M87* Black Hole Ring Confirmed by Independent AI-Based Reanalysis of EHT Data

M87, a giant elliptical galaxy about 55 million light-years from Earth, hosts the supermassive black hole M87*. When an orange ring of light surrounding it was unveiled in 2019, the image drew worldwide attention. What the Event Horizon Telescope (EHT) captured was a structure thought to arise when the black hole's intense gravity bends light, leaving a dark central "shadow." It became a historic visual confirmation of black holes and one of the defining images of modern astronomy.

But the image is not a photograph taken with an ordinary camera. It was reconstructed computationally from limited observational data gathered by combining radio telescopes around the world. Several image-reconstruction methods were used for verification when the image was released in 2019, but because the data are extremely sparse, there remained a possibility that fine details could change depending on the initial model or analysis conditions. Some independent studies had even reported that changing the analysis conditions made the ring structure itself unclear.

A research team led by Samuel Lai of Australia's Commonwealth Scientific and Industrial Research Organisation (CSIRO) has now published new verification results. According to the paper "Independent interferometric observable confirms black hole shadow in M87*", published in Nature on November 11, 2026, the team developed an original image-reconstruction method combining "closure invariants" with generative AI and reanalyzed the EHT's public data. Even with an approach different from the conventional ones, a ring structure about 42 microarcseconds (μas) in diameter was stably reproduced. This provides strong, independent evidence that the ring around M87* is not an artifact of particular image processing.

Importantly, this study did not involve new radio observations; it verified existing data by a different method. The team used public data acquired in 2017, 2018, and 2021, and recovered images in a way that is less affected by instrument calibration errors and by artificial settings in the reconstruction process. In CSIRO's announcement, Laurent Loinard, director of the EHT Collaboration, also welcomed the result, stressing the significance of obtaining the same ring structure from multiple years of data using an independent method.

The difficulty of reconstructing images with an Earth-sized virtual telescope

The ring around the M87* black hole has an angular diameter of only about 40 μas, comparable to the angle subtended by a single human hair seen from more than 1,000 km away. Observing such a tiny structure with radio waves of about 1 mm wavelength would require a telescope as large as the Earth. Since building such a huge antenna is impossible, the EHT uses very long baseline interferometry (VLBI), precisely synchronizing radio telescopes around the world. By combining the radio waves received by multiple telescopes and also exploiting the Earth's rotation, it forms a virtual Earth-scale telescope.

This approach has a drawback: the data are extremely limited. Real radio telescopes exist only at limited locations on Earth, so most of the virtual aperture contains no measurements. What a radio interferometer measures is a subset of the "spatial frequencies," the Fourier transform of the brightness distribution of the target, and recovering the original image requires estimating the missing information. This is known mathematically as an ill-posed problem, meaning that multiple images can be consistent with the same observational data.

Atmospheric fluctuations above each observatory further complicate the analysis. Changes in atmospheric water vapor disturb the phase and amplitude of radio waves, and gain variations in receivers also introduce measurement errors. Because these fluctuations occur on short timescales, fully calibrating the data is difficult. Conventional image reconstruction has relied on methods using the CLEAN algorithm, hybrid mapping combined with self-calibration, and regularized maximum likelihood estimation, all of which can yield results that depend on analysis conditions.

Particularly problematic are the "initial image" given at the start of reconstruction and the regularization settings that stabilize the solution. For example, one of the image-reconstruction pipelines published by the EHT used a Gaussian with a full width at half maximum of 40 μas as its initial image. Because this specifies the approximate size of the target in advance, the final image could be drawn toward the initial condition. In fact, using a larger Gaussian as the starting point can produce artificial repeating structures in the image.

Against this background, some independent reanalyses claimed that widening the field of view (FOV) over which the image is recovered blurred the ring structure, yielding an image closer to a single stretched bright spot. That raised the question of whether the ring published by the EHT reflected a real astronomical structure or was formed by the analysis conditions. The present study tests this field-of-view dependence and related reconstruction issues using a different approach.

{
  "type": "bar",
  "title": "解析手法別のM87*リング直径の比較",
  "unit": "μas",
  "categories": ["GenDIReCT（非パラメトリック）", "幾何学的フォワードモデリング", "EHTC公表値（2019年）"],
  "series": [
    { "name": "推定平均直径", "values": [42.2, 42.3, 42.0] }
  ],
  "caption": "GenDIReCTは42.2±1.5μas、幾何学的フォワードモデリングは42.3±1.1μas。2019年のEHTC公表値は42±3μasで、いずれも誤差範囲内で整合する。",
  "source": "Nature (2026), DOI: 10.1038/s41586-026-11241-9; EHT Collaboration (2019), DOI: 10.3847/2041-8213/ab0ec7"
}

The diameter estimates from the new reanalysis agree well with the 42 ± 3 μas reported by the EHT Collaboration in 2019. That different analysis methods yield comparable values indicates that the size of the ring does not depend strongly on the reconstruction conditions.

Closure invariants: removing the effects of calibration errors

A key role in this study was played by observables called "closure invariants." These mathematically combine measurements from multiple radio telescopes so as to cancel calibration errors specific to each station. In radio interferometry, it has long been known that summing the phases around a triangle formed by three telescopes cancels the phase errors at each station. In addition to this "closure phase," the "closure amplitude," which uses ratios of amplitudes from four telescopes to eliminate gain errors, is also widely used.

The closure invariants adopted by the team are a generalization that treats these ideas in a unified way. From a triangle of three stations they construct a quantity called an "advariant," and by applying further mathematical normalization they derive observables that do not depend on phase or gain errors at individual stations. A distinguishing feature is that information equivalent to conventional closure phases and closure amplitudes can be handled within a common mathematical framework.

Closure invariants have a limitation: they cannot determine the absolute position or total radio flux of the target. Because that information depends on station calibration, it cannot in principle be obtained from closure invariants alone. Information about the shape and the relative distribution of brightness, however, remains. The team exploited this property to directly recover the ring structure of M87* from information that is less sensitive to station-by-station calibration errors.

This reduces the need to finely adjust instrument calibration values during reconstruction or to steer the solution toward a presumed shape of the target. However, closure invariants do not eliminate every observational error. Systematic errors including time variability may remain, and gaps in the data caused by the telescope array's geometry are not resolved. The method does not improve the performance of the instruments; it makes use of morphological information that can be extracted from existing data without depending on calibration.

In this way, the team compared results from methods with different characteristics. If both an AI-based image reconstruction and an analysis that fits geometric models directly to the data support the same ring structure, the possibility that it is a bias specific to one algorithm shrinks. However, since both use the same public EHT data, this must be distinguished from confirmation by new, independent observations.

Independent verification with a diffusion model and geometric models

To reconstruct images from closure invariants, the team used a deep-learning algorithm it developed called GenDIReCT (Generative Deep learning Image Reconstruction with Closure Terms).

At the heart of GenDIReCT is a latent diffusion model, widely used in image-generating AI. It generates images conditioned on closure invariants computed from the observational data, and then uses a convolutional neural network (CNN) to merge and adjust the generated images so that they are consistent with the actual observables. Unlike conventional reconstruction methods, it does not require the analyst to manually specify an initial image of a particular size or shape.

Particularly important in this study are the AI's training data and validation procedure. GenDIReCT was trained mainly on processed natural images from CIFAR-10, not on observational or simulated images of black holes. However, the standard training data also include simple geometric patterns such as Gaussians, disks, ellipses, and rings. To examine whether having learned ring structures in advance might have influenced the results, the team therefore ran an additional test excluding the geometric patterns. Even then, no notable difference appeared in the reconstructed images, confirming that the appearance of the ring does not depend solely on ring patterns in the training data.

Before using the actual M87* data, the team also tested the model under a variety of shapes and noise conditions. After the model passed validation and was applied to the real data, they avoided modifying it based on the results. This is a measure to curb confirmation bias, in which researchers tune an algorithm to approach the image they expect. That said, the statistical prior the deep-learning model acquired from general images does not vanish, so this is not image reconstruction with absolutely no prior assumptions.

The analysis used the EHT's public observational data in the 1.3 mm band. A total of eight observation epochs were used: four days in 2017 (April 5, 6, 10, and 11), two days in 2018 (April 21 and 25), and two days in 2021 (April 13 and 18). For these data, the team set the reconstruction field of view to three sizes:

• 128\times128,\mu\mathrm^2
• 225\times225,\mu\mathrm^2
• 450\times450,\mu\mathrm^2

The field of view acts as an indirect constraint on the maximum size of the reconstructed structure. For example, a narrower field limits the size of structures that fit in the image, which may make a compact ring more likely to appear. The team therefore prepared independent models for each combination of observation date and field of view, and tested whether the ring persists when the field is greatly widened. This directly addresses the problem raised in earlier independent reanalyses that "the ring disappears when the field of view is widened."

Separately from the AI-based reconstruction, the team also performed "geometric forward modeling" based on Bayesian statistics. In this method, predefined geometric shapes are fit to the observational data and the parameters that best explain the data are estimated. Specifically, centered on the "xsringauss" model representing the ring structure, they added one large Gaussian and two offset elliptical Gaussians so that extended emission other than the ring could also be represented.

Parameter estimation used dynamic nested sampling with DYNESTY. This explores candidate combinations of parameters broadly and derives probability distributions from their consistency with the observed closure invariants. Whereas GenDIReCT reconstructs images directly, this approach estimates the ring's diameter and shape from the observables, so comparing the two results allows biases specific to the image-reconstruction method to be tested. However, the results of geometric forward modeling retain constraints stemming from the shape models adopted.

A ring of about 42 μas reproduced across eight epochs

The analysis consistently yielded a compact ring structure with a scale of about 40 μas across the eight observation epochs. In some data with poorer observing conditions, the ring appeared blurred in individual reconstructions, but overall the results agree well with the conventional EHT images.

In particular, for the data with good observing conditions in each year, April 11, 2017, April 21, 2018, and April 18, 2021, the ring structure appeared stably regardless of the field-of-view size. Merging the images generated by GenDIReCT produced a nearly circular ring with a relatively constant diameter and orientation. The brightness shows an asymmetry that strengthens from the south toward the southwest, a feature also consistent with the conventional EHT images.

Regarding the field-of-view dependence flagged in earlier independent reanalyses, the team pointed out that the quality of the observational data may have played a role. Gaps in spatial-frequency coverage due to the station layout and low signal-to-noise ratios may have made image reconstruction unstable. In the present study, at least when high-quality data were used, the ring structure was not lost within the range of fields of view examined. This result is an important rebuttal to the suspicion that the ring was artificially created by a particular field-of-view setting.

Quantitatively, the two independent methods also agreed well. The mean ring diameter from GenDIReCT's nonparametric image reconstruction was 42.2\pm1.5,\mu\mathrm, and the estimate from geometric forward modeling was 42.3\pm1.1,\mu\mathrm. The difference between the central values is only 0.1 μas, well consistent within the estimation errors. It also agrees with the 42\pm3,\mu\mathrm published by the EHT Collaboration in 2019, reaffirming the robustness of the ring-diameter measurement.

Determining the black hole's mass from this diameter requires calibration with additional theoretical models. The size of the observed ring depends not only on the black hole's gravitational radius but also on the process by which radio waves emitted from the surrounding plasma undergo gravitational lensing. The team used a conversion factor obtained from general relativistic simulations of accretion flows, together with a measured distance to M87 of 16.8^0.8_-0.7,\mathrm. As a result, the mass of M87* was estimated as follows:

\[ M_{\mathrm{BH}}=(6.5\pm0.5)\times10^9\,M_{\odot} \]

Here M_ denotes the solar mass. In other words, the mass of M87* is estimated at about 6.5 billion times that of the Sun. However, this number is not obtained directly from the ring diameter alone; it presupposes the distance to the object and a conversion factor based on simulations.

This result is also important for evaluating past measurements of M87*'s mass. The mass estimated from stellar motions was (6.6\pm0.4)\times10^9,M_, whereas the value obtained from gas motions was $3.5^0.9-0.7\times 10^9,M$, a large discrepancy between the two. The present estimate agrees well with the stellar-motion measurement and reconfirms the EHT's 2019 mass estimate with an independent image-analysis method.

On the other hand, if the smaller mass from gas motions were adopted, the expected ring diameter would be about 23 μas, which would be difficult for the EHT, with a resolution of about 20 μas, to resolve clearly. The present result supports the larger mass, but it does not directly resolve the cause of the discrepancy with the gas-motion estimate. That discrepancy will need to be examined further with future high-precision kinematic observations.

The ring is more firmly established, but uncertainty remains in fine structure

This study has more strongly corroborated the existence of the ring structure of about 40 μas surrounding M87*. That does not mean, however, that the details of the reconstructed image have been determined accurately. Indeed, the paper reports that while the ring's diameter and overall orientation are relatively stable across analysis methods, the ring's width and the darkness of its center differ considerably.

For example, conventional reconstructions with DIFMAP are convolved with a 20 μas circular beam, which tends to make the ring look wider than it really is. In contrast, GenDIReCT and regularized maximum likelihood methods yield a sharper ring. The "central brightness ratio," which expresses how dark the center of the ring is relative to its surroundings, also varies with the analysis method. The team points out that such metrics do not necessarily reflect the true physical properties around the black hole accurately, and may be strongly influenced by how the reconstruction is processed.

Uncertainty also remains regarding the brightness asymmetry along the ring. From 2017 through 2018 and 2021, the orientation of the bright part tended to shift counterclockwise. For the change seen between 2017 and 2018 in particular, the team suggests that turbulent motion in the accretion flow around the black hole may be involved. However, the day-to-day variation within 2017 carries uncertainty due to insufficient data, and the specific motion of the accretion flow cannot be uniquely determined from changes in brightness.

Also important are the constraints on the black hole's spin. The team combined general relativistic magnetohydrodynamic (GRMHD) simulations with the radiative transfer code IPOLE to create synthetic images assuming different spin rates and viewing angles. They then compared their similarity to the GenDIReCT reconstructions to see which physical model came closest to the observations.

As a result, models in which the angle between the observer's line of sight and the black hole's spin axis is about 20 degrees showed higher image similarity. This value is also consistent with the line-of-sight angle of M87's jet estimated from past radio observations. A tendency was also seen for models of spinning black holes to show higher similarity than non-spinning ones.

However, a specific spin rate for the black hole could not be determined from this result. What the team evaluated was the similarity between images generated from particular theoretical models and the observed images; they did not directly measure the magnitude of the spin. The estimated inclination of about 20 degrees likewise comes from comparison with simulation models and is not a value uniquely obtained from the observational data alone.

Space VLBI, which places radio telescopes in space, is expected to go beyond these limits. The ground-based EHT is limited in baseline length by the size of the Earth, but combining it with space-based telescopes could allow even longer baselines. If this could resolve the extremely fine structure of the "photon ring" formed by light circling around the black hole, it would open a path to measuring spin more directly and precisely.

The study's greatest achievement is confirming, through independent analysis, that the ring structure of M87* released in 2019 does not depend on a particular image-reconstruction method or initial condition. At the same time, uncertainty remains about finer physical properties such as the ring's width, the darkness of its center, and the black hole's spin. Now that confidence in the basic shape of M87* has grown, the next challenge is to uncover in more detail the structure of spacetime hidden within it, through improved observational resolution and verification of theoretical models.