Conditional Image Diffusion with Interferometric Closure Invariants: Independent EHT Imaging of Centaurus~A and 3C~279

1. Problem Statement

Very-long baseline interferometry (VLBI), particularly the Event Horizon Telescope (EHT), achieves the highest angular resolution in astronomy but faces a severely ill-posed inverse problem. Reconstructing images from sparse aperture coverage is complicated by:

• Calibration Systematics: Traditional imaging relies on visibilities, which require precise calibration of station-based gains (amplitude and phase). Errors in calibration (due to atmospheric effects or electronics) can lead to divergent image reconstructions depending on the assumptions and initial models used.
• Methodological Bias: Different deconvolution algorithms and regularization priors often produce varying fine-scale features (e.g., ring thickness, hot spot locations), limiting the scientific interpretability of the results.
• Lack of Independent Verification: While EHT data is public, independent verification of EHT Collaboration (EHTC) reconstructions using alternative, calibration-independent techniques has been limited.

The core challenge is to develop a robust imaging method that bypasses traditional calibration, relies solely on physical observables immune to station-based errors, and avoids the need for fine-tuned hyperparameters or subjective priors.

2. Methodology: GENDIRECT

The authors apply GENDIRECT (Generative Deep learning Image Reconstruction with Closure Terms), a novel machine-learning framework based on conditional diffusion models.

• Primary Observables (Closure Invariants): Instead of using raw visibilities, GENDIRECT uses co-polar closure invariants. These are constructed from triangular loops of stations (advariants) and are mathematically immune to arbitrarily large station-based multiplicative corruptions (calibration errors). They preserve robust morphological information limited only by additive thermal noise.
• Generative Architecture:
  • Conditional Latent Diffusion Model: A pre-trained model learns the effective prior and regularisation from a diverse dataset (natural images, shapes, etc.). It generates candidate images conditioned on the observed closure invariants. This eliminates the need for user-input hyperparameters or manual regularization terms.
  • Convolutional Neural Network (CNN) Refinement: A separate CNN optimizes the diffusion-generated sample to maximize the likelihood of consistency with the input closure invariants (minimizing the $\chi^2_{CI}$ metric).
• Training Strategy: Models are trained separately for each observing day using augmented natural images and simple shapes. The system is designed to be "blind" to the specific source, relying on the generative nature of the diffusion model to explore the solution space.
• Multimodality Sampling: Unlike deterministic methods, GENDIRECT produces a cluster of plausible image solutions. The authors analyze these clusters to identify dominant morphologies and potential degeneracies (e.g., phantom structures caused by sparse sampling).

3. Key Contributions

• First Real-Data Application: This is the first application of GENDIRECT to real EHT observations, moving beyond synthetic validation.
• Calibration-Independent Imaging: The study demonstrates that high-fidelity imaging can be achieved using only closure invariants, effectively bypassing the calibration pipeline and its associated systematics.
• Independent Verification: The authors provide an independent reconstruction of two EHT targets (3C 279 and Centaurus A) without relying on EHTC calibration solutions or priors, serving as a robust blind test.
• Quantitative Comparison: The paper establishes a rigorous framework for comparing generative reconstructions against reference images using metrics like reduced $\chi^2_{CI}$ (data fidelity) and maximum normalized cross-correlation ($\rho_{NX}$).

4. Results

The authors reconstructed images for 3C 279 (a blazar) and Centaurus A (a radio galaxy) and compared them to EHTC reference models.

3C 279 (Blazar)

• Morphology: The GENDIRECT median reconstruction reveals two prominent emission regions (assumed "core" and "ejecta") separated by $\approx 90 \, \mu\text{as}$.
• Quantitative Agreement:
  • Correlation: $\rho_{NX} = 0.98$ with the EHTC reference.
  • Brightness Ratio: $1.53 \pm 0.02$ (Core/Ejecta), consistent with EHTC's $1.52 \pm 0.01$.
  • Separation & Angle: Measured separation ($90.2 \pm 0.2 \, \mu\text{as}$) and position angle ($-75.6 \pm 0.1^\circ$) match EHTC results within $1\sigma$.
• Proper Motion: By analyzing data from April 5, 6, 10, and 11, 2017, the authors detected proper motion of the southern ejecta component relative to the core.
  • Displacement: $4.6 \pm 1.0 \, \mu\text{as}$ over $\approx 5.39$ days.
  • Velocity: Corresponds to an apparent superluminal velocity of $\beta_{app} \approx 10 \pm 2$ times the speed of light, consistent with EHTC findings.

Centaurus A (Radio Galaxy)

• Morphology: The reconstruction identifies two prominent emission ridges along the jet sheath.
• Quantitative Agreement:
  • Correlation: $\rho_{NX} = 0.91$ with the EHTC reference.
  • Brightness Ratio: $1.4 \pm 0.1$ (South/North), consistent with EHTC's $1.5 \pm 0.2$.
  • Opening Angle: $12.3 \pm 0.3^\circ$, consistent with EHTC's $12.2 \pm 0.1^\circ$.
• Limitations: The counter-jet and faint extended features ($\sim 200 \, \mu\text{as}$ from the apex) were fainter or absent in the GENDIRECT reconstruction compared to the EHTC model, likely due to the limited dynamic range of closure-only imaging and the sparsity of the array.

Multimodality Analysis

• The generative approach identified clusters of solutions. The dominant cluster for both targets contained $>50\%$ of reconstructions and matched the EHTC morphology.
• Minor clusters showed "phantom" repetitions or alternative brightness ratios, highlighting the inherent degeneracies in sparse VLBI imaging that regularization methods might suppress or bias.

5. Significance

• Validation of Closure Invariants: The study proves that closure invariants preserve sufficient morphological information to reconstruct complex astrophysical structures without traditional calibration, even in extremely sparse arrays like the EHT.
• Robustness to Systematics: By avoiding calibration, GENDIRECT provides a fundamentally independent view of spatial structure, confirming that the key features reported by the EHTC (ring structures, jet ridges, proper motion) are robust against calibration assumptions.
• Generative Imaging Paradigm: The work demonstrates that deep generative models can effectively solve ill-posed inverse problems in astronomy by learning priors from data rather than relying on manual regularization.
• Future Applications: GENDIRECT is presented as a robust, blind-imaging tool for current and future VLBI experiments, including the upcoming imaging of M87 and Sgr A* event horizons, and has potential for dynamic and polarimetric imaging extensions.

In conclusion, the paper successfully validates GENDIRECT as a powerful, calibration-independent tool that independently confirms the structural findings of the EHT Collaboration, thereby strengthening the reliability of VLBI imaging results.