Abstract. This paper extends the multilevel Monte Carlo variance reduction technique to nonlinear filtering. In particular, multilevel Monte Carlo is applied to a certain variant of the particle filter, the ensemble transform particle filter (EPTF). A key aspect is the use of optimal transport methods to re-establish correlation between coarse and fine ensembles after resampling; this controls the variance of the estimator. Numerical examples present a proof of concept of the effectiveness of the proposed method, demonstrating significant computational cost reductions (relative to the single-level ETPF counterpart) in the propagation of ensembles.Key words. multilevel Monte Carlo, sequential data assimilation, optimal transport AMS subject classifications. 65C05, 62M20, 93E11, 93B40, 90C05 DOI. 10.1137/15M10382321. Introduction. Data assimilation is the process of incorporating observed data into model forecasts. In data assimilation, one is interested in computing statistics E η [X] of solutions X to random dynamical systems with respect to a posterior measure (η) given partial observations of the system. In particle filtering [7,4], this is done by using an empirical ensemble representing the posterior distribution η at any one time. The propagation in time of the members (particles) of this ensemble can be computationally expensive, especially in high dimensional systems.Recently, the multilevel Monte Carlo (MLMC) method has been developed for achieving significant cost reductions in Monte Carlo simulations [9]. It has been applied to areas such as Markov chain Monte Carlo [13] and quasi-Monte Carlo [10] to return computational cost reductions from existing techniques. It has also been applied to uncertainty quantification within PDEs [6]. The idea is to consider a hierarchy of discretized models, balancing numerical error in cheap/coarse models against Monte Carlo variance in expensive/fine models. It is desirable to adapt MLMC to sequential Monte Carlo methods such as particle filters, and some first steps have been taken in this direction. First, the authors of [11] have developed a multilevel ensemble Kalman filter (EnKF), using MLMC estimators to calculate the mean and covariance of the posterior, in the case where the underlying distributions are Gaussian and the model is linear. However, for non-Gaussian distributions and nonlinear models, the EnKF is biased. The method does, however, converge to a "mean-field limit" [14]. Second, the authors of [3] proposed a multilevel sequential Monte Carlo method for Bayesian inference problems to give significant computational cost reductions from standard techniques. Our goal in this paper is to take a step further by applying MLMC to nonlinear filtering problems.In general, MLMC works by computing statistics from pairs of coarser and finer