20.05.2014 Views

link to my thesis

link to my thesis

link to my thesis

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

120 CHAPTER 7. NUMERICAL EXPERIMENTS<br />

Figure 7.4: Eigenvalue spectra of the matrices A p1<br />

and A xi ,i=1,...,4<br />

third columns contain the number of matrix-vec<strong>to</strong>r products of the method with<br />

the matrix A p1 and the matrices A xi respectively. The fourth column shows the<br />

amount of floating point operations (flops) needed <strong>to</strong> perform all the matrix-vec<strong>to</strong>r<br />

products; the amount of flops is computed as the sum of the products of the number<br />

of multiplications with the number of non-zeros of the corresponding matrices: i.e.,<br />

the JDCOMM method which iterates with the matrix A x1 in the inner loop and with<br />

the matrix A p1 in the outer loop requires 69 iterations with the matrix A p1 and 459<br />

iterations with the matrix A x1 . This costs 69 × 1326556 + 459 × 63186 = 120534738<br />

flops. Finally, column five shows the required computation time.<br />

It is easy <strong>to</strong> see that all the methods compute the same leftmost real eigenvalue,<br />

and moreover, that the performance of the JDCOMM method is much better than the<br />

performance of the JD method, in terms of computation time as well as in required<br />

floating point operations.<br />

Figure 7.5, shows the plot of the norms of the residuals in the eigenvalue equation<br />

(at each JD outer step), against the number of matrix-vec<strong>to</strong>r products.<br />

To put the performance of the JD approach of this paper in<strong>to</strong> perspective, we<br />

discuss briefly the outcomes of the computation of the global minimum of polynomial<br />

(7.6) by the software packages SOSTOOLS and GloptiPoly, which employ <strong>to</strong>tally different<br />

approaches. Both, SOSTOOLS and GloptiPoly, compute the same global optimum<br />

and locations of polynomial p 1 as shown in Table 7.11 (using the default parameter<br />

settings). To determine these, SOSTOOLS uses 19.3 seconds and GloptiPoly uses 17.2<br />

seconds. This shows that the JDCOMM method has the potential <strong>to</strong> outperform these

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

Saved successfully!

Ooh no, something went wrong!