An Independent Look at the Event Horizon ✛
Context
Very long baseline interferometry (VLBI) provides the angular resolution needed to study the vicinity of the event horizon of supermassive black holes. The Event Horizon Telescope (EHT) collaboration has obtained the first images of M87* and Sgr A*, but image reconstruction from VLBI data remains a challenging inverse problem due to sparse aperture coverage, low signal-to-noise ratio, and significant calibration errors. Traditional methods rely on numerous hyperparameters and prior assumptions, leading to ambiguous interpretations. This work aligns with efforts to develop independent analysis methods for EHT data, alongside approaches using closure phases and amplitudes, as well as machine learning for imaging.
Methods
The authors use closure invariants—combinations of visibilities that are independent of calibration errors at individual stations. These observables contain information about the source structure but are robust to systematic errors. For image reconstruction, they use GenDIReCT, a generative diffusion model that operates in the latent space of images and is conditioned on closure invariants. The model is trained on non-astronomical images (CIFAR-10) and simple geometric shapes to avoid introducing astrophysical priors. The algorithm generates an ensemble of possible solutions, allowing for uncertainty and multimodality assessment. Before applying to target data, blind validation was performed on synthetic ngEHT datasets.
Results
On synthetic ngEHT data, GenDIReCT successfully recovered source structures in all cases except one (Sgr A* at 345 GHz with the EHT 2022 configuration). For real 3C279 data, the algorithm identified three components consistent with the EHTC image and measured the proper motion of the southern jet component: 4.6±1.0 microarcseconds over 5.39 days, corresponding to an apparent superluminal speed of about (10±2)c. For Centaurus A, two bright ridge lines along the jet were obtained, with orientation and opening angle matching EHTC results. These results confirm the method's ability to recover morphology without calibration.
Implications
This work demonstrates that closure invariants combined with generative modeling can serve as an independent path for interpreting VLBI data. Unlike related works that use calibrated visibilities or individual closure quantities, here a unified form of copolar closure invariants and a latent-space diffusion model are applied. This reduces dependence on hyperparameters and allows quantitative uncertainty assessment of the reconstruction. Such an approach complements traditional methods and strengthens confidence in scientific conclusions about black hole physics.
Future development
The method can be extended to arbitrary antenna configurations, frequencies, and aperture coverages, reducing the need for retraining for new instruments. With the advent of ngEHT and space VLBI missions, requirements for sensitivity and resolution will increase, and calibration-robust approaches will become critical. It may also incorporate polarimetric data and dynamic imaging to study rapid variations in accretion flows.
Impact
The method will impact high-resolution radio astronomy, particularly studies of supermassive black holes and relativistic jets. It may also be useful for analyzing data from future interferometric arrays.
Next steps
Apply GenDIReCT to public EHT data for M87* and Sgr A* and compare results with published images. Investigate the model's generalizability to other arrays and frequency bands.
Key open problems
This work addresses the unresolved problem of independent verification of event horizon images and quantitative assessment of systematic errors in VLBI. It also touches on the question of how strongly prior assumptions affect the reconstruction of accretion flow structures.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The paper introduces GenDIReCT, a method for reconstructing images from VLBI data that uses closure invariants instead of calibrated visibilities. The algorithm, based on a diffusion generative model, is trained on non-astronomical images and applied to synthetic ngEHT data as well as real observations of 3C279 and Centaurus A. The results are consistent with published EHTC images.
Where its strength is
The strength of the work lies in using closure invariants, which are insensitive to calibration errors, and in applying a diffusion model that can generate an ensemble of solutions to assess uncertainty. This provides an independent path for interpreting VLBI data.
Where it could go next
- It might be interesting to see how the closure invariant approach can be extended to polarimetric observations, leveraging experience from neural network polarization recovery.based on: A neural network learns to see invisible polarization arXiv:2606.16922
- One could consider applying the method to data with strong gravitational lensing, where source image distortions can be accounted for during generation, as in the work on dark matter distorting black hole portraits.based on: The Invisible Artist: How Dark Matter Distorts Portraits of Black Holes arXiv:2605.05635v3
- Another direction is adapting ideas from the work on image collection from flashes to improve reconstruction under sparse aperture coverage, which is typical for VLBI.based on: World in Flickers: How a Camera Assembles an Image from Flashes arXiv:2607.15073
Worth trying
- Test GenDIReCT's robustness to incomplete or noisy closure invariants by simulating partial data loss.
- Use the ensemble of generated images for statistical comparison with accretion disk models.
- Apply the method to data at different frequencies to study the spectral dependence of jet structure.
Why this may matter
Independent reconstruction methods that do not rely on calibration are important for increasing the reliability of interpreting EHT data and future ngEHT observations, especially when studying physics near the event horizon.
Nearby works in our archive
- Archipelago in the Ocean of Data: Red Points Without Color Boundariessimilarity 0.59
- Neural Network Clears the Gravitational Lens from Ancient Lightsimilarity 0.58
- Eccentricity: a magnifying glass for gravitational wavessimilarity 0.58
- The Invisible Artist: How Dark Matter Distorts Portraits of Black Holessimilarity 0.58
- World in Flickers: How a Camera Assembles an Image from Flashessimilarity 0.57
- A neural network learns to see invisible polarizationsimilarity 0.57
The archive is sparse around this work: the closest review is only 0.59, against a usual 0.64. It looks like few have drilled nearby.
Original abstract
as written by the authors arXiv:2608.19675v1
The Event Horizon Telescope (EHT) Collaboration's images of the supermassive black holes in M87 and the Milky Way have provided the first event-horizon-scale views of these objects, opening new avenues for studies of gravitation, accretion physics, and black hole astrophysics. Achieving these results, however, requires imaging under some of the most challenging conditions in radio astronomy, including low signal-to-noise ratios, severe calibration uncertainties, and sparse aperture coverage. With the aim of presenting independent analyses of the public EHT datasets for M87* and Sgr A*, we adopt an approach that is independent in observables, and reconstruction methodology. Our framework is based on closure invariants, a class of interferometric observables that are intrinsically immune to station-based calibration errors and therefore provide robust constraints on source structure. We combine these observables with Generative Deep learning Image Reconstruction with Closure Terms (GenDIReCT), a diffusion-based image reconstruction framework that operates in the latent space of images conditioned on closure invariants. We present independent reconstructions obtained using GenDIReCT on synthetic challenge data sets as well as real EHT data on 3C279 and Centaurus A, and compare them with previously reported results. This work demonstrates the potential of closure-invariant-driven generative imaging as a calibration-resilient framework for Very Long Baseline Interferometry (VLBI) and provides an independent and complementary avenue for interpreting horizon-scale black hole observations.
Closure invariants were used as early as the 1960s to discover the double structure of the radio galaxy Cygnus A without precise antenna calibration. In this work, the algorithm is trained on photographs of cats and airplanes from the CIFAR-10 dataset, yet it still recovers astrophysical jets.
In Carl Sagan's novel 'Contact', the protagonist uses interferometry to detect an extraterrestrial signal; here a similar principle is applied to imaging black holes. In the movie 'Interstellar', the image of Gargantua was created using simulations, not real observations — GenDIReCT brings us closer to obtaining such pictures from data.
Key numbers
- superluminal speed of 3C279 component: ≈(10±2)c
- proper motion of the jet component: 4.6±1.0 μas over 5.39 days
- distance to Centaurus A: 3.8±0.1 Mpc
- EHT observation frequencies: 227.1 and 229.1 GHz