confusing warning with applying projs

Describe the bug

Using apply_proj after marking bad channels should automatically do the right thing instead of raising a confusing warning.

Steps to reproduce

import os
import numpy as np
import mne

sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = os.path.join(sample_data_folder, 'MEG', 'sample',
                                    'sample_audvis_filt-0-40_raw.fif')
raw = mne.io.read_raw_fif(sample_data_raw_file)
raw = raw.pick_types(meg='mag', eeg=False)

print(raw.info['projs'])

raw.info['bads'] = raw.info['ch_names'][:16]
raw2 = raw.copy()

raw.apply_proj()
raw2.info.normalize_proj()
raw2.apply_proj()
np.testing.assert_allclose(raw.get_data(), raw2.get_data())

As per the warning, normalize_proj should be applied but as you can see, it makes no difference except getting rid of the warning.

Expected results

There should be no warning

Actual results

We see a warning:

/Users/mainak/Desktop/mne_proj_issue.py:16: RuntimeWarning: Projection vector "PCA-v1" has magnitude 0.99 (should be unity), applying projector with 86/102 of the original channels available may be dangerous, consider recomputing and adding projection vectors for channels that are eventually used. If this is intentional, consider using info.normalize_proj()
  # raw.apply_proj()
/Users/mainak/Desktop/mne_proj_issue.py:16: RuntimeWarning: Projection vector "PCA-v2" has magnitude 0.95 (should be unity), applying projector with 86/102 of the original channels available may be dangerous, consider recomputing and adding projection vectors for channels that are eventually used. If this is intentional, consider using info.normalize_proj()
  # raw.apply_proj()
/Users/mainak/Desktop/mne_proj_issue.py:16: RuntimeWarning: Projection vector "PCA-v3" has magnitude 0.85 (should be unity), applying projector with 86/102 of the original channels available may be dangerous, consider recomputing and adding projection vectors for channels that are eventually used. If this is intentional, consider using info.normalize_proj()

Additional information

Additionally, the code path is somewhat confusing to follow. There is an if inplace clause:

https://github.com/mne-tools/mne-python/blob/93a3be5d21028331a88c42ef9dfa10566bcd6d1f/mne/io/proj.py#L618-L625

which makes the function return without performing an SVD. I see that _make_projector is never called with inplace=False. Is this dead code? We discussed this with @mshamalainen and the right thing to do is:

• combine the projectors into a matrix with bad channels zeroed out
• compute SVD of this matrix and use it to make the new projector

There is nothing “dangerous” per se in doing this as far as we can see. It only makes the projection operator mathematically correct. That is: P.P = P

[larsoner]

I see that _make_projector is never called with inplace=False. Is this dead code?

One last point to clarify here – note that the default is inplace=False:

https://github.com/mne-tools/mne-python/blob/258b462f06a3a5ef4c6d8ded6ca03f519b498831/mne/io/proj.py#L542-L543

So for example this bit of code in mne.io.proj.make_projector should hit the inplace=False path:

https://github.com/mne-tools/mne-python/blob/258b462f06a3a5ef4c6d8ded6ca03f519b498831/mne/io/proj.py#L539

[jasmainak]

Also:

applying projector with 86/102 of the original channels available may be dangerous

seems to suggest that subselecting channels in projection operator is dangerous. Rather, what’s dangerous is if there was a bad channel when computing the original projector which is a different problem