Convergence, Consistency, and Stability

Denition A one-step nite dierence scheme approximating a partial dierential equation is a convergent scheme if for any solution to the partial dierential equation, n 0 u(t, x), and solutions to the nite dierence scheme, vi , such that vi converges n to u0 (x) as ix converges to x, then vi converges to u(t, x) as (nt, ix) converges to (t, x) as t, x converge to 0. Denition Given a partial dierential equation P u = f and a nite dierence scheme, Pt,x v = f , we say that the nite dierence scheme is consistent with the partial dierential equation if for any smooth function (x, t) P Pt,x 0 as t, x 0. Example Consider the one-way wave equation given by the operator P = /t + /x P = t + x with greater than 0. We will evaluate the consistency of the forward-time forward-space scheme with dierence operator Pt,x given by Pt,x = where n i = (nt, ix). We begin by taking the Taylor expansion of the function in t and x about (tn , xi ). We have that 1 2 +1 3 n = n i + tt + t tt + O (t ) i 2 1 n 2 3 n i+1 = i + xx + x xx + O (x ). 2 This gives us 1 1 Pt,x = t + x + ttt + xxx + O(t2 ) + O(x2 ) 2 2 Thus
1 2 2 ttt + 1 P Pt,x = 2 2 xxx + O (t ) + O (x ) +1 n n n n i i i + i+1 t x

(1)

0 as (t, x) 0. Thus, this scheme is consistent.

Denition The L2-norm of a grid function w, denoted by w

x , 1 /2

is dened as .

x
m=

|wm |

Denition n A nite dierence scheme Pt,x vi = 0 for a rst-order equation is stable in a stability region if there is an integer J such that for any positive time T , there is a constant CT such that
J

vn

CT
j =0

vj

for 0 nt T , with (t, x) .

Von Neumann Analysis

Proving stability directly from the denition is quite dicult, in general. Instead, it is easier to use tools from Fourier analysis to evaluate the stability of nite dierence schemes. In particular, it can be shown that, for some solution to a nite dierence scheme v n , there is a simple mathematical relationship between the Fourier transforms v n ( ) and v 0 ( ) given by v n ( ) = g (x, t, x)n v 0 ( ) where g (x, t, x) = g (, t, x) is called the amplication factor and v n ( ) n is the amplitude of the frequency in the solution v (note that the second superscript n in the above equation is a power and not an index). This quantity is so named because it represents the amount that each frequency in the solution is amplied in advancing the solution one time step. That this relationship arises from such analysis should not be surprising: the fundamental power of Fourier transforms is that they transform dierentiation in the temporal domain to multiplication in the frequency domain. Denition A one-step nite dierence scheme with constant coecients is stable in a stability region if and only if there is a constant K (independent of , t, and x) such that |g (, t, x)| 1 + K t with (t, x) . If g (, t, x) is independent of t and x, the stability condition may be replaced with the restricted stability condition |g ()| 1.

Example Consider again the one-way wave equation and let us evaluate the stability of the forward-time forward-space scheme given in equation (1). It can be shown that n all solutions of any one-step dierence scheme will have the form vj = g n eij . To analyze the stability of such a scheme, we may therefore substitute accordingly and solve for g . g n ei(j +1) g n eij g n+1 eij g n eij + =0 t x g n eij (g 1) g n eij (ei 1) = t x t (g 1) = (ei 1) x t t i g =1+ e x x g = 1 + ei where = t/x and is positive. We can then calculate 1 |g |2 = 1 + 4(1 + ) sin2 . 2 Since is constant, we may use the restricted stability condition and we see that |g | is greater than 1 for = 0. Therefore, this scheme is unstable.

The Lax-Richtmyer Equivalence Theorem

The Lax-Richtmyer Equivalence Theorem is often called the Fundamental Theorem of Numerical Analysis, even though it is only applicable to the small subset of linear numerical methods for well-posed, linear partial dierential equations. Along with Dahlquists equivalence theorem for ordinary dierential equations, the notion that the relationship consistency + stability convergence always holds has caused a great deal of confusion in the numerical analysis of dierential equations. In the case of PDEs, mathematicians are most often interested in nonlinear phenomena, for which Lax-Richtmyer does not apply. More damningly, the forward implication that consistency + stability = convergence is trivial for linear schemes, and thus it is only the converse notion that convergence = stability that the theorem contributes. The intuition that the theorem gives for problems that fall outside the scope of Lax-Richtmyer, however, is faulty, since consistency and stability are often insucient for convergence, and convergence need not imply stability in general.

Suppose we have a stable, linear numerical method that is consistent with a well-posed linear partial dierential equation. Suppose further that the global error at time T = N t is the grid function denoted by E N = QN q N where Qn is the numerical approximation to the solution and q n is the exact solution n N . We will denote our method by N (), such that Qn+1 = N (Qn ). We claim that it is straightforward to show that the method is convergent, which amounts to proving that E N 0 as t 0 and N t T . Since N is stable, we know that there exists a constant CT depending only on T such that N n CT , n N = T /t. We can also express the local trunctation error as the one step error divided by the time step, or 1 n = N (q n ) q n+1 , t and note that consistency implies that n 0 as t 0. We can derive a simple recurrence for the error at time n + 1: E n+1 = Qn+1 q n+1 = N (Qn ) q n+1 = N (E n + q n ) q n+1 = N (E n + q n ) N (q n ) + N (q n ) q n+1 = N (E n + q n ) N (q n ) + = N (E ) +
n

t n

where we use the linearity of N () in the last equality. From this formulation, it is simple to examine the norm of the error and argue that the requirements for convergence are met by stability and consistency.