Proof of Von Neumann Stability Analysis

Von Neumann stability analysis is the standard tool for determining stability of finite difference schemes for linear PDEs with constant coefficients. The method uses Fourier analysis to decompose the error into independent modes.

Statement

Theorem8.3Von Neumann Stability Criterion

Consider a finite difference scheme for a linear PDE with constant coefficients on a uniform grid, with periodic boundary conditions. The scheme is stable (in the $\ell^2$ norm) if and only if the amplification factor $g(\xi)$ satisfies $|g(\xi)| \leq 1 + C\Delta t$ for all frequencies $\xi \in [-\pi, \pi]$ and some constant $C$ independent of $\Delta t$ and $\Delta x$ . For practical purposes, the condition $|g(\xi)| \leq 1$ for all $\xi$ (the von Neumann condition) is equivalent to stability for dissipative schemes.

Proof

Fourier decomposition. On a periodic grid with $N$ points and spacing $\Delta x$ , any grid function $u_j^n$ can be written as a discrete Fourier series: $u_j^n = \sum_{k=-N/2}^{N/2-1} \hat{u}_k^n e^{i\xi_k j\Delta x}$ where $\xi_k = 2\pi k/(N\Delta x)$ .

Linearity and constant coefficients. For a linear scheme with constant coefficients, Fourier modes decouple: if $u_j^0 = e^{i\xi j\Delta x}$ , then $u_j^n = g(\xi)^n e^{i\xi j\Delta x}$ where $g(\xi)$ is the amplification factor, obtained by substituting $u_j^n = g^n e^{i\xi j\Delta x}$ into the scheme.

Stability $\Leftrightarrow$ bounded amplification. By Parseval's theorem: $\|u^n\|_{\ell^2}^2 = N\Delta x \sum_k |\hat{u}_k^n|^2 = N\Delta x \sum_k |g(\xi_k)|^{2n}|\hat{u}_k^0|^2$ .

(i) Necessity. If $|g(\xi_0)| > 1 + C\Delta t$ for some $\xi_0$ , choose $u_0 = e^{i\xi_0 jh}$ . Then $\|u^n\|_{\ell^2} = |g(\xi_0)|^n \|u^0\|_{\ell^2}$ . For $n = T/\Delta t$ : $|g(\xi_0)|^n > (1 + C\Delta t)^{T/\Delta t} \to e^{CT}$ , which is bounded. But if $|g(\xi_0)| > 1 + C\Delta t$ with $C$ growing as $\Delta t \to 0$ , the bound diverges. Specifically, if $|g(\xi_0)| \geq 1 + \alpha/\Delta t$ for some $\alpha > 0$ , then $|g|^n \sim e^{\alpha n/\Delta t} = e^{\alpha T/(\Delta t)^2} \to \infty$ , violating stability.

(ii) Sufficiency. If $|g(\xi)| \leq 1 + C\Delta t$ for all $\xi$ , then $|g(\xi)|^n \leq (1 + C\Delta t)^{T/\Delta t} \leq e^{CT}$ for $n\Delta t \leq T$ . By Parseval: $\|u^n\|_{\ell^2} \leq e^{CT}\|u^0\|_{\ell^2}$ , confirming $\ell^2$ stability.

Example: FTCS for heat equation. Substituting $u_j^n = g^n e^{i\xi j\Delta x}$ : $g = 1 + r(e^{i\xi\Delta x} - 2 + e^{-i\xi\Delta x}) = 1 - 4r\sin^2(\xi\Delta x/2)$ . For $|g| \leq 1$ : we need $-1 \leq 1 - 4r\sin^2(\xi\Delta x/2) \leq 1$ . The right inequality is automatic. The left requires $4r\sin^2(\xi\Delta x/2) \leq 2$ for all $\xi$ , i.e., $4r \leq 2$ , giving $r \leq 1/2$ . $\square$

■

ExampleStability of Advection Schemes

For $u_t + au_x = 0$ ( $a > 0$ ):

FTCS (central space): $g = 1 - i\nu\sin(\xi\Delta x)$ where $\nu = a\Delta t/\Delta x$ . Then $|g|^2 = 1 + \nu^2\sin^2(\xi\Delta x) > 1$ for all $\nu > 0$ . FTCS is unconditionally unstable for advection.
Upwind: $g = 1 - \nu(1 - e^{-i\xi\Delta x}) = 1 - \nu + \nu\cos(\xi\Delta x) - i\nu\sin(\xi\Delta x)$ . Stability requires $0 \leq \nu \leq 1$ (CFL condition).
Lax-Wendroff: $g = 1 - i\nu\sin(\xi\Delta x) - \nu^2(1-\cos(\xi\Delta x))$ . Stable for $|\nu| \leq 1$ .

RemarkLimitations of Von Neumann Analysis

Von Neumann analysis applies only to linear schemes with constant coefficients on uniform grids with periodic (or whole-space) boundary conditions. For variable coefficients, boundaries, and nonlinear problems, stability must be established by energy methods, maximum principles, or other techniques. Nevertheless, the von Neumann condition is a necessary condition even for variable-coefficient problems (the "frozen coefficient" principle).