Large-scale log-determinant computation through stochastic Chebyshev expansions

Insu Han, Dmitry Malioutov, Jinwoo Shin
Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:908-917, 2015.

Abstract

Logarithms of determinants of large positive definite matrices appear ubiquitously in machine learning applications including Gaussian graphical and Gaussian process models, partition functions of discrete graphical models, minimum-volume ellipsoids and metric and kernel learning. Log-determinant computation involves the Cholesky decomposition at the cost cubic in the number of variables (i.e., the matrix dimension), which makes it prohibitive for large-scale applications. We propose a linear-time randomized algorithm to approximate log-determinants for very large-scale positive definite and general non-singular matrices using a stochastic trace approximation, called the Hutchinson method, coupled with Chebyshev polynomial expansions that both rely on efficient matrix-vector multiplications. We establish rigorous additive and multiplicative approximation error bounds depending on the condition number of the input matrix. In our experiments, the proposed algorithm can provide very high accuracy solutions at orders of magnitude faster time than the Cholesky decomposition and Shur completion, and enables us to compute log-determinants of matrices involving tens of millions of variables.

Cite this Paper


BibTeX
@InProceedings{pmlr-v37-hana15, title = {Large-scale log-determinant computation through stochastic Chebyshev expansions}, author = {Han, Insu and Malioutov, Dmitry and Shin, Jinwoo}, booktitle = {Proceedings of the 32nd International Conference on Machine Learning}, pages = {908--917}, year = {2015}, editor = {Bach, Francis and Blei, David}, volume = {37}, series = {Proceedings of Machine Learning Research}, address = {Lille, France}, month = {07--09 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v37/hana15.pdf}, url = { http://proceedings.mlr.press/v37/hana15.html }, abstract = {Logarithms of determinants of large positive definite matrices appear ubiquitously in machine learning applications including Gaussian graphical and Gaussian process models, partition functions of discrete graphical models, minimum-volume ellipsoids and metric and kernel learning. Log-determinant computation involves the Cholesky decomposition at the cost cubic in the number of variables (i.e., the matrix dimension), which makes it prohibitive for large-scale applications. We propose a linear-time randomized algorithm to approximate log-determinants for very large-scale positive definite and general non-singular matrices using a stochastic trace approximation, called the Hutchinson method, coupled with Chebyshev polynomial expansions that both rely on efficient matrix-vector multiplications. We establish rigorous additive and multiplicative approximation error bounds depending on the condition number of the input matrix. In our experiments, the proposed algorithm can provide very high accuracy solutions at orders of magnitude faster time than the Cholesky decomposition and Shur completion, and enables us to compute log-determinants of matrices involving tens of millions of variables.} }
Endnote
%0 Conference Paper %T Large-scale log-determinant computation through stochastic Chebyshev expansions %A Insu Han %A Dmitry Malioutov %A Jinwoo Shin %B Proceedings of the 32nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2015 %E Francis Bach %E David Blei %F pmlr-v37-hana15 %I PMLR %P 908--917 %U http://proceedings.mlr.press/v37/hana15.html %V 37 %X Logarithms of determinants of large positive definite matrices appear ubiquitously in machine learning applications including Gaussian graphical and Gaussian process models, partition functions of discrete graphical models, minimum-volume ellipsoids and metric and kernel learning. Log-determinant computation involves the Cholesky decomposition at the cost cubic in the number of variables (i.e., the matrix dimension), which makes it prohibitive for large-scale applications. We propose a linear-time randomized algorithm to approximate log-determinants for very large-scale positive definite and general non-singular matrices using a stochastic trace approximation, called the Hutchinson method, coupled with Chebyshev polynomial expansions that both rely on efficient matrix-vector multiplications. We establish rigorous additive and multiplicative approximation error bounds depending on the condition number of the input matrix. In our experiments, the proposed algorithm can provide very high accuracy solutions at orders of magnitude faster time than the Cholesky decomposition and Shur completion, and enables us to compute log-determinants of matrices involving tens of millions of variables.
RIS
TY - CPAPER TI - Large-scale log-determinant computation through stochastic Chebyshev expansions AU - Insu Han AU - Dmitry Malioutov AU - Jinwoo Shin BT - Proceedings of the 32nd International Conference on Machine Learning DA - 2015/06/01 ED - Francis Bach ED - David Blei ID - pmlr-v37-hana15 PB - PMLR DP - Proceedings of Machine Learning Research VL - 37 SP - 908 EP - 917 L1 - http://proceedings.mlr.press/v37/hana15.pdf UR - http://proceedings.mlr.press/v37/hana15.html AB - Logarithms of determinants of large positive definite matrices appear ubiquitously in machine learning applications including Gaussian graphical and Gaussian process models, partition functions of discrete graphical models, minimum-volume ellipsoids and metric and kernel learning. Log-determinant computation involves the Cholesky decomposition at the cost cubic in the number of variables (i.e., the matrix dimension), which makes it prohibitive for large-scale applications. We propose a linear-time randomized algorithm to approximate log-determinants for very large-scale positive definite and general non-singular matrices using a stochastic trace approximation, called the Hutchinson method, coupled with Chebyshev polynomial expansions that both rely on efficient matrix-vector multiplications. We establish rigorous additive and multiplicative approximation error bounds depending on the condition number of the input matrix. In our experiments, the proposed algorithm can provide very high accuracy solutions at orders of magnitude faster time than the Cholesky decomposition and Shur completion, and enables us to compute log-determinants of matrices involving tens of millions of variables. ER -
APA
Han, I., Malioutov, D. & Shin, J.. (2015). Large-scale log-determinant computation through stochastic Chebyshev expansions. Proceedings of the 32nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 37:908-917 Available from http://proceedings.mlr.press/v37/hana15.html .

Related Material