A Time-Varying Approach for the Blind Source Separation of Linear Quadratic Mixtures

This paper evaluates nonlinear blind source separation (NLBSS) methods by interpreting the linear-quadratic (LQ) model over a time-varying local linear approximation. Two complementary approaches are explored: decomposing LQ mixtures via local linear approximations and explicitly reformulating them as time-varying system. Simulations validate the method under three scenarios: single-trial separation, Monte Carlo analysis with random matrices, and robustness tests across nonlinearity levels. Results highlight the N-EASI-R algorithm's superior performance for all quadratic terms tested. The study proposes whether combining both strategies could enhance robustness, potentially bridging theory and practical BSS applications.