Xinshuo Yang

Gregory Beylkin, Lucas Monzon, Xinshuo Yang

We present a new adaptive method for electronic structure calculations based on novel fast algorithms for reduction of multivariate mixtures. In our calculations, spatial orbitals are maintained as Gaussian mixtures whose terms are selected in the process of solving equations. Using a fixed basis leads to the so-called "basis error" since orbitals may not lie entirely within the linear span of the basis. To avoid such an error, multiresolution bases are used in adaptive algorithms so that basis functions are selected from a fixed collection of functions, large enough as to approximate solutions within any user-selected accuracy. Our new method achieves adaptivity without using a multiresolution basis. Instead, as a part of an iteration to solve nonlinear equations, our algorithm selects the "best" subset of linearly independent terms of a Gaussian mixture from a collection that is much larger than any possible basis since the locations and shapes of the Gaussian terms are not fixed in advance. Approximating an orbital within a given accuracy, our algorithm yields significantly fewer terms than methods using multiresolution bases. We demonstrate our approach by solving the Hartree-Fock equations for two diatomic molecules, HeH+ and LiH, matching the accuracy previously obtained using multiwavelet bases.

Gregory Beylkin, Lucas Monzon, Xinshuo Yang

We consider fast deterministic algorithms to identify the "best" linearly independent terms in multivariate mixtures and use them to compute, up to a user-selected accuracy, an equivalent representation with fewer terms. One algorithm employs a pivoted Cholesky decomposition of the Gram matrix constructed from the terms of the mixture to select what we call skeleton terms and the other uses orthogonalization for the same purpose. Importantly, the multivariate mixtures do not have to be a separated representation of a function. Both algorithms require $O(r^2 N + p(d) r N) $ operations, where $N$ is the initial number of terms in the multivariate mixture, $r$ is the number of selected linearly independent terms, and $p(d)$ is the cost of computing the inner product between two terms of a mixture in $d$ variables. For general Gaussian mixtures $p(d) \sim d^3$ since we need to diagonalize a $d\times d$ matrix, whereas for separated representations $p(d) \sim d$. Due to conditioning issues, the resulting accuracy is limited to about one half of the available significant digits for both algorithms. We also describe an alternative algorithm that is capable of achieving higher accuracy but is only applicable in low dimensions or to multivariate mixtures in separated form. We describe a number of initial applications of these algorithms to solve partial differential and integral equations and to address several problems in data science. For data science applications in high dimensions,we consider the kernel density estimation (KDE) approach for constructing a probability density function (PDF) of a cloud of points, a far-field kernel summation method and the construction of equivalent sources for non-oscillatory kernels (used in both, computational physics and data science) and, finally, show how to use the new algorithm to produce seeds for subdividing a cloud of points into groups.

Zhihui Zhu, Qiuwei Li, Xinshuo Yang et al.

We study the convergence of a variant of distributed gradient descent (DGD) on a distributed low-rank matrix approximation problem wherein some optimization variables are used for consensus (as in classical DGD) and some optimization variables appear only locally at a single node in the network. We term the resulting algorithm DGD+LOCAL. Using algorithmic connections to gradient descent and geometric connections to the well-behaved landscape of the centralized low-rank matrix approximation problem, we identify sufficient conditions where DGD+LOCAL is guaranteed to converge with exact consensus to a global minimizer of the original centralized problem. For the distributed low-rank matrix approximation problem, these guarantees are stronger---in terms of consensus and optimality---than what appear in the literature for classical DGD and more general problems.