ODE, Boundary-Value, and Spectral Computation

Relativistic calculations repeatedly reduce to ordinary differential equations: worldline evolution, radial boundary problems, elliptic reductions, and complex resonance equations. These problems may share notation while demanding very different algorithms. This chapter organizes the choice around directionality, stiffness, boundary sensitivity, regularity, spectral conditioning, and the evidence needed to support claimed digits.

Helpful background. Formulations of General Relativity identifies the continuum systems and conventions; Ordinary Differential Equations supplies the essential initial- and boundary-value theory; and Partial Differential Equations and Green Functions supplies operator domains, separated boundaries, and radiative conditions.

Choose the problem before the solver

An initial-value problem supplies all data at one point and asks for propagation. A two-point boundary problem distributes conditions across the interval. An eigenvalue or resonance problem adds an unknown spectral parameter and often non-self-adjoint boundary behavior. Treating all three as “integrate an ODE” hides the main numerical difficulty.

The recurring evidence tuple is: exact equations and domain; boundary or initial data; solver and arithmetic environment; resolution, tolerance, and precision sequence; solution and differential residuals; condition estimate; failure criteria; and an independent comparison. Boundary conditioning is developed in Ascher, Mattheij, and Russell 1995, Chapter 3 § 3.4, pp. 110–126, while collocation and its implementation are treated in Chapter 5 § 5.4, pp. 218–225, and §§ 5.6.2–5.6.4, pp. 247–265. The page sequence below turns those ideas into concrete diagnostics.

The six-page route

1. Initial-Value Integration and Stiffness distinguishes local error control from stability restrictions using an exact stiff problem and Robertson kinetics; the relevant stiff stability and implementation theory is developed in Hairer and Wanner 1996, Chapter IV §§ 1–3, pp. 2–55.
2. Shooting for Boundary-Value Problems turns missing initial data into a root and shows how unstable initial-value propagation can spoil a well-conditioned boundary problem.
3. Collocation and Relaxation Methods enforces equations and boundaries globally, using the two branches of Bratu’s problem to test nonlinear convergence.
4. Chebyshev and Multi-Domain Spectral Methods connects rapid convergence to regularity, mappings, and complete interface conditions.
5. Generalized Eigenvalue and Resonance Problems constructs matrix pencils and filters discretization modes before approaching black-hole resonances.
6. Conditioning, Residuals, and Convergence explains why a small backward residual need not imply a small forward error and assembles the final evidence hierarchy.

The ordering is intentional. Adaptive IVP machinery supports shooting; shooting motivates global collocation; collocation leads naturally to spectral bases; spectral matrices produce generalized eigenproblems; and nonnormal eigenproblems make conditioning unavoidable.

Benchmarks that expose different failures

No single benchmark covers the chapter. The exact stiff cosine problem separates stability from accuracy. The exponential boundary problem exposes unstable shooting directions. Bratu’s equation tests branch selection and mesh adaptation. The nonsmooth function $x\lvert x\rvert$ distinguishes single-domain algebraic convergence from an exact multi-domain representation. The Dirichlet Laplacian supplies known eigenvalues, while the Schwarzschild gravitational fundamental frequency tests a complex resonance workflow without rederiving its continuum theory.

Spectral approximation and multi-domain interfaces are reviewed in Grandclément and Novak 2009, §§ 2.1–2.6 and § 4. Their central computational message is that fast convergence depends on representing the domain and regularity correctly. Adding order to a poorly placed domain boundary is not a substitute for modeling the singularity or interface.

Each benchmark pairs a positive expectation with a negative test. A boundary residual may be small while the interior is wrong; a collocation residual may vanish at its own nodes but not between them; a matrix eigenpair may have machine-size residual while drifting with precision. The negative test determines whether the diagnostic actually discriminates the failure it is meant to find.

Reading a convergence table

Change one control at a time. A mesh or polynomial-order sequence tests discretization. A tolerance sequence tests adaptive termination. A precision sequence tests roundoff and conditioning. Endpoint and mapping changes test domain truncation. A different method tests implementation dependence. At least three points are needed before an observed order is meaningful, and no order should be fitted through an error plateau.

Report stable digits, not every digit printed by an eigensolver. Preserve solver versions, scaling, norms, and raw residual definitions so another implementation can reproduce the comparison. If two candidate branches approach one another, match eigenfunctions or invariant subspaces rather than sorting only by eigenvalue.

Chapter review

A spectral resonance calculation is stable from $N=60$ to $N=80$ at double precision, but changes at $N=80$ when rerun with 100-digit arithmetic. Which next tests separate discretization error from conditioning?

Solution

Continue a two-dimensional study: raise precision at fixed $N$ until the result stabilizes, then raise $N$ at that adequate precision. Inspect matrix and differential residuals, coefficient decay, and eigenvalue conditioning. Vary domain maps and endpoints, and compare an independent shooting, continued-fraction, or time-domain calculation. Fixed double-precision agreement alone may be a shared roundoff plateau.

References

• Ascher, U. M., Mattheij, R. M. M., and Russell, R. D. (1995), Numerical Solution of Boundary Value Problems for Ordinary Differential Equations, SIAM. DOI.
• Grandclément, P., and Novak, J. (2009), “Spectral Methods for Numerical Relativity,” Living Reviews in Relativity 12, 1. DOI.
• Hairer, E., and Wanner, G. (1996), Solving Ordinary Differential Equations II: Stiff and Differential-Algebraic Problems, 2nd revised edition, Springer. DOI.