Weird unpredictable behaviour about pyscatwave HOT 3 CLOSED

Can you try with larger signals? If this is still the case, then there is indeed something missing/a bug. I guess you're running into round-off errors, as I did not specifically intend to support very small signals (better to pad them), and there might be some dependancies in the length that lead to what you do observe.

from pyscatwave.

Round-off errors are violent!
You are right, I tried other power of two sizes (20 run, taking the max of standard deviation over coordinates), and 4 is the only one for which I get absurd results. From 8 onwards, I get standard deviation of the order of 10^{-7}.
(Of course, nothing for 2, since your code works only if the size can be divided by 4, but this sounds like a reasonable requirement).

A+ !

from pyscatwave.

Well, all of this is due to conditioning & mantissa: with single floats on GPUs, your precision is about 10^-6 (which is around the normalized std you reported), and I guess a very small number like that occurs, then inverting such numbers is pretty much ill-conditioned! Indeed, there are many improvements to do, feel free to push...

from pyscatwave.