Search for: All records

Abstract Obtaining lightweight and accurate approximations of discretized objective functional Hessians in inverse problems governed by partial differential equations (PDEs) is essential to make both deterministic and Bayesian statistical large-scale inverse problems computationally tractable. The cubic computational complexity of dense linear algebraic tasks, such as Cholesky factorization, that provide a means to sample Gaussian distributions and determine solutions of Newton linear systems is a computational bottleneck at large-scale. These tasks can be reduced to log-linear complexity by utilizing hierarchical off-diagonal low-rank (HODLR) matrix approximations. In this work, we show that a class of Hessians that arise from inverse problems governed by PDEs are well approximated by the HODLR matrix format. In particular, we study inverse problems governed by PDEs that model the instantaneous viscous flow of ice sheets. In these problems, we seek a spatially distributed basal sliding parameter field such that the flow predicted by the ice sheet model is consistent with ice sheet surface velocity observations. We demonstrate the use of HODLR Hessian approximation to efficiently sample the Laplace approximation of the posterior distribution with covariance further approximated by HODLR matrix compression. Computational studies are performed which illustrate ice sheet problem regimes for which the Gauss–Newton data-misfit Hessian is more efficiently approximated by the HODLR matrix format than the low-rank (LR) format. We then demonstrate that HODLR approximations can be favorable, when compared to global LR approximations, for large-scale problems by studying the data-misfit Hessian associated with inverse problems governed by the first-order Stokes flow model on the Humboldt glacier and Greenland ice sheet. 

Abstract The replacement of a nonlinear parameter-to-observable mapping with a linear (affine) approximation is often carried out to reduce the computational costs associated with solving large-scale inverse problems governed by partial differential equations (PDEs). In the case of a linear parameter-to-observable mapping with normally distributed additive noise and a Gaussian prior measure on the parameters, the posterior is Gaussian. However, substituting an accurate model for a (possibly well justified) linear surrogate model can give misleading results if the induced model approximation error is not accounted for. To account for the errors, the Bayesian approximation error (BAE) approach can be utilised, in which the first and second order statistics of the errors are computed via sampling. The most common linear approximation is carried out via linear Taylor expansion, which requires the computation of (Fréchet) derivatives of the parameter-to-observable mapping with respect to the parameters of interest. In this paper, we prove that the (approximate) posterior measure obtained by replacing the nonlinear parameter-to-observable mapping with a linear approximation is in fact independent of the choice of the linear approximation when the BAE approach is employed. Thus, somewhat non-intuitively, employing the zero-model as the linear approximation gives the same approximate posterior as any other choice of linear approximations of the parameter-to-observable model. The independence of the linear approximation is demonstrated mathematically and illustrated with two numerical PDE-based problems: an inverse scattering type problem and an inverse conductivity type problem. 

Abstract. We consider the problem of inferring the basal sliding coefficientfield for an uncertain Stokes ice sheet forward model from syntheticsurface velocity measurements. The uncertainty in the forward modelstems from unknown (or uncertain) auxiliary parameters (e.g., rheologyparameters). This inverse problem is posed within the Bayesianframework, which provides a systematic means of quantifyinguncertainty in the solution. To account for the associated modeluncertainty (error), we employ the Bayesian approximation error (BAE)approach to approximately premarginalize simultaneously over both thenoise in measurements and uncertainty in the forward model. We alsocarry out approximative posterior uncertainty quantification based ona linearization of the parameter-to-observable map centered at themaximum a posteriori (MAP) basal sliding coefficient estimate, i.e.,by taking the Laplace approximation. The MAP estimate is found byminimizing the negative log posterior using an inexact Newtonconjugate gradient method. The gradient and Hessian actions to vectorsare efficiently computed using adjoints. Sampling from theapproximate covariance is made tractable by invoking a low-rankapproximation of the data misfit component of the Hessian. We studythe performance of the BAE approach in the context of three numericalexamples in two and three dimensions. For each example, the basalsliding coefficient field is the parameter of primary interest whichwe seek to infer, and the rheology parameters (e.g., the flow ratefactor or the Glen's flow law exponent coefficient field) representso-called nuisance (secondary uncertain) parameters. Our resultsindicate that accounting for model uncertainty stemming from thepresence of nuisance parameters is crucial. Namely our findingssuggest that using nominal values for these parameters, as is oftendone in practice, without taking into account the resulting modelingerror, can lead to overconfident and heavily biased results. We alsoshow that the BAE approach can be used to account for the additionalmodel uncertainty at no additional cost at the online stage. 

Newton's method is usually preferred when solving optimization problems due to its superior convergence properties compared to gradient-based or derivative-free optimization algorithms. However, deriving and computing second-order derivatives needed by Newton's method often is not trivial and, in some cases, not possible. In such cases quasi-Newton algorithms are a great alternative. In this paper, we provide a new derivation of well-known quasi-Newton formulas in an infinite-dimensional Hilbert space setting. It is known that quasi-Newton update formulas are solutions to certain variational problems over the space of symmetric matrices. In this paper, we formulate similar variational problems over the space of bounded symmetric operators in Hilbert spaces. By changing the constraints of the variational problem we obtain updates (for the Hessian and Hessian inverse) not only for the Broyden-Fletcher-Goldfarb-Shanno (BFGS) quasi-Newton method but also for Davidon--Fletcher--Powell (DFP), Symmetric Rank One (SR1), and Powell-Symmetric-Broyden (PSB). In addition, for an inverse problem governed by a partial differential equation (PDE), we derive DFP and BFGS ``structured" secant formulas that explicitly use the derivative of the regularization and only approximates the second derivative of the misfit term. We show numerical results that demonstrate the desired mesh-independence property and superior performance of the resulting quasi-Newton methods.