VLA imaging of the XMM-LSS / VIDEO deep field at 1-2 GHz
I. Heywood, C. L. Hale, M. J. Jarvis, S. Makhathini, J. A. Peters, M. L. L. Sebokolodi, O. M. Smirnov
MMNRAS in press, 1–13 (2020) Preprint 16th June 2020 Compiled using MNRAS L A TEX style file v3.0
VLA imaging of the XMM-LSS / VIDEO deep field at 1–2 GHz
I. Heywood , , (cid:63) , C. L. Hale , , M. J. Jarvis , , S. Makhathini , , J. A. Peters ,M. L. L. Sebokolodi , and O. M. Smirnov , Astrophysics, Department of Physics, University of Oxford, Keble Road, Oxford, OX1 3RH, UK Department of Physics and Electronics, Rhodes University, PO Box 94, Makhanda, 6140, South Africa South African Radio Astronomy Observatory, 2 Fir Street, Black River Park, Observatory, Cape Town, 7925, South Africa CSIRO Astronomy and Space Science, PO Box 1130, Bentley, WA 6102, Australia Physics Department, University of the Western Cape, Private Bag X17, Bellville, 7535, South Africa National Radio Astronomy Observatory, PO Box 0, Soccoro, NM 87801, USA
Accepted 2020 June 10. Received 2020 June 09; in original form 2019 June 25
ABSTRACT
Modern radio telescopes are routinely reaching depths where normal starforming galaxiesare the dominant observed population. Realising the potential of radio as a tracer of starformation and black hole activity over cosmic time involves achieving such depths overrepresentative volumes, with radio forming part of a larger multiwavelength campaign. Inpursuit of this we used the Karl G. Jansky Very Large Array (VLA) to image ∼ of theVIDEO/XMM-LSS extragalactic deep field at 1–2 GHz. We achieve a median depth of 16 µ Jybeam − with an angular resolution of 4.5 (cid:48)(cid:48) . Comparisons with existing radio observations ofXMM-LSS showcase the improved survey speed of the upgraded VLA: we cover 2.5 timesthe area and increase the depth by ∼
20% in 40% of the time. Direction-dependent calibrationand wide-field imaging were required to suppress the error patterns from off-axis sources ofeven modest brightness. We derive a catalogue containing 5,762 sources from the final mosaic.Sub-band imaging provides in-band spectral indices for 3,458 (60%) sources, with the averagespectrum becoming flatter than the canonical synchrotron slope below 1 mJy. Positional andflux-density accuracy of the observations, and the differential source counts are in excellentagreement with those of existing measurements. A public release of the images and catalogueaccompanies this article.
Key words: radio continuum: galaxies – techniques: interferometric – astronomical databases: surveys
Traditionally, the focus of radio continuum surveys has been on find-ing and studying the radio-loud active galactic nuclei (AGN) popu-lations over the history of the Universe, and more recently the impactthat such radio-powerful objects have on their host galaxy and im-mediate environment. However, this focus is beginning to changeas we move towards ever-deeper radio continuum surveys. This isdue to the fact that at S . < ∼ µ Jy the composition of the radiosource population begins to change from being AGN-dominated tobeing composed predominantly of star-forming galaxies (SFGs) andradio-quiet AGN (e.g. Jarvis & Rawlings 2004; White et al. 2015).Deep radio-continuum surveys are therefore opening up a newwindow on what is usually considered the ‘normal’ galaxy pop-ulation. Radio observations of such star-forming galaxies are im- (cid:63)
Contact author: [email protected] portant, as they have the potential to provide a relatively unbiasedview of the time-averaged star-formation rate (SFR). This is due tothe fact that supernovae are co-located with regions of massive starformation, and when electrons traverse their ageing shock frontsthey decelerate rapidly, producing synchrotron radiation (Condon1992).Thus, over the past decade there have been many efforts toobtain deep radio continuum data over representative volumes ofthe Universe. Leading the way was the original VLA-COSMOSsurvey (Schinnerer et al. 2007) which used the VLA at L-band( ∼ . © 2020 The Authors a r X i v : . [ a s t r o - ph . GA ] J un Heywood et al.
The key aims of surveys such as those listed above, as well asfuture approved and proposed surveys with the SKA and its pre-cursors (e.g. Jarvis et al. 2016; Norris et al. 2011; Jarvis et al. 2015;Prandoni & Seymour 2015), are to understand the link between AGNactivity and the host galaxy properties, to trace the star-formationhistory of the Universe and the evolution in the star-formation mainsequence (e.g. Noeske et al. 2007; Daddi et al. 2007; Whitakeret al. 2012; Johnston et al. 2015). Radio observations are free fromobscuration by dust (cf. ultraviolet tracers), and are generally notconfused and/or suffer from low-angular resolution (cf.
Herschel and SCUBA-2 surveys).In order to achieve these goals the radio observations musttarget fields in the sky with excellent multi-wavelength coverage.This is because the radio data alone do not provide any informationon the stellar mass of the galaxies or their redshifts, although thismay soon change at least at z < . ∼ α ) image and a sourcecatalogue generated using the ProFound software. We comparethese observations with previous radio surveys in Section 4 andpresent the source counts derived from the survey. In Section 5 webriefly summarise our findings and provide a link from where thedata products may be freely downloaded. The observations were conducted using the VLA in B-configuration. A single 1.5 h Scheduling Block (SB) was submittedfor each of the 32 pointings, containing the necessary calibratorscans, as well as scans of the science target. The on-source observa-tion time for each target pointing was 67.5 m, with 3 s per correlatorintegration. The correlator was configured in standard wide-bandcontinuum mode, with 0.994–2.018 GHz of spectral coverage splitup into 16 ×
64 MHz spectral windows (SPWs) for a total of 1,024channels.Data were delivered from the observatory in a CASA format We adopt the convention that the flux density S is proportional to theobserving frequency ν according to: S ∝ ν α . Project codes: 13B-308 and 15A-477, corresponding to observations tak-ing place between 28 November – 24 December 2013 (all pointings except18), and on 21 April 2015 (pointing 18).
Measurement Set (MS) containing visibility data for the target andthe primary and secondary calibrators. The primary calibrator was3C147, and was used to determine the absolute flux density scaleusing the models derived by Perley & Butler (2013) and solve for thebandpass shape. The secondary calibrator was J0217+0144 and thiswas used to determine time-dependent complex gain corrections.Visits to this source were somewhat infrequent, reasoning that self-calibration of the target data would be both feasible and necessary.The description of the steps that follow were applied to each SB in-dividually. The referenced calibrator corrections were derived andapplied using the NRAO CASA pipeline . The pipeline also appliedHanning smoothing to the data, and made a first pass of automaticradio frequency interference (RFI) excision. Following this we ex-amined the scans of the calibrator sources for any remaining RFI.Any gross features were flagged, and the pipeline was re-run fromscratch. RFI was rife, with SPWs 5 and 6 (1.314–1.412 GHz) lost inmany pointings, and spectral windows 8 and 9 (1.506–1.634 GHz)discarded outright for all 32 pointings. Once the reference calibra-tion steps were complete, the corrected visibilities for the target fieldwere split into a single source MS ready for imaging and furthercalibration.All imaging was performed using using the wsclean package(Offringa et al. 2014), which makes use of the efficient w-snapshotalgorithm (Humphreys & Cornwell 2011) to correct for the effects ofusing non-coplanar arrays to conduct wide-field imaging. Imagingparameters were the same for each run, using 12,000 × (cid:48)(cid:48) to cover 2.33 ◦ × ◦ . Images of this sizewere necessary to deconvolve and model confusing sources in thesidelobes of the primary beam. Briggs (1995) weighting was usedin order to suppress the sidelobes in the point spread function (PSF),with the robustness parameter set to 0.0. The frequency dependenceof the sky brightness distribution was captured by deconvolving in4 ×
256 MHz sub-bands. When searching for peaks of emissionduring the minor cycle wsclean uses the full-band image, howeverdeconvolution takes place in each of the sub-bands independently.A polynomial with a user-defined order (in this case 3) is fitted to theclean components and inverted into a visibility model for subtractionduring the major cycle. At the end of the cleaning process thefinal model is (optionally) inverted and written to the
MODEL_DATA column of the MS for use in self-calibration.An initial imaging run was performed with unconstraineddeconvolution terminating at 50,000 iterations. We then used thePyBDSF (Mohan & Rafferty 2015) source finder to locate regionsof significant emission in the image. PyBDSF works by estimat-ing the spatial variation in the background noise level, and thenidentifying peaks that are some threshold (in this case 5) times thebackground. Once these are identified a flood fill operation takesplace down to a secondary threshold (in this case 3) times the back-ground. These islands of emission are then decomposed into groupsof point and Gaussian components.The resulting catalogue was manually examined, and spuriousfeatures were removed. Essentially all of such features were associ-ated with residual PSF-like structures in the image which were notsuccessfully deconvolved due to calibration deficiencies. The pos-itions and shapes of the components in the pruned catalogue werewritten into a blank FITS image for subsequent use as a Booleancleaning mask.Imaging was repeated with the mask being used to constrainthe locations of the deconvolution. Having examined the behaviour https://science.nrao.edu/facilities/vla/data-processing/pipeline MNRAS in press, 1–13 (2020)
LA imaging of the XMM-LSS / VIDEO deep field of the value of the peak residual during the initial cleaning runsthe termination threshold was set at 35,000 iterations. Followingthis imaging run the spectral visibility model derived from thepolynomial fits to the clean component model was used to determinea set of complex gain corrections for both LL and RR polarisations(only the diagonal terms the G-Jones matrix) via self-calibration.All calibration steps were performed using the MeqTrees package(Noordam & Smirnov 2010) using its implementation of the fastStefCal solver (Salvini & Wijnholds 2014).Phase-only solutions were derived for every 300 s ×
64 MHztile of data, the frequency interval corresponding to each SPW.Solutions were forbidden from extending over the gaps in the datawhere the secondary calibrator observations had taken place. Thecalibrated data were re-imaged and the cleaning mask was refinedat this stage, necessary for example if the reduced error patternsaround bright sources following self-calibration revealed new genu-ine emission. This procedure produced acceptable images for 5/ 32 pointings. Amplitude and phase self-calibration, with addi-tional direction-dependent calibration was required for the rest ofthe data. Traditional direction-independent self-calibration provedinadequate for removing the error patterns associated with off-axissources of even modest brightness. Off-axis sources are subjectedto time, frequency and direction-dependent complex gain perturb-ations due to the antenna primary beam response, and effect whichis exacerbated by the large fractional bandwidth of the VLA, andpotentially by second-order effects such as antenna pointing errorsor wind loading (e.g. Smirnov 2011b; Heywood et al. 2013a).Direction-dependent calibration was performed using the dif-ferential gains method (Smirnov 2011a). This is an inverse-modelling approach that can be thought of as a form of simultaneouspeeling (e.g. Noordam 2004) that does not require an iterative ap-proach, and is less prone to instabilities in the presence of confusingsources of similar brightness. Antenna-based complex gain terms(G) are derived as per traditional self-cal, based on an all-inclusivesky model, however additional solvable complex gain terms are de-rived for a subset of ‘problem’ sources in the model using a longersolution interval. These additional differential gain (dE) terms arefixed to unity for all other sources. A hybrid sky model was construc-ted for this purpose. Having identified the positions of the sources towhich the dEs are to be applied, component-based models for thesesources were derived by using PyBDSF to characterise the emissionat those positions in each of the four sub-band images produced bywsclean. The components at these positions in the model imageswere then masked, and visibilities based on these model imageswith the problem sources removed were written to the
MODEL_DATA column of the MS by running wsclean in predict mode. This stepmakes the process computationally cheaper, as computation of thedirection- independent portion of the sky model is a one-time opera-tion. MeqTrees was then used to solve for G and dE terms based onthe pre-computed model, plus the component models which werepredicted on the fly.Solutions were derived on a per-SPW basis, with the sameboundary conditions used for the phase-only solutions, and using therelevant component model for the dE terms. The solution intervalswere 162 and 324 s for the G- and dE-terms respectively. Thesewere extended by a factor of 2 for SPWs 10–12 inclusive, and by afactor of 3 for SPWs 13–15 inclusive, in order to boost the signalto noise in the solutions. Note that SPWs are zero-indexed. Thecleaning masks were again refined at this stage, if required.Figure 1 shows the improvements in image quality gained byapplying the directional calibration. The four rows correspond tofour different sources in pointing 1. Top to bottom, the sources
Figure 1.
Three generations of calibration: the results of applying differentcalibration schemes are shown above for four sources (one per row) selectedfrom pointing number 1. The left hand column shows the sources as imagedfollowing execution of the standard VLA pipeline which applies the refer-enced calibration. The central column shows the subsequent improvementafforded by traditional (amplitude and phase) self-calibration, and the righthand column shows the final image achieved with self-calibration with ad-ditional solvable differential gain terms applied to the four sources. Thesesources are ordered top to bottom by increasing radial separation from thephase centre, and all four are located somewhere between the flank of themain lobe of the primary beam and the first sidelobe, depending on thefrequency. Note the degradation in the performance of self-calibration withincreasing distance from the phase centre, where the primary beam relateddirection-dependent effects can be expected to become more pronounced.The colour scale in this image saturates black at ± − , withwhite being zero. are presented in increasing distance from the phase centre. Allfour sources are situated between the edge of the main lobe of theprimary beam and the first sidelobe, depending on the frequency.Even with the primary beam attenuation these sources are of com-parable apparent brightness, and are amongst the brightest sourcesin the image. The first column in Figure 1 shows the deconvolved image produced following the application of the referenced calib-ration by the VLA pipeline. The second column shows the result ofapplying (amplitude and phase) self-calibration based on a modelderived from the spectral component fitting performed by wsclean.The third column shows the final image following the application ofdifferential gain terms to these four sources. Note that the solutionintervals for the G terms are the same for the second and third scen-arios. The additional dE terms are required here to account for thediffering time, frequency and direction-dependent corruptions that MNRAS in press, 1–13 (2020)
Heywood et al. these sources are subjected to due to their locations in the primarybeam.Once satisfactory calibration had been performed, the datawere subjected to the final imaging procedure. This made use of thefinal cleaning masks, with an initial constrained clean, followed bya shallower (10,000–20,000 iterations, depending on the presenceof low-level extended structure) blind clean of the residual mapwith the mask removed. A thorough investigation of potential cleanbias effects on broadband VLA snapshot data has been made byHeywood et al. (2016). Briefly, clean (or snapshot) bias is a system-atic error in the photometry measurements that is dependent on thebrightness of the source being measured (e.g. Becker et al. 1995;Condon 1997; Huynh et al. 2005). It is thought to be related to theuse of the clean algorithm for deconvolution, exacerbated by thestrong linear features in the PSF of the VLA, and can even affectsources below the noise floor of the survey (White et al. 2007). Thelarge-scale simulation conducted by Heywood et al. (2016) showedthat contraining the deconvolution using masks significantly lessensthe effect, but we can expect clean bias to exist at the few percentlevel close to the catalogue threshold, rapidly becoming negligiblefor brighter sources. Since a 5 σ source will be subject to statisticalfluctuations at the 20% level by definition, no corrections have beenmade to the catalogue for these comparatively small clean bias ef-fects. However, persons extracting photometric measurements closeto the noise floor of the survey should be mindful that clean biasmay be present at the tens of percent level, comparable to the noise-induced statistical uncertainties.A circular 2D Gaussian restoring beam with a FWHM of4.5 (cid:48)(cid:48) was applied to each image. This is marginally broader thanthe generally achievable angular resolution afforded by using thefitted restoring beam, however it accounts for the variations inducedin the PSF by the dynamic scheduling of the observations, and im-parts a desirable uniformity to the mosaicked image. Image-planeprimary beam corrections were applied to the final full-band images,as well as each of the four sub-band images, by dividing each by amodel image of the VLA primary beam computed at the appropriatefrequency, and masked beyond the 30% value. Linear mosaics ofthe 32 images were made using the Montage package, with eachpointing weighted by the assumed spatial noise variance image, inthis case assumed to be represented by the square of the primarybeam pattern. The total intensity mosaic is shown in the upper panel of Figure2, with each of the 32 pointing positions marked. The greyscale islinear and runs from −
20 (white) to 20 µ Jy beam − (black). Thelower panel shows a 1.2 × The frequency behaviour of the sky brightness distribution I ( ν ) ismost commonly modelled as a power law in frequency I ( ν ) = I (cid:18) νν (cid:19) α (1) http://montage.ipac.caltech.edu/ where I is the brightness at reference frequency ν , and the expo-nent α is the spectral index. For most sources over our frequencyrange this is a reasonable assumption. Expressing this in log-spacegivesln I ( ν ) = ln I + α ln (cid:18) νν (cid:19) . (2)Defining x = ln νν (3)and y = ln I ( ν ) (4)allows us to compute the spectral index α from N multi-frequencybrightness measurements according to α = (cid:205) i ( x i − ¯ x )( y i − ¯ y ) (cid:205) i ( x i − ¯ x ) (5)and with standard deviation σ α = (cid:115) (cid:205) i ( y i − ¯ y ) N , (6)where ¯ x and ¯ y are the mean values of x and y .Sub-band imaging of the final calibrated data is used in orderto produce the multi-frequency brightness measurements requiredto produce an in-band spectral index image of the survey area.The 1–2 GHz band is divided up into three sections. Since SPWs8 and 9 are discarded due to RFI in all of the 32 pointings, theLOW, MED and HIGH sub-bands are formed from SPWs 0–3, 4–7 and 10–15 inclusive. These correspond to frequency ranges of0.994–1.25, 1.25–1.506 and 1.634–2.018 GHz, and approximatelyequivalent fractional bandwidths of 23%, 19% and 21%. Each sub-band mosaic is formed in the same way as the full-band mosaicdescribed in Section 3.1, with the primary beam correction andmosaic weighting functions set by patterns appropriate to the centralfrequency of the sub-band. The LOW and MED mosaics are croppedto only include the high sensitivity region of HIGH, and the threeimages are placed into a cube with three frequency planes. The pixelsin this cube are masked below 100 µ Jy beam − , corresponding toapproximately 3–4 σ for a single sub-band, and following this apixel-wise linear fit in log-flux/log-frequency space is performed.The best fitting gradients to each three-point spectrum are thenrecorded as the value of spectral index ( α ) at that position, and thestandard deviation in the measurements is recorded as an estimateof the spectral index error, as per Equation 6. The end products ofthis process are maps of the spectral index and spectral index error,which we make further use of when constructing the componentcatalogue in Section 3.3. The package ProFound (Robotham et al. 2018) was used to gener-ate an associated source catalogue from the total intensity mosaic.Although designed for optical/near-IR surveys, ProFound has beenshown to be able to successfully model radio emission (Hale et al.2019a) for sources of different morphologies. As ProFound doesattempt to fit to any particular morphology (e.g 2D Gaussians),complex morphologies (e.g. AGN with extended jets) may be morefaithfully modelled.To extract the source catalogue, the method of Hale et al.(2019a) is followed. We use a skycut value of 3.5, which only
MNRAS in press, 1–13 (2020)
LA imaging of the XMM-LSS / VIDEO deep field Figure 2.
Total intensity image formed from a linear mosaic of the 32 primary beam-corrected images (upper panel), the locations of which are indicated. Thegrey scale is linear and runs from −
20 (white) to 200 µ Jy beam − (black). A 1.2 × includes pixels that have a value of 3 . × the sky RMS value at thatpixel within a source segment. The segment defines all the pixels ofa source that contribute to the model for the source. As the sourcedensity does not approach that of the classical confusion limit, the groupstats=TRUE setting is used to force neighbouring segmentsthat share a segment boundary to be combined into a single source.This is especially important for resolved extended sources that for example have connected lobe emission, and ensures that (providedthe emission is connected) these sources can be identified as asingle source.Following the method of Hale et al. (2019a) we apply a (restor-ing) beam correction to ensure that emission within the wings ofthe source (especially for faint point-like sources) is not missed. To MNRAS in press, 1–13 (2020)
Heywood et al.
Table 1.
The first ten rows from the radio source catalogue, presented here in order to show the table structure. Please refer to the text for a detaileddescription of each column. The full version of this table is available online as supplementary material.ID RA Dec σ RA σ Dec RA peak Dec peak S int σ S int [deg] [deg] [arcsec] [arcsec] [deg] [deg] [mJy] [mJy](1) (2) (3) (4) (5) (6) (7) (8) (9) J022143.11-041344.6 35.42963 -4.22905 2.59 2.65 35.43002 -4.2294 469.39263 0.01424 J022255.74-051817.5 35.73225 -5.30485 2.29 2.07 35.73219 -5.3048 269.69099 0.0188 J022632.54-051328.8 36.63557 -5.22467 2.13 1.98 36.63563 -5.22469 71.48008 0.00953 J022915.86-044216.7 37.31609 -4.70464 3.94 3.38 37.31561 -4.70498 272.04369 0.04092 J021640.74-044404.4 34.16974 -4.73456 2.06 2.08 34.1698 -4.73445 60.58129 0.00945 J021705.51-042253.1 34.27297 -4.38143 1.91 2.29 34.273 -4.38142 59.63392 0.01103 J022310.19-042306.4 35.79245 -4.38512 1.95 1.98 35.79253 -4.38508 39.94858 0.00794 J022754.85-045705.5 36.97856 -4.95152 2.24 2.0 36.9785 -4.95146 35.40523 0.00853 J022357.09-044112.5 35.98789 -4.68682 1.99 2.35 35.98794 -4.68674 42.21077 0.00876 J022505.11-053648.1 36.27128 -5.61335 2.2 2.44 36.27124 -5.61345 161.54529 0.03284 S peak σ S peak RMS_Peak RMS_Mean θ maj θ min PA α σ α ID2 ID3[mJy b − ] [mJy b − ] [mJy b − ] [mJy b − ] [arcsec] [arcsec] [deg](10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) do this, we take all segments below a given pixel threshold limitand investigate what fraction of the total flux contained within thePSF beam and centred on the RA/Dec position of the source iscontained within the source segment. We apply this correction tothose sources which have a value of N100 (the number of pixels inthe segment found by ProFound) less than 225 pixels. This limitis chosen as 225 pixels in a 15 x 15 pixel box around a central PSFshould contain ∼
99% of the total flux within a PSF beam.Using ProFound with these settings resulted in a catalogueof 7,185 sources. We subsequently discard fitted regions wherethe peak flux density is below five times the noise value at thepeak position of the source, resulting in a final catalogue of 5,780sources. After a visually examining images, we identified 13sources for which multiple components (a total of 30) were actuallya single association. For these sources, the associated componentsare recorded within the final table. In addition to these, 18 sourceswere deemed to be artefacts and were subsequently removed fromthe catalogue. Following this, a total of 5,762 sources remainedwithin the final catalogue. The properties of the first ten sourcesfrom our final catalogue are shown in Table 1. The columns aredefined as follows:(1) Identifier for the component in HHMMSS.S+/-DDMMSSformat, formed from the right ascension and declination position inthe J2000 epoch.(2-3) Flux-weighted right ascension and declination of the com-ponent in degrees taken from the
RAcen and
Deccen columns fromProFound.(4-5) Flux-weighted standard deviations in the right ascension anddeclination of the component, taken from the xsd and ysd columnsfrom ProFound and converted into angular units using the pixelsizes. Note that this is significantly larger than the statistical uncertainty that can be obtained by fitting a point or Gaussiancomponent, and is included here mainly for completeness.(6-7) Right ascension and declination of the peak of the sourcein degrees taken from the
RAmax and
Decmax columns fromProFound.(8) Integrated flux density of the component in mJy. This iscalculated using the ProFound flux column, converted to Jy(from Jy beam − ), with an appropriate beam correction applied tocompensate for the flux density contribution from the outer wingsof the emission (see text).(9) Error in the integrated flux density of the component in mJy.It is calculated similar to (3) but using flux_err instead of flux and applying the square root of the beam correction.(10) Peak intensity of the component in mJy beam − . This isconstructed from the ProFound catalogue as flux × cenfrac .(11) Error in the peak intensity of the component in mJy beam − .It is calculated similar to (7) but using flux_err instead of flux .(12) RMS value in the map at the peak position of the source (givenby columns 6-7).(13) Mean rms over the source segment using the skyRMS_mean column from ProFound.(14) Major axis size of the segment and is quoted here as the2 × R100 column from ProFound and converted to arcseconds. (15) Minor axis size of the segment and is quoted here as the2 × R100 × axrat from ProFound and converted to arcseconds.(16) Positional angle of the source in degrees given by the ang As this (and the minor axis size) are calculated based on the segment size,for faint sources comparable to the noise, the segment will be small and thissize will be underestimated. These are also not comparable (in many cases)to sizes in previous radio catalogues, which are often quoted as full widthhalf maximum values from Gaussian components.MNRAS in press, 1–13 (2020)
LA imaging of the XMM-LSS / VIDEO deep field column from ProFound.(17) Spectral index ( α ) estimate formed by extracting pixels fromthe spectral index map (Section 3.2) over the region correspondingto a given source as determined by ProFound. The mean of thespectral index value of the extracted pixels is determined, weightedby the total intensity values over the same area.(18) Total intensity weighted standard deviation of α , measuredover the corresponding ProFound region.(19)-(20) IAU Source IDs of components that together with theentry in column (1) are part of a single radio source. In the sections that follow we compare the results presented inSections 3.1 and 3.3 with existing radio data in order to validatethese data products. For positional and flux density checks (Sec-tions 4.2 and 4.3) we make use of existing data covering the samefield, namely the Faint Images of the Radio Sky at Twenty-cm(FIRST) survey (Becker et al. 1995), and radio imaging of theVLA-VIRMOS Deep Survey field (VVDS; Bondi et al. 2003), andSubaru-XMM/Newton Deep Field (SXDF; Simpson et al. 2006).For astrometric and photometric checks (Sections 4.2 and 4.3) werestrict the cross-match to sources that have no clear evidence ofhaving extended morphology in order to minimise the effects ofangular resolution differences.
The sensitivity (or background noise level) of a radio mosaic atthese frequencies is generally position dependent. This can be dueto a range of factors, e.g. the increase of the noise at the peripheryof the mosaic due to primary beam correction, calibration deficien-cies leading to error patterns associated within bright sources (e.g.Figure 1), residual sidelobe confusion due to incomplete deconvo-lution, and particularly problematic RFI in some pointings causinghigher than normal data loss for that region. A convenient way tocapture the sensitivity of the mosaic as a function of position is tomake use of the RMS noise map that is produced by the sourcefinder in order to set its internal local detection thresholds. Figure 3shows the a normalised histogram of the pixels in this RMS image.The median RMS noise is 16 µ Jy beam − , with 80% of the mosaicarea having a noise value of < µ Jy beam − . The accuracy to which the position of a component in a radio imagecan be measured depends on two factors (Condon 1997). The firstis a statistical effect related to the signal to the noise ratio (SNR)of the detection and the angular resolution of the instrument. Thesecond is a systematic component coupled to accuracy of the astro-metric reference frame that is applied to the data via the calibration.Both of these effects can be gauged by cross-matching positionalmeasurements with those of suitable reference data, if such data areavailable. Ideally the reference set should have superior depth andangular resolution such that the statistical uncertainties are domin-ated by those of the survey under test. The systematic calibration-related component is best investigated by using the strongest sources(e.g. phase calibrators) for which the statistical contribution in bothdata sets is negligible. In practice, and with many modern radio
Figure 3.
Normalised histogram of the pixel values in the RMS image ofthe mosaic, taken to be a measurement of the background noise across thesurvey. The median noise is 16 µ Jy beam − . observations breaking new ground, the availability of suitable refer-ence sets is limited, and typically relies on using a large-area surveyto investigate the brighter sources that are common to both. The useof bright calibrator sources with excellent positional measurementsis generally not feasible for deep and relatively narrow surveys suchas the one presented here, however the astrometry of surveys suchas FIRST and NVSS is validated against calibrator sources, so witha large enough sample of common sources any systematic offsetsshould be apparent.We calculate offsets in right ascension and declination betweenthe peak positions in our catalogue and the matched position inan external reference catalogue. Three external catalogues are em-ployed, namely FIRST, VVDS and SXDS. The distribution of theseoffsets is shown in Figure 4. The inner ellipse is centred on the meanpositional offset, and has minor and major axes showing ± ∼
90% of them were found tolie within the bounds of pointing 7 of the SXDF mosaic, suggestingan issue either with the calibration or image regridding for thatparticular pointing. The offsets between our catalogue positionsand those of VVDS is noticeable compared to those of FIRST andSXDF, however given that we are consistent with the latter two weassume this is related to the VVDS calibration.
The VLA has very accurate absolute flux calibration (of order 1percent; Perley & Butler 2013) due to the use of well-modelledprimary calibrator sources, in this case 3C147. However, additionalfactors (e.g. subsequent referenced calibration and self-calibrationproblems, deconvolution biases, RFI) can skew the flux calibration.In Figure 5 we compare the peak flux densities of our cataloguescomponents with matched components drawn from the SXDS and
MNRAS in press, 1–13 (2020)
Heywood et al.
Figure 4.
Differences in the peak right ascension and declination of components matched between the ProFound catalogue derived in Section 3.3 and anexternal reference data set. Left to right, the external references are FIRST (Becker et al. 1995), VVDS (Bondi et al. 2003) and SXDF (Simpson et al. 2006).The inner ellipse is centred on the mean positions of the offsets, and its major and minor axes are ± ∼ ± ± ± ± ± ± ± VVDS catalogues. As with the positional checks there were 690 and724 mutually-compact sources for SXDF and VVDS respectively.Matched components are scattered about the 1:1 line where thecatalogued and external component are equal, as shown as the diag-onal on Figure 5. The usual increase in scatter with decreasing peakintensity is seen. As the noise level becomes an increasingly largefraction of the component brightness temporally separate measure-ments of the same source will exhibit larger amounts of scatter.There is no obvious biasing of e.g. the fainter sources, as would beseen by a curve in the distribution of points about the 1:1 line.A potential source of bias in the recovered flux density ofsources is the application of inappropriate self-calibration. Sincethe sky model against which the instrument is calibrated is neverfully complete, the contribution to the visibility function made bythe unmodelled sources can potentially be absorbed by the antenna-based gain solutions, resulting in these unmodelled sources beingsuppressed in the final image. Mitigation of this effect can take theform of conservative time-frequency solution intervals, and minim-ising the degrees of freedom. The latter issue is automatically ad-dressed to some extent by virtue of the VLA having a high ratio ofbaselines (351) to antennas (27), which results in a correspondinglyhigh ratio of equations to solvable parameters during calibration.However the application of differential gains introduces two solv-able parameters into the measurement equation for every additionaldirection that is being solved for, and extra care must be taken.We check for the presence of systematic flux density biasesintroduced by the directional calibration process by comparing theflux densities of matched components in the images formed fromthe NRAO pipeline (i.e. maps for which no self-calibration has beenapplied) and those formed following the full direction-dependentcalibration procedure. These are plotted in Figure 6.As with Figure 5, this plot shows increased broadening of thedistribution away from the diagonal line with decreasing values ofcomponent flux density. Note however that in this case the diagonalline is not simply the 1:1 line, but rather a fit to the data that is indis-tinguishable from the diagonal. A noise-like scattering of the pointsis to be expected, as self-calibration modifies the noise properties of the images. Systematic biasing of the flux density measurements(for example the often-seen suppression of faint sources that arenot in the calibration model) would manifest itself as a curve in thedistribution, for which no evidence is seen.
The catalogued spectral index measurements (having their originsdescribed in Sections 3.2 and 3.3) are shown in Figure 7, withthe corresponding integrated flux density plotted against them. Thedashed line on this plot shows the limit where a source with apeak intensity in mJy beam − , concordant with the y-axis valuesand measured from the full-band mosaic, would drop below thethreshold of any one of the three sub-bands used, and therefore nothave a spectral index measurement in this survey. This line is evalu-ated for the full spectral index range of Figure 7, and demonstratesthat in-band spectral indices are subject to a spectral index depend-ent selection bias for sources with flux densities that approach thesurvey detection threshold. The population of sources that lie abovethis line can thus be considered to be complete for plausible andtypical spectral indices.The mean spectral index of the sources in integrated flux dens-ity bins is measured, and these values are plotted on Figure 7 witherror bars that show ± of the Extended Chandra Deep Field South to a depth of 8.6 µ Jybeam − , from which they match 167 sources with counterparts from MNRAS in press, 1–13 (2020)
LA imaging of the XMM-LSS / VIDEO deep field Figure 5.
Comparison of the peak intensities from two external radio sur-veys, namely VVDS (Bondi et al. 2003) and SXDF (Simpson et al. 2006),plotted against the catalogued peak intensity from our survey. µ Jy beam − (Miller et al.2013). A flattening of the median spectral index (-0.58) is observedin the sub-mJy population, however a substantial fraction of flat orinverted spectrum radio sources is present. Next, we measure the differential source counts for these observa-tions and compare them to previous work. Before comparisons canbe made with previous studies it is important to correct the measureddifferential source counts, which will be underestimated especiallyat the faintest flux densities. This underestimation is due to sev-eral factors. Firstly the variations in the image sensitivity across
Table 2.
Mean spectral index values of the N components with integratedflux densities within the range defined by S min (inclusive) and S max (withbin centre S ). These values are plotted as black markers on Figure 7. Furtherdetails are provided in Section 4.4.N S min S max S α med σ α [ mJy b − ] [ mJy b − ] [ mJy b − ] the survey area means that faint sources will not be detectable inall regions of the image. Secondly, false detections whereby noisepeaks are interpreted as true emission will affect the source countsin the faintest bins. Furthermore, sources for which any (positive ornegative) coincident noise peak represents an appreciable fractionof their total flux density may be redistributed into an adjacent bin(Eddington bias). The methods for determining the factors requiredto correct for these effects are described below. Firstly, we correct the measured source counts from the output cata-logue for non-uniform detection across the field of view as well asresults that may arise from source fluxes being influenced by noisepeaks or troughs. To determine the necessary corrections, simula-tions are used to correct source counts as in Hale et al. (2019b) andthe corrections determined are applied to this work. For this, simu-lated sources are injected into the image and then source extractionis run as in Section 3.3, this can be used to to determine the recov-ery as a function of flux density. For each simulation, 1000 sourceswere injected at random positions within the image. Each of thesesources has an associated flux density with the distribution drawnfrom the SKA Simulated Skies continuum simulation ( S Wilmanet al. 2008, 2010), which provides realistic catalogues containingsimulated extragalactic radio sources of various population typesdown to a flux limit of 10 nJy. The shapes of the injected sources areelliptical components, with the associated sizes also drawn from the S simulations. For each simulated source randomly chosen, eachelliptical component associated with the source are convolved withclean beam of these observations and then injected into the image .As this uses the size distribution of S , this technique should alsoaccount for resolution effects where, for the same total flux dens-ity, larger sources will have a lower peak flux density per beamand therefore will be more challenging to detect above the noisethreshold. Both single (radio quiet AGN and star-forming galaxies)and multi-component (FR-I and FR-II radio galaxies) sources areinjected into the image provided that the total flux of the source is ≥ × σ , where σ was taken as the typical rms of the observations The catalogues used to determine these corrections also have the 5 σ cri-terion described in Section 3.3 imposed before the corrections is calculated. S sources which had components of the largest sizes were not included,to ensure the lobes are not cut off when injecting into the image. Sources withsizes>50" were not included. Only a small fraction of S sources were notincluded due to this limit and so are unlikely to have made a big differenceto the corrections derived.MNRAS in press, 1–13 (2020) Heywood et al. (converted to a total flux assuming a point source), and was takento be 16 µ Jy beam − .To determine the completeness from the simulation we calcu-late the ratio of the output source count distribution to the inputdistribution. As the simulated sources are injected into the image,the observed (real) source count distribution measured from the im-age must first be subtracted. The completeness correction in a givenflux density bin is therefore given by: C COMP ( S i , S i + dS i ) = N sim , out ( S i , S i + dS i ) − N im ( S i , S i + dS i ) N sim , in ( S i , S i + dS i ) (7)where N sim , out is the number of sources detected in the outputsimulated image above 5 σ (as defined in Section 3.3, N im is thenumber of sources within the original image, again above 5 σ andfinally N sim , in is the number of simulated sources within the givenflux density bin that are injected into the image. As N sim , out willbe the combination of both the sources already in the image aswell as those simulated sources that are recovered, the value of N sim , out − N im will quantify those simulated sources that arerecovered from the image. As sources that are injected into theimage, these simulations may also take into account the fact thatsources may merge with others in the image and only be detectableas an individual source.We generated 100 realisations of the simulation and calculatedthe completeness corrections as the median value of these. Theassociated uncertainties that we quote with this are generated fromthe 16 th and 84 th percentiles of the completeness corrections forthe 100 simulations. The inverse of these corrections will need tobe applied to the measured source counts in order to correct forincompleteness. To quantify the fraction of false detections in the image, we makethe assumption that the noise across the image is symmetric andtherefore every positive noise spike will on average have a corres-ponding noise decrement. As such, the number of falsely detectedsources within a given flux density bin can be calculated throughinvestigating how many sources would be detected within the neg-ative image (i.e. where the image is multiplied by -1). The samedetection parameters of ProFound (as described in Section 3.3),5 σ threshold and beam correction method (as described in Haleet al. 2019a) are used to extract the catalogue of sources in thenegative image. As this correction aims to account for the fact thatsome sources within the measured catalogue may be false, this cor-rection will act to decrease the measured source counts. This is inthe opposite direction to the corrections described in Section 4.5.1.The correction that is applied to account for these false detectionsis given by: C FDR ( S i , S i + dS i ) = − N neg ( S i , S i + dS i ) N cat ( S i , S i + dS i ) (8)where N cat is the number of sources within the flux density bin S i , S i + dS i and N neg is the number of sources within the same fluxdensity bin that are detected within the negative image. Figure 6.
Peak intensities from the final calibrated maps of some indi-vidual pointings plotted against the peak intensities of matched componentsmeasured from maps that have not been subjected to any self-calibration,with only the referenced calibration applied. There is no evidence for anycalibration-induced photometry biases. Note that the diagonal line here is afit to the data points.
Figure 7.
Integrated flux density measurement against source spectral indexfor the 3,458 sources that have spectral index estimates. The black markersshow the mean spectral index for the flux density bins listed in Table 2,with the error bars showing 1 standard deviation. The dashed line shows thetheoretical peak flux density limit (as would be measured in the full-bandimage) below which a source of a given spectral index would drop out of oneof the three sub-bands used to form the spectral index map. Sources abovethe entirety of this line can be assumed to be mostly free from signal-to-noiserelated selection biases for most common or plausible astrophysical radiospectra. MNRAS in press, 1–13 (2020)
LA imaging of the XMM-LSS / VIDEO deep field Table 3.
The data corresponding to Figure 8 for measurements below 500 mJy. For each flux density bin we list the differential source counts in bothraw and Euclidean-normalised form. The final column lists the source counts following the application of the corrections described in Section 4.5.1 and4.5.2. Bin Bin Mid Counts Raw d NdS S . Corrected d NdS S . [mJy] [mJy] [sr − Jy . ] [sr − Jy . ]0.10 - 0.13 0.11 636 ±
25 1.45 ± . + . − . ±
26 2.22 ± . + . − . ±
24 2.75 ± . + . − . ±
21 2.89 ± . + . − . ±
19 3.34 ± . + . − . ±
18 4.36 ± . + . − . ±
15 4.35 ± . + . − . ±
13 4.56 ± . + . − . ±
12 5.75 ± . + . − . ±
10 6.03 ± . + . − . ±
10 7.36 ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . ± ± . + . − . To obtain the corrected source counts, which should be a true estim-ate of the underlying flux density distribution, the corrections fromthe completeness simulations and false detections are combined to-gether multiplicatively to the source counts determined from themeasured output catalogue described in Section 3.3. This is appliedto the source count from the catalogue where artefacts have not beenremoved. This is because these artefacts may also be apparent in theinverted image for which the false detection rate is determined from.The associated uncertainties from the measured source counts andthe corrections are then combined together in quadrature in orderto quantify the total uncertainty on these corrected source counts.The corresponding measurements of the source counts and theiruncertainties are given in Table 3.A comparison of the uncorrected and corrected source countsare presented in Figure 8, for which observations of previous meas-ured source counts are also presented. These previous source countmeasurements are from the VLA 3 GHz COSMOS Survey (Smolčićet al. 2017), the compilation of 1.4 GHz source counts presented byde Zotti et al. (2010) and finally the S extragalactic simulated skiesat 1.4 GHz (Wilman et al. 2008). These are all scaled to 1.4 GHz assuming a spectral index of -0.7. The source counts are plotted forthose flux density bins that have a minimum value greater than the5 σ limit (where σ is taken as 16 µ Jy). As can be seen from Table 3,the corrections to the source counts become important at flux dens-ities of S (cid:46) . MNRAS in press, 1–13 (2020) Heywood et al.
Figure 8.
Euclidean normalised differential source counts and their associated uncertainties for these observations are shown above via the pink markerscircles. Also shown are the comparisons to previous work from simulations at 1.4 GHz of Wilman et al. (2008) (black markers) and also the compilationof observations from de Zotti et al. (2010) (grey markers), the VLA 3 GHz COSMOS Survey from Smolčić et al. (2017) (blue markers), and the 1.28 GHzMeerKAT observations from Mauch et al. (2020). These are all scaled to 1.4 GHz assuming a spectral index of -0.7 where necessary.
We have described the production and validation of the reduceddata products associated with a VLA survey covering ∼ ofthe XMM-LSS / VIDEO field. The data we present enhance themulti-wavelength view of one of the best-studied extragalactic deepfields, and we make these products publicly available for use bythe community, downloadable from http://tiny.cc/vla-xmm, or byemailing the contact author.Direction-dependent calibration has been used to produce abroadband radio mosaic that reaches a thermal noise-limited me-dian depth of 16 µ Jy beam − with an angular resolution of 4.5 (cid:48)(cid:48) .Our survey improves on the existing matched-frequency radio data,expanding the area by a factor of 2.5 to encompass the entire regionfor which the deep near-infrared VIDEO data (Jarvis et al. 2013) areavailable, and further increase the depth of the radio data availableover this region at these frequencies by 40%.A source catalogue with 5,762 entries has been produced usingthe ProFound source finder, recently demonstrated to have excel-lent performance for the characterisation of extended radio sourcesby Hale et al. (2019a). The photometric and astrometric perform-ance of the resulting catalogue (and thus the radio mosaic fromwhich it is derived) have been validated by comparison to exist-ing narrowband observations. The bias-corrected differential sourcecounts are also in excellent agreement with simulations and obser-vations. The 66% fractional bandwidth of the VLA allows in-bandspectral indices to be estimated for sources detected at sufficientlyhigh signal to noise ratios, and the catalogue contains spectral in-dex measurements for 60% of sources. The mean spectral index asa function of integrated flux density resembles the canonical syn- chrotron values at the bright end, tending towards flatter spectrumsources below about 1 mJy.Looking forwards, a second data release will be forthcomingwith observations from the compact C and D configurations of theVLA being added in order to improve the sensitivity to the manydiffuse structures that are evident in the mosaic. Observations usingnew mid-frequency SKA precursor instruments will also target thisfield, with the MIGHTEE survey (Jarvis et al. 2016) specificallyplanning deep observations of XMM-LSS to greater depths at com-parable frequencies. The superior angular resolution of our VLAdata will prove useful not only for validation of the MIGHTEEdata, but also potentially for disentangling confused sources for theoptical cross-identification at 100 µ Jy beam − and above. ACKNOWLEDGEMENTS
We thank the anonymous referee and the MNRAS editorial staff fortheir comments. The National Radio Astronomy Observatory is afacility of the National Science Foundation operated under cooper-ative agreement by Associated Universities, Inc. We thank the VLADirector for awarding us 1.5 h of Director’s Discretionary Time inorder to complete this project. This work was supported by resourcesprovided by the Pawsey Supercomputing Centre with funding fromthe Australian Government and the Government of Western Aus-tralia. IH thanks the Rhodes Centre for Radio Astronomy Tech-niques and Technologies (RATT), South Africa, for the provisionof computing facilities. This research has made use of NASA’s As-trophysics Data System. This research made use of Montage. Itis funded by the National Science Foundation under Grant Number
MNRAS in press, 1–13 (2020)
LA imaging of the XMM-LSS / VIDEO deep field ACI-1440620, and was previously funded by the National Aeronaut-ics and Space Administration’s Earth Science Technology Office,Computation Technologies Project, under Cooperative AgreementNumber NCC5-626 between NASA and the California Institute ofTechnology. This research made use of APLpy, an open-source plot-ting package for Python (Robitaille & Bressert 2012). This work wassupported by the Oxford Hintze Centre for Astrophysical Surveyswhich is funded through generous support from the Hintze FamilyCharitable Foundation. CLH acknowledges the Science Technologyand Facilities Council (STFC) for their support through an STFCStudentship.
References