Linear and Multilinear Algebra

ISSN: 0308-1087 (Print) 1563-5139 (Online) Journal homepage: http://www.tandfonline.com/loi/glma20

Reciprocal Lie–Trotter formula

Koenraad M.R. Audenaert & Fumio Hiai

To cite this article: Koenraad M.R. Audenaert & Fumio Hiai (2015): Reciprocal Lie–Trotter
formula, Linear and Multilinear Algebra, DOI: 10.1080/03081087.2015.1082957

To link to this article: http://dx.doi.org/10.1080/03081087.2015.1082957

Published online: 07 Sep 2015.

Submit your article to this journal

Article views: 4

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=glma20

Download by: [Florida State University] Date: 13 September 2015, At: 08:17
 Linear and Multilinear Algebra, 2015
http://dx.doi.org/10.1080/03081087.2015.1082957

Reciprocal Lie–Trotter formula

Koenraad M.R. Audenaertab and Fumio Hiaic∗
a Department of Mathematics, Royal Holloway University of London, Egham, United Kingdom;
b Department of Physics and Astronomy, Ghent University, Ghent, Belgium; c Graduate School of
Information Sciences, Tohoku University, Sendai, Japan
Communicated by N.-C. Wong
Downloaded by [Florida State University] at 08:17 13 September 2015

(Received 30 December 2014; accepted 15 July 2015)

Let A and B be positive semidefinite matrices. The limit of the expression

Z p := (A p/2 B p A p/2 )1/ p as p tends to 0 is given by the well-known
Lie–Trotter formula.Asimilar formula holds for the limit of G p := (A p # B p )2/ p
as p tends to 0, where X # Y is the geometric mean of X and Y . In this paper we
study the limit of Z p and G p as p tends to ∞ instead of 0, with the ultimate goal
of finding an explicit formula, which we call the reciprocal Lie–Trotter formula.
We show that the limit of Z p exists and find an explicit formula in a special case.
The limit of G p is shown for 2 × 2 matrices only.

Keywords: Lie–Trotter formula; reciprocal Lie–Trotter formula; positive semi-

definite matrix; operator mean; geometric mean; log-majorization; antisymmetric
tensor power; Grassmannian manifold
AMS Subject Classifications: Primary: 15A42; 15A16; 47A64

1. Introduction
For any matrices X and Y the well-known Lie–Trotter formula expresses the convergence
lim (e X/n eY/n )n = e X +Y .
n→∞

The symmetric form with a continuous parameter is also well-known for positive semidef-
inite matrices A, B ≥ 0 as
lim (A p/2 B p A p/2 )1/ p = P0 exp(log A+̇ log B), (1.1)
p0

where P0 is the orthogonal projection onto the intersection of the supports of A, B and
log A+̇ log B is defined as P0 (log A)P0 + P0 (log B)P0 .
When σ is an operator mean [1] corresponding to an operator monotone function f
on (0, ∞) such that α := f  (1) is in (0, 1), the operator mean version of the Lie–Trotter
formula holds as
lim (A p σ B p )1/ p = P0 exp((1 − α) log A+̇α log B) (1.2)
p0

∗ Corresponding author. Email: hiai.fumio@gmail.com

© 2015 Taylor & Francis

 2 K.M.R. Audenaert and F. Hiai

for matrices A, B ≥ 0. (This version follows from [2, Theorem 4.11] if one of A, B is
positive definite, but it can be shown for general positive semidefinite matrices based on
[3, Section 4].) In particular, let σ be the geometric mean A # B (introduced first in [4] and
further discussed in [1]), corresponding to the operator monotone function f (x) = x 1/2
(hence α = 1/2). Then (1.2) yields
lim (A p # B p )2/ p = P0 exp(log A+̇ log B), (1.3)
p0

which has the same right-hand side as (1.1).

It turns out that the convergence of both (1.1) and (1.3) is monotone in the log-
majorization order. For d × d matrices X, Y ≥ 0, the log-majorization relation X ≺(log) Y
means that
k k
λi (X ) ≤ λi (Y ), 1 ≤ k ≤ d,
Downloaded by [Florida State University] at 08:17 13 September 2015

i=1 i=1
with equality for k = d, where λ1 (X ) ≥ · · · ≥ λd (X ) are the eigenvalues of X sorted in
decreasing order and counting multiplicities. The Araki–Lieb–Thirring inequality can be
written in terms of log-majorization as
(A p/2 B p A p/2 )1/ p ≺(log) (Aq/2 B q Aq/2 )1/q if 0 < p < q, (1.4)
for matrices A, B ≥ 0, see [5, p. 301–302] and [6,7]. One can also consider the comple-
mentary version of (1.4) in terms of the geometric mean. Indeed, for A, B ≥ 0 we have [7]

(Aq # B q )2/q ≺(log) (A p # B p )2/ p if 0 < p < q. (1.5)

Hence, for matrices A, B ≥ 0, we see that Z p := (A p/2 B p A p/2 )1/ p and G p :=
(A p # B p )2/ p both tend to P0 exp(log A+̇ log B) as p  0, with the former decreasing
(by (1.4)) and the latter increasing (by (1.5)) in the log-majorization order.
The main topic of this paper is the question about what happens to the limits of Z p and
G p as p tends to ∞ instead of 0. Although this seems a natural mathematical problem, we
have not been able to find an explicit statement of concern in the literature. It is obvious that
if A and B are commuting then G p = AB = Z p , independently of p > 0. However, if A
and B are not commuting, then the limit behaviour of Z p and its eigenvalues as p → ∞ is
of a rather complicated combinatorial nature, and that of G p seems even more complicated.
The problem of finding an explicit formula, which we henceforth call the reciprocal Lie–
Trotter formula, also emerges from recent developments of new Rényi relative entropies
relevant to quantum information theory. Indeed, the recent paper [8] proposed to generalize
the Rényi relative entropy as
1  z
Dα,z (ρ
σ ) := log Tr ρ α/2z σ (1−α)/z ρ α/2z
α−1
for density matrices ρ, σ with two real parameters α, z, and discussed the limit formulas
when α, z converge to some special values. The limit case of Dα,z (ρ
σ ) as z → 0 with α
fixed is exactly related to our reciprocal Lie–Trotter problem.
The rest of the paper is organized as follows. In Section 2, we prove the existence of the
limit of Z p as p → ∞ when A, B are d × d positive semidefinite matrices. In Section 3 we
analyse the case when the limit eigenvalue list of Z p becomes λi (A)λi (B) (1 ≤ i ≤ d), the
maximal case in the log-majorization order. Finally, in Section 4, we treat G p ; however, we
 Linear and Multilinear Algebra 3

can prove the existence of the limit of G p as p → ∞ only when A, B are 2 × 2 matrices,
and the general case must be left unsettled. The appendix is a proof of a technical lemma
stated in Section 2.

2. Limit of ( A p/2 B p A p/2 )1/ p as p → ∞

Let A and B be d × d positive semidefinite matrices having the eigenvalues a1 ≥ · · · ≥
ad (≥ 0) and b1 ≥ · · · ≥ bd (≥ 0), respectively, sorted in decreasing order and counting
multiplicities. Let {v1 , . . . , vd } be an orthonormal basis of eigenvectors of A such that
Avi = ai vi for i = 1, . . . , d, and {w1 , . . . , wd } an orthonormal basis of eigenvectors of B
in a similar way. Then A and B are diagonalized as

d
∗
A = V diag(a1 , . . . , ad )V = ai vi vi∗ , (2.1)
Downloaded by [Florida State University] at 08:17 13 September 2015

i=1
d
B = W diag(b1 , . . . , bd )W ∗ = bi wi wi∗ . (2.2)
i=1

For each p > 0 define a positive semidefinite matrix

Z p := (A p/2 B p A p/2 )1/ p , (2.3)
whose eigenvalues are denoted as λ1 ( p) ≥ λ2 ( p) ≥ · · · ≥ λd ( p), again in decreasing order
and counting multiplicities.

Lem m a 2.1 For every i = 1, . . . , d the limit

λi := lim λi ( p) (2.4)
p→∞

exists, and a1 b1 ≥ λ1 ≥ · · · ≥ λd ≥ ad bd .

Proof Since (a1 b1 ) p I ≥ A p/2 B p A p/2 ≥ (ad bd ) p I , we have a1 b1 ≥ λi ( p) ≥ ad bd

for all i = 1, . . . , d and all p > 0. By the Araki–Lieb–Thirring inequality [6] (or the
log-majorization [7]), for every k = 1, . . . , d we have

k 
k
λi ( p) ≤ λi (q) if 0 < p < q. (2.5)
i=1 i=1
k
Therefore, the limit ηk of i=1 λi ( p) as p → ∞ exists for any k = 1, . . . , d so that
η1 ≥ · · · ≥ ηd ≥ 0. Let m (0 ≤ m ≤ d) be the largest k such that ηk > 0 (with m := 0
if η1 = 0). When 1 ≤ k ≤ m, we have λk ( p) → ηk /ηk−1 (where η0 := 1) as p → ∞.
When m < d, λm+1 ( p) → ηm+1 /ηm = 0 as p → ∞. Hence λk ( p) → 0 for all k > m.
Therefore, the limit of λi ( p) as p → ∞ exists for any i = 1, . . . , d. The latter assertion is
clear now. 

Lem m a 2.2 The limit of the first eigenvalue in (2.4) is given by

λ1 = max{ai b j : (V ∗ W )i j = 0},
where (V ∗ W )i j denotes the (i, j) entry of V ∗ W .
 4 K.M.R. Audenaert and F. Hiai

Proof Write V ∗ W = [u i j ]. We observe that

  
d
V ∗ A p/2 B p A p/2 V
p/2 p/2 p
ij
= u ik u jk ai a j bk .
k=1

In particular,
  
d
∗ p p
V A p/2 p
B A p/2
V ii
= |u ik |2 ai bk
k=1

and hence we have


d 
d
p p p p
λ1 ( p) p ≤ Tr A p/2 B p A p/2 = |u ik |2 ai bk ≤ d 2 max{ai bk : u ik = 0},
Downloaded by [Florida State University] at 08:17 13 September 2015

i=1 k=1

where Tr is the usual trace functional on d × d matrices. Therefore,

λ1 ( p) ≤ d 2/ p max{ai bk : u ik = 0}. (2.6)

On the other hand, we have

p p
dλ1 ( p) p ≥ Tr A p/2 B p A p/2 ≥ min{|u ik |2 : u ik = 0} max{ai bk : u ik = 0}

so that
 1/ p
min{|u ik |2 : u ik = 0}
λ1 ( p) ≥ max{ai bk : u ik = 0}. (2.7)
d
Estimates (2.6) and (2.7) give the desired expression immediately. In fact, they prove the
existence of the limit λ1 in (2.4) as well apart from Lemma 2.1. 

In what follows, for each k = 1, . . . , d we write I d (k) for the set of all subsets
I of {1, . . . , d} with |I | = k. For I, J ∈ Id (k), we denote by (V ∗ W ) I,J the k × k
submatrix of V ∗ W corresponding to rows in I and columns in J ; hence det(V ∗W )
I,J
∗
denotes the corresponding minor of V W . We also write a I := i∈I ai and b I := i∈I bi .
Since det(V ∗ W ) = 0, note that for any k = 1, . . . , d and any I ∈ Id (k) we have
det(V ∗ W ) I,J = 0 for some J ∈ Id (k), and that for any J ∈ Id (k) we have
det(V ∗ W ) I,J = 0 for some I ∈ Id (k).

Lem m a 2.3 For every k = 1, . . . , d,

λ1 λ2 · · · λk = max{a I b J : I, J ∈ Id (k), det(V ∗ W ) I,J = 0}. (2.8)

Proof For each k = 1, . . . , d the antisymmetric tensor powers A∧k and B ∧k (see [9]) are
given in the form of diagonalizations as

A∧k = V ∧k diag(a I ) I ∈Id (k) V ∧k ,

B ∧k = W ∧k diag(b I ) I ∈Id (k) W ∧k ,
   
and the corresponding representation of the nk × nk unitary matrix V ∗∧k W ∧k is given by

(V ∗∧k W ∧k ) I,J = det(V ∗ W ) I,J , I, J ∈ Id (k).

 Linear and Multilinear Algebra 5

Note that the largest eigenvalue of

 ∧k p/2 ∧k p ∧k p/2 1/ p  p/2 p p/2 1/ p ∧k

(A ) (B ) (A ) = (A B A )

is λ1 ( p)λ2 ( p) · · · λk ( p), whose limit as p → ∞ is λ1 λ2 · · · λk by Lemma 2.1. Apply

Lemma 2.2 to A∧k and B ∧k to obtain expression (2.8). 

Let H be a d-dimensional Hilbert space (say, Cd ), k be an integer with 1 ≤ k ≤

d, and H∧k be the k-fold antisymmetric tensor of H. We write x 1 ∧ · · · ∧ xk (∈ H∧k )
for the antisymmetric tensor of x1 , . . . , xk ∈ H (see [9]). The next lemma says that the
Grassmannian manifold G(k, d) is realized in the projective space of H∧k . Although the
lemma might be known to specialists, we cannot find a precise explanation in the literature.
Downloaded by [Florida State University] at 08:17 13 September 2015

So, for the convenience of the reader, we will present its sketchy proof in Appendix 1 based
on [10].

Lem m a 2.4 There are constants α, β > 0 (depending on only d and k) such that
√
−1θ
α
P − Q
≤ inf
u 1 ∧ · · · ∧ u k − e v1 ∧ · · · ∧ vk
≤ β
P − Q

θ ∈R

for all orthonormal sets {u 1 , . . . , u k } and {v1 , . . . , vk } and the respective orthogonal projec-
tions P and Q onto span{u 1 , . . . , u k } and span{v1 , . . . , vk }, where
P − Q
is the operator
norm of P − Q and
·
inside infimum is the norm on H∧k .

The main result of the paper is the next theorem showing the existence of limit for the
reciprocal version of (1.1).

Theorem 2.5 For every d × d positive semidefinite matrices A and B, the matrix Z p in
(2.3) converges as p → ∞ to a positive semidefinite matrix.

Proof By replacing A and B with V AV ∗ and V BV ∗ , respectively, we may assume that

V = I and so

A = diag(a1 , . . . , ad ), B = W diag(b1 , . . . , bd )W ∗ .

Choose an orthonormal basis {u 1 ( p), . . . , u d ( p)} of Cd for which we have Z p u i ( p) =

λi ( p)u i ( p) for 1 ≤ i ≤ d. Let λi be given in Lemma 2.1, and assume that 1 ≤ k < d and
λ1 ≥ · · · ≥ λk > λk+1 . Moreover, let λ1 (Z ∧k ∧k
p ) ≥ λ2 (Z p ) ≥ . . . be the eigenvalues of
∧k
Z p in decreasing order. We note that

lim λ1 (Z ∧k
p ) = lim λ1 ( p) · · · λk−1 ( p)λk ( p)
p→∞ p→∞
= λ1 . . . λk−1 λk
> λ1 · · · λk−1 λk+1 = lim λ2 (Z ∧k
p ). (2.9)
p→∞
 6 K.M.R. Audenaert and F. Hiai

Hence it follows that λ1 (Z ∧k ∧k

p ) is a simple eigenvalue of Z p for every p sufficiently large.
Letting w I,J := det W I,J for I, J ∈ Id (k), we compute

(Z ∧k
p ) = (A )
p ∧k p/2 ∧k
W ((diag(b1 , . . . , bd ))∧k ) p (W ∧k )∗ (A∧k ) p/2
p/2
p
p/2
= diag(a I ) I w I,J I,J diag(b I ) I w J,I I,J diag(a I ) I
⎡ ⎤

=⎣ w I,K w J,K a I a J b K ⎦
p/2 p/2 p

K ∈Id (k) I,J

⎡ ⎤
  1/2 1/2 p
aI a J bK
= ηk ⎣ ⎦
p
w I,K w J,K ,
ηk
K ∈Id (k) I,J
Downloaded by [Florida State University] at 08:17 13 September 2015

where ηk := λ1 λ2 · · · λk > 0 so that

ηk = max{a I b K : I, K ∈ Id (k), w I,K = 0}

due to Lemma 2.3. We now define

 
k := (I, K ) ∈ Id (k)2 : w I,K = 0 and a I b K = ηk .

Then we have
⎡ ⎤
 p  p
Z ∧k  1/2 1/2
aI a J bK
=⎣ ⎦
p
w I,K w J,K
ηk ηk
K ∈Id (k) I,J
⎡ ⎤

−→ Q := ⎣ w I,K w J,K δ I,J,K ⎦ ,
K ∈Id (k) I,J

where 
1 if (I, K ), (J, K ) ∈ k ,
δ I,J,K :=
0 otherwise.

Since Q I,I ≥ |w I,K |2 > 0 when (I, K ) ∈ k , note that Q = 0. Furthermore, since
the eigenvalue λ1 (Z ∧k p ) is simple (if p large), it follows from (2.9) that the limit Q of
 ∧k p
Z p /ηk must be a rank one projection ψψ ∗ up to a positive scalar multiple, where
ψ is a unit vector in (Cd )∧k . Since the unit eigenvector u 1 ( p) ∧ · · · ∧ u k ( p) of Z ∧k
 p p
corresponding to the largest (simple) eigenvalue coincides with that√of Z ∧k p /ηk , we
conclude that u 1 ( p) ∧ · · · ∧ u k ( p) converges ψ up to a scalar multiple e −1θ . Therefore, by
Lemma 2.4, the orthogonal projection onto span{u 1 ( p), . . . , u k ( p)} converges as p → ∞.
Assume now that

λ1 = · · · = λk1 > λk1 +1 = · · · = λk2 > · · · > λks−1 +1 = · · · = λks (ks = d).

From the fact proved above, the orthogonal projection onto span{u 1 ( p), . . . , u kr ( p)} con-
verges for any r = 1, . . . , s − 1, and this is trivial for r = s. Therefore, the orthogonal
projection onto span{u kr −1 +1 ( p), . .
. , u kr ( p)} converges to a projection Pr for any r =
1, . . . , s, and thus Z p converges to rs =1 λkr Pr . 
 Linear and Multilinear Algebra 7

For 1 ≤ k ≤ d define ηk by the right-hand side of (2.8). Then Lemma 2.3 (see also the
proof of Lemma 2.1) implies that, for k = 1, . . . , d,
ηk
λk = if ηk > 0
ηk−1
(where η0 := 1), and λk = 0 if ηk = 0. So one can effectively compute the eigenvalues
of Z := lim p→∞ Z p ; however, it does not seem that there is a simple algebraic method to
compute the limit matrix Z .
We end the section with a brief exposition on the generalization of Theorem 2.5 to the
case of more than two matrices without proof. Let A1 , . . . , Am be d ×d positive semidefinite
matrices with diagonalizations
 (l) (l) 
Al = Vl Dl Vl∗ , Dl = diag a1 , . . . , ad , 1 ≤ l ≤ m.
Downloaded by [Florida State University] at 08:17 13 September 2015

For each p > 0 consider the positive semidefinite matrix

 p/2 p/2 p/2 p p/2 p/2 p/2 1/ p
Z p := A1 A2 · · · Am−1 Am Am−1 · · · A1 A1 (2.10)
 p/2 p ∗ p/2 1/ p ∗
Dm−1 · · · W1∗ D1
p/2 p/2
= V1 D1 W1 · · · Dm−1 Wm−1 Dm Wm−1 V1 ,

where
 
(l) d
Wl := Vl∗ Vl+1 = wi j , 1 ≤ l ≤ m − 1.
i, j=1

The eigenvalues of Z p are denoted as λ1 ( p) ≥ λ2 ( p) ≥ · · · ≥ λd ( p) in decreasing order.

Although the log-majorization in (2.5) is no longer available in the present situation, we
can extend Lemma 2.2 in such a way that the limit λ1 := lim p→∞ λ1 ( p) exists and
 (1) (2) (m) 
λ1 = max ai1 ai2 · · · aim : w(i 1 , i 2 , . . . , i m ) = 0 ,

where
 (1) (2) (m)
w(i 1 , i 2 , . . . , i m ) := wi1 j2 w j2 j3 · · · w jm−1 im : 1 ≤ j2 , . . . , jm−1 ≤ d,

(2) (m−1) (2) (m−1)
a j2 · · · a jm−1 = ai2 · · · aim−1 .

This can be applied to the antisymmetric tensors of Al ’s to show that the limit λi :=
lim p→∞ λi ( p) exists for every i = 1, . . . , d. Then the next theorem can be shown by
extending the proof of Theorem 2.5.

Theorem 2.6 For every d × d positive semidefinite matrices A1 , . . . , Am the matrix Z p

in (2.10) converges as p → ∞.

3. The maximal case

Let A and B be d × d positive semidefinite matrices with diagonalizations (2.1) and (2.2).
For each d × d matrix X we write s1 (X ) ≥ s2 (X ) ≥ · · · ≥ sd (X ) for the singular values
of X in decreasing order
k and counting multiplicities.
 For each p > 0 and k = 1, . . . , d,
k p/2 B p/2 ) 2/ p , by the majorization results of Gel’fand and
since i=1 λi ( p) = i=1 si (A
 8 K.M.R. Audenaert and F. Hiai

Naimark and of Horn (see, e.g. [9,11,12]), we have


k 
k 
k
ai j bn+1−i j ≤ λ j ( p) ≤ ajbj
j=1 j=1 j=1

for any choice of 1 ≤ i 1 < i 2 < · · · < i k ≤ d, and for k = d


d 
d
λi ( p) = det A · det B = ai bi .
i=1 i=1

That is, for any p > 0,

(ai bn+1−i )i=1
d
≺(log) (λi ( p))i=1
d
≺(log) (ai bi )i=1
d
(3.1)
Downloaded by [Florida State University] at 08:17 13 September 2015

with the notation of log-majorization, see [7]. Letting p → ∞ gives

(ai bn+1−i )i=1
d
≺(log) (λi )i=1
d
≺(log) (ai bi )i=1
d
(3.2)
for the eigenvalues λ1 ≥ · · · ≥ λd of Z = lim p→∞ Z p . In general, we have noth-
ing to say about the position of (λi )i=1 d in (3.2). For instance, when V ∗ W becomes the
permutation matrix corresponding to a permutation ( j1 , . . . , jd ) of (1, . . . , d), we have
Z p = V diag(a1 b j1 , . . . , ad b jd )V ∗ independently of p > 0 so that (λi ) = (ai b ji ).
In this section we clarify the case when (λi )i=1 d is equal to (ai bi )i=1
d , the maximal case

in the log-majorization order in (3.2). To do this, let 0 = i 0 < i 1 < · · · < il−1 < il = d
and 0 = j0 < j1 < · · · < jm−1 < jm = d be taken so that
a1 = · · · = ai1 > ai1 +1 = · · · = ai2 > · · · > ail−1 +1 = · · · = ail ,
b1 = · · · = b j1 > b j1 +1 = · · · = b j2 > · · · > b jm−1 +1 = · · · = b jm .

Theorem 3.1 In the above situation, the following conditions are equivalent:

(i) λi = ai bi for all i = 1, . . . , d;

(ii) for every k = 1, . . . , d so that ir −1 < k ≤ ir and js−1 < k ≤ js , there are
Ik , Jk ∈ Id (k) such that
{1, . . . , ir −1 } ⊂ Ik ⊂ {1, . . . , ir }, {1, . . . , js−1 } ⊂ Jk ⊂ {1, . . . , js },
det(V ∗ W ) Ik ,Jk = 0;
(iii) the property in (ii) holds for every k ∈ {i 1 , . . . , il−1 , j1 , . . . , jm−1 }.

Proof (i) ⇔ (ii). By Lemma 2.3 condition (ii) means that


k 
k
λi = ai bi , k = 1, . . . , d.
i=1 i=1

It follows (see the proof of Lemma 2.1) that this is equivalent to (i). (ii) ⇒ (iii) is trivial.
(iii) ⇒ (i). By Lemma 2.3 again condition (iii) means that

h 
h
λi = ai bi for all h ∈ {i 1 , . . . , il−1 , j1 , . . . , jm−1 }. (3.3)
i=1 i=1
 Linear and Multilinear Algebra 9
k k
This holds also for h = d thanks to (3.2). We need to prove that i=1 λi = i=1 ai bi for
all k = 1, . . . , d. Now, let ir −1 < k ≤ ir and js−1 < k ≤ js as in condition (ii). If k = ir or
k = js , then the conclusion has already been stated in (3.3). So assume that ir −1 < k < ir
and js−1 < k < js . Set h 0 := max{ir −1 , js−1 } and h 1 := min{ir , js } so that h 0 < k < h 1 .
By (3.3), for h = h 0 , h 1 , we have

h0 
h0 
h1 
h1
λi = ai bi > 0, λi = ai bi .
i=1 i=1 i=1 i=1
h 1
Since ai = ah 1 and bi = bh 1 for h 0 < i ≤ h 1 , we have i=h 0 +1
λi = (ah 1 bh 1 )h 1 −h 0 . By
h 0 +1 h 0 +1
(3.2) we furthermore have i=1 λi ≤ i=1 ai bi and hence
ah 1 bh 1 ≥ λh 0 +1 ≥ λh 0 +2 ≥ · · · ≥ λh 1 .
Downloaded by [Florida State University] at 08:17 13 September 2015

k k
Therefore, λi = ah 1 bh 1 for all i with h 0 + 1 < i ≤ h 1 , from which i=1 λi = i=1 ai bi
follows for h 0 < k < h 1 . 

Proposition 3.2 Assume that the equivalent conditions of Theorem 3.1 hold. Then, for
each r = 1, . . . , l, the spectral projection of Z corresponding to the set of eigenvalues
{air −1 +1 bir −1 +1 , . . . , air bir } is equal to the spectral projection ii=i
r
r −1 +1
vi vi∗ of A corre-
sponding to air . Hence Z is of the form

d
Z= ai bi u i u i∗
i=1
ir ∗
ir ∗
for some orthonormal basis {u 1 , . . . , u d } such that i=ir −1 +1 u i u i = i=ir −1 +1 vi vi for
r = 1, . . . , l.

Proof In addition to Theorem 2.5 we may prove that, for each k ∈ {i 1 , . . . , il−1 }, the
spectral projection of Z p corresponding to {λ1 ( p), . . . , λk ( p)} converges to i=1 k
vi vi∗ .
Assume that k = ir with 1 ≤ r ≤ l − 1. When js−1 < k < js , by condition (iii)
of Theorem 3.1, we have det(V ∗ W ){1,...,k},{1,..., js−1 , js ,..., jk } = 0 for some { js , . . . , jk } ⊂
{ js−1 + 1, . . . , js }. By exchanging w js , . . . , w jk with w js−1 +1 , . . . , wk we may assume that
det(V ∗ W ){1,...,k},{1,...,k} = 0. Furthermore, by replacing A and B with V AV ∗ and V BV ∗ ,
respectively, we may assume that V = I . So we end up assuming that
A = diag(a1 , . . . , ad ), B = W diag(b1 , . . . , bd )W ∗ ,
and det W (1, . . . , k) = 0, where W (1, . . . , k) denotes the principal k × k submatrix of the
top-left corner. Let {e1 , . . . , ed } be the standard basis of Cd . By Theorem 3.1, we have

k 
k−1
lim λ1 (Z ∧k
p )= ai bi > ai bi · ak+1 bk+1 = lim λ2 (Z ∧k
p )
p→∞ p→∞
i=1 i=1

so that the largest eigenvalue of Z ∧k p is simple for every sufficiently large p. Let {u 1 ( p), . . . ,
u d ( p)} be an orthonormal basis of Cd for which Z p u i ( p) = λi ( p)u i ( p) for 1 ≤ i ≤ d.
Then u 1 ( p) ∧ · · · ∧ u k ( p) is the unit eigenvector of Z ∧k
p corresponding to the eigenvalue
λ1 (Z ∧k d ∧k
p ). We now show that u 1 ( p) ∧ · · · ∧ u k ( p) converges to e1 ∧ · · · ∧ ek in (C ) . We
 10 K.M.R. Audenaert and F. Hiai

observe that
 
 p/2 
(A∧k ) p/2 = diag a I I = a{1,...,k} diag 1, α2 , . . . , α d
p/2 p/2 p/2
(k )
 
with respect to the basis ei1 ∧ · · · ∧ eik : I = {i 1 , . . . , i k } ∈ Id (k) , where the first diagonal
 
entry 1 corresponds to e1 ∧ · · · ∧ ek and 0 ≤ αh < 1 for 2 ≤ h ≤ dk . Similarly,
 
 
∧k p p p p
(diag(b1 , . . . , bd )) = b{1,...,k} diag 1, β2 , . . . , β d ,
(k )
 
where 0 ≤ βh ≤ 1 for 2 ≤ h ≤ dk . Moreover, W ∧k is given as
⎡ ⎤
w11 · · · w1(d )

⎢ . . .. ⎥
Downloaded by [Florida State University] at 08:17 13 September 2015

k
W ∧k = w I,J I,J = ⎢ ⎣ .. .. . ⎦,
⎥
w(d )1 · · · w(d )(d )
k k k

where w I,J = det W I,J and so w11 = det W (1, . . . , k) = 0. As in the proof of Theorem 2.5
we now compute
 p
(Z ∧k
p ) = (A )
p ∧k p/2 ∧k
W (diag(b1 , . . . , bd ))∧k (W ∧k )∗ (A∧k ) p/2
⎡ ⎤(d )
(dk ) k
 p ⎢ p/2 p/2 p ⎥
= a{1,...,k} b{1,...,k} ⎣ wi h w j h αi α j βh ⎦ ,
h=1
i, j=1

where α1 = β1 = 1. As p → ∞ we have
⎡ ⎤
(dk )   
⎢ p/2 p/2 p ⎥
⎣ w w α
ih jh i α j β h⎦ −→ diag |w 1h |2
, 0, . . . , 0
h=1 h:βh =1

Since the
 unit eigenvector of Z ∧kp corresponding
 to the largest eigenvalue coincides with
(k )
d
p/2 p/2 p
that of h=1 wi h w j h αi α j βh , it follows that u 1 ( p) ∧ · · · ∧ u k ( p) converges to
√
e1 ∧ · · · ∧ ek up to a scalar multiple e −1θ , θ ∈ R. By Lemma 2.4, this implies the desired
assertion. 

Corollary 3.3 If the eigenvalues a1 , . . . , ad of A are all distinct and the conditions of
Theorem 3.1 hold, then

lim (A p/2 B p A p/2 )1/ p = V diag(a1 b1 , a2 b2 , . . . , ad bd )V ∗ .

p→∞

In particular, when the eigenvalues of A are all distinct and so are those of B, the
conditions of Theorem 3.1 mean that all the leading principal minors of V ∗ W are non-zero.
Let E be a k-dimensional subspace of Cd and E be the orthogonal projection onto E.
Let B be given with diagonalization in (2.2). In [13, Theorem 1.2] Bourin proved that if
E ∩ span{wi : i > k} = {0} then
 Linear and Multilinear Algebra 11

lim λi ((E B p E)1/ p ) = bi for all i = 1, . . . , k. (3.4)

p→∞

Since E B p E is of at most rank k, it is obvious that λi ((E B p E)1/ p ) = 0 for all i > k and
all p > 0. The following is a slight refinement of Bourin’s result.

Corollary 3.4 Let E and B be as stated above. Then the following conditions are
equivalent:

(a) the above (3.4) holds;

(b) E ∩ span{w j : j > k} = {0} for some choice of W in (2.2);
(c) Ew1 , . . . , Ewk are linearly independent for some choice of W in (2.2).

Proof Let {v1 , . . . , vk } be an orthonormal basis of E. In the particular situation of A = E

Downloaded by [Florida State University] at 08:17 13 September 2015

(so a1 = · · · = ak = 1 and ak+1 = · · · = ad = 0), condition (i) of Theorem 3.1

becomes (a) here, and it is immediate to see that (iii) of Theorem 3.1 is equivalent to
det[vi , w j ]1≤i, j≤k = 0 for some choice of W . Since vi , w j  = vi , Ew j  for 1 ≤ i, j ≤
k, it is obvious that the latter condition is equivalent to (c). Hence (a) ⇔ (c). Next, we prove
(b) ⇔ (c). For a given W let F be the orthogonal projection onto F := span{w1 , . . . , wk }.
Consider the operators E F := E|F : F → E and FE := F|E : E → F. Since

FE x, y = x, y = x, E F y, x ∈ E, y ∈ F,

it follows that FE is the adjoint of E F . Since the kernel of FE is E ∩ F ⊥ , (b) means that
FE is injective. This is equivalent to that E F is surjective, which means (c). 

4. Limit of ( A p # B p )1/ p as p → ∞
Another problem, seemingly more interesting, is to know what is shown on the convergence
(A p σ B p )1/ p as p → ∞, the reciprocal version of (1.2). For example, 1/when σ = , the
p
arithmetic mean, the increasing limit of (A p  B p )1/ p = (A p + B p )/2 as 1 ≤ p → ∞
exists and
A ∨ B := lim (A− p  B − p )−1/ p = lim (A p + B p )1/ p (4.1)
p→∞ p→∞

is the supremum of A, B with respect to some spectral order among Hermitian matrices,
see [14] and [15, Lemma 6.15]. When σ = !, the harmonic mean, we have the infimum
counterpart A ∧ B := lim p→∞ (A p ! B p )1/ p , the decreasing limit as 1 ≤ p → ∞.
In this section we are interested in the case where σ = #, the geometric mean. For each
p > 0 and d × d positive semidefinite matrices A, B with the diagonalizations in (2.1) and
(2.2), we define
G p := (A p # B p )2/ p , (4.2)
 2/ p
which is given as A p/2 (A− p/2 B p A− p/2 )1/2 A p/2 if A > 0. The eigenvalues of G p are
denoted as λ1 (G p ) ≥ · · · ≥ λd (G p ) in decreasing order.

Proposition 4.1 For every i = 1, . . . , d the limit


λi := lim λi (G p )
p→∞
 12 K.M.R. Audenaert and F. Hiai

exists, and a1 b1 ≥ 
λ1 ≥ · · · ≥ 
λd ≥ ad bd . Furthermore,
 d
(ai bd+1−i )i=1
d
≺(log) λi i=1 ≺(log) (ai bi )i=1
d
. (4.3)

Proof Since (a1 b1 ) p/2 I ≥ A p # B p ≥ (ad bd ) p/2 I , we have a1 b1 ≥ λi (G p ) ≥ ad bd for

all i = 1, . . . , d and p > 0. By the log-majorization result in [7, Theorem 2.1], for every
k = 1, . . . , d we have

k 
k
λi (G p ) ≥ λi (G q ) if 0 < p < q. (4.4)
i=1 i=1
k
This implies that the limit of i=1 λi (G p ) as p → ∞ exists for every k = 1, . . . , d, and
hence the limit λi (G p ) exists for i = 1, . . . , d as in the proof of Lemma 2.1.
Downloaded by [Florida State University] at 08:17 13 September 2015

To prove the latter assertion, it suffices to show that

(ai bd+1−i )i=1

d
≺(log) (λi (G 1 ))i=1
d
≺(log) (ai bi )i=1
d
(4.5)

for G 1 = (A # B)2 . Indeed, applying this to A p and B p we have

(ai bd+1−i )i=1

d
≺(log) (λi (G p ))i=1
d
≺(log) (ai bi )i=1
d

so that (4.3) follows by letting p → ∞. To prove (4.5), we may by continuity assume that
A > 0. By [7, Corollary 2.3] and (3.1) we have
 d
(λi (G 1 ))i=1
d
≺(log) λi (A1/2 B A1/2 ) i=1 ≺(log) (ai bi )i=1
d
.

Since G 1 A−1 G 1 = B, there exists a unitary matrix V such that A−1/2 G 1 A−1/2 =
1/2 1/2

V BV ∗ and hence G 1 = A1/2 V BV ∗ A1/2 . Since λi (V BV ∗ ) = bi , by the majorization of

Gel’fand and Naimark we have

(ai bd+1−i )i=1

d
≺(log) (λi (G 1 ))i=1
d
,

proving (4.5) 

In view of (2.5) and (4.4) we may consider G p as the complementary counterpart of

Z p in some sense; yet it is also worth noting that G p is symmetric in A and B while Z p
is not. Our ultimate goal is to prove the existence of the limit of G p in (4.2) as p → ∞
similarly
 d to Theorem 2.5 and to clarify, similarly to Theorem 3.1, the minimal case when

λi i=1 is equal to the decreasing rearrangement of (ai bd+1−i )i=1
d . However, the problem

seems much more difficult, and we can currently settle the special case of 2 × 2 matrices
only.

Proposition 4.2 Let A and B be 2 × 2 positive semidefinite matrices with the diagonal-
izations (2.1) and (2.2) with d = 2. Then G p in (4.2) converges as p → ∞ to a positive
semidefinite matrix whose eigenvalues are



 (a1 b1 , a2 b2 ) if (V ∗ W )12 = 0,
λ1 , 
λ2 =
(max{a1 b2 , a2 b1 }, min{a1 b2 , a2 b1 }) if (V ∗ W )12 = 0.
 Linear and Multilinear Algebra 13

Proof Since
 2/ p ∗
G p = V (diag(a1 , a2 )) p # (V ∗ W diag(b1 , b2 )V ∗ W ) p V ,
we may assume without loss of generality that V = I (then V ∗ W = W ).
First, when W12 = 0 (hence W is diagonal), we have for every p > 0
G p = diag(a1 b1 , a2 b2 ).
 
0 w1
Next, when W11 = 0 (hence W = with |w1 | = |w2 | = 1), we have for every
w2 0
p>0
G p = diag(a1 b2 , a2 b1 ).
 
w11 w12
In the rest it suffices to consider the case where W = with wi j = 0 for
Downloaded by [Florida State University] at 08:17 13 September 2015

w21 w22
all i, j = 1, 2. First, assume that det A = det B = 1 so that a1 a2 = b1 b2 = 1. For every
p > 0, since det A p = det B p = 1, it is known [16, Proposition 3.11] (also [17, Proposition
4.1.12]) that
Ap + B p
Ap # B p = √
det(A p + B p )
so that
(A p + B p )2/ p
Gp =  1/ p .
det(A p + B p )
Compute
 p p p p p 
a + |w11 |2 b1 + |w12 |2 b2 w11 w21 b1 + w12 w22 b2
Ap + B p = 1 p p p p p (4.6)
w11 w21 b1 + w12 w22 b2 a2 + |w21 |2 b1 + |w22 |2 b2
and
det(A p + B p ) = 1 + |w21 |2 (a1 b1 ) p + |w22 |2 (a1 b2 ) p + |w11 |2 (a2 b1 ) p + |w12 |2 (a2 b2 ) p
+ |w11 w22 − w12 w21 |2 . (4.7)
Hence, we have
 1/ p  1/ p
lim det(A p + B p ) = a 1 b1 , lim Tr (A p + B p ) = max{a1 , b1 }.
p→∞ p→∞

Therefore, thanks to (4.1), we have

(A ∨ B)2
lim G p = .
p→∞ a 1 b1
Since
1   p/2
Tr (A p # B p ) ≤ λ1 (G p ) ≤ Tr (A p # B p ),
2
we obtain
 2/ p
 2/ p Tr (A p + B p )

λ1 = lim Tr (A # B )
p p
= lim  1/ p
p→∞ p→∞
det(A p + B p )
 
max{a12 , b12 } a 1 b1
= = max , = max{a1 b2 , a2 b1 }.
a 1 b1 b1 a 1
 14 K.M.R. Audenaert and F. Hiai

Furthermore, 
λ2 = min{a1 b2 , a2 b1√ λ1
} follows since  λ2 = 1.
√
For general A, B > 0 let α := det A and β := det B. Since
 2/ p
G p = αβ (α −1 A) p # (β −1 B) p ,
we see from the above case that G p converges as p → ∞ and

λ1 = αβ max{(α −1 a1 )(β −1 b2 ), (α −1 a2 )(β −1 b1 )} = max{a1 b2 , a2 b1 },
and similarly for 
λ2 .
The remaining is the case when a2 and/or b2 = 0. We may assume that a1 , b1 > 0
since the case A = 0 or B = 0 is trivial. When a2 = b2 = 0, since a1−1 A and b1−1 B are
non-commuting rank one projections, we have G p = 0 for all p > 0 by [1, (3.11)]. Finally,
assume that a2 = 0 and B > 0. Then we may assume that a1 = 1 and det B = 1. For ε > 0
Downloaded by [Florida State University] at 08:17 13 September 2015

set Aε := diag(1, ε 2 ). Since det(ε −1 Aε ) = 1, we have

  (ε −1 Aε ) p + B p
Aεp # B p = ε p/2 (ε −1 Aε ) p # B p = ε p/2   .
det (ε −1 Aε ) p + B p

By use of (4.6) and (4.7) with a1 = ε −1 and a2 = ε we compute

 p p −1/2
A p # B p = lim Aεp # B p = |w21 |2 b1 + |w22 |2 b2 diag(1, 0)
ε0

so that
lim G p = diag(b1−1 , 0) = diag(b2 , 0),
p→∞
which is the desired assertion in this final situation. 

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
The work of FH was supported in part by Grant-in-Aid for Scientific Research (C)26400103.

References

[1] Kubo F, Ando T. Means of positive linear operators. Math. Ann. 1980;246:205–224.
[2] Hiai F. Log-majorizations and norm inequalities for exponential operators. In: Janas J, Szafraniec
FH, Zemánek J, editors. Linear operators. Banach Center Publications 38. Warsaw: Polish Acad.
Sci.; 1997. p. 119–181.
[3] Hiai F, Petz D. The Golden–Thompson trace inequality is complemented. Linear Algebra Appl.
1993;181:153–185.
[4] Pusz W, Woronowicz SL. Functional calculus for sesquilinear forms and the purification map.
Rep. Math. Phys. 1975;8:159–170.
[5] Lieb EH, Thirring W. Inequalities for the moments of the eigenvalues of the Schrödinger
Hamiltonian and their relaion to Sobolev inequalities. In: Lieb EH, Simon B, Wightman AS,
editors. Studies in mathematical physics. Princeton (NJ): Princeton University Press; 1976. p.
269–303.
 Linear and Multilinear Algebra 15

[6] Araki H. On an inequality of Lieb and Thirring. Lett. Math. Phys. 1990;19:167–170.
[7] Ando T, Hiai F. Log majorization and complementary Golden–Thompson type inequalities.
Linear Algebra Appl. 1994;197:113–131.
[8] Audenaert KMR, Datta N. α-z-Rényi relative entropies. J. Math. Phys. 2015;56:022202.
[9] Bhatia R. Matrix analysis. New York (NY): Springer; 1996.
[10] Ferrer J, García MI, Puerta F. Differentiable families of subspaces. Linear Algebra Appl.
1994;199:229–252.
[11] Marshall AW, Olkin I, Arnold BC. Inequalities: theory of majorization and its applications. 2nd
ed. New York (NY): Springer; 2011.
[12] Hiai F. Matrix analysis: matrix monotone functions, matrix means, and majorization.
Interdisciplinary Inf. Sci. 2010;16:139–248.
[13] Bourin J-C. Convexity or concavity inequalities for Hermitian operators. Math. Ineq. Appl.
2004;7:607–620.
[14] Kato T. Spectral order and a matrix limit theorem. Linear Multilinear Algebra. 1979;8:15–19.
Downloaded by [Florida State University] at 08:17 13 September 2015

[15] Ando T. Majorizations, doubly stochastic matrices, and comparison of eigenvalues. Linear
Algebra Appl. 1989;118:163–248.
[16] Moakher M. A differential geometric approach to the geometric mean of symmetric positive-
definite matrices. SIAM J. Matrix Anal. Appl. 2005;26:735–747.
[17] Bhatia R. Positive definite matrices. Princeton (NJ): Princeton University Press; 2007.

Appendix 1. Proof of Lemma 2.4

We may assume that H = Cd by fixing an orthonormal basis of H. Let G(k, d) denote the Grass-
mannian manifold consisting of k-dimensional subspaces of H. Let Ok,d denote the set of all
u = (u 1 , . . . , u k ) ∈ Hk such that u 1 , . . . , u k are orthonormal in H. Consider Ok,d as a metric
space with the metric

k 1/2
d2 (u, v) :=
u i − vi
2 , u = (u 1 , . . . , u k ), v = (v1 , . . . , vk ) ∈ Hk .
i=1

k,d be the set of projectivized vectors u = u 1 ∧ . . . ∧u k in H∧k of norm 1, i.e. the

Moreover, let H
quotient space of Hk,d := {u ∈ H∧k : u = u 1 ∧ · · · ∧ u k ,
u
= 1} under the equivalent relation
u ∼ v on Hk,d defined as u = eiθ v for some θ ∈ R. We then have the commutative diagram:

where π and 
π are surjective maps defined for u = (u 1 , . . . , u k ) ∈ Ok,d as

π(u) := span{u 1 , . . . , u k },
π (u) := [u 1 ∧ . . . ∧u k ], the equivalence class ofu 1 ∧ . . . ∧u k ,


and φ is the canonical representation of G(k, d) by the kth antisymmetric tensors (or the kth exterior
products).
As shown in [10], the standard Grassmannian topology on G(k, d) is the final topology (the
quotient topology) from the map π and it coincides with the topology induced by the gap metric:

dgap (U, V) :=
PU − PV

 16 K.M.R. Audenaert and F. Hiai

for k-dimensional subspaces U, V of H and the orthogonal projections PU , PV onto them. On the
k,d induced from the norm on Hk,d ⊂ H∧k , which
other hand, consider the quotient topology on H
is determined by the metric
√
d( π (v)) := inf
u 1 ∧ · · · ∧u k − e −1θ v1 ∧ · · · ∧vk
,
 π (u),  u, v ∈ Ok,d .
θ∈R

It is easy to prove that 

π : (Ok,d , d2 ) → (H k,d , d)
 is continuous. Since (Ok,d , d2 ) is compact, it
thus follows that the final topology on H k,d from the map  
π coincides with the d-topology.
It is clear from the above commutative diagram that the final topology on G(k, d) from π is
homeomorphic via φ to that on H k,d from π . Hence φ is a homeomorphism from (G(k, d), dgap ) onto
(Hk,d , d).
 From the homogeneity of (G(k, d), dgap ) and (H k,d , d)
 under the unitary transformations
there exist constant α, β > 0 (depending on only k, d) such that
 π (u), 
α
Pπ(u) − Pπ(v)
≤ d( π (v)) ≤ β
Pπ(u) − Pπ(v)
, u, v ∈ Ok,d ,
Downloaded by [Florida State University] at 08:17 13 September 2015

which is the desired inequality.