Computational Relativity

Computational relativity turns covariant field equations into finite, testable numerical claims. Between the continuum problem and a scientific conclusion lie choices of variables, gauge, coordinates, boundary conditions, discretization, solver, precision, mesh, hardware, diagnostics, and interpretation. This volume develops that entire chain—from symbolic tensor algebra and elementary wave tests through black-hole evolution, relativistic matter, radiation extraction, synthetic observations, reduced models, and independent computational evidence.

Helpful background. Formulations of General Relativity supplies the continuum decompositions, gauge systems, constraints, and hyperbolicity used here; Differential Equations for Relativity supplies initial- and boundary-value problems, stability, and convergence.

Start with one wave you can solve

If this is your first numerical-relativity course, begin with a wave rather than a binary black hole. You need derivatives, vectors and matrices, elementary ODEs, and enough programming to store an array and repeat an update. The first calculation uses $u(t,x)=\sin(2(x-t))$: the exact answer is known, so every numerical discrepancy can be measured directly.

Follow this short route before choosing a specialization:

1. Set up and run the scalar-wave experiment. Identify the equation, initial data, boundary, grid and reported error.
2. Derive the finite-difference operator. Understand why halving the spacing can reduce a fourth-order error by about sixteen.
3. Separate space and time discretization. Check that a more accurate derivative still needs a stable time integrator.
4. Interpret a convergence sequence. Distinguish the error in the computed field from a small equation residual.

You do not need to master every verification technique before starting. Run this one example, predict what a refinement should change, and compare the prediction with the result. The later chapters explain what must be added when geometry, constraints, shocks, boundaries and uncertain physical models enter.

From a continuum problem to a finite claim

Write a relativistic problem schematically as

$$\mathcal F[u;\mu]=0, \qquad \mathcal B[u]=g, \qquad Q=\mathcal O[u],$$

where $u$ denotes fields, $\mu$ physical parameters, $\mathcal B$ initial and boundary data, and $Q$ an observable. A computation replaces this problem by finite objects,

$$F_{h,\tau,u_{\rm fp}}(U)=0, \qquad Q_{h,\tau,u_{\rm fp}}=O_h(U),$$

with resolution $h$, solver tolerance $\tau$, and floating-point scale $u_{\rm fp}$. The discrete residual can be tiny while $Q_h$ remains wrong because the mesh is coarse, the boundary is inappropriate, the continuum formulation is mis-transcribed, or the observable uses an inconsistent tetrad or normalization.

A defensible result is therefore a tuple:

$$(\text{problem},\text{method},\text{environment}, \text{refinement},\text{checks},\text{uncertainty}, \text{domain}).$$

It records equations, physical data, conventions, algorithm, precision, software and hardware, resolution or tolerance sequence, residuals and invariants, independent benchmark, uncertainty components, and the regime in which the conclusion applies. A bare number, animation, or source repository is not this object.

The volume uses the site’s metric signature $(-,+,+,+)$, explicit $G$ unless a sustained geometrized-unit switch is declared, and Fourier dependence $e^{-i\omega t}$. Pages state every additional coordinate, gauge, hypersurface, tetrad, harmonic, flux, frame, polarization, and data-layout convention before use. When a source uses different signs, the complete convention package is translated; changing one equation in isolation is unsafe.

Baumgarte and Shapiro develop the connection from $3+1$ equations to numerical initial data and evolution in Baumgarte and Shapiro 2010, Chapters 2 and 8–11. This volume assumes the continuum formulations developed in the referenced theory volumes and concentrates on how to implement, test, compare, and interpret them.

Four recurring computational laboratories

The same examples return across chapters so that methods can be compared rather than hidden behind changing physics.

A scalar wave. An exact wave on flat or Schwarzschild spacetime tests symbolic reduction, ODE and PDE discretization, time integration, boundaries, convergence, mesh interfaces, quasinormal ringing, tails, and extraction. Because an analytic or independently converged answer exists, it exposes implementation error cleanly.

A single black hole. Schwarzschild or Kerr tests coordinate transformations, curvature invariants, constraint-satisfying data, gauge evolution, excision or punctures, horizon finding, null radiation, geodesics, tetrads, and synthetic images. The geometry is nontrivial while many invariants and charges are known.

A relativistic shock tube. A Riemann problem tests conservative fluxes, reconstruction, primitive-variable recovery, positivity, causality, equations of state, divergence control, atmosphere treatment, and refinement boundaries. Smooth-wave order is not expected at a discontinuity; conservation and physically admissible states become central.

A compact-binary waveform family. Versioned high-fidelity waveforms connect initial data, evolution, mesh design, horizon and flux budgets, finite-radius extraction, reduced bases, empirical interpolation, Gaussian processes, neural surrogates, out-of-domain warnings, and uncertainty propagation.

Each prose benchmark gives equations, inputs, an algorithmic sequence, expected convergence, failure criteria, and an independent check sufficient to evaluate a conforming implementation. A linked executable or data product, when provided, has its own version and requires independent checks; prose alone does not establish that a calculation has been run successfully.

The thirteen-chapter route

The volume proceeds in the order a trustworthy calculation is built.

1. Computational Experiments and Reproducible Workflows turns a physical question into a finite claim with versioned inputs, environments, data contracts, provenance, deterministic or statistical rerun rules, and preservation.
2. Symbolic Tensor and Algebraic Computation controls dummy indices, tensor symmetries, coordinate domains, tetrads, variational boundary terms, curvature invariants, branches, and independent component checks.
3. ODE, Boundary-Value, and Spectral Computation treats stiffness, shooting, collocation, multidomain Chebyshev methods, generalized eigenvalue pencils, conditioning, and precision-aware mode filtering.
4. Discretizing Relativistic PDEs compares finite differences, finite volumes, finite elements, discontinuous Galerkin methods, time integration, CFL limits, and numerical boundaries.
5. Computational Black-Hole Perturbation Theory implements separated boundary problems, time-domain master equations, continued fractions, spectral resonances, branch cuts, and point sources without rederiving the continuum perturbation theory.
6. Numerical Initial Data and Equilibria solves elliptic constraints for black holes, binaries, and stars and distinguishes solver convergence from physical quality of the free data.
7. Spacetime Evolution, Gauge, and Boundaries compares BSSN, generalized-harmonic, and Z4-type evolutions together with damping, gauge drivers, outer boundaries, excision, and moving punctures.
8. Meshes, Adaptivity, and High-Performance Computing develops coordinates, mapped grids, refinement, multidomain decomposition, task parallelism, accelerators, mixed precision, and performance at a matched error target.
9. Relativistic Hydrodynamics, GRMHD, and Radiation implements conservative matter evolution, primitive recovery, microphysical tables, magnetic-divergence control, transport approximations, and atmosphere or vacuum interfaces.
10. Horizons, Charges, and Gravitational-Wave Extraction locates marginal surfaces, computes quasi-local quantities, extracts curvature radiation, extrapolates finite-radius waveforms, performs characteristic extraction, and closes flux and memory balances.
11. Ray Tracing, Transport, and Synthetic Observations maps camera pixels to null data, integrates geodesics and covariant transfer, handles emission and polarization, and separates numerical agreement from plasma and instrument systematics.
12. Reduced Models, Surrogates, and Machine Learning builds reduced bases, empirical operators, Gaussian-process and neural emulators, physics constraints, calibration, and out-of-distribution failure tests.
13. Verification, Validation, Uncertainty, and Computational Evidence brings these methods together through code verification, solution verification, manufactured solutions, independent comparisons, uncertainty propagation, and reproducible results.

The route is cumulative but not rigid. Readers can enter through a problem-specific path once they understand how to check a calculation and state its uncertainty and limits.

Five ways to read the volume

Graduate numerical foundations. Read chapters 1–4 and 13. This path starts with a finite claim, implements symbolic and discrete operators, couples space and time stability, and ends by distinguishing verification, validation, and uncertainty.

Numerical-relativity production. Read chapters 4 and 6–10, then 13. It follows constraint-satisfying data through free evolution, multiscale meshes, matter or vacuum dynamics, horizons, radiation, and global balance checks.

Perturbations and waveforms. Read chapters 3–5, 10, and 12–13. Continued fractions, collocation, time evolution, extraction, and surrogates are compared through residual, resolution, precision, and independent-method tests.

Matter and synthetic observations. Read chapters 4 and 8–11, then 13. This route connects conservative GRMHD and radiation transport to geodesic cameras and instrument-facing products while keeping equation-of-state, emission, calibration, and numerical systematics separate.

Symbolic and reduced computation. Read chapters 1–3 and 12–13. It emphasizes canonical algebra, domain and branch assumptions, compact bases, learned parameter maps, failure detection, and reconstructible evidence.

A binary result through the full chain

Consider a claim about the phase of the dominant gravitational-wave mode from a binary black-hole evolution. The initial-data chapter requires converged Hamiltonian and momentum constraints, stable ADM quantities, horizon masses and spins, and a record of free-data choices. The evolution chapter fixes formulation, gauge, damping, boundary treatment, puncture or excision method, and monitored constraint norms. The mesh chapter shows that resolution, refinement boundaries, domain maps, and parallel decomposition do not change the phase beyond tolerance.

Extraction then declares the null tetrad, spin-weighted harmonics, retarded time, extraction radii, extrapolation order or characteristic worldtube, and integration constants. A credible phase uncertainty includes resolution, finite-radius or characteristic-extraction, alignment, and reference-time components with their dependence retained. Horizon and radiated energy and angular momentum should close against initial charges within the combined budget. Bishop and Rezzolla survey curvature extraction, finite-radius effects, characteristic methods, and their checks in Bishop and Rezzolla 2016, §§ 3, 6, and 8.

A surrogate may then compress the versioned waveform family, but it cannot erase upstream uncertainty. Holdout and boundary mismatches, basis and interpolation error, predictive calibration, symmetry residuals, and out-of-domain warnings are added to the record. A fast prediction is scientifically usable only if its total error remains below the downstream requirement over the validated parameter domain.

The environment for every stage includes source revision, operating system or immutable image, compiler or interpreter, numerical libraries, precision, hardware, process and thread counts, workflow, inputs, and checksums. A clean rerun regenerates raw products and comparison tables, not just the final figure.

What each computational check establishes

Evidence layer	Central question	Typical check	What it does not establish
symbolic identity	was the intended expression manipulated correctly?	independent components, special cases, symmetry and branch tests	numerical stability of an evolution
code verification	does the implementation correctly approximate the intended mathematical equations?	unit tests, analytic benchmarks, manufactured solutions, mutations	accuracy for one production run
solution verification	how large is numerical error in this calculation?	residual, grid/tolerance/precision/boundary refinement, conservation	adequacy of the physical model
validation	does the model represent the target physical system for its use?	versioned experimental or observational comparisons appropriate to the intended use	universal truth outside the tested regime
reproducibility	can the finite result be reconstructed?	clean environment, checksums, workflow, independent rerun	correctness merely because it repeats
uncertainty propagation	how do declared input and error components affect the observable?	convergence of sampling or quadrature and held-out coverage	missing model discrepancy unless included

Oberkampf and Roy separate code verification, solution verification, validation, and uncertainty in Oberkampf and Roy 2010, Chapters 2–4 and 8–12. The National Academies likewise distinguishes computational reproducibility from independently obtained agreement in National Academies 2019, Summary pp. 6–8 and Chapter 3. The terminology matters because the checks answer different questions.

Scientific boundaries and failure modes

Continuum formulations, black-hole geometry, perturbative spectra, matter microphysics, waveform interpretation, and observational likelihoods remain with their dedicated volumes. Implementing an equation here does not transfer its theoretical or phenomenological conclusions. This volume asks whether the accepted system was discretized, solved, checked, and reported correctly.

The most common failure modes cross chapter boundaries: inconsistent signature or tensor layout; a residual evaluated with the same faulty operator; two resolutions mistaken for convergence; a stable interior scheme coupled to an unstable boundary; a constraint floor hidden by a prettier observable; primitive recovery that silently clips unphysical states; a waveform extracted with an undeclared tetrad; a ray-tracing comparison with different camera or emissivity inputs; a surrogate validated by leaked data; and performance quoted without a matched accuracy target.

Frontier and observational assessments are dated through 2026-08-18 where required. Mixed precision, radiation transport, learned surrogates, physics-informed methods, out-of-distribution uncertainty, computational uncertainty propagation, and synthetic-observation interfaces remain conditional on their tested algorithms, data, hardware, closures, and calibration. A new benchmark, correction, dataset, or failure can change that assessment without invalidating the stable numerical principles around it.

Review the volume

A binary-black-hole project reports a smooth waveform and 50-fold accelerator speedup. It uses two grid resolutions, a mixed-precision evolution, a finite outer boundary, one extraction radius, a neural surrogate trained on all available waveforms before splitting, and a ray-traced image with an unspecified camera tetrad and electron prescription. Constraints are small in the evolution code’s own norm, and a clean rerun reproduces the final plots. What can be concluded, and what is the minimum repair?

Solution

The clean rerun establishes reconstruction of one pipeline, not numerical or physical accuracy. Two resolutions do not robustly determine an observed order; add at least a third, refine time and solver tolerances, and evaluate continuum constraints through an independent operator. Compare all-high-precision and mixed-precision runs at the same resolutions, reporting waveform phase, horizon and flux budgets, constraint norms, hardware, and total resources at the accepted error target.

Move the outer boundary or use a tested constraint-preserving or absorbing treatment, and show that the claimed time window is causally uncontaminated. Extract on several radii, vary extrapolation order and radius window, or compare with characteristic extraction. Declare the tetrad, harmonics, retarded time, strain integration, and sign convention, then close initial, radiated, and final energy and angular momentum within a combined uncertainty.

Rebuild the surrogate after splitting by physical parameter point; reserve boundary and contiguous-hole tests, report median and worst mismatch, and test out-of-domain refusal. For the image, declare camera location and orthonormal tetrad, null normalization, fluid snapshot and interpolation, emission and electron prescriptions, transfer coefficients, instrument transformation, and calibration version. Separate numerical ray/transfer convergence from plasma-model and observational systematics.

Only after these repairs can the project claim a finite waveform or image accuracy over a stated domain. The accelerator speedup is meaningful only at that matched accuracy; neither smooth plots nor reproducibility alone establish the continuum or observational result.

References

• Baumgarte, Thomas W., and Stuart L. Shapiro. Numerical Relativity: Solving Einstein’s Equations on the Computer. Cambridge: Cambridge University Press, 2010. DOI.
• Bishop, Nigel T., and Luciano Rezzolla. “Extraction of Gravitational Waves in Numerical Relativity.” Living Reviews in Relativity 19 (2016): 2. DOI.
• National Academies of Sciences, Engineering, and Medicine. Reproducibility and Replicability in Science. Washington, DC: National Academies Press, 2019. DOI.
• Oberkampf, William L., and Christopher J. Roy. Verification and Validation in Scientific Computing. Cambridge: Cambridge University Press, 2010. DOI.