Suppose at step $t$ the particle approximation of SMC in $d$ dimensions is given by $\sum_{k=1}^N w_k\delta(\vec{x}-\vec{x}_k)$, and at the subsequent step, $t+1$ (after using Bayes' law to update with a new data point), the particle approximation is given by $\sum_{k=1}^N w_k'\delta(\vec{x}-\vec{x}_k)$. Suppose also, to be concrete, that both of these distributions are good at approximating multivariate normal distributions with parameters $(\vec\mu,\Sigma)$ and $(\vec\mu',\Sigma')$, respectively.

I can think of two ways of defining the KL divergence:

  1. The probability of particle $\vec{x}_k$ is $w_k$ (or $w_k'$), so use the formula for discrete distributions, $D_{KL}=\sum_{k=1}^N w_k'\log(w_k'/w_k)$.
  2. Use the formula for KL divergence of two normal distributions, where we use estimates of the multivariate normal parameters derived from the particles (ex. $\mu=\sum_{k=1}^N w_k\vec{x}_k$): $$ D_\text{KL} = { 1 \over 2 } \left\{ \mathrm{tr} \left( \Sigma^{-1} \Sigma' \right) + \left( \mu - \mu'\right)^{\rm T} \Sigma^{-1} ( \mu - \mu' ) - d +\ln { | \Sigma | \over | \Sigma' | } \right\} $$

Would someone be kind enough to interpret both of these definitions? Moreover, in the context of adaptive experiment design, explain which one is correct (if either) and why?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.