Math. Ann.

246, 205-224 (1980)

Annllm
(~1 by Springcr-Verlag 1980

Means of Positive Linear Operators

Fumio Kubo and Tsuyoshi Ando
Division of Applied Mathematics,Research Instituteof Applied Electricity,
Hokkaido University,Sapporo, Japan

I. Introduction
Motivated by a study of electrical network connections, Anderson and Duffin [1]
introduced a binary operation A:B, called parallel addition, for pairs of positive
(semi-definite) matrices. Subsequently Anderson and Trapp [3] have extended this
notion to positive (linear) operators on a Hilbert space and demonstrated its
importance in operator theory. On the other hand, Pusz and Woronowicz [12]
considered a binary operation A # B, called geometric mean, for pairs of positive
operators.
In this paper, taking an axiomatic approach we introduce the notions of
connection and mean. A connection a is a binary operation AaB for positive
operators, which is monotone and continuous from above in each variable and
satisfies the trans]brmer inequality, in the sense
C(AaB) C <=(CA C) a(CBC).
A connection a is a mean if it is normalized, i.e. l a l = 1 where 1 denotes the
identity operator. Such an axiomatic approach is already found in the paper of
Nishio and Ando [11], in which parallel addition is given an axiomatic characteri-
zation. Addition, parallel addition and geometric mean are examples of con-
nections. Though Bhagwat and Subramanian [6] introduced power means, for
instance, {½(Ap + BP)}l/P, these means are not monotone in general, nor satisfy the
transformer inequality.
The plan of the paper is as follows. In Sect. 2 the exact definition of connection
is given. To each connection a there corresponds its transpose a', defined by
(A,B)~BcrA. A natural order relation is introduced for pairs of connection.
Section 3 is the principal part. The key is the existence of an affine order-
isomorphism between the class of connections and the class of (positive) operator-
monotone functions on (0, oe), those functions that are monotone of order n for
every n in the sense of LSwner. This isomorphism a*--~f is characterized by the
relation
AaB = At/2f[A- 1/2BA- 1/2]A1/2

0025-5831/80/0246/0205/$04.00
206 F. Kubo and T. Ando

for invertible A and B. A similar idea of definiting a binary operation with use of a
numerical function is in Pusz and Woronowicz [12]. Thanks to the LSwner theory
on operator-monotone functions, each connection a admits an integral
representation
A a B = a A + b B + ~o Tl + t {(tA) :B} din(t).

When translated to the language of network theory, this means that each
connection is realized by (infinite) series-parallel network connections. An impor-
tant consequence is the concavity
AaB + CaD <=(A + C) a(B + D}.
In Sect. 4 symmetric i.e. a = a ' , means are studied. Arithmetic mean, i.e., the
normalized addition, is shown the maximum of all symmetric means while
harmonic mean, i.e., the normalized parallel addition, is the minimum. The notion
of dual a ± is introduced. Addition and parallel addition are mutually dual.
Section 5 is devoted to characterizations of geometric mean in relation to other
means. As a typical result, if means a and z satisfy the inequality
(AaB) z(Aa±B) <=A~B
then ~ is greater than geometric mean.
In Sect. 6 convergence of means is defined by pointwise convergence of
corresponding operator-monotone functions. Starting with a 0 = a and z o = z the
sequences of means a. and z,, defined by
Aa,B = (Aa n_ t B) a(A z n_ xB)
and
A z,B = (A a,_ t B) z(Av,_ aB),
are shown to converge to the same limit, called the composition. Some special cases
were already considered by Fujii [9,10]. The composition of a mean and its dual
always coincides with geometric mean. As a consequence, geometric mean is
obtained as the limit of alternative iterates of arithmetic and harmonic means.

2. Connections and Means

All operators A, B, C .... in this paper are bounded, linear and positive (i.e. positive
semi-definite) on an infinite dimensional Hilbert space. A scalar multiple of the
identity operator is denoted by the scalar itself; in particular, 1 is the identity
operator. Order relation A ~ B always means the positivity of B - A . That
A 1 ->A 2 > - . . . and A n converges strongly to A is denoted by A,$A.
A binary operation a on the class of positive operators, (A, B ) ~ A a B , is called
a connection if the following requirements are fulfilled.
(I) A < C and B<D imply AaB<CaD.
(II) C(AaB)C ~(CAC)a(CBC).
(III) A,~A and B,~B imply (A,aB,)$AaB.
Means of PositiveOperators 207

An immediate consequence of (II) is

(II)o C(AaB)C=(CAC)c(CBC) for invertible C.
In particular, every connection is positively homogeneous in the sense
a(AaB)=(aA)a(aB) for a>0. (2.1)
A mean is a connection with normalization condition
(IV) t~rl= 1.
The simplest example of a connection is the usual sum A + B. The normalized
sum is called arithmetic mean, in symbol V;
A VB = ½(A + B). (2.2)

Left-trivial mean o9e and right-trivial mean o9, are by definition

AcoeB= A and At~,B= B. (2.3)

Anderson and Trapp [3] introduced the notion of parallel sum A :B and showed
(cf. [11]) that it satisfies (I), (II), and (tlI). For invertible A, B the parallel sum A :B
is given by
A :B=(A -1 + B - l ) - t (2.4)

The normalized parallel sum is called harmonic mean, in symbol !;

A !B = 2(A :B). (2.5)

Remark that in [4] harmonic mean A !B was denoted by A?B.

Pusz and Woronowicz [12] and Ando [4] introduced the notion of geometric
mean A 4~B and showed that it satisfies (I), (1I), and (III). For invertible A, B the
geometric mean A :~ B is given by
A # B = A1/2(A- 1/2BA- U2)l/2AIf2, (2.6)

A positive linear combination of two connections is defined in a natural way; if

a, z are connections and a, b are positive numbers the connection aa + bz is defined
by
A(aa + bz)B = a(A~rB) + b(AzB). (2.7)
Then the class of means becomes a convex set.
A partial order > between two connections is introduced in a natural way;
a > T means by definition that AaB>__AzB for all A and B. Then an important
inequality is ! < 4t=_-<V (see [4]), but this will be discussed later in full generality.
To each connection a corresponds its transpose a' defined by
Aa'B = B a A . (2.8)
A connection a is symmetric by definition if a = a'. Arithmetic, geometric and
harmonic means are symmetric while trivial means o9e and o9, are the transpose of
each other. For any connection a the sum a + a' is always symmetric.
208 F. Kubo and T. Ando

Each connection tr gives rise to a binary operation, denoted by (a), in the class
of connections; (Q, z) ~-->Q(tr)z,defined by
Ap(a) zB = (AeB) a(AzB). (2.9)
Then the following relations are immediate.
o~e(tr)o~,=a and o9,(a)oJe=a'. (2.10)
The non-trivial relation V ( # ) ! = # will be discussed later in full generality.

3. Canonical Representations
A non-negative function f on the open half-interval (0, ~ ) is said to be operator-
monotone if
n m b
~.f(a,)P i< ~ f ( )Qj
1 1

whenever
n
Ea,P,< ~bjQ~,
1 1

where ai>O, bj>0, and Pi and Qj are projections such that

n m

1 1

This definition already implies the continuity, non-decreasingness and concavity

o f f (see [4, 5] and [7]). Let f(0) be defined to be equal to inff(x). Then for each
positive operator A the operator f(A) is defined by functional calculus for
Hermitian operators (see [8]). The condition for f to be operator-monotone can
be stated in the following way
f(A)<f(B) whenever A<B.
The cIass of operator-monotone functions is closed under addition and multipli-
cation by non-negative scalars.
A variant of the L6wner theory on operator-monotone functions is for-
mulated, in the next lemma, in the form convenient to our purpose (see I-4,7]). To
make expressions compact, let us introduce a notation

~b(x,t)= x(l+t----~) for x > 0 and t>=0. (3.1)

x-t- t
For fixed x > 0 the function ~b(x,t) is bounded and continuous on the extended
half-line [0, ~ ] .
Lemma 3.1. The map, m ~ f, defined by
f(x)= S ck(x,t)dm(t) for x > 0 , (3.2)
[ 0 , oo]
Means of PositiveOperators 209

establishes an affine isomorphism from the class of positive Radon measures on

[0, ~ ] onto the class of operator-monotone functions.
Remark that in representation (3.2)

f(0) = inff(x) = m({0}), (3.3)

and
inff(x)/x = m({ oo }). (3.4)

Theorem 3.2. For each connection a and x > 0 the operator lax becomes a scalar.
The map, a ~+f, defined by

f(x)=lax for x>0, (3.5)

establishes an affine order-isomorphism from the class of connections onto the class
of operator-monotone functions.
Proof Let tr be a connection. Following an idea of [11] we shall show first that ifa
projection P commutes with positive operators A and B then it commutes with
AaB and

((AP) a(BP))P = (AaB) P. (3.6)

Since commutativity implies

PAP = AP < A and PBP = BP < B ,

it follows from (I) and (II) that

P(AaB) P <=(PAP) tT(PBP)
-- (AP) (7(BP) __<AaB.
Then the operator A a B - P ( A a B ) P is positive and has a vanishing diagonal block,
hence
(1 - P) { A a B - P(AaB)P} P = O,
which implies the commutativity of P and AaB. Analogously P commutes with
(AP)a(BP). These together prove (3.6).
Now since each scalar commutes with all projections, so does the operator 1ox,
hence is a scalar. Thus (3.5) determines a non-negative function f Let us show that
f is operator-monotone. Consider
t~

A = ZaiP,<=B= ~ b i Q J
1 1

where a t > O, bj > O, and Pi and Qj are projections such that

n m

Z ,=ZQj=I.
1 1
210 F. K u b o and T. A n d o

It follows by (3.6) that

n
laA = ~ {Pia(aiPi)} Pi
1
n

= ~, (laai)Pi
1
n

= Zf(a~)Pi=f(A)
1

and
lab =f(B).
Then inequality f(A)<f(B) follows from (I). Thus f is operator-monotone, and
continuity property (III) further guarantees the relation
taA =f(A). (3.7)
This implies, with help of (II), the relation
AaB = A 1/2f[A - 1/2BA- l/z] A 1/2. (3.8)
In view of continuity property (lid formula (3.8) shows that the connection a is
uniquely recovered from the function f By (3.5) and (3.8) it is almost obvious that
the map a ~ f is an affine order-isomorphism.
It remains only to prove that every operator-monotone function is obtained in
the form (3.5). Take an operator-monotone function f with integral representation
(3.2). Recall that A :B is the parallel sum of A and B. Define a binary operation a
by
AtrB=aA+bB+ S l+tt{(tA):B}dm(t),
(0,~)

where
a=m({O}) and b=m({o©}).

Since parallel sum and trivial means satisfy conditions (I), (II), and (III), the
operation ~ satisfies (I) and (II). Continuity condition (III) for a is proved with use
of the monotone convergence theorem in integration theory, hence ,z is a
connection. Finally for x > 0

lax=a+bx+ ~ x(l+t---~)dm(t)
(o,~) x + t
= I ~x,t)dm(t)=f(x).
[O, co]

Thus the function f is obtained from the connection a. This completes the proof.
The operator-monotone function f produced from a connection a by (3.5) is
called the representing function of a. If a is a mean, then f is normalized in the sense
f(1)=L
Means of Positive Operators 211

The representing functions of left-trivial mean coe and right trivial mean co, are
1 and x, respectively. The representing function of arithmetic mean V is (1 +x)/2
while that of harmonic mean ! is 2x/(1 +x), as seen from (2.4) and (3.7). The
representing function of geometric mean # is x 1/z.
Formula (3.8) already reveals an important property of a mean.

Theorem 3.3. Every mean a possesses the property

(IV') AaA = A for every A .
Proof. Identity (IV') for invertible A follows immediately from formula (3.8)
because o f f ( l ) = I. Appeal to continuity condition (III) to prove (IV') for general
A.
The following variant of Theorem 3.2 is already contained in the proof of
Theorem 3.2.
Theorem 3.4. The map, m~-~a, defined by
l+t
A a B = a A + b B + ~ --~--{(tA):B}dm(t) (3.9)
(0, ~ )
where
a=m({0}) and b=m({~}),
establishes an affine isomorphism from the class of positive Radon measures on
[0, ~ ] onto the class of connections.
The measure m that produces a connection a by (3.9) is called the representing
measure of a. The map m ~ a is order-preserving, but not order-isomorphic. In
fact, to harmonic mean corresponds the Dirac measure on the single point 1 while
the representing measure of arithmetic mean is distributed on two points 0 and ~ .
A connection becomes a mean, that is, satisfies normalization condition (IV) if
and only if its representing measure is a probability measure. Theorem 3.4 tells us
that extreme points of the convex set of means are trivial means and those means
a t (0 < t < ~ ) defined by

Aa, B = 1 + t {(tA):B}. (3.10)

t
The integral representation reveals further that each connection satisfies
conditions much stronger than (I) and (II). The most important is the concavity
property (I') in the next theorem.
Theorem 3.5. Every connection a possesses the following properties.
(I') (AaB) + (CAD) < (A + C)a(B + D).
(II') S*(AcrB)S <(S*AS)cr(S*BS)
where S is not assumed Hermitian.
Proof Since trivial means ¢oe and w, and the means at, defined by (3.10), possess
properties (I') and (II') (see [2] and [4]), so does the connection a by repre-
sentation (3.9).
212 F. Kubo and T. Ando

Condition (I'), when viewed as a property of a special connection, sum + in

this case, in relation to other connections, will single out the class of positive linear
combinations of left- and right-trivial means.
Theorem 3.6. Let z be a connection. Inequality
(AaB) z( CtrD) <=(AzC) a(BzD)
holds for every connection a and all positive operators A, B, C, and D if and only if z
is a non-negative linear combination of left-trivial mean toe and rioht-trivial mean to,.
Proof Obviously each of tot and to, possesses the property in question. Then any
combination atot + bto, with non-negative a and b has the property. This follows
from the concavity (I'). Supposing conversely the property for a connection z, let f
be its representing function. Choose tr to be arithmetic mean. Then for any positive
scalar x and y, substitutions A = B = 1, C = x and D = y will show by (3.7)

Therefore the concave f u n c t i o n / i s convex, too, hence is linear. By Lemma 3.1 and
Theorem 3.2 this implies that z is a non-negative linear combination of tot and to,.
We close this section with explicit determination of PGQ for projections P and
Q. Recall that P ^ Q denotes the projection to the intersection of the range of P
and that of Q.
Theorem 3.'/. If a is a mean, then for every pair of projections P and Q
PaQ = a ( P - P A Q) + b ( Q - P A Q) + P A Q,
where
a=la0 and b= l i m ( l a x ) / x .
x - - ~ oO

Proof Since P A Q commutes with P and Q and the range of (tP):Q is included in
that of P/x Q (see [3], Theorem 11), it follows from (3.6) applied to harmonic mean
that
(tP) :Q = {(tn):Q} (P ^ Q)
= {(tP) (P A Q)} :{Q(P A Q)}
=(t: 1)(e ^ Q)= t(1 +t)-1 (e ^ Q).
Then representation (3.9) becomes, with the representing probability measure m,
l+t
P~Q=m({O})P+m({oo})O+ S ~-{(tP):Q}dm(t)
(0, ~ )

= a(P- e ^ Q)+b(Q-P ^ Q)+P ^ Q.

Here to derive the last identity, formulas (3.3) and (3.4) are used. This completes
the proof.
Means of PositiveOperators 213

An immediate consequence of Theorem 3.7 is the following relation for

projections P and Q
P!Q=PAQ and P#Q=PAQ. (3.11)
The first identity was proved by Anderson and Schreiber [1] while the second was
pointed out by Fujii [9].

4. Transpose and Adjoint

This section is devoted to the study of the transpose and related notions, adjoint
and dual. Let us start with an easy but important technical temma.
Lemma 4.1. If f, g, and h are the representing functions of connections a, z, and 0
respectively, then the representing function of the connection a(r)0 is
f(x) .9[h(x)/f(x)] with convention 0 x oe = 0.
Proof. If k(x) denotes the representing function of a(z)O, then by Theorem 3.2 and
formula (3.8)
k(x) = 1cr(z)~ox=(lox)z(l~x)
= f (x) zh(x) = f (x) .g[h(x)/f (x)] .

Corollary 4.2. / f a is a connection with representing Junction f, then x f ( x - 1) is the

representinff function of the transpose a'. In particular, a is symmetric, if and only if
f ( x ) = x f ( x - ').
Proof The assertion follows immediately from Lemma 4.1 by observing (2.10):
o-' = o ) , ( a ) , %

If a is a non-zero connection with representing function J; the function

f ( x - 1)- 1 is also operator-monotone (see [4]), and can produce by Theorem 3.2 a
connection, called the adjoint of o and denoted by o*. Formula (3.8) gives an
explicit form to the adjoint
A a * B = ( A - l a B - a ) -1 for invertible A,B. (4.1)

The adjoint formation is involutive, (a*)* = a. The adjoint of arithmetic mean

V is harmonic mean !. Trivial means coe and ~o~, and geometric mean 4# are self-
adjoint. Other self-adjoint means are those with representing functions x p where
0 < p < 1. We do not know whether these exhaust all self-adjoint means.
When a is a non-zero connection, its dual, in symbol a ±, is defined by
a l = (a') * =(a*)'. (4.2)
The dual formation is involutive, (a±) ± = a, and the following formulas follow from
(4.2).
a'=(a*)l=(al)* . (4.3)
a* =(a') ± =(a±) ' . (4.4)
214 F, Kubo and T. Ando

Corollary 4.3. If a is a non-zero connection with representing function f, then x/f (x)
is the representing function of the dual a I.
This follows immediately from the definition of a ±, Lemma 4.1 and
Corollary 4.2.
The dual of arithmetic mean V is harmonic mean ! while the dual of left-trivial
mean o)e is right-trivial mean e)~. Corollary 4.3 implies that geometric mean # is
the only self-dual connection.
Now let us turn to the study of symmetric connections.

Theorem 4.4. The map, n~-~cr, defined by

A a B = 2 ( A + B ) + (o,51]-2[-l+t{(tA) :B+A:(tB)}dn(t) (4.5)

where c = n({0}), establishes an affine isomorphism from the class of positive Radon
measures on the unit interval [0, 1] onto the class of symmetric connections.
Proof A proof for that a defined by (4.5) is a symmetric connection is straightfor-
ward. Conversely, let a be a symmetric connection and f its representing function.
It follows from Corollary 4.2 and representation (3.2) that
f(x) = ½{f(x) + x f ( x - 1)}

=½(a+b)(l+x)+ 5 ,l+t)tx + +xt--~-i-}dm(t)

(0, ao) t

=½(a+b)(l+x)+ ~ --2-x+~-[ + dn(t)

(0, t] xt

where
dn(t)=dm(t)+dm(t-1), and n({O})=a+b.
Therefore, with c=a+b, formula (4.5) is valid for invertible A and B, hence for
general A and B by appealing to continuity condition (III). It remains to prove
that a measure n producing a by (4.5) is uniquely determined by a. Consider the
measure m on [0, oo] determined by the relation that dm(t)=½dn(t) or =½dn(t-1)
according as 0 < t < 1 or 1 < t < oe, and m({ 1}) = n({ 1}) and
m({0}) = m({ oo }) = ½n({0}).

Representation (4.5) shows that m produces a by (3.9). Now the uniqueness of m,

hence of n, results from Theorem 3.4. This completes the proof.
In Theorem 4.4 to a symmetric mean corresponds a probability measure. Thus
integral representation (4.5) shows that extreme points of the class of symmetric
means are arithmetic mean, harmonic mean, and ½(at + a't) (0< t < 1), where a t is
defined by (3.10).

Theorem 4.5. Arithmetic mean is the maximum of all symmetric means while
harmonic mean is the minimum.
Means of PositiveOperators 215

Proof The elementary inequality

l+x
for x,t>O
l+x = 2 t +xt = 2
implies
A! B<_
_ 2It -1 ( I + t ) { ( t A ) : B + A : ( t B ) } < A V B ,
the integration of which with respect to the measure n in (4.5) yields the assertion.
Since a + a r and a( :)a' are symmetric for every connection a, the following is an
immediate consequence of Theorem 4.5.
Corollary 4.6. For every mean a and all positive operators A and B the following
inequalities hold.
A :B < (AcrB) :(BOA) < (AaB) + (BaA) < A + B.
Upon replacing ~r by its adjoint or* and appealing to (4.2), Corollary 4.6 has the
following variant.
Corollary 4.7. For every mean a and ever>, positive operators A and B the following
inequalities hoId.
A :B <=(Aa*B) :(AaJ B) < (Air*B) + (A~-LB) < A + B.
We close this section with the dual variants of Theorem 3.5 and 3.6.
Theorem 4.8. Every connection ~ possesses the following property.
(I") (Aae) :(CAD) > (A :C) tr(B :D),
Proof This is true for the zero connection. Let cr be a non-zero connection, by (4.1)
and (2.4) for invertible A and B the inequality in question is equivalent to the
following
(A- 1¢*B- 1)+ (C- 1 a ' D - l) < (A- 1 + C - 1)a*(B- 1 + D - 1),

which is guaranteed by Theorem 3.5. The proof for the assertion is completed by
appealing to continuity condition (III).

Theorem 4.9. Let z be a connection. I f

(AaB)~(C~D) ~ (A,C)~r(B,D)
for every connection a and all positive operators A, B, C, and D, then z is the parallel
sum (aoae):(boo,) for some a ~_0 and b >0.
A proof is modeled on that of Theorem 3.6.

5. Geometric Mean
This section is devoted to the study of geometric mean in relation to other means.
We start with elementary observations on operator-monotone functions. A
normalized operator-monotone function f is said to be trivial if f ( x ) - 1 or
216 F. Kubo and T. Ando

f(x) = x. A trivial o p e r a t o r - m o n o t o n e function is just the representing function of

either one of trivial means coe and co,.
1..emma 5.1. If f is a normalized operator-monotone function, then
x<f(x)<l for 0<x<l
and
x>f(x)>l for l<x<oo.
All inequalities are strict unless f is trivial.
Proof. All the assertions are immediate from the non-negativity and concavity of f
a n d f ( 1 ) = 1.
L e m m a 5.2. I f f is operator-monotone, then the function xf(x) is convex.
Proof. In view of L e m m a 3.1 the assertion follows from the convexity of the
functions
x2(1 + t)
x4,(x, t ) = - -
x+t
L e m m a 5.3. Let f and g be normalized operator-monotone functions. If one oJf and g
is non-trivial the sequence, starting with any initial choice x o and defined successively
by
x. +1 = g(X.)/f(x.),
converges to 1.
Proof. If x, = 1 for some n then x m= 1 for all m > n and the convergence to 1 is
obvious. Suppose that all x,'s are different from 1. Consider a condition, n a m e d
(Mk), that for each n inequalities x , + i < x ~ for i = 1 , 2 ..... k or x , + i > x , for
i = 1, 2 ..... k, are valid according as x, > 1 or x. < 1. Let us prove by induction that
the sequence {x,} satisfies property (Mk) for all k > 1. R e m a r k first of all that
L e m m a 5.1 implies
1 / x ~ > x , + l > x n if 0 < x , < l (5.1)
and
1 / x ~ x n + l < x n if l<x,<oo. (5.2)
These inequalities already show the validity of (M1). Suppose that (Mi) i = 1,2 ..... k
are true for the sequence {x.}. T a k e arbitrary n. If x , > t and x , + i > l for some
1 < i < k, then by induction assumption

Xn+k+ 1 -'~'Xn+i+(k+ 1 - i ) ~ X n + i ~ X n "

If x . > 1 and x.+~ < 1 for i = 1, 2 ..... k then by induction assumption

l > Xn+k =Xn+ l +k- l ~ X n + l

hence by (5.1) and (5,2)
x.+k+ l <--1/x,+~ < 1/x,+ 1 ~ x , .
Means of Positive Operators 217

Therefore x. > 1 implies x. > x.+ k+ 1- Analogously it is shown that x. < 1 implies
X,,<X,,+k+ 2" These establish the validity of (Mk+ 1), completing induction.
It follows from conditions (Mk) k = 1, 2.... that the sequence {x.} is split into
two subsequences {x,,~} and {x,k } such that
x,,,1 "<xm2 < ... <- 1 <- ... <-x,2 ~ x,l .
(One of the subsequences may consist of only finitely many elements or may be
empty.) Suppose, for definiteness, that {x,~} is an infinite sequence, and let
c = klim
~o2
x, k .

The assertion of the theorem will be established if c = 1. For if {x,,~} is an infinite

sequence
lim xmk= g(c)/f(c) = 1.
k---* o~

Suppose, by contradiction, that c > 1 for definiteness. As a limit of successive

iterates, c satisfies the identity
g[g(c)/f(c)] = c. f[g(c)/j(c)] . (5.3)
Consider the functions H(x) and K(x) defined by
H(x) = f(x)" g[g(x)/f(x)]
and
K(x) = x f (x)" f [ g ( x ) / f (x)] .
Then
H(1) = K(1) = 1 and H(c) = K(c). (5.4)

By Lemmas 3.1 and 4.1 H(x) is the representing function of a mean, hence is
concave while K(x) is convex by Lemma 5.2. Then (5.4) implies that H(x) ~ K(x) on
[1,c]. Since H ( 1 ) = K ( 1 ) = I , it follows that K ( x ) > H ( x ) on [0,1]. But then K(0)
=/4(0) = 0, hence H(x) = K(x) = x on [0, 1]. Again by convexity and concavity H(x)
<_x < K ( x ) on [1, c], hence by (5.4) H(c)= K(c)= c. This implies by (5.3)

c / f (c) = g[g(c)/ f (c)] (5.5)

and
1/f(c) = f [g(c)/f(c)]. (5.6)
By assumption one of f and g is non-trivial. In case f is non-trivial, Lemma 5.1
yields g(c)/f(c)> 1 from (5.5) and g(c)/f(c)< 1 from (5.6), a contradiction. In case
- 1 and g is non-trivial, Lemma 5.1 yields from (5.5)
f(x) =

c = g i g ( c ) ] < g(c) < c ,

a contradiction. Finally in case f ( x ) = - x and g is non-trivial, Lemma 5.1 yields from

(5.5) g(c)/f(c) = 1, hence c = 1 by (5.6), a contradiction. Thus assumption c > 1 leads
to a contradiction. This completes the proof.
218 F. Kubo and T. Ando

Let a be a non-trivial connection with representing function f Since the

representing function of the dual ~rI is x/j(x), it follows from (3.8) and (2.6) that
Aa±B = (A ~ B) (AcrB)- 1 (A #¢B) for invertible A, B. (5.7)
This already shows a close relationship between geometric mean and duals.
Lemma 4.1 shows further that ~r(#)a ± coincides with geometric mean;
( AaB) 4~(Aa±B) = A 4~B. (5.8)
Tl~orem 5.4. Let ~ be a mean and a a non-trivial mean. I f the inequality
(A aB) z(A a±B) ~ A zB (5.9)
holds for all positive operators A and B, then "c is greater than geometric mean, i.e. for
all C and D
C#D<CzD.
Proof Let f, g, and h be the representing functions of or, cr±, and t, respectively. The
inequality (5.9), expressing tr(t)cr± < t, can he written
h[g(x)//(x)] ~ h(x)/f(x).
Since g(x)=x/f(x) by Corollary 4.3, this inequality can be written further
H[g(x)/f(x)] < H(x), (5.10)
where by definition
H(x)= h(x)/x if2 .
Since f is non-trivial along with the mean a, it follows from iterative use of (5.10)
and Lemma 5.3 that 1 NH(x), or equivalently x 1/2 <h(x), which means by Theorem
3.2 that z is greater than geometric mean. This completes the proof.
Corollary 5.5. Let z be a mean and ~r a non-trivial mean. I f the inequality
(AaB)z(AalB) ~ AzB (5.11)
holds for all positive operators A and B, then z is smaller than geometric mean, i.e. for
all C and D
CrD<=C~D.
Proof This is reduced to Theorem 5.4 by using a general formula; for non-zero
means a, z, Q
(a(z)@)l = O±(zl)a t , (5.12)

which is a consequence of Lemma 4.1.

Corollary 5.6. Let a be a non-trivial mean. Then geometric mean is the unique
solution of the equation

with unknown z.
Means of Positive Operators 219

Theorem 5.7. Let z be a mean with representing function h. I f inequality (5.9) holds
for every mean a, then the function h(x)/x 1/2 is non-decreasing on the interval [1, co)
and non-increasing on the interval (0, I].
Proof Let H(x) = h(x)/x 1/2. If t < d < c, there is 0 < q < 1 such that d = cL Consider
the m e a n cr with representing function f ( x ) = x p where p = (1 - q ) / 2 . Since the dual
a ± has representing function g(x)=x l-p, as in the proof of Theorem 5.4
n(d) = H[g(c)/f(c)] < H(c).
If d < c < 1, take 0 < q < 1 such that s q = c to prove H(d) > H(c). This completes the
proof.
R e m a r k that T h e o r e m 5.7 admits a partial converse. If H(x) is non-decreasing
on [ I , co) and non-increasing on (0, 1], then inequality (5.9) holds for those means
a whose representing functions f(x) satisfy
{x'/2- f(x)} {x- 1}>0.
In fact, the restriction on f(x) and L e m m a 5.1 imply that 1 < g ( x ) / f ( x ) < x or
x < g ( x ) / f ( x ) ~ 1 according as 1 < x or x < 1, where g(x) is the representing function
of the dual crI. Then the assumption on H implies (5.10), which is equivalent to
(5.9).

Corollary 5.8. Let z be a mean with representing function h. I f inequality (5.11) holds
for every mean a, then the function h(x)/x 1/2 is non-increasing on the interval [1, co)
and non-decreasing on the interval (0,1].
This is reduced to T h e o r e m 5.7 by appealing to (5.12).
Theorem 5.9. For every mean a and all positive operators A and B the following
inequalities hold
(AaB) + (Aa±B) <=A + B
and
(AaB) :(A~r±B) >- A :B.
Proof. The theorem asserts that (5.9) holds when z is arithmetic mean and that
(5.11) does when ~ is harmonic mean. Let f be the representing function of a. T h e n
(5.9) with z, arithmetic mean, is equivalent to
f(x). {1 + x/f(x) 2} <=1 + x
or
{f(x)-- 1} {x -- f(x)} ~ O,
which is valid by L e m m a 5.1. Inequality (5.11) with T, harmonic mean, is proved
analogously.
T h e o r e m 5.9 is considered as a relative of Corollary 4.6. In this connection
remark that the inequalities of T h e o r e m 5.9 cease to hold if the dual a ± is replaced
by the adjoint a*, as seen from the case of the mean with representing function
X2/3.
220 F, Kubo and T. Ando

6. Composition
This section is devoted to the study of construction of a new mean as the limit of
successive iterations of a simple operation. To speak about convergence, we have
to introduce a topology on the class of means. The class of positive Radon
measures on the interval [0, oo] is endowed with the weak-*topology of the dual of
the space C[0, oo] (see [8]). This topology defines the topology we need on the
class of means through the affine isomorphism in Theorem 3.4.

Lemma 6.1. For iz (generalized) sequence of means a, with representing functions f,

the following statements are mutually equivalent.
(i) The sequence % converges to a mean (r~.
(ii) For all invertible A and B the sequence of operators Aa,B is weakly
convergent.
(iii) For each x > 0 the numerical sequence f,(x) is convergent.
The limit in (ii) coincides with Aa®B while that in (iii) is equal to f~(x).

Proof. (i)~(ii). Let m, be the representing measure of % Then by definition of the

topology the sequence m, weak* converges to moo. Since for each fixed x > 0 the
function ~(x,t) of (3.1) is continuous with respect to t, the sequence

f.(x)= j qS(x,t)dm~(t)
[o, col

converges to f~(x). For fixed invertible A and B, take positive constants c and d
such that c<A-lr2BA-llZ<d. Since in view of formula (3.9) the analytic
continuation of f, to a neighborhood of the interval [c, d] is bounded uniformly
with respect to n, the sequence of analytic functions f,(z) converges uniformly to
f~(z) on a neighborhood of the interval [c, d]. Then the norm convergence of Aa,B
to Aa~B results from (3.8) and the Riesz-Dunford representation (see [8])

Z.[A- 1/2BA- 1/2] = 7 1 1 fn (z) ( Z -- A - I/2BA- ,I2)- 'dz

L~l F

where F is a simply closed rectifiablecontour in the neighborhood containing the

interval [c,~ in its interior.
(ii)=~(iii) is immediate by taking A = 1 and B = x and using (3.5).
(iii)=~(i). Since the class of probability measures is weak*-compact (see [8]), the
sequence {an} is relatively compact. Therefore it suffices to prove that it has only
one limiting point. Take an arbitrary limiting point (mean) a with representing
measure m. Denoting the limit of f,,(x) by fix), Lemma 3.1 shows

f(x)= f c~(x,t)dm(t),
[0, ~1

hence f is operator-monotone and the measure m is uniquely determined. Thus

is uniquely determined, too. This completes the proof.
Remark that the restriction x > 0 in (iii) [or equivalently the invertibility of A
and B in (ii)] can not be removed. In fact, if fn(x)= x 1/", the sequence a, converges
to the left-trivial mean o9t with representing function f(x)= 1 but f,(0)= 0 for all n.
Means of PositiveOperators 221

Theorem 6.2. Let a and z be means, one of which is non-trivial. Then the sequences,
starting with arbitrary initial data ~o and z o and defined successively by
a,+t=oja)% and z,+l=o~(z)z,, (6.1)
converge to the same limit,
Proof Let f and g be the representing functions of a and ~, and f , and g, those of
a~ and z., respectively. Then by Lemma 4.1 relation (6.1) is written
f . +1(x) = f.(x), f[o.(x)/f.(x)] (6.2)
and
g. + l(x) = f j x ) . g [g.(x)/f.(x)]. (6.3)

With x.=g.(x)/f.(x) for all n these imply

x. + 1 = g(x.)/f(x.)
hence the sequence x. converges to 1 by Lemma 5.3. Let
a. = f.(x) and b. = g,(x).
We claim that the interval with extreme points a.+ 1 and b.+ ~ is included in the
interval with extreme points a. and b,. Consider the case a. > b.. Since (6.2) and
(6.3) are written respectively
a. + l/a. = f[b./a.] (6.4)
and
b.+ t/a. = g [ b j a . ] , (6.5)

Lemma 5.t yields from (6.4) and (6.5)

bja.<a.+ l/a.< l
and
b,,/a.<=b.+ l/a~<= l ,

which implies that each of a.+ ~ and b.+ 1 is contained in the interval with extreme
points a. and b.. The same conclusion is established analogously for the case
a. < b.. The claim has been proved.
Now in view of Lemma 6.1 the assertion of the theorem is equivalent to the
relation
infla.-b.[ =0.

Since x . = b . / a , converges to 1 and the sequence {a,} is bounded, the relation in

question results from
inf ] a . - b.[ ~ sup ]a,,[ • lim ll - x.[ = 0.

This completes the proof.

222 F. Kubo and T. Ando

W h e n a o = a and t o = Z , the c o m m o n limit in T h e o r e m 6.2 is called the

composition of cr and ~, a n d is denoted by a x z. Fujii [9] defined the composition
o f arithmetic mean and geometric mean and named Gauss mean. The composition
of geometric mean and harmonic mean was considered in [10].
To determine the exact form of composition is not easy in general. The
following theorem gives a convenient rule, a consequence of which is that
geometric mean is obtained as the limit of alternative successive iterates of
arithmetic mean and harmonic mean.
Theorem 6.3. The composition o f a non-trivial mean and its dual always coincides
with geometric mean
O" X O ' ± = ~ .

P r o o f R e m a r k that if both o - a n d , are non-trivial then by (5.12) successive iterates

G ' s and z,'s in Theorem 6.2 satisfy

o ']", + l -_ _- Z ,i ( a ]_ )a,£ and ]-

Tn+ 1
__ £ £
--Tn(Z )*.. ±

This implies by T h e o r e m 6.2

(a x z)± = z ± x a ± . (6.6)
N o w apply formula (6.6) to the case z = cr± to see that 6 x a ± is a self-dual mean.
Geometric mean is the only self-dual mean. This completes the proof,
By construction a x a = a is always valid. In this connection the following
theorem is of interest.

Theorem 6.4. Let a be a non-trivial mean. Then a is the unique solution o f the
equation
O" X " C = O "

with unknown z.

P r o o f We use the notations in the p r o o f of T h e o r e m 6.2. If z is a non-trivial mean

different from a, by T h e o r e m 3.2 there is a point x such that f ( x ) 4 : g(x). Then it is
seen from the p r o o f that each of a t and b 1 is strictly between a o - - f ( x ) and
bo = O(x), hence lima, is different from %, which implies a x z 4=a. If z is left-trivial
mean ~e, ao 4=bo implies that each of a z and b 2 is strictly between a o and b 0, hence
a x z 4=a as above. Finally
~r X O 9 ~ = ( D , 4 = ~r .

This completes the proof,

Composition formation is not associative. In fact, (4t: x 4~)x V does not
coincide with # x (4~ x V).

Theorem 6.5. Let a be a non-trivial mean. The sequence, startin 9 with arbitrary non-
trivial mean a o and defined successively by

ffn+l = G X fin,

converges to a.
Means of Positive Operators 223

Proof Let f and f , be the representing functions of cr and %, respectively. It is seen

from the proof of Theorem 6.2 that for fixed x the value f,(x) is between f(x) and
f , _ l(x). Therefore the sequence f,(x) converges increasingly or decreasingle accord-
ing as fo(X) Nf(x) or Jo(X) >f(x). Then by L e m m a 6.1 the sequence or, converges to
some mean, say ~r~ with representing function f~. Remark that the non-triviality
of % implies the non-triviality of a , . Suppose that f i x ) < f , ( x ) at a point x. Then
by L e m m a 5.1

f(x)" f[f~(x)/f(x)] <foo(x)

and

f(x). f~,[fo~(x)/f(x)] <fo~(x) •

Since, as seen in the p r o o f of L e m m a 6.1, f , converges to f ~ uniformly on each
c o m p a c t subset of (0, oe), there is m such that

f(x). f[fm(x)/f(x)] <To(x)

and

f(x)" fm[jm(x)/f(x)] <f~o(X).

Then as in the p r o o f of Theorem 6.2

f~(x) Nfm+ l(x)

< max {f(x).f[f,,(x)/J(x)], f(x) .L[f,,(x)/f(x)] }
<f~(x),
which is a contradiction. The case f(x)>f®(x) also yields a contradiction. These
establish f ~ =j~ hence a~ = a, completing the proof.
Corollary 6.6. Let a be a non-trivial mean. The sequence, starting with arbitrary
non-trivial mean a o and defined successively by
o-n+ 1 = o'n x ~r,
converges to a.
This follows immediately from Theorem 6.5 by using (6.6).

References
1. Anderson, W.N., Jr., Duffin, R.J. : Series and parallel addition of matrices. J. Math. Anal. Appl. 26,
576-594 (1969)
2. Anderson, W.N.,Jr, Schreiber, M. : The infimum of two projections. Acta Sci. Math. 33, 165-168
(t972)
3. Anderson, W.N.,Jr., Trapp, G.E.: Shorted operators II. SIAM J. Appl. Math. 28, 60-71 (1975)
4. Ando, T. : Topics on operator inequalities, Lecture Note, Sapporo, 1978
5. Bendat, J., Sherman, S. : Monotone and convex operator functions. Trans. Amer. Math. Soc. 79,
58-71 (1955)
6. Bhagwat, K.V., Subramanian, A. : Inequalities between means of positive operators. Math. Proc.
Camb. Phil. Soc. 83, 393-401 (1978)
224 F. Kubo and T. Ando

7. Donogue, W. : Monotone matrix functions and analytic continuation. Berlin, Heidelberg, New
York: Springer 1974
8. Dunford, N., Schwartz, J.T. : Linear operators I. New York, London : Wiley 1963
9. Fujii, J.I.: Arithmetico-geometric mean of operators. Math. Japon. 6, 667-669 (1979)
10. Fuiii, J.l.: On geometric and harmonic means of operators. Math. Japon. (to appear)
11. Nishio, K., Ando, T. : Characterizations of operations derived from network connections. J. Math.
Anal. Appl. 53, 539-549 (1976)
12. Pusz, W., Woronowicz, S.L. : Functional calculus for sesquilinear forms and the purification map.
Rep. Math. Phys. 8, 159-170 (1975)

Received April 5, 1979