Statistical Analysis of Sample-Size Effects in ICA
Independent component analysis (ICA) solves the blind source separation problem by evaluating higher-order statistics, e.g. by estimating fourth-order moments. While estimation errors of the kurtosis can be shown to asymptotically decay with sample size according to a square-root law, they are subject to two further effects for finite samples. Firstly, errors in the estimation of kurtosis increase with the deviation from Gaussianity. Secondly, errors in kurtosis-based ICA algorithms increase when approaching the Gaussian case. These considerations allow us to derive a strict lower bound for the sample size to achieve a given separation quality, which we study analytically for a specific family of distributions and a particular algorithm (fastICA). We further provide results from simulations that support the relevance of the analytical results.
KeywordsIndependent Component Analysis Fisher Information Independent Component Analysis Blind Source Separation Recovery Error
Unable to display preview. Download preview PDF.
- 2.Amari, S., Cichocki, A., Yang, H.H.: A new learning algorithm for blind signal separation. Advances in Neural Information Processing Systems 8, 757–763 (1996)Google Scholar
- 6.Dodel, S., Herrmann, J.M., Geisel, T.: Comparison of temporal and spatial ica in fmri data analysis. In: ICA 2000. Proc. Second Int. Workshop on Independent Component Analysis and Blind Signal Separation, vol. 13, pp. 543–547 (2000)Google Scholar
- 8.Hyvärinen, A., Karhunen, J., Oja, E.: Independent component analysis. John Wiley & Sons, Chichester (2001)Google Scholar