Some asymptotic properties of the spectrum of the Jacobi

ensemble
Holger Dette
Ruhr-Universitat Bochum
Fakultat f ur Mathematik
44780 Bochum, Germany
e-mail: [email protected]
FAX: +49 234 3214 559
Jan Nagel
Ruhr-Universitat Bochum
Fakultat f ur Mathematik
44780 Bochum, Germany
e-mail: [email protected]
June 20, 2008
Abstract
For the random eigenvalues with density corresponding to the Jacobi ensemble
c

i<j
|
i

j
|

i=1
(2
i
)
a
(2 +
i
)
b
I
(2,2)
(
i
)
(a, b > 1, > 0) a strong uniform approximation by the roots of the Jacobi polynomials
is derived if the parameters a, b, depend on n and n . Roughly speaking, the
eigenvalues can be uniformly approximated by roots of Jacobi polynomials with parameters
((2a+2)/ 1, (2b +2)/ 1), where the error is of order {log n/(a+b)}
1/4
. These results
are used to investigate the asymptotic properties of the corresponding spectral distribution
if n and the parameters a, b and vary with n. We also discuss further applications
in the context of multivariate random F-matrices.
AMS Subject Classication: Primary 60F15; Secondary 15A52
Keywords and Phrases: Jacobi ensemble, strong approximations, random F-matrix, orthogonal
polynomials, random matrices
1 Introduction
The three classical ensembles of random matrix theory are the Hermite, Laguerre and Jacobi en-
sembles. The Hermite or Gaussian ensembles arise in physics and are obtained as the eigenvalue
1
distribution of a symmetric matrix with Gaussian entries. The Laguerre or Wishart ensembles
appear in statistics and correspond to the distribution of the singular values of a Gaussian matrix.
Similarly, the Jacobi ensembles are objects of statistical interest and are motivated by multivari-
ate analysis of variance (MANOVA; see Muirhead (1982)). Associated with each ensemble there
is a real positive parameter which is usually considered for three values. The case = 1
corresponds to real matrices, while the ensembles for = 2 and = 4 arise from complex and
quaternion random matrices, respectively, according to Dysons (1962) threefold classication.
Dyson also observed that the eigenvalue distributions correspond to the Gibbs distribution for
the classical Coloumb gas at three dierent temperatures. In other words starting from the
physical interpretation for many decades it was only known that there exist random matrix
models for the Coloumb gas at three dierent temperatures. Recently Dumitriu and Edelman
(2002) provided tridiagonal random matrix models for the -Hermite and -Laguerre ensembles
for all > 0. Dette and Imhof (2007) derived strong uniform approximations of the random
eigenvalues of these ensembles by the (deterministic) roots of Hermite and Laguerre polynomials.
The development of a matrix model corresponding to the Jacobi ensemble was an open problem,
which was recently considered by Lippert (2003) and Killip and Nenciu (2004) and Edelman and
Sutton (2006). In particular, Killip and Nenciu (2004) found a tridiagonal matrix model for the
-Jacobi ensemble with density
f() = c

i<j
|
i

j
|

i=1
(2
i
)
a
(2 +
i
)
b
I
(2,2)
(
i
), (1.1)
(a, b > 1, > 0), where the entries are simple functions of independent random variables with
a beta distribution on the interval [1, 1] [see equations (2.2) and (2.3) in Section 2 for more
details]. It is the purpose of the present paper to obtain further insight in the stochastic properties
of the random eigenvalues with density given by (1.1). In Section 2 we introduce the triangular
matrix proposed by Killip and Nenciu (2004) and present a uniform approximation of the random
eigenvalues with density (1.1) by roots of Jacobi polynomials. Roughly speaking, if n ,
the random eigenvalues can be uniformly approximated by roots of the Jacobi polynomials
P
((2a+2)/1,(2b+2)/1)
n
(x/2), where the error of this approximation is of order {log n/(a +b)}
1/4
.
These results are used in Section 3 to investigate the asymptotic properties of the spectrum if
the parameters a, b and vary simultaneously with n and n . In Section 4 we study as an
application the eigenvalue distribution of a multivariate F-matrix which was also investigated
by Silverstein (1985b) and Collins (2005). We present several extensions of these results. In
particular, we consider almost sure convergence [Silverstein (1985b) discussed convergence in
probability while Collins (2005) considered the expectation of the empirical spectral distribution]
and investigate the case, where the parameters and the sample size converge to innity with
dierent order [Silverstein (1985b) and Collins (2005) discussed the case where a
1
n, b
2
n].
Finally, some technical results are given in an appendix.
2
2 Strong approximation of the Jacobi ensemble
Recall the denition of the Jacobi ensemble in (1.1) and let for p, q > 0 B(p, q) denote a
Beta-distributed random variable on the interval [1, 1] with density
2
1pq
(p + q)
(p)(q)
(1 x)
p1
(1 + x)
q1
I
(1,1)
(x). (2.1)
It was shown by Killip and Nenciu (2004) that if
0
,
1
, . . . ,
2n2
are independent random
variables with distribution

k

_
B(
2nk2
4
+ a + 1,
2nk2
4
+ b + 1) if k even,
B(
2nk3
4
+ a + b + 2,
2nk1
4
) if k odd,
(2.2)
then the joint density of the (real) eigenvalues of the tridiagonal matrix
J :=
_
_
_
_
_
_
b
1
a
1
a
1
b
2
.
.
.
.
.
.
.
.
.
a
n1
a
n1
b
n
_
_
_
_
_
_
(2.3)
with entries
b
k+1
= (1
2k1
)
2k
(1 +
2k1
)
2k2
,
a
k+1
=
_
(1
2k1
)(1
2
2k
)(1 +
2k+1
)
_
1/2
(
2n1
=
1
=
2
= 1) is given by the Jacobi ensemble (1.1). In the following discussion we
consider the asymptotic properties of the eigenvalues of the random matrix (2.3) [or equivalently
of the Jacobi ensemble (1.1)], where n and the parameters a, b and in (1.1) also vary
with n. An important tool of our analysis are the Jacobi polynomials P
(,)
n
(x), which are
dened as the polynomials of degree n with leading coecient (n + + + 1)
n
/2
n
n! satisfying
the orthogonality relation
_
1
1
P
(,)
n
(x)P
(,)
m
(x)(1 x)

(1 + x)

dx = 0 if m = n (2.4)
[see Szego (1975)]. Here and throughout this paper (a)
n
= (a +n)/(a) denotes the Pochham-
mer symbol. The following results provide an almost sure uniform approximation of the random
eigenvalues of the Jacobi ensemble by the roots of orthogonal polynomials, if n . We begin
with a statement of an exponential bound for the probability of a maximal deviation between
the roots of the Jacobi polynomials P
((2a+2)/1,(2b+2)/1)
n
(x/2) and random eigenvalues of the
Jacobi ensemble.
3
Theorem 2.1. Let
(n)
1

(n)
n
denote the ordered eigenvalues with density given by the
Jacobi ensemble (1.1) and x
(n)
1
< < x
(n)
n
denote the ordered roots of the Jacobi polynomial
P
(
2a+2

1,
2b+2

1)
n
(
1
2
x) ,
then the following inequality holds for any (0, 1]
P
_
max
1jn
|
(n)
j
x
(n)
j
| >
_
4(2n 1) exp
__
log(1 +

2
648 + 2
2
)

2
648 + 2
2
_
(a + b + 2)
_
.
(2.5)
Proof of Theorem 2.1. Interchanging the rows and columns of the matrix J dened in (2.3)
it follows that the matrix

J :=
_
_
_
_
_
_
b

1
a

1
a

1
b

2
.
.
.
.
.
.
.
.
.
a

n1
a

n1
b

n
_
_
_
_
_
_
(2.6)
with entries
b

k+1
:= b
nk
= (1
2n2k3
)
2n2k2
(1 +
2n2k3
)
2n2k4
, (2.7)
a

k+1
:= a
nk1
=
_
(1
2n2k5
)(1
2
2n2k4
)(1 +
2n2k3
)
_
1/2
(2.8)
has the same eigenvalues as the matrix J. This implies that the joint density of the eigenvalues
of the matrix

J is also given by the -Jacobi ensemble dened in (1.1). We now consider the
deterministic matrix
D :=
_
_
_
_
_
_
d
1
c
1
c
1
d
2
.
.
.
.
.
.
.
.
.
c
n1
c
n1
d
n
_
_
_
_
_
_
, (2.9)
where we essentially replace the random variables in (2.7) and (2.8) by their corresponding
expectations, that is
d
k+1
:= (1 E[
2n2k3
])E[
2n2k2
] (1 + E[
2n2k3
])E[
2n2k4
] ,
c
k+1
:=
_
(1 E[
2n2k5
])(1 E[
2n2k4
]
2
)(1 + E[
2n2k3
])
_
1/2
.
4
A straightforward calculation [observing that the expectation and variance of a random variable
with density (2.1) are given by (q p)/(p +q) and 4pq/((p +q +1)(p +q)
2
), respectively], yields
d
k+1
=
2(

b
2
a
2
)
(2k + a +

b)(2k + a +

b + 2)
, k = 0, . . . , n 2 ; (2.10)
c
k+1
=
4
2k + a +

b + 2
_
(k + a +

b + 1)(k + a + 1)(k +

b + 1)(k + 1)
(2k + a +

b + 3)(2k + a +

b + 1)
_
1/2
(2.11)
(k = 0, . . . , n 3) where a =
2a+2

,

b =
2b+2

, and (observing that

1
=
2
= 1)
d
n
=
2(

b a)
2n + a +

b 2
, (2.12)
c
n1
=
4
2n + a +

b 2
_
(n + a 1)(n +

b 1)(n 1)
(2n + a +

b 3)
_
1/2
. (2.13)
For the calculation of the eigenvalues of the matrix D we consider the determinant det(x I
n
D),
then it follows by an expansion with respect to the last row that
det(x I
n
D) = (x d
n
) G
n1
(x) c
2
n1
G
n2
(x) , (2.14)
where the polynomials G
0
(x), . . . , G
n1
(x) are dened recursively by the three term recurrence
relation
G
k
(x) = (x s
k
) G
k1
(x) r
2
k1
G
k2
(x) (2.15)
with coecients
s
k
=
2(

b
2
a
2
)
(2k + a +

b)(2k + a +

b 2)
, (2.16)
r
2
k1
=
16(k 1)( a +

b + k 1)( a + k 1)(

b + k 1)
(2k + a +

b 3)(2k + a +

b 2)
2
(2k + a +

b 1)
(2.17)
(k = 1, . . . , n 1; G
0
(x) := 1, G
1
(x) := 0). Now a straightforward calculation and a
comparison with the three term recurrence relation for the monic Jacobi polynomials [see e.g.
Chihara (1978), p. 220] yields
G
k
(x) = 2
k

P
( a,

b)
k
(
1
2
x); k = 0, . . . , n 1, (2.18)
where

P
( a,

b)
k
(x) denotes the kth monic Jacobi polynomial, i.e.

P
( a,

b)
k
(x) =
2
k
k!
(k + a +

b + 1)
k
P
( a,

b)
k
(x) . (2.19)
5
Combining equations (2.15), (2.18) and (2.19) we obtain by a tedious but straightforward calcu-
lation
det(I
n
x D) = G
n
(x) + (s
n
d
n
) G
n1
(x) + (r
2
n1
c
2
n1
) G
n2
(x)
=
4
n
n!
(n + a +

b + 1)
n
_
P
( a,

b)
n
(
1
2
x)
(

b a)(2n + a +

b 1)
(2n + a +

b 2)(n + a +

b)
P
( a,

b)
n1
(
1
2
x)

(2n + a +

b)(n + a 1)(n +

b 1)
(2n + a +

b 2)(n + a +

b)(n + a +

b 1)
P
( a,

b)
n2
(
1
2
x)
_
=
4
n
n!
(n + a +

b + 1)
n
_
P
( a,

b)
n
(
1
2
x)
n +

b
n + a +

b
P
( a,

b)
n1
(
1
2
x)
+
(2n + a +

b)(n + a 1)
(n + a +

b)(n + a +

b 1)
P
( a1,

b)
n1
(
1
2
x)
_
,
where we have used the identity
(n +

b 1)P
( a,

b)
n2
(x) = (n + a +

b 1)P
( a,

b)
n1
(x) (2n + a +

b 2)P
( a1,

b)
n1
(x) (2.20)
in the last step [see Abramovich and Stegun (1965), equation (22.7.18)]. A further application
of this identity to the second polynomial yields
det(I
n
x D) =
4
n
n!
(n + a +

b + 1)
n
_
2n + a +

b
n + a +

b
P
( a1,

b)
n
(
1
2
x)
+
(2n + a +

b)(n + a 1)
(n + a +

b)(n + a +

b 1)
P
( a1,

b)
n1
(
1
2
x)
_
=
4
n
n!
(n + a +

b + 1)
n
_
2n + a +

b
n + a +

b
P
( a1,

b)
n
(
1
2
x)
2n + a +

b
n + a +

b
P
( a1,

b)
n
(
1
2
x)
+
(2n + a +

b)(2n + a +

b 1)
(n + a +

b)(n + a +

b 1)
P
( a1,

b1)
n
(
1
2
x)
_
=
4
n
n!
(n + ( a 1) + (

b 1) + 1)
n
P
( a1,

b1)
n
(
1
2
x) ,
where we have used the identity
(n + a 1)P
( a1,

b)
n1
(x) = (2n + a +

b 1)P
( a1,

b1)
n
(x) (n + a +

b 1)P
( a1,

b)
n
(x) (2.21)
for the second equality [see Abramovich and Stegun (1965), equation (22.7.19)]. Consequently,
the eigenvalues of the matrix D are given by the roots x
(n)
1
< < x
(n)
n
of the Jacobi polynomial
P
( a1,

b1)
n
(
1
2
x). A similar argument as in Silverstein (1985a) now shows that
max
1jn
|
(n)
j
x
(n)
j
| max
1kn
n

j=1
|

J
kj
D
kj
| 4 {3X
n
}
1/2
+ 6X
n
, (2.22)
6
where the elements of the matrices

J and D are denoted by

J
ij
and D
ij
, respectively, and the
random variable X
n
is dened by
X
n
:= max
0k2n2
|
k
E[
k
]| . (2.23)
This implies for the probability in Theorem 2.1
P
_
max
1jn
|
(n)
j
x
(n)
j
| >
_
P
_
2{3X
n
}
1/2
+ 3X
n
>

2
_
P
_
3{3X
n
}
1/2
>

2
_
(2.24)
= P
_
X
n
>

2
108
_

2n2

k=0
P
_
|
k
E[
k
]| >

2
108
_
whenever (0, 1]. Observing Lemma A.1 in the Appendix and that
k
= 1 2
k
, where
k
is
the corresponding Beta distribution on the interval [0, 1], it follows that
P
_
max
1jn
|
(n)
j
x
(n)
j
| >
_
4
2n2

k=0
exp
_
c (
2n 2 k
2
+ a + b + 2)
_
4(2n 1) exp (c (a + b + 2))
where the constant c = c() is given by
c = log
_
1 +

2
648 + 2
2
_

2
648 + 2
2
. (2.25)
This proves the assertion of the theorem. 2
Note that the constant c in (2.25) is negative, and Theorem 2.1 therefore indicates that the
random eigenvalues of the Jacobi ensemble can be approximated by the deterministic roots of
the Jacobi polynomial P
( a1,

b1)
n
(
1
2
x) with a high probability if a + b is large. The following
result makes this statement more precise and provides a strong uniform approximation of the
eigenvalues of the Jacobi ensemble by the roots of the Jacobi polynomial P
( a1,

b1)
n
(
1
2
x), if the
parameters in (1.1) converge suciently fast to innity. The proof follows by similar arguments
as the proof of Theorem 2.2 in Dette and Imhof (2007) and is therefore omitted.
Theorem 2.2. Let
(n)
1

(n)
n
denote the ordered random eigenvalues with density given
by the Jacobi ensemble (1.1) with parameters a = a
n
, b = b
n
, =
n
and x
(n)
1
< < x
(n)
n
denote
the ordered roots of the Jacobi polynomial P
( a
n
1,

b
n
1)
n
(
1
2
x) where
a
n
=
2a
n
+2

n
,

b
n
=
2b
n
+2

n
. (2.26)
If
lim
n
a
n
+ b
n
log n
= , (2.27)
7
then the inequality
max
1jn

(n)
j
x
(n)
j

_
log n
a
n
+ b
n
_
1/4
S
holds for all n 2, where S denotes an a.s. nite random variable. In particular, if
(2.28) liminf
n
a
n
+ b
n
n
> 0,
then there exists an a.s. nite random variable S

such that the inequality

max
1jn

(n)
j
x
(n)
j

_
log n
n
_
1/4
S

holds a.s. for all n 2.

3 Asymptotic spectral properties
In this section we will apply Theorem 2.2 to derive the asymptotic properties of the empirical
spectral distribution
F
J
n
() :=
1
n
n

i=1
I{
(n)
i
} , (3.1)
where
(n)
1

(n)
n
denote the ordered eigenvalues of the Jacobi ensemble dened by (1.1)
with parameters a = a
n
, b = b
n
and =
n
. The results of Section 2 indicate that the empirical
distribution function in (3.1) should exhibit a similar asymptotic behaviour as the empirical
distribution function
F
P
n
() :=
1
n
n

i=1
I{x
(n)
i
} (3.2)
of the ordered roots x
(n)
1
< < x
(n)
n
of the Jacobi polynomial P
( a
n
1,

b
n
1)
n
(
1
2
x), where a
n
=
(2a
n
+ 2)/
n
,

b
n
= (2b
n
+ 2)/
n
. The asymptotic zero distribution of Jacobi polynomials has
been studied by several authors [see e.g. Gawronski and Shawyer (1991), Elbert, Laforgia and
Rodono (1994), Dette and Studden (1995) or Kuijlaars and Van Assche (1999) among many
others], and we can use these results and Theorem 2.2 to derive the asymptotic properties of the
spectrum of the Jacobi ensemble. The following result makes this statement more precise.
Theorem 3.1. Consider the empirical distribution functions of the eigenvalues of the Jacobi
ensemble (1.1) and the roots of the Jacobi polynomial P
( a
n
1,

b
n
1)
n
(x/2) dened by (3.1) and
(3.2), respectively, and let (
n
)
nN
, (
n
)
nN
denote real sequences with
n
> 0 such that the limit
lim
n
F
P
n
(
n
+
n
) = F()
8
exists at every coninuity point of F. If the conditions
a
n
+ b
n
log n

n
,

4
n
(a
n
+ b
n
)
log n

n
(3.3)
are satised, then
lim
n
F
J
n
(
n
+
n
) = F()
almost surely at every continuity point of F.
Proof. Let G
P
n
and G
J
n
denote the empirical distribution functions of the rescaled roots
x
(n)
1

n

n
, . . . ,
x
(n)
n

n

n
corresponding to the Jacobi polynomial P
( a
n
1,

b
n
1)
n
(
1
2
x) and of the eigenvalues

(n)
1

n

n
, . . . ,

(n)
n

n

n
corresponding to the Jacobi ensemble dened by (1.1), respectively. The arguments presented in
the proof of Theorem 2.1 show that G
P
n
and G
J
n
are the empirical distribution functions of the
eigenvalues of the matrices
A
n
:=
1

n
(D
n
I
n
),
B
n
:=
1

n
(J
n
I
n
),
respectively. Observing Lemma 2.3 in Bai (1999) we obtain for the Levy-distance between the
distribution functions G
P
n
and G
J
n
L
3
(G
P
n
, G
J
n
)
2
n
1
n
n

i=1
|
(n)
i
x
(n)
i
|
2

2
n
( max
1in
|
(n)
i
x
(n)
i
|)
2
S
__
log n
(a
n
+ b
n
)
4
n
_
1/2
_
,
where S denotes an a.s. nite random variable. Consequently, we obtain from the assumptions
L(G
P
n
, G
J
n
)
a.s.

n
0 ,
and the assertion of Theorem 3.1 follows observing the identities
G
P
n
() = F
P
n
(
n
+
n
) ,
G
J
n
() = F
J
n
(
n
+
n
) .
2
9
Theorem 3.2. Let
(n)
1

(n)
n
denote the ordered random eigenvalues with density given
by the Jacobi ensemble (1.1). If the assumptions of Theorem 3.1 are satised and that there exist
constants a
1
, a
2
R b
1
, b
2
R
+
such that
(i) lim
n
1

n
_
n +

b
n
1
2n + a
n
+

b
n
2

n
_
=
a
1
2
;
(ii) lim
n
1

n
_
n(n + a
n
1) + (n +

b
n
1)(n + a
n
+

b
n
2)
(2n + a
n
+

b
n
2)
2

n
_
=
a
2
2
;
(iii) lim
n
1

2
n
(n +

b
n
1)(n + a
n
1)n
(2n + a
n
+

b
n
2)
3
=
b
1
4
;
(iv) lim
n
1

2
n
(n +

b
n
1)(n + a
n
1)(n + a
n
+

b
n
2)n
(2n + a
n
+

b
n
2)
4
=
b
2
4
;
then the empirical distribution of the scaled eigenvalues

(n)
1
2(2
n
1)
2
n
, . . . ,

(n)
n
2(2
n
1)
2
n
converges almost surely to a non-degenerate distribution function, i.e.
lim
n
F
J
n
(2
n
+ 2(2
n
1))
a.s.
=
_

a
2
2

b
2
f
(a
1
,a
2
,b
1
,b
2
)
(x)dx ,
where
f
(a
1
,a
2
,b
1
,b
2
)
(x) =
_
_
_
b
1
2

4b
2
(xa
2
)
2
(b
2
b
1
)x
2
+(b
1
a
2
+b
1
a
1
2b
2
a
1
)x+b
2
a
2
1
a
1
a
2
b
1
+b
2
1
if | x a
2
| 2

b
2
,
0 else.
Proof. By Theorem 2.2 in Dette and Studden (1995) it follows that the empirical distribution
of the roots of the Jacobi polynomial P
( a
n
1,

b
n
1)
n
(
1
2
x) has a non-degenerate limit, that is
F
P
n
(2
n
+ 2(2
n
1))
n
_

a
2
2

b
2
f
(a
1
,a
2
,b
1
,b
2
)
(x)dx .
The assertion is now an immediate consequence of Theorem 3.1 2
Example 3.3. Assume that
lim
n
a
n
n
=
0
and lim
n

b
n
n
=
0
for some constants
0
,
0
0. If additionally (a
n
+ b
n
)/ log n , it is easy to see that the
assumptions of Theorem 3.2 are satised with
n
=
1
2
,
n
=
1
2
(n N). Consequently, it follows
10
that the empirical spectral distribution of the Jacobi ensemble (1.1) converges almost surely to
a distribution function with density
f

0
,
0
(x) =
2 +
0
+
0
2
_
(2r
2
x)(x 2r
1
)
4 x
2
I
(2r
1
,2r
2
)
(x) ,
where
r
1
:=

2
0

2
0
4
_
(
0
+ 1)(
0
+ 1)(
0
+
0
+ 1)
(2 +
0
+
0
)
2
,
r
2
:=

2
0

2
0
+ 4
_
(
0
+ 1)(
0
+ 1)(
0
+
0
+ 1)
(2 +
0
+
0
)
2
.
For example, if a
n
= b
n
= 3n and
n
= 2 we have
lim
n
a
n
n
= lim
n

b
n
n
= 3 , (3.4)
and the limiting spectral distribution has the density
f
3,3
(x) =
4

_
7/4 x
2
4 x
2
I
[

7
2
,

7
2
]
(x) ,
which is depicted in the left part of Figure 1. Similarly, if a
n
= b
n
=

n and
n
= 2n we have
lim
n
a
n
n
= lim
n

b
n
n
= 0 ,
and the limiting distribution is given by the arc-sine law on the interval [2, 2] with density
f
0,0
(x) =
1

4 x
2
I
(2,2)
(x) ,
displayed in the right part of Figure 1.
Example 3.4. If the parameters a
n
and

b
n
converge to innity such that
lim
n
a
n
n
= , lim
n

b
n
n
= , lim
n
a
n

b
n
= > 0 ,
and additionally

n
C a
n
= O (n
1+
) , (3.5)
for some C R
+
, (0, 1) it follows that the scaled empirical distribution of the eigenvalues
of the Jacobi ensemble (1.1) with parameters a
n
, b
n
and
n
converges almost surely to Wigners
semi circle law, that is
lim
n
F
J
n
_
2
_
n
a
n
1
2
a
n

b
n
a
n
+

b
n
2
_
a.s.
=
2

2
_

2
x
2
dx | | ,
11
2.5 2 1.5 1 0.5 0 0.5 1 1.5 2 2.5
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45


2.5 2 1.5 1 0.5 0 0.5 1 1.5 2 2.5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure 1: Density of the limiting distribution of the random eigenvalues corresponding to the
Jacobi ensemble and a histogram based on n = 5000 eigenvalues from the Jacobi ensemble (1.1).
Left panel: a
n
= b
n
= 3n,
n
= 2; right panel: a
n
= b
n
=

n,
n
= 2n.
where = 4/(1 +)
3/2
. Note that assumption (3.5) guarantees that the second condition (3.3)
in Theorem 3.1 is satised, i.e.
lim
n
_
n
a
n
1
4
a
n
+ b
n
log n
= .
This situation is also of particular interest in the case, where the inverse temperature
n
converges
to 0. For example, if a
n
= b
n
= n1 and
n
= 2n
1/4
we have a
n
=

b
n
= n
5/4
and the condition
(3.3) is satised. The limiting distribution of the scaled eigenvalues
1
2
_
n
5/4
1
n

(n)
1
< <
1
2
_
n
5/4
1
n

(n)
n
is then given by
f(x) =
1

2 x
2
I
[

2,

2]
(x).
Example 3.5. We now consider the case, where the sequences a
n
and

b
n
converge to innity
with dierent rates. If
lim
n
a
n
n
= , lim
n

b
n
n
=
0
0 ,
and the sequences (
n
)
nN
, (a
n
)
nN
satisfy

n
C, a
n
= O (n
1+
)
12
for constants C R
+
and (0, 1/3), then it is easy to see that the conditions of Theorem 3.2
are satised with

n
= 0 ;
n
=
n
a
n
1
,
and it follows by a straightforward calculation that
lim
n
F
J
n
_
2n
a
n
1
2
_
a.s.
=
1
4
_

s
1
_
(s
2
x)(x s
1
)
x
dx s
1
s
2
,
where
s
1
:= 2(2 +
0
) 4
_
1 +
0
,
s
2
:= 2(2 +
0
) + 4
_
1 +
0
.
Similarly, if
lim
n
a
n
n
= , lim
n

b
n
n
= , lim
n
a
n

b
n
= ,
and for some (0, 1) there exist constants C, C
1
, C
2
> 0 and a > 3/2 1/2 such that

n
C, a
n
C
1
n
1+
,

b
n
C
2
n
1+
, (3.6)
then we obtain for any 2 6
lim
n
F
J
n
_
_
2
_
n(

b
n
1)
a
n
1
2
a
n
+ 2
_
n(

b
n
1)

b
n
2n + a
n
+

b
n
2
_
_
a.s.
=
1
8
_

2
_
(6 x)(x + 2)dx.
4 An application to the multivariate F distribution
Let X
ij
(i = 1, . . . , n; j = 1, . . . , n
1
; n
1
n) and Y
ij
(i = 1, . . . , n; j = 1, . . . , n
2
; n
2
n)
denote independent standard normal distributed random variables and consider the random
matrices
X
n
= (X
ij
)
j=1,...,n
1
i=1,...,n
R
nn
1
; Y
n
= (Y
ij
)
j=1,...,n
2
i=1,...,n
R
nn
2
. (4.1)
The matrix
F
n
:=
_
1
n
1
X
n
X
T
n
__
1
n
2
Y
n
Y
T
n
_
1
R
nn
(4.2)
is called multivariate F-matrix and plays a prominent role in the multivariate analysis of variance
[see e.g. Muirhead (1982)]. Silverstein (1985b) showed that under the conditions
lim
n
n
n
1
= y > 0 lim
n
n
n
2
= y

(0, 1)
13
the empirical distribution of the eigenvalues of a multivariate F-matrix converges in probability
to a non-random distribution function, say F
y,y
, with density
f
y,y
(x) =
1 y

2x(xy

+ y)
_
(x s
1
)(s
2
x)I
(s
1
,s
2
)
(x) , (4.3)
where the bounds of the support are given by
s
1
=
_
1
_
1 (1 y)(1 y

)
1 y

_
2
,
s
2
=
_
1 +
_
1 (1 y)(1 y

)
1 y

_
2
.
Moreover, if y > 1 the limiting distribution has mass 1 1/y at the point 0. A corresponding
result for the expectation of spectral distribution of the matrix F
n
can be found in Collins
(2005). In the following discussion we will extend these results in two dierent directions using
the methodology developed in Section 2. On the one hand, we prove that in the case y (0, 1]
these results are also correct, if almost sure convergence is considered. On the other hand, we
extend these results to the case where n, n
1
, n
2
are not necessarily of the same order.
Theorem 4.1. Consider the multivariate F-matrix dened by (4.2). If
lim
n
n
n
1
= y (0, 1] , lim
n
n
n
2
= y

(0, 1), (4.4)

then the empirical distribution function of the eigenvalues of the matrix F
n
converges a.s. to to
a distribution function with density (4.3).
Proof. Consider the matrix
A
n
:= 2
_
Y
n
Y
T
n
X
n
X
T
n
_ _
Y
n
Y
T
n
+ X
n
X
T
n
_
1
, (4.5)
then it follows from Muirhead (1982) that the joint density of the eigenvalues of A
n
is given by
the Jacobi ensemble (1.1) with = 1 and
a
n
=
1
2
(n
1
n 1) ,
b
n
=
1
2
(n
2
n 1) .
If
F
denotes an eigenvalue of the matrix F
n
we obtain with some appropriate constant C R
the identity
det
_
2
n
2
n
1

F
n
2
n
1
+
F
I
n
2
_
Y
n
Y
T
n
X
n
X
T
n
_ _
Y
n
Y
T
n
+ X
n
X
T
n
_
1
_
=det
_

F
I
n

n
2
n
1
_
X
n
X
T
n
_ _
Y
n
Y
T
n
_
1
_
det
_
Y
n
Y
T
n
_
C = 0 ,
14
which shows that

J
= 2
n
2
n
1

F
n
2
n
1
+
F
(4.6)
is an eigenvalue of the matrix A
n
. Consequently, the empirical distribution function F
F
n
of the
eigenvalues of the matrix F
n
satises the relation
F
F
n
()
a.s.
= 1 F
J
n
_
2
n
2
n
1

n
2
n
1
+
_
, 0 , (4.7)
where F
J
n
denotes the empirical distribution function corresponding to the Jacobi ensemble (1.1)
with parameter = 1 (note that only the case 0 is of interest here). We now use Theorem
3.2 to derive the limiting spectral distribution. For this purpose we identify
a
n
1 =
2a
n
+ 2

1 = n
1
n ,

b
n
1 =
2a
n
+ 2

1 = n
2
n ,
and obtain the limits
lim
n
a
n
n
= lim
n
n
1
n + 1
n
= y
1
1 0 ,
lim
n

b
n
n
= lim
n
n
2
n + 1
n
= y
1
1 > 0 .
Moreover, the assumption
lim
n
a
n
+ b
n
log n
=
is obviously satised. From Example 3.3 it therefore follows that the empirical distribution
function F
J
n
of the eigenvalues of the matrix A
n
converges a.s. to a distribution function F with
density
f(x) =
y
1
+ y
1
2
_
(2r
2
x)(x 2r
1
)
4 x
2
I
(2r
1
,2r
2
)
(x) ,
where r
1
and r
2
are given by
r
1
:=
(y yy

)
2
(y

yy

)
2
4
_
y
2
y
2
(y + y

yy

)
(y + y

)
2
,
r
2
:=
(y yy

)
2
(y

yy

)
2
+ 4
_
y
2
y
2
(y + y

yy

)
(y + y

)
2
Observing the relation (4.7) we obtain
F
F
n
()
a.s.
1 F
_
2
y y

y + y

_
,
15
and the assertion of the theorem now follows by a straightforward but tedious calculation of the
density of the limiting distribution. 2
While the preceding theorem essentially provides an alternative proof of the results of Silverstein
(1985b), the following three theorems extend Silversteins ndings to the case where y, y

= 0.
Theorem 4.2. Consider the multivariate F-matrix dened in (4.2) and denote by
F
1

F
n
the corresponding eigenvalues. If
lim
n
n
n
1
= 0 , lim
n
n
n
2
= 0 , lim
n
n
1
n
2
= > 0
and
n
1
= O(n
1+
)
with (0, 1), then the empirical distribution function of the transformed eigenvalues

i
= 2
_
n
1
n
1
_
n
2
n
n
1
+ n
2
2n

n
2
n
1

F
i
+ n
2
_
i = 1, . . . , n
converges a.s. to a distribution function with density
f

(x) =
2

2
x
2
I { < x < } ,
where = 4/(1 + )
3/2
.
Proof. Recall the denition of the matrix A
n
in (4.5), which corresponds to the Jacobi ensemble
(1.1) with = 1, a
n
=
1
2
(n
1
n 1), b
n
=
1
2
(n
2
n 1). Using the notation (2.26) we obtain
a
n
1 = n
1
n,

b
n
1 = n
2
n. By the assumption of the theorem we have
lim
n
a
n
n
= , lim
n

b
n
n
= , lim
n
a
n

b
n
=
and a
n
= O(n
1+
). Therefore it follows from Example 3.4 that
lim
n
F
J
n
_
2
_
n
n
1
n
2
n
1
n
2
n
1
+ n
2
2n
_
=
2

2
_

2
x
2
I { < x < } dx
a.s., where F
J
n
denotes the empirical distribution of the eigenvalues of the matrix A
n
. The
identity (4.7) implies for > 2
F
J
n
()
a.s.
= 1 F
F
n
_
n
2
n
1
2
2 +
_
16
and therefore it follows
lim
n
F
F
n
_
_
n
2
n
1
_
_
2
2
n
2
n
n
1
+n
2
2n

_
n
n
1
n

1
_
_
_
_
=
2

2
_

2
x
2
I { < x < } dx
a.s., which proves the assertion of the theorem. 2
Theorem 4.3. Consider the multivariate F-matrix and denote by
F
1

F
n
the corre-
sponding eigenvalues. If
lim
n
n
n
1
= 0 , lim
n
n
n
2
= y

(0, 1]
and
n
1
= O(n
1+
) ,
with (0, 1/3), then the empirical distribution function of the scaled eigenvalues

i
=
n
2(n
1
n)
_

F
i
n
1
n
2
+ 1
_
i = 1, . . . , n
converges a.s. to a distribution function F with density
f
y
(x) =
1
4
_
(xs
2
1)(1 xs
1
)
x
2
I
(s
1
2
,s
1
1
)
(x) ,
where the bounds of the support of the density are given by
s
1
:= 2(y
1
+ 1) 4
_
y
1
,
s
2
:= 2(y
1
+ 1) + 4
_
y
1
.
The proof is analogous to the proof of Theorem 4.2, using the rst result of Example 3.5.
Similarly, the following theorem can be proven using the last statement in Example 3.5.
Theorem 4.4. Consider the multivariate F-matrix and denote by
F
1

F
n
the corre-
sponding eigenvalues. If
lim
n
n
n
1
= 0 , lim
n
n
n
2
= 0, lim
n
n
1
n
2
=
and for some (0, 1) there exist constants C
1
, C
2
> 0 and a > 3/2 1/2 such that
n
1
C
1
n
1+
, n
2
C
2
n
1+
,
17
then the empirical distribution function of the scaled eigenvalues

i
= 2
n
1
n
n
1
+ n
2
n
1
n
2
(n
2

_
n(n
2
n))
F
i
(n
1
+
_
n(n
2
n))
_
n(n
2
n)(1 +
n
1
n
2

F
i
)
i = 1, . . . , n
converges a.s. to a distribution function F with density
f(x) =
1
8
_
(6 + x)(2 x)I
[6,2]
(x) .
5 Appendix: auxiliary results
Lemma A.1. Let Z denote a Beta-distributed random variable on the interval [0, 1] with density
(p + q)
(p)(q)
x
p1
(1 x)
q1
I
(0,1)
(x) (p, q > 0) , (5.1)
then for any > 0
P( | Z E[Z] | > ) 4e
c(p+q)
,
where the constant c is dened by
c = log
_
1 +

3 + 2
_

3 + 2
.
Proof. If X (p, p + q), Y (q, p + q) denote independent Gamma-distributed random
variables, it is well known that the ratio Z = X/(X + Y ) has a Beta-distribution with density
(5.1). Because E[Z] = E[X] = p/(p + q) it follows that
P(|Z E[Z]| > ) = P
_

X
X + Y
E[X]

>
_
(5.2)
Dene

= /(3 + 2) and assume that

| X E[X] |

, | Y E[Y ] |

,
then it is easy to see that | X + Y 1 | = | X + Y E[X + Y ] | 2

; and

X
X + Y
E[X]

1
1 2

(|X E[X]| + E[X]|X + Y E[X + Y ]|)

1
1 2

= .
This implies for the probability in (5.2)
P
_

X
X + Y
E[X]

>
_
P(X > E[X] +

) + P(X < E[X]

) (5.3)
+ P(Y > E[Y ] +

) + P(Y < E[Y ] +

) .
18
Using similar arguments as in Dette and Imhof (2007) we obtain the estimates
P(U > E[U] +

) exp{(p + q)(log(1 +

)} ,
P(U < E[U]

) exp{(p + q)(log(1

) +

)} ,
where U is either X or Y . The assertion of Lemma A.1 now follows from (5.3) and the denition
of

observing that log(1 +

> log(1

) +

. 2
Acknowledgements. The authors are grateful to Martina Stein who typed parts of this paper
with considerable technical expertise. The work of the authors was supported by the Sonder-
forschungsbereich Tr/12 (project C2, Fluctuations and universality of invariant random matrix
ensembles) and in part by a NIH grant award IR01GM072876:01A1.
References
M. Abramovich, I. Stegun (1965). Handbook of Mathematical Functions. Dover Publications
Inc., N.Y.
Z. D. Bai (1999). Methodologies in spectral analysis of large dimensional random matrices, a
review. Statist. Sinica 9, 611-677.
T. S. Chihara (1978). An Introduction to Orthogonal Polynomials. Gordon and Breach, New
York.
B. Collins (2005). Product of random projections, Jacobi ensembles and universality problems
arising from free probability. Probability Theory and Related Fields 133, 315-344
H. Dette and L. A. Imhof (2007). Uniform approximation of eigenvalues in Laguerre and Hermite
-ensembles by roots of orthogonal polynomials. Transactions of the American Mathematical
Society 359, 4999-5018.
H. Dette and W.J. Studden (1995). Some new asymptotic properties for the zeros of the Jacobi,
Laguerre and Hermite polynomials. Constructive Approximation, Vol. 11, 227-238.
I. Dumitriu and A. Edelman (2002). Matrix models for beta-ensembles, J. Math. Phys. 43,
5830-5847.
F. J. Dyson (1962). The threefold way. Algebraic structure of symmetry groups and ensembles
in quantum mechanics. J. Math. Phys. 3, 1199-1215.
Edelman and Sutton (2006). The beta-Jacobi matrix model, the CS decomposition, and general-
ized singular value problems. Found. Comput. Math. In press. http://dx.doi.org/10.1007/s10208-
006-0215-9.
A. Elbert, A. Laforgia, L. G. Rodono (1994). On the zeros of Jacobi polynomials. Acta Math.
Hungar. 64, 351-359.
19
W. Gawronski, B. Shawyer (1991). Strong asymptotics and the limit distribution of the zeros of
Jacobi polynomials P
(a
n
+,b
n
+)
n
(x). In: Progress in Approximation Theory (P. Nevai, A. Pinkus
eds.). New York: Academic Press 379-404.
R. Killip and I. Nenciu I (2004). Matrix models for circular ensembles. Int. Math. Res. Not.
50, 2665-2701.
A. B. J. Kuijlaars and W. van Assche (1999). The asymptotic zero distribution of orthogonal
polynomials with varying recurrence coecients, J. Approx. Theory 99 (1999), 167-197.
R. A. Lippert (2003). A matrix model for the beta-Jacobi ensemble. J. Math. Phys. 44
4807-4816.
R. J. Muirhead (1982). Aspects of Multivariate Statistical Theory. Wiley, New York.
J. W. Silverstein (1985a) The smallest eigenvalue of a large dimensional Wishart matrix, Ann.
Probab. 13, 1364-1368.
J. Silverstein (1985b). The limiting eigenvalue distribution of a multivariate F-matrix. SIAM J.
Math. Anal. 16, 641-646.
G. Szego (1975). Orthogonal Polynomials. 4th ed. Amer. Math. Soc., Providence, RI.
20