What would you expect the behavior to be if n != nfft?

A very good question, I now suspect this issue is more of a misunderstanding on my part about what should be expected from the function than any real problem with the code. A standard case of pebkec I suppose :-/

To expand on what I was trying to do - I've got samples of a non-stationary time series of varying lengths (actually the pitch of an aircraft). I've estimated the power spectrum using welch_pgram() to understand some vibration issues we're having. I don't do a lot of signal processing, so I got the windowing wrong to start with. In fact, I ended up having to implement my own my_welch() function to understand how to use the optimized and nicely abstracted version in this package.

So, what went wrong?

• It's quite likely that n != nfft, due to the way that defaults are chosen for the function, which means there will be some of zero padding on average; the amount of zero padding can change a lot as n is changed.
• In my case the signal has a long term trend, so chunks of the signal have a large DC offset, and sensible windowing is particularly important to avoid the spectrum being rubbish in the high frequencies. In particular, this seems to mean choosing a window which drops to zero (ruling out the Hamming window as a reasonable choice, and certainly ruling out the default of no window).

I'll double check this tomorrow.

It's likely that we should apply some window by default if none is specified, since, as you say, this is what other implementations do.

The corresponding code in octave and scipy uses the Hamming and Hann(ing) windows respectively. Octave has some good commentary on the effect of window choice, zero padding, etc, and also mentions removal of the mean by default, which seems like something that might matter a lot in some cases.

https://sourceforge.net/p/octave/signal/ci/default/tree/inst/pwelch.m#l197

It looks like choosing good default settings is a tricky matter.