... deconvolution¹

Often the term 'blind equalization' is used in the same sense.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... any²

The functions must be assumed measurable. We shall, however, omit any questions of measurability in this paper.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... voices³

This application is to be taken rather as an illustrative example than a real application. In practice, the situation is much more complicated than described here due to echos and, above all, time delays, see Section 7.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...[50,56]⁴

Note that it is possible, at least in theory, to consider the robustness of estimators even in the case of the ICA model without any noise. This is because the distribution of ${\bf s}$ may be $\epsilon$-contaminated, or contain outliers, even if ${\bf x}$ were generated from ${\bf s}$ exactly according to the noise-free model. A more realistic analysis of robustness would require, however, the introduction of noise in the model.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... optimal⁵

Here we consider robustness to be one form of optimality, in particular, minimax-optimality in the neighborhood of the assumed model in a space of statistical models [56].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.