Determination

Communicated

of the Inertia of a Partitioned

Hermitian

Matrix

by Alan J. Hoffman

1. INTRODUCTION

1. The
is defined

inertia

of an n x n matrix

trijde

to he the ordered

integer

A with complex

elements

triple,

In A = (z, v, n),
where the nonnegative

integers

the number of eigenvalues

and on the imaginary
matrices
number

rank of a matrix
n(H)

axis,

are considered,
of positive,

k, then there

eigenvalues
without

Y = y(A), and 6 = 6(A) give

exists

matrix

the triple simply gives the

and zero eigenvalues.

We shall denote

Then if H is an Hermitian
Hermitian

matrix.

a permutation

of H remain

invariant

matrix,

matrix,

if

P, such that

PTHP

left corner.

As the

under such a transformation,

we can,

assume

H=
where H,, is nonsingular

the

p(H) =

It is well known that

of order k in the upper

loss of generality,

the left half-plane,

In this paper only Hermitian

and for such matrices

negative,

+ y(H).
S uppose H is an N

p(H) 2

respectively.

A by p(A).

has a nonsingular

7c = n(A),

of A in the right half-plane,

(1)

of order k, and HI*, is the conjugate

transpose

of H,,.
2. Theorem
the inertia

if H is partitioned

as in (l),

as the sum of the inertias

of the two

1 of this paper shows that

of H can be determined
Linear

Copyright

1968 by

,4lgehru

amd Ifs

American

Applications

Elsevier

Publishing

1, 73-81

Company,

(1968)
Inc.

EMILIE

74
Hermitian

matrices

of orders

k and n -

K,, = H,, This method of determining

amount of computation
2n were partitioned
Uhile

k, respectively:

(2,

then

formula,

H,,.

any nonsingular

det H = det H,, det Ii,,,

has

of Theorem

K,,

principal

permuting

we define the Schur

submatrix

the rows and columns

some

formula

in particular,

Hermitian

matrix

principal

Schurs

is the product

submatrix

3. Corollary

1 to Theorem
of the matrix

any principal

submatrix

determine

states

that

1 formulates

in (1).

K,,

of (2).

the determinant

of an

of any nonsingular

of its Schur complement.

explicitly

the condition

Again it is necessary

of H together

of

obtained
so that

the matrix

of the determinant

of H with the determinant

the positivity

complement

of H cogrediently

Then,

with

its

for

only to examine

Schur

complement

to

this property.

4. Theorem

2 gives another

of the Schur complement,

positive

the matrix

G is easier

about the inertia

judicious

to compute

that if H,,

is positive

k columns

and has inertia

this matrix,
principal

of Theorems

7. In Theorems

Schur

definite,

Algebra

of the diagonal
information

of the Schur complement,

semidefivzite,

then

H is nonsingular

k, 0).

submatrix.
1, 73-

and the last

5 and 6 we use one of the classical separation

and Its Applications

by a

of H.

1 and 2, we prove in Theorem

to find lower bounds for the inertia triple of an Hermitian

principal

of

cases

complement.

its size can be reduced

independent,

In H = (k, n -

in terms

In many

submatrices

N,, is negative

of H are linearly
triple,

H,,.

of H can yield

the calculation

of certain

6. As a useful application

the

submatrices

to compute

examination

than

in which, instead

G that is formed

in some cases examination

principal

triple without

or if it is necessary

for the inertia,

of the matrix

parts

3 shows that

and other

formula

we use a matrix

and negative

5. Theorem
elements

Linenr

known

complement

K of H to be the matrix

and then computing

singular

been

1 connecting

defined by (2) the Schur

More generally,

is put into the H,, position,

n -

could

of H with those of H,, and K,, appears not to have been noticed.

of the matrix

the

of order

all the computation

of order a.

Ve shall call the matrix

by first

and

of large order, for, if a matrix

and used for several decades, the statement

properties

Hi,

the inertia of H would save a considerable

into n x n blocks,

Schurs

HAYNSWORTH

H;H;lH,,.

for matrices

be done with matrices

V.

81 (l968)

matrix

theorems
having a

INERTIA
.

OF PARTITIONED

TWO BASIC INERTIA

8. THEOREM

If H is an Hermitian

1.

75

MATRIX

FORMULAS

is nonsingular

where H,,

HERMITIAN

matrix partitioned

as in (I),

of order k, then

In H = In H,, + In K,,,

where K,,

is the Schur complement

Proof.
Let

I,

where
matrix

H is the

Suppose

Q be the

partitioned

in general

verify

Hermitian

matrix

partitioned

matrix

of order

the identity

0 represents

immediately

given in the formula

(2).

as in

(1).

matrix,

represents

Q the block

of H,,

(3)

an (n -

k)

m, and in the

k zero matrix.

Then

we

that

ffll

( 1

Q=Q*HQ=

22

K,,

where

is the matrix

defined

in (2)

[2].

Since

the inertia

same as that of H, and the roots of i? are those of H,, and K,,,
in

(3)

is established.

9.

COROLLARY

1.

definite if and only

definite,
Proof.
Thus,

The
while

H,, and K,,,

that
can

if

be put,

position,

we

follows

are both

complement
is

by
can

positive

immediately

where

the relation
that

the

submatrix

of H is positive

from

for the positivity

of H to find a single
it is a necessary

definite,

of every principal
definite.

a cogredient

(3).

positive

That

in H be positive

submatrix

is, for every

permutative

pair,

condition

submatrix

transformation,

into

which
the

H,,

write
H,,

means

the equality

with its Schur complement.

it is sufficient

which

the Schur

definite

proof

is the

Under the conditions of Theorem 1, H is positive

if at least one principal

simultaneously

of A

> H1*2H$H,2,

H > K for two Hermitian

matrix

H -

K is positive

Linear

Algebra

and

matrices

of the same order

definite.
Its

Applications

1, 73%-81 (1968)

EMILIE

76

10.

V. HAYNSWORTR

If, under the hypotheses of Theorem 1, In N,, = (~5,k - $, 0),

then H,, is congruent to a block diagonal matrix,

P,HllP1*

= D =

(4)

where D, and D, are positive definite matrices of orders 9 and k - fi,

respectively. If H,, is positive (negative) definite, the matrix D, (Dl)
will not appear in (4) nor in the subsequent formulas.
11.

Using the decomposition

THEOREM2.

(4), we can prove

S+NXX H is a matrix

of order

as i72 (1)
and D, are +xitive

n fiaditioned

+ s,
n -

(5)

k, with p(R) 2

Proof.
If P, is the transfo~ation matrix in (4), let P be the direct
sum, P = P, $ I,_,.
Then the matrix

H= PHP*=
has the same inertia as H.

Plf42

f$;pl*

H,,

Applying Theorem 1 to Z?, we have

In Ei = In H = In D -t In G,
where G = Ha, -

(H~P~*)D-l(P~~~~).

Now we let PrH,, =

define R = C,*D,-lC,,

cr
c

, where C, is of order 9.

S =C,*D2-1C2,

(H&P,*)D-l(PIHI,)

= Cl*D,-1C,

Then we can

and we have
-

C2D,-lC,

= R -

S,

where R and S are positive semidefinite, since D,-l and D,-l are positive
definite. Also it is clear that p(R) 5 p(D,), p(S) 2 p(D,), and the theorem
is proved.
Li+aearAlgebra

and Its

Applications

1, 73--81

(1968)

INERTIA

OF PARTITIONED

3. SOME

APPLICATIOSS

12.

Theorems

the separation
However,

HERMITIAN

OF THE

INERTIA

submatrix

Hy In H 2
THEOREM

discussed

in a later section
from Theorems

of H as the matrix

(a, b, c), we shall

3.

FORMULAS

3 and 4 which follow could have been proved

theorem

they follow immediately

a particular

77

MATRIX

If H, and H,

orders k and m respectively,

mean

H,,

x(H)

using

(cf. formula

(13)).

1 and 2 by choosing
of the theorems.

2 a, Y(H) 2 b, 6(H) 2 c.

are two principal

submatrices

of H, of

and

111Hk = (fi,, qfi>O),

In H, = (P,, qm, O),

then
In H 2 (P, q, O),
Proof.
then H,

By

p = max(pAj pm),

a cogredient

into the position

It follows immediately
and one negative
negative

of (1).

max(qk,

qrn).

transformation,

Then

the result

put

Hk and

follows from (2).

from the above result that if H has one positive

diagonal

eigenvalue.

permutative

H,,

q=

element,

then

More generally

H has one positive

we have

and one

the following

If an Hermitian matrix H has a positive definite flrincipal

COROLL~~RY.

submatrix of order p, and a +aegative definite principal

submatrix of order q,

then
In H 2 (P, 4,
13.
matrix

As a special

case of the above corollary,

if H is the partitioned

in (1) and H,, is positive definite of order k, H,, is negative

of order n we have

k, then it is not necessary

The following
difficult

the matrix

definite
K,,,

for

theorem

(k,n-k,O).

is a little

(7)

more general,

and is only slightly

to prove.
4.

THEOREM

as in (1).

to compute

immediateI!,
InH=

more

0).

Sztppose ?pie have an Hermitian

matrix,

H, partitioned

If the submatrix H1, is positive definite of order k, H,, is negative

semidefinite
indePendent,

of order n -

k, and if the last n -

k coltimns of H aye linearly

tken (7) holds.

Linear

Algebra

md

Its

Applications

1, 73-81

(1968)

78

EMILIE

14.

For the proof

LEMMA.

If

of Theorem

the (n +

4 we need the following

m) x m matrix

has rank m, and P is an m x m positive

P i
Proof.

semidefinite matrix,

SPS

I,

(8)

and I, is the identity

matrix

o ,

matrix

of order m + B, T = I, q S*.

m -

Then

This implies that the n x (m -

Y, so that Q2*Qa is a positive

definite

In this matrix

definite

matrices, the rank

Y) matrix Qz has rank

matrix.

to show that the m x m matrix,

In order to prove (8), it is sufficient

S*(P + Q*Q)S, is a positive

Let T be the

of order r.

where QS = (Q1 Q2). Since T and S are nonsingular

of M is also m.

S such that

i
square matrix

then

Q*Q > 0.

Since P 1 0, there exists a nonsingular

where r = p(P),

V. HAYNSWORTH

matrix.

the Schur complement

Now

of Q2*Q2 is

(Ql*Ql+1,)- Q,*Qz(Qz*QPQz*Q1=
1,.
Thus, by the corollary to Theorem 1, the matrix S*(P
definite
15.

Proof of Theorem 4.

then H can be transformed

is positive

If H satisfies the conditions

of Theorem

4,

to the matrix A in (6), where the submatrix

definite, and thus, without loss of generality,

to be the identity
are linearly
Linear

+ Q*Q)S is positive

and (8) is proved.

Algebra

matrix of order k.

independent,
and Its

since they

Applications

may be assumed

Also, the last n -

k columns

are obtained

those of H by

1, 73-81

(1968)

from

of l?

INERTIA

OF

PARTITIONED

multiplication

HERMITIA?;

on the left by the nonsingular

the matrix

G of Theorem

semidefinite,

we have,

2 is G = H,, -

by the above
-- G = (-

Hence

G has inertia

4. HERiVITIAN

16.

(0, n -

blATRICES

In this

section

classical
those

of H.

principal

submatrix

and V(H) 2 y(K).

than 6(K).
results,
exact
17.

However,

which

have

that

the eigenvalues

(13) below.)

theorem

or nonsingular,

6(H) may be greater

If n -

(15) for the inertia

Theorem

that if K is any

then n(H) 2 n(K),

than,

equal to, or less

n -

6(H),

of H, given the inertia

1 of the Hermitian
Su#xxe

matrix

H is Hermitian,

of H of order n -

are new

k, in the orders of H and I(

k = 6(K) -

we obtain

the

of K.

5 deals with the case where K is a principal

of order k = n THEOREMS 5.

to 8(K).

of

matrix H of order n interlace

in formula

if the difference,

or equal

at least

one of the

It is in this last case that we have what we believe

formula

submatrix

precisely

formulas

principal submatrix.

5 we apply

states

from the separation

of H, singular

particularly

is less than

which

SUBJIATRIX

the inertia

matrices

1 of an Hermitian

(This is stated

It follows immediately

applied

In Theorem

[l],

is negative

follows.

PRINCIPAL

we have

Hermitian

theorems

of order M -

I,_ k. Thus

> 0.

having a nonsingular

submatrix.

P i

Since H,,

lemma,

H,,) + C,*Ci

sections

we deal with

separation

a submatrix

matrix

Ci*C,.

A SINGULAR

1 and 2 to matrices

one singular principal

79

k, 0), and the theorem

WITH

In the previous

of Theorems

MATRIX

submatrix

H of order n.

of order n, and K is a principal

1, having the inertia triple,

In K = (P, q, 4.

(9)

Then
InH>($,q,zP~l).
If 6(H) = z -

(10)

1, then
In H = (p + 1, q + 1, z -

Proof.

Suppose

the roots

1).

(11)

of K are

cl,~a,~~~*2a,_~,Linear

Algebra

und Its

(14
App2iratzons

1, 73-81

(1968)

80

EMILIE

and the roots,

Then,

pi, i = 1, . . ., n, of H are ordered

by the separation

theorem

mentioned

from

(9), (12),

in the same

fashion.

above,

(i=l,...,n-1).

pi>=Ei~pici
Therefore,

V. HAYNSWORTH

(13)

and (13),
(i = 1,.

pi 2 ui > 0

.,$)

and
(i=p+z+1,...,12-1).

PiTi 5 CXi< (1
Thus n(H) >= 9, V(H) 2 q.
Also, since we have
P~+l~~=u~+l=Pp+2=~~~=Pp+z=O(p+z=~~~~;-z+l,

it follows

immediately

that

S(H) 2 z -

root can be lost in proceeding

order m + 1.
matrix

(14)

1.

That

from a submatrix

Also, if 6(H) = z -

is, at most

one zero

of order m to one of

1, it can be seen from (14) that

of order n gains one additional

positive and one additional

the

negative

root.
18.

Applying

THEOREM

Hermitian

6.

Theorem
Suppose

5 repeatedly,

K is a $winci$al submatrix

matrix H of order n.

V(H) 1 v(K)

d, d > 0, then

dsn-k,
If d = n -

of order k of the

Then

n(H) 2 Z(K),
If d(H) = d(K) -

we obtain

x(H) 2 n(K) + d,

V(H) 2 y(K) + d.

k, and In K = (~5,q, z), we have the exact formula,

InH=(t,++z-k,q+lz-k,z-n-fk).

If H is nonsingular,

then z = n -

InH=@+nii.

BLOCK

19.

MATRICES

OF

HIGHER

It is clear that Theorems

k, and we have
k,q+fI-k,O).

ORDER

1 and2 could be applied repeatedly

tain formulas for the inertia of an Hermitian

Linear

Algebra

and

Its

Applications

(W

1, 73-81

matrix partitioned
(1964)

to ob-

into a t x t

INERTIA

block

OF PARTITIONED

matrix.

nonsingular

HERMITIAN

For suppose

H = (Hij), i, j = 1, . . . , b, with at least one

diagonal block, which we may assume without loss of generality

to be Hi,.

We repartition

H so that

H=
where

B = (HJ,

Theorem

of H,, in H, and has block

i, j = 2, . . . , t.

be computed

B*H<lB

order t -

by block

1.

For

example,

Then

bl

is the Schur

The submatrices,

multiplication:

G, = Hzj -

i, j = 2,. . ., t. Then if G has at least one nonsingular

we can continue

20.

J3
C
1

j = 2, . . . , t, and C = (HJ,

of G can easily

H;H,H,,,
block,

H*,
H*

1, In H = In H,, + In G, where G = C -

complement
G,,

81

MATRIX

diagonal

in the same manner.

if

H=

with

H,,

nonsingular,

where G,, = H, -

then

H,*iHllHlj

the

Schur

complement

(i, i = 2, 3).

and we let K,, be the Schur complement

Finally,

of H,,

is

if G,, is nonsingular

of G,, in G, Ii,, = Gas -

G,*,G.;G,,,

we haye
In H = In H,, + In G,, + In K,,.
ACKSOWLEDGMENTS
This work was done under contract
with the Institute of Mathematics,
Prof.

A. M. Ostrowski

DA-9-591-EUC-3686

University

for very helpful

of Basel.

of the U.S. Army

The author wishes to thank

discussions.

REFERENCES
1 Alston A. Householder,
York, 1964, p. 76.

Theory of .Watrices in Xwnerical Amzlysis, Blaisdell, New

2 W. V. Parker and J. C. Eaves, Matrices,

l<eceived August

Ronald

Press, New York,

1960, p. 69.

76, 1967

Linear Algebra

and Its Applications

1, 73-81

(1968)