Mathematical Communications 2(1997), 35-39 35

The logarithmic mean is a mean

B. Mond∗, Charles E. M. Pearce†and J. Pečarić‡

Abstract. The fact that the logarithmic mean of two positive num-
bers is a mean, that is, that it lies between those two numbers, is shown
to have a number of consequences.

Key words: logarithmic mean, potential means, operator entropy

Sažetak.Logaritamska sredina je sredina. Činjenica da je log-

aritamska sredina dva pozitivna broja sredina, tj. da leži izmed̄u ta dva
broja ima niz primjena kako za potencijalne i generalizirane logaritamske
sredine tako i u teoriji operatora.

Ključne riječi: logaritamska sredina, potencijalna sredina, opera-

torska entropija

AMS subject classifications: 26D15, 47A63

Received February 25, 1997, Accepted May 27, 1997

1. Introduction
The logarithmic mean of positive numbers x, y is given by
x−y
L(x, y) =
ln x − ln y
if x 6= y and by L(x, x) = x if x = y. It is a mean, that is, we have

min(x, y) ≤ L(x, y) ≤ max(x, y). (1)

Here we shall show that this fact alone leads to a variety of interesting results.
Take x 6= 1. Setting y = 1 in (1) provides
x−1
min(x, 1) ≤ ≤ max(x, 1),
ln x
∗ School of Mathematics, La Trobe University, Bundoora, Vic. 3083, Australia, e-mail:
[email protected]
† Department of Applied Mathematics, University of Adelaide, Adelaide SA 5005, Australia,

e-mail: [email protected]
‡ Faculty of Textile Technology, University of Zagreb, Pierottijeva 6, HR-10 000 Zagreb, Croatia,

e-mail: [email protected]
36 B. Mond, C. E. M. Pearce, and J. Pečarić

so that ln x ≤ x − 1 ≤ x ln x for x > 1. Similarly, the same relation holds for

0 < x < 1. In fact, equality occurs only for x = 1. So we have
ln x ≤ x − 1 ≤ x ln x for x > 0. (2)
Schaumberger [7] has shown that (2) provides a strikingly direct route to several
results involving power means. We may use similar development to obtain results
for integral means but with nonuniform rather than uniform weights.
In Section 2 we derive a key inequality for potential means in this way. A discrete
version is also given. In Section 3 we note the existence of analogous results for
other means which possess integral representations. Finally, in Section 4, we observe
applications to the relative operator entropy of two positive operators on a Hilbert
space.

2. Potential means
Let f, w : [a, b] → R be positive, integrable functions. The potential mean of order
r of a function f with weight function w is given by
Z 1/r
b

 

 w(t)f (t) dt 
r 

a
Mr (f, w) = Z b , r 6= 0,

 


 w(t)dt  
a
Z 
b

 

 w(t) ln f (t)dt 


a
M0 (f, w) = exp Z b
, r = 0.

 


 w(t)dt 

a
For convenience we write f for f (t), w for w(t), M for Mr (f, w) (r 6= 0) and M0
for M0 (f, w). Set x = f r /M r in (2). We get after multiplication by wM r that
f f
rwM r ln ≤ wf r − wM r ≤ rwf r ln .
M M
Integration gives
Z b Z b
f f
rM r w ln dt ≤ 0 ≤ r wf r ln dt,
a M a M
that is, for r > 0,
Z b Z b Z b Z b
Mr w ln f dt − M r ln M w dt ≤ 0 ≤ wf r ln f dt − ln M wf r dt.
a a a a
This is equivalent to
Z b
wf r ln f dt
r r
M ln M0 − M ln M ≤ 0 ≤ a
Z b − M r ln M
w dt
a
 The logarithmic mean is a mean 37

or Z b
wf r ln f dt
M r ln M0 ≤ M r ln M ≤ 0
Z b
· Mr
r
wf dt
a
Therefore, we have
M0 ≤ M ≤ M0 (f, wf r ),
that is, in extenso,

M0 (f, w) ≤ Mr (f, w) ≤ M0 (f, wf r ) for r > 0. (3)

If r < 0, the inequalities are reversed.

Assuming all integrals exist, a consequence of (3) is that

lim Mr (f, w) = M0 (f, w).

r→0

A similar development is available in the discrete case. If a, w are positive n–

tuples, a potential mean of order r with weights w is given by
( n
)1/r
[r] 1 X r
Mn (a, w) = w i ai , r 6= 0,
Wn i=1
( n
)1/Wn
Y
Mn[0] = aw
i
i
,
i=1
n
X
where Wn = wi . With the notation war = (w1 ar1 , w2 ar2 , . . . , wn arn ) we can
i=1
obtain
Mn[0] (a, w) ≤ Mn[r] (a, w) ≤ Mn[0] (a, war ),
which is a generalization of a result from [6] for the case of an unweighted mean.
We have again that
lim Mn[r] (a, w) = Mn[0] (a, w).
r→0

3. Analogues
We now consider the generalized logarithmic mean of order r of positive numbers
x, y. For x 6= y this is defined by
½ r ¾1/(r−1)
x − yr
Lr (x, y) = , r 6= 0, 1,
r(x − y)

L0 (x, y) = L(x, y),

1 x y 1/(x−y)
L1 (x, y) = I(x, y) = (x /y )
e
38 B. Mond, C. E. M. Pearce, and J. Pečarić

and for x = y by Lr (x, x) = x.

These means have an integral representation

Lr (x, y) = Mr−1 (e1 , e0 ),

where e1 (t) = t, e0 (t) = 1 for all t ∈ [min(x, y), max(x, y)]. So, a consequence of (3)
for r > 1 is that
I(x, y) ≤ Lr (x, y) ≤ I(xr , y r )1/r . (4)
Reverse inequalities apply for r < 1 (r 6= 0), while for r = 0 we have
√
G(x, y) ≡ xy ≤ L(x, y) ≤ I(x, y).

We have immediately from (4) that

lim Lr (x, y) = I(x, y).

r→1

In place of the integral potential mean of the function f : [a, b] → R we could

consider more general potential means of functions g : Ω → R, where Ω is an
arbitrary set. For such means, (3) follows in the same way. As special cases we can
again consider, for example, the logarithmic means on n variables (see Pittenger
[5]), the hypergeometric mean (see Brenner and Carlson [1]) and other means which
have representations in the form of integral means (see [1]).

4. Operator theory
A further application lies in the theory of operators. Fujii and Kamei [2] introduced
the notion of the relative operator entropy S(A|B) for positive operators A, B on a
Hilbert space. For A and B invertible, this is given by
³ ´
S(A|B) = A1/2 ln A−1/2 BA−1/2 A1/2 .

Fujii and Kamei [3] proved that

A − AB −1 A ≤ S(A|B) ≤ B − A. (5)

We can derive this (see e.g. [6] pp. 269-273) as a consequence of (2), which may be
re–expressed as
1 − 1/x ≤ ln x ≤ x − 1 for all x > 0. (6)
Since A−1/2 BA−1/2 is a positive operator, we may substitute it for x to obtain

I − A1/2 B −1 A1/2 ≤ ln A−1/2 BA−1/2 ≤ A−1/2 BA−1/2 − I.

Pre– and postmultiplication by the positive operator A1/2 now provide (5).
Recently Furuta [4] proved a generalization of this inequality, namely, if A and
B are positive invertible operators, then for any positive number x0 we have

(ln x0 − 1)A + Bx−1

0 ≥ S(A|B) ≥ (1 − log x0 )A − AB −1 Ax−1
0 . (7)
 The logarithmic mean is a mean 39

We show that this follows from (6). For on substituting x/x0 for x in (6) we
obtain

ln x0 + 1 − (x0 /x) ≤ ln x ≤ ln x0 + (x/x0 ) − 1 for all positive x, x0 .

(This inequality is equivalent to (1) for y = x0 ). Again we replace x by the positive

operator A−1/2 BA−1/2 to derive

(ln x0 +1)I −x0 A1/2 B −1 A1/2 ≤ ln A−1/2 BA−1/2 ≤ (ln x0 −1)I +A−1/2 BA−1/2 x−1
0 .

Pre– and postmultiplication by A1/2 as before yields

(ln x0 + 1)A − x0 AB −1 A ≤ S(A|B) ≤ (ln x0 − 1)A + Bx−1

0 . (8)

We can replace x0 by x−1

0 to give

(1 − ln x0 )A − x−1
0 AB
−1
A ≤ S(A|B) ≤ Bx−1
0 − (ln x0 + 1)A. (9)

The first inequality in (9) taken with the second in (8) provides (7).

References
[1] J. L. Brenner, B. C. Carlson, Homogeneous mean values: weights and as-
ymptotics, J. Math. Anal. Appl. 123(1987), 265–280.
[2] J. I. Fujii, E. Kamei, Relative operator entropy in noncommutative informa-
tion theory, Math. Japon. 34(1989), 341–348.
[3] J. I. Fujii, E. Kamei, Uhlmann’s interpolational method for operator means,
Math. Japon. 34(1989), 541–547.
[4] T. Furuta, Furuta’s inequality and its application to the relative operator en-
tropy, J. Operator Theory 30(1993), 21–30.
[5] A. O. Pittenger, The logarithmic mean in n variables, Amer. Math. Monthly
92(1985), 99–104.
[6] F. Riesz, B. Sz-Nagy, Functional analysis, Translated from the 2nd French
edition by Leo F. Boron, Frederick Ungar Pub. Co., 1955.
[7] N. Schaumberger, Extending a familiar inequality, Pi Mu Epsilon Journal
10(1992), 384–385.