tpcf jackknife covariance matrix normalization error

[duncandc]

The covariance matrices in the tpcf_jackknife() mock_observables function aren't correctly normalized.

The tpcf_jackknife() function uses the numpy.cov() function to calculate the covariance matrix which normalizes by default using 1/(N-1) where N is the number of samples. For jackknife samples, it should instead be normalized by (N-1)/N.

Thanks to @yymao for the consultation on this matter.

[aphearin]

CC @tmcclintock - you should be aware of this bug. We are fixing asap. I will re-release v0.6 of the code very soon, possibly even this week, next week for sure. Until then, after the bugfix PR, you can use the bleeding edge version by cloning master and building from source. Apologies if this bug caused you problems in your work.

[tmcclintock]

I'm pretty sure this is the line, as well as the following two, that need fixing.

According to the numpy docs you need to use either the bias argument or the ddof argument. If you use bias then (without having checked this myself...) you will reproduce what you have now if you use bias = Nsub**3 - 1, and the correct behavior is obtained if you use bias = (Nsub**3 - 1.)/Nsub**3.

Apologies for being too lazy to PR.

[aphearin]

No apologies necessary, Tom, this is super helpful. Many thanks!

[tmcclintock]

Oops it isn't Nsub**3, it is the product of all the the elements of Nsub.

[duncandc]

Right, I believe bias must be a boolean. Take a look at the current fix. I believe we are in agreement.

[tmcclintock]

Ah true. I misread the numpy docs.

[aphearin]

Resolved by PR #818