Jensen's operator inequality and its converses

JENSEN'S OPERATOR INEQUALITY AND ITS CONVERSES

FRANK HANSEN, JOSIP PECARIC, AND IVAN PERIC

Abstract

We give a general formulation of Jensen's operator inequality for unital fields of positive linear
mappings, and we consider different types of converse inequalities.

1. Introduction

Let / be a real interval of any type. A continuous function / : / ^ R is said to be
operator convex if

(1) f{Xx+{l^X)y)<Xf{x) + {l-\)f{y)

holds for each A G [0, 1] and every pair of self-adjoint operators x and y (acting) on
an infinite dimensional Hilbert space H with spectra in / (the ordering is defined
by setting x < y if y — x is positive semi-definite) .

Let f be an operator convex function defined on an interval /. Ch. Davis [3]
provecu a Schwarz inequality

(2) f{^ix))<^{fix)),

where cj): A ^ B{H) is a unital completely positive linear map from a C*-algebra
A to linear operators on a Hilbert space -ff, and a; is a self-adjoint element in A with
spectrum in /. Subsequently M. D. Choi [3] noted that it is enough to assume that
4> is unital and positive. In fact, the restriction of 4> to the commutative C*-algebra
generated by x is automatically completely positive by a theorem of Stinespring.
F. Hansen and G. K. Pedersen [7] proved a Jensen type inequality

Cn \ n

^ a*x^a^ \ < a*f{xi)ai

for operator convex functions / defined on an interval / — [0, a) (with a < oo and
/(O) < 0) and self-adjoint operators xi,...,Xn with spectra in / assuming that
X]i=i o-l.'^i — 1- The restriction on the interval and the requirement /(O) < was
subsequently removed by B. Mond and J. Pecaric in [13], cf. also [S].

The inequality ^ is in fact just a reformulation of ^ 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 x n matrices, holds
for the much richer class of 2n x 2n matrix convex functions. Hansen and Pedersen
used Q 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 Lowner's theorem applying convexity theory and Krein-Milman's
theorem.

^There is small typo in the proof. Davis states that (j) by Stinespring's theorem can be written
on the form = Pp(x)P where p is a *-homomorphism to B{H) and P is a projection acting
on H. In fact, H may be embedded in a Hilbert space K on which p and P acts. The theorem then
follows by the calculation f{<t){x)) = f{Pp{x)P) < Pf(p{x))P = Pp{f{x))P = 4>{f{x)), where the
pinching inequality, proved by Davis in the same paper, is applied.

1

2

FRANK HANSEN, JOSIP PECARIC, AND IVAN PERIC

Finally B. Mond and J. Pecaric fT2] proved the inequality

(71 \ n

i=l I i=l

for operator convex functions / defined on an interval /, where (f)i : B{H) B{K)
are unital positive linear maps, xi, . . . ,Xn are self-adjoint operators with spectra in
/ and wi, . . . , Wn are non-negative real numbers with sum one.

The aim of this paper is to find an inequality which contains (l3|), ([2]) and ([4])
as special cases. Since the inequality in ^ was the motivating step for obtaining
converses of Jensen's inequality using the so called Mond-Pecaric 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 ^ be a C* -algebra. We
say that a field {xt)t£T of operators in A is continuous if the function t xt i&
norm continuous on T. If in addition /i is a Radon measure on T and the function
t — > \\xt\\ is integrable, then we can form the Bochner integral Jj,xtdfi{t), which is
the unique element in A such that

J^Xtdfl{t)^ =

(p[ I xtdfi{t)] = I ip{xt)d^{t)

T

for every linear functional ip in the norm dual A*, cf. [SJ Section 4.1].

Assume furthermore that there is a field {(j)t)teT of positive linear mappings
(j)t : A ^ B from A to another C*-algebra B. We say that such a field is continuous
if the function t 4't{x) is continuous for every x £ A.li the C*-algebras are unital
and the field t — > 0t(l) is integrable with integral 1, we say that {<j)t)t£T is unital.

Theorem 1. Let f : I —> R be an operator convex function defined on an
interval I, and let A and B be unital C* -algebras. If (<f>t)t<=:T is a unital field of
positive linear mappings 4>t '. A ^ B defined on a locally compact Hausdorff space
T with a bounded Radon measure fi, then the inequality

(5) f\^J^Mxt)dtiit)j< J^Mfixt))dfi{t)

holds for every bounded continuous field {xt)t^T of self-adjoint elements in A with
spectra contained in I.

Proof. We first note that the function t 4>t{xt) G is continuous and
bounded, hence integrable with respect to the bounded Radon measure /x. 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

\\iyt)teT\\=snp\\yt\\ {yt)teT e CB{T,A),

t<£T

and it is not difficult to verify that the norm is already complete and satisfy the
C*-identity. In fact, this is a standard construction in C*-algebra theory. It follows
that f{{xt)teT) — {f{xt))teT- We then consider the mapping

it: CB{T,A)^B

defined by setting

IT {{xt)teT) = / 4't{xt) dfi{t) ,

JT

JENSEN'S OPERATOR INEQUALITY AND ITS CONVERSES

3

and note that it is a unital positive linear map. Setting x = {xt)t£T G CB{T,A)^
we use inequality ^ to obtain

/ (tt {{xtWr)) = fi^x)) < niJix)) = tt (/((xt)teT))

= ^((/(a;t))teT) '
which is the statement of the theorem.

To illustrate various techniques in proving Jensen's operator inequality, we give
two proofs of Theorem [T] in the discrete case T = {1, . . . , n}.

Proof. By using the continuity of / and uniform approximation of self-adjoint
operators by simple operators, we may assume that Xi has a spectral resolution on
the form

Xi — ^ ^ ^ li ■ ■ ■ ;

J 6 J.

where each Ji is a finite set and X^jej ^i-j ~ then have

1=1 jeji I \i=ijGJi

i=i jGJi i=i yeJi

n

The second proof: We use an the idea from [9], confer also [13^. If / is operator
convex in / = [0, 1) and /(O) < 0, then there is a connection a such that —f{t) =
ta {1 —t). We use the following two properties of a connection, cf. [11 [3].

(1) (j){aab) < (j){a) a (l){b)

for positive linear maps cf) and positive operators a and b.

(2) Er=i a« < (Eili «o (Er=i b^)

for positive n-tuples (ai, a„) and (&i, 6„). (subadditivity)
We then obtain

- E <^«(/(^»)) = E cr (1 - Xj))

i=l 1=1

n / n \ / n \

< '^(f>i{xi)a(l)i{l - Xi) < a E'^i(l ^^i)

i=l

Consider now an arbitrary operator convex function / defined on [0, 1). The func-
tion f{x) — f{x) — /(O) satisfies /(O) = hence

/ n \ n n

(6) / E'^'(^') -/(o)i^E'^«(/(^^))-/(o)E'^'(i)

4

FRANK HANSEN, JOSIP PECARIC, AND IVAN PERIC

from which the statement follows. We finally obtain ([5]) in this setting for a func-
tion / defined on an arbitrary interval [a, (3) by considering the function g{x) =
f {{/3 — a)x + a) on the interval [0, 1).

Inequality ^ is obviously a generalization of the inequalities ([2|) and (j4]).

3. Converses of Jensen's inequality

The following theorem should be compared with Theorem 2.3 in [lOj . For a function
/ : [m, M] ^ R we use the standard notation:

f{M) ~ fjm) Mf{m)-mf{M)
at — and Pf — -

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^^i^/^s^^, we may set F{x, y) — y~^x^^^y~'^ . Otherwise we may use the
functional calculus on tensor products, see for example [B].

Theorem 2. Let {xt)t£T be a hounded continuous field of self-adjoint elements in
a unital C* -algebra A with spectra in [m, M] defined on a locally compact Hausdorjf
space T equipped with a Radon measure /i, and let {(j)t)t£T be a unital field of
positive linear maps (pt : A B from A to another unital C* -algebra B. Let
f,g : [m, M] — )■ R and F : U x V R be functions such that f ([to, Af]) C U,
g {[m, M]) C V and F is bounded. If F is operator monotone in the first variable
and f is convex in the interval [to, Af], then

F

(7)

(l>t{f{xt))d^i{t),g[ I (j3t{xt)d^{t)

< sup F [afZ -h l3f, g{z)]l.

m<z<M

In the dual case (when f is concave) the opposite inequality holds in ^ with inf
instead of sup.

Proof. For convex / the inequality f{z) < afZ~\-/3f holds for every z G [m, M].
Thus, by using functional calculus, f{xt) < ctfXt + /3/1 for every t G T. Applying
the positive linear maps 0i and integrating, we obtain

(l>t{f{xt))dn{t) <af [ Pt{xt)diM{t)+Pfl.

T JT

Now, using operator monotonicity of F{-,v), we obtain
F 1^ cj)t ifixt)) dfi{t), g (^J^ Mxt)dfi{t)^

< F

"/ / 4't{xt)dfi{t) + Pfl,g [ / (j)t{xt)d^{t)

< sup F [ajz -\- (3f,g{z)]l.

m<z<M

Numerous applications of the previous theorem can be given (see |10]). We give
generalizations of some results from (14) .

Theorem 3. Let {At)t£T be a continuous field of positive operators on a Hilbert
space H defined on a locally compact Hausdorff space T equipped with a Radon mea-
sure II. We assume the spectra are in [to, Af ] for some < m < M. Let furthermore

JENSEN'S OPERATOR INEQUALITY AND ITS CONVERSES

5

{xt)teT be a continuous field of vectors in H such that ||a:t|p(i/i(<) = 1. Then for
any X > 0, p > I and q > 1 we have

(8) (^j {Alxt,xt)d^i{t)^ -\ j {AtXt,xt)dii{t)<C{\m,M,p,q),

where the constant

C{X,m,AI,p,q)

M ( Mi-i - aV < a < ^MP(-^'^)

(9) I \

q \ap J Qp'g — — ?

m(mf-i-A), ^TOP(i-i)<A

and Clip and (3p are the constants a/ and (3f associated with the function f{z) = .

Proof. Applying Theorem [2] for the functfons

f{z)^zP, F{u,v) ^u^/'^ - Xv,

and unital fields of positive linear maps (pt : B{H) — > C defined by setting (j>t{A) =
{Axt,Xt) for t G T, the problem is reduced to determine sup„<^<jvj (z) where
H{z) ^ {apz + Pp^/'J - Xz.

The following Corollary is a generalization of Theorem 5 in [H]. The r— geometric
mean A^rB introduced by F. Kubo and T. Ando in fTT] is defined by setting

Ai^rB = A^l^ ( A-^I^BA-^"'] A^

1/2

for positive invertible operators A and B.

Corollary 1. Let (^t)tgT md (Bt)t<^T be continuous fields of positive invertible
operators on a Hilbert space H defined on a locally compact HausdorfJ space T
equipped with a Radon measure /i such that

mil <At< Mil and m2l < Bt < M2I

for allt€T for some < mi < Mi and < m2 < M2. Then for any A > 0, s > 1,
p > 1 and any continuous field {xt)teT of vectors in H such that Jj, WxtW^dnit) = 1
we have

{A\xt,xt)dyi(t)\ U {B^xt,xt)diiit)
(10) -A / {B^#i/sAPxt,Xt)dnit)

T

where the constant C is defined in Theorem\3[ and 1/p + l/q = 1.

6

FRANK HANSEN, JOSIP PECARIC, AND IVAN PERIC

Proof. By using Theorem [3] we obtain for any A > 0, for any continuous field
{Ct)t£T of positive operators with ml < Ct < Ml and a square integrable contin-
uous field {yt)t^T of vectors in H the inequality

{CtyuVt)d^x{t) \ ( [ {yuyt)dii{i)

(11) -X {Ctyt,yt)dtiit)

JT

< C{X,m,M,s,p) ( {yt,yt)dii{t).

Set now Ct = [b^ '''^A^^B^ '^'^^ and yt = B^'^xt for t G T in HI]) and observe
that

^r<{Br'^AiB;^'^\"<^i.

By using the definition of the 1/s— geometric mean and rearranging (jlip we obtain

l/P / /• \ 1/9

X / {B^#,^,APxt,xt)dfi{t)
It

Blxt.xt) dfj.{t)

p/s , fp/s

9

which gives ([T

In the present context we may obtain results of the Li-Mathias type by using
Theorem [2] and the following result which is a simple consequence of Theorem [1]

Theorem 4. Let {xt)teT be a bounded continuous field of self-adjoint elements
in a unital C* -algebra A defined on a locally compact Hausdorff space T equipped
with a Radon measure fi. We assume the spectra are in [m,M]. Let furthermore
{4't)teT be a unital field of positive linear maps (j)t '■ A ^ B from A to another
unital C* -algebra B. Let f,g : [m, M] — > R and F : U x V ^ H be functions such
that f {[m, M]) C U, g {[m, M]) C V and F is bounded. If F is operator monotone
in the first variable and f is operator convex in the interval [m,M], then

(12)

(l>t{f{xt)) d^i{t),g[ / (j>t{xt) dfi{t)

> M F[f{z),giz)]l.

m<z<M

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 5. Let f be a convex function on [0, oo) and let || • || be a normalized
unitarily invariant norm on B{H) for some finite dimensional Hilbert space H. Let
{(j)t)teT be a unital field of positive linear maps (j)t : B{H) — > B{K), where K is a

JENSEN'S OPERATOR INEQUALITY AND ITS CONVERSES

7

Hilbert space, defined on a locally compact Hausdorff space T equipped with a Radon
measure fi. Then for every continuous field of positive operators {At)ti£T we have

(13)

MfiAt))d^^{t)<f [0)1

fi\\A\\)~f{0)
IIAII

MAt)d^lit).

Especially, for /(O) < 0, the inequality

(14)

is valid.

c^tU{At))dii{t) <

fiWM)

\\A,\\

(j)t{At)dii{t).

Proof. Since / is a convex function, f{x) < -^^^f{rn) + ^1™ /(-M) for every
X € [to, M] where to < M . Since || • || is normalized and unitarily invariant, we have
< At < \\At\\l and thus

\AM-Af.

r/(ll^tll)

il^tl! ' \\M

for every t G T. Applying positive linear maps and integrating we obtain

(15)

or

(16)

Uf{At))dfi{t)

</(0)

d^i(t)

fiWM)

\\AA\

(j)t{At)dii{t)

MfiAt))di,{t)<f [0)1

f{\\At\\)-f{0)

MAt)dti{t).

Note that since ^^d^i{t) < M|MjpOd^(<) = i we obtain, for /(O) < 0,
inequality HH) from (fT5|) .

Remark 1. Setting T — {1} the inequality gives

cbifiA)) < ^j^HA).

Furthermore, setting (f) — 1, we get the inequality f{\\A\\) > ||/(A)|| obtained in [S]
under the assumption that f is a nonnegative convex function with /(O) = 0.

Related inequalities may be obtained by using subdifferentials. If / : R — > R is a
convex function and [to, M] is a closed bounded real interval, then a subdifferential
function of / on [m, M] is any function k: [to, Af] — > R such that

k{x)G[fL{x),f'4x)] xeim,M),

where f'_ and are the one-sided derivatives of / and ^(to) = f+{m) and k{M) =
f'_{M). Since these functions are Borel measurable, we may use the Borel functional
calculus. Subdifferential function for concave functions is defined in analogous way.

Theorem 6. Let {xt)teT be a bounded continuous field of self-adjoint elements in
a unital C* -algebra A with spectra in [m, M] defined on a locally compact Hausdorff
space T equipped with a Radon measure /i, and let {(j)t)t£T be a unital field of positive
linear maps (j)t '■ A ^ B from A to another unital C* -algebra B. If f : R ~* H is a

8

FRANK HANSEN, JOSIP PECARIC, AND IVAN PERIC

convex function then

f{y)l + k{y) (^J^M^Mt)~yl

(17) < / Mfi^t))df^it)

JT

<f{x)l-x I MHxt)) dfiit) + I (j)t{k{xt)xt) dfi{t)

for every x,y ^ [to, Af], where k is a subdifferential function of f on [m,M]. In the
dual case (f is concave) the opposite inequality holds.

Proof. Since / is convex we have f{x) > fijj) + k{y){x — y) for every x,y €
[m,M]. By using the functional calculus it then follows that f{xt) > f{y)l +
k{y){xt — yl) for t gT. Applying the positive linear maps (pt and integrating, LHS
of (fT7)l follows. The RHS of PT| follows similarly by using the functional calculus
in the variable y.

Numerous inequalities can be obtained from ^T7\ . For example, LHS of p7|) may
be used to obtain an estimation from below in the sense of Theorem [5] Namely,
the following theorem holds.

Theorem 7. Let {xt)ti£T be a bounded continuous field of self-adjoint elements in
a unital C* -algebra A with spectra in [m, M] defined on a locally compact Hausdorff
space T equipped with a Radon measure /x, and let {4>t)t!^T be a unital field of positive
linear maps 4>t'. A ^ B from A to another unital C* -algebra B. Let / : R — > R,
g : [to, Af] R and F : U x V ^ Tl be functions such that /([m, A/]) C U,
(/([to, Af]) (1 V, F is bounded, f is convex and f(jj) + k{y){t — y) U for every
y,t G [to, Af], where k is a subdifferential function of f on [to,, Af]. If F is operator
monotone in the first variable, then

F

(18)

<Pt{f{xt)) d^.{t),g[ I (j}t{xt) d^.{t)

T \JT

> M F[fiy) + k{y)iz-y),giz)]l

rn<z<M

for every y £ [to, Af]. In the dual case (when f is concave) the opposite inequality
holds in i f _?<$)) with sup instead o/inf.

Using LHS of (jl7p we can give generalizations of some dual results from [5].

Theorem 8. Let {xt)t£T be a bounded continuous field of positive elements in a
unital C* -algebra A defined on a locally compact Hausdorff space T equipped with a
Radon measure /i, and let {(j)t)teT be a unital field of positive linear maps (j)t: A ^ B
from A to another unital C* -algebra B acting on a finite dimensional Hilbert space
K. Let II • II be a unitarily invariant norm on B{K) and let f: [0, cxd) R be an
increasing function.

(1) // ||1|| = 1 and f is convex with /(O) < then

(19) f[\\ j^ctyt{xt)dpi{t)\\ j < II J^Mfi^t))d^i{t)\\.

(2) // Jj, (j)t{xt) dfi{t) < \\ Jrp (l)t{xt) dfj,{t)\\l and f is concave then

(20) ^ Mfi^t)) dfi{t) <f{\\J^ M^t) dijL{t)\

JENSEN'S OPERATOR INEQUALITY AND ITS CONVERSES

9

Proof. Since /(O) < and / is increasing we have k{y)y — f{y) > and
^(y) ^ 0. From ()17p and the triangle inequahty we have

k{y)\\ I 4>t{xt)di,{t)\\<\\ I ct^t{f{xt))\\ + {k{y)y-f{y)).

Now (|19p follows by setting y = || ^j,(t>t{xt) d^{t)\\. Inequality ([201) follows imme-
diately from the assumptions and from the dual case of LHS in ITT)) by setting
y = II /T'/'t(a;t)rfAi(i)ll-

Finally, to illustrate how RHS of (fT7)) works, we set

_ II Jrp (t)tik{xt)xt) d^i{t)\\
\\J^MH^t))dfim
and obtain a Slater type inequality

II Jj, (j)tik{xt)xt) d^{t)\\

IT \ \\UMHxt))d^i{t)\\

under the condition

Jj, (j)t{k{xt)xt) d^i{t) ^ ^^(j)t{k{xt))dii{t)

\\^^4)t{k{xt)xt)d^i{t)\\ - \\j^(j)t{k{xt))d^i{t)\\ ■
References

1. J. S. Aujla, H. I. Vasudeva, Operator inequalities related to means of operators, Math. Japon.
41 (1995) 383-388.

2. R. Bhatia, Matrix Analysis, Springer, Germany, 1996.

3. M. D. Choi, A Schwarz inequality for positive linear maps on C* -algebras, lUinois J. Math.
18 (1974) 565-574.

4. Ch. Davis, A Schwarz inequality for convex operator functions, Proc. Amer. Math. Soc. 8
(1957) 42-44.

5. R. Drnovsek, T. Kosem, Inequalities between /(||^||) and ||/(|^|)|i, Math. Inequal. and Appl.
8 (2005) 1-6.

6. F. Hansen. Operator monotone functions of several variables. Math. Ineq. Appl., 6:1-17, 2003.

7. F. Hansen, G. K. Pedersen, Jensen's inequality for operators and Lowner's theorem. Math.
Ann. 258 (1982) 229-241.

8. F. Hansen, G. K. Pedersen, Jensen's operator inequality. Bull. London Math. Soc. 35 (2003)
553-564.

9. M. Fujii, F. Kubo, Around Jensen's inequality II, Math. Japon. 27 (1982) 495-499.

10. T. Furuta, J. Micic-Hot, J. Pecaric, Y. Seo, Mond-Pecaric Method in Operator Inequalities,
Element, Zagreb, 2005.

11. F. Kubo, T. Ando, Means of positive linear operators. Math. Ann. 246 (1980) 205-224.

12. B. Mond, J. Pecaric, Converses of Jensen's inequality for several operators. Revue d'Analyse
Numer. et de Th eorie de I'Approxim. 23 (1994) 179-183.

13. B. Mond, J. Pecaric, On Jensen's inequality for operator convex functions, Houston Journal
of Mathematics 21 (1995) 739-754.

14. M. Tominaga, The upper bound of a reverse Holder's type operator inequality and its appli-
cations, J. Inequal. & Appl. 7 (2002) 633-645.

Department of Economics, Copenhagen University, Studiestraede 6, 1455 Copen-
hagen K, DENMARK
E-mail: frank.hansen@econ.ku.dk

Faculty of Textile Technology, University of Zagreb, Pierottijeva 6, 10000 Zagreb,
CROATIA

E-mail: pecaric@hazu.hr

Faculty of Chemical Engineering and Technology, University of Zagreb, Marulicev
TRG 19, 10000, Zagreb, CROATIA
E-mail: iperic@pbf.hr