You are on page 1of 9

SIAM J. ALG. DISC. METH.

(C) 1987 Society for Industrial and Applied Mathematics


Vol. 8, No. 4, October 1987 014

THE CONTRAHARMONIC MEAN OF HSD MATRICES*


Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

WILLIAM N. ANDERSON, JR/f, MICHAEL E. MAYS, THOMAS D. MORLEY


AND GEORGE E. TRAPP

Abstract. For positive scalars a and b the contraharmonic mean of a and b, C(a, b), is defined by
C(a, b) (a + b2)/(a + b).
In this paper we consider a natural matrix generalization of the contraharmonic mean, fit this into the matrix
analogue of some of the classical scalar inequalities for means, develop computational procedures which let us
generate the matrix analogues of an infinite family of scalar means, and study fixed point problems. Finally,
we mention a relationship between least squares problems and the contraharmonic mean.

Key words, harmonic, arithmetic, contraharmonic

AMS(MOS) subject classifications. 15A24, 15A45


1. Introduction. IfA and B are Hermitian positive semidefinite matrices we define
C(A, B) A + B 2(A :B),
where A B is the operation of parallel addition introduced by Anderson and Duffln ].
Many of the properties of C(A, B) are related to those of the harmonic mean 2(A B)
and the arithmetic mean (A + B)/2. For example,
A + B >= C(A, B) >= (A + B)/2
and

or as
The dual mean C’(A, B)
-
(C(A,B)+ 2(A B))/2 =(A + B)/2.
C(A -1, B-) can be written as
C’(A, B) (A B) + 2(A B)C(A, B)-I(A B),

C’(A, B) [A(A B)-IAI [B(A "9)-19].


We will also present variational characterizations of both the contraharmonic mean
and its dual. These variational characterizations arise from the representation of the
contraharmonic mean using the operation of parallel subtraction.
The equation
C(A,X) A + B
is equivalent to the fixed point problem
X 2(A X) + B.
If B is positive definite this latter problem has a unique solution according to the work
of Anderson, Kleindorfer, Kleindorfer and Woodroofe [2], and in fact is the same as the
solution to
C(Y,B)=A+B.

Received by the editors December 29, 1986; accepted for publication (in revised form) May 8, 1987.

" Department of Mathematics and Computer Science, Fairleigh Dickinson University, Teaneck, New
Jersey 07666.
Department of Mathematics, West Virginia University, Morgantown, West Virginia 26506.
School of Mathematics, Georgia Institute of Technology, Atlanta, Georgia 30332.
Department of Statistics and Computer Science, West Virginia University, Morgantown, West Virginia
26506.
674
THE CONTRAHARMONIC MEAN OF HSD MATRICES 675

In the scalar case the solution would be


x {(a + b) + (a 2 + 6ab + b2)1/2}/2,
Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

but in an HSD setting the fractional power needs to be interpreted using the geometric
mean operation #, defined by Pusz and Woronowicz [14] as
A # B AI/2(A-I/2BA-1/2)I/2A 1/2.
Another interesting problem is that of inverse means: given HSD matrices E
and F and means Ml and M2, when can we find HSD matrices A and B satisfying E
M(A, B) and F M2(A, B)? We consider this problem when one of the specified means
is the contraharmonic mean.
In the final section we exhibit a relationship between some classical least squares
problems and the contraharmonic mean. We show that there are two natural ways to
define the contraharmonic mean of three HSD matrices.

2. Preliminaries. The contraharmonic mean is one of the classical means of Greek


mathematics. Its name arises from the fact that, just as solving for x in the equation
(a- x)/(x- b)= a/b
yields the harmonic mean, so solving for x in the equation
(a- x)/(x- b) b/a
yields
x C(a, b) (a 2 + b2)/(a + b).
Many investigators have linked means together by introducing one or more param-
eters into the definition which, when varied, generate some of the "named" means. Thus
the harmonic mean is the case s 0 and the contraharmonic mean is the case s 2 for
the mean studied by Gini [8], Beckenbach [6] and Lehmer [9]"
Gs(a, b) (a + bS)/(a + b 1).
The arithmetic mean and geometric mean can also be associated with this mean, by
letting s and s 1/2. Other parametrizations are discussed in Mays 11 ].
To generalize from the scalar case, we consider matrices on a finite-dimensional
inner product space. The inner product is denoted by (,). A matrix is called Hermitian
positive semidefinite (HSD) if A A*, where A* is the adjoint (conjugate transpose) of
x
A, and (Ax, >= 0 for all vectors x. If A and B are both HSD, we write A >= B if A B
is HSD.
To generalize the contraharmonic mean to the case of HSD matrices, we require
three special HSD matrix operations: the parallel sum (harmonic mean), the geometric
mean and the parallel difference. If A and B are invertible HSD matrices, the parallel
sum, denoted A B, and the geometric mean, denoted A # B, are defined by
A" B A(A + B)-IB
and
A # B A 1/2(A-1/2BA-I/2)l/2A 1/2.
In the case that the inverses do not exist, the limit of nonsingular approximations may
be used as the definition.
Another operation that we will use is a subtraction operation related to A B. We
seek HSD X so that A X C for given A and C. We refer to [3] and 13] for additional
background; here we only require the following result from [13].
676 ANDERSON, MAYS, MORLEY AND TRAPP

If there is an X so that A X C, then one such X is given by the variational formula

min[XIXisHSDand[A-CA A+xA]isHSD}"
Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

The minimum X is called the parallel difference of C and A and written C + A. The
minimum X is the only solution of A X C such that the range of X is contained in
the range ofA. We use the parallel difference to characterize the dual of the contraharmonic
mean. This X may also be obtained from the formula
X= A(A C)+C
where the superscript "+" denotes the Moore-Penrose generalized inverse.
A fundamental theorem used in our investigation is the arithmetic-geometric-
harmonic inequality
(A + B)/2 >-A # B >= 2(A B),
which is valid for all HSD matrices A and B. Trapp [15] has more information and
background on these operations and inequalities.
3. The contraharmonie mean and its dual. IfA and B are HSD matrices, we define
the contraharmonic mean, denoted C(A, B), by
(1) C(A, B) A + B 2(A" B).
This definition is motivated by an identity in the scalar case. Since both + and are
commutative, this operation is commutative. When Q- exists we can write
Q(C(A, B))Q* C(QA Q*, QBQ* ),
a matrix homogeneity property stronger than the standard requirement for homogeneity,
in which Q would have to be a scalar.
Clearly A + B >-_ C(A, B), and since (A + B)/2 >= 2(A B) we see that
C(A, B) >= (A + B)/2.
This guarantees that C(A, B) is HSD.
A direct computation shows that
(C(A, B) + 2(A B))/2 (A + B)/2,
i.e., the arithmetic mean of the contraharmonic and harmonic means is the arithmetic
mean. When A + B is invertible, another direct computation using the equivalent parallel
formula
A B A -A(A + B)-A
yields
(A + B)/2 2(A" B) (A B)(A + B)-(A B)/2;
hence
C(A, B) (A + B)/2 + (A B)(A + B)-(A B)/2.
As a corollary, note that
(A + B)/2 >= (A B)(A + B)-(A B)/2.
because
A + B (A + B)/2 + (A + B)/2 >= C(A, B).
THE CONTRAHARMONIC MEAN OF HSD MATRICES 677

Duality is a natural matrix mean concept. The dual of the contraharmonic mean,
which we will denote C’(A, B), is defined by
Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

C’(A,B)=C(A-1,B-I) -1
when A and B are invertible.
THEOREM 1. C’(A, B) A" B + 2(A B)C(A, B)-I(A B).
Proof. Let D A-1/2BA -1/2. Then
C(A, B) C(A l/2IA I/2,A I/2DA 1/)
A l/2C(I, D)A 1/2
and
C(A -l, B-l) -l A l/2C(I, D-l)-lA
Since I and D commute, as in the scalar case we have

Thus
C(A-,B-I)
-
C(I,D-I) -l I:D+ 2(I:D)C(1,D)-I(I:D).
A1/2{I" D + 2(I" D)C(I,D)-I(I D)}A 1/2
A /2(I" D)A 1/2
+2A 1/2(I D)A 1/2A-1/2C(I, D)-A-I/2A 1/2(i. D)A
A" B + 2(A" B)(A /2C(I, D)A l/2)-l(A "B)
A" B + 2(A" B)(C(A, B))-I(A "B),
as desired.
Another representation of the dual is given by the following theorem:
THEOREM 2. C’(A, B) (A(A B)-IA) (B(A B)-IB).
Proof
C’(A, B) (A -1 + B-l 2(A-l B-l)) -l
(A -1 -A-I. B-I + B-1 A-I. B-1)-l.
Now let
X -l A -I (A -l" B-l) A -l (A + B) -I
and
y-I B-I (A-I B-l) B-I -(A + B) -l.
The proof is complete upon noting that C’(A, B) X: Y, and that
AX-IA A A(A + B)-lA A B
implies X A(A B)-IA, with a similar result for Y.
Now we begin with the original formula for the contraharmonic mean and take the
duals of each side to generate the following sequence of equations. Note we are using the
fact that the arithmetic and harmonic means are duals.
C(A, B) A + B 2(A :B)
or
2(A- B) + C(A, B) A + B.
Taking duals yields
((A + B)/2) C’(A, B) A B,
and thus C’(A, B) exists for all invertible A and B. Since the range of (A + B)/2 contains
678 ANDERSON, MAYS, MORLEY AND TRAPP

the range ofA B, according to a result of [3], the equation ((A + B)/2) :X A B has
a solution, and hence a minimum solution. This minimum solution should be defined
-,
as C’(A, B) in the most general situation; it is in accordance with C(A B-1) for invertible
Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

A and B. Since the range of the contraharmonic mean is contained in the range of the
arithmetic mean, the dual contraharmonic mean may be written
C’(A, B) (A :B) + (A + B)/2.
Alternatively, we can use the variational characterization of parallel subtraction to write
(A + B)/2 A
{
C’(A, B) inf X X is HSD and
[ (A+B)/2
(A + B)/2 + X
+ B)/2
is
] HSD}.
We multiply the composite matrix by 2 (which does not change its HSD character)
to obtain the following theorem.
THEOREM 3.

C’(A’B)=inf{XIXisHSDand[ C(A’B)A+B A+B+2xA+B]isHSD}.


Green and Morley [7] have developed other variational representations of the con-
traharmonic and the dual contraharmonic means. We list their results here for com-
pleteness and comparison. The proofs are in [7].
THEOREM 4. (a)
(C(A,B)z,z) sup
x+y=z
{((A B)x,x) + ((B-A)y, y) + 2((A + B)x, y)}.
(b)

(c)
C(A,B) inf
{X>= 0
(A+B) X-(B-A)
4. Fixed point problems. We now consider two fixed point problems. Given the
HSD matrices A and B, find X and Y so that
(2) C(A,X)=A+B,
(3) C(Y,B)=A+B.
The two fixed point problems may be rewritten as
(2’) X= 2(A :X) + B,
(3’) Y= 2(B: Y) + A.
Equations (2’) and (3’) are special cases of a fixed point problem studied by Anderson,
Kleindorfer, Kleindorfer and Woodroofe [2]. Their form was
(4) Z Q(M: Z)Q* + N,
and they showed that (4) has a unique HSD solution when N is invertible. Assuming A
and B are invertible, we have that (2’) and (3’) have unique solutions. We now show that
the solutions must be equal. Multiply (2’) by A -/ on each side. Then for
X’ A-/zXA -/2
THE CONTRAHARMONIC MEAN OF HSD MATRICES 679

and
B’ A-/ZBA-I/2
we have
Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

(5) X’= 2(I" X’) + B’.


Here we use again the fact that
Q(A’B)Q* =QAQ*’QBQ*.
With the problem reduced to (5), we can use the fact that I and B’ commute to
treat (5) as a scalar problem and write
S’= (I+ B’)/2 + SQRT{(B’)2 + 6B’+ I}.
Therefore
X AI/2X’A 1/2 (A + B)/2 + AI/2SQRT{(B’)2 + 6B’ + I}A 1/2.
The term inside SQRT may be factored as
(aB’ + bI)(B’/a + 1/b)
for any scalars a and b such that a/b + b/a 6. Since we wish to keep all matrices HSD
we will also require that a > 0 and b > 0. The SQRT term may then be written as
(aB’ + bI) # (B’/a + 1/b).
Then using the fact that
O(A # B)Q* QAQ* # QBQ*
we can write
A /2(aB’ + bI) # (B’]a + I]b)A 1/2 (aB + bA) # (B/a + A/b).
Therefore
X (A + B)/2 + (aB + bA) # (B/a + A/b).
A similar analysis is possible for (3’), and we see that Y has an analogous form. The
only question that remains is whether the expression
(aB + bA) # (B/a + A/b)
depends on the choice of a and b. That the expression does not depend on this choice
follows immediately from the results in [2], since the solution is unique.
We close this section by noting some properties of the fixed point viewed as a
function of A and B. Let
T(A,B) (A + B)/2 + {(A + bB)#(A + B/b)}/2,
where
b+ 1/b= k>=2.
Then
T(A,B)>=A+B
because the work of Ando [5] gives that
{(A + bB # (A + B/b) } /2 >- (A # A + bB # B/b)/2 (A + B )/2.
We also have that

and if

then
A
- T(O,B)=B,

A2 and B >- B2,


T(A , B l) > T(A2, B2).
680 ANDERSON, MAYS, MORLEY AND TRAPP

5. Loewner’s theorem and monotonicity. Given a binary operation & on HSD ma-
trices, define a function F(z) by
F(z)I zI & L
Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Such a function F(z) is called a "Pick function." Loewner 10] shows that under certain
restrictions on & the operation I & B is monotone if and only if F(z) is analytic in the
upper half plane and satisfies Im (F(z)) > 0 for Im (z) > 0.
For example, parallel addition is monotone. In this case
F(z)I zI I (z](1 + z))I.
We have for z a + bi that
Im (F(z))= b/((1 + a)2 + b2),
and the result holds.
For the contraharmonic mean C(A, B),
F(z) + z2)/(1 + z).
Since this function has a zero at z i, it will have a sign change and we can conclude
that C(A, B) is not monotone.
6. Related means of HSD matrices. We have seen that several scalar means in the
family
Gs(a, b) (a + b)/(a + b )
-
have interpretations as means of HSD matrices. A set of scalar means arising from a
graphical representation of Moskovitz 12],
M(a, b) (ab + ba)/(a + bS),
may be interpreted in this way as well.
Three algebraic identities connecting these means provide a recurrence relation that
easily extends to the HSD matrix case while avoiding problems of commutativity, so
that we have a family of HSD matrix means
Gs(A,B)
defined for s, an arbitrary integer. These identities are
(6) Gs(A,B)=Ms(A-,B-) -,
(7) Ms(A,B)+ Gs+ (A,B)=A + B
and
(8) Ms(A,B)=G,-(A,B).

Thus we have means and their duals arising not only for s 2 (the contraharmonic
mean) but for s 3, 4, as well. In none of these cases, however, is monotonicity
preserved.
To see why G(A, B) is HSD ifA and B are HSD, note that for both symmetry and
positivity we may use induction. First,
A + B >= Gz(A,B) > (A + B)/2,
and if
A + B >= Gs(A,B)>=(A + B)/2
for all HSD A and B then
A
--
+B >__ G(A -1, B -1) >__ (A -1 + B-I)/2,
THE CONTRAHARMONIC MEAN OF HSD MATRICES 681
SO
-,
- - -
-(A -1 + B-l) -1 >__ -Gs(A B-) __> -((A + B- 1)/2)
Now add A + B to each part, and note that the left-hand part is bounded above by
Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

A + B, the fight-hand part is C(A, B), which is bounded below by (A + B)/2, and the
middle is Gs / I(A, B), by (6) and (7).

7. Inverse mean problems. In this section we are interested in the follow-


ing type of problem: Given E and F HSD, when can we find HSD A and B so that
E (A + B)/2 and F 2(A B)9. This question and similar questions involving the geo-
metric mean are answered in [4]; here we wish to consider questions of the same form
involving the contraharmonic mean. For example, let E and F be HSD. When can we
find HSD A and B so that E A + B and F C(A, B)9.
We know that such a representation can exist only when E and F satisfy 2F >=
E >= F. In fact this inequality is also sufficient for such a representation to exist. This
result is presented in the next theorem.
THEOREM 5. Given HSD E and F with 2F >= E >= F, let A and B be defined by
A (E + E # (2F- E))/2,
B=(E-E#(2F-E))/2.
Then A and B are HSD, A + B E, and C(A, B) F.
Proof It is obvious that A is HSD and A+B=E. Since E>-F, we have
E >= 2F- E and hence
E=E#E>=E#(2F-E)
and B is HSD. To complete the proof we need that C(A, B) F.
Let X E # (2F- E). Recall from [3] that 2F- E is then equal to XE-X. Write
C(A,B) A + B- Z(A B)
E- 2((E + X)/2"(E-X)/2)
E- (E + X)(2E)-’(E- X)
E-(E-XE-X)/2
=E-(E-(2F-E))/2
=F.
Similar techniques lead to the following.
THEOREM 6. Given HSD E and F with 2F >- E >= F, the following A and B are
HSD with E C(A, B) and F (A + B )/2:
A=F+F#(E-F),
B--F-F#(E-F).
We leave unresolved the question of finding A and B such that E C(A, B) and
F A # B. Even the scalar version of this problem is difficult to solve because a quartic

-
equation arises.

8. The contraharmonic mean and least squares problems. If we wish to solve the
Ax d
system B= as a least squares problem, x (A 2 -l- B2)-I(A B)d, If instead we use
x C(A, B)-d_, these agree when there is an x which simultaneously satisfies the equations
Ax d and Bx d. The two forms also agree if AB BA.
682 ANDERSON, MAYS, MORLEY AND TRAPP

A natural extension of the contraharmonic mean to three variables would seem to


be via a least squares problem for a system with three components:
Downloaded 01/04/13 to 152.3.102.242. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Ax=d,
A2_x =_d,
A3_x= d.
The least squares solution of this system is
x (A2 + A2 + A)-’(A, + A2 + A)d.
In the scalar case,
a2 + a + a] a + a2 + a3- 2 ala2 + aa3 + a2a3
al + a2 + a3 a + a2 + a3
a + a2 + a3- 2M.
There is more than one analogous identity in the matrix case because there is more than
one matrix generalization of M. It is noted in 15] that the following are not equivalent

-
in general:
M ((AI" (A2 +A3))+(A2"(A +A3))+(A3"(A1 +A2)))/2
and
M2 2((A (A2" A 3))" (A2 -Ji- (A 1" A 3))" (A q" (A 1" A2))).

Therefore there are at least these two candidates for the contraharmonic mean of three
HSD matrices:
C(A,A2,A3)=A +A2 +A3 2Ml
and
C2(A,A2,A3)=A +A2 +A3- 2M2.

REFERENCES
W. N. ANDERSON, JR. AND R. J. DUFFIN, Series and parallel addition of matrices, J. Math. Anal. Appl.,
26 (1969), pp. 576-594.
[2] W. N. ANDERSON, JR., G. D. KLEINDORFER, P. R. KLEINDORFER AND M. B. WOODROOFE, Consistent
estimates of the parameters of a linear system, Ann. Math. Statist., 40 (1969), pp. 2064-2075.
[3] W. N. ANDERSON, JR., T. D. MORLEY AND G. E. TRAPP, A characterization ofparallel subtraction, Proc.
Natl. Acad. Sci. U.S.A., 76 (1979), pp. 3599-3601.
[4] W. N. ANDERSON, JR. AND G. E. TRAPP, Inverse problems for means of matrices, this Journal, 7 (1986),
pp. 188-192.
[5] T. ANDO, Topics on operator inequalities, Lecture notes, Sapporo, Japan, 1978.
[6] E. F. BECKENBACH, A class of mean value functions, Amer. Math. Monthly, 57 (1950), pp. 1-6.
[7] W. L. GREEN AND T. D. MORLEY, Four variationalformulations ofthe contraharmonic mean ofoperators,
this Journal, 8 (1987), pp. 670-673.
[8] C. GINI, Di unaformula comprensiva delle medie, Metron, 13 (1938), pp. 3-22.
[9] D. H. LEHMER, On the compounding of certain means, J. Math. Anal. Appl., 36 (1971), pp. 183-200.
10] K. LOEWNER, Uber monotone Matrixfunktionen, Math. Z., 38 (1933), pp. 177-216.
11] M. E. MAYS, Functions which parametrize means, Amer. Math. Monthly, 90 (1983), pp. 677-683.
12] D. MOSKOVITZ, An alignment chart for various means, Amer. Math. Monthly, 40 (1933), pp. 592-596.
13] E.L. PEKAREV AND JU. t. SMULJAN, Parallel addition and subtraction ofoperators, Math. USSR Izvestija,
10 (1976), pp. 351-370.
14] W. Pusz AND S. L. WORONOWICZ, Functional calculus for sesquilinear forms and the purification map,
Rep. Math. Phys., 8 (1975), pp. 159-170.
15] G.E. TRAPP, Hermitian semidefinite matrix means and inequalities--an introduction, Linear and Multilinear
Algebra, 16 (1984), pp. 113-123.

You might also like