Minimal Krylov Subspaces for Dimension Reduction by Alexander M. Breuer ARL-RP-0480 May 2014 A reprint from ProQuest/UMI, Indiana University, 2013; ATT 3552604 Approved for public release; distribution is unlimited. NOTICES Disclaimers The findings in this report are not to be construed as an official Department of the Army position unless so designated by other authorized documents. Citation of manufacturer’s or trade names does not constitute an official endorsement or approval of the use thereof. Destroy this report when it is no longer needed. Do not return it to the originator. Army Research Laboratory Aberdeen Proving Ground, MD 21005-5068 ARL-RP-0480 May 2014 Minimal Krylov Subspaces for Dimension Reduction Alexander M. Breuer Survivability/Lethality Analysis Directorate, ARL A reprint from ProQuest/UMI, Indiana University, 2013; ATT 3552604 Approved for public release; distribution is unlimited. REPORT DOCUMENTATION PAGE Form Approved OMB No. 0704-0188 Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing the burden, to Department of Defense, Washington Headquarters Services, Directorate for Information Operations and Reports (0704-0188), 1215 Jefferson Davis Highway, Suite 1204, Arlington, VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to any penalty for failing to comply with a collection of information if it does not display a currently valid OMB control number. PLEASE DO NOT RETURN YOUR FORM TO THE ABOVE ADDRESS. 1. REPORT DATE (DD-MM-YYYY) 2. REPORT TYPE 3. DATES COVERED (From - To) May 2014 Reprint 1 August 2008–31 December 2012 4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER Minimal Krylov Subspaces for Dimension Reduction 5b. GRANT NUMBER 5c. PROGRAM ELEMENT NUMBER 6. AUTHOR(S) 5d. PROJECT NUMBER Alexander M. Breuer 5e. TASK NUMBER 5f. WORK UNIT NUMBER 7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION REPORT NUMBER U.S. Army Research Laboratory ATTN: RDRL-SLB-W ARL-RP-0480 Aberdeen Proving Ground, MD 21005-5068 9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S) 11. SPONSOR/MONITOR'S REPORT NUMBER(S) 12. DISTRIBUTION/AVAILABILITY STATEMENT Approved for public release; distribution is unlimited. 13. SUPPLEMENTARY NOTES A reprint from ProQuest/UMI, Indiana University, 2013; ATT 3552604 14. ABSTRACT Krylov subspaces may be used as an alternative to singular vector spaces or eigenvector spaces for projective dimension reduction for low-rank matrix approximation. Though the truncated spectral or singular value decomposition is optimal for minimizing Frobenius norm error of a low-rank approximation, substituting a Krylov subspace projection may result in marked compute time savings. Previous efforts to apply Krylov subspaces to low-rank approximation problems are extended to block Krylov subspaces. Closely related random projection methods are compared to block Krylov subspaces, and new hybrid approaches are developed. Hybrid random projection Krylov subspace methods offer faster compute times than random projection methods and lower approximation errors when sufficient conditions are met. A novel adaptively blocked Krylov subspace algorithm is developed that offers superior compute times to random projection methods. Stationary inner iteration is considered for accelerating convergence of Krylov subspaces and applied to the low-rank approximation problem; a generalization of eigenvalue approximation bounds is presented for Krylov subspaces augmented with inner iteration. All aforementioned methods are evaluated in terms offloating-point operations and applied to numerous problems. 15. SUBJECT TERMS Krylov subspaces, random projections, dimension reduction, low-rank approximation, iterative methods 17. LIMITATION 18. NUMBER 19a. NAME OF RESPONSIBLE PERSON 16. SECURITY CLASSIFICATION OF: OF ABSTRACT OF PAGES Alexander M. Breuer a. REPORT b. ABSTRACT c. THIS PAGE 19b. TELEPHONE NUMBER (Include area code) Unclassified Unclassified Unclassified UU 200 410-278-9157 Standard Form 298 (Rev. 8/98) Prescribed by ANSI Std. Z39.18 MINIMAL KRYLOV SUBSPACES FOR DIMENSION REDUCTION Alex Breuer Submitted to the faculty of the University Graduate School in partial fulfillment of the requirements for the degree Doctor of Philosophy in the Division of Computer Science of the School of Informatics and Computing, Indiana University January 2013 AcceptedbytheGraduateFaculty,IndianaUniversity,inpartialfulfillmentoftherequirementsfor thedegreeofDoctorofPhilosophy. Doctoral Committee Andrew LUMSDAINE,Ph.D. Gonzalo ARCE,Ph.D. David CRANDALL,Ph.D. Esfandiar HAGHVERDI,Ph.D. Predrag RADIVOJAC,Ph.D. December13,2012 ii Acknowledgments I would like to gratefully acknowledge my research advisor, Professor Andrew Lumsdaine, for his support throughout my graduate study. Without his patience and accommodation, this would not have been possible. I would also like to express my appreciation for my entire graduate committee, ProfessorsGonzaloArce,DavidCrandall,EsfandiarHaghverdiandPredragRadivojac. Iwouldalso liketorecognizetheinvaluablesupportofLisaRoach,ChiefoftheWarfighterSurvivabilityBranch in the U. S. Army Research Laboratory, as well as Director Dr. Paul Tanenbaum and Division Chief Robert Bowen for their encouragement and support during this process. I would also like to thank mywife,StephanieBreuer,whosetirelessfaithinmehasbeenaconstantblessing. iii AlexBreuer MinimalKrylovSubspacesforDimensionReduction Krylov subspaces may be used as an alternative to singular vector spaces or eigenvector spaces for projective dimension reduction for low-rank matrix approximation. Though the truncated spec- tral or singular value decomposition is optimal for minimizing Frobenius norm error of a low-rank approximation, substituting a Krylov subspace projection may result in marked compute time sav- ings. Previous efforts to apply Krylov subspaces to low-rank approximation problems are extended toblockKrylovsubspaces. Closely-relatedrandomprojectionmethodsarecomparedtoblockKrylov subspaces, and new hybrid approaches are developed. Hybrid random-projection Krylov subspace methods offer faster compute times than random projection methods, and lower approximation er- rors when sufficient conditions are met. A novel adaptively-blocked Krylov subspace algorithm is developed that offers superior compute times to random projection methods. Stationary inner it- eration is considered for accelerating convergence of Krylov subspaces and applied to the low-rank approximationproblem;ageneralizationofeigenvalueapproximationboundsispresentedforKrylov subspaces augmented with inner iteration. All aforementioned methods are evaluated in terms of floating-pointoperationsandappliedtonumerousproblems. iv Table of Contents 1 Introduction 1 1.1 Thelow-rankapproximationproblem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 ApproximationoftheSVDforlow-rankapproximation . . . . . . . . . . . . . . . . . . . . 7 2 Low-rankapproximationbackground 12 2.1 Genericdimensionreductionbackground . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.2 Krylovsubspacebackground . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.3 ConvergenceofeigenvaluesinKrylovsubspaces . . . . . . . . . . . . . . . . . . . . . . . . 16 2.4 Low-rankapproximationerrorandeigenvalueapproximations . . . . . . . . . . . . . . . 18 2.5 Krylovsubspacealgorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.6 Randomprojectionmethods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 3 ShortblockKrylovsubspacesforlow-rankapproximation 25 3.1 Randomprojectionsforlow-rankmatrixapproximation . . . . . . . . . . . . . . . . . . . 29 3.2 BlockLanczosforlow-rankapproximation . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 3.2.1 BlockLanczosforwithrefinementforimprovedstability . . . . . . . . . . . . . . 32 3.2.2 Theshrink-and-iterateapproach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 3.2.3 Thehybridrandomprojection-Krylovsubspacemethod . . . . . . . . . . . . . . . 35 3.3 ConvergenceAnalyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.3.1 ReviewofexistingboundsforrandomprojectionsandKrylovsubspaces . . . . . 38 3.3.2 Derivationofnewbounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 v 3.3.3 Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.3.4 Convergenceexample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.4 StabilityAnalysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.4.1 Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 4 ApplicationsofshortblockKrylovsubspaces 58 4.1 ExperimentswiththeYalefacedata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 4.2 Low-rankapproximationinInformationRetrieval . . . . . . . . . . . . . . . . . . . . . . . 64 4.3 ExperimentswiththeColoradoroadnetworkLaplacian . . . . . . . . . . . . . . . . . . . 70 4.4 Stiffnessmatrixexperiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 5 GrABL:AGreedyAdaptiveBlockLanczos method 85 5.1 Typicalspectraofgenericdimensionreductionproblems . . . . . . . . . . . . . . . . . . . 87 5.2 ClassicblockLanczos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 5.3 ABLE:adaptivelyblockedLanczosfortheeigenproblem . . . . . . . . . . . . . . . . . . . 88 5.3.1 ConvergenceofeigenvaluesinblockKrylovsubspaces . . . . . . . . . . . . . . . . 89 5.3.2 Convergenceviablocksizeversusconvergenceviaiterations . . . . . . . . . . . . 91 5.4 GreedyAdaptively-blockedLanczos. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 5.5 RandomprojectionsandorthogonalLanczositeration. . . . . . . . . . . . . . . . . . . . . 92 5.5.1 Choosing mautomatically . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 5.6 TheGrABLalgorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 6 ApplicationsofGrABL 104 6.1 ExperimentswiththeYalefacedata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 6.2 ExperimentswiththeColoradoroadnetworkLaplacian . . . . . . . . . . . . . . . . . . . 110 6.3 ExperimentswiththeBagofWordsterm-documentmatrix . . . . . . . . . . . . . . . . . 116 vi