FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM

PERMUTATIONS UNDER A CONSERVATIVE MEASURE

ALESSANDRA CIPRIANI AND DIRK ZEINDLER

Abstract. In this work we are considering the behavior of the limit shape of
Young diagrams associated to random permutations on the set {1, . . . , n} under
a particular class of multiplicative measures. Our method is based on generating
functions and complex analysis (saddle point method). We show that fluctua-
tions near a point behave like a normal random variable and that the joint fluctu-
ations at different points of the limiting shape have an unexpected dependence
structure. We will also compare our approach with the so-called randomization
of the cycle counts of permutations and we will study the convergence of the
limit shape to a continuous stochastic process.

Contents

1. Introduction 2
Notation 5
2. Preliminaries 5
2.1. Cycle counts 5
2.2. Generating functions 6
2.3. Approximation of sums 7
3. Randomization 8
3.1. Grand canonical ensemble 9
3.2. Limit shape and mod-convergence 10
3.3. Functional CLT 13
4. Saddle point method 14
4.1. Log-n-admissibility 16

Date: January 27, 2014.

1
2 A. CIPRIANI AND D. ZEINDLER

4.2. Limit shape for polynomial weights 18

4.3. Functional CLT for wn (·) 26
4.4. Large deviations estimates 30
4.5. An example: the case gΘ (t) = (1 − t)−1 31
Acknowledgments 32
Appendix A. Euler Maclaurin formula with non integer boundaries 32
References 33

1. Introduction

The aim of this paper is to study the limit shape of a random permutation under
the generalised Ewens measure with polynomial growing cycle weights and the
fluctuations at each point of the limit shape. The study of such objects has a
long history, which started with the papers of Temperley [24] and Vershik [25].
Later on Young diagrams have been approached under a different direction, as
in the independent works of [26] and [20], which first derived the limit shape
when the underpinned measure on partitions is the so-called Plancherel measure.
We will not handle this approach here, even though it presents remarkable con-
nections with random matrix theory and random polymers, among others (see
for example [10]).
We first specify what we define as the limit shape of a permutation. We denote
by Sn the set of permutations on n elements and write each permutation σ ∈ Sn
as σ = σ1 · · · σ` with σj disjoint cycles of length λ j . Disjoint cycles commute and
we thus can assume λ1 ≥ λ2 ≥ · · · ≥ λ` . This assigns to each permutation
σ ∈ Sn in a unique way a partition of n and this partition λ = (λ1 , λ2 , . . . , λ` )
is called the cycle type of σ. We will indicate that λ is such a partition with
the notation λ ` n. We define the size |λ| := ∑i λi (so obviously if λ ` n then
|λ| = n). λ features a nice geometric visualisation by its Young diagram Υλ . This
is a left- and bottom-justified diagram of ` rows with the j−th row consisting
of λ j squares, see Figure 1(a). It is clear that the area of Υλ is n if λ ` n.
After introducing a coordinate system as in Figure 1(b), we see that the upper
boundary of a Young diagram Υλ is a piecewise constant and right continuous
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 3

(a) The Young diagram (b) The shape function wn (·)

Figure 1. Illustration of the Young diagram and the shape of

σ = (3578)(129)(4)(6) ∈ S9

function wn : R+ → N+ with
n
wn ( x ) : = ∑ 1{ λ j ≥ x } (1.1)
j =1

with the convention that C0 := 0. The cycle type of a permutation becomes

a random partition if we endow the space Sn with a probability measure Pn .
What we are then interested in studying is the now random shape wn (·) as
n → +∞, and more specifically to determine its limit shape. The limit shape
with respect to a sequence of probability measures Pn on Sn (and sequences of
positive real numbers An and Bn with An · Bn = n) is understood as a function
w∞ : R+ → R+ such that for each e, δ > 0
"( )#
lim Pn σ ∈ Sn : sup | A− 1
n wn ( Bn x ) − w∞ ( x )| ≤ e = 1. (1.2)
n→+∞ x ≥δ

The assumption An · Bn = n ensures that the area under the rescaled Young
diagram is 1. One of the most frequent choices is An = Bn = n1/2 , but we
will see that it’s useful to adjust the choice of An and Bn to the measures Pn .
Equation (1.2) can be viewed as a law of large numbers for the process wn (·).
The next natural question is then whether fluctuations satisfy a central limit
theorem, namely whether

An wn ( Bn x ) − w∞ ( x )
4 A. CIPRIANI AND D. ZEINDLER

converges (after centering and normalization) in distribution to a Gaussian pro-

cess on the space of càdlàg functions, for example. Of course the role of the
probability distribution with which we equip the set of partitions will be crucial
to this end.

In this paper, we work with the following measure on Sn :

1 `
hn n! ∏
Pn [ σ ] = ϑλ j . (1.3)
j =1

where (λ1 , . . . , λ` ) is the cycle type of σ, (ϑm )m≥1 is a sequence of non-negative

weights and hn is a normalization constant (h0 is defined to be 1). From time to
time we will also use ϑ0 := 0 introduced as convention.
This measure has recently appeared in mathematical physics for a model of the
quantum gas in statistical mechanics and has a possible connection with the
Bose-Einstein condensation (see e.g. [6] and [12]). Classical cases of this mea-
sure are the uniform measure (ϑm ≡ 1) and the Ewens measure (ϑm ≡ ϑ). The
uniform measure is well studied and has a long history (see e.g. the first chap-
ter of [3] for a detailed account with references). The Ewens measure originally
appeared in population genetics, see [14], but has also various practical appli-
cations through its connection with Kingman’s coalescent process, see [18]. The
measure in (1.3) also has some similarities to multiplicative measure for parti-
tions, see for instance [8]. It is clear that we have to make some assumptions on
the sequence (ϑm )m≥1 to be able study the behaviour as n → +∞. We use in this
paper polynomial growing cycle weights ϑm as considered in the recent work
Ercolani and Ueltschi [12] and of Maples, Nikeghbali and Zeindler [22]. More
precisely, we use the weights
mα  
ϑm = + O mβ (1.4)
Γ ( α + 1)
with some α > 0 and 0 ≤ β < α/2. We would like to point out that the require-
ment 0 ≤ β < α/2 and the normalisation constant Γ(α + 1) are not essential and
it only simplifies the notation and the computations. Our argumentation indeed
works also for ϑm ∼ const · mα . Note that the limit shape and the fluctuations
at points of the limit shape with the weights (1.4) have already been studied by
Erlihson and Granovsky in [13] in the context of Gibbs distributions on integer
partitions. However, the approach in this paper is slightly different and allows
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 5

to simplify the computations and to get at the same time the behaviour of the
cumulants and to give some large deviation estimates for the fluctuations at the
limit shape. We may mention two popular methods in the literature to study
the asymptotic behaviour of the function wn ( x ) under such assumptions. The
first one is is complex analytic and uses the saddle-point method as described
in Section 4. This method was used in [12] and [22] and an introduction can be
found for instance in [15, Section VIII]. The second one is stochastic and based
on randomisation and was used in [8] and [13]. We present in Section 3 an argu-
mentation similar to [8] and give at the begin of Section 3 an idea how to deduce
the behaviour for n → ∞ from the randomised setting. It is typically expected
that the ‘randomised’ wn ( x ) has the same asymptotic behaviour as the ‘unran-
domized’ wn ( x ). We will see here that this is not the case. More precisely, we
show that the ‘randomized’ and the ‘unrandomized’ wn ( x ) have different limit
shapes and different behaviours of the fluctuations around the limit shape.

Notation. We denote with R+ := R \ (−∞, 0) and with N+ := N \ {0}. With

bold fonts we will always indicate vectors and h ·, ·i will be the Euclidean scalar
product on R` . M` (R) is the set of ` × ` matrices with real coefficients. L( X )
L
stands for the law of a random variable X and thus → is the symbol representing
convergence in distribution.

2. Preliminaries

We introduce in this section the notation of the cycle counts and the notation of
generating functions.

2.1. Cycle counts. The notation λ = (λ1 , λ2 , . . . , λ` ) is very useful for the illus-
tration of λ via its Young diagram, but in the computations it is better to work
with the cycle counts Ck . These are defined as

Ck (σ) = Ck := # j ≥ 1; λ j = k (2.1)

for k ≥ 1 and λ = (λ1 , λ2 , . . . , λ` ) the cycle type of σ ∈ Sn . We obviously have

for k ≥ 1
n
Ck ≥ 0 and ∑ kCk = n. (2.2)
k =1
6 A. CIPRIANI AND D. ZEINDLER

It is also clear that the cycle type of permutation (or a partition) is uniquely
determined by the vector (C1 , C2 , . . . ). The function wn ( x ) and the measure
Pn [ · ] in (1.1) and (1.3) can now be written as
1 n Ck
wn ( x ) = ∑ Ck and Pn [σ ] =
hn n! k∏
ϑk . (2.3)
k≥ x =1

Our aim is to study the behaviour of wn ( x ) as n → ∞. It is thus natural to

consider the asymptotic behaviour of Ck with respect to the measure Pn [ · ].

Lemma 2.1 ([12], Corollary 2.3). Under the condition hn−1 /hn → 1 the random
variables Ck converge for each k ∈ N in distribution to a Poisson distributed random
variable Yk with E [Yk ] = ϑk
k . More generally for all b ∈ N the following limit in
distribution holds:

lim (C1 , C2 . . . , Cb ) = (Y1 , Y2 . . . , Yb )

n→+∞

with Yk independent Poisson random variables with mean E [Yk ] = ϑk

k .

One might expect at this point that wn ( x ) is close to ∑nk= x Yk . Unfortunately we

will see in Section 4 that the asymptotic behaviour of wn ( x ) is more complicate.

2.2. Generating functions. The (ordinary) generating function of a sequence

( ak )k≥0 of complex numbers is defined as the formal power series
∞
g(z) := ∑ ak zk . (2.4)
j =0

As usual, we define the extraction symbol [zk ] g(z) := ak , that is, as the coefficient
of zk in the power series expansion (2.4) of g(z).
A generating function that plays an important role in this paper is
ϑm m
gΘ ( t ) : = ∑ m
t . (2.5)
m ≥1
mα


As mentioned in the introduction, we will use ϑm = + O m β . We stress
Γ ( α +1)
that generating functions of the type (1 − t)−α fall also in this category, and for
them we will recover the limiting shape as previously done in [13]. We will see
in particular this case in Section 4.
The reason why generating functions are useful is that it is often possible to write
down a generating function without knowing an explicitly. In this case one can
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 7

try to use tools from analysis to extract information about an , for large n, from
the generating function. It should be noted that there are several variants in
the definition of generating functions. However, we will use only the ordinary
generating function and thus call it ‘just’ generating function without risk of
confusion.
The following well-known identity is a special case of the general Pólya’s Enumer-
ation Theorem [23, 16, p. 17] and is the main tool in this paper to obtain generating
functions.

Lemma 2.2. Let ( am )m∈N be a sequence of complex numbers. We then have as formal
power series in t
∞
!
n
tn Cj 1 am m
∑ n! ∑ ∏ a j = ∑ tn ∑ zλ ∏ = exp ∑
C
ak k t
n ∈N σ ∈Sn j =1 n ∈N λ`n k =1 m ≥1
m

where zλ := ∏nk=1 kCk Ck !. If one series converges absolutely, so do the others.

We omit the proof of this lemma, but details can be found for instance in [21,
p. 5].

2.3. Approximation of sums. We require for our argumentation the asymptotic

behaviour of the generating function gΘ (t) as t tends to the radius of conver-
gence, which is 1 in our case.

Lemma 2.3. Let (vn )n∈N a sequence of postive numbers with vn ↓ 0 as n → +∞. We
have for all δ ∈ R \ {−1, −2, −3, . . . }
∞
∑ kδ e−kvn = Γ(δ + 1)v−n δ−1 + ζ (−δ) + O(vn ). (2.6)
k =1

ζ (·) indicates the Riemann Zeta function.

This lemma can proven with Euler Maclaurin summation formula or with the
Mellin transformation. The computations with Euler Maclaurin summation are
straightforward and the details of the proof with the Mellin transformation can
be found for instance in [15, Chapter VI.8]. We thus omit it.
We require also the behaviour of partial sum ∑∞
k= x
θm m
mt as x → ∞ and as t → 1.
We have
8 A. CIPRIANI AND D. ZEINDLER

Lemma 2.4 (Approximation of sums). Let vn , zn be given with zn → +∞ and

zn vn = a0 + a1 n− β for β > 0, a0 > 0 and a0 , a1 ∈ R. We then have for all δ ∈ R and
all q ∈ N
∞
!
 δ +1 q k
Γ(δ + k + 1, a0 )
 
zn a  
∑ kδ e−kvn =
a0 ∑ k!
− 1 n− β
a0
+ O n−(q+1) β
k=bzn c k =0
Z +∞
− B1 (zn − bzn c)zδn e−a0 + B1 (y − byc)(δ − vn y)yδ−1 e−vn y dy.
zn
R +∞
with Γ( a, x ) := x s a−1 e−s ds the incomplete Gamma function and B1 ( x ) := x − 1
2
the first Bernoulli polynomial.

Proof. The proof of this lemma is based on the Euler Maclaurin summation for-
mula, see [2] or [1, Theorem 3.1]. We use the here the following version: let
f : R+ → R have a continuous derivative and suppose that f and f 0 are inte-
grable. Then
Z +∞ Z +∞
∑ f (k) =
c
f ( x ) dx − B1 (c − bcc) f (c) +
c
B1 ( x − b x c) f 0 ( x ) dx. (2.7)
k≥bcc

We substitute f ( x ) := x δ e− xvn , c := zn and notice that f and all its derivatives

tend to zero exponentially fast as x → +∞. Now by the change of variables
zn
x := a0 y
Z +∞   δ +1 Z +∞ a
−vn x δ zn − a1 n − β y
e x dx = yδ e−y e 0 dy =
zn a0 a0
!
 δ +1 q k
Γ(δ + k + 1, a0 )
 
zn a  
=
a0 ∑ k!
− 1 n− β
a0
+O n −(q+1) β
(2.8)
k =0
where we have swapped integral and series expansion of the exponential by
Fubini’s theorem. 

Remark. One can obtain more error terms in the expansion in Lemma 2.4 by
using more terms in the Euler Maclaurin summation formula. We have stated
in Appendix A a version of the Euler Maclaurin summation formula with non-
integer boundaries, which is more suitable for this than the usual one.

3. Randomization
. .
We introduce in this section a probability measure Pt [ · ] on ∪n≥1 Sn , where ∪
denotes the disjoint union, dependent on a parameter t > 0 with Pt [ · |Sn ] =
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 9

Pn [ · ] and consider the asymptotic behaviour of wn ( x ) with respect to Pt [ · ] as

t → 1.

3.1. Grand canonical ensemble. Computations on Sn can turn out to be diffi-

cult and many formulas can not be used to study the behaviour as n → ∞. A
possible solution to this problem is to adopt to a suitable randomization. This
has been successfully introduced by [16] and used also by [8] as a tool to investi-
gate combinatorial structures, and later applied in many contexts. The main idea
of randomization is to define a one-parameter family of probability measures on
.
∪n≥1 Sn for which cycle counts turn out to be independent. Then one is able to
study their behavior more easily, and ultimately the parameter is tuned in such
a way that randomized functionals are distributed as in the non-randomized
context. Let us see how to apply this in our work. We define


GΘ (t) = exp gΘ (t) (3.1)

with gΘ (t) as in (2.5). If GΘ (t) is finite for some t > 0, then for each σ ∈ Sn let
us define the probability measure

1 tn n Ck
GΘ (t) n! k∏
Pt [ σ ] : = ϑk . (3.2)
=1
.
Lemma 2.2 shows that Pt is indeed a probability measure on ∪n≥1 Sn . The
induced distribution on cycle counts Ck can easily be determined.

Lemma 3.1. Under Pt [ · ] the Ck ’s are independent and Poisson distributed with
ϑk k
Et [Ck ] = t .
k

Proof. From Pólya’s enumeration theorem (Lemma 2.2) we obtain

h i 1 tn
Et e−sCk = ∑ ∑ e−sCk Pt [σ] = ∑ ∑ (ϑk e−s )Ck ∏(ϑ j )Cj
n ≥0 σ ∈Sn G Θ ( t ) n ≥0 σ ∈Sn n! j≤n
j6=k

+∞ ϑ
!  
1 j
ϑk k
= exp ∑ t j exp e−s − 1 t
GΘ (t) j =0
j k
 
−s

ϑk k
= exp e − 1 t .
k
Analogously one proves the pairwise independence of cycle counts. 
10 A. CIPRIANI AND D. ZEINDLER

Obviously the following conditioning relation holds:

Pt [ · | Sn ] = Pn [ · ] .

A proof of this fact is easy and can be found for instance in [17, Equation (1)].
We note that wn ( x ) is Pt -a.s. finite, since Et [wn ( x )] < +∞. Now since the
conditioning relation holds for all t with GΘ (t) < +∞, one can try to look
for t satisfying “Pn [ · ] ≈ Pt [ · ]”, which heuristically means that we choose
a parameter for which permutations on Sn weigh as most of the mass of the
measure Pt . We have on Sn
` n
n= ∑ λ j = ∑ kCk .
j =1 k =1

A natural choice for t is thus the solution of

∞ ∞
" #
n = Et ∑ kCk = ∑ ϑk t k . (3.3)
k =1 k =1

which is guaranteed to exist if the series on the right-hand side is divergent at

the radius of convergence (we will see this holds true for our particular choice
kα +O(k β )
of weights). We write t = e−vn and use Lemma 2.3 in our case ϑk = Γ(α+1) to
obtain
!
n = (vn )−α−1 + O (vn )− β−1 =⇒ vn = (n∗ )−1 + O (n∗ ) β−α−1



(3.4)
1
with n∗ := n 1+α . We will fix this choice for the rest of the section.

3.2. Limit shape and mod-convergence. In order to derive our main results
from the measure Pt we will use a tool developed by [19], the mod-Poisson con-
vergence.

Definition 3.2. A sequence of random variables ( Zn )n∈N converges in the mod-Poisson

sense with parameters (µn )n∈N if the following limit
h i
lim exp(µn (1 − eiu ))E eiuZn = Φ(u)
n→+∞

exists for every u ∈ R, and the convergence is locally uniform. The limiting function Φ
is then continuous and Φ(0) = 1.

This type of converge gives stronger results than a central limit theorem, indeed
it implies a CLT (and other properties we will see below). Our goal will then be
to prove the following
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 11

1
Proposition 3.3. Let x ≥ 0 be arbitrary and x ∗ := xn∗ with n∗ = n 1+α . Furthermore,
let t = e−vn with vn as in (3.4). Then the random variables (wn ( x ∗ ))n∈N converge in
the mod-Poisson sense with parameters µn = (n∗ )α w∞r ( x ) + o ( n∗ )α/2 , where

r Γ(α, x )
w∞ ( x ) := . (3.5)
Γ ( α + 1)

Γ(α, x ) is the upper incomplete Gamma function.

Proof. We have
 
∞
is ∑∞ ϑj j
h ∗)
i h i  
Et eiswn ( x = Et e `=b x ∗ c C` = exp  eis − 1 ∑ j
t . (3.6)
j=b x ∗ c

This is the characteristic function of Poisson distribution. We thus obviously

have mod-Poisson convergence with limiting function Φ(t) ≡ 1. It remains to
compute the parameter µn . Applying Lemma 2.3 for x = 0 and Lemma 2.4 for
x > 0 together with (3.4) gives
+∞ j α −1 + O j β −1 j Γ(α, x )


∑ t = (n∗ )α + O (n∗ ) β .


(3.7)
j=b x ∗ c
Γ ( α + 1) Γ ( α + 1)

Since β < α/2 by assumption, we deduce that λn := (n∗ )α w∞

r ( x ) + o ( n∗ )α/2 .

This completes the proof. 

This yields a number of interesting consequences. In first place we can prove a

CLT and detect the limit shape accordingly.

Corollary 3.4 (CLT and limit shape for randomization). With the notation as above,
we have as n → ∞ with respect to Pt

wn ( x ∗ ) − ( n∗ )α w∞
r (x)
L
enr ( x ) :=
w α
r
→ N (0, (σ∞ ( x ))2 ). (3.8)
(n∗ ) 2
r ( x ) (with scaling A = ( n∗ )α and
Furthermore the limit shape of wn ( x ) is given by w∞ n
Bn = n∗ , see (1.2)). In particular, we can choose δ = 0 in (1.2).

Proof. The CLT follows immediately from [19, Prop. 2.4], but also can be de-
duced easily from (3.6) by replacing s by s(n∗ )−α/2 . It is also straightforward to
r ( x ) is the limit shape. For a given e > 0, we choose 0 = x < x <
show that w∞ 0 1
12 A. CIPRIANI AND D. ZEINDLER

r (x
· · · < x` such that w∞ r r
j+1 ) − w∞ ( x j ) < e/2 for 1 ≤ j ≤ ` − 1 and w∞ ( x` ) < e/2.
It is now easy to see that for each x ∈ R+

|(n∗ )−α wn ( x ∗ ) − w∞
r
( x )| > e =⇒ ∃ j with |(n∗ )−α wn ( x ∗j ) − w∞
r
( x j )| > e/2.

Thus
" #
` h i
Pt sup |(n∗ )α wn ( x ∗ ) − w∞
r
( x )| ≥ e ≤ ∑ t
P |( n ∗ α
) w (
n jx ∗
) − w r
(
∞ j x )| ≥ e/2
x ≥0 j =1
(3.9)
It now follows from (3.8) that each summand in (3.9) tends to 0 as n → ∞. This
completes the proof. 

Another by-product of mod-Poisson convergence of a sequence ( Zn )n∈N is that

one can approximate Zn with a Poisson random variable with parameter µn , see
[19, Prop. 2.5]. However in our situation this is trivial since wn ( x ∗ ) is already
Poisson distributed.
As we are going to do in the next section, we are also interested in the behavior
of increments and their joint behaviour.

Proposition 3.5. For all x, y ∈ R, y > x, set

r Γ(α, x ) − Γ(α, y)
wn ( x, y) := wn ( x ) − wn (y) and w∞ ( x, y) := .
Γ ( α + 1)
Then
wn ( x ∗ , y∗ ) − ( n∗ )α w∞
r ( x, y )
L
enr ( x, y) :=
w αp → N (0, 1) (3.10)
∗ r
(n ) 2 w∞ ( x, y)
1 1
as n → ∞ with x ∗ = xn α+1 and with y∗ = yn α+1 .
enr ( x ) and w
Furthermore, w enr ( x, y) are asymptotically independent.

Remark. As we will see, the proof of independence relies on the independence

of cycles coming from Lemma 3.1. Therefore it is easy to generalize the above
result to more than two points.

Proof. The proof of (3.10) almost the same as the proof of (3.8) and we thus omit
it. Since
y ∗ −1 ∞
wn ( x, y) = ∑ Ck and wn (y) = ∑ Ck
k= x∗ k =y∗
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 13

enr ( x ) and w
and all Ck are independent, we have that w enr ( x, y) are independent
for each n ∈ N. Thus w
enr ( x ) and w
enr ( x, y) are also independent in the limit. 

3.3. Functional CLT. The topic of this section is to prove a functional CLT for the
profile wn ( x ) of the Young diagram. Similar results were obtained in a different
framework by [17, 11] on the number of cycle counts not exceeding nb xc , and by
[5] for Young diagrams confined in a rectangular box. We show

enr : R+ → R (see (3.8)) converges weakly with respect to

Theorem 3.6. The process w
Pt as n → ∞ to a continuous process w
e∞r : R+ → R with w
e∞r ( x ) ∼ N (0, σ r ( x )) and
∞
independent increments.

The technique we will exploit is quite standardized (see [17]). We remark that,
unlike in this paper where the Ewens measure is considered, we do not obtain
here a Brownian process, as the variance of w
e∞r (t) − w
e∞r ( s ) for r ≥ s is more

complicated than in the case of the Wiener measure.

We know from Proposition 3.5 the finite dimensional marginals of the process.
More specifically we have for x` ≥ x`−1 ≥ · · · ≥ x1 ≥ 0 that

(n∗ )−α/2 wn ( x`∗ ), wn ( x`−

∗ ∗ ∗ ∗
1 ) − wn ( x` ), . . . , wn ( x1 ) − wn ( x2 ) ∼ N 0, Σ
0



(3.11)

where Σ0 is a diagonal matrix with

0
Σ11 r
= w∞ ( x` ) and Σ0jj = w∞
r
( x`− j+1 , x`− j+2 ) for j ≥ 2.
Now all we need to show to complete the proof of Theorem 3.6 is the tightness
enr . In order to do so, we will proceed similarly to [17], namely
of the process w
we will show that

Lemma 3.7. We have for 0 ≤ x1 < x ≤ x2 < K with K arbitrary

h i  
Et ( w enr ( x1 ))2 (w
enr ( x ) − w enr ( x ))2 = O ( x2 − x1 )2
enr ( x2 ) − w (3.12)
1 1 1
with x ∗ := xn α+1 , x1∗ := x1 n α+1 and x2∗ := x2 n α+1 .

enr is tight.
Lemma 3.7 together with [7, Theorem 15.6] implies that the process w
This and the marginals in (3.11) prove Theorem 3.6.

Proof of Lemma 3.7. We define

h i
E ∗ : = Et ( w
enr ( x ∗ ) − w
enr ( x1∗ ))2 (w
enr ( x2∗ ) − w
enr ( x ∗ ))2 . (3.13)
14 A. CIPRIANI AND D. ZEINDLER

Centering with Et [wn (·)] and the independence of the cycle counts leads us to
 
x ∗ −1 x2∗ −1
!
θ
 ∑ (n∗ )−α k tk  · ∑ (n∗ )−α k tk θ
E∗ =
k= x1∗
k k= x∗
k
Z x∗ Z x∗
(n∗ )−α (n∗ )−α
  
Lem. 2.4 2
α −1 − t α −1 − t
∼ t e dt t e dt
Γ(α + 1) x1∗ Γ ( α + 1) x ∗
Γ(α, x1 ) − Γ(α, x ) Γ(α, x ) − Γ(α, x2 )
  
=
Γ ( α + 1) Γ ( α + 1)
 
∼ O (( x − x1 )( x2 − x )) = O ( x2 − x1 )2 .

Here we have used the fact that Γ(α, ·) is a Lipschitz function and the assump-
tion that x1 < x ≤ x2 < K. 

4. Saddle point method

The aim of this section is to study the asymptotic behaviour of wn ( x ) with re-
spect to Pn [·] as n → ∞ and to compare the results with the results in Section 3.
There are at least two approaches with which to tackle this problem: one is more
probabilistic and was employed by [13] in their paper. The second one was first
developed in [22] from the standard saddle point method.
The first method to study the asymptotic statistics of wn ( x ) with respect to Pn [·]
as n → ∞ is the so called Khintchine method. We illustrate this method briefly
with the normalisation constant hn (see (1.3)). The first step is to write down a
Khintchine’s type representation for the desired quantity. For hn this is given by
! " #
n n
ϑk k
hn = t exp ∑ t Pt ∑ kCk = n
−n
(4.1)
k =1
k k =1

with t > 0 and Pt [ · ] as in Section 3. The second step is to choose the free
parameter t in such a way that Pt [∑nk=1 kCk = n] gets large. Here one can choose
t to be the solution of the equation ∑nk=1 ϑk tk = n.
This argumentation is very close to the argumentation relying on complex anal-
ysis and generating functions. Indeed, it is easy to see that (4.1) is equivalent
to

hn = [tn ] [exp ( gΘ (t))] (4.2)

 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 15

with gΘ (t) as in (2.5). Furthermore, the choice of t is (almost) the solution of the
0 ( t ) = n. We have of course to justify (4.2) (or (4.1)).
saddle point equation tgΘ
But this follows immediately from the definition of hn and Lemma 2.2.
We prefer at this point to work with the second approach. We begin by writing
down the generating functions of the quantities we would like to study.

Lemma 4.1. We have for x ≥ 0 and s ∈ R

  
∞
1 n  ϑ
En exp −swn ( x ) = [t ] exp  gΘ (t) + (e−s − 1) ∑ k tk  .


(4.3)
hn k =b x c
k

Remark. Although the expressions in Lemmas 4.1 and 4.2 hold in broader gener-
ality, starting from Subsection 4.1 we will calculate moment generating functions
on the positive half-line, namely we can assume all parameters s1 , . . . , s` etc to
be non-negative, according to [9, Theorem 2.2].

Proof. It follows from the definitions of Pn [ · ] and wn ( x ) (see (2.3)) that

 
n n
1
∑ ∑ k ∏ ϑk k
C
hn En exp −swn ( x ) =


exp  − s C  (4.4)
n! σ∈ Sn k=b x c k =1

b x c−1 ∞
1
∑ ∏ ∏
C
= ϑk k (ϑk e−s )Ck
n! σ∈ Sn k =1 m=b x c

Applying now Lemma 2.2, we obtain

 
∞ b x c−1 ∞
tn ϑk k ϑ
∑ hn En exp −swn ( x ) ∑ t + e−s ∑ k tk 


= exp  (4.5)
n =0 n! k =1
k k=b x c
k
 
∞
ϑ
= exp  gΘ (t) + (e−s − 1) ∑ k tk  (4.6)
k=b x c
k

Equation (4.3) now follows by taking [tn ] on both sides. 

We are also interested in the joint behaviour at different points of the limit shape.
The results in Section 3 suggest that the increments of wn ( x j+1 ) − wn ( x j ) are
independent for x` ≥ x`−1 ≥ · · · ≥ x1 ≥ 0. It is thus natural to consider

wn (x) = wn ( x`∗ ), wn ( x`−

∗ ∗ ∗ ∗


1 ) − w n ( x ` ) , . . . , w n ( x 1 ) − w n ( x )
2 . (4.7)

We obtain
16 A. CIPRIANI AND D. ZEINDLER

Lemma 4.2. We have for x = ( x1 , . . . , x` ) ∈ R` with x` ≥ x`−1 ≥ · · · ≥ x1 ≥ 0 and

s = ( s 1 , . . . , s ` ) ∈ R`
  
` b x j +1 −1 c
1 n  ϑk k  
En exp −hs, wn (x)i [t ] exp  gΘ (t) + ∑ (e−s j − 1) ∑


= t
hn j =1 k=b x j c
k
(4.8)

with the convention x`+1 := +∞. The proof of this lemma is almost the same as
for Lemma 4.1 and we thus omit it.

4.1. Log-n-admissibility. The approach with which we first addressed the study
of the limit shape is derived from the saddle point method for approximating
integrals in the complex plane. We want to introduce the definition of log-n-
admissible function, generalizing the analogous concept introduced in [22]. We
stress that here, in comparison to the definition of log- (or equivalently Hayman)
admissibility used there, we consider a family of functions parametrized by n
for which log-admissibility holds simultaneously. The definition is therefore a
natural extension.

with gn (t) = ∑∞ k


Definition 4.3. Let gn (t) n ∈N k=0 gk,n t be given with radius of


convergence ρ > 0 and gk,n ≥ 0. We say that gn (t) n∈N is log-n-admissible if there
exist functions an , bn : [0, ρ) → R+ , Rn : [0, ρ) × (−π/2, π/2) → R+ and a sequence
(δn )n∈N s. t.

Saddle-point: For each n there exists rn ∈ [0, ρ) with

a n (r n ) = n (4.9)

Approximation: For all | ϕ| ≤ δn we have the expansion

ϕ2
gn (rn eiϕ ) = gn (rn ) + iϕan (rn ) − bn ( r n ) + R n ( r n , ϕ ) (4.10)
2
where Rn (rn , ϕ) = o ( ϕ3 δn−3 ).
Divergence: bn (rn ) → ∞ and δn → 0 as n → ∞.
Width of convergence: We have δn2 bn (rn ) − log bn (rn ) → +∞ as n → +∞.
Monotonicity: For all | ϕ| > δn , we have
   
Re gn (rn eiϕ ) ≤ Re g(rn e±iδn ) . (4.11)
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 17

The approximation condition allows us to compute the functions a and b exactly.

We have

an (r ) = rgn0 (r ), (4.12)
bn (r ) = rgn0 (r ) + r2 gn00 (r ) (4.13)

Clearly an and bn are strictly increasing real analytic functions in [0, ρ). The error
in the approximation can similarly be bounded, so that
 
Rn (r, ϕ) = ϕ3O rgn0 (r ) + 3r2 gn00 (r ) + r3 gn000 (r )

Having proved Lemma 4.1 we are now able to write down in a more explicit way
generating functions. What we are left with is trying to extract the coefficients
of the expansion given therein. This is the content of


Theorem 4.4. Let gn (t) n ∈N
be log-n-admissible with associated functions an , bn and
constants rn . Call
Gn := [tn ]e gn (t) .
Then

(1) Gn has the asymptotic expansion

1
Gn = √ (rn )−n bn (rn )−1/2 e gn (rn ) (1 + o (1)). (4.14)
2π
(2) Recall hn defined in (1.3). For the class of functions with weights as in (1.4),
1 α
1+ α
hn = q e2n (1 + o (1))
α +2
2π (α + 1)n 1+ α

respectively as n → +∞.

Remark. As it is explained in [15, Chapter VIII] it is possible to take into account

more error terms in the expansion of gn .

Proof of Theorem 4.4. The proof is exactly the same as in [22, Prop. 2.2] and we
thus give only a quick sketch of it, referring the reader to this paper for more
details. As in the well-known saddle point method, we want to evaluate the
integral
1 dz
I
exp ( gn (z)) .
2πi γ z n +1
18 A. CIPRIANI AND D. ZEINDLER

We choose as contour the circle γ := rn eiϕ with ϕ ∈ [−π, π ]. On ϕ ∈ [−δn , δn ]

after changing to polar coordinates we can expand the function g as
ϕ2
Z δn  
3 −3
exp gn (r ) + iϕan (r ) − bn (r ) + o ( ϕ δn ) − inϕ dϕ
−δn 2
We now choose rn such that a(rn ) = rn gn0 (rn ) = n in order to cancel the linear
terms in n. This allows us to approximate the integral on the minor arc with
a Gaussian. One shows that away from the saddle point (so for | ϕ| > δn ) the
contribution is exponentially smaller than on the minor arc and thus it can be
neglected. 

We would like to emphasize also that it will be not always possible to solve the
saddle point equation (4.9) exactly. However it is enough to find an rn such that
q 
a (r n ) − n = o b (r n ) (4.15)

holds.

4.2. Limit shape for polynomial weights. In this section we will derive the limit
shape for Young diagrams for the class of measures given by the weights. We
will not go into all the details to prove the log-n-admissibility for the most gen-
eral case, but will try to give a precise overview of the main steps nonetheless.
One important remark we have to make is that our parameter s will not be fixed,
but will be scaled and hence dependent on n. This comes from the fact that for
a fixed s (4.9) becomes a fixed point equation whose solution cannot be given
constructively, but has only an implicit form. We were not able to use this infor-
mation for our purposes, and hence preferred to exploit a less general, but more
explicit parameter to calculate asymptotics.

4.2.1. Limit shape. The main goal of this subsection is to prove that the weights
(1.4) induce a sequence of log-n-admissible functions of which we can recover
the asymptotics of gn (rn ). This will give us the limit shape of the Young diagram
according to Theorem 4.4. More specifically
1
Theorem 4.5. For the scaling n∗ = n α+1 , x ∗ = xn∗ and s∗ := s(n∗ )−α/2 , define the
functions
s Γ(α, x )
w∞ ( x ) := ,
Γ ( α + 1)
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 19

2 Γ(α + 1, x )2 Γ(α, x )
σ∞ ( x ) := − + .
2Γ(α + 1)Γ(α + 2) Γ(α + 1)
Then
wn ( x ∗ ) − ( n∗ )α w∞
s (x)
L

2

ens ( x ) :=
w −→ N 0, σ∞ ( x ) .
(n∗ )α/2
s ( x ) is the desired limit shape.
In particular w∞

Remark. We note that the limit shape matches the one obtained in [13, Thm. 4.8]
and also the one obtained in the present paper in the randomized case (cf. the
r ( x ) of Prop. 3.3).
definition of w∞

Theorem 4.6. Define the cumulant generating function as

h i tm
Λ(s) := log En e sw̃n ( x )
= ∑ qm .
m ≥1
m!

We then have for m ≥ 2

qm = κm (1 + o (1)). (4.16)

with
−α
(n∗ )α ∗ Γ ( α + 1,

m x)
κm = [s ] 1−s
α Γ ( α + 2)
 ∗
 
e−s − 1 Γ(α + k, x )

Γ(α + 1, x ) ∗
k
+
Γ ( α + 1 ) k ≥0 ∑ k! Γ ( α + 2)
s . (4.17)

We can also determine the behavior of the increments of the function wn (·).
We will consider first the more general case and then give the example of the
two-increment case (refer to (4.7) with ` = 2).

Theorem 4.7. (1) For ` ≥ 2 and x` ≥ x`−1 ≥ · · · ≥ x1 ≥ 0, let

e sn (x) = w
ens ( x` ), w
ens ( x`−1 ) − w
ens ( x` ), . . . , w
ens ( x1 ) − w
ens ( x2 ) .


w

Set x`+1 = +∞. For 1 ≤ j < i < ` we have that

s
ens ( x j ) − w
ens ( x j+1 ), w
ens ( xi ) − w
ens ( xi+1 )


w
e∞ ( xi , x j ) := lim Cov w
n→+∞
(Γ(α + 1, xi ) − Γ(α + 1, xi+1 )) Γ(α + 1, x j ) − Γ(α + 1, x j+1 )


= .
Γ ( α + 1) Γ ( α + 2)
20 A. CIPRIANI AND D. ZEINDLER

Remark. Let us comment briefly on Thm. 4.7. What we obtained in this result is
most unexpected: cycle counts are asymptotically independent under very mild
assumptions (see Lemma 2.1). The assumption of the lemma holds in our case as
the growth of the parameters ϑn is algebraic. The fact that the increments depend
on disjoint sets of cycles would have suggested the asymptotic independence of
wn (y∗ ) from wn ( x ∗ ) − wn (y∗ ). We are aware of the work of [4] handling this
issue in the case of the Ewens sampling formula, in particular showing that
partial sums of cycle counts need not converge to processes with independent
increments. Our result extends this idea in the sense that it shows the explicit
covariance matrix for a whole category of generating functions. It would be
interesting to provide a heuristic explanation for this theorem.

4.2.2. Log-n-admissibility. In order to determine the limit shape we would like

to prove the log-n-admissibility of the function explicited in (4.3). To be more
precise, what we have to prove is

1
Lemma 4.8. Let s ≥ 0, and recall n∗ = n α+1 , s∗ = s(n∗ )α/2 . The function

∞ k α −1 + O k β −1 k


−s∗
gΘ ( t ) + ( e − 1) ∑ t
k=b x ∗ c
Γ ( α + 1)

is log-n-admissible for all x ≥ 0, with gΘ (t) as in (2.5) and

r n : = e −vn (4.18)

with

∗ −α/2 Γ ( α + 1,
 
∗ −1 x)
vn := (n ) 1 − s(n ) .
Γ ( α + 2)
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 21

Proof of Lemma 4.8. Saddle-point and approximation: We start first with the
case β = 0. By doing so one obtains that
+∞ +∞
kα ∗ kα
a (r n ) = ∑ Γ ( α + 1 )
e−kvn + (e−s − 1) ∑
Γ ( α + 1 )
e−kvn
k =1 k=b x ∗ c

∗ Γ(α + 1, x ∗ vn )
= ( v n ) − α −1 + O (1 ) + ( e − s − 1 )
Γ ( α + 1)
∗
+(e−s − 1)O v− α


n (4.19)
Γ(α + 1, x )
 
= n 1 + (α + 1)s(n∗ )−α/2 + O (n∗ )−α


Γ ( α + 2)
Γ(α + 1, x )
  2   
s s 
∗ −α/2
+n − ∗ α/2 + O + O (n )
(n ) (n∗ )α Γ ( α + 1)
 
+O 1 + v− n
α ∗ −α/2
( n )
= n + O (n∗ ) . (4.20)

We also have that

+∞
!
kα+1 −kvn
b (r n ) = O ∑ Γ ( α + 1) e ∼ (α + 1)(n∗ )α+2 + O (n) . (4.21)
k =1

Therefore (4.15) holds true for all α. In the case where β is turned on, we
obtain by performing similar steps that
 
∗ β +1
a (r n ) = n + O ( n ) .

Then (4.15) is satisfied if

β+1 α+2 α
< ⇐⇒ β < (4.22)
α+1 2( α + 1) 2

which holds by assumption.

Divergence: By the above calculations we set δn := (n∗ )−ξ with α +3
3 <ξ<
α +2
2 . This position holds also in the case β > 0.
Monotonicity: In the region | ϕ| = o (1) we wish to show that
 
iϕ
g rn e = g(rn )(1 + o (1)). (4.23)

First remember that gn rn ei±δn = O ((n∗ )α ) by Lemma 2.3. Thus here

we have:
22 A. CIPRIANI AND D. ZEINDLER

(1) if ϕ = o (vn ), then by a change of variable t (vn − iϕ)t

kα−1 −k(vn −iϕ)
∑ Γ ( α + 1)
e
k≥b x ∗ c
Z +∞
(vn − iϕ)−α Γ(α, x )
∼ tα−1 e−t dt = (vn − iϕ)−α
Γ ( α + 1) x Γ ( α + 1)
∗
which is asymptotic to (n∗ )α . Considering the factor e−s − 1 we
obtain that the summand is negligible with respect to Re g(rn e±iδn ) .

(2) If ϕ 6= o (vn ), then

kα−1 −k(vn −iϕ)
∑ Γ ( α + 1) e
k≥b x ∗ c
Z +∞
(vn − iϕ)−α Γ(α, x − ixϕn∗ )
∼ tα−1 e−t dt = (vn − iϕ)−α
Γ ( α + 1) ∗
x −ixϕn +o (1) Γ ( α + 1)
and afterwards use the fact that Γ(α, x + iy) = O yα−1 for |y| large.

Hence the RHS of (4.24) becomes

   
∗ α −1 1− α ∗ α −1 −1
O (n ) (vn − iϕ) = O (n ) ϕ

As ϕ 6= o (vn ), we obtain that O (n∗ )α−1 ϕ−1 = O ((n∗ )α ) o (1)

which is enough to show (4.23).

(3) To conclude we consider the case | ϕ| > C: the function gn rn eiϕ

is bounded there by a constant uniform in n, and then by bounding

gn rn eiϕ through its modulus we have


     
iϕ ±iδn ∗ −α/2
Re gn (rn e ) ≤ Re g(rn e ) 1 + O (n ) . (4.24)

In order to show Thms. 4.5, 4.6 and 4.7 we need to prove first an auxiliary
proposition.
1
Proposition 4.9. For the scaling n∗ := n α+1 , x ∗ := xn∗ and s∗ := s(n∗ )−α/2 the
equality
∗
( e − s − 1) ∑ kα−1 rnk
k≥b x ∗ c

s2 Γ(α + 1, x )2 2
 
∗ α/2
= −s(n ) Γ(α, x ) + Γ(α, x ) − s + o (1) (4.25)
2 Γ ( α + 2)
holds asymptotically as n → +∞.
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 23

Proof. We apply Lemma 2.4 with

f (t) := tα−1 e−tvn ,

∗ −α/2 Γ ( α + 1, x )
 
∗ 1
zn = xn , vn = ∗ 1 − s(n ) and
n Γ ( α + 2)
sx Γ(α + 1, x )
zn vn = x − .
(n∗ )α/2 Γ(α + 2)

The first term of the expansion is

∗
(e−s − 1) (n∗ )α Γ(α, x )
s2
= −s(n∗ )α/2 Γ(α, x ) + Γ(α, x ) + o (1)
2

because (s∗ )3 (n∗ )α = o (1). If β > 0 instead we obtain

∗
(e−s − 1) (n∗ )α Γ(α, x )
s2
= −s(n∗ )α/2 Γ(α, x ) + Γ(α, x ) + o (1)
2

To calculate the expansion up to a O (1) term it is sufficient to consider for k = 1

Γ(α + 1, x ) − β
 
−s∗ ∗ α
(e − 1) ( n ) sn
Γ ( α + 2)
Γ(α + 1, x ) 2
= −Γ(α + 1, x ) s + o (1)
Γ ( α + 2)

This tells us that

∗
( e − s − 1) ∑ kα−1 rnk
k≥b x ∗ c

s2 Γ(α + 1, x )2 2
 
∗ α/2
= −s(n ) Γ(α, x ) + Γ(α, x ) − s + o (1) (4.26)
2 Γ ( α + 2)

As for the remainder, we can find an a priori bound on the Bernoulli polynomials
independent of n on x ∈ [0, 1]. Furthermore,

α −2
∗
   
(e−s − 1) f (b x ∗ c) = O s(n∗ )−α/2 (b xn∗ c)α−1 e−x+o(1) = O s(n∗ ) 2 ,
24 A. CIPRIANI AND D. ZEINDLER

which is small compared to the magnitude of the leading coefficient in s. More-

over
Z +∞
B1 ( x 0 − b x 0 c) f 0 ( x 0 )dx 0 (4.27)
xn∗
Z +∞ Z +∞
0 0 0 0
≤C | f ( x )|dx = C e− x vn (−vn ( x 0 )α−1
xn∗ xn∗
0 α −2 0
+(α − 1)( x ) )dx . (4.28)
zn
With the same substitution x 0 := a0 y we can interchange limit and integral by
the dominated convergence theorem to obtain
 α/2 !
zn
(4.28) = O Γ(1 + α, a0 ) .
a0
∗
Combining this with the first order expansion of (e−s − 1) we obtain
 
∗ α/2−α/2
(4.27) = O s(n ) = s O (1) .


Proof of Thms. 4.5 and 4.6. To determine the behavior of Gn we would like to use
Lemma 4.1. By (4.3)
  
+∞
1 n  ∗ ϑ
En exp −s∗ wn ( x ∗ ) [t ] exp  gΘ (t) + (e−s − 1) ∑ k tk  .


=
hn k=b x ∗ c
k

∞ ∗
We have shown that gn (t) = gΘ (t) + (e−s − 1) ∑+
k=b x ∗ c
ϑk k
k t is log-n-admissible.
Therefore Thm. 4.4 tells us how Gn behaves, and we have more precisely to
recover three terms. In first place we collect the terms for the asymptotic of
e gn (rn ) : one is
−α
Γ(α) (n∗ )α ∗ −α/2 Γ ( α + 1,

−α x)
gΘ (r n ) ∼ v = 1 − s(n )
Γ ( α + 1) n α Γ ( α + 2)
(n∗ )α sΓ(α + 1, x )
= + (n∗ )α/2
α Γ ( α + 2)
s2 Γ(α + 1, x )2  
+ + O (n∗ ) β
2Γ(α + 2)Γ(α + 1)
kα−1 +O(k β−1 ) k
given by Lemma 2.3. The other is ∑k≥b x∗ c Γ ( α +1)
rn which we can approxi-
mate through Prop. 4.9. Secondly we obviously have
Γ(α + 1, x )
−n log(rn ) = (n∗ )α − (n∗ )α/2 s .
Γ ( α + 2)
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 25

Thirdly the behavior of b(rn ) was determined in (4.21). All in all

Γ(α, x )
  
∗ α 1
e g(rn )−n log(rn )
= exp (n ) + s(n∗ )α/21+
Γ ( α + 1)
α
s2 Γ(α + 1, x )2 Γ(α, x )
 
+ − +
2 2Γ(α + 1)Γ(α + 2) Γ(α + 1)
 
3 ∗ −3α/2 ∗ α
+s O (n ) (1 + ( n ) ) . (4.29)

Theorem 4.4 yields the behavior of hn , and the same theorem allows us to con-
clude plugging in (4.14) the expressions obtained in (4.21), (4.29) and hn of (2)
s is the limit shape, in the same fashion the result
therein. It is also clear that w∞
followed in the proof of Corollary 3.4.
For cumulants what we have to do is considering the logarithm of the expansion
(rn )−n b(rn )−1/2 e gn (rn )−n log(rn ) . We claim that it suffices to consider simply the
logarithm of the expression (4.29). In fact,

−α−2 !!
Γ(α, x )

∗ α +2 ∗
log(b(rn )) = log O (n ) 1−s
Γ ( α + 1)
kα
= C1 log(n) + C2 ∑ sk (n∗ ) 2
k ≥0

α (2− k )
whilst each coefficient of sk in gn (rn ) − n log(rn ) is of order (n∗ ) 2 (compare
(4.29)). This confirms that the main contribution stems from (4.29). 

Proof of Thm. 4.7. For multiple increments repeating the proof of Thm. 4.5 tells
us that for a vector wn (x∗ ) as in (4.7) with length ` > 2 we can set

(n∗ )−α/2
v n : = ( n ∗ ) −1 1− (s Γ(α + 1, x` )
Γ ( α + 2) `
!!
`−1
+ ∑ s`−k (Γ(α + 1, x`−1−k ) − Γ(α + 1, x`−k )) .
k =1
26 A. CIPRIANI AND D. ZEINDLER

We deduce from this that

v−
n
α (n∗ )α (n∗ )α/2
gΘ (r n ) ∼ = −
α α Γ ( α + 2)
!
`−1
s` Γ(α + 1, x` ) + ∑ (s`−k−1 (Γ(α + 1, x`−k−1 ) − Γ(α + 1, x`−k ))
k =1
1
+ (s Γ(α + 1, x` )
2Γ(α + 2)Γ(α + 1) `
!2
`−1
+ ∑ (s`−k−1 (Γ(α + 1, x`−k−1 ) − Γ(α + 1, x`−k ))
k =1
+ o (1) . (4.30)
 ∗  x ∗ −1
−s j +1 ϑk k
Since the coefficients of the form e j − 1 ∑k= x∗ k rn do not give a contribu-
j
tion to covariances, the mixed terms will stem from the expansion of the square
in (4.30). In particular we see that the coefficient of si s j , for 1 ≤ j < i < `, is
(Γ(α + 1, xi ) − Γ(α + 1, xi+1 )) Γ(α + 1, x j ) − Γ(α + 1, x j+1 )


.
2Γ(α + 1)Γ(α + 2)


4.3. Functional CLT for wn (·). As in the randomized setting, a functional CLT
can be obtained here too. Unlike the previous case though we do not have the
independence of cycle counts, hence we will have to show the tightness of the
fluctuations as in Sec. 3.3 in two steps (cf. [17]). The result we aim at is, precisely
as before,

ens : R+ → R (see Thm.

Theorem 4.10. The process w 4.5) converges weakly with
respect to Pn as n → ∞ to a continuous process w
e∞s : R+ → R with w
e∞s (x) ∼

N (0, (σ∞s ( x ))2 ) and whose increments are not independent. The covariance structure

is given in Thm. 4.7.

Proof. We will proceed as in the proof of Thm. 3.6. Having shown already the
behavior of the increments in Thm. 4.7 what we have to tackle now is their
tightness. The proof’s goal is again, analogoulsy as Lemma 3.7. However the
evaluation of the LHS of (3.12) is more difficult this time; one possible approach
is present in [11] and is based on Pólya’s enumeration lemma and the calculation
of factorial moments of cycle counts. We prefer rather to follow again [17]. We
will proceed in two main steps.
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 27

i) We define for 0 < t < 1 the measure Pt as in Section 3. By repeating the

proof of [17, Lemma 2.1] we see that

Pt ∑ kCk = n = t n h n e gΘ ( t ) .
 

Mimicking Hansen’s strategy one can also prove that for arbitrary func-
tions Ψ : S → C, where S := ∪n Sn and Ψn : Sn → C s. t. Ψn =
Ψ(C1 , . . . , Cn , 0, 0, . . .)

Et [ Ψ ] e gΘ ( t ) = ∑ tn hn E [Ψn ] + Ψ(0, 0, 0, . . .). (4.31)

n ≥1

ii) As a formal power series identity (4.31) holds for |t| < 1, thus we decide
to set, for x1 , x2 as in the assumptions,

Ψ(k1 , k2 , . . .) :=
 2  !2
 
x∗  x2∗ 
ϑ ϑj j
=  n−γ ∑ k i − i rni  n−γ ∑ k j − rn
i = x ∗ +1
i j = x ∗ +1
j
1
 2
x∗ x∗
ϑ ϑi i
 
=  n −γ ∑ k i − i ti + n −γ ∑ t − e−ivn 
i = x ∗ +1
i i = x ∗ +1
i
1 1

x2∗ x2∗
!2
ϑj ϑj  j 
n−γ ∑ k j − t j + n−γ ∑ t − e− jvn
j = x ∗ +1
j j = x ∗ +1
j

for γ > 0 to be tuned appropriately later. We now calculate, using the

independence of cycle counts under the randomized measure and the fact
28 A. CIPRIANI AND D. ZEINDLER

that VarPt [Ci ] = Et [Ci ] = ϑi i

i t,
 
x∗ x2∗
!
ϑi i  ϑj
Et [Ψ] = n−4γ  ∑ t ∑∗ j t j
i = x ∗ +1
i j = x +1
1
 2 ∗ !2
x∗ x2
ϑj j
ϑi i
+n−4γ  ∑ (t − e−ivn ) ∑ (t − e− jvn )
i = x1∗ +1
i j = x ∗ +1
j
  
x∗ x∗
ϑi i   ϑi i
 
+2n−4γ  ∑ t ∑ t − e−ivn 
i = x ∗ +1
i i = x ∗ +1
i
1 1

x2∗
!2
ϑj
· ∑∗ j (t j − e− jvn )
j = x +1
+...
  ! ∗ 
x∗ x2∗ x
ϑi i  ϑj ϑ  
+2n−4γ  ∑∗ i
t ∑∗ j t j  ∑∗ ii ti − e−ivn 
i = x1 +1 j = x +1 i = x +1 1

x2∗
!
ϑj j
· ∑∗ j
(t − e− jvn )
j= x +1
(1) (2) (9)
=: GΘ (t, n) + GΘ (t, n) + . . . + GΘ (t, n).
ϑj j
Let us define gba (z) := ∑bj=a j z . From (4.31) we obtain

1 n  gΘ ( t ) (1)  1 
(9)

E [Ψn ] = [t ] e GΘ (t, n) + . . . + [tn ] e gΘ (t) GΘ (t, n) .
hn hn

We therefore
 obtain several terms and we will analyze them one by one.
1 n (1)
(a) hn [ t ] e gΘ (t) GΘ (t, n) . One has

n−4γ n  gΘ (t) x∗ x2∗


[t ] e g x ∗ +1 ( t ) g x ∗ +1 ( t )
hn 1
   ∗ !
∗ x
n−4γ n gΘ (t)+log gxx∗ +1 (t) +log gx2∗ +1 (t)
= [t ] e 1 (4.32)
hn

We want to apply the saddle-point method to the sequence of func-

x∗
   
∗
gΘ (t)+log gxx∗ +1 (t) +log gx2∗ +1 (t)
tions gn (t) := e 1 to extract coefficients.
Our first target is to show the log-n-admissibility. We consider again
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 29

the radius rn := e−vn with vn := (n∗ )−1 . In this case as in (4.20)

∗ x∗
+∞
kα ∑ xx1∗ +1 kα
Γ ( α +1)
e−kvn ∑ x2∗ +1 kα
Γ ( α +1)
e−kvn
a (r n ) = ∑ Γ ( α + 1)
e −kvn
+ ∗
gxx∗ +1 (rn )
+ x∗
=
k =1 1
gx2∗ +1 (rn )
x∗ x∗
∑ x1∗ +1
kα
Γ ( α +1)
e−kvn ∑ x2∗ +1 kα
Γ ( α +1)
e−kvn
− α −1
= (vn ) + O (1) + x∗ kα −1 −kvn
+ x∗
∑ x1∗ +1 Γ ( α +1)
e ∑ x2∗ +1 kα −1 −kvn
e
Γ ( α +1)

(vn )−α−2 Cα+1, x, x1 (vn )−α−2 Cα+1, x, x2

= n+ +
v−n
α −1
Cα, x, x1 v−n
α −1
Cα, x, x2
 
= n + O v− n
1
.

where Cα+1, x, x1 , Cα, x, x1 , Cα+1, x, x2 and Cα, x, x2 are constants inde-

pendent of n. Very little changes also in the computions for b(rn )
which lead to b(rn ) = O (n∗ )α+2 , yielding the saddle point equa-

tion (4.15). As far as monotonicity is concerned,

 heuristically
 one can
∗
prove it using the fact that the order of log gxx∗ +1 (t) is smaller that
1
that of the leading term gΘ (t) (as one can already notice for example
in the computations for a(rn ) and b(rn ) above). Since calculations are
straightforward we omit them. Then by Thm. 4.4 one has that (recall
that hn = [tn ]e gΘ (t) )
  
−4γ  1 n gn (t)  x2∗
   x∗ 
n  hn [t ]e  =  gx1∗ +1 (rn ) gx∗ +1 (rn )(1 + o (1)) =


≤ Cn−4γ (vn )−α (Γ(α, x ) − Γ(α, x1 )) (vn )−α (Γ(α, x2 ) − Γ(α, x ))

 

≤ Cn−4γ (vn )−2α |( x − x1 ) ( x2 − x )|

 
= O (( x − x1 )( x2 − x )) = O ( x2 − x1 )2

provided that n−4γ (n∗ )2α = O (1) iff γ := α

2( α +1)
. We highlight that
in this case nγ is precisely the variance of the process (cf. Thm. 4.5).
Here we have also used the fact that the incomplete Gamma function
is continuous
 on a compact
 [0, K ] for some K large.
1 n ( j)
(b) hn [ t ] e gΘ (t) GΘ (t, n) , 2 ≤ j ≤ 9. We want to show that all these
(3)
terms are O ( x2 − x1 )2 as well. We take for example GΘ (t, n) :=


 ∗ 2  ∗ 
ϑj x ϑj
∑ xj= x1∗ +1 i (t j − e− jvn ) ∑ j=2 x∗ +1 j t j . We define the auxiliary func-
ϑj j
tion hba (t) := ∑bj=a j (t − e− jvn ). We wish to apply again the saddle
30 A. CIPRIANI AND D. ZEINDLER

point method. In fact we decompose h as

x∗ ϑ j − jvn
∗ ∗
h xx∗ +1 (t)
1
= gxx∗ +1 (t) −
1
∑
∗ j
e .
x1 +1

We now have
 
2 x∗
x2∗ x2∗ ϑ j − jvn
 ∗ ∗
(t)  ∑
(3)
GΘ (t, n) = gxx∗ +1 (t) g x ∗ +1 (t) − 2gxx∗ +1 (t) g x ∗ +1 e 
1 1 ∗
x +1
j
1
 2
x∗ ϑ j − jvn ∗
+ ∑ e  g x2∗ (t).
j x +1
∗
x1 +1

(3)
It is clear then that in the first-order asymptotics GΘ (as well as all
j
other terms involving t j − rn ) will not give any contribution, because
(3)
GΘ (rn , n) = 0. We ask then ourselves if admissibility holds true for
each one of these terms, but this is fairly easy because of the previous
computations. Indeed we can start for example with the middle one.
We have already shown in (a) that
1  ∗ x∗

n−4γ [tn ] e gΘ (t) gxx∗ +1 (t) gx2∗ +1 (t)
hn 1
 ∗ ϑ 
is log-n-admissible and the term ∑ xx∗ +1 jj e− jvn is a constant inde-
1
pendent of t. Both the other two summands are log-n-admissible
∗
with rn = e−vn := e−1/n : calculations can be performed in the same
fashion as (a) and since they are direct we skip them.

4.4. Large deviations estimates. We are able to prove large deviations estimates
for wn (·) thanks to our method as well. In fact, knowing the behavior of the
Laplace transform enables us to compute the asymptotics of the Young diagram
in the limit. More precisely, let σn be the limit variance as in Thm. 4.5. Define
the normalized moment generating function and its logarithm as
( wn ( x ) − ( n∗ )α w∞
s ( x ) 
 
M (s) := E exp s ,
σn
Λ(s) := log M (s).

The strategy we adopt was first exploited in [22, Theorem 4.1], and relies on the
fact that
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 31

Proposition 4.11. There exist functions ξ (n) = O ((n∗ )α ), σ (n) = O (n∗ )α/2 such

that for all s = O (σ (n)) we obtain

s2  
Λ(s) = + O ξ ( n ) σ ( n ) −3 s 3 .
2
It follows than that
 
0 −3
Λ (s) = O ξ (n)σ(n) s2 ,
 
Λ00 (s) = O ξ (n)σ(n)−3 s.

From this we derive

Proposition 4.12. For all a = O (σn ) let δ := O ξ (n)σ(n)−3 . Then we have

 ( wn ( x ) − ( n ∗ )α w∞
s (x)
      
P  − a < e = 1 − e−2 (1 + δ) exp − a2 /2 + O (δ + ea) .

 
σn
The error terms are absolute.

Proof. The proof can be performed analogously as [22], as we know that (4.29)
holds. 

At this juncture we would like to apply our method to a simple but illustrative
case.

4.5. An example: the case gΘ (t) = (1 − t)−1 . We would like to begin by the
easiest case, in other words to derive the limit shape for one point. We remark
that here all our computations were performed using the function gΘ (t) = t(1 −
t)−1 . This does not affect the computations of the limit shape as it will “only”
make a constant appear, which will be later simplified in all calculations.

Proposition 4.13. For all x ∈ R+

√ √
wn ( x n) − ne− x L
  
−x 1 −x 2
→ N 0, e 1 − e ( x + 1)
n1/4 2
s ( x ) : = e− x (cf. Thm. 4.5 plugging in α = 1).
In particular, the limit shape is w∞

We now pass to the joint behavior of (wn ( x1 ), . . . , wn ( x` )) which can be recov-

ered from
32 A. CIPRIANI AND D. ZEINDLER

Proposition 4.14. Let ` ∈ N+ . For all x1 , . . . , x` ∈ R+ , set xk∗ := xk n1/2 ; then we

have
!
wn xk∗ − n1/2 e− xk


L
→ N (0, Σ)
n1/4
k=1, ..., `

with Σ ∈ M` (R) defined through

Σk, k = − 12 e− xk e− xk xk 2 + 2 e− xk xk + e− xk − 1


(4.33)
Σk, j = − 21 e− xk e− x j xk x j + e− x j xk + e− x j x j + e− x j − 1 , j 6= k.

Acknowledgments

We thank Sabine Jansen for pointing out a mistake in a previous version of the
paper.

Appendix A. Euler Maclaurin formula with non integer boundaries

We prove in this section a slight extension of Euler Maclaurin formula, which

allows to deal also with non-integer summation limits.

Theorem A.1. Let f : R → R be a smooth function, Bk ( x ) be the Bernoulli polynomials

and c < d with c, d ∈ R. We then have for p ∈ N
Z d
∑ f (k) =
c
f ( x ) dx − B1 (d − bdc) f (d) − B1 (c − bcc) f (c) (A.1)
bcc≤k<d
p
Bk+1 (d − bdc) f (k) (d) − Bk+1 (c − bcc) f (k) (c)
+ ∑ (−1)k+1 k!
k =1
(−1) p+1
Z d
+ B p+1 ( x − b x c) f ( p+1) ( x ) dx
( p + 1) ! c

Proof. The proof of this theorem follows the same lines as the proof of the Euler-
Maclaurin summation formula with integer summation limits, see for instance
[1, Theorem 3.1]. We give it here though for completeness. Our proof considers
/ Z. The argumentation for d ∈ Z is completely similar. One
only the case d ∈
possible definition of the Bernoulli polynomials is by induction:

B0 (y) ≡ 1, (A.2)
Z 1
Bk0 (y) = kBk−1 (y) and Bk (y) dy = 1 for k ≥ 1. (A.3)
0
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 33

In particular, we have B1 (y) = y − 12 . We now have for m ∈ Z

Z m +1 Z m +1
f (y) dy = B0 (y − m) f (y) dy
m m
Z m +1
= [ B1 (y − m) f (y)]|m +1
y=m − B1 (y − m) f 0 (y) dy
m
Z m +1
1 1
= f ( m ) + f ( m + 1) − B1 (y − byc) f 0 (y) dy.
2 2 m

since B1 (0) = − 21 and B1 (1) = 21 . We obtain

bdc Z bdc Z bdc

1 1
∑ f (k) =
bcc
f ( x ) dx + f (bcc) + f (bdc) +
2 2 bcc
B1 (y − byc) f 0 (y) dy.
k=bcc

Furthermore, we use
Z d Z d
1
f (y) dy = f (bdc) + B1 (d − bdc) f (d) − B1 (y − byc) f 0 (y) dy.
bdc 2 bdc

and get
bdc Z d Z d
1
∑ f (k) =
bcc
f ( x ) dx + f (bcc) − B1 (d − bdc) f (d) +
2 bcc
B1 (y − byc) f 0 (y) dy.
k=bcc

The argumentation for replacing bcc by c is similar. One gets

Z d
∑ f (k) =
c
f ( x ) dx − B1 (c − bcc) f (c) − B1 (d − bdc) f (d)
bcc≤k<d
Z d
+ B1 (y − byc) f 0 (y) dy.
c
Rd
The theorem now follows by successive partial integration of c B1 (y − byc) f 0 (y) dy.


References
[1] Apostol, T. Introduction to analytic number theory. Springer-Verlag, New York, 1984.
[2] Apostol, T. M. An elementary view of Euler’s summation formula. Amer. Math. Monthly
106, 5 (1999), 409–418.
[3] Arratia, R., Barbour, A., and Tavaré, S. Logarithmic combinatorial structures: a probabilistic
approach. EMS Monographs in Mathematics. European Mathematical Society (EMS), Zürich,
2003.
34 A. CIPRIANI AND D. ZEINDLER

[4] Babu, G. J., Manstavičius, E., and Zacharovas, V. Limiting processes with dependent
increments for measures on symmetric group of permutations. In Probability and number
theory—Kanazawa 2005, vol. 49 of Adv. Stud. Pure Math. Math. Soc. Japan, Tokyo, 2007,
pp. 41–67.
[5] Beltoft, D., Boutillier, C., and Enriquez, N. Random young diagrams in a rectangular
box. e-print (Dec. 2010). http://arxiv.org/abs/1008.0846.
[6] Betz, V., Ueltschi, D., and Velenik, Y. Random permutations with cycle weights. Ann.
Appl. Probab. 21, 1 (2011), 312–331.
[7] Billingsley, P. Convergence of probability measures, second ed. Wiley Series in Probability
and Statistics: Probability and Statistics. John Wiley & Sons Inc., New York, 1999. A Wiley-
Interscience Publication.
[8] Bogachev, L. V. Unified derivation of the limit shape for a ”conservative” ensemble of
random partitions. ArXiv e-prints (Nov. 2011).
[9] Chareka, P. A finite-interval uniqueness theorem for bilateral laplace transforms. Inter-
national Journal of Mathematics and Mathematical Sciences 2007 (2007), Article ID 60916, 6
p.–Article ID 60916, 6 p.
[10] Deift, P. Integrable systems and combinatorial theory. Notices Amer. Math. Soc. 47 (2000),
631–640.
[11] DeLaurentis, J. M., and Pittel, B. G. Random permutations and Brownian motion. Pacific
J. Math. 119, 2 (1985), 287–301.
[12] Ercolani, N., and Ueltschi, D. Cycle structure of random permutations with cycle
weights. Preprint, 2011.
[13] Erlihson, M. M., and Granovsky, B. L. Limit shapes of Gibbs distributions on the set of
integer partitions: the expansive case. Ann. Inst. Henri Poincaré Probab. Stat. 44, 5 (2008),
915–945.
[14] Ewens, W. J. The sampling theory of selectively neutral alleles. Theoret. Population Biology 3
(1972), 87–112; erratum, ibid. 3 (1972), 240; erratum, ibid. 3 (1972), 376.
[15] Flajolet, P., and Sedgewick, R. Analytic Combinatorics. Cambridge University Press, New
York, NY, USA, 2009.
[16] Fristedt, B. The structure of random partitions of large integers. Trans. Am. Math. Soc. 337,
2 (1993), 703–735.
[17] Hansen, J. C. A functional central limit theorem for the Ewens sampling formula. J. Appl.
Probab. 27, 1 (1990), 28–43.
[18] Hoppe, F. M. The sampling theory of neutral alleles and an urn model in population genet-
ics. J. Math. Biol. 25, 2 (1987), 123–159.
[19] Kowalski, E., and Nikeghbali, A. Mod-Poisson convergence in probability and number
theory. Int. Math. Res. Not. IMRN 2010, 18 (2010), 3549–3587.
[20] Logan, B. F., and Shepp, L. A. A variational problem for random Young tableaux. Advances
in Math. 26, 2 (1977), 206–222.
 FLUCTUATIONS NEAR THE LIMIT SHAPE OF RANDOM PERMUTATIONS 35

[21] Macdonald, I. G. Symmetric functions and Hall polynomials, second ed. Oxford Mathematical
Monographs. The Clarendon Press Oxford University Press, New York, 1995. With contri-
butions by A. Zelevinsky, Oxford Science Publications.
[22] Maples, K., Nikeghbali, A., and Zeindler, D. The number of cycles in a random permu-
tation. Electron. Commun. Probab. 17 (2012), no. 20, 1–13.
[23] Pólya, G. Kombinatorische anzahlbestimmungen für gruppen, graphen, und chemische
verbindungen. Acta Mathematica 68 (1937), 145–254.
[24] Temperley, H. N. V. Statistical mechanics and the partition of numbers. ii. the form of
crystal surfaces. Proc. Camb. Phil. Soc. 48 (1952), 683–697.
[25] Vershik, A. M. Statistical mechanics of combinatorial partitions, and their limit configura-
tions. Funktsional. Anal. i Prilozhen. 30, 2 (1996), 19–39, 96.
[26] Vershik, A. M., and Kerov, S. V. Asymptotic behavior of the Plancherel measure of the
symmetric group and the limit form of Young tableaux. Dokl. Akad. Nauk SSSR 233, 6 (1977),
1024–1027.

Weierstraß-Institut, Mohrenstraße 39, 10117 Berlin, Germany

E-mail address: [email protected]

Sonderforschungsbereich 701, Fakultät für Mathematik, Universität Bielefeld, Post-

fach 10 01 31, 33501 Bielefeld, Germany

E-mail address: [email protected]