Past Computational Mathematics and Applications Seminar

Work with Jemima Tabeart, Sarah Dance, Nancy Nichols, Joanne Waller (University of Reading) and Stefano Migliorini, Fiona Smith (Met Office). 
In environmental prediction variational data assimilation (DA) is a method for using observational data to estimate the current state of the system. The DA problem is usually solved as a very large nonlinear least squares problem, in which the fit to the measurements is balanced against the fit to a previous model forecast. These two terms are weighted by matrices describing the correlations of the errors in the forecast and in the observations. Until recently most operational weather and ocean forecasting systems assumed that the errors in the observations are uncorrelated. However, as we move to higher resolution observations then it is becoming more important to specify observation error correlations. In this work we look at the effect this has on the conditioning of the optimization problem. In the context of a linear system we develop bounds on the condition number of the problem in the presence of correlated observation errors. We show that the condition number is very dependent on the minimum eigenvalue of the observation error correlation matrix. We then present results using the Met Office data assimilation system, in which different methods for reconditioning the correlation matrix are tested. We investigate the effect of these different methods on the conditioning and the final solution of the problem.
 

Inverse problems are ubiquitous in many areas of Science and Engineering and, once discretised, they lead to ill-conditioned linear systems, often of huge dimensions: regularisation consists in replacing the original system by a nearby problem with better numerical properties, in order to find meaningful approximations of its solution. In this talk we will explore the regularisation properties of many iterative methods based on Krylov subspaces. After surveying some basic methods such as CGLS and GMRES, innovative approaches based on flexible variants of CGLS and GMRES will be presented, in order to efficiently enforce nonnegativity and sparsity into the solution.

We consider the Cauchy (or steepest descent) method with exact line search applied to a strongly convex function with Lipschitz continuous gradient. We establish the exact worst-case rate of convergence of this scheme, and show that this worst-case behavior is exhibited by a certain convex quadratic function. We also give worst-case complexity bound for a noisy variant of gradient descent method. Finally, we show that these results may be applied to study the worst-case performance of Newton's method for the minimization of self-concordant functions.

The proofs are computer-assisted, and rely on the resolution of semidefinite programming performance estimation problems as introduced in the paper [Y. Drori and M. Teboulle.  Performance of first-order methods for smooth convex minimization: a novel approach. Mathematical Programming, 145(1-2):451-482, 2014].

Joint work with F. Glineur and A.B. Taylor.

 Functions defined by evaluation programs involving smooth  elementals and absolute values as well as max and min are piecewise smooth. For this class we present first and second order, necessary and sufficient conditions for the functions to be locally optimal, or convex, or at least possess a supporting hyperplane. The conditions generalize the classical KKT and SSC theory and are constructive; though in the case of convexity they may be combinatorial to verify. As a side product we find that, under the Mangasarin-Fromowitz-Kink-Qualification, the well established nonsmooth concept of subdifferential regularity is equivalent to first order convexity. All results are based on piecewise linearization and suggest corresponding optimization algorithms.

We propose a multilevel paradigm for the global optimisation of polynomials with sparse support. Such polynomials arise through the discretisation of PDEs, optimal control problems and in global optimization applications in general. We construct projection operators to relate the primal and dual variables of the SDP relaxation between lower and higher levels in the hierarchy, and theoretical results are proven to confirm their usefulness. Numerical results are presented for polynomial problems that show how these operators can be used in a hierarchical fashion to solve large scale problems with high accuracy.

I will present a broad family of stochastic algorithms for inverting a matrix, including specialized variants which maintain symmetry or positive definiteness of the iterates. All methods in the family converge globally and linearly, with explicit rates. In special cases, the methods obtained are stochastic block variants of several quasi-Newton updates, including bad Broyden (BB), good Broyden (GB), Powell-symmetric-Broyden (PSB), Davidon-Fletcher-Powell (DFP) and Broyden-Fletcher-Goldfarb-Shanno (BFGS). After a pause for questions, I will then present a block stochastic BFGS method based on the stochastic method for inverting positive definite matrices. In this method, the estimate of the inverse Hessian matrix that is maintained by it, is updated at each iteration using a sketch of the Hessian, i.e., a randomly generated compressed form of the Hessian. I will propose several sketching strategies, present a new quasi-Newton method that uses stochastic block BFGS updates combined with the variance reduction approach SVRG to compute batch stochastic gradients, and prove linear convergence of the resulting method. Numerical tests on large-scale logistic regression problems reveal that our method is more robust and substantially outperforms current state-of-the-art methods.

Most current methods of Magnetic Resonance Imaging (MRI) reconstruction interpret raw signal values as samples of the Fourier transform of the object. Although this is computationally convenient, it neglects relaxation and off–resonance evolution in phase, both of which can occur to significant extent during a typical MRI signal. A more accurate model, known as Parameter Assessment by Recovery from Signal Encoding (PARSE), takes the time evolution of the signal into consideration. This model uses three parameters that depend on tissue properties: transverse magnetization, signal decay rate, and frequency offset from resonance. Two difficulties in recovering an image using this model are the low SNR for long acquisition times in single-shot MRI, and the nonlinear dependence of the signal on the decay rate and frequency offset. In this talk, we address the latter issue by using a second order approximation of the original PARSE model. The linearized model can be solved using convex optimization augmented with well-stablished regularization techniques such as total variation. The sensitivity of the parameters to noise and computational challenges associated with this approximation will be discussed.

We consider the problem of computing a nonnegative low rank factorization to a given nonnegative input matrix under the so-called "separabilty condition".  This assumption makes this otherwise NP hard problem polynomial time solvable, and we will use first order optimization techniques to compute such a factorization. The optimization model use is based on sparse regression with a self-dictionary, in which the low rank constraint is relaxed to the minimization of an l1-norm objective function.  We apply these techniques to endmember detection and classification in hyperspecral imaging data.

 The matrix logarithm, when applied to symmetric positive definite matrices, is known to satisfy a notable concavity property in the positive semidefinite (Loewner) order. This concavity property is a cornerstone result in the study of operator convex functions and has important applications in matrix concentration inequalities and quantum information theory.
In this talk I will show that certain rational approximations of the matrix logarithm remarkably preserve this concavity property and moreover, are amenable to semidefinite programming. Such approximations allow us to use off-the-shelf semidefinite programming solvers for convex optimization problems involving the matrix logarithm. These approximations are also useful in the scalar case and provide a much faster alternative to existing methods based on successive approximation for problems involving the exponential/relative entropy cone. I will conclude by showing some applications to problems arising in quantum information theory.

This is joint work with James Saunderson (Monash University) and Pablo Parrilo (MIT)

Rational Krylov methods are applicable to a wide range of scientific computing problems, and ​the rational Arnoldi algorithm is a commonly used procedure for computing an ​orthonormal basis of a rational Krylov space. Typically, the computationally most expensive component of this​ ​algorithm is the solution of a large linear system of equations at each iteration. We explore the​ ​option of solving several linear systems simultaneously, thus constructing the rational Krylov​ ​basis in parallel. If this is not done carefully, the basis being orthogonalized may become badly​ ​conditioned, leading to numerical instabilities in the orthogonalization process. We introduce the​ ​new concept of continuation pairs which gives rise to a near-optimal parallelization strategy that ​allows to control the growth of the condition number of this nonorthogonal basis. As a consequence we obtain a significantly more accurate and reliable parallel rational Arnoldi algorithm.
​ ​
The computational benefits are illustrated using several numerical examples from different application areas.
​ ​
This ​talk is based on joint work with Mario Berljafa  available as an Eprint at http://eprints.ma.man.ac.uk/2503/
 

We present global rates of convergence for a general class of methods for nonconvex smooth optimization that include linesearch, trust-region and regularisation strategies, but that allow inaccurate problem information. Namely, we assume the local (first- or second-order) models of our function are only sufficiently accurate with a certain probability, and they can be arbitrarily poor otherwise. This framework subsumes certain stochastic gradient analyses and derivative-free techniques based on random sampling of function values. It can also be viewed as a robustness
assessment of deterministic methods and their resilience to inaccurate derivative computation such as due to processor failure in a distribute framework. We show that in terms of the order of the accuracy, the evaluation complexity of such methods is the same as their counterparts that use deterministic accurate models; the use of probabilistic models only increases the complexity by a constant, which depends on the probability of the models being good. Time permitting, we also discuss the case of inaccurate, probabilistic function value information, that arises in stochastic optimization. This work is joint with Katya Scheinberg (Lehigh University, USA).
 

For linear dynamical systems, model reduction has achieved great success. In the case of linear dynamics,  we know how to construct, at a modest cost, (locally) optimal, input-independent reduced models; that is, reduced models that are uniformly good over all inputs having bounded energy. In addition, in some cases we can achieve this goal using only input/output data without a priori knowledge of internal  dynamics.  Even though model reduction has been successfully and effectively applied to nonlinear dynamical systems as well, in this setting,  bot the reduction process and the reduced models are input dependent and the high fidelity of the resulting approximation is generically restricted to the training input/data. In this talk, we will offer remedies to this situation.

To predict the behaviour of a dynamical system using a mathematical model, an accurate estimate of the current state of the system is needed in order to initialize the model. Complete information on the current state is, however, seldom available. The aim of optimal state estimation, known in the geophysical sciences as ‘data assimilation’, is to determine a best estimate of the current state using measured observations of the real system over time, together with the model equations. The problem is commonly formulated in variational terms as a very large nonlinear least-squares optimization problem. The lack of complete data, coupled with errors in the observations and in the model, leads to a highly ill-conditioned inverse problem that is difficult to solve.

To understand the nature of the inverse problem, we examine how different components of the assimilation system influence the conditioning of the optimization problem. First we consider the case where the dynamical equations are assumed to model the real system exactly. We show, against intuition, that with increasingly dense and precise observations, the problem becomes harder to solve accurately. We then extend these results to a 'weak-constraint' form of the problem, where the model equations are assumed not to be exact, but to contain random errors. Two different, but mathematically equivalent, forms of the problem are derived. We investigate the conditioning of these two forms and find, surprisingly, that these have quite different behaviour.

The requirement to compute Jordan blocks for multiple eigenvalues arises in a number of physical problems, for example panel flutter problems in aerodynamical stability, the stability of electrical power systems, and in quantum mechanics. We introduce a general method for computing a 2-dimensional Jordan block in a parameter-dependent matrix eigenvalue problem based on the so called Implicit Determinant Method. This is joint work with Alastair Spence (Bath).

In many applications we need to find or estimate the $p \ge 1$ largest elements of a matrix, along with their locations. This is required for recommender systems used by Amazon and Netflix, link prediction in graphs, and in finding the most important links in a complex network, for example. 

Our algorithm uses only matrix vector products and is based upon a power method for mixed subordinate norms. We have obtained theoretical results on the convergence of this algorithm via a comparison with rook pivoting for the LU  decomposition. We have also improved the practicality of the algorithm by producing a blocked version iterating on $n \times t$ matrices, as opposed to vectors, where $t$ is a tunable parameter. For $p > 1$ we show how deflation can be used to improve the convergence of the algorithm. 

Finally, numerical experiments on both randomly generated matrices and real-life datasets (the latter for $A^TA$ and $e^A$) show how our algorithms can reliably estimate the largest elements of a matrix whilst obtaining considerable speedups when compared to forming the matrix explicitly: over 1000x in some cases.

For many years, sum-factored algorithms for finite elements in rectangular reference geometry have combined low complexity with the mathematical power of high-order approximation.  However, such algorithms rely heavily on the tensor product structure inherent in the geometry and basis functions, and similar algorithms for simplicial geometry have proven elusive.

Bernstein polynomials are totally nonnegative, rotationally symmetric, and geometrically decomposed bases with many other remarkable properties that lead to optimal-complexity algorithms for element wise finite element computations.  The also form natural building blocks for the finite element exterior calculus bases for the de Rham complex so that H(div) and H(curl) bases have efficient representations as well.  We will also their relevance for explicit discontinuous Galerkin methods, where the element mass matrix requires special attention.

We will review the basic building blocks of iterative solvers, i.e. sparse matrix-vector multiplication, in the context of GPU devices such 
as the cards by NVIDIA; we will then discuss some techniques in preconditioning by approximate inverses, and we will conclude with an 
application to an image processing problem from the biomedical field.

The aim of this talk is to design an efficient multigrid method for constrained convex optimization problems arising from discretization  of  some  underlying  infinite  dimensional  problems. Due  to problem  dependency  of this approach, we only consider bound constraints with (possibly) a linear equality constraint. As our aim is to target large-scale problems, we want to avoid computation of second 
derivatives of the objective function, thus excluding Newton like methods. We propose a smoothing operator that only uses first-order information and study the computational efficiency of the resulting method. In the second part, we consider application of multigrid techniques to more general optimization problems, in particular, the topology design problem.

Quadrature is the term for the numerical evaluation of integrals.  It's a beautiful subject because it's so accessible, yet full of conceptual surprises and challenges.  This talk will review ten of these, with plenty of history and numerical demonstrations.  Some are old if not well known, some are new, and two are subjects of my current research.