PaperlayerView exact revision on arXiv

Source-linked AI summary

Faceting for direction-dependent spectral deconvolution

C. Tasse, B. Hugo, M. Mirmont, O. Smirnov, M. Atemkeng, L. Bester, M. J. Hardcastle, R. Lakhoo, S. Perkins, T. Shimwell

arXiv:1712.02078v2astro-ph.IM

TL;DR

Wide-field, wide-band radio imaging must account for direction-dependent effects while estimating sky spectra to produce deep, high-dynamic-range images. The paper presents a faceted framework and two wide-band deconvolution approaches; it reports that the framework accounts for combined imaging effects that existing imagers leave unaccounted for.

  • Problem

    Deep, high-dynamic-range imaging requires accounting for direction-dependent effects while estimating intrinsic continuum-source spectra.

    Source: §1 · p. 1

  • Method

    The paper uses image-plane faceting to combine generic DDE handling with wide-band spectral deconvolution, including optimized matching pursuit and subspace deconvolution.

    Source: §1 · pp. 1–2

  • Results

    The authors report that their imaging and deconvolution framework is the first to account for the combined effects described, including DDEs and decorrelation-related morphological bias.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The framework is intended to handle generic DDEs and externally defined Jones matrices or beam patterns in wide-band imaging and deconvolution.

    Source: §1 · pp. 1–2

    For full polarization, the implementation computes the Stokes I PSF and assumes the other Stokes PSFs are the same, treating leakages as negligible during the minor cycle.

    Source: deconvolution · p. 8

Abstract

from arXiv · show

The new generation of radio interferometers is characterized by high sensitivity, wide fields of view and large fractional bandwidth. To synthesize the deepest images enabled by the high dynamic range of these instruments requires us to take into account the direction-dependent Jones matrices, while estimating the spectral properties of the sky in the imaging and deconvolution algorithms. In this paper we discuss and implement a wide-band wide-field spectral deconvolution framework (DDFacet) based on image plane faceting, that takes into account generic direction-dependent effects. Specifically, we present a wide-field co-planar faceting scheme, and discuss the various effects that need to be taken into account to solve for the deconvolution problem (image plane normalization, position-dependent PSF, etc). We discuss two wide-band spectral deconvolution algorithms based on hybrid matching pursuit and sub-space optimisation respectively. A few interesting technical features incorporated in our imager are discussed, including baseline dependent averaging, which has the effect of improving computing efficiency. The version of DDFacet presented here can account for any externally defined Jones matrices and/or beam patterns.

1. Introduction

Wide fields, large fractional bandwidth, and high sensitivity make direction-dependent effects and spectral estimation central to deep radio imaging. The paper presents a faceted framework that combines generic DDE handling with wide-band spectral deconvolution.

1. Introduction

  • DDEs must be handled in imaging and deconvolution because modern interferometers’ wide fields and high dynamic range are affected by direction-dependent beam, pointing, and ionospheric effects.These effects can include Faraday rotation.Source: §1 · p. 1
  • The framework combines imaging and deconvolution that account for generic DDEs with wide-band spectral estimation using image-plane faceting.It addresses Jones matrices and beam patterns that vary with direction, baseline, time, and frequency.Source: §1 · pp. 1–2
  • Existing imagers can produce biased source morphologies even when decorrelation is minimized and DDEs are corrected, with unresolved sources appearing larger toward the field edge.The paper attributes this to preferential attenuation of higher-order spatial frequencies by the beam response.Source: §1 · p. 2
  • The paper presents two wide-band deconvolution algorithms: an optimized matching-pursuit variation and joint deconvolution over pixel subsets called subspace deconvolution.One subspace approach is implemented using a genetic algorithm.Source: §1 · p. 2
  • The implementation includes baseline-dependent averaging and irregular tessellation, and the paper discusses a simulation and its main results.The introduction identifies these as implementation features and later-paper topics, without reporting their outcomes here.Source: §1 · p. 2

2. Faceting for wide field imaging

The section develops a faceting approach for wide-field imaging on a shared tangential plane, then improves its treatment of noncoplanar phase variation with facet-dependent fitted kernels.

2. Faceting for wide field imaging

  • Polyhedron faceting requires reprojecting facets and correcting overlaps after synthesis, whereas the described alternative places facets on a common tangential plane.The common-plane approach uses w-dependent uv-coordinate transformations and facet-dependent W-projection-type kernels to address phase errors from noncoplanarity.Source: §2 · p. 2

2.1. Formalism for the faceting

  • The formalism models direction-dependent effects with Jones matrices in the RIME and expresses sky directions relative to each facet’s phase center.The Jones matrices represent linear transformations such as beam, ionosphere, and Faraday-rotation effects; the sky coherency is the underlying source term.Source: §2.1 · pp. 2–3
  • The position-dependent phase includes a second exponential analogous to the w-term, whose first-order approximation can be expressed as a uv-coordinate change.The approximation is identified as a first-order Taylor expansion of the facet-relative δn term.Source: §2.1 · p. 3

2.2. Accurate noncoplanar faceting

  • A low-order two-dimensional polynomial fit to δn yields facet-dependent fw-kernels, applied by analogy with W-projection; a blind 3rd–5th-order polynomial works best among the Fig. 1 comparisons.In practice, a 3rd- to 5th-order polynomial is sufficient to represent w-related phase variation accurately across a facet.Source: Figure 1 · p. 3, §2.2 · p. 3
  • The fit’s first-order coefficient is equivalent to a w-dependent uv scaling that removes the main w-related phase gradient and reduces fw-kernel support.The support depends on w, facet diameter, and facet location; for sufficiently small facets it is only marginally larger than the spheroidal-only kernel.Source: §2.2 · p. 3

3. Imaging in the presence of Direction Dependent Effects

This section frames direction-dependent imaging as a linear mapping from a spatio-spectral sky to dirty images and explains how normalization and local PSFs support its treatment as a local convolution.

3. Imaging in the presence of Direction Dependent Effects

  • The section describes the forward and backward imaging steps in linear-algebra terms, including construction of dirty images and associated PSFs from direction-time-frequency-dependent Jones matrices.The passage defines the normalized mapping as a local convolution operator.Source: §3 · p. 3
  • With baseline-time-frequency-dependent Jones effects, the sky-to-dirty-image mapping is generally not a convolution; under approximate baseline-time-frequency independence of Mueller matrices, image-plane correction can make it a local convolution operator.The effective PSF can still vary with direction, so the normalized mapping is treated as a local convolution.Source: §3 · p. 3
  • The notation includes vectors for the true sky and dirty image, a Fourier transform matrix, visibility-selection and Mueller-like matrices, and a direction-dependent transfer function.The table is identified as an overview of mathematical notation; the supplied definitions also specify several vector and operator dimensions.Source: §3 · p. 4, Table 1

3.1. Forward and backward mappings

The paper represents imaging as a linear mapping from sky pixels to visibilities and uses its adjoint to form dirty images. Faceting approximates the ideal mapping with an FFT per facet, while direction-dependent effects generally make the dirty-image response depart from a global convolution.

3.1. Forward and backward mappings

  • The forward mapping stacks visibility samples across baselines and frequencies, while the adjoint maps them back to the image domain to form the dirty image.The sky vector packs four polarization products for each pixel; the sampling matrix selects the four visibility products for a baseline-frequency sample.Source: §3.1 · p. 4, §3.2 · p. 4
  • The facet-based forward operator applies an FFT per facet with a constant direction-dependent effect, approximating the ideal pixel-by-pixel response and becoming exact for single-pixel facets.The ideal mapping applies a unique direction-dependent effect at every pixel, but implementing it directly is computationally prohibitive.Source: §3.1 · p. 4

3.2. Forming the dirty image

  • Without direction-dependent effects, the dirty-image mapping is a weighted sum of convolution operators; with such effects, the PSF generally becomes direction-dependent instead.This matters because minor-cycle deconvolution methods often assume a global convolution, though faceted methods can accommodate a position-dependent PSF.Source: §3.2 · p. 5

3.3. Toeplitz matrices and convolution

The section characterizes convolution through a Toeplitz-like translation invariance in the image-space response. Direction-dependent effects generally break that invariance and make the PSF position-dependent, except for linear phase slopes.

3.3. Toeplitz matrices and convolution

  • A convolution matrix has entries determined by pixel separation, so equal separations produce equal responses; for polarized images, each entry is a 4 × 4 matrix.Each matrix column represents the PSF shifted to the corresponding image pixel.Source: §3.3 · p. 5
  • A direction-independent image attenuation can be factored from the mapping, preserving a constant PSF shape, whereas unaccounted nontrivial effects can change PSF shape across the image.In the attenuation case, the effect can be absorbed into the apparent sky before convolution.Source: §3.3 · p. 5
  • Direction-dependent effects generally make the PSF direction-dependent even when perfectly known and included in the adjoint; linear phase slopes are the stated exception.This also means W-projection and W-stacking, as approximations to the same mapping, still yield a direction-dependent PSF.Source: §3.3 · pp. 5–6

3.4. Loss of local convolution property and nonuniform noise reponse

The local convolution approximation depends on how the Mueller matrices vary across the image. Nonunitary effects also couple the PSF response to the effect at both source and response positions, and the dirty-image thermal noise is spatially nonuniform.

3.4. Loss of local convolution property and nonuniform noise reponse

  • When Mueller matrices vary approximately linearly over a local region, the response is locally a convolution, which facet imaging can represent with a PSF per facet.The direction-dependence of a PSF sidelobe is tied to Mueller-matrix variation over the corresponding pixel-separation scale.Source: §3.4 · p. 6
  • For nonunitary effects, the PSF sidelobe response is coupled to the Mueller matrices at both the source and response positions.The section identifies decoupling the response from the effect at the source position as desirable.Source: §3.4 · p. 6
  • The thermal noise response in the dirty image is spatially nonuniform.The passage derives this from the covariance transformation applied to independent, identically distributed Gaussian visibility noise.Source: §3.4 · p. 6

3.5. Image-plane corrections

An image-plane correction can restore an approximately direction-dependent convolution mapping when the Mueller matrices are approximately baseline-time independent. The correction yields spatially uniform normalized-image covariance, while the implementation uses frequency chunks and a user-defined, ideally small bandwidth.

3.5. Image-plane corrections

  • When Mueller matrices are approximately baseline-time independent at a frequency, an image-plane correction can make the mapping a direction-dependent convolution operator.The assumption is considered reasonable when the data chunk has a sufficiently small fractional bandwidth.Source: §3.5 · p. 6
  • The normalized mapping has PSFs whose columns differ from one another only by a matrix product, and the corrected image can be viewed as a convolution of the apparent, beam-attenuated sky.The mapping is not strictly a convolution matrix, but is referred to as a direction-dependent convolution matrix.Source: §3.5 · p. 6
  • After applying the correction, the normalized image-plane pixel covariance is spatially uniform.In practice, the Mueller blocks are assumed diagonally dominant and reduced to scalar matrices when computing the correction; DDFacet applies it over a user-defined, ideally small bandwidth.Source: §3.5 · p. 6

3.6. Direction-dependent PSFs

Direction-dependent decorrelation and frequency-varying beams make the effective PSF vary across the field, biasing source morphology. DDFacet accounts for these effects by computing a PSF per facet, with a stated approximation in full-polarization deconvolution.

3.6.1. Effect of decorrelation

  • Direction-dependent decorrelation makes the effective PSF vary with baseline coordinates and source distance from the phase center, biasing morphology more at larger distances.The error is minimized by accepting a small decorrelation at the field edge.Source: §3.6.1 · p. 7
  • Applying the decorrelation factor to unit visibilities during PSF computation and to model visibilities in the major-cycle forward step accounts for decorrelation in both cycles.The factor depends on the time and frequency averaging intervals and the baseline’s speed in the uv domain.Source: §3.6.1 · p. 7

3.6.2. Per-facet PSFs

  • Per-facet PSFs account for local convolution changes that remain even after the image-plane transformation.The local PSF is evaluated for a point source in each facet.Source: deconvolution · p. 7, Figure 3 · p. 8
  • In full polarization, the method computes only the Stokes I PSF, assuming the other Stokes PSFs are the same and leakages negligible during the minor cycle.A fully accurate local convolution description would require 16 PSFs; the impact of the approximation on polarization deconvolution is left for future study.Source: deconvolution · p. 8

3.7. In-facet errors

Faceting assumes the Jones matrix is spatially constant within each facet even though direction-dependent effects vary continuously. A smooth image-plane term can partially account for the resulting in-facet flux error.

3.7. In-facet errors

  • Within a facet, the Jones matrix is assumed spatially constant, although direction-dependent effects vary continuously and can change quickly, especially near the beam’s half-power point.A spatially smooth image term can partially account for this effect.Source: §3.7 · p. 8
  • The in-facet flux error is analyzed under an L2-norm constraint, with the Jones matrix represented as direction-independent within the facet.The estimated sky is denoted b xϕ in the passage.Source: §3.7 · p. 8
  • When no direction-dependent effects are applied during deconvolution, fluxes can be corrected with a smooth beam correction in the image domain.The passage identifies this as the traditional radio-astronomy case.Source: §3.7 · p. 8

4. Wideband deconvolution

The wideband framework estimates sky spectra using a spectral cube and two deconvolution approaches: hybrid matching pursuit and joint subspace optimization. HMP fits local spectro-spatial models, while SSD performs joint fits over pixel subsets and iteratively updates the global residual.

4. Wideband deconvolution

  • The framework forms a spectral cube from frequency-chunk dirty images and presents two algorithms for estimating flux densities and intrinsic spectra while taking direction-dependent Jones effects into account.The algorithms use a matching-pursuit variation and a genetic algorithm, respectively.Source: §4 · p. 9

4.1. hmp deconvolution

  • HMP decomposes the spectral cube around its brightest pixel using multiple spatial and spectral components, estimating all coefficients over a pixel set at each iteration.It grids the data once and natively uses direction-dependent residual images and associated PSFs; the authors report HMP-NNLS works best for extended emission.Source: §4.1 · pp. 9–10

4.2.1. Subspace optimisation for deconvolution

  • SSD makes joint deconvolution practical by optimizing a subset of pixels at each iteration rather than inverting the expensive global operator.Bright regions are isolated into pixel groups called islands and deconvolved independently.Source: §4.2.1 · p. 10
  • Neglecting contributions between islands can bias their estimated sky models, while recomputing the global residual and revisiting islands improves estimates over iterations.A bright source’s sidelobe can initially cause a faint source’s flux to be over- or underestimated.Source: §4.2.1 · p. 10
  • In the two-single-pixel-island example with equal flux and cross-contamination, SSD always converges.The convergence conditions are otherwise described as hard to find and dependent on the structure of the operators.Source: §4.2.1 · pp. 10–11

4.2.2. An example of ssd using genetic algorithm

  • SSDGA uses a genetic algorithm for the island fit, evolving spectro-spatial source models with an L2 fitness function.The method is combinatorial and nondeterministic; the models can include fluxes, spectral index, and Gaussian shape parameters.Source: §4.2.2 · p. 11
  • The example residual panels show iterations 0 through 3; SSD jointly deconvolves pixel subsets and fully removes the estimated flux at each iteration.The passage describes SSD as having remarkable convergence properties.Source: Figure 4 · p. 11

5. Implementation, performance and features

The implementation combines parallel imaging with baseline-dependent averaging, sparsification, flexible facet tessellation, and support for direction-dependent gains and primary beams. Its performance strategies target gridding, the dominant computational cost.

5.1. Parallelization

  • A 12× speedup is achieved with 16 processes, while performance scales approximately linearly through 16 physical cores and declines with hyperthreading.Gridding accounts for 94% of the measured wall time.Source: §5.1 · p. 12

5.2. Baseline-dependent averaging

  • Baseline-dependent averaging trades information loss and conditioning for efficiency; smaller facets permit more aggressive averaging, while gridding speedup levels off near 4×.The reported speedup plateau is attributed to memory-access costs, and the savings can partly offset the extra gridding and FFT work from more facets.Source: §5.2 · pp. 12–14

5.3. Sparsification

  • Sparsification randomly omits visibility blocks in early major cycles, accelerating gridding; a typical five-cycle schedule uses factors 100, 100, 30, 10, and 1.The final cycle uses the full data, and the authors report that factors above 100 are not useful once runtime is dominated by I/O.Source: §5.3 · p. 14

5.4. Nonregular tessellation

  • For nonregular calibration directions, ddfacet builds Voronoi facets from supplied nodes and masks facet images when forming residuals.Degridding uses the converse operation.Source: §5.4 · p. 14

5.5. Primary beam models

  • The imager accepts direction-dependent gain solutions or known primary-beam patterns, applying the relevant effects per facet.Primary-beam-only imaging uses square facets; supported beam modes include FITS patterns and a LOFAR station response.Source: §5.5 · pp. 14–15

6. Simulations

The simulations compare direction-independent and direction-dependent imaging with narrow- and wideband deconvolution. The reported results show that HMP and SSD reduce flux-density and spectral-index errors, with SSD doing slightly better on spectral index.

Simulation setup

  • The test uses a simulated LOFAR LBA observation with 630 baselines, five channels from 30 to 70 MHz, and a bright 10^4 Jy off-axis source.The simulated sky contains about 100 point sources with spectral indices drawn uniformly from −1 to 1.Source: §6 · p. 15

Evaluation modes

  • The runs compare direction-independent and direction-dependent faceting and narrowband imaging with wideband HMP or SSD; the wideband residual cube has three channels.The simulation applies time-, frequency-, and direction-dependent Jones matrices together with the LOFAR station beam.Source: §6 · p. 15

Results

  • HMP and SSD wideband deconvolution reduce flux-density and spectral-index errors; both recover spectral indices better than direction-independent wideband imaging, with SSD slightly better.The comparisons assess error as a function of distance from the beam center.Source: §6 · p. 15

7. Overview and future work

The overview presents DDFacet as a wideband, wide-field deconvolution framework that handles generic direction-dependent effects. It combines coplanar faceting, direction-dependent PSFs, and two spectral deconvolution algorithms, alongside implementation features and planned extensions.

Framework

  • The framework uses a wide-field coplanar faceting scheme with facet-dependent w-kernels, direction-dependent PSFs, and explicit treatment of time- and bandwidth-averaging effects.It supports spatially discrete, time-frequency-baseline-direction-dependent full-polarization Jones matrices.Source: §7 · p. 15

Wideband algorithms

  • The paper introduces hybrid matching pursuit and Subspace Deconvolution for wideband spectral deconvolution; SSD jointly deconvolves pixel subsets as independent islands.A genetic algorithm performs the nonlinear optimization step described for SSD.Source: §7 · p. 15

Simulation figures

  • The simulation figures compare restored images with no beam or wideband correction, with DDE correction alone, and with SSDGA; the flux and spectral-index error figures plot errors against distance from the beam center.The figure descriptions identify gray comparison marks for estimates made without DDE correction.Source: Figure 8 · p. 16, §7 · p. 16, Figure 10 · p. 17

Technical features

  • The implementation supports regular and Voronoi facets, on-the-fly baseline-dependent averaging, sparsification, and asynchronous multiprocessing.The authors report a factor of several speedup for gridding and degridding from baseline-dependent averaging.Source: §7 · p. 16

Scope and future work

  • The reported version supports full-polarization deconvolution and externally defined Jones matrices or beam patterns, and has been tested with LOFAR, VLA, MeerKAT AR1, and ATCA data.The overview also identifies further development areas, including extended-emission bases, other SSD algorithms, GPU gridders, and distributed implementation.Source: §7 · pp. 16–17
Loading 1712.02078v2…