Simple Monte Carlo and the Metropolis algorithm (Q2465298)

Integrals of the form \( \int_{\Omega}f(x) \cdot c\rho(x)\mu(dx) \) with respect to an unknown density \(c \rho(x),\) where \(c > 0,\) \(x \in \Omega,\) \(\mu\) is a probability measure and \((f, \rho)\) belongs to a class \({\mathcal F}(\Omega)\) are used in many applications. The problem is rewritten as a solution operator \( S(f, \rho) = \int_{\Omega}f(x)\rho(x) \mu(dx) / \int_{\Omega}\rho(x) \mu(dx). \) Randomized methods \(S_{n}\) that use \(n\) function evaluations of \(f\) and \(\rho\) as a non-adaptive simple Monte Carlo method and an adaptive Metropolis-Hastings method are considered. On a class \({\mathcal F}(\Omega)\) of input data the error criteria as the individual error of the problem \(e(S_{n}, (f, \rho)),\) the worst-case error on the class \({\mathcal F}(\Omega)\) \(e(S_{n}, {\mathcal F}(\omega))\) and the complexity of the problem \(e_{n}({\mathcal F}(\Omega))\) are defined and analyzed. In section 2 the concepts of the simple Monte Carlo algorithm \(S_{n}^{\text{ simple}}(f, \rho)\) and the Metropolis-Hastings algorithm \(S_{n}^{\text{ mh}}(f, \rho)\) based on using a Markov chain of points of \(\Omega\) are developed. A functional class \({\mathcal F}_{\mathcal C}(\Omega)\) and a class \({\mathcal F}^{\alpha}(\Omega)\) with log-concave densities are introduced. In section 3 the error of the integration on the class \({\mathcal F}_{\mathcal C}(\Omega)\) is analyzed. In Theorem 1 a lower bound of the worst-case error \(e(S_{n}, {\mathcal F}_{\mathcal C}(\Omega))\) for all adaptive or non-adaptive methods \(S_{n}\) is obtained. Theorem 2 provides an upper bound of the worst-case error \(e(S_{n}^{\text{ simple}}, {\mathcal F}_{\mathcal C}(\Omega))\) of the simple Monte Carlo method. For sufficiently big \(n\) both estimations have the same order \(n^{- {1 \over 2}}.\) In section 4 the error of the integration on the class \({\mathcal F}^{\alpha}(\Omega)\) is analyzed. Restrictions on the input data, in particular on the density, in order to improve the complexity are proposed. In Theorem 3 a lower bound for all non-adaptive methods, in particular for the simple Monte Carlo method, of the worst-case error \(e(S_{n}, {\mathcal F}^{\alpha}(\Omega))\) on the class \({\mathcal F}^{\alpha}(\Omega))\) is proved. The procedure, called ``ball-walk-step'', for a construction of a Markov chain \((K, \pi)\) with an invariant distribution \(\pi\) is recalled. The properties of this procedure as reversibility, uniform ergodicity, local conductance and conductance are explained. Theorem 4 gives a lower bound for the conductance of the Metropolis chain, constructed by the ball-walk-step method. Here the density \(\rho\) belongs to the class of log-concave weights on \(\Omega.\) This lower bound is exposed in terms of the local conductance. Theorem 5 gives an upper bound of the individual error of the problem of the Metropolis algorithm based on the Markov chain for the class \({\mathcal F}^{\alpha}(B^{d}),\) where \(B^{d}\) is the \(d\)-dimensional unit ball.