Martin J. Mohlenkamp

Algorithms for Numerical Analysis in High Dimensions

Nearly every numerical analysis algorithm has computational complexity that scales exponentially in the underlying physical dimension. The separated representation, introduced previously, allows many operations to be performed with scaling that is formally linear in the dimension. In this paper we further develop this representation by (i) discussing the variety of mechanisms that allow it to be surprisingly efficient; (ii) addressing the issue of conditioning; (iii) presenting algorithms for solving linear systems within this framework; and (iv) demonstrating methods for dealing with antisymmetric functions, as arise in the multiparticle Schrodinger equation in quantum mechanics. Numerical examples are given.

Numerical operator calculus in higher dimensions

When an algorithm in dimension one is extended to dimension d, in nearly every case its computational cost is taken to the power d. This fundamental difficulty is the single greatest impediment to solving many important problems and has been dubbed the curse of dimensionality. For numerical analysis in dimension d, we propose to use a representation for vectors and matrices that generalizes separation of variables while allowing controlled accuracy. Basic linear algebra operations can be performed in this representation using one-dimensional operations, thus bypassing the exponential scaling with respect to the dimension. Although not all operators and algorithms may be compatible with this representation, we believe that many of the most important ones are. We prove that the multiparticle Schrödinger operator, as well as the inverse Laplacian, can be represented very efficiently in this form. We give numerical evidence to support the conjecture that eigenfunctions inherit this property by computing the ground-state eigenfunction for a simplified Schrödinger operator with 30 particles. We conjecture and provide numerical evidence that functions of operators inherit this property, in which case numerical operator calculus in higher dimensions becomes feasible.

Multivariate Regression and Machine Learning with Sums of Separable Functions

We present an algorithm for learning (or estimating) a function of many variables from scattered data. The function is approximated by a sum of separable functions, following the paradigm of separated representations. The central fitting algorithm is linear in both the number of data points and the number of variables and, thus, is suitable for large data sets in high dimensions. We present numerical evidence for the utility of these representations. In particular, we show that our method outperforms other methods on several benchmark data sets.

Approximating a wavefunction as an unconstrained sum of Slater determinants

The wavefunction for the multiparticle Schrodinger equation is a function of many variables and satisfies an antisymmetry condition, so it is natural to approximate it as a sum of Slater determinants. Many current methods do so, but they impose additional structural constraints on the determinants, such as orthogonality between orbitals or an excitation pattern. We present a method without any such constraints, by which we hope to obtain much more efficient expansions and insight into the inherent structure of the wavefunction. We use an integral formulation of the problem, a Green’s function iteration, and a fitting procedure based on the computational paradigm of separated representations. The core procedure is the construction and solution of a matrix-integral system derived from antisymmetric inner products involving the potential operators. We show how to construct and solve this system with computational complexity competitive with current methods.

Trigonometric identities and sums of separable functions

Modern computers have made commonplace many calculations that were impossible to imagine a few years ago. Still, when you face a problem with a high physical dimension, you immediately encounter the Curse of Dimensionality [1, p.94]. This curse is that the amount of computing power that you need grows exponentially with the dimension. The simplest manifestation of this curse appears when you try to represent a function by its sample values on a grid. If a function of one variable requires N samples, then an analogous function of n variables will need a grid of N samples. Thus, even relatively small problems in high dimensions are still unreasonably expensive. A method has been proposed in [2] to address this problem, based on approximating a function by a sum of separable functions:

Learning to Predict Physical Properties using Sums of Separable Functions

We present an algorithm for learning the function that maps a material structure to its value on some property, given the value of this function on several structures. We pose this problem as one of learning (regressing) a function of many variables from scattered data. Each structure is first converted to a weighted set of points by a process that removes irrelevant translations and rotations but otherwise retains full information about the structure. Then, incorporating a weighted average for each structure, we construct the multivariate regression function as a sum of separable functions, following the paradigm of separated representations. The algorithm can treat all finite and periodic structures within a common framework, and in particular does not require all structures to lie on a common lattice. We show how the algorithm simplifies when the structures do lie on a common lattice, and we present numerical results for that case.

A center-of-mass principle for the multiparticle Schrödinger equation

The center-of-mass principle is the key to the rapid computation of the interaction of a large number of classical particles. Electrons governed by the multiparticle Schrodinger equation have a much more complicated interaction mainly due to their spatial extent and the antisymmetry constraint on the total wave function of the combined electron system. We present a center-of-mass principle for quantum particles that accounts for this spatial extent, the antisymmetry constraint, and the potential operators. We use it to construct an algorithm for computing a size-consistent approximate wave function for large systems with simple geometries.

Function space requirements for the single-electron functions within the multiparticle Schrödinger equation

Our previously described method to approximate the many-electron wavefunction in the multiparticle Schrodinger equation reduces this problem to operations on many single-electron functions. In this work, we analyze these operations to determine which function spaces are appropriate for various intermediate functions in order to bound the output. This knowledge then allows us to choose the function spaces in which to control the error of a numerical method for single-electron functions. We find that an efficient choice is to maintain the single-electron functions in L2 ∩ L4, the product of these functions in L1 ∩ L2, the Poisson kernel applied to the product in L4, a function times the Poisson kernel applied to the product in L2, and the nuclear potential times a function in L4/3. Due to the integral operator formulation, we do not require differentiability.