Reversible density-ratio propagation (source code)

= Reversible density-ratio propagation
{title2=$d(K^t\nu)/d\pi=K^t(d\nu/d\pi)$}

Let $K$ be a reversible <Markov kernel> with stationary <Probability measure> $\pi$, and let the initial <Probability measure> $\nu$ have <Radon-Nikodym derivative> $f=d\nu/d\pi$. After $t$ steps, <Detailed balance> gives
$$
\frac{d(\nu K^t)}{d\pi}(x)=(K^tf)(x)=\int K^t(x,dy)f(y)
$$
for $\pi$-almost every $x$. To prove the identity, integrate a bounded measurable test <function> $g$ against the evolved measure: $\int f K^tg\,d\pi=\int g K^tf\,d\pi$, because reversibility makes $K^t$ <self-adjoint operator>[self-adjoint] in the <integral> pairing.

To express the result with ordinary <probability density functions>, fix a reference measure $m$, write $\pi(dx)=h(x)m(dx)/Z$, and let $r=d\nu/dm$. On the support where $h>0$, $f=Zr/h$, so the evolved density is
$$
q_t(x)=h(x)\,\mathbb E_{Y\sim K^t(x,\cdot)}\left[\frac{r(Y)}{h(Y)}\right].
$$
This formula uses the initial density $r$, not the <Probability measure> $\nu$ evaluated at a point. The unknown <normalizing constant> $Z$ cancels. Reverse simulation from $x$ can estimate this expression without evaluating the <Transition density>, provided the displayed integrand can be evaluated.

For example, the symmetric two-state kernel with holding <Probability> $3/4$ has <Stationary distribution> $(1/2,1/2)$. Starting at state zero gives $f=(2,0)$, then $Kf=(3/2,1/2)$; multiplying by the stationary masses recovers the one-step distribution $(3/4,1/4)$.