uf_timeexpandDesignmat() crashes with certain combinations of EEG.srate, expansion window size, & event latencies

[olafdimigen]

Concrete scenario causing the crash:

• An EEG dataset has an EEG.srate == 256
• and the script uses this window: cfgTimeexpand.timelimits = [-0.1, 0.7];
• Matlab 2023b (irrelevant)

uf_timeexpandDesignmat() then crashes for many participants when function sparse() is called (line 173 :

Xdc = sparse( round(indrow_all(~removeIdx)), round(indcol_all(~removeIdx)), val_all(~removeIdx), dimFullX(1), dimFullX(2) );

with this error:

Error using sparse
Index into matrix must be positive.

Apparent cause of problem

Row indices for the sparse matrix are computed here:

shiftvec = (1:cfg.windowlength) + cfg.timelimits(1) * EEG.srate - 1;
tmp = bsxfun(@plus, rowix, shiftvec);
indrow_all = tmp(:)';

For EEG.srate = 256 and cfg.timelimits(1) = -0.1,
cfg.timelimits(1) * EEG.srate should be -25.6, but due to floating-point precision becomes values such as:

-25.599999999999998

This generates fractional indices like: 0.23, 0.41, ...
For certain integer event latencies rowix, rowix + (-25.599999999999998) lands in (0, 1), e.g. 26 + (-25.6) ≈ 0.4.

These values are larger than zero and therefore (unfortunately) survive this exclusion:
removeIdx = indrow_all > dimFullX(1) | indrow_all <= 0;

...however, they later become zeros via round(), causing the error (non-positive index).

Proposed fix:

Replace

shiftvec = [1:cfg.windowlength] + cfg.timelimits(1)*EEG.srate - 1;

with

shiftvec = (1:cfg.windowlength) + cfg.reallims(1) - 1;

because cfg.reallims(1) is already rounded and was already created as:
cfg.reallims(1) = round(cfg.timelimits(1) * EEG.srate);

@behinger: Wanna take a second look at this to confirm?

[behinger]

looks good to me. Floats suck - we should all just limit ourselves to Int's!

[olafdimigen]

Should we reactive the development branch or should I just push minor bugfixes into "main"?