Post-processing for Boltzmann Sampling
Hi, I was wondering if there are references I can follow to learn more about how the post-processing is done when D-Wave is used as a Boltzmann sampler. I see this documentation on post-processing where it says that the problem graph is first decomposed and then the smaller graph is sampled to match the exact distribution, conditioned on the variables outside this smaller graph. Should I see this as a blocked Gibbs sampling scheme as described in https://arxiv.org/pdf/1606.00919.pdf?
Comments
The sampling post processing procedure is equivalent to the blocked Gibbs sampling scheme if the subgraphs consist of independent variables. The sampling post processing procedure can find larger subgraphs than the blocked Gibbs sampling though, so it is slightly more powerful yet also more complex. With enough iterations both methods sample from the Boltzmann distribution.
I am not aware of any documentation that outlines the post processing in more detail, but will keep looking!
Please sign in to leave a comment.