A noisy data model for MaxNG

A noise-insensitive Euclidean distance function is derived for the MaxNG algorithm. This is a dependent-component-analysis source-separation algorithm based on the maximization of a nongaussianity measure, and has recently been developed for a noiseless mixture model. It is shown that, in the case of observations corrupted by signal-independent stationary Gaussian noise, the probability density function of the output process can be easily made independent of noise if it is approximated via the Parzen-windows method with Gaussian kernels. The role assumed by the aperture parameter is shown to be similar to the one of the regularization parameter in any inverse problem.