Home Search Collections Journals About Contact us My IOPscience

Explicit formulae for the Bures metric

This article has been downloaded from IOPscience. Please scroll down to see the full text article.

1999 J. Phys. A: Math. Gen. 32 2663

(http://iopscience.iop.org/0305-4470/32/14/007)

View the table of contents for this issue, or go to the journal homepage for more

Download details:
IP Address: 66.194.72.152
The article was downloaded on 09/07/2013 at 16:41

Please note that terms and conditions apply.

J. Phys. A: Math. Gen. 32 (1999) 2663–2670. Printed in the UK PII: S0305-4470(99)97156-1

Explicit formulae for the Bures metric

J Dittmann†
Mathematisches Institut, Universität Leipzig, Augustusplatz 10/11, 04109 Leipzig, Germany

Received 1 September 1998

Abstract. The aim of this paper is to derive explicit formulae for the computation of the
Riemannian Bures metric g on the manifold D of (finite-dimensional) nonsingular density matrices
%. This Riemannian metric introduced by Uhlmann generalizes the Fubini–Study metric to mixed
states and is the infinitesimal version of the Bures distance. Several formulae are known for
computing the Bures metric in low dimensions. The formulae presented in this paper allow for
computing in finite dimensions without
P any diagonalization procedures. The first equations we give
j −1
% =
are, essentially, of the form gP aij Tr d% %i−1
P d% % , where aij is a matrix of invariants of
%. A further formula, g% = cij dpi ⊗ dpj + bij Tr d% %i−1 d% %j −1 , is adapted to the local
orthogonal decomposition D ≈ Rn × U(n)/Tn at generic points.

1. Introduction

In recent years many authors considered the Bures metric and the Bures distance because of
their importance in quantum statistics and for the understanding of the geometry of quantum
state spaces. Explicit computations in this area meet some technical difficulties, since, for
example, the Bures metric is defined rather implicitly. The aim of this paper is to provide
several equations for computing the Bures metric in any finite dimension using only matrix
products, determinants and traces.
Let D be the manifold of all positive, Hermitian n×n-matrices. The submanifold of
trace-one matrices is the space of so-called completely entangled mixed states of a finite-
dimensional quantum system. The tangent space T% D consists of all Hermitian n×n-matrices
and the Riemannian Bures metric on D is given by [1],
g% (X 0 , X) = 1
2
Tr X0 G X, X0 ∈ T% D (1)
where G is the (unique) solution of
%G + G% = X. (2)
It should be mentioned, that (1) also defines a metric on the manifolds Dk of rank k densities,
k < n. In this case equation (2) has solutions G for X ∈ T% Dk . This solution is not unique,
but the right-hand side of (1) is still well defined for X 0 ∈ T% Dk . However, we will deal here
with the maximal rank k = n, only. This metric appears quite naturally on the background
of purifications of mixed states and is used in quantum statistics to describe the statistical
distance of mixed states [13]. It is an extremal monotone metric, [12], and seems to be quite
distinguished for physical and mathematical reasons, see e.g. [3].
† E-mail address: [email protected]

0305-4470/99/142663+08$19.50 © 1999 IOP Publishing Ltd 2663

2664 J Dittmann

Several formulae and approaches for computing the Bures metric have been given,
e.g. [1, 4–8]. Based on the integral representation (cf [9]) of the solution of (2) the metric
takes the form [1],
Z ∞
0 1
g(X , X) = Tr X0 e−t% Xe−t% dt.
2 0
Moreover, if |αi, α = 1, 2, . . ., are eigenvectors of % with eigenvalues λα , then a simple
calculation shows that (1) yields [4],
1 X |hα|X|βi|2
g% (X, X) = .
2 α,β λα + λβ
Both formulae are not explicit in the sense that they need the knowledge of eigenvalues of %.
Instead we are looking for equations in a finite dimension similar to
1 1
g= d(det %) ⊗ d(det %) + Tr d% d% (see [1]) (3)
4 det % 2
 
1 1
= Tr d% d% + (d% − % d%)(d% − % d%) (4)
4 det %
which hold for normalized (Tr % = 1), nonsingular 2 × 2-density matrices (see also [5] for
n = 3). They do not require any diagonalization procedure. In section 2 we provide such
expressions which generalize (4) to arbitrary n < ∞ using a method of the theory of matrix
equations. In section 3 we give an equation similar to (3), but adapted to the local isometric
decomposition of the manifold D. We will not suppose the normalization Tr % = 1, this case
is included in our expressions by setting p1 := Tr % = 1 and dp1 = 0.
Notations. The following quantities depend on a positive (resp. non-negative), Hermitian
n × n-matrix %. In order to simplify the notation the dependence on % will be suppressed. By
λ1 6 · · · 6 λn , 0 6 λi , we denote the eigenvalues of % and by 3 the corresponding diagonal
matrix. % is called a generic point of D if % > 0 and all eigenvalues are different. Moreover,
j −1
V will be the Vandermonde matrix (λi ) [11]. Operators acting on matrices are denoted by
bold italic letters, in particular, if not indicated otherwise, L and R denote the left and right
multiplication by %. The Bures metric now takes the form
1 1
g = Tr d% d%. (5)
2 L+R
We set
χ (t) := det(t1 − %) = t n + k1 t n−1 + · · · + kn
k0 := 1 and ki := 0 for k > n or k < 0. Hence, ei := (−1)i ki is the elementary invariant of
degree i and (−1)n χ(t) the characteristic polynomial of %. We set pi := Tr %i = λi1 + · · · + λin ,
then the differential dpi applied to a tangent vector X yields dpi (X) = i Tr X%i−1 . Finally,
we will make use of the following matrix several times:
 
p1 p2 . . . pn
 p2 p3 . . . pn+1 
P :=  .. .. .. .. . (6)
. . . .
pn pn+1 . . . p2n−1
For instance, we have the following.
Criterion. A Hermitian n×n-matrix % > 0 is a generic point of D if det P 6= 0.
Q
Indeed, P = V T 3V . Therefore, det P = det(3) det(V )2 = det(%) i<j (λi − λj )2 .
 Explicit formulae for the Bures metric 2665

2. General formulae

From now on we suppose that % is positive. In order to calculate g% (X0 , X) one has to solve
the matrix equation (2). The matrix (resp. operator) equation EG − GF = X was intensively
studied (for a review, see [9]). Basically, it has a unique solution G if E and F have disjoint
spectra (Sylvester–Rosenblum theorem [9]). In our matrix case, E = % = −F , this is fulfilled
because % is positive. The uniqueness is also clear, not appealing to this theorem, since we may
suppose w.l.o.g. that % is diagonal. Then equation (2) reads (λi + λj )Gij = Xij ; i, j = 1, . . . n.
A further, simple but nice, tool in this theory is the use of similarities of block matrices, in our
case e.g.
     
1 −G −% X 1 G −% 0
= .
0 1 0 % 0 1 0 %
If we apply the polynomial χ as an operator function to both sides we get in the upper-right
box the identity
χ (−%)G + M = 0
where M is the upper-right box of χ applied to the above inner matrix containing X;
X
n X
i−1
M= kn−i (−%)j X%i−j −1 . (7)
i=1 j =0

But χ (−%) is invertible. This can be seen as follows. The characteristic polynomial of −%
equals (−1)n χ (−t) and the positivity of % implies that χ(t) and χ(−t) have no common
divisors. Hence, by the Euclidean algorithm there exists two polynomials p, q such that
p(t)χ (t) + q(t)χ (−t) = 1, and inserting % gives q(%)χ(−%) = 1. Therefore, the solution G
of (2) is given by
X
n X
i−1
G = −χ(−%)−1 kn−i (−%)j X%i−j −1 (8a)
i=1 j =0

or, in a more compact form,

     
−1 G χ(−%)−1 0 −% X
=− χ . (8b)
0 0 0 0 0 %
The first explicit formula we get for the Bures metric is the following.
Proposition 1.
    
1 0 0 χ(−%)−1 0 −% X
g(X0 , X) := − Tr χ . (9)
2 X0 0 0 0 0 %

The inverse of χ(−%) is again a polynomial expression in %.PTherefore, we can rewrite (8a)
using the Cayley–Hamilton theorem and G will have the form 16i,j 6n aij %i−1 X%j −1 , where
the coefficients aij are invariants of %. That means,
1 X
= aij Li−1 Rj −1 (10a)
L+R 16i,j 6n

= (Id, L, . . . , Ln−1 )A(Id, R, . . . , Rn−1 )T A := (aij ). (10b)

Of course, one can also directly see the existence of such a representation. The operator L + R
acts on a finite-dimensional space. Therefore, its inverse is a polynomial in L + R, which we
can reduce to the above form using χ (L) = Lχ (%) = 0 and similarly for R. The solution of (2)
2666 J Dittmann

given by Smith in [10] is, essentially, of this form. The formulae for the coefficients in terms
of the invariants of % that one reads off from [10] will be given at the end of this section.
The representation
P (10) is unique provided % is generic. Indeed, if there were coefficients
such that 16i,j 6n aij0 Li−1 Rj −1 = 0 then applying this operator to all vectors of a common
eigenbasis of L and R would result in V T A0 V = 0. But the Vandermonde matrix V is
nonsingular for a generic % and we would conclude aij0 = 0. Moreover, in the generic case the
matrix A is necessarily symmetric.
In order to get a compact expression for the coefficient matrix A in (10) we define the
n×n-matrix K by
 
0 1 0 ... 0
 0 0 1 ... 0 
 . .. .. .. 

K :=  . . . 
. . . (11)
 0 0 0 ... 1 
−kn −kn−1 −kn−2 . . . −k1
K carries out the reduction of powers of % by χ(%) = 0. Indeed we have
   
% 1
% 2
 % 
 . =K . 
 .   . 
. .
%n %n−1
and similarly for the reduction of powers of L and R. Thus, the multiplication of (10b) by
L + R leads to
K T A + AK = C
where
 
1 0 ... 0
0 0 ... 0
C := 
 ... .. .. 
. .
0 0 ... 0
represents the identity operator Id = L0 R0 . Note that K has the same characteristic
polynomial as %. Now we may proceed as above to find A. For this purpose we apply
χ (t) to
     
1 −A −K T C 1 A −K T 0
= .
0 1 0 K 0 1 0 K
Instead of (8) and (7) we obtain
A = −χ (−K T )−1 N
where
  X
n X
i−1
−K T C
N =χ = kn−i (−K T )j CK i−j −1
0 K 12 i=1 j =0
 
kn−1 kn−2 1
. . . k1
 −kn−2 −kn−3 . . . −1
0
=
 .. .. .. 
 = [(−1) kn+1−i−j ]i,j =1 .
..
i+1 n
(12)
. . . .
(−1)n−1 0 ... 0 0
Ti j
To see the last equation note that K CK has only one in the (i + 1, j + 1)-position and zero
otherwise; K moves the ‘1’ coming from C to the right and K T moves it down. Hence we get
the following.
 Explicit formulae for the Bures metric 2667

Proposition 2. The Bures metric equals

Xn
g = 21 aij Tr d% %i−1 d% %j −1 (13)
i,j =1

where (aij ) = −χ (−K T )−1 N, K and N are given by (12) and (11).
The probably ‘most explicit’ form of the coefficients aij is given by the following.
Proposition 3. (Smith [10]):
 
(−1)i Xn−i X
n−j
i+j +r +s
aij = (−1)r kr ks 8 (14)
2 det H r=0 s=0 2
where
 
k1 k3 ...
k2n−1
 k0 k2 k2n−2 
...
H =
 .. .. .. 
.. = [k2j −i ]i,j =1
n
. . . .
0 0 ... kn
and 8(m) = 0 if m is not an integer, and otherwise 8(m) is the cofactor in det H of the
element in the first row and mth column of H .

Remark. The determinant of H is not equal to zero, more precisely:

n(n+1) Y Y
det H = (−1) 2 λi (λi + λj ) 6= 0.
i i<j

n(n+1)
Indeed, changing the order of rows and columns yields det H = (−1) 2 det[en+1−2i+j ].
But the last determinant is just the (symmetric) Schur function of the eigenvalues of % (cf [11],
I.3) related to the partition (n, n − 1, . . . , 1) = (1, . . . , 1) + (n − 1, . . . , 0) leading to the above
product.

3. A formula adapted to D ≈ Rn × U(n)/Tn

Every % ∈ D can be diagonalized with a suitable unitary u;

% = uµ2 u∗ µ = diag(µ1 , . . . , µn ) µi ∈ R+
∗ ∗ ∗
and we have d% = 2uµ dµ u + u [u du, µ ]u . By a straightforward calculation we find
2
 
1 1
(d%) = u µ−1 dµ + ([u∗ du, µ2 ]) u∗
L % + R% Lµ2 + Rµ2
and
1 1
g% = Tr dµ dµ + Tr[ u∗ du, µ2 ] 2 ([u∗ du, µ2 ]). (15)
2 Lµ + Rµ2
Therefore, in a neighbourhood of a generic point the Riemannian manifold D locally looks like
Rn × U(n) /Tn , where Rn is equipped with the standard metric and the metric on the homogeneous
space U(n)/Tn depends on the first parameter. The tangent space at % splits into
T% D = T%|| + T%||⊥ (16)
where T%|| is the subspace of Hermitian matrices commuting with %. Its orthogonal complement
(wrt. the Bures metric) is the space of all [a, %], a is anti-Hermitian. If % is diagonal then (16)
is the decomposition into diagonal and off-diagonal Hermitian matrices.
2668 J Dittmann

From now on let % be a generic point of D. By P and P ⊥ = Id − P we denote the

(orthogonal) projectors onto the subspaces in (16).

Lemma.
X
n
P (X) = %i (P −1 )ij Tr X%j −1 (17a)
i,j =1
X
n
= (P −1 )ij %i X%j −1 (17b)
i,j =1

where P is the matrix of power invariants given by (6).

Proof. To show (17a) we use that for a generic % the powers %, %2 , . . . , %n form a basis of
the vector space of all Hermitian matrices commuting with %. Moreover, Pij = Tr %i+j −1 =
4 g% (%i , %j ) and Tr X%j −1 = 4 g% (X, %j ). Bearing this in mind (17a) is just the usual formula
for the orthogonal projection onto a subspace with a given basis;
X
P(v) = bi (hbα |bβ i)−1
ij hbj |vi.
i,j

To see (17b) let Xαβ , α, β = 1, . . . , n be a common eigenbasis of L and R, %Xαβ = λα Xαβ ,

Xαβ % = λβ Xαβ (Xαβ may not be Hermitian). Then the complex span of T%|| resp. T%||⊥ is
generated by all Xαβ with α = β resp. α 6= β. For X = Xαβ the right-hand side of (17b)
yields ηαβ Xαβ , where
X
n
j −1
ηαβ = (P −1 )ij λiα λβ = (3V P −1 V T )αβ .
i,j =1

V is the Vandermonde matrix of eigenvalues of %. But P = V T 3V implies ηαβ = δαβ . 

From (17) we now get

1 1 X n
dpj
P (X) = %i−1 (P −1 )ij (X)
L+R 2 i,j =1 j
X n  
1 1
P ⊥ (X) = aij − (P −1 )ij %i−1 X%j −1
L+R i,j =1
2

where we used Tr j X%j −1 = dpj (X) and L+1R (X) = 21 %−1 X for X ∈ T|| . The matrix (aij ) is
given by proposition 2 or 3. Inserting these equations into
 
1 1 1 ⊥
g = Tr P (d%) P (d%) + d% P (d%)
2 L+R L+R
yields the following.

Proposition 4. The decomposition g = gT|| + gT||⊥ of the Bures metric is given by

1 X n
dpi −1 dpj 1 X n
g= (P )ij + (2aij − (P −1 )ij ) Tr d% %i−1 d% %j −1 . (18)
4 i,j =1 i j 4 i,j =1
 Explicit formulae for the Bures metric 2669

4. Examples

Propositions 1–4 involve elementary and power invariants, which can be expressed by each
other (cf [11]). Rewriting the identity
Xm
mkm + pr km−r = 0 m = 1, 2, . . .
r=1
as a system of linear equations for the ei resp. the pi one obtains the relations
 
p1 1 0 ... 0
 p2 p1 2 ... 0 
1  . .. 
ei = det   .
.
..
.
..
. . 

i!  
pi−1 pi−2 pi−3 . . . i−1
pi pi−1 pi−2 . . . p1
 
e1 1 0 ... 0
 2e2 e1 1 ... 0 
 .. 

pi = det 
.. .. .. 
. . . . .
 (i−1)e ei−3 . . . 1 
i−1 ei−2
iei ei−1 ei−2 . . . e1
This identity also allows for expressing a power invariant pm , m > n, by invariants of a
degree less or equal to n. Especially the i + 1-row of our matrix P equals (pi+1 , . . . , pn+i ) =
(p1 , . . . , pn )K i .
The number of terms in propositions 1–4 rapidly increases with the dimension n. Thus
we give only certain expressions for n = 2, 3 in terms of power resp. elementary invariants.
Concerning proposition 1 we get
 0 2 −1
 Tr X (% + e1 % + e2 1) (%X − X% + e1 X) for n = 2
0
g(X , X) := 2 Tr X (% + e1 % + e2 % + e3 1)−1
1 0 3 2

×(%2 X − %X% + X%2 + e1 (%X − X%) + e2 X) for n = 3.
The following terms appear in propositions 2–4:

n = 2:
   dp1 
1 p3 −p2
gT|| = ( dp1 1 , dp2
) 1
4(p1 p3 − p22 ) 2 −p2 p1 2 dp
  2 
1 e1 e2 −2e2 de1
= ( de1 , de2 )
4e2 (e12 − 4e2 ) −2e2 e1 de2
 2 
1 e1 + e2 −e1
A=
2e1 e2 −e1 1
   
2 −2e2 e1 2 p2 − p12 p1
2A − P −1 = =
e1 (e12 − 4e2 ) e1 −2 p1 (2p2 − p12 ) p1 −2

n = 3:
det P = −p33 + 2p2 p3 p4 − p1 p42 − p22 p5 + p1 p3 p5
" #
1 p3 p5 − p42 p3 p4 − p2 p5 p2 p4 − p32
−1
P = ∗ p1 p5 − p32 p2 p3 − p1 p4
det P ∗ ∗ p1 p3 − p22
2670 J Dittmann
" #
1 e1 e22 + e12 e3 − e2 e3 −e12 e2 e1 e2 − e3
A= ∗ e13 + e3 −e12 .
2e3 (e1 e2 − e3 ) ∗ ∗ e1

Acknowledgments

I would like to thank A Uhlmann for stimulating discussions and valuable remarks. Moreover
I am grateful to P M Alberti and K Schmüdgen for advice concerning the literature on the
subject.

References

[1] Uhlmann A 1992 The metric of Bures and the geometric phase Quantum Groups and Related Topics ed R Gielerak
et al (Dordrecht: Kluwer) pp 267–74
[2] Uhlmann A 1986 Parallel transport and ‘quantum holonomy’ Rep. Math. Phys. 24 229–40
[3] Dittmann J and Uhlmann A 1998 Connections and metrics respecting purification of quantum states J. Math.
Phys. to appear
[4] Hübner M 1992 Explicit computation of the Bures distance for density matrices Phys. Lett. A 163 239–42
[5] Dittmann J 1993 On the Riemannian geometry of finite dimensional mixed states Sem. S. Lie 3 73–87
[6] Twamley J 1996 Bures and statistical distance for squeezed thermal states J. Phys. A: Math. Gen. 29 3723
[7] Slater P B 1998 Bures metrics for certain high-dimensional quantum systems Phys. Lett. A 244 35–42
[8] Slater P B 1996 Quantum Fisher–Bures information and two-level systems and a tree-level extension J. Phys.
A: Math. Gen. 29 L271–5
[9] Bhatia R and Rosenthal P 1997 How and why to solve the operator equation AX − XB = Y Bull. Lond. Math.
Soc. 29 1–21
[10] Smith R A 1966 Matrix calculations for Lyapunov quadratic forms J. Diff. Eqns 2 208–17
[11] MacDonald I G 1995 Symmetric Functions and Hall Polynomials (Oxford: Clarendon)
[12] Petz D 1996 Monotone metrics on matrix spaces Linear Algebr. Appl. 244 81–96
[13] Braunstein S L and Caves C M 1994 Statistical distance and the geometry of quantum states Phys. Rev. Lett. 72
3439–43