Nonpositive asymptotic-variance change for the observable-gradient coupling

Let π\pi be the invariant measure, let fL02(π)f\in L_0^2(\pi), and let ϕL02(π)\phi\in L_0^2(\pi) be the corresponding solution of the overdamped Poisson equation. Let αf\alpha_f be the matrix-valued coupling function defined from the normalised gradients of ff, and let Γα\Gamma_{\alpha} be the associated coupling operator. Observable-gradient conjecture.

δσF2(Γα)=Rd×Rd(ϕ(y)αf(x,y)ϕ(x))e(V(x)+V(y))dxdy0.\delta\sigma_F^2(\Gamma_{\alpha})=\int_{\mathbb{R}^d\times\mathbb{R}^d}\left(\nabla\phi(y)\cdot\alpha_f(x,y)\nabla\phi(x)\right)e^{-(V(x)+V(y))}\,\mathrm{d}x\,\mathrm{d}y\leq 0.

The claim predicts that using the observable gradient as a surrogate for the Poisson-solution gradient improves, or at least does not worsen, the asymptotic variance for small perturbations. It is motivated by an integration-by-parts identity and numerical experiments, but the source gives no proof.

Sources & referencesView supporting material

Primary source

N. Nuesken and G. A. Pavliotis, “Constructing sampling schemes via coupling: Markov semigroups and optimal transport”, arXiv:1806.11026 (2018).

Progress summary

Never refreshed

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Solutions 0

No solutions have been posted yet.