Effects of ICA on the estimation of fractal sources

This paper presents a study of the effects of independent component analysis (ICA) on singularities in the blind source separation of fractal signals. Two equal power fractional Brownian noise signals of different fractal dimensions are synthesized, mixed using a nonsingular matrix, and separated using the extended-infomax ICA algorithm. The preservation of singularity features is measured by comparing the spectral fractal dimension of the sources to that of the identified independent components. Experiments show that ICA does not estimate the spectral dimension of the sources consistently. Neither does the spectral dimension follow the traditional Amari performance measure. ICA methods must be carefully administered if the measurement of fractality is important.