Volumes of Orthogonal Groups and Unitary Groups

Lin Zhang∗
Institute of Mathematics, Hangzhou Dianzi University, Hangzhou 310018, PR China
arXiv:1509.00537v5 [math-ph] 3 Nov 2017

Abstract
The matrix integral has many applications in diverse fields. This review article begins
by presenting detailed key background knowledge about matrix integral. Then the volumes
of orthogonal groups and unitary groups are computed, respectively. As a unification, we
present Mcdonald’s volume formula for a compact Lie group. With this volume formula, one
can easily derives the volumes of orthogonal groups and unitary groups. Applications are also
presented as well. Specifically, The volume of the set of mixed quantum states is computed
by using the volume of unitary group. The volume of a metric ball in unitary group is also
computed as well.
There are no new results in this article, but only detailed and elementary proofs of existing
results. The purpose of the article is pedagogical, and to collect in one place many, if not all,
of the quantum information applications of the volumes of orthogonal and unitary groups.

∗ E-mail: [email protected]; [email protected]

1
Contents

1 Introduction 3

2 Volumes of orthogonal groups 5

2.1 Preliminary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 The computation of volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 Volumes of unitary groups 28

3.1 Preliminary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2 The computation of volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4 The volume of a compact Lie group 55

5 Applications 73
5.1 Hilbert-Schmidt volume of the set of mixed quantum states . . . . . . . . . . . . . . 73
5.2 Area of the boundary of the set of mixed states . . . . . . . . . . . . . . . . . . . . . 77
5.3 Volume of a metric ball in unitary group . . . . . . . . . . . . . . . . . . . . . . . . . 79

6 Appendix I: Volumes of a sphere and a ball 85

7 Appendix II: Some useful facts 89

7.1 Matrices with simple eigenvalues form open dense sets of full measure . . . . . . . 89
7.2 Results related to orthogonal groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.3 Results related to unitary groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

8 Appendix III: Some matrix factorizations 107

9 Appendix IV: Selberg’s integral 112

2
1 Introduction

Volumes of orthogonal groups and unitary groups are very useful in physics and mathematics
[3, 4]. In 1949, Ponting and Potter had already calculated the volume of orthogonal and unitary
group [27]. A complete treatment for group manifolds is presented by Marinov [17], who ex-
tracted the volumes of groups by studying curved path integrals [18]. There is a general closed
formula for any compact Lie group in terms of the root lattice [?]. Clearly the methods used
previously are not followed easily. Życzkowski in his paper [38] gives a re-derivation on the
volume of unitary group, but it is still not accessible. The main goal of this paper is to compute
the volumes of orthogonal groups and unitary groups in a systematic and elementary way. The
obtained formulae is more applicable.
As an application, by re-deriving the volumes of orthogonal groups and unitary groups, the
authors in [38] computes the volume of the convex (n2 − 1)-dimensional set D (C n ) of the density
matrices of size n with respect to the Hilbert-Schmidt measure. Recently, the authors in [30] give
the integral representation of the exact volume of a metric ball in unitary group; and present
diverse applications of volume estimates of metric balls in manifolds in information and coding
theory.
Before proceeding, we recall some notions in group theory. Assume that a group G acts on
the underlying vector space X via g| xi for all g ∈ G and x ∈ X . Let | xi be any nonzero vector in
X . The subset G| xi := { g| xi : g ∈ G} is called the G -orbit of | xi ∈ X . Denote

xG := { g ∈ G : g| xi = | xi}.

We have the following fact:

G| xi ∼ G /xG .
If xG = {e}, where e is a unit element of G , then the action of G on | xi is called free. In this case,

G| xi ∼ G /{e} ⇐⇒ G ∼ G| xi.

Now we review a fast track of the volume of a unitary group. Let us consider the unitary groups
G = U (n + 1), n = 1, 2, . . . To establish their structure, we look at spaces in which the group acts
transitively, and identify the isotropy subgroup. The unitary group U (n + 1) acts naturally in
the complex vector space X = C n+1 through the vector or "defining" representation; the image
of any nonzero vector | xi = |ψi ∈ C n+1 is contained in the maximal sphere S2n+1 of radius
k ψ k since k ψ k = k Uψ k for U ∈ U (n + 1), and in fact it is easy to see that it sweeps the whole
sphere when U runs though U (n + 1), i.e. the group U (n + 1) acts transitively in this sphere.
The isotropy group of the vector |n + 1i is easily seen to be the unitary group with an entry less,
that is xG = U (n) ⊕ 1. Indeed, the following map is onto:

ϕ : U (n + 1) −→ U (n + 1)|n + 1i = S2n+1 .

3
If we identity U (n) with U (n) ⊕ 1, as a subgroup of U (n + 1), then

U (n)|n + 1i = |n + 1i,

implying that ker ϕ = U (n), and thus

U (n + 1)/ ker ϕ ∼ U (n + 1)|n + 1i = S2n+1 = (U (n + 1)/ ker ϕ) |n + 1i.

Therefore we have the equivalence relation

U (n + 1)/U (n) = S2n+1 . (1.1)

This indicates that

 
vol (U (n + 1)) = vol S2n+1 · vol (U (n)) , n = 1, 2, . . . (1.2)

That is,
 
 
vol (U (n)) = vol S1 × vol S3 × · · · × vol S2n−1 . (1.3)

We can see this in [8]. The volume of the sphere of unit radius embedded in R n (n > 1), is
calculated from the Gaussian integral (see also Appendix I for the details):
√ Z +∞
2
π= e−t dt.
−∞

Now
Z n Z
√
n +∞
− t2 2
π = e dt = e−k v k dv
−∞ Rn
Z +∞ Z Z +∞
2 2
= e−r dσdr = σn−1 (r)e−r dr,
0 S n −1 ( r ) 0
R
where σn−1 (r) = S n −1 (r ) dσ is the volume of sphere S n−1 (r) of radius r. Since

σn−1 (r) = σn−1 (1) × rn−1 ,

it follows that
√
n Z +∞ 2
π = σn−1 (1) × rn−1 e−r dr,
0

implying that

  2π 2
n
n −1
.
vol S := (1.4)
Γ n2

4
 Finally, we get the volume formula of a unitary group:

n ( n +1 )
n
2π k 2n π 2
vol (U (n)) := ∏ = . (1.5)
k=1
Γ ( k) 1!2! · · · (n − 1)!

2 Volumes of orthogonal groups

2.1 Preliminary

The following standard notations will be used [19]. Scalars will be denoted by lower-case letters,
vectors and matrices by capital letters. As far as possible variable matrices will be denoted by
X, Y, . . . and constant matrices by A, B, . . ..
Let A = [ aij ] be a n × n matrix, then Tr ( A) = ∑nj=1 a jj is the trace of A, and det( A) is the
determinant of A, and T
over a vector or a matrix will denote its transpose. Let X = [ xij ] be a
m × n matrix of independent real entries xij ’s. We denote the matrix of differentials by dX, i.e.
 
dx11 dx12 ··· dx1n
 
 dx21 dx22 · · · dx2n 
 
dX := [dxij ] =  . .. .. .. .
 .. . . . 
 
dxm1 dxm2 · · · dxmn

Then [dX ] stands for the product of the m × n differential elements

m n
[dX ] := ∏ ∏ dxij (2.1)
i=1 j=1

and when X is a real square symmetric matrix, that is, m = n, X = X T , then [dX ] is the product
of the n(n + 1)/2 differential elements, that is,
n n
[dX ] := ∏ ∏ dxij = ∏ dxij . (2.2)
j=1 i= j i> j

Throughout this paper, stated otherwise, we will make use of conventions that the signs will
be ignored in the product [dX ] of differentials of independent entries. Our notation will be the
following: Let X = [ xij ] be a m × n matrix of independent real entries. Then

[dX ] = ∧m n
i=1 ∧ j=1 dxij (2.3)

5
when [dX ] appears with integrals or Jacobians of transformations;

[dX ] = ∏m n
i=1 ∏ j=1 dxij (2.4)

when [dX ] appears with integrals involving density functions where the functions are nonnegative
and the absolute value of the Jacobian is automatically taken.
As Edelman said [7] before, many researchers in Linear Algebra have little known the fact that
the familiar matrix factorizations, which can be viewed as changes of variables, have simple Ja-
cobians. These Jacobians are used extensively in applications of random matrices in multivariate
statistics and physics.
It is assumed that the reader is familiar with the calculation of Jacobians when a vector of
scalar variables is transformed to a vector of scalar variables. The result is stated here for the
sake of completeness. Let the vector of scalar variables X be transformed to Y, where
   
x1 y1
 .   . 
X= .   . 
 .  and Y =  .  ,
xn yn

by a one-to-one transformation. Let the matrix of partial derivatives be denoted by

 
∂Y ∂yi
= .
∂X ∂x j
h i
∂y
The the determinant of the matrix ∂x ij is known as the Jacobian of the transformation X going
to Y or Y as a function of X it is written as
 
∂yi
J (Y : X ) = det or [dY ] = J (Y : X )[dX ], J 6= 0
∂x j
and
1
J (Y : X ) = or 1 = J (Y : X ) J ( X : Y ).
J (X : Y )
Note that when transforming X to Y the variables can be taken in any order because a permu-
tation brings only a change of sign in the determinant and the magnitude remains the same, that
is, | J | remains the same where | J | denotes the absolute value of J. When evaluating integrals
involving functions of matrix arguments one often needs only the absolute value of J. Hence
in all the statements of this notes the notation [dY ] = J [dX ] means that the relation is written
ignoring the sign.

Proposition 2.1. Let X, Y ∈ R n be of independent real variables and A ∈ R n×n be a nonsingular matrix
of constants. If Y = AX, then

[dY ] = det( A)[dX ]. (2.5)

6
Proof. The result follows from the definition itself. Note that when Y = AX, A = [ aij ] one has

yi = ai1 x1 + · · · + ain xn , i = 1, . . . , n

where x j ’s and y j ’s denote the components of the vectors X and Y, respectively. Thus the partial
derivative of yi with respect to x j is aij , and then the determinant of the Jacobian matrix is
det( A).

In order to see the results in the more complicated cases we need the concept of a tensor
product.

Definition 2.2 (Tensor product). Let A = [ aij ] ∈ R p×q and B = [bij ] ∈ R m×n . Then the tensor
product, denoted by ⊗, is a pm × qn matrix in R pm×qn, formed as follows:
 
a11 B a12 B · · · a1q B
 
 a21 B a22 B · · · a2q B 
 
B⊗ A =  . .. .. ..  (2.6)
 .. . . . 
 
a p1 B a p2 B · · · a pq B

and
 
b11 A b12 A ··· b1n A
 
 b21 A b22 A · · · b2n A 
 
A⊗B =  . .. .. .. . (2.7)
 .. . . . 
 
bm1 A bm2 A · · · bmn A

Definition 2.3 (Vector-matrix correspondence). Let X = [ xij ] ∈ R m×n matrix. Let the j-th column
of X be denoted by X j . That is, X = [ X1 , . . . , Xn ], where
 
x1j
 . 
Xj =   . .
. 
xmj

Consider an mn-dimensional vector in R mn , formed by appending X1 , . . . , Xn and forming a long

string. This vector will be denoted by vec( X ). That is,
 
X1
 . 
vec( X ) =  . 
 . . (2.8)
Xn

From the above definition, we see that the vec mapping is a one-to-one and onto correspon-
dence from R m×n to R n ⊗ R m . We also see that vec( AXB) = ( A ⊗ BT ) vec( X ) if the product AXB
exists.

7
Proposition 2.4. Let X, Y ∈ R m×n be of independent real variables and A ∈ R m×m and B ∈ R n×n
nonsingular matrices of constants. If Y = AXB, then

[dY ] = det( A)n det( B)m [dX ]. (2.9)

Proof. Since Y = AXB, it follows that vec(Y ) = ( A ⊗ BT ) vec( X ). Then by using Proposition 2.1,
we have
  
∂Y ∂(vec(Y ))
J (Y : X ) = det := det
∂X ∂(vec( X ))
= det( A ⊗ B ) = det( A ⊗ 1 n ) det(1 m ⊗ BT )
T

= det( A)n det( B)m ,

implying that
[dY ] = J (Y : X )[dX ] = det( A)n det( B)m [dX ].

This completes the proof.

Remark 2.5. Another approach to the proof that [dZ ] = det( A)n [dX ], where Z = AX, is de-
scribed as follows: we partition Z and X, respectively, as: Z = [ Z1 , . . . , Zn ], X = [ X1 , . . . , Xn ].
Now Z = AX can be rewritten as Zj = AX j for all j. So
 
A
 
∂Z ∂( Z1 , . . . , Zn )  A 
 
= = .. , (2.10)
∂X ∂ ( X1 , . . . , X n )  . 
 
A

implying that  
A
 
 A 
 
[dZ ] = det  ..  [dX ] = det( A)n [dX ].
 . 
 
A

Proposition 2.6. Let X, A, B ∈ R n×n be lower triangular matrices where A = [ aij ] and B = [bij ] are
constant matrices with a jj > 0, b jj > 0, j = 1, . . . , n and X is a matrix of independent real variables. Then

Y = X + X T =⇒ [dY ] = 2n [dX ], (2.11)

!
n
j
Y = AX =⇒ [dY ] = ∏ ajj [dX ], (2.12)
j=1
!
n
n − j+1
Y = XB =⇒ [dY ] = ∏ b jj [dX ]. (2.13)
j=1

8
Thus
!
n
j n − j+1
Y = AXB =⇒ [dY ] = ∏ a jj b jj [dX ]. (2.14)
j=1

Proof. Y = X + X T implies that

     
x11 0 · · · 0 x11 x21 · · · xn1 2x11 x21 ··· xn1
     
 x21 x22 · · · 0     x21 
   0 x22 · · · x2n   2x22 · · · xn2 
 . .. .. 
..  + .. .. .. .. = . .. .. .. .
 .. . . .   .   .. . 
  . . .   . . 
xn1 xn2 · · · xnn 0 0 ··· xnn xn1 xn2 ··· 2xnn

When taking the partial derivatives the n diagonal elements give 2 each and others unities
and hence [dY ] = 2n [dX ]. If Y = AX, then the matrices of the configurations of the par-
tial derivatives, by taking the elements in the orders (y11 , y21 . . . , yn1 ); (y22 , . . . , yn2 ); . . . ; ynn and
( x11 , x21 . . . , xn1 ); ( x22 , . . . , xn2 ); . . . ; xnn are the following:
∂(y11 , y21 . . . , yn1 ) ∂(y22 , . . . , yn2 )
= A, = A[1̂|1̂], . . . ,
∂( x11 , x21 . . . , xn1 ) ∂( x22 , . . . , xn2 )
∂ynn
= A[1̂ · · · n[− 1|1̂ · · · n[
− 1] = ann ,
∂xnn

where A[î1 · · · iˆµ | ĵ1 · · · ĵν ] means that the obtained submatrix from deleting both the i1 , . . . , iµ -th
rows and the j1 , . . . , jν -th columns of A. Thus
 ∂(y ,y ,...,y ) 
11 21 n1
∂( x11 ,x21 ,...,xn1 )
0 ··· 0
 ∂( y22 ,...,yn2 )

∂Y  0 ··· 0 
 ∂( x22 ,...,xn2 ) 
=  .. .. .. .. 
∂X  . . . . 
 
∂ynn
0 0 ··· ∂xnn
 
A 0 ··· 0
 
 0 A[1̂|1̂] · · · 0 
 
=  .. .. .. .. .
 . . . . 
 
0 0 ··· A[1̂ · · · n[
− 1|1̂ · · · n[
− 1]

We can also take another approach to this proof. In fact, we partition X, Y by columns, respec-
tively, Y = [Y1 , . . . , Yn ] and X = [ X1 , . . . , Xn ]. Then Y = AX is equivalent to Yj = AX j , j = 1, . . . , n.
Since Y, X, A are lower triangular, it follows that
   
y11 x11    
    y22 x22
 y21   x21  
    ..   
 = A[1̂|1̂]  ...  , . . . , ynn = A[1̂ · · · n[
 .  = A . , . − 1|1̂ · · · n[
− 1] xnn = ann xnn .
 ..   ..     
   
yn2 xn2
yn1 xn1

9
Now
n n
[dY ] = ∏[dYj ] = det( A) det( A[1̂|1̂]) · · · det( A[1̂ · · · n[
− 1|1̂ · · · n[
− 1]) ∏[dX j ]
j=1 j=1
!
n
j
= ∏ ajj [dX ].
j=1

Next if Y = XB, that is,

  
x11 0 ··· 0 b11 0 ··· 0
  
 x21 x22 · · · 0   ··· 
   b21 b22 0 
Y = XB =  . .. .. 
..   .. .. .. .. 
 .. . . .  . 
  . . . 
xn1 xn2 · · · xnn bn1 bn2 · · · bnn
 
x11 b11 0 ··· 0
 
 x21 b11 + x22 b21 x22 b22 ··· 0 
 
=  . .. .. .. 
 .. . . . 
 
n n
∑ j=1 xnj b j1 ∑ j=1 xnj b j2 ··· xnn bnn

The matrices of the configurations of the partial derivatives, by taking the elements in the orders
y11 ; (y21 , y22 ); . . . ; (yn1 , . . . , ynn ) and x11 ; ( x21 , x22 ); . . . ; ( xn1 , . . . , xnn ) are the following:
" # " #T
∂y11 ∂(y21 , y22 ) b11 b21 b11 0
= b11 , = = ,
∂x11 ∂( x21 , x22 ) 0 b22 b21 b22
   T
b11 b21 b31 b11 0 0
∂(y31 , y32 , y33 )    
= 0 b22 b32  
 =  b21 b22 0   ,...,
∂( x31 , x32 , x33 )
0 0 b33 b31 b32 b33
 T
b11 0 ··· 0
 
 b22 · · · 0 
∂(yn1 , yn2 , . . . , ynn )  b21 
= . .. .. ..  .
∂( xn1 , xn2 , . . . , xnn )  .. . . . 
 
bn1 bn2 · · · bnn

Thus
 ∂y11

∂x11 0 ··· 0
 ∂( y21 ,y22 ) 
∂Y  0 ··· 0 
 ∂( x21 ,x22 ) 
= .. .. .. .. 
∂X  . . . . 
 
∂( yn1 ,yn2 ,...,ynn )
0 0 ··· ∂( xn1 ,xn2 ,...,xnn )

Denote by B[i1 . . . iµ | j1 . . . jν ] the sub-matrix formed by the i1 , . . . , iµ -th rows and j1 , . . . , jν -th

10
columns of B. Hence
 T
B [ 1| 1] 0 ··· 0
 
∂Y  0 B[12|12] · · · 0 
 
= .. .. .. ..  .
∂X  . . . . 
 
0 0 ··· B [1 · · · n |1 · · · n ]

The whole configuration is a upper triangular matrix with b11 appearing n times and b22 appear-
ing n − 1 times and so on in the diagonal. Also we give another approach to derive the Jacobian
for Y = XB. Indeed, we partition Y, X by rows, respectively,
   
Y1 X1
   
 Y2   X2 
   
Y =  . , X =  . ,
 ..   .. 
   
Yn Xn

where Yj , X j are row-vectors. So Y = XB is equivalent to Yj = X j B, j = 1, . . . , n. Moreover

y11 = x11 B[1|1] = x11 a11 , [y21 , y22 ] = [ x21 , x22 ] B[12|12], . . . ,
[yn1 , . . . , ynn ] = [yn1 , . . . , ynn ] B[1 . . . n|1 . . . n].

Therefore
n n
[dY ] = ∏[dYj ] = det( B[1|1]) det( B[12|12]) · · · det( B[1 · · · n|1 · · · n]) ∏[dXj ]
j=1 j=1
!
n
n + j−1
= ∏ bjj [dX ].
j=1

We are done.

Proposition 2.7. Let X be a lower triangular matrix of independent real variables and A = [ aij ] and
B = [bij ] be lower triangular matrices of constants with a jj > 0, bij > 0, j = 1, . . . , n. Then
!
n
j
Y = AX + X T AT =⇒ [dY ] = 2n ∏ ajj [dX ], (2.15)
j=1
!
n
n − j+1
Y = XB + BT X T =⇒ [dY ] = 2n ∏ bjj [dX ]. (2.16)
j=1

Proof. Let Z = AX. Then Y = Z + ZT . Thus [dY ] = 2n [dZ ]. Since Z = AX, it follows from
Proposition 2.6 that !
n
j
[dZ ] = ∏ a jj [dX ],
j=1

implying the result. The proof of the second identity goes similarly.

11
Proposition 2.8. Let X and Y be n × n symmetric matrices of independent real variables and A ∈ R n×n
nonsingular matrix of constants. If Y = AX AT , then

[dY ] = det( A)n+1 [dX ]. (2.17)

Proof. Since both X and Y are symmetric matrices and A is nonsingular we can split A and AT
as products of elementary matrices and write in the form

Y = · · · E2 E1 XE1T E2T · · ·

where Ej , j = 1, 2, . . . are elementary matrices. Write Y = AX AT as a sequence of transformations

of the type

Y1 = E1 XE1T , Y2 = E2 Y1 E2T , . . . , =⇒
[dY1 ] = J (Y1 : X )[dX ], [dY2 ] = J (Y2 : Y1 )[dY1 ], . . .

Now successive substitutions give the final result as long as the Jacobians of the type J (Yk :
Yk−1 ) are computed. Note that the elementary matrices are formed by multiplying any row (or
column) of an identity matrix with a scalar, adding a row (column) to another row (column) and
combinations of these operations. Hence we need to consider only these two basic elementary
matrices. Let us consider a 3 × 3 case and compute the Jacobians. Let E1 be the elementary
matrix obtained by multiplying the first row by α and E2 by adding the first row to the second
row of an identity matrix. That is,
   
α 0 0 1 0 0
   
E1 =    
 0 1 0  , E2 =  1 1 0 
0 0 1 0 0 1

and
 
α2 x11 αx12 αx13
 
E1 XE1T =  αx21 x22 x23 ,
αx31 x32 x33
 
u11 u11 + u12 u13
 
E2 Y1 E2 = 
T
 u11 + u21 u11 + u21 + u12 + u22 u13 + u23
,

u31 u31 + u32 u33

where Y1 = E1 XE1T and Y2 = E2 Y1 E2T and the elements of Y1 are denoted by uij ’s for convenience.

12
The matrix of partial derivatives in the transformation Y1 written as a function of X is then
 
α2 0 0 0 0 0
 
 0 α 0 0 0 0 
 
 
∂Y1  0 0 1 0 0 0 
=  .
∂X 
 0 0 0 α 0 0 
 
 0 0 0 0 1 0 
 
0 0 0 0 0 1

This is obtained by taking the xij ’s in the order x11 ; ( x21 , x22 ); ( x31 , x32 , x33 ) and the uij ’s also in
the same order. Thus the Jacobian is given by

J (Y1 : X ) = α4 = α3+1 = det( E1 )3+1 .

Or by definition it follows directly that

[dY1 ] = d(α2 x11 )d(αx12 )d(αx13 )dx22 dx23 dx33 = α4 [dX ].

For a n × n matrix it will be αn+1 . Let the elements of Y2 be denoted by vij ’s. Then again taking
the variables in the order as in the case of Y1 written as a function of X the matrix of partial
derivatives in this transformation is the following:
 
1 0 0 0 0 0
 
 1 1 0 0 0 0 
 
 
∂Y2  1 2 1 0 0 0 
= 
.

∂Y1  0 0 0 1 0 0 
 
 0 0 0 1 1 0 
 
0 0 0 0 0 1

The determinant of this matrix is 1 = 13+1 = det( E2 )3+1 . In general such a transformation gives
the Jacobian, in absolute value, as 1 = 1n+1 . Thus the Jacobian is given by

J (Y : X ) = det(· · · E2 E1 )n+1 = det( A)n+1 .

We are done.

Example 2.9. Let X ∈ R n×n be a real symmetric positive definite matrix having a matrix-variate
gamma distribution with parameters (α, B = BT > 0). We show that

Z
1 n +1 n−1
det( B)−α = [dX ] det( X )α− 2 e− Tr (BX ) , Re(α) > . (2.18)
Γn (α) X >0 2

13
 Indeed, since B is symmetric positive definite there exists a nonsingular matrix C such that
B = CCT . Note that

Tr ( BX ) = Tr (CT XC ) .

Let

U = CT XC =⇒ [dU ] = det(C )n+1 [dX ]

from Proposition 2.8 and det( X ) = det( B)−1 det(U ). The integral on the right reduces to the
following:
Z Z
n +1 n +1
[dX ] det( X )α− 2 e− Tr ( BX ) = det( B)−α [dU ] det(U )α− 2 e− Tr (U ) .
X >0 U >0

But
Z
n +1
[dU ] det(U )α− 2 e− Tr (U ) = Γn (α)
U >0
n −1
for Re(α) > 2 . The result is obtained.

Proposition 2.10. Let X, Y ∈ R n×n skew symmetric matrices of independent real variables and A ∈
R n×n nonsingular matrix of constants. If Y = AX AT , then

[dY ] = det( A)n−1 [dX ]. (2.19)

Note that when X is skew symmetric the diagonal elements are zeros and hence there are
n ( n − 1)
only 2 independent variables in X.

Proposition 2.11. Let X, A, B ∈ R n×n be lower triangular matrices where A = [ aij ] and B = [bij ]
are nonsingular constant matrices with positive diagonal elements, respectively, and X is a matrix of
independent real variables. Then
!
n
j
Y = AT X + X T A =⇒ [dY ] = 2n ∏ ajj [dX ], (2.20)
j=1
!
n
n n − j+1
Y = XB + BX T T
=⇒ [dY ] = 2 ∏ b jj [dX ]. (2.21)
j=1

Proof. Consider Y = AT X + X T A. Premultiply by ( AT )−1 and postmultiply by A−1 to get the

following:

Y = AT X + X T A =⇒ ( AT )−1 YA−1 = ( AT )−1 X T + X A−1 .

Let
!
 T n
−1 −1 n −( n− j+1)
Z = XA + XA =⇒ [dZ ] = 2 ∏ a jj [dX ]
j=1

14
by Proposition 2.7 and
 T
Z = A−1 YA−1 =⇒ [dZ ] = det( A)−(n+1) [dY ]

by Proposition 2.8. Now writing [dY ] in terms of [dX ] one has

! ! !
n n n
−( n− j+1) j
[dY ] = ∏ ajj 2n ∏ anjj+1 [dX ] = 2n ∏ ajj [dX ]
j=1 j=1 j=1

since det( A) = ∏nj=1 a jj because A is lower triangular. Thus the first result follows. The second
is proved as follows. Clearly,
 T
Y = BX T + XBT =⇒ B−1Y ( BT )−1 = B−1 X + B−1 X := Z.
 
−j
Thus [dZ ] = det( B)−(n+1) [dY ] and [dZ ] = 2n ∏ nj=1 a jj [dX ]. Therefore expressing [dY ] in
terms of [dX ] gives the second result.

Proposition 2.12. Let X ∈ R n×n be a symmetric positive definite matrix of independent real variables
and T = [tij ] a real lower-triangular matrix with t jj > 0, j = 1, . . . , n, and tij , i > j independent. Then
!
n
n j
T
X = T T =⇒ [dX ] = 2 ∏ t jj [dT ], (2.22)
j=1
!
n
n − j+1
X = TT T =⇒ [dX ] = 2n ∏ tjj [dT ]. (2.23)
j=1

Proof. By considering the matrix of differentials one has

X = TT T =⇒ dX = dT · T T + T · dT T .

Now treat this as a linear transformation in the differentials, that is, dX and dT as variables and
T a constant. This completes the proof.

n −1
Example 2.13. Let X ∈ R n×n symmetric positive definite matrix and Re(α) > 2 . Show that

Z
n +1
Γn (α) := [dX ] det( X )α− 2 e− Tr ( X )
X >0
   
n ( n −1 ) 1 n−1
= π 4 Γ(α)Γ α − ···Γ α− . (2.24)
2 2

15
 Let T be a real lower triangular matrix with positive diagonal elements. Then the unique
representation (see Theorem 8.5)
!
n
n − j+1
X = TT T =⇒ [dX ] = 2n ∏ t jj [dT ].
j=1

Note that


Tr ( X ) = Tr ( TT T ) = t211 + (t221 + t222 ) + · · · + t2n1 + · · · + t2nn ,
n
det( X ) = det( TT T ) = ∏ t2jj .
j=1

When X > 0, we have TT T > 0, but t jj > 0, j = 1, . . . , n which means that −∞ < tij < ∞, i >
n ( n − 1)
j, 0 < t jj < ∞, j = 1, . . . , n. The integral splits into n integrals on t jj ’s and 2 integrals on tij ’s,
i > j. That is,
Z ∞
! Z ∞
!
n α− n+2 1
n − j+1 − t2jj − t2ij
Γn (α) = ∏2 t2jj t jj e dt jj × ∏ e dtij .
j=1 0 i> j −∞

But
Z ∞  
j 2 j−1
2 (t2jj )α− 2 e−t jj dt jj = Γ α − ,
0 2
j−1
for Re(α) > 2 ,j = 1, . . . , n and
Z ∞ √
− t2jj
e dt jj = π.
−∞

Multiplying them together the result follows. Note that

j−1 n−1
Re(α) > , j = 1, . . . , n =⇒ Re(α) > .
2 2
The next result is extremely useful and uses the fact (see Theorem 8.5) that any positive
definite n × n matrix X has a unique decomposition as X = T T T, where T is an upper-triangular
n × n matrix with positive diagonal elements.

Proposition 2.14. If X is an n × n positive definite matrix and X = T T T, where T is upper-triangular

with positive diagonal elements, then
n
n − j+1
[dX ] = 2n ∏ t jj [dT ].
j=1

Proof. Since X = T T T, now express each of the elements of X on and above the diagonal in
terms of each of the elements of T and take differentials. Remember that we are going to take
the exterior product of these differentials and that products of repeated differentials are zero;

16
hence there is no need to keep track of differentials in the elements of T which have previously
occurred. We get

x11 = t211 , dx11 = 2t11 dt11 ,

x12 = t11 t12 , dx12 = t11 dt12 + · · ·
..
.
x1n = t11 t1n , dx1n = t11 dt1n + · · ·
x22 = t212 + t222 , dx22 = 2t22 dt22 + · · ·
..
.
x2n = t12 t1n + t22 t2n , dx2n = t22 dt2n + · · ·
..
.
xnn = t21n + · · · + t2nn , dxnn = 2tnn dtnn + · · ·

Hence taking exterior products gives

n
^ n
^ n
n n −1 n − j+1
[dX ] = dxij = 2n t11 t22 · · · tnn dtij = 2n ∏ t jj [dT ],
i6 j i6 j j=1

as desired.

2.2 The computation of volumes

Definition 2.15 (Stiefel manifold). Let A be a n × m(n > m) matrix with real entries such that
AT A = 1 m , that is, the m columns of A are orthonormal vectors. The set of all such matrices A is
known as the Stiefel manifold, denoted by O(m, n). That is, for all n × m matrices A,

O(m, n) = { A ∈ R n×m : AT A = 1 m } , (2.25)

where R n×m denotes the set of all n × m real matrices.

The equation AT A = 1 m imposes m(m + 1)/2 conditions on the elements of A. Thus the
number of independent entries in A is mn − m(m + 1)/2.
If m = n, then A is an orthogonal matrix. The set of such orthogonal matrices form a group.
This group is known as the orthogonal group of m × m matrices.

Definition 2.16 (Orthogonal group). Let B be a n × n matrix with real elements such that BT B =
1 n . The set of all B is called an orthogonal group, denoted by O(n). That is,

O(n) = { B ∈ R n×n : BT B = 1 n } . (2.26)

17
 Clearly O(n, n) = O(n). Note that BT B = 1 n imposes n(n + 1)/2 conditions and hence the
number of independent entries in B is only n2 − n(n + 1)/2 = n(n − 1)/2.

Definition 2.17 (A symmetric or a skew symmetric matrix). Let A ∈ R n×n . If A = AT , then A is

said to be symmetric and if AT = − A, then it is skew symmetric.

Proposition 2.18. Let V ∈ O(n) with independent entries and the diagonal entries or the entries in the
first row of V all positive. Denote dG = V T dV where V = [v1 , . . . , vn ]. Then
n −1 n
[dG ] = ∏ ∏ hvi , dv j i (2.27)
i=1 j= i+1

= 2n(n−1)/2 det(1 n + X )−(n−1) [dX ], (2.28)

where X is a skew symmetric matrix such that the first row entries of (1 n + X )−1 , except the first entry,
are negative.

Proof. Let the columns of V be denoted by v1 , . . . , vn . Since the columns are orthonormal, we
have hvi , v j i = δij . Then
hvi , dv j i + hdvi , v j i = 0,

implying hv j , dv j i = 0 since hv j , dv j i is a real scalar. We also have

hvi , dv j i = −hv j , dvi i for i 6= j.

Then V T dV is a skew symmetric matrix. That is,

 
vT1
 
 vT 
T  2 
dG = V dV =  .  [dv1 , dv2 , · · · , dvn ]
 .. 
 
T
vn
 
hv1 , dv1 i hv1 , dv2 i · · · hv1 , dvn i
 
 hv2 , dv1 i hv2 , dv2 i · · · hv2 , dvn i 
 
=  .. .. .. .. .
 . . . . 
 
hvn , dv1 i hvn , dv2 i · · · hvn , dvn i

This indicates that

 
0 hv1 , dv2 i ··· hv1 , dvn i
 
 −hv1 , dv2 i 0 ··· hv2 , dvn i 
 
dG =  . . .. .. 
 .. .. . . 
 
−hv1 , dvn i −hv2 , dvn i · · · 0

18
Then there are only n(n − 1)/2 independent entries in G. Then [dG ] is the wedge product of the
entries upper the leading diagonal in the matrix V T dV:

[dG ] = ∧in=−11 ∧nj=i+1 hvi , dv j i

This establishes (2.27). For establishing (2.28), take a skew symmetric matrix X, then V = 2(1 n +
X )−1 − 1 n is orthonormal such that VV T = 1 n . Further the matrix of differentials in V is given
by dV = −2(1 n + X )−1 · dX · (1 n + X )−1 , i.e.

1
dV = − (1 n + V ) · dX · (1 n + V ).
2
Thus
1
dG = V T dV = − (1 n + V T ) · dX · (1 n + V )
2
and the wedge product is obtained
  n −1 √  n −1
1n + V T
[dG ] = det √ [dX ] = det 2( 1 n + X ) − 1 [dX ].
2
Therefore the desired identity (2.28) is proved.

Proposition 2.19. Let X be a n × n symmetric matrix of independent real entries and with distinct and
nonzero eigenvalues λ1 > · · · > λn and let D = diag(λ1 , . . . , λn ). Let V ∈ O(n) be a unique such that
X = VDV T . Then
  
[dX ] = ∏ in=−11 ∏ nj=i+1  λi − λ j  [dD ][dG ], (2.29)

where dG = V T dV.

Proof. Take the differentials in X = VDV T to get

dX = dV · D · V T + V · dD · V T + V · D · dV T ,

implying that
V T · dX · V = V T dV · D + dD + D · dV T V.

Let dY = V T · dX · V for fixed V. Since dG = V T dV, it follows that [dX ] = [dY ] and

dY = dG · D + dD + D · dGT
= dD + dG · D − D · dG
= dD + [dG, D ]

19
where we used the fact that dGT = −dG and the concept of commutator, defined by [ M, N ] :=
MN − NM. Clearly the commutator of M and N is skew symmetric. Now

dy jj = dλ j and dyij = (λ j − λi )dgij , i < j = 1, . . . , n.

Then
! ! ! !
n n n  
[dY ] = ∏ dyjj ∏ dyij = ∏ dλ j ∏  λi − λ j  dgij
j=1 i< j j=1 i< j
!
 
= ∏  λi − λ j  [dD ][dG ],
i< j

where [dD ] = ∏nj=1 dλ j and [dG ] = ∏ i< j dgij , and the desired conclusion is obtained.

Clearly X = VDV T , where D = diag(λ1 , . . . , λn ) (λ1 > · · · > λn ), VV T = 1 n , is not a one-

to-one transformation from X to ( D, V ) since X determines 2n matrices [±v1 , . . . , ±vn ], where
v1 , . . . , vn are the columns of V, such that X = VDV T .
This transformation can be shown to be unique if one entry from each row and column are
of a specified sign, for example, the diagonal entries are positive. Once this is done we are
integrating with respect to dG, where dG = V T dV over the full orthogonal group O(n), the
result must be divided by 2n to get the result for a unique transformation X = VDV T .

Remark 2.20. Now we can try to compute the following integral based on Eq. (2.29):
Z Z
! Z
n   n
[dX ] = δ ∑ λ j − 1 ∏  λi − λ j  ∏ dλ j × [dG ]
X >0:Tr ( X )=1 λ1 >···> λn >0 j=1 i< j j=1 O1 ( n )
Z ∞
! Z
n   n
1   1
=
n! 0
δ ∑ λ j − 1 ∏ λi − λ j ∏ dλ j × 2n O(n) [dG]
j=1 i< j j=1
1 1
= vol (O(n)) ,
2n n! (1,1)
Cn
that is,
Z n ( n −1 )
 
n π 4 Γ n+ 2
1 n
j
vol (D (R )) := [dX ] =  
∏Γ 2 .
X >0:Tr ( X )=1 Γ n(n4+1) Γ 21 j=1

(1,1)
See below for the notation Cn and vol (O(n)).

Proposition 2.21. Let X be a p × n( p 6 n) matrix of rank p and let X = TU1T , where T is a p × p

lower triangular matrix with distinct nonzero diagonal entries and U1 is a unique n × p semi-orthogonal
matrix, U1T U1 = 1 p , all are of independent real entries. Let U2 be an n × (n − p) semi-orthogonal matrix

20
such that U1 augmented with U2 is a full orthogonal matrix. That is, U = [U1 U2 ], U T U = 1 n , U2T U2 =
1 n− p , U1T U2 = 0. Let u j be the j-th column of U and du j its differential. Then
 
p  n− j
[dX ] = ∏ j=1  t jj  [dT ][dU1 ], (2.30)

where
p
[dU1 ] = ∧ j=1 ∧ni= j+1 hui , du j i.

Proof. Note that

" # " # " #
U1T U2T U2 U1T U2 1p 0
U TU = [U1 U2 ] = = .
U2T U2T U1 U2T U2 0 1 n− p

Take the differentials in X = TU1T to get

dX = dT · U1T + T · dU1T .

Then

dX · U = dT · U1T U + T · dU1T U
= dT · U1T [U1 , U2 ] + T · dU1T [U1 , U2 ]
= [dT + T · dU1T · U1 , T · dU1T · U2 ]

since U1T U1 = 1 p , U1T U2 = 0. Make the substitutions

dW = dX · U, dY = dU1T · U1 , dS = dU1T · U2 , dH = T · dS.

Now we have
dW = [dT + T · dY, dH ].

Thus
    
dt11 0 ··· 0 t11 0 ··· 0 0 dy12 · · · dy1p
 ..   ..   
 dt   .   
 21 dt22 · · · .   t21 t22 ···   −dy12 0 ··· dy2p 
dT + T · dY =  . .. + .. ..  .. .. .. .. .
 .. ..   ..
 . . 0   . . . 0  
 . . . . 

dt p1 dt p2 · · · dt pp t p1 t p2 · · · t pp −dy1p −dy2p · · · 0

Let us consider, for example, the case where p = 2, 3 in order for computing the wedge product
of dT + T · dY. Now for p = 2, we have
" # " #" #
dt11 0 t11 0 0 dy12
dT + T · dY = +
dt21 dt22 t21 t22 −dy12 0
" #
dt11 t11 dy12
=
dt21 − t22 dy12 dt22 + t21 dy12

21
Thus the wedge product of dT + T · dY is:

[dT + T · dY ] = dt11 ∧ (t11 dy12 ) ∧ (dt21 − t22 dy12 ) ∧ (dt22 + t21 dy12 )
= t11 dt11 ∧ dy12 ∧ dt21 ∧ dt22 = t11 [dT ][dY ]
!
2  
2− j
= ∏  tjj  [dT ][dY ].
j=1

For p = 3, we have
    
dt11 0 0 t11 0 0 0 dy12 dy13
    
dT + T · dY = 
 dt 21 dt 22 0  +  t21 t22 0   −dy12
   0 dy 23


dt31 dt32 dt33 t31 t32 t33 −dy13 −dy23 0
 
dt11 t11 dy12 t11 dy13
 

=  dt21 − t22 dy12 dt22 + t21 dy12 t21 dy13 + t22 dy23 ,

dt31 − t32 dy12 − t33 dy13 dt32 + t31 dy12 − t33 dy23 dt33 + t31 dy13 + t32 dy23

implying the wedge product of dT + T · dY is:

!
3   3− j
t211 t22 [dT ][dY ] = ∏  tjj  [dT ][dY ].
j=1

For the general p, by straight multiplication, and remembering that the variables are only

dy12 , . . . , dy1p , dy23 , . . . , dy2p , . . . , dy p−1p .

Thus the wedge product of dT + T · dY gives

!
p
  p− j
∏  tjj  [dY ][dT ],
j=1

ignoring the sign, and

p −1 p
[dY ] = ∧ j=1 ∧i= j+1 hui , du j i.

Now consider dH = T · dS. Since dS is a p × (n − p) matrix, we have

!
p  
n − p
[dH ] = det( T ) [dS] = ∏  t jj 
n− p
[dS].
j=1

The wedge product in dS is the following:

p
[dS] = ∧ j=1 ∧ni= p+ j hui , du j i.

22
Hence from the above equations,

[dX ] = [dW ] = ∧[dT + TdY ] ∧ [dH ]

! !
p   p  
p − j n − p
= ∏  tjj  [dY ][dT ] ∏  t jj  [dS]
j=1 j=1
!
p  n− j
= ∏  tjj  [dY ][dT ][dS].
j=1

Now
p −1 p p
[dY ][dS] = ∧ j=1 ∧i= j+1 hui , du j i ∧ j=1 ∧ni= p+ j hui , du j i.

Substituting back one has

!
p n− j p
[dX ] = ∏  tjj  [dT ] ∧ j=1 ∧ni= j+1 hui , du j i
j=1

which establishes the result.

If the triangular matrix T is restricted to the one with positive diagonal entries, that is, t jj >
0, j = 1, . . . , p, then while integrating over T using Proposition 2.21, the result must be multiplied
by 2 p . Without the factor 2 p , the t jj ’s must be integrated over −∞ < t jj < ∞, j = 1, . . . , p. If the
expression to be integrated contains both T and U, then restrict t jj > 0, j = 1, . . . , p and integrate
U over the full Stiefel manifold. If the rows of U are u1 , . . . , u p , then ±u1 , . . . , ±u p give 2 p choices.
Similarly t jj > 0, t jj < 0 give 2 p choices. But there are not 22p choices in X = TU. There are only
2 p choices. Hence either integrate out the t jj ’s over −∞ < t jj < ∞ and a unique U or over
0 < t jj < ∞ and the U over the full Stiefel manifold. For uniqueness of matrix factorization, see
Theorem 8.6.

Proposition 2.22. Let X1 be an n × p(n > p) matrix of rank p of independent real entries and let
X1 = U1 T1 where T1 is a real p × p upper triangular matrix with distinct nonzero diagonal entries and
U1 is a unique real n × p semi-orthogonal matrix, that is, U1T U1 = 1 p . Let U = [U1 U2 ] such that
U T U = 1 n , U2T U2 = 1 n− p , U1T U2 = 0. Let u j be the j-th column of U and du j its differential. Then
 
p  n− j
[dX1 ] = ∏ j=1  t jj  [dT1 ][dU1 ], (2.31)

where
p n
[dU1 ] = ∏ ∏ hui , du j i.
j=1 i= j+1

Proposition 2.23. If X1 , T1 and U1 are as defined in Proposition 2.22, then the surface area of the full
p
Stiefel manifold O( p, n) or the total integral of the wedge product ∧ j=1 ∧ni= j+1 hui , du j i over O( p, n) is
given by

23
 Z pn
p 2pπ 2
∧ j=1 ∧ni= j+1 hui , du j i =
, (2.32)
O( p,n) Γ p n2

where    
p ( p −1 ) 1 p−1
Γ p (α) = π 4 Γ(α)Γ α − ··· α−
2 2
p −1
for Re(α) > 2 .

Proof. Note that since X1 is n × p, the sum of squares of the np variables in X1 is given by
n p
Tr ( X1T X1 ) = ∑ ∑ x2ij .
i=1 j=1

Then
Z Z ∞ Z ∞ p n p
− ∑ni=1 ∑ j =1 x2ij
[dX1 ]e− Tr ( X1 X1 ) =
T

X1 −∞
···
−∞
e ∏ ∏ dxij
i=1 j=1
n p Z ∞
2
= ∏∏ e− xij dxij
i=1 j=1 − ∞
np
= π 2

by direct evaluation of the exponential integrals. Make the transformation as in Proposition 2.22:

X1 = U1 T1 =⇒ X1T X1 = T1T T1 ,

where T1 is a real p × p upper triangular matrix with distinct nonzero diagonal entries and U1 is
a unique real n × p semi-orthogonal matrix—U1T U1 = 1 p , implying

Tr ( X1T X1 ) = Tr ( T1T T1 ) = ∑ t2ij .

i6 j

Note that dX1 is available from Proposition 2.22. Now

Z Z
! Z 
p n− j −
− Tr ( X1T X1 ) t2 p
X1
[dX1 ]e = ∏  tjj  e ∑i6j ij [dT1 ]
T1 j=1 O( p,n)
∧ j=1 ∧ni= j+1 hui , du j i .

But for 0 < t jj < ∞, −∞ < tij < ∞(i < j) and U1 unrestricted,
Z p   n− j n
− ∑i6 j t2ij
∏  tjj 
T1 j=1
e [dT1 ] = 2− p Γ p
2

24
observing that for j = 1, . . . , p, the p integrals
Z ∞  
 2 n j−1
 t jj  n− j e−t jj dt jj = 2−1 Γ − , n > j − 1,
0 2 2
and each of the p( p − 1)/2 integrals
Z ∞ √
2
e−tij dtij = π, i < j.
−∞

Thus the result that follows.

Theorem 2.24. Let X be a full-ranked and n × n matrix of independent real entries and let X = UT,
where T is a real n × n upper triangular matrix with distinct nonzero diagonal entries and U is a unique
real orthogonal matrix. Let u j be the j-th column of U and du j its differential. Then the volume content of
the full orthogonal group O(n) is given by

Z n2 n ( n +1 )
2n π 2 2n π 4
vol(O(n)) = ∧[U T dU ] =
=  . (2.33)
O( n) Γn n2 n
∏k=1 Γ 2k

Proposition 2.25. Let X be a p × n( p 6 n) matrix of rank p and let X = TU1T , where T is a p × p lower
triangular matrix with distinct positive diagonal entries t jj > 0, j = 1, . . . , p and U1 is a unique n × p
semi-orthogonal matrix, U1T U1 = 1 p , all are of independent real entries. Let A = XX T = TT T . Then
n p +1 p
[dX ] = 2− p det( A) 2 − 2 [dA] ∧ j=1 ∧ni= j+1 hui , du j i.

Proof. Since A = TT T , it follows that

!
p
p p + 1− j
[dA] = 2 ∏ t jj [dT ],
j=1

i.e. !
p
−p − p − 1+ j
[dT ] = 2 ∏ t jj [dA].
j=1

But
!
p n− j
[dX ] = ∏  tjj  [dT ][dU1 ],
j=1

where
p n
[dU1 ] = ∏ ∏ hui , du j i.
j=1 i= j+1
p
Note that det( A) = det( T )2 = ∏ j=1 t2jj . The desired conclusion is obtained.

25
Proposition 2.26. Let X be a m × n matrix of rank m(m 6 n), T = [t jk ] a m × m lower triangular
matrix with t jj > 0, j = 1, . . . , m and L a n × m matrix satisfying LT L = 1 m , where the matrices are of
independent real entries. Then show that, if X = TLT , then
 
n− j
[dX ] = ∏m t
j=1 jj [dT ][d L̂ ] (2.34)

where
m n
d L̂ = ∏ ∏ hli , dlj i,
j=1 i= j+1

l j is the j-th column of L̂ = [ L L1 ] ∈ O(n); dli the differential of the i-th column of L.

Proposition 2.27 (Polar decomposition). Let X be a m × n(m 6 n) matrix, S a m × m symmetric

positive definite matrix and L a n × m matrix with LT L = 1 m , all are of independent real entries. Then
√
show that, if X = SLT , then
 m
1 n − m −1
[dX ] = (det(S)) 2 [dS][d L̂ ] (2.35)
2

where d L̂ is defined in Proposition 2.26.

√
Proof. Now if X = SLT and LT L = 1 m , then XX T = S. By Proposition 2.26, we have X = TLT ,
where T = [t jk ] is a m × m lower triangular matrix with t jj > 0, j = 1, . . . , m and L a n × m matrix
satisfying LT L = 1 m . Denote L̂ = [ L L1 ] ∈ O(n). Hence
!
m
n− j
[dX ] = ∏ tjj [dT ][d L̂].
j=1

It also holds that S = TT T implies

!
m
m m + 1− j
[dS] = 2 ∏ t jj [dT ].
j=1

Both expressions indicate that

!
m
−m
[dX ] = 2 ∏ tnjj−m−1 [dS][d L̂].
j=1

Since det(S) = det( TT T ) = ∏ m 2

j=1 t jj , it follows that
 m
1 n − m −1
[dX ] = (det(S)) 2 [dS][d L̂].
2
We are done.

Proposition 2.28. With the same notations as in Proposition 2.27, it holds that

26
 Z mn
2m π 2
[d L̂] =
(2.36)
LT L =1 m Γm n2

and for m = n
Z n2
2n π 2
[dV̂ ] =
. (2.37)
O( n) Γn n2
Define the normalized orthogonal measures as

!
!
Γn n2 Γn n
2
dµ(V ) := n2
[dV̂ ] = n2 ∏hvi , dvj i (2.38)
2n π 2 2n π 2 i> j
or
n

!
Γn 2
dµ(V ) := n2
[dG ], (2.39)
2n π 2

where dG = V T dV for V = [v1 , . . . , vn ] ∈ O(n). It holds that [dV̂ ] = ∧(V T dV ) is invariant under
simultaneous translations V → UVW, where U, W ∈ O(n). That is, dµ(V ) is an invariant measure
under both left and right translations, i.e. Haar measure over O(n).

Proof. We know that Z

[dX ]e− Tr ( XX ) = π 2 .
T mn

X
√
From Proposition 2.27, via the transformation X = SLT , we see that
Z  m Z Z
− Tr ( XXT ) 1 n − m −1
− Tr (S)
[dX ]e = det(S) 2 e [dS] × [d L̂].
X 2 S >0 LT L =1 m

We also see from the definition of Γ p (α) that

n Z n m +1
Γm = det(S) 2 − 2 e− Tr (S) [dS].
2 S >0
Then
Z mn
2m π 2
[d L̂] =
.
LT L =1 m Γm n2
For m = n, the result follows easily.
For fixed U, W ∈ O(n), we have

(UVW )T d(UVW ) = (W T V T U T ) (U · dV · W ) = W T · dG · W,

implying that
∧ [(UVW )T d(UVW )] = [dG ].
That is, dµ(V ) = dµ(UVW ) for all U, V, W ∈ O(n), dµ(V ) is an invariant measure under both
left and right translations over U (n).

27
3 Volumes of unitary groups

3.1 Preliminary

In Section 2, we dealt with matrices where the entries are either real constants or real variables.
Here we consider the matrices whose entries are complex quantities. When the matrices are real,
we will use the same notations as in Section 2. In the complex case, the matrix variable X will
be denoted by X e to indicate that the entries in X are complex variables so that the entries of
theorems in Section 3 will not be confused with those in Section 2. The complex conjugate of a
e will be denoted by A
matrix A e and the conjugate transpose by A e∗ . The determinant of A e will
e). The absolute value of a scalar a will also be denoted by | a | . The wedge
be denoted by det( A
e will be denoted by [dX
product of differentials in X e ] and the matrix of differentials by dX.
e
It is assumed that the reader is familiar with the basic properties of real and complex matrices.
Some properties of complex matrices will be listed here for convenience.
√
A matrix X e with complex elements can always be written as X e = X1 + −1X2 where
X1 = Re( X e ) and X2 = Im( X e ) are real matrices. Let us examine the wedge product of the
e In general, there are n2 real variables in X1 and another n2 real variables and
differentials in X.
the wedge product of the differentials will be denoted by the following:

Notation 1:
h  i h  i
e ] := [dX1 ][dX2 ] or [dX
[d X e ] := d Re( X
e) e)
d Im( X (3.1)

where [dX1 ] is the wedge product in dX1 and [dX2 ] is the wedge product in dX2 . In this notation
an empty product is interpreted as unity. That is, when the matrix Xe is real then X2 is null and
e ] := [dX1 ]. If X
[d X e is a hermitian matrix, then X1 is symmetric and X2 is skew symmetric, and in
this case

( 1) ( 2)
[dX1 ] = ∧ j>k dx jk and [dX2 ] = ∧ j>k dx jk (3.2)
h i h i √
( 1) ( 2) e is a scalar function of X
e = X1 + −1X2 then Y e can be
where X1 = x jk and X2 = x jk . If Y
√
e = Y1 + −1Y2 where Y1 and Y2 are real. Thus if Y
written as Y e = F(Xe ) it is a transformation
of ( X1 , X2 ) to (Y1 , Y2 ) or where (Y1 , Y2 ) is written as a function of ( X1 , X2 ) then we will use the
following notation for the Jacobian in the complex case.

Notation 2: (Jacobians in the complex case). J (Y1 , Y2 : X1 , X2 ): Jacobian of the transforma-

√
tion where Y1 and Y2 are written as functions of X1 and X2 or where Y e = Y1 + −1Y2 is a
√
e = X1 + −1X2 .
function of X

28
 e ∈ C n×n and (2n) × (2n) matrices B and C where
Lemma 3.1. Consider a matrix A
" # " #
√ A 1 A 2 A 1 − A 2
Ae = A1 + −1A2 , B = ,C =
− A2 A1 A2 A1
where A1 , A2 ∈ R n×n . Then for det( A1 ) 6= 0
 
 e)  = | det( B)| 2 = | det(C )| 2 .
1 1
 det( A (3.3)
√
Proof.
q Let √ det( A) = a + −1b where a and b are real scalars. Then the absolute value is available
√ √ √ √
as ( a + −1b)( a − −1b). If det( A1 + −1A2 ) = a + −1b, then det( A1 − −1A2 ) = a −
√
−1b. Hence
√ √ √ √
( a + −1b)( a − −1b) = det( A1 + −1A2 ) det( A1 − −1A2 )
" √ #!
A1 + −1A2 0
= det √ .
0 A1 − −1A2
Adding the last n columns to the first n columns and then adding the last n rows to the first n
rows we have
" √ #! " √ #!
A1 + −1A2 0 2A1 A1 − −1A2
det √ = det √ √ .
0 A1 − −1A2 A1 − −1A2 A1 − −1A2
Using similar steps we have
" √ #! " √ #!
2A1 A1 − −1A2 2A1 A1 − −1A2
det √ √ = det √ √
A1 − −1A2 A1 − −1A2 − −1A2 21 A1 − 21 −1A2
" √ #!
2A1 − −1A2
= det √ 1
− −1A2 2 A1
= det( A1 ) det( A1 + A2 A1−1 A2 )
" #! " #!
A1 A2 A1 − A2
= det = det .
− A2 A1 A2 A1
e) is
by evaluating as the determinant of partitioned matrices. Thus the absolute value of det( A
given by
  q
 e)  = det( A1 ) det( A1 + A2 A−1 A2 ) = | det( B)| 2 = | det(C )| 2 .
1 1
det( A 1

This establishes the result.

Remark 3.2. Now we denote A1 = Re( A e) and A2 = Im( A

e). Clearly both Re( A
e) and Im( Ae) are
√
e := Re( A
real matrices. Each complex matrix A e) + −1Im( Ae) can be represented faithfully as a
block-matrix
" #
e) −Im( A
Re( A e)
e −→
A . (3.4)
e) Re( A
Im( A e)

29
Thus
" #
e)T Im( A
Re( A e)T
e∗ −→
A . (3.5)
−Im( Ae)T Re( A
e)T

e=A
Then Y eXe can be rewritten as, via block-matrix technique,
" # " #" #
e) −Im(Y
Re(Y e) e) −Im( A
Re( A e) e ) −Im( X
Re( X e)
= . (3.6)
Im(Ye) Re(Y e) e) Re( A
Im( A e) e ) Re( X
Im( X e)

From the above, we see that the mentioned representation is an injective ring homomorphism
which is continuous. sometimes we use the following representation:
" # " #" #
Re(Ye) e) −Im( A
Re( A e) e)
Re( X
= . (3.7)
Im(Ye) e) Re( A
Im( A e) e)
Im( X

Lemma 3.1 can be reexpressed as

 " #!
 e) −Im( A
Re( A e)   2  
   e)  =  det( A
eAe∗ )  .
 det  =  det( A (3.8)
 e) Re( A
Im( A e) 

Proposition 3.3. Let X, e ∈ C n be of n independent complex variables each, A

e Y e ∈ C n×n a nonsingular
e=A
matrix of constants. If Y eX,
e then
 
[dYe] =  det( A
eAe∗ )  [dX
e ]. (3.9)

e∗ = X
If Y e∗ , then
e∗ A
 
e∗ ] = (−1)n  det( A
[dY eAe∗ )  [dX
e ]. (3.10)
√ √
e = X1 +
Proof. Let X e = Y1 + −1Y2 , where Ym ∈
−1X2 , where Xm ∈ R n , m = 1, 2. Let Y
e = A
R n , m = 1, 2 are real. Y eX e = A is real. This
e implies that Ym = AXm , m = 1, 2 if A
transformation is such that the 2n real variables in (Y1 , Y2 ) are written as functions of the 2n real
variables in ( X1 , X2 ). Let

X1T = [ x11 , . . . , xn1 ], X2T = [ x12 , . . . , xn2 ],

Y1T = [y11 , . . . , yn1 ], Y2T = [y12 , . . . , yn2 ].

Then the Jacobian is the determinant of the following matrix of partial derivatives:
∂(Y1 , Y2 ) ∂(y11 , . . . , yn1 , y12 , . . . , yn2 )
= .
∂ ( X1 , X2 ) ∂( x11 , . . . , xn1 , x12 , . . . , xn2 )
Note that
∂(Y1 , Y2 ) ∂(y11 , . . . , yn1 , y12 , . . . , yn2 )
= .
∂ ( X1 , X2 ) ∂( x11 , . . . , xn1 , x12 , . . . , xn2 )

30
Note that
∂Y1 ∂(y11 , . . . , yn1 )
= = A,
∂X1 ∂( x11 , . . . , xn1 )
∂Y1 ∂Y2
= 0= ,
∂X2 ∂X1
∂Y2
= A.
∂X2
Thus the Jacobian is
" #!
A 0
J = det = det( A)2 .
0 A
√
e is complex, then let A
If A e = A1 + −1A2 where A1 and A2 are real. Then
√ √ √
Ye = Y1 + −1Y2 = ( A1 + −1A2 )( X1 + −1X2 )
√
= ( A 1 X1 − A 2 X2 ) + − 1 ( A 1 X2 + A 2 X1 )

implies that

Y1 = A1 X1 − A2 X2 , Y2 = A1 X2 + A2 X1 .

Then
∂Y1 ∂Y1
= A1 , = − A2 ,
∂X1 ∂X2
∂Y2 ∂Y2
= A2 , = A1 .
∂X1 ∂X2
Thus the Jacobian is
" #!
A1 − A2  2  
 e)  =  det( A
eAe∗ )  ,
J = det =  det( A
A2 A1

e∗ ] = [dY
which establishes the result. The second result follows by noting that [dY e1 ](−1)n [dY
e2 ] =
e ].
(−1)n [dY

Proposition 3.4. Let X, e Ye ∈ C m×n of mn independent complex variables each. Let A e ∈ C m×m and
e ∈ C n×n nonsingular matrices of constants. If Y
B e=A eX
e B,
e then
 n  m
[dY e] =  det( A eAe∗ )   det( B
eBe∗ )  [dX
e ]. (3.11)
√ √
e = Y1 + −1Y2 and X e = X1 + −1X2 . Indeed, let Y e = [Y e1 , . . . , Y
e e
Proof. Let Y   n ] and X =
e1 , . . . , X ej = A
e n ], then Y eX
e j , j = 1, . . . , n when Y e= A eX.
e Thus [dY  eA
ej ] =  det( A 
e∗ )  [ d Xe j ] for each
[X
j, therefore, ignoring the signs,
n  n
[dY ej ] =  det( A
e] = ∏[dY eAe∗ )  [dX
e ].
j=1

31
 √
e = A1 +
Denoting A −1A2 , the determinant is
" #! n
A1 − A2
J = det .
A2 A1
√
e = B1 +
Hence the Jacobian in this case, denoting B −1B2 , is given by
" #!m
B1 B2
J = det .
− B2 B1

For establishing our result, write Ye=A eZ

e where Z e=X e B.e That is,
 n  n  m
e] =  det( A
[dY eAe∗ )  [dZe ] =  det( A
eAe∗ )   det( B
eBe∗ )  [dX
e ].

This completes the proof.

 n
e] =  det( A
Remark 3.5. Another approach to the fact that [dY eAe∗ )  [dZ e = A
e ], where Y eZ,
e is
described as follows: 
Re(Y
e) e)Re( Z
= Re( A e)Im( Z
e ) − Im( A e)
Im(Y
e) e)Re( Z
= Im( A e)Im( Z
e ) + Re( A e)

leading to  
e), Im(Y e) " #
∂ Re(Y e)(n) −Im( A
Re( A e)(n)
  = ,
e ), Im( Z
∂ Re( Z e) e)(n) Re( A
Im( A e)(n)

where
   
e)
Re( A e)
Im( A
 ..   .. 
e)(n) := 
Re( A .  , Im( A
e)(n) :=  . .
   
e)
Re( A e)
Im( A

Then the Jacobian of this transformation can be computed as

" #!
  Re( Ae)(n) −Im( A
e)(n)
e), Im(Y
J Re(Y e) : Re( Z
e ), Im( Z
e) = det (3.12)
Im( Ae)(n) Re( A
e)(n)
  √ 2
 e)(n) + −1Im( Ae)(n)  .
= det Re( A (3.13)

That is,
    2  2n  n
e), Im(Y
J Re(Y e) : Re( Z e ) = det A
e ), Im( Z e(n)  = det( A
e)  = det( A
eAe∗ )  .

32
 e B
e A, e ∈ C n×n be lower triangular matrices where X
e is matrix of n ( n + 1)
Proposition 3.6. Let X, 2 indepen-
e B
dent complex variables, A, e are nonsingular matrices of constants. Then

e=X
Y e+X e] = 22n [dX
e T =⇒ [dY e ], (3.14)
e ] = 2n [ d X
=⇒ [dY e ] if the xejj ’s are real; (3.15)
!
n  2j
e=A
Y eX e] = ∏  e
e =⇒ [dY a jj  [d Xe ], (3.16)
j=1
!
n 2j−1
e] =
=⇒ [dY ∏  eajj  e ] if the e
[d X a jj ’s and xejj ’s are real; (3.17)
j=1
!
n  2( n − j + 1)
 
e=X
Y eB e] =
e =⇒ [dY ∏ ebjj  e ],
[d X (3.18)
j=1
!
n 2(n− j)+1
 
e] =
=⇒ [dY ∏ ebjj  e ] if the e
[d X b jj ’s and xejj ’s are real; (3.19)
j=1

Proof. Results (3.14) and (3.15) are trivial. Indeed, note that

2e
x jj , if j = k
yejk = .
 xe , if j > k
jk

By the definition, ignoring the sign, we have

e] = ∧ j>k de
[dY y jk = ∧nj=1 de
y jj ∧ j>k de
y jk ,

( 1) ( 2) ( 1) √ ( 2) (m) (m)
y jk := dy jk dy jk for yejk = y jk +
where de −1y jk . So for j = 1, . . . , n, we get y jj = 2x jj , m =
1, 2. Hence the result. If xejj ’s are real, the result follows easily by definition. Let
√ √ √ √
e = Y1 +
Y e = X1 +
−1Y2 , X e = A1 +
−1X2 , A e = B1 +
−1A2 , B −1B2 ,
(m) (m) (m) (m)
Ym = [y jk ], Xm = [ x jk ], Am = [ a jk ], Bm = [b jk ], m = 1, 2.

where Ym , Xm , Am , Bm , m = 1, 2 are all real.

When Y e = A eX
e we have Y1 = A1 X1 − A2 X2 and Y2 = A1 X2 + A2 X1 . The matrix of partial
∂Y1
derivative of Y1 with respect to X1 , that is ∂X1 , can be seen to be a lower triangular matrix with
( 1)
a jj repeated j times, j = 1, . . . , n, on the diagonal. Let this matrix be denoted by G1 . That is,
 
A1
 
∂Y1 ∂Y2  A1 [1̂|1̂] 
 
= := G1 =  .. .
∂X1 ∂X2  . 
 
A1 [1̂ · · · n[
− 1|1̂ · · · n[
− 1]

33
 ( 2)
Let G2 be a matrix of the same structure with a jj ’s on the diagonal. Similarly,
 
A2
 
∂Y1 ∂Y2  A2 [1̂|1̂] 
 
− = := G2 =  .. .
∂X2 ∂X1  . 
 
A2 [1̂ · · · n[
− 1|1̂ · · · n[
− 1]

Then the Jacobian matrix is given by

" #
∂(Y1 , Y2 ) G1 − G2
= .
∂ ( X1 , X2 ) G2 G1
√
e = G1 +
Let G −1G2 . Then
 
e
A
 
 e[1̂|1̂]
A 
e=
G 

,
 .. 
 . 
e[1̂ · · · n[
A − 1|1̂ · · · n[
− 1]

where
√ √
e = A1 +
A e[1̂|1̂] = A1 [1̂|1̂] + −1A2 [1̂|1̂], . . . ,
−1A2 , A
√
e[1̂ · · · n[
A − 1|1̂ · · · n[
− 1] = A1 [1̂ · · · n[ − 1] + −1A2 [1̂ · · · n[
− 1|1̂ · · · n[ − 1|1̂ · · · n[
− 1] .

Thus
  " #!
∂(Y1 , Y2 ) G1 − G2
det = det .
∂ ( X1 , X2 ) G2 G1
 2 √
 e)  where G e is
e = G1 + −1G2 . Since G
From Lemma 3.1, the determinant is available as det( G
triangular the absolute value of the determinant is given by
 2  2  2  2
 e[1̂|1̂])  · · ·  det( A
e)   det( A
e)  =  det( A e[1̂ · · · n[ 
 det( G − 1|1̂ · · · n[
− 1]) 
n  2
= ∏( ajj  )j .
j=1

This establishes (3.16). Another approach is presented also: Let Y e = [Y e1 , . . . , Y

en ], where Y
ej , j =
e Similarly for X
1, . . . , n, is the j-th column of the matrix Y. e = [X
e1 , . . . , X
e n ]. Now Y e=A eXe implies
that

ej = A
Y eXe j , j = 1, . . . , n.

34
That is,
           
ye11 xe11 0 0 0 0
           
 ye21   xe21   ye22   xe22   0   0 
  e     e     e  
 .  = A . , .  = A . ,..., .  = A . .
 ..   ..   ..   ..   ..   .. 
           
yen1 xen1 yen2 xen2 yenn xenn

e X,
Since Y, e are all lower triangular, it follows that
e A
   
ye11 xe11    
    ye xe22
 ye21   xe21   22   
    . e[1̂|1̂]  ...  , . . . ,
 .  = A e , . =A
 ..   ...   .   
   
yen2 xen2
yen1 xen1
e[1̂ . . . n[
yenn = A 1 . . . n[
− 1|b − 1] xenn ,

e
where A[î ĵ |î ĵ ] stands for a matrix obtained from deleting the i, j-th rows and columns of A,
respectively. We can now draw the conclusion that
 
[dYej ] =  det( A[1̂ . . . [
j − 1]|1̂ . . . [
j − 1]) A[1̂ . . . [
j − 1]|1̂ . . . [
 e
j − 1])∗  [dX j ],

that indicates that

n 
n 
e] =  [ [ [ [ ∗
[dY ∏[dYej ] = ∏ det( A[1̂ . . . j − 1]|1̂ . . . j − 1]) A[1̂ . . . j − 1]|1̂ . . . j − 1])  [dXe j ]
j=1 j=1
  
 e∗ )  det( A e[1̂|1̂]∗ )  · · · | e
eA
= det( A e[1̂|1̂] A a∗nn | [dX
ann e e]

= |e ann |2 × | e
a22 · · · e
a11 e a33 · · · e
a22 e ann |2 [dX
ann |2 × · · · × | e e]
!
n  2j
= ∏  eajj  [dXe ].
j=1

If the xejj ’s and e

a jj ’s are real then note that the x jk ’s for j > k contribute e
a jj twice that is, corre-
( 1) ( 2) ( 1)
sponding to x jk and x jk , whereas the e
a jj ’s appear only once corresponding to the x jj ’s since the
( 2) e=X
eBe and if a matrix H1 is defined corresponding to
x jj ’s are zeros. This establishes (3.17). If Y
( 1)
G1 then note that the b jj ’s appear n − j + 1 times on the diagonal for j = 1, . . . , n. Results (3.18)
and (3.19) are established by using similar steps as in the case of (3.16) and (3.17).

Proposition 3.7. Let X e ∈ C n×n be a

e ∈ C n×n be hermitian matrix of independent complex entries and A
e=A
nonsingular matrix of constants. If Y eXeAe∗ , then
 n
e] = det( A
[dY eAe∗ )  [dX
e ]. (3.20)

35
 e is nonsingular it can be written as a product of elementary matrices. Let E
Proof. Since A e1 , . . . , E
ek
be elementary matrices such that

e=E
A ek E
ek−1 · · · E e∗ = E
e1 =⇒ A e2∗ · · · E
e1∗ E e∗ .
k

e1 be such that the j-th row of an identity matrix is multiplied by a scalar

For example, let E
√
c = a + −1b where a, b ∈ R. Then E
e e1 X
eE e is multiplied by
e∗ means that the j-th row of X
1
√ √
a + −1b and the j-th column of Xe is multiplied by a − −1b. Let

e1 = E
U e1 X
eEe1∗ , U
e2 = E
e2 U e2∗ , . . . , U
e1E ek = E
ek U ek∗ .
e k−1 E

e written as a function X
Then the Jacobian of Y e is given by

e:X
J (Y e:U
e ) = J (Y e k−1 ) · · · J (U
e1 : X
e ).

e 1 ] in terms of [dX
Let us evaluate [dU e ] by direct computation. Since X
e is hermitian its diagonal
elements are real and the elements above the leading diagonal are the complex conjugates of
√
e 1 is also of the same structure as X.
those below the leading diagonal, and U e Let U e 1 = U + −1V
√
and Xe = Z + −1W where U = [u jk ], V = [v jk ], Z = [z jk ], W = [w jk ] are all real and the diagonal
elements of V and W are zeros. Take the u jj ’s and z jj ’s separately. The matrix of partial derivatives
of u11 , . . . , unn with respect to z11 , . . . , znn is a diagonal matrix with the j-th element a2 + b2 and
all other elements unities. That is,
 
1
 .. 
 . 
 
∂(diag(U )) ∂(u11 , . . . , unn )  
= =
 c |2
a2 + b 2 = | e  := C,

∂(diag( Z )) ∂(z11 , . . . , znn )  
 .. 
 . 
1

where diag( X ) means the diagonal matrix, obtained by keeping the diagonal entries of X and
ignoring the off-diagonal entries.
n ( n − 1) n ( n − 1)
The remaining variables produce a 2 × 2 matrix of the following type
" #
∂(U0 , V0 ) A0 B0
=
∂( Z0 , W0 ) − B0 A0

where U0 , V0 , Z0 , W0 mean that the diagonal elements are deleted, A0 is a diagonal matrix with
n − 1 of the diagonal elements equal to a and the remaining unities and B0 is a diagonal matrix
such that corresponding to every a in A0 there is a b or −b with j − 1 of them equal to −b and

36
n − j of them equal to b. Thus the Jacobian of this transformation is:
 
C 0 0
∂(diag(U ), U0 , V0 )  
e e
J (U1 : X ) =  
= det  0 A0 B0  
∂(diag( Z ), Z0 , W0 ) 
0 − B0 A0
" #!
A0 B0
= det(C ) det
− B0 A0
 √ 2
From Lemma 3.1, the determinant is  det( A0 + −1B0 )  . That is,
  " #!
∂(U0 , V0 ) A0 B0
det = det
∂( Z0 , W0 ) − B0 A0
 √ √ 
 
=  det(( A0 + −1B0 )( A0 + −1B0 )∗ ) 
= ( a2 + b 2 ) n − 1 .

Thus
 n
[ dU e ] =  det( E
e 1 ] = ( a2 + b 2 ) n [d X e1∗ )  [dX
e1 E e ].

Note that interchanges of rows and columns can produce only a change in the sign in the deter-
minant, the addition of a row (column) to another row (column) doesn not change the determinant

e1 will produce det( E e∗ )  in the Jacobian. Thus by com-
e1 E
and elementary matrices of the type E 1
e1 : X
puting J (U e ), J (U
e2 : U
e 1 ) etc we have
 n
e] = det( A
[dY eAe∗ )  [dX
e ].

As a specific example, the configuration of the partial derivatives for n = 3 with j = 2 is the
following: If
   
z11 ∗ ∗ 1 0 0
 √   
e =  z21 + −1w21 ∗  e  c 0 
X  √
z22
√  and E1 =  0 e ,
z31 + −1w31 z32 + −1w32 z33 0 0 1

then
 
z11 ∗ ∗
 √ 
e1
U =  c(z21 + −1w21 )
 e c |2 z22
|e ∗ 

√ √
z31 + −1w31 ce(z32 + −1w32 ) z33
 
u11 ∗ ∗
 √ 
= 
 u21 + √ −1v21 u22 ∗ 
,
√
u31 + −1v31 u32 + −1v32 u33

37
thus

c |2 z22 , u33 = z33 ,

u11 = z11 , u22 = | e
u21 = az21 − bw21 , u31 = z31 , u32 = az32 + bw32 ,
v21 = aw21 + bz21 , v31 = w31 , v32 = aw32 − bz32 .

Now
    
u11 1 0 0 0 0 0 0 0 0 z11
    
 u22   c |2 0 0   z22 
   0 |e 0 0 0 0  0 
    
 u33   0 0 1 0 0 0 0 0 0  z33 
    
    
 u21   0 0 0 a 0 0 −b 0 0   z21 
    
 u31   0 0 0   z31 
 = 0 0 0 0 1 0  .
    
 u32   a 0 0 b   z32 
   0 0 0 0 0  
    
 v21   0 0 0 b 0 0 a 0 0  w21 
    
 v31   0 1 0   w31 
   0 0 0 0 0 0  
v32 0 0 0 0 0 −b 0 0 a w32

Now
     
1 a −b
     
C=
 | ce|2  , A0 = 
  1  , B0 = 
  0 .

1 a b

We are done.

e is skew hermitian then the diagonal elements are purely imaginary, that is, the
Remark 3.8. If X
real parts are zeros. It is easy to note that the structure of the Jacobian matrix for a transformation
of the type Y e = A eXeA e∗ , where Xe ∗ = − X,e remains the same as that in the hermitian case of
Proposition 3.7. The roles of (u jj , z jj )’s and (v jj , w jj )’s are interchanged. Thus the next theorem
will be stated without proof.

Proposition 3.9. Let X e ∈ C n×n

e ∈ C n×n skew hermitian matrix of independent complex entries. Let A
e=A
be a nonsingular matrix of constants. If Y eXeAe∗ , then
 n
e] =  det( A
[dY eAe∗ )  [dX
e ]. (3.21)

Some simple nonlinear transformations will be considered here. These are transformations
which become linear transformations in the differentials so that the Jacobian of the original trans-
formation becomes the Jacobian of the linear transformation where the matrices of differentials
are treated as the new variables and everything else as constants.

38
 e ∈ C n×n be hermitian positive definite matrix of independent complex variables.
Proposition 3.10. Let X
e ∈ C n×n be upper triangular matrices of independent complex
e ∈ C n×n be lower triangular and Q
Let T
variables with real and positive diagonal elements. Then
!
n
e = TT e e∗ e] = 2 n 2( n − j)+1 e ],
X =⇒ [dX ∏ t jj [d T (3.22)
j=1
!
n
eQ
e=Q e ∗ =⇒ [dX
e ] = 2n 2( j−1)+1 e ].
X ∏ t jj [d Q (3.23)
j=1

Proof. When the diagonal elements of the triangular matrices are real and positive there exist
√ √
unique representations X e = T eTe∗ and Xe = QeQ
e ∗ . Let Xe = X1 + −1X2 and Te = T1 + −1T2 ,
e = [ xejk ], T
e = [et jk ], e (m) (m)
where X t jk = 0, j < k, Xm = [ x ], Tm = [t ], m = 1, 2. Note that X1 is
jk jk
symmetric and X2 is skew symmetric. The diagonal elements of X2 and T2 are zeros. Hence when
considering the Jacobian we should take xejj , j = 1, . . . , p and xejk , j > k separately.
√ √ √
e=T
X eTe∗ =⇒ X1 + −1X2 = ( T1 + −1T2 )( T1T − −1T2T )

 X1 = T1 T T + T2 T T
1 2
=⇒ ,
X = T TT − T TT
2 2 1 1 2

( 1) ( 2)
with t jj = t jj , t jj = 0, j = 1, . . . , n. Note that
        2 
( 1) ( 1) 2 ( 1) 2 ( 2) 2 ( 2)
x jj = t j1 +···+ t jj + t j1 +···+ t j,j−1

implies that
( 1)
∂x jj ( 1)
( 1)
= 2t jj = 2t jj , j = 1, . . . , n.
∂t jj

So
 
( 1) ( 1)
2t11
∂( x11 , . . . , xnn )  .. 
( 1) ( 1)
=
 .  := Z.

∂(t11 , . . . , tnn )
2tnn

( 1)
Now consider the x jk ’s for j > k. It is easy to note that
" #
∂( X10 , X20 ) U V
= ,
∂( T10 , T20 ) W Y

( 1)
where a zero indicates that the x jj ’s are removed and the derivatives are taken with respect to the
( 1) ( 2)
t jk ’s and t jk ’s for j > k. U and Y are lower triangular matrices with t jj repeated n − j times along

39
 ( 2) ( 1)
the diagonal and V is of the same form as U but with t jj = 0 along the diagonal and the t jk ’s
( 2) ( 1) ( 1) ( 1) ( 1) ( 1) ( 1)
replaced by the t jk ’s. For example, take the x jk ’s in the order x21 , x31 , . . . , xn1 , x32 , . . . , xn,n−1
( 1) n ( n − 1) n ( n − 1)
and t jk ’s also in the same order. Then we get the 2 × 2 matrix
 
t11 0 ··· 0
 
∂X10  ∗ t11 · · · 0 
 
U= = . .. .. 
∂T10  .. . . 0 
 
∗ ∗ ··· tn−1,n−1

where the ∗’s indicate the presence of elements some of which may be zeros. Since U and V are
lower triangular with the diagonal elements of V being zeros, one can make W null by adding
suitable combinations of the rows of (U, V ). This will not alter the lower triangular nature or the
diagonal elements of Y. Then the determinant is given by
" #!
U V n
2( n − j )
det = det(U ) det(Y ) = ∏ t jj .
W Y j=1

Multiply with the 2t jj ’s for j = 1, . . . , n to establish the result. As a specific example, we consider
√
the case where n = 3. Let X e ∈ C3×3 . Denote X e = X1 + −1X2 . Thus
   
( 1) ( 1) ( 1) ( 2) ( 2)
x11 x21 x31 0 − x21 − x31
 ( 1) ( 1)

( 1) 
 ( 2) 
( 2) 
X1 =  x
 21 x 22 x 32 
, X 2 =  x
 21 0 − x 32 
.
( 1) ( 1) ( 1) ( 2) ( 2)
x31 x32 x33 x31 x32 0
√
e = T1 +
Similarly, let T −1T2 . We also have:
   
( 1)
t 0 0 0 0 0
 11 ( ) ( )
  
T1 =  1 1  , T2 =  t(2) 0 0  .
 t21 t22 0   21 
( 1) ( 1) ( 1) ( 2) ( 2)
t31 t32 t33 t31 t32 0

Now X1 = T1 T1T + T2 T2T can be expanded as follows:

 
( 1) ( 1) 2 ( 1) ( 1) ( 1) ( 1) ( 1) ( 1)
x11 = , x21 = t21 t11 , x31 = t31 t11 ;
t11
  2  2  2
( 1) ( 1) ( 1) ( 2) ( 1) ( 1) ( 1) ( 1) ( 1) ( 2) ( 2)
x22 = t21 + t22 + t21 , x32 = t31 t21 + t32 t22 + t31 t21 ;
  2  2  2   2  2
( 1) ( 1) ( 1) ( 1) ( 2) ( 2)
x33 = t31 + t32 + t33 + t31 + t32 .

Then X2 = T2 T1T − T1 T2T can be expanded as follows:

( 2) ( 2) ( 1) ( 2) ( 2) ( 1) ( 2) ( 2) ( 1) ( 2) ( 1)
x21 = t21 t11 , x31 = t31 t11 , x32 = t31 t21 + t32 t22 .

40
From the above, we see that
 ( 1)
  ( 1)  ( 1)

dx11 2t 0 0 0 0 0 0 0 0 dt11
   11  ( 1) 
 dx(1)   0 ( 1)
2t22 0
( 1)
2t21 0 0
( 2)
2t21 0 0  dt22 
 22    
 ( 1)   ( 1) ( 1) ( 1) ( 2) ( 2)  ( 1) 
 dx33   0 0 2t33 0 2t31 2t32 0 2t31 2t32  dt33 
    
 ( 1)   ( 1) ( 1)  ( 1) 
 dx21   t21 0 0 t11 0 0 0 0 0  dt21 
    
 dx(1)  =  t(1) ( 1)  ( 1)
dt31 
 31   31 0 0 0 t11 0 0 0 0  .
 ( 1)   ( 1) ( 1) ( 1) ( 1) ( 2) ( 2)  ( 1) 
 dx32   0 t32 0 t31 t21 t22 t31 t21 0  dt32 
    
 ( 2)   ( 2) ( 1)  ( 2) 
 dx21   t21 0 0 0 0 0 t11 0 0  dt21 
   ( 2)  
 dx ( 2 )   t 0 0 0 0 0 0
(
t11
1 )
0  dt31 
( 2)
 31   31  
( 2) ( 2) ( 2) ( 1) ( 1) ( 2)
dx32 0 t32 0 t31 0 0 0 t21 t22 dt32

In what follows, we compute the Jacobian of this transformation:

  ( 1) 
2t11 0 0 0 0 0 0 0 0
 ( 1) ( 1) ( 2) 
 0 2t22 0 2t21 0 0 2t21 0 0  
 
 ( 1) ( 1) ( 1) ( 2) ( 2)  
 0 0 2t33 0 2t31 2t32 0 2t31 2t32 
 
  ( 1) ( 1) 
 21t 0 0 t 11 0 0 0 0 0 
  ( 1) ( 1)

e e  
J ( X : T ) = det  t31 0 0 0 t11 0 0 0 0  

 ( 1) ( 1) ( 1) ( 1) ( 2) ( 2)  
 0 t32 0 t31 t21 t22 t31 t21 0  
 
  ( 2) ( 1) 
 t21 0 0 0 0 0 t11 0 0 
  ( 2) 
 t 0 0 0 0 0 0 t
( 1)
0 
 31 11 
( 2) ( 2) ( 1) ( 1)
0 t32 0 t31 0 0 0 t21 t22
 
1 0 0 0 0 0 0 0 0
 
 0 2t(1) 0
( 1)
2t21 0 0
( 2)
2t21 0 0  
 22 
 ( 1) ( 1) ( 1) ( 2) ( 2)  
 0 0 2t33 0 2t31 2t32 0 2t31 2t32 
 
  ( 1) ( 1) 
 t21 0 0 t11 0 0 0 0 0 
( 1)   
= 2t11 det   ( 1) ( 1)
0  
 t31 0 0 0 t11 0 0 0  ,
 ( 1) ( 1) ( 1) ( 1) ( 2) ( 2) 
 0 t32 0 t31 t21 t22 t31 t21 0  
 
  ( 2) ( 1) 
 t21 0 0 0 0 0 t11 0 0 
  ( 2) ( 1)

 t 0 0 0 0 0 0 t11 0  
 31 
( 2) ( 2) ( 1) ( 1)
0 t32 0 t31 0 0 0 t21 t22

by adding the corresponding multiples of the first row to the second row through the last one,

41
respectively, we get
 
1 0 0 0 0 0 0 0 0
 ( 1) ( 1) ( 2) 
 0 2t22 0 2t21 0 0 2t21 0 0 
 
 ( 1) ( 1) ( 1) ( 2) ( 2) 
 0 0 2t33 0 2t31 2t32 0 2t31 2t32 
 
 ( 1) 
 0 0 0 t11 0 0 0 0 0 
( 1)  ( 1)

J ( X : T ) = 2t11 det 
e e

 0 0 0 0 t11 0 0 0 0  ,

 ( 1) ( 1) ( 1) ( 1) ( 2) ( 2) 
 0 t32 0 t31 t21 t22 t31 t21 0 
 
 ( 1) 
 0 0 0 0 0 0 t11 0 0 
 ( 1)

 0 0 0 0 0 0 0 t11 0 
 
( 2) ( 2) ( 1) ( 1)
0 t32 0 t31 0 0 0 t21 t22

iteratively, finally we get the final result. We also take a simple approach (i.e. by definition) with
a tedious computation as follows:

( 1) ( 1) ( 1) ( 1) ( 1) ( 1) ( 1)
dx11 = 2t11 dt11 , dx21 = t21 dt11 + t11 dt21 , dx31 = t31 dt11 + t11 dt31 ;
( 1) ( 1) ( 1) ( 2) ( 2)
dx22 = 2t22 dt22 + 2t21 dt21 + 2t21 dt21 ,
( 1) ( 1) ( 1) ( 1) ( 1) ( 1) ( 1) ( 2) ( 2) ( 2) ( 2)
dx32 = t31 dt21 + t21 dt31 + t32 dt22 + t22 dt32 + t31 dt21 + t21 dt31 ;
( 1) ( 1) ( 1) ( 1) ( 1) ( 2) ( 2) ( 2) ( 2)
dx33 = 2t31 dt31 + 2t32 dt32 + 2t33 dt33 + 2t31 dt31 + 2t32 dt32 ,

and

( 2) ( 2) ( 2) ( 2) ( 2) ( 2)
dx21 = t21 dt11 + t11 dt21 , dx31 = t31 dt11 + t11 dt31 ,
( 2) ( 2) ( 1) ( 1) ( 2) ( 2) ( 2)
dx32 = t31 dt21 + t21 dt31 + t22 dt32 + t32 dt22 .

Hence we can compute the Jacobian by definition as follows:

( 1) ( 1)
e ] = dx ∧ dx ∧ dx ∧ dx ∧ dx ∧ dx ∧ dx ∧ dx ∧ dx ( 1) ( 2) ( 2) ( 1) ( 2) ( 1) ( 1)
[d X 11 21 22 32 32 33
!31 21 31
3
2( n − j)+1 e ].
= 23 ∏ t [d T
jj
j=1

eQ
e=Q
The proof in the case of X e ∗ is similar but in this case it can be seen that the triangular
matrices corresponding to U and Y will have t jj repeated j − 1 times along the diagonal for
j = 1, . . . , n.

e∗ = X
Example 3.11. If X e ∈ C n×n is positive definite , and Re(α) > n − 1,
Z    α−n
e
Γn (α) := e ]  det X
[d X e  e
e− Tr ( X )
e >0
X
n ( n −1 )
= π 2 Γ ( α ) Γ ( α − 1) · · · Γ ( α − n + 1) . (3.24)

42
Indeed, let T e = [et jk ], e
t jk = 0, j < k be a lower triangular matrix with real and positive diagonal
elements t jj > 0, j = 1, . . . , n such that Xe=T eT
e∗ . Then from Proposition 3.10
!
n
e ] = 2n 2( n − j)+1 e]
[d X ∏ tjj [d T
j=1

Note that
    n  2  2  2
e = Tr T
Tr X e∗ =
eT ∑ t2jj + et21  + · · · + etn1  + · · · + etn,n−1 
j=1

and
!
  α−n n
 e )  e ] = 2n 2α−2j+1 e ].
det( X [d X ∏ t jj [d T
j=1

e splits into n integrals over the t jj ’s and n ( n − 1)

The integral over X integrals over the e
2 t jk ’s, j > k.
( 1) ( 2) ( 1 ) √ ( 2 )
Note that 0 < t jj < ∞, −∞ < t jk < ∞, −∞ < t jk < ∞, where et jk = t jk + −1t jk . But
Z ∞
2α−2j+1 − t2jj
2 t jj e dt jj = Γ(α − j + 1), Re(α) > j − 1,
0

for j = 1, . . . , n, so Re(α) > n − 1 and

    
Z Z ∞ Z ∞ (1) 2 (2) 2
2 − t jk + t jk
e−| t jk | det jk =
e ( 1) ( 2)
e dt jk dt jk = π.
et jk −∞ −∞

The desired result is obtained.

Definition 3.12 (e
Γn (α): complex matrix-variate gamma). It is defined as stated in Example 3.11.
We will write with a tilde over Γ to distinguish it from the matrix-variate gamma in the real case.

Example 3.13. Show that

 α  α−n
 e)   det( X
e )  ee
 det( B e− Tr ( BX )
e) =
f (X
e
Γn (α)
e∗ = B
for B e > 0, X
e∗ = Xe > 0, Re(α) > n − 1 and f ( Xe ) = 0 elsewhere, is a density function for Xe
where B e is a constant matrix, with Γ e) > 0
en (α) as given in Definition 3.12. Indeed, evidently f ( X
e and for all X
for all X e and it remains to show that the total integral is unity. Since B
e is hermitian
e such that B
positive definite there exists a nonsingular C e=C e Then
e∗ C.
   
Tr BeX
e = Tr C eX
eCe∗ .

Hence from Proposition 3.10

  n  n
e=C
Y eXeC e] = det C
e∗ =⇒ [dY e∗  [dX
eC e ] = det( B
e)  [dX
e ],

43
and
   
e ∗−1 =⇒ det( X
e = C −1 YC
X e )  = | det(CC ∗ )| −1 det(Y
e) 

Then
  α−n
  e
Z Z det( e
Y )  e− Tr (Y)
e )[dX
f (X e] = e]
[dY = 1.
e >0
X e >0
Y e
Γn (α)
e ) is known as
But from Example 3.11, the right side is unity for Re(α) > n − 1. This density f ( X
e
the complex matrix-variate density with the parameters α and B.

3.2 The computation of volumes

Definition 3.14 (Semiuniatry and unitary matrices). A p × n matrix Ue is said to be semiunitary

eU
if U e ∗ = 1 p for p < n or U e = 1 n for p > n. When n = p and U
e ∗U eUe ∗ = 1 n , then U
e is called a
unitary matrix. The set of all n × n unitary matrices is denoted by U (n). That is,
n o
U (n) := Ue ∈ C n×n : U
eUe ∗ = 1n , (3.25)

where C n×n denotes the set of all n × n complex matrices.

Definition 3.15 (A hermitian or a skew hermitian matrix). Let A e ∈ C n×n . If A

e= A
e∗ , then A
e is
e∗ = − A,
said to be hermitian and if A e then it is skew hermitian.

When dealing with unitary matrices a basic property to be noted is the following:

eU
U e ∗ = 1 =⇒ U
e ∗ dU e ∗ U.
e = − dU e
 ∗
But U e ∗ dU
e = dU e which means that U
e ∗ U, e ∗ dU
e is a skew hermitian matrix. The wedge product
 
e ∗ dU,
of U e namely, ∧ U e ∗ dU
e enters into the picture when evaluating the Jacobians involving
unitary transformations. Hence this will be denoted by dG e for convenience. Starting from
e ∗U
U e = 1 n one has dU
e ·U e ∗.
  √
Assume that U e = [ueij ] ∈ U (n) where ueij ∈ C. Let uejj =  uejj  e −1θ j by Euler’s formula, where
θ j ∈ [−π, π ]. Then
 √ √  √ 
| ue11 | ue12 e− −1θ2 ··· ue1n e− −1θn e −1θ1 0 ··· 0
 √ √  √ 
 ue21 e− −1θ1 | ue22 | ··· ue2n e− −1θn  0 e −1θ2 ··· 0 
e=   
U .. .. .. ..  .. .. .. ..  . (3.26)
 . . . .  . . . . 
 √ √  √ 
uen1 e− −1θ1 uen2 e− −1θ2 ··· | uenn | 0 0 ··· e −1θn

e ∈ U (n) can be factorized into a product of a unitary matrix with

This indicates that any U
diagonal entries being nonnegative and a diagonal unitary matrix. It is easily seen that such

44
factorization of a given unitary matrix is unique. In fact, we have a correspondence which is
one-to-one:

U (n) ∼ (U (n)/U (1)×n ) × U (1)×n .

Notation. When Ue is a n × n unitary matrix of independent complex entries, Ue ∗ its conjugate

transpose and dU e : = dU
e the matrix of differentials then the wedge product in dG e ·Ue ∗ will be
e]. That is, ignoring the sign,
denoted by [dG
   
e ] : = ∧ dU
[d G e ·U
e∗ = ∧ Ue · dU
e∗ .

If the diagonal entries or the entries in one row of this unitary matrix Ue are assumed to be real,
then the skew hermitian matrix dU e ·U e1 and its wedge product by
e ∗ will be denoted by dG
 
e1 ] := ∧ dU
[d G e∗ .
e ·U

e] here means the wedge product over U (n), but however [dG
Indeed, [dG e1 ] means the wedge
e ] = [d G
product over U (n)/U (1)×n . Therefore, for any measurable function f over U (n), [dG e1 ][d D
e ],
Z Z Z
f (U e] =
e )[dG eD
f (V e1 ][d D
e )[dG e ],
U (n) U1 ( n ) U ( 1) × n

where U e = V eDe for V e = U

e ∈ U1 ( n ) , d G e ∗ dU
e and dG e1 = V e ∗ dV.
e Furthermore, let Ve = [veij ] for
veij ∈ C and vejj = v jj ∈ R + . Then it holds still that V e ∈ U (n), thus v jj is not an independent
q
variable, for example, v11 = 1 − | ve21 |2 − · · · − | ven1 |2 . From this, we see that
   
[d Ge1 ] = ∏ d Re(V e ∗ dV
e )ij d Im(V e ∗ dV
e )ij
i< j

and !
n
e] =
[d G ∏ Im(Ue ∗ dUe ) jj e ∗ dU
× ∏ Re(U e ∗ dU
e )ij Im(U e )ij .
j=1 i< j

e ∈ C n×n be lower triangular and U

Proposition 3.16. Let T e ∈ U (n) be of independent complex variables.
e=T
Let X eU.
e Then:

e being real and positive,

(i) for all the diagonal entries t jj , j = 1, . . . , n of T
 
e ] = ∏nj=1 t2(n− j)+1 [dT
[d X e]
e][dG (3.27)
jj

e = dU
where dG e ·U
e ∗ ; and

e being real,
(ii) for all the diagonal entries in U
!
n
  2( n − j )
e] =
[d X ∏ etjj  e1 ]
e][dG
· [d T (3.28)
j=1

e1 = dU
where dG e ·U
e ∗.

45
 e=T
Proof. Taking differentials in X eUe one has

e = dT
dX e·U
e+T
e · dU.
e

e ∗ and observing that U

Postmultiplying by U eUe ∗ = 1 n we have

dX e ∗ = dT
e·U e+T
e · dU e ∗.
e ·U (3.29)

e be real and positive and all other elements in T

(i). Let the diagonal elements in T e and U
e be
complex. Let

e = dX
dV e ∗ =⇒ [dV
e·U e ] = [d X
e] (3.30)

ignoring the sign, since U e = dU

e is unitary. Let dG e ·U e]. Then
e ∗ and its wedge product be [dG

e = dT
dV e+T e
e · dG (3.31)

e is skew hermitian. Write

where dG

( 1) √ ( 2)
e = [vejk ],
V vejk = v jk + −1v jk ,
√
e = [et jk ], et jk = t(1) + −1t(2) , j > k, t(1) = t jj > 0, t(2) = 0,
T jk jk jj jj
( 1) √ ( 2 ) ( 1 ) ( 1 ) ( 2) ( 2)
e = [de
dG gjk ], de gjk = dgjk + −1dgjk , dgjk = −dgkj , dgjk = dgkj .

From Eq. (3.31),



det + e g1k + · · · + e
t j1 de t jj de
gjk , j>k
jk
v jk =
de

 et deg + · · · + et de g , j < k.
j1 1k jj jk

(m) (m) (m) (m) (m) (m)

The new variables are dv jk = vbjk , dt jk = bt jk , dgjk = gbjk , j > k, m = 1, 2. The matrices of
partial derivatives are easily seen to be the following:
   
( 1) (m)
v jj
∂b v jk
∂b
  = 1,  , j > k = 1, m = 1, 2,
∂bt jj (m)
∂bt jk
   
( 1) ( 2)
vkj
∂b ∂bvkj
 , j > k = A,  (2) , j > k = B
( 1)
gkj
∂b ∂bgkj

where A and B are triangular matrices with t jj repeated n − j times,

 
( 2)
v jj
∂b
  = diag(t11 , . . . , tnn ).
( 2)
gjj
∂b

46
By using the above identity matrices one can wipe out other submatrices in the same rows and
columns and using the triangular blocks one can wipe out other blocks below it when evaluating
the determinant of the Jacobian matrix and finally the determinant in absolute value reduces to
the form
n
2( n − j)+1
det( A) det( B)t11 · · · tnn = ∏ tjj .
j=1

Hence the result. As a specific example, we consider the case where n = 3. We expand the
e = dT
expression: dV e+T e That is,
e · dG.
   
e) − T2 d Im( G
dV1 = dT1 + T1 d Re( G e) ,
   
e) + T1 d Im( G
dV2 = dT2 + T2 d Re( G e) .

Furthermore,
 
( 1) ( 1) ( 1)
dv11 dv12 dv13
 
 dv(1) dv(1) dv(1) 
 21 22 23 
( 1) ( 1) ( 1)
dv31 dv32 dv33
    
( 1) ( 1) ( 1) ( 1)
dt11 0 0 t11 0 0 0 dg12 dg13
 ( 1) ( 1)
  ( 1) ( 1)
 

=  dt21 dt22  
0  +  t21 t22 0   −dg(1) 0 dg
( 1) 
 12 23 
( 1) ( 1) ( 1) ( 1) ( 1) ( 1) ( 1) ( 1)
dt31 dt32 dt33 t31 t32 t33 −dg13 −dg23 0
  
( 2) ( 2) ( 2)
0 0 0 dg11 dg12 dg13
 ( 2)  ( 2) ( 2) ( 2)

−  t21 0 0  
  dg12 dg22 dg23  
( 2) ( 2) ( 2) ( 2) ( 2)
t31 t32 0 dg13 dg23 dg33

and
 
( 2) ( 2) ( 2)
dv11 dv12 dv13
 
 dv(2) dv(2) dv(2) 
 21 22 23 
( 2) ( 2) ( 2)
dv31 dv32 dv33
    
( 1) ( 1)
0 0 0 0 0 0 0 dg12 dg13
 ( 2)   ( 2)  ( 1) ( 1)

= dt21 0 0  
 +  t21 0 0  
  −dg12 0 dg23 

( 2) ( 2) ( 2) ( 2) ( 1) ( 1)
dt31 dt32 0 t31 t32 0 −dg13 −dg23 0
  
( 1) ( 2) ( 2) ( 2)
t 0 0 dg11 dg12 dg13
 11 ( 1) ( 1)
  
+ t t 0   dg(2) dg(2) dg(2)  .
 21 22  12 22 23 
( 1) ( 1) ( 1) ( 2) ( 2) ( 2)
t31 t32 t33 dg13 dg23 dg33

47
Thus

( 1) ( 1) ( 1) ( 1) ( 1)
dv11 = dt11 , dv12 = t11 dg12 , dv13 = t11 dg13 ,
( 1) ( 1) ( 1) ( 2) ( 2) ( 1) ( 1) ( 1) ( 2) ( 2)
dv21 = dt21 − t22 dg12 − t21 dg11 , dv22 = dt22 + t21 dg12 − t21 dg12 ,
( 1) ( 1) ( 1) ( 1) ( 2) ( 2)
dv23 = t21 dg13 + t22 dg23 − t21 dg13 ,
( 1) ( 1) ( 1) ( 1) ( 1) ( 2) ( 2) ( 2) ( 2)
dv31 = dt31 − t32 dg12 − t33 dg13 − t31 dg11 − t32 dg12 ,
( 1) ( 1) ( 1) ( 1) ( 1) ( 2) ( 2) ( 2) ( 2)
dv32 = dt32 + t31 dg12 − t33 dg23 − t31 dg12 − t32 dg22 ,
( 1) ( 1) ( 1) ( 1) ( 1) ( 2) ( 2) ( 2) ( 2)
dv33 = dt33 + t31 dg13 + t32 dg23 − t31 dg13 − t32 dg23

and

( 2) ( 2) ( 2) ( 2) ( 2) ( 2)
dv11 = t11 dg11 , dv12 = t11 dg12 , dv13 = t11 dg13 ,
( 2) ( 2) ( 1) ( 2) ( 2) ( 2) ( 1)
dv21 = dt21 + t21 dg11 + t22 dg12 − t21 dg12 ,
( 2) ( 2) ( 1) ( 1) ( 2) ( 2)
dv22 = t21 dg12 + t21 dg12 + t22 dg22 ,
( 2) ( 2) ( 1) ( 1) ( 2) ( 2)
dv23 = t21 dg13 + t21 dg13 + t22 dg23 ,
( 2) ( 2) ( 2) ( 1) ( 1) ( 2) ( 1) ( 2) ( 2)
dv31 = dt31 − t32 dg12 + t31 dg11 + t32 dg12 + t33 dg13 ,
( 2) ( 2) ( 2) ( 1) ( 1) ( 2) ( 1) ( 2) ( 2)
dv32 = dt32 + t31 dg12 + t31 dg12 + t32 dg22 + t33 dg23 ,
( 2) ( 2) ( 1) ( 2) ( 1) ( 1) ( 2) ( 1) ( 2) ( 2)
dv33 = t31 dg13 + t32 dg23 + t31 dg13 + t32 dg23 + t33 dg33 .
( 1) ( 2)
v jk = dv jk ∧ dv jk , then
According to the definition, we now have de

e ] = de
[ dV v11 ∧ dev12 ∧ de v13 ∧ de
v21 ∧ de
v22 ∧ de
v23 ∧ de
v31 ∧ de
v32 ∧ de
v33
!
3
2(3− j)+1 e ].
e][dG
= ∏t [d Tjj
j=1

e be real and all other elements in U

(ii). Let the diagonal elements of U e and T
e complex. Starting
from Eq. (3.31), observing that dGe is dG
e1 in this case with the wedge product [dG
e1 ], and taking
( 1) ( 1) ( 2) ( 2)
v jk ’s in the order dv jk , j > k, dv jk , j < k, dv jk , j > k, dv jk , j < k and the other
the variables de
( 1) ( 1) ( 2) ( 2)
variables in the order dt jk , j > k, dgjk , j > k, dt jk , j > k, dgjk , j > k, we have the following
configuration in the Jacobian matrix:
 
1 ∗ ∗ ∗
 
 0 −A 0 A 
 1 2 
 
 0 ∗ 1 ∗ 
 
0 − A2 0 − A1

where the matrices marked by ∗ can be made null by operating with the first and third column
submatrices when taking the determinant. Thus they can be taken as null matrices, and A1 and

48
 ( 1) ( 2)
A2 are triangular matrices with respectively t jj and t jj repeated n − j times in the diagonal. The
Jacobian matrix can be reduced to the form
" # " # " #
A B 1 0 0 0
,A= ,B = .
−B A 0 − A1 0 A2

Then the determinant is given by

 
 A B   √ √ 
   
  =  det(( A + −1B)( A + −1B)∗ ) 
 −B A 
 √ √ 
 
=  det((− A1 + −1A2 )(− A1 + −1A2 )∗ ) 
p   2( n − j )
= ∏ etjj 
j=1

√ ( 1) √ ( 2)
since − A1 + −1A2 is triangular with the diagonal elements −t jj + −1t jj repeated n − j
 2  2  2
( 1) ( 2)
times, giving t jj + t jj = et jj  repeated n − j times in the final determinant and hence
the result.

By using Proposition 3.16 one can obtain expressions for the integral over U e of [dGe] and
e1 ]. These will be stated as corollaries here and the proofs will be given after stating both the
[d G
corollaries.

e = dU
Theorem 3.17. Let dG e ·U
e ∗ , where U
e ∈ U (n). Then

Z 2 n ( n +1 )

e] = 2n π n 2n π 2
vol (U (n)) = [d G = . (3.32)
U (n) e
Γn (n) 1!2! · · · (n − 1)!

e be a n × n matrix of independent complex variables. Let

Proof. Let X
Z Z 2
B= e ]e− Tr ( Xe Xe ∗ ) =
[d X e ]e− ∑ j,k | xejk | = π n
[d X
2

e
X e
X

since
    
Z Z +∞ Z +∞ (1) 2 (2) 2
2 − x jk + x jk
e−| xejk | de
( 1) ( 2)
x jk = e dx jk dx jk = π.
xejk −∞ −∞

Consider the transformation used in Proposition 3.16 with t jj ’s real and positive. Then

  
e e ∗ e e ∗
Tr X X = Tr T T

49
and let
Z Z Z
!
n 2
eX
− Tr ( X e∗)
e− ∑ j>k | t jk | [dT
e ]e 2( n − j)+1 e e ].
e][dG
B=
e
X
[d X =
e
T e
U
∏ t jj
j=1

But
Z ∞
2( n − j)+1 − t2jj 1
t jj e dt jj = Γ(n − j + 1) for n − j + 1 > 0
0 2
and for j > k
Z 2
e−| t jk | det jk = π.
e
et jk

e gives
Then the integral over T
n ( n −1 )
n
2− n π 2
∏ Γ(n − j + 1) = 2−n Γen (n).
j=1

Hence
Z n n 2
e] = 2 π .
[d G
U (n) e
Γn (n)

Theorem 3.18. Let U e1 = dU

e 1 ∈ U (n) with the diagonal elements real. Let dG e1 · U
e ∗ . Let the full unitary
1
group of such n × n matrices Ue 1 be denoted by U1 (n) = U (n)/U (1)×n . Then

Z n ( n − 1)
n ( n −1 )

vol (U1 (n)) = vol U (n)/U (1) ×n

= e1 ] = π
[d G =
π 2
. (3.33)
U1 ( n ) e
Γn (n) 1!2! · · · (n − 1)!
e com-
Proof. Now consider the transformation used in Proposition 3.16 with all the elements in T
e real. Then
plex and the diagonal elements of U
Z Z
!
n 2
2( n − j ) − ∑ j>k |et jk | e1 ].
e][dG
B= ∏ t
e
T
e
e1
U
[d T jj
j=1

Note that
Z
!
2 n ( n −1 )
− |et jk |
∏ e det jk =π 2 .
et jk
j> k
√
Let et jj = et = t1 + −1t2 . Put t1 = r cos θ and t2 = r sin θ. Let the integral over et jj be denoted by
a j . Then
Z   Z +∞ Z +∞
e 2( n − j) − | et |2 2 2
aj = t e det = (t21 + t22 )n− j e−(t1 +t2 ) dt1 dt2
et −∞ −∞
Z +∞ Z +∞
2 2
= 4 (t21 + t22 )n− j e−(t1 +t2 ) dt1 dt2
0 0
Z π Z ∞
2 2
= 4 (r2 )n− j · e−r · r · drdθ
θ =0 r =0
= πΓ(n − j + 1) for n − j + 1 > 0.

50
Then
Z
en ( n )
B = πnΓ e1 ] = π n2
[d G
e1
U

implies that
Z n ( n − 1)
e1 ] = π
[d G ,
e1
U e
Γn (n)
which establishes the result.

Example 3.19. Evaluate the integral

Z  α e
∆( α) = e ] det( X
[d X eXe ∗ )  · e− Tr ( Xe Xe ∗ ) = π n2 Γn (α + n) (3.34)
e
X e
Γn (n)

for Re(α) > −1, where X e ∈ C n×n matrix of independent complex variables. Indeed, put Xe=T eU,
e
where Ue ∈ U ( n ), T
e is lower triangular with real distinct and positive diagonal elements. Then
 α n
 eXe ∗ )  =
 det( X ∏ t2α
jj ,
j=1
     2 n  2
eX
Tr X e∗ = Tr T e∗ =
eT ∑ etjk  = ∑ t2jj + ∑ etjk  ,
j> k j=1 j> k
!
n
e] = 2( n − j)+1 e ],
e][dG
[d X ∏ t jj [d T
j=1
Z 2
e] = 2n π n
[d G .
e
U e
Γn (n)
Thus
Z
!
n 2
2α+2( n − j)+1 − t2jj
e− ∑ j>k | t jk |
e]
e][dG e
∆( α) =
eU
T, e
[d T ∏t e
j=1
Z ∞
! Z +∞
! Z
n 2
2α+2( n − j)+1 − t2jj − |et jk | e]
= ∏ t e dt jj · ∏ e det jk · [dG
−∞ e
U
j=1 0 j> k
! 2
n
−n n ( n −1 ) 2n π n
= 2 ∏ Γ ( α + n − j + 1) ·π 2 ·
en (n)
Γ
j=1

for Re(α + n) > n − 1 or Re(α) > −1. That is,

2 e
Γn (α + n)
∆( α) = π n for Re(α) > −1.
en ( pn)
Γ

51
 e ∈ C n×n be a hermitian matrix of independent complex entries with real distinct
Proposition 3.20. Let X
e ∈ U (n) with real diagonal entries and let X
eigenvalues λ1 > λ2 > · · · > λn . Let U e = UD
e U e ∗ , where
D = diag(λ1 , . . . , λn ). Then
   
e ] = ∏ j>k  λk − λ j 2 · [dD ][dG
[d X e1 ], (3.35)

e1 = U
where dG e ∗ · dU.
e

e = UD
Proof. Take the differentials in X e Ue ∗ to get

e = dU
dX e∗ + U
e ·D·U e∗ + U
e · dD · U e ∗.
e · D · dU

e ∗ , postmultiply by U
Premultiply by U e1 is skew hermitian. Then one has
e and observe that dG

e = dG
dW e1 · D + dD − D · dG
e1

e = U
where dW e ∗ · dX
eU e ] = [d X
e with [dW e ]. Using the same steps as in the proof of Proposi-
tion 7.10, we have
!
 2
e] =
[dW ∏  λk − λ j  e1 ].
[dD ][dG
j> k

Hence the result follows.

Example 3.21. Let D = diag(λ1 , . . . , λn ) where the λ j ’s are real distinct and positive or let λ1 >
· · · > λn > 0. Show that

(i)
!  2 !2
Z  2 e
Γn (n) n

λ1 >···> λn >0
[dD ] ∏  λk − λ j  e − Tr ( D )
=
π n ( n − 1)
= ∏ Γ ( j) ;
j> k j=1

(ii)
Z
! !
 2 n
 α−n en ( α) e
Γ Γn (n)
[dD ] ∏  λk − λ j  · ∏  λj  · e− Tr (D ) = .
λ1 >···> λn >0 j> k j=1 π − 1)
n ( n

e be a n × n hermitian positive definite matrix, U

In fact, let Y e a unitary matrix with real diagonal
elements such that

e ∗Y
U eUe = D = diag(λ1 , . . . , λn ).

From the matrix-variate gamma integral

Z   α−n
e
Γn (α) = [dY e) 
e]  det(Y e
· e− Tr (Y) for Re(α) > n − 1.
e =Y
Y e ∗ >0

52
Hence
Z
e
Γn (n) = e]e− Tr (Ye) .
[dY
e =Y
Y e ∗ >0

e = UD
Then Y e Ue ∗ implies that
!
 2
e] =
[dY ∏  λk − λ j  e1 ],
[dD ][dG e ∗ · dU.
e1 = U
dG e
j> k

Note that
   
Tr Y e e U
= Tr UD e ∗ = Tr ( D ) ,
!α−n
  α−n n  
 e) 
det(Y = ∏  λj  .
j=1

Then
Z
! Z
 2
e
Γn (n) = [dD ] ∏  λk − λ j  e− Tr (D ) × e1 ].
[d G
λ1 >···> λn >0 e
U
j> k

Clearly
Z n ( n − 1)
e1 ] = π
[d G .
e
U e
Γn (n)
Substituting this, results (i) and (ii) follow.

Remark 3.22. We can try to compute the following integral in (3.35):

Z Z
! Z
n n
e
[d X ] = 2
δ 1 − ∑ λ j ∏(λi − λ j ) ∏ dλ j × e1 ]
[d G
e (X
X:Tr e ) =1 λ1 > λ2 >···> λn >0 U ( n )
j=1 i< j j=1 1

which is equivalent to the following

Z Z
! Z
n n
e] = 1 e1 ]
e (X
X:Tr e ) =1
[d X
n!
δ 1 − ∑ λj ∏(λi − λ j )2 ∏ dλ j × U1 ( n )
[d G
j=1 i< j j=1
n −1 n ( n −1 )
1 ∏ j = 0 Γ ( n − j ) Γ ( n − j + 1) π 2
= × n
n! Γ ( n2 ) ∏ j=1 Γ ( j )
1 n ( n −1 ) Γ ( 1) Γ ( 2) · · · Γ ( n + 1)
= π 2 ,
n! Γ ( n2 )
where we used the integral formula:
Z
!
n n ∏nj=−01 Γ(n − j)Γ(n − j + 1)
δ 1 − ∑ λj ∏(λi − λ j )2 ∏ dλ j = Γ ( n2 )
.
j=1 i< j j=1

This means that

53
 n ( n −1 ) Γ ( 1) Γ ( 2) · · · Γ ( n )
vol (D (C n )) = π 2 . (3.36)
Γ ( n2 )

We make some remarks here: to obtain the volume of the set of mixed states acting on C n ,
e ]. By definition, the first integral gives 1 1HS , where
one has to integrate the volume element [dX n! Cn
(1,2)
CnHS = Cn and
n
zZ }| Z { !
∞ ∞ n n n
1
( α,β)
= ··· δ 1 − ∑ λ j ∏ λαk −1 ∏(λi − λ j ) β ∏ dλ j
Cn 0 0 j=1 k i< j j=1
   
β β
1 n Γ 1+ j2 Γ α + ( j − 1) 2
= 
n ( n − 1)
 ∏ 
β
 ,
Γ αn + β 2 j=1 Γ 1+ 2

while the second is equal to the volume of the flag manifold. To make the diagonalization
unique, one has to restrict to a certain order of eigenvalues, say λ1 > λ2 > · · · > λn (a generic
density matrix is not degenerate), which corresponds to a choice of a certain Weyl chamber of the
eigenvalue simplex ∆n−1 . In other words, different permutations of the vector of n generically
1
different permutations (Weyl chambers) equals to n!. This is why the factor n! appears in the
right hand side in the above identity. In summary, the transformation

e 7→ ( D, U
X e)

e = UD
such that X e U e ∗ , is one-to-one if and only if D = diag(λ1 , . . . , λn ), where λ1 > · · · > λn
e ∈ U ( n ) / U ( 1) × n .
and U

Remark 3.23. In this remark, we will discuss the connection between two integrals [14]: for
α, β > 0,
n
zZ }| Z { !
∞ ∞ n n n
( 1)
In (α, β) = ··· δ 1 − ∑ λ j ∏ λαk −1 ∏ (λi − λ j ) β ∏ dλ j ,
0 0 j=1 k=1 16i< j6n j=1
n
z }| Z {
Z
!
∞ ∞ n n n
( w)
In (α, β) =
0
···
0
exp − ∑ x j ∏ xkα−1 ∏ ( xi − x j ) β ∏ dx j .
j=1 k=1 16i< j6n j=1

( 1)
We introduce an auxiliary variable t in the expression of In (α, β), and define I (t) as
n
z }| Z {
Z
!
∞ ∞ n n n
I ( t) =
0
···
0
δ t − ∑ λj ∏ λαk −1 ∏ (λi − λ j ) β ∏ dλ j . (3.37)
j=1 k=1 16i< j6n j=1

54
 ( 1)
Then I (1) = In (α, β). Taking the Laplace transform, denoted by L , of I (t), we obtain
n
zZ }| Z { "Z ! #
∞ ∞ n n n
− st
L ( I ) :=
0
···
0
δ t − ∑ λj e dt ∏ λαk −1 ∏ (λi − λ j ) β ∏ dλ j (3.38)
j=1 k=1 16i< j6n j=1
n
z }| Z {
Z
!
∞ ∞ n n n
=
0
···
0
exp −s ∑ λ j ∏ λαk −1 ∏ (λi − λ j ) β ∏ dλ j (3.39)
j=1 k=1 16i< j6n j=1
n
z }| Z {
Z
!
∞ ∞ n n n
− αn− β(n2 )
= s
0
···
0
exp − ∑ x j ∏ xkα−1 ∏ ( xi − x j ) β ∏ dx j , (3.40)
j=1 k=1 16i< j6n j=1

which means that

n
I (s) := L ( I ) = s−αn− β(2 ) · I (w) (α, β).
e (3.41)

Therefore
n
−1 −1 tαn+ β(2 )−1
I ( t) = L (L ( I )) = L (e
I) =
· I (w) (α, β). (3.42)
Γ αn + β( n2 )
Letting t = 1 gives the conclusion:
( 1) 1
In (α, β) =
· I (w) (α, β). (3.43)
Γ αn + β( n2 )
( 1)
In order to calculate the integral In (α, β), it suffices to calculate the integral I (w) (α, β), which is
derived in Corollary 9.4 via Selberg’s integral (See Appendix):
   
β β
n Γ 1+j Γ α + ( j − 1 )
2 2
I (w) (α, β) = ∏   . (3.44)
β
j=1 Γ 1+ 2

Matrix integrals, especially over unitary groups, are very important. We recently present
some results of this aspect [32]. Generally speaking, the computation of matrix integrals is very
difficult from the first principle. Thus frequently we need to perform variable substitution in
computing integrals. The first step in substitution is to compute Jacobians of this transformation,
this is what we present. We also apply the matrix integrals over unitary groups to a problem
[33, 34] in quantum information theory.

4 The volume of a compact Lie group

The content of this section is mainly from [22]. Some missing details are provided. Note that
Macdonald’s result presents unifying treatment for computing the volume of a compact Lie
group. Before proceeding, it is necessary to recall the notion of root system and its related
properties.

55
Definition 4.1. A root system ( E, R) is a finite-dimensional real vector space E with an inner
product h·, ·i, together with a finite collection R of nonzero vectors in E satisfying the following
properties:

(i) span{ R} = E.

(ii) If α ∈ R, then Lα ∩ R = {±α}, where Lα = {rα : r ∈ R }.

(iii) If α, β ∈ R, then sα ( β) ∈ R, where sα is the linear transformation of E defined by

h x, αi
sα ( x ) = β − 2 α, x ∈ E.
hα, αi

h β,αi
(iv) For all α, β ∈ R, 2 hα,αi ∈ Z.

The dimension of E is called the rank of the root system and the elements of R are called roots.

Definition 4.2. If ( E, R) is a root system, the Weyl group W of R is the subgroup of the orthogonal
group of E generated by the reflections sα (α ∈ R). That is, for arbitrary positive integer k ∈ N
and any non-negative integer n1 , . . . , nk , snα11 · · · snαkk ∈ W, i.e.,
n o
W = snα11 · · · snαkk : sα j ∈ R for j = 1, . . . , k ∈ N .

Definition 4.3. If ( E, R) is a root system, then for each root α ∈ R, the co-root α∨ is the vector
given by
α
α∨ := 2 .
hα, αi
The set of all co-roots is denoted R∨ and is called the dual root system to R.

It is easily seen that h β, α∨ i = 2 hh β,βii ∈ Z for α, β ∈ R. Besides, we see that

β,α

 4
α∨ , α∨ = .
hα, αi

In fact, if R is a rooty system, then R∨ is also a root system and the Weyl group for R∨ is the
same as the Weyl group for R. Furthermore, R∨∨ = R. Thus

α∨
α=2 .
hα∨ , α∨ i

Definition 4.4. If ( E, R) is a root system, a subset ∆ of R is called a base if the following conditions
hold:

(i) ∆ is a basis for E as a vector space.

56
 (ii) Each root α ∈ R can be expressed as a linear combination of elements of ∆ with integer
coefficients and in such a way that the coefficients are either all non-negative or all non-
positive.

The roots for which the coefficients are non-negative are called positive roots and the others are
called negative roots (relative to the base ∆). The set of positive roots relative to a fixed base ∆ is
denoted by R+ and the set of negative roots is denote by R− . Thus R = R+ ⊔ R− . The elements
of ∆ are called the positive simple roots.

Note that if ∆ is a base for R, then the set of all co-roots α∨ (α ∈ ∆) is a base for the dual root
system R∨ . We also see that if ∆ is a base, then W is generated by the reflections sα with α ∈ ∆.

Definition 4.5. An element µ of E is an integral element if for all α ∈ R, hµ, α∨ i ∈ Z. If ∆ is a base

for R, an element µ of E is dominant (relative to ∆) if hµ, αi > 0 for all α ∈ ∆ and strictly dominant
if hµ, αi > 0 for all α ∈ ∆.

Definition 4.6. Let ∆ = {α1 , . . . , αn } be a base. Then the fundamental weights (relative to ∆) are
the elements µ1 , . . . , µn with the property that
D E
µi , α∨j = δij , i, j = 1, . . . , n.

That is, {µ1 , . . . , µn } can be viewed as the dual base of ∆∨ := {α1∨ , . . . , α∨

n }.

Definition 4.7. Let ∆ = {α1 , . . . , αn } be a base for R and R+ the associated set of positive roots.
We then let ρ denote half the sum of the positive roots and let σ denote half the sum of the positive
co-roots:
1 1
α∨ .
2 α∈∑ 2 α∈∑
ρ= α, σ=
R+ R+

They are often called the Weyl vectors.

One important result for Weyl vector is that ρ are strictly dominant integral element; indeed,


ρ, α∨ = 1, ∀ α∨ ∈ ∆∨ . (4.1)

Similarly,

hσ, αi = 1, ∀α ∈ ∆. (4.2)

It is well-known that the Weyl vector can be expressed as the sum of fundamental weights:
n
ρ= ∑ µj . (4.3)
j=1

57
Definition 4.8. Let ∆ = {α1 , . . . , αn } be a base for R. For a positive root α ∈ R+ , the height of α is
defined by
n
ht(α) = ∑ k j ∈ Z+ ,
j=1

where α = ∑nj=1 k j α j for non-negative integers k1 , . . . , kn . That is,

!
n n
ht ∑ kj αj = ∑ k j ∈ Z+ .
j=1 j=1

Apparently, ht(α) = 1 if and only if α ∈ ∆.

With these notations, we can show that hρ, α∨ i = ht(α∨ ) for any α ∈ R+ . Note that the base
∆ = {α1 , . . . , αn } for the root system R is given, then ∆∨ = {α1∨ , . . . , α∨
n } is known for the dual
root system R∨ . Now for any α ∈ R+ , α∨ is also positive co-root and it is can be written as
n
α∨ = ∑ k j α∨j , k j ∈ N, j = 1, . . . , n.
j=1

By using (4), we see that

* +

 n n n D E n n
ρ, α∨ = ∑ µi , ∑ k j α∨j = ∑ k j µi , α∨j = ∑ k j δij = ∑ k j = ht(α∨ ).
i=1 j=1 i,j=1 i,j=1 j=1

Similarly, hσ, αi = ht(α) for α ∈ R+ .

We also have to recall the following polynomial function P : E → R given by

P( X ) = ∏ hα, X i .
α∈ R+

Note that P has an important property:

P(w · X ) = det(w) P( X ), ∀w ∈ W, ∀ X ∈ E.

Indeed, for any w ∈ W, we have

D E
P(w · X ) = ∏ hα, w · X i = ∏ w−1 · α, X .
α∈ R+ α∈ R+

Suppose first that w = w−1 = s β , where β is a positive simple root. Since s β permutes the positive
roots different from β, whereas s β · β = − β. Thus

P(s β · X ) = ∏ hα, X i × h− β, X i = − ∏ hα, X i .

α∈ R+ \{ β } α∈ R+

58
This because the determinant of a reflection is −1. Note W is generated by all positive simple
roots α. Thus for w ∈ W, it can be written as the product of some positive simple roots: w =
sα1 · · · sαk . Then

P ( w · X ) = P ( s α 1 · · · s α k · X ) = − P ( s α 1 · · · s α k −1 · X )
= ···
= (−1)k P( X ) = det(sα1 · · · sαk ) P( X )
= det(w) P( X ).

This means that P is alternating, or skew-symmetric.

Next we can start our focus, i.e., computing the volume of a compact Lie group. Let G be a
compact Lie group and let g be its Lie algebra, thought of as the tangent space Te G to G at the
identity element e. Choose a Lebesgue measure µ L on the vector space g. By means of a chart
of G at e, we can construct from µ L a translation-invariant measure on a neighborhood of e in
G, and then we can extend this by translation in G to a Haar measure µHaar on G. The purpose
of this section is to establish a formula (Eq. (4.4) below), obtained by Macdonald in 1980, for
µHaar ( G ), the volume of G relative to the measure µHaar , as a function of µ L .
There are two ingredients in the formula. Firstly, from a suitably chosen Chevalley basis (i.e.
root vectors) of the complexification of g, we can construct an "integer lattice" gZ , which is a
lattice in g and a Lie algebra over Z. By abuse of notation, let µ L (g/gZ ) denote the volume
(with respect to µ L ) of a fundamental parallelepiped for gZ in g. Secondly, it is well-known that
the manifold G has the same cohomology, apart from torsion, as a product of odd-dimensional
spheres, say of dimensions r1 , . . . , rn . We shall prove the following result:

Theorem 4.9 ([22]). It holds that

n
µHaar ( G ) = µ L (g/gZ ) ∏ vol (Sr j ) (4.4)
j=1

where vol (Sr j ) is the superficial measure of the unit sphere Sr j in Rr j +1 , that is (since r j = 2m j + 1 is

m j +1
odd) vol (Sr j ) = vol S2m j +1 = Γ2π ( m + 1)
. By abuse of notation, we record the above fact as
j

n
vol( G ) = vol(g/gZ ) ∏ vol (Sr j ) .
j=1

Proof. Let T be a maximal torus in G and let t ⊂ g be its Lie algebra. Let d = dim( G ), n = dim( T ).
The Lebesgue measure µ L on g determines a Lebesgue measure (also denoted by µ L ) and hence
a Haar measure (also denoted by µHaar ) on T 1 . Let tZ be the lattice in t such that the kernel of
1 Note: an inner product on g does determine Lebesgue measures on both g and t.

59
exp : t → T is ker exp = 2πtZ , then clearly

µHaar ( T ) = µ L (t/2πtZ ) = (2π )n µ L (t/tZ ), (4.5)

vol( T ) = vol(t/2πtZ ) = (2π )n vol(t/tZ ). (4.6)

Let R be the set of roots of G relative to T, and let W be the Weyl group. The roots are real linear
forms on t, integer-valued on the lattice tZ . Fix a system of positive roots, and let P = ∏α>0 α, a
homogeneous polynomial function on t of degree N = 12 (d − n), i.e., the number of all positive
roots.
Let hξ, η i be a positive definite inner product on g which is invariant under the adjoint action
of G and such that the cube generated by an orthonormal basis of g has unit volume relative to
p
µ L . Let k ξ k = hξ, ξ i .
If φ is a suitable G-invariant function on g, we have [6]
Z Z
vol( G/T )
φ(ξ )dµ L (ξ ) = φ(τ ) P(τ )2 dµ L (τ ) (4.7)
g |W | t


be the counterpart for g of Weyl’s integration formula. By taking φ(ξ ) = exp − 12 hξ, ξ i , we
obtain
q Z   Z  
1 1 vol( G ) 1
(2π )d = exp − hξ, ξ i dµ L (ξ ) = exp − hτ, τ i P(τ )2 dµ L (τ ) (4.8)
g 2 | W | vol( T ) t 2

To calculate this integral, we proceed as follows. Let h x, yi = ∑ j x j y j be the usual inner product
on R n . This extends to a scalar product on the algebra S(R n ) of polynomial functions on R n ,


such that xα , x β = α!δαβ for any two multi-indices α = (α1 , . . . , αn ), β = ( β 1 , . . . , β n ) ∈ N n . Here
α! := ∏ nj=1 α j ! and δαβ := ∏nj=1 δα j β j . Indeed, such scalar product is equivalently defined [12] by

h p, qi ∂ := p(∂)q|x=0 ,

where p, q ∈ R [ x1 , . . . , xn ] are homogeneous polynomials of the same order and p(∂) means that
x j is replaced by ∂ j in p( x) ≡ p( x1 , . . . , xn ). For instance, a generic polynomial in R [ x1 , . . . , xn ]
can be written as
n
αj
p( x ) = ∑ cα xα , xα := ∏ xj .
α j=1


∂
αj β 
Thus we know that αj xj j  = α j !δα j β j , and
∂x j 
x j =0
  !
D E n
∂ αj n
βj 
n
∂α j β
n
xα , x β = ∏ α j  ∏ xj  =∏ αj xj j = ∏ α j !δα β
j j
= α!δαβ .
j=1 ∂x j j=1  j=1 ∂x j j=1
( x1 ,...,xn )=0

60
Let γ be the (Gaussian) measure on R n defined by

1 1
dγ( x) = p exp − h x, xi [dx]
(2π )n 2

where [dx] is Lebesgue measure. For a function f on R n , let f ∗ = f ∗ γ, i.e.,

Z
f ∗ ( x) = f ( x − y)dγ(y)
Rn

whenever the integral is defined. Then for all f , g ∈ S(R n ) we have

∆
f∗ = e2 f (4.9)

where ∆ is the Laplace operator, and

Z
f ∗ (ix) g∗ (ix)dγ( x) = h f , gi ∂ . (4.10)
Rn

ξα
To prove Eq. (4.9), we may assume by linearity that f is a monomial xα , i.e. the coefficient of α!
in eh x,ξ i . Indeed, !
h x,ξ i
n ∞ (∑nj=1 x j ξ j )m
e = exp ∑ xj ξ j = ∑ m!
,
j=1 m =0

where !m
n  
m α α
∑ xj ξ j = ∑
α
x ξ
j=1 α: | α |= m

for (mα ) = m!
α1 !···αn ! , x
α = x1α1 · · · xnαn , and ξ α = ξ 1α1 · · · ξ nαn ; | α | = ∑nj=1 α j and α! = α1 ! · · · αn !. Thus
∞
ξα α
eh x,ξ i = ∑ ∑ x .
m =0 α: | α |= m α!

This means that eh x,ξ i is a linear combination of monomials xα . Hence it is enough to verify
Eq. (4.9) when f ( x) = eh x,ξ i ; indeed,
Z
f ∗ ( x) = f ( x − y)dγ(y)
Rn
Z  
h x − y,ξ i p1 1
= e exp − hy, yi [dy]
Rn (2π )n 2
Z  
h x,ξ i h− y,ξ i p 1 1
= e e exp − hy, yi [dy]
Rn (2π )n 2

implying
n Z  
∗ h x,ξ i 11 2
f ( x) = e
Rn
∏ √2π
exp − (y j + 2ξ j y j ) dy j
2
j=1
n Z  
1 2 1 1
= eh x,ξ i ∏ e 2 ξ j √ exp − (y j + ξ j )2 d(y j + ξ j )
j=1 2π R n 2

61


then a simple calculation shows that f ∗ ( x) = eh x,ξ i exp 1
2 hξ, ξ i . Then
! ! !!
n 2 n n n
∂ ∂2
∆eh x,ξ i = ∑ 2 exp ∑ ξ j xj = ∑ exp ∑ ξ j xj
j=1 ∂x j j=1 j=1 ∂x2j j=1
n
= ∑ ξ 2j ehx,ξ i = hξ, ξ i eh x,ξ i ,
j=1

∆
which means that f ∗ ( x) = e 2 f ( x) since
∞  m
∆ ∆
h x,ξ i 1 ∆
e f ( x) = e e
2 2 = ∑ eh x,ξ i
m =0 m! 2
∞ ∞
1 1
= ∑ m
∆m eh x,ξ i = ∑ m
hξ, ξ i m eh x,ξ i
m =0 m!2 m =0 m!2
1
= e 2 hξ,ξ i eh x,ξ i = f ∗ ( x).

Likewise, it is enough to verify Eq. (4.10) when f ( x) is replaced by eh x,ξ i and g( x) by eh x,η i (and
h f , gi ∂ by ehξ,η i ). Indeed, for this case,
* +
D E ∞
ξα ∞
ηβ
h f , gi ∂ = eh x,ξ i , eh x,η i = ∑ ∑ α! xα , ∑ ∑ β! x β
∂
i=0 α: | α |= i j=0 β: | β |= j
∂
∞ ∞
ξα ηβ D α βE
= ∑ ∑ ∑ ∑ x ,x
i=0 α: | α |= i j=0 β: | β |= j
α! β! ∂

∞
ξα ηα
= ∑ ∑ α! = ehξ,η i
j=0 α: | α |= j

and
Z Z
1 1
f ∗ (ix) g∗ (ix)dγ( x) = e 2 hξ,ξ i ehix,ξ i e 2 hη,η i ehix,η i dγ( x)
Rn Rn Z
1 1
= e 2 hξ,ξ i e 2 hη,η i eih x,ξ −η i dγ( x)
Rn
1 1 1
= e 2 hξ,ξ i 2 hη,η i − 2 hξ − η,ξ − η i
e e = ehξ,η i

where we used the fact that

Z ∞  2
1 − at2 −ity 1 y
√ e e dt = √ exp − .
2π −∞ 2a 4a

Suppose now that f is a harmonic homogeneous polynomial, i.e. ∆ f = 0. Then ∆k f = 0 for k > 1
and
∞
∆ ∆k
f∗ = e2 f = ∑ 2k k! f = f
k=0

62
by Eq. (4.9), and therefore from Eq. (4.10) we obtain
Z Z
h f , f i∂ = f ∗ (ix) f ∗ (ix)dγ( x) = f (ix) f (ix) dγ( x)
n Rn
ZR
= f ( x)2 dγ( x).
Rn

Now we have already proven that if f is a harmonic homogeneous polynomial, then we have
Z Z
1 1
h f , f i∂ = f ( x)2 dγ( x) = p f ( x)2 e− 2 h x,x i [dx]. (4.11)
Rn (2π )n Rn

The polynomial P is skew-symmetric with respect to the Weyl group W, and the inner product
(on t or t∗ ) is W-invariant. It follows that ∆P = 0, and hence from Eq. (4.11) that
Z   q
1
exp − hτ, τ i P(τ )2 dµ L (τ ) = (2π )n h P, Pi∂ ,
t 2

where
* + * +
N N
h P, Pi∂ = ∏ α, ∏ α = ∏ αj , ∏ αj
α >0 α >0 j=1 j=1
D E
def
= ∑ α 1 ⊗ · · · ⊗ α N , α σ −1 ( 1) ⊗ · · · ⊗ α σ −1 ( N ) .
σ∈S N

Now substituting this integral value in Eq. (4.8), we obtain

q q
1 vol( G )
(2π )d = (2π )n h P, Pi∂ .
| W | vol( T )
That is,

vol( G ) d−n |W | |W |
= (2π ) 2 = (2π ) N . (4.12)
vol( T ) h P, Pi∂ h P, Pi∂
d−n
Note here that N = 2 . Now the number h P, Pi has already been calculated by Steinberg [9]:
n
h P, Pi∂ = | W | P(ρ) = 2− N | W | ∏ m j ! ∏ hα, αi
j=1 α >0

1
where ρ := 2 ∑α>0 α is half the sum of the positive roots, and the m j are the exponents of G.
Indeed, from Weyl’s denominator formula:
 1 1

∑ sign(w)ewρ = ∏ e 2 α − e− 2 α ,
w ∈W α >0

we see that for any chosen X such that hα, X i 6= 0 for all α > 0, we have that
 
∑ sign(w)ehwρ,tXi = e−hρ,tXi ∏ ehα,tXi − 1 , ∀t ∈ R.
w ∈W α >0

63
That is,
 
∑ sign(w)ethwρ,X i = e−thρ,X i ∏ ethα,X i − 1
w ∈W α >0
ethα,X i − 1
= t N e−thρ,X i ∏ .
α >0 t

Now using the Taylor expansion of ex :

∞
xk
ex = ∑ ,
k=0
k!

we see that
ethα,X i − 1 ∞ k−1
t hα, X ik
= hα, X i + ∑ = hα, X i + o(t).
t k=2
k!
And thus

ethα,X i − 1
t N e−thρ,X i ∏ = t N (1 + o(t))( ∏ hα, X i + o(t)) = t N ( P( X ) + o(t)).
α >0 t α >0

Taking the N-th derivative with respect to t on both sides, we get that

∑ sign(w) hwρ, X i N ethwρ,X i = N!P( X ) + o(t).

w ∈W

Now taking the limit for t → 0, we get that

∑ sign(w) hwρ, X i N = N!P( X ) ⇐⇒ ∑ sign(w)(wρ) N = N!P.

w ∈W w ∈W

Thus
* + * +
1 N 1 N
h P, Pi∂ =
N! ∑ sign(w)(wρ) , ∏ α =
N! ∑ sign(w) (wρ) , ∏ α
w ∈W α >0 w ∈W α >0
* +
1 1
= ∑ sign(w)2 ρN , ∏ α = | W | N! ∏ hρ, αi = | W | P(ρ)
N! w ∈W α >0 N! α >0

where we used the fact that the polynomial P is skew-symmetric with respect to the Weyl group
W. We can also present another approach to the identity h P, Pi∂ = | W | P(ρ). Denote qρ ( X ) =
∑w∈W sign(w)ehwρ,X i Note that

 
h wρ,X i  h wρ,X i 
P ( ∂ ) q ρ ( 0) = ∑ sign(w) P(∂)e 
X =0
= ∑ sign(w) ∏ ∂α e 

w ∈W w ∈W α >0 X =0
d
= ∑ sign(w) ∏ t=0 ehwρ,tαi = ∑ sign(w) ∏ hwρ, αi ,
w ∈W α>0 dt w ∈W α >0

64
i.e.,

P ( ∂ ) q ρ ( 0) = ∑ sign(w) ∏ hα, wρi = ∑ sign(w) P(wρ)

w ∈W α >0 w ∈W
2
= ∑ sign(w) P(ρ) = | W | P(ρ).
w ∈W

In addition,

 
1
2 hα,X i − 21 h α,X i 
P ( ∂) ∏ e −e  = P(∂) P(0) = h P, Pi∂ .
α >0

X =0

Indeed,
!
 1
 ∞
α2n+1
2α − 21 α
∏ e −e = ∏ α + ∑ 2n
2 (2n + 1)!
= ∏ α + terms of degree at least N + 1.
α >0 α >0 n =1 α >0

Note that P(∂) is the linear partial operator of order N, hence


 1 
1 
P(∂) ∏ e 2 hα,X i − e− 2 hα,X i  = P(∂) P(0) + 0 = h P, Pi∂ .
α >0

X =0

Thus
h P, Pi∂ = | W | P(ρ).
Now for any root α,
2 hρ, αi

= ρ, α∨ = ht(α∨ ),
hα, αi
the height of α∨ , and it is known that if d1 6 d2 6 · · · 6 dn are the degrees of the basic invariants,
where d j = m j + 1 [11], the number of roots of height k minus the number of roots of height k + 1
is just the number of m j ’s equal to k (or the number of d j ’s equal to k + 1) [15, 28, 21, 11]. That is,
 ∨       
 α > 0 : ht(α∨ ) = k  −  α∨ > 0 : ht(α∨ ) = k + 1  =  j ∈ [n] : m j = k  .

Indeed, in order to show this result, we partition the number N of all positive roots via the height
of all positive co-roots:
 −1 
hG
∨
α |α > 0 = α∨ > 0 : ht(α∨ ) = j ,
j=1

where h is the Coxeter number [11] for Weyl group W. Denote by nk = |{α∨ > 0 : ht(α∨ ) = k}|.
Thus we see that
h −1
N= ∑ nk , n1 > n2 > · · · > n h −1 .
k=1

It is shown that the Poincaré polynomial W (t) := ∑w∈W tℓ(w) , where ℓ(w) is the length of w ∈ W,
can be factorized by two ways [21]:
1 − t1+ht(α) n
1 − t m j +1
W ( t) = ∏ 1 − tht(α) = ∏ 1−t
. (4.13)
α >0 j=1

65
This implies that
n
1 + ht(α)
| W | = lim W (t) =
t →1
∏ ht(α) = ∏ ( m j + 1) .
α >0 j=1

Note the above identity holds also for dual root system since W corresponds to both the original
root system R and its dual root system R∨ , it follows that
n
1 + ht(α∨ )
∏ ht(α∨ )
= ∏ ( m j + 1) .
α >0 j=1

Now partition [n] by the number of times k as exponents of W:

−1
hG 
[n ] = Ik , I k := j ∈ [n ] : m j = k .
k=1

Then   n2   n h −1
1 + ht(α∨ ) 3 h
∏ ht(α∨ ) = 2n1 ··· = 2n 1 − n 2 3n 2 − n 3 · · · h n h −1
α >0 2 h−1
and
n
∏(m j + 1) = 2|I | 3|I | · · · h|I − | .
1 2 h 1

j=1

Therefore we have

nk − nk+1 = |Ik | , k = 1, . . . , h − 2; nh−1 = |Ih−1 | .

In what follows, we calculate the value of ∏α>0 ht(α∨ ). In fact, we have ht(α∨ ) ∈ [1, h − 1] ∩ N
and

∏ ht(α∨ ) = 1n 2n 1 2
· · · ( h − 1) n h −1 . (4.14)
α >0

Clearly,

n1 = |Ih−1 | + · · · + |I2 | + |I1 | ,

n2 = |Ih−1 | + · · · + |I2 | ,
..
.
nh−1 = |Ih−1 | .

Furthermore,

∏ ht(α∨ ) = (1!)|I1 | (2!)|I2 | · · · [(h − 1)!]|Ih−1 |

α >0
= ∏ mj ! × ∏ mj ! × · · · × ∏ mj!
j∈I1 j∈I2 j∈I h −1
n
= ∏ mj! = ∏ mj! = ∏ m j !.
j∈I1 ⊔···⊔I h −1 j∈[n ] j=1

66
Thus

 n n
2 h ρ, αi
∏ = ∏ ρ, α∨ = ∏(dj − 1)! = ∏ m j !.
α>0 h α, α i α >0 j=1 j=1

Therefore,
n
2 hρ, αi hα, αi
P ( ρ) = ∏ hα, ρi = ∏ ∏ = 2− N ∏ m j ! ∏ hα, αi .
α >0 α>0 h α, α i α >0 2 j=1 α >0

That is,
n
P(ρ) = 2− N ∏ m j ! ∏ hα, αi . (4.15)
j=1 α >0

Since also | W | = ∏nj=1 d j = ∏nj=1 (m j + 1) [11], it follows that

n
h P, Pi = | W | P(ρ) = 2− N | W | ∏ m j ! ∏ hα, αi .
j=1 α >0

That is,
n
h P, Pi = 2− N ∏ d j ! ∏ hα, αi . (4.16)
j=1 α >0

Hence we have
vol( G ) (2π ) N 2π
vol( T )
=
P ( ρ)
= ∏ hα, ρi (4.17)
α >0

a formula due to Harish-Chandra, and also, using Eq. (4.6),

(2π ) N vol( T ) (2π ) N +n vol(t/tZ )

vol( G ) = = −N n
P ( ρ) 2 ∏ j=1 m j ! ∏ α>0 hα, αi
4 n
2π m j +1
= vol(t/tZ ) ∏ ∏ mj !
α>0 h α, α i j=1

 n
= vol(t/tZ ) ∏ α∨ , α∨ ∏ vol (Sr j )
α >0 j=1

where r j = 2m j + 1 and ∑nj=1 m j = N and 4

hα,αi
= hα∨ , α∨ i for α∨ is the coroot corresponding to α.
Let gC , tC be the complexifications of g, t. For X = ξ + iη ∈ gC (ξ, η ∈ g) Let X̄ = ξ − iη. There
exists a Chevalley basis of gC relative to tC in which the root vectors Xα satisfy X−α = X̄α . Let
ξ α = Xα + X−α , ηα = i( Xα − X−α ), then the "integer lattice" gZ is spanned by the ξ α and ηα for
α > 0, together with a basis of tZ ; the 2N vectors ξ α , ηα are orthogonal to each other and to t, and
k ξ α k = k ηα k = k α∨ k. It follows that


vol(g/gZ ) = vol(t/tZ ) ∏ α∨ , α∨ ;
α >0

67
substituting this, we derive the desired formula:
n
vol( G ) = vol(g/gZ ) ∏ vol (Sr j ) .
j=1

We are done.

Remark 4.10. Let (W, S) be an irreducible finite Coxeter system of rank n with S = {sα1 , . . . , sαn }
its set of simple reflections. The element c := sα1 · · · sαn ∈ W is called Coxeter transformation.
Since all Coxeter transformations are conjugate, they have the same order, characteristic polyno-
mial and eigenvalues. The order h of Coxeter elements is called the Coxeter number of W. For a
fixed Coxeter element c ∈ W, if its eigenvalues are of the form:

exp (2πim1 /h) , , exp (2πimn /h)

with 0 < m1 6 · · · 6 mn < h, then the integers m1 , . . . , mn are called the exponents of W. We have
already known that for any irreducible root system R of rank n,

(i) m j + mn− j+1 = h for 1 6 j 6 n;

(ii) m1 = 1, mn = h − 1;

(iii) the height of the highest root in R+ is given by mn = h − 1.

Let k j := |{α ∈ R+ |ht(α) = j}|. The height distribution of R+ is defined as a multiset of positive
integers:
{ k 1 , . . . , k h −1 } .

Apparently, (k1 , . . . , kh−1 ) ⊢ N = | R+ |. It is well-known that the exponents of the Weyl group
W are given by the dual partition of the height distribution of R+ is given by a multiset of non-
negative integers [15, 28, 21, 1]:
n o
( 0) n − k 1 , ( 1) k 1 − k 2 , . . . , ( h − 2) k h −2 − k h −1 , ( h − 1) k h −1 ,

where ( a)b means the integer a appears exactly b times. If k j − k j+1 > 0, then m j appears exactly
k j − k j+1 times. Otherwise, m j does not appear if k j = k j+1 .
Note that in the Macdonald’s method to the volume of a compact Lie group, we see that
d−n
d = ∑nj=1 r j = 2 ∑ nj=1 m j + n. It follows that ∑nj=1 m j = 2 = N. Besides, since m j + mn− j+1 = h,
it follows that
n
1
N= ∑ m j = 2 nh.
j=1

68
Remark 4.11. Denote all polynomials in x = ( x1 , . . . , xd ) with coefficients in F, where F = C or
R, by F [ x1 , . . . , xd ]. A polynomial is called homogeneous if all the monomials appearing in this
polynomial have the same total degree. Denote the space of homogeneous polynomials of degree
n in d variables by Pn (F d ). That is,
( )
P n (F d ) = P | P( x) = ∑ cα xα , where x ∈ F d .
α: | α |= n

Note that dim Pn (R d ) = (n+nd−1). For any two homogeneous polynomials P, Q ∈ Pn (F d ), a scalar
product may be defined as follows:

h P, Qi := ∑ p̄α qα α!,
α: | α |= n


where P( x) = ∑α:| α |=n pα xα and Q( x) = ∑α:| α |=n qα xα . From this definition, we see that xα , x β =
α!δαβ . In particular, if P, Q ∈ Pn (R d ), the scalar product can be rewritten as [12, 13]

h P, Qi = ( P(∂) Q)(0).

Indeed, since (∂α x β ) x =0 = α!δαβ and

! !
 
P ( ∂) Q = ∑ pα ∂α ∑ qβ xβ = ∑ pα q β ∂α x β ,
α: | α |= n β: | β |= n α,β: | α |=| β |= n

it follows that


( P(∂) Q)(0) = ∑ pα q β α!δαβ = ∑ pα qα α!,
α,β: | α |=| β |= n α: | α |= n

which means that h P, Qi = ( P(∂) Q)(0).

As an example, let P( x) = hα1 , xi hα2 , xi and Q( x) = h β1 , xi h β 2 , xi be real polynomials in
P2 (R3 ), where α j = ( a1j , a2j , a3j ) and β j = (b1j , b2j , b3j ) for j = 1, 2. In what follows, we calculate the
scalar product h P, Qi. Apparently,
3 3  
j i j j i
P ( ∂) = ∑ a1i a2 ∂i ∂ j = a i i 2
∑ 1 2 i ∑ 1 2 1 2 ∂i ∂ j ,
a ∂ + a a + a a
i,j=1 i=1 i< j
3 3  
j i j j i
Q( x) = ∑ b1i b2 xi x j = b i i 2
∑ 1 2 i ∑ 1 2 1 2 xi x j .
b x + b b + b b
i,j=1 i=1 i< j

It follows that
3   
j j j j
P(∂) Q( x) = 2 ∑ a1i a2i b1i b2i + ∑ a1i a2 + a1 a2i b1i b2 + b1 b2i
i=1 i< j

69
Now we see that

h α1 , β 1 i h α2 , β 2 i + h α1 , β 2 i h α2 , β 1 i
! ! ! !
3 3 3 3
j j j j
= ∑ a1i b1i ∑ a2 b2 + ∑ a1i b2i ∑ a2 b1
i=1 j=1 i=1 j=1
" # " #
3 3
j j j j
= ∑ a1i b1i a2i b2i +∑ a1i b1i a2 b2 + ∑ a1i b2i a2i b1i +∑ a1i b2i a2 b1
i=1 i6 = j i=1 i6 = j
3
j j j j
= 2 ∑ a1i a2i b1i b2i + ∑ a1i b1i a2 b2 + ∑ a1i b2i a2 b1 .
i=1 i6 = j i6 = j

That is,
3
j j j j
hα1 , β1 i hα2 , β2 i + hα1 , β2 i hα2 , β1 i = 2 ∑ a1i a2i b1i b2i + ∑ a1i b1i a2 b2 + ∑ a1i b2i a2 b1
i=1 i6 = j i6 = j
3    
j j j j j j j j
= 2 ∑ a1i a2i b1i b2i + ∑ a1i b1i a2 b2 + a1 b1 a2i b2i + ∑ a1i b2i a2 b1 + a1 b2 a2i b1i
i=1 i< j i< j
3  
j j j j j j j j
= 2 ∑ a1i a2i b1i b2i + ∑ a1i b1i a2 b2 + a1 b1 a2i b2i + a1i b2i a2 b1 + a1 b2 a2i b1i
i=1 i< j
3   
j j j j
= 2 ∑ a1i a2i b1i b2i + ∑ a1i a2 + a1 a2i b1i b2 + b1 b2i .
i=1 i< j

Therefore, we can conclude that

( P(∂) Q)(0) = hα1 , β1 i hα2 , β2 i + hα1 , β2 i hα2 , β1 i

D E
= ∑ α 1 ⊗ α 2 , β σ −1 ( 1) ⊗ β σ −1 ( 2) .
σ ∈ S2



We can generalize this result to the case where P( x) = ∏nj=1 α j , x and Q( x) = ∏kj=1 β j , x , in
short, P = ∏kj=1 α j and Q = ∏kj=1 β j . We have the following:
D E
h P, Qi = ( P(∂) Q)(0) = ∑ α 1 ⊗ · · · ⊗ α k , β σ −1 ( 1) ⊗ · · · β σ −1 ( k ) .
σ ∈ Sk

Remark 4.12. Consider the real vector space V generated by the roots. h·, ·i denotes the inner
product on V which is invariant under the operation of the Weyl group W. The given inner
product on V extends to one on the symmetric algebra of V by the formula:
* +
k k k D E D E D E
∏ j∏ j
α , β = ∑ ∏ j σ ( j)
α , β − 1 = ∑ 1 σ ( 1)
α , β − 1 · · · α , β − 1
k σ ( k) (4.18)
j=1 j=1 σ ∈ Sk j =1 σ ∈ Sk
D E
= ∑ α 1 ⊗ · · · ⊗ α k , β σ −1 ( 1) ⊗ · · · β σ −1 ( k ) (4.19)
σ ∈ Sk

70
where α j , β j ∈ V. In view of this, we see that
* +
∏ α, ρN = ∑ ∏ hα, ρi = N! ∏ hα, ρi .
α >0 σ ∈ S N α >0 α >0

In the following, we show that [23]

* +
n
∨
∏ α, ∏ α = ∏ d j !, (4.20)
α >0 α >0 j=1

2ασ−1 ( j )
where α∨ = 2α
hα,αi such that hα, α∨ i = 2. Indeed, by replacing each α∨
σ −1 ( j )
by D E in the
ασ−1 ( j ) ,ασ−1 ( j )
definition, we see that
* +
D E
∏ α, ∏ α∨ α1 ⊗ · · · α N , α ∨ ∨
def
= ∑ σ −1 ( 1) ⊗ · · · α σ −1 ( N )
α >0 α >0 σ∈S N
D E 2N
= ∑ α 1 ⊗ · · · α N , α σ −1 ( 1) ⊗ · · · α σ −1 ( N ) D E
σ∈S N ∏N j = 1 α σ −1 ( j ) , α σ −1 ( j )

2N D E
= ∑
∏α>0 hα, αi σ∈S
α 1 ⊗ · · · α N , α σ −1 ( 1) ⊗ · · · α σ −1 ( N ) .
N

Therefore
* +
n
∨ 2N
∏ α, ∏ α =
∏ α>0 hα, αi
h P, Pi = ∏ dj !.
α >0 α >0 j=1

Note that in deriving the volume formula of a compact Lie group, we used the following fact
which is necessarily recorded here:

Theorem 4.13 ([6]). The mapping that assigns to φ in C0 (g), the space of continuous functions with
compact support on g, the function on G/T × t
 
( gT, X ) 7−→ φ(Adg ( X ))  det(adX )g/t  (4.21)

extends to a topological isomorphism: L1 (g) → L1 ( G/T × t)W , where now sT ∈ W acts on ( gT, X ) by
sending it to ( gs−1 T, Ads ( X )). Moreover, if φ ∈ L1 (g), then:
Z Z Z 
 
1
φ( X )dX = φ(Adg ( X ))d( gT )  det(adX )g/t  dX. (4.22)
g | W | t G/T
It is easily seen that if φ is G-invariant, i.e., φ(Adg ( X )) = φ( X ) for all g ∈ G, then
Z
vol( G )
φ(Adg ( X ))d( gT ) = φ( X )vol( G/T ) = φ( X ) .
G/T vol( T )
Thus
Z Z  
1 vol( G )
φ( X )dX = φ( X )  det(adX )g/t  dX,
g | W | vol( T ) t

71
  
where Jacobian  det(adX )g/t  = P( X )2 for P = ∏ α>0 α.
More simpler proof about Macdonald’s volume formula can be found in [10]. With Macdon-
ald’s volume formula, one can derive all volume formulas for orthogonal groups and unitary
groups. Although these methods is a digression to the subject of the present paper, I also collect
these material together for reference.
Once again, we will use the Macdonald’s volume formula for a compact Lie group to derive
the volume formula for the special unitary group G := SU(n). The special unitary group SU(n)
is of rank n − 1. Its root system consists of 2( n2 ) roots spanning a (n − 1)-dimensional Euclidean
space. Here, we use n redundant coordinates instead of n − 1 to emphasize the symmetries of
the root system (the n coordinates have to add up to zero). In other words, we are embedding
this n − 1 dimensional vector space in an n-dimensional one. The maximal torus T of G is
equivalently identified with SU(1)×n , and its Lie algebra is t ∼
= R n−1 of rank n − 1. The lattice
tZ ∼
= Z n−1 . Thus vol(t/tZ ) = 1.
The set of all positive roots of g is given by

R+ = αij ∈ t∗ : αij ( X ) = xi − x j for any X = diag( x1 , . . . , xn ), i < j . (4.23)

Thus for any α ∈ R+ , we have hα, αi = 2. Note that N = | R+ | = (n2 ). Then the Coxeter number
2N
h= n −1 = n. The exponents of the special unitary group SU(n) are: {1, . . . , n − 1}. Note that
n −1
4
volHS ( G ) = vol(t/tZ ) ∏ ∏ vol (Sr ) , j

α>0 h α, α i j=1

where r j = 2j + 1. It follows that

n
n −1  
volHS ( G ) = 2( 2 ) ∏ vol S2j+1 .
j=1


Since U(n) = U(1) × SU(n), it follows from volHS (U(1)) = vol S1 that
  n
n  
volHS (U(n)) = vol S1 volHS (SU(n)) = 2( 2 ) ∏ vol S2j−1 .
j=1

That is,
n ( n +1 )
(2π ) 2
volHS (U(n)) = n .
∏ j=1 Γ ( j )
Besides, we can also employ Harish-Chandra’s volume formula for flag manifold, we get that

2π (2π ) N (2π ) N
volHS (U(n)/T n ) = ∏ hα, ρi = = −N n
α >0 P ( ρ) 2 ∏ j=1 m j ! ∏ α>0 hα, αi
n ( n −1 )
(2π ) N (2π ) 2
= −N N
= n ,
2 0!1! · · · (n − 1)!2 ∏ j=1 Γ ( j )

72
implying that
n ( n −1 ) n ( n −1 )
(2π ) 2 (2π ) 2
volHS (U(n)) = vol(T n ) n = (2π )n n .
∏ j=1 Γ ( j ) ∏ j=1 Γ ( j )

Again, we also see that

n ( n +1 )
(2π ) 2
volHS (U(n)) = n .
∏ j=1 Γ ( j )
Remark 4.14. We can rewrite the above results as two partitions of N, the number of all positive
roots, (n1 , . . . , nh−1 ) ⊢ N and (mn , . . . , m1 ) ⊢ N, where n j denotes the number of j-th row; mi the
number of n − j + 1-th column or the n − j + 1-th row of the conjugate partition of the original
partition (mn , . . . , m1 ) ⊢ N.

5 Applications

The present section is directly written based on [30, 37, 38]. The results were already obtained.
We just here add some interpretation, from my angle, about them since the details concerning
computation therein are almost ignored.

5.1 Hilbert-Schmidt volume of the set of mixed quantum states

Any unitary matrix may be considered as an element of the Hilbert-Schmidt space of operators
 
e V
with the scalar product hU, e iHS = Tr Ue ∗V
e . This suggests the following definition of an
e := U
invariant metric of the unitary group U (n): denote dG e ∗ dU,
e then
D E  
e dG
ds2 := dG, e = − Tr dGe2 , (5.1)
HS

implying
n    n  n  
 e 2  e 2  e 2
ds2 = ∑  dGij  = ∑ jj
 d G  + 2 ∑ ij  .
 d G (5.2)
i,j=1 j=1 i< j

e∗ = −dG,
Since dG e it follows that
n   n  2 n  2
 e 2 e eij )) .
ds2 = ∑  dG jj  + 2 ∑ d(Re( Gij )) + 2 ∑ d(Im( G (5.3)
j=1 i< j i< j

This indicates that the Hilbert-Schmidt volume element is given by

n ( n −1 )
n   n ( n −1 )
dν = 2 2
∏ d Im ( e
G jj ) eij ))d(Im( G
× ∏ d(Re( G eij )) = 2 2 [dG e ], (5.4)
j=1 i< j

73
that is,
Z Z
n ( n −1 )
volHS (U (n)) := dν = 2 2 e]
[d G (5.5)
U (n) U (n)
n ( n +1 )
n ( n −1 ) 2n π 2
= 2 2 × (5.6)
1!2! · · · (n − 1)!
n ( n +1 )
(2π ) 2
= . (5.7)
1!2! · · · (n − 1)!

Finally we have obtained the Hilbert-Schmidt volume of unitary group:

n ( n +1 )
n ( n −1 ) (2π ) 2
volHS (U (n)) = 2 2 vol (U (n)) = . (5.8)
1!2! · · · (n − 1)!

We compute the volume of the convex (n2 − 1)-dimensional set D (C n ) of density matrices of
size n with respect to the Hilbert-Schmidt measure.
The set of mixed quantum states D (C n ) consists of Hermitian, positive matrices of size n,
normalized by the trace condition

D (C n ) = {ρe : C n → C n |ρe∗ = ρe, ρe > 0, Tr (ρe) = 1}. (5.9)

It is a compact convex set of dimensionality (n2 − 1). Any density matrix may be diagonalized
e U
by a unitary rotation, ρe = UΛ e ∗ , where U
e ∈ U (n) and Λ = diag(λ1 , . . . , λn ) for λ j ∈ R + .
Since Tr (ρe) = 1, it follows that ∑nj=1 λ j = 1, so the spectra space is isomorphic with a (n − 1)-
n o
dimensional probability simplex ∆n−1 := p ∈ R + : ∑nj=1 p j = 1 .
Let Be be a diagonal unitary matrix. Since ρe = U e BΛ
e Be∗ U
e ∗ , in the generic case of a non-
degenerate spectrum (i.e. with distinct non-negative eigenvalues), the unitary matrix U e is deter-
e On the other hand, the matrix Λ is defined up to a
mined up to n arbitrary phases entering B.
permutation of its entries. The form of the set of all such permutations depends on the character
of the degeneracy of the spectrum of ρe.
e BΛ
Representation ρe = U e Be∗ U
e ∗ makes the description of some topological properties of the
(n2 − 1)-dimensional space D (C n ) easier. Identifying points in ∆n−1 which have the same com-
ponents (but ordered in a different way), we obtain an asymmetric simplex ∆ e n−1 . Equivalently,
one can divide ∆n−1 into n! identical simplexes and take any one of them. The asymmetric
e n−1 can be decomposed in the following natural way:
simplex ∆
[
e n −1 =
∆ δd1 ,...,dk , (5.10)
d1 +···+ dk = n

74
 e n − 1 , d1
where k = 1, . . . , n denotes the number of different coordinates of a given point of ∆
the number of occurrences of the largest coordinate, d2 the number of occurrences of the sec-
ond largest etc. Observe that δd1 ,...,dk is homeomorphic with the set Gk , where G1 is a single
point, G2 is a half-closed interval, G3 an open triangle with one edge but without corners and,
generally, Gk is an (k − 1)-dimensional simplex with one (k − 2)-dimensional hyperface without
boundary (the latter is homeomorphic with an (k − 2)-dimensional open simplex). There are n
ordered eigenvalues: λ1 > λ2 > · · · > λn , and n − 1 independent relation operators "larger(>) or
e n−1 consists of 2n−1 parts,
equal(=)", which makes altogether 2n−1 different possibilities. Thus, ∆
out of which (mn− 1
−1) parts are homeomorphic with Gm , when m ranges from 1 to n.
Let us denote the part of the space D (C n ) related to the spectrum in δd1 ,...,dk (k different
eigenvalues; the largest eigenvalue has d1 multiplicity, the second largest d2 etc) by Dd1 ,...,dk . A
mixed state ρe with this kind of the spectrum remains invariant under arbitrary unitary rotations
performed in each of the d j -dimensional subspaces of degeneracy. Therefore the unitary matrix
e has a block diagonal structure with k blocks of size equal to d1 , . . . , dk and
B

Dd1 ,...,dk ∼ [U (n)/(U (d1 ) × · · · × U (dk ))] × Gk , (5.11)

where d1 + · · · + dk = n and d j > 0 for j = 1, . . . , k. Thus D (C n ) has the structure

[ [
D (C n ) ∼ Dd1 ,...,dk ∼ [U (n)/(U (d1 ) × · · · × U (dk ))] × Gk , (5.12)
d1 +···+ dk = n d1 +···+ dk = n

e
where the sum ranges over all partitions (d1 , · · · , dk ) ⊢ n of n. The group of rotation matrices B
equivalent to U (d1 ) × · · · × U (dk ) is called the stability group of U (n).
Note also that the part of D1,...,1 represents a generic, non-degenerate spectrum. In this case
all elements of the spectrum of ρe are different and the stability group is equivalent to an n-torus
 
D1,...,1 ∼ U (n)/(U (1)×n × Gn . (5.13)

The above representation of generic states enables us to define a product measure in the space
D (C n ) of mixed quantum states. To this end, one can take the uniform (Haar) measure on U (n)
and a certain measure on the simplex ∆n−1 .
The other 2n−1 − 1 parts of D (C n ) represent various kinds of degeneracy and have measure
zero. The number of non-homeomorphic parts is equal to the number P(n) of different represen-
tations of the number n as the sum of positive natural numbers. Thus P(n) gives the number of
different topological structures present in the space D (C n ).
To specify uniquely the unitary matrix of eigenvectors U, e it is thus sufficient to select a point
on the coset space
(n)
FlC := U (n)/U (1)×n ,

called the complex flag manifold. The volume of this complex flag manifold is:

75
   volHS (U (n)) (2π ) 2
n ( n −1 )
(n)
volHS FlC = n = . (5.14)
volHS (U (1)) 1!2! · · · (n − 1)!

The generic density matrix is thus determined by (n − 1) parameters determining eigenvalues

and (n2 − n) parameters related to eigenvectors, which sum up to the dimensionality (n2 − 1)
of D (C n ). Although for degenerate spectra the dimension of the flag manifold decreases, these
cases of measure zero do not influence the estimation of the volume of the entire set of density
matrices. In this subsection, we shall use the Hilbert-Schmidt metric. The infinitesimal distance
takes a particularly simple form

ρ k2HS = hde
ds2HS = k de ρiHS
ρ, de (5.15)

e U
valid for any dimension n. Making use of the diagonal form ρe = UΛ e † , we may write
 
ρ=U
de e dΛ + [Ud e U,
e Λ] U e∗ (5.16)

Thus the infinitesimal distance can be rewritten as

n n D   E
  e ∗ e  2
ds2HS = ∑ dλ2j + 2 ∑(λi − λ j )2  i  U dU  j  (5.17)
j=1 i< j
n n  D   E2
  e 
= ∑ dλ2j 2
+ 2 ∑(λi − λ j )  i dG j  , (5.18)
j=1 i< j

e=U
where dG e ∗ dU.
e Apparently, ∑nj=1 dλ j = 0 since ∑nj=1 λ j = 1. Thus

n −1 n  
 e 2
ds2HS = ∑ dλi mij dλ j + 2 ∑(λi − λ j )2  dG ij  . (5.19)
i,j=1 i< j
p
The corresponding volume element gains a factor det( M ), where M = [mij ] is the metric in the
(n2 − n)-dimensional simplex ∆n−1 of eigenvalues. Note that
 
1 1 ··· 1
 
 1 1 ··· 1 
 
M = 1n +  . . .. ..  .
 .. .. . . 
 
1 1 ··· 1

Therefore the Hilbert-Schmidt volume element is given by

 
√ n −1     
2 eij )  .
eij ) d Im( G
dVHS = n ∏ dλ j ∏ (λi − λ j )  ∏ 2d Re( G (5.20)
j=1 i< j
 i< j


76
Then
Z √ Z n −1 Z
n ( n −1 )
dVHS = n2 2
∏(λi − λ j )2 ∏ dλ j × e1 ]
[d G
i< j j=1
√ n ( n −1 )
= n2 2 vol (D (C n )) .

That is, respect to Hilbert-Schmidt measure, the volume of the set of mixed quantum states is
√ n ( n −1 )
volHS (D (C n )) = n2 2 vol (D (C n )) ,

i.e.

√ n ( n −1 ) Γ ( 1) Γ ( 2) · · · Γ ( n )
volHS (D (C n )) = n(2π ) 2 . (5.21)
Γ ( n2 )

We see from the above discussion that the obtained formula of volume depends the used
measure. If we used the Hilbert-Schmidt measure, then we get the Hilbert-Schmidt volume of
the set of quantum states [38]; if we used the Bures measure, then we get the Bures volume of
the set of quantum states [29].
A special important problem is to compute the volume of the set of all separable quantum
states, along this line, some investigation on this topic had already been made [35, 36]. There are
some interesting topics for computing volumes of the set of some kinds of states, for instance,
Milz also considered the volumes of conditioned bipartite state spaces [24], Link gave the ge-
ometry of Gaussian quantum states [16] as well. We can also propose some problems like this.
Consider the following set of all states being of the form:
   
1 1 2 2

1
C 12 , 12 := ρe12 ∈ D C ⊗ C : Tr1 (ρe12 ) = 12 = Tr2 (ρe12 ) .
2 2 2
Paratharathy characterized the extremal points of this set [26]. He obtained that all the extremal
points of this convex set is maximal entangled states. That is,
|0i|ψ0 i + |1i|ψ1 i
√ .
2
1 1


It remains open to compute the volume of this convex set C 2 12 , 2 12 .

5.2 Area of the boundary of the set of mixed states

The boundary of the set of mixed states is far from being trivial. Formally it may be written as a
solution of the equation
det(ρe) = 0

77
which contains all matrices of a lower rank. The boundary ∂D (C n ) contains orbits of different
dimensionality generated by spectra of different rank and degeneracy. Fortunately all of them
are of measure zero besides the generic orbits created by unitary rotations of diagonal matrices
with all eigenvalues different and one of them equal to zero;

Λ = {0, λ2 < · · · < λn } .

Such spectra form the (n − 2)-dimensional simplex ∆n−2 , which contains (n − 1)! the Weyl
chambers—this is the number of possible permutations of elements of Λ which all belong to
the same unitary orbits.
Hence the hyper-area of the boundary may be computed in a way analogous to (5.21):
Z Z
! Z
n  2 n 2
e] =
[d X δ ∑ λj − 1 ∏  λ i − λ 
j ∏ ( λ dλ j ) × e1 ]
[d G
e )= n −1 j
rank( X 0< λ2 <···< λn j=2 2 = i < j6 n j=2 U1 ( n )

n −1
zZ }| Z { ! Z
∞ ∞ n   n
1  λi − λ j 2 ∏(λ2 dλ j ) × e1 ]
=
( n − 1) ! 0
··· δ
0
∑ λj − 1 ∏ j
U1 ( n )
[d G
j=2 2 = i < j6 n j=2
n ( n −1 )
1 Γ ( 1) · · · Γ ( n ) Γ ( 1) · · · Γ ( n + 1) π 2
= 2
Γ(n) Γ ( n − 1) Γ ( 1) · · · Γ ( n )
n ( n −1 ) Γ ( 1) · · · Γ ( n + 1)
  
( n − 1)
=π 2 = vol D Cd ,
Γ ( n ) Γ ( n 2 − 1)
i.e.

√ n ( n −1 )
   √ n ( n −1 ) Γ ( 1) · · · Γ ( n + 1)
( n − 1)
volHS = n − 12 2 vol(n−1) D C d = n − 1(2π ) 2 .
Γ ( n ) Γ ( n 2 − 1)

In an analogous way, we may find the volume of edges, formed by the unitary orbits of the
vector of eigenvalues with two zeros. More generally, states of rank N − n are unitarily similar
to diagonal matrices with n eigenvalues vanishing,

Λ = {λ1 = · · · = λm = 0, λm+1 < · · · < λn } .

These edges of order m are n2 − m2 − 1 dimensional, since the dimension of the set of such spectra
is n − m − 1, while the orbits have the structure of U (n)/ (U (m) × U (1)n−m ) and dimensionality
n2 − m2 − (n − m). We obtain the volume of the hyperedges
 
√ vol Fl
(n)
(n−m) n−m 1 HS C
volHS =  
(n − m)! C (1+2m,2) vol Fl
(m)
n−m HS C

78
5.3 Volume of a metric ball in unitary group

Consider a metric ball around the identity 1 n in the n-dimensional unitary group U (n) with
Euclidean distance ǫ,
n o
Bǫ : = Ue ∈ U (n) :
Ue − 1n
6ǫ , (5.22)
2

where k∗k p is the p-norm for p = 2. We consider the invariant Haar-measure µ, a unform
√
e by e −1θ j . The joint density of the
distribution defined over U (n). Denote the eigenvalues of U
angles θ j is given by

1  √ √ 2
 −1θi −1θ j 
p ( θ1 , . . . , θ n ) = ∏
(2π )n n! 16i< j6n
e − e  , (5.23)

where θ j ∈ [−π, π ], j = 1, . . . , n. In what follows, we check the correctness of the integral formula:
Z
p(θ )dθ = 1.

 √ √ 2 √
 
Indeed, set J (θ ) = ∏16i< j6n  e −1θi − e −1θ j  and ζ j = e −1θ j , so
 2
J (θ ) = ∏  ζi − ζ j  = ∏(ζ i − ζ j )(ζ i−1 − ζ − 1
j ) (5.24)
i< j i< j

= (sign τ )(ζ 1 · · · ζ n )−(n−1) ∏ (ζ i − ζ j )2 , (5.25)

i< j

where τ = (n · · · 21), i.e. τ ( j) = n + 1 − j or τ is written as

!
1 2 ··· n
τ := .
n n−1 ··· 1

n ( n −1 )
Note that sign τ = (−1) 2 . We see that the integral is the constant term in

Cn (sign τ )(ζ 1 · · · ζ n )−(n−1) ∏(ζ i − ζ j )2 . (5.26)

i< j

Thus our task is to identify the constant term in this Laurent polynomial. To work on the last
factor, we recognize

V (ζ ) = ∏(ζ i − ζ j ) (5.27)
i< j

as a Vandermonde determinant; hence

σ(1)−1 σ( n )−1
V (ζ ) = ∑ (sign σ)ζ1 · · · ζn . (5.28)
σ ∈ Sn

79
Hence
σ(1)+ π (1)−2 σ( n )+ π ( n )−2
∏ ( ζ i − ζ j )2 = V ( ζ )2 = ∑ (sign σ)(sign π )ζ 1 · · · ζn . (5.29)
i< j σ,π ∈Sn

Let us first identify the constant term in

J (θ ) = (sign τ )(ζ 1 · · · ζ n )−(n−1) V (ζ )2 . (5.30)

We see this constant term is equal to

n
z
Z 2π
}| Z {
1 2π
··· J (θ )dθ
(2π )n 0 0
n
z
Z 2π
}| Z { !
1 2π
σ(1)+ π (1)− n −1 σ( n )+ π ( n )− n −1
= (sign τ )
(2π )n 0
···
0
∑ (sign σ)(sign π )ζ 1 · · · ζn dθ
σ,π ∈Sn
 Z 2π   Z 2π 
1 σ(1)+ π (1)− n −1 1 σ( n )+ π ( n )− n −1
= (sign τ ) ∑ (sign σ)(sign π )
2π 0
ζ1 dθ1 ×···×
2π 0
ζn dθn
σ,π ∈Sn
= (sign τ ) ∑ (sign σ)(sign π ) = (sign τ ) ∑ (sign σ)(sign π ).
( σ,π )∈Sn × Sn : ∀ j,σ( j)+ π ( j)= n+1 ( σ,π )∈Sn × Sn :π = τσ

Note that the sum is over all (σ, π ) ∈ Sn × Sn such that σ( j) + π ( j) = n + 1 for each j ∈ {1, . . . , n}.
In other words, we get π ( j) = n + 1 − σ( j) = τ (σ( j)) for all j ∈ {1, . . . , n}, i.e. π = τσ. Thus the
sum is equal to
(sign τ ) ∑ (sign σ)(sign τσ) = n!,
σ ∈ Sn

which gives rise to

n
}| Z { z
Z 2π
1 2π 1
n
··· J (θ )dθ = .
(2π ) 0 0 n!

e
Now the condition on the distance measure U − 1 n 6 ǫ is equivalent to
2
n  √ 2
 
∑  e −1θj − 1  6 ǫ2 .
j=1
√ √

Using Euler’s formula e −1θ = cos θ + −1 sin θ and the fact that (cos θ − 1)2 + sin2 θ = 4 sin2 θ
2 ,
we get
n  
e 2 θj ǫ2
U − 1 n 6 ǫ ⇐⇒
2
∑ sin 2
6
4
. (5.31)
j=1

Thus the (normalized) volume of the metric ball Bǫ equals the following:
Z
vol( Bǫ ) := µ( Bǫ ) = e ).
dµ(U (5.32)
Bǫ

80
By spectral decomposition of unitary matrix, we have
 
θ1 0 ··· 0
 
√  0 θ2 · · · 0 
Ue =V eDeV e = e −1Θ , Θ = 
e ∗, D  .

 , θ j ∈ [−π, π ]( j = 1, . . . , n).
 .. .. . .
 . . 0 
0 0 ··· θn

Hence h i
e ∗ · dU
V e ·V
e = dD
e+ Ve ∗ dV,
e De ,

implying that
 h i
e∗ · U
V e ∗ dU
e ·V e ∗ dD
e=D e+ Ve ∗ dV,
e De .

e=U
Let dG e ∗ dU, e1 = V
e dG e ∗ dV
e and dX
e=V e · V.
e ∗ · dG e Thus

[d X e]
e ] = [d G (5.33)

e ∈ U (n)/U (1)×n . We also have

because of V
 h i
e e ∗ e e e e∗ · V
e ∗ · dU
e · V,
e
dX = D · d D + dG1 , D = D

it follows that

e ] = [ dU
[d X e ]. (5.34)

Apparently,
n o
Bǫ = VeD
eVe ∗ ∈ U (n) : D e ∈ U ( n ) / U ( 1) × n ,
e ∈ U ( n ), V De − 1n
6ǫ . (5.35)
2

But
 √ √ 2
 −1θi 
e] =
[ dU ∏ e − e −1θ j  [d D e1 ],
e ][dG (5.36)
16i< j6n

therefore
 √ √ 2
 −1θi 
e] =
[d G ∏ e − e −1θ j  [d D e1 ],
e ][dG (5.37)
16i< j6n

e ∈ U ( n ) / U ( 1) × n ,
together with the facts that the region in which (θ1 , . . . , θn ) lies is symmetric and V
implying
Z Z Z  √ √ 2 Z
e] = 1  −1θi −1θ j  e e1 ].
Bǫ
[d G
n!
··· θ j ∈[− π,π ]( j =1,...,n ); ∏ e − e  [ d D ] ×
U (n)
[d G
2 16i< j6n
∑nj=1 sin2 ( θ j /2)6 ǫ4

81
That is,
Z Z Z  √ √ 2
1  −1θi 
e) =
dµ(U ··· θ j ∈[− π,π ]( j =1,...,n ); ∏ e − e −1θ j  [d D
e ],
Bǫ (2π )n n! 2 16i< j6n
∑ nj=1 sin2 ( θ j /2)6 ǫ4

where Z Z
[d Ge]
e) = R
dµ(U and e] = (2π )n
[d G e1 ].
[d G
[ d e]
G U (n) U1 ( n )
U (n)
From this, we get
Z Z n
vol ( Bǫ ) = ··· θ j ∈[− π,π ]( j =1,...,n ); p(θ1 , . . . , θn ) ∏ dθ j , (5.38)
2 j=1
∑ nj=1 sin2 ( θ j /2)6 ǫ4
√ √ ǫ2
where ǫ ∈ [0, 2 n]. For the maximal distance ǫ = 2 n, the restriction ∑nj=1 sin2 (θ j /2) 6 4
 
becomes irrelevant and vol B2√n = 1.
We start by rewriting the n-dimensional integral (5.38), with the help of a Dirac delta function,
as
Z ǫ2 Z π Z π n  ! n
4 θj
vol ( Bǫ ) = ··· δ t − ∑ sin2 p(θ1 , . . . , θn ) ∏ dθ j dt. (5.39)
0 −π −π j=1
2 j=1

We know that Z ∞
δ(t − a) f (t)dt = f ( a).
−∞
If f (t) ≡ 1 {t:α6t6 β}, then
Z β Z ∞
δ(t − a)dt = δ(t − a)1 {t:α6t6 β}dt = 1 {t:α6t6 β}.
α −∞

By using the Fourier representation of Dirac Delta function

Z ∞ √
1 − 1( t − a ) s
δ ( t − a) = e ds,
2π −∞

we get
Z ǫ2 Z ∞ Z ǫ2
!
4 1 4
√
− 1( t − a ) s
δ(t − a)dt = e dt ds
0 2π −∞ 0
√  √ ǫ2 
1
Z ∞ −1 1 − e −1 4 s
= √ ds.
2π −∞ se −1as
 
θj
Let a = ∑nj=1 sin2 2 = n
2 − ∑nj=1 12 cos θ j . Indeed,
n n n
n n n
− ∑ sin2 (θ j /2) = − + ∑ (1 − sin2 (θ j /2)) = − + ∑ cos2 (θ j /2)
2 j=1 2 j=1 2 j=1
n
n
n 1 1
= − +∑ 1 + cos θ j = ∑ 2 cos θj .
2 j=1 2 j=1

82
Therefore
√  √ ǫ2 
Z ǫ2 n  ! Z ∞ −1 1 − e −1 4 s √
4 θj 1 −1s ∑nj=1
cos θ j
δ t − ∑ sin2 dt = √ n e 2 ds
0 j=1
2 2π −∞ se −1 2 s
√  √ ǫ2  !
1
Z ∞ −1 1 − e −1 4 s n √ cos θ j
−1s
=
2π
√
−1 n2 s ∏e 2 ds
−∞ se j=1

Inserting this formula into (5.39) and performing the integration over t first, we have
√  √ ǫ2 
−1 4 s !
Z ∞ − 1 1 − e Z n √
1 − 1s
cos θ j
vol( Bǫ ) = √ n p ( θ1 , . . . , θ n ) ∏ e 2 dθ
j
2π −∞ se −1 2 s [− π,π ]n j=1
√  √ ǫ2 
−1 4 s
Z ∞ − 1 1 − e
1
= √ n Dn (s)ds,
2π −∞ se −1 2 s
where
Z n √ cos θ j
−1s
Dn ( s) = p ( θ1 , . . . , θ n ) ∏ e 2 dθ j .
[− π,π ]n j=1

Since
 √ √ 2  √   √ 
 −1θi 
∏ e − e −1θ j  = det e −1(i−1)θk det e −1(i−1)θk
16i< j6n

is a product of two Vandermonde determinants where i, k = 1, . . . , n. The following fact will be

used.

Proposition 5.1 (Andréief’s identity). For two n × n matrices M (x) and N (x), defined by the following:
   
M1 ( x 1 ) M1 ( x 2 ) · · · M1 ( x n ) N1 ( x1 ) N1 ( x2 ) · · · N1 ( xn )
   
 M2 ( x 1 ) M2 ( x 2 ) · · · M2 ( x n )   N2 ( x1 ) N2 ( x2 ) · · · N2 ( xn ) 
   
M ( x) =  .. .. .. ..  , N ( x) =  .. .. .. .. 
 . . . .   . . . . 
   
M n ( x1 ) M n ( x2 ) · · · M n ( x n ) Nn ( x1 ) Nn ( x2 ) · · · Nn ( xn )

and a function w(·) such that the integral

Z b
Mi ( x) Nj ( x)w( x)dx
a

exists, then the following multiple integral can be evaluated as

Z Z n Z b

··· det( M (x)) det( N (x)) ∏ w( x j )dx j = det Mi (t) Nj (t)w(t)dt , (5.40)
∆ a,b j=1 a

where ∆a,b := {x = ( x1 , . . . , xn ) ∈ R n : b > x1 > x2 > · · · > xn > a}.

83
 By invoking this identity, we know that
Z Z  √   √  n √ cos θ j
1
Dn ( s) = ··· det e −1(i−1)θk det e− −1(i−1)θk ∏ e −1s 2 dθ j
(2π )n n! [− π,π ]n j=1
Z π √ √

1 − 1( i − j ) θ −1s cos2 θ
= det e e dθ .
(2π )n −π

In what follows, we need the Bessel function of the first kind. The definition is

Z π √ ∞  x 2j+n
1 −1( nθ − x sin θ ) (−1) j
Jn ( x) =
2π −π
e dθ = ∑ Γ( j + n + 1) j! 2 .
j=0

We can list some properties of this Bessel function:

Rπ √
−1( nθ + x sin θ ) dθ
Rπ √
−1(− nθ − x sin θ ) dθ
(i) −π e = −π e

1
Rπ √
−1( nθ − x sin θ ) dθ 1
Rπ √
−1(− nθ + x sin θ ) dθ
(ii) Jn ( x) = 2π −π e = 2π −π e

1
Rπ √
−1(− nθ − x sin θ ) dθ 1
Rπ √
−1( nθ + x sin θ ) dθ
(iii) J−n ( x) = 2π −π e = 2π −π e

We claim that

Z π √ √
−1( nθ + x cos θ ) −1 nπ
e dθ = 2πe 2 Jn ( x). (5.41)
−π

Indeed, since
Z π √ √ Z π √ √
−1( nθ + x cos θ ) −1 nπ −1 nπ
e dθ = e− 2 e −1( nθ + x sin θ )
dθ = e− 2 · 2π J−n ( x)
−π −π

and J−n ( x) = (−1)n Jn ( x), the desired claim is obtained.

Thus
Z π √ √ Z π √
− 1( i − j ) θ −1s cos2 θ −1[( i− j) θ + 2s cos θ ]
e e dθ = e dθ
−π −π
√ s
−1 i −2 j π
= 2πe Ji− j .
2
We now have

84
 1  √ i− j  s  1   s 
−1 2 π
Dn ( s) = det 2πe Ji− j = det Ji− j .
(2π )n 2 (2π )n 2

By the definition of Jn ( x), we have Jn (− x) = (−1)n Jn ( x), and furthermore,

Dn (−s) = Dn (s).

Therefore
√  √ ǫ2 
1
Z ∞ −1 1 − e −1 4 s
vol( Bǫ ) = √ n Dn (s)ds
2π −∞ se −1 2
s

 
ǫ2
1
Z ∞ sin ns
2 + sin 4 − n
2 s
  s 
= det Ji− j ds,
π 0 s 2
where we used Euler’s formula. In fact, we have
√  
√ 2 √ n ǫ2 n
−1 ǫ4 s −1 4 −2 s
1−e e − −1 2 s −e
√ =
se −1 n2 s s

  
  
n ǫ2 ǫ2
cos 2s − cos 4 − n2 s √ − sin n
2s − sin 4 − n
2 s
= + −1 .
s s
Finally, we get the volume of a metric ball in unitary group U (n):

  
Z ns ǫ2 n   s 
1 ∞ sin 2 + sin 4 − 2 s
vol( Bǫ ) = det Ji− j ds. (5.42)
π 0 s 2

6 Appendix I: Volumes of a sphere and a ball

The following result is very important in deriving the volumes of a sphere and a ball in F n where
F = R, C.

Proposition 6.1. It holds that

Z
π n n −1
δ(t − hψ|ψi)[dψ] = t ( t > 0) . (6.1)
Cn Γ(n)

85
 In particular,
Z
πn
δ(1 − hψ|ψi)[dψ] = . (6.2)
Cn Γ(n)
Proof. Recall that
n
[dψ] = [d(Re(ψ))][d(Im(ψ))] = ∏ dx j dy j
j=1

for |ψi = ∑nj=1 ψj | ji and ψj = x j + iy j ∈ R + iR ( j = 1, . . . , n). We now have

!
n n
δ(1 − hψ|ψi)[dψ] = δ 1 − ∑ ( x2j + y2j ) ∏ dxj dyj .
j=1 j=1

Let !
Z n n
Γ(n)
F ( t) = δ t− ∑ ( x2j + y2j ) ∏ dxj dyj .
πn j=1 j=1

Then the Laplace transform of F (t) is given by:

"
Z Z ∞
! #
n n
Γ(n) 2 2 − st
L ( F )(s) =
πn 0
δ t − ∑ j j
( x + y ) e dt ∏ dxj dyj
j=1 j=1
"Z # "Z #
Γ(n) ∞ n ∞ n
− sx2j − sy2
= N ∏ e dx j ∏ e j dy j = s−n Γ(n)
π − ∞ j=1 − ∞ j=1

implying via the inverse Laplace transform that F (t) = tn−1 . This indicates that
Z
π n n −1
δ(t − hψ|ψi)[dψ] = t .
Γ(n)
In particular, for t = 1, we get the desired result.

In the following, we will give a method via Dirac delta function to calculate the volume of
n-dimensional ball in R n . Moreover, based on this formula, we then give the surface of (n − 1)-
dimensional sphere in R n . Denote the ball of radius R in R n as
( )
n
n
B n ( R) := ( x1 , . . . , x n ) ∈ R : ∑ x2j 6R 2

j=1

and the sphere of radius R in R n as

( )
n
S n −1 ( R ) : = ( x1 , . . . , x n ) ∈ R n : ∑ x2j = R2 .
j=1

Now the volume of the ball of radius R is given by

Z R2 Z
!
n n
vol(B n ( R)) = dt δ t−∑ x2j ∏ dxj . (6.3)
0 Rn j=1 j=1

86
Denote !
Z n n
F ( t) = δ t−∑ x2j ∏ dxj .
j=1 j=1

Then its Laplace transform is given by

n n
L ( F )(s) = π 2 s− 2 .

This implies that

n
π 2 n −1
F ( t) =
t2 .
Γ n2
Thus
Z R2 n n
π 2 Rn π2
vol(B n ( R)) = F (t)dt = n

n = n

Rn . (6.4)
0 Γ 2 2 Γ 2 +1

By differentiating the above formula with respect to the radius R, we get that
  dvol(B n ( R)) π2 2π 2 n−1
n n
n −1
nRn−1 =
vol S ( R) = = 2R · F ( R2 ) = n

R . (6.5)
dR Γ 2 +1 Γ n2

From this, we see that

Z
!
  n n
n −1 2
vol S ( R) = 2R · δ R −∑ x2j ∏ dxj , (6.6)
j=1 j=1

where
Z
! n
n n
π2
δ R2 − ∑ x2j ∏ dxj =
R n −2 . (6.7)
j=1 j=1 Γ n2

In particular, for R = 1, we see that

Z
! Z
  n n
n −1
vol S ( 1) = 2 δ 1− ∑ x2j ∏ dxj = 2 δ (1 − hu|ui) [du], (6.8)
j=1 j=1 Rn

where
Z
! Z n
n n
π2
δ 1− ∑ x2j ∏ dxj = Rn δ (1 − hu|ui) [du] = Γ n
. (6.9)
j=1 j=1 2

We also see that the volume of the ball of radius R in C n ∼

= R2n is given by
Z R2 Z
πn
vol(B n ( R, C )) = dt δ (t − hψ|ψi) [dψ] = R2n . (6.10)
0 Cn Γ ( n + 1)
The surface of the sphere of radius R in C n is given by
  Z
2n −1

2π n 2n−1
vol S ( R, C ) = 2R · δ R2 − hψ|ψi [dψ] = R , (6.11)
Cn Γ (n)

87
where
Z
π n 2n−2
δ R2 − hψ|ψi [dψ] = R . (6.12)
Cn Γ (n )
Recall a property of Dirac delta function:

1
δ a2 − x 2 = (δ( a − x) + δ( a + x)). (6.13)
2 | a|
From this, we see that

  1
δ R2 − h ψ |ψ i = δ R2 − k ψ k2 = (δ( R − k ψ k) + δ( R + k ψ k)). (6.14)
2R
Clearly δ( R + k ψ k) = 0 since R + k ψ k > 0. Therefore we obtain

2R · δ R2 − hψ|ψi = δ( R − k ψ k). (6.15)

In summary, we have the following result.

Proposition 6.2. The surface of sphere of radius R in C n can be represented by

  Z 2π n 2n−1
vol S2n−1 ( R, C ) = δ ( R − k ψ k) [dψ] = R . (6.16)
Cn Γ (n)

We remark here that

Z Z ∞ Z
f ( x)[dx] = dr [du]rn−1 f (r · u)δ(1 − k u k). (6.17)
Rn 0 Rn

In particular, if f is independent of u, i.e., f (ru) = f (r), then we have that

Z Z ∞ Z n Z ∞
2π 2
f ( x)[dx] = rn−1 f (r)dr × δ(1 − k u k)[du] =
rn−1 f (r)dr. (6.18)
Rn 0 Rn Γ n2 0

Indeed, For a Gaussian random vector ω = [ω1 , . . . , ωn ]T ∈ R n with i.i.d. standard normal
random variable ω j ∼ N (0, 1), we have
Z  
− n2 1 2
1 = (2π ) exp − k ω k [dω ]. (6.19)
Rn 2

88
By using polar coordinate of ω, we see ω = k ω k · u = r · u, where k u k = 1. Thus

[dω ] = dr × δ(r − k ω k)[dω ] = rn−1 dr × δ(1 − k u k)[du], (6.20)

where

δ(1 − k u k)[du] = 2δ(1 − hu|ui)[du].

Indeed, the truth is checked as follows,

Z  
− n2 1 2
(2π ) exp − k ω k [dω ]
Rn 2
Z ∞   Z
n 1
= (2π )− 2 exp − r2 rn−1 dr δ(1 − k u k)[du]
0 2 Rn
n n
n Z
= (2π )− 2 2 2 −1 Γ δ(1 − k u k)[du]
2 Rn

Z
n
Γ n2 Γ n2 2π 2
= n δ(1 − k u k)[du] = n
= 1.
2π 2 R n 2π 2 Γ n2

7 Appendix II: Some useful facts

7.1 Matrices with simple eigenvalues form open dense sets of full measure

The present subsection is written based on the Book by Deift and Gioev [5].
Let Hn (C ) be the set of all n × n Hermitian matrices with simple spectrum, i.e. the multi-
plicity of each eigenvalue are just one. Next we show that Hn (C ) is an open and dense set of
2
full measure (i.e. the Lebesgue measure of the complement is vanished) in H (C n ) ≃ R n , the
2
set of all n × n Hermitian matrices, that is, R n \Hn (C ) is a set of zero-measure in the sense of
Lebesgue measure.
Assume that M e is an arbitrary Hermitian matrix in Hn (C ). with simple spectrum µ1 <
e have simple
· · · < µn , then by standard perturbation theory, all matrices in a neighborhood of M
e ∈ H (C n ) with eigenvalues {h j : j = 1, . . . , n}, then by spectral theorem,
spectrum. Moreover if H
He = UΛ
e Ue ∗ for some unitary U
e and Λ = diag(h1 , . . . , hn ). Now we can always find ǫj arbitrarily
small for all j so that h j + ǫj are distinct for all j. Thus

e ǫ := Udiag
H e e∗
(h1 + ǫ1 , . . . , hn + ǫn )U

e The above two facts show

is a Hermitian matrix with simple spectrum, arbitrarily close to H.
that Hn (C ) is open and dense. In order to show that Hn (C ) is of full measure, consider the
discriminant:
∆( λ) : = ∏ ( λ j − λ i )2 .
16i< j6n

89
By the fundamental theorem of symmetric functions, ∆ is a polynomial function of the ele-
mentary symmetric functions of the λ j ’s, and hence a polynomial function of the n2 entries
e11 , . . . , H
H e nn , Re( H
e ij ), Im( H
e ij )(i < j) of H.
e Now if Hn (C ) were not of full measure, then ∆ would
2 2
vanish on a set of positive measure in R n . It follows that ∆ ≡ 0 because ∆ is polynomial in R n .
Let H = diag(1, 2, . . . , n) is a Hermitian matrix with distinct spectrum and ∆( H ) 6= 0. This
gives a contradiction and so Hn (C ) is of full measure in H (C n ).
Similarly, the set Sn (R ) of all n × n real symmetric matrices with simple spectrum is an open
and dense set of full measure in the set S (R n ) of all real symmetric matrices.
From the above discussion, we conclude that the set of all density matrices with distinct
positive eigenvalues is an open and dense set of full measure in the set of all density matrices.
We begin by considering real symmetric matrices of size n, which is the simplest case. Because
Sn (R ) is of full measure, it is sufficient for the purpose of integration to restrict our attention
to matrices M ∈ Sn (R ). Let µ1 < · · · < µn denote the eigenvalues of M. By spectral theorem,
M = UΛU T , where Λ = diag(µ1 , . . . , µn ). Clearly the columns of U are defined only up to
multiplication by ±1, and so the map M 7→ (Λ, U ):

Sn (R ) ∋ M 7→ (Λ, U ) ∈ R n↑ × O(n)

is not well-defined, where

R n↑ = {(µ1 , . . . , µn ) ∈ R n : µ1 < · · · < µn } , O(n) = n × n orthogonal group.

We consider instead the map

b ) ∈ R n↑ × (O(n)/K1 )
φ1 : Sn (R ) ∋ M 7→ (Λ, U (7.1)

where K1 is the closed subgroup of O(n) containing 2n elements of the form diag(±1, . . . , ±1),
O(n)/K1 is the homogeneous manifold obtained by factoring O(n) by K1, and U b = UK1 is the
coset containing U. The map φ1 is now clearly well-defined.
The differentiable structure on O(n)/K1 is described in the following general result about
homogeneous manifolds:

Proposition 7.1. Let K be a closed subgroup of a Lie group G and let G/K be the set { gK : g ∈ G } of
left cosets module K. Let π : G → G/K denote the natural projection π ( g) = gK. Then G/K has a
unique manifold structure such that π is C ∞ and there exist local smooth sections of G/K in G, i.e., if
gK ∈ G/K, there is a neighborhood W of gK and a C ∞ map τ : W → G such that π ◦ τ = id.

In other words, for each gK, it is possible to choose a g′ ∈ gK ⊂ G such that the map
gK 7→ g′ ≡ τ ( gK )) is locally defined and smooth.

90
 For example, if G = R × , the multiplication group of nonzero real numbers, and if K is the
subgroup {±1}, then G/K ∼ = { x > 0} and π ( a) = | a | ∼
= {± a}. If gK = {± a} for some a > 0,
then τ ({± a}) = a. Also, if G = O(2) and K1 consists of four elements of the form
" # " #
±1 0 u11 u12
, g= ∈ O(2),
0 ±1 u21 u22

then (" # " # " # " #)

u11 u12 −u11 u12 u11 −u12 −u11 −u12
gK1 = , , , .
u21 u22 −u21 u22 u21 −u22 −u21 −u22
Since u211 + u221 = 1 = u212 + u222 , each column of g contains at least one nonzero number: For
example, suppose u21 and u22 are nonzero; then there exists a unique g′ ∈ gK such that the
elements in the second row of g′ are positive. The same is true for all g′′ K close to gK. Then
g′′ K 7→ g′′ is the desired (local map) τ. We can generalize the above construction on O(n)/K1 .
We will prove the following result:

Proposition 7.2. φ1 is a diffeomorphism from Sn (R ) onto R n↑ × (O(n)/K1 ).

Proof. Note that (here and below we always speak of the real dimensions)
  N ( N − 1) N ( N + 1)
dim R n↑ × (O(n)/K1 ) = N + = (7.2)
2 2
as it should.
b ) lies in R n × (O(n)/K1 ) →
Define the map φ̂1 : R n↑ × (O(n)/K1 ) → Sn (R ) as follows: If (Λ, U ↑
Sn (R ), then

b ) = UΛU T
φ̂1 (Λ, U (7.3)

b If U ′ is another such matrix, then U ′ = Uh for some h ∈ K1

where U is any matrix in the coset U.
and so
U ′ Λ(U ′ )T = UhΛhT U T = UΛU T .

Hence φ̂1 is well-defined. We will show that

φ1 ◦ φ̂1 = idR n↑ ×(O(n)/K1) (7.4)

and

φ̂1 ◦ φ1 = idSn (R ) . (7.5)

Indeed,
  
φ1 φ̂1 b
Λ, U b
= φ1 (UΛU T ) , U ∈ U (7.6)
 
= (Λ, UK1 ) = Λ, Ub . (7.7)

91
Conversely, if M = UΛU T ∈ Sn (R ), then
 
b = UK1 = UΛU T = M.
φ̂1 (φ1 ( M )) = φ̂1 Λ, U (7.8)

This proves Eq. (7.4) and Eq. (7.5). In order to prove that φ1 is a diffeomorphism, it suffices to
show that φ1 and φ̂1 are smooth.
The smoothness of φ1 follows from perturbation theory: Fix M0 = U0 Λ0 U0T ∈ Sn (R ), where
Λ0 = diag(µ01 , . . . , µ0n ) for µ01 < · · · < µ0n , and U0 ∈ O(n). Then for M near M0 , M ∈
Sn (R ), the eigenvalues of M, µ1 ( M ) < · · · < µn ( M ), are smooth functions of M. Moreover, the
associated eigenvectors u j ( M ),

Mu j ( M ) = λ j ( M )u j ( M ), 1 6 j 6 n,

can be chosen orthogonal


ui ( M ), u j ( M ) = δij , (1 6 i, j 6 n),

and smooth in M. Indeed, for any j with 1 6 j 6 n, let Pj be the orthogonal projection
I
1 1
Pj ( M ) = ds (7.9)
Γj s − M
2πi
 
where Γ j is a small circle of radius ǫ around µ0j ,  µ0i − µ0j  > ǫ for i 6= j. Then, where u j ( M0 ) is
the j-th column of M0 ,

Pj ( M )u j ( M0 )
u j ( M ) = q
 , 1 6 j 6 n,
u j ( M0 ), Pj ( M )u j ( M0 )

is the desired eigenvector of M.

Set U ( M ) = [u1 ( M ), . . . , un ( M )] ∈ O(n). Then M 7→ (Λ( M ), U ( M )) is smooth and hence

φ1 ( M ) = (Λ( M ), U ( M )) ≡ π (U ( M ))

is smooth, as claimed.  
b 0 ∈ R n × (O(n)/K1 ) and let τ be the lifting
Finally, we show that φ̂1 is smooth. Fix Λ0 , U ↑
b
map from some neighborhood W of U0 to O(n). Now for all U b ∈ W, τ (U
b) ∈ U
b by Proposition 7.1.
Hence
 
b = τ (U
φ̂1 Λ, U b )Λτ (U
b )T (7.10)
 
b 0 , and hence everywhere on R n × (O(n)/K1 ).
from which it is clear that φ̂1 is smooth near Λ0 , U ↑
This completes the proof.

92
 Now consider Hn (C ). The calculations are similar to Sn (R ). Define the map
 
φ2 : Hn (C ) ∋ M 7→ Λ, U b ∈ R n↑ × (U (n)/K2 ). (7.11)


Here K2 is the closed subgroup of U (n) given by T × · · · × T = diag eiθ1 , . . . , eiθn : θ j ∈ R ,
Λ = diag(λ1 , . . . , λn ), λ1 < · · · < λn as before, and

M = UΛU ∗ , U
b = π (U ),

where π : U (n) → U (n)/K2 is the natural projection U 7→ π (U ) = UK2 of the unitary

group U (n) onto the homogeneous manifold U (n)/K2 . Let τ : U (n)/K2 → U (n) denote the
(locally defined) lifting map as above. As before, φ2 is a well-defined map that is one-to-
one from Hn (C ) to U (n) → U (n)/K2 . Note that because dim(U (n)) = n2 , dim(U (n)/K2 ) =
dim(U (n)) − dim(K2 ) = n2 − n, as it should.
As before, the proof of the following result is similar to that of Proposition 7.2.

Proposition 7.3. φ2 is a diffeomorphism from Hn (C ) onto R n↑ × (U (n)/K2 ).

7.2 Results related to orthogonal groups

Proposition 7.4. Let Y, X, T ∈ R n×n , where Y and T are nonsingular, X is skew symmetric and T is
lower triangular of independent real entries yij ’s, xij ’s and tij ’s respectively. Let t jj > 0, j = 1, . . . , n −
1, −∞ < tnn < ∞; −∞ < t jk < ∞, j > k; −∞ < x jk < ∞, j < k or −∞ < t jk < ∞, j > k and
the first row entries, except the first one, of (1 n + X )−1 are negative. Then the unique representation
 
Y = T 2(1 n + X )−1 − 1 n = T (1 n − X )(1 n + X )−1 implies that
!
n  
n − j
[dY ] = 2 n ( n −1) /2
∏  tjj  (det(1 n + X )) −(n−1) [dT ][dX ]. (7.12)
j=1

Proof. Take the differentials to get

   
dY = dT · 2(1 n + X )−1 − 1 n + T · −2(1 n + X )−1 · dX · (1 n + X )−1
   
1
= dT · 2(1 n + X )−1 − 1 n + T · − (1 n + Z ) · dX · (1 n + Z )
2
1
= dT · Z − T (1 n + Z ) · dX · (1 n + Z ).
2
where Z = 2(1 n + X )−1 − 1 n . Then we have
1
T −1 · dY · ZT = T −1 · dT − (1 n + Z ) · dX · (1 n + ZT ).
2
The Jacobian of the transformation of T and X going to Y is equal to the Jacobian of dT, dX going
to dY. Now treat dT, dX, dY as variables and everything else as constants. Let

dU = T −1 · dY · ZT , dV = T −1 · dT, dW = (1 n + Z ) · dX · (1 n + ZT ).

93
Thus

[dU ] = det( T )−n det( ZT )n [dY ] = det( T )−n [dY ]

!
n  
n
=⇒ [dY ] = det( T )n [dU ] = ∏  t jj  [dU ].
j=1

Note that det( ZT ) = ±1 since Z is orthogonal. Since X is skew symmetric, one has

[dW ] = det(1 n + Z )n−1 [dX ] = 2n(n−1) det(1 n + X )−(n−1) [dX ].

One also has

!
n − j
[dV ] = ∏  tjj  [dT ].
j=1

Now we see that

1 1
dU = dV − dW =⇒ U = V − W.
2 2
Let U = [uij ], V = [vij ], W = [wij ]. Then since T is lower triangular tij = 0, i < j and thus V is
lower triangular, and since X is skew symmetric x jj = 0 for all j and xij=− x ji for i 6= j and thus W
is skew symmetric. Thus we have
1
uii = vii , uij = − wij , i < j
2
1
uij = vij + wij , i > j.
2
Take the u-variables in the order uii , i = 1, . . . , n; uij , i < j; uij , i > j and vii , i = 1, . . . , n; vij , i <
j; vij , i > j. Then the Jacobian matrix is of the following form:
 
1n 0 0


 0 − 12 1 (n2 ) 0 


1
0 2 1(2)
n 1 (2 )
n

n(n−1)/2
and the determinant, in absolute value, is 12 . That is,
 n(n−1)/2
1
[dU ] = [dV ][dW ].
2
Now substitute for [dU ], [dW ] and [dV ], respectively to obtain the result.

Proposition 7.5. Let Y, X, D ∈ R n×n be of independent real entries, where Y is symmetric with distinct
and nonzero eigenvalues, X is skew symmetric with the entries of the first row of (1 n + X )−1 , except
the first entry, negative and D = diag(λ1 , . . . , λn ), with λ1 > · · · > λn . Then, excluding the sign,
   
Y = 2(1 n + X )−1 − 1 n D 2(1 n + X )−1 − 1 n implies that
!
−( n−1)  
[dY ] = 2 n ( n −1) /2
(det(1 n + X )) ∏  λi − λ j  [dX ][dD]. (7.13)
i< j

94
Proof. Let Z = 2(1 n + X )−1 − 1 n . Take the differentials and reduce to get
1 1
ZT · dY · Z = − (1 n + ZT ) · dX · (1 n + Z ) · D + dD + D · (1 n + ZT ) · dX · (1 n + Z ).
2 2
Put

dU = ZT · dY · Z, dW = dD, dV = (1 n + ZT )dX (1 n + Z ).

Thus
1 1
dU = − dV · D + D · dV + dW.
2 2
But since dY is symmetric and Z is orthogonal, [dU ] = det( Z )n+1 [dY ] = [dY ], excluding the
sign. Clearly [dW ] = [dD ]. Since X is skew symmetric we have

dV = det(1 n + Z )n−1 [dX ] = 2n(n−1) det(1 n + X )−(n−1) [dX ].

We see that

duii = dwii ,
1
duij = (λi − λ j )dvij , i < j.
2
Take the u-variables in the order uii , i = 1, . . . , n; uij , i < j and the w and v-variables in the order
wii , i = 1, . . . , n; vij , i < j. Then the matrix of partial derivatives is of the following form:
" #
1 0
,
0 M

where M is a diagonal matrix with the diagonal elements 21 (λi − λ j ), i < j. There are n(n − 1)/2
 
elements. Hence the determinant of the above matrix, in absolute value, is 2−n(n−1)/2 ∏i< j  λi − λ j .
That is,
!
 
[dU ] = 2−n(n−1)/2 ∏  λi − λ j  [dV ][dD ].
i< j

Hence
!
 
[dY ] = 2n(n−1)/2 det(1 n + X )−(n−1) ∏  λi − λ j  [dX ][dD ].
i< j

Remark 7.6. When integrating over the skew symmetric matrix X using the transformation in
Proposition 7.5, under the unique choice for Z = 2(1 n + X )−1 − 1 n , observe that
Z n2
n ( n −1) /2 −( n−1) π 2
2 det(1 n + X ) [dX ] = n

. (7.14)
X Γn 2
Note that the λ j ’s are to be integrated out over ∞ > λ1 > · · · > λn and X over a unique choice
of Z.

95
7.3 Results related to unitary groups

When Xe is skew hermitian, that is, X e ∗ = − X,

e it is not difficult to show that 1 ± X e are both
e = 2( 1 + X
nonsingular and Z e )−1 − 1 is unitary, that is, Z
eZe ∗ = 1. This property will be made use
of in the first result that will be discussed here. Also note that

e ) −1 − 1 = (1 + X
2( 1 + X e ) −1 (1 − X
e ) = (1 − X e ) −1 .
e )(1 + X

When X e is skew hermitian and of functionally independent complex variables, then there
n ( n − 1) e Let T
e be a lower triangular matrix of functionally
are p + 2 2 = n2 real variables in X.
independent complex variables with the diagonal elements being real. Then there are n2 real
e also. Thus combined, there are 2n2 real variables in T
variables in T e and X.
e It can be shown that
 
Ye=T e 2( 1 + Xe ) −1 − 1

can produce a one-to-one transformation when the t jj ’s are real and positive.

e=T
Y eZ,
e

eZ
where Z e ∗ = 1, T
e = [et jk ], e
t jj = t jj > 0, j = 1, . . . , n. Then
n
eY
Y e∗ = T
eTe∗ =⇒ t211 = ∑ | ye1k |2 .
k=1

Note that when t11 is real and positive it is uniquely determined in terms of Y. e Now consider the
first row elements of TeT
e∗ that is t2 , t11e
t21 , . . . , t11etn1 . Hence et21 , . . . , e
tn1 , that is, the first column
11
e is uniquely determined in terms of Y.
of T e Now consider the second row of T eT
e∗ and so on.
Thus T e But Z
e is uniquely determined in terms of Y. e=T e and hence Z,
e −1 Y e thereby X
e is uniquely
determined in terms of Ye with no additional restrictions imposed on the elements of Z.
e
In this chapter the Jacobians will also be written ignoring the sign as in the previous chapters.

e X
Proposition 7.7. Let Y, e and Te = [et jk ] be n × n matrices of functionally independent complex variables
e is nonsingular, X
where Y e is skew hermitian and T e is lower triangular with real and positive diagonal
elements. Ignoring the sign, if
 
e=T
Y e + 1 ) −1 − 1 = T
e 2( X e (1 − X e ) −1 ,
e )(1 + X

then

!
n  −n
n2
e] = 2 · 2( n − j)+1  e ))  · [dX
e )(1 − X e ][dT
e ].
[dY ∏ tjj ·  det((1 + X (7.15)
j=1

96
  
e=T
Proof. Taking differentials in Y e 2( 1 + X
e )−1 − 1 , one has
   
e=T
dY e ) −1 · d X
e − 2( 1 + X e ) −1 + d T
e · (1 + X e ) −1 − 1 .
e · 2( 1 + X

Let

e = 2( 1 + X
Z e ) −1 = 1 (1 + Z
e )−1 − 1 =⇒ (1 + X e)
2
eZ
and observe that Z e ∗ = 1. Then

e−1 · dY
T e ∗ = − 1 (1 + Z
e·Z e ) · dX e∗ ) + T
e · (1 + Z e−1 · dT.
e (7.16)
2
e = (1 + Z
Let dW e ) · dX
e · (1 + Ze ∗ ). Then
 n  −n
e ] =  det((1 + Z
[dW e )(1 + Z e ] = 22n2 ·  det((1 + X
e ∗ ))  · [dX e ))  · [dX
e )(1 − X e ]. (7.17)

e=T
Let dU e−1 · dT,
e then
!
n
e] = −(2j−1) e ].
[ dU ∏ t jj · [d T (7.18)
j=1

e=T
Let dV e·Z
e−1 · dY e ∗ , then
 −n
e ] =  det( T
[ dV eTe∗ )  · [dY
e ]. (7.19)

Eq. (7.16) reduces to

e = − 1 dW
dV e + dU.
e (7.20)
2
e is skew hermitian. Denote
Note that dW
√
e = [vejk ] = [v(1) ] +
V −1[v jk ],
( 2)
jk
( 1) √ ( 2)
e = [w
W e jk ] = [w jk ] + −1[w jk ],
√
e = [uejk ] = [u(1) ] + −1[u(2) ],
U jk jk

(m) (m) (m) e = −1We + U.

e
where v jk , w jk , u jk , m = 1, 2 are all real. Then from Eq. (7.20), we see that V 2
Thus
(m) 1 (m) (m)
v jk = w + u jk , j > k, m = 1, 2,
2 kj
(m) 1 (m)
v jk = − w jk , j < k, m = 1, 2,
2
( 1) ( 1)
v jj = u jj ,
( 2) 1 ( 2)
v jj = − w jj .
2

97
The matrices of partial derivatives are the following:
 
( 1) ( 2) ( 1) ( 2) ( 1) ( 2)
∂ v jj , v jj ; v jk , v jk ( j < k); v jk , v jk ( j > k)
 
( 1) ( 2) ( 1) ( 2) ( 1) ( 2)
∂ u jj , w jj ; w jk , w jk ( j < k); u jk , u jk ( j > k)
 
1n 0 0 0 0 0
 
 0 − 1 1n 0 0 0 0 
 2 
 1 n 
 0 0 − 1
2 (2 ) 0 0 0 
=  

 0 0 0 − 21 1 (n2 ) 0 0 
 
 0 0 1 n
0 1 (n2 ) 0 
 2 1 (2 ) 
1 n
0 0 0 2 1 (2 ) 0 1 (n2 )

where
 
( 1)
∂v jj
A11 =  ( 1)
, j = 1, . . . , n = 1 n ,
∂u jj
 
( 2)
∂v jj 1
A22 =  ( 2)
, j = 1, . . . , n = − 1 n
∂w jj 2

and
 
( 1)
∂v jk
A33 =  , j < k  = − 1 1 (n) ,
( 1) 2 2
∂w jk
 
( 2)
∂v jk 1
A44 =  (2) , j < k = − 1 (n2 )
∂w jk 2
 
( 1)
∂v jk
A55 =  , j > k = 1 (n2 ) ,
( 1)
∂u jk
 
( 2)
∂v jk
A66 =  (2) , j > k = 1 (n2 ) .
∂u jk
   
( 1) ( 2)
∂v jk 1 ∂v jk
1
A53 =  ( 1)
: j > k = 1 (n2 ) , A64 =  (2) : j > k = 1 (n2 ) .
∂wkj 2 ∂wkj 2

The determinants of A11 , . . . , A66 contribute towards the Jacobian and the product of the deter-
2
minants, in absolute value, is 2−n(n−1)−n = 2−n . Without going through the above procedure
one may note from (7.16) that since dX e has n2 real variables, multiplication by 1 produces the
2

98
 2
factor 2−n in the Jacobian. From (7.17),(7.18), (7.19) and (7.20) we have
! !
n n
2
e] =
[dY ∏ tjj · [dVe ] = ∏ tjj · 2−n · [dW
2n 2n e ][dU
e]
j=1 j=1
! !
n  −n n
2n2 − n2  e ))  ·
e )(1 − X −(2j−1) e ][dT
e]
= ∏ t2n
jj ·2 · det((1 + X ∏ t jj · [d X
j=1 j=1
!
n  −n
2 2( n − j)+1  e ))  · [dX
e )(1 − X e ][dT
e ].
= 2n ∏ tjj ·  det((1 + X
j=1

This completes the proof.

e ∈ C n×n skew hermitian matrix of independent complex variables. Then

Example 7.8. Let X
show that

Z  −n 2
e ]  det((1 + X e ))  =
e )(1 − X πn
[d X .
e
X 2n ( n − 1 ) e
Γn (n)

e = [yejk ] be a n × n matrix of independent complex variables. Consider the

Indeed, let Y
integral
Z n Z +∞ 2
e]e− Tr (YeYe∗ ) =
[dY ∏ e−| yejk | de
y jk = π n .
2
(7.21)
e
Y j,k=1 − ∞
 
Now consider a transformation Y e=T e 2( 1 + X
e )−1 − 1 , where T
e = [et jk ] is lower triangular with
t jj ’s realand positive e
  and no restrictions on X other than that it is skew hermitian. Then since
Tr Y eY
e∗ = Tr T eTe∗ , from Proposition 7.7 and Eq. (7.21), we have

Z
!
n  −n
n2 e ][dT
e ]e eT
− Tr ( T e∗) n2 2( n − j)+1  e )) 
e )(1 − X
π =
eX
T, e
[d X ·2 · ∏ t jj · det((1 + X
j=1
Z
! Z
n  −n
= e ]e
[d T − Tr ( Te Te ∗ ) n2
·2 · ∏
2( n − j)+1
t jj × e ] det((1 + X
[d X e )) 
e )(1 − X
e
T e
X
j=1

Note that
!
n
 2
eT
− Tr ( T e∗ )
e = exp − ∑ t2jj − ∑ et jk  .
j=1 j> k

But
Z +∞ Z +∞
− |et jk |
2
2( n − j)+1 − t2jj 1
e det jk = π and t jj e dt jj = Γ ( n − j + 1) .
−∞ 0 2

99
Hence
Z
!
n
− Tr ( Te T
e∗) 2( n − j)+1
e]e
[d T ∏ t jj = 2− n e
Γ n ( n ).
e
T j=1

Substituting this the result follows.

Remark 7.9. When a skew hermitian matrix X e is used to parameterize a unitary matrix such as
e in Proposition 7.7, can we evaluate the Jacobian by direct integration? This will be examined
Z
here. Let
Z  −n
In : = e ]  det((1 + X
[d X e ))  .
e )(1 − X
e
X

e as follows:
Partition 1 + X
" #
1 + xe11 e12
X
e=
1+X ,
−Xe∗ 1+Xe1
12

where Xe12 represents the first row of 1 + Xe excluding the first element 1 + xe11 , and 1 + X
e 1 is
e by deleting the first row and the first column. Note that
obtained from 1 + X
 
e ) = det(1 + X
det(1 + X e1 ) 1 + xe11 + X e 1 ) −1 X
e12 (1 + X ∗
e12 .

Similarly,
" #
e12
1 − xe11 − X
e=
1−X
e∗
X 1−X e1
12

and
 
e ) = det(1 − X
det(1 − X e1 ) 1 − xe11 + X e 1 ) −1 X
e12 (1 + X ∗
e12 .

e1 ) let U
For fixed (1 + X e 12 := X
e 12 (1 + X
e1 )−1 , then
  −1
e  e e  e12 ]
[dU12 ] = det((1 + X1 )(1 − X1 ))  · [dX

e ∗ = −X
and observing that X e1 we have
1

X e 1 ) −1 X
e 12 (1 + X ∗
e12 e 12 (1 − X
=U e 1 )U ∗
e 12 .

e be a unitary matrix such that

Let Q
√ √
e∗ X
Q e = diag( −1λ1 , . . . , −1λn−1 )
e1 Q

where λ1 , . . . , λn−1 are real. Let

e12 = U
V e = [ve1 , . . . , ven−1 ].
e 12 Q

100
Then
√ √
e 12 (1 − X
U e 1 )U ∗
e 12 = (1 − −1λ1 ) | ve1 |2 + · · · + (1 − −1λn−1 ) | ven−1 |2

and
√
1 + xe11 + X e 1 ) −1 X
e12 (1 + X ∗
e12 = a− −1b

where

a = 1 + | ve1 |2 + · · · + | ven−1 |2
( 2)
b = − x11 + λ1 | ve1 |2 + · · · + λn−1 | ven−1 |2
√ ( 2) ( 2)
observing that xe11 is purely imaginary, that is, xe11 = −1x11 , where x11 is real. Thus
 −n
 e )) 
e )(1 − X
det((1 + X

yields the factor

√ √
[( a − −1b)( a + −1b)]−n = ( a2 + b2 )−n .

So
Z Z Z  √ √ −n
In = det(1 + X e1 )( a − −1b) det(1 − X e1 )( a + −1b) [d Xe1 ][dX
e12 ]de
x11
e1 X
X e 12 xe11
Z Z Z    −n
 
=  det ( 1 + e
X 1 )( 1 − e
X 1 )  ( a2 + b 2 ) − n [d X
e1 ][dX
e12 ]de
x11 .
e1
X e 12
X xe11

Since   −1
e12 ] = [dU
[ dV e 12 ] =  det (1 + X
e 12 ] and [dU e1 )  [dX
e1 )(1 − X e12 ],

it follows that   
e12 ] = det (1 + X
[d X e1 )  [dV
e1 )(1 − X e12 ].

Based on this, we have

Z Z Z    −(n−1)
 
In =  det ( 1 + e
X 1 )( 1 − e
X 1 )  ( a2 + b 2 ) − n [d X
e1 ][dV
e12 ]de
x11 .
e1
X e12
V xe11

Then
Z Z
( 2)
In = In −1 (2)
( a 2 + b 2 ) − n [ dV
e12 ]dx
11
e12
V x11
Z Z  −n !
e12 ] a−2n b2 ( 2)
= In −1 [ dV 1+ 2 dx11 .
e12
V
(2)
x11 a

101
 ( 2) ( 2) ( 2)
Consider the integral over x11 , −∞ < x11 < +∞. Change x11 to b and then to c = b/a. Then
Z  −n Z  −n Z +∞
b2 ( 2) b2
1 + dx 11 = 1 + db = a (1 + c2 )− p dc
(2)
x11 a2 b a2 c=− ∞
Z ∞ 1

1


Γ Γ n −
= 2a (1 + c2 )−n dc = a 2 2
:= k,
0 Γ(n)
by evaluating using a type-2 beta integral after transforming u = c2 . Hence
Z
In = kIn−1 e12 ] a−(2n−1)
[ dV
e12
V
Z +∞ Z +∞  −(2n−1)
= kIn−1 ··· 1 + | ve1 |2 + · · · + | ven−1 |2 v1 · · · de
de vn −1 .
−∞ −∞
( 1) √ ( 2) ( 1)
For evaluating the integral use the polar coordinates. Let vej = v j + −1v j , where v j and
( 2)
vj are real. Let

v(1) = r j cos θ j ,
j
0 6 r j < ∞, 0 6 θ j 6 2π.
v 2) = r sin θ ,
(
j j j

Then denoting the multiple integral by In−1 , we have

Z +∞ Z +∞  −(2n−1)
I n −1 = ··· 1 + | ve1 |2 + · · · + | ven−1 |2 v1 · · · de
de vn −1
−∞ −∞
Z +∞ Z +∞
−(2n−1)
= (2π )n−1 ··· r1 · · · rn−1 1 + r12 + · · · + r2n−1 dr1 · · · drn−1 .
r1 =0 r n −1 = 0

Evaluating this by a Dirichlet integral we have

Γ(n)
I n −1 = π n −1 for n > 2.
Γ(2n − 1)
Hence for n > 2,


n −1
√ Γ(n)Γ n − 21
In = In −1 π π .
Γ(2n − 1)Γ(n)
By using the duplication formula for gamma functions
 
√ 2n−2 1
Γ(2n − 1) = π2 Γ n− Γ ( n ).
2
Hence
πn
In = In −1 .
22n−2 Γ(n)
Repeating this process we have
2
πn π n −1 π πn
In = 2n−2 · · · = .
2 Γ(n) 22(n−1)−2 Γ( p − 1) 22−2 Γ(1) 2n ( n − 1 ) e
Γn (n)
This is what we obtained in Example 7.8.

102
 e in terms of a skew
Next we consider a representation of a hermitian positive definite matrix Y
e and a diagonal matrix D such that
hermitian matrix X
   ∗
e ) − 1 − 1 D 2( 1 + X
e = 2( 1 + X
Y e ) −1 − 1

where D = diag(λ1 , . . . , λ p ) with the λ j ’s real distinct and positive, and the first row elements of
e )−1 real and of specified signs, which amounts to require 2(1 + X
(1 + X e ) − 1 − 1 ∈ U ( n ) / U ( 1) × n .
In this case it can be shown that the transformation is unique. Note that

e = ZD
Y e Ze ∗ = λ1 Z e1∗ + · · · + λ p Z
e1 Z en∗
en Z

it indicates that

e − λj 1)Z
(Y ej = 0, j = 1, . . . , n

where Z e1 , . . . , Z
en are the columns of Z e such that h Z
ej , Z
ek i = δjk . Since λ1 , . . . , λn are the eigen-
values of Y, e which are assumed to be real distinct and positive, D is uniquely determined in
terms of Y.e Note that Z ej is an eigenvector corresponding to λ j such that h Zej , Z ej i = 1, j = 1, . . . , n.
√
Hence Z e j is uniquely determined in terms of Y e except for a multiple of ±1, ± −1. If any par-
ticular element of Z e j is assumed to be real and positive, for example the first element, then Z ej is
uniquely determined. Thus if the first row elements of Z e are real and of specified signs, which is
e )−1 are real and of specified signs, then
equivalent to saying that the first row elements of (1 + X
the transformation is unique.

Proposition 7.10. Let Y e and X e be n × n matrices of functionally independent complex variables such that
e is hermitian positive definite, X
Y e is skew hermitian and the first row elements of (1 + X
e )−1 are real and
of specified signs. Let D = diag(λ1 , . . . , λn ), where the λ j ’s are real distinct and positive. Ignoring the
sign, if
   ∗
e e −1 e −1
Y = 2( X + 1 ) − 1 D 2( X + 1 ) − 1 ,

then

!
 2  −n
 
e] = 2
[dY n ( n − 1)
· ∏  λk − λ j  e e e ][dD ].
·  det((1 + X )(1 − X ))  · [dX
j> k

e = 2( 1 + X
Proof. Let Z e )−1 − 1, X
e ∗ = − X. e = ZD
e Taking the differentials in Y e Ze ∗ we have

e = dZ
dY e∗ + Z
e·D·Z e∗ + Z
e · dD · Z e∗.
e · D · dZ (7.22)

103
But

e ) −1 · d X
e = − 2( 1 + X
dZ e ) −1
e · (1 + X
1 e ) · dX e · (1 + Z
e)
= − (1 + Z
2
and

e ∗ = 2( 1 − X
dZ e ) −1 · d X e ) −1
e · (1 − X
1 e ∗ ) · dX e ∗ ).
e · (1 + Z
= (1 + Z
2
From (7.22), one has

e ∗ · dY
Z e = − 1 (1 + Z
e·Z e ∗ ) · dX e ) · D + dD + 1 D · (1 + Z
e · (1 + Z e ∗ ) · dX
e · (1 + Z
e)
2 2
e∗ Z
observing that Z e = 1. Let

dU e ∗ · dY
e = Z e·Z
e =⇒ [dU
e ] = [dY e∗ Z
e] since Z e = 1, (7.23)
dVe = (1 + Z e ∗ ) · dX
e · (1 + Ze ) = (1 + X e ∗ )−1 · 4dX e )−1 =⇒
e · (1 + X
  −n
[ dVe ] = 4n2 ·  det((1 + X e )(1 − Xe ))  · [dX e] (7.24)

e But in our case there are only n2 − n real variables in X

if there are n2 free real variables in X. e
when X e is uniquely chosen and hence
 −n
e ] = 4n 2 −n  e e  e ],
[ dV det (( 1 + X )( 1 − X ))  · [d X (7.25)

and

e = − 1 dV
dU e · D + 1 D · dV
e + dD. (7.26)
2 2
e is skew hermitian and dU
From (7.26) and using the fact that dV e is hermitian we have

(m) 1 (m)
du jj = dλ j , du jk = ± (λk − λ j )dv jk , j > k, m = 1, 2.
2
Thus the determinant of the Jacobian matrix, in absolute value, is
!2
1    2
∏ 2 λ k − λ j  = 2− n ( n − 1 ) ∏  λ k − λ j  .
j> k j> k

That is,
 
e ] = 2− n ( n − 1 ) · ∏  λ k − λ j  2 · [ dV
[ dU e ][dD ].
j> k

e ] and [dV
Substituting for [dU e ] from (7.23) and (7.24) the result follows.

104
 e an n × n skew hermitian matrix with the first row elements of (1 + X
Example 7.11. For X e ) −1
real and of specified signs show that
Z  −n e
e ] det((1 + X
[d X e ))  = Γn (n) ,
e )(1 − X
e
X ∆
where
Z
" #
 2 − Tr (D )
∆ : = 2n ( n − 1 ) [dD ] ∏  λi − λ j  e ,
λ1 >···> λn >0 16i< j6n

with D = diag(λ1 , . . . , λn ), λ1 > · · · > λn > 0. Consider a n × n hermitian positive definite

e of functionally independent complex variables. Let
matrix Y
Z Z   n−n
B = e]e− Tr (Y) =
[dY
e
e] det(Y
[dY e)  e
e− Tr (Y)
e =Y
Y e ∗ >0 e >0
Y
n ( n −1 )
= e
Γn (n) = π 2 Γ ( n ) Γ ( n − 1) · · · Γ ( 1)
n ( n −1 )
= π 2 (n − 1)!(n − 2)! · · · 1!

evaluating the integral by using a complex matrix-variate gamma integral. Put

e = ZD
Y e Ze∗ , e ) −1 − 1
e = 2( 1 + X
Z

as in Proposition 7.10. Then

!
 2  −n
 e ))  · [dX
e ] = 2n ( n − 1 ) ·
[dY ∏  λk − λ j  e )(1 − X
·  det((1 + X e ][dD ]
j> k

and
Z  −n
B = e ]  det((1 + X
[d X e )) 
e )(1 − X
e
X
Z
!
 2
×
λ1 >···> λn >0
[dD ]2 n ( n − 1)
· ∏  λk − λ j  e− Tr ( D ) .
j> k

Hence the result. From (i) in Example 3.21, we see that

  n ( n − 1)  2
2 e
∆= Γn (n) ,
π

which implies that, when X e is taken over all skew hermitian under the restriction that the first
e )−1 are real and of specified signs,
row elements of (1 + X

105
 Z   − n   n ( n − 1) 1
e ] det((1 + X
[d X e ))  = π
e )(1 − X .
e
X 2 e
Γn (n)

Remark 7.12. In fact, we can derive the volume formula (3.36) from Proposition 7.10. The rea-
soning is as follows:
Z Z
! !
n  2
e >0:Tr (Y
Y e ) =1
e ] = 2n ( n − 1 )
[dY
λ1 >···> λn >0
[dD ]δ ∑ λj − 1 ∏  λk − λ j 
j=1 j> k
Z  −n
× e ]  det((1 + X
[d X e )) 
e )(1 − X
Xe

2 − 1)
n ( n Γ ( 1) · · · Γ ( n ) Γ ( 1) · · · Γ ( n + 1)  π  n ( n − 1 ) 1
= × ×
n! Γ ( n2 ) 2 e
Γn (n)
n ( n −1 ) Γ ( 1) · · · Γ ( n )
= π 2 .
Γ ( n2 )
e ∈ C n×n be a hermitian matrix of independent complex variables. Show that
Example 7.13. Let X
Z 2
n ( n −1 )
e ]e− Tr ( Xe Xe ∗ ) = 2− 2 π n2 .
[d X
e
X

e∗ = X
Indeed, X e implies that
  n  2
Tr X e∗ =
eX ∑ x2jj + 2 ∑  xeij  .
j=1 i< j

Thus
Z Z ∞
! Z
!
n 2
e ]e eX
− Tr ( X e∗ ) − x2jj −2| xeij |
e
[d X = ∏ e dx jj × ∏ e xij
de
X j=1 − ∞ i< j
Z
!
2
− n (n2−1)
e−| xeij | de
n
= π ×2 2
∏ xij
i< j
n2
n
− n (n2−1) n ( n −1 ) n ( n −1 )
= π ×2 2 ×π 2 = 2− 2 π2.

Example 7.14. By using Example 7.13 or otherwise show that

Z
! !
 2 n n ( n −1 )
n −1
n
∏  λk − λ j  exp − ∑ λ2j dλ1 · · · dλn = 2− 2 π 2
∞ > λ1 >···> λn >− ∞
∏ j!.
j> k j=1 j=1

e = UD
Indeed, in Example 7.13 letting X e Ue ∗ , where D = diag(λ1 , . . . , λn ) and U
e ∈ U1 (n), gives
rise to !
e] =
[d X ∏ (λi − λ j ) 2 e1 ], λ1 > · · · > λn ,
[dD ][dG
i< j

106
which means that
Z
n ( n −1 ) n2
2− 2 π 2 = e ]e− Tr ( Xe Xe ∗ )
[d X
e
X
Z
! ! Z
n
= ∏(λi − λ j ) 2
exp − ∑ λ2j [dD ] × e1 ].
[d G
∞ > λ1 >···> λn >− ∞ i< j j=1 U1 ( n )

That is,
Z
! !
n n n ( n −1 )
n −1
n

∞ > λ1 >···> λn >− ∞

∏ (λi − λ j ) 2
exp − ∑ λ2j ∏ dλ j = 2− 2 π2 ∏ j!.
i< j j=1 j=1 j=1

Therefore
Z
! !
n n n ( n −1 )
n
n
∏ ( λ i − λ j )2 exp − ∑ λ2j ∏ dλ j = 2− 2 π2 ∏ j!.
i< j j=1 j=1 j=1

8 Appendix III: Some matrix factorizations

The materials in this section are collected from Muirhead’s book [25]. It is the necessary under-
lying basis for computing some Jacobians .
Firstly, we recall the Gram-Schmidt orthogonalization process which enables us to construct
an orthonormal basis of R m given any other basis X1 , . . . , Xm of R m . We define


Y1 = X1 ,







 Y2 = X2 − hhYY1 ,X 2i
Y1 ,
 1 ,Y1 i

Y3 = X3 − hhYY2 ,X 3i
Y2 − hhYY1 ,X 3i
Y1 ,
 2 ,Y2 i 1 ,Y1 i



 ············





Ym = Xm − ∑m−1 hYj ,Xm i Yj ,
j=1 hY ,Y i j j

1
and put Zj = Y,
hYj ,Yj i j
where j = 1, . . . , m. Then Z1 , . . . , Zm form an orthonormal basis for R m .
Next matrix factorization utilizes this process.

Proposition 8.1. If A is a real m × m matrix with real characteristic roots, then there exists an orthogonal
matrix H such that H T AH is an upper-triangular matrix whose diagonal elements are the characteristic
roots of A.

Proof. Let λ1 , . . . , λm be the characteristic roots of A := A1 and let X1 be a characteristic vector

of A corresponding to λ1 . This is real since the characteristic roots are real. Let X2 , . . . , Xm be
any other vectors such that X1 , X2 , . . . , Xm for a basis for R m . Using the Gram-Schmidt orthog-
onalization process, construct from X1 , X2 , . . . , Xm an orthonormal basis given as the columns of

107
the orthogonal matrix H1 , where the first column h1 is proportional to X1 , so that h1 is also a
characteristic vector of A corresponding to λ1 . Then the first column of AH1 is Ah1 = λ1 h1 , and
hence the first column of H1T A1 H1 is λ1 H1T h1 . Since this is the first column of λ1 H1T H1 = λ1 1 m , it
is (λ1 , 0, . . . , 0)T . Hence
" #
T
λ1 B1
H1 A1 H1 = ,
0 A2
where A2 is (m − 1) × (m − 1). Since

det( A1 − λ1 m ) = (λ1 − λ) det( A2 − λ1 m−1 )

and A1 and H1T A1 H1 have the same characteristic roots, the characteristic roots of A2 are λ2 , . . . , λm .
Now, using a construction similar to that above, find an orthogonal (m − 1) × (m − 1) matrix
H2 whose first column is a characteristic vector of A2 corresponding to λ2 . Then
" #
T
λ2 B2
H2 A2 H2 = ,
0 A3

where A3 is (m − 2) × (m − 2) with characteristic roots λ3 , . . . , λm .

Repeating this procedure an additional m − 3 times we now define the orthogonal matrix

H = H1 (1 ⊕ H2 )(12 ⊕ H3 ) · · · (1 m−2 ⊕ Hm−1 )

and note that H T AH is upper-triangular with diagonal elements equal to λ1 , . . . , λm .

Proposition 8.2. If A is an m × m non-negative definite matrix of rank r then:

(i) There exists an m × r matrix B of rank r such that A = BBT .

(ii) There exists an m × m nonsingular matrix C such that

" #
1r 0
A=C CT .
0 0

Proof. As for (i), let D1 = diag(λ1 , . . . , λr ) where λ1 , . . . , λr are the nonzero characteristic roots
of A, and let H be an m × m orthogonal matrix such that H T AH = diag(λ1 , . . . , λr , 0, . . . , 0).
Partition H as H = [ H1 , H2 ],where H1 is m × r and H2 is m × (m − r); then
" #
D1 0
A=H H T = H1 D1 H1T .
0 0
√ √ √
Putting D1 = diag( λ1 , . . . , λr ), we then have
p p
A = H1 D1 D1 H1T = BBT ,

108
 √
where B = H1 D1 is m × r of rank r. As for (ii), let C bee an m × m nonsingular matrix whose
first r columns are the columns of the matrix B in (i). Then
" #
1r 0
A=C CT .
0 0

The following result is used often in the text.

Proposition 8.3 (Vinograd, 1950). Suppose that A and B are real matrices, where A is k × m and B is
k × n, with m 6 n. Then AAT = BBT if and only if there exists an m × n matrix H with HH T = 1 m
such that AH = B.

Proof. First suppose there exists an m × n matrix H with HH T = 1 m such that AH = B. Then
BBT = AHH T AT = AAT .
Now suppose that AAT = BBT . Let C be a k × k nonsingular matrix such that
" #
T T
1r 0
AA = BB = C CT ,
0 0

where rank( AAT ) = r. Now put D = C −1 A, E = C −1 B and partition these as

" # " #
D1 E1
D= , E= ,
D2 E2

where D1 is r × m, D2 is (k − r) × m, E1 is r × n, and E2 is (k − r) × n. Then

" # " #
E1 E1T E1 E2T 1r 0
EET = = C −1 BBT C −1,T =
E2 E1T E2 E2T 0 0

and " # " #

D1 D1T D1 D2T 1r 0
DDT = = C −1 AAT C −1,T =
D2 D1T D2 D2T 0 0
which imply that E1 E1T = D1 D1T = 1r and D2 = 0, E2 = 0, so that
" # " #
D1 E1
D= , E= .
0 0

e2 be an (n − r) × n matrix such that

Now let E
" #
E1
e=
E
e2
E

109
 e 2 and an (n − r) × (n − m)
is an n × n orthogonal matrix, and choose an (n − r) × m matrix D
matrix De 3 such that
" #
D1 0
e=
D
e2 D
D e3

is an n × n orthogonal matrix. Then

" # " # " # " #
E1 1r 0 D1 0 1r 0
E= = e
E, [ D, 0] = = e
D,
0 0 0 0 0 0 0

eTE
and hence E = [ D, 0] D e = [ D, 0] Q, where Q = D
e TE
e is n × n orthogonal. Partitioning Q as
" #
H
Q= ,
P

where H is m × n and P is (n − m) × n, we then have HH T = 1 m and

C −1 B = E = DH = C −1 AH

so that B = AH, completing the proof.

Proposition 8.4. Let A be an n × m real matrix of rank m(6 n). Then:

(i) A can be written as A = H1 B, where H1 is n × m with H1T H1 = 1 m and B is m × m positive

definite.

(ii) A can be written as " #

1m
A=H B,
0
where H is n × n orthogonal and B is m × m positive definite.
√
Proof. As for (i), let B := AT A be the positive definite square root of the positive definite matrix
AT A, so that
AT A = B2 = BT B.

Now by using Theorem 8.3, A can be written as A = H1 B, where H1 is n × m with H1T H1 = 1 m .

As for (ii), let H1 be the matrix in (i) such that A = H1 B and choose an n × (n − m) matrix H2 so
that H = [ H1 , H2 ] is n × n orthogonal. Then
" #
1m
A = H1 B = H B.
0

We are done.

We now turn to decompositions of positive definite matrices in terms of triangular matrices.

110
Theorem 8.5. If A is an m × m positive definite matrix, then there exists a unique m × m upper-triangular
matrix T with positive diagonal elements such that A = T T T.

Proof. An induction proof can easily be constructed. The stated result holds trivially for m = 1.
Suppose the result holds for positive definite matrices of size m − 1. Partition the m × m matrix
A as " #
A11 a12
A= ,
aT12 a22
where A11 is (m − 1) × (m − 1). By the induction hypothesis there exists a unique (m − 1) ×
(m − 1) upper-triangular matrix T11 with positive diagonal elements such that A11 = T11
T
T11 .
Now suppose that
" # " #" # " #
T T T
A11 a12 T11 0 T11 x T11 T11 T11 x
A= = = ,
aT12 a22 xT y 0 y xT T11 xT x + y2

T −1
where x is (m − 1) × 1 and y ∈ R1 . For this to hold we must have x = ( T11 ) a12 , and then

−1 −1 −1
y2 = a22 − xT x = a22 − aT12 T11 T
( T11 ) a12 = a22 − aT12 A11 a12 .

Note that this is positive, and the unique y > 0 satisfying this is
q
−1
y = a22 − aT12 A11 a12 .

This completes the proof.

Theorem 8.6. If A is an n × m real matrix of rank m(6 n), then A can be uniquely written as A = H1 T,
where H1 is n × m with H1T H1 = 1 m and T is m × m upper-triangular wit positive diagonal elements.

Proof. Since AT A is m × m positive definite it follows form Theorem 8.5 that there exists a unique
m × m upper-triangular matrix with positive diagonal elements such that AT A = T T T. By The-
orem 8.3, there exists an n × m matrix H1 with H1T H1 = 1 m such that A = H1 T. Note that H1 is
unique because T is unique and rank( T ) = m.

Theorem 8.7. If A is an m × m positive definite matrix and B is an m × m symmetric matrix, there exists
an m × m nonsingular matrix L such that A = LLT and B = LDLT , where D = diag(d1 , . . . , dm ), with
d1 , . . . , dm being the characteristic roots of A−1 B. If B is positive definite and d1 , . . . , dm are all distinct,
L is unique up to sign changes in the first row of L.
√
Proof. Let A be the positive definite square root of A. There exists an m × m orthogonal matrix
H such that
A−1/2 BA1/2 = HDH T ,

111
where D = diag(d1 , . . . , dm ). Putting L = A1/2 H, we now have A = LLT and B = LDLT . Note
that d1 , . . . , dm are the characteristic roots of A−1 B.
Now suppose that B is positive definite and the d j are all distinct. Assume that as well
as A = LLT and B = LDLT , we also have A = MMT and B = MDMT , where M is m × m
nonsingular. Then
  T
M −1 L M −1 L = M −1 LLT M −1,T = M −1 AM −1,T = M −1 MMT MT,−1 = 1 m

so that the matrix Q = M −1 L is orthogonal and QD = DQ. If Q = (qij ) we then have qij di = qij d j
e = diag(±1, . . . , ±1),
so that qij = 0 for i 6= j. Since Q is orthogonal it must then have the form Q
e
and L = M Q.

Theorem 8.8 (SVD). If A is an m × n real matrix (m 6 n), there exist an m × m orthogonal matrix H
and an n × n orthogonal matrix Q such that

H AQT = [Σm , 0],

where Σm = diag(d1 , . . . , dm ) for d j > 0, j = 1, . . . , m and d21 , . . . , d2m are the characteristic roots of AAT .

Proof. Let H be an orthogonal m × m matrix such that AAT = H T D2 H, where D2 = diag(d21 , . . . , d2m ),
with d2j > 0 for j = 1, . . . , m because AAT is non-negative definite. Let D = diag(d1 , . . . , dm ) with
d j > 0 for j = 1, . . . , m; then AAT = ( H T D )( H T D )T , and by Theorem 8.3, there exists an m × n
matrix Q1 with Q1 QT1 = 1 m such that A = H T DQ1 . Choose an (n − m) × n matrix Q2 so that the
n × n matrix " #
Q1
Q=
Q2
is orthogonal; we now have A = H T DQ1 = H T [Σm , 0] Q so that H AQT = [Σm , 0], and the proof is
complete.

Theorem 8.9. If Z ∈ SO(m), i.e., Z is an orthogonal matrix with determinant one, then there exists an
m × m skew-symmetric X such that
Z = eX .

9 Appendix IV: Selberg’s integral

This section is rewritten based on Mehta’s book [20]. The well-known Selberg’s integral is calcu-
lated, and some variants and consequences are obtained as well.

112
Theorem 9.1 (Selberg’s integral). For any positive integer N, let [dx] = dx1 · · · dx N ,

∏
16i< j6 N ( xi − x j ), if N > 1,
∆ ( x ) ≡ ∆ ( x1 , . . . , x N ) = (9.1)
1, if N = 1,

and
!
N
Φ ( x ) ≡ Φ ( x1 , . . . , x N ) = ∏ xαj −1 (1 − xj ) β −1
| ∆( x)|2γ . (9.2)
j=1

Then
Z 1 Z 1 N −1
Γ(α + γj)Γ( β + γj)Γ(γ + 1 + γj)
S N (α, β, γ) ≡
0
···
0
Φ( x)[dx] = ∏ Γ(α + β + γ( N + j − 1))Γ(1 + γ)
, (9.3)
j=0

and for 1 6 K 6 N,
Z 1 Z 1
! Z 1 Z 1
K K
α + γ ( N − j)
0
···
0
∏ xj Φ( x)[dx] = ∏ α + β + γ(2N − j − 1) 0
···
0
Φ( x)[dx], (9.4)
j=1 j=1

valid for integer N and complex α, β, γ with

 
1 Re(α) Re( β)
Re(α) > 0, Re( β) > 0, Re(γ) > − min , , . (9.5)
N N−1 N−1
Aomoto’s proof. For brevity, let us write
R
[0,1] N f ( x1 , . . . , x N )Φ( x1 , . . . , x N )dx1 · · · dx N
h f ( x1 , . . . , x N )i := R . (9.6)
[0,1] N Φ( x1 , . . . , x N )dx1 · · · dx N

Firstly, we note that

d
( x a x2 · · · x K Φ ) (9.7)
dx1 1
N
x1a x2 · · · xK x a x2 · · · x K
= ( a + α − 1) x1a−1 x2 · · · xK Φ − ( β − 1) Φ + 2γ ∑ 1 Φ. (9.8)
1 − x1 j=2
x1 − x j

Indeed,
d d
( x a x2 · · · xK Φ) = ax1a−1 x2 · · · xK Φ + x1a x2 · · · xK Φ ( x1 , . . . , x N ) , (9.9)
dx1 1 dx1
where
!
N
d d 2γ
dx1
Φ=
dx1
| ∆( x)| ∏ xαj −1 (1 − xj ) β−1 . (9.10)
j=1

Since
=1
[ x1a x2 · · · xK Φ( x1 , . . . , x N )]xx11 = 0 = x2 · · · xK Φ(1, x2 , . . . , x N ) − 0 = 0, (9.11)

113
it follows from integrating between 0 and 1 in (9.7), we get
   
x1a x2 · · · xK N x1a x2 · · · xK
0 = (a + α − 1)h x1a−1 x2 · · · x K i − ( β − 1) + 2γ ∑ . (9.12)
1 − x1 j=2
x1 − x j

Now for a = 1 and a = 2, we get


 N  
x1 x2 · · · x K x1 x2 · · · x K
0 = α h x2 · · · x K i − ( β − 1) + 2γ ∑ , (9.13)
1 − x1 j=2
x1 − x j
 2  N  2 
x1 x2 · · · x K x1 x2 · · · x K
0 = (α + 1)h x1 x2 · · · xK i − ( β − 1) + 2γ ∑ . (9.14)
1 − x1 j=2
x1 − x j

Clearly
     
x1 x2 · · · x K x12 x2 · · · xK x1 x2 · · · x K x 2 x2 · · · x K
− = − 1
x1 − x j x1 − x j x1 − x j x1 − x j
= h x1 x2 · · · x K i ; (9.15)

interchanging x1 and x j and observing the symmetry,


    0,
x1 x2 · · · x K x j x2 · · · x K if 2 6 j 6 K,
=− = (9.16)
x1 − x j x1 − x j  1 h x · · · x i, if K < j 6 N,
2 2 K

and
* + 
  x2j x2 · · · xK  1 h x1 x2 · · · xK i, if 2 6 j 6 K,
x12 x2 · · · xK
=− = 2 (9.17)
x1 − x j x1 − x j  h x x · · · x i, if K < j 6 N.
1 2 K

Performing the difference: (9.13) − (9.14) and using the above, we get:

(α + 1 + γ(K − 1) + 2γ( N − K ) + β − 1)h x1 x2 · · · xK i = (α + γ( N − K ))h x2 · · · xK i, (9.18)

or repeating the process

(α + γ( N − K ))
h x1 x2 · · · x K i = h x1 · · · x K − 1 i = · · · (9.19)
α + β + γ(2N − K − 1)
K
α + γ ( N − j)
= ∏ α + β + γ(2N − j − 1) . (9.20)
j=1

For K = N, (9.4) can be written as

K
S N (α + 1, β, γ) α + γ ( N − j)
S N (α, β, γ)
= h x1 x2 · · · x N i = ∏ α + β + γ(2N − j − 1) (9.21)
j=1

114
or for a positive integer α,

S N (α, β, γ) S N (α, β, γ) S (2, β, γ)

= ··· N (9.22)
S N (1, β, γ) S N (α − 1, β, γ) S N (1, β, γ)
K K
α − 1 + γ ( N − j) 1 + γ ( N − j)
= ∏ α − 1 + β + γ(2N − j − 1) ∏ 1 + β + γ(2N − j − 1)
· · · (9.23)
j=1 j=1
K
(α − 1 + γ( N − j)) · · · (1 + γ( N − j))
= ∏ (α − 1 + β + γ(2N − j − 1)) · · · (1 + β + γ(2N − j − 1)) . (9.24)
j=1

Note that Γ(m + ζ ) = (ζ )m Γ(ζ ), where (ζ )m := ζ (ζ + 1) · · · (ζ + m − 1). By using this identity,

we get
Γ(α + γ( N − j))
(α − 1 + γ( N − j)) · · · (1 + γ( N − j)) =
Γ(1 + γ( N − j))
and
Γ(α + β + γ(2N − j − 1))
(α − 1 + β + γ(2N − j − 1)) · · · (1 + β + γ(2N − j − 1)) = .
Γ(1 + β + γ(2N − j − 1))

It follows that
N
Γ(α + γ( N − j))Γ(1 + β + γ(2N − j − 1))
s N (α, β, γ) = S N (1, β, γ) ∏
j=1
Γ(α + β + γ(2N − j − 1))Γ(1 + γ( N − j))
!
N N
Γ(1 + β + γ(2N − j − 1)) Γ(α + γ( N − j))Γ( β + γ( N − j))
= S N (1, β, γ) ∏ ∏ .
j=1
Γ( β + γ( N − j))Γ(1 + γ( N − j)) j=1 Γ(α + β + γ(2N − j − 1))

As S N (α, β, γ) = S N ( β, α, γ) by the fact that ∆(1 − x) = ±∆( x), that is, S N (α, β, γ) is a symmetric
function of α and β, at the same time, the following factor is also symmetric in α and β:
N
Γ(α + γ( N − j))Γ( β + γ( N − j))
∏ Γ(α + β + γ(2N − j − 1))
.
j=1

It follows that the factor

N
Γ(1 + β + γ(2N − j − 1))
S N (1, β, γ) ∏
j=1
Γ( β + γ( N − j))Γ(1 + γ( N − j))

should be a symmetric function of α and β. But, however, this factor is independent of α, therefore
it should be also independent of β by the symmetry. Denote this factor by c(γ, N ), we get
N
Γ(α + γ( N − j))Γ( β + γ( N − j))
S N (α, β, γ) = c(γ, N ) ∏
j=1
Γ(α + β + γ(2N − j − 1))
N
Γ(α + γ( j − 1))Γ( β + γ( j − 1))
= c(γ, N ) ∏ (9.25)
j=1
Γ(α + β + γ( N + j − 2))

where c(γ, N ) is independent of α and β.

115
 To determine c(γ, N ), put α = β = 1;
Z N
Γ(1 + γ( j − 1))2
S N (1, 1, γ) = | ∆( x)|2γ dx = c(γ, N ) ∏ . (9.26)
[0,1] N j=1
Γ(2 + γ( N + j − 2))

Let y be the largest of the x1 , . . . , x N and replace the other x j by x j = yt j , where 0 6 t j 6 1.

Without loss of generality, we assume that y = x N . Then x j = yt j for j = 1, . . . , N − 1. Then

N −1
| ∆( x1 , . . . , x N )|2γ = yγN ( N −1) · | ∆(t1 , . . . , t N −1 )|2γ · ∏ (1 − tj )2γ , (9.27)
j=1

and the Jacobian of change of variables is

 
 
 t1 t2 · · · t N −1 1  
 
 y 0 0 · · · 0  
    
 ∂ ( x1 , . . . , x N )   
 det  =  0 y 0 · · · 0   = y N −1 . (9.28)
 ∂(y, t1 , . . . , t N −1 )   
 .. .. .. .. ..  
 . . . . . 
 
 0 0 ··· y 0 

Now we have
Z
S N (1, 1, γ) = N! | ∆( x)|2γ dx
06 x1 6···6 x N 61
Z 1 Z N −1
= N! y γN ( N −1) N −1
y dy · | ∆(t1 , . . . , t N −1 )|2γ ∏ (1 − tj )2γ dt1 . . . dtN −1
0 06t1 6···6t N −1 61 j=1
 Z 1  Z
!
N −1
γN ( N −1) N −1 2γ 2γ
= N
0
y y dy
[0,1] N −1
| ∆(t1 , . . . , t N −1 )| ∏ (1 − t j ) dt1 . . . dt N −1
j=1
1
= S N −1 (1, 2γ + 1, γ),
γ ( N − 1) + 1

that is
1
S N (1, 1, γ) = S N −1 (1, 2γ + 1, γ) (9.29)
γ ( N − 1) + 1
c(γ, N − 1) N −1 Γ(1 + γ( j − 1))Γ(2γ + 1 + γ( j − 1))
γ ( N − 1) + 1 ∏
= · (9.30)
j=1
Γ(1 + 2γ + 1 + γ( N − 1 + j − 2))

c(γ, N − 1) N −1 Γ(1 + γ( j − 1))Γ(1 + γ + γj)

γ ( N − 1) + 1 ∏
= · . (9.31)
j=1
Γ(2 + γ( N + j − 1))

Thus
N
Γ(1 + γ( j − 1))2 c(γ, N − 1) N −1 Γ(1 + γ( j − 1))Γ(1 + γ + γj)
c(γ, N ) ∏
γ ( N − 1) + 1 ∏
= S N (1, 1, γ) = · .
j=1
Γ(2 + γ( N + j − 2)) j=1
Γ(2 + γ( N + j − 1))

116
This implies that
c(γ, N ) Γ(1 + γN )
= , (9.32)
c(γ, N − 1) Γ (1 + γ )
or
   
c(γ, N ) c(γ, 2) Γ(1 + γN ) Γ(1 + γ2)
c(γ, N ) = c(γ, 1) · ··· = c(γ, 1) · ··· . (9.33)
c(γ, N − 1) c(γ, 1) Γ (1 + γ ) Γ (1 + γ )
Finally we get
N N
Γ(1 + γj) Γ(1 + γj)
c(γ, N ) = c(γ, 1) · ∏ = ∏ , (9.34)
j=2
Γ (1 + γ ) j=2
Γ (1 + γ )

where the fact that c(γ, 1) = 1 is trivial.

Remark 9.2. Note that the conclusion is derived here for integers α, β and complex γ.
A slight change of reasoning due to Askey gives (9.25) directly for complex α, β and γ as fol-
lows. (9.88) and the symmetry identity that S N (α, β, γ) = S N ( β, α, γ) give the ratio of S N (α, β, γ)
and S N (α, β + m, γ) for any integer m,
S N (α, β + m, γ) S N (α, β + m, γ) S (α, β + 1, γ)
= ··· N (9.35)
S N (α, β, γ) S N (α, β + m − 1, γ) S N (α, β, γ)
N N
( β + m − 1) + γ ( N − j ) β + γ ( N − j)
=∏ ···∏ (9.36)
j=1
α + ( β + m − 1) + γ(2N − j − 1) j=1
α + β + γ(2N − j − 1)
N
( β + γ( N − j))m
=∏ , (9.37)
j=1
( α + β + γ(2N − j − 1))m

where we have used the notation

Γ( a + m)
( a) m = ; m > 0;
Γ(m)
i.e. ( a)0 = 1 and ( a)m := a( a + 1) · · · ( a + m − 1) for m > 1. Now
Z N
S N (α, β + m, γ) = | ∆( x)|2γ ∏ xαj −1 (1 − x j ) β+m−1dx j (9.38)
[0,1] N j=1
Z N   β + m −1
− αN − γN ( N −1) 2γ xj
= m | ∆( x)| ∏ xαj −1 1− dx j . (9.39)
[0,m ] N j=1
m

Thus
N
(α + β + γ(2N − j − 1))m
S N (α, β, γ) = S N (α, β + m, γ) · ∏ (9.40)
j=1
( β + γ( N − j))m
Z  
− αN − γN ( N −1) 2γ
N
α −1 x j β + m −1 N
(α + β + γ(2N − j − 1))m
=m | ∆( x)| ∏ x j 1− dx j ∏ (9.41)
[0,m ] N j=1
m j=1
( β + γ( N − j))m
N  Z N  
x j β + m −1
(α + β + γ(2N − j − 1))m −α−γ( N −1)
=∏ m | ∆( x)|2γ ∏ xαj −1 1 − (9.42)
dx j .
j=1
( β + γ( N − j))m [0,m ] N j=1
m

117
Denote a j = α + γ( N − j), b j = β + γ( N − j) and c j = α + β + γ(2N − j − 1), we have

Γ(c+m)
(c)m b−c Γ(b) Γ(c + m) b−c Γ(b) Γ(m)mc
m = m = Γ(b+m)
.
(b)m Γ(c) Γ(b + m) Γ(c) b
Γ(m)m

By using the fact that

Γ(m + c)
lim = 1 (∀c ∈ R ),
m→∞ Γ (m)mc

it follows that
(c)m b−c Γ(b)
lim m = , (9.43)
m→∞ (b)m Γ(c)

therefore
N (c j )m N Γ(bj )
lim
m→∞
∏ (b j )m mb −c j j
=∏
Γ(c j )
. (9.44)
j=1 j=1

Taking m → ∞ in (9.40) gives rise to the following:

!Z Z ∞
N Γ(b ) ∞ N

j
S N (α, β, γ) = ∏ ··· | ∆( x)|2γ ∏ xαj −1 exp − x j dx j . (9.45)
j=1
Γ(c j ) 0 0 j=1

Furthermore,
!R∞ R∞

NΓ( a j )Γ(bj ) 0 ··· 0 | ∆( x)|2γ ∏ N α −1
j=1 x j exp − x j dx j
S N (α, β, γ) = ∏ Γ(c j ) ∏N
. (9.46)
j=1 j=1 Γ ( a j )

By the symmetry of α and β, it follows that the factor

R∞ R∞

0 ··· 0 | ∆( x)|2γ ∏ N α −1
j=1 x j exp − x j dx j
(9.47)
∏N
j=1 Γ ( a j )

is a symmetric function of α and β. Since it is independent of β, it is also independent of α by the

symmetry of α and β. Thus
R∞ R∞
R∞ R∞ β −1

0 ··· 0 | ∆( x)|2γ ∏ N α −1
j=1 x j exp − x j dx j 0 ··· 0 | ∆( x)|2γ ∏ N
j=1 x j exp − x j dx j
= ,(9.48)
∏N
j=1 Γ ( a j ) ∏N
j=1 Γ ( b j )

which is denoted by c(γ, N ). This indicates that

Z ∞ Z ∞ N
N
··· | ∆( x)|2γ ∏ xαj −1 exp − x j dx j = c(γ, N ) ∏ Γ( a j ) (9.49)
0 0 j=1 j=1
N N N
Γ(1 + γj) Γ(α + γ( j − 1))Γ(1 + γj)
=∏ ∏ Γ( a j ) = ∏ . (9.50)
j=2
Γ (1 + γ ) j=1 j=1
Γ (1 + γ )

118
Remark 9.3. Now we show that
Γ(n + α)
lim = 1 (∀α ∈ R + ∪ {0}).
n→∞ Γ(n)nα
Indeed, in order to prove this fact, we need a limit representation of the gamma function given
n
(1+ 1k )
by Carl Friedrich Gauss via Euler’s representation of n! = ∏ ∞
k = 1 1+ n : k

n!nz
Γ(z) = lim .
n → ∞ z ( z + 1) · · · ( z + n )

Γ(n + α) Γ ( α ) α ( α + 1) · · · ( α + n − 1)
lim = lim (9.51)
n→∞ Γ(n)nα n→∞ (n − 1)!nα
α ( α + 1) · · · ( α + n ) n
= Γ(α) lim (9.52)
n→∞ n!nα n+α
α ( α + 1) · · · ( α + n ) n
= Γ(α) lim lim (9.53)
n→∞ n!nα n→∞ n + α
1
= Γ(α) · · 1 = 1. (9.54)
Γ(α)
Another short and elementary proof of this fact can be derived from a result related to inequalities
for Gamma function ratios [31]:
Γ ( x + a)
x ( x + a ) a−1 6 6 x a (∀ a ∈ [0, 1]). (9.55)
Γ( x)
Indeed,
Γ( x + α)
∼ xα (9.56)
Γ( x)
as x → ∞ with α fixed.(This was the objective of Wendel’s article.) To show this, first suppose
that α ∈ [0, 1]. Then Eq. (9.55) gives
 α  α −1 Γ( x + α)
1+ 6 6 1,
x Γ( x) xα
Γ( x +α)
α −1
leasing to limx →∞ Γ( x ) x α
= 1 since lim x→∞ 1 + αx = limx→∞ 1 = 1. For α > 1, the statement
now follows from the fact that
Γ( x + α) Γ ( x + α − 1)
= ( x + α − 1) .
Γ( x) Γ( x)
We are done.

Corollary 9.4 (Laguerre’s integral). By letting x j = y j /L and β = L + 1 in Selberg’s integral and

taking the limit L → ∞, we obtain
Z ∞ Z ∞ N
N
Γ(α + γ( j − 1))Γ(1 + γj)
··· | ∆( x)|2γ ∏ xαj −1 exp − x j dx j = ∏ . (9.57)
0 0 j=1 j=1
Γ (1 + γ )

119
Corollary 9.5 (Hermite’s integral). By letting x j = y j /L and α = β = λL2 + 1 in Selberg’s integral
and taking the limit L → ∞, we obtain
Z +∞ Z +∞ N   N
Γ(1 + γj)
··· | ∆( x)|2γ ∏ exp −λx2j dx j = (2π ) N/2 (2λ)− N (γ( N −1)+1)/2 ∏ . (9.58)
−∞ −∞ j=1 j=1
Γ (1 + γ )

Remark 9.6. For an integer γ, the last equation can also be written as a finite algebraic identity.
Firstly, we note that
Z +∞  n Z
2 2

n
+∞ i d

exp − a x − 2iaxλ x dx = exp − a2 x2 − 2iaxλ dx (9.59)
−∞ −∞ 2a dλ
 n  Z +∞ 
i d 2 2


= exp(−λ ) exp −( ax + iλ) dx (9.60)
2a dλ −∞
√  
π i d n
= exp(−λ2 ), (9.61)
a 2a dλ
letting λ = 0 in the above reasoning, we get
Z +∞ √    √   
2 2

n π i d n 
2  π i d n 
2 
exp − a x x dx = exp(−λ ) = exp(− x ) , (9.62)
−∞ a 2a dλ λ =0 a 2a dx x =0
√
we replace a by a in the last equation, we get
Z +∞ r   
2

n π i d n 
2 
exp − ax x dx = √ exp(− x ) , (9.63)
−∞ a 2 a dx x =0
thus (9.58) therefore takes the form:
 γN ( N −1)   !
i ∂ ∂ 2γ N  N
( jγ)!
2 
√ ∏ − exp − ∑ j 
x = (2a)−γN ( N −1)/2 ∏ (9.64)
.
2 a 16 p< q6 N
∂x p ∂xq j=1 j=1
γ!
( x1 ,...,x N )=0

Replacing the exponential by its power series expansion, one notes that the term (− ∑ N 2 ℓ
j=1 x j )
gives zero on differentiation if ℓ < γN ( N − 1)/2, and leaves a homogeneous polynomial of order
ℓ − γN ( N − 1)/2 in the variables x1 , . . . , x N , if ℓ > γN ( N − 1)/2. On setting x j = 0, j = 1, . . . , N,
one sees that therefore that there is only one term, corresponding to ℓ = γN ( N − 1)/2, which
gives a non-zero contribution. So
 2γ !ℓ
N N
∂ ∂ ( jγ)!
∏ − ∑ x2j = 2ℓ ℓ ! ∏ , (9.65)
16 p< q6 N
∂x p ∂xq j=1 j=1
γ!

where ℓ = γN ( N − 1)/2. If P( x) := P( x1 , . . . , x N ) and Q( x) := Q( x1 , . . . , x N ) are homoge-

neous polynomials in x := ( x1 , . . . , x N ) of the same degree, then a little reflection shows that
P(∂/∂x) Q( x) is a constant which is also equal to Q(∂/∂x) P( x). Thus one can interchange the
roles of x j and ∂/∂x j to get
!ℓ
N N
∂2 ( jγ)!
∑ 2 ∏ ( xi − x j )2γ = 2ℓ ℓ! ∏ , (9.66)
k=1 ∂xk 16i< j6 N j=1
γ!

where ℓ = γN ( N − 1)/2.

120
Corollary 9.7. It holds that
Z  2γ N dθ Z 2π Z 2π  2γ N dθ
  k  iθi iθ j  k
 ∆(eiθ1 , . . . , eiθ N )  ∏ = ··· ∏  e − e  ∏ (9.67)
[0,2π ] N k=1
2π 0 0 16i< j6 N k=1
2π
( Nγ)!
= , (9.68)
(γ!) N
where γ is non-negative integer.

Corollary 9.8. It holds that when there is no overlap in the two sets of factors, the result is
Z 1 Z 1 K1 K1 + K2
B(K1 , K2 ) =
0
···
0 i=1
∏ xi ∏ (1 − x j )Φ( x)dx (9.69)
j = K1 + 1
K K
∏ i=1 1 (α + γ( N − i )) ∏ j=2 1 ( β + γ( N − j))
= I N (α, β, γ) K +K
, (9.70)
∏k=1 1 2 (α + β + γ(2N − k − 1))
where K1 , K2 > 0, K1 + K2 6 N, and when there is overlap
Z 1 Z 1 K1 K1 + K2 − K3
C(K1 , K2, K3 ) =
0
···
0 i=1
∏ xi ∏ (1 − x j )Φ( x)dx (9.71)
j = K1 + 1− K3
K3
α + β + γ ( N − k − 1)
= B(K1 , K2 ) ∏ , (9.72)
k=1
α + β + 1 + γ(2N − k − 1)

where K1 , K2 , K3 > 0, K1 + K2 − K3 6 N.

Remark 9.9. Still another integral of interest is the average value of the product of traces of the
matrix in the circular ensembles. For example,
Z 2π Z 2π  N
 2p
1 (γ!) N   2γ N
 iθk   iθi iθ j 
∑ e 
 16i∏
S N ( p, γ) := ··· e − e  ∏ dθk (9.73)
(2π ) N ( Nγ)! 0 0  k=1 < j6 N k=1

is known for γ = 1 that S N ( p, 1) gives the number of permutations of (1, . . . , k) in which the
length of the longest increasing subsequence is less than or equal to N. One has in particular,

S N ( p, 1) = p!, 0 6 p 6 N.

It is desirable to know the integrals S N ( p, γ) for a general γ.

The following short proof of Selberg’s formula is from [2].

Anderson’s proof of Selberg’s Integral. Anderson’s proof depends on Dirichlet’s generalization of

the beta integral given in the following: For Re(α j ) > 0,
Z Z
∏nj=1 Γ(α j )
··· p0α0 −1 p1α1 −1 · · · pαnn −1 dp0 · · · dpn−1 =  , (9.74)
V Γ ∑nj=1 α j

121
where V is the set p j > 0, ∑nj=0 p j = 1. The formula is used after a change of variables. To see
this, first consider Selberg’s integral, which may be written as
Z 1 Z xn Z x2  γ
Sn = n!An (α, β, γ) := n! ··· | φ(0)|α−1 | φ(1)| β−1  ∆φ  dx1 · · · dxn , (9.75)
0 0 0

where 0 < x1 < x2 < · · · < xn < 1,

n
φ ( t) = ∏(t − xj ) = tn − φn−1tn−1 + · · · + (−1)n φ0 (9.76)
j=1

and ∆φ is the discriminant of φ, so that

   2
   n ′   
 ∆φ  =  ∏ φ ( x j )  =  ∏ ( xi − x j )  .
 j=1   16i< j6n 

We now change the variables from x1 , . . . , xn to φ0 , . . . , φn−1, which are the elementary symmetric
functions of the xi ’s. In fact, we have:
Z  γ− 1
An (α, β, γ) = | φ(0)|α−1 | φ(1)| β−1  ∆φ  2 dφ0 dφ1 · · · dφn−1, (9.77)

where the integration is over all points (φ0 , φ1 , . . . , φn−1 ) in which the φj are elementary sym-
metric functions of x1 , . . . , xn with 0 < x1 < · · · < xn . Indeed, it is sufficient to prove that the
Jacobian   q 
∂φi  ∆φ .
det =
∂x j
Observe that two columns of the Jacobian are equal when xi = x j . Thus ∏i< j ( xi − x j ) is a factor
of the determinant. Moreover, the Jacobian and ∏i< j ( xi − x j ) are homogeneous and of the same
degree. This proves the above Jacobian.
We make a similar change of variables in Eq. (9.74). To accomplish this, set
n
ϕ ( t) = ∏(t − ζ j ) ( 0 6 ζ 0 < ζ 1 < · · · < ζ n < 1)
j=0

and let
( )
n
D= ∏(t − xj ) : ζ j−1 < xj < ζ j ; j = 1, . . . , n . (9.78)
j=1

Next, we show that: For all φ(t) = tn − φn−1 tn−1 + · · · + (−1)n φ0 ∈ D , the following map
 
φ(ζ 0 ) φ(ζ n )
(φ0 , φ1 , . . . , φn−1 ) 7→ ,..., ′ ≡ ( p0 , p1 , . . . , p n ) ∈ R n + 1
ϕ′ (ζ 0 ) ϕ (ζ n )

where ϕ′ (t) denotes the derivative of ϕ(t), is a bijection and p j > 0 with ∑nj=0 p j = 1.

122
 Observe that
φ(ζ j ) (ζ j − x1 )(ζ j − x2 ) · · · (ζ j − xn )
pj = ′
= >0 (9.79)
ϕ (ζ j ) (ζ j − ζ 0 ) · · · (ζ j − ζ j−1 )(ζ j − ζ j+1 ) · · · (ζ j − ζ n )
ϕ(t)
since the numerator and denominator have exactly n − j negative factors. Now let ϕ j (t) = t−ζ j .
By Lagrange’s interpolation formula
n n ϕ j ( t)
φ ( t) = ∑ p j ϕ j ( t) ≡ ∑ ϕ′ (ζ j )
φ ( ζ j ). (9.80)
j=0 j=0

One can directly verify this by checking that both sides of the equation are polynomials of degree
n and are equal at n + 1 points t = ζ j , j = 0, . . . , n. Equate the coefficients of tn on both sides to
get ∑nj=0 p j = 1. Now for a given point ( p0 , p1 , . . . , pn ) with ∑nj=1 p j = 1 and p j > 0, j = 1, . . . , n,
define φ(t) by Eq. (9.80). The expressions

φ(ζ j ) = p j ϕ j (ζ j ) = p j (ζ j − ζ 0 ) · · · (ζ j − ζ j−1 )(ζ j − ζ j+1 ) · · · (ζ j − ζ n )

and

φ(ζ j+1 ) = p j+1 ϕ j+1 (ζ j+1 ) = p j+1 (ζ j+1 − ζ 0 ) · · · (ζ j+1 − ζ j )(ζ j+1 − ζ j+2 ) · · · (ζ j+1 − ζ n )

show that φ(ζ j ) and φ(ζ j+1 ) have different signs and φ vanishes at some point x j+1 between ζ j
and ζ j+1 . Thus φ ∈ D . This proves the bijection.
We can now restate Dirichlet’s formula Eq. (9.74) as:

Z
 α − 1
n   α j −1 ∏nj=0  ϕ′ (ζ j )  j 2 Γ(α j )
∏  φ(ζ j )  dφ0 · · · dφn−1 =   . (9.81)
φ( t)∈D j=0 Γ ∑nj=1 α j

Indeed,  
  α j −1  φ ( ζ j )  α j −1
α −1 φ(ζ j )
pj j = =  ,
ϕ′ (ζ j )  ϕ ′ ( ζ j )  α j −1

hence
Z Z
··· p0α0 −1 p1α1 −1 · · · pαnn −1 dp0 · · · dpn−1 (9.82)
V
Z n   α j −1
1
=  
 ϕ ′ ( ζ j )  α j −1
∏  φ(ζ j )  dp0 · · · dpn−1 . (9.83)
∏nj=0 φ( t)∈D j=0

We need to verify that the Jacobian

  n
∂ ( p0 , . . . , p n − 1 )  − 1
det
∂(φ0 , . . . , φn−1 )
= ∏  ϕ′ (ζ j )  2
,
j=0

123
  − 1  1
that is, dp0 · · · dpn−1 = ∏nj=0  ϕ′ (ζ j )  2 dφ0 · · · dφn−1 or dφ0 · · · dφn−1 = ∏nj=0  ϕ′ (ζ j )  2 dp0 · · · dpn−1 .
Since
1  n n −1 n

pj = ζ j − φ n − 1 ζ j + · · · + (− 1 ) φ0 ,
ϕ′ (ζ j )
the Jacobian is
       
    j   det ζ j   1
 ∂ ( p , . . . , p )   det ζ i   i n  ∏nj=0  ϕ′ (ζ j )  2
 det 0 n − 1 =  n −1   
 ∂(φ0 , . . . , φ ) 
 n −1 ′ =
 ∏ j=0 ϕ ( ζ j )   ∏ n ϕ′ (ζ )  = ∏n  ϕ′ (ζ )  .
n −1
   j=0 j  j=0 j

The numerator is a Vandermonde determinant and therefore the result follows:

Z Z
··· p0α0 −1 p1α1 −1 · · · pαnn −1 dp0 · · · dpn−1
V
Z
!
n
  α −1 n 1
1
=  α − 1 ∏  φ(ζ j )  j ∏  ϕ′ (ζ j )  2 dp0 · · · dpn−1
∏ nj=0  ϕ′ (ζ j )  j 2 φ( t)∈D j=0 j=0
Z n   α j −1
1
=   1
 ϕ′ (ζ j ) α j − 2
∏  φ(ζ j )  dφ0 · · · dφn−1.
∏ nj=0 φ( t)∈Dj=0

The final step is to obtain the (2n − 1)-dimensional integral. Let φ(t) and Φ(t) be two polynomials
such that
n −1 n
φ ( t) = ∏ ( t − xi ) and Φ(t) = ∏ ( t − y j ), (9.84)
i=1 j=1
0 < y1 < x1 < y2 < · · · < xn−1 < yn < 1. (9.85)

The resultant of φ and Φ, denoted R(φ, Φ), is given by

     
   n   n −1 
     
| R(φ, Φ)| =  ∏ ( xi − yj )  =  ∏ φ(yj )  =  ∏ Φ( xi )  . (9.86)
 i∈[n−1];j∈[n]  j=1 i=1

The absolute value of the discriminant of φ can be written as | R(φ, φ′ )|. That is,

  n −1
 ∆φ  = | ∆( x)| 2 = ∏ φ ′ ( x j ).
j=1

The (2n − 1)-dimensional integral is

Z
| Φ(0)| α−1 | Φ(1)| β−1 | R(φ, Φ)|γ−1 dφ0 · · · dφn−2dΦ0 · · · dΦn−1
( φ,Φ)
  γ −1
Z  n 
α −1 β −1  
= | Φ ( 0 )| | Φ ( 1 )| ∏ φ ( y )
j  dφ0 · · · dφn−2 dΦ0 · · · dΦn−1 . (9.87)
( φ,Φ)  j=1


124
Here the integration is over all φ and Φ defined by Eq. (9.84). Then we show that Selberg’s
integral An (α, β, γ) satisfies the recurrence relation:
Γ(α)Γ( β)Γ(γn)
An (α, β, γ) = An−1 (α + γ, β + γ, γ) . (9.88)
Γ(α + β + γ(n − 1))
In fact, integrate the (2n − 1)-dimensional integral Eq. (9.87) with respect to dφ0 · · · dφn−2 and
use Φ(t) instead of ϕ(t) in Eq. (9.81) to get
Z
  γ −1
 n 
 
| Φ(0)| α−1 | Φ(1)| β−1  ∏ φ(y j )  dφ0 · · · dφn−2dΦ0 · · · dΦn−1
( φ,Φ)  j=1 
Z Z n 
!
α −1 β −1  γ −1
=
Φ
| Φ(0)| | Φ(1)| ∏  φ(yj )  dφ0 · · · dφn−2 dΦ0 · · · dΦn−1
φ j=1
  γ− 1 
Z n  ′
Φ ( y )  2 Γ(γ)
∏ j = 1 j
= | Φ(0)| α−1 | Φ(1)| β−1   dΦ0 · · · dΦn−1
Φ Γ(γn)
 γ − 1
Z  n  2
Γ(γ)n   Γ(γ)n
= | Φ(0)| α−1 | Φ(1)| β−1  ∏ Φ′ (y j )  dΦ0 · · · dΦn−1 = An (α, β, γ).
Γ(γn) Φ  j=1  Γ(γn)

e(t) = t ∏ nj=1 (t − x j ), and

It remains to compute Eq. (9.87) in another way, set φ

α0 = α, α j = γ( j = 1, . . . , n − 1), αn = β; x0 = 0, xn = 1

so that Eq. (9.87) is equal to

  γ −1
Z  n −1 
 
| Φ(0)| α−1 | Φ(1)| β−1  ∏ Φ( x j )  dΦ0 · · · dΦn−1 dφ0 · · · dφn−2
( φ,Φ)  j=1 
Z n   α j −1
=
( φ,Φ)
∏  Φ( x j )  dΦ0 · · · dΦn−1 dφ0 · · · dφn−2. (9.89)
j=0

e instead of ϕ(t) in Eq. (9.81) to

Now integrate Eq. (9.89) with respect to dΦ0 · · · dΦn−1 and use φ
obtain

Z  n −1
γ− 1
Γ ( α ) Γ ( β ) Γ ( γ ) n −1  2  1  ′  β− 21
 e′ ( x j )  e′ (0) α− 2  φ
φ e ( 1) 
∏ φ dφ0 · · · dφn−2 .
Γ(α + β + γ(n − 1)) φ  j=1 

Since
 
 ′   n −1 
e (0)  =  ∏ x j  ,
φ 
 j=1 
 
 ′   n −1 
e (1)  =  ∏ (1 − x j )  ,
φ
 j=1 
n  n −1  n −1   
∏  φ′ ( x j )  = ∏ j  ∏  1 − x j   ∆φ  ,
 x
j=1 j=1 j=1

125
the last integral can be written as
Z
!
n −1
Γ ( α ) Γ ( β ) Γ ( γ ) n −1 α + γ −1  γ − 1
 ∆φ  2 dφ0 · · · dφn−2
∏ xj (1 − x j ) β + γ −1
Γ(α + β + γ(n − 1)) φ j=1

Γ ( α ) Γ ( β ) Γ ( γ ) n −1
= An−1 (α, β, γ).
Γ(α + β + γ(n − 1))

Equate the two different evaluations of the (2n − 1)-dimensional integral to obtain the result.
Finally, Selberg’s formula is obtained by iterating Eq. (9.88) (n − 1) times.

Acknowledgement

Partial material of the present work is completed during a research visit to Chern Institute of
Mathematics, at Nankai University. The author would like, in particular, to thank Seunghun
Hong for his remarks about the approach toward the volume of unitary group via Macdonald’s
method for the volume of a compact Lie group. Both Zhen-Peng Xu and Zhaoqi Wu are acknowl-
edged for valuable comments for the earlier version of the paper. LZ is grateful to the financial
support from National Natural Science Foundation of China (No.11301124).

References

[1] T. Abe, M. Barakat, M. Cuntz, T. Hoge, and H. Terao, The freeness of ideal subarrangements of
Weyl arrangements, J. Eur. Math. Soc. 18, 1339-1348 (2016).

[2] G.E. Andrews, R. Askey, and R. Roy, Special Functions, Cambridge University Press (1999).

[3] L.J. Boya, The geometry of compact Lie groups, Rep. Math. Phys. 30(2), 149-162 (1991).

[4] L.J. Boya, E.C.G. Sudarshan, T. Tilma, Volumes of compact manifolds, Rep. Math. Phys.
52(3), 401-422 (2003).

[5] P. Deift, D. Gioev, Random Matrix Theory: Invariant Ensembles and Universality, AMS (2009).

[6] J.J. Duistermaat, J.A.C Kolk, Lie Groups, Springer-Verlage New York, Inc. (2000).

[7] A. Edelman, Eigenvalues and condition numbers of random matrices, Ph.D. Thesis (1989).

[8] K. Fujii, Introduction to Grassmann manifolds and quantum computation, J. Appl. Math.
8, 371-405 (2002).

[9] G. Harder, A Gauss-Bonnet formula for discrete arithmetically defined groups, Ann. Ec. Norm.
Sup. 4, 409-445 (1971).

126
[10] Y. Hashimoto, On Macdonald’s formula for the volume of a compact Lie group, Comment. Math.
Helv. 72, 660-662 (1997).

[11] J.E. Humphreys, Reflection groups and Coxeter groups, Cambridge University Press (1990).

[12] K. Iwasaki, Basic invariants of finite feflection groups, J. Algebra 59 538-547 (1997).

[13] A.W. Knapp, Lie groups beyound an introduction, 2nd, Birkhäuser (2002).

[14] S. Kumar and A. Pandey, Entanglement in random pures states: spectral density and average von
Neumann entropy, J. Phys. A : Math. Theor. 44, 445301 (2011).

[15] B. Kostant, The principal three-dimensional subgroup and the Betti numbers of a complex simple
Lie group, Amer. J. Math. 81, 973-1032 (1959).

[16] V. Link and W.T. Strunz, Geometry of Gaussian quantum states, J. Phys. A : Math. Theor.
48, 275301 (2015).

[17] M.S. Marinov, Invariant voulmes of compact groups, J. Phys. A : Math. Gen.
13, 3357-3366 (1980); Correction to ‘Invariant voulmes of compact groups’, J. Phys. A : Math.
Gen. 14, 543-544 (1981).

[18] M.S. Marinov and M.V. Terentyev, Dynamics of the group manifold and path integrals, Fort. D.
Phys. 27, 511-545 (1979).

[19] A.M. Mathai, Jacobians of Matrix Transformations and Functions of Matrix Arguments, World
Scientific (1997).

[20] M.L. Mehta, Random Matrices, 3nd Edition, Elsevier Academic Press (2004).

[21] I.G. Macdonald, The Poincaré series of a Coxeter group, Math. Ann. 199, 161-174 (1972).

[22] I.G. Macdonald, The volume of a compact Lie group, Inven. Math. 56: 93-95 (1980).

[23] I.G. Macdonald, Some conjectures for root systems, SIAM J. Math. Anal. 13(6), (1982).

[24] S. Milz and W.T. Strunz, Volumes of conditioned bipartite state spaces, J. Phys. A : Math. Theor.
48, 035306 (2015).

[25] R.J. Muirhead, Aspects of Multivariate Statistics Theory, John Wiley & Sons, New York (2005).

[26] K.R. Parthasarathy, Extremal quantum states in coupled systems, Ann. I. H. Poincaré–PR
41, 257-268 (2005).

127
[27] F.W. Ponting, H.S.A. Potter, The volume of orthognal and unitary space, Q. J. Math
os-20(1): 146-154 (1949).

[28] R. Steinberg, Finite reflection groups, Trans. Amer. Math. Soc. 91(3), 493-504 (1959).

[29] H.J. Sommers and K. Życzkowski, Bures volume of the set of mixed quantum states, J. Phys. A :
Math. Gen. 36, 10083-10100 (2003).

[30] L. Wei, R.A. Pitaval, J. Corander, and O. Tirkkonen, From random matrix theory to coding
theory: volume of a metric ball in unitary group, arXiv: 1506.07259

[31] J.G. Wendel, Note on the gamma function, Amer. Math. Monthly 55, 563-564 (1948).

[32] L. Zhang, Matrix integrals over unitary groups: An application of Schur-Weyl duality,
arXiv: 1408.3782

[33] L. Zhang, A.K. Pati, and J. Wu, Interference visibility, entanglement, and quantum correlation,
Phys. Rev. A 92, 022316 (2015).

[34] L. Zhang, L. Chen, and K. Bu, Fidelity between a bipartite state and another one undergoing local
unitary dynamics, Quant. Inf. Processing 14, 4715 (2015).

[35] K. Życzkowski, P. Horodecki, A. Sanpera, and M. Lewenstein, Volume of the set of separable
states, Phys. Rev. A 58, 883 (1998).

[36] K. Życzkowski, Volume of the set of separable states. II, Phys. Rev. A 60, 3496 (1999).

[37] K. Życzkowski, W. Slomczynski, The Monge metric on the sphere and geometry of quantum states,
J. Phys. A : Math. Gen. 34, 6689 (2001).

[38] K. Życzkowski, H.-J. Sommers, Hilbert-Schmidt volume of the set of mixed quantum states, J.
Phys. A : Math. Gen. 36, 10115 (2003).

128