Towards independent event horizon imaging of the supermassive black holes in M87 and the Milky Way

arXiv:2608.19675 · astro-ph.IM, astro-ph.CO · Submitted 2026-08-20 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: Next we'll be talking about the paper "Towards independent event horizon imaging of the supermassive black holes in M87 and the Milky Way".

Jocelyn: The paper was written by the authors from Commonwealth Scientific and Industrial Research Organisation (CSIRO).

Vera: Stay tuned as we take you through the paper and discuss its implications.

The Core Problem and Solution: Jocelyn: Now, looking at the initial summary of "Towards independent event horizon imaging of the supermassive black holes in M87 and the Milky Way," we really grasp why this research is so crucial for a deep understanding of space.

Vera: The paper lays out that Very Long Baseline Interferometry, or VLBI, is inherently difficult because we're dealing with sparse aperture coverage and massive calibration uncertainties that plague these observations.

Jocelyn: It's not just the data quality; the observational conditions themselves make a definitive image reconstruction almost impossible using standard methods. The signal-to-noise ratios are so low that it feels like we are looking for a needle in a giant haystack.

Subrahmanyanyan: The authors address this by focusing on closure quantities—specifically closure phases and closure amplitudes—which allow us to isolate the true structure of the source without needing perfectly calibrated gains at each individual telescope station.

Vera: That’s a clever way of saying that we can accurately determine what the object looks like even when we don't know exactly how every single piece of equipment is performing at any given time. The closure quantities bypass those systemic errors entirely, which is a massive gain for us.

Jocelyn: I am fascinated by this concept because it sounds like the the mathematical properties of the data itself are being used to overcome major engineering challenges that traditional methods simply cannot handle.

Subrahmanyanyan: This approach allows us to move beyond just checking if a single image is possible; instead, we get a set of robust constraints on what the source structure must be, based purely on these invariants.

Vera: We've seen how they define this problem and its solution in the summary, and now we can see how this leads into their specific methodology in the next segment.

The GenDIReCT Framework: Jocelyn: Moving past that initial setup, when we examine the core of "Towards independent event horizon imaging of the supermassive black holes in M87 and the Milky Way," we see how they propose a revolutionary solution called GenDIReCT. This is Generative Deep learning Image Reconstruction with Closure Terms.

Subrahmanyanyan: The core idea here is that this generative framework operates in the latent space of images, allowing us to use learned image priors from vast datasets while being strictly conditioned on those closure invariants we just discussed.

Vera: It’s a powerful combination, using a deep learning model not to guess the data, but to reconstruct plausible solutions directly from these mathematically defined observables. That's a huge shift in perspective for us all in the field of imaging.

Jocelyn: I’m trying to picture this process; it seems like they are using AI not just as an image generator, but as a sophisticated decoder for the raw, invariant data that we get from our survey equipment.

Subrahmanyanyan: The paper emphasizes that this approach allows us to sample the posterior distribution of images rather than just one single reconstruction, which is critical for quantifying uncertainty in such an under-constrained problem.

Vera: That’s incredibly useful; we are moving from a single "best guess" toward understanding the entire range of physically plausible solutions based on our data.

Jocelyn: I think this framework provides a necessary mathematical structure for our survey data, ensuring that even when we find subtle features, we know exactly how much uncertainty there is around them.

Subrahmanyanyan: We need to see how they apply this rigorous methodology to the real-world targets in the next segment to prove its capability.

Testing and Results on Real Data: Vera: Now, looking at the results presented in "Towards independent event horizon imaging of the supermassive black holes in M87 and the Milky Way," we see that they rigorously tested GenDIReCT first on synthetic challenge data sets created by the next-generation EHT. This was a vital step to ensure that our method wasn't just an academic exercise.

Jocelyn: It’s impressive how they validated the framework against these synthetic VLBI data, confirming its success across multiple scenarios before applying it to real observations. That is a huge amount of confidence built up upfront for our community.

Subrahmanyanyan: This systematic approach ensures that our theoretical foundation is solid, allowing us to confidently move forward with real-world observations without having to worry about the method itself failing under pressure.

Vera: The results from the synthetic data are very encouraging, and we see that they successfully recovered structures in virtually all cases. This suggests the framework is robust enough to handle anticipated future array upgrades as well.

Jocelyn: And when they move onto real targets like 3C279 and Centaurus A, we see that GenDIReCT identifies morphological features consistent with the established EHT results. That's a major achievement for verification of the method.

Subrahmanyanyan: This consistency is important because it tells us that our new, calibration-independent approach aligns perfectly with the observational evidence of fundamental physics in these extreme systems.

Vera: We are seeing this work provide a strong basis for setting a new standard for how we interpret VLBI data moving forward, especially as we look toward those next-generation arrays.

Jocelyn: It seems like this methodology allows us to quantify uncertainty in a way traditional methods simply cannot handle, which is absolutely huge for our wide-field survey work.

Subrahmanyanyan: We are genuinely getting closer to having a reliable laboratory for fundamental physics by being able to test general relativity with this level of precision.

Vera: It’s incredibly exciting to think about all the discoveries we have ahead as we put this framework to more real, complex targets in the Milky Way and beyond.

Jocelyn: I know that when our next set of results comes in, we'll be able to interpret them with unprecedented confidence because of this robust method.

Subrahmanyanyan: It’s definitely about building a stable, foundational tool for the entire field that will allow us to interpret the most challenging data sets in astronomy.

Conclusion and Future Outlook: Vera: So, as we wrap up our discussion of "Towards independent event horizon imaging of the supermassive black holes in M87 and the Milky Way," it’s clear that this research represents a monumental shift in how we approach high-resolution astronomy.

Jocelyn: I think what's most exciting is that this method's success across different targets like 3C279 and Centaurus A proves it isn't just one technical fluke, but a genuinely reliable tool for complex structure identification across the sky.

Subrahmanyanyan: The fact that GenDIReCT achieves results comparable to or even superior to existing analyses shows a major breakthrough in our ability to confirm the underlying physics with rigorous methodology.

Vera: It’s not just about matching previous work, though; it's about providing an an entirely new, orthogonal pathway for us all to agree on what's happening at these extreme scales without being constrained by old methods.

Jocelyn: That is huge for our survey work because it confirms that this methodology isn't just a technical gimmick but reliably identifies complex structures across different astronomical targets.

Subrahmanyanyan: The ability to support various array configurations makes the entire field more consistent and theoretically sound, regardless of how the instrument is configured in terms of geometry.

Vera: I hope this sets the standard for how we interpret all future VLBI data, especially as we look toward those next-generation arrays that will be coming online.

Jocelyn: I find it fascinating that this approach allows us to quantify uncertainty in a way traditional methods simply cannot handle, which is vital for our massive datasets.

Subrahmanyanyan: We are genuinely getting closer to having a reliable laboratory for fundamental physics by being able to test general relativity with this level of precision.

Vera: It’s incredible to think about all the discoveries we have ahead as we put this framework on more real, complex targets in the Milky Way and beyond.

Jocelyn: And I know that when our next set of results comes in, we'll be able to interpret them with much higher confidence because of this robust method.

Subrahmanyanyan: This is all about building a stable, foundational tool for the entire field that will allow us to interpret the most challenging data sets in astronomy.

Vera: We've covered so much ground today regarding Nithyanandan Thyagarajan and his team's work on "Towards independent event horizon imaging of the supermassive black holes in M87 and the Milky Way." It’s a truly exciting time for our field.

Jocelyn: We definitely have some huge implications for how we plan future observations to consider based on this groundbreaking methodology.

Subrahmanyanyan: I hope this provides the theoretical framework we need to understand the extreme environments near black holes better in the coming years.

Commonwealth Scientific and Industrial Research Organisation (CSIRO)

astro-ph.IM, astro-ph.CO

Submitted: 2026-08-20

Updated: 2026-08-20

Comments: 7 pages excluding references, 5 figures, submitted to Proceedings of SPIE Astronomical Telescopes + Instrumentation 2026

Journal ref: Proc. SPIE 14153, Radio Telescopes, Technologies, and Methods, 141530O (17 Aug 2026)

DOI: 10.1117/12.3106005

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 87/100

The gist: The Event Horizon Telescope (EHT) Collaboration’s images of M87 and Sgr A have opened new avenues for studies of gravitation and accretion physics, but achieving these results requires imaging

Key concepts

Very Long Baseline Interferometry (VLBI): VLBI is difficult because it involves sparse aperture coverage and massive calibration uncertainties. The signal-to-noise ratios are low, making definitive image reconstruction challenging using standard methods.
Closure Quantities: These are specific mathematical properties, like closure phases and amplitudes, that allow researchers to isolate the true structure of a source. They enable accurate determination of the object's appearance without needing perfectly calibrated gains at every telescope station.
GenDIReCT Framework: This is Generative Deep learning Image Reconstruction with Closure Terms. It uses a generative framework operating in image latent space, conditioned strictly on closure invariants to reconstruct plausible solutions rather than just one single image guess.

Terminology

Calibration — 25× in this paper · explained in 7 episodes
Event Horizon Telescope — 13× in this paper
Aperture — 11× in this paper · explained in 1 episode
Very Long Baseline Interferometry — 6× in this paper
Latent Space — 5× in this paper · explained in 7 episodes
Core — 4× in this paper · explained in 2 episodes
Accretion Physics — 3× in this paper · explained in 2 episodes
Diffusion Model — 3× in this paper · explained in 8 episodes
Robustness — 3× in this paper · explained in 5 episodes
Commonwealth Scientific and Industrial Research Organisation — 2× in this paper
Consistency — 2× in this paper · explained in 3 episodes
Latent Diffusion Model — 2× in this paper · explained in 2 episodes
Quantifying Uncertainty — 2× in this paper · explained in 5 episodes
Continuous Ranked Probability Score — 1× in this paper
Diffusion Models — 1× in this paper · explained in 6 episodes
Inverse Problem — 1× in this paper · explained in 2 episodes
Latent Diffusion Models — 1× in this paper · explained in 2 episodes
Verification — 1× in this paper · explained in 2 episodes

Summary

The Event Horizon Telescope (EHT) Collaboration’s images of M87 and Sgr A have opened new avenues for studies of gravitation and accretion physics, but achieving these results 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. This inherent difficulty means that image reconstruction often depends on a range of algorithmic choices and hyperparameters—such as initial models, regularisation strategies, deconvolution windows, and assumed fields of view—leading to alternative image reconstructions that differ from published EHT results.

This work aims to provide an independent analysis using a methodology based on closure invariants. The authors adopt an approach where the observables are intrinsically immune to station-based calibration errors, allowing them to isolate source-structure information while mitigating one of the dominant sources of systematic uncertainty from calibration in VLBI data analysis.

To reconstruct images from these observables, they employ Generative Deep learning Image Reconstruction with Closure Terms (GenDIReCT), a generative diffusion framework. GenDIReCT operates in the latent space of images and is conditioned on closure invariants, leveraging learned image priors while reducing dependence on arbitrary imaging hyperparameters.

The development of GenDIReCT was rigorously validated to minimize confirmation bias. Training was performed primarily on non-astronomical images (e.g., CIFAR-10), a strategy adopted to avoid embedding astrophysical expectations into the learned image priors. The framework's robustness was assessed using complementary metrics, including chi squared and the Continuous Ranked Probability Score (CRPS), ensuring that scientific conclusions remained qualitatively consistent across independent diagnostics.

The GenDIReCT methodology was then applied to several datasets:

Synthetic Challenge Datasets: Applying GenDIReCT to the ngEHT Analysis Challenges demonstrated its effectiveness, as it successfully recovered the structures in all cases except one (Sgr A* at 345 GHz with the EHT 2022 array configuration).
3C 279 Data: The application of GenDIReCT to this real EHT observation yielded results consistent with the EHTC reference image. By analyzing the data separately for each observation day, they found that a southern jet ejecta on subparsec scale exhibited a proper motion of 4.6 plus or minus 1.0 µas over about 5.39 days, corresponding to an apparent superluminal velocity of (10 plus or minus 2) c.
Centaurus A Data: GenDIReCT consistently reconstructed the two bright ridge-lines along the sheath and near the base of the approaching jet, finding that both the orientation and the opening angle of the ridge lines... are consistent with those reported by the EHTC.

The successful application of these methods provides a complementary pathway for interpreting horizon-scale VLBI observations. The authors conclude that GenDIReCT offers a robust approach to imaging under challenging VLBI conditions. Future developments will focus on improving the framework’s ability to generalize across arbitrary array configurations, observing frequencies, and aperture coverages, which is essential as the astronomical community moves toward next-generation facilities like ngEHT.

Improvements for AI systems

Based on a rigorous analysis of this paper's methodology—specifically the marriage of closure invariants with Generative Deep Learning (GenDIReCT)—I have identified three highly specific areas for improvement in current AI systems. These improvements move generative models from relying solely on learned data priors to incorporating fundamental physical constraints, leading to vastly more robust and reliable outputs.

Current Limitation: Most state-of-the-art Generative Adversarial Networks (GAN) and Diffusion Models use raw or pre-calibrated input features, making them highly susceptible to systematic noise, sensor drift, or calibration errors in the input data stream.

The Improvement: Develop a dedicated feature extraction layer that operates not on absolute visibility values, but on closure invariants (i.e., combinations of visibilities that are mathematically immune to station-based gain corruptions). This invariant set is then used as the primary conditional input (C) for the latent diffusion model.

What the Improved AI System Can Do:

Achieve Calibration-Resilient Reconstruction: The system can reliably reconstruct high-fidelity outputs even when deployed in environments (e.g, remote sensors, distributed networks) where perfect calibration is impossible or highly uncertain. It bypasses the need for complex, error-prone preprocessing pipelines.
Enable Robust Cross-Domain Deployment: By decoupling the required input fidelity from the output quality, this system can be applied to other domains (e.g., satellite imagery in variable weather conditions, medical imaging from non-standard detectors) with significantly reduced failure rates due to environmental noise.

Current Limitation: Standard generative models (like Diffusion) learn image priors based on statistical likelihood from data. They are data-driven and lack inherent physical laws, often producing outputs that are visually plausible but physically impossible (e.g, violating conservation of energy or angular momentum).

Current Limitation: Most AI systems provide a single, deterministic output (e.g., one image or one classification). This obscures the inherent ambiguity or degeneracy present in under-constrained data problems—the system reports a single truth when multiple valid solutions may exist.

Abstract

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.