Skip to content

Stability, Consistency, and Convergence

The reliability of a finite difference scheme depends on three fundamental properties: consistency (local accuracy), stability (bounded error growth), and convergence (approach to the true solution). The Lax Equivalence Theorem connects these concepts.


The Fundamental Principle

For linear, well-posed problems:

\[ \boxed{ \text{Consistency} + \text{Stability} \Longrightarrow \text{Convergence} } \]

This is the Lax Equivalence Theorem: a consistent scheme converges if and only if it is stable.


Consistency

Definition

A scheme is consistent if the local truncation error vanishes as the mesh is refined.

For a PDE \(\mathcal{L}u = 0\) and discrete approximation \(\mathcal{L}_{\Delta}u_{\Delta} = 0\):

\[ \text{Local truncation error} = \mathcal{L}_{\Delta}u - \mathcal{L}u \]

where \(u\) is the exact solution.

Consistency: LTE \(\to 0\) as \(\Delta\tau, \Delta S \to 0\).

Order of Consistency

If LTE \(= O((\Delta\tau)^p + (\Delta S)^q)\), the scheme is: - \(p\)-th order in time - \(q\)-th order in space

Example: Crank-Nicolson

For smooth solutions:

\[ \text{LTE} = O((\Delta\tau)^2 + (\Delta S)^2) \]

Crank-Nicolson is second-order consistent in both time and space.


Stability

Definition

A scheme is stable if errors remain bounded as the computation proceeds.

For a linear scheme \(\mathbf{u}^{n+1} = B\mathbf{u}^n\), stability requires:

\[ \boxed{ \|B^n\| \leq C \quad \text{for all } n\Delta\tau \leq T } \]

for some constant \(C\) independent of the mesh.

Von Neumann (Fourier) Analysis

The standard method for analyzing stability of constant-coefficient schemes.

Idea: Decompose errors into Fourier modes \(e^{ik\xi}\) and analyze their growth.

Amplification factor: For mode \(k\), let \(G(k)\) be the ratio of amplitudes at successive time steps.

Stability criterion:

\[ \boxed{ |G(k)| \leq 1 \quad \text{for all } k } \]

Example: Explicit Scheme for Heat Equation

For \(u_\tau = \frac{1}{2}u_{xx}\) with explicit discretization:

\[ G = 1 - 2\lambda(1 - \cos(k\Delta x)) \]

where \(\lambda = \frac{\Delta\tau}{(\Delta x)^2}\).

Stability requires: \(|G| \leq 1 \Rightarrow \lambda \leq \frac{1}{2}\)

\[ \boxed{ \Delta\tau \leq \frac{(\Delta x)^2}{2} \quad \text{(CFL condition)} } \]

Example: Implicit Scheme

\[ G = \frac{1}{1 + 2\lambda(1 - \cos(k\Delta x))} \]

Since the denominator \(> 1\), we have \(|G| < 1\) for all \(\lambda > 0\).

Unconditionally stable: No restriction on \(\Delta\tau\).

Example: Crank-Nicolson

\[ G = \frac{1 - \lambda(1 - \cos(k\Delta x))}{1 + \lambda(1 - \cos(k\Delta x))} \]

Unconditionally stable: \(|G| \leq 1\) for all \(\lambda\).


Convergence

Definition

A scheme converges if the numerical solution approaches the exact solution as the mesh is refined:

\[ \lim_{\Delta\tau, \Delta S \to 0} \max_{n,j} |u_j^n - u(\tau_n, S_j)| = 0 \]

Convergence Rate

If the error satisfies:

\[ \|u_{\Delta} - u\| = O((\Delta\tau)^p + (\Delta S)^q) \]

the scheme has convergence rate \((p, q)\).

The Lax Equivalence Theorem

Theorem: For a well-posed linear initial value problem and a consistent finite difference scheme, stability is equivalent to convergence.

Proof sketch: 1. Let \(e^n = u^n - u(\tau_n)\) be the global error 2. Error evolution: \(e^{n+1} = Be^n + \text{LTE}\) 3. By induction: \(e^n = B^n e^0 + \sum_{k=0}^{n-1} B^{n-1-k}\text{LTE}_k\) 4. If stable: \(\|B^n\| \leq C\), so errors remain bounded 5. If consistent: LTE \(\to 0\), so total error \(\to 0\)


Practical Stability Analysis

Matrix Criterion

For \(\mathbf{u}^{n+1} = B\mathbf{u}^n\), stability requires:

\[ \rho(B) \leq 1 + O(\Delta\tau) \]

where \(\rho(B)\) is the spectral radius (largest eigenvalue magnitude).

Energy Methods

For some problems, multiply by \(u\) and sum to get energy estimates:

\[ \|u^{n+1}\|^2 \leq \|u^n\|^2 \]

This directly proves stability.

Comparison Principles

Monotone schemes preserve the discrete maximum principle:

\[ \min_j u_j^n \leq u_j^{n+1} \leq \max_j u_j^n \]

This implies \(L^\infty\) stability.


Issues with Non-Smooth Data

The Problem

Option payoffs like \((S-K)^+\) are not smooth at \(S = K\). This affects:

  1. Consistency: Taylor expansions assume smoothness
  2. Convergence rate: Observed order may be lower than theoretical

Observed vs Theoretical Convergence

Data Smoothness Theoretical Order Observed Order
\(C^\infty\) 2 (C-N) 2
\(C^0\) (continuous) 2 1-1.5
Kink at strike 2 0.5-1

Remedies

  1. Rannacher smoothing: Implicit steps near maturity
  2. Payoff smoothing: Replace \((S-K)^+\) with smooth approximation
  3. Grid adaptation: Fine grid near strike
  4. Richardson extrapolation: Improve convergence from multiple grids

The CFL Condition

The Courant-Friedrichs-Lewy (CFL) condition is a necessary condition for stability of explicit schemes.

For Advection Equation u_t + cu_x = 0

\[ \boxed{ |c|\frac{\Delta t}{\Delta x} \leq 1 } \]

Interpretation: The numerical domain of dependence must contain the physical domain of dependence.

For Diffusion Equation u_t = Du_xx

\[ \boxed{ D\frac{\Delta t}{(\Delta x)^2} \leq \frac{1}{2} } \]

For Black-Scholes (Variable Coefficients)

The CFL condition involves the local diffusion coefficient \(\frac{1}{2}\sigma^2 S^2\):

\[ \frac{\sigma^2 S_{\max}^2 \Delta\tau}{2(\Delta S)^2} \leq \frac{1}{2} \]

Convergence Study (Practical)

To verify implementation, perform a convergence study:

  1. Compute solutions on grids with \(M, 2M, 4M, \ldots\) points
  2. Compare to analytical solution (if available) or finest grid
  3. Plot error vs grid size on log-log scale
  4. Slope gives convergence order

Example Results

Grid Error Ratio
\(M = 50\) 0.0412
\(M = 100\) 0.0103 4.0
\(M = 200\) 0.0026 4.0

Ratio \(\approx 4\) confirms second-order convergence (error \(\sim M^{-2}\)).


Summary

\[ \boxed{ \text{Consistency} + \text{Stability} \Longleftrightarrow \text{Convergence} } \]
Concept Definition Verification
Consistency LTE \(\to 0\) Taylor expansion
Stability \(\|B^n\| \leq C\) Von Neumann analysis
Convergence \(u_\Delta \to u\) Lax theorem + above
Scheme Stability CFL Restriction
Explicit Conditional \(\Delta\tau \leq C(\Delta S)^2\)
Implicit Unconditional None
Crank-Nicolson Unconditional None

A well-designed finite difference scheme must be both consistent and stable to guarantee convergence to the correct solution.


Exercises

Exercise 1. Consider the explicit scheme for the heat equation \(u_\tau = \frac{1}{2}u_{xx}\) with mesh ratio \(\lambda = \Delta\tau / (\Delta x)^2\). Show that the amplification factor is \(G = 1 - 2\lambda(1 - \cos(k\Delta x))\) and derive the stability condition \(\lambda \leq 1/2\) directly from \(|G| \leq 1\).

Solution to Exercise 1

The explicit scheme for \(u_\tau = \frac{1}{2}u_{xx}\) with mesh ratio \(\lambda = \Delta\tau / (\Delta x)^2\) is:

\[ u_j^{n+1} = u_j^n + \lambda(u_{j+1}^n - 2u_j^n + u_{j-1}^n) \]

Substitute a single Fourier mode \(\varepsilon_j^n = G^n e^{ikj\Delta x}\) into the difference equation:

\[ G^{n+1} e^{ikj\Delta x} = G^n e^{ikj\Delta x} + \lambda G^n \bigl(e^{ik(j+1)\Delta x} - 2e^{ikj\Delta x} + e^{ik(j-1)\Delta x}\bigr) \]

Dividing both sides by \(G^n e^{ikj\Delta x}\):

\[ G = 1 + \lambda(e^{ik\Delta x} + e^{-ik\Delta x} - 2) = 1 + \lambda(2\cos(k\Delta x) - 2) \]

Using the identity \(\cos\theta - 1 = -2\sin^2(\theta/2)\):

\[ G = 1 - 4\lambda\sin^2\!\left(\frac{k\Delta x}{2}\right) = 1 - 2\lambda(1 - \cos(k\Delta x)) \]

This confirms \(G = 1 - 2\lambda(1 - \cos(k\Delta x))\).

For stability we need \(|G| \leq 1\). Since \(\sin^2(k\Delta x / 2) \in [0, 1]\), the amplification factor ranges from \(G = 1\) (when \(k\Delta x = 0\)) to \(G = 1 - 4\lambda\) (when \(k\Delta x = \pi\)).

The condition \(|G| \leq 1\) requires:

\[ -1 \leq 1 - 4\lambda \leq 1 \]

The right inequality \(1 - 4\lambda \leq 1\) is always satisfied for \(\lambda > 0\). The left inequality gives \(1 - 4\lambda \geq -1\), i.e., \(4\lambda \leq 2\), hence \(\lambda \leq 1/2\).

Note that here the diffusion coefficient is \(D = 1/2\), so \(\lambda = \Delta\tau / (\Delta x)^2\) and the CFL condition \(D\lambda' \leq 1/2\) with \(\lambda' = \Delta\tau / (\Delta x)^2\) gives \(\Delta\tau \leq (\Delta x)^2 / (2D) = (\Delta x)^2\), consistent with \(\lambda \leq 1/2\) under the convention used.


Exercise 2. Verify the consistency of the Crank-Nicolson scheme by substituting the exact solution into the discrete equation and performing a Taylor expansion. Show that the local truncation error is

\[ \text{LTE} = O((\Delta\tau)^2 + (\Delta S)^2) \]

Identify the leading-order error terms explicitly.

Solution to Exercise 2

The Crank-Nicolson scheme for \(u_\tau = \frac{1}{2}\sigma^2 u_{SS}\) is:

\[ \frac{u_j^{n+1} - u_j^n}{\Delta\tau} = \frac{1}{2}\sigma^2 \cdot \frac{1}{2}\left[\frac{u_{j+1}^n - 2u_j^n + u_{j-1}^n}{(\Delta S)^2} + \frac{u_{j+1}^{n+1} - 2u_j^{n+1} + u_{j-1}^{n+1}}{(\Delta S)^2}\right] \]

Substitute the exact solution \(u(S_j, \tau_n)\) and expand in Taylor series around \((S_j, \tau_{n+1/2})\), the midpoint in time.

Temporal discretization: With \(\tau_{n+1/2} = \tau_n + \Delta\tau/2\):

\[ u_j^{n+1} = u_{n+1/2} + \frac{\Delta\tau}{2}u_\tau + \frac{1}{2}\left(\frac{\Delta\tau}{2}\right)^2 u_{\tau\tau} + O((\Delta\tau)^3) \]
\[ u_j^n = u_{n+1/2} - \frac{\Delta\tau}{2}u_\tau + \frac{1}{2}\left(\frac{\Delta\tau}{2}\right)^2 u_{\tau\tau} + O((\Delta\tau)^3) \]

So \((u_j^{n+1} - u_j^n)/\Delta\tau = u_\tau\big|_{n+1/2} + O((\Delta\tau)^2)\).

Spatial discretization: The central difference \(\delta_{SS}^2 u = (u_{j+1} - 2u_j + u_{j-1})/(\Delta S)^2 = u_{SS} + \frac{1}{12}(\Delta S)^2 u_{SSSS} + O((\Delta S)^4)\). The Crank-Nicolson average at levels \(n\) and \(n+1\) evaluates to \(u_{SS}\big|_{n+1/2} + O((\Delta S)^2 + (\Delta\tau)^2)\).

Combining, the local truncation error is:

\[ \text{LTE} = \left[u_\tau - \frac{1}{2}\sigma^2 u_{SS}\right]_{n+1/2} + O((\Delta\tau)^2) + O((\Delta S)^2) \]

The bracketed term vanishes because \(u\) satisfies the PDE. The leading-order error terms are:

  • Time: \(\frac{1}{24}(\Delta\tau)^2 u_{\tau\tau\tau}\)
  • Space: \(\frac{1}{12}\sigma^2(\Delta S)^2 u_{SSSS}\)

Therefore LTE \(= O((\Delta\tau)^2 + (\Delta S)^2)\), confirming second-order consistency in both time and space.


Exercise 3. The Lax Equivalence Theorem states that for a consistent, well-posed linear problem, stability is equivalent to convergence. Explain why each of the three hypotheses (linearity, well-posedness, consistency) is necessary by giving a counterexample or argument for what can go wrong when each is removed.

Solution to Exercise 3

Linearity: The Lax Equivalence Theorem applies to linear problems. Without linearity, the superposition principle fails: errors cannot be decomposed into independent modes, and the relationship between stability and convergence breaks down. A simple counterexample is the nonlinear ODE \(u' = u^2\) discretized with forward Euler. The scheme can be consistent and stable for small perturbations yet fail to converge for large perturbations because the error evolution is inherently nonlinear.

Well-posedness: The PDE must have a unique solution that depends continuously on the data. The backward heat equation \(u_t = -u_{xx}\) is ill-posed: arbitrarily small perturbations in the initial data lead to unbounded growth in the exact solution. No consistent, stable numerical scheme can converge to a meaningful solution of an ill-posed problem, because the "exact solution" itself is not well-defined under perturbation.

Consistency: Without consistency, the scheme may be stable (errors do not grow) but converge to the wrong answer. For example, consider the trivial scheme \(u_j^{n+1} = u_j^n\) for the heat equation \(u_\tau = u_{xx}\). This scheme is perfectly stable (\(\|B^n\| = 1\)) but has a non-vanishing truncation error as \(\Delta\tau, \Delta x \to 0\) (it approximates \(u_\tau = 0\), not \(u_\tau = u_{xx}\)). The numerical solution converges — but to the initial condition, not to the solution of the heat equation.


Exercise 4. A convergence study on four successively refined grids produces the following maximum errors for a European call priced via Crank-Nicolson:

Grid (\(M\)) Max Error
50 0.0820
100 0.0205
200 0.0051
400 0.0013

Compute the error ratios and confirm second-order convergence. If the error were instead \(O((\Delta S)^{3/2})\), what ratios would you expect?

Solution to Exercise 4

The error ratios between successive grids are:

\[ \frac{0.0820}{0.0205} = 4.0, \quad \frac{0.0205}{0.0051} = 4.02, \quad \frac{0.0051}{0.0013} = 3.92 \]

All ratios are approximately 4. Since doubling \(M\) halves \(\Delta S\), and the error decreases by a factor of 4, we have Error \(\propto (\Delta S)^2 \propto M^{-2}\), confirming second-order convergence.

If the error were \(O((\Delta S)^{3/2})\) instead, then halving \(\Delta S\) would reduce the error by a factor of \(2^{3/2} = 2\sqrt{2} \approx 2.83\). The expected ratios would be approximately 2.83 rather than 4.


Exercise 5. For the Black-Scholes PDE with \(\sigma = 0.25\) and \(S_{\max} = 400\), compute the CFL restriction on \(\Delta\tau\) for the explicit scheme when \(\Delta S = 2\). How many time steps are needed for \(T = 1\) year? Repeat the calculation in log-price coordinates with \(\Delta x = 0.02\) and compare.

Solution to Exercise 5

In original coordinates with \(\sigma = 0.25\), \(S_{\max} = 400\), and \(\Delta S = 2\):

\[ \Delta\tau \leq \frac{(\Delta S)^2}{\sigma^2 S_{\max}^2} = \frac{(2)^2}{(0.25)^2 \times (400)^2} = \frac{4}{0.0625 \times 160000} = \frac{4}{10000} = 4 \times 10^{-4} \]

For \(T = 1\) year:

\[ N \geq \frac{T}{\Delta\tau} = \frac{1}{4 \times 10^{-4}} = 2{,}500 \text{ time steps} \]

In log-price coordinates with \(\Delta x = 0.02\):

\[ \Delta\tau \leq \frac{(\Delta x)^2}{\sigma^2} = \frac{(0.02)^2}{(0.25)^2} = \frac{4 \times 10^{-4}}{0.0625} = 6.4 \times 10^{-3} \]

For \(T = 1\) year:

\[ N \geq \frac{1}{6.4 \times 10^{-3}} \approx 157 \text{ time steps} \]

The log-price formulation requires roughly 16 times fewer time steps (157 vs. 2,500), because the CFL restriction is independent of \(S_{\max}\).


Exercise 6. The payoff \((S - K)^+\) has a kink at \(S = K\). Explain, using the concept of local truncation error, why the convergence rate of a second-order scheme may degrade near \(S = K\). Describe two remedies and state how each restores the expected convergence order.

Solution to Exercise 6

The local truncation error (LTE) of a second-order finite difference scheme is derived via Taylor expansion of the exact solution. For the central difference approximation of \(u_{SS}\):

\[ \frac{u(S+\Delta S) - 2u(S) + u(S-\Delta S)}{(\Delta S)^2} = u_{SS} + \frac{(\Delta S)^2}{12}u_{SSSS} + O((\Delta S)^4) \]

This expansion requires \(u\) to have at least four continuous derivatives. At \(S = K\), the payoff \((S-K)^+\) has a kink: \(u_S\) is discontinuous (it jumps from 0 to 1 for a call), and \(u_{SS}\) contains a Dirac delta. The higher derivatives \(u_{SSS}\), \(u_{SSSS}\) do not exist in the classical sense.

Consequently, the Taylor expansion underlying the LTE analysis is invalid near \(S = K\). The actual truncation error at nodes near the kink is \(O(1)\) rather than \(O((\Delta S)^2)\), and the global error degrades from \(O((\Delta S)^2)\) to \(O((\Delta S)^{1/2})\) or \(O(\Delta S)\) depending on how close grid points are to the kink.

Remedy 1: Rannacher smoothing. Perform 2-4 fully implicit (backward Euler) time steps near maturity before switching to Crank-Nicolson. The implicit scheme has strong damping that smooths out the kink in 1-2 time steps, after which the solution is \(C^\infty\) and Crank-Nicolson achieves its full second-order rate.

Remedy 2: Payoff smoothing. Replace the non-smooth payoff with a smooth approximation over a small interval \([K - \epsilon, K + \epsilon]\), for example using a polynomial or Gaussian kernel. With an appropriate choice of \(\epsilon = O(\Delta S)\), this restores the full second-order convergence rate.


Exercise 7. Suppose a numerical scheme has iteration matrix \(B\) with spectral radius \(\rho(B) = 1 + 2\Delta\tau\). Is this scheme stable in the sense required by the Lax Equivalence Theorem? Justify your answer by relating \(\rho(B)\) to the bound \(\|B^n\| \leq C\) for \(n\Delta\tau \leq T\).

Solution to Exercise 7

Yes, the scheme is stable in the Lax sense. The Lax stability condition requires \(\|B^n\| \leq C\) for all \(n\Delta\tau \leq T\), where \(C\) is independent of the mesh parameters.

With \(\rho(B) = 1 + 2\Delta\tau\), the spectral radius satisfies \(\rho(B) \leq 1 + C'\Delta\tau\) with \(C' = 2\). For a normal matrix (or using the spectral radius as a bound on the matrix norm), we have:

\[ \|B^n\| \leq \rho(B)^n = (1 + 2\Delta\tau)^n \]

Since \(n\Delta\tau \leq T\), we have \(n \leq T / \Delta\tau\), and:

\[ (1 + 2\Delta\tau)^n \leq (1 + 2\Delta\tau)^{T/\Delta\tau} \leq e^{2T} \]

where the last inequality uses \((1 + a/m)^m \leq e^a\). Since \(e^{2T}\) is a finite constant depending only on \(T\) (not on \(\Delta\tau\) or \(n\)), we have \(\|B^n\| \leq e^{2T} = C\) for all \(n\Delta\tau \leq T\).

This is precisely the Lax stability condition. A spectral radius of \(1 + O(\Delta\tau)\) is permissible — it is only \(\rho(B) > 1 + O(1)\) (growth independent of mesh refinement) that causes instability.