Past Computational Mathematics and Applications Seminar

One of the major steps in the adaptive finite element methods (AFEM) is the adaptive selection of the next partition. The process is usually governed by a strategy based on carefully chosen local error indicators and aims at convergence results with optimal rates. One can formally relate the refinement of the partitions with growing an oriented graph or a tree. Then each node of the tree/graph corresponds to a cell of a partition and the approximation of a function on adaptive partitions can be expressed trough the local errors related to the cell, i.e., the node. The total approximation error is then calculated as the sum of the errors on the leaves (the terminal nodes) of the tree/graph and the problem of finding an optimal error for a given budget of nodes is known as tree approximation. Establishing a near-best tree approximation result is a key ingredient in proving optimal convergence rates for AFEM.

The classical tree approximation problems are usually related to the so-called h-adaptive approximation in which the improvements a due to reducing the size of the cells in the partition. This talk will consider also an extension of this framework to hp-adaptive approximation allowing different polynomial spaces to be used for the local approximations at different cells while maintaining the near-optimality in terms of the combined number of degrees of freedom used in the approximation.

The problem of conformity of the resulting partition will be discussed as well. Typically in AFEM, certain elements of the current partition are marked and subdivided together with some additional ones to maintain desired properties of the partition like conformity. This strategy is often described as “mark → subdivide → complete”. The process is very well understood for triangulations received via newest vertex bisection procedure. In particular, it is proven that the number of elements in the final partition is limited by constant times the number of marked cells. This hints at the possibility to design a marking procedure that is limited only to cells of the partition whose subdivision will result in a conforming partition and therefore no completion step would be necessary. This talk will present such a strategy together with theoretical results about its near-optimal performance.

In this talk, I am going to give an introduction to operator preconditioning as a general and robust strategy to precondition linear systems arising from Galerkin discretization of PDEs or Boundary Integral Equations. Then, in order to illustrate the applicability of this preconditioning technique, I will discuss the simple case of weakly singular and hypersingular integral equations, arising from exterior Dirichlet and Neumann BVPs for the Laplacian in 3D. Finally, I will show how we can also tackle operators with a more difficult structure, like the electric field integral equation (EFIE) on screens, which models the scattering of time-harmonic electromagnetic waves at perfectly conducting bounded infinitely thin objects, like patch antennas in 3D.

We present a novel approach to the solution of time-dependent PDEs via the so-called monolithic or all-at-once formulation.

This approach will be explained for simple parabolic problems and its utility in the context of PDE constrained optimization problems will be elucidated.

The underlying linear algebra includes circulant matrix approximations of Toeplitz-structured matrices and allows for effective parallel implementation. Simple computational results will be shown for the heat equation and the wave equation which indicate the potential as a parallel-in-time method.

This is joint work with Elle McDonald (CSIRO, Australia), Jennifer Pestana (Strathclyde University, UK) and Anthony Goddard (Durham University, UK)

ABSTRACT

We approximate the solution of the stationary Stokes equations with various conforming and nonconforming inf-sup stable pairs of finite element spaces on simplicial meshes. Based on each pair, we design a discretization that is quasi-optimal and pressure robust, in the sense that the velocity H^1-error is proportional to the best H^1-error to the analytical velocity. This shows that such a property can be achieved without using conforming and divergence-free pairs. We bound also the pressure L^2-error, only in terms of the best approximation errors to the analytical velocity and the analytical pressure. Our construction can be summarized as follows. First, a linear operator acts on discrete velocity test functions, before the application of the load functional, and maps the discrete kernel into the analytical one.

Second, in order to enforce consistency, we  possibly employ a new augmented Lagrangian formulation, inspired by Discontinuous Galerkin methods.

Small block overlapping, and non-overlapping, Schwarz methods are theoretically highly attractive as multilevel smoothers for a wide variety of problems that are not amenable to point relaxation methods.  Examples include monolithic Vanka smoothers for Stokes, overlapping vertex-patch decompositions for $H(\text{div})$ and  $H(\text{curl})$ problems, along with nearly incompressible elasticity, and augmented Lagrangian schemes.

 While it is possible to manually program these different schemes,  their use in general purpose libraries has been held back by a lack   of generic, composable interfaces. We present a new approach to the   specification and development such additive Schwarz methods in PETSc  that cleanly separates the topological space decomposition from the  discretisation and assembly of the equations. Our preconditioner is  flexible enough to support overlapping and non-overlapping additive  Schwarz methods, and can be used to formulate line, and plane smoothers, Vanka iterations, amongst others. I will illustrate these new features with some examples utilising the Firedrake finite element library, in particular how the design of an approriate computational interface enables these schemes to be used as building blocks inside block preconditioners.

This is joint work with Patrick Farrell and Florian Wechsung (Oxford), and Matt Knepley (Buffalo).

Classical methods for X-ray computed tomography (CT) are based on the assumption that the X-ray source intensity is known. In practice, however, the intensity is measured and hence uncertain. Under normal circumstances, when the exposure time is sufficiently high, this kind of uncertainty typically has a negligible effect on the reconstruction quality. However, in time- or dose-limited applications such as dynamic CT, this uncertainty may cause severe and systematic artifacts known as ring artifacts.
By modeling the measurement process and by taking uncertainties into account, it is possible to derive a convex reconstruction model that leads to improved reconstructions when the signal-to-noise ratio is low. We discuss some computational challenges associated with the model and illustrate its merits with some numerical examples based on simulated and real data.

 Multiscale methods based on coupled partial differential equations defined on bulk and embedded manifolds are still poorly explored from the theoretical standpoint, although they are successfully used in applications, such as microcirculation and flow in perforated subsurface reservoirs. This work aims at shedding light on some theoretical aspects of a multiscale method consisting of coupled partial differential equations defined on one-dimensional domains embedded into three-dimensional ones. Mathematical issues arise because the dimensionality gap between the bulk and the inclusions is larger than one, named as the high dimensionality gap case. First, we show that such model derives from a system of full three-dimensional equations, by the application of a topological model reduction approach. Secondly, we rigorously analyze the problem, showing that the averaging operators applied for the model reduction introduce a regularization effect that resolves the issues due to the singularity of solutions and to the ill-posedness of restriction operators. Then, we discretize the problem by means of the finite element method and we analyze the approximation error. Finally, we exploit the structure of the model reduction technique to analyze the modeling error. This study confirms that for infinitesimally small inclusions, the modeling error vanishes.

This is a joint work with Federica Laurino, Department of Mathematics, Politecnico di Milano.

Driven by its numerous applications in computational science, the approximation of smooth, high-dimensional functions via sparse polynomial expansions has received significant attention in the last five to ten years.  In the first part of this talk, I will give a brief survey of recent progress in this area.  In particular, I will demonstrate how the proper use of compressed sensing tools leads to new techniques for high-dimensional approximation which can mitigate the curse of dimensionality to a substantial extent.  The rest of the talk is devoted to approximating functions defined on irregular domains.  The vast majority of works on high-dimensional approximation assume the function in question is defined over a tensor-product domain.  Yet this assumption is often unrealistic.  I will introduce a method, known as polynomial frame approximation, suitable for broad classes of irregular domains and present theoretical guarantees for its approximation error, stability, and sample complexity.  These results show the suitability of this approach for high-dimensional approximation through the independence (or weak dependence) of the various guarantees on the ambient dimension d.  Time permitting, I will also discuss several extensions.

When the linear system in Newton’s method is approximately solved using an iterative method we have an inexact or truncated Newton method. The outer method is Newton’s method and the inner iterations will be the iterative method. The Inexact Newton framework is now close to 30 years old and is widely used and given names like Newton-Arnoldi, Newton-CG depending on the inner iterative method. In this talk we will explore convergence properties when the outer iterative method is Gauss-Newton, the Halley method or an interior point method for linear programming problems.

We discuss the key role that bespoke linear algebra plays in modelling PDEs with random coefficients using stochastic Galerkin approximation methods. As a specific example, we consider nearly incompressible linear elasticity problems with an uncertain spatially varying Young's modulus. The uncertainty is modelled with a finite set of parameters with prescribed probability distribution.  We introduce a novel three-field mixed variational formulation of the PDE model and and  assess the stability with respect to a weighted norm. The main focus will be  the efficient solution of the associated high-dimensional indefinite linear system of equations. Eigenvalue bounds for the preconditioned system can be  established and shown to be independent of the discretisation parameters and the Poisson ratio.  We also  discuss an associated a posteriori error estimation strategy and assess proxies for the error reduction associated with selected enrichments of the approximation spaces.  We will show by example that these proxies enable the design of efficient  adaptive solution algorithms that terminate when the estimated error falls below a user-prescribed tolerance.

This is joint work with Arbaz Khan and Catherine Powell

Just like optimization needs derivatives, shape optimization needs shape derivatives. Their definition and computation is a classical subject, at least concerning first order shape derivatives. Second derivatives have been studied as well, but some aspects of their theory still remains a bit mysterious for practitioners. As a result, most algorithms for shape optimization are first order methods.

To understand this situation better and in a general way, we consider first and second order shape sensitivities of integrals on smooth submanifolds using a variant of shape differentiation. Instead of computing the second derivative as the derivative of the first derivative, we choose a one-parameter family of perturbations  and compute first and second derivatives with respect to that parameter. The result is a  quadratic form in terms of a perturbation vector field that yields a second order quadratic model of the perturbed functional, which can be used as the basis of a second order shape optimization algorithm. We discuss the structure of this derivative, derive domain expressions and Hadamard forms in a general geometric framework, and give a detailed geometric interpretation of the arising terms.

Finally, we use our results to construct a second order SQP-algorithm for shape optimization that exhibits indeed local fast convergence.

Standard quadratic programs have numerous applications and play an important role in copositivity detection. We consider reformulating a standard quadratic program as a mixed integer linear programming (MILP) problem. We propose alternative MILP reformulations that exploit the specific structure of standard quadratic programs. We report extensive computational results on various classes of instances. Our experiments reveal that our MILP reformulations significantly outperform other global solution approaches. 
This is joint work with Jacek Gondzio.

In this talk I will present two recent findings concerning the preconditioning of elliptic problems. The first result concerns preconditioning of elliptic problems with variable coefficient K by an inverse Laplacian. Here we show that there is a close relationship between the eigenvalues of the preconditioned system and K. 
The second results concern the problem on mixed form where K approaches zero. Here, we show a uniform inf-sup condition and corresponding robust preconditioning. 

In many applications requiring the solution of a linear system Ax=b, the matrix A has been shown to have a low-rank property: its off-diagonal blocks have low numerical rank, i.e., they can be well approximated by matrices of small rank. Several matrix formats have been proposed to exploit this property depending on how the block partitioning of the matrix is computed.
In this talk, I will discuss the block low-rank (BLR) format, which partitions the matrix with a simple, flat 2D blocking. I will present the main characteristics of BLR matrices, in particular in terms of asymptotic complexity and parallel performance. I will then discuss some recent advances and ongoing research on BLR matrices: their multilevel extension, their use as preconditioners for iterative solvers, the error analysis of their factorization, and finally the use of fast matrix arithmetic to accelerate BLR matrix operations.

A posteriori error estimators are a key tool for the quality assessment of given finite element approximations to an unknown PDE solution as well as for the application of adaptive techniques. Typically, the estimators are equivalent to the error up to an additive term, the so called oscillation. It is a common believe that this is the price for the `computability' of the estimator and that the oscillation is of higher order than the error. Cohen, DeVore, and Nochetto [CoDeNo:2012], however, presented an example, where the error vanishes with the generic optimal rate, but the oscillation does not. Interestingly, in this example, the local $H^{-1}$-norms are assumed to be computed exactly and thus the computability of the estimator cannot be the reason for the asymptotic overestimation. In particular, this proves both believes wrong in general. In this talk, we present a new approach to posteriori error analysis, where the oscillation is dominated by the error. The crucial step is a new splitting of the data into oscillation and oscillation free data. Moreover, the estimator is computable if the discrete linear system can essentially be assembled exactly.
 

The FEniCS system [1] allows the description of finite element discretisations of partial differential equations using a high-level syntax, and the automated conversion of these representations to working code via automated code generation. In previous work described in [2] the high-level representation is processed automatically to derive discrete tangent-linear and adjoint models. The processing of the model code at a high level eases the technical difficulty associated with management of data in adjoint calculations, allowing the use of optimal data management strategies [3].

This previous methodology is extended to enable the calculation of higher order partial differential equation constrained derivative information. The key additional step is to treat tangent-linear
equations on an equal footing with originating forward equations, and in particular to treat these in a manner which can themselves be further processed to enable the derivation of associated adjoint information, and the derivation of higher order tangent-linear equations, to arbitrary order. This enables the calculation of higher order derivative information -- specifically the contraction of a Kth order derivative against (K - 1) directions -- while still making use of optimal data management strategies. Specific applications making use of Hessian information associated with models written using the FEniCS system are presented.

[1] "Automated solution of differential equations by the finite element method: The FEniCS book", A. Logg, K.-A. Mardal, and  G. N. Wells (editors), Springer, 2012
[2] P. E. Farrell, D. A. Ham, S. W. Funke, and M. E. Rognes, "Automated derivation of the adjoint of high-level transient finite element programs", SIAM Journal on Scientific Computing 35(4), C369--C393, 2013
[3] A. Griewank, and A. Walther, "Algorithm 799: Revolve: An implementation of checkpointing for the reverse or adjoint mode of computational differentiation", ACM Transactions on Mathematical Software 26(1), 19--45, 2000

$$
\def\curl#1{\left\{#1\right\}}
\def\vek#1{\mathbf{#1}}
$$
lll-posed problems arise often in the context of scientific applications in which one cannot directly observe the object or quantity of interest. However, indirect observations or measurements can be made, and the observable data $y$ can be represented as the wanted observation $x$ being acted upon by an operator $\mathcal{A}$. Thus we want to solve the operator equation \begin{equation}\label{eqn.Txy} \mathcal{A} x = y, \end{equation} (1) often formulated in some Hilbert space $H$ with $\mathcal{A}:H\rightarrow H$ and $x,y\in H$. The difficulty then is that these problems are generally ill-posed, and thus $x$ does not depend continuously on the on the right-hand side. As $y$ is often derived from measurements, one has instead a perturbed $y^{\delta}$ such that ${y - y^{\delta}}_{H}<\delta$. Thus due to the ill-posedness, solving (1) with $y^{\delta}$ is not guaranteed to produce a meaningful solution. One such class of techniques to treat such problems are the Tikhonov-regularization methods. One seeks in reconstructing the solution to balance fidelity to the data against size of some functional evaluation of the reconstructed image (e.g., the norm of the reconstruction) to mitigate the effects of the ill-posedness. For some $\lambda>0$, we solve \begin{equation}\label{eqn.tikh} x_{\lambda} = \textrm{argmin}_{\widetilde{x}\in H}\left\lbrace{\left\|{b - A\widetilde{x}} \right\|_{H}^{2} + \lambda \left\|{\widetilde{x}}\right\|_{H}^{2}} \right\rbrace. \end{equation} In this talk, we discuss some new strategies for treating discretized versions of this problem. Here, we consider a discreditized, finite dimensional version of (1), \begin{equation}\label{eqn.Axb} Ax =  b \mbox{ with }  A\in \mathbb{R}^{n\times n}\mbox{ and } b\in\mathbb{R}^{n}, \end{equation} which inherits a discrete version of ill conditioning from [1]. We propose methods built on top of the Arnoldi-Tikhonov method of Lewis and Reichel, whereby one builds the Krylov subspace \begin{equation}
\mathcal{K}_{j}(\vek A,\vek w) = {\rm span\,}\curl{\vek w,\vek A\vek w,\vek A^{2}\vek w,\ldots,\vek A^{j-1}\vek w}\mbox{ where } \vek w\in\curl{\vek b,\vek A\vek b}
\end{equation}
and solves the discretized Tikhonov minimization problem projected onto that subspace. We propose to extend this strategy to setting of augmented Krylov subspace methods. Thus, we project onto a sum of subspaces of the form $\mathcal{U} + \mathcal{K}_{j}$ where $\mathcal{U}$ is a fixed subspace and $\mathcal{K}_{j}$ is a Krylov subspace. It turns out there are multiple ways to do this leading to different algorithms. We will explain how these different methods arise mathematically and demonstrate their effectiveness on a few example problems. Along the way, some new mathematical properties of the Arnoldi-Tikhonov method are also proven.

Since the legendary 1972 encounter of H. Montgomery and F. Dyson at tea time in Princeton, a statistical correspondence of the non-trivial zeros of the Riemann Zeta function with eigenvalues of high-dimensional random matrices has emerged. Surrounded by many deep conjectures, there is a striking analogyto the energy levels of a quantum billiard system with chaotic dynamics. Thanks 
to extensive calculation of Riemann zeros by A. Odlyzko, overwhelming numerical evidence has been found for the quantum analogy. The statistical accuracy provided by an enormous dataset of more than one billion zeros reveals distinctive finite size effects. Using the physical analogy, a precise prediction of these effects was recently accomplished through the numerical evaluation of operator determinants and their perturbation series (joint work with P. Forrester and A. Mays, Melbourne).
 

The present work concerns the approximation of the solution map associated to the parametric Helmholtz boundary value problem, i.e., the map which associates to each (real) wavenumber belonging to a given interval of interest the corresponding solution of the Helmholtz equation. We introduce a single-point Least Squares (LS) rational Padé-type approximation technique applicable to any meromorphic Hilbert space-valued univariate map, and we prove the uniform convergence of the Padé approximation error on any compact subset of the interval of interest that excludes any pole. We also present a simplified and more efficient version, named Fast LS-Padé, applicable to Helmholtz-type parametric equations with normal operators.

The LS-Padé techniques are then employed to approximate the frequency response map associated to various parametric time-harmonic wave problems, namely, a transmission/reflection problem, a scattering problem and a problem in high-frequency regime. In all cases we establish the meromorphy of the frequency response map. The Helmholtz equation with stochastic wavenumber is also considered. In particular, for Lipschitz functionals of the solution, and their corresponding probability measures, we establish weak convergence of the measure derived from the LS-Padé approximant to the true one. Two-dimensioanl numerical tests are performed, which confirm the effectiveness of the approximation method.As of the dates

 Joint work with: Francesca Bonizzoni and  Ilaria Perugia (Uni. Vienna), Davide Pradovera (EPFL)

Random matrices now play a role in many areas of theoretical, applied, and computational mathematics. Therefore, it is desirable to have tools for studying random matrices that are flexible, easy to use, and powerful. Over the last fifteen years, researchers have developed a remarkable family of results, called matrix concentration inequalities, that balance these criteria. This talk offers an invitation to the field of matrix concentration inequalities and their applications.

We first consider multilevel Monte Carlo and stochastic collocation methods for determining statistical information about an output of interest that depends on the solution of a PDE with inputs that depend on random parameters. In our context, these methods connect a hierarchy of spatial grids to the amount of sampling done for a given grid, resulting in dramatic acceleration in the convergence of approximations. We then consider multifidelity methods for the same purpose which feature a variety of models that have different fidelities. For example, we could have coarser grid discretizations, reduced-order models, simplified physics, surrogates such as interpolants, and, in principle, even experimental data. No assumptions are made about the fidelity of the models relative to the “truth” model of interest so that unlike multilevel methods, there is no a priori model hierarchy available. However, our approach can still greatly accelerate the convergence of approximations.

In the past few decades, power grids across the world have become dependent on markets that aim to efficiently match supply with demand at all times via a variety of pricing and auction mechanisms. These markets are based on models that capture interactions between producers, transmission and consumers. Energy producers typically maximize profits by optimally allocating and scheduling resources over time. A dynamic equilibrium aims to determine prices and dispatches that can be transmitted over the electricity grid to satisfy evolving consumer requirements for energy at different locations and times. Computation allows large scale practical implementations of socially optimal models to be solved as part of the market operation, and regulations can be imposed that aim to ensure competitive behaviour of market participants.

Questions remain that will be outlined in this presentation.

Firstly, the recent explosion in the use of renewable supply such as wind, solar and hydro has led to increased volatility in this system. We demonstrate how risk can impose significant costs on the system that are not modeled in the context of socially optimal power system markets and highlight the use of contracts to reduce or recover these costs. We also outline how battery storage can be used as an effective hedging instrument.

Secondly, how do we guarantee continued operation in rarely occuring situations and when failures occur and how do we price this robustness?

Thirdly, how do we guarantee appropriate participant behaviour? Specifically, is it possible for participants to develop strategies that move the system to operating points that are not socially optimal?

Fourthly, how do we ensure enough transmission (and generator) capacity in the long term, and how do we recover the costs of this enhanced infrastructure?