22.01.2014 Views

download searchable PDF of Circuit Design book - IEEE Global ...

download searchable PDF of Circuit Design book - IEEE Global ...

download searchable PDF of Circuit Design book - IEEE Global ...

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.

Complex Zeros <strong>of</strong> Complex Polynomials 41<br />

axis intersections and spaced between these at 45 degrees. Running the<br />

program shows that these roots have coordinates equal to either unity or<br />

I/.fi, with agreement through eight significant figures on most computers.<br />

Adding the temporary statements in Table 3.1 and running the program again<br />

show how few times the algorithm needs to reduce the step length in a chosen<br />

search direction. Such reductions usually occur early in the search at some<br />

distance from the root (minimum) location.<br />

The roots are printed whenever changes in x and yare less than I.E - 5 or,<br />

following 10 step-size reductions, when F is no greater than I.E-8. The<br />

algorithm is aborted if the latter condition fails or when there have been more<br />

than 50 iterations (search directions). Little memory is required; there are two<br />

vectors (single-subscript arrays) for the coefficient's rectangular components<br />

a k and b k<br />

and two more vectors for X k and Yk in (3.26). These are dimensioned<br />

to hold N elements, where N is the maximum polynomial degree.<br />

However, on computers with exponent ranges <strong>of</strong> about 1Oexp( + / - 37),<br />

numerical overflow occurs for polynomials <strong>of</strong> degree greater than 20. Exponent<br />

ranges to lOexp( + / -99) usually solve polynomials up to degree 35. The<br />

difficulty occurs in the large polynomial value because <strong>of</strong> the poor initial root<br />

guess <strong>of</strong> z=O.1 +jl.<br />

Gradient root finders such as Moore's suffer from a chronic problem with<br />

multiple roots. Consideration <strong>of</strong> a function such as y=(x-I)' and its derivative<br />

shows that repeated (multiple) roots cause gradients (coordinate derivatives)<br />

that tend to zero in the neighborhood <strong>of</strong> the root. This causes some<br />

inaccuracy in repeated root values, because Moore's method depends on<br />

gradient scaling in the step length formulas (3.9)-(3.11). The code in Table 3.2<br />

can be added to print the value <strong>of</strong> (3.11).<br />

Table 3.2.<br />

Temporary Code to Print the<br />

Squared Length <strong>of</strong> a Gradieut<br />

2085 PRINT'GRAD MAG SQD-"; PM<br />

Example 3.4. Add the program code m Table 3.2 to root-finder Program<br />

B3-1 and solve the polynomial<br />

1080 + 2466z+ 2025z' + 765z 3 + 135z 4 + 9z' = 9(z+ 1)(z+ 2)(z+ 3)(z+4)(z+ 5).<br />

Note that the "GRAD MAG SQD" value (3.11) is well scaled. Then solve the<br />

polynomial<br />

54+ 135z+ 126z' + 56z 3 + 12z 4 +z'=(z+ 1)(z+2)(z+3)3.<br />

Note that the squared gradient length used as a divisor in the search step<br />

adjustment is well behaved until the z= - 3 repeated root is encountered.<br />

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

Saved successfully!

Ooh no, something went wrong!