380 Electronic Communications in Probability
where by equations 1, 3 and 4 ε
n
α := −P n
Π−α, x , n
2 3
λ
max
− 2 y o
− P ¦
Π−α, x , ˜ Π y, α
© − P
n n
2 3
λ
min
+ 2 x , ˜ Π y, α
o + P
¦ N ∆
1,n
= 0 ©
P ¦ ˜
Π y, α ©
+ P {Π−α, x} P ¦
N ∆
2,n
= 0 ©
− P {Π−α, x} P ¦ ˜
Π y, α ©
. Using the triangular inequality, we obtain:
|ε
n
α| ≤ 6 max
P {Π−α, x} , P
¦ ˜ Π y, α
© .
As { Π−α, x } ⊂ {n
2 3
λ
min
+ 2 −α}, we have P
{Π−α, x } ≤ P{n
2 3
λ
min
+ 2 −α} −−→
n →∞
F
− GU E
−α −−→
α→∞
0 . We can apply the same arguments to
{ ˜ Π y, α } ⊂ {n
2 3
λ
max
− 2 α}. We thus obtain: lim
α→∞
lim sup
n →∞
|ε
n
α| = 0 . 6
The difference P ¦
N ∆
1,n
= 0 , N ∆
2,n
= 0 ©
−P ¦
N ∆
1,n
= 0 ©
P ¦
N ∆
2,n
= 0 ©
in the right -
hand side of 5 converges to zero as n → ∞ by Theorem 1 for every α large enough. We
therefore obtain lim sup
n →∞
|u
n
| = lim sup
n →∞
|ε
n
α| . The lefthand side of the above equation is a constant w.r.t. α while the second term whose
behaviour for small α is unknown converges to zero as α → ∞ by 6. Thus, lim
n →∞
u
n
= 0. The mere definition of u
n
together with Tracy and Widom fluctuation results yields lim
n →∞
P n
n
2 3
λ
min
+ 2 x , n
2 3
λ
max
− 2 y o
=
1 − F
− GU E
x
F
+ GU E
y . This completes the proof of Corollary 1.
2.2 Application: Fluctuations of the ratio of the extreme eigenvalues in
the GUE
As a simple consequence of Corollary 1, we can easily describe the fluctuations of the ratio
λ
max
λ
min
. The counterpart of such a result to Gaussian Wishart matrices is of interest in digital communication see
[ 4
] for an application in digital signal detection.
Corollary 2. Let M be a nn matrix from the GUE. Denote by λ
min
and λ
max
its smallest and largest eigenvalues, then
n
2 3
λ
max
λ
min
+ 1
D
−−→
n →∞
− 1
2 λ
−
+ λ
+
, where
D
− → denotes convergence in distribution, λ
−
and λ
+
are independent random variable with respective distribution F
− GU E
and F
+ GU E
.
Asymptotic Independence in the Spectrum of the GUE 381
Proof. The proof is a mere application of Slutsky’s lemma see for instance [
18 , Lemma 2.8
]. Write
n
2 3
λ
max
λ
min
+ 1 =
1 λ
min
h n
2 3
λ
max
− 2 + n
2 3
λ
min
+ 2 i
. 7
Now, λ
min −1
goes almost surely to -2 as n → ∞, and n
2 3
λ
max
− 2 + n
2 3
λ
min
+ 2 converges in distribution to the convolution of F
− GU E
and F
+ GU E
by Corollary 1. Thus, Slutsky’s lemma yields the convergence in distribution of the right-hand side of 7 to
−
1 2
λ
−
+ λ
+
with λ
−
and λ
+
independent and distributed according to F
− GU E
and F
+ GU E
. Proof of Corollary 2 is completed.
3 Proof of Theorem 1
3.1 Useful results
3.1.1 Kernels
Let {H
k
x}
k ≥0
be the classical Hermite polynomials H
k
x := e
x
2
−
d d x
k
e
−x
2
and consider the function ψ
n k
x defined for 0 ≤ k ≤ n − 1 by: ψ
n k
x :=
n 2
1 4
e
−
nx2 4
2
k
k p
π
1 2
H
k
Ç n 2
x
. Denote by K
n
x, y the following kernel on R
2
K
n
x, y :=
n −1
X
k=0
ψ
n k
xψ
n k
y , 8
= ψ
n n
xψ
n n
−1
y − ψ
n n
yψ
n n
−1
x x
− y .
9 Equation 9 is obtained from 8 by the Christoffel-Darboux formula. We recall the two well-
known asymptotic results
Proposition 1. a Bulk of the spectrum. Let µ
∈ −2, 2. ∀x, y ∈ R
2
, lim
n →∞
1 n
K
n
µ +
x n
, µ + y
n
= sin πρµx
− y πx − y
, 10
where ρµ = p
4 −µ
2
2π
. Furthermore, the convergence 10 is uniform on every compact set of R
2
. b Edge of the spectrum.
∀x, y ∈ R
2
, lim
n →∞
1 n
23
K
n
2 +
x n
23
, 2 + y
n
23
=
AixAi
′
y − Ai yAi
′
x x
− y ,
11 where Aix is the Airy function. Furthermore, the convergence 11 is uniform on every
compact set of R
2
.
382 Electronic Communications in Probability
We will need as well the following result on the asymptotic behavior of functions ψ
n k
.
Proposition 2. Let µ ∈ −2, 2, k ∈ {0, 1} and denote by K a compact set of R.
a Bulk of the spectrum. There exists a constant C such that sup
x ∈K
ψ
n n
−k
µ +
x n
≤ C .
12 b Edge of the spectrum. There exists a constant C such that
sup
x ∈K
ψ
n n
−k
2
x n
23
≤ n
16
C . 13
The proof of these results can be found in [
11 , Chapter 7
], see also [ 1
, Chapter 3 ].
3.1.2 Determinantal representations, Fredholm determinants
There are determinantal representations using kernel K
n
x, y for the joint density p
n
of the eigenvalues λ
n i
; 1 ≤ i ≤ n, and for its marginals see for instance [
10 , Chapter 6
]: p
n
x
1
, · · · , x
n
= 1
n det
¦ K
n
x
i
, x
j
©
1 ≤i, j≤n
, 14
Z
R
n −m
p
n
x
1
, · · · , x
n
d x
m+1
· · · d x
n
= n − m
n det
¦ K
n
x
i
, x
j
©
1 ≤i, j≤m
m ≤ n . 15
Definition 1. Consider a linear operator S defined for any bounded integrable function f : R → R
by S f : x
7→ Z
R
Sx, y f yd y , where Sx, y is a bounded integrable Kernel on R
2
→ R with compact support. The Fredholm determinant Dz associated with operator S is defined as follows
∀z ∈ C, Dz := det1
− zS = 1 +
∞
X
k=1
−z
k
k Z
R
k
det ¦
Sx
i
, x
j
©
1 ≤i, j≤k
d x
1
· · · d x
k
. 16
It is in particular an entire function and its logarithmic derivative has a simple expression [ 17
, Section 2.5] given by
D
′
z Dz
= −
∞
X
k=0
T k + 1z
k
, 17
where T k =
Z
R
k
Sx
1
, x
2
Sx
2
, x
3
· · · Sx
k
, x
1
d x
1
· · · d x
k
. 18
For details related to Fredholm determinants, see for instance [
14 ,
17 ].
The following kernel will be of constant use in the sequel S
n
x, y; λ, ∆ :=
p
X
i=1
λ
i
1
∆
i
xK
n
x, y, 19
where λ = λ
1
, · · · , λ
p
∈ R
p
or λ ∈ C
p
, depending on the need, and ∆ = ∆
1
, · · · , ∆
p
is a collection of p bounded Borel sets in R.
Asymptotic Independence in the Spectrum of the GUE 383
Remark 3. The kernel K
n
x, y is unbounded and one cannot consider its Fredholm determinant without caution. The kernel S
n
x, y is bounded in x since the kernel is zero if x is outside the compact closure of the set
∪
p i=1
∆
i
, but a priori unbounded in y. In all the forthcoming compu- tations, one may replace S
n
with the bounded kernel ˜ S
n
x, y = P
p i,ℓ=1
λ
i
1
∆
i
x1
∆
ℓ
yK
n
x, y and get exactly the same results. For notational convenience, we keep on working with S
n
.
Proposition 3. Let p ≥ 1 be a fixed integer, ℓ = ℓ
1
, · · · , ℓ
p
∈ N
p
and denote ∆ = ∆
1
, · · · , ∆
p
, where every ∆
i
is a bounded Borel set. Assume that the ∆
i
’s are pairwise disjoint. Then the following identity holds true
P ¦
N ∆
1
= ℓ
1
, · · · , N ∆
p
= ℓ
p
© =
1 ℓ
1
· · · ℓ
p
− ∂
∂ λ
1 ℓ
1
· · ·
− ∂
∂ λ
p
ℓ
p
det 1 − S
n
λ, ∆
λ
1
=···=λ
p
=1
, 20
where S
n
λ, ∆ is the operator associated to the kernel defined in 19.
Proof of Proposition 3 is postponed to Section 4.1.
3.1.3 Useful estimates for kernel S
n
x, y; λ, ∆ and its iterations Consider µ, ∆ and ∆
n
as in Theorem 1. Assume moreover that n is large enough so that the Borel sets ∆
i,n
; 1 ≤ i ≤ p are pairwise disjoint. For i ∈ {1, · · · , p}, define κ
i
as κ
i
= ¨
1 if − 2 µ
i
2
2 3
if µ
i
= 2 .
21 Otherwise stated, κ
1
= κ
p
=
2 3
and κ
i
= 1 for 1 i p.
Let λ ∈ C
p
. With a slight abuse of notation, denote by S
n
x, y; λ the kernel
S
n
x, y; λ := S
n
x, y; λ, ∆
n
. 22
For 1
≤ m, ℓ ≤ p and Λ ⊂ C
p
, define M
mℓ,n
Λ := sup
λ ∈Λ
sup
x, y∈∆
m,n
∆
ℓ,n
S
n
x, y; λ ,
23 where S
n
x, y; λ is given by 22. Proposition 4. Let Λ