This action might not be possible to undo. Are you sure you want to continue?

# On convergence rates in the central limit theorems

**for combinatorial structures
**

Hsien-Kuei Hwang

∗

Institute of Statistical Science

Academia Sinica

11529 Taipei

Taiwan

August 15, 1998

Abstract

Flajolet and Soria established several central limit theorems for the parameter “number of

components” in a wide class of combinatorial structures. In this paper, we shall prove a simple

theorem which applies to characterize the convergence rates in their central limit theorems. This

theorem is also applicable to arithmetical functions. Moreover, asymptotic expressions are derived

for moments of integral order. Many examples from diﬀerent applications are discussed.

Key words. Central limit theorem, convergence rate, moments, combinatorial construction, arithmeti-

cal function.

1 Introduction

This paper examines the occurrence of Gaussian laws in large random combinatorial structures. We are

mainly concerned with the rate of normal approximation to discrete random distributions. Examples

that we discuss include such diverse problems as cycles, inversions and involutions in permutations,

the construction of heaps in the ﬁeld of data structures, set partitions, and the problem of “factorisatio

numerorum” in number theory.

Problems in combinatorial enumeration and in number theory often lead to the decomposition of

the underlying structure into more elementary ones by suitable algebraic operations. For instance,

a permutation can be decomposed into a set of cycles, a binary tree can be deﬁned to be either

empty or the union of a root-node and two binary trees, and a factorization of a natural number n

is a decomposition of n into a product of prime factors. It is well-known that operations like union,

sequence, cycle, set and multiset on the structural side correspond to explicitly expressible forms on

∗

This work was partially supported by the ESPRIT Basic Research Action No. 7141 (ALCOM II) while the author

was at LIX, Ecole Polytechnique.

1

the generating function side. Hence it is useful to establish general theorems from which one can

conclude certain asymptotic results (especially, statistical properties) for parameters in the underlying

structure by verifying some basic analytic properties of the generating function. Bender initiated this

line of investigation; see [1, 2, 3, 7, 13, 14, 15, 19].

Applying singularity and probabilistic analysis on bivariate generating functions, Flajolet and

Soria [13, 14] proved a series of central limit theorems for the parameter “number of components” in

a wide class of combinatorial structures issuing principally from combinatorial constructions. Brieﬂy,

they obtain their results by showing that the characteristic function ϕ

n

(t) of the normalized random

variable Ω

∗

n

in question (to be explained below) tends, as n tends to inﬁnity, to e

−

1

2

t

2

; and this

establishes the weak convergence of the distribution function F

n

(x) of Ω

∗

n

to the standard normal law

by Levy’s continuity theorem. With the aid of the Berry-Esseen inequality, we shall make explicit the

convergence rates in their central limit theorems.

In the next section, we propose a simple theorem on convergence rate which turns out to have wide

applications. In particular, it is applicable to the central limit theorems of Flajolet and Soria [13, 14]

which we shall discuss in section 3. Moreover, as a direct consequence of our general assumptions on the

moment generating function, we derive an asymptotic expression for moments of integral order. Then,

we state an eﬀective version of the central limit theorem of Haigh [16] (by establishing the convergence

rate). This theorem is borrowed from [21] where we applied it to derive the asymptotic normality

of the cost of constructing a random heap. Examples from combinatorial enumerations, computer

algorithms, probabilistic number theory, arithmetical semigroups and orthogonal polynomials will be

discussed in the ﬁnal section.

Throughout this paper, we denote by Φ(x) the standard normal distribution:

Φ(x) =

1

√

2π

_

x

−∞

e

−

1

2

t

2

dt (x ∈ R).

All limits (including O, o and ∼), whenever unspeciﬁed, will be taken as n → ∞. All generating

functions (ordinary or exponential) will denote functions analytic at 0 with non-negative coeﬃcients.

The symbol [z

n

]f(z) represents the coeﬃcient of z

n

in the Taylor expansion of f(z).

2 Main results

Let ¦Ω

n

¦

n≥1

be a sequence of random variables. Deﬁne µ

n

= E(Ω

n

), σ

2

n

= Var(Ω

n

) and F

n

(x) =

Pr¦Ω

n

< µ

n

+xσ

n

¦, x ∈ R. If the distribution of Ω

n

is asymptotically normal, then F

n

(x) satisﬁes

sup

x

[F

n

(x) −Φ(x)[ →0, (1)

the pointwise convergence being uniform with respect to x in any ﬁnite interval of R. The ﬁrst

possible reﬁnement to (1) is to determine the convergence rate, which, in most cases, is of order σ

−1

n

.

2

The most general method for establishing the convergence rate (not restricted to normal limiting

distribution) is the Berry-Esseen inequality (see (2 below) which relates the estimate of the diﬀerence

of two distribution functions to that of corresponding characteristic functions, the latter being usually

more manageable, especially when F

n

(x) is a step-function.

Since in probability theory, this subject is almost thoroughly studied when Ω

n

can be decomposed

as a sum of (independent or dependent) random variables, we content here with presenting two simple

theorems which, for most of our applications, turn out to be suﬃcient. Our object is not to search

for results of the most general kind but for those which are easy to apply for combinatorial and

number-theoretic problems, especially, when probability generating function or characteristic function

are available.

The following theorem is motivated by the observation that many asymptotically normal distribu-

tions have mean and variance of the same orders.

Let ¦Ω

n

¦

n≥1

be a sequence of integral random variables.

Suppose that the moment generating function satisﬁes the asymptotic expression:

M

n

(s) := E(e

Ω

n

s

) =

m≥0

Pr¦Ω

n

= m¦e

ms

= e

H

n

(s)

_

1 +O

_

κ

−1

n

__

,

the O-term being uniform for [s[ ≤ τ, s ∈ C, τ > 0, where

(i) H

n

(s) = u(s)φ(n)+v(s), with u(s) and v(s) analytic for [s[ ≤ τ and independent of n; u

(0) ,= 0;

(ii) φ(n) →∞;

(iii) κ

n

→∞.

Theorem 1 Under these assumptions, the distribution of Ω

n

is asymptotically Gaussian:

Pr¦

Ω

n

−u

(0)φ(n)

_

u

(0)φ(n)

< x¦ = Φ(x) +O

_

1

κ

n

+

1

_

φ(n)

_

,

uniformly with respect to x, x ∈ R.

Proof. Let µ

n

= u

(0)φ(n) and σ

2

n

= u

**(0)φ(n). Deﬁne the random variable Ω
**

∗

n

= (Ω

n

−µ

n

)/σ

n

with

distribution function (characteristic function) F

n

(x) (ϕ

n

(t)) respectively.

To obtain an upper bound for the diﬀerence [F

n

(x) −Φ(x)[, we shall estimate the ratio [(ϕ

n

(t) −

e

−t

2

/2

)/t[ and apply the following Berry-Essen inequality [29, p. 109]:

Let F(x) be a non-decreasing function, G(x) a diﬀerentiable function of bounded variation on the

real line, ϕ(t) and γ(t) the corresponding Fourier-Stieltjes transforms:

ϕ(t) =

_

∞

−∞

e

itx

dF(x), γ(t) =

_

∞

−∞

e

itx

dG(x).

Suppose that F(−∞) = G(−∞), F(∞) = G(∞), T is an arbitrary positive number, and [G

(x)[ ≤ A.

Then for every b > 1/(2π) we have

sup

−∞<x<∞

[F(x) −G(x)[ ≤ b

_

T

−T

¸

¸

¸

¸

ϕ(t) −γ(t)

t

¸

¸

¸

¸

dt +r(b)

A

T

, (2)

3

where r(b) is a positive constant depending only on b.

Take G(x) = Φ(x) (so that A = 1/

√

2π) and T = T

n

= cσ

n

, where 0 < c < τ is a suﬃciently small

constant, we shall show that

J

n

=

_

T

n

−T

n

¸

¸

¸

¸

¸

ϕ

n

(t) −e

−

1

2

t

2

t

¸

¸

¸

¸

¸

dt = O

_

σ

−1

n

+κ

−1

n

_

,

so that Theorem 1 follows from the inequality (2).

Let us write M

n

(s) = e

u(s)φ(n)+v(s)

(1 +E

n

(s)), so that E

n

(0) = 0 and E

n

(s) = O

_

κ

−1

n

_

for [s[ ≤ τ.

By taking a small circle around the origin we easily deduce that E

n

(s) = O

_

[s[κ

−1

n

_

for [s[ ≤ c. Since

[ϕ

n

(t)[ ≤ 1, we have

ϕ

n

(t) = exp

_

−

µ

n

σ

n

it +u(

it

σ

n

)φ(n) +v(

it

σ

n

)

__

1 +O

_

[t[

σ

n

κ

n

__

= exp

_

−

t

2

2

+O

_

[t[ +[t[

3

σ

n

__

+O

_

[t[

σ

n

κ

n

_

,

for [t[ ≤ T

n

. Using the inequality [e

w

−1[ ≤ [w[e

|w|

for all complex w, we obtain

¸

¸

¸

¸

¸

ϕ

n

(t) −e

−

1

2

t

2

t

¸

¸

¸

¸

¸

= O

__

1 +t

2

σ

n

_

exp

_

−

t

2

2

+O

_

[t[ +[t[

3

σ

n

__

+

1

σ

n

κ

n

_

= O

__

1 +t

2

σ

n

_

e

−

1

4

t

2

+

1

σ

n

κ

n

_

([t[ ≤ T

n

),

for c suﬃciently small. Hence

J

n

=

_

T

n

−T

n

¸

¸

¸

¸

¸

ϕ

n

(t) −e

−

1

2

t

2

t

¸

¸

¸

¸

¸

dt

= O

_

1

σ

n

_

T

n

−T

n

(1 +t

2

)e

−

1

4

t

2

dt +

1

κ

n

_

= O

_

1

σ

n

+

1

κ

n

_

.

This completes the proof of Theorem 1. 2

The assumption that the moment generating function M

n

(s) exists in the neighbourhood of 0

is not a necessary condition for the validity of Theorem 1. As is evident from the proof, a weaker

suﬃcient condition is that [u

(it)[ and [v

**(it)[ are bounded for s close to 0. But the following theorem
**

does require the analyticity of M

n

(s).

Theorem 2 The k

th

moment of Ω

n

satisﬁes

1

k!

m≥0

m

k

Pr¦Ω

n

= m¦ = π

k

(φ(n)) +O

_

φ(n)

k

κ

n

_

(k = 1, 2, 3, . . .),

where π

k

(X) is a polynomial in X of degree k deﬁne by

π

k

(X) = [s

k

]e

u(s)X+v(s)

(k = 1, 2, 3, . . .). (3)

4

In particular, the mean and the variance of Ω

n

satisfy

E(Ω

n

) = u

(0)φ(n) +v

(0) +O

_

κ

−1

n

_

, (4)

Var(Ω

n

) = u

(0)φ(n) +v

(0) +O

_

κ

−1

n

_

. (5)

Proof. Since M

n

(s) is analytic for [s[ ≤ τ, by Cauchy’s formula, we obtain

m≥0

Pr¦Ω

n

= m¦

m

k

k!

= π

k

(φ(n)) +R

k

(n),

where the remainder term R

n

(n) satisﬁes

R

k

(n) = O

_

1

κ

n

_

|s|=r

[s[

−k−1

¸

¸

¸e

u(s)φ(n)

¸

¸

¸ [ds[

_

(0 < r ≤ τ).

Clearly, the degree of the polynomial π

k

(X) is k. Taking r = φ(n)

−1

, we obtain

R

k

(n) = O

_

φ(n)

k

κ

n

_

,

for k = 1, 2, 3, . . .

For the more reﬁned results (4) and (5), let us consider the cumulant generating function of Ω

n

:

λ

n

(s) := log M

n

(s) = u(s)φ(n) +v(s) +Y

n

(s),

where Y

n

(s) = O

_

κ

−1

n

_

. By the analyticity of M

n

and the relation M

n

(0) = 1, the function λ

n

(s) is

well-deﬁned in a neighbourhood of the origin and is analytic there. Similar arguments as above yield

the required results (4) and (5) since the mean and the variance of Ω

n

are nothing but λ

n

(0) and

λ

n

(0), respectively. This completes the proof. 2

Writing u

j

:= u

(j)

(0) and v

j

:= v

(j)

(0), j = 1, 2, 3, . . ., we have

π

1

(X) = u

1

X +v

1

, π

2

(X) =

u

2

1

2

X

2

+

u

2

+ 2u

1

v

1

2

X +

v

2

+v

2

1

2

.

In general, some of the terms in π

k

(φ(n)) may be absorbed by the remainder term O

_

φ(n)

k

κ

−1

n

_

.

Corollary 1 We have

m≥0

m

k

Pr¦Ω

n

= m¦ = u

(0)

k

φ(n)

k

_

1 +O

k

_

1

φ(n)

+

1

κ

n

__

,

for each ﬁxed k = 1, 2, 3, . . .

3 Combinatorial schemes of Flajolet and Soria

We now apply Theorem 1 to establish eﬀective versions of the central limit theorems considered in

[13, 14]. The developments here are made possible since the singularity analysis provides uniform

bounds on the coeﬃcients of analytic functions.

5

3.1 Exp-log class

The exp-log class covers such problems as cycles in permutations, cycles in random mappings and

random mapping patterns, irreducibles in polynomials over ﬁnite ﬁelds (and, more generally, over an

additive arithmetic semigroup under Knopfmacher’s axiom A

#

), proﬁles of increasing trees, etc. See

[13, 23, 19].

Recall that a generating function C(z) analytic at 0 is called logarithmic [13] if there exists a

constant a > 0, such that for z ∼ ζ, ζ > 0 being the radius of convergence of C,

C(z) = a log

1

1 −z/ζ

+H(z),

where the function H(z) is analytic in the region

∆ := ¦z : [z[ ≤ ζ + and [ arg(z −ζ)[ ≥ δ¦ ( > 0, 0 < δ <

π

2

),

(with possibly a ﬁnite number of other branch cuts starting from the circle [z[ = ζ), and as z → ζ in

∆,

H(z) = K +O

_

log

−

1

2

1

1 −z/ζ

_

,

uniformly in z, where K is some constant. For brevity, we shall refer to as saying that C(z) is

logarithmic with parameters (ζ, a, K).

Remark. The condition that we imposed here on H(z) is weaker than that required in [14]. They need

H(z) = K +o

_

log

−1

1

1 −z/ζ

_

(z →ζ, z ∈ ∆),

to ensure that both the mean and the variance are a log n + O(1). But by our Theorem 2, only the

weaker condition H(z) = K + o(1) for z ∈ ζ, z ∈ ∆, suﬃces to guarantee the later. This was the

original condition used in [13]. 2

We ﬁrst consider generating functions of the form

P(w, z) =

n,m≥0

p

nm

w

m

z

n

= exp (wC(z) +S(w, z)) (C(0) = 0),

where C(z) is logarithmic with parameters (ζ, a, K) and the function S(w, z) is analytic for [z[ < ζ +

and [w[ < 1 +

, ,

**> 0. By singularity analysis [14, 12] (see [19] for details), we obtain
**

M

n

(s) =

m≥0

Pr¦Ω

n

= m¦e

ms

:=

[z

n

]P(e

s

, z)

[z

n

]P(1, z)

= exp

_

(e

s

−1)a log n +K(e

s

−1) +S(e

s

, ζ) −S(1, ζ) + log

Γ(a)

Γ(ae

s

)

_

_

1 +O

_

1

√

log n

__

, (6)

uniformly for s in some complex neighbourhood of the origin.

6

Theorem 3 Let F

n

(x) = Pr¦Ω

n

< a log n + x

√

a log n¦, where Ω

n

and a are deﬁned by (6). Then

F

n

(x) satisﬁes, uniformly with respect to x,

F

n

(x) = Φ(x) +O

_

1

√

log n

_

.

Proof. The result follows from Theorem 1, since 1/Γ(ae

s

) ,= 0 for ['(s)[ < [ log a[. 2

Theorem 4 The k

th

moment of Ω

n

, k ∈ N

+

, satisﬁes

1

k!

m≥0

m

k

Pr¦Ω

n

= m¦ =

(a log n)

k

k!

_

1 +O

_

(log n)

−1/2

__

,

In particular, the mean and the variance of Ω

n

satisfy

E(Ω

n

) = a log n −aψ(a) +K +S

1

+O

_

(log n)

−

1

2

_

Var(Ω

n

) = a log n −aψ(a) −a

2

ψ

(a) +K +S

2

+S

1

+O

_

(log n)

−

1

2

_

where ψ denotes the logarithmic derivative of the Gamma function and the two constants S

1

and S

2

are deﬁned by

S

1

=

∂

∂w

S(w, ζ)

¸

¸

¸

¸

w = 1

, S

2

=

∂

2

∂w

2

S(w, ζ)

¸

¸

¸

¸

w = 1

.

Proof. This follows from (6) and Theorem 2. 2

The above theorems apply to the following four classes of combinatorial structures when C(z) is

logarithmic:

1. Partitional complex construction: P(w, z) = exp(wC(z));

2. Partitional complex construction in which no two components are order-isomorphic:

P(w, z) =

k≥1

_

1 +

wz

k

k!

_

c

k

(c

k

:= k![z

k

]C(z));

3. Multiset construction: (i) total number of components (i.e., counted with multiplicity)

P(w, z) = exp

_

_

k≥1

w

k

k

C(z

k

)

_

_

;

(ii) number of distinct components (counted without multiplicity)

P(w, z) =

k≥1

_

1 +

wz

k

1 −z

k

_

c

k

(c

k

:= [z

k

]C(z)),

when the dominant singularity of C is strictly inferior to 1;

4. Set construction:

P(w, z) = exp

_

_

k≥1

(−1)

k−1

k

w

k

C(z

k

)

_

_

,

when the dominant singularity of C is strictly less than 1.

Further reﬁnement of Theorem 1 is possible under slightly stronger conditions on H(z). This will

be discussed in the companion paper [20]; cf. also [19, Ch. 2].

7

3.2 Algebraic-logarithmic class

Next, let us consider generating functions of the form

P(w, z) =

n,m≥0

p

nm

w

m

z

n

=

1

(1 −wC(z))

α

_

log

1

1 −wC(z)

_

β

(C(0) = 0),

where β ∈ N, α ≥ 0, and α + β > 0. Deﬁne the random variable Ω

n

by its moment generating

function:

M

n

(s) =

m≥0

Pr¦Ω

n

= m¦ e

ms

=

[z

n

]P(e

s

, z)

[z

n

]P(1, z)

,

for suﬃciently large n.

Deﬁnition. (1-regular function [14]) A generating function C(z) ,≡ z

q

(q = 0, 1, 2, . . .) analytic at

z = 0 is called 1-regular if (i) its Taylor expansion at z = 0 involves only non-negative coeﬃcients; (ii)

there exists a positive number ρ < , being the radius of convergence of C(z), such that C(ρ) = 1.

Assume, without loss of generality, that C(z) is aperiodic, namely, C(z) ,≡ z

a

D(z

b

) for some generating

function D(z) and integers a, b ≥ 2.

To start with, let us ﬁrst suppose that C(z) is 1-regular and consider the equation

1

w

= C(z).

Since C(z) is 1-regular, we have C

**(ρ) > 0, where ρ satisﬁes C(ρ) = 1. By the inverse function
**

theorem, the above equation is locally invertible, viz, the equation has a unique solution ρ(w) for

w ∼ 1 and ρ(w) is analytic at w = 1. Of course ρ(1) = ρ.

The following lemma is fundamental to further investigations.

Lemma 1 Let P

n

(w) := [z

n

]P(w, z). The formula

P

n

(w) = A

α

(w)ρ(w)

−n

n

α−1

(log n)

β

0≤k≤β

_

β

k

_

(log n)

−k

0≤j≤k

g

j

(α) (log A(w))

k−j

+

+O

_

[ρ(w)[

−n

n

α−2

(log n)

β

_

,

holds uniformly for w ∼ 1, where

A(w) =

1

wρ(w)C

(ρ(w))

, g

j

(α) =

d

j

dx

j

1

Γ(x)

¸

¸

¸

¸

¸

x = α

.

Proof. Let us ﬁx a w, w ∼ 1. Expanding C(z) at z = ρ(w), we get

C(z) =

1

w

−ρ(w)C

(ρ(w))

_

1 −

z

ρ(w)

_

+

ρ

2

(w)

2

C

(ρ(w))

_

1 −

z

ρ(w)

_

2

+ .

Thus

1

1 −wC(z)

= A(w)

_

1 −

z

ρ(w)

_

−1

_

1 +A

1

(w)

_

1 −

z

ρ(w)

_

+A

2

(w)

_

1 −

z

ρ(w)

_

2

+

_

,

8

where

A

1

(w) =

ρ(w)C

(ρ(w))

2C

(ρ(w))

, A

2

(w) = ρ

2

(w)

_

C

(ρ(w))

2

4C

(ρ(w))

2

−

C

(ρ(w))

6C

(ρ(w))

_

.

For P(w, z), we get

_

1

1 −wC(z)

_

α

_

log

1

1 −wC(z)

_

β

= A

α

(w)

0≤j≤β

_

β

j

_

(log A(w))

β−j

_

1 −

z

ρ(w)

_

−α

_

log

1

1 −z/ρ(w)

_

j

+

+O

_

_

1 −

z

ρ(w)

_

−α+1

_

log

1

1 −z/ρ(w)

_

β

_

.

By the remark before the lemma, we can apply the singularity analysis [12]. Thus transferring to

coeﬃcients, we obtain

P

n

(w) = A

α

(w)

0≤j≤β

_

β

j

_

(log A(w))

β−j

0≤k≤j

g

j−k

(α)(log n)

k

+ O

_

[ρ(w)[

−n

n

α−2

(log n)

β

_

.

The double sum can be rearranged in descending powers of log n:

0≤j≤β

_

β

j

_

(log A(w))

β−j

0≤k≤j

g

j−k

(α)(log n)

k

= (log n)

β

0≤k≤β

_

β

k

_

(log n)

−k

0≤j≤k

g

j

(α) (log A(w))

k−j

.

This completes the proof. 2

In particular, when β = 0 and α > 0, we obtain

P

n

(w) = A

α

(w)ρ(w)

−n

n

α−1

_

1 +O

_

n

−1

__

;

when α = 0, β = 1,

P

n

(w) = ρ(w)

−n

n

−1

_

1 +O

_

n

−1

__

;

and when α = 0 and β > 0,

P

n

(w) = ρ(w)

−n

n

−1

(log n)

β

1≤k≤β

_

β

k

_

(log n)

−k

1≤j≤k

g

j

(0) (log A(w))

k−j

+

+O

_

[ρ(w)[

−n

n

−2

(log n)

β

_

.

Note that the indices of the last two sums both begin at 1 since g

0

(0) = 1/Γ(0) = 0.

Remark. From a computational point of view, the following expressions for g

k

(0) are useful.

g

k

(0) = k![x

k

]

1

Γ(x)

= k!(−1)

k−1

[x

k−1

]

1

Γ(1 −x)

9

= k!(−1)

k−1

[x

k−1

]

Γ(x) sin πx

π

= k(−1)

k−1

0≤j≤

k−2

2

(−1)

j

π

2j

_

k −1

2j + 1

_

Γ

(k−2−2j)

(0)

=

0≤j≤

k−1

2

(−1)

k−1+j

π

2j

_

k

2j + 1

_

Γ

(k−1−2j)

(1).

Thus, with M

n

(s) := P

n

(e

s

)/P

n

(1), we have, by Lemma 1,

M

n

(s) =

_

e

s

ρ(e

s

)C

(ρ(e

s

))

ρC

(ρ)

_

−α

ρ(e

s

)

−n

ρ

−n

(1 +ε

α,β

(n)) , (7)

where

ε

α,β

(n) =

_

O

_

n

−1

_

, if (α > 0 and β = 0) or (α = 0 and β = 1);

O

_

(log n)

−1

_

, if (α = 0 and β ≥ 2) or (α > 0 and β > 0),

uniformly for [s[ ≤ τ, τ > 0.

Deﬁne two constants

α

1

=

1

ρC

(ρ)

, and α

2

=

1

ρ

2

C

(ρ)

2

+

C

(ρ)

ρC

(ρ)

3

−

1

ρC

(ρ)

.

Theorem 5 Let F

n

(x) = Pr¦Ω

n

< α

1

n +x

√

α

2

n¦. Then F

n

(x) satisﬁes

F

n

(x) = Φ(x) +O

_

1

√

n

+ε

α,β

(n)

_

,

uniformly in x, x ∈ R.

Proof. Using (7) and Applying Theorem 1. 2

Theorem 6 The k

th

moment of Ω

n

satisﬁes

1

k!

m≥0

m

k

Pr¦Ω

n

= m¦ = Π

k

(n) +O(n

k

ε

α,β

(n)),

where

k

(n) is a polynomial in n of degree k deﬁned by

Π

k

(n) := [s

k

] exp

_

−nlog

ρ(e

s

)

ρ

__

e

s

ρ(e

s

)C

(ρ(e

s

))

ρC

(ρ)

_

−α

,

for k = 1, 2, 3, . . . In particular, we have

E(Ω

n

) = α

1

n +αL

1

+O(ε

α,β

(n))

Var(Ω

n

) = α

2

n +αL

2

+O(ε

α,β

(n)) ,

where the two constants L

1

, L

2

are given by

L

1

:=

1

ρC

(ρ)

+

C

(ρ)

C

(ρ)

2

−1

L

2

:=

2αρC

1

C

2

+αρ

2

C

2

2

−ρ

2

C

2

1

C

2

−ρC

3

1

+ρC

1

C

2

−2αρC

3

1

+2ρ

2

C

2

2

−ρ

2

C

1

C

3

+αρ

2

C

4

1

−2αρ

2

C

2

1

C

2

+αC

2

1

+C

2

1

2ρ

2

C

4

1

,

(C

j

:= C

(j)

(ρ) for j = 1, 2, 3.)

10

Proof. This follows from (7) and Theorem 2. 2

These results apply to sequence constructions of both labelled and unlabelled structures

P(w, z) =

1

1 −wC(z)

(C(0) = 0);

and to cycle constructions

P(w, z) = log

1

1 −wC(z)

, P(w, z) =

k≥1

ϕ(k)

k

log

1

1 −w

k

C(z

k

)

,

provided that the radius of convergence of C(z) in the last case is less than 1, where ϕ(k) denotes

Euler’s totient function. It should be noted that for sequence constructions, we often have exponential

convergence rate for M

n

(s).

Remark. For Bender’s central limit theorem [1, Thm 1], the convergence rate is O

_

n

−

1

2

_

, provided

that the function A(s) in that theorem has bounded derivative for s near 0.

4 A classical central limit theorem

The following theorem is borrowed from [21] where we used it to derive the asymptotic normality of

the cost of constructing a random heap. Since the proof has already been given there, it is omitted

here. Without the explicit error term it is due to Haight [16].

The object of study here is sums of independent but not identically distributed random variables.

This problem has been extensively studied in the probability literature; see, for instance, [6, ¸27] [25,

Ch. VI] [17, 29].

Theorem 7 Let ¦Ω

n

¦

n

be a sequence of random variables taking only non-negative integral values

with mean µ

n

and variance σ

2

n

. Suppose that the probability generating function P

n

(z) of Ω

n

can be

decomposed as

P

n

(z) =

1≤j≤k

n

P

nj

(z),

for some sequence ¦k

n

¦

n

, where the P

nj

(z) are polynomials such that (i) each P

nj

(z) is itself a prob-

ability generating function of some random variable, say, Ω

nj

(1 ≤ j ≤ k

n

); and (ii)

M

n

σ

n

→0,

where M

n

= max

1≤j≤k

n

deg P

nj

(z). Then the distribution of Ω

n

is asymptotically Gaussian:

Pr¦

Ω

n

−µ

n

σ

n

< x¦ = Φ(x) +O

_

M

n

σ

n

_

,

uniformly with respect to x.

Proof. See [21].

11

Corollary 2 Let P

n

(z) be a probability generating function of the random variable Ω

n

, n ≥ 1. Suppose

that, for each ﬁxed n ≥ 1, P

n

(z) is a polynomial whose roots are all located in and on the half-plane

'z ≤ 0. If σ

n

→∞, σ

2

n

:= Var(Ω

n

), then Ω

n

satisﬁes

Pr¦

Ω

n

−µ

n

σ

n

< x¦ = Φ(x) +O

_

1

σ

n

_

,

for −∞< x < ∞, where µ

n

= E(Ω

n

).

Proof. M = 2 in Theorem 7. 2

In particular, if all roots of P

n

(z) are non-positive, then the results of the corollary hold, cf. [18].

5 Examples

Let us discuss some typical examples from diﬀerent applications.

Example 1. Cycles in permutations. A permutation is a set of cycles. The number of permutations

having m cycles is enumerated by the (signless) Stirling number of the ﬁrst kind s(n, m):

s(n, m) := [w

m

z

n

] exp

_

wlog

1

1 −z

_

= [w

m

] (w(w + 1) (w +n −1)) (1 ≤ m ≤ n, n ≥ 1).

Both Theorems 3 and 7 apply to the random variable Ω

n

, deﬁned as the number of cycles in a random

permutation of ¦1, 2, . . . , n¦, where each of the n! permutations is equally likely. We obtain

Pr¦

Ω

n

−log n

√

log n

< x¦ = Φ(x) +O

_

1

√

log n

_

,

a result originally due to Goncharov (without error term).

Many other examples in this same (exp-log) class can be found in [5, 13, 23, 33] [26, ¸2.4 and 2.5]

and [19, Ch. 5].

Example 2. Stirling numbers of the second kind. The generating function for the number of blocks

(marked by w) in set partitions satisﬁes

n,m≥0

S(n, m) w

m

z

n

n!

= e

w(e

z

−1)

,

where S(n, m) denotes Stirling numbers of the second kind. Theorem 1 is not applicable since, as we

shall see, the mean and the variance are not of the same order. So we turn to Theorem 7. Harper [18]

showed that the polynomials

P

n

(z) :=

0≤m≤n

S(n, m) z

m

(n = 1, 2, 3, . . .),

have only simple non-positive roots. It remains to show that σ

n

→∞, namely,

σ

2

n

:= Var(Ω

n

) =

b

n+2

b

n

−

b

n+1

b

n

−1 →∞,

12

where b

n

:= n! [z

n

]e

e

z

−1

are the Bell numbers. Using the asymptotic expression [9, Ch. 6]

b

n

=

e

n(ρ+ρ

−1

−1)−1

√

ρ + 1

_

1 −

ρ

2

(2ρ

2

+ 7ρ + 10)

24 n(ρ + 1)

3

+O

_

ρ

2

n

2

__

,

where ρe

ρ

= n, it is not hard to show that

σ

2

n

=

n

ρ(ρ + 1)

+O(1) ∼

n

(log n)

2

→∞;

so that Theorem 7 gives, deﬁning Pr¦Ω

n

= m¦ := b

−1

n

S(n, m),

Pr¦

Ω

n

−µ

n

σ

n

< x¦ = Φ(x) +O

_

log n

√

n

_

,

uniformly with respect to x, where

µ

n

= E(Ω

n

) =

b

n+1

b

n

−1 =

n

ρ

+O(1) ∼

n

log n

.

The asymptotic normality is due to Harper [18]. Alternatively, we can write

Pr¦

Ω

n

−n/ρ

_

n/(ρ(ρ + 1))

< x¦ = Φ(x) +O

_

log n

√

n

_

,

but this formula cannot be further simpliﬁed by using the asymptotics of ρ.

For other examples, see [31, 28].

Example 3. Cost of heap construction. Doberkat [11] shows that the probability generating functions

for the number of exchanges and for the number of comparisons used by Floyd’s algorithm to construct

a heap of size n have the form

1≤j≤n

p

n,j

(z), where p

n,j

(z) are polynomials of degree ≤ log

2

n.

Moreover, the mean and the variance are all linear. Thus Theorem 7 applies with M

n

= log n to

both cases [21]. More precisely, given a random permutation of size n, deﬁning Ω

n

as the number of

exchanges used by Floyd’s algorithm to construct a heap from this permutation, we have [21, Thm 4]

Pr¦

Ω

n

−c

1

n

√

c

2

n

< x¦ = Φ(x) +O

_

log n

√

n

_

,

uniformly with respect to x, where c

1

= −2+

j≥1

j(2

j

−1)

−1

= 0.744033... and c

2

= 2−

j≥1

j

2

(2

j

−

1)

2

= 0.261217... Similar result holds for the number of comparisons and other construction algorithms

preserving randomness in each step, cf. [27].

For other examples on algorithms, see [19, Ch. 1].

Example 4. Ordered set-partitions. The generating function for the number of blocks (marked by w)

in an ordered set-partition is (1 −w(e

z

−1))

−1

. Obviously, e

z

−1 is 1-regular. Theorem 5 applies. We

can also consider other constraints like parity on the size of each block and on the number of blocks,

cf. [8, p. 225–226]. Many other examples can be found in [1, 14, 22] and [19, Ch. 7].

13

Example 5. Inversions in permutations. Given a permutation π = (π

1

, π

2

, . . . , π

n

) of ¦1, 2, 3, . . . , n¦,

a pair π

i

, π

j

is called an inversion if i < j and π

i

> π

j

. Let Ω

n

denote the number of inversions in a

random permutation of size n. Then it is easily seen that [24, ¸5.1.1]

Pr¦Ω

n

= m¦ =

1

n!

[z

m

]

_

(1 +z)(1 +z +z

2

) (1 +z +z

2

+ +z

n−1

)

_

.

From which we obtain

E(Ω

n

) =

n(n −1)

4

, and Var(Ω

n

) =

n(n −1)(2n + 5)

72

, (8)

for n ≥ 1. Theorem 7 applies and we get, in view of (8),

Pr¦

Ω

n

−n

2

/4

n

3/2

/6

< x¦ = Φ(x) +O

_

1

√

n

_

,

for all x.

Example 6. Erd˝os-Kac Theorem. Deﬁne a random variable ξ

n

by its probability distribution:

Pr¦ξ

n

= m¦ =

1

n

[¦k : 1 ≤ k ≤ n, ω(k) = m¦[ =

1

n

[z

m

]

1≤k≤n

z

ω(k)

,

where ω(n) denotes the number of distinct prime divisors of n, e.g., ω(23

4

5

6

7

8910

11

12

13

141516

) = 6.

It is known that E(ξ

n

) = log log n + O(1), and Var(ξ

n

) = log log n + O(1). The Erd˝ os-Kac Theorem

states that

Pr¦

ξ

n

−log log n

√

log log n

< x¦ →Φ(x).

The convergence rate was ﬁrst conjectured by Le Veque and then proved by R´enyi and Tur´ an [30] to

be O((log log n)

−

1

2

). Let us apply Theorem 1 to establish this fact.

Starting with the well-known Selberg’s formula [32]

M

n

(z) := E(e

ξ

n

z

) = V (e

z

)(log n)

e

z

−1

_

1 +O

_

1

log n

__

,

uniformly for [·z[ ≤ π, where V (t) is an entire function in the t-plane, we readily obtain

Pr¦

ξ

n

−log log n

√

log log n

< x¦ = Φ(x) +O

_

1

√

log log n

_

,

uniformly with respect to x.

For many other examples, see [10] and [19, Ch. 9].

Example 7. Number of factors in square-free factorizations. Let f

n,k

denote the number of factor-

izations of n into k distinct factors greater than 1, n ≥ 2, k ≥ 1. Deﬁne f

1,k

= δ

0,k

, where δ

a,b

is

Kronecker’s symbol. Consider the bivariate generating function:

P(w, s) := 1 +

n≥2

n

−s

m≥1

f

n,m

w

m

,

14

which equals to

P(w, s) =

n≥2

_

1 +

w

n

s

_

('s > 1, w ∈ C).

Deﬁne a polynomial P

n

(w) := 1 +

2≤k≤n

m≥1

f

k,m

w

m

and a random variable ξ

n

for which

Pr¦ξ

n

= m¦ =

[w

m

] P

n

(w)

P

n

(1)

.

We show in [19, Ch. 10] that

P

n

(w) =

ne

2

√

wlog n

2

√

π(log n)

3/4

w

3/4

(1 +w)Γ(w)

_

1 +O

_

1

√

log n

__

,

the error term being uniform with respect to w ∈ C¸ (−∞, 0]. Thus we obtain

M

n

(z) := E(e

ξ

n

z

) =

2 e

2

√

log n(e

z/2

−1)−3z/4

Γ(e

z

)(1 +e

z

)

_

1 +O

_

1

√

log n

__

,

and Theorem 1 gives

Pr¦

ξ

n

−

√

log n

4

_

1

4

log n

< x¦ = Φ(x) +O

_

1

4

√

log n

_

,

for all x.

For other examples, see [19, Ch. 10].

Example 8. Hermite polynomials. The number of ﬁxed points (cycles of length 1), marked by w, in

the involutions is enumerated by

exp

_

wz +

z

2

2

_

=

n≥0

h

n

(w)

n!

z

n

,

where h

n

(w) are Hermite polynomials. The asymptotic behaviour of h

n

(w) is well-known [34, p. 200]:

h

n

(w) = 2

−

1

2

n

n/2

e

−n/2+

√

n+1 w−w

2

/4

_

1 +O

_

1

√

n

__

,

uniformly for all ﬁnite w. Deﬁning Ω

n

as the number of ﬁxed points in a random involution, we obtain

M

n

(s) = E(e

Ω

n

s

) = e

√

n+1 (e

s

−1)−

1

4

(e

2s

−1)

_

1 +O

_

1

√

n

__

,

uniformly for [·s[ ≤ π. We obtain

Pr¦

Ω

n

−

√

n + 1

4

√

n + 1

< x¦ = Φ(x) +O

_

1

4

√

n

_

,

the O being uniform with respect to x, x ∈ R.

15

References

[1] E. A. Bender. Central and local limit theorems applied to asymptotic enumeration. Journal of

Combinatorial Theory, Series A, 15:91–111, 1973.

[2] E. A. Bender and L. B. Richmond. Central and local limit theorems applied to asymptotic

enumeration. II. multivariate generating functions. Journal of Combinatorial Theory, Series A,

34:255–265, 1983.

[3] E. A. Bender, L. B. Richmond and S. G. Williamson. Central and local limit theorems applied

to asymptotic enumeration. III. matrix recursions. Journal of Combinatorial Theory, Series A,

35:263–278, 1983.

[4] E. A. Bender and L. B. Richmond. A generalisation of Canﬁeld’s formula. Journal of Combina-

torial Theory, Series A, 41:50–60, 1986.

[5] F. Bergeron, P. Flajolet, and B. Salvy. Varieties of increasing trees. In Proceedings CAAP’92,

Lecture Notes in Computer Science, 581, pages 24–48, Springer-Verlag, New York, 1992.

[6] P. Billingsley, Probability and Measure, Second edition, Wiley, New York, 1986.

[7] E. R. Canﬁeld. Central and local limit theorems for the coeﬃcients of polynomials of binomial

type. Journal of Combinatorial Theory, Series A, 23:275–290, 1977.

[8] L. Comtet. Advanced Combinatorics, The Art of Finite and Inﬁnite Expansions. D. Reidel

Publishing Company, Dordrecht-Holland, revised and enlarged edition, 1974.

[9] N. G. de Bruijn. Asymptotic Methods in Analysis. North-Holland, 1958. Reprinted by Dover,

1981.

[10] H. Delange. Sur des formules de Atle Selberg. Acta Arithmetica, 19:105–146, 1971.

[11] E. E. Doberkat. An average case analysis of Floyd’s algorithm to construct heaps. Information

and Control, 61:114–131, 1984.

[12] P. Flajolet and A. M. Odlyzko. Singularity analysis of generating functions. SIAM Journal on

Discrete Mathematics, 3:216–240, 1990.

[13] P. Flajolet and M. Soria. Gaussian limiting distributions for the number of components in com-

binatorial structures. Journal of Combinatorial Theory, Series A, 53:165–182, 1990.

[14] P. Flajolet and M. Soria. General combinatorial schemes: Gaussian limit distributions and expo-

nential tails. Discrete Mathematics, 114:159–180, 1993.

16

[15] Z. Gao and L. B. Richmond. Central and local limit theorems applied to asymptotic enumera-

tion IV: multivariate generating functions. Journal of Computational and Applied Mathematics,

41:177–186, 1992.

[16] J. Haigh. A neat way to prove asymptotic normality. Biometrika, 58:677–678, 1971.

[17] P. Hall, Rates of Convergence in the Central Limit Theorem, Research Notes in Mathematics, 62,

Pitman Advanced Publlishing Company, Boston, 1982.

[18] L. H. Harper. Stirling behaviour is asymptotically normal. Annals of Mathematical Statistics,

38:410–414, 1967.

[19] H.-K. Hwang. Th´eor`emes limites pour les structures combinatoires et les fonctions arithm´etiques,

Th`ese, Ecole polytechnique, 1994.

[20] H.-K. Hwang. On asymptotic expansions in the central and local limit theorems for combinatorial

structures. submitted.

[21] H.-K. Hwang and J.-M. Steyaert. On the number of heaps and the cost of heap construction.

Theoretical Computer Science, to appear.

[22] A. Knopfmacher, J. Knopfmacher, and R. Warlimont. “Factorisatio numerorum” in arithmetical

semigroups. Acta Arithmetica, 61:327–336, 1992.

[23] J. Knopfmacher. Analytic Arithmetic of Algebraic Function Fields, volume 50 of Lecture Notes

in Pure and Applied Mathematics. Marcel Dekker, Inc., New York and Basel, 1979.

[24] D. E. Knuth. The Art of Computer Programming, Volume 3: Sorting and Searching. Addison

Wesley, Reading, Massachusetts, 1973.

[25] M. Lo`eve, Probability Theory, Forth edition, Springer-Verlag, New York, 1977.

[26] H. M. Mahmoud. Evolution of Random Search Trees. John Wiley & Sons, New York, 1992.

[27] C. J. H. McDiamid and B. A. Reed. Building heaps fast. Journal of Algorithms, 10:352–365,

1989.

[28] L. Mutafˇciev. Limit distributions related to the weighted Stirling numbers of the ﬁrst and second

kind. Collection: Limit Theorems in Probability and Statistics. Colloquia Mathematica Societatis

J´anos Bolyai, 36:823–842, 1982.

[29] V. V. Petrov. Sums of Independent Random Variables. Springer-Verlag, Berlin-Heidelberg, New-

York, 1975. Translated from the Russian by A. A. Brown.

17

[30] A. R´enyi and P. Tur´ an. On a theorem of Erd˝ os-Kac. Acta Arithmetica, 4:71–84, 1958.

[31] A. Ruci´ nski and B. Voigt. A local limit theorem for generalized Stirling numbers. Revue Roumaine

de Math´ematiques Pures et Appliqu´ees, 35:161–172, 1990.

[32] A. Selberg. Note on a paper by L. G. Sathe. Journal of the Indian Mathematical Society, 18:83–87,

1954.

[33] M. Soria. M´ethode d’Analyse pour les Constructions Combinatoires et les Algorithmes. Th`ese

d’

´

Etat, Universit´e de Paris-Sud, 1990.

[34] G. Szeg¨o. Orthogonal Polynomials, volume XXIII of American Mathematical Society Colloquium

Publications. AMS, Providence, Rhode-Island, forth edition, 1977.

18

Hence it is useful to establish general theorems from which one can conclude certain asymptotic results (especially. we derive an asymptotic expression for moments of integral order. we state an eﬀective version of the central limit theorem of Haigh [16] (by establishing the convergence rate). With the aid of the Berry-Esseen inequality. o and ∼). All limits (including O. Flajolet and Soria [13. Deﬁne µn = E(Ωn ). 3. Brieﬂy. Then. as a direct consequence of our general assumptions on the moment generating function. which. as n tends to inﬁnity. statistical properties) for parameters in the underlying structure by verifying some basic analytic properties of the generating function. All generating functions (ordinary or exponential) will denote functions analytic at 0 with non-negative coeﬃcients. it is applicable to the central limit theorems of Flajolet and Soria [13. 2 Main results 2 Let {Ωn }n≥1 be a sequence of random variables.the generating function side. arithmetical semigroups and orthogonal polynomials will be discussed in the ﬁnal section. we shall make explicit the convergence rates in their central limit theorems. will be taken as n → ∞. If the distribution of Ωn is asymptotically normal. to e− 2 t . 15. 14] proved a series of central limit theorems for the parameter “number of components” in a wide class of combinatorial structures issuing principally from combinatorial constructions. we denote by Φ(x) the standard normal distribution: 1 Φ(x) = √ 2π x −∞ 1 2 e− 2 t dt 1 2 (x ∈ R). in most cases. 19]. σn = Var(Ωn ) and Fn (x) = Pr{Ωn < µn + xσn }. 13. computer algorithms. x (1) the pointwise convergence being uniform with respect to x in any ﬁnite interval of R. In the next section. Applying singularity and probabilistic analysis on bivariate generating functions. The symbol [z n ]f (z) represents the coeﬃcient of z n in the Taylor expansion of f (z). The ﬁrst −1 possible reﬁnement to (1) is to determine the convergence rate. In particular. 14. Bender initiated this line of investigation. Examples from combinatorial enumerations. they obtain their results by showing that the characteristic function ϕn (t) of the normalized random variable Ω∗ in question (to be explained below) tends. 2 . Moreover. and this n establishes the weak convergence of the distribution function Fn (x) of Ω∗ to the standard normal law n by Levy’s continuity theorem. whenever unspeciﬁed. see [1. 7. 14] which we shall discuss in section 3. we propose a simple theorem on convergence rate which turns out to have wide applications. Throughout this paper. x ∈ R. 2. This theorem is borrowed from [21] where we applied it to derive the asymptotic normality of the cost of constructing a random heap. then Fn (x) satisﬁes sup |Fn (x) − Φ(x)| → 0. is of order σn . probabilistic number theory.

Our object is not to search for results of the most general kind but for those which are easy to apply for combinatorial and number-theoretic problems. where (i) Hn (s) = u(s)φ(n)+v(s). Suppose that F (−∞) = G(−∞). u (0) = 0. turn out to be suﬃcient. (ii) φ(n) → ∞. T is an arbitrary positive number. the distribution of Ωn is asymptotically Gaussian: Pr{ Ωn − u (0)φ(n) < x} = Φ(x) + O u (0)φ(n) 1 + κn 1 . F (∞) = G(∞). t T (2) 3 . p. Suppose that the moment generating function satisﬁes the asymptotic expression: Mn (s) := E(eΩn s ) = m≥0 Pr{Ωn = m}ems = eHn (s) 1 + O κ−1 n . the O-term being uniform for |s| ≤ τ . ∞ γ(t) = −∞ eitx dG(x). 2 Proof. this subject is almost thoroughly studied when Ωn can be decomposed as a sum of (independent or dependent) random variables. when probability generating function or characteristic function are available. (iii) κn → ∞. To obtain an upper bound for the diﬀerence |Fn (x) − Φ(x)|. the latter being usually more manageable. τ > 0. s ∈ C. G(x) a diﬀerentiable function of bounded variation on the real line. The following theorem is motivated by the observation that many asymptotically normal distributions have mean and variance of the same orders. and |G (x)| ≤ A. φ(n) uniformly with respect to x.The most general method for establishing the convergence rate (not restricted to normal limiting distribution) is the Berry-Esseen inequality (see (2 below) which relates the estimate of the diﬀerence of two distribution functions to that of corresponding characteristic functions. with u(s) and v(s) analytic for |s| ≤ τ and independent of n. 109]: Let F (x) be a non-decreasing function. Theorem 1 Under these assumptions. especially. Then for every b > 1/(2π) we have T sup −∞<x<∞ |F (x) − G(x)| ≤ b −T ϕ(t) − γ(t) A dt + r(b) . ϕ(t) and γ(t) the corresponding Fourier-Stieltjes transforms: ∞ ϕ(t) = −∞ eitx dF (x). we content here with presenting two simple theorems which. Since in probability theory. Let µn = u (0)φ(n) and σn = u (0)φ(n). especially when Fn (x) is a step-function. we shall estimate the ratio |(ϕn (t) − e −t2 /2 )/t| and apply the following Berry-Essen inequality [29. x ∈ R. for most of our applications. Deﬁne the random variable Ω∗ = (Ωn − µn )/σn with n distribution function (characteristic function) Fn (x) (ϕn (t)) respectively. Let {Ωn }n≥1 be a sequence of integral random variables.

. . 3. . where 0 < c < τ is a suﬃciently small constant. . a weaker suﬃcient condition is that |u (it)| and |v (it)| are bounded for s close to 0. . (3) φ(n)k κn (k = 1. 2 σn σ n κn for |t| ≤ Tn .). Since n |ϕn (t)| ≤ 1. we obtain ϕn (t) − e− 2 t t 1 2 = O = O 1 + t2 σn 1 + t2 σn exp − 1 2 t2 +O 2 1 σ n κn |t| + |t|3 σn + 1 σ n κn e− 4 t + (|t| ≤ Tn ). . 2. 2. But the following theorem does require the analyticity of Mn (s). n By taking a small circle around the origin we easily deduce that En (s) = O |s|κ−1 for |s| ≤ c. . we shall show that Tn Jn = −Tn ϕn (t) − e− 2 t t 1 2 −1 dt = O σn + κ−1 . so that En (0) = 0 and En (s) = O κ−1 for |s| ≤ τ .). 2 The assumption that the moment generating function Mn (s) exists in the neighbourhood of 0 is not a necessary condition for the validity of Theorem 1. As is evident from the proof. Using the inequality |ew − 1| ≤ |w|e|w| for all complex w. σ n κn = O This completes the proof of Theorem 1.where r(b) is a positive constant depending only on b. n so that Theorem 1 follows from the inequality (2). Hence Tn Jn = −Tn ϕn (t) − e− 2 t t 1 2 dt 1 2 1 Tn 1 (1 + t2 )e− 4 t dt + σn −Tn κn 1 1 = O + . Let us write Mn (s) = eu(s)φ(n)+v(s) (1 + En (s)). 3. Theorem 2 The k th moment of Ωn satisﬁes 1 mk Pr{Ωn = m} = πk (φ(n)) + O k! m≥0 where πk (X) is a polynomial in X of degree k deﬁne by πk (X) = [sk ]eu(s)X+v(s) 4 (k = 1. √ Take G(x) = Φ(x) (so that A = 1/ 2π) and T = Tn = cσn . we have ϕn (t) = exp − µn it it |t| it + u( )φ(n) + v( ) 1+O σn σn σn σ n κn 2 3 t |t| + |t| |t| = exp − + O +O . for c suﬃciently small.

3. let us consider the cumulant generating function of Ωn : λn (s) := log Mn (s) = u(s)φ(n) + v(s) + Yn (s). the function λn (s) is n well-deﬁned in a neighbourhood of the origin and is analytic there. where Yn (s) = O κ−1 . . some of the terms in πk (φ(n)) may be absorbed by the remainder term O φ(n)k κ−1 . . respectively. n Var(Ωn ) = u (0)φ(n) + v (0) + O κ−1 . |s|=r Clearly. 3 Combinatorial schemes of Flajolet and Soria We now apply Theorem 1 to establish eﬀective versions of the central limit theorems considered in [13. . n Proof. we have π1 (X) = u1 X + v1 . . Similar arguments as above yield the required results (4) and (5) since the mean and the variance of Ωn are nothing but λn (0) and λn (0). . we obtain Rk (n) = O for k = 1. . For the more reﬁned results (4) and (5). 5 . 2 2 2 φ(n)k κn . 14]. for each ﬁxed k = 1. we obtain Pr{Ωn = m} m≥0 (4) (5) mk = πk (φ(n)) + Rk (n).. . Taking r = φ(n)−1 . 2. the mean and the variance of Ωn satisfy E(Ωn ) = u (0)φ(n) + v (0) + O κ−1 . 2 In general. This completes the proof.In particular. j = 1. 2. by Cauchy’s formula. the degree of the polynomial πk (X) is k. Since Mn (s) is analytic for |s| ≤ τ . The developments here are made possible since the singularity analysis provides uniform bounds on the coeﬃcients of analytic functions. 3. n Corollary 1 We have mk Pr{Ωn = m} = u (0)k φ(n)k 1 + Ok m≥0 1 1 + φ(n) κn . Writing uj := u(j) (0) and vj := v (j) (0). 2. . k! where the remainder term Rn (n) satisﬁes Rk (n) = O 1 κn |s|−k−1 eu(s)φ(n) |ds| (0 < r ≤ τ ). π2 (X) = 2 u2 2 u2 + 2u1 v1 v2 + v1 1 X + X+ . 3. By the analyticity of Mn and the relation Mn (0) = 1. .

uniformly in z. only the weaker condition H(z) = K + o(1) for z ∈ ζ. a. etc. cycles in random mappings and random mapping patterns. C(z) = a log 1 + H(z). K). The condition that we imposed here on H(z) is weaker than that required in [14]. z ∈ ∆. 0 < δ < π ).1 Exp-log class The exp-log class covers such problems as cycles in permutations. This was the original condition used in [13]. For brevity. 6 . we shall refer to as saying that C(z) is logarithmic with parameters (ζ. z ∈ ∆). ζ) − S(1. proﬁles of increasing trees.3. z) = n. such that for z ∼ ζ. a. See [13. 1 − z/ζ where the function H(z) is analytic in the region ∆ := {z : |z| ≤ ζ + and | arg(z − ζ)| ≥ δ} ( > 0. By singularity analysis [14. z) is analytic for |z| < ζ + and |w| < 1 + . where K is some constant. ζ) + log × 1+O √ 1 log n . . Mn (s) = m≥0 > 0. z)) (C(0) = 0). K) and the function S(w. to ensure that both the mean and the variance are a log n + O(1). over an additive arithmetic semigroup under Knopfmacher’s axiom A# ). Remark. where C(z) is logarithmic with parameters (ζ. and as z → ζ in ∆. we obtain Pr{Ωn = m}ems := [z n ]P (es . suﬃces to guarantee the later. z) [z n ]P (1. irreducibles in polynomials over ﬁnite ﬁelds (and. z) Γ(a) Γ(aes ) (6) = exp (es − 1)a log n + K(es − 1) + S(es . ζ > 0 being the radius of convergence of C. 2 (with possibly a ﬁnite number of other branch cuts starting from the circle |z| = ζ). Recall that a generating function C(z) analytic at 0 is called logarithmic [13] if there exists a constant a > 0. 23. more generally. But by our Theorem 2. We ﬁrst consider generating functions of the form P (w. H(z) = K + O log− 2 1 1 1 − z/ζ .m≥0 2 pnm wm z n = exp (wC(z) + S(w. 12] (see [19] for details). They need H(z) = K + o log−1 1 1 − z/ζ (z → ζ. 19]. uniformly for s in some complex neighbourhood of the origin.

ζ) . z) = exp(wC(z)). satisﬁes 1 (a log n)k mk Pr{Ωn = m} = 1 + O (log n)−1/2 k! m≥0 k! In particular.. z) = exp (−1)k−1 k wk C(z k ) . Var(Ωn ) = a log n − aψ(a) − a2 ψ (a) + K + S2 + S1 + O (log n)− 2 1 where ψ denotes the logarithmic derivative of the Gamma function and the two constants S1 and S2 are deﬁned by S1 = ∂ S(w. counted with multiplicity) wk P (w. z) = k≥1 wz k 1+ k! ck (ck := k![z k ]C(z)). Set construction: k≥1 P (w. The result follows from Theorem 1. 4. 2]. z) = k≥1 wz k 1+ 1 − zk ck (ck := [z k ]C(z)). 3. Partitional complex construction: P (w.e. The above theorems apply to the following four classes of combinatorial structures when C(z) is logarithmic: 1. Ch. ∂w w=1 S2 = ∂2 S(w. Fn (x) = Φ(x) + O √ 1 . Multiset construction: (i) total number of components (i. ∂w2 w=1 2 Proof. Then Fn (x) satisﬁes. where Ωn and a are deﬁned by (6). This will be discussed in the companion paper [20]. Further reﬁnement of Theorem 1 is possible under slightly stronger conditions on H(z). Partitional complex construction in which no two components are order-isomorphic: P (w. uniformly with respect to x. k k≥1 (ii) number of distinct components (counted without multiplicity) P (w. log n 2 Proof. This follows from (6) and Theorem 2. k ∈ N+ . z) = exp C(z k ) .√ Theorem 3 Let Fn (x) = Pr{Ωn < a log n + x a log n}. cf. since 1/Γ(aes ) = 0 for | (s)| < | log a|. the mean and the variance of Ωn satisfy E(Ωn ) = a log n − aψ(a) + K + S1 + O (log n)− 2 1 . 2. Theorem 4 The k th moment of Ωn . when the dominant singularity of C is strictly less than 1. ζ) . when the dominant singularity of C is strictly inferior to 1. also [19. 7 .

C(z) ≡ z a D(z b ) for some generating function D(z) and integers a. 1 + A1 (w) 1 − 8 z ρ(w) + A2 (w) 1 − z ρ(w) 2 + ··· . . where β ∈ N. viz. where A(w) = 1 . we get C(z) = Thus 1 z = A(w) 1 − 1 − wC(z) ρ(w) −1 1 z − ρ(w)C (ρ(w)) 1 − w ρ(w) + ρ2 (w) z C (ρ(w)) 1 − 2 ρ(w) 2 + ···. z). 2. Let us ﬁx a w. let us consider generating functions of the form P (w. Deﬁne the random variable Ωn by its moment generating function: Mn (s) = m≥0 Pr{Ωn = m} ems = [z n ]P (es . z) for suﬃciently large n. 1. such that C(ρ) = 1. By the inverse function theorem. . where ρ satisﬁes C(ρ) = 1. Deﬁnition. b ≥ 2. let us ﬁrst suppose that C(z) is 1-regular and consider the equation 1 = C(z). (1-regular function [14]) A generating function C(z) ≡ z q (q = 0. Of course ρ(1) = ρ. w ∼ 1. Lemma 1 Let Pn (w) := [z n ]P (w. w Since C(z) is 1-regular. being the radius of convergence of C(z). and α + β > 0.) analytic at z = 0 is called 1-regular if (i) its Taylor expansion at z = 0 involves only non-negative coeﬃcients. namely. dxj Γ(x) x = α Proof. To start with. . The formula Pn (w) = Aα (w)ρ(w)−n nα−1 (log n)β 0≤k≤β β (log n)−k gj (α) (log A(w))k−j + k 0≤j≤k +O |ρ(w)|−n nα−2 (log n)β . Expanding C(z) at z = ρ(w). (ii) there exists a positive number ρ < . that C(z) is aperiodic.m≥0 pnm wm z n = 1 1 log α (1 − wC(z)) 1 − wC(z) β (C(0) = 0). .2 Algebraic-logarithmic class Next. holds uniformly for w ∼ 1. The following lemma is fundamental to further investigations. α ≥ 0. the above equation is locally invertible. [z n ]P (1. z) = n. wρ(w)C (ρ(w)) gj (α) = dj 1 . we have C (ρ) > 0.3. z) . Assume. without loss of generality. the equation has a unique solution ρ(w) for w ∼ 1 and ρ(w) is analytic at w = 1.

where A1 (w) = For P (w. β (log n)−k gj (0) (log A(w))k−j + k 1≤j≤k +O |ρ(w)| −n −2 n (log n)β . β = 1. Remark. Note that the indices of the last two sums both begin at 1 since g0 (0) = 1/Γ(0) = 0. gk (0) = k![xk ] 1 1 = k!(−1)k−1 [xk−1 ] Γ(x) Γ(1 − x) 9 . Pn (w) = ρ(w)−n n−1 (log n)β 1≤k≤β . we get 1 1 − wC(z) = Aα (w) 0≤j≤β α ρ(w)C (ρ(w)) . The double sum can be rearranged in descending powers of log n: β (log A(w))β−j gj−k (α)(log n)k j 0≤k≤j β (log n)−k gj (α) (log A(w))k−j . From a computational point of view. 4C (ρ(w))2 6C (ρ(w)) log 1 1 − wC(z) β β z (log A(w))β−j 1 − j ρ(w) −α+1 −α log . when β = 0 and α > 0. 2C (ρ(w)) A2 (w) = ρ2 (w) C (ρ(w))2 C (ρ(w)) − . In particular. we obtain Pn (w) = Aα (w) 0≤j≤β β (log A(w))β−j gj−k (α)(log n)k j 0≤k≤j + O |ρ(w)|−n nα−2 (log n)β . we can apply the singularity analysis [12]. . the following expressions for gk (0) are useful. Thus transferring to coeﬃcients. Pn (w) = ρ(w)−n n−1 1 + O n−1 and when α = 0 and β > 0. we obtain Pn (w) = Aα (w)ρ(w)−n nα−1 1 + O n−1 when α = 0. 1 1 − z/ρ(w) j + +O 1− z ρ(w) log 1 1 − z/ρ(w) β By the remark before the lemma. z). k 0≤j≤k 2 0≤j≤β = (log n)β 0≤k≤β This completes the proof.

O (log n)−1 . 2 3 (ρ) ρC (ρ) ρC (ρ) √ Theorem 5 Let Fn (x) = Pr{Ωn < α1 n + x α2 n}. ρ−n (7) uniformly for |s| ≤ τ . 3. where the two constants L1 . if (α = 0 and β ≥ 2) or (α > 0 and β > 0). L2 are given by L1 := L2 := 1 C (ρ) + −1 ρC (ρ) C (ρ)2 2 2 3 3 2 4 2 2 2 2αρC1 C2 +αρ2 C2 −ρ2 C1 C2 −ρC1 +ρC1 C2 −2αρC1 +2ρ2 C2 −ρ2 C1 C3 +αρ2 C1 −2αρ2 C1 C2 +αC1 +C1 . In particular.) 10 . n uniformly in x. if (α > 0 and β = 0) or (α = 0 and β = 1). 3. 2. .β (n) .= k!(−1)k−1 [xk−1 ] = k(−1)k−1 0≤j≤ Γ(x) sin πx π (−1)j π 2j k−1 Γ(k−2−2j) (0) 2j + 1 k−2 2 = 0≤j≤ k−1 2 (−1)k−1+j π 2j k Γ(k−1−2j) (1). we have E(Ωn ) = α1 n + αL1 + O (εα.β (n)) . k! m≥0 where k (n) 2 is a polynomial in n of degree k deﬁned by Πk (n) := [sk ] exp −n log ρ(es ) ρ es ρ(es )C (ρ(es )) ρC (ρ) −α .β (n) = O n−1 . . 2j + 1 Thus. Then Fn (x) satisﬁes 1 Fn (x) = Φ(x) + O √ + εα.β (n)). 4 2ρ2 C1 (Cj := C (j) (ρ) for j = 1. Theorem 6 The k th moment of Ωn satisﬁes 1 mk Pr{Ωn = m} = Πk (n) + O(nk εα.β (n)) . 2. for k = 1. Proof. x ∈ R. with Mn (s) := Pn (es )/Pn (1). we have. τ > 0. . Using (7) and Applying Theorem 1. Mn (s) = where εα. ρC (ρ) and α2 = ρ2 C 1 C (ρ) 1 + − . es ρ(es )C (ρ(es )) ρC (ρ) −α ρ(es )−n (1 + εα. Deﬁne two constants α1 = 1 .β (n)) Var(Ωn ) = α2 n + αL2 + O (εα. by Lemma 1.

Proof. z) = log 1 . the convergence rate is O n− 2 . 2 provided that the radius of convergence of C(z) in the last case is less than 1. [6. For Bender’s central limit theorem [1. where the Pnj (z) are polynomials such that (i) each Pnj (z) is itself a probability generating function of some random variable. say. 29]. 11 Mn Ωn − µn < x} = Φ(x) + O . VI] [17. we often have exponential convergence rate for Mn (s). It should be noted that for sequence constructions. The object of study here is sums of independent but not identically distributed random variables. where ϕ(k) denotes Euler’s totient function. See [21]. 1 4 A classical central limit theorem The following theorem is borrowed from [21] where we used it to derive the asymptotic normality of the cost of constructing a random heap.Proof. see. 1 − wC(z) P (w. Remark. z) = and to cycle constructions P (w. This follows from (7) and Theorem 2. Then the distribution of Ωn is asymptotically Gaussian: Pr{ uniformly with respect to x. σn σn . and (ii) Mn → 0. k C(z k ) k 1−w k≥1 1 1 − wC(z) (C(0) = 0). This problem has been extensively studied in the probability literature. Thm 1]. Theorem 7 Let {Ωn }n be a sequence of random variables taking only non-negative integral values 2 with mean µn and variance σn . σn where Mn = max1≤j≤kn deg Pnj (z). §27] [25. Without the explicit error term it is due to Haight [16]. Suppose that the probability generating function Pn (z) of Ωn can be decomposed as Pn (z) = 1≤j≤kn Pnj (z). z) = ϕ(k) 1 log . for some sequence {kn }n . Ωnj (1 ≤ j ≤ kn ). provided that the function A(s) in that theorem has bounded derivative for s near 0. These results apply to sequence constructions of both labelled and unlabelled structures P (w. for instance. it is omitted here. Ch. Since the proof has already been given there.

m) denotes Stirling numbers of the second kind. 13. 5]. then Ωn satisﬁes Pr{ Ωn − µn 1 < x} = Φ(x) + O . Pn (z) is a polynomial whose roots are all located in and on the half-plane 2 z ≤ 0.). Ch. for each ﬁxed n ≥ 1. . It remains to show that σn → ∞. . Many other examples in this same (exp-log) class can be found in [5. . Cycles in permutations. have only simple non-positive roots. 23. 2. The number of permutations having m cycles is enumerated by the (signless) Stirling number of the ﬁrst kind s(n. Harper [18] showed that the polynomials Pn (z) := 0≤m≤n S(n. Example 1. Proof. So we turn to Theorem 7. 2 In particular. If σn → ∞. cf. σn σn for −∞ < x < ∞. The generating function for the number of blocks (marked by w) in set partitions satisﬁes S(n. . Theorem 1 is not applicable since.5] and [19. M = 2 in Theorem 7. m) z m (n = 1. Example 2. 5 Examples Let us discuss some typical examples from diﬀerent applications. Suppose that. bn bn 12 .Corollary 2 Let Pn (z) be a probability generating function of the random variable Ωn . 33] [26. m): s(n. n ≥ 1). n ≥ 1. m) := [wm z n ] exp w log 1 1−z = [wm ] (w(w + 1) · · · (w + n − 1)) (1 ≤ m ≤ n. . where each of the n! permutations is equally likely. deﬁned as the number of cycles in a random permutation of {1. σn := Var(Ωn ). the mean and the variance are not of the same order. We obtain Ωn − log n 1 Pr{ √ < x} = Φ(x) + O √ . then the results of the corollary hold. m) wm n. §2. Both Theorems 3 and 7 apply to the random variable Ωn . n}. log n log n a result originally due to Goncharov (without error term). . 2 σn := Var(Ωn ) = bn+2 bn+1 − − 1 → ∞.4 and 2. namely. [18]. 2. 3. if all roots of Pn (z) are non-positive. .m≥0 zn z = ew(e −1) . where µn = E(Ωn ). A permutation is a set of cycles. as we shall see. Stirling numbers of the second kind. n! where S(n.

given a random permutation of size n. We can also consider other constraints like parity on the size of each block and on the number of blocks. c2 n n uniformly with respect to x.744033. see [19. Using the asymptotic expression [9. Ch.j (z) are polynomials of degree ≤ log2 n.. Ch. σn n uniformly with respect to x. cf. Moreover. Similar result holds for the number of comparisons and other construction algorithms preserving randomness in each step. 225–226]. More precisely. it is not hard to show that 2 σn = 1− ρ2 (2ρ2 + 7ρ + 10) +O 24 n (ρ + 1)3 ρ2 n2 . Alternatively. see [31. Cost of heap construction.. Many other examples can be found in [1. Obviously. 1]. we can write Pr{ Ωn − n/ρ log n < x} = Φ(x) + O √ . 14. Thus Theorem 7 applies with Mn = log n to both cases [21]. deﬁning Ωn as the number of exchanges used by Floyd’s algorithm to construct a heap from this permutation. n n + O(1) ∼ → ∞. Example 4. p. cf. 13 . where pn. [8.. For other examples on algorithms. Example 3. 7]. For other examples. ez − 1 is 1-regular. Doberkat [11] shows that the probability generating functions for the number of exchanges and for the number of comparisons used by Floyd’s algorithm to construct a heap of size n have the form 1≤j≤n pn. [27].261217. Ch. m).j (z). 28].where bn := n! [z n ]ee z −1 are the Bell numbers. we have [21. Theorem 5 applies. deﬁning Pr{Ωn = m} := b−1 S(n. bn ρ log n The asymptotic normality is due to Harper [18]. Ordered set-partitions.. the mean and the variance are all linear. where c1 = −2+ j −1 j≥1 j(2 −1) = 0. The generating function for the number of blocks (marked by w) in an ordered set-partition is (1 − w(ez − 1))−1 . where µn = E(Ωn ) = n n bn+1 − 1 = + O(1) ∼ . and c2 = 2− j≥1 j 2 (2j − 1)2 = 0. 6] −1 en(ρ+ρ −1)−1 √ bn = ρ+1 where ρeρ = n. ρ(ρ + 1) (log n)2 so that Theorem 7 gives. n Pr{ Ωn − µn log n < x} = Φ(x) + O √ . n n/(ρ(ρ + 1)) but this formula cannot be further simpliﬁed by using the asymptotics of ρ. 22] and [19. Thm 4] log n Ωn − c1 n Pr{ √ < x} = Φ(x) + O √ .

It is known that E(ξn ) = log log n + O(1). log log n The convergence rate was ﬁrst conjectured by Le Veque and then proved by R´nyi and Tur´n [30] to e a be O((log log n)− 2 ). k ≥ 1.. and Var(ξn ) = log log n + O(1). Number of factors in square-free factorizations. π2 . we readily obtain 1 ξn − log log n < x} = Φ(x) + O √ . The Erd˝s-Kac Theorem o states that ξn − log log n Pr{ √ < x} → Φ(x). 9]. in view of (8). e. see [10] and [19. uniformly for | z| ≤ π. πn ) of {1. . 72 (8) 1 m [z ] (1 + z)(1 + z + z 2 ) · · · (1 + z + z 2 + · · · + z n−1 ) . Consider the bivariate generating function: P (w. Example 6.b is Kronecker’s symbol. Pr{ √ log log n log log n uniformly with respect to x. n ≥ 2. Deﬁne a random variable ξn by its probability distribution: o Pr{ξn = m} = 1 1 |{k : 1 ≤ k ≤ n. Pr{ for all x. Starting with the well-known Selberg’s formula [32] Mn (z) := E(eξn z ) = V (ez )(log n)e z −1 1 1+O 1 log n .k = δ0. ω(k) = m}| = [z m ] z ω(k) . 14 . n}.Example 5. . Deﬁne f1. Example 7.m wm . ω(2·34 ·56 ·78910 ·1112 ·13141516 ) = 6. 4 and Var(Ωn ) = n(n − 1)(2n + 5) . s) := 1 + n≥2 n−s m≥1 fn. 3. Theorem 7 applies and we get.g. where δa. For many other examples. Then it is easily seen that [24. . πj is called an inversion if i < j and πi > πj . a pair πi . n! for n ≥ 1. Given a permutation π = (π1 . . Let Ωn denote the number of inversions in a random permutation of size n. . . where V (t) is an entire function in the t-plane.k denote the number of factor- izations of n into k distinct factors greater than 1. Erd˝s-Kac Theorem. §5.1] Pr{Ωn = m} = From which we obtain E(Ωn ) = n(n − 1) . Inversions in permutations. n n 1≤k≤n Ωn − n2 /4 1 < x} = Φ(x) + O √ .k . Let fn. Let us apply Theorem 1 to establish this fact. . n n3/2 /6 where ω(n) denotes the number of distinct prime divisors of n. .1. Ch. 2.

w ∈ C). log n where hn (w) are Hermite polynomials. 0]. Γ(ez )(1 + ez ) log n and Theorem 1 gives Pr{ for all x. 200]: hn (w) = 2− 2 nn/2 e−n/2+ 1 √ n+1 w−w2 /4 1 1+O √ n .which equals to P (w. the error term being uniform with respect to w ∈ C \ (−∞. x ∈ R. Ch. m Deﬁne a polynomial Pn (w) := 1 + 2≤k≤n m≥1 fk. 1 √ . The asymptotic behaviour of hn (w) is well-known [34. 10] that [wm ] Pn (w) . We obtain √ Ωn − n + 1 √ Pr{ < x} = Φ(x) + O 4 n+1 the O being uniform with respect to x. Deﬁning Ωn as the number of ﬁxed points in a random involution. uniformly for all ﬁnite w. p. Example 8. n! n≥0 ξn − 4 √ log n 1 4 < x} = Φ(x) + O log n √ 4 1 . 4 n √ n+1 (es −1)− 1 (e2s −1) 4 1 1+O √ n . The number of ﬁxed points (cycles of length 1). Pn (1) √ 1 n e2 w log n 1+O √ Pn (w) = √ 3/4 w 3/4 (1 + w)Γ(w) log n 2 π(log n) . see [19. 10]. marked by w. Hermite polynomials. Thus we obtain √ z/2 1 2 e2 log n (e −1)−3z/4 ξn z Mn (z) := E(e ) = 1+O √ . s) = n≥2 1+ w ns ( s > 1. For other examples. 15 . in the involutions is enumerated by exp wz + z2 2 = hn (w) n z . Ch.m w and a random variable ξn for which Pr{ξn = m} = We show in [19. we obtain Mn (s) = E(eΩn s ) = e uniformly for | s| ≤ π.

The Art of Finite and Inﬁnite Expansions. Canﬁeld. B. Advanced Combinatorics. Richmond. 1983. Williamson. Central and local limit theorems applied to asymptotic enumeration. New York. Gaussian limiting distributions for the number of components in combinatorial structures. [9] N. [8] L. Richmond and S. Reidel Publishing Company. [4] E. A. matrix recursions. Flajolet. de Bruijn. 1974. 1973. Soria. 3:216–240. Bender. Series A. 1986. II. 1971. Odlyzko. 581. Comtet. Series A. Flajolet and M. Varieties of increasing trees. New York. 23:275–290. Central and local limit theorems for the coeﬃcients of polynomials of binomial type. Discrete Mathematics. revised and enlarged edition. Reprinted by Dover. B. [6] P. 15:91–111. Lecture Notes in Computer Science. 61:114–131. Series A. Bender and L. Flajolet and A. Series A. Series A. Central and local limit theorems applied to asymptotic enumeration. 1981. Springer-Verlag. M. R. [12] P. Richmond. Journal of Combinatorial Theory. A generalisation of Canﬁeld’s formula. Central and local limit theorems applied to asymptotic enumeration. 114:159–180. Journal of Combinatorial Theory. [14] P. Series A. North-Holland. Soria. General combinatorial schemes: Gaussian limit distributions and exponential tails. 1977. Journal of Combinatorial Theory. multivariate generating functions. G. 1984. III. Doberkat. [5] F. 1958. Journal of Combinatorial Theory. Delange. A. L. 16 . Probability and Measure. 1983. D. B. 1986. P. 1993. Information and Control. G. Billingsley. SIAM Journal on Discrete Mathematics. 1992. [11] E. Singularity analysis of generating functions. Salvy. Asymptotic Methods in Analysis. Journal of Combinatorial Theory. [10] H.References [1] E. Bender and L. A. [13] P. pages 24–48. Bergeron. [7] E. 53:165–182. In Proceedings CAAP’92. Journal of Combinatorial Theory. Dordrecht-Holland. Flajolet and M. [3] E. A. and B. Bender. 1990. 34:255–265. Wiley. E. [2] E. An average case analysis of Floyd’s algorithm to construct heaps. Acta Arithmetica. 41:50–60. 19:105–146. 1990. Second edition. Sur des formules de Atle Selberg. 35:263–278.

e [26] H. e e e Th`se. New York. Colloquia Mathematica Societatis J´nos Bolyai. 1982. M. 61:327–336.[15] Z. 1992. [28] L. Steyaert. [27] C. Warlimont. 41:177–186. Theoretical Computer Science. V. New York and Basel. Haigh. The Art of Computer Programming. [24] D. Springer-Verlag. Pitman Advanced Publlishing Company. Brown.-K. Acta Arithmetica. [17] P. Inc. Sums of Independent Random Variables. Forth edition. J. New York. 1971. 17 . and R. B. 1977. A. 62. Lo`ve. Probability Theory. NewYork. Collection: Limit Theorems in Probability and Statistics. Limit distributions related to the weighted Stirling numbers of the ﬁrst and second c kind. 1992.-M. 1967. Rates of Convergence in the Central Limit Theorem. submitted. “Factorisatio numerorum” in arithmetical semigroups. Marcel Dekker. Harper. 1989. Springer-Verlag. Petrov. [23] J. [22] A. John Wiley & Sons. Analytic Arithmetic of Algebraic Function Fields. Annals of Mathematical Statistics. Th´or`mes limites pour les structures combinatoires et les fonctions arithm´tiques. Mahmoud. Journal of Computational and Applied Mathematics. [19] H. H.-K. E. 36:823–842. 1975. Massachusetts. [21] H. [18] L. [25] M. Mutafˇiev. to appear.. Evolution of Random Search Trees. Hwang. Knopfmacher.-K. Stirling behaviour is asymptotically normal. Knuth. J. Hwang and J. H. Berlin-Heidelberg. On asymptotic expansions in the central and local limit theorems for combinatorial structures. Ecole polytechnique. Volume 3: Sorting and Searching. Journal of Algorithms. Knopfmacher. Richmond. McDiamid and B. volume 50 of Lecture Notes in Pure and Applied Mathematics. a [29] V. 1973. Reading. 1994. Building heaps fast. 38:410–414. Biometrika. On the number of heaps and the cost of heap construction. 1992. 10:352–365. Central and local limit theorems applied to asymptotic enumeration IV: multivariate generating functions. Knopfmacher. e [20] H. Research Notes in Mathematics. Hwang. Hall. Gao and L. A neat way to prove asymptotic normality. 1982. [16] J. Translated from the Russian by A. Reed. 1979. Boston. Addison Wesley. 58:677–678. A.

M´thode d’Analyse pour les Constructions Combinatoires et les Algorithmes. Rhode-Island. Acta Arithmetica. Voigt.[30] A. Orthogonal Polynomials. Note on a paper by L. Th`se e e ´ d’Etat. R´nyi and P. e [34] G. G. 1990. Journal of the Indian Mathematical Society. forth edition. 1977. AMS. 1958. Sathe. e a o [31] A. A local limit theorem for generalized Stirling numbers. Soria. 1954. [33] M. Universit´ de Paris-Sud. On a theorem of Erd˝s-Kac. 35:161–172. volume XXIII of American Mathematical Society Colloquium o Publications. 18:83–87. Ruci´ski and B. Selberg. e e [32] A. 18 . 4:71–84. Revue Roumaine n de Math´matiques Pures et Appliqu´es. 1990. Tur´n. Providence. Szeg¨.