11.08.2013 Views

Untitled - Cdm.unimo.it

Untitled - Cdm.unimo.it

Untitled - Cdm.unimo.it

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.

Derivative Matrices 149<br />

Only in a very few large-scale computations n is bigger than 100 (over 200 in extreme<br />

cases). In most of the applications n does not exceed 30. In this case <strong>it</strong> is convenient<br />

to use a direct method, such as Gauss elimination w<strong>it</strong>h maximal pivots, rather than<br />

an <strong>it</strong>erative method. Direct factorization (such as LU decompos<strong>it</strong>ion) is recommended<br />

when the same system has to be solved several times w<strong>it</strong>h different data sets. It is well<br />

known that the global cost of these procedures is proportional to n 3 .<br />

As we shall see in chapter thirteen, for partial differential equations, the derivative<br />

matrices in each variable have basically the same structure as those described here. This<br />

is because spectral methods are usually applied to boundary value problems, defined<br />

on domains obtained by direct product of intervals (rectangles, cubes, and so forth). In<br />

this context, to save both computer memory and CPU time, an <strong>it</strong>erative procedure is<br />

often preferable to a direct approach. A survey of the most popular <strong>it</strong>erative algor<strong>it</strong>hms<br />

in spectral methods is given in canuto, hussaini, quarteroni and zang(1988),<br />

chapter 5, and boyd(1989), chapter 12. The development of these techniques is based<br />

on the analysis of problems in one variable. Therefore, a detailed study of the simplest<br />

s<strong>it</strong>uations will be helpful in finding new strategies.<br />

A lot of problems are defined in the physical space (see section 7.4). The major<br />

drawback, however, is that the corresponding matrices are full and non symmetric. W<strong>it</strong>h<br />

the exception of the Chebyshev case, where the FFT algor<strong>it</strong>hm considerably reduces<br />

the amount of operations (see sections 4.3 and 7.2), the cost of a matrix-vector multi-<br />

plication is proportional to n 2 . To be compet<strong>it</strong>ive, an <strong>it</strong>erative method has to reach an<br />

accurate solution in very few <strong>it</strong>erations. To achieve this goal, <strong>it</strong> is imperative to work<br />

w<strong>it</strong>h precond<strong>it</strong>ioned matrices. How to construct and implement this kind of matrix is<br />

discussed in the next chapter. For the moment, let us briefly recall some basic facts.<br />

Let D be a n×n matrix (n ≥ 1) and ¯q a given vector in R n . We want to determine<br />

the solution ¯p ∈ R n of the linear system<br />

(7.6.1) D¯p = ¯q, where det(D) = 0.<br />

A one-step <strong>it</strong>erative method to evaluate ¯p is obtained by introducing a new n × n non<br />

singular matrix R and defining an in<strong>it</strong>ial guess ¯p (0) ∈ R n . Successively, one constructs

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

Saved successfully!

Ooh no, something went wrong!