Skip to main content

Subsection B.1.1 Nomenclature

2026101400 | numerical methods, differential equations, numerical analysis

Consistency, Stability, and Convergence.

Whenever we approximate differential equations (ODEs and PDEs) via numerical discretization (e.g., finite differences), we can think about the accuracy of the simulated solution with respect to the ground truth continuous solution using three fundamental concepts:

Definition B.1.1 Consistency.

A numerical scheme is said to be consistent with respect to a differential equation if the local truncation error \(\tau(\Delta t, \Delta x)\) vanishes to zero as the discretization step sizes tend towards zero:
\begin{equation*} \lim_{\Delta t, \Delta x \to 0} \tau(\Delta t, \Delta x) = 0 \end{equation*}

Remark B.1.2.

Consistency is a measure of how accurately a discretization scheme approximates the exact, smooth analytical solution. Specifically, given
\begin{equation*} \tau = \operatorname{O}(\Delta t^p) + \operatorname{O}(\Delta x^q)\text{,} \end{equation*}
the scheme is then said to be consistent with order \(p\) in \(t\) and order \(n\) in \(x\text{.}\)

Definition B.1.3 Stability.

A numerical scheme is said to be stable if small perturbations (e.g., round-off errors, small fluctuations in initial/boundary data) remain bounded and do not blow up towards infinity as the scheme advances.

Definition B.1.4 Convergence.

A numerical method is said to be convergent if the global error between the ground truth analytical/continuous solution \(u(x,t)\) and the numerical approximation \(u_j^n\) tends to zero at any fixed point \((x,t)\) as the grid spacings are decreased towards zero:
\begin{equation*} \lim_{\Delta t, \Delta x \to 0} \| u(\cdot, t_n) - u^n \| = 0 \end{equation*}
Table B.1.6. Comparison of Numerical Scheme Foundations
Property Question Answered Evaluation Target Common Diagnostic Tool
Consistency Does the discrete formulation of the PDE match the analytical ground truth PDE in the limit? Local Truncation Error (\(\tau\)) on exact solution Taylor series expansion
Stability Do numerical errors and perturbations remain bounded? Growth factor of numerical modes / perturbation propagation Von Neumann analysis, Matrix eigenvalues, CFL condition
Convergence Does the computed simulation approach the ground truth analytical solution? Global error (\(\epsilon_j^n = u(x_j, t_n) - u_j^n\)) Lax Equivalence Theorem (well-posed linear systems only), grid refinement studies