MAINT: signal.csd: port away from using _spectral_helper
#22460
Conversation
* This allows to mark the internal functions `_spectral_helper`, `fft_helper` and `_triage_segements` as legacy, since they are only used by the legacy functions `spectrogram` and `stft`. * Some small additions to the unit tests to achieve 100% coverage for `csd()`. * Remove csd related functions from file `_scipy_spectral_tst_shim.py` * An explanation of the `ShortTimeFFT` usage was added to the `csd()` docstr.
_spectral_helper_spectral_helper
… relationship between squared magnitude/cross spectrum and PSD/cross PSD.
… between squared magnitude/cross spectrum and PSD/cross PSD.
|
Rendered docs for commit 5b00e15: |
|
It must be frustrating to have a ready PR just sit there for more than a month. Sorry about that @DietBru ! |
Well, my available time for working on SciPy is also quite unevenly distributed (as you probably noted)—so I have the same problem of not being able to review PRs as much as I would like. First to the rationale of this PR: The class Concerning the spectral analysis functionality, I'll just discuss calculating (cross-) spectral power density (PSD) aspects here: From an algorithmic view, the signal module is not in bad shape. Using them properly is not straightforward though, since the choice of parameters can change the underlying modelling assumptions significantly. Hence, I think the main challenge lies in providing useful guidance in the documentation. To elaborate: The classical Welch algorithm view assumes a stationary stochastic process and provides little insights on choosing window and overlap. This PR illustrates (again) that the Welch algorithm is equivalent to determining the frequency marginal of a For many practical applications, it is sufficient to calculate a reasonable PSD, which allows comparing signals to each other—hence the normal heuristics suffice. On the other hand, when interpreting spectral cross-densities, the shape of the dual window isn't always negligible. Unfortunately, I am not aware of any literature, which discusses such problems for practical applications. A chapter or two in a textbook is probably more suited for discussing these kind of modeling problems than the SciPy documentation. Sorry, for not being able to provide a more useful answer here. |
|
@ev-br let me know if that's enough for you to look. If not, I can do it (and sorry for not looking/noticing this PR earlier!) |
|
Thanks @DietBru! This is a really helpful summary. I'm not claiming detailed knowledge or understanding of details here, and I'm not going to ask you for a couple of chapters worth of textbook for the SciPy docs :-). My impression is that the direction makes total sense, the PR is mostly maintenance + doc improvements with no hard backwards compat breaks and a clear path forward. That said, if you want to take a closer look @larsoner, I happily leave it all to you. |
scipy/signal/_spectral_py.py
Outdated
|
|
||
| If `noverlap` is 0, this method is equivalent to Bartlett's method | ||
| [2]_. | ||
| The ratio of the squared magnitude (``scaling='spectrum'``) divided the spectral |
|
Thanks for the input, @tylerjereddy. I addressed your remarks in commit 11c0c45. |
|
Looks like my docs-related comments were addressed and there are three core devs in favor of a merge so in it goes, thanks. |
* Remove phrase "(consult the code snippet above)." * Use `csd` instead of `csd()` everywhere in docstr. * Follow up to PR scipy#22460. [docs only]
Fixes scipy#23036 This bug was introduced in 1.16.0rc1 with PR scipy#22460.
This allows to mark the internal functions
_spectral_helper,fft_helperand_triage_segementsas legacy, since they are only used by the legacy functionsspectrogramandstftnow. Further changes:csd()._scipy_spectral_tst_shim.pyShortTimeFFTusage was added to thecsd()docstr.Related Issues and PRs:
#12544: Close as wont fix, since only
_triage_segementsis affected.#13577: Close as wont fix, since only
_spectral_helperis affected . Furthermore,ShortTimeFFTsupports "chunking".#15593: Close, since only
_triage_segementsis changed.#15587: Close as wont fix, since only
_triage_segementsis affected.#19001: Edits relevant parts of
spectral_py.pyCloses#22674 (check comment).