Francoise Tisseur is Professor of Numerical Analysis in the School of Mathematics, University of Manchester, UK. Tell us about yourself and your work. ‘Tam a numerical analyst specializing in numerical linear algebra. My work spans the full range from developing fundamental theory to deriving algorithms and implementing them in software Tam particulary interested in the theory and numerical solution of algebraic cigenvalue problems, that i, to find scalars and non-zero vectors x and y satisfying N(a)x = 0, where N : > Crxn is an analytic function on an open set 2 ¢ €, the standard eigenvalue problem (A — Ai)x = 0 studied in Chapter 7 being the simplest «ase, However, in practical applications, such as inthe dynamic analysis of mechanical systems (where the eigenvalues represent vibrational frequencies), inthe linear stability ‘of flows in fluid mechanics, or in the stability analysis of time-delay systems, the clements of N(2) can be polynomial, rational or exponential functions of & and the corresponding non-linear eigenvalue problems can be very difficult to solve. Tam also interested in tropical algebra (also known as max-plus algebra), where smatices and vectors have entries in RU (0) and where the addition a + bis replaced bya maximization max(a, 6) and the multiplication ab is replaced by an addition a +B. [As in section 1.4 of this book, we ean add two tropical matrices and also multiply them. In fact, many of the tools of linear algebra described in this book are available in tropical algebra. A major difference with classical linear algebra is that the maximum operation lacks inverses. Tropical algebra allows us to describe, ina linear way, a phenomenon that is non-linear in the conventional algebra, These inciude, for example, parallel computation, transportation networks and scheduling, What are the challenges connected with the subject? ‘Non-linear eigenvalue problems present many mathematical challenges, For some there isa lack of underlying theory. For others, numerical methods struggle to provide any accuracy or to solve very large problems in a reasonable time,Petros Drineas is Associate Professor of Computer Science at the Rensselaer Polytechnic Institute in New York. ho Tell us about yourself and your work. incar algebra isa fundamental tool in the design and analysis of algorithms for data ‘mining and statistical data analysis applications: numerous modern, massive datasets are ‘modelled as matrices, Naturally a solid background in linear algebra isa must if your ‘work will eventually involve processing large-scale data, and many highly rewarding obs in the fature will involve the analysis of Big Data. Think for example, ofthe following «data matrix: the rows ofthe matsix corzespond to customers of your favourite online store, and the columns correspond to the products availble atthe store, Let the ()-th entry of the matrix be the ‘uty’ of produetj for customer i. Then, a recommender system like the ones used by many such online stores, seeks to xecommend high utility products to customers, with the obvious hope that the customer will purchase the product. Since itis not possible to ask every customes to reveal his or her preference for ‘every produc, recommender system tres to infer products of high utility for a ‘customer from a small sample of known entries of the matrix. lot of my work focuses on problems ofthis type: how ean we reconstruct missing entries of a matrix from a small sample of rows and/or columns andor elements ofthe matrix? Solving this problem better than the current tate ofthe art would have made you a millionaire in the [Netflix challenge a couple of years ago! ‘Areated question focuses on identifying inluemtial zows and/or columns and/or clements ofa matrix, Imagine, for example, that you ate intrested in approximating the product of two matrices by selecting afew rows ofthe frst matrix and a few columns of ‘the second matrix. Which rows and or columns will return the best approximation? Such problems are of particular interest when dealing with massive matrices that cannot be ‘asiy stored in random access memory (RAM). Similar sampling based algorithms are very useful in approximating singular values and singular vectors of matrices. These operations are of particular importance in data mining, since they le atthe heart of principal components analysis (PCA), a fundamental dimensionality reduction technique that helps visualize, denoise and interpret high-dimensional datasets ‘The challenges in my line of work lie atthe intersection of linear algebra and probability theory. This ‘marriage’ ofthe wo domains as well asthe formal study of the ‘effects of sampling rows/columns/elements of matrices is relatively recent, and you can trace ito afew influential papers in the late 1990s. On the one hand, probability theoryprovides the fundamental tools that help quantify the ‘distance’ between the original ‘matrix and the sampled one. On the other hand, linear algebraic tools (for example, ‘matrix perturbation theory), allow us to quantify how characteristic properties (cg, singular values, singular vectors, etc) of the original matrix can be approximated using the sampled matrix. Have you any particular messages that you would like to give to students starting off studying linear algebra? ‘You need to understand singular value decomposition (SVD)! This will probably be tovards the end of an introductory linear algebra course, but it rally should be the take-home message of any such course. It ies atthe heart of so many things, from ‘computing the rank of a matrix, to solving least-square problems and computing the inverse or pseudoinverse of a matrix. I will quote Prof. Dianne O'Leary ofthe University ‘of Maryland: SVD is the Swiss Army knife and the Rolls Royce of numerical linear algebra. Ithas numerous uses (ike a Swiss Army knife), bu itis computationally expensive (like a Rolls Royce). Fhope that some of my own work on randomized algorithms for the SVD has helped make it bit more affordable!Anshul Gupta is a Research Staff Member in the HPC Group, Business Analytics and Mathematical Sciences, at the IBM Thomas J. Watson Research Center in New York. Tell us about yourself and your work. Although my background is in computer science, my work i atthe confluence of ‘computer science and numerical linear algebra, It often involves applying computer science techniques to develop scalable parallel algorithms and software for numerical ‘methods used in scientific computing and optimization, How do you use linear algebra in your job, how important is linear algebra and what are the challenges connected with the subject? Thoth use and develop linear algebra algorithms and software. The biggest challenge that ace is that there is no software for many numerical linear algebra algorithms that are published in linear algebra journals such ae SIMAX, SISC, etc. The corresponding papers typically present limited experimental evidence of the effectiveness ofthe proposed techniques anda reliable comparison with competing techniques is even harder to find. What are the key issues in your area of linear algebra research? ‘There are two important issues 1) lack of resilience of linear algebra algorithms and software to soft errors on massively parallel machines, 2) scalability himiting effects of global communication and synchronization on linear algebra software on massively parallel machines Addressing both these issues requires the developing of novel algorithmic techniques. Have you any particular messages that you would like to give to students starting off studying linear algebra? Always pay attention to how a linear algebra algorithm will perform on real problems on real computers. Asymptotic bounds based on operation counts are almost obsolete when the operations ae virtually fre and almost ll the cost isin data transfer. Different algorithms with similar complexities can have different spatial and temporal locakty properties, which have a huge impact on the behaviour ofthe algorithm ina realistic settingJanet Drew is Professor of Astrophysics and Assistant Dean of School (Research) at the University of Hertfordshire, UK. Tell us about yourself and your work: Tam a research astrophysicist and university professor who qualified via a physics undergraduate degree and astrophysies PhD. Over the past 30 years Thave worked in the UK successively at University College London, Cambridge, Oxford and Imperial College - moving to Hertfordshire a few years ago. For two-thirds of my research career Ihave operated as a theoretical astronomer specialising in the modelling of ionised ‘outflows from hot stare and accretion disks in various settings. More recently the ‘emphasis has shifted to observational astronomy through leading the international consortium of scientists carrying out digital optical broadband surveys, with narrowband H-alpha, of the Milky Way (see e.g. ttp:/www-iphas.org). How do you use linear algebra in your job? It was more important to me asa theoretical astrophysicist, because then the modelling, involved the numerical solution of mildly non-linear systems of simullancous equations, ‘generally rendered in linear form to make them tractable, Although less important now, ithasn’t gone away because it turns out that one of the best methods of establishing a global uniform calibration to wide-field astronomical survey data is to set up and solve a large system of simultancous equations (of dimension several thousand), How important is linear algebra? ‘Very. The need to obtain reliable solutions to systems of linear equations is extremely ‘common in the world of physical science, Its everywhere, frankly. As Iset out on my research career, after oblaining my PAD, one of the first book purchases I needed to make ‘was a textbook in numerical analysis (the real-world computational application of linear algebra, essentially) to sit by my side as I found ways to model radiation transport effects in the statistical equilibrium of many-level hydrogen atoms. Hydrogen, incidentally, is far and away the most abundant and therefore most important element in the Universe - ithas to be described well in understanding its gas content both in stars and in the space in between them,What are the challenges connected with the subject? 1€T may subtly redefine the subject as numerical analysis, one of the challenges isto find ‘ways to test that you really are solving the system(s) of equations you want to solve in non-trivial cases. The issue of convergence (and efficiency) arises too, for schemes that aim to deal with non-linearity through iteration, It doesn't usually take too long to set up an algorithm, but it takes a whole heap longer to validate them, Have you any particular messages that you would like to give to students starting off studying linear algebra? Just that it really has more applications than you can even begin to imagine when you first meet the nuts-and-bolts mathematics at school and undergraduate level. Matrices are your friends: treat them nicely and they will do so much for you.Chao Yang is a Staff Scientist in the Scientific Computing Group at the Lawrence Berkeley National Laboratory (LBNL) in California. Tell us about yourself and your work. Thave been working on numerical linear algebra since graduate school (Rice University) [My expertise is in solving large-scale eigenvalue problems. At LBNL, Twork on a number of interesting ecience projects. All of them require using linear algebra to obtain desirable solutions How do you use linear algebra in your job? use inear algebra to solve quantum mechanics problems. One of the fundamental problems in chemistry and material science is to understand the electronic structure of atoms, molecules and solids. This problem can be solved by computing the solutions to Schréidinger’s equation, which isan eigenvalue problem, There are various ways to obtain approximations to the solution. All of them ultimately result in solving a large sparse matrix eigenvalue problem. How important is linear algebra? Linear algebra is indispensable in my research, Every problem eventually boils down to a linear algebra problem, What are the challenges connected with the subject? ‘The main challenge is the size ofthe problem, which nowadays can be in the order of millions or even billions. However, these large problems often have structures (eg sparsity). Therefore, to solve these problems, one has to take advantage ofthe structure What are the key issues in your area of linear algebra research? ‘Convergence of iterative solvers for linear equations, least squares problems and cigenvalue problems, methods that can take advantage of the sparsity or other structures of matricesDes Higham is the 1966 Professor of Numerical Analysis in the Department of Mathematics and Statistics at the University of Strathclyde in Glasgow, UK. Tell us about yourself and your work. I develop and apply mathematical models and computational algorithms. As a ‘mathematician, Iam mostly interested in proving results, for example, showing that a ‘model can reproduce known behaviour of a physical system and then using it to make predictions that can be tested experimentally. On the algorithmic side, I aim to prove that a computational technique solves the required problem accurately and effectively However, [like to work alongside colleagues from other areas, including the life sciences and e-business. This isa great way for me to get hold of realistic data sets, and it often ‘throws up challenges that inspire new mathematical ideas, So I find it very rewarding to ‘have my research stimulated and stress-tested by outside applications. How do you use linear algebra in your job? ‘A good example is my research in network science. Here we are typically given data about pairwise interactions (who phoned who, who emailed who, who follows who on ‘Twittr). This typeof information i naturally represented by a matrix, and many ofthe interesting questions (who are the key players, where are the bottlenecks, where isthe best place to learn the latest rumour, and can we categorize people into groups?) can be ‘thought of as problems in linear algebra How important is linear algebra? ‘Tasks such as solving linear systems and linear least squares problems and computing cigenvalues/eigenvectors are the nuts and bolts that go into most ofthe algorithms used in scientific computing. What are the challenges connected with the subject? In today’s high-tech digital world, we are being swamped with data In fact, a lot of itis about us! There is behavioural information about our on-line activities, our digital ‘communication patterns and our supermarket spending habits, and new experimental techniques are generating more fundamental biological information about our DNAsequences and the behaviour of our genes and our neurons, We need bigger computers and smarter algorithms to keep up with the data deluge. What are the key issues in your area of linear algebra research? At the moment, Iam focusing on two aspects First, modelling issues for networks: what laws of motion’ are responsible for the patterns of interactions that we see in digital social media; ie, how do matrix entries change over time? Seconé, what can we usefully do with large scale interaction data; ie. how do we summarize a giant matrix? Have you any particular messages that you would like to give to students starting off studying linear algebra? You only have to look atthe success of Google, and its underlying PageRank algorithm, to geta fel forthe relevance and power of linear algebra. Ifyou want to understand {ideas in mathematics and statistics, and develop a useful et of sil, linea algebra isthe place to start