14.02.2013 Views

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Chapter 4. Neurocomput<strong>in</strong>g 64:223-234, 2005 95<br />

f1, f2<br />

2<br />

1.5<br />

1<br />

0.5<br />

-2 -1.5 -1 -0.5<br />

-0.5<br />

0.5 1 1.5 2<br />

-1<br />

-1.5<br />

-2<br />

1<br />

0.5<br />

0<br />

-0.5<br />

-1<br />

f(AS) A −1 f(AS)<br />

0.5 1 1.5<br />

0.5 1 1.5<br />

1<br />

0.5<br />

0<br />

-0.5<br />

-1<br />

1<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

0<br />

0.2 0.4 0.6 0.8<br />

0.2 0.4 0.6 0.8<br />

Fig. 1. Example of a postnonl<strong>in</strong>ear transformation us<strong>in</strong>g an absolutely degenerate<br />

matrix A and <strong>in</strong> [0, 1] 2 uniform sources S.<br />

f ◦ A(x1, 0) = �<br />

x1 + 1<br />

2π s<strong>in</strong>(πx1), 2x1 + 1 �<br />

s<strong>in</strong>(πx1)<br />

π<br />

= (1, 2) �<br />

x1 + 1 �<br />

s<strong>in</strong>(πx1)<br />

2π<br />

f ◦ A(0, x2) = (1, −2) �<br />

x2 + 1 �<br />

s<strong>in</strong>(πx2)<br />

2π<br />

f ◦ A(x1, 1) = (1, −2) + (1, 2) �<br />

x1 − 1 �<br />

s<strong>in</strong>(πx1)<br />

2π<br />

f ◦ A(1, x2) = (1, 2) + (1, −2) �<br />

x2 − 1 �<br />

s<strong>in</strong>(πx2)<br />

2π<br />

So we have constructed a situation <strong>in</strong> which two uniform sources are mixed<br />

by f ◦ A, see figure 1. They can be separated either by A −1 ◦ f −1 or by A −1<br />

alone. We have shown that the latter also preserves the boundary, although<br />

it conta<strong>in</strong>s a different postnonl<strong>in</strong>earity (namely identity) <strong>in</strong> contrast to f −1 <strong>in</strong><br />

the former model. Nonetheless this is no <strong>in</strong>determ<strong>in</strong>acy of the model itself,<br />

s<strong>in</strong>ce A −1 f(AS) is obviously not <strong>in</strong>dependent. So by look<strong>in</strong>g at the boundary<br />

alone, we sometimes cannot detect <strong>in</strong>dependence if the whole system is highly<br />

symmetric. This is the case if A is absolutely degenerate. In our example f<br />

was chosen such that the non-trivial postnonl<strong>in</strong>ear mixture looks l<strong>in</strong>ear (at<br />

the boundary), and this was possible due to the <strong>in</strong>herent symmetry <strong>in</strong> A.<br />

If we however assume that A is mix<strong>in</strong>g and not absolutely degenerate, then we<br />

will show for all fully-bounded sources S that except for scal<strong>in</strong>g <strong>in</strong>terchange<br />

between f and A no more <strong>in</strong>determ<strong>in</strong>acies than <strong>in</strong> the aff<strong>in</strong>e l<strong>in</strong>ear case exist.<br />

Note that if f is only assumed to be cont<strong>in</strong>uously differentiable, then additional<br />

<strong>in</strong>determ<strong>in</strong>acies come <strong>in</strong>to play.<br />

6<br />

1<br />

0.8<br />

0.6<br />

0.4<br />

0.2<br />

0

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!