计算机科学

计算机科学 ›› 2023, Vol. 50 ›› Issue (6): 45-51.doi: 10.11896/jsjkx.230200209

• 高性能计算 • 上一篇    下一篇

基于双倍双精度施密特正交化方法的QR分解算法

金洁茜1, 谢和虎2, 杜配冰3, 全哲1, 姜浩4   

  1. 1 湖南大学信息科学与工程学院 长沙 410082
    2 中国科学院数学与系统科学研究院 北京 100190
    3 西北核技术研究所 西安 710024
    4 国防科技大学计算机学院 长沙 410073
  • 收稿日期:2023-02-27 修回日期:2023-04-11 出版日期:2023-06-15 发布日期:2023-06-06
  • 通讯作者: 姜浩(haojiang@nudt.edu.cn)
  • 作者简介:(jinjiexi@hnu.edu.cn)
  • 基金资助:
    国家自然科学基金(61907034);国家重点研发计划(2020YFA0709803);科学挑战专题(TZ2016002)

QR Decomposition Based on Double-double Precision Gram-Schmidt Orthogonalization Method

JIN Jiexi1, XIE Hehu2, DU Peibing3, QUAN Zhe1, JIANG Hao4   

  1. 1 College of Computer Science and Electronic Engineering,Hunan University,Changsha 410082,China
    2 Academy of Mathematics and Systems Science,Chinese Academy of Sciences,Beijing 100190,China
    3 Northwest Insititute of Nuclear Technology,Xi'an 710024,China
    4 College of Computer Science and Technology,National University of Defense Technology,Changsha 410073,China
  • Received:2023-02-27 Revised:2023-04-11 Online:2023-06-15 Published:2023-06-06
  • About author:JIN Jiexi,born in 1999,postgraduate,is a member of China Computer Federation.Her main research interest is high performance computing.JIANG Hao,born in 1983,Ph.D,professor,assoicate research fellow.His main research interests include high perfor-mance computing and numerical analysis.
  • Supported by:
    National Natural Science Foundation of China(61907034),National Key Research and Development Program of China(2020YFA0709803) and Science Challenge Project(TZ2016002).

PDF (PC)

3819

摘要： 当矩阵的规模较大或者条件数较高时,格拉姆-施密特(Gram-Schmidt)正交化算法和其相关修正算法时常表现出数值不稳定性的现象。为了解决该问题,探索了修正Gram-Schmidt算法(MGS)中舍入误差的累积效应,然后基于无误差变换技术和双倍双精度算法,设计并实现了双倍双精度修正Gram-Schmidt正交化算法(DDMGS)。该算法的精度测试中显示所提算法较分块施密特正交化(BMGS_SVL,BMGS_CWY,BCGS_PIP与BCGS_PIO)的变体算法具有更好的数值稳定性,证明了DDMGS算法能够有效地减少矩阵的正交性损失,提升数值精度,展示了所提算法的可靠性。在算法的性能测试中,首先计算并比较了不同算法的浮点计算量(flops),随后将所提DDMGS算法与修正施密特正交化算法在ARM和Intel两款处理器上作比较,虽然DDMGS算法的运行时间分别是MGS的5.03倍和18.06倍左右,但获得了明显的精度提升效果。

关键词: 施密特正交化算法, QR分解, 无误差变换, 双倍双精度算法, 舍入误差, 浮点算术

Abstract: The Gram-Schmidt orthogonalization algorithm and its related modified algorithms often show numerical instability when computing ill-conditioned or large-scale matrices.To solve this problem,this paper explores the cumulative effect of round-off errors of modified Gram-Schmidt algorithm(MGS),and then designs and implements a double-double precision modified Gram-Schmidt orthogonalization algorithm(DDMGS) based on the error-free transformation technology and double-double precision algorithm.A variety of accuracy tests illustrate that DDMGS algorithm has better numerical stability than the varients of BMGS_SVL,BMGS_CWY,BCGS_PIP and BCGS_PIO algorithms,which proves that DDMGS algorithm can effectively reduce the loss of orthogonality of matrix,improve the numerical accuracy,and demonstrate the stability of our algorithm.In the performance test, the floating point computations(flops) of different algorithms are calculated and then compared DDMGS algorithm with the modified Gram-Schmidt algorithm on ARM and Intel processors,the runtime of the DDMGS algorithm proposed in this paper isabout 5.03 and 18.06 times that of MGS respectively,but the accuracy is improved significantly.

Key words: Gram-Schmidt algorithm, QR decomposition, Error-free transformation, Double-double precision algorithm, Roundoff error, Floating-point arithmetic

中图分类号: 

  • TP391
[1]LEON S J,BJORCK A,GANDER W.Gram-Schmidt orthogonalization:100 years and more[J].Numerical Linear Algebra with Applications,2013,20(3):492-532.
[2]HUANG J,GU H.QuickBird image fusion based on Gram-Schmidt transform[C]//Geographic Information and Internet of Things Forum and Proceedings of the 2010 Annual Academic Conference of the Jiangsu Society of Surveying and Mapping.2010.
[3]ZHANG K,HSIEH M H,LIU L,et al.Quantum Gram-Schmidt processes and their application to efficient state readout for quantum talgorithms[J].Physical Review Research,2021,3(4):043095.
[4]BAILEY D H.A fortran-90 double-double library[J/OL].https://www.davidhbailey.com/dhbsoftware,2001.
[5]BAILEY D H,KRASNY R,PELZ R.Multiple precision,multiple processor vortex sheet roll-up computation[R].Society for Industrial and Applied Mathematics(SIAM),Philadelphia,PA (United States),1993.
[6]OGITA T,RUMP S M,OISHI S.Accurate sum and dot product[J].SIAM Journal on Scientific Computing,2005,26(6):1955-1988.
[7]DEKKER T J.A floating-point technique for extending theavailable precision[J].Numerische Mathematik,1971,18(3):224-242.
[8]KAHAN W.Pracniques:further remarks on reducing truncation errors[J].Communications of the ACM,1965,8(1):40.
[9]KNUTH D E.The art of computer programming:Vol.3[M].Newyork:Pearson Education,1997.
[10]STRANG G.Introduction to linear algebra(5th ed)[M].Wellesley-Cambridge Press,1993.
[11]CARSON E,LUND K,ROZLOZNIK M.The Stability of Block Variants of Classical Gram-Schmidt[J].SIAM Journal on Matrix Analysis and Applications,2021,42(3):1365-1380.
[12]CARSON E,LUND K,ROZLOZNÍK M,et al.Block Gram-Schmidt algorithms and their stability properties[J].Linear Algebra and its Applications,2022,638:150-195.
[13]SCHMIDT E.Zur Theorie der linearen und nichtlinearen Integralgleichungen.III.Teil[J].Mathematische Annalen,1908,65(3):370-399.
[14]GRAM J P.Ueber die Entwickelung reeller Functionen inReihen mittelst der Methode der kleinsten Quadrate[J].J Reine Angew Math,1883,94:41-73.
[15]WONG Y K.An application of orthogonalization process to the theory of least squares[J].The Annals of Mathematical Statistics,1935,6(2):53-75.
[16]DE LAPLACE P S.Théorie analytique des probabilités:Vol.7[M].Courcier,1820.
[17]DANIEL J W,GRAGG W B,KAUFMAN L,et al.Reorthogonalization and stable algorithms for updating the Gram-Schmidt QR factorization[J].Mathematics of Computation,1976,30(136):772-795.
[18]JALBY W,PHILIPPE B.Stability analysis and improvement of the block Gram-Schmidt algorithm[J].SIAM Journal on Scientific and Statistical Computing,1991,12(5):1058-1073.
[19]BARLOW J L,SMOKTUNOWICZ A.Reorthogonalized blockclassical Gram-Schmidt[J].Numerische Mathematik,2013,123(3):395-423.
[20]RIVAZ A,MOGHADAM M M,SADEGHI D,et al.An approach of orthogonalization within the Gram-Schmidt algorithm[J].Computational and Applied Mathematics,2018,37(2):1250-1262.
[21]Standard for binary floating point arithmetic:ANSI/IEEEstandard 754-2008[S].IEEE,2008.
[22]BJORCK A.Numerics of gram-schmidt orthogonalization[J].Linear Algebra and Its Applications,1994,197:297-316.
[23]HIDA Y,LI X S,BAILEY D H.Algorithms for quad-double precision floating point arithmetic[C]//Proceedings 15th IEEE Symposium on Computer Arithmetic.IEEE,2001:155-162.
[24]LAUCHLI P.Jordan-elimination und Ausgleichung nach kleinsten Quadraten[J].Numerische Mathematik,1961,3(1):226-240.
[25]PEI M L.A test matrix for inversion procedures[J].Communications of the ACM,1962,5(10):508.
[26]SMOKTUNOWICZ A,BARLOW J L,LANGOU J.A note on the error analysis of classical Gram-Schmidt[J].Numerische Mathematik,2006,105(2):299-313.
[27]NISHI T,RUMP S M,OISHI S.On the generation of very ill-conditioned integer matrices[J].Nonlinear Theory and Its Applications,IEICE,2011,2(2):226-245.
[28]GRAILLAT S,JÉZÉQUEL F,PICOT R.Numerical validationof compensated summation algorithms with stochastic arithmetic[J].Electronic Notes in Theoretical Computer Science,2015,317:55-69.
[29]JIANG H.Study on reliable computing and rounding error ana-lysis in floating-point arithmetic[D].Changsha:National University of Defense Technology,2013.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
渝ICP备16013121号-4
地址：重庆市渝北区洪湖西路18号    邮编：401121  
电话：023-63500828（编辑部） 023-67039612（会议/专题） 023-67039623（财务/发票）
E-mail：jsjkx12@163.com
本系统由北京玛格泰克科技发展有限公司设计开发