Some interesting results for the Neo3 presented by Linkwitz' recent large signal tests: http://linkwitzlab.com/frontiers_6.htm#Y -
This is very interesting. The plot literally shows what I called for a long time as memory effect. What SL shows is the heat up process. If he continue with a signal with far less strength (so it really reflect the high dynamic range we have in music) and the transducer begins to cool down, he would see the reverse process of "compression". In real world, the music notes are arranged in a non-repetitive manner, so these compression and reverse compression goes up and down in unpredictable way. But it is absolutely not random. Actually it is correlated to the input signal and the correlation cannot be better described than as memory effect.
BTW, this is exact what the servo subwoofer resolves -- no thermal compression and reverse compression, nor thermal memory effect. I wrote to SL some years ago (2006 or earlier) and was very enthusiastic about it and only received a reply to the effect it is a secondary effect
As one can imagine, when you can see anomalies in waveform, the distortion level is already pretty high. I have come up a way to even detect a small variation in amplitude modulation due to memory effect. The idea is to use burst tone (say of waveform length x), and record a single response of length 16x or even longer. Then time shift it to multiples of x to construct multiple delayed versions of response and add them up. It reconstructs the response back to a pure sine wave. We can then do FFT, remove the fundamental component, and then observe the distortion between time 0 and x. The same distortion will also repeat between x and 2x, 2x to 3x, ...etc as artifacts. The distortion contains more than just normal harmonic distortion. It has a higher resolution than what SL has used.
This basically uses the principle of superimposition. If a system is linear and time invariant, the reconstructed waveform should be pure single tone sine. The selection of x and the response measurement window should be carefully selected so that it captures the complete transient response at the beginning and end of the burst tone.