An algorithm for automatically selecting a suitable verification method for linear systems

  • Authors:

  • Katsuhisa Ozaki;Takeshi Ogita;Shin'Ichi Oishi

  • Affiliations:

  • Department of Mathematical Sciences, College of Systems Engineering and Science, Shibaura Institute of Technology, Saitama, Japan 337-8570 and JST, CREST, Tokyo, Japan 102-0075;Department of Mathematical Sciences, Tokyo Woman's Christian University, Tokyo, Japan 167-8585 and JST, CREST, Tokyo, Japan 102-0075;Faculty of Science and Engineering, Waseda University, Tokyo, Japan 169-0072 and JST, CREST, Tokyo, Japan 102-0075

  • Venue:
  • Numerical Algorithms
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

Several methods have been proposed to calculate a rigorous error bound of an approximate solution of a linear system by floating-point arithmetic. These methods are called `verification methods'. Applicable range of these methods are different. It depends mainly on the condition number and the dimension of the coefficient matrix whether such methods succeed to work or not. In general, however, the condition number is not known in advance. If the dimension or the condition number is large to some extent, then Oishi---Rump's method, which is known as the fastest verification method for this purpose, may fail. There are more robust verification methods whose computational cost is larger than the Oishi---Rump's one. It is not so efficient to apply such robust methods to well-conditioned problems. The aim of this paper is to choose a suitable verification method whose computational cost is minimum to succeed. First in this paper, four fast verification methods for linear systems are briefly reviewed. Next, a compromise method between Oishi---Rump's and Ogita---Oishi's one is developed. Then, an algorithm which automatically and efficiently chooses an appropriate verification method from five verification methods is proposed. The proposed algorithm does as much work as necessary to calculate error bounds of approximate solutions of linear systems. Finally, numerical results are presented.