PaperlayerView exact revision on arXiv

Source-linked AI summary

Interferometric Imaging Directly with Closure Phases and Closure Amplitudes

Andrew A. Chael, Michael D. Johnson, Katherine L. Bouman, Lindy L. Blackburn, Kazunori Akiyama, Ramesh Narayan

arXiv:1803.07088v1astro-ph.IMastro-ph.HE

TL;DR

Interferometric imaging depends on calibrated visibilities, yet high-frequency calibration and self-calibration assumptions can strongly affect reconstructed images. This paper instead images directly from closure amplitudes and phases, which cancel station-based calibration errors. Closure-only imaging produces calibration-insensitive results and matches the best iterative CLEAN plus self-calibration results on ALMA and VLBA datasets.

  • Problem

    Traditional interferometric imaging requires calibrated visibilities, but high-frequency calibration is difficult and self-calibration outcomes depend on initial models and other assumptions.

  • Method

    The paper uses regularized maximum-likelihood imaging directly with closure amplitudes and closure phases, supplemented by constraints for otherwise unconstrained flux and centroid information.

  • Results

    Closure-only images remain consistent in fidelity across amplitude-gain miscalibration levels and can match the best iterative multi-scale CLEAN plus self-calibration results on ALMA and VLBA data.

  • Takeaways & Limitations

    Closure-only imaging can bypass self-calibration and provide minimal-calibration-assumption source-structure estimates or starting models for calibrated-data pipelines.

  • Takeaways & Limitations

    Closure-only imaging can converge poorly from uninformative initial images and become trapped in incorrect local minima, especially for datasets with more baselines.

Abstract

from arXiv · show

Interferometric imaging now achieves angular resolutions as fine as 10 microarcsec, probing scales that are inaccessible to single telescopes. Traditional synthesis imaging methods require calibrated visibilities; however, interferometric calibration is challenging, especially at high frequencies. Nevertheless, most studies present only a single image of their data after a process of "self-calibration," an iterative procedure where the initial image and calibration assumptions can significantly influence the final image. We present a method for efficient interferometric imaging directly using only closure amplitudes and closure phases, which are immune to station-based calibration errors. Closure-only imaging provides results that are as non-committal as possible and allows for reconstructing an image independently from separate amplitude and phase self-calibration. While closure-only imaging eliminates some image information (e.g., the total image flux density and the image centroid), this information can be recovered through a small number of additional constraints. We demonstrate that closure-only imaging can produce high fidelity results, even for sparse arrays such as the Event Horizon Telescope, and that the resulting images are independent of the level of systematic amplitude error. We apply closure imaging to VLBA and ALMA data and show that it is capable of matching or exceeding the performance of traditional self-calibration and CLEAN for these data sets.

1. INTRODUCTION

Interferometric imaging requires calibrated visibilities, but high-frequency calibration is difficult and self-calibration results depend on modeling and processing assumptions. The paper presents efficient closure-only imaging to reduce those calibration assumptions while remaining computationally practical.

  • High-frequency interferometry makes calibration difficult because atmospheric phase variations and amplitude errors can change rapidly.Pointing errors from small antenna beams can also introduce large, time-varying visibility-amplitude errors.
  • Traditional self-calibration alternates imaging and gain solving, but its convergence and final image depend on initialization and other procedural choices.These assumptions include the initial source model, CLEAN regions, gain-derivation method, and recalibration cadence.
  • The paper reconstructs images directly from closure amplitudes and closure phases without requiring absolute phase or amplitude calibration beyond integration-time stability.Closure-only imaging is presented as a less calibration-committed alternative to imaging calibrated complex visibilities.
  • Analytic gradients for closure-quantity chi-squared terms make closure-only imaging nearly as computationally efficient as standard complex-visibility imaging.The method remains feasible on a personal computer for large connected-element datasets such as ALMA.
  • The framework extends regularized maximum-likelihood interferometric imaging and is implemented in the eht-imaging software library.The paper describes simulations and applications to EHT, VLBA, and ALMA data.

2. VISIBILITIES AND CLOSURE QUANTITIES

Interferometers measure noisy, gain-corrupted Fourier samples of an image, motivating closure quantities that cancel station-based calibration errors. Closure phases and amplitudes preserve source information, but reduce information content and require care under low signal-to-noise.

  • 2.1. Interferometric Visibilities: Interferometric visibilities are Fourier components of the source image sampled at projected baseline coordinates, with thermal noise and station-dependent instrumental effects.Earth rotation changes the projected baselines and supplies additional Fourier-plane samples over time.
  • 2.2. Closure Phases and Closure Amplitudes: Closure quantities are robust to arbitrarily large station-based complex gains but contain less source information than the full complex visibilities.Because they mix different Fourier components, their physical interpretation can also be difficult.
  • 2.2. Closure Phases and Closure Amplitudes: Station-based phase errors cancel when three visibilities around a triangle are multiplied, leaving a closure phase that is preserved apart from thermal noise.At low signal-to-noise, the bispectrum amplitude and phase require debiasing before imaging.
  • 2.2. Closure Phases and Closure Amplitudes: Closure amplitudes use ratios of visibility amplitudes on four stations to cancel station-based amplitude gains.Three quadrangles can be formed per four-station set, but only two are independent because their product is unity.
  • 2.3. Redundant and Trivial Closure Quantities: Redundant sites add closure-amplitude information and can improve calibration and imaging, while trivial closure-phase triangles add no new source-structure measurements.The two redundant sites in the 2017 EHT array more than doubled its closure-amplitude information.
  • 2.4. Thermal Noise on Closure Quantities: At moderately low signal-to-noise, Gaussian noise approximations and measured-SNR uncertainty estimates for closure quantities become less reliable.Log closure amplitudes mitigate reciprocal-amplitude tails and provide more robust imaging observables.

3. IMAGING WITH REGULARIZED MAXIMUM LIKELIHOOD

The section formulates interferometric imaging as regularized maximum likelihood and develops closure-based data terms that avoid reliance on calibrated complex visibilities. It also addresses information absent from closure quantities through image-level constraints and accelerates optimization with analytic gradients.

  • Closure-only data: Closure phases and amplitudes provide calibration-insensitive combinations of visibilities, enabling imaging without absolute phase or amplitude calibration beyond stability during integration.The method reconstructs images directly from closure quantities rather than separately self-calibrating amplitude and phase.
  • Closure information: Redundant EHT sites more than double closure-amplitude information, while adding no closure-phase information beyond reducing thermal noise.The closure-amplitude increase occurs because some closure amplitudes involving co-located baselines remain non-trivial.
  • Imaging with calibrated visibilities: CLEAN requires well-calibrated complex visibilities, while self-calibration iteratively derives gains from an initial image and can propagate early source-model assumptions into later rounds.The initial source model and selected clean boxes can affect the final image and convergence time.
  • Objective function: Regularized maximum likelihood minimizes an objective combining data-term goodness of fit with regularizers that constrain images when measurements alone are insufficient.Hyperparameters control the relative weighting of data terms and regularizers, with ideal data-term weights tied to measurement counts.
  • Data-term gradients: Analytic objective gradients reduce computational scaling from roughly O(M^2 × N) to O(M × N) for DTFT and from O(M × (M log M + N)) to O(M log M + N) for FFT.In typical reconstructions, the resulting imaging-speed improvement is comparable to the number of free parameters.
  • Regularizer terms: Because closure amplitudes leave image normalization unconstrained and closure phases leave centroid position unconstrained, flux and centroid regularizers restore these properties.The total flux can be supplied from a flux-calibrated single station, while the centroid can be fixed to a chosen field-of-view position.

4. IMPLEMENTATION

The implementation uses constrained optimization and repeated, progressively smoothed initialization, with Fourier computation selected according to dataset size. It evaluates reconstructions using NRMSE after accounting for resolution, grid, and centroid differences.

  • Software and optimization: The eht-imaging implementation minimizes the imaging objective with Limited-Memory BFGS and supports combinations of data terms and regularizers.The framework was implemented in the eht-imaging software library using the Scipy implementation of Limited-Memory BFGS.
  • Image representation: Positive pixel flux is enforced through the transformation I_i = exp ξ_i, while each pixel array represents a continuous image formed with a pixel pulse.Gradients are adjusted for the log-intensity parameterization.
  • Initialization: Multiple imaging runs use a blurred version of the previous result as the next initial image to improve convergence and avoid local minima caused by unconstrained high-frequency artifacts.Hyperparameter weights are also adjusted across rounds before being restored to the prescribed relative weighting.
  • Fourier computation: DTFT is used for small datasets and narrow fields of view, whereas larger VLBA and ALMA datasets use a nonequispaced fast Fourier transform when the DTFT becomes too large or slow.The implementation choice is driven by memory and repeated trial-visibility extraction costs.
  • Synthetic data: Synthetic VLBI tests vary station-based gain errors, thermal noise, elevation, and opacity before comparing reconstructed images with the true source images.The simulated gain model includes time-independent offsets and time-dependent fluctuations, with identical Gaussian standard deviations defining the gain-error level.
  • Fidelity evaluation: NRMSE evaluates pixel-to-pixel similarity after matching resolution and image grids and shifting reconstructions to maximize cross-correlation.These adjustments address unresolved fine-scale structure, differing pixel grids, and centroid offsets in uncalibrated-phase reconstructions.

5. RESULTS

Simulated EHT tests show closure-only imaging remains faithful across amplitude gain errors, while redundant sites improve accuracy and direct closure-amplitude imaging can outperform its logarithmic variant. Applications to VLBA and ALMA data produce broadly consistent, high-quality images, though convergence and multiscale structure remain practical limitations.

  • 5.1. Results: Simulated EHT Images: Closure-only imaging remains insensitive to amplitude gain error, whereas methods using the bispectrum or visibility amplitudes lose fidelity as gain error increases.The simulations used gain uncertainties of 0%, 5%, 10%, 25%, 50%, 75%, and 100%.
  • 5.1. Results: Simulated EHT Images: Direct closure-amplitude imaging is more faithful than log closure-amplitude imaging, but it often converges more slowly and depends more strongly on data-term weights and initial field of view.These results are reported for the simulated Sgr A∗, M87, and related reconstructions.
  • 5.1. Results: Simulated EHT Images: For the narrow, high-dynamic-range jet image, closure-only methods preserve overall structure across gain errors but achieve less dynamic range at zero gain error.Spurious low-luminosity features were attributed to a local minimum in the objective function and sparse Fourier coverage.
  • 5.1. Results: Simulated EHT Images: Closure-only reconstructions can approach the fidelity of gain-calibrated methods when redundant EHT sites provide additional short-baseline closure amplitudes.Removing redundant sites makes the reconstructions substantially less accurate.
  • 5.2. Results: VLBA and ALMA Images: Closure imaging produces VLBA and ALMA images broadly comparable to expert CLEAN and self-calibration reconstructions, while post-hoc self-calibration can sharpen resolution and disk gaps.The ALMA closure reconstruction also avoids prominent periodic dark spots associated with dirty-beam sidelobes in the CLEAN image.
  • 5.2. Results: VLBA and ALMA Images: Closure imaging tends to avoid high-frequency artifacts, but it recovers less extended jet structure than multiscale CLEAN in the VLBA example.A multiscale approach is identified as a potential improvement for closure imaging.

6. DISCUSSION

Closure-only imaging reduces dependence on calibration assumptions while producing competitive reconstructions for simulated EHT, VLBA, and ALMA data. Its main practical challenge is early convergence, especially for real datasets and narrow structures.

  • Closure-only imaging is less biased by calibration assumptions and can serve as a minimal-assumption source-structure estimate or self-calibration model.The method does not eliminate all tunable parameters, but it avoids dependence on the self-calibration loop’s calibration assumptions.
  • Closure-only reconstructions maintain consistent fidelity across amplitude gain-error levels in simulated EHT data.With redundant stations, fidelity approaches that of images made from perfectly calibrated data using conventional algorithms.
  • One self-calibration round after closure-only imaging matches the best iterative, multi-scale CLEAN plus self-calibration results on ALMA and VLBA datasets.The resulting complex-visibility images are smooth and high resolution.
  • Early convergence is difficult, with closure-only optimization sometimes becoming trapped in incorrect local minima, particularly for real ALMA and VLBA data.A weak uncalibrated-visibility-amplitude constraint, as low as 1–10% of the closure-amplitude χ2 term, helped guide convergence.
  • The eht-imaging framework supports flexible combinations of data terms and regularizers, self-calibration, synthetic-data generation, and interferometric-data analysis.

7. CONCLUSIONS

The paper presents regularized maximum-likelihood imaging with closure quantities and implements it in eht-imaging. Closure-only imaging performs robustly on simulated EHT, VLBA, and ALMA data, while additional amplitude constraints can improve convergence.

  • The framework extends regularized maximum-likelihood imaging by fitting closure amplitudes or their logarithms directly instead of relying on amplitude self-calibration.It is implemented in the eht-imaging software library and supports arbitrary data products.
  • Analytic gradients for closure phases, closure amplitudes, and log closure amplitudes enable efficient imaging directly from uncalibrated data.
  • Closure-only imaging bypasses self-calibration, minimizes calibration assumptions, and remains insensitive to the level of amplitude gain miscalibration in simulated EHT data.With redundant sites, its fidelity approaches that of conventional imaging using perfectly calibrated data.
  • Closure imaging produces millimeter-wavelength VLBA and ALMA images comparable in quality to expert multi-scale CLEAN and self-calibration reconstructions.
  • A single subsequent self-calibration and complex-visibility imaging pass produces further-refined images with fewer suspicious features attributable to CLEAN artifacts.
  • Calibration-insensitive closure quantities are especially valuable at millimeter and submillimeter wavelengths, where calibration uncertainties are a large and variable part of the error budget.A soft constraint from uncalibrated visibility amplitudes can dramatically aid convergence.

APPENDIX

The appendix gives gradient expressions used by the imaging software for visibility and closure-based data terms. These expressions rely on a sparsely sampled discrete Fourier-transform representation.

  • The implementation assumes a DTFT matrix A_ij and uses its conjugate transpose A† as the adjoint operator.Because the visibility data are sparsely sampled, A†A ≠ 1.
  • The appendix presents the gradient of the complex-visibility χ2 term introduced as Equation 16 and also shown as Equation 22.
  • It provides the gradient of the visibility-amplitude χ2 term used in the imaging objective.
  • It provides gradients for the closure-amplitude χ2 and log-closure-amplitude terms used by closure-only imaging.
Loading 1803.07088v1…