next up previous
Next: What is independence? Up: Independent Component Analysis Previous: Ambiguities of ICA

Illustration of ICA

To illustrate the ICA model in statistical terms, consider two independent components that have the following uniform distributions:

 \begin{displaymath}
p(s_i)= \begin{cases}\frac{1}{2\sqrt{3}}& \text{if } \vert s_i\vert\leq\sqrt{3}\\
0 & \text{otherwise} \end{cases}\end{displaymath} (5)

The range of values for this uniform distribution were chosen so as to make the mean zero and the variance equal to one, as was agreed in the previous Section. The joint density of s1 and s2 is then uniform on a square. This follows from the basic definition that the joint density of two independent variables is just the product of their marginal densities (see Eq. 10): we need to simply compute the product. The joint density is illustrated in Figure 5 by showing data points randomly drawn from this distribution.


  
Figure 5: The joint distribution of the independent components s1and s2 with uniform distributions. Horizontal axis: s1, vertical axis: s2.
\resizebox{.50\textwidth}{!}{
\includegraphics{Suni}}


  
Figure 6: The joint distribution of the observed mixtures x1 and x2. Horizontal axis: x1, vertical axis: x2.
\resizebox{.50\textwidth}{!}{
\includegraphics{Xuni}}

Now let as mix these two independent components. Let us take the following mixing matrix:

\begin{displaymath}{\bf A}_0=\begin{pmatrix}2 & 3\\ 2 & 1 \end{pmatrix}\end{displaymath} (6)

This gives us two mixed variables, x1 and x2. It is easily computed that the mixed data has a uniform distribution on a parallelogram, as shown in Figure 6. Note that the random variables x1 and x2 are not independent any more; an easy way to see this is to consider, whether it is possible to predict the value of one of them, say x2, from the value of the other. Clearly if x1 attains one of its maximum or minimum values, then this completely determines the value of x2. They are therefore not independent. (For variables s1 and s2 the situation is different: from Fig. 5 it can be seen that knowing the value of s1 does not in any way help in guessing the value of s2.)

The problem of estimating the data model of ICA is now to estimate the mixing matrix ${\bf A}_0$ using only information contained in the mixtures x1 and x2. Actually, you can see from Figure 6 an intuitive way of estimating ${\bf A}$: The edges of the parallelogram are in the directions of the columns of ${\bf A}$. This means that we could, in principle, estimate the ICA model by first estimating the joint density of x1 and x2, and then locating the edges. So, the problem seems to have a solution.

In reality, however, this would be a very poor method because it only works with variables that have exactly uniform distributions. Moreover, it would be computationally quite complicated. What we need is a method that works for any distributions of the independent components, and works fast and reliably.

Next we shall consider the exact definition of independence before starting to develop methods for estimation of the ICA model.


next up previous
Next: What is independence? Up: Independent Component Analysis Previous: Ambiguities of ICA
Aapo Hyvarinen
2000-04-19