ksddescent.mmd_lbfgs

ksddescent.mmd_lbfgs(x0, target_samples, bw=1, max_iter=10000, tol=1e-05, store=False)

Sampling by optimization of the MMD

This uses target samples from a base distribution and returns new samples by minimizing the maximum mean discrepancy.

Parameters

x0torch.tensor, size n_samples x n_features

initial positions

target_samplestorch.tensor, size n_samples x n_features

Samples from the target distribution

bwfloat

bandwidth of the stein kernel

max_iterint

max numer of iters

tolfloat

tolerance for L-BFGS

Returns

x: torch.tensor

The final positions

References

M.Arbel, A.Korba, A.Salim, A.Gretton. Maximum mean discrepancy gradient flow, Neurips, 2020.