Numerical Linear Algebra Group

The supercomputers in the TOP500 list are ranked using the High-Performance Linpack (HPL) Benchmark, which gauges the performance of a computer by solving a large dense linear system by Gaussian elimination with partial pivoting. The size of the coefficient matrix depends on the computational power of the machine being assessed, because more powerful systems require larger benchmark problems in order to reach their peak performance.

The test matrices used by HPL have random entries uniformly distributed on the interval . The 2-norm condition number of such matrices depends on their size, and can potentially be very large for the matrices required by today’s most powerful computers: the largest linear systems solved on Summit, the machines that leads the November 2019 TOP500 ranking, have order , and even larger systems will be needed to benchmark the coming generations of exascale systems.

An matrix with specified 2-norm condition number can be generated as , where and are random real orthogonal matrices from the Haar distribution (a natural uniform probability distribution on the orthogonal matrices) and is diagonal with nonnegative entries . It is well known that has 2-norm condition number if and otherwise. This technique, which is used by the gallery('randsvd', ...) function, requires floating-point operations to generate a test matrix of order , and would thus be impractical in an extreme-scale setting.

In our recent EPrint Generating extreme-scale matrices with specified singular values or condition numbers, Nick Higham and I present four methods that, by giving up the requirement that the matrices and be from the Haar distribution, reduce the cost of the approach above from cubic to quadratic. The matrices generated retain a number of desirable properties that make them a suitable choice for testing linear solvers at scale.

These cheaper algorithms are particularly well suited to distributed-memory environments, since all communication between the nodes can be avoided at the price of a negligible increase in the overall computational cost. They are appealing to MATLAB users too, as the following example demostrates.

n = 10000; kappa = 1e6; mode = 2; rng(1)
% gallery('randsvd',...)
fprintf('gallery(''randsvd'',...):                %5.2f seconds elapsed.\n',...
timeit(@()gallery('randsvd',n,kappa,mode,[],[],1)));
% Algorithm 3.1 in our EPrint.
method = 1; matrix = 0;
fprintf('Alg. 3.1 with Haar U:                  %5.2f seconds elapsed.\n',...
        timeit(@()randsvdfast(n,kappa,mode,method,matrix)));
matrix = 2;
fprintf('Alg. 3.1 with U=gallery(''orthog'',n,2): %5.2f seconds elapsed.\n',...
        timeit(@()randsvdfast(n,kappa,mode,method,matrix)));
% Algorithm 4.1 in our EPrint.
method = 3; matrix = 0;
fprintf('Alg. 4.1 with Haar U:                  %5.2f seconds elapsed.\n',...
        timeit(@()randsvdfast(n,kappa,mode,method,matrix)));
matrix = 2;
fprintf('Alg. 4.1 with U=gallery(''orthog'',n,2): %5.2f seconds elapsed.\n',...
        timeit(@()randsvdfast(n,kappa,mode,method,matrix)));

In the listing above, randsvdfast is an implementation of our algorithms available on the MATLAB Central File Exchange. Setting the argument matrix to 0 tells our function to pick from the Haar distribution, whereas setting it to 2 causes to be chosen in a non-random way. Running this code in MATLAB 2019b on a machine equipped with an Intel processor I5-6500 running at 3.20GHz the script produces the output

gallery('randsvd',...):                79.52 seconds elapsed.
Alg. 3.1 with Haar U:                  19.70 seconds elapsed.
Alg. 3.1 with U=gallery('orthog',n,2):  1.90 seconds elapsed.
Alg. 4.1 with Haar U:                  19.28 seconds elapsed.
Alg. 4.1 with U=gallery('orthog',n,2):  1.43 seconds elapsed.

Therefore, for randsvdfast is up to 56 times faster than gallery('randsvd',...).

Most backward error bounds for numerical linear algebra algorithms are of the form , for a machine precision and a problem size . The dependence on of these bounds is known to be pessimistic: together with Nick Higham, our recent probabilistic analysis [SIAM J. Sci. Comput., 41 (2019), pp. A2815–A2835], which assumes rounding errors to be independent random variables of mean zero, proves that can be replaced by a small multiple of with high probability. However, even these smaller bounds can still be pessimistic, as the figure below illustrates.

The figure plots the backward error for summation (in single precision) of floating-point numbers randomly sampled from a uniform distribution. For numbers in the distribution, the bound is almost sharp and accurately predicts the error growth. However, for the distribution, the error is much smaller, seemingly not growing with . This strong dependence of the backward error on the data cannot be explained by the existing bounds, which do not depend on the values of the data.

In our recent preprint, we perform a new probabilistic analysis that combines a probabilistic model of the rounding errors with a second probabilistic model of the data. Our analysis reveals a strong dependence of the backward error on the mean of the data : indeed, our new backward error bounds are proportional to . Therefore, for data with small or zero mean, these new bounds are much sharper as they bound the backward error by a small multiple of the machine precision independent of the problem size .

Motivated by this observation, we also propose new algorithms that transform the data to have zero mean, so as to benefit from these more favorable bounds. We implement this idea for matrix multiplication and show that our new algorithm can produce significantly more accurate results than standard matrix multiplication.

Embedded systems are based on low-power, low-performance processors and can be found in various medical devices, smart watches, various communication devices, cars, planes, mobile phones and many other places. These systems come in a hardware and software package optimized for specific computational tasks and most commonly have real-time constraints. As these systems usually have energy usage and cost constraints too, sophisticated numerical hardware that can process floating-point data is not included, but rather only integer arithmetic, which is simpler in terms of area and power of the processors.

ISO 18037:2008 is a standard for embedded C programming language support. It lays out various rules that C compilers should support to make embedded systems easier to program using a high-level language. One of the most important definitions in this standard is fixed-point arithmetic data types and operations. Support for fixed-point arithmetic is highly desirable, since if it is not provided integers with scaling factors have to be used, which makes code hard to maintain and debug and most commonly requires assembler level changes or completely new implementations for each different platform.

The GCC compiler provides some support of the fixed-point arithmetic defined in this standard for ARM processors. However, in my recent technical report (https://arxiv.org/abs/2001.01496) I demonstrated various numerical pitfalls that programmers of embedded systems based on ARM and using GCC can get into. The issues demonstrated include

• larger than half machine epsilon errors in rounding decimal constants to fixed-point data types,
• errors in conversions between different data types,
• incorrect pre-rounding of arguments of mixed-format arithmetic operations before the operation is performed, and
• lack of rounding of the outputs of arithmetic operations.

These findings can be used to improve the accuracy of various embedded numerical libraries that might be using this compiler. To demonstrate one of the issues, here is a piece of test code:

The multiplication operation is a mixed-format operation, since it multiplies an unsigned long fract argument with an accum argument, therefore it is subject to prerounding of the unsigned long fract argument as described in the report. Since the comparison step in the if () sees that the argument a is larger than zero and b larger than 1, the code is executed with a hope that c will not be set to zero. However, in the arithmetic operation, a is incorrectly pre-rounded to 0, which causes c = 0*b, an unexpected outcome and a bug that is hard to detect and fix.

In many applications requiring the solution of a linear system , the matrix possesses a block low-rank (BLR) property: most of its off-diagonal blocks are of low numerical rank and can therefore be well approximated by low-rank matrices. This property arises for example in the solution of discretized partial differential equations, because the numerical rank of a given off-diagonal block is closely related to the distance between the two subdomains associated with this block. BLR matrices have been exploited to significantly reduce the cost of solving , in particular in the context of sparse direct solvers such as MUMPS, PaStiX, and STRUMPACK.

However, the impact of these low-rank approximations on the numerical stability of the solution in floating-point arithmetic has not been previously analyzed. The difficulty of such an analysis lies in the fact that, unlike for classical algorithms without low-rank approximations, there are two kinds of errors to analyze: floating-point errors (which depend on the unit roundoff ), and low-rank truncation errors (which depend on the the low-rank threshold , the parameter controlling the accuracy of the blockwise low-rank approximations). Moreover, the two kinds of error cannot easily be isolated in the analysis, because the BLR compression and factorization stages are often interlaced.

In our recent preprint, with Nick Higham, we present rounding error analysis for the solution of a linear system by LU factorization of BLR matrices. Assuming that a stable pivoting scheme is used, we prove backward stability: the relative backward error is bounded by a modest constant times , and does not depend on the unit roundoff as long as is safely smaller than . This is a very desirable theoretical guarantee that can now be given to the users, who can therefore control the numerical behavior of BLR solvers simply by setting to the target accuracy. We illustrate this key result in the figure below, which shows a strong correlation between the backward error and the threshold for 26 matrices from the SuiteSparse collection coming from a wide range of real-life applications.

The adjective “Bohemian” was used for the first time in a linear algebra context by Robert Corless and Steven Thornton to describe the eigenvalues of matrices whose entries are taken from a finite discrete set, usually of integers. The term is a partial acronym for “BOunded Height Matrix of Integers”, but the origin of the term was soon forgotten, and the expression “Bohemian matrix” is now widely accepted.

As Olga Taussky observed already in 1960, the study of matrices with integer elements is “very vast and very old”, with early work of Sylvester and Hadamard that dates back to the second half of the nineteenth century. These names are the first two in a long list of mathematicians that worked on what is now known as the “Hadamard conjecture”: for any positive integer multiple of 4, there exists an by matrix , with entries and , such that .

If this is the best-known open problem surrounding Bohemian matrices, it is far from being the only one. During the 3-day workshop “Bohemian Matrices and Applications” that our group hosted in June last year, Steven Thornton released the Characteristic Polynomial Database, which collects the determinants and characteristic polynomials of billions of samples from certain families of structured as well as unstructured Bohemian matrices. All the available data led Steven to formulate a number of conjectures regarding the determinants of several families of Bohemian upper Hessenberg matrices.

Gian Maria Negri Porzio and I attended the workshop, and set ourselves the task of solving at least one of these open problems. In our recent preprint, we enumerate all the possible determinants of Bohemian upper Hessenberg matrices with ones on the subdiagonal. We consider also the special case of families with main diagonal fixed to zero, whose determinants turn out to be related to some generalizations of Fibonacci numbers. Many of the conjectures stated in the Characteristic Polynomial Database follow from our results.

Summing numbers is a key computational task at the heart of many numerical algorithms. When performed in floating-point arithmetic, summation is subject to rounding errors: for a machine precision , the error bound for the most basic summation algorithms, such as recursive summation, is proportional to .

Nowadays, with the growing interest in low floating-point precisions and ever increasing in applications, such error bounds have become unacceptably large. While summation algorithms leading to smaller error bounds are known (compensated summation is an example), they are computationally expensive.

In our recent preprint, Pierre Blanchard, Nick Higham and I propose a class of fast and accurate summation algorithms called FABsum. We show that FABsum has an error bound of the form , where is a block size, which is independent of to first order. As illustrated by the figure below, which plots the measured error using single precision as a function of , FABsum can deliver substantially more accurate results than recursive summation. Moreover, FABsum can be easily incorporated in high-performance numerical linear algebra kernels in order to boost accuracy with only a modest drop in performance, as we demonstrate in the paper with the PLASMA library.

The wave-kernel functions and arise in the solution of second order differential equations such as with initial conditions at . Here, is an arbitrary square matrix and . The square root in these formulas is illusory, as both functions can be expressed as power series in , so there are no questions about existence of the functions.

How can these functions be computed efficiently? In Computing the Wave-Kernel Matrix Functions (SIAM J. Sci. Comput., 2018) Prashanth Nadukandi and I develop an algorithm based on Padé approximation and the use of double angle formulas. The amount of scaling and the degree of the Padé approximant are chosen to minimize the computational cost subject to achieving backward stability for in exact arithmetic.

In the derivation we show that the backward error of any approximation to can be explicitly expressed in terms of a hypergeometric function. To bound the backward error we derive and exploit a new bound for in terms of the norms of lower powers of ; this bound is sharper than one previously obtained by Al-Mohy and Higham.

Numerical experiments show that the algorithm behaves in a forward stable manner in floating-point arithmetic and is superior in this respect to the general purpose Schur–Parlett algorithm applied to these functions.

The solution of a linear system is a fundamental task in scientific computing. Two main classes of methods to solve such a system exist.

• Direct methods compute a factorization of matrix , such as LU factorization, to then directly obtain the solution by triangular substitution; they are very reliable but also possess a high computational cost, which limits the size of problems that can be tackled.
• Iterative methods compute a sequence of iterates converging towards the solution ; they are inexpensive but their convergence and thus reliability strongly depends on the matrix properties, which limits the scope of problems that can be tackled.

A current major challenge in the field of numerical linear algebra is therefore to develop methods that are able to tackle a large scope of problems of large size.

To accelerate the convergence of iterative methods, one usually uses a preconditioner, that is, a matrix ideally satisfying three conditions: (1) is cheap to compute; (2) can be easily inverted; (3) is a good approximation to . With such a matrix , the preconditioned system is then cheap to solve with an iterative method and often requires a small number of iterations only. An example of a widely used class of preconditioners is when is computed as a low-accuracy LU factorization.

Unfortunately, for many important problems it is quite difficult to find a preconditioner that is both of good quality and cheap to compute, especially when the matrix is ill conditioned, that is, when the ratio between its largest and smallest singular values is large.

In our paper A New Preconditioner that Exploits Low-rank Approximations to Factorization Error, with Nick Higham, which recently appeared in SIAM Journal of Scientific Computing, we propose a novel class of general preconditioners that builds on an existing, low-accuracy preconditioner .

This class of preconditioners is based on the following key observation: ill-conditioned matrices that arise in practice often have a small number of small singular values. The inverse of such a matrix has a small number of large singular values and so is numerically low rank. This observation suggests that the error matrix is of interest, because we may expect to retain the numerically low-rank property of .

In the paper, we first investigate theoretically and experimentally whether is indeed numerically low rank; we then describe how to exploit this property to accelerate the convergence of iterative methods by building an improved preconditioner , where is a low-rank approximation to . This new preconditioner is equal to and is thus almost a perfect preconditioner if . Moreover, since is a low-rank matrix, can be cheaply computed via the Sherman–Morrison–Woodbury formula, and so the new preconditioner can be easily inverted.

We apply this new preconditioner to three different types of approximate LU factorizations: half-precision LU factorization, incomplete LU factorization (ILU), and block low-rank (BLR) LU factorization. In our experiments with GMRES-based iterative refinement, we show that the new preconditioner can achieve a significant reduction in the number of iterations required to solve a variety of real-life problems.

According to Moler and Van Loan, truncating the Taylor series expansion to the exponential at 0 is the least effective of their nineteen dubious ways to compute the exponential of a matrix. Such an undoubtedly questionable approach can, however, become a powerful tool for evaluating matrix functions when used in conjunction with other strategies, such as, for example, the scaling and squaring technique. In fact, truncated Taylor series are just a special case of a much larger class of rational approximants, known as the Padé family, on which many state-of-the-art algorithms rely in order to compute matrix functions.

A customary choice for evaluating these approximants at a matrix argument is the Paterson–Stockmeyer method, an evaluation scheme that was originally proposed as an asymptotically optimal algorithm for evaluating polynomials of matrices, but generalizes quite naturally to rational functions, which are nothing but the solutions of a linear system whose coefficients and right-hand side are polynomials of the same matrix. This technique exploits a clever rewriting of a polynomial in as a polynomials in , for some positive integer , and overall requires about matrix products to evaluate a polynomial of degree .

As shown in the figure, when the Paterson–Stockmeyer scheme is used the number of matrix multiplications required to evaluate a polynomial of degree grows slower than itself, with the result that evaluating polynomials of different degree will asymptotically have the same cost. For example, evaluating a polynomial of any degree between 43 and 49 requires 12 matrix multiplications, thus there is little point in considering an approximant of degree 43 when evaluating that of degree 49 has roughly the same cost but will in all likelihood deliver a more accurate result.

When designing algorithms to evaluate functions of matrices, one is interested in finding the optimal degrees, those marked with a red circle in the figure above, since they guarantee maximal accuracy for a given computational cost. When fixed precision is considered, finding all such degrees is not a problem: a backward error analysis can be used to determine the maximum degree that will ever be needed, say, and then looking at the plot is enough to find all the optimal degrees smaller than . In order to deal with arbitrary precision arithmetic, however, a different strategy is needed, as depending on the working precision and the required tolerance, approximants of arbitrarily high degree may be needed. The new Eprint Optimality of the Paterson–Stockmeyer Method for Evaluating Matrix Polynomials and Rational Matrix Functions studies the cost of the Paterson–Stockmeyer method for polynomial evaluation and shows that a degree is optimal if and only if it is a quarter-square, that is, a number of the form for some nonnegative integer , where is the floor function.

Similar results can be obtained for Paterson–Stockmeyer-like algorithms for evaluating diagonal Padé approximants, rational functions whose numerator and denominator have same degree. In that case, one can show that an approximant is optimal if and only if the degree of numerator and denominator is an eight-square, an integer of the form for some . In particular, for diagonal Padé approximants to the exponential, due to a symmetry of the coefficients of numerator and denominator, faster algorithms can be developed, and an explicit formula—not as nice as that in the two previous cases—can be derived for the optimal orders of these approximants.