11.08.2013 Views

Untitled - Cdm.unimo.it

Untitled - Cdm.unimo.it

Untitled - Cdm.unimo.it

SHOW MORE
SHOW LESS

Create successful ePaper yourself

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

Eigenvalue Analysis 169<br />

Next, following orszag(1980), we consider the matrix R := {rij} 0≤i≤n , w<strong>it</strong>h entries<br />

0≤j≤n<br />

⎪⎨<br />

(8.4.1) rij :=<br />

⎧<br />

1 if i = j = 0 or i = j = n,<br />

−1<br />

h (n)<br />

i ˆ h (n)<br />

i<br />

2<br />

h (n)<br />

i+1h(n) i<br />

−1<br />

h (n)<br />

i+1 ˆ h (n)<br />

i<br />

⎪⎩<br />

0 elsewhere.<br />

if 1 ≤ i = j + 1 ≤ n − 1,<br />

if 1 ≤ i = j ≤ n − 1,<br />

if 1 ≤ i = j − 1 ≤ n − 1,<br />

The matrix R represents the second-order centered fin<strong>it</strong>e-differences operator defined<br />

on the grid points η (n)<br />

j , 0 ≤ j ≤ n. If f is a given regular function, the matrix R<br />

applied to the vector {f(η (n)<br />

j )}0≤j≤n furnishes an approximation of −f ′′ (η (n)<br />

i ),<br />

1 ≤ i ≤ n − 1. We remark that this kind of discretization is of local type (see section<br />

7.2), hence <strong>it</strong> does not possess the good convergence behavior of spectral methods, even<br />

if the nodes of a Gauss formula are involved. However, R is tridiagonal, thus, the<br />

computation of R −1 can be performed in a fast way. Therefore, R satisfies the first<br />

requirement to be a good precond<strong>it</strong>ioning matrix. Now, we must study the spectrum of<br />

R −1 D. We denote by Λn,m, 0 ≤ m ≤ n, the eigenvalues obtained after precond<strong>it</strong>ioning.<br />

Two of them, say Λn,0 and Λn,n, are equal to 1. Numerical experiments indicate the<br />

existence of a constant c > 1 (depending on α and β) such that, for any n ≥ 2, one<br />

has<br />

(8.4.2) 1 ≤ |Λn,m| < c, 1 ≤ m ≤ n − 1.<br />

Comparing this w<strong>it</strong>h (8.3.6), the improvement is impressive. Furthermore, the Λn,m’s<br />

are real and pos<strong>it</strong>ive and c is in general less than 2.5 . We can be more precise in the<br />

Chebyshev case. In fact, in haldenwang, labrosse, abboudi and deville(1984),

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

Saved successfully!

Ooh no, something went wrong!