This is at least a relatively new typo: #7547

BTW, this is pretty bad, so any back porting that is possible would be great.

While we don't like mlab if we are going to keep offering it, we should probably have a test to check that it gives sensible results. @duncanmmacleod if you wanted to add a test to this PR, that'd be great (something like compute the coherence and then assert that its equal to a hard-coded coherence values). If you aren't up to it, I can easily do a second test PR.

@jklymak, I would be happy to add simple tests, at least cohere(x, x) gives an ones. But, when I try to execute the test suite from my git clone I get:

$ py.test-2.7
============================= test session starts ==============================
platform darwin -- Python 2.7.14, pytest-3.2.5, py-1.5.2, pluggy-0.4.0
rootdir: /Users/duncan/git/matplotlib-cohere-typo, inifile: pytest.ini
plugins: xdist-1.20.0, forked-0.2, cov-2.5.1
collected 0 items / 1 errors

==================================== ERRORS ====================================
______________________________ ERROR collecting  _______________________________
...
lib/matplotlib/testing/conftest.py:12: in pytest_configure
    matplotlib._init_tests()
lib/matplotlib/__init__.py:1462: in _init_tests
    from matplotlib import ft2font
E   ImportError: cannot import name ft2font

Is there something special I need to do to get the tests going (macOS 10.13.2, macports pytest).

OK, you may not have a development environment set up? http://matplotlib.org/devel/contributing.html
If you did, you'd have freetype I think.

I do everything via conda.

If you want to post what you think is a working test, I could test for you.

My only problem w/ cohere(x,x) is I don't think it'd have caught this bug because Pxy = Pxx. OTOH hand I think cohere(x, x.roll(2)) or something else to introduce a phase shift should make Pxy neq Pxx. That won't come out to exactly one coherence for a short time series, however, so you'd need to get the values. If you want to use random numbers, note that we use a specific seed: np.random.seed(19680801).

http://matplotlib.org/devel/testing.html
http://matplotlib.org/devel/contributing.html

thanks @jklymak, the problem with the install was that I hadn't installed mpl in developer mode, the prompt to the instructions was what I needed.

However, I don't think I'm going to be much help any time soon on a unit test, since the setup for test_mlab.py is ... complicated. I could probably write a test without using any of the cls data, but it might not site well with the parametrize loop etc. I'm probably best placed to hand off to someone else for testing.

Suggest we merge, and then prevail upon @QuLogic to give us a hand writing a test.

ping @anntzer (ahem, #7547)

oops.