Redundant Data Elimination in Independent Component Analysis

Abstract

Independent component analysis involves a lot of data in statistical calculations. This paper studies the model by examining which part of the data is essential and which part is redundant for defining the mixing system and proposes an idea called redundant data elimination. Statistical properties change in the direction of uniform distribution as redundant data are eliminated, yet the model still holds and the solution still exists. A theoretical explanation is given of the geometrical transformation of independent sources. The above reasoning is verified by separation experiment. It is shown that this idea can also improve model match for unsymmetrical sources.