Effect of signals´ probabilistic distributions on performance of adaptive noise canceling algorithms

In adaptive noise canceling applications, the Least Mean Square (LMS) algorithm has been widely used due to its theoretical and implementation simplicities. Recently, Independent Component Analysis (ICA)-based algorithms are applied in speech or echo cancellation applications. Utilizing higher order statistics, ICA achieves better performance than the conventional LMS in these applications. This paper studies the performance of the two adaptive noise cancellation approaches with different signals´ probabilistic distributions. Our research indicates that the ICA-based approach works better for super-Gaussian signals, while LMS-based method is preferable for sub-Gaussian signals. Therefore, an appropriate choice between the LMS- and ICA- based approaches can be made if prior information about the signal´s probabilistic distribution is available.