Ballot Theorems, Old and New

Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

Ballot theorems, old and new

L. Addario-Berry∗ B.A. Reed†

January 9, 2007

“There is a big difference between a fair game and a game it’s wise to play.”
-Bertrand (1887b).

1 A brief history of ballot theorems

1.1 Discrete time ballot theorems

We begin by sketching the development of the classical ballot theorem as it first appeared
in the Comptes Rendus de l’Academie des Sciences. The statement that is fairly called the
first Ballot Theorem was due to Bertrand:
Theorem 1 (Bertrand (1887c)). We suppose that two candidates have been submitted to
a vote in which the number of voters is µ. Candidate A obtains n votes and is elected;
candidate B obtains m = µ − n votes. We ask for the probability that during the counting of
the votes, the number of votes for A is at all times greater than the number of votes for B.
This probability is (2n − µ)/µ = (n − m)/(n + m).

Bertrand’s “proof” of this theorem consists only of the observation that if Pn,m counts the
number of “favourable” voting records in which A obtains n votes, B obtains m votes and
A always leads during counting of the votes, then

Pn+1,m+1 = Pn+1,m + Pn,m+1 ,

the two terms on the right-hand side corresponding to whether the last vote counted is
for candidate B or candidate A, respectively. This “proof” can be easily formalized as

Department of Statistics, University of Oxford, UK.

School of Computer Science, McGill University, Canada and Projet Mascotte, I3S (CNRS/UNSA)-
INRIA, Sophia Antipolis, France.

1
follows. We first note that the binomial coefficient Bn,m = (n + m)!/n!m! counts the total
number of possible voting records in which A receives n votes and B receives m, Thus, the
theorem equivalently states that for any n ≥ m, Bn,m − Pn,m , which we denote by ∆n,m ,
equals 2mBn,m /(n + m). This certainly holds in the case m = 0 as Bn,0 = 1 = Pn,0 ,
and in the case m = n, as Pn,n = 0. The binomial coefficients also satisfy the recurrence
Bn+1,m+1 = Bn+1,m + Bn,m+1 , thus so does the difference ∆n,m . By induction,
∆n+1,m+1 = ∆n+1,m + ∆n,m+1
2m 2(m + 1) 2(m + 1)
= Bn+1,m + Bn,m+1 = Bn+1,m+1 ,
n+m+1 n+m+1 n+m+2
as is easily checked; it is likely that this is the proof Bertrand had in mind.

After Bertrand announced his result, there was a brief flurry of research into ballot theorems
and coin-tossing games by the probabilists at the Academie des Sciences. The first formal
proof of Bertrand’s Ballot Theorem was due to André and appeared only weeks later (André,
1887). André exhibited a bijection between unfavourable voting records starting with a vote
for A and unfavourable voting records starting with a vote for B. As the latter number is
clearly Bn,m−1 , this immediately establishes that Bn,m − Pn,m = 2Bn,m−1 = 2mBn,m /(n + m).

A little later, Barbier (1887) asserted but did not prove the following generalization of the
classical Ballot Theorem: if n > km for some integer k, then the probability candidate A
always has more than k-times as many votes as B is precisely (n − km)/(n + m). In response
to the work of André and Barbier, Bertrand had this to say:

“Though I proposed this curious question as an exercise in reason and calcu-


lation, in fact it is of great importance. It is linked to the important question
of duration of games, previously considered by Huygens, [de] Moivre, Laplace,
Lagrange, and Ampere. The problem is this: a gambler plays a game of chance
in which in each round he wagers n1 ’th of his initial fortune. What is the proba-
bility he is eventually ruined and that he spends his last coin in the (n + 2µ)’th
round?” (Bertrand, 1887a)

He notes that by considering the rounds in reverse order and applying Theorem 1 it is clear
n n+2µ −(2µ+n)
that the probability that ruin occurs in the (n+2µ)’th round is nothing but n+2µ µ
2 .
By informal but basic computations, he then √ derives that the probability ruin occurs before
2/πn
the (n + 2µ)’th round is approximately 1 − √n+2µ , so for this probability to be large, µ must
be large compared to n2 . (Bertrand might have added Pascal, Fermat, and the Bernoullis
(Hald, 1990, pp. 226-228) to his list of notable mathematicians who had considered the game
of ruin; (Balakrishnan, 1997, pp. 98-114) gives an overview of prior work on ruin with an
eye to its connections to the ballot theorem.)

Later in the same year, he proved that in a fair game (a game in which, at each step, the
average change in fortunes is nil) where at each step, one coin changes hands, the expected

2
number of rounds before ruin is infinite. He did so using the fact that by the above formula,
the probability of ruin in the t’th round (for t large) is of the order 1/t3/2 , so the expected
time to ruin behaves as the sum of 1/t1/2 , which is divergent. He also stated that in a fair
game in which player A starts with a dollars and player B starts with b dollars, the expected
time until the game ends (until one is ruined) is precisely ab (Bertrand, 1887b). This fact
is easily proved by letting ea,b denote the expected time until the game ends and using the
recurrence ea,b = 1 + (ea−1,b + ea,b−1 )/2 (with boundary conditions ea+b,0 = e0,a+b = 0).
Expanding on Bertrand’s work, Rouché provided an alternate proof of the above formula for
the probability of ruin (Rouché, 1888a). He also provided an exact formula for the expected
number of rounds in a biased game in which player A has a dollars and bets a0 dollars each
round, player B has b dollars and bets b0 dollars each round, and in each round player A
wins with probability p satisfying a0 p > b0 (1 − p) (Rouché, 1888b).

All the above questions and results can be restated in terms of a random walk on the set
of integers Z. For example, let S0 = 0 and, for i ≥ 0, Si+1 = Si ± 1, each with probability
1/2 and independently of the other steps - this is called a symmetric simple random walk.
(For the remainder of this section, we will phrase our discussion in terms of random walks
instead of votes, with Xi+1 = Si+1 − Si constituting a step of the random walk.) Then
Theorem 1 simply states that given that St = h > 0, the probability that Si > 0 for all
i = 1, 2, . . . , t (i.e. the random walk is favourable for A) is precisely h/t. Furthermore, the
time to ruin when player A has a dollars and player B has b dollars is the exit time of the
random walk S from the interval [a, −b]. The research sketched above constitutes the first
detailed examination of the properties of a random walk S0 , S1 , . . . , Sn conditioned on the
value Sn , and the use of such information to study the asymptotic properties of such a walk.

In 1923, Aeppli proved Barbier’s generalized ballot theorem by an argument similar to that
used by André’s. This proof is presented in Balakrishnan (1997, pp.101-102), where it is also
observed that Barbier’s theorem can be proved using Bertrand’s original recurrence in the
same fashion as above. A simple and elegant technique was used by Dvoretzky and Motzkin
(1947) to prove Barbier’s theorem; we use it to prove Bertrand’s theorem as an example of
its application, as it highlights an interesting perspective on ballot-style results.

We think of X = (X1 , . . . , Xn+m , X1 ) as being arranged clockwise around a cycle (so that
Xn+m+1 = X1 ). There are precisely n + m walks corresponding to this set, obtained by
choosing a first step Xi , so to establish Bertrand’s theorem it suffices to show that how-
ever X1 , . . . , Xn+m are chosen such that Sn = n − m > 0, precisely n − m of the walks
Xi+1 , . . . , Xn+m , X1 , . . . , Xi are favourable for A. Let Sij = Xi+1 + . . . + Xj (this sum in-
cludes Xn+m if i < j). We say that Xi , . . . , Xj is a bad run if Sij = 0 and Si0 j < 0 for all
i0 ∈ {i + 1, . . . , j} (this set includes n + m if i > j). In words, this condition states that
i is the first index for which the reversed walk starting with Xj and ending with Xi+1 is
nonnegative. It is immediate that if two bad runs intersect then one is contained in the
other, so the maximal bad runs are pairwise disjoint. (An example of a random walk and its
bad runs is shown in Figure 1).
Pn
If Xi = 1 and Sij = 0 for some j then Xi begins a bad run, and since Sn = i=1 Xi > 0, if

3
10 10

Figure 1: On the left appears the random walk corresponding to the voting sequence
(1, −1, −1, 1, 1, −1, −1, 1, 1, 1), doubled to indicate the cyclic nature of the argument. On
the right is the reversal of the random walk; the maximal bad runs are shaded grey.

Xi = −1 then Xi ends a bad run. As Sij = 0 for a maximal bad run and Xi = 1 for every
Xi not in a bad run, there must be precisely n − m values of i for which Xi is not in a bad
run. If the walk starting with Xi is favourable for A then for all i 6= j, Sij is positive, so
Xi is not in a bad run. Conversely, if Xi is not in a bad run then Xi = 1 and for all j 6= i,
Sij > 0, so the walk starting with Xi is favourable for A. Thus there are precisely (n − m)
favourable walks corresponding to X , which is what we set out to prove.

With this technique, the proof of Barbier’s theorem requires nothing more than letting the
positive steps have value 1/k instead of 1. This proof is notable as it is the first time the idea
of cyclic permutations was applied to prove a ballot-style result. This “rotation principle”
is closely related to the strong Markov property of the random walk: for any integer t ≥ 0,
the random walk St − St , St+1 − St , St+2 − St , . . . has identical behavior to the walk S0 , S1 , S2
and is independent of S0 , S1 , . . . , St . (Informally, if we have examined the behavior of the
walk up to time S, we may think of restarting the random walk at time t, starting from
a height of St ; this will be important in the generalized ballot theorems to be presented
later in the paper.) This proof can be rewritten in terms of lattice paths by letting votes
for A be unit steps in the positive x-direction and votes for B be unit steps in the positive
y-direction. When conceived of in this manner, this proof immediately yields several natural
generalizations (Dvoretzky and Motzkin, 1947; Grossman, 1950; Mohanty, 1966).

Starting in 1962, Lajos Takács proved a sequence of increasingly general ballot-style results
and statements about the distribution of the maxima when the ballot is viewed as a random
walk (Takács, 1962a,b,c, 1963, 1964a,b, 1967). We highlight two of these theorems below; we
have not chosen the most general statements possible, but rather theorems which we believe
capture key properties of ballot-style results.

A family of random variables Xi , . . . , Xn is interchangeable if for all (r1 , . . . , rn ) ∈ Rn and all

4

permutations σ of {1, . . . , n}, P {Xi ≤ ri ∀1 ≤ i ≤ n} = P Xi ≤ rσ(i) ∀1 ≤ i ≤ n . We say
X1 , . . . , Xn is cyclically interchangeable if this equality holds for all cyclic permutations σ.
A family of interchangeable random variables is cyclically interchangeable, but the converse
is not always true. The first theorem states:
Theorem 2. Suppose that X1 , . . . , Xn are integer-valued, cyclically interchangeable random
variables with maximum value 1, and for 1 ≤ i ≤ n, let Si = X1 + . . . + Xi . Then for any
integer 0 ≤ k ≤ n,
k
P {Si > 0 ∀1 ≤ i ≤ n|Sn = k} = .
n

This theorem was proved independently by Tanner (1961) and Dwass (1962) – we note
that it can also be proved by Dvoretzky and Motzkin’s approach. (As a point of historical
curiosity, Takács’ proof of this result in the special case of interchangeable random variables,
and Dwass’ proof of the more general result above, appeared in the same issue of Annals of
Mathematical Statistics.) Theorem 2 and the “bad run” proof of Barbier’s ballot theorem
both suggest that the notion of cyclic interchangeability or something similar may lie at the
heart of all ballot-style results.
Theorem 3 (Takács (1967), p. 12). Let X1 , X2 , . . . be an infinite sequence of iid integer
random variables with mean µ and maximum value 1 and for any i ≥ 1, let Si = X1 +. . .+Xi .
Then (
µ if µ > 0,
P {Sn > 0 for n = 1, 2, . . .} =
0 if µ ≤ 0.

The proof of Theorem 3 proceeds by applying Theorem 2 to finite subsequences X1 , X2 , . . . , Xn ,


so this theorem also seems to be based on cyclic interchangeability. Takács has generalized
these theorems even further, proving similar statements for step functions with countably
many discontinuities and in many cases finding the exact distribution of maxni=1 (Si − i).

(Takács originally stated his results in terms of non-negative integer random variables –
his original formulation results if we consider the variables (1 − X1 ), (1 − X2 ), . . . and the
corresponding random walk.) Theorem 3 implies the following classical result about the
probability of ever returning to zero in a biased simple random walk:
Theorem 4 (Feller (1968), p. 274). In a biased simple random walk 0 = R0 , R1 , . . . in which
P {Ri+1 − Ri = 1} = p > 1/2, P {Ri+1 − Ri = −1} = 1 − p, the probability that there is no
n ≥ 1 for which Rn = 0 is 2p − 1.

Proof. Observe that the expected value of Ri − Ri−1 is 2p − 1 > 0, so if R1 = −1 then with
probability 1, Ri = 0 for some i ≥ 2. Thus,

P {Rn 6= 0 for all n ≥ 1} = P {Rn > 0 for all n ≥ 1} .

The latter probability is equal to 2p − 1 by Theorem 3.

5
We close this section by presenting the beautiful “reflection principle” proof of Bertrand’s
theorem. We think of representing the symmetric simple random walk as a sequence of
points (0, S0 ), (1, S1 ), . . . , (n, Sn ) and connecting neighbouring points. If S1 = 1 and the
walk is unfavourable, then letting k be the smallest value for which Sk = 0 and “reflecting”
the random walk S0 , . . . , Sk in the x-axis yields a walk from (0, 0) to (n, t) whose first step
is negative – this is shown in Figure 2. This yields a bijection between walks that are
unfavourable for A and start with a positive step, and walks that are unfavourable for A
and start with a negative step. As all walks starting with a negative step are unfavourable
for A, all that remains is rote calculation. This proof is often incorrectly attributed to
André (1887), who established the same bijection in a different way - its true origins remain
unknown.

Figure 2: The dashed line is the reflection of the random walk from (0,0) to the first visit of
the x-axis.

1.2 Continuous time ballot theorems

The theorems which follow are natural given the results presented in Section 1.1; how-
ever, their statements require slightly more preliminaries. A stochastic process is simply a
nonempty set of real numbers T and a collection of random variables {Xt , t ∈ T } defined
on some probability space. The collection of random variables {X1 , . . . , Xn } seen in Section
1.1 is an example of a stochastic process for which T = {1, 2, . . . , n}. In this section we
are concerned with stochastic processes for which T = [0, r] for some 0 < r < ∞ or else
T = [0, ∞).

A stochastic process {Xt , 0 ≤ t ≤ r} has (cyclically) interchangeable increments if for all


n = 2, 3, . . ., the finite collection of random variables {Xrt/n − Xr(t−1)/n , t = 1, 2, . . . , n} is
(cyclically) intechangeable. A process {Xt , t ≥ 0} has interchangeable increments if for all
r > 0 and n > 0, {Xrt/n − Xr(t−1)/n , t = 1, 2, . . . , n} is interchangeable, and is stationary if

6
this latter collection is composed of independent identically distributed random variables.
As in the discrete case, these are natural sorts of prerequisites for a ballot-style theorem to
apply.

There is an unfortunate technical restriction which applies to all the ballot-style results
we will see in this section. The stochastic process {Xt , t ∈ T } is said to be separable if
there are almost-everywhere-unique measurable functions X + , X− such that almost surely
X− ≤ Xt ≤ X + for all t ∈ T , and there are countable subsets S− , S + of T such that almost
surely X + = supt∈S + Xt and X− = inf t∈S− Xt . The results of this section only hold for
separable stochastic processes. In defense of the results, we note that there are nonseparable
stochastic processes {Xt , 0 ≤ t ≤ r} for which sup{Xt − t, 0 ≤ t ≤ r} is non-measurable. As
the distribution of this random variable is one of the key issues with which we are concerned,
the assumption of separability is natural and in some sense necessary in order for the results
to be meaningful. Moreover, in very general settings it is possible to construct a separable
stochastic process {Yt |t ∈ T } such that for all t ∈ T , Yt and Xt are almost surely equal (see,
e.g., Gikhman and Skorokhod, 1969, Sec.IV.2); in this case it can be fairly said that the
assumption of separability is no loss.

The following theorem is the first example of a continuous-time ballot theorem. A sample
function of a stochastic process is a function xω : T → R given by fixing some ω ∈ Ω and
letting xω (t) = Xt (ω).
Theorem 5 (Takács (1965a)). If {Xt , 0 ≤ t ≤ r} is a separable stochastic process with
cyclically interchangeable increments whose sample functions are almost surely nondecreasing
step functions, then
(
t−s
t
if 0 ≤ s ≤ t,
P {Xt − X0 ≤ t for 0 ≤ t ≤ r|Xr − X0 = s} =
0 otherwise.

This theorem is a natural continuous equivalent of Theorem 2 of Section 1.1; it can be used to
prove a theorem in the vein of Theorem 3 which applies to stochastic processes {Xt , t ≥ 0}.
Takács’ other ballot-style results for continuous stochastic processes are also essentially step-
by-step extensions of his results from the discrete setting; see Takács (1965a,b, 1967, 1970b).

In 1957, Baxter and Donsker derived a double integral representation for sup{Xt − t, t ≥ 0}
when this process has stationary independent increments. Their proof proceeds by analyzing
the zeros of a complex-valued function associated to the process. They are able to use their
representation to explicitly derive its distribution when the process is a Gaussian process, a
coin-tossing process, or a Poisson process. This result was rediscovered by Takács (1970a),
who also derived the joint distribution of Xr and sup{Xt − t, 0 ≤ t ≤ r} for r finite, using
a generating function approach. Though these results are clearly related to the continuous
ballot theorems, they are not as elegant, and neither their statements nor their proofs bring
to mind the ballot theorem. It seems that considering separable stationary processes in their
full generality does not impose enough structure for it to be possible to prove these results
via straightforward generalization of the discrete equivalents.

7
A beautiful perspective on the ballot theorem appears by considering random measures
instead of stochastic processes. Given an almost surely nondecreasing separable stochastic
process {Xt , 0 ≤ t ≤ r}, fixing any element ω of the underlying probability space Ω yields
a sample function xω . By our assumptions on the stochastic process, almost every sample
function xω yields a measure µω on [0, r], where µω [0, b] = xω (b) − xω (a). This allows us to
define a “random” measure µ on [0, r]; µ is a function with domain Ω, µ(ω) = µω , and for
almost all ω ∈ Ω, µ(ω) is a measure on [0, r]. If xω is a nondecreasing step function, then µω
has countable support, so is singular with respect to the Lebesgue measure (i.e. the set of
points which have positive µω -measure has Lebesgue measure 0); if this holds for almost all
ω then µ is an “almost surely singular” random measure.

We have just seen an example of a random measure; we now turn to a more precise definition.
Given a probability space S = (Ω, Σ, P), a random measure on a possibly infinite interval
T ⊂ R is a function µ with domain Ω × T satisfying that for all r ∈ T , µ(·, r) is a random
variable in S, and for almost all ω ∈ Ω, µ(ω, ·) is a measure on T . A random measure µ
is almost surely singular if for almost all ω ∈ Ω, µ(ω, ·) is a measure on T singular with
respect to the Lebesgue measure. (This definition hides some technicality; in particular,
for the definition to be useful it is key that the set of ω for which µ is singular is itself
a measurable set! See Kallenberg (1999) for details.) A random measure µ on R+ , say,
is stationary if for all t, letting Xt,i = µ(·, (i + 1)/t) − µ(·, i/t), the family {Xt,i |i ∈ N}
is composed of independent identically distributed random variables; stationarity for finite
intervals is defined similarly.

This perspective can be used to generalize Theorem 5. Konstantopoulos (1995) has done so,
as well as providing a beautiful proof using a continuous analog of the reflection principle.
The most powerful theorem along these lines to date is due to Kallenberg. To a given
stationary random measure µ defined on T ⊆ R+ we associate a random variable I called
the sample intensity of µ. (Intuitively, I is the random average number of points in an
arbitrary measurable set B ⊂ T of positive finite measure, normalized by the measure of B.
For a formal definition, see (Kallenberg, 2003, p. 189).)
Theorem 6 (Kallenberg (1999)). Let µ be an almost surely singular, stationary random
measure on T = R+ or T = (0, 1] with sample intensity I and let Xt = µ(·, t) − µ(·, 0) for
t ∈ T . Then there exists a uniform [0, 1] random variable U independent from I such that
Xt I
sup = almost surely.
t∈T t U

It turns out that if T = (0, 1] then conditional upon the event that X1 = m, I = m almost
Xt

surely. It follows that in this case P supt∈T t ≤ 1|X1 = max{1 − X1 , 0}. Similarly, if
T = R+ and Xtt → m almost surely as t → ∞, then I = m almost surely, so in this case
Xt
P supt∈T t ≤ 1 = max{1 − m, 0}. This theorem can thus be seen to include continuous
generalizations of both Theorem 2 and Theorem 3.

Kallenberg has also proved the following as a corollary of Theorem 6 (this is a slight refor-
mulation of his original statement, which applied to infinite sequences):

8
Theorem 7. If X is a real random variable with maximum value 1 and {X1 , X2 , . . . , Xn }
are iid copies of X with corresponding partial sums {0 = S0 , S1 , . . . , Sn }, then
Sn
P {Si > 0∀1 ≤ i ≤ n|Sn } ≥ .
n

It is worth comparing this theorem with Theorem 2; the theorems are almost identical, but
Theorem 7 relaxes the integrality restriction at the cost of replacing the equality of Theorem
2 with an inequality.

1.3 Outline

To date, Theorem 7 is the only ballot-style result which has been proved for random walks
that may take non-integer values. Paraphrasing Harry Kesten (1993), the goal of our research
is to move towards making ballot theorems part of “the general theory of random walks”
– part of the body of results that hold for all random walks (with independent identically
distributed steps), regardless of the precise distribution of their steps. We succeed in proving
ballot-style theorems that hold for a broad class of random walks, including all random walks
that can be renormalized to converge in distribution to a normal random variable. A truly
general ballot theorem, however, remains beyond our grasp.

In Section 2 we discuss in what sense existing ballot theorems such as those presented in
Section 1 are optimal, and what sorts of “general ballot theorems” it makes sense to search
for in light of this optimality. In Section 3 we demonstrate our approach in a restricted
setting and prove a weakening of our main result. This allows us to highlight the key ideas
behind our general ballot theorems without too much notational and technical burden. In
Section 4, we sketch the main ideas required to strengthen the presented result. Finally, in
Section 5 we address the limits of our approach and suggest some avenues for future research.

2 General ballot theorems

The aim of our research is to prove analogs of the discrete-time ballot theorems of Section
1 for more general random variables. The Theorems of Section 1.1 all have two restrictions:
(1) they apply only to integer-valued random variables, and (2) they apply only to random
variables that are bounded from one or both sides. (In the continuous-time setting, the
restriction that the stochastic processes are almost surely integer-valued, increasing step
functions is of the same flavour.) In this section we investigate what ballot-style theorems
can be proved when such restrictions are removed.

The restrictions (1) and (2) are necessary for the results of Section 1.1 to hold. Suppose,
for example, that we relax the condition of Theorem 2 requiring that the variables are

9
bounded from above by +1. If X takes every value in N with positive probability, then
P {Si > 0∀1 ≤ i ≤ n|Sn = n} < 1, so the conclusion of the theorem fails to hold. For a
more explicit example, let X be any random variable taking values ±1, ±4 and define the
corresponding cyclically interchangeable sequence and random walk. For S3 = 2 to occur,
we must have {X1 , X2 , X3 } = {4, −1, −1}. In this case, for Si > 0, i = 1, 2, 3 to occur, X1
must equal 4. By cyclic interchangeability, this occurs with probability 1/3, and not 2/3, as
Theorem 2 would suggest. This shows that the boundedness condition (2) is required. If we
relax the integrality condition (1), we can construct a similar example where the conclusions
of Theorem 2 do not hold.

Since the results of Section 1.1 can not be directly generalized to a broader class of random
variables, we seek conditions on the distribution of X so that the bounds of that section have
the correct order, i.e., so that P {Si > 0 ∀ 1 ≤ i ≤ n|Sn = k} = Θ(k/n). (When we consider
random variables that are not necessarily integer-valued, the right conditioning will in fact
be on an event such as {k ≤ Sn < k + 1} or something similar.) How close we can come to
this conclusion will depend on what restrictions on X we are willing to accept. It turns out
that a statement of this flavour holds for the mean zero random walk Sn0 = Sn − nEX as
long as there is a sequence {an }n≥0 for which (Sn − nEX)/an converges to a non-degenerate
normal distribution (in this case, we say that X is in the range of attraction of the normal
distribution and write X ∈ D; for example, √ the classical central limit theorem states that
2
if E {X } < ∞ then we may take an = n for all n.) For the purposes of this expository
article, however, we shall impose a slightly stronger condition than that stated above.

From this point on, we restrict our attention to sums of mean zero random variables. We
note this condition is in some sense necessary in order for the results we are hoping for to
hold. If EX 6= 0 – say EX > 0 – then it is possible that X is non-negative, so the only way
for Sn = 0 to occur is that X1 = . . . = Xn = 0, and so P {Si > 0 ∀ 1 ≤ i ≤ n|Sn = 0} = 0,
and not Θ(1/n) as we would hope from the results of Section 1.

3 Ballot theorems for closely fought elections

One of the most basic questions a ballot theorem can be said to answer is: given that an
election resulted in a tie, what is the probability that one of the candidates had the lead
at every point aside from the very beginning and the very end. In the language of random
walks, the question is: given that Sn = 0, what is the probability that S does not return to 0
or change sign between time 0 and time n? Erik Sparre Andersen has studied the conditional
behavior of random walks given that Sn = 0 in great detail, in particular deriving beautiful
results on the distribution of the maximum, the minimum, and the amount of time spent
above zero. Much of the next five paragraphs can be found in Andersen (1953), for example,
in slightly altered terminology.

We call the event that Sn does not return to zero or change sign before time n, Lead n . We

10
can easily bound P {Lead n |Sn = 0} using the fact that X1 , . . . , Xn are interchangeable. If
we condition on the multiset of outcomes {X1 , . . . , Xn } = {xσ(1) , . . . , xσ(n) }, and then choose
a uniformly random cyclic permutation σ and a uniform element i of {1, . . . , n}, then the
interchangeability of X1 , . . . , Xn implies that (xσ(i) , . . . , xσ(n) , xσ(1) , . . . , xσ(i−1) ) has the same
distribution as if we had sampled directly from (X1 , . . . , Xn ).
Pj−1
Letting sj = k=1 xσ(k) , in order for Lead n to occur given that Sn = 0, it must be the
case that si is either the unique maximum or the unique minimum among {s1 , . . . , sn }. The
probability that this occurs is at most 2/n as it is exactly 2/n if there are unique maxima
and minima, and less if either the maximum or minimum is not unique. Therefore,
2
P {Lead n |Sn = 0} ≤ . (1)
n
On the other hand, the sequence certainly has some maximum (resp. minimum) si , and if
X1 = xi then Sj is always non-positive (resp. non-negative). Denoting this event by Nonpos n
(resp. Nonneg n ), we therefore have
1 1
P {Nonpos n |Sn = 0} ≥ and P {Nonneg n |Sn = 0} ≥ (2)
n n
If Sn = 0 then the (n − 1) renormalized random variables given by Xi0 = Xi+1 + X1 /(n − 1)
0 n−1 n
= (n − 1) i=1 Xi0 = (n − 1) i=1 Xi = 0. If X1 > 0 and none of
P P
satisfy (n − 1)Sn−1
the renormalized partial sums are negative, then Lead n occurs. The renormalized random
variables are still interchangeable (see Andersen (1953, Lemma 2) for a proof of this easy
fact), so we may apply the second bound of (2) to obtain
1
P {Lead n |Sn = 0, X1 > 0} ≥ .
n−1
An identical argument yields the same bound for P {Lead n |Sn = 0, X1 < 0}, and combining
these bounds yields

P {Lead n |Sn = 0} ≥ P {Lead n |Sn = 0, X1 6= 0} P {X1 6= 0|Sn = 0}


1 − P {X1 = 0|Sn = 0}
≥ .
n−1
As long as P {X1 = 0|Sn = 0} < 1, this yields that P {Lead n |Sn = 0} ≥ α/n for some α > 0.
By interchangeability, it is easy to see that P {X1 = 0|Sn = 0} is bounded uniformly away
from 1 for large n, as long as Sn = 0 does not imply that X1 = . . . = Xn = 0 almost surely.
(Note, however, that there are cases where P {X1 = 0|Sn = 0} = 1, for example √ if the Xi
only take values in the non-negative integers and in the negative multiples of 2.)

Sparre Andersen’s approach gives a necessary and sufficient, though not terribly explicit,
condition for P {Lead n |Sn = 0} = Θ(1/n) to hold. Philosophically, in order to make ballot
theorems part of the “general theory of random walks”, we would like necessary and sufficient
conditions on the distribution of X1 for P {Lead n |Sn = k} = Θ(k/n) for all k = O(n). Even
more generally, we may ask: what are sufficient conditions on the structure of a multiset S

11
of n numbers to ensure that if the elements of the multiset sum to k, then in a uniformly
random permutation of the set, all partial sums are positive with probability of order k/n?
In the remainder of the section, we focus our attention on sets S whose elements are sampled
independently from a mean-zero probability distribution, i.e., they are the steps of a mean-
zero random walk. (We remark that it is possible to apply parts of our analysis to sets S
that do not obey this restriction, but we will not pursue such an investigation√ here.) We will
derive sufficient conditions for such bounds to hold in the case that k = O( n); it turns out
that for our approach to succeed it suffices that the step size X is in the range of attraction of
the normal distribution, though our best result requires slightly stronger moment conditions
on X than those of the classical central limit theorem.

Before stating our generalized ballot theorems, we need one additional definition. We say a
variable X has period d > 0 if dX is an integer random variable and d is the smallest positive
real number for which this holds; in this case X is called a lattice random variable, otherwise
X is non-lattice. We can prove the following:
Theorem 8. Suppose X satisfies EX = 0, Var {X} > 0, E {X 2+α } < ∞ for some α >
0, and X is non-lattice. Then for any fixed A > 0, given independent
Pi random variables
X1 , X2 , . . . distributed as X with associated partial sums Si = j=1 Xj , for all k such that

0 ≤ k = O( n),
 
k+1
P {k ≤ Sn ≤ k + A, Si > 0 ∀ 0 < i < n} = Θ .
n3/2
Theorem 9. Suppose X satisfies EX = 0, Var {X} > 0, E {X 2+α } < ∞ for some α > 0,
and X is a lattice random variable with period d. Then given independent
Pi random variables
X1 , X2 , . . . distributed as X with associated partial sums Si = j=1 Xj , for all k such that

0 ≤ k = O( n) and such that k is a multiple of 1/d,
 
k+1
P {Sn = k, Si > 0 ∀ 0 < i < n} = Θ .
n3/2

From these theorems, we may√ derive “true” (conditional) ballot theorems as corollaries, at
least in the case that k = O( n). The following result was proved by Stone (1965), and
is the tip of an iceberg of related results. Let Φ be the density function a N (0, 1) random
variable.
Theorem 10. Suppose Sn is a sum of independent, identically distributed√random variables
distributed as X with EX = 0, and there is a constant a such that Sn /a n converges to a
N (0, 1) random variable. If X is non-lattice let B be any bounded set; then for any h ∈ B
and x ∈ R √
hΦ(x/a n) √
P {|Sn − x| ≤ h/2} = √ + o(1/ n).
a n
Furthermore, if X is a lattice random variable with period d, then for any x ∈ {n/d | n ∈ Z},

Φ(x/a n) √
P {Sn = x} = √ + o(1/ n).
a n

12
√ √
In both cases, no(1/ n) → 0 as n → ∞ uniformly over all x ∈ R and h ∈ B.

Together with Theorem 8 this immediately yields:


Corollary 11. Under the conditions of Theorem 8,
 
k+1
P {Si > 0 ∀ 0 < i < n|k ≤ Sn ≤ k + A} = Θ .
n

Similarly, combining Theorem 9 with Theorem 10, we have


Corollary 12. Under the conditions of Theorem 9,
 
k+1
P {Si > 0 ∀ 0 < i < n|Sn = k} = Θ .
n

As we remarked above, the approach we are about to sketch can also be used to prove a
ballot theorem under the weaker restriction that X is in the
 range of attraction of the normal
distribution, at the cost of replacing the bound Θ k+1 n
by the bound k+1
n1+o(1)
; for the sake
of brevity and clarity we will not discuss the rather minor modifications to our approach
that are needed to handle this case. Furthermore, for the purposes of this expository article,
we shall not prove Theorems 8 or 9 in their full generality or strength, instead restricting
our attention to a special case which allows us to highlight the key elements of our proofs.
Finally, we shall provide a detailed explanation of only the upper bound, after which we
shall briefly discuss our approach to the lower bound. We will prove:
Theorem 13. Suppose X satisfies EX = 0, Var {X} > 0, |X| < C, and X is non-lattice.
Then for any fixed A > 0, given independent random variablespX1 , X2 , . . . , distributed as X
with associated partial sums Si = ij=1 Xj , for all 0 ≤ k = o( n/ log n),
P

 
(k + 1) log n
P {k ≤ Sn ≤ k + A, Si > 0 ∀ 0 < i < n} = O .
n3/2

Of course, a conditional ballot theorem that is correspondingly weaker than Corollary 11


follows by combining Theorem 13 with Theorem 10. We remark that in cases where Theorem
7 applies, it provides a lower bound on P {k ≤ Sn ≤ k + A, Si > 0 ∀ 0 < i < n} of the same
order as the upper bound of Theorem 13. From this point forward, X will always be a random
variable satisfying the conditions in Theorem 13, and X1 , X2 , . . . , will be independent copies
of X with corresponding partial sums S1 , S2 , . . ..

To begin providing an intuition of our approach, we first remark that if Si > 0 ∀0 < i < n
is to occur, then for any r, letting T be the first time t ≥ 1 that St > r or St ≤ 0, we have
either ST > 0 or T > n. (We will end up choosing the value r so that T = o(n) except with
negligibly small probability, so to bound the previous probability we shall essentially need to

13
bound the probability that ST > 0, i.e., that the walk “stays positive”. We will see shortly
that Wald’s identity implies that P {ST > 0} = O(1/r).

We may impose a similar constraint on the “other end” of the random walk S, by letting
S 0 be the negative reversed random walk given by S00 = 0, and for i > 0, Si+10
= Si0 − Xn−i
0
(it will be useful to think of Si as being defined even for i > n, which we may do by letting
X0 , X−1 , . . . be independent copies of X). If Si > 0 ∀0 < i < n and k ≤ Sn ≤ k + C are to
occur, then letting T 0 be the first time t that St0 ≤ −(k + A) or St0 > r − (k + A), it must
be the case that either ST0 0 > 0 or T 0 > n. (Again, we will choose r so that T 0 = o(n) with
extremely high probability.)

Finally, in order for k ≤ Sn ≤ k + A to occur, the two ends of the random walk must “match
up”. We may make this mathematically precise by noting that as long as T < n − T 0 , we
may write Sn as ST + (Sn−T 0 − ST ) − ST0 0 , and may thus write the condition k ≤ Sn ≤ k + A
as
k + ST0 0 − ST ≤ (Sn−T 0 − ST ) ≤ k + A + ST0 0 − ST .
If T + T 0 is at most n/2, say, then Sn−T 0 − ST is the sum of at least n/2 random variables. In
√ limit theorem suggests that Sn−T − ST should “spread itself
this case, the classical central 0

out” over a range of order n, and essentially


√ this fact will allow us to show that the two
ends “meet up” with probability O(1/ n).

3.1 Staying positive

To begin formalizing the above sketch, let us first turn to bounds on the probabilities of the
events ST > 0 and ST0 0 > 0.

Lemma 14. Fix r > 0 and s ≥ 0, and let Tr,s be the first time t > 0 that either St > r or
St ≤ −s. Then P STr,s > 0 ≤ (s + C)/(r + s + C).

Proof. We first remark that ETr,s is finite; this is a standard result that can be found in,
e.g., (Feller, 1968, Chapter 14.4), and we shall also rederive this result a little later. Thus,
by Wald’s identity, we have that ESTr,s = ETr,s EX1 = 0, and letting Pos r denote the event
{STr,s > 0}; we may therefore write
  
0 = ESTr,s = E STr,s |Pos r P {Pos r } + E STr,s |Pos r P Pos r . (3)

 {ST |Pos r } ≥ r, and by our assumption that X has absolute value at most
By definition E
C, we have E ST |Pos r ≥ −(s + C). Therefore

0 ≥ rP {Pos r } − (s + C)P Pos r = rP {Pos r } − (s + C)(1 − P {Pos r }),

and rearranging the latter inequality yields that P {Pos r } ≤ (s + C)/(r + s + C).

14
As an aside, we note that may easily derive
 a lower bound of the same order for P {Pos r } in a
similar fashion; we first observe that E STr,s |Pos r < r +C. Similarly, E STr,s |Pos r ≤ −s,
and using the fact that X has zeromean and positive variance, it is also easy to see that
there is  > 0 such that in fact E STr,s |Pos r ≤ − max{, s}. Combining (3) these two
bounds, we thus have

0 < (r + C)P {Pos r } − max{, s}P Pos r = (r + C)P {Pos r } − max{, s}(1 − P {Pos r }),
so P {Pos r } ≥ max{, s}/(r + C + max{, s}). Lemma 14 immediately yields the bounds we
require for P {ST > 0} and P {ST0 0 > 0}; next we show that for a suitable choice of r, with
extremely high probability, both T and T 0 are o(n).

3.2 The time to exit a strip.

For r ≥ 0, we consider the first time t for which |St | ≥ r, denoting this time Tr . We prove
Lemma 15. There is B > 0 such that for all r ≥ 1, ETr ≤ Br2 and for all integers k ≥ 1,
P {Tr ≥ kBr2 } ≤ 1/2k .

This is an easy consequence of a classical result on how “spread out” sums of independent
identically distributed random variables become (which we will also use later when bounding
the probability that the two ends of the random walk “match up”). The version we present
can be found in Kesten (1972):
Theorem 16. For any family of independent identically distributed real random variables
X1 , X2 , . . . with positive, possibly infinite variance and associated partial sums S1 , S2 , . . . ,
there is a constant c depending only on the distribution of X1 such that for all n,

sup P {x ≤ Sn ≤ x + 1} ≤ c/ n.
x∈R

Proof of Lemma 15. Observe that the expectation bound follows directly from the probabil-
ity bound, since if the probability bound holds then we have
∞ ∞ ∞
X X
2 2
X dBr2 e
= 2dBr2 e,

ETr ≤ P {Tr ≥ j} ≤ dBr eP Tr > idBr e ≤
j=0 i=0 i=0
2i

which establishes the expectation bound with a slightly changed value of B. It thus remains
to prove the probability bound. By Theorem 16, there is c > 0 (and we can and will assume
c > 1) such that
b2rc
 X 
P |Sd128c2 r2 e | ≤ 2r ≤ P i ≤ Sd128c2 r2 e ≤ i + 1
i=b−2rc
c 1
≤ (4r + 1) p < , (4)
d128c2 r2 e 2

15
the last inequality holding as c > 1 and r > 1. Let t∗ = d128c2 r2 e - then P {Tr > t∗ } ≤ 1/2.
We use this fact to show that for any positive integer k, P {Tr > kt∗ } ≤ 1/2k , which will
establish the claim with B = 128c2 + 1, for example. We proceed by induction on k, having
just proved the claim for k = 1. We have

P {Tr > (k + 1)t∗ } = P {Tr > (k + 1)t∗ ∩ T > kt}


= P {Tr > (k + 1)t∗ |Tr > kt∗ } P {Tr > kt}
1
= k · P {Tr > (k + 1)t∗ |Tr > kt∗ } ,
2
by induction. It remains to show that P {Tr > (k + 1)t∗ |Tr > kt∗ } ≤ 1/2. If Tr > kt∗ then
by the strong Markov property we may think of restarting the random walk at time kt∗ .
Whatever the value of Skt∗ , if the restarted random walk exits [−2r, 2r] then the original
random walk exits [−r, r], so this inequality holds by (4). This proves the lemma.

This bound on the time to exit a strip is the last ingredient we need; we now turn to the
proof of Theorem 13.

3.3 Proof of Theorem 13

Fix A > 0 as in the statement of the theorem. For r ≥ 1 we denote by Tr the first time t that
|St | ≥ r. We let S 0 be the negative reversed random walk given by S00 = 0, and for i > 0,
0
Si+1 = Si0 −Xn−i (again as above, we define Si0 for i > n by letting X0 , X−1 , . . . be independent
copies of X), and let Tr0 be the first time t that |St0 | ≥ r. We choose B such that for all r ≥ 1
and and for all integers k ≥ 1, P {Tr ≥ kBr2 } ≤ 1/2k and P {Tr0 ≥ kBr2 } ≤ 1/2k – such a
choice exists by Lemma 15.
p
Choose r∗ = b n/9B log nc – then with k = d2 log ne < 2 log n + 1, it is the case that
kBn (2 log n + 1)n n
kB(r∗ )2 ≤ < < ,
9B log n 9 log n 4
so P {Tr∗ ≥ n/4} ≤ 1/2k ≤ 1/n2 , and similarly P {Tr0∗ ≥ n/4} ≤ 1/n2 .

Next let T be the first time t that St > r∗ or St ≤ 0, and let T 0 be the first time t that
St0 > rp∗
− (k + A) or St0 ≤ −(k + A). It is immediate that T < Tr∗ . Furthermore, since
k = o( n/ log n), (k+A) < r∗ for n large enough, so r∗ > r∗ −(k+A) > 0 > −(k+A) > −r∗ ;
it follows that T 0 < Tr0∗ . These two inequalities, combined with the bounds for Tr∗ and Tr0∗ ,
yield n no 1 n
0 no 1
P T ≥ ≤ 2 and P T ≥ ≤ 2 (5)
4 n 4 n

Let E be the event that k ≤ Sn ≤ k + A, and Si > 0 for all 0 < i < n – we aim to show that
P {E} = O((k + 1) log n/n3/2 ). In order that E occur, it is necessary that either T ≥ n/4 or

16
T 0 ≥ n/4 (we denote the union of these two events by D), or that the following three events
occur (these events control the behavior of the beginning, end, and middle of the random
walk, respectively):

E1 : ST > 0 and T < n/4,

E2 : ST0 0 > 0 and T 0 < n/4,


0
E3 : letting ∆ = Sbn/4c − Sbn/4c , we have k + ∆ ≤ Sn−bn/4c − Sbn/4c ≤ k + ∆ + A

It follows that
P {E} ≤ P {D} + P {E1 , E2 , E3 } .
Furthermore, P {D} ≤ P {T ≥ n/4}+P {T 0 ≥ n/4} ≤ 2/n2 by (5), so to show that P {E} =
O(log n/n3/2 ), it suffices to show that P {E1 , E2 , E3 } = O(log n/n3/2 ); we now demonstrate
that this latter bound holds, which will complete the proof.

The events E1 and E2 are independent, as E1 is determined by the random variables


X1 , . . . , Xbn/4c , and E2 is determined by the random variables Xn−bn/4c+1 , . . . , Xn . Fur-
thermore, in the notation of Lemma 14, T is an event of the form Tr,s with r = r∗ , s = 0; it
follows that P {ST > 0} ≤ C/(r∗ + C). Since S 0 has step size −X and | − X| < C, we may
also apply Lemma 14 to the walk S 0 with the choice r = r∗ − k + C, s = k + C, to obtain
the bound P {ST0 0 > 0} ≤ (k + 2C)/(r + k + 2C). Therefore

P {E1 , E2 , E3 } = P {E3 |E1 , E2 } P {E1 } P {E2 }


≤ P {E3 |E1 , E2 } P {ST > 0} P {ST0 0 > 0}
C(k + 2C) 2C 2 (k + 1)
≤ P {E3 |E1 , E2 } · ∗ ∗ < P {E3 |E1 , E2 } · (6)
r (r + k + 2C) (r∗ )2
To bound P {E3 |E1 , E2 }, we observe that

P {E3 |E1 , E2 } ≤ sup P {E3 |E1 , E2 , ∆ = x}


x∈R

= sup P k + x ≤ Sn−bn/4c − Sbn/4c ≤ k + x + A|E1 , E2 , ∆ = x . (7)
x∈R

Furthermore, the event that k + x ≤ Sn−bn/4c − Sbn/4c ≤ k + x + A is independent from


E1 , E2 , and from the event that ∆ = x, as the former event is determined by the random
variables Xbn/4c+1 , . . . , Xn−bn/4c , and the latter events are determined by the random variables
X1 , . . . , Xbn/4c , Xn−bn/4c+1 , . . . , Xn . It follows from this independence, (7), and the strong
Markov property that

P {E3 |E1 , E2 } ≤ sup P k + x ≤ Sn−bn/4c − Sbn/4c ≤ k + x + A
x∈R

= sup P k + x ≤ Sn−2bn/4c ≤ k + x + A .
x∈R

≤ (A + 1) sup P k + x ≤ Sn−2bn/4c ≤ k + x + 1 , (8)
x∈R

17
the last inequality holding by a union bound. By Theorem 16, there is c > 0 depending only
on X, such that

 c 2c
sup P x ≤ Sn−2bn/4c ≤ x + 1 ≤ p ≤ √ ,
x∈R n − 2bn/4c n

and it follows from this fact and from (8) that



2c(A + 1)
P {E3 |E1 , E2 } ≤ √ .
n

Combining this bound with (6) yields


√ √
2c(A + 1) 2C 2 (k + 1) 2 2c(A + 1)C 2 (k + 1)
P {E1 , E1 , E3 } ≤ √ · = √ .
n (r∗ )2 (r∗ )2 n
p √
Since r∗ = b n/9B log nc, (r∗ )2 ≥ n/10B log n for n large enough, so letting a = 2 2(A +
1)cC 2 · 10B = O(1), we have
 
a(k + 1) log n (k + 1) log n
P {E1 , E1 , E3 } < =O , (9)
n3/2 n3/2
as claimed.

4 Strengthening Theorem 13

There are two key ingredients needed to move from the upper bound in Theorem 13 to the
stronger and more general upper bound in Theorem 8. The first concerns stopping times Tr,s
of the form seen in Lemma 14. Without
 the assumption
that the step size
X is bounded,
we have no a priori bound on E STr,s |STr,s > r or on E STr,s |STr,s ≤ −s , so we can not
straightforwardly apply Wald’s identity to bound P STr,s > 0 as we did above.

Griffin and McConnell (1992) have proven bounds on E |STr,r | − r (a quantity they call
the overshoot at r), for random walks with step size X in the domain of attraction of the
normal distribution; their results are the best possible in the setting they consider. Their
bounds do not directly imply the bounds we need, but we are able to use their results to
obtain such bounds using a bootstrapping technique we refer to as a “doubling argument”.
The key idea behind this argument can be seen by considering a symmetric simple random
walk S and a stopping time T3k,0 , for some positive integer k. Let T be the first time t > 0
that |St | = k. If the event ST3k,0 > 0 is to occur, it must be the case that ST = k. Next, let
T 0 be the first time t > T that |St − ST | ≥ 2k – if ST3k,0 > 0 is to occur, it must also be the
case that ST 0 − ST = 2k. By the independence of disjoint sections of the random walk, it
follows that 
P ST3k,0 > 0 ≤ P {ST = k} P {ST 0 − ST = 2k} .

18
In the notation of Lemma 14, T is a stopping time of the form Tk,k , and T2 is a stopping
time of the form T2k,2k , so we have
  
P ST3k,0 > 0 ≤ P STk,k > 0 P ST2k,2k > 0 .

Furthermore, for a general random walk S we can use the Griffin and McConnell’s  bounds

on the overshoot together with the approach of Lemma 14 to prove bounds P STk,k > 0 .
In general, we consider a sequence of stopping times T1 , T2 , . . ., where Ti+1 is the first time
after Ti that |STi+1 − STi | ≥ 2i , and apply Griffin and McConnell’s results to bound the
probability that the random walk goes positive at each step. By applying their results
to such a sequence of stopping times, we are able to ensure that the error in our bounds
resulting from the “overshoot” does not accumulate, and thereby prove the stronger bounds
we require.

The second difficulty we must overcome is due to the fact that in order to remove the
superfluous log n factor
p in the bound of Theorem 13, we need to replace
√ the stopping time
Tr∗ (with r∗ = O( n/ log n)) by a stopping time Tr0 (with r0 = Θ( n)). However, for such
a value r0 , ETr0 = Θ(n), and our upper tail bounds on Tr0 are not strong enough to ensure
that Tr0 ≤ bn/4c with sufficiently high probability.

To deal with this problem, we apply the ballot theorem inductively. Instead of stopping the
walk at a stopping time Tr0 , we stop the walk deterministically at time t1 = bn/4c. In order
that Sn > 0 for all 0 < i < n occur, it must be the case that either Tr0 ≤ t1 and STr0 > 0, or
there is 0 < k ≤ r0 − C such that k ≤ St1 ≤ k + A and additionally, Si > 0 for all 0 < i < t1 .
We bound the probability of the former event using our strengthening of Lemma 14, and
bound the probability of the latter event by inductively applying the ballot theorem. Of
course, an identical analysis applies to the negative reversed random walk S 0 , and allows us
to strengthen our control of the end of the random walk correspondingly.

Finally, we give some idea of our lower bound. We fix some value r0 of order Θ( n);
paralleling the proof of Theorem 13, we let T be the first time t that St > r0 or St ≤ 0, and
let T 0 be the first time t that St0 > r0 − k or St0 ≤ −k. In order that k ≤ Sn ≤ k + A, and
Si > 0 for all 0 < i < n, it suffices that the following three events occur (these events control
the behavior of the beginning, end, and middle of the random walk, respectively):

E1 : ST > 0, ST ≤ 2r0 , and T < n/4,

E2 : ST0 0 > 0, ST0 0 ≤ 2r0 − k, and T 0 < n/4,

E3 : letting ∆ = ST0 0 − ST , we have k + ∆ ≤ Sn−T 0 − ST ≤ k + ∆ + A and Si − ST > −r0


for all T < i < n − T 0 .

Using an approach very similar to that of Theorem 13, we are able to show that in fact
P {E1 } = Θ(1/r0 ) and that P {E2 } = Θ((k + 1)/r0 ). The two key observations of that allow
us to prove a lower bound on P {E3 } are the following:

19

• Given that E1 and E2 occur, k + ∆ = O( n), and so it is not hard √ to show using
Theorem 10 that P {k + ∆ ≤ Sn−T 0 − ST ≤ k + ∆ + A|E1 , E2 } = Θ(1/ n).

• Since n − T − T 0 =√ O(n), we expect the random walk ST , ST +1 , . . . , Sn−T 0 to have a



0
spread of order O( n). Since r = Θ( n), it easy to see (again using Theorem 10, or
by the classical central limit theorem) that Si − ST > −r0 for all T < i < n − T 0 with
probability Ω(1).

Based on these two observations, we trust that the reader will find it plausible
√ that given E1
and E2 , the intersection of the events in E3 occurs with probability Θ(1/ n); in this case,
combining our bounds √ much as in Theorem 13 yields a lower bound on P {E1 , E2 , E3 } of
0 2 3/2
order (k + 1)/(r ) n = Θ((k + 1)/n ).

5 Conclusion

In writing this survey, we hoped to convince the reader the theory of ballots is not only rich
and beautiful, in-and-of itself, but is also very much alive. Our new results are far from
conclusive in terms of when ballot-style behavior can be expected of sums of independent
random variables, and more generally of permutations of sets of real numbers. In the final
paragraphs, we highlight some of the questions that remain unanswered.

The results of Section 3 are√unsatisfactory in that they only yield “true” (conditional) ballot
theorems when Sn = O( n). Ideally, we would like such results to hold whatever the
range of Sn . Two key weaknesses of our approach are that it (a) relies on estimates for
P {x ≤ Sn ≤ x + c} that are based√ on the central limit theorem, and these estimates are not
good enough when Sn is not O( n), and (b) relies on bounds on the “overshoot” that only
hold when the step size X is in the range of attraction of the normal distribution, Kesten and
Maller (1994) and, independently, Griffin
 and McConnell (1994), have derived necessary and
sufficient conditions in order that P STr,r → 1/2 as r → ∞; in particular
 they show that
α
for any α < 2, there are distributions with E {X } = ∞ for which P STr,r → 1. Therefore,
we can not expect to use a doubling argument in this case, which seriously undermines our
approach.

As we touched upon at various points in the paper, aspects of our technique seem as though
they should work for analyzing more general random permutations of sets of real numbers.
Since Andersen observed the connection between conditioned random walks and random
permutations (Andersen, 1953, 1954), and Spitzer (1956) pointed out the full generality
of Andersen’s observations, just about every result on conditioned random walks has been
approached from the permutation-theoretic perspective sooner or later. There is no reason
our results should not benefit from such an approach.

20
References
Erik Sparre Andersen. Fluctuations of sums of random variables. Mathematica Scandinavica,
1:263–285, 1953.

Erik Sparre Andersen. Fluctuations of sums of random variables ii. Mathematica Scandi-
navica, 2:195–223, 1954.

Désiré André. Solution directe du probleme resolu par m. bertrand. Comptes Rendus de
l’Academie des Sciences, 105:436–437, 1887.

N. Balakrishnan. Advances in Combinatorial Methods and Applications to Probability and


Statistics. Birkhäuser, Boston, MA, first edition, 1997.

Émile Barbier. Generalisation du probleme resolu par m. j. bertrand. Comptes Rendus de


l’Academie des Sciences, 105:407, 1887.

J. Bertrand. Observations. Comptes Rendus de l’Academie des Sciences, 105:437–439, 1887a.

J. Bertrand. Sur un paradox analogue au problème de saint-pétersburg. Comptes Rendus de


l’Academie des Sciences, 105:831–834, 1887b.

J. Bertrand. Solution d’un probleme. Comptes Rendus de l’Academie des Sciences, 105:369,
1887c.

Aryeh Dvoretzky and Theodore Motzkin. A problem of arrangements. Duke Mathematical


Journal, 14:305–313, 1947.

Meyer Dwass. A fluctuation theorem for cyclic random variables. The Annals of Mathemat-
ical Statistics, 33(4):1450–1454, December 1962.

William Feller. An Introduction to Probability Theory and Its Applications, Volume 1, vol-
ume 1. John Wiley & Sons, Inc, third edition, 1968.

I.I. Gikhman and A.V. Skorokhod. Introduction to the Theory of Random Processes. Saun-
ders Mathematics Books. W.B. Saunders Company, 1969.

Philip S. Griffin and Terry R. McConnell. On the position of a random walk at the time of
first exit from a sphere. The Annals of Probability, 20(2):825–854, April 1992.

Philip S. Griffin and Terry R. McConnell. Gambler’s ruin and the first exit position of
random walk from large spheres. The Annals of Probability, 22(3):1429–1472, July 1994.

Howard D. Grossman. Fun with lattice-points. Duke Mathematical Journal, 14:305–313,


1950.

A. Hald. A History of Probability and Statistics and Their Applications before 1750. John
Wiley & Sons, Inc, New York, NY, 1990.

21
Olav Kallenberg. Ballot theorems and sojourn laws for stationary processes. The Annals of
Probability, 27(4):2011–2019, 1999.

Olav Kallenberg. Foundations of Modern Probability. Probability and Its Applications.


Springer Verlag, second edition, 2003.

Harry Kesten. Sums of independent random variables–without moment conditions. Annals


of Mathematical Statistics, 43(3):701–732, June 1972.

Harry Kesten. Frank spitzer’s work on random walks and brownian motion. Annals of
Probability, 21(2):593–607, April 1993.

Harry Kesten and R.A. Maller. Infinite limits and infinite limit points of random walks and
trimmed sums. The Annals of Probability, 22(3):1473–1513, 1994.

Takis Konstantopoulos. Ballot theorems revisited. Statistics & Probability Letters, 24(4):
331–338, September 1995.

Sri Gopal Mohanty. An urn problem related to the ballot problem. The American Mathe-
matical Monthly, 73(5):526–528, 1966.

Émile Rouché. Sur la durée du jeu. Comptes Rendus de l’Academie des Sciences, 106:
253–256, 1888a.

Émile Rouché. Sur un problème relatif à la durée du jeu. Comptes Rendus de l’Academie
des Sciences, 106:47–49, 1888b.

Frank Spitzer. A combinatorial lemma and its applications to probability theory. Transac-
tions of the American Mathematical Society, 82(2):323–339, July 1956.

Charles J. Stone. On local and ratio limit theorems. In Proceedings of the Fifth Berkeley
Symposium on Mathematical Statistics and Probability, pages 217–224, 1965.

Lajos Takács. Ballot problems. Zeitschrift für Warscheinlichkeitstheorie und verwandte


Gebeite, 1:154–158, 1962a.

Lajos Takács. A generalization of the ballot problem and its application in the theory of
queues. Journal of the American Statistical Association, 57(298):327–337, 1962b.

Lajos Takács. The time dependence of a single-server queue with poisson input and general
service times. The Annals of Mathematical Statistics, 33(4):1340–1348, December 1962c.

Lajos Takács. The distribution of majority times in a ballot. Zeitschrift für Warschein-
lichkeitstheorie und verwandte Gebeite, 2(2):118–121, January 1963.

Lajos Takács. Combinatorial methods in the theory of dams. Journal of Applied Probability,
1(1):69–76, 1964a.

Lajos Takács. Fluctuations in the ratio of scores in counting a ballot. Journal of Applied
Probability, 1(2):393–396, 1964b.

22
Lajos Takács. A combinatorial theorem for stochastic processes. Bulletin of the American
Mathematical Society, 71:649–650, 1965a.

Lajos Takács. On the distribution of the supremum for stochastic processes with interchange-
able increments. Transactions of the American Mathematical Society, 119(3):367–379,
September 1965b.

Lajos Takács. Combinatorial Methods in the Theory of Stochastic Processes. John Wiley &
Sons, Inc, New York, NY, first edition, 1967.

Lajos Takács. On the distribution of the maximum of sums of mutually independent and
identically distributed random variables. Advances in Applied Probability, 2(2):344–354,
1970a.

Lajos Takács. On the distribution of the supremum for stochastic processes. Annales de
l’Institut Henri Poincaré B, 6(3):237–247, 1970b.

J.C. Tanner. A derivation of the borel distribution. Biometrika, 48(1-2):222–224, June 1961.

23

You might also like