760

MA THEMA TICS: K. FAN

PROC. N. A. S.

always solvable uniquely for the integer y since (i) 01(Xx)/d is an integer,
(ii) X/d is prime to r/d.
Algorithmically, to apply (19.1), we may di'tide by d in those cycles of 01
which consist of multiples of d (simply omitting the other cycles, in which
none of the numerals is divisible by 1), and then multiply each numeral by
X', where (X/d)X' =- 1 (mod rld), reducing the result to its least non-negative residue mod rid.
If 0 is an exponential substitution, its derivative 01 is linear: 01(x) =,Bx
(mod r), and the condition: dj0i(dx) for all x, is obviously fulfilled. We
have for the derivative of 0' = 0X:80(x) = ,Bx (mod r/d). Hence:
THEOREM XXVII. Any power 0 of an exponential substitution 0 is a
special substitution, exponential if 1a 1 (mod r/d), linear if 13 = 1 (mod
rld), where d = (X, r).
Of course, any power of a linear substitution is linear. But the property
of the last theorem is not characteristic of exponential and linear substitutions. E.g., in the case of the following subsitution on the range 32:
0 = (0) (16) (1, 7, 9, 31, 17, 23, 25, 15) (3, 29, 27, 5, 19, 13, 11, 21)
(2, 6, 18, 22) (10, 30, 26, 14) (4, 12). (8, 24) (20, 28),
every power of 0 is special, but 0, having the derivative
01 = (0) (2) (4) (6) (1 3 5 7),
of type 3, is itself of type 4, neither exponential nor linear.
The- necessary and sufficient condition that every power e of a given
special substitution 0 be itself special is that, for every divisor d of r, the
cycle of 01 containing d shall consist entirely of multiples of d. Thi equirement is met, in particular, if r = pa, p a prime, and p is a fixed numral of
01. In the example just given, r = 23, and the numeral 2 is fixed in 01.
1 Vol. 37, nos. 9, 10 (September, October, 1951).
2

The parenthesis notation is employed for greatest common divisor.

The contrary case has already been disposed of in Note II, 13.

MAXIMUM PROPERTIES AND INEQUALITIES FOR .THE

EIGENVALUES OF COMPLETELY CONTINUOUS OPERATORS*
By KY FAN
DEPARTMENT OF MATHEMATICS, UNIVERSITY OF NOTRE DAME

Communicated by John von Neumann, September 8, 1951

1. Although several extremum properties of the eigenvalues of completely continuous (abbreviated to c. c.) operators in Hilbert space D are
know;n, nearly all of them deal with a single operator. In this respect, the

VOL. 37, 1951

MA THEMA TICS: K. FAN

761

following theorem of J. von Neumann' is of particular significance, as it involves two matrices: Let A, B be two square matrices of order n (with complex
coefficients). If { Xi }, { } are the eigenvalues2 of A *A, B*B respectively, then
E VXKis the maximum oft tr (UAVB)I, when U, V run over all unitary

IKc

matrices of order n. In the present note, Theorem 1 contains two maximum

properties (4), (5) for any finite number m of c. c. operators in Hilbert
space !. (4) is a generalization of von Neumann's theorem, while (5) is a
multiplicative companion of (4). The case m = 1 of (5) includes the important inequality of H. Weyl3 k12. . j2 < X1X2 .. . Xn (n = 1, 2, *)
between the eigenvalues { b } of a c. c. operator A in p and the eigenvalues { Xi
of A *A. The case m = 2 of (5) includes the following theorem of A. Horn4
concerning two c. c. operators A, B in ,: If Xi , { Ki }, {,/1I are the eigenvalues of A*A, B*B and (AB)*(AB), respectively, then . . . .n < X1X2 * X
XnKjK2 ... icn for n = 1, 2, .... The proof of our Theorem 1 is based on
this result of Horn and the elementary Lemma 1. Thus, in the case of
finite matrices, our proof of (4) differs from von Neumann's more geometric
proof.
For any two c. c. operators A, B in ,, inequality (10) in Theorem 2 gives
a comparison, essentially different from Horn's, of the eigenvalues of
(AB)*(AB) with those of A*A and B*B, while inequality (11) compares
the eigenvalues of (A + B)*(A + B) with those of A*A and B*B. From
(11) we derive Theorem 3 on asymptotic distribution of the eigenvalues of
(A + B)*(A + B). In the case of linear integral equations with symmetric
kernels, both (11) and Theorem 3 are well-known results of Weyl.5
As an application of the maximum property (4), we have in Theorem 5
another inequality comparing the eigenvalues of (A + B)*(A + B) with
those of A *A, B *B and involving an arbitrary symmetric gauge function.
This generalizes a result of von Neumann' concerning finite matrices. To
derive Theorem 5 from (4), we first prove Theorem 4 concerning a general
inequality for symmetric gauge functions. The geometric meaning of
Theorem 4 is clear. Analytically it is analogous to a classical theorem of
R. F. Muirhead,6 although the class of functions involved in his theorem is
entirely different from the class of symmetric gauge functions.
... > bx 2
2. LEMMA 1A. Let a, 2 a2 2 . . .> a. 0, bi
b2
0. If pij (1 < i, j < n) are n2 non-negative numbers such that

{I

Z pi < l forallj,
i=1

E pj < l forall i,

j=l

(1)

then
n

EjE p-

i=l jl

bj <

i
i t1

as b.

(2)

MA THEMA TICS: K. FAN

762

PROC. N. A. S.

Let ci 2 E pj b, (1 < i < n), For a fixed positive integer

ji
n
k
1 j n). ThenO<qj<1and qj<k. Since
k <n,letqj= p2 (1<
b, . bj+1 0, we have

Proof:

JS

k-1\

k-1

EE q.
jmi

k-1

q,)bk<
E(bJ-bk) + kbk
bj + (ki-i
i
\j

or
k

Sbi
E cj:5 ji1
ji-

(3)

(1 < k<n).
x

By Abel's lemma, ai 2 aj+j 2(0 and (3) imply E ajcj< E atbi

i-1
i=1
We shall use the above lemma in the following form:
LEMMA 1B. Let {at }, { bi } be two non-increasing infinite sequences of nonnegative numbers such that , a,bi < + . If pv (i, j = 1, 2, .. .) are noni=l

negatwe numbers such that E pt, S 1 for all j and

go

co

pq :< 1 for all i, then

co

~
i=l}l

~~

1i

THEOREM 1. Let A1, A2, .. . , Am be m (.1) c. c. operators in Hilbert

space ,. For each j = 1, 2, ., ., m, let X,j Xj2,.. ., j, . .. be the eigenvalues2
of A;Aj. Then for any positive integer n:
n1
n
Max Ejn~~~~~~~~~~
(U1A1.. UmAm xi, xi) =jj(IX1fX2. XmJ)"',2 (4)
Max

Det (U1A1.. . UmAm xi,) 2 = II

1 .i, k 5n

ijI

(NA>q

. XMi); (5)

U2, Ur independently run over all unitary

where, for both maxima, U .,
operators and { Xi, x2, .., xX } runs over all sets of n orthonormal elements in $!.
Proof: For each j = 1, 2, . . ., m, let Hj be the non-negative square root of
Ai A. Then Hi is also c. c. and its eigenvalues are VJ{ /j} . Let y1t (1 . i
< n) be n orthonormal elements in f such that Hjyji = VXyji (1 . i < n).
Let Ai = WjH1 be the polar factorization of von Neumann.7 Here Wj is a
partially isometric operator and its initial set is the closed linear manifold
determined by the range of Hj. We have then W,*A, = Hi. Since W: is
also partially isometric and its initial set is the closed linear manifold determined by the range of Aj, we can find a unitary operator Vj in such that
VAjyji = W*Ajyii = H1y,, (1 < i < n). Let Zj be a unitary operator
in t such that Zjyj. = Yj-I, i (1 < i . n), where we put yoi = ymi. Let now

763

MA THEMATICS: K. FAN

VOL. 37, 1951

Uj = Zj Vj (1 < j < m). Then U1A . ... UmAryr.n = (X12B2t . . . Xmt) /Ymt
(1< i < n). Hence, for the particular choice of Uj = Zj V, (1< j < m) and
Xi= Ym (1 < i < n), the relations

(U1A,

UmAnxi, xi)

<

... UmAmXi, Xk)

...

Det (UA
i

1i, kxn

(X1tX2t *..* j)1/2

(6)

II (X2t .. Xrn)

(7)

<

i=1

are satisfied and actually become equalities.

It remains to show that inequalities (6), (7) hold for any m unitary operators U1, . . ., Um and any n orthonormal elements xi, . . ., x. in t. We first
prove the case m = 1 of (6):
E (UA ixi, xi) < E V;.
(8)
i-1
5=1
Consider the polar factorization of A1= W1H1. Let
be a complete
orthonormal set such that H1Ylk = A/X y,. (k = 1, 2, 3, .. . Then,
since (U1A1xj, xi) = (H1xj, WU*4xj), we have

Iy I}

E
i =1

(U1A xi, xi)

OD

E E A/it (x, Y.)(YL W;U*1 Xi)

i = 1 ke1

and therefore

(U1A1x,, Xj).<

(Xi

I V

yR)

2+

i=l k=1

\ /ZL

(wiYLko lX,) 2.
AVkX. Thus (8) is

By Lemma lB, each of the last two double sums is .

:
i=l

proved.
Let now B = U1A, . .. UmAm and let
be the eigenvalues of B*B.
Since (UfA ) *(UjAj) = A4A j, repeated application of Horn's theorem gives

IKil

II
i=1

Ki <

II
i=l

(X1iX2i .

Xmf), (n

1, 2, .)

(9)

n
n
which, according to a lemma of Polya,8 implies (i V< lijX2... .Xmi) 1/2
Hence (6) will result when we apply (8) to B.
According to Lemmas 2, 3 in an earlier note9 by the author, we have for
any n orthonormal elements xl, x2, .. ., xx:

Det (Bxi, Xk) 2 < Det (B*Bxi,

kds
l (7 i,
wci, k<n 9
which, together with (9), yields (7).

Xk) <

II Ki,

764

MA THEMA TICS: K. FAN

PROC. N. A. S.

3. THEOREM 2. Let A, B be two c. c. operators in Sg. If {xj, {Ki},

{ }, o I} are the eigenvalues of A *A, B*B, (AB)*(AB), (A + B) *(A + B),
respectively, then for any two non-negative integers m, n:

(10)

/'m+n+1 < Xm+i*Kn+l

VOm+n+l < \/m+l + VKn+l1

( 1)
Proof: Let { xs }, { yi be two complete orthonormal sets such that
AA*xi = Xixi, B *Byi = Kjyi (i = 1, 2, ... .). Let AB = WH be the polar
factorization of AB, where H is the non-negative square root of (AB)*(AB)
and Wis partially isometric. For any z of ID, we have (Hz, z)2 = (W*ABz,
If z is subject to conditions |IzlI = 1,
Z)2 < (AA*Wz, Wz)(B*Bz, z).
=
<
<
and
i
0
(z, yj) = 0 (1 < j < n), then (AA*Wz, Wz)
m)
(1
(z, W*xi)
<
l2Xm+1 < Xm+i, (B*Bz, z) < Kn+l and therefore (Hz, Z)2< Xm+1.Kn+I.
Thus we get (10) by applying the minimum-maximum principle to H. The
proof of (11) is similar.
The next theorem deals with asymptotic distribution of eigenvalues.
{X }, at be
THEOREM 3. Let A, B be two c. c. operators in ?. Let {Ix,
B*B
and
If, for
A
respectively.
B)
+
*(A
+
B)
(A
*A,
the eigenvalues of
we
r
have
0,
>
some exponent
(12)
limnTX7 = Xo < + o and lim nrK, = 0,

IIWZl

{I

71-0n-ao

then
lim nron = W.

(13)

Proof: Let h be a fixed positive integer. By (11), we have

\1T(h+l)n+j < V/Xhx+j + V'K,+l (n = 1, 2, 3, . . .; j = O, 1, . . ., h),
which combined with (12), gives
(14)
lim nr/2 a.'/2 < [(h + 1)/h]T/2 .0/2.
If we apply (11) to A = (A + B) + (-B), we get inequality
VX(h+ l)n+j - VKn +l, which together with (12) yields
Jim nr/2 On '/2 >[h/(h+ 1)]r/2 1/2

Va+,+j >
(15)

Then (13) results from (14), (15) by letting h -- co.

4. We turn now to a general inequality for symmetric gauge functions.'0
THEOREM 4. Let a, 2 a2 2 ... > an> 0, bl2 b22 . .. . bn 2 0. Then a
necessary and sufficient condition that
'1'(al, a2, . ., an) < '(bi, b2, . . ., bn) (16)
for all symmetric gauge functions 4' of n variables, is that

VOL. 37, 1951

MA THEMA TICS: K. FAN

k

765

(1< k <n).

Eat < bi

(17)

The necessity of (17) is immediate, when we consider, for each k, the

particular symmetric gauge function 4' defined by
4(41 {2, - -.,

Max
il<h<

...

<ik

(I0fjj

+ 1021 +

-+Jkj)

For the proof of the sufficiency of (17), we only indicate that it results from
the following two lemmas.
LEMMA 2. Let { bi } be a set of n real numbers. Let O < a < 1 and let h, j be
two distinctfixed indices. If

ch = abh + (1- a)bj,

cj = (1- a)bh + abj,

cj

b1fori $ h, j,
(18)

then 4'(c1, C2, . . ., Cn) < 4(b, b2,.. ., bn) for all symmetric gauge functions 4' of n
variables.
LEMMA 3. Let {as }, { bi } be two sets each of n real numbers satisfying (17)
and a1 2 a +i. Then by the successive application of a finite number of substitutions ofform (18), {b1} can be reduced to a set {diJ such that as < di (1 < i <
n) 11
5. THEOREM 5. Let A, B be two c. c. operators in !. If
{Ki} and
{e} are the eigenvalues of A*A, B*B and (A + B) * (A + B) respectively, then

{IXi,

4'(V'

V/J2,..., -V-.) < 4(D

V2.,

/Xn) + (4D(VKQ VK,. VK. )

(19)

holds for every symmetric gauge function 4b of any number n of variables.

Proof: By Theorem 4, we need only to prove the following special case of
(19):

ZV'.ZV
< X +ZV
i-'

i-i

Ki.

i=l

(n

1, 2, 3. ).

(20)

Consider the polar factorization of A + B = WH, where H is the nonnegative square root of (A + B)*(A + B) and W is partially isometric.
Let zl, z2, . . ., z2 be n orthonormal elements such that Hz1 = V/aZ1 (1 < i <
n). We can find a unitary operator U such that U(A + B);z = W*(A +
B)z1 = Hz1 (1 < i < n). Then
n

E "i =
_

(U(A + B)zi, z) = j (UAzi, zf) + E (UBzt, Zi)

and (20) follows from

766

MATHEMATICS: A. D. MICHAL

(UAzi, zi)

i=l

.
< E,

~~i=l

Zif

il

(UBzf, z)

PiLoc. N. A. S.

.
<

i=

K,

which are true by (4).

* This work was supported in part by the Office of Naval Research.
1 Von Neumann, J., "Some Matrix-Inequalities and Metrization of Matric-Space,"
Tomsk Univ. Rev., 1, 286-300 (1937).
2 The eigenvalues will always be arranged in order of decreasing absolute value.
3 Weyl, H., "Inequalities Between the Two Kinds of Eigenvalues of a Linear Transformation," these PROCEEDINGS, 35, 408-411 (1949).
4 Horn, A., "On the Singular Values of a Product of Completely Continuous Operators," Ibid., 36, 374-375 (1950).
6 Weyl, H., "Das asymptotische Verteilungsgesetz der Eigenwerte linearer partieller
Differentialgleichungen," Math. Annalen, 71, 441-479 (1912).
6 Hardy, G. H., Littlewood, J. E., and Polya, G., Inequalities, Cambridge University Press, Cambridge, 1934, pp. 4 -49.
7Von Neumann, J., "Uber adjungierte Funktionaloperatoren," Ann. of Math., 33,
294-310 (1932).
8 Polya, G., "Remark on Weyl's Note:* Inequalities Between .the Two Kinds of
Eigenvalues of a Linear Transformation," these PRocsBwNGs, 36, 49-51 (1950).
* Fan, K., "On a Theorem of Weyl concerning Eigenvalues of Linear Transformations, II," Ibid., 36, 31-35 (1950).
10 Following von Neumann, a gauge function 0 (in the sense of Minkowki) is called
symmetric, if '(h,, , ...,tn) = I(eEj1i, j *--, entj,.) for any combination of signs
ei = i 1 and for any permutation (j1, j2, .... jjn) of (1, 2, ..., n). For general properties
of symmetric gauge functions, see Schatten, R., A Theory of Cross-Spaces, Princeton
University Press, Princeton, 1950, pp. 84-92.
11 The proof of Lemma 3 is similar to the corresponding part of the proof of Muirhead's theorem (see the reference in footnote 6). Using Lemma 3, one can also prove:
For two sets a, bt I each of n non-negative numbers arranged in descending order,
condition (17) is equivalent to the existence of ns non-negative numbers PjN satisfying

(1) and at =I

pjbj (1 < i < n).

INVARIANT DIFFERENTIAL FORAMS IN SEVERAL GROUP

VARIABLES AS SOLUTIONS OF PARTIAL DIFFERENTIAL
EQUATIONS IN FRJECHET DIFFERENTIALS*
BY ARISTOTLE D. MICHAL
CALIFORNIA INSTITUTE OF TECHNOLOGY

Communicated by G. C. Evans, September 4, 1951

1. Introduction.-There is a particularly interesting class of partial differential equations in Frkhet differentials that characterize abstract invariants and in particular differential forms that are left invariant by the
parameter groups vr and r2 of a differentiable abstract group A. We shall