SFA is a technique that grabs the slow parts of input signals, which are basically the components with the biggest one-lag autocorrelation. When mixing things up in a nonlinear way, we can lose the original signal’s time structure and correlation due to other sources getting involved. But don’t fret! The slowest component with the largest autocorrelation extracted by SFA should be tied to one possible source. It’s been proven that this super-duper slow feature taken from nonlinear mixtures is either a source signal or its invertible transformation. And if you wanna go deeper, extended SFA uses SFA to get each source signal per iteration until they’ve all been nabbed.
