2D bimodal distribution

We aim to generate samples from a toy bi-modal distribution. We compare sampling algorithms based on gradient flows and accelerated gradient flows under Wasserstein metric, Kalman-Wasserstein metric and Stein metric. For the approximation of gradient log density term, we use a Gaussian kernel and the kernel bandwidth is selected by the BM method. We apply the restart technique for discrete-time algorithms of AIG flows.

The convergence rate of the particle system depends on the metric. For a fixed metric, samples generated by accelerated gradient flows always converge faster than the ones generated by gradient flows.