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.

-- -~---<br />

Conjugate Gradient Search 133<br />

the columns in matrix P; then (5.43) defines the substitutions<br />

Using these in (5.22) produces<br />

x,=y'+Y2'<br />

x2= -y, +Y2'<br />

Q(Y) = 72y, + 32Y2'<br />

(5.45)<br />

(5.46)<br />

(5.47)<br />

so that the cross terms are indeed removed, and the minimum could be found<br />

in no more than two linear searches.<br />

It is straightforward to show that changes in the gradient vectors <strong>of</strong> a<br />

quadratic function are mapped by a constant linear transformation to the<br />

corresponding changes in the variable vectors. As Figure 5.18 illustrates,<br />

points A and B in the x space have gradient values (perpendicular to their<br />

level curve), and these gradient vectors can be plotted in their own space.<br />

There may be more than one x with the same g. Apply the gradient expression<br />

(5.13) <strong>of</strong> a quadratic function to points x; and X;+1 and their corresponding<br />

gradients ri and ri+'; the two equations may be subtracted to yield<br />

(ri+1-g;)=A(x;+I-xl (5.48)<br />

Using tJ. to indicate the differences and inverting (5.48), the mapping result is<br />

tJ.x=A-'tJ.g. (5.49)<br />

This result was anticipated by Newton's step in (5.34), which went to a<br />

minimum where ri+ 1=0 was required. The importance <strong>of</strong> (5.49) is that it<br />

shows the invariance <strong>of</strong> that mapping, independent <strong>of</strong> locations on any<br />

quadratic surface.<br />

Finally, it is shown that the altitude above the minimum value <strong>of</strong> a<br />

quadratic surface at some point p is equal to a quadratic form composed <strong>of</strong><br />

the gradient at the point in question, g(p), and the inverse <strong>of</strong> its constant<br />

lal<br />

Ihl<br />

Figure S.lS. A mapping <strong>of</strong> variable space to gradient space. (a) Constant objective function<br />

curves in the variable space; (b) corresponding loci and points in the gradient space. [From<br />

Davidon, 1959.] .

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

Saved successfully!

Ooh no, something went wrong!