Mathematics in Independent Component Analysis
Mathematics in Independent Component Analysis
Mathematics in Independent Component Analysis
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
124 Chapter 7. Proc. ISCAS 2005, pages 5878-5881<br />
Bl<strong>in</strong>d signal separation <strong>in</strong>to groups of dependent<br />
signals us<strong>in</strong>g jo<strong>in</strong>t block diagonalization<br />
Abstract— Multidimensional or group <strong>in</strong>dependent component<br />
analysis describes the task of transform<strong>in</strong>g a multivariate observed<br />
sensor signal such that groups of the transformed signal<br />
components are mutually <strong>in</strong>dependent - however dependencies<br />
with<strong>in</strong> the groups are still allowed. This generalization of<br />
<strong>in</strong>dependent component analysis (ICA) allows for weaken<strong>in</strong>g<br />
the sometimes too strict assumption of <strong>in</strong>dependence <strong>in</strong> ICA.<br />
It has potential applications <strong>in</strong> various fields such as ECG,<br />
fMRI analysis or convolutive ICA. Recently we could calculate<br />
the <strong>in</strong>determ<strong>in</strong>acies of group ICA, which f<strong>in</strong>ally enables us,<br />
also theoretically, to apply group ICA to solve bl<strong>in</strong>d source<br />
separation (BSS) problems. In this paper we <strong>in</strong>troduce and<br />
discuss various algorithms for separat<strong>in</strong>g signals <strong>in</strong>to groups<br />
of dependent signals. The algorithms are based on jo<strong>in</strong>t block<br />
diagonalization of sets of matrices generated us<strong>in</strong>g several signal<br />
structures.<br />
Fabian J. Theis<br />
Institute of Biophysics, University of Regensburg<br />
93040 Regensburg, Germany, Email: fabian@theis.name<br />
of commut<strong>in</strong>g symmetric n × n matrices Mi, to f<strong>in</strong>d an<br />
orthogonal matrix E such that E ⊤ MiE is diagonal for all i.<br />
In the follow<strong>in</strong>g we will use a generalization of this technique<br />
as algorithm to solve MBSS problems. Instead of fully<br />
diagonaliz<strong>in</strong>g Mi <strong>in</strong> jo<strong>in</strong>t block diagonalization (JBD) we<br />
want to determ<strong>in</strong>e E such that E ⊤ MiE is block-diagonal<br />
(after fix<strong>in</strong>g the block-structure).<br />
Introduc<strong>in</strong>g some notation, let us def<strong>in</strong>e for r, s =1,...,n<br />
the (r, s) sub-k-matrix of W =(wij), denoted by W (k)<br />
rs ,to<br />
be the k × k submatrix of W end<strong>in</strong>g at position (rk, sk).<br />
Denote Gl(n) the group of <strong>in</strong>vertible n×n matrices. A matrix<br />
W ∈ Gl(nk) is said to be a k-scal<strong>in</strong>g matrix if W (k)<br />
rs =0<br />
for r �= s, andW is called a k-permutation matrix if for<br />
each r =1,...,n there exists precisely one s such that W (k)<br />
rs<br />
equals the k × k unit matrix.<br />
Hence, fix<strong>in</strong>g the block-size to k, JBD tries to f<strong>in</strong>d E<br />
such that E ⊤ MiE is a k-scal<strong>in</strong>g matrix. In practice due<br />
to estimation errors, such E will not exist, so we speak of<br />
approximate JBD and imply m<strong>in</strong>imiz<strong>in</strong>g some error-measure<br />
on non-block-diagonality.<br />
Various algorithms to actually perform JBD have been<br />
proposed, see [5] and references there<strong>in</strong>. In the follow<strong>in</strong>g we<br />
will simply perform jo<strong>in</strong>t diagonalization (us<strong>in</strong>g for example<br />
the Jacobi-like algorithm from [6]) and then permute the<br />
columns of E to achieve block-diagonality — <strong>in</strong> experiments<br />
this turns out to be an efficient solution to JBD [5].<br />
I. INTRODUCTION<br />
In this work, we discuss multidimensional bl<strong>in</strong>d source<br />
separation (MBSS) i.e. the recovery of underly<strong>in</strong>g sources<br />
s from an observed mixture x. As usual, s has to fulfill<br />
additional properties such as <strong>in</strong>dependence or diagonality of<br />
the autocovariances (if s possesses time structure). However<br />
<strong>in</strong> contrast to ord<strong>in</strong>ary BSS, MBSS is more general as some<br />
source signals are allowed to possess common statistics. One<br />
possible solution for MBSS is multidimensional <strong>in</strong>dependent<br />
component analysis (MICA) — <strong>in</strong> section IV we will discuss<br />
other such conditions. The idea MICA is that we do not require<br />
full <strong>in</strong>dependence of the transform y := Wx but only mutual<br />
<strong>in</strong>dependence of certa<strong>in</strong> tuples yi1 ,...,yi2 . If the size of all<br />
III. MULTIDIMENSIONAL ICA (MICA)<br />
tuples is restricted to one, this reduces to ord<strong>in</strong>ary ICA. In<br />
Let k, n ∈ N. We call an nk-dimensional random vec-<br />
general, of course the tuples could have different sizes, but<br />
tor y k-<strong>in</strong>dependent if the k-dimensional random vectors<br />
for the sake of simplicity we assume that they all have the<br />
(y1,...,yk)<br />
same length k.<br />
Multidimensional ICA has first been <strong>in</strong>troduced by Cardoso<br />
[1] us<strong>in</strong>g geometrical motivations. Hyvär<strong>in</strong>en and Hoyer then<br />
presented a special case of multidimensional ICA which they<br />
called <strong>in</strong>dependent subspace analysis [2]; there the dependence<br />
with<strong>in</strong> a k-tuple is explicitly modelled enabl<strong>in</strong>g the authors<br />
to propose better algorithms without hav<strong>in</strong>g to resort to the<br />
problematic multidimensional density estimation.<br />
II. JOINT BLOCK DIAGONALIZATION<br />
Jo<strong>in</strong>t diagonalization has become an important tool <strong>in</strong> ICAbased<br />
BSS (used for example <strong>in</strong> JADE [3]) or <strong>in</strong> BSS rely<strong>in</strong>g<br />
on second-order time-decorrelation (for example <strong>in</strong> SOBI<br />
[4]). The task of (real) jo<strong>in</strong>t diagonalization is, given a set<br />
⊤ ,...,(ynk−k+1,...,ynk) ⊤ are mutually <strong>in</strong>dependent.<br />
A matrix W ∈ Gl(nk) is called a k-multidimensional<br />
ICA of an nk-dimensional random vector x if Wx is k<strong>in</strong>dependent.<br />
If k =1, this is the same as ord<strong>in</strong>ary ICA.<br />
Us<strong>in</strong>g MICA we want to solve the (noiseless) l<strong>in</strong>ear MBSS<br />
problem x = As, where the nk-dimensional random vector x<br />
is given, and A ∈ Gl(nk) and s are unknown. In the case of<br />
MICA s isassumedtobek-<strong>in</strong>dependent.<br />
A. Indeterm<strong>in</strong>acies<br />
Obvious <strong>in</strong>determ<strong>in</strong>acies are, similar to ord<strong>in</strong>ary ICA, <strong>in</strong>vertible<br />
transforms <strong>in</strong> Gl(k) <strong>in</strong> each tuple as well as the fact<br />
that the order of the <strong>in</strong>dependent k-tuples is not fixed. Indeed,<br />
if A is MBSS solution, then so is ALP with a k-scal<strong>in</strong>g<br />
matrix L and a k-permutation P, because <strong>in</strong>dependence is