Reduced basis methods for nonlocal diffusion problems with random input data

Abstract

The construction, analysis, and application of reduced-basis methods for uncertainty quantification problems involving nonlocal diffusion problems with random input data is the subject of this work. Because of the lack of sparsity of discretized nonlocal models relative to analogous local partial differential equation models, the need for reducedorder modeling is much more acute in the nonlocal setting. In this effort, we develop reduced-basis approximations for nonlocal diffusion equations with affine random coefficients. Efficiency estimates of the proposed greedy reduced-basis methods is provided. Numerical examples are used to illustrate the effect varying various model parameters have on the efficiency and accuracy of the reduced-basis method relative to the sparse grid interpolation using the full finite element method. It is shown that the proposed reduced-basis approach can indeed provide substantial savings over the standard sparse grid method when combined with full finite element solvers. is used to describe long-range spatial interactions. For example, in solid mechanics, the strain energy density at a point x can depend on points in a neighborhood of x having non-zero volume. This is in contrast to local PDE models where interactions occur only in an infinitesimal neighborhood of the point x. Further, without some interventions, PDEs cannot model discontinuous behavior such as crack nucleation and propagation. Nonlocal models that are free of spatial derivatives can gracefully handle continuous and discontinuous behaviors such as fracture [1,2]. In image denosing, it is often reasonable to assume that the noise at any point is related directly to that any other point in the image, or at least to points in relevant regions, rather than just its immediate neighbors. This leads to the study of nonlocal operators in image analyses [3][4][5][6] that, similar to the nonlocal model in continuum mechanics, can process both structures (continuous) and textures (discontinuous) within the same framework. In the case considered here, i.e., diffusion, nonlocal interaction give rise to anomalous diffusion, i.e., to diffusive behavior that does not follow Fickian laws or that can be modeled by Brownian motions. In [10], a finite element framework and a thorough variational analysis is provided for nonlocal problems based on the nonlocal vector calculus recently developed in [11].On the other hand, the inputs to a model, be they local or nonlocal, are often subject to uncertainties. For example, in the PDE setting, coefficients, forcing terms, and data in initial and boundary conditions may be uncertain. Thus, uncertainty quantification, i.e., the analysis and quantification of the uncertainties in the outputs of a model due to uncertainties in its inputs, has been of growing interest in recent years. For local problems modeled by differential equations with random input data, the classical Monte Carlo method is widely used. The appealing properties of this method are its ease of implementation based on existing det...

Numerical methods for nonlocal and fractional models

“…Thus the task at hand is to construct a ROM basis such that N rom ≪ N h , which results in ROM solutions u p,rom (x) that are acceptably accurate approximations of u δ,h,p (x) for any p φ ∈ Γ φ and p θ ∈ Γ θ , or at least for subsets of the parameter domains that are of interest. Two examples of the use of reduced-order models for nonlocal diffusion are given by Guan, Gunzburger, Webster and Zhang (2017), who use a greedy reduced basis (GRB) approach, and Witman, Gunzburger and Peterson (2017), who use a proper orthogonal decomposition (POD) approach. 6 Both approaches involve only the parameter vector p θ .…”

Model reduction for fractional elliptic problems using Kato's formula

“…. We apply these rules to the integrals in (25), resulting in the fully discrete approximation, 27) We emphasize that the y-discretization does not suffer from any numerical instabilities as s ↑ 1 or s ↓ 0: the weights τ j,± are positive, no larger than 1, and independent of s, β 0 (1 − s) and β 0 (s) are just sinc functions, and w ± (y) has bounded L 2 norm for all y ≥ 0, i.e., for all inputs. The following codifies this stability.…”

“…However, low-rank structure in solution sets to fractional problems has been empirically noted even earlier [44]. For problems involving nonlocal integral kernels, the authors in [25] also proposed a reduced basis approach, but use a different strategy to perform model reduction. Our contributions in this article are as follows:…”

An adaptive reduced basis ANOVA method for high-dimensional Bayesian inverse problems

“…This method is to adaptively select parameter samples, where errors between the reduced approximation and the finite element approximation are large. To assess the errors, we use the residual error indicator which is also adopted by [46,33,47,48]. Following our notation in [46], when considering linear PDEs, the algebraic system associated with (13) can be written as A ξ t u ξ t = f where A ξ t ∈ R N h ×N h , and u ξ t , f ∈ R N h .…”