Available online at www.sciencedirect.

com

Journal of Functional Analysis 262 (2012) 1498–1514

www.elsevier.com/locate/jfa

Matrix power means and the Karcher mean

Yongdo Lim a,∗ , Miklós Pálfia b
a Department of Mathematics, Kyungpook National University, Taegu 702-701, Republic of Korea
b Department of Automation and Applied Informatics, Budapest University of Technology and Economics,
1117 Budapest, Magyar Tudósok krt. 2, Qép., Hungary
Received 17 September 2011; accepted 9 November 2011
Available online 23 November 2011
Communicated by Alain Connes

Abstract
We define a new family of matrix means {Pt (ω; A)}t∈[−1,1] , where ω and A vary over all positive prob-
ability vectors in Rn and n-tuples of positive definite matrices resp. Each of these means except t = 0
arises as a unique positive definite solution of a non-linear matrix equation, satisfies all desirable properties
of power means of positive real numbers and interpolates between the weighted harmonic and arithmetic
means. The main result is that the Karcher mean coincides with the limit of power means as t → 0. This
provides not only a sequence of matrix means converging to the Karcher mean, but also a simple proof
of the monotonicity of the Karcher mean, conjectured by Bhatia and Holbrook, and other new properties,
which have recently been established by Lawson and Lim and also Bhatia and Karandikar using probabilis-
tic methods on the metric structure of positive definite matrices equipped with the trace metric.
© 2011 Elsevier Inc. All rights reserved.

Keywords: Positive definite matrix; Geometric mean; Monotonicity; Riemannian trace metric; Metric nonpositive
curvature; Thompson metric; Power mean; Riemannian barycenter

1. Introduction

The Riemannian trace metric on the convex cone P = Pm of m × m positive definite Hermitian
matrices plays an important role in many applied areas involving matrix interpolation, filtering,
estimation, optimization and averaging, where it has been increasingly recognized that the Eu-
clidean distance is often not the most suitable for the set P and that working with the appropriate

* Corresponding author.
E-mail addresses: [email protected] (Y. Lim), [email protected] (M. Pálfia).

0022-1236/$ – see front matter © 2011 Elsevier Inc. All rights reserved.
doi:10.1016/j.jfa.2011.11.012
 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514 1499

geometry does matter in computational problems. (Recall the trace metric distance between two
 1
positive definite matrices is given by δ(A, B) = ( ki=1 log2 λi (A−1 B)) 2 , where λi (X) denotes
the i-th eigenvalue of X in ascending order.) It turns out that the Riemannian geometry plays a
key role particularly in the study of inversion invariant data averaging procedures in image pro-
cessing, in radar detection and in brain-computer interfacing [5,4,29]. An attractive candidate of
data averaging procedures is the least squares mean [24] of positive definite matrices. This mean
has appeared under a variety of other designations: Frechet mean, Cartan mean, Riemannian
center of mass [18], Riemannian geometric mean [29], or frequently, Karcher mean [14], the
terminology we adopt. The Karcher mean of n positive definite matrices A1 , . . . , An is defined
as the unique minimizer (provided it exists) of the sum of squares of the Riemannian trace metric
distances to each of the Ai , i.e.,

n
Λ(A1 , . . . , An ) = arg min δ 2 (X, Ai ). (1.1)
X∈P i=1

This idea had been anticipated by Élie Cartan (see, for example, Section 6.1.5 of [6]), who
showed among other things such a unique minimizer exists if the points all lie in a convex ball in
a Riemannian manifold; see also Karcher’s paper [18]. Using Karcher’s formula for the gradient
of the objective function (Theorem 2.1 of [18]) or computing appropriate derivatives as in [10,28]
yields that the Karcher mean coincides with the unique positive definite solution of the Karcher
equation

n
 
log X 1/2 A−1
i X
1/2
= 0. (1.2)
i=1

Various numerical methods for the solution of (1.1) or (1.2) have been introduced in the liter-
ature: fixed point methods, optimization algorithms like Newton’s method or a gradient descent
method, and iterative methods; see [14] and references therein. Unfortunately neither an ex-
plicit expression nor an explicit sequence of matrix means converging directly to the Karcher
mean is known. Nevertheless the monotonicity of the Karcher mean, conjectured by Bhatia and
Holbrook [11] and one of key axiomatic properties of matrix geometric means, was recently
established by Lawson and Lim [24] via a probabilistic convergence of approximations and by
Bhatia and Karandikar [12] via some probabilistic counting arguments, both arguments depend-
ing heavily on basic inequalities for the Riemannian metric. In this paper we provide a more
direct, non-probabilistic proof of the monotonicity of the Karcher mean that depends on finding
a sequence of matrix means satisfying monotonicity that converge directly to the Karcher mean.
The principal goal of this paper is to construct a particular family of matrix means, each with nu-
merous desirable properties such as monotonicity, that converges to the Karcher mean and show
that these properties are preserved in the limit.
a t +···+a t 1
The basic family of means we consider are the power means. The power mean ( 1 n n ) t
of n positive real
numbers a1 , . . . , an arises as the unique positive solution of the elementary
equation x = n1 ni=1 x 1−t ait , and converges to the geometric mean of a1 , . . . , an as t → 0. In

this paper we consider a matrix analogue of x = n1 ni=1 x 1−t ait , namely

1
n
X= X #t Ai , (1.3)
n
i=1
1500 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514

where A #t B = A1/2 (A−1/2 BA−1/2 )t A1/2 , the t-weighted geometric mean of A and B. We
prove that for each t ∈ (0, 1], Eq. (1.3) has a unique positive definite solution, denoted by
Pt (A1 , . . . , An ), and show that each of these matrix means (called a power mean) arises as
a unique fixed point of a strict contraction for the Thompson metric. We show these power
means vary continuously with t and satisfy analogues of basic properties of power means
of positive real numbers (e.g., monotonicity and joint concavity). We then establish that the
Karcher mean is the limit of power means as t → 0. This gives, in particular, a simple and
non-probabilistic proof of monotonicity, joint concavity and other new properties of the Karcher
mean recently established by Bhatia and Karandikar [12], and a globally convergent method for
obtaining the Karcher mean by taking the limit of Xk = P 1 (A1 , . . . , An ). Moreover, together
k
with P−t (A1 , . . . , An ) := Pt (A−1 −1 −1
1 , . . . , An ) , this provides a complete extension of the power
means of positive reals to positive definite matrices in the sense that the family of matrix means
{Pt (A1 , . . . , An )}t∈[−1,1] interpolates continuously between the harmonic (t = −1) and arith-
metic (t = 1) means with the Karcher mean appearing at t = 0.

2. Riemannian and Thompson metrics

Let H be the space of Hermitian matrices of a fixed size m, and P the convex cone of positive
definite Hermitian matrices. For X, Y ∈ H, we write that X  Y if Y − X is positive semidefinite,
and X < Y if Y − X is positive definite. The Frobenius norm  · 2 gives rise to the Riemannian
structure on P: X, Y A = Tr(A−1 XA−1 Y ), where A ∈ P, X, Y ∈ TA (P) = H. The Rieman-
 −1 1
nian metric distance is given by δ(A, B) = [ m 2
i=1 log λi (A B)] , where the λi (X) denote
2

the eigenvalues of X. Then P becomes a Cartan–Hadamard manifold, a simply connected com-

plete Riemannian manifold with non-positive sectional curvature [20]. For A, B ∈ P and t ∈ R,
the t-weighted geometric mean of A and B is defined by A #t B = A1/2 (A−1/2 BA−1/2 )t A1/2 .
The curve t → A #t B yields the unique geodesic from A to B for the Riemannian metric and
A # B = A #1/2 B is the unique midpoint between A and B. The following properties for the
weighted geometric mean are well known [19,24,22].

Lemma 2.1. Let A, B, C, D ∈ P and let t ∈ R. Then

(i) A #t B = A1−t B t for AB = BA, and (aA) #t (bB) = a 1−t bt (A #t B) for a, b > 0;
(ii) (Löwner–Heinz inequality) A #t B  C #t D for A  C, B  D and t ∈ [0, 1];
(iii) M(A #t B)M ∗ = (MAM ∗ ) #t (MBM ∗ ) for any non-singular M;
(iv) A #t B = B #1−t A, (A #t B)−1 = A−1 #t B −1 ;
(v) (λA + (1 − λ)B) #t (λC + (1 − λ)D)  λ(A #t C) + (1 − λ)(B #t D) for λ, t ∈ [0, 1];
(vi) det(A #t B) = det(A)1−t det(B)t ; and
(vii) ((1 − t)A−1 + tB −1 )−1  A #t B  (1 − t)A + tB for t ∈ [0, 1].

The Thompson metric on P is defined by d∞ (A, B) = log(A−1/2 BA−1/2 )∞ , where X∞
denotes the spectral norm of X. It is known that d∞ is a complete metric on P and d∞ (A, B) =
max{log M(B/A), log M(A/B)}, where M(B/A) = inf{α > 0: B  αA} = λ1 (A−1/2 BA−1/2 ),
the largest eigenvalue of A−1/2 BA−1/2 . See [32,16].

Lemma 2.2. (See [9,16].) We have

(i) d∞ (A, B) = d∞ (A−1 , B −1 ) = d∞ (MAM ∗ , MBM ∗ ) for any M ∈ GL(m, C);

 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514 1501

(ii) d∞ (A # B, A) = d∞ (A # B, B) = 12 d∞ (A, B);

(iii) d∞ (A #t B, C #t D)  (1 − t)d∞ (A, C) + td∞ (B, D), t ∈ [0, 1].

The following non-expansive property of addition for the Thompson metric will be useful for
our purpose.

Lemma 2.3. (See [23].) Let Ai , Bi ∈ P and let ti > 0, i = 1, 2, . . . , n. Then

 n 
 
n

d∞ ti Ai , ti Bi  max d∞ (Ai , Bi ) .
1in
i=1 i=1

3. Matrix power means

We denote by n the simplex of positive probability vectors in Rn convexly spanned by the

unit coordinate vectors.

Theorem 3.1. Let A1 , . . . , An ∈ P and let ω = (w1 , . . . , wn ) ∈ n . Then for each t =

(t1 , . . . , tn ) ∈ (0, 1]n , the following equation has a unique positive definite solution:


n
X= wi (X #ti Ai ). (3.4)
i=1

Furthermore, the solution varies continuously over t ∈ (0, 1]n .


Proof. We will show that the map f : P → P defined by f (X) = ni=1 wi (X #ti Ai ) is a strict
contraction with respect to the Thompson metric. Let X, Y > 0. By Lemma 2.2 and Lemma 2.3,
    
d∞ f (X), f (Y )  max d∞ wi (X #ti Ai ), wi (Y #ti Ai )
1in

 max d∞ (X #ti Ai , Y #ti Ai )
1in
 
 max (1 − ti )d∞ (X, Y ) = max (1 − ti ) d∞ (X, Y ).
1in 1in

Since max1in {1 − ti } ∈ [0, 1), f is a strict contraction.

By the continuity of fixed points of strict contractions (see e.g. [30]), the solution of (3.4)
varies continuously over t ∈ (0, 1]n . 2

Definition 3.2 (Matrix power means). Let A = (A1 , . . . , An ) ∈ Pn and ω ∈ n . For t ∈ (0, 1], we
denote by Pt (ω; A) the unique solution of


n
X= wi (X #t Ai ). (3.5)
i=1

For t ∈ [−1, 0), we define Pt (ω; A) = P−t (ω; A−1 )−1 , where A−1 = (A−1 −1
1 , . . . , An ). We call
Pt (ω; A) the ω-weighted power mean of order t of A1 , . . . , An . To simplify the notation we write
Pt (A) = Pt (1/n, . . . , 1/n; A).
1502 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514

 
Remark 3.3. We note that P1 (ω; A) = ni=1 wi Ai and P−1 (ω; A) = ( ni=1 wi A−1 −1
i ) , the ω-
weighted arithmetic and harmonic means of A1 , . . . , An , respectively. For t ∈ [−1, 0), Pt (ω; A)
is the unique positive definite solution of

−1

n
X= wi (X #−t Ai )−1 . (3.6)
i=1

n
Indeed, X −1 = i=1 wi (X
−1 #−t A−1
i ) if and only if X
−1 = P (ω; A−1 ).
−t


Remark 3.4. Let f : P → P defined by f (X) = ni=1 wi (X #t Ai ), t ∈ (0, 1]. Then by the
Löwner–Heinz inequality, f is monotone: X  Y implies that f (X)  f (Y ). By Theorem 3.1,
f is a strict contraction for the Thompson metric with the least contraction coefficient less than
or equal to 1 − t.

For A = (A1 , . . . , An ) ∈ Pn , M ∈ GL(m, C), a = (a1 , . . . , an ) ∈ (0, ∞)n , ω = (w1 , . . . , wn ) ∈

n , and for a permutation σ on n-letters, we set
 
MAM ∗ = MA1 M ∗ , . . . , MAn M ∗ , Aσ = (Aσ (1) , . . . , Aσn ),
1
A(k) = (A, . . . , A) ∈ Pnk , ω(k) = (ω, . . . , ω) ∈ nk ,
  k  
k k
  1
at = a1t , . . . , ant , ω a = n (w1 a1 , . . . , wn an ) ∈ n ,
i=1 wi ai
1
ω̂ = (w1 , . . . , wn−1 ) ∈ n−1 , a · A = (a1 A1 , . . . , an An ).
1 − wn

We list some basic properties of Pt (ω; A).

Proposition 3.5. Let A = (A1 , . . . , An ), B = (B1 , . . . , Bn ) ∈ Pn , ω ∈ n , a = (a1 , . . . , an ) ∈

(0, ∞)n and let s, t ∈ [−1, 1] \ {0}.
 1
(1) Pt (ω; A) = ( ni=1 wi Ati ) t if the Ai ’s commute;
n 1
(2) Pt (ω; a · A) = ( i=1 wi ait ) t Pt (ω at ; A);
(3) Pt (ωσ ; Aσ ) = Pt (ω; A) for any permutation σ ;
(4) Pt (ω; A)  Pt (ω; B) if Ai  Bi for all i = 1, 2, . . . , n;
(5) d∞ (Pt (ω; A), Pt (ω; B))  max1in {d∞ (Ai , Bi )};
(6) (1 − u)P|t| (ω; A) + uP|t| (ω; B)  P|t| (ω; (1 − u)A + uB) for any u ∈ [0, 1];
(7) Pt (ω; MAM ∗ ) = MPt (ω; A)M ∗ for any invertible matrix M;
(8) Pt (ω; A−1 )−1 = P−t (ω; A);
(9) Det(P−|t| (ω; A))  ni=1 Det(Ai )wi  Det(P|t| (ω; A));
 n
(10) ( ni=1 wi A−1 i )
−1  P (ω; A) 
t i=1 wi Ai ;
(11) Pt (ω(k) ; A(k) ) = Pt (ω; A) for any k ∈ N;
(12) Pt (ω; A1 , . . . , An−1 , X) = X if and only if X = Pt (ω̂; A1 , . . . , An−1 ). In particular,
Pt (A1 , . . . , An , X) = X if and only if X = Pt (A1 , . . . , An );
 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514 1503

(13) For s ∈ (0, 1], Pt (ω; X #s A1 , . . . , X #s An ) = X if and only if X = Pst (ω; A);
(14) If t ∈ (0, 1], then Φ(Pt (ω; A))  Pt (ω; Φ(A)) for any positive unital linear map Φ, where
Φ(A) = (Φ(A1 ), . . . , Φ(An )). If t ∈ [−1, 0), then Φ(Pt (ω; A))  Pt (ω; Φ(A)) for any
strictly positive unital linear map Φ; and
(15) For any unitarily invariant norm ||| · ||| and t ∈ (0, 1],
1
− 1t
  
n t
  
n
 t
Pt (ω; A)  wi |||Ai |||t and P−t (ω; A)  wi A−1  .
i
i=1 i=1


Proof. (1) Suppose that the Ai ’s commute. Let t ∈ (0, 1] and X = ( ni=1 wi Ati )1/t . Then
  
X #t Ai = X 1−t Ati and ni=1 wi (X #t Ai ) = nj=1 wi X 1−t Ati = X 1−t ni=1 Ati = X 1−t X t = X.

By uniqueness, ( ni=1 wi Ati )1/t = X = Pt (ω; A). Furthermore, P−t (ω; A) = Pt (ω; A−1 )−1 =

( ni=1 wi A−t
i )
−1/t .
 1
(2) Let t ∈ (0, 1]. Set β = ( ni=1 wi ait ) t , ζ = ω at ∈ n and X = Pt (ω at ; A). Then

ζi = (ω at )i = n 1w a t wi ait and X = ni=1 ζi (X #t Ai ). Therefore,
i=1 i i


n
  n 
n
wi (βX) #t (ai Ai ) = β 1−t wi ait (X #t Ai ) = β 1−t ζi β t (X #t Ai )
i=1 i=1 i=1

n
=β ζi (X #t Ai ) = βX.
i=1

 1
By uniqueness, ( ni=1 wi ait ) t Pt (ω at ; A) = βX = Pt (ω; a · A).
For t ∈ [−1, 0), we have
 − 1 −1
 −1 
n t
 
Pt (ω; a · A) = P−t ω; (a · A)−1 = wi ait P−t ω a ;A
t −1

i=1
 1  1

n t
  
n t
 
−1 −1
= wi ait P−t ω a ;A
t
= wi ait Pt ω at ; A .
i=1 i=1

(3) Follows from the defining Eqs. (3.5) and (3.6).

(4) Suppose that Ai  Bi for all i = 1, 2, . . . , n. Let t ∈ (0, 1]. Define


n 
n
f (X) = wi (X #t Ai ) and g(X) = wi (X #t Bi ).
i=1 i=1

Then Pt (ω; A) = limk→∞ f k (X) and Pt (ω; B) = limk→∞ g k (X) for any X ∈ P, by the Ba-
nach fixed point theorem. By the Löwner–Heinz inequality, f (X)  g(X) for all X ∈ P,
and f (X)  f (Y ), g(X)  g(Y ) whenever X  Y . Let X0 > 0. Then f (X0 )  g(X0 ) and
f 2 (X0 ) = f (f (X0 ))  g(f (X0 ))  g 2 (X0 ). Inductively, we have f k (X0 )  g k (X0 ) for all
k ∈ N. Therefore, Pt (ω; A) = limk→∞ f k (X0 )  limk→∞ g k (X0 ) = Pt (ω; B).
1504 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514

Let t ∈ [−1, 0). Then A−1  B−1 and thus P−t (ω; A−1 )  P−t (ω; B−1 ). Therefore,
Pt (ω; A) = P−t (ω; A−1 )−1  P−t (ω; B−1 )−1 = Pt (ω; B).
(5) Let t ∈ (0, 1]. Let X = Pt (ω; A) and Y = Pt (ω; B). Then by Lemma 2.2 and Lemma 2.3,
 

n 
n
d∞ (X, Y ) = d∞ wi (X #t Ai ), wi (Y #t Bi )
i=1 i=1
 
 max d∞ (X #t Ai , Y #t Bi )  max (1 − t)d∞ (X, Y ) + td∞ (Ai , Bi )
1in 1in

= (1 − t)d∞ (X, Y ) + t max d∞ (Ai , Bi ) ,
1in

which implies that d∞ (X, Y )  max1in {d∞ (Ai , Bi )}. Since d∞ is invariant under inversion,
we also have
    −1  −1 
d∞ P−t (ω; A), P−t (ω; B) = d∞ Pt ω; A−1 , Pt ω; B−1
    
= d∞ Pt ω; A−1 , Pt ω; B−1
  −1  
 max d∞ A−1 i , Bi = max d∞ (Ai , Bi ) .
1in 1in

(6) Let t ∈ (0, 1]. Let X = Pt (ω; A) and Y = Pt (ω; B). For u ∈ [0, 1], we set Zu =
(1 − u)X + uY . Let f (Z) = ni=1 wi (Z #t ((1 − u)Ai + uBi )). Then by the joint concavity
of the two-variable geometric mean


n
 
Zu = (1 − u)X + uY = wi (1 − u)(X #t Ai ) + u(Y #t Bi )
i=1

n
   
 wi (1 − u)X + uY #t (1 − u)Ai + uBi = f (Zu ).
i=1

Inductively, Zu  f k (Zu ) for all k ∈ N. Therefore, (1 − u)Pt (ω; A) + uPt (ω; B) = Zu 

Pt (ω; (1 − u)A + uB).
(7) Follows from the defining equation of Pt (ω; A) and the uniqueness of the positive definite
solution.
(8) True by definition. 
(9) Let t ∈ (0, 1]. Let X = Pt (ω; A). Then X = ni=1 wi (X #t Ai ) implies that
 

n 
n
Det(X) = Det wi (X #t Ai )  Det(X #t Ai )wi
i=1 i=1
t

n 
n
= Det(X) (1−t)wi
Det(Ai ) twi
= Det(X) (1−t) wi
Det(Ai ) ,
i=1 i=1

n = 2 and by an appropriate sym-

where the inequality follows by Corollary 7.6.9 of [17] for
metrization method for n > 2, and thus Det(Pt (ω; A))  ni=1 Det(Ai )wi . From this, we also
have
 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514 1505

    −1     −1
Det P−t (ω; A) = Det Pt ω; A−1 = Det P−t ω; A−1

n
  −1 −wi n
 Det Ai = Det(Ai )wi .
i=1 i=1

(10) Let t ∈ (0, 1]. Let X = Pt (ω; A). By using the two-variable weighted arithmetic-
geometric mean inequality, we obtain


n 
n
  
n
X= wi (X #t Ai )  wi (1 − t)X + tAi = (1 − t)X + t wi Ai ,
i=1 i=1 i=1
n
which implies that X  i=1 wi Ai . Similarly,

−1 −1

n 
n 
n
 
X= wi (X #t Ai )  wi (X #t Ai )−1 = wi X −1 #t A−1
i .
i=1 i=1 i=1

Taking inverses of both sides leads to


n
  n
 
n
−1 
X −1  wi X −1 #t A−1
i  w i (1 − t)X −1
+ tAi = (1 − t)X −1
+ t wi A−1
i ,
i=1 i=1 i=1

which implies that X  ( ni=1 wi A−1 −1
i ) .
The case t ∈ [−1, 0) holds by duality.
(11) Let t ∈ (0, 1] and let X = Pt (ω; A). Then
 n 
n
1  n
X= wi (X #t Ai ) = wi (X #t Ai ) + · · · + wi (X #t Ai )
k
i=1 i=1 i=1
 
k

and therefore X = Pt (ω(k) ; A(k) ). The case t ∈ [−1, 0) is similar. 

(12) Let t ∈ (0, 1]. Then Pt (ω; A1 , . . . , An−1 , X) = X if and only if X = n−1 i=1 wi (X #t Ai ) +
1 n−1
wn X if and only if X = 1−wn i=1 wi (X #t Ai ) if and only if X = Pt (ω̂; A1 , . . . , An−1 ). By
duality, (12) holds for t ∈ [−1, 0). If ω = n+1 1
(1, 1, . . . , 1) ∈ n+1 , then ω̂ = n1 (1, . . . , 1) ∈ n ,
and thus, Pt (A1 , . . . , An , X) = X if and only if X = Pt (A1 , . . . , An ).
(13) Note that X #t (X #s Ai ) = X #st Ai . Let  s ∈ (0, 1]. Suppose that t ∈ (0, 1]. Then X =
Pt (ω; X #s A1 , . . . , X #s An ) if and only if X = ni=1 wi (X #st Ai ) if and only if X = Pst (ω; A).
If t ∈ [−1, 0), then X = Pt (ω; X #s A1 , . . . , X #s An ) if and only if X −1 = P−t (ω; X −1 #s
A−1
1 ,...,X
−1 # A−1 ) if and only if X −1 = P
s n
−1
−st (ω; A ) if and only if X = Pst (ω; A), since
st ∈ (0, 1].
(14) Note that Φ(A #t B)  Φ(A) #t Φ(B) for any A, B > 0 and t ∈ [0, 1] (cf. Theorem 4.1.5
of [10]). Let t ∈ (0, 1] and Xt = Pt (ω; A). Then


n 
n
 
Φ(Xt ) = wi Φ(Xt #t Ai )  wi Φ(Xt ) #t Φ(Ai ) . (3.7)
i=1 i=1
1506 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514


Define f (X) = ni=1 wi (X #t Φ(Ai )). Then limk→∞ f k (X) = Pt (ω; Φ(A)) for any X > 0.
By (3.7), f (Φ(Xt ))  Φ(Xt ). Since f is monotonic, f k (Φ(Xt ))  Φ(Xt ) for all k ∈ N. Thus,
Pt (ω; Φ(A)) = limk→∞ f k (Φ(Xt ))  Φ(Xt ) = Φ(Pt (ω; A)).
Let t ∈ [−1, 0) and let Φ be a strictly positive unital linear map. By Choi’s inequality (The-
orem 2.3.6 of [10]), Φ(A)−1  Φ(A−1 ) for all A > 0. By (4) and the preceding paragraph,
Φ(P−t (ω; A−1 ))  P−t (ω; Φ(A−1 ))  P−t (ω; Φ(A)−1 ). This implies that Φ(Pt (ω; A)) =
Φ(P−t (ω; A−1 )−1 )  Φ(P−t (ω; A−1 ))−1  P−t (ω; Φ(A)−1 )−1 = Pt (ω; Φ(A)).
(15) Let t ∈ (0, 1] and X = Pt (ω; A). Then


n 
n 
n
|||X|||  wi |||X #t Ai |||  wi |||X|||1−t |||Ai |||t = |||X|||1−t wi |||Ai |||t ,
i=1 i=1 i=1

where the second inequality follows from Theorem 2.10 of [27] and Corollary IX.5.3 [8], and
 1
hence |||Pt (ω; A)|||  ( ni=1 wi |||Ai |||t ) t . Since |||A−1 |||  |||A|||−1 for any A > 0, |||P−t (ω; A)||| =

|||Pt (ω; A−1 )−1 |||  |||Pt (ω; A−1 )|||−1  [ ni=1 wi |||A−1 t − 1t . 2
i ||| ]

Remark 3.6. From the (AGH) inequalities (Proposition 3.5(10)) we can obtain other inequalities
from operator monotone functions on the positive reals. Let f : (0, ∞) → (0, ∞) be an operator
monotone increasing (resp. decreasing) function. Then
 
  n 
n
Pt ω; f (A)  wi f (Ai )  f wi Ai ,
i=1 i=1
−1  
  
n 
n
Pt ω; f (A)  wi f (Ai )−1 f wi Ai ,
i=1 i=1

respectively; these follow from the equivalence between operator monotonicity and operator log-
concavity by Ando and Hiai [1].

Property (12) implies in particular that Pt (ω̂; A1 , . . . , An−1 ) is the unique fixed point of the
map f (X) = Pt (ω; A1 , . . . , An−1 , X). By Proposition 3.5(5), f is a non-expansive map for the
Thompson metric.

Corollary 3.7. Let t ∈ [−1, 1] \ {0}, ω ∈ n and let A1 , . . . , An−1 ∈ P. Then there exists
X0 ∈ P such that limk→∞ f k (X0 ) = Pt (ω̂; A1 , . . . , An−1 ), where f : P → P is defined by
f (X) = Pt (ω; A1 , . . . , An−1 , X). Furthermore, for B ∈ P,

Pt (ω : A1 , . . . , An−1 , B)  B implies Pt (ω̂; A1 , . . . , An−1 )  B.

In particular, Pt (A1 , . . . , An−1 , B)  B implies that Pt (A1 , . . . , An−1 )  B.

Proof. Let f (X) = Pt (ω; A1 , . . . , An−1 , X). By Proposition 3.5(4), f is monotonic. Let B ∈ P.
Pick α, β > 0 such that B, Ai ∈ [βI, αI ] = {X ∈ P: βI  X  αI } for all i = 1, . . . , n − 1. Then
by Proposition 3.5(10), f maps [βI, αI ] into itself. Indeed,
 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514 1507

βI = Pt (ω; βI, . . . , βI )  Pt (ω; A1 , . . . , An−1 , X)  Pt (ω; αI, . . . , αI ) = αI

for any X ∈ [βI, αI ]. So, f k (X0 ) ∈ [βI, αI ] for all k ∈ N and for any X0 ∈ [βI, αI ]. Let
X0 ∈ [βI, αI ] such that f (X0 )  X0 . Then by induction f k+1 (X0 )  f k (X0 ) for all k ∈ N.
That is, {f k (X0 )}∞ k=1 is a decreasing sequence bounded below by βI and thus converges to
Pt (ω̂; A1 , . . . , An−1 ), which is the unique fixed point of f . In particular for X0 = αI , we have
that f (αI )  αI and limk→∞ f k (αI ) = Pt (ω̂; A1 , . . . , An−1 ).
Suppose that f (B)  B. Then f k+1 (B)  f k (B)  B for all k ∈ N and hence Pt (ω̂; A1 , . . . ,
An−1 ) = limk→∞ f k (B)  f (B)  B. 2

The problem of finding an explicit form of Pt (ω; A) is non-trivial, except for n = 2.

Proposition 3.8. For t ∈ (0, 1], we have

    t  1
Pt (w1 , w2 ; A, B) = A # 1 w1 A + w2 (A #t B) = A1/2 w1 I + w2 A−1/2 BA−1/2 t A1/2 .
t

n  n  k n−k
In particular, P 1 (w1 , w2 ; A, B) = k=0 k w1 w2 (B # k A).
n n

Proof. Let X = Pt (w1 , w2 ; A, B). Then by definition, X = w1 (X #t A) + w2 (X #t B). Set-

ting U = A−1/2 XA−1/2 and Z = A−1/2 BA−1/2 yields U = w1 U 1−t + w2 (U #t Z), which
is equivalent to I = w1 U −t + w2 (U −1/2 ZU −1/2 )t , that is, Z = ( U −w
t 1
1I t
w2 ) . This implies that
1
U = [w1 I + w2 Z t ] t and

 1   t  1
X = A1/2 U A1/2 = A1/2 w1 I + w2 Z t t A1/2 = A1/2 w1 I + w2 A−1/2 BA−1/2 t A1/2
   t  1/2  
= A1/2 I #1/t w1 I + w2 A−1/2 BA−1/2 A = A #1/t w1 A + w2 (A #t B) .

If t = 1
n for some n ∈ N, then

n  

 1 n n n−k
A−1/2 XA−1/2 = U = w1 I + w2 Z n = w1k w2n−k Z n
k
i=1

and hence
 n  
 n  
 n n−k  n  n−k 
X=A 1/2
w1k w2n−k Z n A1/2
= w1k w2n−k A1/2 Z n A1/2
k k
i=1 i=1
n 
  n 
 
n n
= w1k w2n−k (A # n−k B) = w1k w2n−k (B # k A). 2
k n k n
i=1 i=1

Remark 3.9. We observe that for any (w1 , w2 ) ∈ 2 ,

lim Pt (w1 , w2 ; A, B) = A #w2 B.

t→0
1508 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514

Indeed, setting Z = A−1/2 BA−1/2 , we have for t > 0

    1
A−1/2 A # 1 w1 A + w2 (A #t B) A−1/2 = w1 I + w2 Z t t → Z w2 . (3.8)
t

That is, Pt (w1 , w2 ; A, B) → A1/2 Z w2 A1/2 = A1/2 (A−1/2 BA−1/2 )w2 A1/2 = A #w2 B. This fur-
ther implies that

 −1
lim Pt (w1 , w2 ; A, B) = lim P−t w1 , w2 ; A−1 , B −1
t→0− t→0−
 −1
= A−1 #w2 B −1 = A #w2 B.

Remark 3.10. We note that the power mean Pt (A, B) coincides with the operator mean aris-
ing from ft : (0, ∞) → (0, ∞) defined by ft (x) = ( x 2+1 ) t . It is called the quasi-arithmetic
t 1

(power) mean of order t. Its operator monotonicity (cf. Proposition 3.5), infinite divisibility,
and the complete positivity of an associated linear operator have been studied by Bhatia and
Kosaki [13] and Besenyei and Petz [7]. It turns out [31] that the power mean Pt (A, B) arises
as the midpoint operation of a manifold equipped with an affine connection. One can also see
1
that Pt (w1 , w2 ; A, B)(= A # 1 [w1 A + w2 (A #t B)]) = (w1 At + w2 B t ) t for non-commuting A
t
At +···+At 1
and B. In fact, = exp( log A1 +···+log
limt→0 ( 1 n n ) t n
An
) and is known as the Log-Euclidean
mean [3]. We note that the Log-Euclidean mean is far from the geometric mean A # B for
n = 2.

4. The Karcher mean via power means

Let A = (A1 , . . . , An ) ∈ Pn , ω = (w1 , . . . , wn ) ∈ n . The ω-weighted Karcher mean Λ(ω; A)

of A is defined to be the unique positive definite solution of the equation


n
 
wi log X −1/2 Ai X −1/2 = 0. (4.9)
i=1

We note from (4.9) that Λ(ω; A−1 )−1 = Λ(ω; A), the self-duality of the Karcher mean.

Lemma 4.1. Let D ⊂ R be an open interval and let 0 > 0. Let F : D × (− 0 , 0) → R be a map
satisfying

(i) F is an increasing function in the first variable,

(ii) there exists a continuous and increasing function f : D → R such that for all a ∈ D, f (a) =
∂
∂t F (a, t)|t=0 .

Then for any a ∈ D and any sequence an of D converging to a,

   
1
lim n F an , − F (a, 0) = f (a).
n→∞ n
 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514 1509

Proof. Let a ∈ D and let an be a sequence converging to a. Let > 0. Since f is a continuous
and increasing function, there exists δ > 0 such that

f (a) − < f (a − δ)  f (a)  f (a + δ) < f (a) + . (4.10)

Since an → a, there exists N1 > 0 such that a − δ < an < a + δ for all n  N1 . Since F is an
increasing function in the first variable, F (a − δ, 1/n)  F (an , 1/n)  F (a + δ, 1/n) for all
n  N1 . That is, for all n  N1 ,
     
n F (a − δ, 1/n) − F (a, 0)  n F (an , 1/n) − F (a, 0)  n F (a + δ, 1/n) − F (a, 0) .

By (4.10), f (a) − < f (a − δ) and f (a + δ) < f (a) + , and by (ii),

 
lim n F (a ± δ, 1/n) − F (a, 0) = f (a ± δ)
n→∞

one can find N2 > N1 such that for all n  N2 , f (a) − < n(F (a ± δ, 1/n) − F (a, 0)) and
f (a) + > n(F (a + δ, 1/n) − F (a, 0)). This completes the proof. 2

Since the map F (x, t) = x t on (0, ∞) × R satisfies the conditions in Lemma 4.1, we have the
following result.

Lemma 4.2. Let x0 > 0 and let xn be a sequence of positive real numbers converging to x0 . Then
1/n
limn→∞ n(xn − 1) = log x0 .

The main result of this paper is the following.

Theorem 4.3. We have

lim Pt (ω; A) = Λ(ω; A).

t→0

Proof. By Proposition 3.5(8) and the self-duality of the Karcher

 mean, it suffices to show that
limt→0+ Pt (ω; A) = Λ(ω; A). Set Xt = Pt (ω; A). Since Xt = ni=1 wi (Xt #t Ai ),


n
 −1/2 −1/2 

n
 −1/2 −1/2 t
I= wi X t (Xt #t Ai )Xt = wi X t Ai Xt .
i=1 i=1

In particular for all t ∈ (0, 1],


n  −1/2 −1/2 t 
(Xt Ai Xt ) −I
0= wi . (4.11)
t
i=1

Let {tk }∞
k=1 be a sequence in (0, 1] converging to 0. Since Xt lies in the order interval
determined by the ω-weighted harmonic mean and arithmetic mean, which is compact, the
sequence {Xtk } has at least one limit point. Suppose that X0 is a limit point of {Xtk }. We
will show that X0 = Λ(ω; A). Passing to a subsequence, we may assume that Xtk → X0 , as
1510 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514

−1/2 −1/2 −1/2 −1/2 −1/2 −1/2

k → ∞. Then Xtk Ai Xtk → X0 Ai X0 for all i. Setting Ytk = Xtk Ai Xtk and
−1/2 −1/2
Y0 = X0 Ai X0 yields Ytk → Y0 . Let Utk be a unitary matrix such that := Dtk is Utk Ytk Ut∗k
a diagonal matrix. Since Ytk converges to Y0 and the unitary group is compact, we may assume
that Utk → U0 and Dtk → D0 for some unitary matrix U0 and a diagonal matrix D0 . Indeed, first
consider a subsequence of Utk converging to a unitary matrix U0 , second the corresponding sub-
sequence of Ytk , which always converges to Y0 , and then finally consider the corresponding sub-
t
Dtkk −I
sequence of Dtk , which converges to D0 := U0 Y0 U0∗ . By Lemma 2.3, limk→∞ tk = log D0 .
This implies that

 −1/2 −1/2 tk     ∗ 
(Xtk Ai Xtk ) −I (Ytk )tk − I Utk (Dtk )tk Utk − Im
lim = lim = lim
k→∞ tk k→∞ tk k→∞ tk
 tk 
D tk − I  
= lim Ut∗k Utk = log U0∗ D0 U0
k→∞ tk
 −1/2 −1/2 
= log Y0 = log X0 Ai X0 .

n −1/2 −1/2
This together with (4.11) yields 0 = i=1 wi log(X0 Ai X0 ). That is, X0 = Λ(ω; A). 2

From Theorem 3.1 and Theorem 4.3 we obtain

Corollary 4.4. With P0 (ω; A) = Λ(ω; A), the map t → Pt (ω; A) is continuous on [−1, 1].

The basic properties of power means in Proposition 3.5 together with Theorem 4.3 provide
simple proofs of some important properties of the Karcher mean.

Corollary 4.5. (Cf. [24,12].) The Karcher mean satisfies the following properties:

(P1) (Consistency with scalars) Λ(ω; A) = Aw 1 · · · Aw

1 wn
n if the Ai ’s commute;
(P2) (Joint homogeneity) Λ(ω; a1 A1 , . . . , an An ) = a1 1 · · · anwn Λ(ω; A);
(P3) (Permutation invariance) Λ(ωσ ; Aσ ) = Λ(ω; A), where ωσ = (wσ (1) , . . . , wσ (n) );
(P4) (Monotonicity) If Bi  Ai for all 1  i  n, then Λ(ω; B)  Λn (ω; A);
(P5) d∞ (Λ(ω; A), Λ(ω; B))  max1in {d∞ (Ai , Bi )};
(P6) (Invariancy) Λ(ω; M ∗ AM) = M ∗ Λ(ω; A)M for any invertible M;
(P7) (Joint concavity) Λ(ω; (1 − u)A + uB)  (1 − u)Λ(ω; A) + uΛ(ω; B) for 0  u  1;
(P8) (Self-duality) Λ(ω; A−1 −1 −1 = Λ(ω; A , . . . , A );
1 , . . . , An )  1 n
(P9) (Determinant identity) DetΛ(ω; A) = ni=1 (DetAi )wi ; and
 n
(P10) (AGH weighted mean inequalities) ( ni=1 wi A−1 i )
−1  Λ(ω; A) 
i=1 wi Ai ;
(P11) Λ(ω(k) ; A(k) ) = Λ(ω; A) for any k ∈ N;
(P12) Λ(ω; A1 , . . . , An−1 , X) = X if and only if X = Λ(ω̂; A1 , . . . , An−1 ). In particular,
Λ(A1 , . . . , An , X) = X if and only if X = Λ(A1 , . . . , An );
(P13) for any t ∈ (0, 1], X = Λ(ω; X #t A1 , . . . , X #t An ) if and only if X = Λ(ω; A);
(P14) Φ(Λ(ω; A))  Λ(ω; Φ(A)) for any positive unital linear map Φ. If Φ is strictly positive,
then
 Φ(Λ(ω;−w A)) = Λ(ω; Φ(A)); 
(P15) ni=1 |||A−1i ||| i  |||Λ(ω; A)|||  n
i=1 |||Ai ||| for any unitarily invariant norm ||| · |||.
wi
 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514 1511

Proof. By Proposition 3.5, Theorem 4.3 and by the Karcher equation, (P1)–(P13) are immediate.
For instance, since each Pt (ω; ·) is monotonic, its limit Λ(ω; ·) also is.
(P14) By Proposition 3.5(13), Φ(Pt (ω; A))  Pt (ω; Φ(A)) for all t ∈ (0, 1]. As t → 0, we
have Φ(Λ(ω; A))  Λ(ω; Φ(A)). If Φ is strictly positive, then Φ(Pt (ω; A))  Pt (ω; Φ(A)) for
all t ∈ [−1, 1) by Proposition 3.5(13). Then Φ(Λ(ω; A))  Λ(ω; Φ(A)).
 1
(P15) By Proposition 3.5(11), |||Pt (ω; A)|||  ( ni=1 wi |||Ai |||t ) t for all t ∈ (0, 1]. As t → 0,
 1 
we have |||Λ(ω; A)|||  limt→0 ( ni=1 wi |||Ai |||t ) t = ni=1 |||Ai |||wi , where the equality follows
from the fact that weighted power means of positive real numbers converge to the weighted
geometric mean. The other inequality follows similarly. 2

Ando, Li and Mathias [2] listed the ten properties (P1)–(P10) for the unweighted case
ω = (1/n, . . . , 1/n) as properties that a geometric mean of n positive definite matrices should
satisfy, and their mean, called the ALM geometric mean, possesses all of them. The BMP
geometric mean of Bini, Meini and Poloni [15] is also a matrix geometric mean in this ax-
iomatic sense. In fact, there are infinitely many matrix geometric means: fixed point means of the
ALM and BMP geometric means [26] and their weighted geometric means ALM(A1 , . . . , An ) #t
BMP(A1 , . . . , An ), t ∈ [0, 1]. The properties (P11)–(P13) are special for the Karcher mean. Some
parts of the properties (P14) and (P15) have been established by Bhatia and Karandikar [12]. For
the weighted case, there are also infinitely many weighted geometric means of n positive def-
inite matrices: the weighted Karcher mean, the weighted BMP geometric mean [25] and their
weighted geometric means. We note that there has been no successful weighted extension of the
ALM geometric mean.
Next we investigate some other properties of the power mean that hold for the Karcher
mean.

Corollary 4.6. If A  B and Ai < Bi for some i, then Λ(ω; A) < Λ(ω; B) and Pt (ω; A) <
Pt (ω; B) for any sufficiently small t.

Proof. Find 0 < α < 1 such that Ai  αBi . Then

Λ(ω; A1 , . . . , Ai , . . . , An )  Λ(ω; B1 , . . . , αBi , . . . , Bn ) = α wi Λ(ω; B1 , . . . , Bi , . . . , Bn )

< Λ(ω; B1 , . . . , Bi , . . . , Bn )

where we used the joint homogeneity and monotonicity of the Karcher mean. Finding 0 < β < 1
such that Λ(ω; A) < βΛ(ω; B) < Λ(ω; B), we have from Theorem 4.3 that

lim Pt (ω; A) = Λ(ω; A) < βΛ(ω; B) < Λ(ω; B) = lim Pt (ω; B)

t→0 t→0

which implies that Pt (ω; A) < Pt (ω; B) for any sufficiently small t. 2

The continuity, indeed Lipschitz continuity, of the Karcher mean (P5) follows from d∞ (Λ(ω;
A), Λ(ω; B))  max{d∞ (Ai , Bi )}, which in turn follows from Proposition 3.5(5) and Theo-
1512 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514

rem 4.3. A stronger result is the non-expansiveness of the Karcher mean;

  n
δ Λ(ω; A), Λ(ω; B)  wi δ(Ai , Bi ). (4.12)
i=1

This nice inequality has been proved by Lawson and Lim [24] and Bhatia and Karandikar [12].

Corollary 4.7. If δ(Λ(ω; A), Λ(ω; B)) = ni=1 wi δ(Ai , Bi ), then for any sufficiently small t,
δ(Pt (ω; A), Pt (ω; B)) < ni=1 wi δ(Ai , Bi ).

Proof. Note that

    n
lim δ Pt (ω; A), Pt (ω; B) = δ Λ(ω; A), Λ(ω; B) < wi δ(Ai , Bi ). 2
t→0
i=1

Property (P12) of the Karcher mean implies that Λ(ω̂; A1 , . . . , An−1 ) is the unique fixed point
of the map f (X) = Λ(ω; A1 , . . . , An−1 , X). By (4.12), f is a strict contraction on P with respect
to the Riemannian metric. The following (Löwer) order behavior around the fixed point of f is
special for the Karcher mean.

Corollary 4.8. We have

Λ(ω; A1 , . . . , An−1 , An ) < An implies Λ(ω̂; A1 , . . . , An−1 ) < Λ(ω; A1 , . . . , An−1 , An ),


and if ni=1 wi log Ai < 0, then Λ(ω; A1 , . . . , An ) < I and Pt (ω; A) < I for any sufficiently
small t.

Proof. Suppose that B := Λ(ω; A1 , . . . , An−1 , An ) < An . Define f : P → P by f (X) =

Λ(ω; A1 , . . . , An−1 , X). Then by (4.12), it is a strict contraction for δ and hence has a
unique positive definite solution. By (P12) and Banach fixed point theorem, limk→∞ f k (B) =
Λ(ω̂; A1 , . . . , An−1 ). It follows from f (An ) = B < An and strict monotonicity of the Karcher
mean (Corollary 4.6) that f (B) = Λ(ω; A1 , . . . , An−1 , B) < Λ(ω; A1 , . . . , An−1 , An ) =
f (An ) = B. By induction, f k (B) < · · · < f (B) < f (An ) = B and therefore Λ(ω̂; A1 , . . . ,
An−1 ) = limk→∞ f k (B) < B = Λ(ω; A1 , . . . , An−1 , An
).
Next, suppose that ni=1 wi log Ai < 0. Set Y = − ni=1 wi log Ai . Then Y > 0 and hence
B := exp(Y ) > I . From
 

n 
n
1 
n
0= wi log Ai + Y = wi log Ai + log B = wi log Ai + log B
2
i=1 i=1 i=1

and ω1 := 12 (w1 , . . . , wn , 1) ∈ n+1 , we have Λ(ω1 ; A1 , . . . , An , B) = I < B. By strict mono-

tonicity of the Karcher mean, Λ(ω1 ; A1 , . . . , An , I ) < Λ(ω1 ; A1 , . . . , An , B) = I . From the first
paragraph and the fact that ω̂1 = ω, we have Λ(ω; A1 , . . . , An )  Λ(ω1 ; A1 , . . . , An , I ) < I .
Finally limt→0 Pt (ω; A1 , . . . , An ) = Λ(ω; A1 , . . . , An ) < I implies that for sufficiently small t,
Pt (ω; A1 , . . . , An ) < I . 2
 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514 1513

One can obtain in a way similar to the preceding that Λ(ω; A1 , . . . , An−1 , An )  An implies
Λ(ω̂; A1 , . . . , An−1 )  Λ(ω; A1 , . . . , An−1 , An ) and


n
(Y ) wi log Ai  0 implies Λ(ω; A1 , . . . , An )  I.
i=1

The property (Y ), which was established by Yamazaki [33], is one of characteristic properties of
the Karcher mean by the following result.

Theorem 4.9. The Karcher mean is uniquely determined by congruence invariancy (P6), self-
duality (P8), and (Y ).

Proof. Let g : n × Pn → P be a map satisfying (P6), (P8) and (Y ). By (P8) and (Y ),

 n
i=1 wi log Ai = 0 implies that g(ω; A1 , . . . , An ) = I . Let X = Λ(ω; A). Then
n −1/2 A X −1/2 ) = 0 and hence g(ω; X −1/2 A X −1/2 , . . . , X −1/2 A X −1/2 ) = I .
i=1 wi log(X i 1 n
By (P6), g(A1 , . . . , An ) = X = Λn (A1 , . . . , An ). 2

Our method of deriving the monotonicity of the Karcher mean is free from any probabilistic
and Riemannian geometric techniques because we have just started from the Karcher equa-
tion (4.9). The Karcher equation can be defined on the convex cone of positive definite operators
on an infinite dimensional Hilbert space. But the existence and uniqueness of a positive definite
solution have not previously been investigated in any depth. The weighted power means exist
since the Thompson metric exists on the cone of positive definite operators, on which Lemma 2.2
and Lemma 2.3 are still valid [21,22]. So if one can show the monotonicity of the power mean
function t → Pt (ω; A), then the strong limit of the sequence Xk = P 1 (ω; A) exists and is prob-
k
ably a solution of the Karcher equation. Note that the power mean Pt (ω; A) is contained in the
order interval determined by the weighted harmonic and arithmetic means.
By a numerical simulation, the following result seems to be true. If 0  t  s  1, then
Pt (ω; A)  Ps (ω; A) for all ω ∈ n and A ∈ Pn . By (3.8), it is true for n = 2.
Changing the weighted arithmetic mean operation in the defining Eq. (3.4) of the power mean
Pt (ω; A) into any weighted geometric mean G(ω; A) of n positive definite matrices, which is
non-expansive for the Riemannian metric or the Thompson metric, yields other matrix geometric
means via the geometric mean equation

X = G(ω; X #t A1 , . . . , X #t An ), t ∈ (0, 1].

For instance, one may take G = BMP and G = ALM for the unweighted case (see [25] for their
non-expansiveness). One can check by the non-expansive property that a unique positive definite
solution exists, denoted by Gt (ω; A). By the self-duality of G, G−t (ω; A) := Gt (ω; A−1 )−1 =
Gt (ω; A). Then by using the fixed point approach in Proposition 3.5, one can see that Gt is a
weighted matrix mean (satisfies (P1)–(P10)) and is also non-expansive. By (P13), Λt = Λ for all
t ∈ (0, 1]. The general convergence of Gt (ω; A) as t → 0 and the monotonicity of t → Gt (ω; A)
are non-trivial and suggest interesting future work.
1514 Y. Lim, M. Pálfia / Journal of Functional Analysis 262 (2012) 1498–1514

Acknowledgments

The authors thank an anonymous referee for his/her insightful comments and suggestions.
This work was supported by the Korea Science and Engineering Foundation (KOSEF) grant
funded by the Korea government (MEST) (No. 2009-0070972).

References

[1] T. Ando, F. Hiai, Operator log-convex functions and operator means, Math. Ann. 350 (2011) 611–630.
[2] T. Ando, C.K. Li, R. Mathias, Geometric means, Linear Algebra Appl. 385 (2004) 305–334.
[3] V. Arsigny, P. Fillard, X. Pennec, N. Ayache, Geometric means in a novel vector space structure on symmetric
positive-definite matrices, SIAM J. Matrix Anal. Appl. 29 (2006) 328–347.
[4] A. Barachant, S. Bonnet, M. Congedo, C. Jutten, Riemannian geometry applied to BCI classification, preprint.
[5] F. Barbaresco, Interactions between symmetric cone and information geometries: Bruhat–Tits and Siegel spaces
models for higher resolution autoregressive doppler imagery, in: Lecture Notes in Comput. Sci., vol. 5416, 2009,
pp. 124–163.
[6] M. Berger, A Panoramic View of Riemannian Geometry, Springer-Verlag, 2003.
[7] A. Besenyei, D. Petz, Completely positive mappings and mean matrices, Linear Algebra Appl. 435 (2011) 984–997.
[8] R. Bhatia, Matrix Analysis, Springer-Verlag, New York, 1997.
[9] R. Bhatia, On the exponential metric increasing property, Linear Algebra Appl. 375 (2003) 211–220.
[10] R. Bhatia, Positive Definite Matrices, Princeton Ser. Appl. Math., Princeton University Press, Princeton, NJ, 2007.
[11] R. Bhatia, J. Holbrook, Riemannian geometry and matrix geometric means, Linear Algebra Appl. 413 (2006) 594–
618.
[12] R. Bhatia, R. Karandikar, Monotonicity of the matrix geometric mean, Math. Ann., doi:10.1007/s00208-011-
0721-9, in press.
[13] R. Bhatia, H. Kosaki, Mean matrices and infinite divisibility, Linear Algebra Appl. 424 (2007) 36–54.
[14] D. Bini, B. Iannazzo, Computing the Karcher mean of symmetric positive definite matrices, preprint.
[15] D. Bini, B. Meini, F. Poloni, An effective matrix geometric mean satisfying the Ando–Li–Mathias properties, Math.
Comp. 79 (2010) 437–452.
[16] G. Corach, H. Porta, L. Recht, Convexity of the geodesic distance on spaces of positive operators, Illinois J. Math. 38
(1994) 87–94.
[17] R. Horn, C. Johnson, Matrix Analysis, Cambridge University Press, 1985.
[18] H. Karcher, Riemannian center of mass and mollifier smoothing, Comm. Pure Appl. Math. 30 (1977) 509–541.
[19] F. Kubo, T. Ando, Means of positive linear operators, Math. Ann. 246 (1980) 205–224.
[20] S. Lang, Fundamentals of Differential Geometry, Grad. Texts in Math., Springer, 1999.
[21] J. Lawson, Y. Lim, Symmetric spaces with convex metrics, Forum Math. 19 (2007) 571–602.
[22] J. Lawson, Y. Lim, Metric convexity of symmetric cones, Osaka J. Math. 44 (2007) 795–816.
[23] J. Lawson, Y. Lim, A general framework for extending means to higher orders, Colloq. Math. 113 (2008) 191–221.
[24] J. Lawson, Y. Lim, Monotonic properties of the least squares mean, Math. Ann. 351 (2011) 267–279.
[25] H. Lee, Y. Lim, T. Yamazaki, Multi-variable weighted geometric means of positive definite matrices, Linear Algebra
Appl. 435 (2011) 307–322.
[26] Y. Lim, On Ando–Li–Mathias geometric mean equations, Linear Algebra Appl. 428 (2008) 1767–1777.
[27] J. Matharu, J. Aujla, Some inequalities for unitarily invariant norms, Linear Algebra Appl., doi:10.1016/
j.laa.2010.08.013, in press.
[28] M. Moakher, A differential geometric approach to the geometric mean of symmetric positive-definite matrices,
SIAM J. Matrix Anal. Appl. 26 (2005) 735–747.
[29] M. Moakher, On the averaging of symmetric positive-definite tensors, J. Elasticity 82 (2006) 273–296.
[30] K.-H. Neeb, Compressions of infinite-dimensional bounded symmetric domains, Semigroup Forum 61 (2001) 71–
105.
[31] M. Pálfia, Classification of affine matrix means, preprint, 2011.
[32] A.C. Thompson, On certain contraction mappings in a partially ordered vector space, Proc. Amer. Math. Soc. 14
(1963) 438–443.
[33] T. Yamazaki, The Riemannian mean and matrix inequalities related to the Ando–Hiai inequality and Chaotic order,
Oper. Matrices, in press.