Golub Memorial Conference  - Abstracts and Transparency Files

James Baglama, URI
Block Krylov Subspace Methods for Solving Large Sparse SVD and Eigenvalue Problems  

Abstract:  For the eigenvalue problem, we will describe the augmented block Householder Arnoldi method and for the SVD problem, we will describe the augmented block Lanczos bidiagonalization method. Both methods combine the advantages of a block routine with augmenting the Krylov subspace. These methods are fast, efficient, and reliable. Public domain MATLAB codes are available for both methods and several numerical experiments will be presented.

 

Tom Bella, University of Connecticut
Quasiseparable matrices and polynomials

Abstract: The interplay between polynomials and dense structured matrices is a classical topic. Structure in this sense is interpreted to mean their n2 entries can be “compressed” to a smaller number O(n) of parameters. Operating directly on these parameters allows one to design efficient fast algorithms for these matrices and for the related applied problems.

In the past decades matrices with structures such as DFT/DCT/DST, Toeplitz, Hankel, Vandermonde or Cauchy structure were the focus of attention. In this talk, some results that demonstrate that a relatively new quasiseparable structure enables substantial generalizations of a number of different algorithms will be presented.

 

Vani Cheruvu , Florida State University, (joint work with Joseph Tribbia)
Adjoint sensitivity analysis for two-layered shallow water model in a limited domain

Abstract:   Sensitivity analysis is defined as the determination of the potential impact on  some quantitative measures of a forecast aspect due to arbitrary perturbations of the model dynamic fields at earlier times. Adjoints are indispensible for efficiently solving several types of optimization problems, which includes sensitivity analysis. Currently we are working on the development of the adjoint tools for a spectral-element based two-layered shallow water model in a limited domain with non-reflecting boundary conditions. This system has much in common with limited area primitive equation models used in numerical weather prediction.

 
Moody T. Chu, North Carolina State University
Orthogonal Polynomials, Moments, Measure Deformation, Dynamical Systems, and SVD Algorithm

Abstract:  Iterates generated from discrete dynamical systems such as the QR algorithm and the SVD algorithm are time-1 samples of solutions to the Toda lattice and the Lotka-Volterra equation, respectively. In this talk we present some recent discoveries that connect diverse topics such as soliton theory, integrable systems, continuous fractions, T functions, orthogonal polynomials, Sylvester identity, moments, and Hankel determinants together. Of particular interest are the three facts that

1. Each of the Toda lattice and the Lotka-Volterra equation governs the evolution of a certain class of orthogonal polynomials whose orthogonality is determined by a specific time-dependent measure.

2. Since the measure deformation is explicitly known, moments can be calculated which, when properly assembled, lead to the conclusion abstractly, but literally, that the iterates of the QR algorithm and the SVD algorithm can be expressed in closed-form!

3. Hankel determinantal solutions are too complicated to be useful. However, a “smart” integrability-preserving discretization of the Lotka-Volterra equation can yield a new SVD algorithm.

(The fact that the SVD algorithm is related to orthogonal polynomials, another subject of Gene’s favorites, has surprised and pleased Gene in our last email communication a few weeks before he passed away.)

 

 Linda Kaufman, William Paterson University, (joint work with Daniel Clark, Serafim Stamatis, Edwin Torres, Ory Diaz, Steve Bang, Phillip Kang, Diamond Flores, Mark Fila, Phillip Nelson and Yomi Idris)

Implementing the Retraction Algorithm for factoring symmetric banded matrices

 

Abstract: Problems arising when modeling structures subject to vibrations or in designing optical fibers wrapped around a spool  give rise to eigenvalue problems and linear systems with symmetric, banded matrices.  We present various variations of the retraction algorithm which factors a matrix  into a product of matrices while preserving symmetry, the bandwidth, and element growth even  when the original matrix is indefinite.  The factorization may be used to solve a system or compute the inertia of a matrix.  It is  based on the algorithm of Bunch and Kaufman, 1977, for symmetric nonbanded systems, which grew out of the first paper that Gene Golub asked Linda Kaufman to referee while she was his graduate student research assistant in 1973.   Theoretically the operation count is about 1/2 that of nonsymmetric banded Gaussian elimination and 2/3 the space. Experimental results will be presented comparing the various versions of the retraction algorithm with LAPACK’s nonsymmetric band routines and the Snap-back code of Irony and Toledo from Tel Aviv University. One variant is based on a theorem by Gill, Golub. Murray and Saunders about the product of Givens transformations.

 

Misha Kilmer, Tufts University, ( joint work with Malena Espanol and Dianne O'Leary)

Multilevel Methods for Image Deblurring Problems   (This presentation was cancelled due to illness)

Abstract:  We present a multilevel method for discrete ill-posed problems formulated
as least squares or total least squares problems. Our approach is meant to preserve edges in the solution, a particularly important feature when applied to signal or image deblurring problems. In this method we use the Haar wavelets as restriction and prolongation operators. We show that the choice of the Haar wavelet operator has the advantage of preserving matrix structure, such as Toeplitz, among grids, and we discuss how this can be exploited to obtain faster solvers on each level.  

 

Steven Leon,  UMass Dartmouth

Remembering Gene Golub, Part I - A Remarkable Career  ---  Part II - Serra House, the Golden Year, 1977 - 78

Abstract:  The talk was presented in two parts. Part I was a brief overview of the career and accomplishments of Gene Golub. This presentation was made as part of the Opening Remarks for the conference. Part II  consisted of remembrances of a very special year at Serra House. The presentation included photos, most of which were provided by Walter Gander. This presentation was part of the Remembrances given at the banquet honoring Gene.

.

.

Svante Littmarck (and Bjorn Sjodin),  Comsol

Multiphysics Simulations

ABSTRACT:  The span of public domain solvers made available by the numerical linear algebra community helps tremendously when developing a software capable of solving general systems of coupled PDEs. The desire for more accurate mathematical models of real-world applications steadily poses new challenges corresponding to open research problems in solver technology and also new demands for integration with existing tools within CAD and other types of engineering software. We will talk about how publicly available solver technology is adapted to fit in a commercial software environment and also what challenges are met in the process.


Michael Neumann, University of Connecticut

Optimization of the Spectral Radius of Sums and Products of Nonnegative Matrices

Abstract:   Let A be an n × n irreducible nonnegative matrix. We will establish that over the set  Ωn,nof all n × n doubly stochastic matrices, the optimum values, minimum and maximum, of the spectral radii of ρ(SA) and ρ(A + S), S in  Ωn , occur at permutation matrices. For the case when A is also symmetric, a by–product of our proof–technique yields a result allowing us to show that ρ(S1A) ≥ ρ(S2A) whenever S1 and S2 are two symmetric matrices such that S1A and S2A and nonnegative matrices and S1S2 is a positive semidefinite matix. This result has several corollaries. One corollary is that ρ(S1A) ≥ ρ(S2A), when S1= (1/n)J and S2= (1/n − 1)(J − I), where J is the n × n matrix of all ones. A second corollary is a comparison theorem for weak regular splittings of two monotone matrices.

      Our proof–techniques are strong enough to show that if all the entries of A are bounded below by 1, then the optimal values of ρ(A + N), as N ranges over all n × n normal matrices with a spectral radius 1, are attained at a real n × n unitary matrix.   The results come from several joint works with J. Axtell, L. Han, D. Hershkowitz, and N. Sze.
 

Vadim Olshevsky , University of Connecticut  (joint work with Tom Bella and Upendra Prasad)
Lipschitz stability of canonical Jordan bases of H-selfadjoint matrices under structure-preserving perturbations

Abstract:  We study Jordan-structure-preserving perturbations of matrices selfadjoint in the indefinite inner product. The main result  is Lipschitz stability of the corresponding so-called similitude matrices. The result can be reformulated as Lipschitz stability, under small perturbations, of canonical Jordan bases (i.e., eigenvectors and generalized eigenvectors enjoying a certain flipped orthonormality relation) of matrices selfadjoint in the indefinite inner product. The proof relies upon the analysis of small perturbations of invariant subspaces, where the size of a permutation of an invariant subspace is measured using the concepts of a gap and of a semigap.

Haesun Park, Georgia Institute of Technology  ( joint work with Jingu Kim and Hyunsoo Kim)
Nonnegative Matrix Factorizations and Clustering
 (This presentation was cancelled due to a family emergency)

Variations of the Nonnegative Matrix Factorization (NMF) are discussed for their application to clustering. We show how interpreting the objective function of K-means as that of a lower rank approximation with special constraints allows comparisons between various formulations of NMF and K-means and provides the insight to design new clustering algorithms. We discuss introduction of additional constraints such as sparsity and orthogonality on the factors in the NMF objective function for the purpose of designing clustering algorithms based on NMF. The experimental results with synthetic and text data shows that these constrained NMF algorithms not only provide an alternative to K-means, but rather give better and consistent solutions to clustering of nonnegative data.
 


Bob Plemmons,
Wake Forest University (joint work with Donghui Chen and Peter Zhang)
Least Squares with Nonnegativity Constraints and Generalizations

Abstract: The work of Gene Golub on least squares computations has greatly influenced statistical applications throughout the mathematical sciences, as well as in engineering.  In the past decade, work by Gene turned to computational, mathematical and statistical techniques for modeling and analyzing modern massive data sets, a promising and flourishing frontier in data and information systems and their applications. In this talk a survey of some of the algorithms for enforcing nonnegativity constraints in least squares computations is given. Generalizations involving approximate low-rank matrix and tensor factorizations are also discussed,  and results from some numerical experiments on massive data sets are provided.

Daniel B Szyld, Temple University

Inexact Krylov Subspace Methods for PDEs and Control Problems

Abstract:  In many circumstances, a known good preconditioner is not easily computable. Instead, an approximation to it is available. This is the case, for example, when the preconditioner has an inverse associated with it, such as in Schur complements (e.g., in saddle point problems), or in the reduced Hessian in some control problems. The application of the preconditioner implies then an iterative solution of a linear system. In these cases, the question is how accurately to solve the (inner) iteration. In our work on Inexact Krylov methods, we have shown that the inner iterations can be solved progressively less accurately, as the underlying Krylov method (e.g, GMRES) converges to the overall solution. Computable inner stopping criteria have been developed to guarantee convergence of the overall method. We will discuss these criteria, and illustrate its application to several problems. In particular, we will report on applying these ideas to parabolic control problems, where the reduced Hessian has two different inverses; and thus two inner iteration criteria are needed.

 

Homer Walker, WPI and US Department of Energy
DIIS Acceleration of Fixed-Point Iteration

Abstract: A certain fixed-point iteration is widely used in electronic structure computations for determining electron densities having minimum total energy. A certain auxiliary procedure, called Direct Inversion on the Iterative Subspace (DIIS), is often used toaccelerate its convergence. I will discuss DIIS as a general acceleration procedure for fixed-point iteration, focussing on a linearly constrained least-squares problem occurring in DIIS and its efficient solution through QR factorization updating.

 

Man-Chung Yeung

An analysis of ML(n)BiCGSTAB

 

Abstract:  ML(n)BiCGSTAB is a Krylov subspace method for the solution of linear systems.It is a natural generalization of BiCGSTAB from the single starting Lanczos to the multiple starting Lanczos. An analysis indicates that the residuals of the approximate solutions in ML(n)BiCGSTAB can be defined in some other ways (other than the definition in Yeung and Chan’s 1997 paper). Some new definitions of the residuals result in new versions of ML(n)BiCGSTAB which have cheaper memory requirements and can converge faster. However, they also tend to diverge from the corresponding true residuals more easily. Numerical experiments will be presented to illustrate this phenomenon.  

.

Pavel Zhlobich

Fast Algorithm for Inversion of  polynomial-Vandermonde matrices related to (H,m)-quasiseparable matrices