − np
P −k < √ < k ≈ A(k),
npq

where A(k) is the area under the normal curve between −k and k.
For k = 2 this area is approximately .95; what does this say about

15
p̄ = Sn /n? Doing a little rearranging, we see that
 
p̄ − p
P −2 < q pq < 2 ≈ .95
n

or √ √ !
pq pq
P −2 < p̄ − p < 2 ≈ .95.
n n
√
Since pq ≤ 21 ,
!
1 1 >
P − √ < p̄ − p < √ ≈ .95.
n n

Thus, if we choose √1n = .01, or n = 10, 000, there is a 95 percent chance

that our estimate p̄ = Sn /n will not be off by more than .01. This is
a large number for rather modest accuracy; in our example we carried
out 10,000 walks from each point and this required about 5 seconds on
the Dartmouth computer. We shall see later, when we obtain an exact
solution, that we did obtain the accuracy predicted.

Exercise 1.2.3 You play a game in which you start a random walk
at the center in the grid shown in Figure 5. When the walk reaches

Figure 5: ♣

16
the boundary, you receive a payment of +1 or −1 as indicated at the
boundary points. You wish to simulate this game to see if it is a
favorable game to play; how many simulations would you need to be
reasonably certain of the value of this game to an accuracy of .01?
Carry out such a simulation and see if you feel that it is a favorable
game.

1.2.4 The original Dirichlet problem; the method of relax-

ations
The Dirichlet problem we have been studying is not the original Dirich-
let problem, but a discrete version of it. The original Dirichlet problem
concerns the distribution of temperature, say, in a continuous medium;
the following is a representative example.
Suppose we have a thin sheet of metal gotten by cutting out a small
square from the center of a large square. The inner boundary is kept
at temperature 0 and the outer boundary is kept at temperature 1 as
indicated in Figure 6. The problem is to find the temperature at
points in the rectangle’s interior. If u(x, y) is the temperature at (x, y),
then u satisfies Laplace’s differential equation
uxx + uyy = 0.
A function that satisfies this differential equation is called harmonic.
It has the property that the value u(x, y) is equal to the average of the
values over any circle with center (x, y) lying inside the region. Thus to
determine the temperature u(x, y), we must find a harmonic function
defined in the rectangle that takes on the prescribed boundary values.
We have a problem entirely analogous to our discrete Dirichlet problem,
but with continuous domain.
The method of relaxations was introduced as a way to get approx-
imate solutions to the original Dirichlet problem. This method is ac-
tually more closely connected to the discrete Dirichlet problem than
to the continuous problem. Why? Because, faced with the continuous
problem just described, no physicist will hesitate to replace it with an
analogous discrete problem, approximating the continuous medium by
an array of lattice points such as that depicted in Figure 7, and search-
ing for a function that is harmonic in our discrete sense and that takes
on the appropriate boundary values. It is this approximating discrete
problem to which the method of relaxations applies.
Here’s how the method goes. Recall that we are looking for a func-
tion that has specified boundary values, for which the value at any

17
Figure 6: ♣

18
Figure 7: ♣

19
interior point is the average of the values at its neighbors. Begin with
any function having the specified boundary values, pick an interior
point, and see what is happening there. In general, the value of the
function at the point we are looking at will not be equal to the average
of the values at its neighbors. So adjust the value of the function to be
equal to the average of the values at its neighbors. Now run through
the rest of the interior points, repeating this process. When you have
adjusted the values at all of the interior points, the function that re-
sults will not be harmonic, because most of the time after adjusting the
value at a point to be the average value at its neighbors, we afterwards
came along and adjusted the values at one or more of those neighbors,
thus destroying the harmony. However, the function that results after
running through all the interior points, if not harmonic, is more nearly
harmonic than the function we started with; if we keep repeating this
averaging process, running through all of the interior points again and
again, the function will approximate more and more closely the solution
to our Dirichlet problem.
We do not yet have the tools to prove that this method works for
a general initial guess; this will have to wait until later (see Exercise
1.3.12). We will start with a special choice of initial values for which
we can prove that the method works (see Exercise 1.2.5).
We start with all interior points 0 and keep the boundary points
fixed.
1 1
1 0 0 1
1 0 0 0 0
1 0 0
After one iteration we have:
1 1
1 .547 .648 1
1 .75 .188 .047 0
1 0 0
Note that we go from left to right moving up each column replacing
each value by the average of the four neighboring values. The compu-
tations for this first iteration are

.75 = (1/4)(1 + 1 + 1 + 0)

.1875 = (1/4)(.75 + 0 + 0 + 0)
.5469 = (1/4)(.1875 + 1 + 1 + 0)

20
 .0469 = (1/4)(.1875 + 0 + 0 + 0)
.64844 = (1/4)(.0469 + .5769 + 1 + 1)
We have printed the results to three decimal places. We continue the
iteration until we obtain the same results to three decimal places. This
occurs at iterations 8 and 9. Here’s what we get:
1 1
1 .823 .787 1
1 .876 .506 .323 0
1 0 0
We see that we obtain the same result to three places after only nine
iterations and this took only a fraction of a second of computing time.
We shall see that these results are correct to three place accuracy. Our
Monte Carlo method took several seconds of computing time and did
not even give three place accuracy.
The classical reference for the method of relaxations as a means
of finding approximate solutions to continuous problems is Courant,
Friedrichs, and Lewy [3]. For more information on the relationship
between the original Dirichlet problem and the discrete analog, see
Hersh and Griego [10].

Exercise 1.2.4 Apply the method of relaxations to the discrete prob-

lem illustrated in Figure 7.

Exercise 1.2.5 Consider the method of relaxations started with an

initial guess with the property that the value at each point is ≤ the
average of the values at the neighbors of this point. Show that the
successive values at a point u are monotone increasing with a limit
f (u) and that these limits provide a solution to the Dirichlet problem.

1.2.5 Solution by solving linear equations

In this section we will show how to find an exact solution to a two-
dimensional Dirichlet problem by solving a system of linear equations.
As usual, we will illustrate the method in the case of the example
introduced in Section 1.2.1. This example is shown again in Figure
8; the interior points have been labelled a, b, c, d, and e. By our
averaging property, we have
xb + x d + 2
xa =
4
21
 Figure 8: ♣

xa + x c + 2
xb =
4
xd + 3
xc =
4
xa + x c + x e
xd =
4
xb + x d
xe = .
4
We can rewrite these equations in matrix form as
1 −1/4 0 −1/4 0 xa 1/2
    
 −1/4 1 0 0 −1/4   xb   1/2 
   
 
 0 0 1 −1/4 0   xc  =  3/4  .
    
    
 −1/4 0 −1/4 1 −1/4   xd   0 
0 −1/4 0 −1/4 1 xe 0
We can write this in symbols as
Ax = u.
Since we know there is a unique solution, A must have an inverse and
x = A−1 u.

22
Carrying out this calculation we find
.823
 
 .787 
 
 .876  .
Calculated x =  
 
 .506 
.323
Here, for comparison, are the approximate solutions found earlier:
.824
 
 .785 
 
Monte Carlo x =  .876  .
 
 .503 
 

.317
.823
 
 .787 
 
Relaxed x =  .876 .
 
 
 .506 
.323
We see that our Monte Carlo approximations were fairly good in that
no error of the simulation is greater than .01, and our relaxed approx-
imations were very good indeed, in that the error does not show up at
all.

Exercise 1.2.6 Consider a random walker on the graph of Figure 9.

Find the probability of reaching the point with a 1 before any of the
points with 0’s for each starting point a, b, c, d.

Exercise 1.2.7 Solve the discrete Dirichlet problem for the graph shown
in Figure 10. The interior points are a, b, c, d. (Hint: See Exercise
1.2.2.)

Exercise 1.2.8 Find the exact value, for each possible starting point,
for the game described in Exercise 1.2.3. Is the game favorable starting
in the center?

1.2.6 Solution by the method of Markov chains

In this section, we describe how the Dirichlet problem can be solved by
the method of Markov chains. This method may be viewed as a more
sophisticated version of the method of linear equations.

23
Figure 9: ♣

Figure 10: ♣

24
 A finite Markov chain is a special type of chance process that may
be described informally as follows: we have a set S = {s1 , s2 , . . . , sr }
of states and a chance process that moves around through these states.
When the process is in state si , it moves with probability Pij to the
state sj . The transition probabilities Pij are represented by an r-by-r
matrix P called the transition matrix. To specify the chance process
completely we must give, in addition to the transition matrix, a method
for starting the process. We do this by specifying a specific state in
which the process starts.
According to Kemeny, Snell, and Thompson [15], in the Land of
Oz, there are three kinds of weather: rain, nice, and snow. There are
never two nice days in a row. When it rains or snows, half the time it
is the same the next day. If the weather changes, the chances are equal
for a change to each of the other two types of weather. We regard the
weather in the Land of Oz as a Markov chain with transition matrix:
R N S
 
R 1/2 1/4 1/4
P = N
 1/2 0 1/2 .
S 1/4 1/4 1/2
When we start in a particular state, it is natural to ask for the
probability that the process is in each of the possible states after a
specific number of steps. In the study of Markov chains, it is shown
that this information is provided by the powers of the transition matrix.
Specifically, if Pn is the matrix P raised to the nth power, the entries
Pijn represent the probability that the chain, started in state si , will,
after n steps, be in state sj . For example, the fourth power of the
transition matrix P for the weather in the Land of Oz is
R N S
 
R .402 .199 .398
P4 = N  .398 .203 .398 .
 

S .398 .199 .402

Thus, if it is raining today in the Land of Oz, the probability that
the weather will be nice four days from now is .199. Note that the
probability of a particular type of weather four days from today is es-
sentially independent of the type of weather today. This Markov chain
is an example of a type of chain called a regular chain. A Markov chain
is a regular chain if some power of the transition matrix has no zeros.
In the study of regular Markov chains, it is shown that the probability

25
of being in a state after a large number of steps is independent of the
starting state.
As a second example, we consider a random walk in one dimension.
Let us assume that the walk is stopped when it reaches either state 0
or 4. (We could use 5 instead of 4, as before, but we want to keep the
matrices small.) We can regard this random walk as a Markov chain
with states 0, 1, 2, 3, 4 and transition matrix given by
0 1 2 3 4
 
0 1 0 0 0 0
1 1/2 0 1/2 0 0 
 
 
P = 2 0 1/2 0 1/2 0 .
3 0

0 1/2 0 1/2 

4 0 0 0 0 1
The states 0 and 4 are traps or absorbing states. These are states
that, once entered, cannot be left. A Markov chain is called absorbing
if it has at least one absorbing state and if, from any state, it is possible
(not necessarily in one step) to reach at least one absorbing state. Our
Markov chain has this property and so is an absorbing Markov chain.
The states of an absorbing chain that are not traps are called non-
absorbing.
When an absorbing Markov chain is started in a non-absorbing
state, it will eventually end up in an absorbing state. For non-absorbing
state si and absorbing state sj , we denote by Bij the probability that
the chain starting in si will end up in state sj . We denote by B the
matrix with entries Bij . This matrix will have as many rows as non-
absorbing states and as many columns as there are absorbing states.
For our random walk example, the entries Bx,4 will give the probability
that our random walker, starting at x, will reach 4 before reaching 0.
Thus, if we can find the matrix B by Markov chain techniques, we will
have a way to solve the Dirichlet problem.
We shall show, in fact, that the Dirichlet problem has a natural
generalization in the context of absorbing Markov chains and can be
solved by Markov chain methods.
Assume now that P is an absorbing Markov chain and that there are
u absorbing states and v non-absorbing states. We reorder the states
so that the absorbing states come first and the non-absorbing states
come last. Then our transition matrix has the canonical form:
I 0
 
P= .
R Q

26
Here I is a u-by-u identity matrix; 0 is a matrix of dimension u-by-v
with all entries 0.
For our random walk example this canonical form is:

0 4 1 2 3
 
0 1 0 0 0 0
4 0 1 0 0 0 
 
 
1 1/2 0 0 1/2 0 .
2 0

0 1/2 0 1/2 

3 0 1/2 0 1/2 0

The matrix N = (I − Q)−1 is called the fundamental matrix for the

absorbing chain P. (Note that I here is a v-by-v identity matrix!) The
entries Nij of this matrix have the following probabilistic interpretation:
Nij is the expected number of times that the chain will be in state sj
before absorption when it is started in si . (To see why this is true,
think of how (I − Q)−1 would look if it were written as a geometric
series.) Let 1 be a column vector of all 1’s. Then the vector t = NI
gives the expected number of steps before absorption for each starting
state.
The absorption probabilities B are obtained from N by the matrix
formula
B = (I − Q)−1 R.
This simply says that to get the probability of ending up at a given
absorbing state, we add up the probabilities of going there from all the
non-absorbing states, weighted by the number of times we expect to be
in those (non-absorbing) states.
For our random walk example
1
 
0 2
0
Q =  12 0 1
 
2 
1
0 2
0

− 21
 
1 0
 1
I − Q =  −2 1 − 12 


0 − 21 1
1 2 3
3 1
 
1 2
1 2
N = (I − Q)−1 = 21 2 1


3 12 1 3
2

27
 3 1
    
2
1 2
1 3
t = N1 = 
1 2 11 = 4
   
1 3
2
1 2
1 3
0 4
3 1 1 3 1
    
2
1 2 2
0 1 4 4
1 1
B = NR =  1 2 10 0 = 2 .
    
2 2
1 3 1 1 3
2
1 2
0 2
3 4 4
Thus, starting in state 3, the probability is 3/4 of reaching 4 before
0; this is in agreement with our previous results. From t we see that
the expected duration of the game, when we start in state 2, is 4.
For an absorbing chain P, the nth power Pn of the transition prob-
abilities will approach a matrix P∞ of the form
I 0
 
P ∞
= .
B 0
We now give our Markov chain version of the Dirichlet problem. We
interpret the absorbing states as boundary states and the non-absorbing
states as interior states. Let B be the set of boundary states and D the
set of interior states. Let f be a function with domain the state space
of a Markov chain P such that for i in D
X
f (i) = Pij f (j).
j

Then f is a harmonic function for P. Now f again has an averaging

property and extends our previous definition. If we represent f as a
column vector f , f is harmonic if and only if
Pf = f .
This implies that
P2 f = P · Pf = Pf = f
and in general
Pn f = f .
Let us write the vector f as
fB
 
f=
fD
where fB represents the values of f on the boundary and fD values on
the interior. Then we have
fB I 0 fB
    
=
fD B Q fD

28
and
fD = BfB .
We again see that the values of a harmonic function are determined
by the values of the function at the boundary points.
Since the entries Bij of B represent the probability, starting in i,
that the process ends at j, our last equation states that if you play a
game in which your fortune is fj when you are in state j, then your
expected final fortune is equal to your initial fortune; that is, fairness
is preserved. As remarked above, from Markov chain theory B = NR
where N = (I − Q)−1 . Thus

fD = (I − Q)−1 RfB .

(To make the correspondence between this solution and the solution of
Section 1.2.5, put A = I − Q and u = RfB .)
A general discussion of absorbing Markov chains may be found in
Kemeny, Snell, and Thompson [15].

Exercise 1.2.9 Consider the game played on the grid in Figure 11.
You start at an interior point and move randomly until a boundary

Figure 11: ♣

point is reached and obtain the payment indicated at this point. Using
Markov chain methods find, for each starting state, the expected value
of the game. Find also the expected duration of the game.

29
1.3 Random walks on more general networks
1.3.1 General resistor networks and reversible Markov chains
Our networks so far have been very special networks with unit resistors.
We will now introduce general resistor networks, and consider what it
means to carry out a random walk on such a network.
A graph is a finite collection of points (also called vertices or nodes)
with certain pairs of points connected by edges (also called branches).
The graph is connected if it is possible to go between any two points
by moving along the edges. (See Figure 12.)

Figure 12: ♣

We assume that G is a connected graph and assign to each edge xy

a resistance Rxy ; an example is shown in Figure 13. The conductance
of an edge xy is Cxy = 1/Rxy ; conductances for our example are shown
in Figure 14.
We define a random walk on G to be a Markov chain with transition
matrix P given by
Cxy
Pxy =
Cx
with Cx = y Cxy . For our example, Ca = 2, Cb = 3, Cc = 4, and
P

Cd = 5, and the transition matrix P for the associated random walk is

a b c d
1 1
a 0 0
 
2 2
1 2
b0 0 3 3


c  14 1
0 1
 
4 2

d 15 2
5
2
5
0

30
Figure 13: ♣

Figure 14: ♣

31
 Figure 15: ♣

Its graphical representation is shown in Figure 15.

Since the graph is connected, it is possible for the walker to go
between any two states. A Markov chain with this property is called
an ergodic Markov chain. Regular chains, which were introduced in
Section 1.2.6, are always ergodic, but ergodic chains are not always
regular (see Exercise 1.3.1).
For an ergodic chain, there is a unique probability vector w that is
a fixed vector for P, i.e., wP = w. The component wj of w represents
the proportion of times, in the long run, that the walker will be in state
j. For random walks determined by electric networks, the fixed vector
is given by wj = Cj /C, where C = x Cx . (You are asked to prove this
P

in Exercise 1.3.2.) For our example Ca = 2, Cb = 3, Cc = 4, Cd = 5,

and C = 14. Thus w = (2/14, 3/14, 4/14, 5/14). We can check that w
is a fixed vector by noting that
1 1
0 0
 
2 2
0 1 2
2 3 4 5 0 3 3

2 3 4 5
( 14 ) 1 = ( 14 ).
 
14 14 14 1 1 14 14 14
0

4 4 2
 
1 2 2
5 5 5
0

32
 In addition to being ergodic, Markov chains associated with net-
works have another property called reversibility. An ergodic chain is
said to be reversible if wx Pxy = wy Pyx for all x, y. That this is true for
our network chains follows from the fact that
Cxy Cyx
Cx Pxy = Cx = Cxy = Cyx = Cy = Cy Pyx .
Cx Cy

Thus, dividing the first and last term by C, we have wx Pxy = wy Pyx .
To see the meaning of reversibility, we start our Markov chain with
initial probabilities w (in equilibrium) and observe a few states, for
example
a c b d.
The probability that this sequence occurs is
2 1 1 2 1
wa Pac Pcb Pbd = · · · = .
14 2 4 3 84
The probability that the reversed sequence

d b c a

occurs is
5 2 1 1 1
wd Pdb Pbc Pca = · · · = .
14 5 3 4 84
Thus the two sequences have the same probability of occurring.
In general, when a reversible Markov chain is started in equilibrium,
probabilities for sequences in the correct order of time are the same as
those with time reversed. Thus, from data, we would never be able to
tell the direction of time.
If P is any reversible ergodic chain, then P is the transition ma-
trix for a random walk on an electric network; we have only to define
Cxy = wx Pxy . Note, however, if Pxx 6= 0 the resulting network will
need a conductance from x to x (see Exercise 1.3.4). Thus reversibility
characterizes those ergodic chains that arise from electrical networks.
This has to do with the fact that the physical laws that govern the
behavior of steady electric currents are invariant under time-reversal
(see Onsager [25]).
When all the conductances of a network are equal, the associated
random walk on the graph G of the network has the property that, from
each point, there is an equal probability of moving to each of the points
connected to this point by an edge. We shall refer to this random walk

33
as simple random walk on G. Most of the examples we have considered
so far are simple random walks. Our first example of a random walk
on Madison Avenue corresponds to simple random walk on the graph
with points 0, 1, 2, . . . , N and edges the streets connecting these points.
Our walks on two dimensional graphs were also simple random walks.

Exercise 1.3.1 Give an example of an ergodic Markov chain that is

not regular. (Hint: a chain with two states will do.)

Exercise 1.3.2 Show that, if P is the transition matrix for a random

walk determined by an electric network, then the fixed vector w is given
by wx = CCx where Cx = y Cxy and C = x Cx .
P P

Exercise 1.3.3 Show that, if P is a reversible Markov chain and a, b, c

are any three states, then the probability, starting at a, of the cycle
abca is the same as the probability of the reversed cycle acba. That
is Pab Pbc Pca = Pac Pcb Pba . Show, more generally, that the probability
of going around any cycle in the two different directions is the same.
(Conversely, if this cyclic condition is satisfied, the process is reversible.
For a proof, see Kelly [13].)

Exercise 1.3.4 Assume that P is a reversible Markov chain with Pxx =

0 for all x. Define an electric network by Cxy = wx Pxy . Show that the
Markov chain associated with this circuit is P. Show that we can allow
Pxx > 0 by allowing a conductance from x to x.

Exercise 1.3.5 For the Ehrenfest urn model, there are two urns that
together contain N balls. Each second, one of the N balls is chosen at
random and moved to the other urn. We form a Markov chain with
states the number of balls in one of the urns. For N = 4, the resulting
transition matrix is
0 1 2 3 4
 
0 0 1 0 0 0
1 3
1 0 0 0

4 1
4
1

P = 20 2
0 2
0.
3 1
30 0 4
0 4


4 0 0 0 1 0
1 4 6 4 1
Show that the fixed vector w is the binomial distribution w = ( 16 , 16 , 16 , 16 , 16 ).
Determine the electric network associated with this chain.

34
1.3.2 Voltages for general networks; probabilistic interpreta-
tion
We assume that we have a network of resistors assigned to the edges of
a connected graph. We choose two points a and b and put a one-volt
battery across these points establishing a voltage va = 1 and vb = 0,
as illustrated in Figure 16. We are interested in finding the volt-

Figure 16: ♣

ages vx and the currents ixy in the circuit and in giving a probabilistic
interpretation to these quantities.
We begin with the probabilistic interpretation of voltage. It will
come as no surprise that we will interpret the voltage as a hitting prob-
ability, observing that both functions are harmonic and that they have
the same boundary values.
By Ohm’s Law, the currents through the resistors are determined
by the voltages by
vx − v y
ixy = = (Vx − vy )Cxy .
Rxy
Note that ixy = −iyx . Kirchhoff’s Current Law requires that the total
current flowing into any point other than a or b is 0. That is, for x 6= a, b
X
ixy = 0.
y

35
This will be true if X
(vx − vy )Cxy = 0
y
or X X
vx Cxy = Cxy vy .
y y

Thus Kirchhoff’s Current Law requires that our voltages have the prop-
erty that
X Cxy X
vx = vy = Pxy vy
y Cx y

for x 6= a, b. This means that the voltage vx is harmonic at all points

x 6= a, b.
Let hx be the probability, starting at x, that state a is reached before
b. Then hx is also harmonic at all points x 6= a, b. Furthermore

va = ha = 1

and
vb = hb = 0.
Thus if we modify P by making a and b absorbing states, we obtain
an absorbing Markov chain P̄ and v and h are both solutions to the
Dirichlet problem for the Markov chain with the same boundary values.
Hence v = h.
For our example, the transition probabilities P̄xy are shown in Figure
17. The function vx is harmonic for P̄ with boundary values va =
1, vb = 0.
To sum up, we have the following:
Intrepretation of Voltage. When a unit voltage is applied be-
tween a and b, making va = 1 and vb = 0, the voltage vx at any point x
represents the probability that a walker starting from x will return to
a before reaching b.
In this probabilistic interpretation of voltage, we have assumed a
unit voltage, but we could have assumed an arbitrary voltage va be-
tween a and b. Then the hitting probability hx would be replaced by
an expected value in a game where the player starts at x and is paid
va if a is reached before b and 0 otherwise.
Let’s use this interpretation of voltage to find the voltages for our
example. Referring back to Figure 17, we see that

va = 1

36
 Figure 17: ♣

vb = 0
1 1
vc = + vd
4 2
1 2
vd = + vc .
5 5
7
Solving these equations yields vc = 16 and vd = 38 . From these
7
voltages we obtain the current ixy . For example icd = ( 16 − 38 ) · 2 = 81 .
The resulting voltages and currents are shown in Figure 18. The
7
voltage at c is 16 and so this is also the probability, starting at c, of
reaching a before b.

1.3.3 Probabilistic interpretation of current

We turn now to the probabilistic interpretation of current. This in-
terpretation is found by taking a naive view of the process of electri-
cal conduction: We imagine that positively charged particles enter the
network at point a and wander around from point to point until they
finally arrive at point b, where they leave the network. (It would be
more realistic to imagine negatively charged particles entering at b and

37
Figure 18: ♣

38
leaving at a, but realism is not what we’re after.) To determine the
current ixy along the branch from x to y, we consider that in the course
of its peregrinations the point may pass once or several times along the
branch from x to y, and in the opposite direction from y to x. We may
now hypothesize that the current ixy is proportional to the expected net
number of movements along the edge from x to y, where movements
from y back to x are counted as negative. This hypothesis is correct,
as we will now show.
The walker begins at a and walks until he reaches b; note that if
he returns to a before reaching b, he keeps on going. Let ux be the
expected number of visits to state x before reaching b. Then ub = 0
and, for x 6= a, b, X
ux = uy Pyx .
y

This last equation is true because, for x 6= a, b, every entrance to x

must come from some y.
We have seen that Cx Pxy = Cy Pyx ; thus
X Pxy Cx
ux = uy
y Cy
or
ux X uy
= Pxy .
Cx y Cy
This means that vx = ux /Cx is harmonic for x 6= a, b. We have also
vb = 0 and va = ua /Ca . This implies that vx is the voltage at x when
we put a battery from a to b that establishes a voltage ua /Ca at a and
voltage 0 at b. (We remark that the expression vx = ux /Cx may be
understood physically by viewing ux as charge and Cx as capacitance;
see Kelly [13] for more about this.)
We are interested in the current that flows from x to y. This is
!
ux uy ux Cxy uy Cyx
ixy = (vx −vy )Cxy = − Cxy = − = ux Pxy −uy Pyx .
Cx Cy Cx Cy

Now ux Pxy is the expected number of times our walker will go from x
to y and uy Pyx is the expected number of times he will go from y to
x. Thus the current ixy is the expected value for the net number of
times the walker passes along the edge from x to y. Note that for any
particular walk this net value will be an integer, but the expected value
will not.

39
 As we have already noted, the currents ixy here are not those of our
original electrical problem, where we apply a 1-volt battery, but they
are proportional to those original currents. To determine the constant
of proportionality, we note the following characteristic property of the
new currents ixy : The total current flowing into the network at a (and
out at b) is 1. In symbols,
X
iay = 1.
y

Indeed, from our probabilistic interpretation of ixy this sum represents

the expected value of the difference between the number of times our
walker leaves a and enters a. This number is necessarily one and so the
current flowing into a is 1.
This unit current flow from a to b can be obtained from the currents
in the original circuit, corresponding to a 1-volt battery, by dividing
through by the total amount of current y iay flowing into a; doing
P

this to the currents in our example yields the unit current flow shown
in Figure 19.

Figure 19: ♣

This shows that the constant of proportionality we were seeking to

determine is the reciprocal of the amount of current that flows through

40
the circuit when a I-volt battery is applied between a and b. This
quantity, called the effective resistance between a and b, is discussed in
detail in Section 1.3.4.
To sum up, we have the following:
Interpretation of Current. When a unit current flows into a
and out of b, the current ixy flowing through the branch connecting x to
y is equal to the expected net number of times that a walker, starting
at a and walking until he reaches b, will move along the branch from x
to y. These currents are proportional to the currents that arise when a
unit voltage is applied between a and b, the constant of proportionality
being the effective resistance of the network.
We have seen that we can estimate the voltages by simulation. We
can now do the same for the currents. We have to estimate the expected
value for the net number of crossings of xy. To do this, we start a large
number of walks at a and, for each one, record the net number of
crossings of each edge and then average these as an estimate for the
expected value. Carrying out 10,000 such walks yielded the results
shown in Figure 20.

Figure 20: ♣

The results of simulation are in good agreement with the theoretical

values of current. As was the case for estimates of the voltages by
simulation, we have statistical errors. Our estimates have the property

41
that the total current flowing into a is 1, out of b is 1, and into any
other point it is 0. This is no accident; the explanation is that the
history of each walk would have these properties, and these properties
are not destroyed by averaging.

Exercise 1.3.6 Kingman [17] introduced a different model for current

flow. Kelly [13] gave a new interpretation of this model. Both authors
use continuous time. A discrete time version of Kelly’s interpretation
would be the following: At each point of the graph there is a black
or a white button. Each second an edge is chosen; edge xy is chosen
with probability Cxy /C where C is the sum of the conductances. The
buttons on the edge chosen are then interchanged. When a button
reaches a it is painted black, and when a button reaches b it is painted
white. Show that there is a limiting probability px that site x has a
black button and that px is the voltage vx at x when a unit voltage is
imposed between a and b. Show that the current ixy is proportional to
the net flow of black buttons along the edge xy. Does this suggest a
hypothesis about the behavior of conduction electrons in metals?

1.3.4 Effective resistance and the escape probability

When we impose a voltage v between points a and b, a voltage va = v
is established at a and vb = 0, and a current ia = x iax will flow
P

into the circuit from the outside source. The amount of current that
flows depends upon the overall resistance in the circuit. We define
the effective resistance Reff between a and b by Reff = va /ia . The
reciprocal quantity Ceff = 1/Reff = ia /va is the effective conductance.
If the voltage between a and b is multiplied by a constant, then the
currents are multiplied by the same constant, so Reff depends only on
the ratio of va to ia .
Let us calculate Reff for our example. When a unit voltage was im-
posed, we obtained the currents shown in Figure 18. The total current
flowing into the circuit is ia = 9/16+10/16 = 19/16. Thus the effective
resistance is
va 1 16
Reff = = 19 = .
ia 16
19
We can interpret the effective conductance probabilistically as an
escape probability. When va = 1, the effective conductance equals the
total current ia flowing into a. This current is
X X Cay X
ia = (va − vy )Cay = (va − vy ) Ca = Ca (1 − Pay vy ) = Ca pesc
y y Ca y

42
where pesc is the probability, starting at a, that the walk reaches b
before returning to a. Thus

Ceff = Ca pesc

and
Ceff
pesc = .
Ca
In our example Ca = 2 and we found that ia = 19/16. Thus
19
pesc = .
32
In calculating effective resistances, we shall use two important facts
about electric networks. First, if two resistors are connected in series,
they may be replaced by a single resistor whose resistance is the sum of
the two resistances. (See Figure 21.) Secondly, two resistors in parallel

Figure 21: ♣

may be replaced by a single resistor with resistance R such that

1 1 1 R1 R2
= + = .
R R1 R2 R1 + R 2
(See Figure 22.)
The second rule can be stated more simply in terms of conductances:
If two resistors are connected in parallel, they may be replaced by a
single resistor whose conductance is the sum of the two conductances.
We illustrate the use of these ideas to compute the effective resis-
tance between two adjacent points of a unit cube of unit resistors, as
shown in Figure 23. We put a unit battery between a and b. Then,
by symmetry, the voltages at c and d will be the same as will those at

43
Figure 22: ♣

Figure 23: ♣

44
 Figure 24: ♣

e and f . Thus our circuit is equivalent to the circuit shown in Figure

24.
Using the laws for the effective resistance of resistors in series and
parallel, this network can be successively reduced to a single resistor
of resistance 7/12 ohms, as shown in Figure 25. Thus the effective
resistance is 7/12. The current flowing into a from the battery will be
ia = R 1 = 12/7. The probability that a walk starting at a will reach
eff
b before returning to a is
12
ia 4
pesc = = 7 = .
Ca 3 7
This example and many other interesting connections between elec-
tric networks and graph theory may be found in Bollobas [2].

Exercise 1.3.7 A bug walks randomly on the unit cube (see Figure
26). If the bug starts at a, what is the probability that it reaches food
at b before returning to a?

Exercise 1.3.8 Consider the Ehrenfest urn model with N = 4 (see

Exercise 1.3.5). Find the probability, starting at 0, that state 4 is
reached before returning to 0.

Exercise 1.3.9 Consider the ladder network shown in Figure 27. Show
that if Rn is the effective resistance of a ladder with n rungs then R1 = 2
and
2 + 2Rn
Rn+1 = .
2 + Rn

45
Figure 25: ♣

46
Figure 26: ♣

Figure 27: ♣

47
 √
Use this to show that limn→∞ Rn = 2.

Exercise 1.3.10 A drunken tourist starts at her hotel and walks at

random through the streets of the idealized Paris shown in Figure 28.
Find the probability that she reaches the Arc de Triomphe before she

Figure 28: ♣

reaches the outskirts of town.

1.3.5 Currents minimize energy dissipation

We have seen that when we impose a voltage between a and b voltages vx
are established at the points and currents ixy flow through the resistors.
In this section we shall give a characterization of the currents in terms of

48
a quantity called energy dissipation. When a current ixy flows through
a resistor, the energy dissipated is

i2xy Rxy ;

this is the product of the current ixy and the voltage vxy = ixy Rxy . The
total energy dissipation in the circuit is
1X 2
E= i Rxy .
2 x,y xy

Since ixy Rxy = vx − vy , we can also write the energy dissipation as

1X
E= ixy (vx − vy ).
2 x,y

The factor 1/2 is necessary in this formulation since each edge is counted
twice in this sum. For our example, we see from Figure 18 that
2 2 2 2 2
9 10 7 2 1 12 1 19
    
E= ·1+ ·1+ ·1+ · + · = .
16 16 16 16 2 16 2 16
If a source (battery) establishes voltages va and vb at a and b, then
the energy supplied is (va − vb )ia where ia = x iax . By conservation
P

of energy, we would expect this to be equal to the energy dissipated.

In our example va − vb = 1 and ia = 19 16
, so this is the case. We shall
show that this is true in a somewhat more general setting.
Define a flow j from a to b to be an assignment of numbers jxy to
pairs xy such that
(a) jxy = −jyx

(b) y jxy = 0 if x 6= a, b
P

(c) jxy = 0 if x and y are not adjacent.

We denote by jx = y jxy the flow into x from the outside. By (b)
P

jx = 0 for x 6= a, b. Of course jb = −ja . To verify this, note that

X XX 1X
ja + j b = jx = jxy = (jxy + jyx ) = 0,
x x y 2 x,y

since jxy = −jyx .

With this terminology, we can now formulate the following version
of the principle of conservation of energy:

49
 Conservation of Energy. Let w be any function defined on the
points of the graph and j a flow from a to b. Then
1X
(wa − wb )ja = (wx − wy )jxy .
2 x,y
Proof.
X X X X X
(wx − wy )jxy = (wx jxy ) − (wy jxy )
x,y x y y x
X X X X
= wa jay + wb jby − wa jxa − wb jxb
y y x x
= wa ja + wb jb − wa (−ja ) − wb (−jb )
= 2(wa − wb )ja .
Thus
1X
(wa − wb )ja = (wx − wy )jxy
2 x,y
as was to be proven. ♦
If we now impose a voltage va between a and b with vb = 0, we
obtain voltages vx and currents ixy . The currents i give a flow from a
to b and so by the previous result, we conclude that
1X 1X 2
v a ia = (vx − vy )ixy = i Rxy .
2 x,y 2 x,y xy
Recall that Reff = va /ia . Thus in terms of resistances we can write this
as
1X 2
i2xy Reff = i Rxy .
2 x,y xy
If we adjust va so that ia = 1, we call the resulting flow the unit
current flow from a to b. The unit current flow from a to b is a particular
example of a unit flow from a to b, which we define to be any flow ixy
from a to b for which ia = −ib = 1. The formula above shows that the
energy dissipated by the unit current flow is just Reff . According to a
basic result called Thomson’s Principle, this value is smaller than the
energy dissipated by any other unit flow from a to b. Before proving
this principle, let us watch it in action in the example worked out above.
Recall that, for this example, we found the true values and some
approximate values for the unit current flow; these were shown in Figure
20. The energy dissipation for the true currents is
16
E = Reff = = .8421053.
19
50
Our approximate currents also form a unit flow and, for these, the
energy dissipation is
1 1
Ē = (.4754)2 ·1+(.5246)2 ·1+(.3672)2·1+(.1082)2 · +(.6328)2 · = .8421177.
2 2
We note that Ē is greater than E, though just barely.
Thomson’s Principle. (Thomson [33]). If i is the unit flow
from a to b determined by Kirchhoff’s Laws, then the energy dissipation
1 P 2 1 P 2
2 x,y ixy Rxy minimizes the energy dissipation 2 x,y jxy Rxy among all
unit flows j from a to b.
Proof. Let j be any unit flow from a to b and let dxy = jxy − ixy .
Then d is a flow from a to b with da = x dax = 1 − 1 = 0.
P

2
(ixy + dxy )2 Rxy
X X
jxy Rxy =
x,y x,y

i2xy Rxy + 2 d2xy Rxy

X X X
= ixy Rxy dxy +
x,y x,y x,y

i2xy Rxy d2xy Rxy .

X X X
= +2 (vx − vy )dxy +
x,y x,y x,y

Setting w = v and j = d in the conservation of energy result above

shows that the middle term is 4(va − vb )da = 0. Thus
2
i2xy Rxy + d2xy Rxy ≥ i2xy Rxy .
X X X X
jxy Rxy =
x,y x,y x,y x,y

This completes the proof. ♦

Exercise 1.3.11 The following is the so-called “dual form” of Thom-

son’s Principle. Let u be any function on the points of the graph G of
a circuit such that ua = 1 and ub = 0. Then the energy dissipation
1X
(ux − uy )2 Cxy
2 x,y

is minimized by the voltages vx that result when a unit voltage is es-

tablished between a and b, i.e., va = 1, vb = 0, and the other voltages
are determined by Kirchhoff’s Laws. Prove this dual principle. This
second principle is known nowadays as the Dirichlet Principle, though
it too was discovered by Thomson.

Exercise 1.3.12 In Section 1.2.4 we stated that, to solve the Dirichlet

problem by the method of relaxations, we could start with an arbitrary

51
initial guess. Show that when we replace the value at a point by the
average of the neighboring points the energy dissipation, as expressed in
Exercise 1.3.11, can only decrease. Use this to prove that the relaxation
method converges to a solution of the Dirichlet problem for an arbitrary
initial guess.

1.4 Rayleigh’s Monotonicity Law

1.4.1 Rayleigh’s Monotonicity Law
Next we will study Rayleigh’s Monotonicity Law. This law from electric
network theory will be an important tool in our future study of random
walks. In this section we will give an example of the use of this law.
Consider a random walk on streets of a city as in Figure 29. Let

Figure 29: ♣

pesc be the probability that a walker starting from a reaches b before

returning to a. Assign to each edge a unit resistance and maintain a
voltage of one volt between a and b; then a current ia will flow into the
circuit and we showed in Section 1.3.4 that
ia ia
pesc = = .
Ca 4
Now suppose that one of the streets (not connected to a) becomes
blocked. Our walker must choose from the remaining streets if he

52
reaches a corner of this street. The modified graph will be as in Figure
30. We want to show that the probability of escaping to b from a is

Figure 30: ♣

decreased.
Consider this problem in terms of our network. Blocking a street
corresponds to replacing a unit resistor by an infinite resistor. This
should have the effect of increasing the effective resistance Reff of the
circuit between a and b. If so, when we put a unit voltage between a
and b less current will flow into the circuit and
ia 1
pesc = =
4 4Reff

will decrease.
Thus we need only show that when we increase the resistance in one
part of a circuit, the effective resistance increases. This fact, known as
Rayleigh’s Monotonicity Law, is almost self-evident. Indeed, the father
of electromagnetic theory, James Clerk Maxwell, regarded this to be
the case. In his Treatise on Electricity and Magnetism ([21], p. 427),
he wrote

If the specific resistance of any portion of the conductor

be changed, that of the remainder being unchanged, the
resistance of the whole conductor will be increased if that

53
 of the portion is increased, and diminished if that of the
portion is diminished. This principle may be regarded as
self-evident . . . .
Rayleigh’s Monotonicity Law. If the resistances of a circuit
are increased, the effective resistance Reff between any two points can
only increase. If they are decreased, it can only decrease.
Proof. Let i be the unit current flow from a to b with the resistors
Rxy . Let j be the unit current flow from a to b with the resistors R̄xy
with R̄xy ≥ Rxy . Then
1X 2 1X 2
R̄eff = jxy R̄xy ≥ j Rxy .
2 x,y 2 x,y xy

But since j is a unit flow from a to b, Thomson’s Principle tells us that

the energy dissipation, calculated with resistors Rxy , is bigger than that
for the true currents determined by these resistors: that is
1X 2 1X 2
jxy Rxy ≥ i Rxy = Reff .
2 x,y 2 x,y xy

Thus R̄eff ≥ Reff . The proof for the case of decreasing resistances is
the same.

Exercise 1.4.1 Consider a graph G and let Rxy and R̄xy be two differ-
ent assignments of resistances to the edges of G. Let R̂xy = R̄xy + Rxy .
Let Reff , R̄eff , and R̂eff be the effective resistances when R, R̄, and R̂,
respectively, are used. Prove that

R̂eff ≥ R̄eff + Reff .

Conclude that the effective resistance of a network is a concave function

of any of its component resistances (Shannon and Hagelbarger [31].)

Exercise 1.4.2 Show that the effective resistance of the circuit in Fig-
ure 31 is greater than or equal to the effective resistance of the circuit
in Figure 32. Use this to show the following inequality for Rij ≥ 0:
1 1 1
1 1 ≥ 1 1 + 1 1 .
R11 +R12
+ R21 +R22 R11
+ R21 R12
+ R22

See the note by Lehman [18] for a proof of the general Minkowski
inequality by this method.

54
 Figure 31: ♣

Figure 32: ♣

Exercise 1.4.3 Let P be the transition matrix associated with an elec-

tric network and let a, b, r, s be four points on the network. Let P̄ be a
transition matrix defined on the state-space S = {a, b, r, s}. Let P̄ii = 0
and for i 6= j let P̄ij be the probability that, if the chain P is started
in state i, then the next time it is in the set S − {i} it is in the state j.
Show that P̄ is a reversible Markov chain and hence corresponds to an
electric network of the form of a Wheatstone Bridge, shown in Figure
33. Explain how this proves that, in order to prove the Monotonicity
Law, it is sufficient to prove that Reff is a monotone function of the
component resistances for a Wheatstone Bridge. Give a direct proof of
the Monotonicity Law for this special case.

1.4.2 A probabilistic explanation of the Monotonicity Law

We have quoted Maxwell’s assertion that Rayleigh’s Monotonicity Law
may be regarded as self-evident, but one might feel that any argument
in terms of electricity is only self-evident if we know what electricity is.
In Cambridge, they tell the following story about Maxwell: Maxwell
was lecturing and, seeing a student dozing off, awakened him, asking,
“Young man, what is electricity?” “I’m terribly sorry, sir,” the student

55
 Figure 33: ♣

replied, ‘I knew the answer but I have forgotten it.” Maxwell’s response
to the class was, “Gentlemen, you have just witnessed the greatest
tragedy in the history of science. The one person who knew what
electricity is has forgotten it.”
To say that our intuition about the Monotonicity Law is only as
solid as our understanding of electricity is not really a valid argument,
of course, because in saying that this law is self-evident we are secretly
depending on the analogy between electricity and the flow of water (see
Feynman [6], Vol. 2, Chapter 12). We just can’t believe that if a water
main gets clogged the total rate of flow out of the local reservoir is
going to increase. But as soon as we admit this, some pedant will ask
if we’re talking about flows with low Reynolds number, or what, and
we’ll have to admit that we don’t understand water any better than we
understand electricity.
Whatever our feelings about electricity or the flow of water, it seems
desirable to have an explanation of the Monotonicity Law in terms of
our random walker. We now give such an explanation.
We begin by collecting some notation and results from previous
sections. As usual, we have a network of conductances (streets) and a
walker who moves from point x to point y with probability
Cxy
Pxy =
Cx
where Cxy is the conductance from x to y and Cx = y Cxy . We choose
P

two preferred points a and b. The walker starts at a and walks until
he reaches b or returns to a. We denote by vx the probability that the
walker, starting at a, reaches a before b. Then va = 1, vb = 0, and the

56
function vx is harmonic at all points x 6= a, b. We denote by pesc the
probability that the walker, starting at a, reaches b before returning to
a. Then X
pesc = 1 − pax vx .
x

Now we have seen that the effective conductance between a and b

is
Ca pesc .
We wish to show that this increases whenever one of the conductances
Crs is increased. If a is different from r or s, we need only show that pesc
increases. The case where r or s coincides with a is easily disposed of
(see Exercise 1.4.4). The case where r or s coincides with b is also easy
(see Exercise 1.4.5). Hence from now on we will assume that r, s 6= a
and r, s 6= b.
Instead of increasing Crs , we can think of adding a new edge of
conductance  between r and s. (See Figure 34.) We will call this

Figure 34: ♣

new edge a “bridge” to distinguish it from the other edges. Note that
the graph with the bridge added will have more than one edge between

57
r and s (unless there was no edge between r and s in the original
graph), and this will complicate any expression that involves summing
over edges. Everything we have said or will say holds for graphs with
“multiple edges” as well as for graphs without them. So far, we have
chosen to keep our expressions simple by assuming that an edge is
determined by its endpoints. The trade-off is that in the manipulations
below, whenever we write a sum over edges we will have to add an extra
term to account for the bridge.
Why should adding the bridge increase the escape probability? The
first thing you think is, “Of course, it opens up new possibilities of
escaping!” The next instant you think, “Wait a minute, it also opens
up new possibilities of returning to the starting point. What ensures
that the first effect will outweigh the second?” As we shall see, the
proper reply is, “Because the walker will cross the bridge more often in
the good direction than in the bad direction.” To turn this blithe reply
into a real explanation will require a little work, however.
To begin to make sense of the notion that the bridge gets used more
often in the good direction than the bad, we will make a preliminary
argument that applies to any edge of any graph. Let G be any graph,
and let rs be any edge with endpoints not a or b. vr > vs . Since the
walker has a better chance to escape from s than from r, this means
that to cross this edge in the good direction is to go from r to s We
shall show that the walker will cross the edge from r to s more often
on the average than from s to r.
Let ux be the expected number of times the walker is at x and uxy
the expected number of times he crosses the edge xy from x to y before
he reaches b or returns to a. The calculation carried out in Section
1.3.3 shows that ux /Cx is harmonic for x 6= a, b with ua /Ca = 1/Ca
and bb /Cb = 0. But the function vx /Ca also has these properties, so by
the Uniqueness Principle
ux vx
= .
Cx Ca
Now
Crs Crs
urs = ur Prs = ur = vr
Cr Ca
and
Csr Csr
usr = us Psr = us = vs .
Cs Ca
Since Crs = Csr , and since by assumption vr ≥ vs , this means that
urs ≥ usr . Therefore, we see that any edge leads the walker more often
to the more favorable of the points of the edge.

58
 Now let’s go back and think about the graph with the bridge. The
above discussion shows that the bridge helps in the sense that, on the
average, the bridge is crossed more often in the direction that improves
the chance of escaping. While this fact is suggestive, it doesn’t tell us
that we are more likely to escape than if the bridge weren’t there; it
only tells us what goes on once the bridge is in place. What we need is
to make a “before and after” comparison.
Recall that we are denoting the conductance of the bridge by . To
distinguish the quantities pertaining to the walks with and without the
bridge, we will put () superscripts on quantities that refer to the walk
()
with the bridge, so that, e.g., pesc denotes the escape probability with
the bridge.
Now let d() denote the expected net number of times the walker
crosses from r to s. As above, we have

ur() u()
!
()   s
d = ur() − us() = − .
Cr +  Cs +  Cr +  Cs + 

Claim.
()
pesc = pesc + (vr − vs )d() .
Why. Every time you use the bridge to go from r to s, you improve
your chances of escaping by

(1 − vs ) − (1 − vr ) = vr − vs

assuming that you would continue your walk without using the bridge.
To get the probability of escaping with the bridge, you take the prob-
ability of escaping without the bridge, and correct it by adding in the
change attributable to the bridge, which is the difference in the origi-
nal escape probabilities at the ends of the bridge, multiplied by the net
number of times you expect to cross the bridge. ♥
Proof. Suppose you’re playing a game where you walk around the
graph with the bridge, and your fortune when you’re at x is vx , which
is the probability that you would return to a before reaching b if the
bridge weren’t there. You start at a, and walk until you reach b or
return to a.
This is not a fair game. Your initial fortune is 1 since you start at
a and va = 1. Your expected final fortune is
() () ()
1 · (1 − pesc ) + 0 · pesc = 1 − pesc .

59
The amount you expect to lose by participating in the game is
()
pesc .

(Note that escaping has suddenly become a bad thing!)

Let’s analyze where it is that you expect to lose money. First of all,
you lose money when you take the first step away from a. The amount
you expect to lose is
()
X
1− Pax vx = pesc .
x

Now if your fortune were given by vx() instead of vx , the game would
be fair after this first step. However, the function vx is not harmonic
for the walk with the bridge; it fails to be harmonic at r and s. Every
time you step away from r, you expect to lose an amount
!
X Crs  
vr − vx + vs = (vr − vs ) .
x Cr +  Cr +  Cr + 
Similarly, every time you step away from s you expect to lose an amount

(vs − vr ) .
Cs + 
The total amount you expect to lose by participating in the game
is:

expected loss at first step +

(expected loss at r) · (expected number of times at r) +
(expected loss at s) · (expected number of times at s)
= pesc +

(vr − vs ) u() +
Cr +  r

(vs − vr ) u()
Cs +  s
= pesc + (vr − vs )d() .

Equating this with our first expression for the expected cost of playing
the game yields the formula we were trying to prove.
According to the formula just established,

ur() us()
!
()
pesc − pesc = (vr − vs ) − .
Cr +  Cs + 

60
For small , we will have

ur() us() ur us vr vs
− ≈ − = − ,
Cr +  Cs +  Cr Cs Ca Ca
so for small 
() 
pesc − pesc ≈ (vr − vs )2.
Ca
This approximation allows us to conclude that
()
pesc ≥ pesc ≥ 0

for small . But this is enough to establish the monotonicity law, since
any finite change in  can be realized by making an infinite chain of
graphs each of which is obtained from the one before by adding a bridge
of infinitesimal conductance.
To recapitulate, the difference in the escape probabilities with and
without the bridge is obtained by taking the difference between the
original escape probabilities at the ends of the bridge, and multiplying
by the expected net number of crossings of the bridge. This quantity
is positive because the walker tends to cross the bridge more often in
the good direction than in the bad direction.

Exercise 1.4.4 Give a probabilistic argument to show that Ca pesc

increases with Car for any r. Give an example to show that pesc by
itself may actually decrease.

Exercise 1.4.5 Give a probabilistic argument to show that Ca pesc

increases with Crb for any r.

Exercise 1.4.6 Show that when vr = vs , changing the value of Crs

does not change pesc .

Exercise 1.4.7 Show that

∂
R = i2rs .
∂Rrs eff
Exercise 1.4.8 In this exercise we ask you to derive an exact formula
for the change in escape probability
()
pesc − pesc ,

in terms of quantities that refer only to the walk without the bridge.

61
 (a) Let Nxy denote the expected number of times in state y for a
walker who starts at x and walks around the graph without the bridge
until he reaches a or b. It is a fact that
 
ur() = ur + ur() (Nsr + 1 − Nrr ) + us() (Nrr − Nsr ).
Cr +  Cs + 
Explain in words why this formula is true.
(b) This equation for ur() can be rewritten as follows:

Cr
ur() = ur + d() (Nsr − Nrr ).
Cr + 
Prove this formula. (Hint: Consider a game where your fortune at x is
Nxr , and where you start from a and walk on the graph with the bridge
until you reach b or return to a.)
(c) Write down the corresponding formula for us() , and use this
formula to get an expression for d() in terms of quantities that refer to
the walk without the bridge.
()
(d) Use the expression for d() to express pesc − pesc in terms of
quantities that refer to the walk without the bridge, and verify that
the value of your expression is ≥ 0 for  ≥ 0.

Exercise 1.4.9 Give a probabilistic interpretation of the energy dissi-

pation rate.

1.4.3 A Markov chain proof of the Monotonicity Law

Let P be the ergodic Markov chain associated with an electric network.
When we add an  bridge from r to s, we obtain a new transition
matrix P() that differs from P only for transitions from r and s. We
can minimize the differences between P and P() by replacing P by the
matrix P̂ corresponding to the circuit without the bridge but with an 
conductance added from r to r and from s to s. This allows the chain
to stay in states r and s but does not change the escape probability
from a to b. Thus, we can compare the escape probabilities for the two
matrices P̂ and P() , which differ only by
 ()
P̂rr = Cr +
Prr =0
Crs
P̂rs = Cr +
()
Prs = CCrsr +
+

 () .
P̂ss = Cs +
Pss =0
Csr
P̂sr = CS +
()
Psr = CCsrs +
+

62
 We make states a and b into absorbing states. Let N̂ and N() be
the fundamental matrices for the absorbing chains obtained from P̂ and
P() respectively. Then N̂ = (I − Q̂)−1 and N() = (I − Q() )−1 where
Q̂ and Q() differ only for the four components involving only r and s.
That is,
r s
 
!
() r − Cr + Cr +
Q = Q̂ +  = Q̂ + hk
s Cs +
− Cs+
where h is the column vector with only components r and s non-zero


!
r Cr +
h=
s − Cs+

and k is a row vector with only components r and s non-zero

 r s
k= −1 1 .

J. G. Kemeny has pointed out to us that if A is any matrix with

inverse N and we add to A a matrix of the form −hk, then A − hk
has an inverse if and only if kNh 6= 1 and, if so, N̄ = (A − hk)−1 is
given by
N̄ = N + c(Nh)(kN)
where c = 1/(1 − kNh). You are asked to prove this in Exercise 1.4.10.
Adding −hk to A = I − Q̂ and using this result, we obtain

N() = N̂ + c(N̂h)(kN̂).

Using the simple nature of h and k we obtain

!
() N̂ir  N̂is 
Nij = N̂ij + − (N̂sj − N̂rj )
Cr +  Cs + 
and
1
c= N̂rr  N̂sr  N̂ss  N̂rs 
.
1+ Cr +
− Cr +
+ Cs +
− Cs +

Since N̂rr is the expected number of times in r starting in r and N̂sr

is the expected number of times in r starting in s, N̂rr ≥ N̂sr . Similarly
N̂ss ≥ N̂rs and so the denominator of c is ≥ 1. In particular, it is
positive.

63
 Recall that the absorption probabilities for state b are given by
X
Bxb = Nxy Pyb .
y

()
Since Pxb = P̂xb ,
!
() N̂xr  N̂xs 
Bxb = B̂xb + c − (B̂sb − B̂rb ).
Cr +  Cs + 
()
Since Pax = P̂ax ,
!
() ûr  ûs 
pesc = p̂esc + c − (B̂sb − B̂rb )
Cr +  Cs + 

where ûx is the expected number of times that the ergodic chain P̂ ,
started at state a, is in state x before returning to a reaching b for
the first time. The absorption probability Bxa is the quantity vx intro-
duced in the previous section. As shown there, reversibility allows us
to conclude that
ûx B̂xa B̂xa
= =
Ĉx Ĉa Ca
so that
() c
pesc = pesc + (B̂sb − B̂rb )2
Ca
and this shows that the Monotonicity Law is true.
The change from P to P̂ was merely to make the calculations easier.
As we have remarked, the escape probabilities are the same for the two
chains as are the absorption probabilities Bib . Thus we can remove the
hats and write the same formula.
() c
pesc = pesc + (Bsb − Brb )2 .
Ca
The only quantity in this final expression that seems to depend upon
quantities from P̂ is c. In Exercise 1.4.11 you are asked to show that c
can also be expressed in terms of the fundamental matrix N obtained
from the original P.

Exercise 1.4.10 Let A be a matrix with inverse N = A−1 Let h be a

column vector and k a row vector. Show that

N̄ = (A − hk)−1

64
exists if and only if kNh 6= 1 and, if so,
(Nh)(kN)
N̄ = N + .
1 − kNh
Exercise 1.4.11 Show that c can be expressed in terms of the funda-
mental matrix N of the original Markov chain P by
1
c= Nrr  Nsr  Nss  Nrs  .
1+ Cr
− Cr
+ Cs
− Cs

2 Random walks on infinite networks

2.1 Polya’s recurrence problem
2.1.1 Random walks on lattices
In 1921 George Polya [26] investigated random walks on certain infinite
graphs, or as he called them, “street networks”. The graphs he con-
sidered, which we will refer to as lattices, are illustrated in Figure 35.

Figure 35: ♣

To construct a d-dimensional lattice, we take as vertices those points

(x1 , . . . , xd ) of Rd all of whose coordinates are integers, and we join each
vertex by an undirected line segment to each of its 2d nearest neighbors.

65
These connecting segments, which represent the edges of our graph,
each have unit length and run parallel to one of the coordinate axes of
Rd . We will denote this d-dimensional lattice by Zd . We will denote
the origin (0, 0, . . . , 0) by 0.
Now let a point walk around at random on this lattice. As usual,
by walking at random we mean that, upon reaching any vertex of the
graph, the probability of choosing any one of the 2d edges leading out
1
of that vertex is 2d . We will call this random walk simple random walk
in d dimensions.
When d = 1, our lattice is just an infinite line divided into segments
of length one. We may think of the vertices of this graph as represent-
ing the fortune of a gambler betting on heads or tails in a fair coin
tossing game. Simple random walk in one dimension then represents
the vicissitudes of his or her fortune, either increasing or decreasing by
one unit after each round of the game.
When d = 2, our lattice looks like an infinite network of streets and
avenues, which is why we describe the random motion of the wandering
point as a “walk”.
When d = 3, the lattice looks like an infinite “jungle gym”, so per-
haps in this case we ought to talk about a “random climb”, but we will
not do so. It is worth noting that when d = 3, the wanderings of our
point could be regarded as an approximate representation of the ran-
dom path of a molecule diffusing in an infinite cubical crystal. Figure
36 shows a simulation of a simple random walk in three dimensions.

2.1.2 The question of recurrence

The question that Polya posed amounts to this: “Is the wandering
point certain to return to its starting point during the course of its
wanderings?” If so, we say that the walk is recurrent. If not, that is,
if there is a positive probability that the point will never return to its
starting point, then we say that the walk is transient.
If we denote the probability that the point never returns to its
starting point by pesc , then the chain is recurrent if pesc = 0, and
transient if pesc > 0.
We will call the problem of determining recurrence or transience of
a random walk the type problem.

66
Figure 36: ♣

67
2.1.3 Polya’s original question
The definition of recurrence that we have given differs from Polya’s
original definition. Polya defined a walk to be recurrent if, with prob-
ability one, it will pass through every single point of the lattice in the
course of its wanderings. In our definition, we require only that the
point return to its starting point. So we have to ask ourselves, “Can
the random walk be recurrent in our sense and fail to be recurrent in
Polya’s sense?”
The answer to this question is, “No, the two definitions of recurrence
are equivalent.” Why? Because if the point must return once to its
starting point, then it must return there again and again, and each
time it starts away from the origin, it has a certain non-zero probability
of hitting a specified target vertex before returning to the origin. And
anyone can get a bull’s-eye if he or she is allowed an infinite number of
darts, so eventually the point will hit the target vertex.

Exercise 2.1.1 Write out a rigorous version of the argument just given.

2.1.4 Polya’s Theorem: recurrence in the plane, transience

in space
In [26], Polya proved the following theorem:
Polya’s Theorem. Simple random walk on a d-dimensional lattice
is recurrent for d = 1, 2 and transient for d > 2.
The rest of this work will be devoted to trying to understand this
theorem. Our approach will be to exploit once more the connections be-
tween questions about a random walk on a graph and questions about
electric currents in a corresponding network of resistors. We hope that
this approach, by calling on methods that appeal to our physical intu-
ition, will leave us feeling that we know “why” the theorem is true.

Exercise 2.1.2 Show that Polya’s theorem implies that if two random
walkers start at 0 and wander independently, then in one and two
dimensions they will eventually meet again, but in three dimensions
there is positive probability that they won’t.

Exercise 2.1.3 Show that Polya’s theorem implies that a random walker
in three dimensions will eventually hit the line defined by x = 2, z = 0.

68
2.1.5 The escape probability as a limit of escape probabilities
for finite graphs
We can determine the type of an infinite lattice from properties of
bigger and bigger finite graphs that sit inside it. The simplest way to
go about this is to look at the lattice analog of balls (solid spheres) in
space. These are defined as follows: Let r be an integer—this will be the
radius of the ball. Let G(r) be the graph gotten from Zd by throwing out
vertices whose distance from the origin is > r. By “distance from the
origin” we mean here not the usual Euclidean distance, but the distance
“in the lattice”; that is, the length of the shortest path along the edges
of the lattice between the two points. Let ∂G(r) be the “sphere” of
radius r about the origin, i.e., those points that are exactly r units
from the origin. In two dimensions, ∂G(r) looks like a square. (See
Figure 37.) In three dimensions, it looks like an octahedron.

Figure 37: ♣

We define a random walk on G(r) as follows: The walk starts at

0 and continues as it would on Zd until it reaches a point on ∂G(r)
and then it stays at this point. Thus the walk on G(r) is an absorbing
Markov chain with every point of ∂G(r) an absorbing state.
(r)
Let pesc be the probability that a random walk on G(r) starting at 0,
(r)
reaches ∂G(r) before returning to 0. Then pesc decreases as r increases

69
 (r)
and pesc = limr→∞ pesc is the escape probability for the infinite graph.
If this limit is 0, the infinite walk is recurrent. If it is greater than 0,
the walk is transient.

2.1.6 Electrical formulation of the type problem

Now that we have expressed things in terms of finite graphs, we can
make use of electrical methods. To determine pesc electrically, we sim-
ply ground all the points of ∂G(r) , maintain 0 at one volt, and measure
the current i(r) flowing into the circuit. (See Figure 38.)

Figure 38: ♣

From Section 1.3.4, we have

(r) i(r)
pesc = ,
2d
where d is the dimension of the lattice. (Remember that we have to
divide by the number of branches coming out of the starting point.)
Since the voltage being applied is 1, i(r) is just the effective conductance

70
between 0 and ∂G(r) , i.e.,
1
i(r) = (r)
.
Reff
(r)
where Reff is the effective resistance from 0 to ∂G(r) . Thus
(r) 1
pesc = (r)
.
2dReff
Define Reff , the effective resistance from the origin to infinity, to be
(r)
Reff = lim Reff .
r→∞
(r)
This limit exists since Reff is an increasing function of r. Then
1
pesc = .
2dReff
Of course Reff may be infinite; in fact, this will be the case if and only
if pesc = 0. Thus the walk is recurrent if and only if the resistance to
infinity is infinite, which makes sense.

Figure 39: ♣

The success of this electrical formulation of the type problem will

come from the fact that the resistance to infinity can be estimated using
classical methods of electrical theory.

71
2.1.7 One Dimension is easy, but what about higher dimen-
sions?
We now know that simple random walk on a graph is recurrent if and
only if a corresponding network of 1-ohm resistors has infinite resistance
“out to infinity”. Since an infinite line of resistors obviously has infinite
resistance, it follows that simple random walk on the 1-dimensional
lattice is recurrent, as stated by Polya’s theorem.

Figure 40: ♣

What happens in higher dimensions? We are asked to decide whether

a d-dimensional lattice has infinite resistance to infinity. The difficulty
is that the d-dimensional lattice Zd lacks the rotational symmetry of
the Euclidean space Rd in which it sits.
To see how this lack of symmetry complicates electrical problems,
we determine, by solving the appropriate discrete Dirichlet problem,
the voltages for a one-volt battery attached between 0 and the points
of ∂G(3) in Z2 . The resulting voltages are:

0
0 .091 0
0 .182 .364 .182 0
0 .091 .364 1 .364 .091 0 .
0 .182 .364 .182 0
0 .091 0
0

The voltages at points of ∂G(1) are equal, but the voltages at points of
∂G(2) are not. This means that the resistance from 0 to ∂G(3) cannot
be written simply as the sum of the resistances from 0 to ∂G(1) , ∂G(1)
to ∂G(2) , and ∂G(2) to ∂G(3) . This is in marked contrast to the case of
a continuous resistive medium to be discussed in Section 2.1.8.
(3) (3)
Exercise 2.1.4 Using the voltages given for G(3) , find Reff and pesc .

72
Exercise 2.1.5 Consider a one-dimensional infinite network with re-
sistors Rn = 1/2n from n to n + 1 for n = . . . , −2, −1, 0, 1, 2, . . .. De-
scribe the associated random walk and determine whether the walk is
recurrent or transient.

Exercise 2.1.6 A random walk moves on the non-negative integers;

when it is in state n, n > 0, it moves with probability pn to n + 1 and
with probability 1 − pn , to n − 1. When at 0, it moves to 1. Determine
a network that gives this random walk and give a criterion in terms of
the pn for recurrence of the random walk.

2.1.8 Getting around the lack of rotational symmetry of the

lattice
Suppose we replace our d-dimensional resistor lattice by a (homoge-
neous, isotropic) resistive medium filling all of Rd and ask for the ef-
fective resistance to infinity. Naturally we expect that the rotational
symmetry will make this continuous problem easier to solve than the
original discrete problem. If we took this problem to a physicist, he or
she would probably produce something like the scribblings illustrated
in Figure 41, and conclude that the effective resistance is infinite for
d = 1, 2 and finite for d > 2. The analogy to Polya’s theorem is ob-
vious, but is it possible to translate these calculations for continuous
media into information about what happens in the lattice?
This can indeed be done, and this would certainly be the most
natural approach to take. We will come back to this approach at the
end of the work. For now, we will take a different approach to getting
around the asymmetry of the lattice. Our method will be to modify the
lattice in such a way as to obtain a graph that is symmetrical enough
so that we can calculate its resistance out to infinity. Of course, we will
have to think carefully about what happens to that resistance when we
make these modifications.

2.1.9 Rayleigh: shorting shows recurrence in the plane, cut-

ting shows transience in space
Here is a sketch of the method we will use to prove Polya’s theorem.
To take care of the case d = 2, we will modify the 2-dimensional
resistor network by shorting certain sets of nodes together so as to get a
new network whose resistance is readily seen to be infinite. As shorting
can only decrease the effective resistance of the network, the resistance

73
 Figure 41: ♣

of the original network must also be infinite. Thus the walk is recurrent
when d = 2.
To take care of the case d = 3, we will modify the 3-dimensional
network by cutting out certain of the resistors so as to get a new network
whose resistance is readily seen to be finite. As cutting can only increase
the resistance of the network, the resistance of the original network must
also be finite. Thus the walk is transient when d = 3.
The method of applying shorting and cutting to get lower and up-
per bounds for the resistance of a resistive medium was introduced
by Lord Rayleigh. (Rayleigh [29]; see also Maxwell [21], Jeans [11],
PoIya and Szego [28]). We will refer to Rayleigh’s techniques collec-
tively as Rayleigh’s short-cut method. This does not do Rayleigh justice,
for Rayleigh’s method is a whole bag of tricks that goes beyond mere
shorting and cutting—but who can resist a pun?
Rayleigh’s method was apparently first applied to random walks
by C. St. J. A. Nash-Williams [24], who used the shorting method to
establish recurrence for random walks on the 2-dimensional lattice.

74
2.2 Rayleigh’s short-cut method
2.2.1 Shorting and cutting
In its simplest form, Rayleigh’s method involves modifying the network
whose resistance we are interested in so as to get a simpler network. We
consider two kinds of modifications, shorting and cutting. Cutting in-
volves nothing more than clipping some of the branches of the network,
or what is the same, simply deleting them from the network. Shorting
involves connecting a given set of nodes together with perfectly con-
ducting wires, so that current can pass freely between them. In the
resulting network, the nodes that were shorted together behave as if
they were a single node.

Figure 42: ♣

2.2.2 The Shorting Law and the Cutting Law; Rayleigh’s

idea
The usefulness of these two procedures (shorting and cutting) stems
from the following observations:
Shorting Law. Shorting certain sets of nodes together can only
decrease the effective resistance of the network between two given nodes.

75
 Cutting Law. Cutting certain branches can only increase the
effective resistance between two given nodes.
These laws are both equivalent to Rayleigh’s Monotonicity Law,
which was introduced in Section 1.4.1 (see Exercise 2.2.1):
Monotonicity Law. The effective resistance between two given
nodes is monotonic in the branch resistances.
Rayleigh’s idea was to use the Shorting Law and the Cutting Law
above to get lower and upper bounds for the resistance of a network.
In Section 2.2.3 we apply this method to solve the recurrence problem
for simple random walk in dimensions 2 and 3.

Exercise 2.2.1 Show that the Shorting Law and the Cutting Law are
both equivalent to the Monotonicity Law.

2.2.3 The plane is easy

When d = 2, we apply the Shorting Law as follows: Short together
nodes on squares about the origin, as shown in Figure 43. The network

Figure 43: ♣

we obtain is equivalent to the network shown in Figure 44.

Now as n 1-ohm resistors in parallel are equivalent to a single resistor
of resistance n1 ohms, the modified network is equivalent to the network

76
 Figure 44: ♣

shown in Figure 45. The resistance of this network out to infinity is

Figure 45: ♣

∞
X 1
= ∞.
n=0 8n + 4

As the resistance of the old network can only be bigger, we conclude

that it too must be infinite, so that the walk is recurrent when d = 2.

Exercise 2.2.2 Using the shorting technique, give an upper bound for
(3)
pesc , and compare this with the exact value obtained in Exercise 2.1.4.

2.2.4 Space: searching for a residual network

When d = 3, what we want to do is delete certain of the branches of
the network so as to leave behind a residual network having manifestly
finite resistance. The problem is to reconcile the “manifestly” with the

77
“finite”. We want to cut out enough edges so that the effective resis-
tance of what is left is easy to calculate, while leaving behind enough
edges so that the result of the calculation is finite.

2.2.5 Trees are easy to analyze

Trees—that is, graphs without circuits—are undoubtedly the easiest to
work with. For instance, consider the full binary tree, shown in Figure
46. Notice that sitting inside this tree just above the root are two

Figure 46: ♣

copies of the tree itself. This self-similarity property can be used to

compute the effective resistance R∞ from the root out to infinity. (See
Exercise 2.2.3.) It turns out that R∞ = 1. We will demonstrate this
below by a more direct method.
To begin with, let us determine the effective resistance Rn between
the root and the set of nth generation branch points. To do this, we
should ground the set of branch points, hook the root up to a 1-volt
battery, and compute
1
Rn = .
current through battery
For n = 3, the circuit that we would obtain is shown in Figure 48.

78
Figure 47: ♣

Figure 48: ♣

79
 In the resulting circuit, all branch points of the same generation are
at the same voltage (by symmetry). Nothing happens when you short
together nodes that are at the same potential. Thus shorting together
branch points of the same generation will not affect the distribution of
currents in the branches. In particular, this modification will not affect
the current through the battery, and we conclude that
1 1
Rn = = .
current in original circuit current in modified circuit
For n = 3, the modified circuit is shown in Figure 49. This picture

Figure 49: ♣

shows that
1 1 1 1
R3 = + + = 1 − 3.
2 4 8 2
More generally,
1 1 1 1
Rn = + + . . . + n = 1 − n.
2 4 2 2
Letting n → ∞, we get
1
R∞ = lim Rn = lim 1 − = 1.
n→∞ n→∞ 2n
Figure 50 shows another closely related tree, the tree homogeneous
of degree three: Note that all nodes of this tree are similar—there is no
intrinsic way to distinguish one from another. This tree is obviously
a close relative of the full binary tree. Its resistance to infinity is 2/3.

80
 Figure 50: ♣

Exercise 2.2.3 (a) Show, using the self-similarity of the full binary
tree, that the resistance R∞ to infinity satisfies the equation
R∞ + 1
R∞ =
2
and conclude that either R∞ = 1 or R∞ = ∞.
(b) Again using the self-similarity of the tree, show that

Rn + 1
Rn+1 =
2
where Rn denotes the resistance out to the set of the nth-generation
branch points. Conclude that

R∞ = lim Rn = 1.
n→∞

2.2.6 The full binary tree is too big

Nothing could be nicer than the two trees we have just described. They
are the prototypes of networks having manifestly finite resistance to
infinity. Unfortunately, we can’t even come close to finding either of
these trees as a subgraph of the three-dimensional lattice. For in these

81
trees, the number of nodes in a “ball” of radius r grows exponentially
with r, whereas in a d-dimensional lattice, it grows like r, i.e., much
slower. (See Figure 51.) There is simply no room for these trees in
any finite-dimensional lattice.

2.2.7 NT3 : a “three-dimensional” tree

These observations suggest that we would do well to look for a nice tree
NT3 where the number of nodes within a radius r of the root is on the
order of r 3 . For we might hope to find something resembling NT3 in
the 3-dimensional lattice, and if there is any justice in the world, this
tree would have finite resistance to infinity, and we would be done.
Before introducing NT3 , let’s have a look at NT2 , our choice for
the tree most likely to succeed in the 2-dimensional lattice (see Figure
52). The idea behind NT2 is that, since a ball of radius r in the graph
ought to contain something like r 2 points, a sphere of radius r ought
to contain something like r points, so the number of points in a sphere
should roughly double when the radius of the sphere is doubled. For
this reason, we make the branches of our tree split in two every time
the distance from the origin is (roughly) doubled.
Similarly, in a 3-dimensional tree, when we double the radius, the
size of a sphere should roughly quadruple. Thus in NT3 , we make the
branches of our tree split in four where the branches of NT2 would have
split in two. NT3 is shown in Figure 53. Obviously, NT3 is none too
happy about being drawn in the plane.

2.2.8 NT3 has finite resistance

To see if we’re on the right track, let’s work out the resistance of our
new trees. These calculations are shown in Figures 54 and 55.
As we would hope, the resistance of NT2 is infinite, but the resis-
tance of NT3 is not.

Exercise 2.2.4 Use self-similarity arguments to compute the resis-

tance of NT2 and NT3 .

2.2.9 But does NT3 fit in the three-dimensional lattice?

We would like to embed NT3 in Z3 . We start by trying to embed NT2
in Z2 . The result is shown in Figure 56.

82
Figure 51: ♣

83
Figure 52: ♣

84
Figure 53: ♣

85
Figure 54: ♣

86
Figure 55: ♣

87
Figure 56: ♣

88
 To construct this picture, we start from the origin and draw 2 rays,
one going north, one going east. Whenever a ray intersects the line
x + y = 2n − 1 for some n, it splits into 2 rays, one going north, and one
going east. The sequence of pictures in Figure 57 shows successively
larger portions of the graph, along with the corresponding portions of
NT2 .

Figure 57: ♣

Of course this isn’t really an embedding, since certain pairs of points

that were distinct in NT2 get identified, that is, they are made to
correspond to a single point in the picture. In terms of our description,
sometimes a ray going north and a ray going east pass through each

89
other. This could have been avoided by allowing the rays to “bounce”
instead of passing through each other, at the expense of embedding
not NT2 but a close relative—see Exercise 2.2.7. However, because the
points of each identified pair are at the same distance from the root of
NT2 , when we put a battery between the root and the nth level they
will be at the same potential. Hence, the current flow is not affected
by these identifications, so the identifications have no effect on Reff .
For our purposes, then, we have done just as well as if we had actually
embedded NT2 .
To construct the analogous picture in three dimensions, we start
three rays off from the origin going north, east, and up. Whenever a
ray intersects the plane x + y + z = 2n − 1 for some n, it splits into three
rays, going north, east, and up. This process is illustrated in Figure
58. Surprisingly, the subgraph of the 3-dimensional lattice obtained
in this way is not NT3 ! Rather, it represents an attempt to embed
the tree shown in Figure 59. We call this tree NT2.5849... because it
is 2.5849 . . .-dimensional in the sense that when you double the radius
of a ball, the number of points in the ball gets multiplied roughly by 6
and
6 = 2log2 6 = 22.5849... .
Again, certain pairs of points of NT2.5849... have been allowed to corre-
spond to the same point in the lattice, but once again the intersections
have no effect on Reff .
So we haven’t come up with our embedded NT3 yet. But why
bother? The resistance of NT2.5849... out to infinity is
1 2 4 1 2 2 1 1
 
+ + +... = 1 + + ( )2 + . . . = 2 = 1.
3 9 27 3 3 3 31− 3

Thus we have found an infinite subgraph of the 3-dimensional lattice

having finite resistance out to infinity, and we are done.

Exercise 2.2.5 This exercise deals with the escape probability pesc
for simple random walk in 3 dimensions. The idea is to turn upper
and lower bounds for the resistance of the lattice into bounds for pesc .
Bounds are the best we can ask for using our method. The determina-
tion of the exact value will be discussed in Section 2.3.5. It is roughly
.66.
(a) Use a shorting argument to get an upper bound for pesc .
(b) We have seen that the resistance of the 3-dimensional lattice is
at most one ohm. Show that the corresponding lower bound for pesc is

90
91
Figure 59: ♣

92
1/6 Show that this lower bound can be increased to 1/3 with no extra
effort.

Exercise 2.2.6 Prove that simple random walk in any dimension d >
3 is transient.

Exercise 2.2.7 Show how the not-quite embeddings of NT2 and NT2.5849...
can be altered to yield honest-to-goodness embeddings of “stretched-
out” versions of these trees, obtained by replacing each edge of the tree
by three edges in series. (Hint: “bounce”.)

2.2.10 What we have done; what we will do

We have finally finished our electrical proof of Polya’s theorem. The
proof of recurrence for d = 1, 2 was straight-forward, but this could
hardly be said of the proof for d = 3. After all, we started out trying
to embed NT3 and ended up by not quite embedding something that
was not quite NT3 !
This is not bad in itself, for one frequently sets out to do something
and in the process of trying to do it gets a better idea. The real prob-
lem is that this explicit construction is just too clever, too artificial.
We seem to be saying that a simple random walk in 3 dimensions is
transient because it happens to contain a beautifully symmetrical sub-
graph that is in some sense 2.5849 . . .-dimensional! Fine, but what if
we hadn’t stumbled upon this subgraph? Isn’t there some other, more
natural way?
We will see that indeed there are more natural approaches to show-
ing transience for d = 3. One such approach uses the same idea of em-
bedding trees, but depends on the observation that one doesn’t need to
be too careful about sending edges to edges. Another approach, based
on relating the lattice not to a tree but to Euclidean space, was already
hinted at in Section 2.1.8. The main goal for the rest of this work
will be to explore these more natural electrical approaches to Polya’s
theorem.
Before jumping into this, however, we are going to go back and take
a look at a classical—i.e., probabilistic—approach to Polya’s theorem.
This will give us something to compare our electrical proofs with.

93
2.3 The classical proofs of Polya’s Theorem
2.3.1 Recurrence is equivalent to an infinite expected num-
ber of returns
For the time being, all of our random walks will be simple. Let u be
the probability that a random walker, starting at the origin, will return
to the origin. The probability that the walker will be there exactly k
times (counting the initial time) is uk (1−u). Thus, if m is the expected
number of times at the origin,
1
∞kuk−1 (1 − u) =
X
m= .
k=1
1−u

If m = ∞ then u = 1, and hence the walk is recurrent. If m < ∞ then

u < 1, so the walk is transient. Thus m determines the type of the
walk.
We shall use an alternate expression for m. Let un be the probability
that the walk, starting at 0, is at 0 on the nth step. Since the walker
starts at 0, u0 = 1. Let en be a random variable that takes on the value
1 if, at time n the walker is at 0 and 0 otherwise. Then
∞
X
T = en
n=0

is the total number of times at 0 and

∞
X
m = E(T ) = E(en ).
n=0

But E(en ) = 1 · un + 0 · (1 − un ) = un . Thus

∞
X
m= un .
n=0

Therefore, the walk will be recurrent if the series diverges and

P∞
n=0
transient if it converges.

Exercise 2.3.1 Let Nxy be the expected number of visits to y for a

random walker starting in x. Show that Nxy is finite if and only if the
walk is transient.

94
2.3.2 Simple random walk in one dimension
Consider a random walker in one dimension, started at 0. To return
to 0, the walker must take the same number of steps to the right as
to the left; hence, only even times are possible. Let us compute u2n .
Any path that returns in 2n steps has probability 1/2n . The number
of possible paths equals the number of ways that we can choose the n
times to go right from the 2n possible times. Thus
!
2n 1
u2n = .
n 22n

We shall show that u2n = ∞ by using Stirling’s approximation:

P
n
√
n! ∼ 2πne−n nn .

Thus √
(2n!) 1 2π · 2ne−2n (2n)2n 1
u2n = 2n
∼ √ =√ .
n!n! 2 ( 2πne n ) 2
−n n 2 2n πn
Therefore,
X X 1
u2n ∼ √ =∞
n n πn
and a simple random walk in one dimension is recurrent.
Recall that this case was trivial by the resistor argument.

Exercise 2.3.2 We showed in Section 1.1.5 that a random walker

starting at x with 0 < x < N has probability x/N of reaching N
before 0. Use this to show that a simple random walk in one dimension
is recurrent.

Exercise 2.3.3 Consider a random walk in one dimension that moves

from n to n+1 with probability p and to n−1 with probability q = 1−p.
Assume that p > 1/2. Let hx be the probability, starting at x, that the
walker ever reaches 0. Use Exercise 1.1.9 to show that hx = (q/p)x for
x ≥ 0 and hx = 1 for x < 0. Show that this walk is transient.

Exercise 2.3.4 For a simple random walk in one dimension, it follows

from Exercise 1.1.7 that the expected time, for a walker starting at x
with 0 < x < N , to reach 0 or n is x(N − x). Prove that for the infinite
walk, the expected time to return to 0 is infinite.

95
Exercise 2.3.5 Let us regard a simple random walk in one dimension
as the fortune of a player in a penny matching game where the players
have unlimited credit. Show that the result is a martingale (see Section
1.1.6). Show that you can describe a stopping system that guarantees
that you make money.

2.3.3 Simple random walk in two dimensions

For a random walker in two dimensions to return to the origin, the
walker must have gone the same number of times north and south and
the same number of times east and west. Hence, again, only even
times for return are possible. Every path that returns in 2n steps has
probability 1/42n . The number of paths that do this by taking k steps
to the north, k south, n − k east and n − k west is
!
2n 2n!
= .
k, k, n − k, n − k k!k!(n − k)!(n − k)!
Thus
n
1 X (2n)!
u2n =
42n k=0 k!k!(n − k)!(n − k)!
n
1 X (2n)! n!n!
= 2n
4 k=0 n!n! k!k!(n − k)!(n − k)!
n
! !2
1 2n X n
= 2n .
4 n k=0 k
 2  
Pn n 2n
But k=0 k
= n
(see Exercise 2.3.6). Hence
!!2
1 2n
u2n = .
22n n
This is just the square of the one dimension result (not by accident,
either—see Section 2.3.6). Thus we have
X X 1
m= u2n ≈ = ∞.
n n πn
Recall that the resistor argument was very simple in this case also.
 2  
Exercise 2.3.6 Show that nk=0 nk = 2n (Hint: Think of choosing
P
n
n balls from a box that has 2n balls, n black and n white.)

96
2.3.4 Simple random walk in three dimensions
For a walker in three dimensions to return, the walker must take an
equal number of steps back and forth in each of the three different
directions. Thus we have
1 X (2n)!
u2n = 2n
6 j,k j!j!k!k!(n − j − k)!(n − j − k)!

where the sum is taken over all j, k with j + k ≤ n. Following Feller

[5], we rewrite this sum as
! !2
1 2n X 1 n!
u2n = 2n n
.
2 n j,k 3 j!k!(n − j − k)!

Now consider placing n balls randomly into three boxes, A, B, C. The

probability that j balls fall in A, k in B, and n − j − k in C is
!
1 n 1 n!
n
= n .
3 j, k, n − j − k 3 j!k!(n − j − k)!
Intuitively, this probability is largest when j, k, and n − j − k are as
near as possible to n/3, and this can be proved (see Exercise 2.3.7).
Hence, replacing one of the factors in the square by this larger value,
we have:
!  
1 2n  1 n! X 1 n!
u2n ≤ 2n n
j k j k j k 
n
,
2 n 3 n ! n ! n ! j,k 3 j!k!(n − j − k)!
3 3 3

where bn/3c denotes the greatest integer ≤ n/3. The last sum is 1 since
it is the sum of all probabilities for the outcomes of putting n balls into
three boxes. Thus
! 
1 2n  1 n! 
u2n ≤ j k .
22n n 3 n n !3
3

Applying Stirling’s approximation yields

K
u2n ≤
n3/2
for suitable constant K. Therefore
X X 1
m= u2n ≤ K < ∞,
n n n3/2

97
and a simple random walk in three dimensions is recurrent.
While this is a complex calculation, the resistor argument was also
complicated in this case. We will try to make amends for this presently.
 
n
Exercise 2.3.7 Prove that j,k,n−j−k is largest when j, k, and n−j −k
are as close as possible to n/3.

Exercise 2.3.8 Find an appropriate value for the ”suitable constant”

K that was mentioned above, and derive an upper bound for m. Use
this to get a lower bound for the probability of escape for simple random
walk in three dimensions.

2.3.5 The probability of return in three dimensions: exact

calculations
Since the probability of return in three dimensions is less than one, it is
natural to ask, “What is this probability?” For this we need an exact
calculation. The first such calculation was carried out in a pioneering
paper of McCrea and Whipple [22]. The solution outlined here follows
Feller [5], Exercise 28, Chapter 9, and Montroll and West [23].
Let p(a, b, c; n) be the probability that a random walker, starting at
0, is at (a, b, c) after n steps. Then p(a, b, c; n) is completely determined
by the fact that
p(0, 0, 0; 0) = 1
and
1 1
p(a, b, c; n) = p(a − 1, b, c; n − 1) + p(a + 1, b, c; n − 1) +
6 6
1 1
p(a, b − 1, c; n − 1) + p(a, b + 1, c; n − 1) +
6 6
1 1
p(a, b, c − 1; n − 1) + p(a, b, c + 1; n − 1).
6 6
Using the technique of generating functions, it is possible to derive
a solution of these equations as

p(a, b, c; n)
1
= ·
(2π)3
n
π π π cos x + cos y + cos z
Z Z Z 
cos(xa) cos(yb) cos(zc)dxdydz.
−π −π −π 3

98
Of course, we can just verify that this formula satisfies our equations
once someone has suggested it. Having this formula, we put a = b =
c = 0 and sum over n to obtain the expected number of returns as
3 π π π 1
Z Z Z
m= dxdydz.
(2π)3 −π −π −π 3 − (cos x + cos y + cos z)
This integral was first evaluated by Watson [35] in terms of elliptic
integrals, which are tabulated. A simpler result was obtained by Glasser
and Zucker [8] who evaluated this integral in terms of gamma functions.
Using this result, we get
√
6 1 5 7 11
       
m= 3
Γ Γ Γ Γ = 1.516386059137 . . . ,
32π 24 24 24 24
where Z ∞
Γ(x) = e−t tx−1 dt
0

is Euler’s gamma function. (Incidentally, the value given by Glasser and

Zucker [8] for the integral above needs to be corrected by multiplying
by a factor of 1/(384π).)
Recall that m = 1/(1 − u) so that u = (m − 1)/m. This gives

u = .340537329544 . . . .

2.3.6 Simple random walk in two dimensions is the same as

two independent one-dimensional random walks
We observed that the probability of return at time 2n in two dimensions
is the square of the corresponding probability in one dimension. Thus
it is the same as the probability that two independent walkers, one
walking in the x direction and the other in the y direction, will, at time
2n, both be at 0. Can we see that this should be the case? The answer
is yes. Just change our axes by 45 degrees to new axes x̄ and ȳ as in
Figure 60.
Look at the possible outcomes for the first step using the x, y coor-
dinates and the x̄, ȳ coordinates. We have

x, y coordinates x̄, ȳ√coordinates

√
(0, 1) (1/ √ 2, 1/ 2)√
(0, −1) (−1/ √ 2, −1/√ 2) .
(1, 0) (1/ √ 2, −1/√2)
(−1, 0) (−1/ 2, 1/ 2)

99
 Figure 60: ♣

Assume that we have two independent walkers, one moving with step
size √12 randomly along the x̄ axis and the other moving with the same
step size along the ȳ axis. Then, if we plot their positions using the
x, y axes, the four possible outcomes for the first step would agree with
those given in the second column of the table above. The probabilities
for each of the four pairs of outcomes would also be (1/2) · (1/2) = 1/4.
Thus, we cannot distinguish a simple random walk in two dimensions
from two independent
√ walkers along the x̄ and ȳ axes making steps of
magnitude 1/ 2. Since the probability of return does not depend upon
the magnitude of the steps, the probability that our two independent
walkers are at (0, 0) at time 2n is equal to the product of the prob- 
abilities that each separately is at 0 at time 2n, namely (1/22n ) 2n n
.
Therefore, the probability
  that the standard walk will be at (0, 0) at
2n 2n 2
time 2n is ((1/2 ) n ) as observed earlier.

2.3.7 Simple random walk in three dimensions is not the

same as three independent random walks
In three dimensions, the probability that three independent walkers are
each back to 0 after time 2n is
! !3
2n 1
u2n = .
n 22n

100
This does not agree with our result for a simple random walk in three
dimensions. Hence, the same trick cannot work. However, it is interest-
ing to consider a random walk which is the result of three independent
walkers. Let (i, j, k) be the position of three independent random walk-
ers. The next position is one of the eight possibilities (i ± 1, j ± 1, k ± 1)
Thus we may regard their progress as a random walk on the lattice
points in three dimensions. If we center a cube of side 2 at (i, j, k),
then the walk moves with equal probabilities to each of the eight cor-
ners of the cube. It is easier to show that this random walk is transient
(using classical methods) than it is for simple random walk. This is be-
cause we can again use the one-dimension calculation. The probability
u2n for return at time 2n is
! !3 !3
2n 1 1
u2n = ∼ √ .
n 22n πn
Thus 3/2
X X 1
m= u2n ∼ <∞
n n πn
and the walk is transient.
The fact that this three independent walkers model and simple ran-
dom walk are of the same type suggests that when two random walks
are “really about the same”, they should either both be transient or
both be recurrent. As we will soon see, this is indeed the case. Thus we
may infer the transience of simple random walk in 3 dimensions from
the transience of the three independent walkers model without going
through the involved calculation of Section 2.3.4.

2.4 Random walks on more general infinite net-

works
2.4.1 Random walks on infinite networks
From now on we assume that G is an infinite connected graph. We
assume that it is of bounded degree, by which we mean that there is
some integer E such that the number of edges from any point is at
most E. We assign to each edge xy of G a conductance Cxy > 0 with
Rxy = C1xy . The graph G together with the conductances C = (Cxy ) is
called a network and denoted by (G, C). Given a network (G, C), we
define a random walk by
Cxy
Pxy =
Cx

101
where Cx = y Cxy . When all the conductances are equal, we obtain
P

a random walk that moves along each edge with the same probability:
In agreement with our previous terminology, we call this walk simple
random walk on G.
We have now a quite general class of infinite-state Markov chains.
As in the case of finite networks, the chains are reversible Markov
chains: That is, there is a positive vector w such that wx Pxy = wy Pyx .
As in the finite case, we can take wx = Cx , since Cx Pxy = Cxy = Cyx =
Cy Pyx .

2.4.2 The type problem

Let (G, C) be an infinite network with random walk P. Let 0 be a
reference point. Let pesc be the probability that a walk starting at 0
will never return to 0. If pesc = 0 we say that P is recurrent, and if
pesc > 0 we say that it is transient. You are asked to show in Exercise
2.4.1 that the question of recurrence or transience of P does not depend
upon the choice of the reference point. The type problem is the problem
of determining if a random walk (network) is recurrent or transient.
In Section 2.1.5 we showed how to rephrase the type problem for
a lattice in terms of finite graphs sitting inside it. In Section 2.1.6 we
showed that the type problem is equivalent to an electrical network
problem by showing that simple random walk on a lattice is recurrent
if and only if the lattice has infinite resistance to infinity. The same
arguments apply with only minor modifications to the more general
infinite networks as well. This means that we can use Rayleigh’s short-
cut method to determine the type of these more general networks.

Exercise 2.4.1 Show that the question of recurrence or transience of

P does not depend upon the choice of the reference point.

2.4.3 Comparing two networks

Given two sets of conductances C and C̄ on G, we say that (G, C̄) <
(G, C) if C̄xy < Cxy for all xy, or equivalently, if R̄xy > Rxy for all
xy. Assume that (G, C̄) < (G, C). Then by the Monotonicity Law,
R̄eff ≥ Reff . Thus if random walk on (G, C̄) is transient, i.e., if R̄eff <
∞, then random walk on (G, C) is also transient. If random walk on
(G, C) is recurrent, i.e., if Reff = ∞, then random walk on (G, C̄) is
also recurrent.

102
 Theorem. If (G, C) and (G, C̄) are networks, and if there exist
constants u, v with 0 < u ≤ v < ∞ such that

uCxy ≤ C̄xy ≤ vCxy

for all x and y, then random walk on (G, C̄) is of the same type as
random walk on (G, C).
Proof. Let Uxy = uCxy and Vxy = vCxy . Then (G, U) ≤ (G, C̄) ≤
(G, V). But the random walks for (G, U) and (G, V) are the same as
random walk on (G, C). Thus random walk for (G, C̄) is of the same
type as random walk on (G, C). ♦
Corollary. Let (G, C) be a network. If for every edge xy of G
we have 0 < u < Cxy < v < ∞ for some constants u and v, then the
random walk on (G, C) has the same type as simple random walk on
G.

Exercise 2.4.2 Consider the two-dimensional lattice. For each edge,

we toss a coin to decide what kind of resistor to put across this edge.
If heads turns up, we put a two-ohm resistor across this edge; if tails
turns up, we put a one-ohm resistor across the edge. Show that the
random walk on the resulting network is recurrent.

Exercise 2.4.3 Consider the analogous problem to Exercise 2.4.2 in 3

dimensions.

2.4.4 The k-fuzz of a graph

For any integer k, the k-fuzz of a graph G is the graph Gk obtained
from G by adding an edge xy if it is possible to go from x to y in at
most k steps. For example, the 2-fuzz of the two-dimensional lattice is
shown in Figure 61; please note that horizontal and vertical edges of
length 2, such as those joining (0, 0) to (0, 2), have not been indicated.

Theorem. Simple random walk on G and on the k-fuzz Gk of G

have the same type.
Proof. Let P be simple random walk on G. Define P̄ = (P + P2 +
. . . + Pk )/k. Then P̄ may be considered to be P, watched at one of the
first k steps chosen at random, then at a time chosen at random from
the next k steps after this time, etc. Thinking of P̄ in this way, we see
that P is in state 0 at least once for every time P̄ in state 0. Hence, if
P̄ is recurrent so is P. Assume now that P̄ is transient. Choose a finite

103
 Figure 61: ♣

set S so that 0 cannot be reached in k steps from a point outside of S.

Then, since the walk P̄ will be outside S from some time on, the walk
P cannot be at 0 after this time, and P is also transient. Therefore, P
and P̄ are of the same type.
Finally, we show that P̄ has the same type as simple random walk
on Gk . Here it is important to remember our restriction that G is of
bounded degree, so that for some E no vertex has degree > E. We
know that P is reversible with wP = w, where wx is the number of
edges coming out of x. From its construction, P̄ is also reversible and
wP̄ = w. P̄ is the random walk on a network (Gk , C̄) with C̄xy =
wx P̄xy . If P̄xy > 0, there is a path x, x1 , x2 , . . . , xm−1 , y in G from x to
y of length m ≤ k. Then
1 1 1 1
P̄xy ≥ ( )m ≥ ( )k .
k E k E
Thus
1 1
0 < ( )k ≤ P̄xy ≤ 1
k E
and
1 1
0 < ( )k ≤ C̄xy ≤ E.
k E
Therefore, by the theorem on the irrelevance of bounded twiddling
proven in Section 2.4.3, P̄ and simple random walk on Gk are of the
same type. So G and Gk are of the same type.
NOTE: This is the only place where we use probabilistic methods
of proof. For the purist who wishes to avoid probabilistic methods,
Exercise 2.4.9 indicates an alternative electrical proof.

104
 We show how this theorem can be used. We say that a graph G
can be embedded in a graph G if the points x of G can be made to
correspond in a one-to-one fashion to points x̄ of G in such a way that
if xy is an edge in G, then x̄ȳ is an edge in G.
Theorem. If simple random walk on G is transient, and if G can
be embedded in a k-fuzz Gk of G then simple random walk on G is
also transient. Simple random walk on G and G are of the same type
if each graph can be embedded in a k-fuzz of the other graph.
Proof. Assume that simple random walk on G is transient and that
G can be embedded in a k-fuzz Gk of G. Since Reff for G is finite and G
can be embedded in Gk , Reff for Gk is finite. By our previous theorem,
the same is true for G and simple random walk on G is transient.
If we can embed G in Gk and G in Gk , then the random walk on G
is transient if and only if the random walk on G is. ♦

Exercise 2.4.4 We have assumed that there is a bound E for the

number of edges coming out of any point. Show that if we do not
assume this, it is not necessarily true that G and Gk are of the same
type. (Hint: Consider a network something like that shown in Figure
62.)

Figure 62: ♣

2.4.5 Comparing general graphs with lattice graphs

We know the type of simple random walk on a lattice Zd . Thus to
determine the type of simple random walk on an arbitrary graph G, it
is natural to try to compare G with Zd . This is feasible for graphs that
can be drawn in some Euclidean space Rd in a civilized manner.
Definition. A graph G can be drawn in a Euclidean space Rd in a
civilized manner if its vertices can be embedded in Rd so that for some
r < ∞, s > 0
(a) The length of each edge is ≤ r.
(b) The distance between any two points is > s.

105
 Note that we make no requirement about being able to draw the
edges of G so they don’t intersect.
Theorem. If a graph can be drawn in Rd in a civilized manner,
then it can be embedded in a k-fuzz of the lattice Zd .
Proof. We carry out the proof for the case d = 2. Assume that G
can be drawn in a civilized manner in R2 . We want to show that G
can be embedded in a k-fuzz of Z2 . We have been thinking of Z2 as
being drawn in R2 with perpendicular lines and adjacent points a unit
distance apart on these lines, but this embedding is only one particular
way of representing Z2 . To emphasize this, let’s talk about L2 instead
of Z2 . Figure 63 shows another way of drawing L2 in R2 . From a

Figure 63: ♣

graph-theoretical point of view, this is the same as Z2 . In trying to

compare G to L, we take advantage of this flexibility by drawing L2 so
small that points of G can be moved onto points of L2 without bumping
into each other.
Specifically, let L2 be a two-dimensional rectangular lattice with
lines a distance s/2 apart. In any square of L2 , there is at most one
point of G. Move each point x of G to the southwest corner x̄ of the
square that it is in, as illustrated in Figure 64.
Now since any two adjacent points x, y in G were within r of each
other in R2 , the corresponding points x̄, ȳ in L2 will have Euclidean
distance < r + 2s. Choose k so that any two points of L2 whose
Euclidean distance is < r + 2s can be connected by a path in L2 of
at most k steps. Then x̄ and ȳ will be adjacent in L2k and—since the

106
 Figure 64: ♣

prescription for k does not depend on x and y—we have embedded G

in the k-fuzz L2k .
Corollary. If G can be drawn in a civilized manner in R1 or R2 ,
then simple random walk on G is recurrent.
Proof. Assume, for example, that G can be drawn in a civilized
manner in R2 . Then G can be embedded in a k-fuzz Z2k of Z2 . If simple
random walk on G were transient, then the same would be true for Z2k
and Z2 . But we know that simple random walk on Z2 is recurrent.
Thus simple random walk on G is recurrent. ♦
Our first proof that random walk in three dimensions is transient
consisted in showing that we could embed a transient tree in Z3 . We
now know that it would have been sufficient to show how to draw a
transient tree in R3 in a civilized manner: This is easier (see Exercise
2.4.5).
The corollary implies that simple random walk on any sufficiently
symmetrical graph in R2 is recurrent. For example, simple random
walk on the regular graph made up of hexagons shown in Figure 65 is
recurrent.
We can even consider very irregular graphs. For example, on the
cover of the January 1977 Scientific American, there is an example due
to Conway of an infinite non-periodic tiling using Penrose tiles of the

107
 Figure 65: ♣

form shown in Figure 66. It is called the cartwheel pattern; part of

Figure 66: ♣

it is shown in Figure 67. A walker walking randomly on the edges of

this very irregular infinite tiling will still return to his or her starting
point.
Assume now that G can be drawn in a civilized manner in R3 . Then
to show that simple random walk on G is of the same type as Z3 , namely
transient, it is sufficient to show that we can embed Z3 in a k-fuzz of
G. This is clearly possible for any regular lattice in R3 . The three
lattices that have been most studied and for which exact probabilities
for return have been found are called the SC, BCC, and FCC lattices.

108
 Figure 67: ♣

The SC (simple cubic) lattice is just Z3 . The walker moves each time
to a new point by adding a random choice from the six vectors

(±1, 0, 0), (0, ±1, 0), (0, 0, ±1).

For the BCC (body-centered cubic) lattice, the choice is one of the eight
vectors
(±1, ±1, ±1).
This was the walk that resulted from three independent one-dimensional
walkers. For the FCC (face-centered cubic) lattice, the random choice
is made from the twelve vectors

(±1, ±1, 0), (±1, 0, ±1), (0, ±1, ±1).

For a discussion of exact calculations for these three lattices, see Mon-
troll and West [23]
As we have seen, once the transience of any one of these three walks
is established, no calculations are necessary to determine that the other
walks are transient also. Thus we have yet another way of establishing
Polya’s theorem in three dimensions: Simply verify transience of the

109
walk on the BCC lattice via the simple three-independent-walkers com-
putation, and infer that walk on the SC lattice is also transient since
the BCC lattice can be embedded in a k-fuzz of it.

Exercise 2.4.5 When we first set out to prove Polya’s theorem for
d = 3, our idea was to embed NT3 in Z3 . As it turned out, what
we ended up embedding was not NT3 but NT2.5849... , and we didn’t
quite embed it at that. We tried to improve the situation by finding
(in Exercise 2.2.7) an honest-to-goodness embedding of a relative of
NT2.5849... , but NT3 was still left completely out in the cold. Now,
however, we are in a position to embed NT3 , if not in Z3 then at least
in a k-fuzz of it. All we need to do is to draw NT3 in R3 in a civilized
manner. Describe how to do this, and thereby give one more proof of
Polya’s theorem for d = 3.

Exercise 2.4.6 Find a graph that can be embedded in a civilized man-

ner in R3 but not in R2 , but is nonetheless recurrent.

Exercise 2.4.7 Assume that G is drawn in a civilized manner in R3 .

To show that simple random walk on G is transient, it is enough to know
that Z3 can be embedded in a k-fuzz of G. Try to come up with a nice
condition that will guarantee that this is possible. Can you make this
condition simple, yet general enough so that it will settle all reasonably
interesting cases? In other words, can you make the condition nice
enough to allow us to remember only the condition, and forget about
the general method lying behind it?

2.4.6 Solving the type problem by flows: a variant of the

cutting method
In this section we will introduce a variant of the cutting method whereby
we use Thomson’s Principle directly to estimate the effective resistance
of a conductor.
Thomson’s Principle says that, given any unit flow through a resis-
tive medium, the dissipation rate of that flow gives an upper bound for
the effective resistance of the medium. This suggests that to show that
a given infinite network is transient, it should be enough to produce a
unit flow out to infinity having finite energy dissipation.
In analogy with the finite case, we say that j is a flow from 0 to
infinity if

(a) jxy = jyx .

110
 (b) y jxy = 0 if x 6= 0.
P

We define j0 = y j0y . If j0 = 1, we say that j is a unit flow to

P

infinity. Again in analogy with the finite case, we call 21 x,y jxy
2
Rxy the
P

energy dissipation of the flow j.

Theorem. The effective resistance Reff from 0 to ∞ is less than
or equal to the energy dissipation of any unit flow from 0 to infinity.
Proof. Assume that we have a unit flow j from 0 to infinity with
energy dissipation
1X 2
E= j Rxy .
2 x,y xy
We claim that Reff ≤ E. Restricting jxy to the edges of the finite graph
G(r) , we have a unit flow from 0 to ∂G(r) in G(r) . Let i(r) be the unit
current flow in G(r) from 0 to ∂G(r) . By the results of Section 1.3.5,

(r) 1 X (r) 2 1X 2 1X 2
Reff = (ixy ) Rxy ≤ jxy Rxy ≤ j Rxy = E,
2 G(r) 2 G(r) 2 x,y xy

where indicates the sum over all pairs x, y such that xy is an edge
P
G(r)
of G(r) .

Exercise 2.4.8 We have billed the method of using Thomson’s Prin-

ciple directly to estimate the effective resistances of a network as a
variant of the cutting method. Since the cutting method was derived
from Thomson’s Principle, and not vice versa, it would seem that we
have got the cart before the horse. Set this straight by giving an infor-
mal (“heuristic”) derivation of Thomson’s Principle from the cutting
method. (Hint: see Maxwell [21], Chapter VIII, Paragraph 307.) For
more on this question, see Onsager [25].

Exercise 2.4.9 Let G be an infinite graph of bounded degree and Gk

the k-fuzz of G. Using electric network arguments, show that Reff < ∞
for G if and only if Reff < ∞ for Gk .

2.4.7 A proof, using flows, that simple random walk in three

dimensions is transient
We now apply this form of the cutting method to give another proof
that simple random walk on the threedimensional lattice is transient.
All we need is a flow to infinity with finite dissipation. The flow we are
going to describe is not the first flow one would think of. In case you are

111
curious, the flow described here was constructed as a side effect of an
unsuccessful attempt to derive the isoperimetric inequality (see Polya
[27]) from the “max-flow min-cut” theorem (Ford and Fulkerson [7]).
The idea is to find a flow in the positive orthant having the property
that the same amount flows through all points at the same distance
from 0.
Again, it is easiest to show the construction for the two-dimensional
case. Let G denote the part of Z2 lying in the first quadrant. The graph
G(4) is shown in Figure 68.

Figure 68: ♣

We choose our flow so that it always goes away from 0. Into each
point that is not on either axis there are two flows, one vertical and
one horizontal. We want the sum of the corresponding values of jxy to
be the same for all points the same distance from 0. These conditions
completely determine the flow. The flow out of the point (x, y) with
x + y = n is as shown in Figure 69. The values for the currents out to
the fourth level are shown in Figure 70. In general, the flow out of a
point (x, y) with x + y = n is
x+1 y+1 1
+ =
(n + 2)(n + 1) (n + 2)(n + 1) n+1

and the flow into this point is

x y 1
+ =
n(n + 1) n(n + 1) n+1

112
Figure 69: ♣

Figure 70: ♣

113
Thus the net flow at (x, y) is 0. The flow out of 0 is (1/2) + (1/2) = 1.
For this two-dimensional flow, the energy dissipation is infinite, as it
would have to be. For three dimensions, the uniform flow is defined as
follows: Out of (x, y, z) with x + y + z = n we have the flow indicated
in Figure 71. The total flow out of (x, y, z) is then

Figure 71: ♣

2(x + 1) 2(y + 1) 2(z + 1)

+ +
(n + 3)(n + 2)(n + 1) (n + 3)(n + 2)(n + 1) (n + 3)(n + 2)(n + 1)
2
= .
(n + 2)(n + 1)
The flow into (x, y, z) comes from the points (x − 1, y, z), (x, y − 1, z),
(x, y, z − 1) and, hence, the total flow into (x, y, z) is
2x 2y 2z 2
+ + = .
(n + 2)(n + 1)n (n + 2)(n + 1)n (n + 2)(n + 1)n (n + 2)(n + 1)
Thus the net flow for (x, y, z) is 0. The flow out of 0 is (1/3) + (1/3) +
(1/3) = 1 We have now to check finiteness of energy dissipation. The
flows coming out of the edges at the nth level are all ≤ 2/(n + 1)2 .
There are (n + 1)(n + 2)/2 points a distance n from 0, and thus there
are (3/2)(n + 1)(n + 2) ≤ 3(n + 1)2 edges coming out of the nth level.
Thus the energy dissipation E has
!2
X 2 2 X 1
E≤ 3(n + 1) = 12 < ∞,
n (n + 1)2 n (n + 1)2
and the random walk is transient.

114
2.4.8 The end
We have come to the end of our labors, and it seems fitting to look
back and try to say what it is we have learned.
To begin with, we have seen how phrasing certain mathematical
questions in physical terms allows us to draw on a large body of physical
lore, in the form of established methods and ways of thought, and
thereby often leads us to the answers to those questions.
In particular, we have seen the utility of considerations involving
energy. In took hundreds of years for the concept of energy to emerge
and take its rightful place in physical theory, but it is now recognized
as perhaps the most fundamental concept in all of physics. By phrasing
our probabilistic problems in physical terms, we were naturally led to
considerations of energy, and these considerations showed us the way
through the difficulties of our problems.
As for Polya’s theorem and the type problem in general, we have
picked up a bag of tricks, known collectively as “Rayleigh’s short-cut
method”, which we may expect will allow us to determine the type of
almost any random walk we are likely to embark on. In the process, we
have gotten some feeling for the connection between the dimensionality
of a random walk and its type. Furthermore, we have settled one of
the main questions likely to occur to someone encountering Polya’s
theorem, namely: “If two walks look essentially the same, and if one
has been shown to be transient, must not the other also be transient?”
Another question likely to occur to someone contemplating Polya’s
theorem is the question raised in Section 2.1.8: “Since the lattice Zd is
in some sense a discrete analog of a resistive medium filling all of Rd ,
should it not be possible to go quickly and naturally from the trivial
computation of the resistance to infinity of the continuous medium to
a proof of Polya’s theorem?” Our shorting argument allowed us to do
this in the two-dimensional case; that leaves the case of three (or more)
dimensions. Again, it is considerations of energy that allow us to make
this connection. The trick is to start with the flow field that one gets
by solving the continuous problem, and adapt it to the lattice, so as to
get a lattice flow to infinity having finite dissipation rate. We leave the
working out of this as an exercise, so as not to rob readers of the fun
of doing it for themselves.

Exercise 2.4.10 Give one final proof of Polya’s theorem in 3 dimen-

sions by showing how to adapt the 1/r 2 radial flow field to the lattice.
(Hint: “cubes”.)

115
References
[1] Edwin Abbott. Flatland. 1899.

[2] B. Bollobás. Graph Theory. 1979.

[3] R. Courant, K. Friedrichs, and H. Lewy. Über die pertiellen

Differenzengleichungen der mathematischen Physik. Math. Ann.,
100:32–74, 1928.

[4] J. L. Doob. Stochastic Processes. 1953.

[5] W. Feller. An Introduction to Probability Theory and Its Applica-

tions, volume I. 1968.

[6] R. P. Feynman. The Feynmann Lectures on Physics. 1964.

[7] L. R. Ford and D. R. Fulkerson. Maximal flow through a network.

Can. J. Math., 8:399–404, 1956.

[8] M. L. Glasser and I. J. Zucker. Extended Watson integrals for the

cubic lattice. Proc. Natl. Acad. Sci., USA, 74:1800–1801, 1977.

[9] D. Griffeath and T. M. Liggett. Critical phenomena for Spitzer’s

reversible nearest particle systems. Ann. Probab., 10:881–895,
1982.

[10] R. Hersh and R. J. Griego. Brownian motion and potential theory.

Scientific American, pages 66–74, March 1969.

[11] J. Jeans. The Mathematical Theory of Electricity and Magnetism,

5th Edition. 1966.

[12] S. Kakutani. Markov processes and the Dirichlet problem. Proc.

Jap. Acad., 21:227–233, 1945.

[13] F. Kelly. Reversibility and Stochastic Networks. 1979.

[14] J. G. Kemeny, J. L. Snell, and A. W. Knapp. Denumerable Markov

Chains. 1966.

[15] J. G. Kemeny, J. L. Snell, and G. L. Thompson. Finite Markov

Chains, 3rd Edition. 1974.

[16] Harry Kesten. Percolation Theory for Mathematicians. 1982.

116
[17] J. F. C. Kingman. Markov population processes. J. Appl. Prob.,
6:1–18, 1969.
[18] A. Lehman. A resistor network inequality, Problem 60-5. SIAM
Review, 4:150–154, 1965.
[19] Paul Lévy. Théorie de l’addition des variable aléatoires. 1937.
[20] T. J. Lyons. A simple criterion for transience of a reversible Markov
chain. Ann. Probab., 11:393–402, 1983.
[21] J. C. Maxwell. Treatise on Electricity and Magnetism, 3rd Edition.
1891.
[22] W. H. McCrea and F. J. W. Whipple. Random paths in two and
three dimensions. Proc. of the Royal Soc. of Edinburgh, 60:281–
298, 1940.
[23] E. W. Montroll and B. J. West. Fluctuation phenomena. In Studies
in Statistical Mathematics, volume 7, pages 61–175. 1979.
[24] C. St. J. A. Nash-Williams. Random walk and electric currents in
networks. Proc. Camb. Phil. Soc., 55:181–194, 1959.
[25] L. Onsager. Reciprocal relations in irreversible processes I. Phys.
Rev., 37:405–426, 1931.
[26] G. Polya. Über eine Aufgabe betreffend die Irrfahrt im Strassen-
netz. Math. Ann., 84:149–160, 1921.
[27] G. Polya. How to Solve It, 2nd Edition. 1957.
[28] G. Polya and G. Szego. Isoperimetric Inequalities of Mathematical
Physics. 1951.
[29] J. W. S. Rayleigh. On the theory of resonance. In Collected sci-
entific papers, volume 1, pages 33–75. 1899.
[30] H. L. Royden. Harmonic functions on open Riemann surfaces.
Trans. Amer. Math. Soc., 73:40–94, 1952.
[31] C. E. Shannon and D. W. Hagelbarger. Concavity of resistance
function. J. of Appl. Phys., 27:42–43, 1956.
[32] J. L. Snell. Probability and martingales. The Mathematical Intel-
ligencer, 4, 1982.

117
[33] W. Thomson and P. G. Tait. Treatise on Natural Philosophy. 1879.

[34] J. Ville. Étude critique de la notion de collectif. 1939.

[35] G. N. Watson. Three triple integrals. Quarterly J. Math., 10:266–

276, 1939.

118