Allow different numbers of points for P and Q in MMD/HSIC

[karlnapf]

and in general! This requires to change some minor bits of the implementation. CQuadraticTimeMMD CHSIC CLinearTimeMMD (last one requires some thought as the expression is not in the reference papers)

[lambday]

@karlnapf for QuadraticTimeMMD should the statistic returned still be m*MMD when m and n are different?

[karlnapf]

Good point! I don't really know to be honest. Have to talk to @sejdino on that.

[sejdino]

Hey @lambday, in the m!=n case, it ends up being (m+n)*MMD that has a nice distribution, so this is what we should be returning. Have a look at this paper: eq (3) is the unbiased MMD, (5) is biased (ignore the square root though--we are always working with the squared MMD) and eq (10) gives the asymptotic distribution. Conveniently, the permutation test will work in exactly the same way: merge everything, split randomly into m and n, recompute. For spectral approximation, you will need to take into account ratios rho_x and rho_y like in eq. (10). Gamma approximation might be messy, so I think we should do that only for m=n for now.

BTW, I would hold off changing anything in CLinearTimeMMD (and in B-test) for now - since it's not clear how best to reconcile m!=n with the streaming facility - we will need to think about this a bit.

Also, no need to do anything in HSIC - independence testing doesn't make sense with different numbers of samples.

[lambday]

Thanks @sejdino. Going through the paper to make sure I get these things right.

[lambday]

@karlnapf @sejdino Hi, I wanted to work on this and gathered some confusion regarding the spectral approximation when m != n for quadratic time MMD. The current implementation uses the formula in this paper for approximating mMMD_b^2 as this 2 * sum_l [1/(2m) * nu_l \ z_l^2] where nu_l are the eigenvalues of centered gram matrix. I thought of using eq. 10 from A kernel two-sample test paper using rho_x and rho_y, but I am not sure what should I use as an estimator of the eigenvalues of the equation when m and n are different. Would that be 1/(m+n) times eigenvalues of the centered matrix?

[sejdino]

Precisely, these are 1/(m+n) times the eigenvalues of the centred kernel matrix on the merged samples. These can then just be plugged into the eq. (10) from "A kernel two-sample test" with rho_x=m/(m+n) and rho_y=n/(m+n).

[lambday]

@sejdino thanks a lot. I am hoping to finish this by tomorrow :)

[karlnapf]

And done! @lambday You will have a long list of resolved issues for your proposal ;)