Optimal scaling and diffusion limits for the Langevin algorithm in high dimensions (Q1931320): Difference between revisions

The authors consider the Metropolis-adjusted algorithm, in short MALA, and studies its efficiency on a natural class of target measures supported on an infinite-dimensional Hilbert space. After an introduction (Section 1), in Section 2, they state the main theorem (Theorem 2.6) of the paper, having defined precisely the setting in which it holds. Thus, this theorem shows that the quantity \(h(l)\) is the asymptotic speed function of the limiting diffusion obtained by resealing the Metropolis-Hastings Markov chain \(x^N= \{x^{k,N}\}_{k\geq 0}\). To this end, the vector space spanned by the first \(N\) eigenfunctions of the covariance operator is introduced. It can be observed that \(X^N\subset{\mathcal H}^r\) for any \(r\in[0;+\infty)\). Then, in Section 3, the proof of the main theorem is given, postponing the proof of a number of key technical estimates to Section 4. Finally, in Section 5, the authors conclude by summarizing and providing the outlook for further research in this area. This is a very good technical and instructive paper.