ON THE MAXIMUM PRINCIPLE OF KY FAN

M. MARCUS AND B. N. MOYLS

1. I n t r o d u c t i o n . In 1951 F a n (1) proved the following interesting extreme

value result: Let Ai, . . . , Am be completely continuous operators on a Hilbert
space § . For v — 1, 2, . . . , m let \ai > X ^ + i be the characteristic roots of
Aff* Aa. Then, for any positive integer k,
k k / m \j
X) (UiAi . . . UmAmXi, xt)
a) i=l 2=1 \ ff=l /

(2) max |det{ (UiAi . . . UmA

i, .7=1,2 k
)ir=nnu i=l <r=l
where both maxima are taken over all unitary operators U\, . . . , Um and all sets
of k orthonormal (o.n.) vectors.
F a n proved (1) for m = 1 and then applied an inequality of Polya (7)
and a recent result of H o r n (3) to obtain the theorem for m > 2. This result
is a generalization of a result of von N e u m a n n (10), which states t h a t if
A and B are n-square complex matrices with singular values (12) at > ai+u
(3t > (3i+i (i = 1, 2, . . . , n — 1), then
n
(3) m a x | t r ( £ L 4 r a ) | = 2 > < 0<,
2=1
where the maximum is taken over all unitary U and V.
In this paper we shall confine our a t t e n t i o n to the case of finite matrices.
W e shall show t h a t both (1) and (2) are special cases of a general m a x i m u m
result for compound operators (Theorem 3). As applications we obtain
inequalities analogous to those of H o r n (3) and Ostrowski on the singular
values of a product. An inequality of S. N . Roy (8) (later published with a
different proof b y B. Sz. N a g y (9)) is a special case of Theorem 3 and Ostrowski's
inequalities (6) connecting Schur-convex functions of singular values and
characteristic roots.

2. F a n ' s first r e s u l t . Before proceeding, we point out t h a t for finite

matrices (1) follows immediately from (3). Following F a n we need only prove
(1) in the case m = 1.
Let Xi, . . . , xk be an o.n. set of vectors, k < n, and let P denote the ortho-
gonal projection into the subspace spanned by Xi, . . . , xk. T h e n

Received May 23, 1956. This research was supported by the U.S. Air Force under contract
No. AF 18 (603)-83.
313

Downloaded from https://www.cambridge.org/core. 25 Oct 2021 at 08:37:34, subject to the Cambridge Core terms of use.
 814 M. MARCUS AND B. N. MOYLS

X) (UiAiXi, xt ) ]£ (UiAixitPxt)
k

J2 (PUiAiXi,Xi)
= |tr(PCMi)|
<max|tr(7Pi/ii4i)|
Ui.V
n
< X Pi V^ii
by (3), where the pj are the singular values of P. Since p ; = 1 for j = 1, . . . , k,
and p:j = 0 for j > fe -f 1, (1) follows for m = 1.

3. Properties of the Grassmann algebra. Let xt{i = 1, . . . , r), 1 < r < n

be vectors in unitary w-space Fw. Then
(4) S = Xi A X2 A . • • A XT

denotes the Grassmann exterior product of the xt] it is a vector in Vm where

m = ("). If -4 is a linear transformation of Vni the rth induced compound of
A is a linear transformation of Vmi defined by
(5) Cr{A) xi A . . . A xr = Axi A . . . A Axr.
The following properties, which we list for later reference, can be found in
(11, pp. 63-67):
(i) Oi A . . . A xr, yi A . . . A ^ r ) = d e t { ( ^ , ^ ) | f i ; , i r.
(ii) cr{AB) = CM) Cr(B), CM'1) = CrM), CM*) = [CM)]*.
where A* is the transposed conjugate of A.
(iii) The characteristic roots of Cr(A) are the (nT) products of the character-
istic roots of A taken r at a time.
(iv) If A has any of the following properties, so does Cr(A): non-singular,
normal, Hermitian, non-negative Hermitian (n.n.h.), unitary.
To prune some of the foliage of indices usually necessary in discussing
these objects we introduce some notation. The set of (*) distinct sequences of
positive integers iu . . . , iT satisfying 1 < ii < . . . < ir < k will be denoted
by Qkn and a typical such sequence will be called w. If #i, . . . , xk is a set of k
vectors in Vn, and œ is the set {iu . . . , ir) in Qkr, we set
X(£ Xjj /\ . . . /\ Xir .

If (a) is a set of n numbers &i, . . . , an, Er(a) = Er(au • • • > #n) will denote
the rth elementary symmetric function of (a).

4. Results.The following theorem is a generalization of the result in (5):

THEOREM 1. Let H be n.n.h. with characteristic roots hj > hj+u j = 1> • • • >
n — 1. Let f(t) = ts, s real, and

Downloaded from https://www.cambridge.org/core. 25 Oct 2021 at 08:37:34, subject to the Cambridge Core terms of use.
 THE MAXIMUM PRINCIPLE OF KY FAN 315

<j)(xlt . . . , * * ) = X) f[(Cr(H)xm Xu)]t

(aeQkr

where 1 < r < k < n. Then, if s > 1,

(6) max0 - Erl/(W,...,/(A*)]î
and if 0 < 5 < 1,
(7) min0 = Er[f(hn-k+1),...,f(hn)];
where the maximum and minimum are taken over all sets of k o.n. vectors
Xi, . . . , X/c in Vn.
Proof. For the case s = 1, the theorem is proved in (5). The present proof
is an application of this result to f(H).
First note that 1 if ||x|| = 1,
(8) f[(Hx,x)] < (f(H)x,x).
For, if Xi, . . . , xn are o.n. characteristic vectors of H corresponding to
hi, . . . , h,n then
MHx,x)]=f\ £,hj\(x,xj)\
L j=l
n

< Ç/(^) l(*,^)|2

= lf(H)x,x).
Secondly,-
(9) " f(Cr(H)) = Crif(H)).
For, if U is a unitary matrix such that U*HU = D, diagonal,
f(CT(II)) =f(Cr(UDU*)) = Cr(U)f(CT(D)) Cr(U*)
= Cr(U) Cr (/(£>)) Cr(C/*) = Cr (J(UDU*)) = Cr(f(H)).
To establish (6) we observe that by (8), (9) and the remark at the beginning
of the proof,
4>< E (KCr(H))x„,x„) = E {Cr{f(H))x„,x„)
<Er(f(h) f(ht)).
Equality can be achieved by choosing Xi, . . . , xk to be an o.n. set of character-
istic vectors of H corresponding to hi, . . . , hk, since by (i),

= z/(det{A,. «,.„}) = z Xn*«.)

(7) is proved similarly.
x
This inequality holds for any continuous convex function / defined on the spectrum of H.
The inequality is reversed if / i s concave.
2
This equality holds for any function / defined on the spectrum of H and satisfying the
relation/(Vy) [ = f(x) f(y).

Downloaded from https://www.cambridge.org/core. 25 Oct 2021 at 08:37:34, subject to the Cambridge Core terms of use.
 316 M. MARCUS AND B. N. MOYLS

Theorem 3 is our main result. Theorem 2 is a special case. However, it is

the first step in the proof of the general theorem, and its proof appears to be
of some interest in itself.
THEOREM 2. Let A be an arbitrary n-square complex matrix with singular
values at > ai+i(i = ! , . . . , « — 1). Let

</>0i, . . . ,xk; U) = £ (Cr(UA)xU9xn)

weQkr

where 1 < r < k < n. Then

(10) max</> = Er(au . . . fak),

where the maximum is taken over all sets of k o.n. vectors Xi, . . . , xk and all
unitary U.
Proof. Since A = VH, where V is unitary and H = (A*A)^ is n.n.h., we
may without loss in generality replace A by H. Set y$ = U~lxj} (j = 1, . . . , fe).
By the notation ^,®Cr(A) we mean the direct sum of Cr{A) taken (kr) times;
and similarly for J^®xu. Since ^®Cr(H) is n.n.h. by (iv),

=
^ I 2 (^r(H)Xœ, Criu^Xu)]

( E © cr(H) E e x*, E © y.)

< ( E ©£>•(#) E ©*». E ©*«)*
(E ©cr(H)Z ey-, E 0y-)é
= E (Cr(#)xM,*w) M E (C,(Jny-.%) f
Lo>eQ^ r J LœeQkr J
< £ r (ai, . . . , ak).
The last inequality follows from (6) applied to each of the square roots.
Equality holds when the Xj are suitable eigenvectors of H and U is the
identity matrix.
THEOREM 3. Let A\, . . . , Am be arbitrary n-square complex matrices with
singular values avi, where a^i > aaii+i (i = 1, . . . , n — 1 ; a = 1, . . . , m).
Let

4>(xu . . . ,xk; Ui, . . . , Um) = X) (Cr{UiAi . . . UmAm)xœ, xa) ,

otQkr

where 1 < r < k < n. Then

I m m \
(11) max
* = ^ i n <**!> • • • » n ^ j »
\ (T=l (7=1 /
where the maximum is taken over all sets of k o.n. vectors xi, . . . , xk, and all
unitary matrices Ui, . . . , Um.

Downloaded from https://www.cambridge.org/core. 25 Oct 2021 at 08:37:34, subject to the Cambridge Core terms of use.
 THE MAXIMUM PRINCIPLE OF KY FAN 317

Proof. For a fixed £/2, . . • , Um, let fa > fa > • • • > &i be the singular
values of 5 = ^41(t/2^42) . . . {UmAm). By Theorem 2
(12) <>
/ < E r (/3i, . . . , & ) .
The singular values of £7,^4* are the aai (i = 1, . . . , n). By a result of
Ostrowski (6, p. 283, equation 106a),
/ m m \

ET{pi &) < EX n «.i n «.*) •

There remains to show that the right side of (11) is taken on for a suitable
choice of the x's and C/'s. As in Theorem 2 we may replace each Aff by
iJcr = 04ff* ^ L ) i Let xai, . . . , xan (a = 1, . . . , w) be an o.n. set of character-
istic vectors of Ha corresponding to OCff l , . . . , (X . Choose the Uff so that an

U(fXai X(j—lfj, (T = 2, . . . , m,
t-^ 1 *^1 i ^mii

for i = 1, . . . , n. Set
Xffo Xfffi / \ . . . / \ Xcrfr

and

3=1

where co is the set {i\, . . . , i r }- Then

utQkr i

/ J Oimoi ^ r \ U \ - k l \ . . . (Jm—l±±m—\) Xm—ii(a, Xma)

:=:
/ f &m<>) • • • OLl(a\pCirwi %mi>))

• / m m \

= 2 1 n«»<i • • - n a**..)
\ <r=l <r=l /
/ m m \
= Er\ Yl OLaU • • • i I T ««r* ) •
\ er=l <r=l /

This completes the proof of the theorem.

Remarks. I. By setting r — 1 in (11) we obtain (1); by setting r ~ k we
obtain (2).
II. Theorem 3 does not follow immediately from (1) applied to the
compound because the lexicographic ordering of the eigenvalues does not
necessarily correspond to the ordering by magnitude.
III. Let Aa, a = 1, . . . , m, have eigenvalues \ai, ordered so that
\Ki\ > |X*,*+i| (i = 1, . . . , « — 1). We can find unitary Va such that
V<r* AaVa = Ta, triangular, with the diagonal elements ordered by absolute

Downloaded from https://www.cambridge.org/core. 25 Oct 2021 at 08:37:34, subject to the Cambridge Core terms of use.
 318 M. MARCUS AND B. N. MOYLS

magnitude. Then if U„ = Va-\ Vff* (a = 2, . . . , m), Ui = Vm UVi* where U

is the diagonal matrix with

exp y - Z) arg ^*j

in the ith row and ith column; and if Xj = F m ^ (j = 1, . . . , fe), where ^ ;
is the unit vector with 1 in the ith position; then
/ _m_ m
\
2-J \^T\\J\A\ . . . UmAmjXa, X u)
coeQkr
Srin ix»ii n iXrti).
It follows from Theorem 3 t h a t
/ _W_ _m_ \ / _W_ //;
\

i3) EA n iXrti, • • •. n iXrti ) < M n «»i. • • •. n «,*).

The case rn = 1, r = k is Weyl's inequality. Actually, (13) also follows easily
from Ostrowski's discussion of Schur-convex functions.

5. Applications. Let X = AB, where A and B are arbitrary w-square

complex matrices. Consider the convex sum
Y = <JX*X + <5XX*,
where a + b = 1, cr > 0, <5 > 0. Let <£* > #*+i, «i > « 2+ i, fit > /?H-I
(i = 1, . . . , n — 1) be respectively the non-negative square roots of the
characteristic roots of Y and the singular values of A and B. In the Theorem 4
we shall use the following concavity property of the elementary symmetric
functions:3
/ / (a) and (b) are sets of n non-negative numbers and 1 < r < n, then
ElTlT{a + b)> E\lr{a) + E\,r(b).
THEOREM 4. For 0 < 5 < 1 and 1 < r < k < n,

Er(<f>n i • • • > 0w-fc+l)

/ r \ 2s
«w-;-+l Pn-j+l) E°r(PnS, • • • » A i - A + l ) Er((XnS', . . . , aJLA+i).

Proof. Let Xi, . . . , x& be an o.n. set of characteristic vectors of F corre-

sponding respectively to
0n» • • • » 0n-*+l»
Then

3
This property follows from a concavity property proved by W. Fenchel (2). This inequality
also follows directly as a corollary of a similar property for symmetric functions proved by
Lopes and Marcus in (4).

Downloaded from https://www.cambridge.org/core. 25 Oct 2021 at 08:37:34, subject to the Cambridge Core terms of use.
 THE MAXIMUM PRINCIPLE OF KY FAN .319

EV'tâ <^l*+i) = E\"[{Yx„ xj)s]

= E\'T[{<T{X*X X„ Xj) + Ô(XX*Xj, Xj) } *]

> E1r/r[2s-1{<rs(x*Xxj,xJy + bs{xx*X],x}y\]

> 2s-1{ExTlr[^{X*Xxj, xj)s] + El/r[Ôs(XX*Xj, xj)']}
1 , s
= 2'- ((; £!"[(I*IÏ/,Ï,)'] + 5 E1r/T[(XX*Xj,XjY]}.

It follows from the Hadamard determinant theorem that

Er[(X*Xxj,Xjy]= £ (U(X*Xxtnxtl))S
Kil<...<ir<k\ *=1 /
> E (cr(x*x)x„,x»y.
aeQkr
Thus

E\'\&, . . . , 0nl*+l) > 2 - 1 < r S | E (Cr(X*X) X„, OC*)']'

+ 2s-1 Ssj D (Cr(***) %,, x„)s}1/r

= 2s-1 ^ J E (c r (^*^) cr{B) xw cr(B) Xuyy

+ 2s-1 5 S | E (Cr(5B*) C,{A*) x*, Cr(A*) xay)

+ 2s-1 5 s (n /£-*-ij ''{Z (C,(^*) *., xj 8 } 1 ''

> «r^- 1 ^ ( n <*^+1)2/r ^"(/a:*,..., A-*+i)}
The last inequality follows by Theorem 1. The result follows by a classical
inequality and by taking rth powers of both sides.
We remark that when a is zero, v* is to be taken as 1.
Theorem 4 may be used to relate the characteristic polynomials of the
matrices involved. For example,

\pr({ABB*A*)s)\ > 2 r ( - 1 ) n «-I** |/>r((5*B)*)|,

where

Xn + E Pr(M) Xn~r

is the characteristic polynomial of the matrix M.

Downloaded from https://www.cambridge.org/core. 25 Oct 2021 at 08:37:34, subject to the Cambridge Core terms of use.
 320 M. MARCUS AND B. N. MOYLS

In (9) B. Sz. Nagy proves that if A and B are complex n-square matrices,
then
(14) «„/?„ < |X«| < a i 0 i ,
where at > a*+i, Pi > /3i+i are the singular values of A and B, and \ t ,
\\i\ > |X i+ i|, are the characteristic roots of AB. By Theorem 3,
(15) ErflAil, • • • , M ) < £ r ( a i /3i, . . . , a* A),
for 1 < r < k < w. By setting r = fe = 1, we obtain the upper inequality of
(14). If an fin = 0, the lower inequality is trivial. Otherwise A and B are non-
singular. The characteristic roots of (AB)'1 are the X*-1, and the singular
values of A'1 and B~l are the a f 1 and jS*-1. Hence

(16) ET(\\n\ , • . • , |Xre_fc+i| ) < Er(an (3n , . . . , aw_£+i ^ w _ fc+ i).

Again setting r = fe = 1, we obtain the lower inequality of (14).
Both (15) and (16) have immediate generalizations in two directions:
first, to a product of more than two matrices, and second, to the more general
class of Schur-convex functions.

REFERENCES

1. K. Fan, Maximum properties and inequalities for the eigenvalues of completely continuous
operators, Proc. N.A.S. (U.S.A.), 37 (1951), 760-766.
2. W. Fenchel, Generalisation du théorème de Brunn-Minkowski concernant les corps convexes,
C.R. des Sci. de l'Acad. des Sri., 203 (1936), 764-766.
3. A. Horn, On the singular values of a product of completely continuous operators, Proc.
N.A.S. (U.S.A.), 36 (1950), 374-375.
4. M. D. Marcus and L. Lopes, Inequalities for symmetric functions and Hermitian matrices,
Can. J. Math., 9 (1957), 305-312.
5. M. D. Marcus and J. L. McGregor, Extremal properties of Hermitian matrices, Can. J.
Math., 8 (1956), 524-531.
6. A. Ostrowski, Sur quelques applications des fonctions convexes et concaves au sens de I.
Schur, J. Math, pures et appl. (9), 31 (1952), 253-292.
7. G. Pôlya, Remark on WeyVs note: Inequalities between the two kinds of eigenvalues of a linear
transformation, Proc. N.A.S. (U.S.A.), 36 (1950), 49-51.
8. S. N. Roy, A useful theorem in matrix theory, Proc. Amer. Math. S o c , 5 (1954), 635-638.
9. B. Sz. Nagy, Remark on S. N. Roy's paper: A useful theorem in matrix theory, Proc. Amer.
Math. S o c , 7 (1956), 1.
10. J. von Neumann, Some matrix-inequalities and metrization of matrix-space, Tomsk Univ.
Rev., 1 (1937), 286-300.
11. J. H. M. Wedderburn, Lectures on Matrices, Amer. Math. Soc. Colloq. Publications, 17
(1934).
12. H. Weyl, Inequalities between the two kinds of eigenvalues of a linear transformation, Proc.
N.A.S. (U.S.A.), 35 (1949), 408-411.

University of British Columbia

Downloaded from https://www.cambridge.org/core. 25 Oct 2021 at 08:37:34, subject to the Cambridge Core terms of use.