Correlation Calibration: A Hybrid Calibration Technique for Radio Interferometric Arrays

Calibrating out per-antenna signal chain effects is an essential step in analyzing radio interferometric data. For drift-scanning arrays, robustly calibrating the data is especially challenging due to the lack of the ability to track a calibration source. Consequently, calibration strategies for drift-scanning arrays are limited by our knowledge of the radio sky at large, as well as the direction-dependent instrument response. In the context of 21 cm cosmology, where small calibration errors can conspire to overwhelm the cosmological signal, it is therefore crucially important to develop calibration strategies that are capable of accurately calibrating the data in the presence of sky or instrument modeling errors. In this paper we present CorrCal, a covariance-based calibration strategy for redundant radio interferometric arrays. CorrCal is a hybrid calibration strategy that leverages the strengths of traditional sky-based calibration and redundant calibration in a computationally efficient framework that is fairly insensitive to modeling errors. We find that the calibration errors from CorrCal are unbiased and far below typical thermal noise thresholds across a wide range of modeling error scenarios. We show that CorrCal is computationally efficient: our implementation is capable of evaluating the likelihood and its gradient in less than a second for 1,000-element class arrays using just a single laptop core. Given CorrCal’s computational efficiency and robustness to modeling errors, we anticipate that it will serve as a useful tool in the analysis of radio interferometric data from current and next-generation experiments targeting the cosmological 21 cm signal.

💡 Research Summary

The paper introduces CorrCal, a covariance‑based hybrid calibration method designed for large, drift‑scan radio interferometric arrays such as HERA, CHORD, and HIRAX. Traditional calibration approaches fall into two categories. Sky‑based calibration relies on accurate sky models and direction‑dependent instrument responses; it works well for steerable dishes but struggles for drift‑scan arrays that lack a bright, tracked calibrator and operate at low frequencies where sky and instrument knowledge is limited. Redundant calibration exploits the fact that baselines with identical geometric separations should measure the same visibility, allowing one to solve for per‑antenna gains without a sky model. However, real arrays are never perfectly redundant, and the method leaves several absolute degeneracies (overall flux scale, overall phase, and tip‑tilt) that must be fixed by an additional absolute calibration step, typically sky‑based again. Both approaches are vulnerable to modeling errors, which can leak into the calibrated data and overwhelm the faint 21 cm cosmological signal.

CorrCal unifies these ideas by treating the data as a set of correlated Gaussian random variables and directly modeling the full baseline‑to‑baseline covariance matrix. The observed complex visibilities are split into real and imaginary parts, forming a 2N‑dimensional real vector d (N = number of baselines). The covariance C(θ) consists of three additive components: (1) thermal noise (diagonal), (2) a term arising from bright point sources that introduces unequal variance in the real and imaginary parts and non‑zero cross‑covariance, and (3) a diffuse sky term that is spatially correlated across baselines. The model parameters θ include per‑antenna complex gains and the intrinsic visibilities of each redundant group.

The likelihood is the multivariate Gaussian L ∝ det(C)^{‑1/2} exp(‑½ dᵀ C^{‑1} d). Maximizing L is equivalent to minimizing the negative log‑likelihood: –log L = log det C + dᵀ C^{‑1} d. The first term regularizes the gain parameters, while the second term forces the model covariance to match the empirical covariance d dᵀ. Taking derivatives yields the familiar result –∂log L/∂C = C^{‑1} – C^{‑1} d dᵀ C^{‑1}, showing that the optimum occurs when C ≈ d dᵀ.

A key insight is that for a redundant array the covariance matrix is extremely sparse: only baselines belonging to the same redundant group have non‑zero off‑diagonal blocks. The authors exploit this sparsity to construct C efficiently, using block‑diagonal structures and automatic differentiation to compute gradients. The implementation employs a quasi‑Newton optimizer (L‑BFGS) and can evaluate both the log‑determinant and the gradient for a 1 000‑antenna array (≈5 × 10⁵ baselines) in under one second on a single laptop core, demonstrating remarkable computational scalability.

To assess robustness, the authors run end‑to‑end simulations where they generate visibilities from a sky model (including point sources and diffuse emission), apply known per‑antenna gains, add thermal noise, and then deliberately introduce modeling errors: mis‑specified source fluxes/positions, incorrect diffuse power spectra, antenna position offsets, and gain non‑linearity. Across a wide range of error amplitudes, CorrCal recovers unbiased gain solutions with residual errors well below the thermal noise level (often an order of magnitude smaller). The sky‑based prior effectively lifts the absolute degeneracies left by pure redundant calibration, while the redundant constraints keep the solution insensitive to sky‑model inaccuracies.

The paper also compares CorrCal to the “unified calibration” framework of Byrne et al. (2021), which similarly combines sky and redundant information but relies on MCMC sampling and does not scale well to thousands of antennas. CorrCal’s likelihood‑maximization approach sidesteps the expensive sampling, achieving near‑real‑time performance without sacrificing accuracy.

Limitations are acknowledged. The current implementation treats each frequency channel and time integration independently; extending to full multi‑frequency, multi‑time calibration will require additional priors (e.g., smoothness constraints) and possibly hierarchical modeling. The Gaussian assumption for the covariance may break down in the presence of strong radio‑frequency interference or highly non‑linear instrument behavior, which would need separate flagging or mitigation steps.

In summary, CorrCal provides a mathematically rigorous, computationally efficient, and robust hybrid calibration technique that leverages both redundancy and sky information. Its open‑source codebase, extensive documentation, and demonstrated performance suggest it will become a valuable tool for current and next‑generation 21 cm experiments, helping to keep calibration‑induced systematics below the stringent thresholds required for detecting the faint cosmological signal.