Lawson, J. and Lim, Y.

Osaka J. Math.
44 (2007), 795–816

METRIC CONVEXITY OF SYMMETRIC CONES

J IMMIE LAWSON and YONGDO LIM

(Received April 24, 2006, revised November 10, 2006)

Abstract
In this paper we introduce a general notion of a symmetric cone, valid for the
finite and infinite dimensional case, and prove that one can deduce the seminegative
curvature of the Thompson part metric in this general setting, along with standard
inequalities familiar from operator theory. As a special case, we prove that every
symmetric cone from a JB-algebra satisfies a certain convexity property for the
Thompson part metric: the distance function between points evolving in time on two
geodesics is a convex function. This provides an affirmative answer to a question of
Neeb [22].

1. Introduction

Let A be a unital C  -algebra with identity e, and let A+ be the set of positive
invertible elements of A. It is known that A+ is an open convex cone in the space
H(A) of hermitian elements. The geometry of A+ has been studied by several au-
thors. One approach has been to endow A+ with a natural Finsler structure and metric
and use these for a substitute for the Riemannian geometry commonly considered in
finite-dimensional examples. One particular focus in this geometry has been the study
of appropriate non-positive curvature properties. One prevalent notion of non-positive
curvature is a purely metric one, that of convexity of the metric. In [3], [4] and [9],
Andruchow-Corach-Stojanoff and Corach-Porta-Recht have shown the convexity of the
distance function along two distinct geodesics and its equivalence to the well-known
Loewner-Heinz inequality. In [22], Neeb established an appropriate differential geomet-
ric notion of seminegative (equal non-positive) curvature for certain classes of Finsler
manifolds.
Our approach is somewhat different from either of the preceding. We replace the
differential geometric structure by the structure of a symmetric space endowed with a
midpoint operation and study seminegative curvature via convexity of the metric. In
[16] we obtained the convexity of the metric for symmetric spaces with weaker metric
assumptions than those enjoyed by the Finsler metric on A+ .
The Finsler distance or length metric on A+ used in the earlier referenced papers
agrees with the Thompson metric, which is widely known and has many applications in
general convex cones of normed spaces ([27], [24]). The geodesic line passing through

2000 Mathematics Subject Classification. 53C35, 47H07, 46B20, 53B40.

796 J. L AWSON AND Y. L IM

a and b is given by a,b (t) = a 1=2 (a 1=2 ba 1=2 )t a 1=2 and the Thompson metric is defined
by d(a, b) = maxflogka 1=2 ba 1=2 k, logka 1=2 b 1 a 1=2 kg. The convexity theorem states
that for a, b, c, d 2 A+ , the real function t 7! d( a,b (t), c,d (t)) is convex. For Riemann-
ian manifolds this convexity condition is equivalent to the manifold having non-positive
curvature in the Riemannian sense. In general, however, it is a weaker notion than the
more prevalent metric notion of a CAT0 -space arising from Alexandrov’s metric notion
of spaces of non-positive curvature; see Section II.1.18 and the following appendix in
[8] (cf. [5], [10]).
The main purpose of this paper is to extend the convexity result on A+ to much
more general cones endowed with a symmetric structure that appropriately interacts
with the conal structure. A special case is the symmetric cone arising as the set of
invertible squares of a Jordan-Banach algebra (JB-algebra). Our general results applied
to this particular case provide an affirmative answer to a question raised by Neeb [22].
A subsidiary goal of the paper is to present a very general framework, that of a general
notion of a symmetric cone, in which one can derive and study various inequalities,
such as those familiar from operator theory.

2. Symmetric spaces with midpoints

We recall from ([14], [15]) the underlying algebraic structure with which we work
and basic properties thereof. A ℄-symmetric set (called dyadic symsets in [14]) con-
sists of a binary system (X , ), with left translation Sx y := x  y representing the point
symmetry through x, satisfying for all a, b, c 2 X :
(S1) a  a = a (Sa a = a);
(S2) a  (a  b) = b (Sa Sa = id X );
(S3) a  (b  c) = (a  b)  (a  c) (Sa Sb = SSa b Sa );
(S4) the equation x  a = b (Sx a = b) has a unique solution x 2 X , called the midpoint
or mean of a and b, and denoted by a ℄ b.
The axioms bear close resemblance to the Loos axioms for a symmetric space [21].
A binary system (X , ) satisfying (S1), (S2), and (S3) also satisfies (S4) if and only
if it is a quasigroup. Thus the preceding structures are also referred to as symmetric
quasigroups. Systems satisfying only Axioms (1)–(3) are called symmetric sets (or in-
volutive quandles in knot theory circles).
A pointed ℄-symmetric set is a triple (X , , " ), where (X , ) is a ℄-symmetric set
and " 2 X is some distinguished point, called the base point. In this setting we define

x 0 = ", x 1
:= S" x, x 2 := Sx " , x 1=2 := " ℄ x

and inductively from these definitions all dyadic powers are defined so that the follow-
ing rules are satisfied:

(x r )s = x r s , x r ℄ x s = x (r +s)=2 .
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 797

If we consider the dyadic rationals D endowed with the ℄-symmetric structure a  b =

2a b (the reflection of b through a), then a ℄ b = (a + b)=2, the usual midpoint, and
the map t 7! x t : D ! X is both a -homomorphism and ℄-homomorphism. From this
fact the preceding rules (and others) easily follow.
The displacement group G(X ) (also called the transvection group) of a ℄-symmetric
set X is the group generated under the composition by all transformations of the form
Sx Sy , x, y 2 X . It follows from Axioms (S2) and (S3) that these are automorphisms
and thus there is a group action (g, x) 7! g.x : G(X )  X ! X with G(X ) acting as
automorphisms. If X is pointed with base point " , then G(X ) is generated by all Sx S"
and X embeds into G(X ) as a twisted subgroup (closed under g  h = gh 1 g) via the
quadratic representation Q : X ! G(X ) defined by Q(x) = Sx S" . The image Q(X )
is a pointed ℄-symmetric set under the preceding -operation and the quadratic repre-
sentation is an isomorphism between X and Q(X ). In particular, Q(X ) is uniquely
2-divisible and Q(x ℄ y) = Q(x) ℄ Q(y), Q(x 1=2 ) = Q(x)1=2 ([14, Theorem 5.4]). For
x, y 2 X , we write interchangeably as convenient

x.y = Q(x)y = Q(x)(y).

R EMARK 2.1. The following useful calculation rules are derived in [14] or can
easily be derived by the methods there:
(1) Q(Q(x)y) = Q(x)Q(y)Q(x) or (x.y).z = x.(y.(x.z)).
(2) (Q(x)) 1 = Q(x 1 ).
(3) (Q(x)y) 1 = Q(x 1 )y 1 or (x.y) 1 = x 1 .y 1 .
(4) Sx r x s = x r  x s = x 2r s , Q(x r )x s = x r .x s = x 2r +s , x r ℄ x s = x (r +s)=2 .

Lemma 2.2 (Riccati lemma). In a pointed ℄-symmetric set X , the geometric

mean a ℄ b is the unique solution in X of the Riccati equation

1
Q(x)a =b

and is given by

a ℄ b = Q(a 1=2 )(Q(a 1=2 )b)1=2 = a 1=2 .(a 1=2 .b)1=2

(cf. the last paragraph in p.21). Furthermore, the geometric mean operation satisfies
(i) a ℄ b = b ℄ a,
(ii) (a ℄ b) 1 = a 1 ℄ b 1 ,
(iii) g.(a ℄ b) = (g.a) ℄ (g.b) for any g 2 G(X ).

Lemma 2.3. In a pointed ℄-symmetric set X ,

a ℄ Q(b)a 1
= b, 8a, b 2 X .
798 J. L AWSON AND Y. L IM

Proof. The lemma follows from the fact that x = Q(b)a 1

if and only if b = a ℄ x,
which follows by the Riccati lemma.

3. Symmetric spaces with convex metrics

We now impose metric and topological structure upon ℄-symmetric sets.

D EFINITION 3.1. A pointed symmetric space with convex metric is a pointed

℄-symmetric set P equipped with a complete metric d(
,
) satisfying for all x, y 2 P
and g 2 G(P)
(i) d(g.x, g.y) = d(x, y),
(ii) d(x 1 , y 1 ) = d(x, y),
(iii) d(x 1=2 , y 1=2 )  (1=2) d(x, y),
(iv) x 7! x 2 : P ! P is continuous.
A symmetric space with convex metric is a ℄-symmetric set equipped with a complete
metric that is a pointed symmetric space with convex metric with respect to some
pointing.

E XAMPLE 3.2. Let R be equipped with the standard ℄-symmetric operation x 

y := 2x y and the usual metric. Then x ℄ y = (x + y)=2, the usual midpoint operation,
and the metric is convex. Thus (R, , 0) is a pointed symmetric space with convex
metric.

We recall some basic results about symmetric spaces with convex metrics from [16].

Theorem 3.3 ([16]). Let P be a symmetric space with convex metric. Then for
distinct x, y 2 P, there exists a unique continuous homomorphism x, y (called an
s-geodesic) of ℄-symmetric sets from R into P satisfying x, y (0) = x and x, y (1) = y.
Furthermore, the maps

(x, y) 7! x  y : P  P ! P, (t, x, y) 7! x, y (t) := x ℄t y : R  P P!P

are continuous.

The element x ℄t y is called the t-weighted mean of x and y. Note that x ℄ y =

x ℄1=2 y.

Theorem 3.4 ([16]). Let P be a symmetric space with convex metric. For every
pair ( , ) of s-geodesics, the real function

t 7! d( (t), (t))
is a convex function.
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 799

R EMARK 3.5. We note that the unique s-geodesic line satisfying x, y (0) = x and
x, y (1) = y is

x, y (t) = x ℄t y = x 1=2 .(x 1=2

.y)t

and  y, x (1 t) = x, y (t), t 2 R ([16]). In particular,

(3.1) (Q(y)x)t = Q(y)Q(x 1=2 )(Q(x 1=2 )y 2 )t 1

.

Indeed,

Q(y 1
)(Q(y)x)t = y 2
℄t x = x ℄1 t y 2

= Q(x 1=2 )(Q(x 1=2 )y 2 1 t

)
= Q(x 1=2 )(Q(x 1=2 )y 2 )t 1
.

4. Convex cones with convex metrics

Let V be a Banach space and let  henceforth denote a non-empty open convex
cone of V : t    for all t > 0,  +   , and  \  = f0g, where  denotes
the closure of . We further assume that  is a normal cone: that is, there exists a
constant K with kx k  K k y k for all x, y 2  with x  y. For a normal cone , the
relation

x y if and only if y x 2
is a partial order. We write x < y if y x 2 .
Any member " of  is an order unit for the ordered space (V , ), and the cone
is normal if and only if the order unit norm determined by " is compatible, i.e., de-
termines the topology of V . In this case 0  x  y implies kx k  k y k with respect to
the order unit norm, that is, we may assume without loss of generality that K = 1. We
henceforth make this assumption. In fact, for y 2 V , "  y = M 2  for a sufficiently
large M, and hence M "  y  M " . Moreover, y = M[(y = M + " ) " ] 2  , i.e.,
V =  . By Proposition 1.1 in [24], for a normal cone , the order unit norm is
compatible. The normality condition of the order unit follows from its definition. See
[24], [12, Section 1.2], and [29, Section 14] for more details.
A.C. Thompson [27] (cf. [23], [24]) has proved that  is a complete metric space
with respect to the Thompson part metric defined by
    
x y
d(x, y) = max log M , log M
y x

where M(x = y) := inff > 0 : x   y g. The Thompson metric can be alternatively real-
ized as an appropriately defined Finsler length metric. Since  is an open subset of
800 J. L AWSON AND Y. L IM

V , it carries a natural structure of (real) differentiable manifold and its tangent space
Tx  can be identified to V = fx g  V (cf. [13]). For x 2  and v 2 V = Tx , we
define the Finsler metric by the order unit norm for v :

jvjx := infft > 0 : tx  v  t x g.

The Thompson part metric d(x, y) agrees with the Finsler distance from x to y:
Z 1 
(4.1) d(x, y) = inf j 0 (t)j (t) dt : 2 S, (0) = x, (1) = y
0

where S denotes the set of piecewise C 1 maps : [0, 1] !  ([24, Theorem 1.1]).

Lemma 4.1. Let  be an open convex normal cone in a Banach space V . Sup-
pose that there is a pointed ℄-symmetric structure on  such that the displacements
Q(x):  !  are positively homogeneous for all x 2 . Then (x) 1 = (1=)x 1 and
t " = (")t for all  > 0 and all dyadic rationals t. Moreover x ℄t x = 1 t t x for
all x 2 , ,  > 0 and all dyadic rationals t. Furthermore, the following conditions
are equivalent:
p
(i) (x)1=2 = x 1=2 for all x 2  and  > 0;
(ii) (x)t = t x t for any dyadic rational t and x 2  and  > 0;
(iii) Q(x) = 2 Q(x) for any x 2  and  > 0.

Proof. Let A be the set of all dyadic rationals t such that t " = (" )t . It is im-
mediate that 0, 1 2 A. Set x = " . Then for any dyadic rational t by homogenity of
Q(x) and Remark 2.1

x t+1 = Q(x t =2 )x = Q(x t =2 )(" ) =  Q(x t =2 )" = x t .

A simple induction then yields for any positive integer n, x t+n = n x t (for example,
x t+2 = Q(x 1=2 )x t+1 = Q(x 1=2 )(x t ) =  Q(x 1=2 )x t = x t+1 = 2 x t ). It follows that x n =
n x 0 = n " and thus A includes all positive integers.
For a positive integer n and t = n, " = x n+n = n x n . Thus x n =  n " and A
includes the negative integers as well.
The preceding results apply to any  > 0, in particular to  = 1=2 . Thus for
m

y = " , y 2 = 2 " = " . It follows that (" )1=2 = y = 1=2 " , i.e, 1=2m 2 A. For any
m m m m

integer n,

(" )n =2 = ((" )1=2 )n = (1=2 " )n = (" )n = n " = n =2 " ,

m m m m

where the penultimate equality follows from the first two paragraphs for  = .
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 801

Suppose that  and  are positive real numbers. We recall from Remark 3.5 that
x ℄t y = x 1=2 .(x 1=2 .y)t = Q(x)1=2 (Q(x) 1=2 y)t . Then the preceding result implies that
 t
1=2 1=2 1=2
" ℄t " = Q(") (Q(") (")) = Q(") Q(" ) 1=2 (" )
t

   t  t
 t  
= Q(" )1=2 " = Q(")1=2 "= Q(" )1=2 "
  
 t

= " = (1 t t )"

and for x 2 ,
x ℄t x = Q(x 1=2 )(" ℄t ") = Q(x 1=2 )((1 t t )") = (1 t t )Q(x 1=2 )" = (1 t t )x.

Next, we show that

(4.2) (x) 1
=  1x 1
.

It follows from (" ) 1

=  1 " , x =  Q(x 1=2 )" = Q(x 1=2 )(" ) and Remark 2.1 that

(x) 1
= (Q(x 1=2 )(" )) 1
= Q(x 1=2 )(" ) 1
= Q(x 1=2 )( 1 " ) =  1
Q(x 1=2 )" =  1 x 1
.

We next prove the equivalence of the conditions in the statement. Assume (x)1=2 =
p
x 1=2 for x 2  and  > 0. To prove (iii), we first calculate
p
x ℄  y = x 1=2 .(" ℄ (x 1=2 .y)) = x 1=2 .((x 1=2 .y))1=2 = x 1=2 .( ((x 1=2 ).y)1=2 )
p p
= x 1=2 .(" ℄ x 1=2
.y) = (x ℄ y).

By the Riccati lemma, z = Q(x)y implies that x = z ℄ y 1 or x = (1=)(z ℄ y 1 ) =

((1=2 )z) ℄ y 1 , and again by the Riccati lemma we have (1=2 )z = Q(x)y or z =
2 Q(x)y. This shows that Q(x) = 2 Q(x) for any x 2  and  > 0.
(iii) implies (ii). Suppose Q(x) = 2 Q(x) for x 2  and  > 0. From (x)2 =
Q(x)" = 2 Q(x)" = 2 x 2 and by a simple induction (x)n = n x n for any positive
integer n. Indeed, if (x)k = k x k for k = 1, 2, : : : , n, then

(x)n+1 = Q(x)(x)n 1
= 2 Q(x)(n 1 n 1
x ) = n+1 x n+1 .

By (4.2), (x) n = ((x) 1 )n = ( 1 x 1 )n =  n x n for any positive integer n. Further-

more, (1=2 x 1=2 )2 = x and hence (x)1=2 = 1=2 x 1=2 for any integer m. For any
m m m m m m

integer n,

(x)n =2 = ((x)1=2 )n = (1=2 x 1=2 )n = n =2 x n =2 .

m m m m m m

Therefore (ii) follows, and the trivial implication (ii) implies (i) completes the proof.
802 J. L AWSON AND Y. L IM

The next theorem gives the main result of this paper. Note that all powers are
computed in the given ℄-symmetric structure of the cone.

Theorem 4.2. Let  be an open convex normal cone in a Banach space V . Sup-
pose that there is a pointed ℄-symmetric structure on  satisfying
(i) x 1=2  (" + x)=2.
(ii) the squaring map x 7! x 2 = Q(x)" is continuous (in the relative norm topology
of ).
(iii) every basic displacement Q(x) is continuous and linear (that is, additive and
positively homogeneous) on .
Then  is a symmetric space with convex metric with respect to the Thompson metric
that satisfies the equivalent conditions (i), (ii) and (iii) of Lemma 4.1. Furthermore, (i)
the order-reversing property of inversion, (ii) the harmonic-geometric-arithmetic mean
inequality, and (iii) the Loewner-Heinz inequality all hold: for a, b 2 ,
(i) b 1  a 1 if a  b,
(ii) 2(a 1 + b 1 ) 1  a ℄ b  (1=2)(a + b), and
(iii) a 1=2  b1=2 if a  b.

Proof. The proof proceeds in steps.

S TEP 1. Each Q(x) extends to an invertible bounded linear operator on V that
is an order-isomorphism. Let T :  !  be linear (additive and positive homogeneous)
and continuous. Let :    ! V be defined by (x, y) = x y. Then is an open
mapping that is surjective since V =   (see the second paragraph of this section).
One verifies directly that T extends to a map, again called T , from V to V defined
by T (x y) := T (x) T (y) and that the following diagram commutes:

It is straightforward to verify that T : V ! V is an additive homomorphism and homo-

geneous with respect to positive scalars. Since an additive homomorphism is homo-
geneous for the scalar 1, it follows that T is linear. Since is open, it is a quotient
map and thus T : V ! V is continuous, hence bounded, since T  T is continuous.
For T = Q(x), we conclude from hypothesis (iii) that Q(x) extends (uniquely) to a
bounded linear map. Since members of G() are compositions of basic displacements,
the same conclusion holds for them.
Since Q(x) 1 = Q(x 1 ), Q(x) is invertible with inverse Q(x 1 ) on . It follows
readily that their extensions are inverses. Since Q(x) preserves , by continuity it
also preserves , and thus is order-preserving on V . Since Q(x 1 ) is similarly order-
preserving, it follows that Q(x) is an order-isomorphism. It follows that any member
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 803

of G() is an order-isomorphism, being a composition of such. Hence each member

of G() is an isometry for the Thompson metric, since the latter is defined from the
order. Thus Axiom 3.1 (i) is satisfied.
S TEP 2. The inversion x 7! x 1 = S" x is order reversing.
Suppose that 0 < x  y. Then by the Riccati lemma 2.2 we have

Q(y ℄ x)y 1
=x  y = Q(x ℄ y)x 1
.

Since Q(y ℄ x) = Q(x ℄ y) is an order-isomormisphism by Step 1, the conclusion follows.

S TEP 3. The harmonic-geometric-arithmetic mean inequality holds:
 1 1 1
(4.3)
x +y
2
 x ℄ y  x +2 y .
Since each displacement is linear and order preserving on  and preserves the geo-
metric mean operation by Lemma 2.2, we have

x ℄ y = (Q(x 1=2 )" ) ℄ (Q(x 1=2 )Q(x 1=2 )y)

= x 1=2 .(" ℄ (x 1=2 .y))
 1=2 
"+x
 x 1=2 . .y
(i)

2
x+y
= ,
2

the geometric-arithmetic mean inequality. The harmonic-geometric mean inequality fol-

lows from the order reversing property of inversion (Step 2) and the geometric-arithmetic
mean inequality

x℄y ℄ y 1)
Lemma 2.2 1 1
= (x
 1 1 1
 x +y
Step 2
.
2

S TEP 4. The squaring map x 7! x 2 is continuous for the Thompson metric. There-
fore Axiom 3.1 (iv) is satisfied. Indeed, this is a consequence of the agreement of the
norm topology with that of Thompson metric ([24, Proposition 1.1]).
S TEP 5. Inversion is an isometry with respect to the Thompson metric. Therefore
Axiom 3.1 (ii) is satisfied. Let x, y 2  and let  > 0 such that x   y. Then since
Q(y 1=2 ) is linear and preserves the order, we have

y 1=2 .x y 1=2
.( y) = " .
804 J. L AWSON AND Y. L IM

The order reversing property of the inversion implies that

(y 1=2 .x) 1
= y 1=2 .x 1
 (") 1 Lemma 4.1
=
1

".

Again by linearity,
 
x 1
y 1=2
.
1
 
1
" = y 1.

This string shows that M(y 1 =x 1 ) = M(x = y) and hence d(x 1 , y 1 ) = d(x, y).
S TEP 6. For a 2 , define f a :  !  by f a (x) = (1=2)(x + Q(a)x 1 ). Then

lim f an (x) = a,
n !1
8 x 2 .
(See [20, Theorem 7] for symmetric cones of Euclidean Jordan algebras.)
First, we observe that the map f a is continuous (inversion, the displacements, and
the linear operations are continuous). Suppose that the iteration has a limit point, say
b. Then by continuity, b = f a (b) = (1=2)(b + Q(a)b 1 ) and then b = Q(a)b 1 . By the
Riccati lemma, a = b ℄ b = b. The convergence is proved by several substeps.
(i) f an (x)  a: By the G-A (geometric-arithmetic mean) inequality and Lemma 2.3,

(x) + Q(a) f an 1 (x) 1 )  f an 1 (x) ℄ Q(a) fan 1 (x)

1 n 1
f an (x) = (f 1
= a.
2 a
(ii) Q(a) f an (x) 1  a: By linearity, the invertibility of Q(a), and the equality
(Q(a)x) 1 = Q(a) 1 x 1 ,
  1
1
Q(a) f an (x) 1 = Q(a) ( f an 1 (x) + Q(a) f an 1 (x) 1 )
2
  1
1 1
= (Q(a) f an 1 (x) + f an 1 (x) 1 )
2
  1
1
= ((Q(a) f an 1 (x) 1 ) 1
+ f an 1 (x) 1 )
2

H-G
(Q(a) fan 1 (x) 1 ) ℄ f an 1 (x)
Lemma 2.3
= a.

(iii) k fan (x) a k  1=2n 1

k fa (x) a k: By (i) and (ii),

0  f an (x) Q(a) f an 1 (x) 1 ) 

1 n 1 1 1 n 1
a= (f (x) a) (a (f (x) a),
2 a 2 2 a
and then by the normality of the cone,

k fan (x) ak  k f (x)

1 n 1
2 a
a k.
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 805

S TEP 7. The Loewner-Heinz inequality holds:

(4.4) 0<a b implies a 1=2  b1=2 .

By Step 6, it is enough to show statement An : f an1=2 (" )  f bn1=2 (" ) for all n = 0, 1, 2, : : :
whenever 0 < a  b (cf. [20, Corollary 9]). However, for the induction to proceed
smoothly, we prove additionally statement Bn : Q(a 1=2 ) f an1=2 (" ) 1  Q(b1=2 ) f bn1=2 (" ) 1 .
The statement A0 reduces to "  " and B0 asserts that

a = Q(a 1=2 )(" 1 )  Q(b1=2 )(" 1 ) = b,

which is true by hypothesis. Suppose that Ak and Bk hold for k = n 1. Then

1 n 1

f an1=2 (" ) = f a1=2 (" ) + Q(a 1=2 ) f an1=21 (" ) 1
2


 12 fbn1=21 (") + Q(a1=2 ) fan1=21 (") 1


 12 fbn1=21 (") + Q(b1=2 ) fbn1=21 (") 1
= f bn1=2 (" ),

where the two inequalities are applications of An 1 and Bn 1 respectively.

Showing

(4.5) Q(a 1=2 ) f an1=2 (" ) 1

 Q(b1=2 ) fbn 1=2 (" ) 1

is equivalent (by inverting) to showing

(4.6) Q(a 1=2 ) f an1=2 (" )  Q(b 1=2 ) f bn1=2 (" ).

However, since
1
Q(a 1=2 ) f an1=2 (" ) = (Q(a 1=2 ) f an1=21 (" ) + f an1=21 (" ) 1 )
2
1
Q(b 1=2 ) f bn1=2 (" ) = (Q(b 1=2 ) f bn1=21 (" ) + f bn1=21 (" ) 1 )
2

and since f an1=21 (" ) 1

 f bn1=21 (" ) 1
(by induction hypothesis An 1 and inversion), (4.6)
follows from

(4.7) Q(a 1=2 ) f an1=21 (" )  Q(b 1=2 ) f bn1=21 (" )

or

(4.8) Q(a 1=2 ) f an1=21 (" ) 1

 Q(b1=2 ) fbn 1=2
1
(" ) 1 ,
806 J. L AWSON AND Y. L IM

p
which is true by the inductive hypothesis Bn 1 .
S TEP 8. (a)1=2 = a 1=2 for any  > 0 and a 2 , and therefore the equivalent
conditions of Lemma 4.1 are satisfied. To prove this we use mainly the facts that
(Q(x)y) 1 = Q(x 1 )y 1 (Remark 2.1) and (x) 1 =  1 x 1 (Lemma 4.1). By Step 6, it
p
is enough to show statement An : f (na)1=2 (" ) =  f an1=2 ( 1=2 " ) for all positive integers n.
p
Set b = (1= )" . The statement A1 follows by a direct computation:
p p
1 1 
f (a)1=2 (" ) = (" + Q(a)1=2 " ) = (" + a) = (b + Q(a 1=2 )b 1 ) =  f a1=2 (b).
2 2 2

To proceed by induction, we need also to include the following in our induction:

(4.9) Bn : Q(a)1=2 ( f an1=2 (b)) 1

= Q(a 1=2 )( f an1=2 (b)) 1 , n = 1, 2, : : : .

Statement Bn is true for n = 1 because


p  1 
p  1
Q(a) 1=2 2
( f a1=2 (b)) = Q(a)
1

1=2
p" + Q(a1=2 )( ") 2
= Q(a)

1=2
p" + a
 
  1
2 (a) 1
=

p + p Q(a) 1=2 (a)
1

1 2
= ( 3=2 a
+  1=2 " ) 1
 
= 2( 1=2 a 1 + 1=2 " ) 1 = 2Q(a 1=2 )( 1=2 " + Q(a 1=2 )(1=2 " )) 1
= Q(a 1=2 )( f a1=2 (b)) 1 .

Next, suppose that Bn holds. This implies that

(4.10)  Q(a 1=2 )Q(a) 1=2

Q(a 1=2 )( f an1=2 (b)) 1
= Q(a 1=2 )( f an1=2 (b)) 1 .

Indeed, Q(a 1=2 )( f an1=2 (b)) 1 = Q(a)1=2 ( f an1=2 (b)) 1 = (1=)Q(a)1=2 ( f an1=2 (b)) 1
implies
that  Q(a) 1=2 Q(a 1=2 )( f an1=2 (b)) 1 = ( f an1=2 (b)) 1 . Then

Q(a)1=2 ( f an+1
1=2 (b))
1
= 2Q(a)1=2 ( f an1=2 (b) +  Q(a 1=2 )( f an1=2 (b)) 1 ) 1

= (Q(a) 1=2 ( f an1=2 (b)) +  Q(a) 1=2 Q(a 1=2 )( f an1=2 (b)) 1 ) 1

= 2(Q(a 1=2 )( f an1=2 (b)) +  Q(a) 1=2 Q(a 1=2 )( f an1=2 (b)) 1 ) 1

= 2Q(a 1=2 )( f an1=2 (b) +  Q(a 1=2 )Q(a) 1=2 Q(a 1=2 )( f an1=2 (b)) 1 ) 1

= 2Q(a 1=2 )( f an1=2 (b) + Q(a 1=2 )( f an1=2 (b)) 1 )

(4.10) 1

= Q(a 1=2 )( f a1=2 ( f an1=2 (b))) 1

= Q(a 1=2 )( f an+1
1=2 (b))
1
,

where the third equality follows from statement Bn and taking inverses.
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 807
p
Finally, suppose additionally that An holds, that is, f (na)1=2 (" ) =  fan1=2 (b). Then
p
a)1=2 (" ) = f (a)1=2 ( f (a)1=2 (" )) = f (a)1=2 (  f a 1=2 (b))
f (n+1 n n

1 p n p
= (  f a1=2 (b) + Q(a 1=2 )(  f an1=2 (b)) 1 )
p
2

= ( f an1=2 (b) + Q(a)1=2 ( f an1=2 (b)) 1 )
p
2

( f n1=2 (b) + Q(a 1=2 )( f an1=2 (b)) 1 )
(4.9)
=
p 2 a
p
=  f a1=2 ( f an1=2 (b)) =  f an+1
1=2 (b).

S TEP 9. For x, y 2 ,
d(x 1=2 , y 1=2 ) 
1
d(x, y).
2

Therefore Axiom 3.1 (iii) is satisfied. It is enough to show that M(x = y)  M(x 1=2 = y 1=2 )2 .
This follows from the Loewner-Heinz inequality and Step 8:

M(x = y) = inff > 0 : x  yg

p
 inff > 0 : x  ( y)1=2 =  y 1=2 g
1=2

= infft 2 : x 1=2  t y 1=2 g

= M(x 1=2 = y 1=2 )2 .

5. Symmetric cones

Our earlier results motivate the following definition.

D EFINITION 5.1. Let  be an open normal convex cone in a Banach space V

equipped with a ℄-symmetric structure making it a ℄-symmetric set. Then  is a
℄-symmetric cone if the following conditions are satisfied:
(i) x ℄ y  (x + y)=2 for all x, y 2 ;
(ii) the following maps are continuous:

(x, y) 7! x  y :    ! , (t, x, y) 7! x, y (t) := x ℄t y : R     ! ;

(iii) Every member of the displacement group G() extends to a bounded linear order-
preserving operator on V .
(iv) x ℄t  y = 1 t t (x ℄ y) for all ,  > 0, x, y 2  and t 2 R.
808 J. L AWSON AND Y. L IM

The next result follows essentially from Theorem 4.2.

Corollary 5.2. Let  be an open convex normal cone in a Banach space V .

Suppose that there is a pointed ℄-symmetric structure on  satisfying
(i) 2x  " + x 2
(ii) the squaring map x 7! x 2 = Q(x)" is continuous (in relative norm topology of ),
(iii) every basic displacement Q(x) is continuous and linear (that is, additive and
positively homogeneous) on .
Then  is a pointed symmetric space with convex metric, the Thompson metric whose
metric topology agrees with the relative topology, and also a ℄-symmetric cone. Converse-
ly, a ℄-symmetric cone satisfies these three conditions with respect to any pointing.

Proof. Assume conditions (i)–(iii). Note that since x 7! x 2 is a bijection, we can

rewrite hypothesis (i) in the form
(i0 ) x 1=2  (" + x)=2.
Thus the hypotheses of Theorem 4.2 are satisfied. Hence the geometric-arithmetic mean
inequality holds, i.e., condition 5.1 (i) is satisfied. Property 5.1 (ii) follows directly
from Theorem 4.2 and Theorem 3.3. By Step 1 of the proof of Theorem 4.2, each
member of G() extends to an invertible bounded linear operator on V that is an order
isomorphism, so Property 5.1 (iii) is valid. Property 5.1(iv) follows for dyadic rationals
t from Step 8 of Theorem 4.2 and Lemma 4.1:

x ℄t  y = Q((x)1=2 )(Q((x) 1=2

) y)t
= Q(1=2 x 1=2 )( Q( 1=2 x 1=2 )y)t
=  Q(x 1=2 )( 1  Q(x 1=2 )y)t
= ( 1 )t Q(x 1=2 )(Q(x 1=2 )y)t
= 1 t t (x ℄t y).

That it holds for all t 2 R then follows from continuity.

Step 9 of the proof of Theorem 4.2 establishes that the Thompson metric is a con-
vex metric, and Step 4 that its metric topology agrees with the relative topology.
Conversely assume that we choose some point " in the ℄-symmetric cone . Then
using 5.1 (i), we have
 
" + x2
2x = 2(" ℄ x )  2
2
= " + x 2,
2

and thus hypothesis (i) is satisfied. Hypotheses (ii) and (iii) follow immediately from
5.1 (ii) and 5.1 (iii) resp., since x 2 = x  " .
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 809

E XAMPLE 5.3. Let A be a unital C  -algebra with identity e, and let A+ be the
set of positive invertible elements of A. It follows readily from Corollary 5.2 and stan-
dard basic facts from the theory of C  -algebras that A+ is a ℄-symmetric cone. To see
this we need the standard basic facts that A+ is an open normal convex cone in the
closed subspace H(A) of hermitian elements, that each element of A+ has a unique
square root in A+ , and that x 2  0 for every x 2 H(A). The set A+ is a twisted sub-
group (closed under (x, y) 7! x y 1 x) with unique square roots of the multiplicative
group of invertible elements of A, hence a pointed ℄-symmetric set with respect to
x  y = x y 1 x and distinguished point the identity e. Furthermore, the powers com-
puted in the algebra agree with those computed in (A+ , , e) [14]. Hence condition (ii)
of Corollary 5.2 holds. Condition 5.2 (i) is equivalent to (e x)2  0, thus valid. Since
Q(x)y = x(y 1 ) 1 x = x yx, condition 5.2 (iii) holds.

The next lemma is elementary, but will prove useful for our purposes.

Lemma 5.4. Let A be a subset of [0, 1] that contains 0 and 1, is closed under
the operation of taking midpoints, and is closed under sequential limits. The A = [0, 1].

Theorem 5.5 (Loewner-Heinz, [2]). Let   V be a ℄-symmetric cone. If x1 

x2 and y1  y2 for x1 , x2 , y1 , y2 2 , then x1 ℄t y1  x2 ℄t y2 for 0  t  1.

Proof. If x1 ℄t y1  x1 ℄t y2 and x1 ℄t y2  x2 ℄t y2 , then we obtain our desired

conclusion by transitivity. Thus (using commutivity of ℄ and x ℄t y = y ℄1 t x) it suffices
to show that b  c implies a ℄t b  a ℄t c. By Corollary 5.2 we may choose any member
of  for our distinguished point, so without loss of generality we assume a = " . For
t = 1=2, ℄t = ℄, and we have by the Loewner-Heinz inequality (Theorem 4.2)

" ℄ b = b1=2  c1=2 = " ℄ c,

so the theorem is valid for t = 1=2.

Let b  c in . There exists a -homomorphism (and hence ℄-homomorphism)
b : R !  such that b (0) = " and b (1) = b; then by definition bt = " ℄t b = b (t).
Consider the set A of all t 2 [0, 1] such that bt = " ℄t b  " ℄t c = ct . For t = 0, we
have "  " and for t = 1 we have b  c, so 0, 1 2 A. Suppose that t1 , t2 2 A. Then
for t = (t1 + t2 )=2,

bt = b(t1 +t2 )=2 = bt1 ℄ bt2  ct 1

℄ c t2 = c t ,

where the inequality follows from the case t = 1=2 established above. By closedness
of the relation  and continuity of ℄t , A is closed under limits of sequential limits.
Thus by Lemma 5.4 A = [0, 1].
810 J. L AWSON AND Y. L IM

Theorem 5.6. The general harmonic-geometric-arithmetic mean inequality holds

in any ℄-symmetric cone , that is, for x, y 2  and t 2 [0, 1]:

((1 t)x 1
+ t y 1) 1
 x ℄t y  (1 t)x + t y.

Proof. Let A be the set of all t 2 [0, 1] for which the HGA-inequality holds. For
t = 0 (resp. t = 1) it reduces to x  x  x (resp. y  y  y) so 0, 1 2 A. By closedness
of the order and continuity of the operations, A is sequentially closed.
Let Ht := ((1 t)x 1 + t y 1 ) 1 , G t := x ℄t y, and At := (1 t)x + t y for 0  t  1.
Suppose that t, s 2 A. Then by elementary computation and the standard HGA-inequality
(Theorem 4.2),
 1 1 1
(Ht ) + (Hs )
H(t+s)=2 =
2
 1 1
s,t 2 A 1
 (G t ) + (G s )
2


H-G
Gt ℄ Gs
2.1 (4)
= G (t+s)=2 .

By an analogous computation, or by taking inverses, one obtains that G (t+s)=2  A(t+s)=2 .

By Lemma 5.4 A = [0, 1], yielding the theorem.

The HGA-inequalities provide an approximation scheme for x ℄ y.

Lemma 5.7. For x, y in a ℄-symmetric cone , x ℄ y = H (x, y) ℄ A(x, y), where

H (x, y) is the harmonic mean and A(x, y) is the arithmetic mean of x and y.

Proof. We have
1 1
x +y 1 1
Q(x ℄ y)H (x, y) 1
= Q(x ℄ y) = (x ℄ y.x 1
+ y ℄ x.y 1
)1 = (y + x) = A(x, y).
2 2 2

From the Riccati lemma it follows that x ℄ y = H (x, y) ℄ A(x, y).

Theorem 5.8. For x, y in a ℄-symmetric cone  in a Banach space V , define

H1 = H (x, y), the harmonic mean, and A1 = A(x, y), the arithmetic mean. Inductively
define Hn+1 = H (Hn , An ) and An+1 = A(Hn , An ). Then for each n,

Hn  Hn+1  x ℄ y  An+1  An ,
and Hn ! x ℄ y, An ! x ℄ y.
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 811

Proof. By Lemma 5.7 and induction, we have Hn ℄ An = x ℄ y for each n. The

asserted inequality then follows from the HGA-inequality. We fix " 2  and endow V
with the order-unit norm for the order unit " ; the topology of this norm agrees with that
of the original Banach space norm. In this norm the arithmetic mean of two points is
halfway between them in distance, and 0  x  y implies kx k  k y k, hence k y z k 
kx z k whenever z  y  x. From these facts we conclude that
k Hn+1 An+1 k  k Hn An+1 k =
1
2
k Hn A n k,

and thus k Hn An k  2 n kx y k for each n. Hence k Hn x ℄ y k  k Hn An k 

2 n kx y k for each n, so Hn ! x ℄ y. Similarly An ! x ℄ y.
The previous results have been obtained in symmetric cones of Euclidean Jordan
algebras [20].

Theorem 5.9 (The Furuta inequality, [11]). Let  be a ℄-symmetric cone in a

Banach space V and let 0 < b  a. If 0  p, q, r 2 R satisfies p + 2r  (1 + 2r )q and
1  q, then

b( p+2r )=q  (br .a p )1=q .

Proof. The proof is the same as given in [28] for Banach -algebras with contin-
uous involution, where the Loewner-Heinz inequality (Theorem 5.5), the order preserv-
ing property of the quadratic representations, the order reversing property of inversion
(Theorem 4.2), and the equality (3.1) are applied as the main tools.

6. JB-algebras and symmetric cones

In this section we illustrate and apply our previously results in the context of
JB-algebras. A basic reference for the theory of JB-algebras, particulary the results
we need in what follows, is the book of Hanche-Olsen and Størmer [12].
A Jordan algebra is a vector space Z with a commutative multiplication x y such
that x(x 2 y) = x 2 (x y) holds for x, y 2 Z . An involution on a complex Jordan algebra Z
is an antilinear involutive map z 7! z  with (z w ) = w  z  for all z, w 2 Z . A JB-algebra
V is a real Jordan algebra with unit e endowed with a complete norm k
k such that

kz wk  kz k kwk, kz 2 k = kz k2 , kz k2  kz 2 + w2 k.
A JB -algebra is a complex Banach space Z endowed with the structure of a Jordan
algebra with involution  such that

kz wk  kz k
kwk, kfzz  z gk = kz k3
812 J. L AWSON AND Y. L IM

for z, w 2 Z , where fx y  z g = (x y  )z + x(y  z) y  (x z). There is a one-to-one corre-

spondence between JB-algebras and JB -algebras: For any JB -algebra Z , the hermit-
ian part V := fx 2 Z: x  = x g is a JB-algebra under the restricted norm. Conversely, for
every JB-algebra V the complexified algebra Z := V + i V has a unique norm making
Z equipped with the canonical involution a JB -algebra ([30], [7], [12], [25]).
Let V be a JB-algebra. For x 2 V we write L(x)(y) = x y, the multiplication op-
erator. We consider the set

 := fx 2 V : Spec(L(x))  (0, 1)g.

Then  is an open convex cone of V (see [29, Section 21], particularly [29, Proposi-
tion 21.19], also [12, Section 3.3]) and is realized as

 = exp(V ) := fexp(x) : x 2 V g.
The Banach algebra norm agrees with the order unit norm

jx je := infft > 0 : te  x  0g,

or equivalently  is a normal cone ([1, Theorem 2.2], [12, Proposition 3.3.10], or [29,
Proposition 21.19]). The quadratic representation of the Jordan algebra is defined by
P(z) = 2L(z)2 L(z 2 ). It is well-known that for each z 2 , P(z) 2 G(), the linear
automorphism group of . In fact, there is a polar decomposition G() = P() Aut(V )
where Aut(V ) denotes the Jordan automorphism group of V (see [29, Corollary 22.29]).
We further note ([29, Proposition 22.27]) that Aut(V ) = fg 2 G(): g(e) = eg. The basic
properties

P(z)z 1
= z, P(z) 1
= P(z 1 ), P(P(z)w ) = P(z)P(w )P(z)

([29, Corollary 19.9 and Proposition 19.18]) yield a pointed symmetric set structure
x  y = P(x)y 1 with " := e as base point on the set of invertible elements, in particular
on the cone  (see p.67 of [21]; see also the discussion in Section 3.9 of [12]). In
symmetric set notation, P(a) = Q(a) and the symmetric set inverse a 1 := e  a agrees
with the Jordan inverse of a.
Next, we show that the pointed symmetric space (, " = e) is ℄-symmetric. Let
x, y 2  such that x 2 = y 2 . Then by the commutativity of Jordan products, 0 = x 2 + y 2 =
L(x + y)(x y). Since L(z) is invertible for all z 2  ([29, Proposition 21.19 and
Corollary 21.22], [12, Lemma 3.2.10]), x y = 0. This implies that each element of 
has a unique square root in . Note that if a = exp(x), x 2 V then a 1=2 = exp((1=2)x).
Moreover, if a, b 2  then the quadratic equation

1
P(x)a =b
 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 813

has a unique solution in . Note that x = P(a 1=2 )(P(a 1=2 )b)1=2 2  solves the equa-
tion (cf. [14] and [19]). Suppose that x and y are solutions in . Then

(P(a 1=2 )x)2 = P(P(a 1=2 )x)" = P(a 1=2 )P(x)P(a 1=2 )"
= P(a 1=2 )(P(x)a 1 )
= P(a 1=2 )b = P(a 1=2 )(P(y)a 1 )
= P(a 1=2 )P(y)P(a 1=2 )"
= (P(a 1=2 )y)2

and hence P(a 1=2 )x = P(a 1=2 )y, so x = y. We conclude that the open convex cone
 is a ℄-symmetric set under the operation x  y = P(x)y 1 . In this case the dyadic
power a t of a = exp(x) agrees with exp(t x) and the geometric mean a ℄ b of a and b is

a ℄ b = P(a 1=2 )(P(a 1=2 )b)1=2 .

Corollary 6.1. Let V be a JB-algebra and let  be the associated symmetric

cone. Then  is a symmetric space with convex metric with respect to the Thompson
metric. In particular, the harmonic-geometric-arithmetic mean inequality (4.3) and the
Loewner-Heinz inequality (4.4) hold and the distance function between points evolving
in time on two geodesics is a convex function.

Proof. Let x 2 . Since the subalgebra generated by e and x is isometrically

(order and algebra)-isomorphic to C(X ) for some compact Hausdorff space ([1, Propo-
sition 2.3]), the inequality

" ℄ x = x 1=2 
e+x
2

holds. The squaring map x 7! x 2 is continuous (Banach algebra product). The qua-
dratic representation P(x) is obviously a bounded linear operator. This shows that the
pointed ℄-symmetric set  satisfies the three conditions of Theorem 4.2. The last as-
sertion then follows from Theorem 3.4.

The last assertion of the preceding corollary provides a positive answer to a ques-
tion of Neeb [22]. Neeb considers a JB -algebra Z and the associated symmetric cone
 in the real JB-subalgebra V = fz 2 Z : z  = z g with the Finsler structure on  given
by jv jex = ke L(x) v k for x, v 2 V ([22, Example 6.6]). The geodesic line passing
through (0) = e x and (1) = e y is given by (t) = e L(x) (e L(x) e y )t . One of Neeb’s
questions concerns the convexity of the real function with respect to the Finsler metric
distance

(6.1) t 7! d(x, (t))

814 J. L AWSON AND Y. L IM

where x 2  and is a geodesic. From P(exp u) = exp 2L(u) for u 2V ([29, Corol-
lary 22.8]), we see that for x, y 2 ,

jvjx = jvje log x = ke L(log x)

v k = ke2L(log x
1=2
)
v k = k P(x 1=2
)v k

(log x is well-defined for any x 2  from the fact that the closed unital subalgebra
generated by x is an abelian C  -algebra, [29, Lemma 20.33]) and the geodesic line
passing through x and y is

(t) = P(x 1=2 )(P(x 1=2

)y)t = x ℄t y.

Since the JB -algebra norm agrees with the order unit norm,

jvjx = infft > 0 : te  P(x 1=2 )v  teg = infft > 0 : tx  v  t x g.

This implies that the Finsler distance is exactly the Thopmson part metric fromd (4.1)
and hence the function (6.1) is convex.

R EMARK 6.2. The harmonic-geometric-arithmetic mean inequality and the

Loewner-Heinz inequality with applications to the Finsler geometry of finite dimensional
symmetric cones are studied in [17], [18], [19] and [20]. It has recently been discovered
by Bhatia [6] that the non-positive curvature property of the convex cone of positive def-
inite matrices holds for metrics inherited from symmetric gauge functions.

E XAMPLE 6.3. The hermitian elements x = x  of any C  -algebra form a

JB-algebra with respect to the symmetric product x Æ y := (x y + yx)=2. In this case 
is the cone of positive elements of Example 5.3. Spin factors [12, Chapter 6], which
arose in the study of anticommutation relations in physics, provide another type of ex-
ample. Given a real Hilbert space H , let A = H  R1 have the norm ka + 1k = ka k + jj
and define a product in A by

(a + 1) Æ (b + 1) = ((a + b) + (ha, bi + )1.

Then A is a JB-algebra, and hence its corresponding cone  satisfies the hypotheses,
and hence conclusions of Theorem 4.2.

7. Hermitian Banach -algebras

D EFINITION 7.1. Let Z be a unital Banach algebra Z with a continuous invo-

lution  and let X consist of the self-adjoint elements of Z . Let e denote the unit
element of Z . The unital Banach algebra Z is called hermitian if  (x)  R and kx k =
supj (x)j for every x = x  . We let

 := fx = x  :  (x)  (0, 1)g.

 M ETRIC C ONVEXITY OF S YMMETRIC C ONES 815

We note that by the Shirali-Ford theorem [26]  (zz  )  [0, 1) for every z 2 Z .
If z is invertible then  (zz  )  (0, 1).

Theorem 7.2. Let Z be a hermitian Banach algebra. Then  is a ℄-symmetric

cone of X .

Proof. It is shown in [29, Corollary 14.16] that  is an open convex cone of X

and the order unit norm with respect to e coincides with the given norm k
k, which
implies the normality of .
For a 2 , we denote a 1=2 := exp((1=2) log a) where log denotes the principal
branch of the complex logarithm. Then (a 2 )1=2 = a ([28, Lemma 6]) and therefore
each element in  has a unique square root in . Moreover, if a, b 2  then ab 1 a =
(ab 1=2 )(ab 1=2 ) is contained in the cone  since it is invertible and  (zz  )  [0, 1).
This shows that  is a uniquely 2-divisible twisted subgroup of G(Z ), the group of
invertible elements of Z .
The conditions of Corollary 5.2 hold; the verification is similar to the case for
C  -algebras (Example 5.3).

ACKNOWLEDGEMENT . The authors thank the referee for a careful reading of the
manuscript and helpful suggestions. This work was supported by grant No. 2006020300
from the Joint Research Program of Kosef.

References

[1] E.M. Alfsen, F.W. Shultz and E. Størmer: A Gelfand-Neumark theorem for Jordan algebras,
Advances in Math. 28 (1978), 11–56.
[2] T. Ando: Topics on Operator Inequalities, Lecture Notes Hokkaido Univ., Sapporo, 1978.
[3] E. Andruchow, G. Corach and D. Stojanoff: Geometrical significance of Löwner-Heinz in-
equality, Proc. Amer. Math. Soc. 128 (2000), 1031–1037.
[4] E. Andruchow, G. Corach, and D. Stojanoff: Löwner’s theorem and the differential geometry
of the space positive operators, preprint.
[5] W. Ballmann: Lectures on Spaces of Nonpositive Curvature, DMV Seminar 25, Birkhäuser,
Basel, 1995.
[6] R. Bhatia: On the exponential metric increasing property, Linear Algebra Appl. 375 (2003),
211–220.
[7] R. Braun, W. Kaup and H. Upmeier: A holomorphic characterization of Jordan C  -algebras,
Math. Z. 161 (1978), 277–290.
[8] M.R. Bridson and A. Haefliger: Metric Spaces of Non-Positive Curvature, Grundlehren der
Math. Wissenschaft 319, Springer, Berlin, 1999.
[9] G. Corach, H. Porta and L. Recht: Convexity of the geodesic distance on spaces of positive
operators, Illinois J. Math. 38 (1994), 87–94.
[10] P. Eberlein: Geometry of Non-Positively Curved Manifolds, Chicago Lectures in Math., U.
Chicago Press, 1996.
[11] T. Furuta: A  B  0 assures (B r A p B r )1=q  B ( p+2r )=q for r  0, p  0, q  1 with (1 + 2r )q 
p + 2r , Proc. Amer. Math. Soc. 101 (1987), 85–88.
816 J. L AWSON AND Y. L IM

[12] H. Hanche-Olsen and E. Størmer: Jordan Operator Algebras, Monographs and Studies in Math.
21, Pitman, Boston, 1984.
[13] S. Lang: Fundamentals of Differential Geometry, Graduate Texts in Math. 191, Springer, Hei-
delberg, 1999.
[14] J. Lawson and Y. Lim: Symmetric sets with midpoints and algebraically equivalent theories,
Results Math. 46 (2004), 37–56.
[15] J. Lawson and Y. Lim: Means on dyadic symmetric sets and polar decompositions, Abh. Math.
Sem. Univ. Hamburg 74 (2004), 135–150.
[16] J. Lawson and Y. Lim: Symmetric spaces with convex metrics, to appear, Forum Math.
[17] J. Lawson and Y. Lim: Solving symmetric matrix word equations via symmetric space machin-
ery, Linear Algebra Appl. 414 (2006), 560–569.
[18] Y. Lim: Finsler metrics on symmetric cones, Math. Ann. 316 (2000), 379–389.
[19] Y. Lim: Geometric means on symmetric cones, Arch. Math. (Basel) 75 (2000), 39–45.
[20] Y. Lim: Applications of geometric means on symmetric cones, Math. Ann. 319 (2001), 457–468.
[21] O. Loos: Symmetric Spaces, I: General Theory, W.A. Benjamin, Inc., New York-Amsterdam,
1969.
[22] K.-H. Neeb: A Cartan-Hadamard theorem for Banach-Finsler manifolds, Geom. Dedicata 95
(2002), 115–156.
[23] R.D. Nussbaum: Hilbert’s Projective Metric and Iterated Nonlinear Maps, Memoirs of Amer.
Math. Soc. 391, 1988.
[24] R.D. Nussbaum: Finsler structures for the part metric and Hilbert’s projective metric and appli-
cations to ordinary differential equations, Differential Integral Equations 7 (1994), 1649–1707.
[25] H. Olsen and E. Størmer: Jordan Operator Algebras, Monographs Stud. Math. 21, Pitman,
London, 1984.
[26] S. Shirali and J.W.M. Ford: Symmetry in complex involutory Banach algebras II, Duke Math.
J. 37 (1970), 275–280.
[27] A.C. Thompson: On certain contraction mappings in a partially ordered vector space, Proc.
Amer. Math. soc. 14 (1963), 438–443.
[28] K. Tanahashi and A. Uchiyama: The Furuta inequality in Banach -algebras, Proc. Amer.
Math. Soc. 128 (2000), 1691–1695.
[29] H. Upmeier: Symmetric Banach Manifolds and Jordan C  -algebras, North-Holland Mathemat-
ics Studies 104 North-Holland, Amsterdam, 1985.
[30] J.D.M. Wright: Jordan C  -algebras, Michigan Math. J. 24 (1977), 291–302.

Jimmie Lawson
Department of Mathematics
Louisiana State University
Baton Rouge, LA70803
USA
e-mail: [email protected]

Yongdo Lim
Department of Mathematics
Kyungpook National University
Taegu 702–701
Korea
e-mail: [email protected]