Calibration of Quasi-Redundant Interferometers

J. L. Sievers

Source-linked AI summary

Calibration of Quasi-Redundant Interferometers

J. L. Sievers

arXiv:1701.01860v1Jan 7, 2017astro-ph.IM

TL;DR

Radio interferometers need calibration despite diffuse foregrounds and imperfectly redundant arrays. The paper introduces covariance-based correlation calibration with partial sky information and sparse computation, finding improved reconstruction over redundant calibration in the tested simulations.

Problem
Radio-array calibration is difficult because traditional redundant calibration assumes exact baseline redundancy, whereas real arrays have antenna and beam non-uniformities and diffuse foregrounds dominate target signals.
Source: §1 · p. 1, §2.1 · p. 3
Method
The paper models expected visibility covariances to relax explicit redundancy, incorporates known source positions with unknown fluxes, and uses a two-level sparse matrix inverse for large arrays.
Source: §1 · p. 2, §2.3 · p. 4, §3 · p. 5
Results
Including bright-source position information, the simulations improve amplitude/phase reconstruction by approximately 2/5 relative to redundant calibration for the adopted noise and position errors.
Source: §4 · p. 8, §6 · p. 10
Takeaways & Limitations
Correlation calibration provides a flexible framework for calibrating imperfect or non-redundant arrays and can recover the overall phase gradient when source positions are included.
Source: §6 · p. 10
The method requires sufficient quasi-redundancy; if visibility blocks cannot be represented by substantially fewer eigenvectors, redundancy-based calibration may be unsatisfactory.
Source: §3.1 · p. 5

Abstract

from arXiv · show

High precison calibration is essential for a new generation of radio interferometers looking for Epoch of Reionization and Baryon Acoustic Oscillation signatures in neutral hydrogen. These arrays have so far been calibrated by redundant calibration, which usually assumes baselines intended to be identical are perfectly so. We present a new calibration scheme that relaxes the assumption of explicit redundancy by calculating the expected covariance of baselines. The technique also allows one to take advantage of partial knowledge of the sky, such as point sources with known positions but unknown fluxes. We describe a 2-level sparse matrix inverse to make the calibration tractable for 1,000-element class interferometers. We provide a reference implementation and use it to test the calibration of simulations of an array with imperfectly located antennas observing Euclidean-distributed point sources. Including position information for a handful of the brightest sources, we find the amplitude/phase reconstruction improves by a factor of $\sim$2/5 over redundant calibration for the noise levels/position errors adopted in the simulations. Inclusion of source positions also allows us to measure the overall phase gradient across the array, information which is lost in traditional redundant calibration.

1. Introduction

Close-packed radio arrays rely on redundant calibration for difficult EoR and BAO observations, but real arrays are imperfect and diffuse foregrounds make accurate sky modeling challenging.

Motivation

EoR and BAO arrays face calibration challenges because diffuse foregrounds can exceed the target signals by factors of 10^3 or more.Traditional self-calibration depends on a useful common sky model, which may be impossible to construct at the required quality.Source: §1 · p. 1

Existing approach

Redundant calibration solves for the sky and relative antenna gains by treating nominally identical baselines as measurements of the same signal.This approach has successfully calibrated redundant arrays such as PAPER.Source: §1 · pp. 1–2

Gap and contribution

Real arrays violate exact redundancy through antenna-position errors and primary-beam variations, limiting the assumptions of traditional redundant calibration.The proposed alternative accounts for these imperfections while retaining the diffuse sky as the dominant signal.Source: §1 · p. 2

2. Likelihood Formulation

The likelihood formulation replaces exact redundancy with covariance modeling, enabling calibration with imperfect arrays and partial sky information while retaining traditional redundant calibration as a limiting case.

2.1. Traditional Redundant Calibration

Traditional redundant calibration models per-antenna complex gains and shared sky values for baselines occupying the same UV point.Its likelihood can be minimized when the number of visibilities exceeds the number of antennas plus unique baselines, but four degeneracies remain.Source: §2.1 · p. 2
Traditional redundant calibration cannot determine overall gain, global phase, or two-dimensional phase-gradient degeneracies from visibilities alone.The phase-gradient degeneracy is equivalent to changing the array pointing center while shifting the sky model.Source: §2.1 · p. 2
Its two key assumptions are an unconstrained sky and explicitly redundant baselines, while UV-gradient corrections can increase noise and fail to represent realistic beam and position variations.For arrays with pointing, positioning, and primary-beam scatter, extending the formalism to all effects rapidly becomes intractable.Source: §2.1 · pp. 2–3

2.2. Correlation Calibration

Correlation calibration removes explicit sky dependence by representing sky contributions as covariance between visibility measurements.For a redundant block, the effective covariance combines diagonal visibility noise with an outer-product sky term.Source: §2.2 · p. 3
The covariance formulation is equivalent to fitting and marginalizing over the sky, including in the infinite-sky-variance limit that recovers traditional redundant calibration.The equivalence follows because both procedures produce the same minimized χ2.Source: §2.2 · pp. 3–4

2.3. Relaxing Redundancy

Finite sky covariance allows non-identical baselines to be related through a spatial power spectrum and modeled effects such as pointing offsets, dish-position errors, and beam variations.The method assumes the sky is a Gaussian random field, an approximation described as robust for realistic skies.Source: §2.3 · p. 4
Known source positions can be incorporated by adding their predicted visibility covariance, while taking the source amplitude to infinity marginalizes over unknown flux.This gives the likelihood less penalty for placing signal where a known source is expected.Source: §2.3 · p. 4

3. Minimizing χ2

This section develops a sparse covariance-based formulation for evaluating and minimizing calibration χ2 in quasi-redundant arrays. A two-level Woodbury inversion and nonlinear conjugate-gradient solver make the method practical for 1,000-antenna-class problems.

3.1. Evaluating χ2

The covariance formulation groups visibilities into quasi-redundant blocks while incorporating known-source contributions and visibility noise.The real formulation uses 2×2 inverse-gain blocks and eigenvectors to approximate imperfectly redundant covariance modes.Source: §3.1 · p. 5, §3 · p. 5
Redundancy-dependent calibration becomes unreliable when covariance blocks cannot be represented by substantially fewer eigenvectors than their visibilities.This condition indicates insufficient redundancy for a satisfactory solution.Source: §3.1 · p. 5
Sparse covariance structure reduces calibration χ2 evaluation to operations scaling as nvis times max(nsrc, nredundant)^2.Quasi-redundant blocks and known-source contributions are represented through vector outer products and inverted with repeated Woodbury identities.Source: §3.1 · pp. 5–6

3.2. Minimizing χ2

Calibration is optimized with a nonlinear conjugate-gradient solver using gradients assembled efficiently by looping over visibilities.Shared matrix-vector products are reused across antenna gains, and each visibility contributes to the two antennas in its block.Source: §3.2 · p. 6

3.3. Reference Implementation

A reference Python/C implementation solves a full 1,000-antenna-class calibration in 15–25 seconds on one laptop-class core under a representative setup.The same setup evaluates χ2 in less than 0.1 seconds and its gradient in about 0.25 seconds.Source: §3.3 · pp. 6–7
The implementation currently requires only a C compiler and Python, with future work targeting minimization, data reuse, and GPU optimization.The authors note that small problems are highly memory-limited.Source: §3.3 · p. 7

4. Application to Sample Data

Simulations of an imperfect 8×8 array compare source-aware covariance calibration, covariance calibration without bright-source positions, and redundant calibration. Source-aware calibration most improves phase reconstruction, while correct antenna-position covariances improve amplitude reconstruction.

4. Application to Sample Data

The comparison used roughly 2,000 simulations and evaluated amplitude scatter relative to the mean and phase scatter after fitting an array-wide offset and gradient.This treatment accounts for poorly constrained overall normalization and phase gradient.Source: §4 · p. 7
8.3 × 10^-4 amplitude and 7.0 × 10^-4 phase standard deviations were achieved with full covariance and bright-source position information.The simulations used an 8×8 quasi-regular array with 0.04-wavelength position noise and typically about 10 individually treated bright sources.Source: §4 · pp. 7–8
Phase error was about 5 times lower with source-aware calibration than with redundant or source-unaware covariance calibration.The source-unaware covariance and redundant cases performed similarly, with covariance calibration about 5% better.Source: §4 · p. 8
Amplitude reconstruction was improved by a factor of 2 when true antenna positions informed the covariance, while bright-source information added only modest amplitude improvement.The corresponding standard deviations were 9.2 × 10^-4 without source positions and 1.7 × 10^-3 for redundant calibration.Source: §4 · p. 8

5. Future Applications

The paper outlines extensions of correlation calibration for non-redundant arrays, antenna-property inference, bandpass calibration, and polarization calibration.

5. Future Applications

The χ2 formulation is sufficiently general to support additional future applications beyond the extensions described here.Source: §5 · p. 8

5.1. Non-redundant Arrays

Correlation calibration can handle sufficiently dense non-redundant arrays by grouping the UV plane into tiles, though ignoring cross-tile covariances is sub-optimal.The approximation should remain unbiased when cross-tile covariances are neglected.Source: §5.1 · p. 9

5.2. Solving for Antenna Properties

Normalized χ2 enables comparisons among antenna models, supporting inference of antenna properties such as positions, pointing offsets, and beam deformations.Markov-chain exploration and partitioned covariance inverses are proposed for varying antenna properties efficiently.Source: §5.2 · p. 9

5.3. Bandpass Calibration

Bandpass calibration can model source spectral behavior and foreground cross-frequency structure, including spectral-index and curvature modes, to promote spectrally smooth sky images.The overall spectral index may remain poorly constrained, while higher-order calibration is expected to be robust.Source: §5.3 · pp. 9–10

5.4. Polarization Calibration

Polarization calibration can enforce consistent sky polarization across antennas and use calibrators with known polarization to lock the solution while accounting for diffuse-emission uncertainty.Source: §5.4 · p. 10

6. Conclusions

Correlation calibration accounts for array non-uniformities and partial sky knowledge while remaining computationally feasible for thousand-element arrays. Simulations show improved amplitude reconstruction, with source-position information also substantially improving phase reconstruction.

Amplitude reconstruction improves for arrays with small antenna-position errors, while including positions of the brightest sources significantly improves phase reconstruction.Source: §6 · p. 10
A 2-level sparse matrix description makes the likelihood fast to calculate for thousand-element-class arrays, and canned solver routines converge well.Source: §6 · p. 10
Correlation calibration provides a flexible framework for sufficiently dense non-redundant arrays, array-property reconstruction, bandpass calibration, and polarization calibration.A single-frequency reference implementation is provided and is intended to gain more features over time.Source: §6 · p. 10

A. Assumption of Gaussianity

The calibration formalism assumes a Gaussian sky, but the paper argues that several forms of non-Gaussianity do not bias calibration when modeled or constrained appropriately. Unmodeled sharp power-spectrum features remain a potential concern.

Known phase correlations can improve calibration, while unmodeled phase correlations do not bias the calibration solution.Source: A. Assumption of Gaussianity · p. 11
The simulations show no ill effects from point-source-dominated low-kurtosis skies, whereas an extreme isolated-plane-wave case could violate the similar-visibility assumption.The paper characterizes that failure case as highly contrived and unlike observed foregrounds.Source: A. Assumption of Gaussianity · p. 11
Smooth errors in the input sky power spectrum generally alter weighting rather than bias calibration, across signal- and noise-dominated limits.Source: A. Assumption of Gaussianity · p. 11
Unmodeled large, sharp features in the true power spectrum could negatively affect calibration, although such features are not expected.Source: A. Assumption of Gaussianity · p. 12