How to derive the symmetry condition for the stretch move in emcee / affine invariant ensemble MCMC samplers?

117 Views Asked by At

The emcee library implements the stretch move proposal in line with the paper by Goodman and Weare. More specifically, for two given walkers $X_k$ and $X_i$, the proposal is $$ X_k \to Y = X_i + Z(X_k - X_i). $$ Here $Z$ is a 1D random variable with pdf $g(z)$. The paper above then proceeds demanding the proposal be symmetric and concludes without proof that this implies that the pdf satisfies the symmetry condition $$ g(1/z) = z g(z). $$

How do we proof this result?

I understand that the inverse proposal is $$ Y \to X_k = X_i + Z^{-1}(Y - X_i). $$ Then, demanding symmetry, the random variable $Z^{-1}$ should have the same distribution as $Z$, which would imply that $$ g(1/z) = z^2 g(z), $$ see for example here.

Am I missing something?

1

There are 1 best solutions below

0
On BEST ANSWER

With the help of the emcee-user mailing group this question got cleared out, see here for the discussion. For future reference, their answer is copied here.

It turns out that both the conditions $g(1/z) = z g(z)$ and $g(1/z) = z^2 g(z)$ are correct if accompanied with a suitable acceptance ratio in the Metropolis-Hastings algorithm. In fact, any condition of the form $$ g(1/z) = z^m g(z) $$ can be chosen if we take for the acceptance probability $$ A(X_k \to Y) = \min\big(1, Z^{m + n - 2} \pi(Y) / \pi(X_k)\big), \tag{1}\label{eq1} $$ where $Z$ is a 1D random variable with pdf $g(z)$ and $n$ is the dimension of the parameter space. To see this, let the proposal distribution in $X$-space be $P(X_k \to Y)$. Since proposals are governed by the scale factor $Z \sim g(z)$ in the stretch move $$ X_k \to Y = X_i + Z(X_k−X_i), $$ the proposal probability to end up in small volume around $Y$ satisfies $$ P(X_k \to Y) r^{n-1} \,dr \,d\Omega_{n-1} = g(z) dz, $$ using spherical coordinates centered around $X_i$ and letting $r = z|X_k - X_i|$. Further simplifying this equation gives $$ P(X_k \to Y) z^{n-1} |X_k - X_i|^n \,d\Omega_{n-1} = g(z). $$ For the opposite move $Y \to X_k$ one obtains $$ P(Y \to X_k) z^{-(n-1)} |Y - X_i|^n \,d\Omega_{n-1} = g(1/z). $$ Lastly, substituting these relations in the Metropolis acceptance ratio $$ A(X_k \to Y) = \min\bigg(1, \frac{P(Y \to X_k) \pi(Y)}{ P(X_k \to Y) \pi(X_k)}\bigg) $$ reproduces \eqref{eq1}.

The Goodman and Weare paper chooses $m = 1$, in which case the factor in the acceptance ratio is $Z^{n-1}$. In a 1D parameter space this reduces to one, and the acceptance probability is similar in form to when a symmetric proposal $P(X_k \to Y)$ would be used.

As already pointed out in the question, a choice of $m = 2$ corresponds to a symmetric distribution for $Z$ in the sense that the probabilities for $Z$ and $Z^{-1}$ are the same.

Lastly, the choice $m = 2 - n$ implies that the factor in the acceptance ratio is one in all dimensions, and could also be considered symmetric.