This interactive element illustrates the performance of two different resampling strategies: multinomial resampling, which simply randomly draws from a weighted list of indices, and stochastic universal resampling, which draws a random offset and then samples the indices from the empirical cumulative density function at regular intervals. Choose a resampling strategy, click "Resample", and observe the following patterns:
Multinomial resampling eventually leads to a collapse of sample diversity. If we resample too often, by random chance, some particles will not be resampled. Unique particles that are not resampled are lost forever. Eventually, only one unique sample will survive.
Stochastic universal resampling (SUR) lowers the risk of losing sample diversity. Once the weights are equalized after the first resampling step, SUR's resampling strategy no longer reduces sample diversity.
Furthermore, you can click the "Rejuvenate" button to add a bit of noise to the particle's states (represented here as color). Observe how rejuvenation re-introduces sample diversity, but not that excessive rejuvenation can drown out the information contained within the ensemble in noise.