arXiv subjects

Dan Crisan

Publications and source records attributed to Dan Crisan.

90 records · Page 5Linked to original sources

arXiv2017

Numerical Method for FBSDEs of McKean-Vlasov Type

This paper is dedicated to the presentation and the analysis of a numerical scheme for forward-backward SDEs of the McKean-Vlasov type, or equivalently for solutions to PDEs on the Wasserstein space. Because of the mean field structure of the equation, earlier methods for classical forward-backward systems fail. The scheme is based on a variation of the method of continuation. The principle is to implement recursively local Picard iterations on small time intervals. We establish a bound for the rate of convergence under the assumption that the decoupling field of the forward-bakward SDE (or equivalently the solution of the PDE) satisfies mild regularity conditions. We also provide numerical illustrations.

math.PR↗

arXiv2016

Particle-kernel estimation of the filter density in state-space models

Sequential Monte Carlo (SMC) methods, also known as particle filters, are simulation-based recursive algorithms for the approximation of the a posteriori probability measures generated by state-space dynamical models. At any given time $t$, a SMC method produces a set of samples over the state space of the system of interest (often termed "particles") that is used to build a discrete and random approximation of the posterior probability distribution of the state variables, conditional on a sequence of available observations. One potential application of the methodology is the estimation of the densities associated to the sequence of a posteriori distributions. While practitioners have rather freely applied such density approximations in the past, the issue has received less attention from a theoretical perspective. In this paper, we address the problem of constructing kernel-based estimates of the posterior probability density function and its derivatives, and obtain asymptotic convergence results for the estimation errors. In particular, we find convergence rates for the approximation errors that hold uniformly on the state space and guarantee that the error vanishes almost surely as the number of particles in the filter grows. Based on this uniform convergence result, we first show how to build continuous measures that converge almost surely (with known rate) toward the posterior measure and then address a few applications. The latter include maximum a posteriori estimation of the system state using the approximate derivatives of the posterior density and the approximation of functionals of it, for example, Shannon's entropy. This manuscript is identical to the published paper, including a gap in the proof of Theorem 4.2. The Theorem itself is correct. We provide an {\em erratum} at the end of this document with a complete proof and a brief discussion.

stat.CO↗

arXiv2016

Pointwise Gradient Bounds for Degenerate Semigroups (of UFG type)

In this paper we consider diffusion semigroups generated by second order differential operators of degenerate type. The operators that we consider do not, in general, satisfy the Hormander condition and are not hypoelliptic. In particular, instead of working under the Hormander paradigm, we consider the so-called UFG condition, introduced by Kusuoka and Strook in the eighties. The UFG condition is weaker than the uniform Hormander condition, the smoothing effect taking place only in certain directions (rather than in every direction, as it is the case when the Hormander condition is assumed). Under the UFG condition, Kusuoka and Strook deduced sharp small time asymptotic bounds for the derivatives of the semigroup in the directions where smoothing occurs. In this paper, we study the large time asymptotics for the gradients of the diffusion semigroup in the same set of directions and under the same UFG condition. In particular, we identify conditions under which the derivatives of the diffusion semigroup in the smoothing directions decay exponentially in time. This paper constitutes therefore a stepping stone in the analysis of the long time behaviour of diffusions which do not satisfy the Hormander condition.

math.AP↗

arXiv2016

Two-dimensional pseudo-gravity model

We analyze a simple macroscopic model describing the evolution of a cloud of particles confined in a magneto-optical trap. The behavior of the particles is mainly driven by self--consistent attractive forces. In contrast to the standard model of gravitational forces, the force field does not result from a potential; moreover, the non linear coupling is more singular than the coupling based on the Poisson equation. We establish the existence of solutions, under a suitable smallness condition on the total mass, or, equivalently, for a sufficiently large diffusion coefficient. When a symmetry assumption is fulfilled, the solutions satisfy strengthened estimates (exponential moments). We also investigate the convergence of the $N$-particles description towards the PDE system in the mean field regime.

math.AP↗

arXiv2016

Uniform convergence over time of a nested particle filtering scheme for recursive parameter estimation in state--space Markov models

We analyse the performance of a recursive Monte Carlo method for the Bayesian estimation of the static parameters of a discrete--time state--space Markov model. The algorithm employs two layers of particle filters to approximate the posterior probability distribution of the model parameters. In particular, the first layer yields an empirical distribution of samples on the parameter space, while the filters in the second layer are auxiliary devices to approximate the (analytically intractable) likelihood of the parameters. This approach relates the this algorithm to the recent sequential Monte Carlo square (SMC$^2$) method, which provides a {\em non-recursive} solution to the same problem. In this paper, we investigate the approximation, via the proposed scheme, of integrals of real bounded functions with respect to the posterior distribution of the system parameters. Under assumptions related to the compactness of the parameter support and the stability and continuity of the sequence of posterior distributions for the state--space model, we prove that the $L_p$ norms of the approximation errors vanish asymptotically (as the number of Monte Carlo samples generated by the algorithm increases) and uniformly over time. We also prove that, under the same assumptions, the proposed scheme can asymptotically identify the parameter values for a class of models. We conclude the paper with a numerical example that illustrates the uniform convergence results by exploring the accuracy and stability of the proposed algorithm operating with long sequences of observations.

stat.CO↗

arXiv2015

A simple scheme for the parallelization of particle filters and its application to the tracking of complex stochastic systems

We investigate the use of possibly the simplest scheme for the parallelisation of the standard particle filter, that consists in splitting the computational budget into $M$ fully independent particle filters with $N$ particles each, and then obtaining the desired estimators by averaging over the $M$ independent outcomes of the filters. This approach minimises the parallelisation overhead yet displays highly desirable theoretical properties. Under very mild assumptions, we analyse the mean square error (MSE) of the estimators of 1-dimensional statistics of the optimal filtering distribution and show explicitly the effect of parallelisation scheme on the convergence rate. Specifically, we study the decomposition of the MSE into variance and bias components, to show that the former decays as $\frac{1}{MN}$, i.e., linearly with the total number of particles, while the latter converges towards $0$ as $\frac{1}{N^2}$. Parallelisation, therefore, has the obvious advantage of dividing the running times while preserving the (asymptotic) performance of the particle filter. Following this lead, we propose a time-error index to compare schemes with different degrees of parallelisation. Finally, we provide two numerical examples. The first one deals with the tracking of a Lorenz 63 chaotic system with dynamical noise and partial (noisy) observations, while the second example involves a dynamical network of modified FitzHugh-Nagumo (FH-N) stochastic nodes. The latter is a large dimensional system ($\approx3,000$ state variables in our computer experiments) designed to numerically reproduce typical electrical phenomena observed in the atria of the human heart. In both examples, we show how the proposed parallelisation scheme attains the same approximation accuracy as a centralised particle filter with only a small fraction of the running time, using a standard multicore computer.

stat.CO↗

arXiv2015

A Probabilistic approach to classical solutions of the master equation for large population equilibria

We analyze a class of nonlinear partial differential equations (PDEs) defined on $\mathbb{R}^d \times \mathcal{P}_2(\mathbb{R}^d),$ where $\mathcal{P}_2(\mathbb{R}^d)$ is the Wasserstein space of probability measures on $\mathbb{R}^d$ with a finite second-order moment. We show that such equations admit a classical solutions for sufficiently small time intervals. Under additional constraints, we prove that their solution can be extended to arbitrary large intervals. These nonlinear PDEs arise in the recent developments in the theory of large population stochastic control. More precisely they are the so-called master equations corresponding to asymptotic equilibria for a large population of controlled players with mean-field interaction and subject to minimization constraints. The results in the paper are deduced by exploiting this connection. In particular, we study the differentiability with respect to the initial condition of the flow generated by a forward-backward stochastic system of McKean-Vlasov type. As a byproduct, we prove that the decoupling field generated by the forward-backward system is a classical solution of the corresponding master equation. Finally, we give several applications to mean-field games and to the control of McKean-Vlasov diffusion processes.

math.PR↗

arXiv2014

A Stable Particle Filter in High-Dimensions

We consider the numerical approximation of the filtering problem in high dimensions, that is, when the hidden state lies in $\mathbb{R}^d$ with $d$ large. For low dimensional problems, one of the most popular numerical procedures for consistent inference is the class of approximations termed particle filters or sequential Monte Carlo methods. However, in high dimensions, standard particle filters (e.g. the bootstrap particle filter) can have a cost that is exponential in $d$ for the algorithm to be stable in an appropriate sense. We develop a new particle filter, called the \emph{space-time particle filter}, for a specific family of state-space models in discrete time. This new class of particle filters provide consistent Monte Carlo estimates for any fixed $d$, as do standard particle filters. Moreover, we expect that the state-space particle filter will scale much better with $d$ than the standard filter. We illustrate this analytically for a model of a simple i.i.d. structure and one of a Markovian structure in the $d$-dimensional space-direction, when we show that the algorithm exhibits certain stability properties as $d$ increases at a cost $\mathcal{O}(nNd^2)$, where $n$ is the time parameter and $N$ is the number of Monte Carlo samples, that are fixed and independent of $d$. Similar results are expected to hold, under a more general structure than the i.i.d.~one. independently of the dimension. Our theoretical results are also supported by numerical simulations on practical models of complex structures. The results suggest that it is indeed possible to tackle some high dimensional filtering problems using the space-time particle filter that standard particle filters cannot handle.

stat.CO↗

arXiv2014

The filtering equations revisited

The problem of nonlinear filtering has engendered a surprising number of mathematical techniques for its treatment. A notable example is the change-of--probability-measure method originally introduced by Kallianpur and Striebel to derive the filtering equations and the Bayes-like formula that bears their names. More recent work, however, has generally preferred other methods. In this paper, we reconsider the change-of-measure approach to the derivation of the filtering equations and show that many of the technical conditions present in previous work can be relaxed. The filtering equations are established for general Markov signal processes that can be described by a martingale-problem formulation. Two specific applications are treated.

math.PR↗

arXiv2014

A second order time discretization of the solution of the non-linear filtering problem

The solution of the continuous time filtering problem can be represented as a ratio of two expectations of certain functionals of the signal process that are parametrized by the observation path. We introduce a new time discretisation of these functionals corresponding to a chosen partition of the time interval and show that the convergence rate of discretisation is proportional with the square of the mesh of the partition.

math.PR↗

arXiv2014

Runge-Kutta schemes for backward stochastic differential equations

We study the convergence of a class of Runge-Kutta type schemes for backward stochastic differential equations (BSDEs) in a Markovian framework. The schemes belonging to the class under consideration benefit from a certain stability property. As a consequence, the overall rate of the convergence of these schemes is controlled by their local truncation error. The schemes are categorized by the number of intermediate stages implemented between consecutive partition time instances. We show that the order of the schemes matches the number $p$ of intermediate stages for $p\le3$. Moreover, we show that the so-called order barrier occurs at $p=3$, that is, that it is not possible to construct schemes of order $p$ with $p$ stages, when $p>3$. The analysis is done under sufficient regularity on the final condition and on the coefficients of the BSDE.

math.PR↗

arXiv2014

A central limit type theorem for Gaussian mixture approximations to the nonlinear filtering problem

Approximating the solution of the nonlinear filtering problem with Gaussian mixtures has been a very popular method since the 1970s. However, the vast majority of such approximations are introduced in an ad-hoc manner without theoretical grounding. This work is a continuation of [4, 5], where we described a rigorous Gaussian mixture approximation to the solution of the filtering problem. We deduce here a refined estimate of the rate of convergence of the approximation. We do this by proving a central limit type theorem for the error process. We also find the optimal variances of the Gaussian measures are of order $1/\sqrt{n}$. This implies, in particular, that the mean square error of the approximation as defined in [4, 5] is of order $1/n$.

math.PR↗

arXiv2013

Kusuoka-Stroock gradient bounds for the solution of the filtering equation

We obtain sharp gradient bounds for perturbed diffusion semigroups. In contrast with existing results, the perturbation is here random and the bounds obtained are pathwise. Our approach builds on the classical work of Kusuoka and Stroock [7],[9],[10],[11], and extends their program developed for the heat semi-group to solutions of stochastic partial differential equations. The work is motivated by and applied to nonlinear filtering. The analysis allows us to derive pathwise gradient bounds for the un-normalised conditional distribution of a partially observed signal. It uses a pathwise representation of the perturbed semigroup in the spirit of classical work by Ocone [14]. The estimates we derive have sharp small time asymptotics.

math.PR↗

arXiv2013

Generalised Particle Filters with Gaussian Mixtures

Stochastic filtering is defined as the estimation of a partially observed dynamical system. A massive scientific and computational effort is dedicated to the development of numerical methods for approximating the solution of the filtering problem. Approximating the solution of the filtering problem with Gaussian mixtures has been a very popular method since the 1970s (see [1],[2],[46],[49]). Despite nearly fifty years of development, the existing work is based on the success of the numerical implementation and is not theoretically justified. This paper fills this gap and contains a rigorous analysis of a new Gaussian mixture approximation to the solution of the filtering problem. We deduce the L^2-convergence rate for the approximating system and show some numerical example to test the new algorithm.

math.PR↗

arXiv2012

On the Stability of Sequential Monte Carlo Methods in High Dimensions

We investigate the stability of a Sequential Monte Carlo (SMC) method applied to the problem of sampling from a target distribution on $\mathbb{R}^d$ for large $d$. It is well known that using a single importance sampling step one produces an approximation for the target that deteriorates as the dimension $d$ increases, unless the number of Monte Carlo samples $N$ increases at an exponential rate in $d$. We show that this degeneracy can be avoided by introducing a sequence of artificial targets, starting from a `simple' density and moving to the one of interest, using an SMC method to sample from the sequence. Using this class of SMC methods with a fixed number of samples, one can produce an approximation for which the effective sample size (ESS) converges to a random variable $\varepsilon_N$ as $d\rightarrow\infty$ with $1<\varepsilon_{N}<N$. The convergence is achieved with a computational cost proportional to $Nd^2$. If $\varepsilon_N\ll N$, we can raise its value by introducing a number of resampling steps, say $m$ (where $m$ is independent of $d$). In this case, ESS converges to a random variable $\varepsilon_{N,m}$ as $d\rightarrow\infty$ and $\lim_{m\to\infty}\varepsilon_{N,m}=N$. Also, we show that the Monte Carlo error for estimating a fixed dimensional marginal expectation is of order $\frac{1}{\sqrt{N}}$ uniformly in $d$. The results imply that, in high dimensions, SMC algorithms can efficiently control the variability of the importance sampling weights and estimate fixed dimensional marginals at a cost which is less than exponential in $d$ and indicate that, in high dimensions, resampling leads to a reduction in the Monte Carlo error and increase in the ESS.

stat.CO↗

arXiv2011

Error Bounds and Normalizing Constants for Sequential Monte Carlo in High Dimensions

In a recent paper Beskos et al (2011), the Sequential Monte Carlo (SMC) sampler introduced in Del Moral et al (2006), Neal (2001) has been shown to be asymptotically stable in the dimension of the state space d at a cost that is only polynomial in d, when N the number of Monte Carlo samples, is fixed. More precisely, it has been established that the effective sample size (ESS) of the ensuing (approximate) sample and the Monte Carlo error of fixed dimensional marginals will converge as $d$ grows, with a computational cost of $\mathcal{O}(Nd^2)$. In the present work, further results on SMC methods in high dimensions are provided as $d\to\infty$ and with $N$ fixed. We deduce an explicit bound on the Monte-Carlo error for estimates derived using the SMC sampler and the exact asymptotic relative $\mathbb{L}_2$-error of the estimate of the normalizing constant. We also establish marginal propagation of chaos properties of the algorithm. The accuracy in high-dimensions of some approximate SMC-based filtering schemes is also discussed.

stat.CO↗

arXiv2010

Second order discretization of Backward SDEs

In [5] the authors suggested a new algorithm for the numerical approximation of a BSDE by merging the cubature method with the first order discretization developed by [3] and [16]. Though the algorithm presented in [5] compared satisfactorily with other methods it lacked the higher order nature of the cubature method due to the use of the low order discretization. In this paper we introduce a second order discretization of the BSDE in the spirit of higher order implicit-explicit schemes for forward SDEs and predictor corrector methods.

math.PR↗

arXiv2005

Approximate McKean-Vlasov Representations for a class of SPDEs

The solution $\vartheta =(\vartheta_{t})_{t\geq 0}$ of a class of linear stochastic partial differential equations is approximated using Clark's robust representation approach (\cite{c}, \cite{cc}). The ensuing approximations are shown to coincide with the time marginals of solutions of a certain McKean-Vlasov type equation. We prove existence and uniqueness of the solution of the McKean-Vlasov equation. The result leads to a representation of $\vartheta $as a limit of empirical distributions of systems of equally weighted particles. In particular, the solution of the Zakai equation and that of the Kushner-Stratonovitch equation (the two main equations of nonlinear filtering) are shown to be approximated the empirical distribution of systems of particles that have equal weights (unlike those presented in \cite{kj1} and \cite{kj2}) and do not require additional correction procedures (such as those introduced in \cite{dan3}, \cite{dan4}, \cite{dmm}, etc).

math.PR↗