compute_nmPLV applies the n:m exponent to the time axis instead of the channel axis, averages the wrong axis, and calls a deprecated helper

[Ramdam17]

Three defects in compute_nmPLV (hypyp/analyses.py:851-931). Found while auditing the duplicated einsum helpers; not related to the hypyp.sync backends.

1. The n:m phase exponentiation is applied along the time axis.

After the reshape at hypyp/analyses.py:914, phase has shape (n_epoch, n_freq, 2*n_ch, n_samp) — axis 2 is channels, axis 3 is time. But hypyp/analyses.py:923-924 slice axis 3:

phase[:, :, :, :n_ch] = phase[:, :, :, :n_ch] ** n
phase[:, :, :, n_ch:] = phase[:, :, :, n_ch:] ** m

Verified on a (2, 3, 8, 100) array with n_ch = 4:

  phase[:, :, :, :n_ch]  -> shape (2, 3, 8, 4)  : 8 channels x 4 TIME SAMPLES (of 100)
  phase[:, :, :n_ch, :]  -> shape (2, 3, 4, 100): 4 CHANNELS x 100 samples   <- intended

So the exponent n is applied to the first n_ch time samples of all channels and m to the remaining samples. Participant 1 and participant 2 are never separated, which is the entire purpose of the function. It should be phase[:, :, :n_ch, :] and phase[:, :, n_ch:, :].

2. The final average collapses the wrong axis.

hypyp/analyses.py:930:

con = np.nanmean(con, axis=1)

con has shape (n_epoch, n_freq, C, C), so axis=1 averages over frequency and returns (n_epoch, C, C). The docstring (analyses.py:876-877) promises (n_freq, 2*n_channels, 2*n_channels). compute_sync does con.swapaxes(0, 1) before its nanmean (analyses.py:556); this function does not. The returned leading axis is epochs, labelled as frequencies.

3. It calls a deprecated private helper.

hypyp/analyses.py:928 calls _multiply_conjugate, which emits a DeprecationWarning and is scheduled for removal in 1.0.0 (analyses.py:1111-1116). A public, non-deprecated function therefore emits a deprecation warning the user cannot trace to their own code, and will hard-break at 1.0.0. It should use hypyp.sync.multiply_conjugate (byte-identical implementation, hypyp/sync/base.py:132-136).

compute_nmPLV also open-codes the PLV formula inline (analyses.py:916-929), bypassing get_metric entirely, so it gets no backend acceleration.

Given defects 1 and 2, this function cannot currently produce a correct n:m PLV. Worth considering whether to fix it or deprecate it — there are no callers in the repository, tests, or tutorials.