In computational mathematics, we often face partial differential equations (PDEs) that are difficult — or impossible — to solve analytically. To approach these problems, we turn to numerical methods like the finite difference method (FDM).
But how do we ensure that these discrete approximations actually behave like the original equation? That’s where key ideas like grid projection, residual, consistency, and stability come into play.
We begin with a general PDE:
\[Lu = f\]Where:
To solve this numerically, we define a discrete grid over space and time:
\[x_m = x_0 + mh, \quad t_n = t_0 + n\tau\]Here:
The grid projection of a continuous function like $u$ or $f$ is:
\[u_{\tau h} = u(x_m, t_n), \quad \phi_{mn} = f(x_m, t_n)\]The function $y_{mn}$, defined only at discrete points, is called a grid function, and it approximates the continuous solution: $y_{mn} \approx u(x_m, t_n)$.
We now discretize the operator $L$ to form a finite difference operator, which we’ll also call $L$ by convention. This gives us the discrete version of the PDE:
\[Ly = \phi\]This is an algebraic system that we can solve numerically. But how do we know if this system faithfully represents the original PDE?
To test how well our finite difference method approximates the PDE, we define the residual:
\[r_{\tau h} = L u_{\tau h} - \phi\]This measures the discrepancy between the finite difference scheme and the exact solution, evaluated on the grid.
If the method is good, this residual should become small as the grid is refined.
A finite difference method is called consistent with the original PDE if the residual tends to zero as the grid is refined:
\[\|r_{\tau h}\| \to 0 \quad \text{as} \quad h, \tau \to 0\]Here, $|\cdot|$ is a suitable norm (e.g. $L^2$, $L^\infty$) over the space of grid functions.
Intuition: Consistency checks if your scheme captures the dynamics of the original PDE correctly.
Stability ensures that small perturbations in input data or the right-hand side do not lead to large errors in the numerical solution.
Formally, the method is stable if, for two problems with perturbed right-hand sides:
\[Ly^{(1)} = f + \varepsilon^{(1)}, \quad Ly^{(2)} = f + \varepsilon^{(2)},\]the solutions satisfy:
\[\|y^{(1)} - y^{(2)}\| \leq C \left( \|\varepsilon^{(1)}\| + \|\varepsilon^{(2)}\| \right)\]for a constant $C$ independent of $h, \tau$, and for small enough perturbations.
Intuition: Stability asks whether the method is sensitive to noise or rounding errors. A stable method won’t “explode” due to small mistakes.
If a finite difference method is consistent and stable, then it is convergent.
That is:
\[y_{mn} \to u(x_m, t_n) \quad \text{as} \quad h, \tau \to 0\]This tells us the numerical solution truly approximates the continuous one.
| Concept | What it means | Why it matters |
|---|---|---|
| Grid projection | Sample continuous functions on a discrete grid | Bridge between continuous and discrete |
| Consistency | Scheme matches PDE as grid refines | Are we solving the right problem? |
| Stability | Errors stay under control | Will it explode? |
| Convergence | Numerical solution approaches the true one | End goal: correctness |
Here are common methods used to prove stability:
For a system $Ay = b$, prove:
\[\|A^{-1}\| \leq C\]Construct a discrete inequality like:
\[\|y^{n+1}\| \leq \|y^n\| + \text{(controlled terms)}\]Use convex combinations of past values to show:
\[\|y^{n+1}\|_\infty \leq \|y^n\|_\infty\]Understanding how finite difference methods transition from the continuous world to the discrete is crucial for building reliable numerical solvers. You now know:
A good scheme must be consistent, stable, and therefore convergent. Without all three, your numerical answers might be fast — but they won’t be right.
Written with curiosity and appreciation for beautiful math 💻✨