Image Reconstruction and Model Fitting

Image reconstruction from horizon-scale VLBI data is an underdetermined inverse problem: many brightness distributions agree with the same sparse, noisy Fourier samples. A scientifically useful feature is therefore one that survives changes of reconstruction family, regularizer, calibration treatment, data subset, and direct visibility-domain model. This page explains that stability criterion for EHT observations of M87* and Sagittarius A*. It does not identify a reconstructed central depression directly with an event horizon.

Helpful background. Black-Hole Geometry and Mechanics supplies critical-curve and lensing scales; Relativistic Matter and Compact Objects supplies radiative source models; Black-Hole Astrophysics supplies accretion variability and opacity; and Computational Relativity supplies synthetic-observation validation.

The inverse problem and its assumptions

Let $d_i$ denote a measured complex visibility or closure quantity and let $\mathcal F_i[I]$ be the same observable predicted from a trial image $I(l,m)$. A regularized maximum-likelihood reconstruction has the schematic form

$$\widehat I= \underset{I\geq0}{\operatorname{argmin}} \left[\chi^2\!\left(d,\mathcal F[I];\boldsymbol{g},\boldsymbol{\Sigma}\right) +\sum_k\alpha_k R_k[I]\right].$$

Here $\boldsymbol{g}$ are nuisance gains, $\boldsymbol{\Sigma}$ is the data covariance, $R_k$ may favor smoothness, sparsity, compactness, or entropy, and $\alpha_k$ sets the strength of each preference. Positivity, field of view, pixel size, compact-flux priors, and the treatment of total flux are assumptions too. A good fit does not erase them: in an incomplete Fourier plane, the regularizer selects among data-compatible images.

The principal method families ask related but distinct questions:

• CLEAN and related deconvolution methods fit point components before restoring them with a chosen beam;
• regularized maximum-likelihood methods, including eht-imaging and SMILI, optimize pixels with explicit penalties;
• flexible Bayesian rasters produce posterior distributions over image and calibration parameters;
• geometric models fit rings, crescents, Gaussians, or splines directly to visibilities; and
• physical libraries compare ray-traced GRMHD images to the sampled data.

Agreement is strongest when it appears in observables before it appears in a picture—for example, a common visibility-amplitude minimum or closure-phase pattern—and when methods with genuinely different parameterizations recover it.

What the 2019 M87* analysis established

The data anchor is EHT product 2019-D01-01, DOI 10.25739/g85n-f134, containing calibrated Stokes-$I$ visibility data for M87* on 2017 April 5, 6, 10, and 11. The associated pipeline product 2019-D01-02 publishes DIFMAP, eht-imaging, and SMILI parameter-survey workflows. For a reproducible snapshot, this page pins the repository at commit 80d230e76c08edd31548e7cfe61dc7cf94300780. Its README states an important calibration boundary: the public files contain Stokes $I$, whereas Paper IV used parallel-hand RR/LL data, so dependency and self-calibration details can produce small differences Event Horizon Telescope Collaboration 2019, pipeline release 2019-D01-02, README.

Paper IV separated blind imaging by four teams from a later common parameter survey. It used DIFMAP, eht-imaging, and SMILI; varied regularizer weights and compact-flux assumptions; reconstructed synthetic sources sampled like the observation; and compared features across days, bands, and pipelines. A roughly circular, asymmetric ring persisted, while small-scale details did not have the same support Event Horizon Telescope Collaboration 2019, §§ 5–9.

Paper VI then compared geometric visibility models, reconstructed-image measurements, and GRMHD image libraries. Its interpretation distinguished the observed emission-ring diameter from the metric critical curve and used external mass–distance information Event Horizon Telescope Collaboration 2019, §§ 3–6 and § 8. Convergence among these routes supports a characteristic angular scale; it does not make emissivity, optical depth, or the simulation library irrelevant.

Validation with held-out information

One should reserve information before choosing hyperparameters or declaring a feature. Useful tests include leaving out one observing day, frequency band, station, baseline family, or scan; fitting on a nonredundant closure set and evaluating the posterior predictive distribution on held-out closures; and repeating the analysis with gain priors widened. Because a station removal changes both sensitivity and Fourier coverage, a disappearing feature is a diagnostic, not automatically evidence that the excluded station was faulty.

Blind synthetic tests provide a known truth. Start with several ground truths—a ring, disk, crescent, double source, and nonring extended source—sample each at the release’s exact $(u,v)$ coordinates, corrupt them with thermal noise and station gains drawn from the declared calibration model, then apply every candidate pipeline without revealing the input label. Recovery should be scored on visibility prediction and prespecified features, not aesthetic similarity. A method that turns every ground truth into a ring fails this adversarial test even if its reconstructions look stable.

The statistical model must accompany the picture. Report the likelihood, covariance or nonredundant closure basis, image and gain priors, parameter intervals, flag and signal-to-noise selections, coherent averaging, frequency combination, and scattering or variability prescription. Thermal scatter within that model is statistical uncertainty. Sparse coverage, residual calibration, regularizer choice, time variability, scattering, and physical-library incompleteness are systematic uncertainty or model dependence.

Results and limitations through 2026-08-18

The evidence cutoff for mutable claims is 2026-08-18. “Stable” below means stable under the listed tests, not unique among all mathematically possible images.

Evidence layer	Result at the cutoff	Scope and limitation
Analytic mechanism	Sparse Fourier sampling admits nonunique images; closure quantities constrain calibration-resistant combinations.	The statement explains why priors are required but does not select a morphology.
Blind simulations	The 2019 parameter surveys recovered broad ring-scale structure across three pipelines and multiple synthetic sources.	Tested truths, noise, gain models, and hyperparameter grids do not exhaust failure cases.
Individual object: M87* 2017	Independent imaging and direct visibility fits support a persistent ring-scale feature in the DOI-bound release.	Fine azimuthal structure depends more strongly on day, method, and calibration; the public Stokes-$I$ product is not exactly the RR/LL working set.
Individual object: M87* 2018	Data product 2024-D01-01, DOI 10.25739/epm5-r371, supplies a later calibrated epoch for persistence tests.	A second epoch tests temporal robustness for one source, not universal image morphology.
Individual object: Sagittarius A*	Multiple static and dynamic imaging methods found a thick ring-scale structure despite rapid variability Event Horizon Telescope Collaboration 2022, §§ 3–7.	Scattering and intrahour evolution make time averaging and variability normalization source-specific.
Population and null evidence	Two resolved objects do not define a population distribution; nonring synthetic inputs and held-out-data failures remain essential null tests.	Selection favors the brightest, largest apparent horizons; there is no complete horizon-scale population sample.

Credible competing positions therefore concern the strength and formulation of the claim. Regularized images emphasize feature stability across priors; geometric or Bayesian fits quantify morphology in visibility space; GRMHD libraries test physical consistency; and closure-only claims remain closest to measured combinations. The evidence supports robust ring-scale structure in the named objects, but the uniqueness of detailed brightness maps and a model-independent horizon interpretation remain unresolved.

The interpretation could change after a new EHT data or calibration release, an added station, a materially different reconstruction family, a failed blind synthetic recovery, or evidence that a claimed feature vanishes under a prespecified day, band, station, or scan holdout.

Reproduction boundary

A faithful teaching reproduction downloads 2019-D01-01, preserves release flags, records exact stations, scans, bands, and averaging, and runs at least two independent algorithms from 2019-D01-02 plus a direct geometric visibility fit. It retains each method’s configuration files and dependency versions, uses identical data selections where possible, reconstructs blinded synthetic sources first, and evaluates held-out visibilities or closures. It should also repeat the result without the longest baselines and with widened gain priors.

This page has not executed those pipelines and claims no fresh fit. The recipe is deliberately pinned and testable, but differences from published figures are expected if the public Stokes-$I$ release, modern dependencies, or altered random seeds are used. Reproduction means explaining those differences, not tuning until a familiar image appears.

Common pitfalls

A common restoring beam does not make algorithms independent. Shared data, priors, initialization, or hyperparameter searches can create correlated agreement. Independence must be established in parameterization and workflow.

Visual stability is not predictive validation. A feature should improve or preserve fit quality on withheld observables and survive synthetic counterexamples.

A physical library cannot test models it omits. A strong likelihood ratio within one GRMHD library is conditional on its electron prescription, magnetic states, spins, inclinations, and temporal sampling.

Exercises

Two reconstructions fit the same data. Model A has $\chi_A^2=100$ and regularizer value $R_A=20$; model B has $\chi_B^2=106$ and $R_B=5$. For objective $J=\chi^2+\alpha R$, find when the preferred image changes. What does the result say about interpreting one optimum as data determined?

Solution

The objectives are

$$J_A=100+20\alpha, \qquad J_B=106+5\alpha.$$

They are equal when $15\alpha=6$, or $\alpha=0.4$. For $\alpha<0.4$, A wins because its likelihood term is better; for $\alpha>0.4$, B wins because its smaller regularizer penalty dominates. At $\alpha=1$, for example, $J_A=120$ while $J_B=111$. The data have not changed, yet the selected image has. A defensible analysis must therefore choose or tune $\alpha$ without inspecting the desired feature, report sensitivity across it, and test both candidates on held-out observables.

References

• Event Horizon Telescope Collaboration. “First M87 EHT Results from the 2018 Observing Campaign: Calibrated Data.” Data product 2024-D01-01 (2024). DOI.
• Event Horizon Telescope Collaboration. “First M87 EHT Results: Calibrated Data.” Data product 2019-D01-01 (2019). DOI.
• Event Horizon Telescope Collaboration. “First M87 EHT Results: Imaging Pipelines.” Data product 2019-D01-02 (2019), pinned commit 80d230e76c08edd31548e7cfe61dc7cf94300780. Repository.
• Event Horizon Telescope Collaboration. “First M87 Event Horizon Telescope Results. IV. Imaging the Central Supermassive Black Hole.” The Astrophysical Journal Letters 875 (2019): L4. DOI.
• Event Horizon Telescope Collaboration. “First M87 Event Horizon Telescope Results. VI. The Shadow and Mass of the Central Black Hole.” The Astrophysical Journal Letters 875 (2019): L6. DOI.
• Event Horizon Telescope Collaboration. “First Sagittarius A* Event Horizon Telescope Results. III. Imaging of the Galactic Center Supermassive Black Hole.” The Astrophysical Journal Letters 930 (2022): L14. DOI.