See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/227258253

Symmetric Positive-Definite Matrices: From Geometry to Applications and

Visualization

Chapter · January 2006

DOI: 10.1007/3-540-31272-2_17

CITATIONS READS

118 2,384

2 authors, including:

Maher Moakher
École Nationale d'Ingénieurs de Tunis
72 PUBLICATIONS   3,219 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Riemannian Means View project

Self-contact in elastic rods View project

All content following this page was uploaded by Maher Moakher on 19 May 2014.

The user has requested enhancement of the downloaded file.

17
Symmetric Positive-Definite Matrices: From
Geometry to Applications and Visualization

Maher Moakher1 and Philipp G. Batchelor2

1
Laboratory for Mathematical and Numerical Modeling in Engineering Science,
National Engineering School at Tunis, ENIT-LAMSIN, B.P. 37, 1002 Tunis
Belvéd`
´ `ere, Tunisia
[email protected]
2
Imaging Sciences Division, 5th floor Thomas Guy House, Guy’s Hospital, King’s
College London, London SE1 9RT, UK
[email protected]

Summary. In many engineering applications that use tensor analysis, such as ten-
sor imaging, the underlying tensors have the characteristic of being positive def-
inite. It might therefore be more appropriate to use techniques specially adapted
to such tensors. We will describe the geometry and calculus on the Riemannian
symmetric space of positive-definite tensors. First, we will explain why the geom-
etry, constructed by Emile Cartan, is a natural geometry on that space. Then, we
will use this framework to present formulas for means and interpolations specific to
positive-definite tensors.

17.1 Introduction

Symmetric positive-definite (SPD) matrices arise in many physical and math-

ematical contexts. This is not surprising as the set of SPD matrices has a very
rich structure and possesses many interesting features. Recent years have seen
an increasing demand for a rigorous framework for dealing with different op-
erations on the set of SPD matrices such as regularization, interpolation, and
averaging of SPD matrices data sets.
In this work we exploit the differential geometric structure of the set of
SPD matrices and use analogy with Euclidean space to give precise definitions
for means of SPD matrices. We also discuss various anisotropy measures as
well as methods for constructing multivariate interpolations of SPD matrices.
286 M. Moakher and P.G. Batchelor

17.2 Geometry of the Space of SPD Matrices

In this section we fix notations and briefly recall some differential-geometric
properties of the space of symmetric positive-definite matrices. Further details
can be found in standard texts such as [4, 6, 13].
Let M(n) denote the space of n × n real matrices. On M(n) we have
the Frobenius inner product A, BF = tr(AT B) and the associated norm
AF = [tr(AT A)]1/2 . On M(n) we define the Euclidean metric by

dF (A, B) = A − B . (17.1)

Let GL(n) be the general linear group of all nonsingular matrices in M(n).
The exponential
∞ of a matrix A ∈ M(n) is given as usual by the power series
exp A = k=0 Ak /k!, which converges for all A ∈ M(n). The exponential is
a differentiable map from M(n) onto GL(n).
The vector space of symmetric matrices in M(n) is denoted by S(n). For
P ∈ S(n) we say that P is positive semidefinite if xT Px ≥ 0 for all x ∈ Rn .
If P is positive semidefinite and invertible we say that P is symmetric posi-
tive definite. The subset of S(n) consisting of all positive-definite matrices is
a convex cone whose interior consists of all positive-definite matrices and is
denoted by P(n). While in most applications n = 3 (or n = 2), the differ-
ent notions dealt with in this work are introduced for arbitrary n > 0. The
graphical illustrations presented here are for the interesting case n = 3.

17.2.1 Riemannian Structure of P(n)

The set P(n) is a manifold whose tangent space at any of its points P is the
space TP P(n) = {P} × S(n). The infinitesimal arclength

ds := (tr(P−1 dP)2 )1/2 = P−1/2 dPP−1/2 F (17.2)

defines a Riemannian metric on P(n). The general linear group GL(n) acts
transitively on the manifold P(n) by congruent transformations defined for
S ∈ GL(n) by [S]P = ST PS. Using properties of the trace one can easily
verify that for any curve P(t) in P(n) we have

([S]P(t))−1/2 [S]dP(t)([S]P(t))−1/2 F = (P(t))−1/2 dP(t)(P(t))−1/2 F ,

and hence, [S] is an isometry for the Riemannian metric.

The exponential of a symmetric matrix is a symmetric positive-definite
matrix. The inverse map, i.e., the principal logarithm, which we denote by Log,
of a symmetric positive-definite matrix is a symmetric matrix. The geodesic
distance between P and Q in P(n) is given by [6, p. 326]
$ %1/2

n
−1 −1
dR (P, Q) =  Log(P Q)F = 2
log λi (P Q) , (17.3)
i=1
 17 Symmetric Positive-Definite Matrices 287

where λi (P−1 Q), 1 ≤ i ≤ n are the eigenvalues of the matrix P−1 Q. Because
P−1 Q is similar to P−1/2 QP−1/2 , the eigenvalues λi (P−1 Q) are all positive
and hence (17.3) is well defined for all P and Q of P(n). The unique geodesic
joining P and Q is the curve

[0, 1]  t → P1/2 (P−1/2 QP−1/2 )t P1/2 . (17.4)

It follows from (17.3) that dR (P−1 , Q−1 ) = dR (P, Q). Hence, the inversion
P → P−1 is an involutive isometry on P(n) for this metric, and therefore,
P(n) becomes a Riemannian symmetric space. It is in fact a typical example
of a symmetric space of non-compact type as classified by E. Cartan [13]. It
is also an example of a Riemannian manifold of nonpositive curvature [6].

17.2.2 The Kullback-Leibler Divergence

In information geometry, closeness between two probability distributions p and

q on an event space Ω is usually measured by the Kullback-Leibler divergence
or ‘relative entropy’ [1],

p(x)
KL(p, q) = p(x) log dx . (17.5)
Ω q(x)

Recall that a divergence on a space X is a non-negative function J(·, ·) on

the Cartesian product space X × X which is zero only on the diagonal, i.e.,
J(x, y) ≥ 0 for all x and y in X and that J(x, y) = 0 if and only if x = y. We
mention here that a symmetrized form of this divergence has been recently
used by Lenglet et al. [7] for the segmentation of probability density fields in
the context of diffusion MRI.
The Kullback-Leibler divergence between the two zero-mean Gaussian dis-
tributions
 
1 1
p(x|P) =  exp − xT P−1 x ,
(2π)n det P 2
 
1 1 T −1
q(x|Q) =  exp − x Q x ,
(2π)n det Q 2

whose covariant matrices are P and Q gives rise to the Kullback-Leibler di-
vergence for the two SPD matrices P and Q

KL(P, Q) = tr(Q−1 P − I) − log det(Q−1 P) . (17.6)

If λi , i = 1, . . . , n denote the (positive) eigenvalues of Q−1 P then


n
KL(P, Q) = (λi − log λi − 1) . (17.7)
i=1
288 M. Moakher and P.G. Batchelor

From this expression, as x − log x − 1 ≥ 0 for all x > 0 with equality holding
only when x = 1, it becomes clear that KL(·, ·) defines a divergence on the
space of SPD matrices.
We emphasize here the fact that the Kullback-Leibler divergence (17.6)
does not define a distance on the space of positive-definite matrices as it
is neither symmetric with respect to its two arguments nor does it satisfy
the triangle inequality. Its symmetrized form KLs (P, Q) = 12 (KL(P, Q) +
KL(Q, P)) can be expressed as

KLs (P, Q) = 1
2 tr(Q−1 P + P−1 Q − 2I), (17.8)

or, in terms of the λi ’s, as

n  2
1  1
KLs (P, Q) = λi − √ . (17.9)
2 i=1 λi

By construction, the symmetrized Kullback-Leibler divergence (17.8) is in-

variant under inversion, i.e., KLs (P−1 , Q−1 ) = KLs (P, Q). It is easy to see
that it is also invariant under congruent transformations, i.e.,

KLs (P, Q) = KLs (ST PS, ST QS), for all S ∈ GL(n).

From (17.9) it becomes clear that the symmetrized Kullback-Leibler di-

vergence (17.8) behaves as the square of a distance. For comparison of
the distance measures dF (·, ·), dR (·, ·) and KLs (·, ·) in Fig. 17.1 we show
‘spheres’ centered at the identity tensor and with radii r = 0.1, 0.5 and
 (a) dF (I, P) =
1: I − PF = r, (b) dR (I, P) =  Log PF = r, and (c)
KLs (I, P) = √12 P1/2 − P−1/2 F = r, where P ∈ P(3).

(a) (b) (c)

Fig. 17.1. ‘Spheres’: Isosurfaces of the distance measure between a SPD tensor in
P(3) with eigenvalues (λ1 , λ2 , λ3 ) and the identity tensor; (a) Euclidean distance,
(b) geodesic distance, and (c) square root of the Kullback-Leibler symmetrized
divergence
 17 Symmetric Positive-Definite Matrices 289

17.3 Anisotropy Indices

In this section we are going to use the distance and divergence functions
discussed above to define properly invariant measures of anisotropy of SPD
tensors. Such measures are a useful tool for the identification of changes in
tissue structure from diffusion tensor magnetic resonance imaging (DT-MRI)
data sets, see for example [2, 10, 11].
Definition 17.3.1 We define the anisotropy index of a matrix relative to a
distance (or a divergence) to be its distance (or square root of the divergence)
to its closest isotropic matrix.
This definition guarantees that the anisotropy index is a non-negative quantity
that is zero only for isotropic matrices and that it inherits all the invariance
properties of the distance (or divergence) it is induced from.
Proposition 17.3.2 Let P be a SPD matrix with eigenvalues λ1 , . . . , λn .
• The anisotropy index relative to the Euclidean distance is
  
tr P 1
AF (P) = dF P, I = tr P2 − tr2 P
n n
2
3
3n − 1  2 2 
n
=4 λ − λi λj .
n i=1 i n
1≤i<j≤n

• The anisotropy index relative to the Riemannian distance is

 √   1
AR (P) = dR P, det PI = tr Log2 P − tr2 Log P
n

n
2
3
3n − 1  n
2 
= 4 ln2 λi − ln λi ln λj .
n i=1 n
1≤i<j≤n

• The anisotropy index relative to the symmetrized Kullback-Leibler diver-

gence is
     √
AKL (P) = dKL P, tr P/ tr P I = 2( tr P tr P−1 − n)
−1

2 2
3 3
3 3 n n
3
= 424 λi 1/λi − 2n .
i=1 i=1

Proof. By straightforward computations we first determine the positive num-

ber α such that the isotropic matrix αI is closest (relative to the distance or
divergence in question) to P, then we compute the corresponding distance or
divergence.
290 M. Moakher and P.G. Batchelor

(a) (b) (c)

Fig. 17.2. Anisotropy indices: Isosurfaces of the anisotropy index of SPD tensors in
P(3) represented in the space (λ1 , λ2 , λ3 ) of eigenvalues; (a) Fractional anisotropy,
(b) geodesic anisotropy, and (c) Kullback-Leibler anisotropy

Note that AR (·) has the same functional form in terms of the eigenvalues
as AF (·) but on a logarithmic scale. We remark that AR (·) and AKL (·) are
invariant under matrix inversion while AF (·) is not.
The range of all of the above anisotropy indices is [0, ∞). In order to
compare these indices with other anisotropy indices with range [0, 1) used in
the literature, we normalize these indices in the following way

F A(P) = AF (P)/PF ,
GA(P) = AR (P)/(1 + AR (P)) ,
KLA(P) = AKL (P)/(1 + AKL (P)) .

The normalized anisotropy index relative to the Euclidean distance coincides

with the fractional anisotropy index (FA) commonly used in the DT-MRI
community [2, 10, 11]. The geodesic anisotropy (GA) was recently introduced
in [3] and applied to a DT-MRI data set.
Figure 17.2 gives isosurfaces for the (normalized) anisotropy index1 with
respect to the Euclidean distance, Riemannian distance, and the Kullback-
Leibler divergence for SPD matrices in P(3). The fractional anisotropy is
defined for all symmetric tensors but only the parts of the isosurfaces that are
inside the positive orthant of the (λ1 , λ2 , λ3 )-space are shown. Both the geo-
desic and the Kullback-Leibler anisotropies are defined only for SPD tensors.
Their isosurfaces always stay inside the positive orthant. For the geodesic and
the Kullback-Leibler anisotropies, the limiting value 1 corresponds to singular
matrices (i.e., the boundary of the positive orthant), whereas for the fractional
anisotropy this value is attained when at least one eigenvalue is infinite. For
nearly isotropic tensors, the geodesic and the Kullback-Leibler anisotropies are
similar to the fractional anisotropy. However, for tensors relatively far form
being isotropic the behavior of the geodesic anisotroppy is similar to that of
1
From now on we will consider only the normalized anisotropy indices.
 17 Symmetric Positive-Definite Matrices 291

(a) (b) (c)

Fig. 17.3. Anisotropy indices: Contours of the anisotropy index of SPD tensors in
P(3) with eigenvalues (λ1 , λ2 , λ3 ) on an octahedral plane; (a) Fractional anisotropy,
(b) geodesic anisotropy, and (c) Kullback-Leibler anisotropy.

the Kullback-Leibler anisotropy, and both behave quite differently than the
fractional anisotropy.
The set of all symmetric tensors in S(3) with a given trace is a plane
in the (λ1 , λ2 , λ3 )-space, called an octahedral plane, which is perpendicular
to the line of isotropic tensors (or in the language of plasticity theory, the
line of hydrostatic pressure). The intersection of this plane with the positive
orthant of SPD tensors is an equilateral triangle. In Fig. 17.3, contours of
the anisotropies on the octahedral plane that passes through the identity
tensor, i.e., plane of tensors with trace equal 3, are presented. Once again, the
fractional anisotropy does not see the boundary of the set of SPD tensors: for
relatively large values of the anisotropy the contour lines go over the limiting
equilateral triangle. The contour lines for the geodesic and Kullback-Leibler
anisotropies, on the other hand, stay inside this triangle and follow it closely
for large values of the anisotropy index.
We recall that the spectral decomposition of a SPD matrix P in P(3) is
P = RDRT , where D = diag(λ1 , λ2 , λ3 ) are the eigenvalues of P and R
is an orthogonal matrix whose columns are the eigenvectors of P. The set
of (positive) eigenvalues λ1 , λ2 , λ3 and the orthogonal matrix R provide a
parametrization for the elements of P(3). It has been customary to use this
parametrization to visualize a SPD matrix P as an ellipsoid whose principal
directions are parallel to the eigenvectors of P and axes proportional to the
eigenvalues of P. Thus the methods discussed in this chapter such as averag-
ing, interpolation can also be used to perform these operations for ellipsoids.
In Fig. 17.4 we use this representation to visualize the diffusion tensors of a
brain region and we use color for representing the indicated anisotropy index.

17.4 Means

The arithmetic and geometric means, usually used to average a finite set of
positive numbers, generalize naturally to a finite set of SPD matrices. This
292 M. Moakher and P.G. Batchelor

(a) (b) (c)

Fig. 17.4. Anisotropies: Diffusion ellipsoids of a brain region colored by the FA (a),
the GA (b) and the KLA (c). (See color plates)

generalization is based on the key observation that a mean has a variational

characterization [8]. The arithmetic mean minimizes the sum of the squared
Euclidean distances to given positive numbers xk , k = 1, . . . , m

m
x̄ = arg min |x − xk |2 .
x>0
k=1
√
Likewise, the geometric mean, x̃ = x1 x2 · · · xm , minimizes the sum of the
m

squared hyperbolic distances to the given positive numbers xk


m
x̃ = arg min | log x − log xk |2 .
x>0
k=1

By analogy with the set of positive numbers we define means of symmetric

positive-definite matrices as follows:
Definition 17.4.1 We define a mean relative to a distance (or a divergence)
of a finite set of SPD matrices P1 , . . . , Pm to be the SPD matrix P that
minimizes

m
d(Pi , P)2 ,
i=1

where d(·, ·) designates the distance (or the square root of the divergence).

17.4.1 Metric-Based Means

Using the Euclidean distance (17.1) and the Riemannian distance (17.3) on
P(n) in Definition 17.4.1 we obtain the arithmetic and geometric means.

Proposition 17.4.2 Given a set of SPD matrices P1 , . . . , Pm .

• Their mean relative to the Euclidean distance (17.1) is the arithmetic mean
given by
 17 Symmetric Positive-Definite Matrices 293

1 
m
A(P1 , . . . , Pm ) = Pi . (17.10)
m i=1

• Their mean relative to the Riemannian distance (17.3) is the geometric

mean G(P1 , . . . , Pm ) which is the unique solution of the nonlinear matrix
equation
m
Log(P−1i P) = 0 . (17.11)
i=1

We note that, in general, equation (17.11) can not be solved in closed

forms. This is basically due to the non-commutative nature of matrix multi-
plication. However, the geometric mean of two matrices P1 and P2 is given
explicitly by

G(P1 , P2 ) = P1 (P−1
1 P2 )
1/2
= P2 (P−1
2 P1 )
1/2
. (17.12)

We remark that the arithmetic and geometric means are invariant under con-
gruent transformations, and that the geometric mean is invariant under inver-
sion. We refer the reader to [8] for further details on the geometric mean and a
proof of its characterization (17.11). Solution of the nonlinear matrix equation
(17.11) can be obtained numerically by different methods. For instance, one
can use Newton’s method on general Riemannian manifolds which is similar
to the classical Newton’s method on a Euclidean space but with the substi-
tution of straight lines by geodesics and vector addition by the exponential
map, see e.g., [12]. We also point out the fixed point algorithm proposed in
[9] specifically to solve (17.11).

17.4.2 Kullback-Leibler Divergence-Based Means

In the previous section we showed that arithmetic and geometric means are
defined, and arise naturally, as the unique minimizers of the sum of squared
distances from a given set of SPD matrices. The following Lemma shows that
the arithmetic and harmonic means arise as minimizers of functions defined
by the Kullback-Leibler divergence and that the geometric mean of those two
means arise as the unique minimizer of a function defined by the symmetrized
Kullback-Leibler divergence.
Lemma 17.4.3 Let Qi , i = 1, . . . , m be m given SPD matrices.
1. The function

m
A(P) := KL(Qi , P)
k=1

is minimized by A(Q1 , . . . , Qm ), i.e., the arithmetic mean of Q1 , . . . , Qm .

294 M. Moakher and P.G. Batchelor

2. The function

m
B(P) := KL(P, Qi )
k=1
m 
−1 −1
is minimized by H(Q1 , . . . , Qm ) := m i=1 Qi , i.e., the harmonic
mean of Q1 , . . . , Qm .
3. The function
m
C(P) := KLs (P, Qi )
k=1

is minimized by G(A(Q1 , . . . , Qm ), H(Q1 , . . . , Qm )), i.e., the geometric

mean of the arithmetic mean of Q1 , . . . , Qm and the harmonic mean of
Q1 , . . . , Qm .
∂ det X
Proof. With the help of the formula = (det X)(X−1 )T , we have
∂X

m
∇A(P) = mP−1 − P−1 Qi P−1 ,
i=1

m
∇B(P) = −mP−1 + Q−1
i ,
i=1
m
 
m
∇C(P) = 1
2 Q−1
i −P −1 −1
Qi P .
i=1 i=1

Equating these gradients to zero and solving for P yield the results.
The following Lemma shows that the mean based on the symmetrized
Kullback-Leibler divergence of two matrices coincides with their geometric
mean.
Lemma 17.4.4 The geometric mean satisfies the identity

G(P, Q) = G(A(P, Q), H(P, Q))

for any two matrices P, Q in P(n).

Proof. By invariance of the arithmetic, geometric and harmonic means under
congruent transformations it suffices to prove this Lemma for the case P = I.
In fact, we have

G(A(I, Q), H(I, Q)) = G( 12 (I + Q), 2(I + Q−1 )−1 )

= (I + Q−1 )−1 ((I + Q−1 )(I + Q))1/2
= (I + Q−1 )−1 (Q + 2I + Q−1 )1/2
= (I + Q−1 )−1 (Q1/2 + Q−1/2 )
= (I + Q−1 )−1 (I + Q−1 )Q1/2 = Q1/2 = G(I, Q) .
 17 Symmetric Positive-Definite Matrices 295

Before we close this section we note that weighted means can also be
defined by analogy with means of positive numbers.
Definition 17.4.5 We define a weighted mean relative to a distance (or a
divergence) of a finite set of SPD matrices P1 , . . . , Pm with (non-negative)
weights w1 , . . . , wm to be the SPD matrix P that minimizes

m
wi d(Pi , P)2 ,
i=1

where d(·, ·) designates the distance (or the square root of the divergence).
Among the applications of weighted means we can cite their use as a smoothing
filter for denoising measured SPD data, see e.g., Chap. 21 by Welk et al. In
the next section we are going to use weighted means to define interpolation
of scattered SPD data.

17.5 Interpolation

One of the emerging problems from the DT-MRI community is the interpo-
lation of scattered diffusion tensor data, see for example Chap. 18 by Pajevic
et al. and Chap. 19 by Weickert and Welk. Given the values of a symmetric
positive-definite matrix field at some points of space what is the natural way
to evaluate the tensor field at some other points? We present here several
methods of multivariate interpolation of SPD tensor data over simplicial do-
mains. These interpolation methods are analogous to multivariate Lagrange
interpolation of scalar or vector data. The main ingredients are the use of
weighted means and barycentric coordinates.

17.5.1 Univariate Interpolation

We start by discussing the univariate interpolation. Given two symmetric

positive-definite matrices P1 and P2 , as the set of SPD matrices P(n) is subset
of the Euclidean space of real matrices M(n), one can define the function

λ → P(λ) = λP1 + (1 − λ)P2 , 0≤λ≤1, (17.13)

as the linear interpolation of the two tensors P1 and P2 . Indeed, as P(n) is

an open convex cone, the tensor P(λ) is SPD for all λ ∈ [0, 1]. We remark in
passing that if λ is outside of the interval [0, 1], the matrix P(λ) can leave the
set P(n), which means that linear extrapolation of SPD matrices might not
be possible.
Another way to look at P(λ) is that it is the weighted arithmetic mean of
P1 and P2 with weights λ and 1 − λ. Note that the weights are exactly the
barycentric coordinates on the line segment (simplex of dimension one) [0, 1].
296 M. Moakher and P.G. Batchelor

Fig. 17.5. Univariate interpolation: Ellipsoidal representation of linear interpola-

tion (left) and geodesic interpolation (right) between two SPD tensors. In both cases,
the colors are based on the values of the geodesic anisotropy index. (See color plates)

This property of the linear interpolation on the line segment [0, 1], which also
holds true for simplices of higher dimension, can be generalized by replacing
the weighted arithmetic mean by other weighted means. For example, if we
use the weighted geometric mean we obtain the geodesic interpolation given
explicitly by
λ → G(λ) = P1 (P−1 λ −1
1 P2 ) = P2 (P2 P1 )
1−λ
, 0≤λ≤1. (17.14)
The geodesic interpolation naturally takes into account the positive-definite
character of the involved matrices. Furthermore, the matrix G(λ) is always in
P(n) even if λ falls outside the interval [0, 1] (extrapolation). If the matrices P1
and P2 commute then G(λ) = Pλ1 P1−λ 2 . Figure 17.5 shows diffusion ellipsoids
for linear interpolation based on (17.13) and geodesic interpolation based on
(17.14) between two SPD tensors.

17.5.2 Multivariate Interpolation

Given d + 1 SPD matrices P1 , . . . , Pd+1 that are the values of a SPD matrix
field at the d + 1 vertices of a d-dimensional simplex, the linear Lagrange
interpolation at a point with barycentric coordinates (λ1 , . . . , λd+1 ) is given
by

d+1
(λ1 , . . . , λd+1 ) → P(λ1 , . . . , λd+1 ) = λi P i , (17.15)
i=1
where the λi ’s satisfy 0 ≤ λi ≤ 1, for 1 ≤ i ≤ d + 1 and λ1 + · · · + λd+1 = 1.
Once again, P(λ1 , . . . , λd+1 ) can be seen as the weighted arithmetic mean
of the SPD matrices P1 , . . . , Pd+1 . Similar to the univariate case, we can also
define the geodesic interpolation of P1 , . . . , Pd+1 at a point with barycentric
coordinates (λ1 , . . . , λd+1 ) as the weighted geometric mean of P1 , . . . , Pd+1
with weights λ1 , . . . , λd+1 . However, unlike the univariate case, for d > 1 this
weighted geometric mean cannot be given explicitly and one has to numerically
solve the nonlinear matrix equation

d+1
λi Log(P−1
i X) = 0 . (17.16)
i=1
 17 Symmetric Positive-Definite Matrices 297

Fig. 17.6. Bivariate interpolation: Ellipsoidal representation of the interpolation

of a 3D tensor field over a 2-dimensional hexagonal region: Euclidean interpolation
(left) and geodesic interpolation (right). In both cases, the colors correspond to the
values of the geodesic anisotropy index. (See color plates)

In the case where all Pi ’s commute this equation yields the solution
λ
Pλ1 1 · · · Pd+1
d+1
.

Otherwise, (17.16) is solved numerically, e.g., by the fixed point algorithm

described in [9].
As an illustration of multivariate interpolation we visualize in Fig. 17.6 the
result of a bivariate interpolation of SPD tensors on a hexagonal region. A ten-
sor field given by its values at the vertices of the blue triangles is interpolated
based on both the Euclidean and geodesic methods.

References
1. Amari, S. (1985) Differential-Geometrical Methods in Statistics. Springer-
Verlag, Berlin Heidelberg.
2. Basser, P. and Pierpaoli, C. (1996) Microstructural and physiological features
of tissues elucidated by quantitative-diffusion-tensor MRI,
I J. Magn. Reson. B
111/3, pp. 209–219.
3. Batchelor, P. G., Moakher, M., Atkinson, D., Calamante, F., Connelly, A. (2004)
A rigorous framework for diffusion tensor analysis using Riemannian geometry.
Magn. Reson. Med., in press.
4. Helgason, S. (1978) Differential Geometry, Lie Groups, and Symmetric spaces.
Academic Press, New York.
5. Kullback, S. L. (1959) Information Theory and Statistics. Wiley, New York.
6. Lang, S. (1999) Fundamentals of Differential Geometry. Springer-Verlag, New
York.
 298 M. Moakher and P.G. Batchelor

7. Lenglet, C., Rousson, M., Deriche, R. (2004) Segmentation of 3D Probabil-

ity Density Fields by Surface Evolution: Application to Diffusion MRI, I in:
MICCAI 2004, Part I, C. Barillot, D. R. Haynor, P. Hellier, eds. Lecture Notes
in Computer Science, Vol. 3216, Springer-Verlag, Berlin, pp. 18–25.
8. Moakher, M. (2003) A differential geometric approach to the geometric mean
of symmetric positive-definite matrices. SIAM J. Matrix Anal. Appl., in press.
9. Moakher, M. (2004) On the averaging of symmetric positive-definite tensors,
submitted to: J. Elasticity.
10. Papadakis, N. G., Xing, D., Houston, G. C., Smith, J. M., Smith, M. I., James,
M. F., Parsons, A. A., Huang, C. L.-H., Hall, L. D., Carpenter, T. A. (1999)
A study of rotationally invariant and symmetric indices of diffusion anisotropy,
Magn. Reson. Imag., 17/6, pp. 881–892.
11. Pierpaoli, C. and Basser, P. J. (1996) Toward a quantitative assessment of dif-
fusion anisotropy, Magn. Reson. Med., 36/6, pp. 893–906.
12. Smith, S. T. (1993) Geometric optimization methods for adaptive filtering. Ph.D.
thesis, Harvard University, Cambridge, Massachussetts.
13. Terras, A. (1988) Harmonic Analysis on Symmetric Spaces and Applications II.
Springer-Verlag, New York.

View publication stats