20.05.2014 Views

link to my thesis

link to my thesis

link to my thesis

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

7.2. COMPUTING THE GLOBAL MINIMUM USING TARGET SELECTION 113<br />

Table 7.4: Polynomial test cases of various complexities<br />

# n 2d Matrix size Global minimum Terms<br />

1 2 8 49 −7.259 11<br />

2 2 8 49 −4.719 7<br />

3 2 10 81 −20.71 19<br />

4 2 10 81 −24.85 19<br />

5 4 4 81 −24.45 13<br />

6 4 4 81 −37.41 17<br />

7 2 12 121 −3.711 × 10 4 22<br />

8 2 12 121 −52.77 23<br />

9 3 6 125 −20.63 21<br />

10 3 6 125 −142.9 19<br />

11 2 14 169 − 6116 36<br />

12 2 14 169 − 1115 33<br />

13 2 16 225 −2.927 × 10 6 46<br />

14 2 16 225 − 2258 40<br />

15 3 8 343 − 2398 34<br />

16 3 8 343 −884.9 42<br />

17 2 24 529 −1.944 × 10 4 78<br />

18 2 24 529 −1.062 × 10 7 88<br />

19 3 10 729 −1.218 × 10 4 57<br />

20 3 10 729 −1.349 × 10 4 72<br />

21 3 12 1331 −8.138 × 10 5 115<br />

22 3 12 1331 −1.166 × 10 6 112<br />

that is used is given in Table 7.5. The heuristic method used here <strong>to</strong> compute the<br />

matrix-vec<strong>to</strong>r products is the least-increments method and the target selection of the<br />

iterative solvers is chosen such that the smallest real eigenvalues are computed first.<br />

Table 7.5: Settings for JD and JDQZ<br />

JD<br />

JDQZ<br />

# Eigenvalues : 1 # Eigenvalues : 1<br />

Tolerance : 5 × 10 −8 · (2d − 1) n Tolerance : 5 × 10 −8 · (2d − 1) n<br />

Imag. threshold : 1 × 10 −8 · (2d − 1) n Imag. threshold : 1 × 10 −8 · (2d − 1) n<br />

Max # iterations : 1000 Max # iterations : 1000<br />

Min dimension : 0.5 · (2d − 1) n Min dimension : 0.5 · (2d − 1) n<br />

Max dimension : 0.7 · (2d − 1) n Max dimension : 0.7 · (2d − 1) n<br />

Start vec<strong>to</strong>r : Random Start vec<strong>to</strong>r : Random<br />

Preconditioning : None Preconditioning : None<br />

Extraction : Refined Schur Decomposition<br />

: No<br />

Expansion : JD Conjugate pairs : No<br />

Linear Solver : GMRES Linear solver : GMRES<br />

Fix target : 0.01 Test space : 3<br />

Krylov space : Yes Lin solv <strong>to</strong>l : 0.7, 0.49<br />

Max # LS it. : 10 Max # LS it. : 15<br />

Thick restart : 1 vec<strong>to</strong>r Track : 1.0 × 10 −4<br />

The mean results of the 20 global minima computations of all the polynomials in<br />

the test set are given in Table 7.6. This table shows for JD and JDQZ the mean time<br />

and its standard deviation needed <strong>to</strong> compute the smallest real eigenvalue (denoted<br />

by mean ± standard deviation), the mean required number of matrix-vec<strong>to</strong>r product<br />

and its standard deviation, and the error. This error is defined as the mean of the

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!