Category CS P03 Optimizing the Deflation Strategy of a Divide and Conquer Eigensolver

Abstract A classic problem in numerical linear algebra is the calculation of the

eigenvalues of a symmetric matrix. These eigenvalues are useful in

calculating critical frequencies in many systems, such as vibrations in

mass spring systems ([1] page 254). Not much research has been done

on the deflating criteria in the divide and conquer eigenvalue algorithm. My

hypothesis is that a better deflating criterion than the currently used

constant threshold exists, and that the deflating threshold will be a

function of the size of the matrix. A genetic algorithm is used to search for

this better function. The genetic algorithm (GA) in 12 out of 18 runs (on

100 random matrices) showed that a logarithmic function is better than a

constant function. A quadratic fit yielded a logarithmic deflating function

that was optimal for general matrices, not just the 100 random matrices

used in each run of the GA. Data collected with this new function showed

acceptable errors (1.73 accurate digits versus the average 1.88

computed with the GA). With this function, the divide and conquer method

yielded 2.61 accurate digits on larger matrices. In conclusion, the

hypothesis was correct, and the new logarithmic deflating function is

better than a constant function.

Bibliography [1] Demmel, James W. Applied Numerical Linear Algebra. 1 ed. Philadelphia:

SIAM,

1997. Print.

[2] Potter, Christopher. "A Parallel Divide and Conquer Eigensolver." EPFL


Supercomputing Review. EPFL, n.d. Web. 14 Feb. 2010.

<http://ditwww.epfl.ch/SIC/SA/publications/SCR95/7-95-27a.html>.

[3] Stewart, G. W. Matrix Algorithms, Volume II: Eigensystems. 1 ed.

Philadelphia:
First Previous Next Last