Numerical Analysis, Verification, and Validation

A numerical result supports a gravity claim only when the continuum problem, discrete problem, computed solution, and physical comparison are kept distinct. This record maps the evidence needed at each layer: implementation checks, residuals and conditioning, convergence, constraints and invariants, benchmarks, uncertainty, and a rerunnable record. It does not certify a particular code or computed value.

Helpful background. Differential Equations for Relativity supplies stability and convergence language; PDEs and Green Functions for Gravitation supplies domains and boundary data; Computational Relativity supplies the gravity-specific algorithms.

Evidence chain for a numerical result

Question	Evidence	What it does not establish
Was the intended equation implemented?	Exact identities, unit and operator tests, manufactured solutions, and deliberate fault injection	Accuracy of one production solution
Was the discrete problem solved?	Algebraic or iterative residuals below a declared tolerance	Closeness to the continuum solution
Does the discretization approach the continuum problem?	Multi-resolution or multi-tolerance convergence, constraints, invariants, and an independent benchmark	Agreement with nature
Is the physical model adequate for its intended use?	Comparison with versioned observations or experiments, including calibration and model discrepancy	A theorem outside the tested regime
Can the finite claim be challenged?	Pinned equations, inputs, environment, raw outputs, analysis, and rejection criteria	Independence merely because files are available

Verification asks whether the equations were solved correctly; validation asks whether the verified model represents the relevant physical system adequately for a declared use. Uncertainty quantification propagates identified numerical, input, calibration, stochastic, and model components without erasing their different meanings. This separation follows Roy and Oberkampf 2011, “Uncertainty Framework” and Conclusions, pp. 2131–2144.

Residual, conditioning, and error

For a nonsingular discrete linear problem $A_hu_h^*=b_h$ with $b_h\ne0$, let $u_h$ be a computed approximation and define

$$r_h=b_h-A_hu_h, \qquad \kappa(A_h)=\lVert A_h\rVert\,\lVert A_h^{-1}\rVert.$$

For any consistent induced norm, the relative forward error obeys

$$\frac{\lVert u_h-u_h^*\rVert}{\lVert u_h^*\rVert} \le \kappa(A_h) \frac{\lVert r_h\rVert}{\lVert b_h\rVert}.$$

Thus an appropriately scaled residual is a backward-error indicator only relative to a declared perturbation model; it is not automatically a small forward error. Singular systems, nonlinear problems, nonnormal evolution, and poorly scaled variables need their own sensitivity analysis. The norm, scaling, and perturbation model are part of every condition number; see Higham 2002, Ch. 7, Thms. 7.1–7.2, pp. 119–137.

The discrete residual also differs from the continuum residual. Evaluate the unfactored continuum equations, constraints, and boundary conditions on a representation that was not used to define the solver’s collocation equations whenever possible.

Convergence and extrapolation

Let $Q_1,Q_2,Q_3$ be the same scalar quantity on successively coarser grids with equal refinement ratio $r\gt1$. In a monotone single-power regime, the observed order is

$$p=\frac{\ln\left\lvert(Q_3-Q_2)/(Q_2-Q_1)\right\rvert}{\ln r}.$$

This estimate is conditional. The three runs must represent identical continuum data; iterative error, roundoff, interpolation, and sampling noise must be smaller than grid differences; and a finer run should confirm the predicted change. Unequal ratios require the corresponding implicit relation. Oscillatory convergence, shocks, moving mesh interfaces, cancellation, or mixed orders can invalidate one fitted $p$. The reporting procedure and its assumptions are stated in Celik et al. 2008, recommended Steps 1–5 and Eqs. (1)–(10), pp. 078001-1–078001-4. A Richardson estimate or grid-convergence index is not a statistical confidence interval.

Gravity-specific checks

A strong calculation combines several diagnostics:

• evolution constraints and gauge conditions converge in declared norms;
• mass, angular momentum, charge, flux, or symplectic balances close within their boundary terms;
• extracted observables are stable under resolution, outer-boundary placement, gauge or frame choice, and extraction procedure;
• a smooth exact or manufactured problem recovers formal order, while a lower-regularity benchmark tests the production regime; and
• an independently formulated calculation agrees after both results are extrapolated to a common physical quantity.

The robust-stability, linear-wave, gauge-wave, shifted-gauge-wave, and Gowdy tests demonstrate why formulation, constraints, solution error, and runtime behavior must be inspected separately Babiuc et al. 2008, §§ 3–8. Passing one test cannot validate an astrophysical model.

The full gravity treatment is Verification, Validation, Uncertainty, and Computational Evidence. A defensible result records equations and conventions, domain and data, code and dependency revisions, precision and tolerances, every resolution, raw diagnostics, analysis scripts, expected convergence, independent checks, and explicit failure criteria. Performance rankings or numerical physics conclusions require a dated, reproducible calculation rather than an evergreen method summary.

References

• Babiuc, Maria C., et al. 2008. “Implementation of Standard Testbeds for Numerical Relativity.” Classical and Quantum Gravity 25, 125012. Open preprint.
• Celik, Ismail B., et al. 2008. “Procedure for Estimation and Reporting of Uncertainty Due to Discretization in CFD Applications.” Journal of Fluids Engineering 130, 078001. DOI.
• Higham, Nicholas J. 2002. Accuracy and Stability of Numerical Algorithms, 2nd ed. Society for Industrial and Applied Mathematics. DOI.
• Roy, Christopher J., and William L. Oberkampf. 2011. “A Comprehensive Framework for Verification, Validation, and Uncertainty Quantification in Scientific Computing.” Computer Methods in Applied Mechanics and Engineering 200, 2131–2144. DOI.