arXiv subjects

Anil Damle

Publications and source records attributed to Anil Damle.

36 records · Page 2Linked to original sources

arXiv2021

Network interpolation

Given a set of snapshots from a temporal network we develop, analyze, and experimentally validate a so-called network interpolation scheme. Our method allows us to build a plausible, albeit random, sequence of graphs that transition between any two given graphs. Importantly, our model is well characterized by a Markov chain, and we leverage this representation to analytically estimate the hitting time (to a predefined distance to the target graph) and long term behavior of our model. These observations also serve to provide interpretation and justification for a rate parameter in our model. Lastly, through a mix of synthetic and real-world data experiments we demonstrate that our model builds reasonable graph trajectories between snapshots, as measured through various graph statistics. In these experiments, we find that our interpolation scheme compares favorably to common network growth models, such as preferential attachment and triadic closure.

cs.SI↗

arXiv2020

Fast Matrix Square Roots with Applications to Gaussian Processes and Bayesian Optimization

Matrix square roots and their inverses arise frequently in machine learning, e.g., when sampling from high-dimensional Gaussians $\mathcal{N}(\mathbf 0, \mathbf K)$ or whitening a vector $\mathbf b$ against covariance matrix $\mathbf K$. While existing methods typically require $O(N^3)$ computation, we introduce a highly-efficient quadratic-time algorithm for computing $\mathbf K^{1/2} \mathbf b$, $\mathbf K^{-1/2} \mathbf b$, and their derivatives through matrix-vector multiplication (MVMs). Our method combines Krylov subspace methods with a rational approximation and typically achieves $4$ decimal places of accuracy with fewer than $100$ MVMs. Moreover, the backward pass requires little additional computation. We demonstrate our method's applicability on matrices as large as $50,\!000 \times 50,\!000$ - well beyond traditional methods - with little approximation error. Applying this increased scalability to variational Gaussian processes, Bayesian optimization, and Gibbs sampling results in more powerful models with higher accuracy.

cs.LG↗

arXiv2020

Entrywise convergence of iterative methods for eigenproblems

Several problems in machine learning, statistics, and other fields rely on computing eigenvectors. For large scale problems, the computation of these eigenvectors is typically performed via iterative schemes such as subspace iteration or Krylov methods. While there is classical and comprehensive analysis for subspace convergence guarantees with respect to the spectral norm, in many modern applications other notions of subspace distance are more appropriate. Recent theoretical work has focused on perturbations of subspaces measured in the $\ell_{2 \to \infty}$ norm, but does not consider the actual computation of eigenvectors. Here we address the convergence of subspace iteration when distances are measured in the $\ell_{2 \to \infty}$ norm and provide deterministic bounds. We complement our analysis with a practical stopping criterion and demonstrate its applicability via numerical experiments. Our results show that one can get comparable performance on downstream tasks while requiring fewer iterations, thereby saving substantial computational time.

math.NA↗

arXiv2020

Over-parametrized neural networks as under-determined linear systems

We draw connections between simple neural networks and under-determined linear systems to comprehensively explore several interesting theoretical questions in the study of neural networks. First, we emphatically show that it is unsurprising such networks can achieve zero training loss. More specifically, we provide lower bounds on the width of a single hidden layer neural network such that only training the last linear layer suffices to reach zero training loss. Our lower bounds grow more slowly with data set size than existing work that trains the hidden layer weights. Second, we show that kernels typically associated with the ReLU activation function have fundamental flaws -- there are simple data sets where it is impossible for widely studied bias-free models to achieve zero training loss irrespective of how the parameters are chosen or trained. Lastly, our analysis of gradient descent clearly illustrates how spectral properties of certain matrices impact both the early iteration and long-term training behavior. We propose new activation functions that avoid the pitfalls of ReLU in that they admit zero training loss solutions for any set of distinct data points and experimentally exhibit favorable spectral properties.

math.NA↗

arXiv2020

Uniform bounds for invariant subspace perturbations

For a fixed symmetric matrix A and symmetric perturbation E we develop purely deterministic bounds on how invariant subspaces of A and A+E can differ when measured by a suitable "row-wise" metric rather than via traditional measures of subspace distance. Understanding perturbations of invariant subspaces with respect to such metrics is becoming increasingly important across a wide variety of applications and therefore necessitates new theoretical developments. Under minimal assumptions we develop new bounds on subspace perturbations under the two-to-infinity matrix norm and show in what settings these row-wise differences in the invariant subspaces can be significantly smaller than the analogous two or Frobenius norm differences. We also demonstrate that the constitutive pieces of our bounds are necessary absent additional assumptions and, therefore, our results provide a natural starting point for further analysis of specific problems. Lastly, we briefly discuss extensions of our bounds to scenarios where A and/or E are non-normal matrices.

math.NA↗

arXiv2019

Incrementally Updated Spectral Embeddings

Several fundamental tasks in data science rely on computing an extremal eigenspace of size $r \ll n$, where $n$ is the underlying problem dimension. For example, spectral clustering and PCA both require the computation of the leading $r$-dimensional subspace. Often, this process is repeated over time due to the possible temporal nature of the data; e.g., graphs representing relations in a social network may change over time, and feature vectors may be added, removed or updated in a dataset. Therefore, it is important to efficiently carry out the computations involved to keep up with frequent changes in the underlying data and also to dynamically determine a reasonable size for the subspace of interest. We present a complete computational pipeline for efficiently updating spectral embeddings in a variety of contexts. Our basic approach is to "seed" iterative methods for eigenproblems with the most recent subspace estimate to significantly reduce the computations involved, in contrast with a naïve approach which recomputes the subspace of interest from scratch at every step. In this setting, we provide various bounds on the number of iterations common eigensolvers need to perform in order to update the extremal eigenspace to a sufficient tolerance. We also incorporate a criterion for determining the size of the subspace based on successive eigenvalue ratios. We demonstrate the merits of our approach on the tasks of spectral clustering of temporally evolving graphs and PCA of an incrementally updated data matrix.

math.NA↗

arXiv2018

Disentanglement via entanglement: A unified method for Wannier localization

The Wannier localization problem in quantum physics is mathematically analogous to finding a localized representation of a subspace corresponding to a nonlinear eigenvalue problem. While Wannier localization is well understood for insulating materials with isolated eigenvalues, less is known for metallic systems with entangled eigenvalues. Currently, the most widely used method for treating systems with entangled eigenvalues is to first obtain a reduced subspace (often referred to as disentanglement) and then to solve the Wannier localization problem by treating the reduced subspace as an isolated system. This is a multi-objective nonconvex optimization procedure and its solution can depend sensitively on the initial guess. We propose a new method to solve the Wannier localization problem, avoiding the explicit use of an an optimization procedure. Our method is robust, efficient, relies on few tunable parameters, and provides a unified framework for addressing problems with isolated and entangled eigenvalues.

physics.comp-ph↗

arXiv2018

Variational formulation for Wannier functions with entangled band structure

Wannier functions provide a localized representation of spectral subspaces of periodic Hamiltonians, and play an important role for interpreting and accelerating Hartree-Fock and Kohn-Sham density functional theory calculations in quantum physics and chemistry. For systems with isolated band structure, the existence of exponentially localized Wannier functions and numerical algorithms for finding them are well studied. In contrast, for systems with entangled band structure, Wannier functions must be generalized to span a subspace larger than the spectral subspace of interest to achieve favorable spatial locality. In this setting, little is known about the theoretical properties of these Wannier functions, and few algorithms can find them robustly. We develop a variational formulation to compute these generalized maximally localized Wannier functions. When paired with an initial guess based on the selected columns of the density matrix (SCDM) method, our method can robustly find Wannier functions for systems with entangled band structure. We formulate the problem as a constrained nonlinear optimization problem, and show how the widely used disentanglement procedure can be interpreted as a splitting method to approximately solve this problem. We demonstrate the performance of our method using real materials including silicon, copper, and aluminum. To examine more precisely the localization properties of Wannier functions, we study the free electron gas in one and two dimensions, where we show that the maximally-localized Wannier functions only decay algebraically. We also explain using a one dimensional example how to modify them to obtain super-algebraic decay.

physics.comp-ph↗

arXiv2017

Fast spatial Gaussian process maximum likelihood estimation via skeletonization factorizations

Maximum likelihood estimation for parameter-fitting given observations from a Gaussian process in space is a computationally-demanding task that restricts the use of such methods to moderately-sized datasets. We present a framework for unstructured observations in two spatial dimensions that allows for evaluation of the log-likelihood and its gradient (i.e., the score equations) in $\tilde O(n^{3/2})$ time under certain assumptions, where $n$ is the number of observations. Our method relies on the skeletonization procedure described by Martinsson & Rokhlin in the form of the recursive skeletonization factorization of Ho & Ying. Combining this with an adaptation of the matrix peeling algorithm of Lin et al. for constructing $\mathcal{H}$-matrix representations of black-box operators, we obtain a framework that can be used in the context of any first-order optimization routine to quickly and accurately compute maximum-likelihood estimates.

stat.ME↗

arXiv2017

Computing localized representations of the kohn-sham subspace via randomization and refinement

Localized representation of the Kohn-Sham subspace plays an important role in quantum chemistry and materials science. The recently developed selected columns of the density matrix (SCDM) method [J. Chem. Theory Comput. 11, 1463, 2015] is a simple and robust procedure for finding a localized representation of a set of Kohn-Sham orbitals from an insulating system. The SCDM method allows the direct construction of a well conditioned (or even orthonormal) and localized basis for the Kohn-Sham subspace. The SCDM algorithm avoids the use of an optimization procedure and does not depend on any adjustable parameters. The most computationally expensive step of the SCDM method is a column pivoted QR factorization that identifies the important columns for constructing the localized basis set. In this paper, we develop a two stage approximate column selection strategy to find the important columns at much lower computational cost. We demonstrate the effectiveness of this process using the dissociation process of a BH$_3$NH$_3$ molecule, an alkane chain and a supercell with 256 water molecules. Numerical results for the large collection of water molecules show that two stage localization procedure can be more than 30 times faster than the original SCDM algorithm and compares favorably with the popular Wannier90 package.

math.NA↗

arXiv2017

Robust and efficient multi-way spectral clustering

We present a new algorithm for spectral clustering based on a column-pivoted QR factorization that may be directly used for cluster assignment or to provide an initial guess for k-means. Our algorithm is simple to implement, direct, and requires no initial guess. Furthermore, it scales linearly in the number of nodes of the graph and a randomized variant provides significant computational gains. Provided the subspace spanned by the eigenvectors used for clustering contains a basis that resembles the set of indicator vectors on the clusters, we prove that both our deterministic and randomized algorithms recover a basis close to the indicators in Frobenius norm. We also experimentally demonstrate that the performance of our algorithm tracks recent information theoretic bounds for exact recovery in the stochastic block model. Finally, we explore the performance of our algorithm when applied to a real world graph.

math.NA↗

arXiv2017

A recursive skeletonization factorization based on strong admissibility

We introduce the strong recursive skeletonization factorization (RS-S), a new approximate matrix factorization based on recursive skeletonization for solving discretizations of linear integral equations associated with elliptic partial differential equations in two and three dimensions (and other matrices with similar hierarchical rank structure). Unlike previous skeletonization-based factorizations, RS-S uses a simple modification of skeletonization, strong skeletonization, which compresses only far-field interactions. This leads to an approximate factorization in the form of a product of many block unit-triangular matrices that may be used as a preconditioner or moderate-accuracy direct solver, with dramatically reduced rank growth. We further combine the strong skeletonization procedure with alternating near-field compression to obtain the hybrid recursive skeletonization factorization (RS-WS), a modification of RS-S that exhibits reduced storage cost in many settings. Under suitable rank assumptions both RS-S and RS-WS exhibit linear computational complexity, which we demonstrate with a number of numerical examples.

math.NA↗

arXiv2016

SCDM-k: Localized orbitals for solids via selected columns of the density matrix

The recently developed selected columns of the density matrix (SCDM) method [J. Chem. Theory Comput. 11, 1463, 2015] is a simple, robust, efficient and highly parallelizable method for constructing localized orbitals from a set of delocalized Kohn-Sham orbitals for insulators and semiconductors with $Γ$ point sampling of the Brillouin zone. In this work we generalize the SCDM method to Kohn-Sham density functional theory calculations with k-point sampling of the Brillouin zone, which is needed for more general electronic structure calculations for solids. We demonstrate that our new method, called SCDM-k, is by construction gauge independent and is a natural way to describe localized orbitals. SCDM-k computes localized orbitals without the use of an optimization procedure, and thus does not suffer from the possibility of being trapped in a local minimum. Furthermore, the computational complexity of using SCDM-k to construct orthogonal and localized orbitals scales as O(N log N ) where N is the total number of k-points in the Brillouin zone. SCDM-k is therefore efficient even when a large number of k-points are used for Brillouin zone sampling. We demonstrate the numerical performance of SCDM-k using systems with model potentials in two and three dimensions.

physics.comp-ph↗

arXiv2015

A geometric approach to archetypal analysis and non-negative matrix factorization

Archetypal analysis and non-negative matrix factorization (NMF) are staples in a statisticians toolbox for dimension reduction and exploratory data analysis. We describe a geometric approach to both NMF and archetypal analysis by interpreting both problems as finding extreme points of the data cloud. We also develop and analyze an efficient approach to finding extreme points in high dimensions. For modern massive datasets that are too large to fit on a single machine and must be stored in a distributed setting, our approach makes only a small number of passes over the data. In fact, it is possible to obtain the NMF or perform archetypal analysis with just two passes over the data.

stat.ME↗

arXiv2015

A technique for updating hierarchical skeletonization-based factorizations of integral operators

We present a method for updating certain hierarchical factorizations for solving linear integral equations with elliptic kernels. In particular, given a factorization corresponding to some initial geometry or material parameters, we can locally perturb the geometry or coefficients and update the initial factorization to reflect this change with asymptotic complexity that is polylogarithmic in the total number of unknowns and linear in the number of perturbed unknowns. We apply our method to the recursive skeletonization factorization and hierarchical interpolative factorization and demonstrate scaling results for a number of different 2D problem setups.

math.NA↗

arXiv2014

Compressed representation of Kohn-Sham orbitals via selected columns of the density matrix

Given a set of Kohn-Sham orbitals from an insulating system, we present a simple, robust, efficient and highly parallelizable method to construct a set of, optionally orthogonal, localized basis functions for the associated subspace. Our method explicitly uses the fact that density matrices associated with insulating systems decay exponentially along the off-diagonal direction in the real space representation. Our method avoids the usage of an optimization procedure, and the localized basis functions are constructed directly from a set of selected columns of the density matrix (SCDM). Consequently, the only adjustable parameter in our method is the truncation threshold of the localized basis functions. Our method can be used in any electronic structure software package with an arbitrary basis set. We demonstrate the numerical accuracy and parallel scalability of the SCDM procedure using orbitals generated by the Quantum ESPRESSO software package. We also demonstrate a procedure for combining SCDM with Hockney's algorithm to efficiently perform Hartree-Fock exchange energy calculations with near linear scaling.

physics.comp-ph↗

arXiv2014

Linear Hamilton Jacobi Bellman Equations in High Dimensions

The Hamilton Jacobi Bellman Equation (HJB) provides the globally optimal solution to large classes of control problems. Unfortunately, this generality comes at a price, the calculation of such solutions is typically intractible for systems with more than moderate state space size due to the curse of dimensionality. This work combines recent results in the structure of the HJB, and its reduction to a linear Partial Differential Equation (PDE), with methods based on low rank tensor representations, known as a separated representations, to address the curse of dimensionality. The result is an algorithm to solve optimal control problems which scales linearly with the number of states in a system, and is applicable to systems that are nonlinear with stochastic forcing in finite-horizon, average cost, and first-exit settings. The method is demonstrated on inverted pendulum, VTOL aircraft, and quadcopter models, with system dimension two, six, and twelve respectively.

math.OC↗

arXiv2014

Pole expansion for solving a type of parametrized linear systems in electronic structure calculations

We present a new method for solving parametrized linear systems. Under certain assumptions on the parametrization, solutions to the linear systems for all parameters can be accurately approximated by linear combinations of solutions to linear systems for a small set of fixed parameters. Combined with either direct solvers or preconditioned iterative solvers for each linear system with a fixed parameter, the method is particularly suitable for situations when solutions to a large number of distinct parameters or a large number of right hand sides are required. The method is also simple to parallelize. We demonstrate the applicability of the method to the calculation of the response functions in electronic structure theory. We demonstrate the numerical performance of the method using a benzene molecule and a DNA molecule.

math.NA↗