Conditional path sampling for stochastic differential equations through drift relaxation (Q646490)

The author presents an algorithm for the efficient sampling of conditional paths of stochastic differential equations (SDEs) \[ dX_t=a(X_t)dt+\sigma(X_t)dB_t, t\in[0,T]. \] To construct sample paths from this equation such that the endpoints are distributed according to the densities \(h(X_0)\) and \(g(X_T)\), respectively, we have to sample the density \[ h(X_{T_0})\prod_{i=1}^{I}p(X_{T_i}|X_{T_{i-1}})g(X_{T_i}), \] where \(0=T_0<T_1<\cdots<T_I=T\) is a partition of the interval \([0,T]\) and \(p(X_{T_i}|X_{T_{i-1}})\) is the transition probability from \(X_{T_{i-1}}\) at time \(T_{i-1}\) to \(X_{T_{i}}\) at time \(T_{i}\) . This density can be sampled using MCMC (Markov Chain Monte Carlo) sampling assuming that the transition densities \(p(X_{T_i}|X_{T_{i-1}})\) can be evaluated. However, the major issue with MCMC sampling is whether it can be performed efficiently. Instead of MCMC sampling directly from the density, the author proposes to provide the MCMC sampler of the density with a better initial condition. He introduces a system of SDEs with modified drift \[ dY_t =b(Y_t)dt + \sigma(Y_t)dB_t, \] where \(b(Y_t)\) is suitably chosen to facilitate the conditional path sampling problem. He considers the collection of \(L+1\) modified systems of SDEs \[ dY^l_t =(1-\epsilon_l)b(Y^l_t)dt +\epsilon_la(Y_t^l)dt+ \sigma(Y^l_t)dB_t, \] where \(0=\epsilon_0<\epsilon_1<\cdots<\epsilon_L=1\). For the \(l\)-th SDE in the sequence, \(p_l(Y^l_{T_i}|Y^l_{T_{i-1}})\) denotes the corresponding transition probability. With this notation \(p_L(Y^L_{T_i}|Y^L_{T_{i-1}})=p(X_{T_i}|X_{T_{i-1}})\). The proposed drift relaxation algorithm is as follows. Sample using MCMC the density \[ h(Y^0_{T_0})\prod_{i=1}^{I}p_0(Y^0_{T_i}|Y^0_{T_{i-1}})g(Y^0_{T_i}). \] For \(l=1,\dots,L\), take the last sample path from the \((l-1)\)-st level and use it as an initial condition for MCMC sampling of the density \[ h(Y^l_{T_0})\prod_{i=1}^{I}p_l(Y^l_{T_i}|Y^l_{T_{i-1}})g(Y^l_{T_i}) \] at the \(l\)-th level. Keep the last sample path at the \(L\)-th level. The levels from 0 to \(L-1\) are auxiliary and only serve the purpose of providing the sampler at level \(L\) with a better initial condition. The final sampling is performed at the \(L\)-th level that corresponds to the original SDE. The drift relaxation algorithm is similar to simulated annealing (SA), used in equilibrium statistical mechanics. The algorithm is particularly suited for filtering/smoothing applications. The author uses the drift relaxation algorithm to modify a popular filtering method called particle filter. The paper contains numerical results of the application of the modified particle filter to the standard example of filtering diffusion in a double-well potential.