Beam-Crossing Technique Calibration

• Beam-Crossing Technique is a calibration method for fixed, wide-field drift-scan arrays that uses crossing points of drifting source tracks to decouple beam gain from source flux.
• It transforms an underconstrained multiplicative problem into an overdetermined linear inversion by leveraging 180° rotational symmetry and HEALPix discretization.
• Simulation studies and PAPER observations demonstrate that the method achieves beam and flux accuracies within 10–15%, effectively mitigating sidelobe contamination and sparse calibrator challenges.

Searching arXiv for the specified paper and closely related work on primary-beam calibration for drift-scanning arrays. I’ll look up the target paper and related calibration papers so the article can be grounded in the arXiv record. The beam-crossing technique is a primary-beam calibration method for fixed, wide-field, drift-scanning antenna elements in which the sky drifts through a stationary primary beam rather than being tracked mechanically. In this setting, calibration means solving for the direction-dependent gain $B(\theta,\phi)$ or $B(l,m)$ of each antenna element, while simultaneously inferring source flux densities from drift-scan measurements. The central idea is to construct an interrelated network of source “crossing points,” namely sky locations at which multiple source tracks sample the same beam-response value. With the single assumption that the beam has $180^\circ$ rotational symmetry about the boresight, the method converts otherwise underconstrained drift-scan measurements into an over-determined inversion for both the beam model and the source catalog, referenced to one calibrator source (Pober et al., 2011).

1. Observing problem and calibration objective

In a conventional tracking dish, beam calibration can be carried out by pointing at a bright calibrator and dithering around it to map the gain pattern. That procedure is not available to a drift-scanning array of fixed elements such as PAPER, MWA, or LOFAR, because the antennas do not physically track sources; instead, the sky drifts through the beam. The beam-crossing technique was developed specifically to address this incompatibility between drift-scan operation and standard beam-calibration routines (Pober et al., 2011).

The unknown to be estimated is the primary-beam response $B(\theta,\phi)$, or equivalently $B_i$ on a discretized sky, together with the intrinsic flux densities $S_j$ of compact sources. For a source $j$ drifting through direction $(\theta_j(t),\phi_j(t))$, the perceived flux is modeled as

$V_j(t)=B(\theta_j(t),\phi_j(t))\cdot S_j + n_j(t),$

where $n_j(t)$ denotes measurement noise. In discretized form, for the $B(l,m)$0 sample of source $B(l,m)$1 falling in pixel $B(l,m)$2, the measurement equation is

$B(l,m)$3

This formulation exposes the core degeneracy: if a single source of unknown flux is observed along one drift track, the beam response along that track cannot be separated from the source flux without an external reference. The calibration problem is therefore not merely one of fitting a beam shape, but of breaking a multiplicative ambiguity between beam gain and source brightness.

The method is motivated by three practical difficulties stated explicitly for low-frequency, wide-field drift-scan arrays: the inability to “point and wiggle” around a calibrator, the complicated and hard-to-parametrize beam shapes of dipoles or tiles, and the sparsity of bright calibrators with accurate source spectra. A plausible implication is that direct parametric beam fitting is often less robust than an inversion strategy that leverages repeated sampling of common beam values by multiple sources.

2. Source tracks and crossing-point geometry

The defining construct of the method is the source crossing point. Each compact source produces a drift track across the stationary beam once per sidereal day. Under the assumption of $B(l,m)$4 rotational symmetry,

$B(l,m)$5

each real track has a mirrored counterpart rotated by $B(l,m)$6 in azimuth. Two distinct sources therefore generate four tracks in total, and these tracks pairwise cross at discrete locations on the sky (Pober et al., 2011).

At a crossing point $B(l,m)$7, two different sources $B(l,m)$8 and $B(l,m)$9 sample the same beam value $180^\circ$0, yielding

$180^\circ$1

This gives two equations involving the common beam response and the two source fluxes, up to an overall scale. When many sources are linked through many such crossings, the result is an over-determined network that constrains both the beam and the relative source fluxes. The technique therefore replaces the need for a large set of independently characterized calibrators with a graph-like structure of interlocking beam samples.

The geometry is important because the method does not require arbitrary source trajectories to intersect directly. The rotational-symmetry assumption effectively doubles the available track set through mirrored trajectories and halves the number of independent beam pixels to be solved. This is what allows significant beam coverage with only a few tens of sources. In the formulation summarized for the method, the final system is referenced to a single calibrator source, so absolute flux scaling remains anchored even though most source spectra are not assumed known a priori.

A common misunderstanding is to equate “beam-crossing” with crossing-angle schemes in colliders. In this context, the term refers instead to the crossing of source tracks through equal-response regions of a stationary radio primary beam. The relevant crossings are in beam-response space and on the celestial sphere, not at an accelerator interaction point.

3. Discretization and linearized inverse problem

The practical implementation begins by extracting source tracks from raw visibilities. The summary specifies two routes: a delay/delay-rate filter (DDR) or an image-based fit. These produce time-binned measurements $180^\circ$2 for each source $180^\circ$3. The sky is then gridded, typically with HEALPix, so that all measurements falling into the same pixel $180^\circ$4 are treated as sampling the same beam response $180^\circ$5. Averaging within each pixel/time bin forms quantities

$180^\circ$6

The crossing relations implied by the sidereal geometry and $180^\circ$7 symmetry are then used to identify which $180^\circ$8 measurements correspond to the same beam pixel (Pober et al., 2011).

Because the basic model is bilinear in $180^\circ$9 and $B(\theta,\phi)$0, the method proceeds by log-linearization:

$B(\theta,\phi)$1

where $B(\theta,\phi)$2 contains the effects of noise and log-domain bias. The corresponding weighted linear system is written as

$B(\theta,\phi)$3

with data vector $B(\theta,\phi)$4 and unknown vector

$B(\theta,\phi)$5

The design matrix $B(\theta,\phi)$6 contains one nonzero entry linking each datum to its beam cell and one linking it to its source. The diagonal weighting matrix satisfies $B(\theta,\phi)$7, so the inversion is explicitly inverse-variance weighted.

This formulation is structurally significant because it transforms a multiplicative calibration problem into a sparse linear least-squares problem. The price of this simplification is the introduction of a small log-domain bias. The summary notes that simulations show this bias to be sub-dominant, at the few percent level, relative to sidelobe contamination in a small array. That observation suggests that the dominant error budget in early implementations is not the linearization itself but imperfect source extraction from raw visibilities.

4. Solution strategy, covariance, and regularization

The unknown vector is solved with weighted least squares or SVD. The estimator is given as

$B(\theta,\phi)$8

In practice, a robust routine such as singular-value decomposition with truncation is used to avoid instabilities (Pober et al., 2011).

The formal covariance of the solution is

$B(\theta,\phi)$9

from which uncertainties in $B_i$0 and $B_i$1 may be propagated back to uncertainties in $B_i$2 and $B_i$3. This is one of the important technical features of the method: it is not only a reconstruction procedure but also an uncertainty-propagating inversion.

The least-squares solution obtained at crossing pixels is not the end product. After solving for the source fluxes, all source tracks are divided by the fitted $B_i$4 values to obtain beam samples along every track, not only at the crossing locations. Those samples are then “CLEANed” in Fourier space, in Högbom style, to fill gaps and produce a continuous $B_i$5. In addition, the solved beam pixels can be regularized by exploiting the fact that electromagnetic-model beams are spectrally and spatially smooth. The specific regularization described is filtering in the Fourier or spherical-harmonic domain, retaining only low-order modes so as to suppress unphysical high-frequency ripples induced by sidelobe leakage.

The method therefore combines three layers of inference: a discrete algebraic solve on crossing points, a flux-normalized extension of the beam solution along entire source tracks, and a smoothness-enforcing interpolation or deconvolution step. This suggests that the technique is best understood not as a single estimator but as a calibration pipeline with a well-defined inversion core.

5. Simulation studies and empirical performance

Three levels of validation are described: simplified simulated tracks, full visibility simulations, and real observations with PAPER. In the simplified simulations, a toy beam such as an elliptical, rotated pattern was combined with $B_i$6–$B_i$7 sources of known $B_i$8, generating measurements of the form $B_i$9 plus Gaussian noise. Running the crossing-point pipeline on these data recovered the beam to $S_j$0 and the source fluxes to $S_j$1 for realistic noise levels of $S_j$2 Jy r.m.s. (Pober et al., 2011).

The full visibility simulations were more demanding. They used a $S_j$3-element PAPER layout with baselines $S_j$4–$S_j$5 m over $S_j$6–$S_j$7 MHz, including thermal noise, the Sun, and sidelobes from $S_j$8 point sources. DDR-filter extraction yielded perceived source fluxes contaminated by sidelobes at up to $S_j$9. Applying the same least-squares solve and smoothness prior recovered the beam to $j$0 of the true beam, improving to $j$1 after Fourier-domain filtering, while source fluxes had average errors of approximately $j$2.

These results are important because they separate the intrinsic behavior of the inversion from the complications of realistic interferometric data. The idealized experiment demonstrates that the crossing-point algebra itself is accurate, while the full simulation shows the extent to which sidelobe contamination and source-extraction errors propagate into the beam solution. A plausible implication is that improvements in source-isolation methods would directly improve beam-crossing performance without altering the basic inversion framework.

6. Application to PAPER observations and resulting products

The observational demonstration used $j$3 h of data from Green Bank PAPER-12, restricted to the north–south polarization, over $j$4–$j$5 MHz in $j$6 kHz channels. Pre-processing consisted of RFI excision, bandpass calibration, a one-day fringe fit on Cygnus A, and gain stabilization using load (“gain-o-meter”) temperature. DDR filters together with MCMC self-calibration produced perceived source tracks $j$7 for $j$8 bright sources, with Cygnus A as the flux anchor (Pober et al., 2011).

For the beam solve, the data were gridded to HEALPix pixels of approximately $j$9 side. Under the $(\theta_j(t),\phi_j(t))$0 symmetry assumption, approximately $(\theta_j(t),\phi_j(t))$1 crossing points were identified among the $(\theta_j(t),\phi_j(t))$2 sources. Weighted least squares was then used to solve for $(\theta_j(t),\phi_j(t))$3 and $(\theta_j(t),\phi_j(t))$4, after which all tracks were divided by the fitted source fluxes to generate beam samples across the accessible sky. The final interpolation step combined CLEAN with smoothness enforcement.

The resulting beam model was smooth and nearly symmetric around zenith, with a half-power width of approximately $(\theta_j(t),\phi_j(t))$5, and it agreed with electromagnetic simulations to within $(\theta_j(t),\phi_j(t))$6. The same inversion produced a source catalog of $(\theta_j(t),\phi_j(t))$7 measured flux densities at approximately $(\theta_j(t),\phi_j(t))$8 MHz, tied to the Baars et al. (1977) scale via Cygnus A, with typical uncertainties of $(\theta_j(t),\phi_j(t))$9.

The PAPER application demonstrates that the technique yields two products simultaneously: an empirical primary-beam model and a low-frequency flux catalog internally consistent with that beam model. This dual output is central to the method’s utility, because drift-scan beam calibration and source characterization are treated as a coupled inverse problem rather than as separate preprocessing stages.

7. Assumptions, limitations, and methodological significance

The method is deliberately based on minimal assumptions. The summary identifies only two such assumptions in the overall formulation: $V_j(t)=B(\theta_j(t),\phi_j(t))\cdot S_j + n_j(t),$0 rotational symmetry and zero response below the horizon (Pober et al., 2011). This economy of assumptions is methodologically important because it avoids dependence on a rigid parametric beam model in regimes where dipole or tile beams are difficult to parametrize accurately.

Its limitations are equally explicit. First, the inversion depends on sufficient sky coverage by a network of tens of sources; sparse coverage leaves parts of the beam weakly constrained. Second, the log-linear formulation introduces a small bias. Third, sidelobe contamination can dominate the error budget in small arrays, and the full-visibility simulations show that source extraction can be contaminated at the $V_j(t)=B(\theta_j(t),\phi_j(t))\cdot S_j + n_j(t),$1 level before regularization. Fourth, the smoothness prior and Fourier-domain CLEAN are not merely cosmetic post-processing; they are required to suppress unphysical ripples and to interpolate between sampled tracks.

A second misconception is that beam calibration for drift-scan arrays necessarily requires a large sample of well-characterized calibrators. The beam-crossing technique explicitly circumvents that requirement by solving source flux densities relative to a single anchor source. Another misconception is that drift-scan geometry precludes high-fidelity empirical beam measurement; the reported PAPER results, together with the simulations, show that beam and flux accuracies at the $V_j(t)=B(\theta_j(t),\phi_j(t))\cdot S_j + n_j(t),$2 level are achievable when the crossing network is sufficiently interlocked and the beam is regularized as a smooth function.

In significance, the technique provides a way to exploit drift-scan geometry rather than treat it as an obstacle. By turning repeated source transits into a coupled system of beam pixels and source fluxes, it reframes primary-beam calibration as a network inversion problem. That perspective is particularly well matched to wide-field, low-frequency interferometers for which beam complexity, fixed pointing, and limited calibrator catalogs make conventional calibration strategies ineffective.