While the formulation of most data assimilation schemes assumes an unbiased observation model error, in real applications model error with nontrivial biases is unavoidable. A practical example is errors in the radiative transfer model (which is used to assimilate satellite measurements) in the presence of clouds. Together with the dynamical model error, the result is that many (in fact 99%) of the cloudy observed measurements are not being used although they may contain useful information. This paper presents a novel nonparametric Bayesian scheme that is able to learn the observation model error distribution and correct the bias in incoming observations. This scheme can be used in tandem with any data assimilation forecasting system. The proposed model error estimator uses nonparametric likelihood functions constructed with data-driven basis functions based on the theory of kernel embeddings of conditional distributions developed in the machine learning community. Numerically, positive results are shown with two examples. The first example is designed to produce a bimodality in the observation model error (typical of “cloudy” observations) by introducing obstructions to the observations that occur randomly in space and time. The second example, which is physically more realistic, is to assimilate cloudy satellite brightness temperature–like quantities, generated from a stochastic multicloud model for tropical convection and a simple radiative transfer model.
22 December 2021 in Communications on Pure and Applied Mathematics
We study the problem of estimating linear response statistics under external perturbations using time series of unperturbed dynamics. Based on the fluctuation-dissipation theory, this problem is reformulated as an unsupervised learning task of estimating a density function. We consider a nonparametric density estimator formulated by the kernel embedding of distributions with "Mercer-type" kernels, constructed based on the classical orthogonal polynomials defined on non-compact domains. While the resulting representation is analogous to Polynomial Chaos Expansion (PCE), the connection to the reproducing kernel Hilbert space (RKHS) theory allows one to establish the uniform convergence of the estimator and to systematically address a practical question of identifying the PCE basis for a consistent estimation. We also provide practical conditions for the well-posedness of not only the estimator but also of the underlying response statistics. Finally, we provide a statistical error bound for the density estimation that accounts for the Monte-Carlo averaging over non-i.i.d time series and the biases due to a finite basis truncation. This error bound provides a means to understand the feasibility as well as limitation of the kernel embedding with Mercer-type kernels. Numerically, we verify the effectiveness of the estimator on two stochastic dynamics with known, yet, non-trivial equilibrium densities.
We study the spectral convergence of a symmetrized graph Laplacian matrix induced by a Gaussian kernel evaluated on pairs of embedded data, sampled from a manifold with boundary, a sub-manifold of $ \mathbb{R}^m $. Specifically, we deduce the convergence rates for eigenpairs of the discrete graph-Laplacian matrix to the eigensolutions of the Laplace-Beltrami operator that are well-defined on manifolds with boundary, including the homogeneous Neumann and Dirichlet boundary conditions. For the Dirichlet problem, we deduce the convergence of the truncated graph Laplacian, whose convergence was recently observed numerically in applications, and provide a detailed numerical investigation on simple manifolds. Our method of proof relies on a min-max argument over a compact and symmetric integral operator, leveraging the RKHS theory for spectral convergence of integral operators and a recent asymptotic result of a Gaussian kernel integral operator on manifolds with boundary.
23 November 2022 in Journal of Scientific Computing