Académique Documents
Professionnel Documents
Culture Documents
CARI
C
(Dedicated to the memory of Gert K. Pedersen)
Abstract
We give a general formulation of Jensens operator inequality for unital elds of positive linear
mappings, and we consider different types of converse inequalities.
1. Introduction
Let I be a real interval of any type. A continuous function f : I R is said
to be operator convex if
(1) f (x + (1 )y) f (x) + (1 )f (y)
holds for each [0, 1] and every pair of self-adjoint operators x and y
(acting) on an innite dimensional Hilbert space H with spectra in I (the
ordering is dened by setting x y if y x is positive semi-denite).
Let f be an operator convex function dened on an interval I. Ch. Davis
[4] proved
1
a Schwarz inequality
(2) f ((x)) (f (x)) ,
where : A B(H) is a unital completely positive linear map from a C
-
algebra A to linear operators on a Hilbert space H, and x is a self-adjoint
element in A with spectrum in I. Subsequently M. D. Choi [3] noted that it is
enough to assume that is unital and positive. In fact, the restriction of to the
commutative C
i=1
a
i
x
i
a
i
_
n
i=1
a
i
f (x
i
)a
i
for operator convex functions f dened on an interval I = [0, ) (with
and f (0) 0) and self-adjoint operators x
1
, . . . , x
n
with spectra in I assuming
that
n
i=1
a
i
a
i
= 1. The restriction on the interval and the requirement f (0)
0 was subsequently removed by B. Mond and J. Pe cari c in [13], cf. also [8].
The inequality (3) is in fact just a reformulation of (2) although this was not
noticed at the time. It is nevertheless important to note that the proof given in
[7] and thus the statement of the theorem, when restricted to n n matrices,
holds for the much richer class of 2n 2n matrix convex functions. Hansen
and Pedersen used (3) to obtain elementary operations on functions, which
leave invariant the class of operator monotone functions. These results then
served as the basis for a new proof of Lwners theorem applying convexity
theory and Krein-Milmans theorem.
Finally B. Mond and J. Pe cari c [12] proved the inequality
(4) f
_
n
i=1
w
i
i
(x
i
)
_
n
i=1
w
i
i
(f (x
i
))
for operator convex functions f dened on an interval I, where
i
: B(H)
B(K) are unital positive linear maps, x
1
, . . . , x
n
are self-adjoint operators with
spectra in I and w
1
, . . . , w
n
are non-negative real numbers with sum one.
The aim of this paper is to nd an inequality which contains (3), (2) and
(4) as special cases. Since the inequality in (4) was the motivating step for
obtaining converses of Jensens inequality using the so called Mond-Pe cari c
method, we also give some results pertaining to converse inequalities in the
new formulation.
2. The main result
Continuous Fields of Operators
Let T be a locally compact Hausdorff space and let A be a C
-algebra. We
say that a eld (x
t
)
t T
of operators in A is continuous if the function t x
t
is norm continuous on T . If in addition is a Radon measure on T and
the function t x
t
is integrable, then we can form the Bochner integral
_
T
x
t
d(t ), which is the unique element in A such that
__
T
x
t
d(t )
_
=
_
T
(x
t
) d(t )
jensens operator inequality and its converses 63
for every linear functional in the norm dual A
t
: A B from A to another C
-algebras. If (
t
)
t T
is a unital
eld of positive linear mappings
t
: A B dened on a locally compact
Hausdorff space T with a bounded Radon measure , then the inequality
(5) f
__
T
t
(x
t
) d(t )
_
_
T
t
(f (x
t
)) d(t )
holds for every bounded continuous eld (x
t
)
t T
of self-adjoint elements in A
with spectra contained in I.
Proof. We rst note that the function t
t
(x
t
) B is continuous and
bounded, hence integrable with respect to the bounded Radon measure .
We may organize the set CB(T, A) of bounded continuous functions on T
with values in A as a normed involutive algebra by applying the point-wise
operations and setting
(y
t
)
t T
= sup
t T
y
t
, (y
t
)
t T
CB(T, A),
and it is not difcult to verify that the norm is already complete and satisfy
the C
-algebra theory. It
follows that f ((x
t
)
t T
) = (f (x
t
))
t T
. We then consider the mapping
: CB(T, A) B
dened by setting
((x
t
)
t T
) =
_
T
t
(x
t
) d(t ),
andnote that it is a unital positive linear map. Settingx = (x
t
)
t T
CB(T, A),
we use inequality (2) to obtain
f (((x
t
)
t T
)) = f ((x)) (f (x)) = (f ((x
t
)
t T
)) = ((f (x
t
))
t T
),
which is the statement of the theorem.
To illustrate various techniques in proving Jensens operator inequality, we
give two proofs of Theorem 2.1 in the discrete case T = {1, . . . , n}.
64 frank hansen, josip pe cari c and ivan peri c
Proof. By using the continuity of f and uniform approximation of self-
adjoint operators by simple operators, we may assume that x
i
has a spectral
resolution on the form
x
i
=
jJ
i
t
i,j
e
i,j
, i = 1, . . . , n,
where each J
i
is a nite set and
jJ
i
e
i,j
= 1. We then have
f
_
n
i=1
i
(x
i
)
_
= f
_
n
i=1
i
_
jJ
i
t
i,j
e
i,j
__
= f
_
n
i=1
jJ
i
t
i,j
i
(e
i,j
)
_
= f
_
n
i=1
jJ
i
_
i
(e
i,j
) t
i,j
_
i
(e
i,j
)
_
i=1
jJ
i
_
i
(e
i,j
)f (t
i,j
)
_
i
(e
i,j
)
=
n
i=1
i
_
jJ
i
f (t
i,j
)e
i,j
_
=
n
i=1
i
(f (x
i
)) .
The second proof. We use an the idea from [9], confer also [13]. If f is
operator convex in I = [0, 1) and f (0) 0, then there is a connection such
that f (t ) = t (1 t ). We use the following two properties of a connection,
cf. [1], [9].
(1) (ab) (a)(b) for positive linear maps and positive operators
a and b.
(2)
n
i=1
a
i
b
i
_
n
i=1
a
i
_
_
n
i=1
b
i
_
for positive n-tuples (a
1
, . . . , a
n
)
and (b
1
, . . . , b
n
). (subadditivity)
We then obtain
i=1
i
(f (x
i
)) =
n
i=1
i
(x
i
(1 x
i
))
n
i=1
i
(x
i
)
i
(1 x
i
)
_
n
i=1
i
(x
i
)
_
_
n
i=1
i
(1 x
i
)
_
=
_
n
i=1
i
(x
i
)
_
_
1
n
i=1
i
(x
i
)
_
= f
_
n
i=1
i
(x
i
)
_
.
jensens operator inequality and its converses 65
Consider now an arbitrary operator convex function f dened on [0, 1). The
function
f (x) = f (x) f (0) satises
f (0) = 0 hence
(6) f
_
n
i=1
i
(x
i
)
_
f (0)1
n
i=1
i
(f (x
i
)) f (0)
n
i=1
i
(1)
from which the statement follows. We nally obtain (5) in this setting for a
function f dened on an arbitrary interval [, ) by considering the function
g(x) = f (( )x + ) on the interval [0, 1).
Inequality (5) is obviously a generalization of the inequalities (3), (2) and
(4).
3. Converses of Jensens inequality
The following theorem should be compared with Theorem 2.3 in [10]. For a
function f : [m, M] R we use the standard notation:
f
=
f (M) f (m)
M m
and
f
=
Mf (m) mf (M)
M m
.
We will apply functions F(t, s) of two real variables to operators. In simple
cases, when it makes sense, we may just replace numbers with operators. If for
example F(t, s) = s
1
t
1/2
s
1
, we may set F(x, y) = y
1
x
1/2
y
1
. Otherwise
we may use the functional calculus on tensor products, see for example [6].
Theorem 3.1. Let (x
t
)
t T
be a bounded continuous eld of self-adjoint
elements in a unital C
t
(f (x
t
)) d(t ), g
__
T
t
(x
t
) d(t )
__
sup
mzM
F
_
f
z +
f
, g(z)
_
1.
In the dual case (when f is concave) the opposite inequality holds in (7) with
inf instead of sup.
66 frank hansen, josip pe cari c and ivan peri c
Proof. For convex f the inequality f (z)
f
z +
f
holds for every
z [m, M]. Thus, by using functional calculus, f (x
t
)
f
x
t
+
f
1 for every
t T . Applying the positive linear maps
t
and integrating, we obtain
_
T
t
(f (x
t
)) d(t )
f
_
T
t
(x
t
) d(t ) +
f
1.
Now, using operator monotonicity of F(, v), we obtain
F
__
T
t
(f (x
t
)) d(t ), g
__
T
t
(x
t
) d(t )
__
F
_
f
_
T
t
(x
t
) d(t ) +
f
1, g
__
T
t
(x
t
) d(t )
__
sup
mzM
F
_
f
z +
f
, g(z)
_
1.
Numerous applications of the previous theorem can be given (see [10]). We
give generalizations of some results from [14].
Theorem 3.2. Let (A
t
)
t T
be a continuous eld of positive operators on
a Hilbert space H dened on a locally compact Hausdorff space T equipped
with a Radon measure . We assume the spectra are in [m, M] for some
0 < m < M. Let furthermore (x
t
)
t T
be a continuous eld of vectors in H
such that
_
T
x
t
2
d(t ) = 1. Then for any 0, p 1 and q 1 we have
(8)
__
T
A
p
t
x
t
, x
t
d(t )
_
1/q
_
T
A
t
x
t
, x
t
d(t ) C(, m, M, p, q),
where the constant
(9)
C(, m, M, p, q) =
M
_
M
p
q
1
_
, 0 <
p
q
M
p
_
1
q
1
_
q1
q
_
q
_ 1
1q
+
p
p
,
p
q
M
p
_
1
q
1
_
p
q
m
p
_
1
q
1
_
m
_
m
p
q
1
_
,
p
q
m
p
_
1
q
1
_
and
p
and
p
are the constants
f
and
f
associated with the function f (z) =
z
p
.
jensens operator inequality and its converses 67
Proof. Applying Theorem 3.1 for the functions
f (z) = z
p
, F(u, v) = u
1/q
v,
and unital elds of positive linear maps
t
: B(H) C dened by set-
ting
t
(A) = Ax
t
, x
t
for t T , the problem is reduced to determine
sup
mzM
H(z) where H(z) = (
p
z +
p
)
1/q
z.
The following Corollary is a generalization of Theorem 5 in [14]. The r-
geometric mean A#
r
B introduced by F. Kubo and T. Ando in [11] is dened
by setting
A#
r
B = A
1/2
_
A
1/2
BA
1/2
_
r
A
1/2
for positive invertible operators A and B.
Corollary 3.3. Let (A
t
)
t T
and (B
t
)
t T
be continuous elds of posit-
ive invertible operators on a Hilbert space H dened on a locally compact
Hausdorff space T equipped with a Radon measure such that
m
1
1 A
t
M
1
1 and m
2
1 B
t
M
2
1
for all t T for some 0 < m
1
< M
1
and 0 < m
2
< M
2
. Then for any
0, s 1, p 1 and any continuous eld (x
t
)
t T
of vectors in H such
that
_
T
x
t
2
d(t ) = 1 we have
(10)
__
T
A
p
t
x
t
, x
t
d(t )
_
1/p
__
T
B
q
t
x
t
, x
t
d(t )
_
1/q
_
T
B
q
t
#
1/s
A
p
t
x
t
, x
t
d(t ) C
_
,
m
p/s
1
M
q/s
2
,
M
p/s
1
m
q/s
2
, s, p
_
M
q
2
where the constant C is dened in Theorem 3.2 and 1/p + 1/q = 1.
Proof. By using Theorem 3.2 we obtain for any 0, for any continuous
eld (C
t
)
t T
of positive operators with m1 C
t
M1 and a square integrable
continuous eld (y
t
)
t T
of vectors in H the inequality
(11)
__
T
C
s
t
y
t
, y
t
d(t )
_
1/p
__
T
y
t
, y
t
d(t )
_
1/q
_
T
C
t
y
t
, y
t
d(t ) C(, m, M, s, p)
_
T
y
t
, y
t
d(t ).
Set now C
t
=
_
B
q/2
t
A
p
t
B
q/2
t
_
1/s
and y
t
= B
q/2
t
x
t
for t T in (11) and
observe that
m
p/s
1
M
q/s
2
1
_
B
q/2
t
A
p
t
B
q/2
t
_
1/s
M
p/s
1
m
q/s
2
1.
68 frank hansen, josip pe cari c and ivan peri c
By using the denition of the 1/s-geometric mean and rearranging (11) we
obtain
__
T
A
p
t
x
t
, x
t
d(t )
_
1/p
__
T
B
q
t
x
t
, x
t
d(t )
_
1/q
_
T
B
q
t
#
1/s
A
p
t
x
t
, x
t
d(t )
C
_
,
m
p/s
1
M
q/s
2
,
M
p/s
1
m
q/s
2
, s, p
__
T
B
q
t
x
t
, x
t
d(t )
C
_
,
m
p/s
1
M
q/s
2
,
M
p/s
1
m
q/s
2
, s, p
_
M
q
2
which gives (10).
In the present context we may obtain results of the Li-Mathias type by
using Theorem 3.1 and the following result which is a simple consequence of
Theorem 2.1.
Theorem 3.4. Let (x
t
)
t T
be a bounded continuous eld of self-adjoint
elements in a unital C
t
(f (x
t
)) d(t ), g
__
T
t
(x
t
) d(t )
__
inf
mzM
F [f (z), g(z)] 1.
In the dual case (when f is operator concave) the opposite inequality holds
with sup instead of inf.
We also give generalizations of some results from [5].
Theorem 3.5. Let f be a convex function on [0, ) and let be a nor-
malized unitarily invariant norm on B(H) for some nite dimensional Hilbert
space H. Let (
t
)
t T
be a unital eld of positive linear maps
t
: B(H)
B(K), where K is a Hilbert space, dened on a locally compact Hausdorff
space T equipped with a Radon measure . Then for every continuous eld of
jensens operator inequality and its converses 69
positive operators (A
t
)
t T
we have
(13)
_
T
t
(f (A
t
)) d(t ) f (0)1 +
_
T
f (A
t
) f (0)
A
t
t
(A
t
) d(t ).
Especially, for f (0) 0, the inequality
(14)
_
T
t
(f (A
t
)) d(t )
_
T
f (A
t
)
A
t
t
(A
t
) d(t ).
is valid.
Proof. Since f is a convex function, f (x)
Mx
Mm
f (m) +
xm
Mm
f (M)
for every x [m, M] where m M. Since is normalized and unitarily
invariant, we have 0 < A
t
A
t
1 and thus
f (A
t
)
A
t
1 A
t
A
t
f (0) +
A
t
A
t
f (A
t
)
for every t T . Applying positive linear maps and integrating we obtain
(15)
_
T
t
(f (A
t
)) d(t )
f (0)
_
1
_
T
t
(A
t
)
A
t
d(t )
_
+
_
T
f (A
t
)
A
t
t
(A
t
) d(t )
or
(16)
_
T
t
(f (A
t
)) d(t ) f (0)1 +
_
T
f (A
t
) f (0)
A
t
t
(A
t
) d(t ).
Note that since
_
T
t
(A
t
)
A
t
d(t )
_
T
A
t
t
(1)
A
t
(x), f
+
(x)], x (m, M),
where f
and f
+
are the one-sided derivatives of f and k(m) = f
+
(m) and
k(M) = f
t
(x
t
)d(t ) y1
_
_
T
t
(f (x
t
)) d(t )
f (x)1 x
_
T
t
(k(x
t
)) d(t ) +
_
T
t
(k(x
t
)x
t
) d(t )
for every x, y [m, M], where k is a subdifferential function of f on [m, M].
In the dual case (f is concave) the opposite inequality holds.
Proof. Since f is convex we have f (x) f (y) + k(y)(x y) for every
x, y [m, M]. By using the functional calculus it then follows that f (x
t
)
f (y)1 + k(y)(x
t
y1) for t T . Applying the positive linear maps
t
and
integrating, LHS of (17) follows. The RHS of (17) follows similarly by using
the functional calculus in the variable y.
Numerous inequalities can be obtained from(17). For example, LHSof (17)
may be used to obtain an estimation from below in the sense of Theorem 3.1.
Namely, the following theorem holds.
Theorem 3.8. Let (x
t
)
t T
be a bounded continuous eld of self-adjoint
elements in a unital C
t
(f (x
t
)) d(t ), g
__
T
t
(x
t
) d(t )
__
inf
mzM
F [f (y) + k(y)(z y), g(z)] 1
for every y [m, M]. In the dual case (when f is concave) the opposite
inequality holds in (18) with sup instead of inf.
Using LHS of (17) we can give generalizations of some dual results from
[5].
Theorem 3.9. Let (x
t
)
t T
be a bounded continuous eld of positive ele-
ments in a unital C
-algebra B acting on a
nite dimensional Hilbert space K. Let be a unitarily invariant norm on
B(K) and let f : [0, ) R be an increasing function.
(1) If 1 = 1 and f is convex with f (0) 0 then
(19) f
__
_
_
_
_
T
t
(x
t
) d(t )
_
_
_
_
_
_
_
_
_
_
T
t
(f (x
t
)) d(t )
_
_
_
_
.
(2) If
_
T
t
(x
t
) d(t )
_
T
t
(x
t
) d(t )1 and f is concave then
(20)
_
T
t
(f (x
t
)) d(t ) f
__
_
_
_
_
T
t
(x
t
) d(t )
_
_
_
_
_
1.
Proof. Since f (0) 0 and f is increasing we have k(y)y f (y) 0
and k(y) 0. From (17) and the triangle inequality we have
k(y)
_
_
_
_
_
T
t
(x
t
) d(t )
_
_
_
_
_
_
_
_
_
T
t
(f (x
t
))
_
_
_
_
+ (k(y)y f (y)).
Now (19) follows by setting y =
_
_
_
T
t
(x
t
) d(t )
_
_
. Inequality (20) follows
immediately from the assumptions and from the dual case of LHS in (17) by
setting y =
_
_
_
T
t
(x
t
) d(t )
_
_
.
Finally, to illustrate how RHS of (17) works, we set
x =
_
_
_
T
t
(k(x
t
)x
t
) d(t )
_
_
_
_
_
T
t
(k(x
t
)) d(t )
_
_
72 frank hansen, josip pe cari c and ivan peri c
and obtain a Slater type inequality
_
T
t
(f (x
t
)) d(t ) f
_
_
_
_
T
t
(k(x
t
)x
t
) d(t )
_
_
_
_
_
T
t
(k(x
t
)) d(t )
_
_
_
1
under the condition
_
T
t
(k(x
t
)x
t
) d(t )
_
_
_
T
t
(k(x
t
)x
t
) d(t )
_
_
_
T
t
(k(x
t
)) d(t )
_
_
_
T
t
(k(x
t
)) d(t )
_
_
.
REFERENCES
1. Aujla, J. S., Vasudeva, H. I., Operator inequalities related to means of operators, Math. Japon.
41 (1995), 383388.
2. Bhatia, R., Matrix Analysis, Springer, Germany, 1996.
3. Choi, M. D., A Schwarz inequality for positive linear maps on C
CEV TRG 19
10000 ZAGREB
CROATIA
E-mail: iperic@pbf.hr