Deflation as a method of variance reduction for estimating the trace of a matrix inverse
Abstract
Many fields require computing the trace of the inverse of a large, sparse matrix. The typical method used for such computations is the Hutchinson method which is a Monte Carlo (MC) averaging over matrix quadratures. To improve its convergence, several variance reductions techniques have been proposed. In this paper, we study the effects of deflating the near null singular value space. We make two main contributions. First, we analyze the variance of the Hutchinson method as a function of the deflated singular values and vectors. Although this provides good intuition in general, by assuming additionally that the singular vectors are random unitary matrices, we arrive at concise formulas for the deflated variance that include only the variance and mean of the singular values. We make the remarkable observation that deflation may increase variance for Hermitian matrices but not for non-Hermitian ones. This is a rare, if not unique, property where non-Hermitian matrices outperform Hermitian ones. The theory can be used as a model for predicting the benefits of deflation. Second, we use deflation in the context of a large scale application of "disconnected diagrams" in Lattice QCD. On lattices, Hierarchical Probing (HP) has previously provided an order of magnitude ofmore »
- Authors:
-
- College of William and Mary, Williamsburg, VA (United States)
- Publication Date:
- Research Org.:
- Thomas Jefferson National Accelerator Facility (TJNAF), Newport News, VA (United States)
- Sponsoring Org.:
- USDOE Office of Science (SC), Nuclear Physics (NP)
- OSTI Identifier:
- 1362120
- Report Number(s):
- JLAB-THY-16-2291; DOE/OR/23177-3870; arXiv:1603.05988
Journal ID: ISSN 1064-8275
- Grant/Contract Number:
- AC05-06OR23100; CCF 1218349; ACI S12-SSE 1440700; FC02-12ER41890; FG02-04ER41302; AC05-06OR23177
- Resource Type:
- Accepted Manuscript
- Journal Name:
- SIAM Journal on Scientific Computing
- Additional Journal Information:
- Journal Volume: 39; Journal Issue: 2; Journal ID: ISSN 1064-8275
- Publisher:
- SIAM
- Country of Publication:
- United States
- Language:
- English
- Subject:
- 97 MATHEMATICS AND COMPUTING; deflation; random unitary matrices; Monte Carlo; trace of matrix inverse; Hutchinson; singular values
Citation Formats
Gambhir, Arjun Singh, Stathopoulos, Andreas, and Orginos, Kostas. Deflation as a method of variance reduction for estimating the trace of a matrix inverse. United States: N. p., 2017.
Web. doi:10.1137/16M1066361.
Gambhir, Arjun Singh, Stathopoulos, Andreas, & Orginos, Kostas. Deflation as a method of variance reduction for estimating the trace of a matrix inverse. United States. https://doi.org/10.1137/16M1066361
Gambhir, Arjun Singh, Stathopoulos, Andreas, and Orginos, Kostas. Thu .
"Deflation as a method of variance reduction for estimating the trace of a matrix inverse". United States. https://doi.org/10.1137/16M1066361. https://www.osti.gov/servlets/purl/1362120.
@article{osti_1362120,
title = {Deflation as a method of variance reduction for estimating the trace of a matrix inverse},
author = {Gambhir, Arjun Singh and Stathopoulos, Andreas and Orginos, Kostas},
abstractNote = {Many fields require computing the trace of the inverse of a large, sparse matrix. The typical method used for such computations is the Hutchinson method which is a Monte Carlo (MC) averaging over matrix quadratures. To improve its convergence, several variance reductions techniques have been proposed. In this paper, we study the effects of deflating the near null singular value space. We make two main contributions. First, we analyze the variance of the Hutchinson method as a function of the deflated singular values and vectors. Although this provides good intuition in general, by assuming additionally that the singular vectors are random unitary matrices, we arrive at concise formulas for the deflated variance that include only the variance and mean of the singular values. We make the remarkable observation that deflation may increase variance for Hermitian matrices but not for non-Hermitian ones. This is a rare, if not unique, property where non-Hermitian matrices outperform Hermitian ones. The theory can be used as a model for predicting the benefits of deflation. Second, we use deflation in the context of a large scale application of "disconnected diagrams" in Lattice QCD. On lattices, Hierarchical Probing (HP) has previously provided an order of magnitude of variance reduction over MC by removing "error" from neighboring nodes of increasing distance in the lattice. Although deflation used directly on MC yields a limited improvement of 30% in our problem, when combined with HP they reduce variance by a factor of over 150 compared to MC. For this, we pre-computated 1000 smallest singular values of an ill-conditioned matrix of size 25 million. Furthermore, using PRIMME and a domain-specific Algebraic Multigrid preconditioner, we perform one of the largest eigenvalue computations in Lattice QCD at a fraction of the cost of our trace computation.},
doi = {10.1137/16M1066361},
journal = {SIAM Journal on Scientific Computing},
number = 2,
volume = 39,
place = {United States},
year = {Thu Apr 06 00:00:00 EDT 2017},
month = {Thu Apr 06 00:00:00 EDT 2017}
}
Web of Science
Works referenced in this record:
Randomized algorithms for estimating the trace of an implicit symmetric positive semi-definite matrix
journal, April 2011
- Avron, Haim; Toledo, Sivan
- Journal of the ACM, Vol. 58, Issue 2
Adaptive Multigrid Algorithm for the Lattice Wilson-Dirac Operator
journal, November 2010
- Babich, R.; Brannick, J.; Brower, R. C.
- Physical Review Letters, Vol. 105, Issue 20
An estimator for the diagonal of a matrix
journal, November 2007
- Bekas, C.; Kokiopoulou, E.; Saad, Y.
- Applied Numerical Mathematics, Vol. 57, Issue 11-12
High-precision calculation of the strange nucleon electromagnetic form factors
journal, August 2015
- Green, Jeremy; Meinel, Stefan; Engelhardt, Michael
- Physical Review D, Vol. 92, Issue 3
A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines
journal, January 1990
- Hutchinson, M. F.
- Communications in Statistics - Simulation and Computation, Vol. 19, Issue 2
Improved stochastic estimation of quark propagation with Laplacian Heaviside smearing in lattice QCD
journal, June 2011
- Morningstar, C.; Bulava, J.; Foley, J.
- Physical Review D, Vol. 83, Issue 11
Random matrix theory and spectral sum rules for the Dirac operator in QCD
journal, July 1993
- Shuryak, E. V.; Verbaarschot, J. J. M.
- Nuclear Physics A, Vol. 560, Issue 1
Hierarchical Probing for Estimating the Trace of the Matrix Inverse on Toroidal Lattices
journal, January 2013
- Stathopoulos, Andreas; Laeuchli, Jesse; Orginos, Kostas
- SIAM Journal on Scientific Computing, Vol. 35, Issue 5
PRIMME: preconditioned iterative multimethod eigensolver—methods and software description
journal, April 2010
- Stathopoulos, Andreas; McCombs, James R.
- ACM Transactions on Mathematical Software, Vol. 37, Issue 2
Computing and Deflating Eigenvalues While Solving Multiple Right-Hand Side Linear Systems with an Application to Quantum Chromodynamics
journal, January 2010
- Stathopoulos, Andreas; Orginos, Konstantinos
- SIAM Journal on Scientific Computing, Vol. 32, Issue 1
Domain-Decomposition-Type Methods for Computing the Diagonal of a Matrix Inverse
journal, January 2011
- Tang, Jok M.; Saad, Yousef
- SIAM Journal on Scientific Computing, Vol. 33, Issue 5
Random Matrix Theory and Chiral Symmetry in QCD
journal, December 2000
- Verbaarschot, J. J. M.; Wettig, T.
- Annual Review of Nuclear and Particle Science, Vol. 50, Issue 1
Estimating the trace of the matrix inverse by interpolating from the diagonal of an approximate inverse
journal, December 2016
- Wu, Lingfei; Laeuchli, Jesse; Kalantzis, Vassilis
- Journal of Computational Physics, Vol. 326
Controlling excited-state contamination in nucleon matrix elements
journal, June 2016
- Yoon, Boram; Gupta, Rajan; Bhattacharya, Tanmoy
- Physical Review D, Vol. 93, Issue 11
Works referencing / citing this record:
Proton and neutron electromagnetic form factors from lattice QCD
text, January 2018
- Alexandrou, C.; Bacchio, S.; Constantinou, M.
- Deutsches Elektronen-Synchrotron, DESY, Hamburg
Complete flavor decomposition of the spin and momentum fraction of the proton using lattice QCD simulations at physical pion mass
text, January 2020
- Alexandrou, C.; Bacchio, S.; Constantinou, M.
- Deutsches Elektronen-Synchrotron, DESY, Hamburg
Proton and neutron electromagnetic form factors from lattice QCD
text, January 2019
- Alexandrou, C.; Bacchio, S.; Constantinou, M.
- Deutsches Elektronen-Synchrotron, DESY, Hamburg