Diffusion limits of the random walk Metropolis algorithm in high dimensions (Q433896)

From MaRDI portal
scientific article
Language Label Description Also known as
English
Diffusion limits of the random walk Metropolis algorithm in high dimensions
scientific article

    Statements

    Diffusion limits of the random walk Metropolis algorithm in high dimensions (English)
    0 references
    0 references
    0 references
    0 references
    8 July 2012
    0 references
    The authors establish a diffusion limit for a random walk Metropolis-Hastings (RWM) algorithm approximating samples from a measure on an infinite-dimensional space, more precisely, for the RWM algorithm applied to distribution families obtained as finite-dimensional approximations to a measure on an separable Hilbert space. In general, diffusion limits are used as a tool to measure the complexity of MCMC methods which have been applied to high-dimensional target measures with product structure. In that case, the individual components of the Markov chain satisfy an invariance principle with respect to a scalar stochastic differential equation. The present paper extends this approach to more general high-dimensional target measures naturally arising in applications. The established invariance principle is such that the entire Markov chain converges to an infinite-dimensional continuous-time stochastic process given by a Hilbert-valued stochastic differential equation, i.e., a stochastic partial differential equations (SPDE). In more detail, the authors consider target measures \(\pi\) on a real separable Hilbert space \(H\) which possess a Radon-Nikodym derivative with respect to a Gaussian measure \(\pi_0\) on \(H\) of the form \[ \frac{d\pi}{d\pi_0}=M_\Psi\exp(-\Psi(x)) \] for some real, measurable functional \(\Psi\). Then, a realisable implementation (necessarily in finite dimensions) of the RWM algorithm for \(\pi\) is obtained by applying an RWM algorithm for a projection \(\pi^N\) of \(\pi\) on an \(N\)-dimensional subspace of \(H\). The main result of the paper is that started in stationarity the piecewise constant linear interpolants of the Markov chains in the \(N\)-dimensional subspaces converge for \(N\to\infty\) weakly to the solution of an SPDE in a suitable subspace of \(H\) started at \(\pi\) which is the invariant measure for this SPDE. This is proved under a list of assumptions on the function \(\Psi\) and the trace class covariance operator of the Gaussian measure \(\pi_0\) which according to the authors are satisfied in many applications. The practical implications of this result are that at stationarity the work to explore that state space scales as \(\mathcal{O}(N)\) and the speed at which the invariant measure is explored can be maximized by tuning the acceptance rate probability to \(0.234\) analogous to the known case for measures of product structure.
    0 references
    Markov chain Monte Carlo
    0 references
    random walk Metropolis-Hastings
    0 references
    difffusion limit
    0 references
    optimal convergence
    0 references
    scaling limits
    0 references
    convergence time
    0 references
    stochastic partial differential equations
    0 references

    Identifiers

    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references
    0 references