Identity Trick#

I think the slides from the MLSS 2018 meeting is the only place that I have encountered anyone actually explicitly mentioning this Identity trick.

Given an integral problem:

p(x)=p(x|z)p(z)dz

I can multiply by an arbitrary distribution which is equivalent to 1.

p(x)=p(x|z)p(z)q(z)q(z)dz

Then I can regroup and reweight the integral

p(x)=p(x|z)p(z)q(z)q(z)dz

This results in a different expectation that we initially had

p(x)=Eq(z)[p(x|z)p(z)q(z)]

Examples:

  • Importance Sampling

  • Manipulate Stochastic gradients

  • Derive Probability bounds

  • RL for policy corrections