Numerical Covariant Radiative Transfer

Numerical covariant transfer evolves radiation along an already verified null ray while converting local emission and absorption coefficients into observer-frame intensity. The invariant formulation prevents artificial frame factors, but accuracy still depends on fluid interpolation, optical-depth stepping, frequency sampling, and stopping rules. A resolved geodesic is therefore necessary but not sufficient for a resolved image or spectrum.

Required background. Affine and Variational Geodesics supplies null trajectories and affine parametrization.

Helpful background. Formulations of General Relativity supplies frames and numerical spacetime data; Differential Equations for Relativity supplies stiff integration and interpolation; Black-Hole Astrophysics supplies emission and material models; and Computational Relativity supplies the common numerical-evidence standard.

Invariant transfer along a ray

Use the future-directed frequency-normalized ray tangent of Observer Cameras and Local Tetrads, with $dx^\mu/d\lambda=p^\mu$ and cyclic frequency $\nu=-p_\mu u^\mu\gt0$ measured in the local material frame. Physical photon momentum is $2\pi p^\mu$ when $\hbar=1$; fixing this normalization prevents an unnoticed factor of $2\pi$ in the affine transfer equation. The invariant intensity, emissivity, and absorption coefficient are

$$\mathcal I=\frac{I_\nu}{\nu^3}, \qquad \mathcal J=\frac{j_\nu}{\nu^2}, \qquad \mathcal A=\nu\alpha_\nu.$$

With a consistently normalized affine parameter, the unpolarized transfer equation can be written

$$\frac{d\mathcal I}{d\lambda} =\mathcal J-\mathcal A\mathcal I.$$

These combinations encode Liouville invariance in vacuum: when $\mathcal J=\mathcal A=0$, $\mathcal I$ is constant even as gravitational or Doppler shift changes $\nu$. Younsi and collaborators derive the invariant coefficients and optical-depth form in Younsi et al. 2012, §2 and §§2.1–2.2, Eqs. (1)–(18).

The equation assumes geometric optics and prescribed local coefficients. Elastic or inelastic scattering that couples directions or frequencies requires an angular/frequency closure or a Monte Carlo treatment; it cannot be represented by silently adding a scalar loss term unless re-emission is modeled consistently. Time-dependent media also require a declared interpolation between snapshots at the spacetime event reached by each ray.

Why the powers of frequency differ

Along a future-directed ray, the infinitesimal path length measured in the fluid frame is $d\ell=-u_\mu dx^\mu=\nu\,d\lambda$. Local emission adds $j_\nu d\ell$ to $I_\nu$. Dividing by $\nu^3$ therefore gives $(j_\nu/\nu^2)d\lambda$, while local absorption removes $\alpha_\nu I_\nu d\ell/\nu^3 =(\nu\alpha_\nu)\mathcal I\,d\lambda$. Vacuum frequency changes are already accounted for by the invariant $\mathcal I$. This derives both powers in the affine equation.

Backward ray tracing finds the path; radiative transfer still needs correct boundary data. One option stores the path and integrates forward from the far boundary. Alternatively, accumulate positive optical depth from the camera into the scene and sum attenuated emission. For a near cell with optical depth $\Delta\tau$ and invariant source $\mathcal S=\mathcal J/\mathcal A$, radiation incident from its far side emerges toward the camera as

$$\mathcal I_{\mathrm{near}} =\mathcal I_{\mathrm{far}}e^{-\Delta\tau} +\mathcal S(1-e^{-\Delta\tau}).$$

Starting with zero intensity at the camera and simply integrating the forward equation backward imposes the wrong boundary condition.

In vacuum, a source intensity is observed as $I_{\nu,\mathrm{obs}}=g_{\mathrm{red}}^3 I_{\nu,\mathrm{em}}(\nu_{\mathrm{obs}}/g_{\mathrm{red}})$. The emitted frequency must be shifted along with the amplitude; multiplying an unchanged frequency-bin value by $g_{\mathrm{red}}^3$ is generally wrong.

Optical depth and stable stepping

For a homogeneous slab of length $L$ in its rest frame, constant $j_\nu$ and $\alpha_\nu>0$ give

$$I_\nu(L)=I_\nu(0)e^{-\tau} +S_\nu\left(1-e^{-\tau}\right), \qquad \tau=\alpha_\nu L, \qquad S_\nu=\frac{j_\nu}{\alpha_\nu}.$$

This exact solution is the first benchmark. It tests sign, path orientation, thin-limit cancellation, and thick-limit saturation. For $\tau\ll1$, evaluate $1-e^{-\tau}$ with a numerically stable exponential-minus-one operation; direct subtraction loses relative accuracy. For $\tau\gg1$, explicit stepping can be stiff and waste effort resolving intensity that has already relaxed to the source function.

An optical-depth integrator may evolve $\tau$ and use a formal solution over each segment. Its segment length must resolve variations in $u^\mu$, $j_\nu$, $\alpha_\nu$, and metric, not merely the geodesic curvature. Limit steps at fluid cell boundaries or account for the interpolation polynomial analytically. Negative interpolated density or temperature can produce unphysical negative opacity; positivity preservation and out-of-domain handling belong in the transfer algorithm.

Vincent and collaborators provide one optically thin numerical formulation in Vincent et al. 2011, §2.2.1, Eqs. (9)–(15). Optically thin integration is a useful limit, not a universal approximation: it breaks down when absorption or scattering makes optical depth order unity.

Benchmarks beyond one frame

Run the homogeneous slab at $\tau=10^{-4}$, $1$, and $10^2$ over at least three transfer tolerances or step sizes. Compare the absolute error in the thin case and the relative departure from $S_\nu$ in the thick case. Boost the entire slab and camera, transform frequency and coefficients, and verify that reconstructed $I_\nu/\nu^3$ is invariant. This frame test can expose a missing power of redshift that resolution refinement alone will not find.

Next integrate through an analytic Schwarzschild or Kerr ray with a prescribed smooth coefficient field. Compare direct affine stepping with optical-depth stepping and reverse the ray while transforming boundary data consistently. For an external benchmark dataset, pin the exact analytic model and software revision. The analytic and shared-snapshot comparison problems in Prather et al. 2023, §§3.1–3.3 and §§4.1–4.4 can supply such inputs, but their full Stokes tests exceed the scalar scope of this page.

In a numerical fluid snapshot, refine six controls independently: geodesic tolerance, transfer step, spatial and temporal fluid interpolation, image pixels, and frequency grid. Image resolution cannot diagnose under-resolved optical depth along each ray. Compare integrated spectra as well as pixel norms, because local positive and negative errors can cancel differently.

Evidence and reproduction

The environment record should pin metric and fluid checksums, coordinate chart, camera tetrad, ray and transfer integrators, coefficient prescription, interpolation libraries, precision, frequency and image grids, and software revision. Inputs include units, affine normalization, boundary intensity, coefficient frames, scattering approximation, snapshot times, masks, and termination conditions. Outputs should retain geodesic residuals, optical-depth histories for diagnostic rays, images, spectra, and refinement differences.

A clean future rerun would regenerate the thin/intermediate/thick slab, a boosted copy, and a smooth curved-spacetime test. It would sweep transfer steps and interpolation resolutions, report analytic errors and observed convergence, and emit machine-readable configurations and checksums. This page specifies the sequence; it does not claim an absent executable has run it.

Reject the result if vacuum changes $\mathcal I$, the slab misses either asymptotic limit, a boosted calculation disagrees after invariant transformation, negative coefficients arise without a physical model, or image changes fail to stabilize under separate ray, transfer, fluid, pixel, and frequency refinement. Optical-depth stiffness and interpolation across sharp fluid features are distinct failure modes and should not be hidden inside one tolerance.

Check your understanding

An image is unchanged when pixel resolution doubles, but diagnostic rays cross cells with $\Delta\tau\approx20$ per transfer step. Halving the transfer step moves the spectrum by $8\%$. Has image convergence been shown?

Solution

Only pixel-sampling convergence at one under-resolved transfer setting has been tested. A step spanning optical depth twenty is not intrinsically inaccurate: a constant-coefficient formal solution is exact for any optical depth. The observed $8\%$ change instead shows unresolved coefficient variation or an unsuitable stepping scheme. Refine or formally integrate the optical-depth segments, repeat the slab benchmark in the thick regime, and then recheck both image and spectrum while varying fluid interpolation independently.

Continue with Synchrotron and Thermal Emission Models.

References

• Prather, Cora, et al. 2023. “Comparison of Polarized Radiative Transfer Codes Used by the Event Horizon Telescope Collaboration.” The Astrophysical Journal 950, 35. arXiv. DOI.
• Vincent, Frédéric H., Thibaut Paumard, Éric Gourgoulhon, and Guy Perrin. 2011. “GYOTO: A New General Relativistic Ray-Tracing Code.” Classical and Quantum Gravity 28, 225011. arXiv.
• Younsi, Ziri, Kinwah Wu, and Sebastian V. Fuerst. 2012. “General Relativistic Radiative Transfer: Formulation and Emission from Structured Tori around Black Holes.” Astronomy & Astrophysics 545, A13. arXiv.