BUG: Inconsistent Welch results by API about mne-python

[larsoner]

To me it looks like "just" a difference in "n_fft" default. And psd_array_welch's default of 256 is pretty small for typical data, and compute_psd's default of 1024 is better. But it's odd that the evoked.compute_psd docs don't mention that the default is different. We should document where the defaults of compute_psd differ from the psd_array_* methods (might just be n_fft?) in compute_psd then we're okay I think. (The default of 1024 is better in my mind than 256 so don't want to change that, and doesn't seem worth it to change psd_array_welch to match by deprecation.)

[alexrockhill]

I only see the 256 default documented in psd_array_welch, I don't see 1024 mentioned anywhere although that seems to be the case from the data but maybe I'm just missing it

[drammock]

The n_fft default for compute_psd() is actually data-dependent and was inherited from the legacy plot_psd() and plot_psd_topo() functions. It is:

mne-python/mne/time_frequency/spectrum.py

Lines 283 to 284 in 43414ba

	tm = _time_mask(self.times, tmin, tmax, sfreq=self.info["sfreq"])
	method_kw["n_fft"] = min(np.sum(tm), 2048)

I'll add a docstring note about the default.

[alexrockhill]

What do you think about changing the default in mne.time_frequency.psd_array_welch to match @drammock ? I agree with @larsoner that that tends to be rather few points for how high sampling rates tend to be nowadays but I'm not overly opinionated. I am a bit opinionated that I don't particularly like how two seemingly synonymous APIs don't have the same defaults, I think ideally they'd be the same.

[alexrockhill]

So no to harmonizing?

[drammock]

So no to harmonizing?

Sorry I didn't see your comment before the PR was merged. My feeling is: the default that uses min(n_times, 2048) is a nice convenience default to avoid folks seeing scipy warnings. Folks who are using the psd_array_* interface are already doing "advanced" things (working outside of our normal MNE objects) so they might actually not appreciate a default that does a bit of magic like that (i.e., they might prefer to see those warnings). So I would want users to be able to turn it off, meaning the psd_array_welch function would need a new n_fft="auto" or so (so that passing an integer would not suppress warnings, but "auto" would adapt the value to the data size if necessary, thereby preventing the warning).

Which all feels a bit too... over-engineered? Folks doing advanced things can just write the min(n_times, 2048) or whatever into their own scripts.

[alexrockhill]

Hmm don't love it but I don't really have a better idea at the moment, I think the documentation will go a long way so thanks for that.