betaFeedback

WorksPhysics and Astronomy

Pushing the frontiers of low-frequency radio astronomy: An ultra-deep sub-arcsecond LOFAR image of ELAIS-N1

Abstract

We present ultra-deep sub-arcsecond radio images of the ELAIS-N1 field covering 2.5x2.5 deg^2, produced from 200 hours of LOw Frequency ARray (LOFAR) observations in the 115-168 MHz band. The radio continuum images are obtained using recent advances in LOFAR long-baseline calibration and imaging. We achieve a central point-source sensitivity of 5.3 μμJy/beam at an angular resolution of 0.33"x0.39" and an area-weighted sensitivity of 7.4 μμJy/beam at 0.36"x0.45", making these the deepest radio images below 1.28 GHz to date and among the deepest at any radio frequency. By combining our data with previously released deep 6" LOFAR images, we compile a catalogue with a total of 62,916 unique radio sources across 2.5x2.5 deg^2. Our calibration strategy, combined with astrometric alignment using 95 Gaia quasars, results in positional offsets of ΔΔRA = -3 ±\pm 27 mas and ΔΔDEC = 2 ±\pm 17 mas for these same sources. We further demonstrate the potential for deep polarisation studies at sub-arcsecond resolution with instrumental leakages from Stokes I into linear polarisation down to 0.12%, which is about a factor of two better than the typical leakage in the LOFAR Two-metre Sky Survey at 6" resolution. Our results demonstrate that imaging a large number of observations over a wide field area at sub-arcsecond resolution is now computationally feasible. This establishes a new benchmark for low-frequency radio imaging in terms of combining depth, resolution, and survey area, which lays the foundation for deeper and wider surveys in the LOFAR2.0 and SKA era.

Introduction

The LOw Frequency ARray (LOFAR; van Haarlem et al. [142]) is currently the only interferometer capable of routinely producing sub-arcsecond imaging of the low-frequency radio sky ([85]). Using LOFAR’s high-band antennas (HBAs), we have over the past few years carried out wide-field (2.5×2.52.5 \times2.5 deg2^2) imaging of several target fields at sub-arcsecond angular resolution in the 115–168 MHz band, by utilising baselines up to 1980 km across the European continent ([126]; [41]; [51]). This unique capability has enabled studies for a broad range of science cases that require sub-arcsecond resolution over wide areas of the sky (for a review, see Morabito et al. [84]). Extending these capabilities to produce ultra-deep sub-arcsecond images of the LOFAR deep fields (see [111]; [130]; [9]; [14]; [114], for current deep fields) is a next step, enabling the detection and study of the faintest and most distant radio source populations.

LOFAR’s sub-arcsecond wide-field imaging capability was for the first time shown to be possible by Sweijen et al. [126], who produced the first wide-field image with one 8-hour observation. This was followed by a demonstration of deeper imaging by de Jong et al. [41] using 32 hours of observations. While these studies highlighted the potential of sub-arcsecond wide-field imaging with LOFAR, scaling up to ultra-deep imaging with hundreds of hours of LOFAR observations remained non-trivial. The primary challenges included processing and imaging 16 TB of visibility data per 8-hour observation1 within a computationally feasible timescale, automatically selecting and precisely calibrating high signal-to-noise ratio (SNR) calibrator sources across the wide field of view (FoV) under varying ionospheric conditions, the difficulty of jointly calibrating the short baselines from the Dutch stations with the long baselines from the European stations, and the need for human-interventions for validation and quality assessment.

Recent work has addressed the aforementioned challenges. These include general developments in the calibration software facetselfcal (van Weeren et al. [146]) that have improved calibration performance and reliability (e.g. [145], Appendix A), also when jointly calibrating short and long baselines (de Jong et al. [42]). Automated self-calibration parameter selection and quality assessment by combining quality metrics with a machine learning algorithm have removed the need for human intervention ([41, 42]). The high computational costs for the final deep imaging with multiple observations are mitigated through sidereal visibility averaging (SVA; de Jong et al. [40]). This approach averages calibrated visibilities from different epochs with similar baseline coordinates, which, after applying all calibration solutions, reduces the total data volume of tens of observations to that of one or two. Combining this with recent improvements in automated processing strategies (de Jong et al. [42]), the overall computational cost of processing for ultra-deep imaging with hundreds of hours of LOFAR data has been lowered by more than an order of magnitude compared to the first successful LOFAR sub-arcsecond wide-field imaging by Sweijen et al. [126]. By making use of the Common Workflow Language (CWL; Amstutz et al. 2016; Crusoe et al. [29]), Toil (Vivian et al. [151]), and Simple Linux Utility for Resource Management (SLURM; Yoo et al. [161]) for data processing on high-performance computing clusters, we have also enabled efficient automated parallel processing of numerous observations, as detailed in van der Wild et al. (subm.). CWL workflows are reusable, interoperable, and portable, which aligns CWL-based pipelines with FAIR principles (e.g. Wilkinson et al. [158]) and supports scalable processing of large numbers of LOFAR observations.

In this work, we use the most recent data processing strategies for LOFAR long-baseline imaging to release ultra-deep sub-arcsecond wide-field images across a FoV of 2.5×2.52.5 \times2.5 deg2^{2} of the European Large Area ISO Survey-North 1 (ELAIS-N1; Oliver et al. [100]) deep field, using 200 hours of LOFAR-HBA observations in the 115–168 MHz band. The highest resolution image represents the deepest LOFAR image to date and is among the deepest radio images ever produced at any frequency. Located at a right ascension (RA) of 16:11:00 and declination (DEC) of +54:57:00, ELAIS-N1 has been extensively studied across a broad range of wavelengths, including X-rays (e.g., Manners et al. [78]), ultraviolet (e.g., Martin et al. [79]; Chaturvedi et al. [20]), optical (e.g., McMahon et al. [82]; Aihara et al. [2]), infrared (e.g., Lawrence et al. [73]; Mauduit et al. [81]; Oliver et al. [101]), and radio (e.g., Sirothia et al. [119]; Ocran et al. [94]; Sinha et al. [118]; Kondapally et al. [68]; Best et al. [9]), and is one of the selected LOFAR deep fields (Kondapally et al. [69]; Duncan et al. [49]; Tasse et al. [130]; Sabater et al. [111]; Bondi et al. [14]). The deepest sub-arcsecond LOFAR image so far of this field was produced by de Jong et al. [41] using 32 hours of LOFAR observations, reaching a central sensitivity of 14 µJy beam−1^{-1}. More recently, Shimwell et al. [114] imaged the same field using 505 hours of all available LOFAR observations containing data from the Dutch stations, achieving a central sensitivity of about 10 µJy beam−1^{-1} at 6′′ resolution. Although they used about 16 times more observing time than de Jong et al. [41], the two images reach comparable point-source sensitivities. This is a consequence of both the larger collecting area of the international stations (van Haarlem et al. [142]) and of confusion noise (Condon et al. [23]), which limits deep 6′′-imaging with LOFAR with only the baselines from the Dutch stations (e.g. Sabater et al. [111]; Shimwell et al. [114]).

Combined with the wealth of existing multi-wavelength data (Kondapally et al. [69]; Duncan et al. [49]) and the forthcoming deep optical–infrared catalogues of ELAIS-N1 (Pirie et al. in prep.), our ultra-deep sub-arcsecond images open a new observational window onto a broad range of science cases. These include separating star-formation and AGN contributions to the radio emission of fainter sources (Morabito et al. [87, 86]), probing the cosmic star-formation history to high redshifts (e.g. Novak et al. [92]; Cochrane et al. [21]; Wang et al. [154]; Thykkattu et al. 2026), extending measurements of the radio luminosity function with a more complete sample of the faintest and least powerful radio jets (e.g. Gendre et al. [56]; Novak et al. [93]; Kondapally et al. [68]; de Jong et al. [38]), among other applications (e.g. Morabito et al. [84]). Our results also lay the foundation for even deeper and wider future surveys that aim to exploit the full international LOFAR during the upcoming LOFAR 2.0 era.

This paper is structured as follows. Section 2 describes the selection of the ELAIS-N1 observations used in this work and the data processing strategy to reduce the data. In Section 3, we outline the cataloguing process for obtaining a single multi-resolution source catalogue. The data quality is presented in Section 4, followed by a discussion and the conclusions in Sections 5 and 6, respectively.

Observations and data reduction

The data reduced in this work cover the ELAIS-N1 deep field (RA = 242.75 deg, DEC = 54.95 deg). Below we describe the data selection and processing workflow used to reduce the data. We also discuss the flux density scale and astrometric corrections applied to the data prior to imaging and analysis.

Data

Table 1 lists the LOFAR observations that we selected for calibration and imaging. Out of the total 505 hours of available LOFAR observations, targeting the ELAIS-N1 field (Shimwell et al. [114]), we selected 200 hours from the LT10_012 and LT14_003 projects (PI: P.N. Best). Hence, 305 hours were excluded based on the following criteria:

Observ. IDProjectCalibrator IDObserv. dateStations (int)Frequencies (MHz)RMS 1.2″ (µJy beam⁻¹)
L686892LT10_012L68688806-12-201870 (12)120-165116
L686906LT10_012L68690208-12-201875 (13)120-165106
L686920LT10_012L68691603-12-201874 (12)120-165102
L686934LT10_012L68694007-12-201874 (12)120-165110
L686948LT10_012L68694405-12-201873 (11)120-167113
L686962LT10_012L68695826-11-201873 (11)120-16588
L686976LT10_012L68697225-11-201873 (11)128-167112
L686990LT10_012L68698624-11-201873 (11)126-167103
L687004LT10_012L68700022-11-201870 (9)128-167106
L760853LT10_012L76084904-12-201967 (12)120-166100
L769393LT10_012L76938925-05-202075 (13)120-16589
L769407LT10_012L76940319-05-202071 (12)120-16584
L769421LT10_012L76941712-05-202073 (13)120-167126
L769435LT10_012L76943115-03-202072 (11)120-16696
L798022LT14_003L79803012-11-202072 (11)117-16692
L798074LT14_003L79808214-11-202074 (11)115-16490
L798092LT14_003L79810020-11-202068 (11)115-16495
L798128LT14_003L79813626-11-202069 (9)115-16681
L811718LT14_003L81172619-03-202172 (11)117-16584
L811736LT14_003L81174420-03-202175 (13)115-16492
L816272LT14_003L81628015-05-202175 (14)115-16493
L816308LT14_003L81631623-05-202175 (13)116-16286
L816344LT14_003L81635227-05-202173 (14)115-16692
L818954LT14_003L81896222-06-202172 (10)115-156117
L833466LT14_003L83347409-10-202173 (13)119-165115

Table 1. Metadata from the 25 ELAIS-N1 observations used in this paper. Notes. All calibrator IDs correspond to 10-minute observations of calibrator 3C 295. The number of international stations is given in brackets. The reported numbers of stations and frequency channels correspond to the data after flagging. The RMS noise values are measured from the 1.2″ wide-field images and reflect the combination of flagging and ionospheric conditions.

  • All observations from projects LC2_024 and LC4_008 were recorded either with Dutch LOFAR stations only, or with international stations excluding the most distant ones (in Poland, Latvia, and Ireland), which were not yet part of the array when these data were taken (Blaszkiewicz et al. 2016; Vrublевskis et al. 2020; Murphy et al. [89]). This makes them unsuitable for sub-arcsecond resolution imaging.

  • Four observations were recorded at 10 sec intervals. This time resolution results in extreme time smearing effects when imaging at sub-arcsecond resolution (see § 4.2).

  • Three observations showed elevated RMS noise levels, exceeding twice the expected values at 6′′ resolution as reported by Shimwell et al. [114], making them unsuitable for calibration with the international stations.

  • Eight observations were excluded after visual inspection of the calibration solutions from the LOFAR Initial Calibration pipeline (LINC; de Gasperin et al. [33]), due to excessive flagging (>40%) of the international stations.

The observations from project LT10_012 were stored with a 1-second time resolution, whereas those from LT14_003 used a 2-second resolution, which increases time-smearing at the edge of the FoV (see § 4.2). Additionally, the LT14_003 pointings were offset by 0.03° from the field centre from LT10_012, requiring a phase shift to align all observations to a common RA of 16:11:00 and DEC of +54:57:00. The observations L686962, L769391, L798074, and L816727 had already been processed by de Jong et al. [41], but were reprocessed in this work at the DD calibration step using the updated calibration strategy by de Jong et al. [42].

Calibration

To calibrate the selected datasets, we adopted the strategy outlined in de Jong et al. [41] up to the direction-independent (DI) calibration stage. Beyond this point, we followed the direction-dependent (DD) calibration strategy of de Jong et al. [42] for Stokes I and the polarisation imaging strategy of van Weeren et al. [144] for deep polarisation imaging. This means that after downloading the data from the LOFAR long-term archive (LTA; Renting & Holties [110]; van Haarlem et al. [142]), we applied the HBA calibrator and target workflows from LINC2, the DDF-pipeline3, workflows from the Pipeline for the International LOFAR Telescope (PILOT)4, and additional calibration steps with facetselfcal5 in the order outlined below. As each processing step has been described in detail in previous work, we refer the reader to the corresponding cited papers for further details. In summary, the overall processing steps were applied as follows:

  1. The first step involved executing the automated LINC workflows, which derived initial calibration corrections for instrumental and ionospheric effects across all LOFAR stations ([148, 33]). These corrections were obtained from a 10-minute scan of the well-known calibrator source 3C 295, while additional calibration solutions for the Dutch LOFAR stations were derived from the ELAIS-N1 target field data.

  1. Next, we ran the DDF-pipeline for DD calibration of the Dutch LOFAR stations ((Tasse 2014a,b; Tasse et al. 2018, 2021; Smirnov & Tasse 2015)). This gave us full-Jones DI calibration solutions for the Dutch LOFAR stations, which we used as a starting point for the calibration of the international LOFAR stations, and DD calibration solutions to subtract sources located outside the 2.5×2.5 deg22.5 \times2.5\ \mathrm{deg}^{2} FoV ([126]), after which the DD solutions were discarded.

  1. After obtaining the calibration corrections for the Dutch stations, we used the delay–calibration workflow from PILOT to obtain full-Jones DI calibration corrections for all LOFAR stations ([85]). We used the Seyfert 2 galaxy JF160607.6+552135, selected by Ye et al. (2024), from the catalogue of calibrators from the Long-Baseline Calibrator Survey (LBCS, [66, 65]), as calibrator source. This calibration step incorporated the skymodel and refined calibration steps described in de Jong et al. [41]. Because the source is unpolarised ([62, 103]), any polarised signal detected towards it is instrumental, so with the full-Jones calibration the polarisation leakage was corrected.

  1. After applying the DI corrections to our data, we performed an additional DD calibration step with facetselfcal to derive solutions for the Dutch LOFAR stations in several high-SNR directions. This is essential for stabilising the subsequent DD calibration of the Dutch and international stations combined. We followed the strategy of de Jong et al. [42], which describes this step in more detail, with one main difference: we first jointly calibrated and imaged six observations (the 4 observations from de Jong et al. [41] with L686892 and L686906) to produce an accurate sky model of the full 2.5×2.5 deg22.5 \times2.5\ \mathrm{deg}^{2} FoV, against which we then calibrated all remaining observations in a single iteration. This step is not yet part of PILOT, but may be added in the future.

  1. Next, with the datasets with the DI calibration solutions for the international stations applied and using the DD calibration solutions for the Dutch stations, we applied the dd-calibration workflow from PILOT to obtain DD calibration solutions for all LOFAR stations, as detailed in de Jong et al. [42]. This workflow pre-applied the nearest calibration solutions from step 4 and then calibrated the international LOFAR stations. See Figure 1 for the DD calibrator selection and corresponding facet layout, where the calibration solutions are assumed to be constant within each polygon from the Voronoi tessellation ([113, 148]).

Facet layout of the 24 calibrators with corresponding facet IDs

Figure 1. Facet layout of the 24 calibrators with corresponding facet IDs. The DI calibrator is indicated by the yellow star and the other DD calibrators are marked with red crosses. The facet boundaries are indicated by the green lines and are based on a Voronoi tessellation, which divides the field into polygon regions centred on calibrator sources, where each point in the field is assigned to its nearest calibrator. The calibration solutions are assumed to be constant within each polygon region ([113, 148]).

  1. With the calibration solutions for all international stations we made the first high-resolution wide-field images with the image-intermediate-resolution workflow from PILOT. This returned images of the full FoV at an angular resolution of 1.2′′1.2^{\prime\prime} (e.g. Ye et al. 2024; de Jong et al. [41]). Most important data products from this step were the fully calibrated Stokes I images for quality assessment (see RMS noise levels in Table 1) and the model images from each frequency channel, as these were used as input for the subsequent processing step (de Jong et al. [41]).

  1. Using the 1.2′′1.2^{\prime\prime} model images, we ran the facet-subtract workflow from PILOT to split each facet into its own visibility dataset, where all sources outside of the facet were subtracted (de Jong et al. [42]).

  1. After the facet-subtraction, we applied a final calibration step, in which the selected calibrator sources from step 5 were calibrated jointly across all observations. This followed the method described in de Jong et al. [42]. This ensured all observations were astrometrically aligned before the final combined imaging and further improved the deep calibration quality as signal from other bright sources in the field was here subtracted outside the facets. The calibration solutions were directly applied to the facet data. The workflow to perform this will be added in the future to PILOT.

  1. Finally, to allow for deep polarisation imaging, we aligned the polarisation angles between the different observations, using the known polarised source 7C 1604+5447 (Herrera Ruiz et al. [62]; Piras et al. [103]). The steps for this alignment procedure and for this particular source are detailed in van Weeren et al. [144].

The advantage of this calibration strategy is that all 25 observations are processed independently up to step 8 (with the exception of the additional step 4). The joint calibration in step 8 is performed only after the preceding processing steps have been completed for all observations. This enables embarrassingly parallel processing of the computationally intensive calibration and imaging steps up to that point. The RMS noise levels of the 1.2′′1.2^{\prime\prime} wide-field images produced in step 6 range from 81 to 162 μ\muJy beam−1^{-1}, as shown in Table 1. This reflects the combined effects of data flagging, ionospheric conditions, and variations in instrumental/observing properties across the years the observations were taken. After all calibration steps, we had datasets with fully DD calibrated visibilities for each of the 24 facets. These were then imaged, as outlined in the next subsection.

All the (self-)calibration steps, with PILOT and during the extra calibration steps, were done with facetselcal. This software package utilises the Default Preprocessing Pipeline (DP3; van Diepen et al. [140]; Dijkema et al. [46]) and WSCLe an (Offringa et al. [97]; Offringa & Smirnov [98]), and functionalities from lofar_helpers (de Jong et al. [37]),6 to perform (self-)calibration on a source. Recent enhancements to facetselcal have improved the reproducibility of calibration through improved metadata logging and features to return configuration files from calibration solutions and images. These developments represent an important step towards aligning the software with FAIR principles (e.g. Wilkinson et al. [158]; de Jong [36]). This is essential to keep track of automatically selected parameters and to reproduce results in case further improvement of the output data or deeper imaging is desired in the future (see § 5.2). All software was executed with Apptainer containers (Kurtzer et al. [70]) from FLoCs,7 (Sweijen et al. [124]).

Imaging

To reduce the computational cost for deep imaging combining the datasets from each individual observation, we applied SVA after calibration (de Jong et al. [40]). This method averages the visibilities from different epochs with similar baseline coordinates. Since the baseline coordinates (uvwu v w) in the datasets are set by the time axis, the accuracy of SVA is determined by the time resolution, meaning that only the time resolution for the output dataset needs to be specified. Before SVA, each facet dataset was averaged in time by a factor set by the size of the facet (de Jong et al. [42]). To minimise the introduction of ad-

Notes. Facet numbers correspond to the numbers shown in Figure 1. The RMS is the standard deviation of the primary-beam corrected residual image, computed after iterative 3σ3\sigma clipping about the median.

additional time-smearing effects, we over-sampled the time axis of the final SVA-applied dataset by a factor of two compared to the averaged facet datasets, based on the analysis from de Jong et al. [40]. In this way, we reduced the total number of calibrated and facet-subtracted visibilities from all observations combined (about 200 TB of Dysco-compressed visibilities) by a factor of 12.5, which decreased the computational cost of gridding and prediction during imaging by about an order of magnitude compared to imaging without SVA.

Imaging was performed with WSClean (Offringa et al. [97]), following the imaging parameters from de Jong et al. [41] and using the wgridder module (Arras et al. [6]; Ye et al. (2022)). Hence, for each facet, we used a minimum uvuv-cut of 80λ80\lambda, and pixel sizes of 0.1′′0.1^{\prime\prime}, 0.2′′0.2^{\prime\prime}, and 0.4′′0.4^{\prime\prime}. These were paired with Gaussian tapers of 0.3′′0.3^{\prime\prime}, 0.6′′0.6^{\prime\prime}, and 1.2′′1.2^{\prime\prime}, respectively. The Briggs weighting (Briggs [19]) was set by de Jong et al. [41] to −1.5-1.5. However, for the highest angular resolution we changed this value in this work to −1.35-1.35 as this improved the point-source sensitivity by about 10%, while still achieving high angular resolution. To ensure efficient deconvolution, we employed automatic masking, multi-scale cleaning (Cornwell [25]), and an RMS box size set to 50 times the synthesised beam (Offringa & Smirnov [98]). We also applied a correction for the primary beam attenuation.

Each individually imaged facet has its own restoring beam and RMS noise, as reported in Table 2. The labels 0.3′′0.3^{\prime\prime}, 0.6′′0.6^{\prime\prime}, and 1.2′′1.2^{\prime\prime} refer to the Gaussian tapers used during imaging, while the actual restoring beams of each facet, as fitted by WSClean, are listed in this table. Facets further from the pointing centre have larger restoring beams at all three resolutions. They reflect the combination of different LOFAR station beams. The international stations are larger than the Dutch stations and therefore have narrower station beams, so away from the phase centre their sensitivity drops faster than that of the Dutch stations. Because the beam correction applied to each facet dataset also rescales the visibility weights, the longest baselines receive relatively lower weights in these facets, which broadens the PSF. The same primary beam attenuation from all stations combined also causes the RMS noise to increase towards the edge of the field. Before mosaicing the facets together, we restored them to a common resolution of 0.36′′×0.45′′0.36^{\prime\prime} \times0.45^{\prime\prime}, 0.58′′×0.62′′0.58^{\prime\prime} \times0.62^{\prime\prime}, and 1.0′′×1.5′′1.0^{\prime\prime} \times1.5^{\prime\prime}, matching the resolutions used by de Jong et al. [41]. The mosaicing was done with Swarp9 (Bertin et al. [8]). For many facets, in particular the ones labelled with 1.2′′1.2^{\prime\prime}, the restoring beam is smaller than the native beam. Because the residuals retain the native PSF, the flux density units are then not consistent within a facet. The mosaics are therefore intended for visualisation purposes only. All catalogues in this work were extracted from the individual facet images at their native resolution (see § 3), and quantitative measurements should be made on these facet images, which we release alongside the mosaics. In Figure 2 we display the full mosaic at 1.0′′×1.5′′1.0^{\prime\prime} \times1.5^{\prime\prime} angular resolution with postage stamps of radio galaxies at 0.36′′×0.45′′0.36^{\prime\prime} \times0.45^{\prime\prime} angular resolution. In Appendix § 10 we demonstrate with a few postage stamp images the resolving power of imaging at different angular resolutions.

Full wide-field radio continuum LOFAR image of the ELAIS-N1 field with postage stamp images of various radio galaxies

Figure 2. Full wide-field radio continuum LOFAR image of the ELAIS-N1 field at 1.0′′×1.5′′1.0^{\prime\prime} \times1.5^{\prime\prime} angular resolution with postage stamp images of various radio galaxies at 0.36′′×0.45′′0.36^{\prime\prime} \times0.45^{\prime\prime} resolution. The optical background is a three-colour composite of the Subaru Hyper Suprime-Cam (HSC; [2]), where gg-, rr-, and ii-band imaging, are mapped to blue, green, and red, respectively. The colour map for radio emission was generated with CMasher ([137]). We show several resolution comparisons of these sources in Appendix § 10.

Facet0.3″ Resolution (arcsec2^{2})0.3″ RMS (µJy beam−1^{-1})0.6″ Resolution (arcsec2^{2})0.6″ RMS (µJy beam−1^{-1})1.2″ Resolution (arcsec2^{2})1.2″ RMS (µJy beam−1^{-1})
00.35 × 0.417.20.56 × 0.609.91.08 × 2.1215.5
10.35 × 0.428.10.57 × 0.6111.11.12 × 2.3818.2
20.33 × 0.395.30.54 × 0.577.71.00 × 1.5013.2
30.33 × 0.395.40.54 × 0.578.11.00 × 1.4814.7
40.33 × 0.396.20.55 × 0.588.91.02 × 1.6615.0
50.35 × 0.427.60.57 × 0.6010.31.09 × 2.2416.7
60.33 × 0.405.60.55 × 0.588.31.02 × 1.6614.0
70.35 × 0.428.60.57 × 0.6111.11.10 × 2.2717.5
80.37 × 0.448.80.62 × 0.6312.31.29 × 2.7618.6
90.34 × 0.416.50.55 × 0.599.21.04 × 1.8215.0
100.36 × 0.4610.30.61 × 0.6513.91.37 × 2.9120.9
110.36 × 0.4610.60.62 × 0.6714.31.44 × 2.9419.3
120.36 × 0.437.80.59 × 0.6111.81.11 × 2.3816.3
130.35 × 0.427.40.58 × 0.6110.11.10 × 2.3216.5
140.35 × 0.438.30.58 × 0.6211.31.14 × 2.4617.8
150.35 × 0.427.60.57 × 0.6110.81.07 × 2.1917.5
160.38 × 0.4510.00.64 × 0.6513.31.20 × 2.6818.6
170.34 × 0.416.90.56 × 0.598.81.03 × 1.8614.4
180.34 × 0.405.40.55 × 0.588.10.99 × 1.5013.3
190.35 × 0.438.10.58 × 0.6111.21.13 × 2.4617.6
200.34 × 0.406.60.55 × 0.599.41.04 × 1.8315.4
210.36 × 0.448.20.59 × 0.6312.11.24 × 2.6919.3
220.34 × 0.417.30.56 × 0.5910.41.05 × 1.9916.2
230.33 × 0.395.40.55 × 0.588.41.01 × 1.6214.2
Mosaic0.36 × 0.457.40.58 × 0.6210.31.0 × 1.516.5

Table 2. Restoring beam resolution, as fitted by WSClean, and RMS noises of each individual facet and for the mosaic.

Flux scale alignment

Because of our extensive DD calibration process (steps 4, 5, and 8 in § 2.2), there is an increased likelihood that the amplitude corrections introduced small drifts in the flux density scale that was accurately set in step 3 from § 2.2. For this reason, and to align our flux density scale with the recent 6′′6^{\prime\prime} results of Shimwell et al. [114], we imaged the fully calibrated facet datasets at 6′′6^{\prime\prime} resolution, including the international stations and using the imaging parameters from § 2.3 with a 6′′6^{\prime\prime} Gaussian taper. We then generated a source catalogue using the source finder PyBDSF (Mohan & Rafferty [83], see § 3 for details). We cross-matched our 6′′6^{\prime\prime} catalogues for each facet with theirs for sources having peak intensities larger than ten times the local RMS and with integrated flux density measurements larger than five times the reported PyBDSF error. This provided us with 5 379 sources which we used to align the flux scales.

Dividing the flux densities from Shimwell et al. [114] by our measurements at the same resolution yielded a median flux density ratio of 0.95±0.020.95 \pm0.02, with values ranging from 0.89±0.010.89 \pm0.01 to 1.07±0.011.07 \pm0.01 across the individual facets. The uncertainties are the standard error of the median, obtained by bootstrap resampling with each flux density perturbed within the uncertainty reported by PyBDSF. Shimwell et al. [114] used different software and imaging parameters (DDFacet; Tasse et al. [129], Dutch stations only, Briggs robust −0.5-0.5, and no taper). Since the integrated flux densities of high SNR compact sources are, to first order, insensitive to the visibility weighting scheme and deconvolution depth, we expect these differences to have only a minor effect. They may contribute to a small global offset, but they cannot explain the facet-to-facet variation, as our imaging parameters are identical for all facets. For most facets, the ratios are also within the ∼9%\sim9\% flux density scale uncertainty of Shimwell et al. [114].

Applying the facet-to-facet scale factors using the median ratios brought our flux density scale into closer agreement with that of Shimwell et al. [114], yielding an average flux density ratio of 1.00±0.011.00 \pm0.01 across the full FoV. This procedure assumes the flux density scale of Shimwell et al. [114] to be free of facet-to-facet variations. Residual DD errors at the percent level cannot be excluded in their data either, so our correction should be understood as placing our images on the same scale as Shimwell et al. [114] rather than as an absolute flux density calibration. To demonstrate the accuracy of our flux density scale, we show in Appendix § 9 the spectra of several of our DD calibrators across measurements from different radio surveys.

Astrometric corrections

The previous work by de Jong et al. [41] reported a median astrometric offset of ΔRA=94±93\Delta\mathrm{RA} = 94 \pm93 mas and ΔDEC=−67±64\Delta\mathrm{DEC} = -67 \pm64 mas, based on comparing their radio sources positions with optical positions from Kondapally et al. [69], which were calibrated to the Two Micron All Sky Survey (2MASS; Skrutskie et al. [120]). Since the astrometric accuracy of 2MASS is around 100 mas, these data were of insufficient quality to further improve the astrometric precision. To obtain a better estimate of the astrometric precision, in this work we instead used the positions of quasars detected by Gaia (Gaia Collaboration et al. [54], which have sub-milliarcsecond positional accuracy.

We used PyBDSF to determine the positions of the sources in our highest resolution facet images (see § 3). To select quasars from our radio detections, we used the optAGN column from Kondapally et al. [69], which contains sources from the Million Quasar Catalog (Flesch [53]) or which were spectroscopically identified as AGN. These sources were cross-matched with our radio positions within 6′′6^{\prime\prime}, which matched with the radio catalogue resolution from the cross-matched catalogue. This left us with 260 sources that were then matched with the nearest Gaia source within 1′′1^{\prime\prime}. This resulted in 95 source matches across 22 of our 24 facets, where we found a global median astrometric offset of ΔRA=22±71\Delta\mathrm{RA} = 22 \pm71 mas and ΔDEC=−17±41\Delta\mathrm{DEC} = -17 \pm41 mas. The uncertainty is the scaled median absolute deviation (SMAD).

In the left panel of Figure 3, we show the ΔRA\Delta\mathrm{RA} and ΔDEC\Delta\mathrm{DEC} for these sources, with different facets indicated by different colours. The offsets cluster together for many of the facets, which can be explained by the fact that each facet received its own phase corrections during DD calibration, resulting in facet-specific systematic astrometric offsets. To correct for this, we applied facet-to-facet astrometric corrections based on the offsets relative to the Gaia positions. For facets 8 and 14, there were no Gaia quasar matches, so we applied to these facets the global median astrometric corrections from all facets together (ΔRA=31\Delta\mathrm{RA} = 31 mas, ΔDEC=−24\Delta\mathrm{DEC} = -24 mas). After applying all the corrections, the median astrometric offsets for the 95 quasars are reduced to ΔRA=−3±27\Delta\mathrm{RA} = -3 \pm27 mas and ΔDEC=2±17\Delta\mathrm{DEC} = 2 \pm17 mas, as shown in the right panel of Figure 3. This residual scatter is consistent with the positional uncertainties reported by PyBDSF, whose median values for this sample are σRA=17±15\sigma_{\mathrm{RA}} = 17 \pm15 mas and σDEC=20±13\sigma_{\mathrm{DEC}} = 20 \pm13 mas.

Astrometric offsets of Gaia quasars before and after correction

Figure 3. Astrometric accuracy of 95 Gaia quasars relative to the nearest radio sources in our newly obtained dataset before (left) and after (right) astrometric corrections. The black cross marks the median offsets in ΔRA\Delta\mathrm{RA} and ΔDEC\Delta\mathrm{DEC}, while different facets are shown in distinct colours to highlight the facet-to-facet offsets. For sign convention we used RAGaia−RALOFAR\mathrm{RA}_{\mathrm{Gaia}}-\mathrm{RA}_{\mathrm{LOFAR}} and DECGaia−DECLOFAR\mathrm{DEC}_{\mathrm{Gaia}}-\mathrm{DEC}_{\mathrm{LOFAR}}.

Computational costs

All data processing was carried out on the Spider cluster from SURF.10 The cluster nodes are equipped with AMD Rome and Bergamo architectures, offering 60 and 120 cores per node, along with 960 and 1440 GB of RAM per node, and local SSD storage of 10 TB and 12 TB, respectively.

The total computational cost of processing all 25 ELAIS-N1 LOFAR observations, covering a total integration time of 200 hours and a bandwidth of 53 MHz, amounted to approximately 750 000 CPU core hours, of which about 10% was spent on imaging. Had all observations been recorded with a 1 sec time resolution (see § 2.1), the processing cost would have increased by an additional ~150 000 CPU core hours. The total cost of processing all 25 observations is only about three times higher than that reported by Sweijen et al. [126], who processed a single comparable dataset recorded with a 2 sec time resolution, and is close to the total computational cost reported by de Jong et al. [41] who processed four of the LOFAR observations included in this work, of which two were recorded at 1 sec and two at 2 sec. These comparisons highlight the result of the computational optimisations, which are discussed in detail by de Jong et al. [42].

Cataloguing

We aim to generate a multi-resolution catalogue containing all available information on each radio source in our images, using all facet images at their native resolution (Table 2), combined with the deep 6′′ catalogue from Shimwell et al. [114]. Below, we first describe how we prepare the individual catalogues at each resolution and remove unreliable low SNR detections, followed by how we combine the catalogues from different resolutions.

Generating initial catalogues

Initial catalogues were generated independently at each resolution (with labels 0.3′′, 0.6′′, and 1.2′′) and for each facet, using PyBDSF (Mohan & Rafferty [83]),11 with identical parameters to what was adopted by de Jong et al. [41]. These include a primary RMS box of 120×120 pixels moved in increments of 15 pixels, supplemented by a smaller 40×40 pixel box moved in increments of 10 pixels for bright sources. À trous wavelet decomposition was applied to break the Gaussian-fitting residual image into wavelet images at multiple spatial scales (Starck & Murtagh [123]), and a group tolerance value of 10 was adopted to enhance the merging of neighboring Gaussians into single sources compared to the default value of 1. Other parameters were kept at their defaults. To the catalogues from our 0.3′′, 0.6′′, and 1.2′′ images, we add the 6′′ catalogue from Shimwell et al. [114].

We add to each catalogue a unique ID for each source, and carry these IDs over to the Gaussian catalogue. These unique IDs are of the format X_Y where the prefix X is the facet number for the source (see Figure 1), and the suffix Y is the source ID from PyBDSF. The facet numbers are unique, while source numbers are only unique within a facet, so the combination provides a unique ID for every source in the entire field. The 6′′ catalogue came from a single image, rather than a faceted one, but we use the facet layout to create consistently formatted unique IDs (and remove sources from the catalogue which are outside of the high-resolution images).

Filtering low SNR sources

To identify and remove low SNR sources where PyBDSF may have had difficulty estimating the local noise, we recalculate the noise from the RMS maps for every source at every resolution, in a 200×200 pixel box centered on the source location. This provides a better estimate for the RMS for every source. The reason we calculate the RMS for every single source using the RMS maps at every resolution is that the additional information at the other resolutions can also provide an indication of whether or not one would expect a source to be detected (e.g., if the peak intensity at 6′′ is less than 5σ5\sigma based on the noise at 1.2′′, we would not expect a detection at 1.2′′). This information is not used in the SNR cuts but is included in the catalogues. For the SNR cuts we use only the noise at the relevant resolution for each source.

We tried several combinations of criteria based on integrated flux density, peak intensity, and their associated uncertainties to filter out low-SNR sources. While using the total flux density to define the signal for SNR keeps the most sources (90% to 99%, depending on the resolution), this is at the cost of having a large number of false detections or improperly fitted sources. Particularly because wavelets were used in source detection, the higher resolution images often have low SNR sources with peak intensity to noise ratio less than unity, but a total flux density to noise ratio in the hundreds. PyBDSF struggles in these cases to find the difference between the background and the FWHM wings of the source. For this reason, we chose, similar to previous high-resolution LOFAR deep field data releases (Sweijen et al. [126]; de Jong et al. [41]; Escott et al. [51]), to use the peak intensity to define SNR, which provides a much cleaner sample at the cost of being less complete.

Applying a threshold requiring peak intensities to be at least 5 times the calculated local RMS, we find that 40% of the sources are filtered out at 0.3′′, 23% at 0.6′′, 27% at 1.2′′, and 20% at 6′′ (see Table 3). We also remove the corresponding Gaussian component(s) from the Gaussian catalogue. The largest removal fraction for the 0.3′′ catalogue is likely due to a combination of stronger smearing effects (see § 4.2) and the resolving power, since for example some of the faint star-forming galaxies are more likely to be resolved at higher resolution (see § 5.1). Given the substantial source filtering, and for reproducibility, we also release the “raw” PyBDSF outputs in case someone wishes to look up individual faint sources. However, we advise caution, as these catalogues include false detections and are unreliable for large population studies. Future work will explore how to improve the source detections across resolutions.

0.3′′0.6′′1.2′′6.0′′
Gaussians58 26144 10639 82071 412
SNR > 535 84234 16829 94358 821
Sources51 77939 41335 07959 993
SNR > 530 96530 17425 65647 882
Single28 48428 58523 81141 219
Multiple2 4811 5891 8456 663
No. rowsNo. sources
0.3″30 97830 965
0.6″30 35530 174
1.2″26 23925 656
6.0″48 76247 882
Combined63 84162 916

Table 3. The number of Gaussians and sources before and after removing sources where the peak intensity has SNR < 5, for all different resolutions. “Single” refers to single component sources and “multiple” to multiple component sources, according to PyBDSF.

Quantifying completeness requires separate simulations at every resolution (due to differences in surface brightness sensitivity and the wide range of source structure at different resolutions), which are beyond the scope of this paper. Such simulations would also need to account for time and bandwidth smearing, which is another factor that affects the completeness for the 0.3′′ and 0.6′′ images, given our peak-based SNR threshold (see § 4.2). Simulations would also provide the false positive

rate, but the fact that we aim to match across different resolutions already helps mitigate this. For example, a marginal detection at 0.3″ is likely a real source if it is also detected at 6″. Future work aiming to quantify the true false-detection rate should therefore focus on the unmatched sources at each resolution, and fold in PyBDSF’s wavelet decomposition.

Cross-matching resolutions

Next, we use positional matching based on the RA and DEC in the catalogues from each resolution. This requires setting a match radius. Typically the FWHM of the beam can be used, but when matching two different resolutions, there are two different FWHMs. After checking that indeed the larger FWHM contains all possible matches using the smaller FWHM, we use only the larger match radius. Figure 4 shows a schematic diagram which helps visualise this process. We first use the 0.3″ catalogue as a reference, and then repeat the process from 0.6″ (matching to 1.2″ and 6.0″), and repeat again starting from 1.2″ (matching to 6.0″). All matches are kept, as are sources where there are no matches at other resolutions. These three resulting catalogues are merged and we remove duplicates and rows whose detections are all contained in another row with detections at more resolutions. These are matches where either a unique single detection, unique pair, or unique triple match exists on a different row with more information.

Schematic diagram of how the match radius changes across different resolutions

Figure 4. A schematic diagram of how we change the match radius across different resolutions. This example is only for the 0.3″ catalogue, but we repeat this process starting from the 0.6″ and 1.2″ to account for all possible combination of resolutions.

In the multi-resolution catalogue, each row contains a unique combination of information across the four resolutions, and can have information at 1, 2, 3, or 4 resolutions, with null values for any resolution where there is not a detection. The differing point source and surface brightness sensitivities at different resolutions, combined with source structure, means that not every source is detected at every resolution. It is important to note that in the final catalogue, one row does not equal one source. For example, if a source has a single component at 6″ but breaks into two components at 0.3″, there will be two rows, one for each of the individual 0.3″ components, each carrying the information from the associated component at 6″ resolution. The complete set of information on this source therefore requires selecting both of these rows to get the information on the individual components.

We identify 1 115 locations for visual inspection. These are locations where sources with a single detection at a lower resolution break into multiple components at higher resolutions; we record a single common position (that from the lowest resolution) for inspection. These will be explored in future work.

Cataloguing summary

To summarise, the final combined catalogue has 63 841 rows with 62 916 unique sources. Table 4 summarises the number of rows and unique source IDs at each resolution. The source density is determined in the FoV (2.5×2.52.5 \times2.5 deg2^2) to be 4 954, 4 828, 4 105, and 7 661 per square degree for 0.3″, 0.6″, 1.2″, and 6.0″, respectively. The 0.3″ catalogue alone has, to our knowledge, the highest source density of any survey with sufficient resolution to identify AGN via their brightness temperature [87].

We show in Figure 5 the number of matches at other resolutions, for each resolution. Of the 62 916 unique sources, 50.5% are detected at more than one resolution and 25.5% at all four resolutions, which highlights the benefit of combining multiple resolutions. The 1.2′′ images have the lowest number of unique detections, as this resolution has the lowest point-source sensitivity and is a balance between the highest and lowest angular resolution in terms of surface brightness sensitivity. Figure 6 shows the fraction of unique sources as a function of distance from the phase centre for sources detected at 1, 2, 3, or 4 resolutions with the fraction of sources detected only at 6′′ for comparison. Within the inner 0.3 deg, about 40% of the sources are detected at a single resolution, mostly at 6′′ only, indicating that most single-resolution detections are faint or extended sources below the detection threshold at higher resolutions. Beyond ∼0.8 deg, the fraction of sources detected at all four resolutions decreases, while the fraction detected only at 6′′ increases, consistent with the increasing time and bandwidth smearing (see § 4.2) and the narrower primary beam of the international stations. The fractions of sources detected at two or three resolutions remain roughly constant with distance, as sources that lose for example their 0.3′′ detection towards the edge of the field move into these categories, while others move from them to the single resolution category.

Number of other resolutions to which sources are matched for each resolution

Figure 5. For each resolution, the number of other resolutions to which sources are matched.

Fraction of unique sources detected at one, two, three, or four angular resolutions as a function of distance from the phase centre

Figure 6. Fraction of unique sources detected at one, two, three, or four of the angular resolutions (labelled with 0.3′′, 0.6′′, 1.2′′, and 6′′) as a function of distance from the phase centre. Grey points show the fraction of sources detected only at 6′′. Horizontal error bars indicate the bin widths and vertical error bars the Poisson uncertainties.

Data quality

With our images and catalogues, we examine in this section several aspects that determine the quality of the data. These include the depth, smearing, confusion noise, and polarisation.

Depth

We achieve a central point-source sensitivity of 5.3 µJy beam−1^{-1} in facet 2 at 0.33′′×0.39′′ and an area-weighted RMS noise of 7.4 µJy beam−1^{-1} over the full 2.5×2.5 deg2^{2} field at 0.36′′×0.45′′ resolution. The central point-source sensitivity is a factor of two lower than that of the central point-source sensitivity from the 6′′ wide-field image by Shimwell et al. [114], who imaged 505 hours of LOFAR data. The improvement is due to the larger collecting area and higher sensitivity of the international LOFAR stations, as well as the fact that our images are not affected by confusion noise at sub-arcsecond resolution (see § 4.3), unlike imaging at 6′′ with LOFAR (e.g. Sabater et al. [111]; ); Shimwell et al. [114]). Compared to de Jong et al. [41], who imaged 32 hours of the same field at the same resolution, we achieve a factor of 2.6 improvement. This is consistent with the expected square-root scaling with total integration time. To compare the sensitivity improvements between imaging with a standard single 8-hour LOFAR observation and our 200-hour imaging, we show in Figure 7 images of the same sky area at both corresponding total integration times at two different resolutions. The deeper images in the right panels recover more diffuse emission from jets and reveals a larger number of faint background point sources, mainly corresponding to distant star-forming galaxies (e.g. Best et al. [9]).

Comparison between imaging a single LOFAR observation and 25 observations for two different sky areas

Figure 7. Comparison between imaging a single LOFAR observation (left) and 25 observations (right) for two different sky areas. The images in the top are restored to a circular beam of 1.5″, the images in the bottom are restored to circular beam of 0.4″. The images in the left panels are created with the observation with ID L686962. The colour map was generated with CMasher [137].

In Figure 8 we compare the point-source sensitivity from our deepest high-resolution image with other surveys’ deep images (Condon et al. [24]; ); Bock et al. [12]; ); Schinnerer et al. [112]; ); Lane et al. [72]; ); Helfand et al. [61]; ); Wayth et al. [156]; ); Smolčić et al. [122]; ); Intema et al. [64]; ); Shimwell et al. [117, 115, 114, 116]; Mauch et al. [80]; ); Lacy et al. [71]; ); Alger et al. (2020); ); van der Vlugt et al. [138]; ); Norris et al. [91]; ); Sabater et al. [111]; ); Hurley-Walker et al. [63]; ); Adams et al. [1]; ); de Gasperin et al. [34]; ); Groenveld et al. (2024); ); Duchesne et al. [48, 47]; ); Hale et al. [60]). As most of the deepest images cover a considerably smaller FoV, we use for this comparison the area-weighted RMS noise within the inner 1.5×1.5 deg2^{2} of our image, which is 5.8 µJy beam−1^{-1} and highlights the significant depth we reach in this work. We find in this figure that our image is the most sensitive radio image below 1.28 GHz. The comparison with higher frequency images depends on the assumed spectral index (α\alpha, where Sν∝ναS_{\nu} \propto\nu^{\alpha}). For a steep index of α=−1\alpha=-1, our image reaches an equivalent point-source sensitivity comparable to that of the MeerKAT DEEP2 image (Mauch et al. [80]), whereas for α∼−0.8\alpha\sim-0.8, typical of synchrotron-dominated radio sources (Böhme et al. [13]), the MeerKAT DEEP2 image remains deeper, while the image from the COSMOS-XS 3 GHz survey (Algera et al. [3]; ); van der Vlugt et al. [138]) reaches a more comparable equivalent point-source sensitivity. For α∼−0.5\alpha\sim-0.5, which An et al. (2024) found to be typical of star-forming galaxies between 150 and 610 MHz in ELAIS-N1, both of these images and the Chandra Deep Field South (CDFS) image from the MeerKAT International GHz Tiered Extragalactic Exploration Survey (MIGHTEE; Hale et al. [60]) all reach a better equivalent sensitivity than our image.

Sensitivity as a function of frequency for various surveys and deep images

Figure 8. Sensitivity as a function of frequency for various surveys and deep images, where the radius of each circle represents the square root of the beam size and the colour represents the survey area. “LOFAR-HD-EN1” corresponds to the deepest image presented in this paper, where we show the area-weighted RMS noise within the inner 1.5×1.51.5 \times1.5 deg2^{2} of the field (5.85.8 µJy beam−1^{-1}) to highlight the central depth in our highest resolution images. From the MIGHTEE images released by Hale et al. [60], the deepest CDFS-DEEP and Cosmic Evolution Survey (COSMOS) images are displayed. The dashed lines show the sensitivity of our image extrapolated to other frequencies using different spectral indices: α=−1\alpha= -1 for steep-spectrum sources, α=−0.8\alpha= -0.8 for a typical mix of LOFAR-detected radio sources according to Böhme et al. [13], and α=−0.5\alpha= -0.5 as measured for star-forming galaxies in ELAIS-N1 (An et al. 2024).

When we compare the deepest observations below an equivalent sensitivity at 150 MHz of 20 µJy beam−1^{-1}, we find the highest number of detections in the LOFAR-Deep-EN1 from Shimwell et al. [114], owing to its FoV being four times larger than ours. However, since our multi-resolution catalogue is combined with theirs for the central 2.5 × 2.5 deg2^{2} of their FoV, our catalogue has a higher source density. Our FoV is significantly larger than those of MeerKAT DEEP2, COSMOS-XS 3 GHz, and MIGHTEE CDFS-DEEP, yielding higher numbers of source detections than theirs, although our point-source sensitivity drops fast towards the edge of the field. This benefits science dependent on larger number statistics. In addition, our 0.3′′ images far surpass the resolution of every image and survey shown in Figure 8, adding scientific value that goes beyond depth alone (Morabito et al. [84]).

Smearing

Bandwidth and time smearing are key factors that degrade image quality and source detections. Bandwidth smearing causes sources to broaden radially, while time smearing produces blurring along the tangential direction relative to the phase centre. This results in a decrease in peak intensity. Additional smearing may also be introduced by ionospheric effects [67, 75], even after applying DD corrections. Furthermore, SVA can introduce additional smearing at the level of a few percent [40]. Although smearing conserves the integrated flux density of a source, the accompanying loss of peak intensity means that part of the emission from faint sources near the detection threshold falls below the noise, so their integrated flux densities are systematically underestimated. Since we have removed sources from our catalogue with low SNR values based on their peak intensities (see § 3.2), smearing may

also affect the completeness of our catalogue. Below we assess the various smearing effects on our data.

According to the formulas of Bridle & Schwab [18], for LOFAR datasets typically stored in the LTA with a time resolution of 1 sec and a channel width of 12.21 kHz, the peak intensity of a point source is already reduced by about a factor of two at 0.3″ resolution, 1.25 degrees from the phase centre. On top of this, as described in § 2.1, 11 of the 25 observations that we used to create our wide-field image were recorded with a 2 sec time resolution. As a result, the total time and bandwidth smearing in our images becomes significant towards the outer edge of the field (e.g. [126, 41]). In addition to the intrinsic smearing, we also apply further averaging to the datasets from each facet prior to imaging [42]. These averaging factors are calculated under the assumption that the already present smearing at the edge of a 2.5×2.52.5 \times2.5 deg2^{2} LOFAR FoV at 1 sec and 12.21 kHz is acceptable, as that keeps the data volumes manageable.12 Hence, a facet that fits in a 0.8×0.80.8 \times0.8 deg2^{2} box will be averaged by a factor of 3 in time and frequency. This means that the bandwidth and time smearing from the full dataset is carried over towards the edges of the facets after averaging.

To illustrate the impact of time and bandwidth smearing effects across the field, Figure 9 shows the estimated peak intensity reduction for baselines corresponding to the three angular resolutions (see § 2.3). We find that, as expected, the 0.3′′0.3^{\prime\prime} images suffer the most from smearing, with peak intensity losses of up to 80% at the edge of the field, whereas the lower resolution images are less affected, as time and bandwidth smearing scale with angular resolution. We can also see the additionally introduced smearing at the edges of each facet. As an alternative check of the effects of time and bandwidth smearing, de Jong et al. [41] measured the ratio of peak intensity to integrated flux density for compact sources as a function of distance from the phase centre, using a subset of the same ELAIS-N1 observations with a similar mixture of 1 sec and 2 sec data. The trends they found agree with our predictions in Figure 9. However, these figures should be interpreted with caution as the

Leftover peak intensities in final wide-field images due to time and bandwidth smearing across the FoV

Figure 9. Leftover peak intensities in our final wide-field images due to time and bandwidth smearing across the FoV, according to the formulas from Bridle & Schwab [18], for each mosaic at their restored resolution: 0.36′′ × 0.45′′ (left), 0.58′′ × 0.62′′ (centre), and 1.0′′ × 1.5′′ (right). These smearing maps take into account the averaging after facet-subtraction and capture the weighted combination of observations recorded at 1 sec and 2 sec (§ 2.1).

SNR threshold biases the source selection and sources are more likely to be resolved at higher resolution, which lowers their peak intensities independently of smearing (see § 5.1). In Appendix § 11 we briefly discuss how finer time and frequency resolution in the initial data reduces time and bandwidth smearing, at increased computational cost.

To estimate the impact of ionospheric smearing, we re-imaged the most central facet (facet 2) and a facet at the edge of the field (facet 10) for one observation corresponding to calm ionospheric conditions (L769407) and one under more disturbed ionospheric conditions (L686948). We selected these observations by visually inspecting the self-calibration images and calibration solutions from all our processed observations. We then selected compact sources (labelled as single components by PyBDSF) with peak intensities exceeding ten times the local RMS noise and used the ratio of peak intensity to integrated flux density as a measure of smearing. For the observation with the more disturbed ionosphere, we find that this ratio is on average 15 ±\pm 5% lower, indicating that ionospheric conditions do introduce peak intensity losses, though at a lower level than the intrinsic time and bandwidth smearing away from the phase centre for observations stored at 1 sec and 12.21 kHz and imaged at 0.3′′0.3^{\prime\prime}.

SVA can also introduce smearing effects due to precession, nutation, and annual aberration, particularly since our observations span a three-year interval (affecting precession and nutation terms) and were taken at different times of the year (introducing annual aberration effects). As mentioned in § 2.3, we therefore over-sampled the time axis by a factor of 2, which reduces these effects to only a few percent at the edge of the field (de Jong et al. [40]). This contribution is negligible compared to the much larger smearing caused by the time and frequency averaging mentioned above.

Confusion noise

Confusion noise arises when faint background radio sources close to or below the detection threshold become so numerous that across the FoV multiple sources fall within a single synthesised beam. In this regime, each faint background source contributes a scaled PSF, and the superposition of many such contributions adds a substantial non-Gaussian noise-like background in addition to the Gaussian thermal noise. This introduces limits to the obtained sensitivity, which is set by the angular resolution and the underlying source counts ([22, 23, 162, 80]). Confusion noise is known to be a dominant limitation for deep low-resolution wide-field imaging. In particular, LOFAR observations using only the Dutch stations start to be dominated by confusion noise when imaging with several hundred hours of data, such that further increases in observing time yield only marginal sensitivity improvements (e.g. [111, 114]).

Since the confusion noise scales approximately as the square of the beam size ([22]), an image at 0.3′′0.3^{\prime\prime} resolution can be made about 400 times deeper than one at 6′′6^{\prime\prime} before confusion becomes limiting. We adopt 0.3′′0.3^{\prime\prime} as a representative resolution for simplicity, although the restoring beams of our highest-resolution facet images are larger (see Table 2). Note that at 144 MHz, LOFAR can reach a resolution of up to ∼0.2′′\sim0.2^{\prime\prime} with its longest baselines ([85]), and the addition of more distant international stations with LOFAR2.0 may push this further (e.g. Dechyev & Kozarev 2023), which would further reduce the confusion noise. Shimwell et al. [114] reached in their 6′′6^{\prime\prime} image a central RMS of 10.7 µJy beam−1^{-1}, against a thermal noise of 7.5 μJy beam−1\mu\mathrm{Jy}\,\mathrm{beam}^{-1}, implying a confusion noise contribution of 10.72−7.52=7.6 μJy beam−1\sqrt{10.7^2 - 7.5^2} = 7.6\,\mu\mathrm{Jy}\,\mathrm{beam}^{-1}. Scaled to 0.3′′, confusion noise would therefore only become relevant when imaging close to a thermal noise of 20 nJy beam−1\mathrm{nJy}\,\mathrm{beam}^{-1}. When accounting for the higher point-source sensitivity of the international LOFAR stations, reaching that thermal noise would require on the order of 20 million hours of observing time. This total integration time extends well beyond the operational lifetime of LOFAR, even if the point-source sensitivity were an order of magnitude better. Hence, further extensions of this work to even deeper imaging (see § 5.2) will not be constrained by confusion noise. However, other effects, such as sidelobe confusion from sources outside the FoV, residual DD calibration errors, and dynamic range limitations around bright sources, are likely to limit the achievable depth before the thermal noise limit is reached.

Polarisation

As described in § 2.2, in step 3 of the calibration process we corrected for polarisation leakage and for differences between the right- and left-handed circular polarisations using an unpolarised calibrator. In step 9, we aligned the polarisation angles across the different observations using a known linearly polarised source (7C 1604+5447), thereby enabling deep polarisation imaging. This follows the same procedure described in van Weeren et al. [144]. Since these are the deepest data for polarisation studies below 200 MHz, we will briefly discuss the quality and refer to future work for more in-depth analysis (de Jong et al. in prep.).

After alignment of the different observations we find for our leakage calibrator source (ICRFJ160607.6+552135) an instrumental leakage of 0.12% for linear polarisation (Q,U) and 0.02% for circular polarisation (V), which is approximately a factor of two better than the typical leakage reported for LoTSS at 6′′ resolution (Shimwell et al. [115]). To demonstrate deep polarisation imaging, we performed RM-synthesis (Brentjens & de Bruyn [17]) with the RM-Tools package (Purcell et al. [106]; Van Eck et al. [141]) on Stokes Q and U images from our polarisation calibrator, using 8 hours and 200 hours of LOFAR data, and following the same imaging settings as van Weeren et al. [144]. This gives us the Faraday dispersion functions (FDF) at the different linearly polarised regions as illustrated in Figure 10. We use contour levels of 13σ13\sigma to stay far above the false-detection level of peaks in the polarised intensity cube across the Faraday depth range searched. This figure show that the FDF is considerably less noisy in the 200-hour dataset than in the single 8-hour observation, demonstrating the effectiveness of the alignment technique of van Weeren et al. [144] for deep linear polarisation imaging. In the Stokes-I images with linearly polarised emission contours, the hotspots appear more prominently linearly polarised in the deeper images, and the extension denoted by “Lobe S” extends more towards the central host galaxy in the deep image. The RM value of “Hotspot S” is 6.29±0.01 rad m−26.29 \pm0.01\,\mathrm{rad}\,\mathrm{m}^{-2}, while slightly lower values are measured for “Hotspot N” (6.06±0.07 rad m−26.06 \pm0.07\,\mathrm{rad}\,\mathrm{m}^{-2}) and “Lobe S” (5.66±0.05 rad m−2^{-2}). We also find polarisation fractions of 2.10±0.02% and 0.54±0.07% for the northern and southern hotspots, respectively. These values are all consistent with what was reported by van Weeren et al. [144].

Stokes I images of the polarisation calibrator with FDF spectra at 0.3′′ resolution shown with 8 and 200 hours of LOFAR data

Figure 10. Stokes I images of the polarisation calibrator (7C 1604+5447) with FDF spectra at 0.3′′ resolution shown with 8 and 200 hours of LOFAR data, with polarised intensity contours overlaid at levels of 13×σ×(1,2,4,8,…)13 \times\sigma\times(1, 2, 4, 8, \ldots), where in the 8-hour image σ=30 μJy beam−1\sigma= 30\,\mu\mathrm{Jy}\,\mathrm{beam}^{-1} and in the 200-hour image σ=6 μJy beam−1\sigma= 6\,\mu\mathrm{Jy}\,\mathrm{beam}^{-1}. Following van Weeren et al. [144] and for visualisation purposes, the spectra of “Hotspot N” and “Lobe S” were scaled by factors of 5 and 2, respectively.

A more detailed study of the linear polarisation properties of the ELAIS-N1 dataset presented in this paper will be described in de Jong et al. (in prep.) and will include a quantification of the polarisation leakage across the field, an investigation of beam depolarisation effects, and a search for newly detected polarised sources. This builds upon the work by Piras et al. [103, 104] at 6″ resolution. Stokes V will also be analysed in future work.

Discussion

With the data presented in the previous sections, we compare in this section source detections across different resolutions. We conclude by touching upon a future outlook for deeper imaging with the SKA and LOFAR2.0.

Source detections across resolutions

There is no single “optimal” angular resolution to perform all types of science, as point-source sensitivity is higher at higher resolutions, whereas the surface brightness sensitivity is better at lower resolutions. Hence, some of the faint compact sources may be detected at high resolution but not at low resolution, while extended sources may fall within the detection limits of a low resolution image but remain undetected in a high resolution image. Other sources that are blended into a single source at a low resolution may be resolved into multiple components at higher resolution. This makes higher resolutions best suited for accurate astrometry and optical identifications (see § 2.5), characterising and resolving detailed source structures (e.g. Varenius et al. [149]; Ramírez-Olivencia et al. [108, 107]; Timmerman et al. [133]; Mahatma et al. [77]; Gloudemans et al. [57]), or separation of different compact source components (e.g. AGN vs. star-formation; Morabito et al. [87, 86]), whereas lower resolutions are better for detecting large extended structures (e.g. Govoni et al. [58]; van Weeren et al. [143, 147]; Botteon et al. [16, 15]; de Jong et al. [39]) or giant radio galaxies (e.g. Dabhade et al. [30]; Oei et al. [95]; Mostert et al. [88]). These differences in source detections for different angular resolutions explain why the combined multi-resolution catalogue has a much higher source count compared to the individual catalogues for each image resolution (see Table 4).

Figure 11 compares the apparent source counts as a function of the integrated flux density across the different angular resolutions. As expected, we find in the upper panel that the sources detected in the 0.3″ images include the faintest radio sources, reflecting the higher point-source sensitivity at the highest resolution. However, at flux densities around 70 μJy and 100 μJy, the 6″ and 0.6″ resolution images, respectively, detected more sources than the 0.3″ image. To limit the role of smearing in this resolution comparison, we show in the lower panel of Figure 11 the same histograms for sources within 0.5 deg of the phase centre, where smearing is small (see Figure 9). In this region, the 6″ and 0.6″ images still detect more sources than the 0.3″ image above ∼70 μJy, while the deficit of 0.3″ detections relative to 6″ and 0.6″ is only slightly smaller than for the full field. This indicates that smearing contributes but that the cross-over is also caused by the loss of surface brightness sensitivity at higher resolution, combined with the fraction of the star-forming galaxy population – the dominant source population at these flux densities (Best et al. [9]) – that is resolved at higher resolution, lowering their peak intensities below the detection threshold. For instance, at z∼1z \sim1, typical star-forming galaxies are large enough to be resolved at ∼0.3″ but remain point sources at 0.6″ (e.g. van der Wel et al. [139]; Ward et al. [155]), so this population may be disproportionately lost from the 0.3″ catalogue. At the higher integrated flux density end, where radio galaxies start to become the more dominant population (Best et al. [9]), the source counts are ordered by resolution, with the lower resolutions detecting more sources than higher resolutions. This is mainly due to the loss of surface brightness sensitivity towards higher resolutions, which either lowers the measured integrated flux density or causes sources to be undetected because most of their extended emission, e.g. from radio jets, is resolved out below our detection thresholds.

Source count histograms as a function of integrated flux density at four angular resolutions

Figure 11. Source count histograms as a function of integrated flux density at four angular resolutions for all unique sources (see Table 4) in the full field (upper) and within 0.5 deg of the phase centre (lower). The 0.3″, 0.6″, and 1.2″ resolutions correspond to the images and catalogues produced in this work, while the 6″ sources are from the deep imaging of Shimwell et al. [114]. Dashed vertical lines mark the median integrated flux density at each resolution. Note: in the upper panel, the medians at 0.6″ and 6″ nearly coincide.

To estimate the loss of surface brightness sensitivity, we compare the median ratio of the integrated flux densities at 0.3″ and the different lower resolutions as a function of the integrated flux density for single component sources detected at both resolutions with integrated flux densities exceeding five times the PyBDSF uncertainty. The selection of single component sources provides a lower limit of the loss of surface brightness sensitivity.

Median ratio of integrated flux densities at 0.3″ and lower resolution as a function of integrated flux density

Figure 12. Median ratio of the integrated flux densities at 0.3″ and lower resolution as a function of the integrated flux density at the lower resolution, for single component sources detected at both resolutions with integrated flux densities exceeding five times the PyBDSF uncertainty. Shaded areas show the bootstrap uncertainties on the median in each flux density bin, and dashed horizontal lines the median ratio over all flux densities for each pair. The dotted line marks a ratio of unity.

ity across the whole population of radio sources. Over all flux densities, we find median ratios of S0.3/S0.6=0.97±0.01S_{0.3}/S_{0.6} = 0.97 \pm0.01, S0.3/S1.2=0.89±0.01S_{0.3}/S_{1.2} = 0.89 \pm0.01, and S0.3/S6.0=0.84±0.01S_{0.3}/S_{6.0} = 0.84 \pm0.01. The uncertainties are the standard error of the median, obtained by bootstrap resampling. The ratios decrease as the resolution difference increases, confirming the loss of surface brightness sensitivity towards higher angular resolution.

Figure 12 shows how these integrated flux density ratios vary with integrated flux density. This behaviour reflects three effects: changing source populations, the higher point-source sensitivity at higher resolutions, and the better surface brightness sensitivity at lower resolutions. At the lowest integrated flux densities, where star-forming galaxies are the dominant source population ([9]), the ratios approach or exceed unity, most strongly for the 1.2″ comparison. This is because the 0.3″ images have the best point-source sensitivity and the 1.2″ images the poorest, so faint compact star-forming galaxies are better recovered at high resolution. Towards ∼500\sim500–800800 µJy the ratios decline. In this regime star-forming galaxies still dominate, but the brighter ones tend to be larger, so a larger fraction of their emission may be partly resolved out in the 0.3″ images, and so the loss of surface brightness sensitivity at high resolution has its largest effect here on the ratios. At higher flux densities the ratios rise again as radio-loud AGN become the dominant population. Because we selected single-component sources, this population is biased towards compact AGN cores, which, owing to their small angular sizes, are expected to have similar flux densities at all resolutions. The minimum of the ratio shifts to higher flux densities as the difference in resolution increases. This is partly because sources have higher integrated flux densities at lower resolution, and partly because the emission resolved out on larger angular scales comes

Future deeper imaging

While confusion noise will not be a limitation for deeper imaging with LOFAR at sub-arcsecond resolutions (see § 4.3), it will become an issue for the upcoming Square Kilometre Array Low (SKA-Low; [45]). The SKA-Low will operate at similar frequencies to LOFAR but with significantly higher sensitivity ([162, 105]). This instrument is therefore expected to become confusion-limited within a few hours (depending on the array configuration), owing to its current maximum baseline length of 73.4 km.13 As a result, SKA-Low will not be able to surpass the sensitivities achieved in this work unless its maximum baselines are extended ([132]).

For LOFAR2.0, the ongoing upgrade of LOFAR (e.g. [50, 102, 153]), deeper observations are planned. One of the future planned large projects will be the LOFAR2.0 Ultra Deep Observation (LUDO; PIs: L.K. Morabito & P.N. Best),14 which aims to image ∼3200 hours of LOFAR data of the Euclid Deep Field North (EDFN) and achieve a depth of 2 µJy beam−1^{-1} across the central 3.7 deg2^{2}. This follows on the shallower LOFAR surveys of the same field at 6″ by [14] and at 0.3″ by de Jong et al. (subm.). The combination of the deep optical and near-infrared data from the Euclid Deep Survey ([52]) with ultra-deep high-resolution low-frequency radio observations from LOFAR will provide a powerful dataset for studying the relationship between radio and optical/near-infrared emission from galaxies across a wide range of redshifts ([10, 76]).

Enabling the imaging of thousands of hours of LOFAR data within feasible wall time will benefit from further reductions in the processing cost of PILOT. A large part of this work will be carried out by the LOFAR Enhanced Network for Sharp Surveys (LENSS) project.15 Potential improvements for LENSS are outlined by de Jong et al. [42], who highlight in particular substantial gains achievable through porting parts of the calibration and imaging pipeline to GPUs (e.g. [135, 136, 150, 134, 43]).

Summary and conclusions

We have presented the release of ultra-deep wide-field sub-arcsecond images and a multi-resolution catalogue of the ELAIS-N1 field using 200 hours of LOFAR HBA observations. Our main results can be summarised as follows:

  • By combining recent advances in calibration, imaging, and data processing strategies, we demonstrate that imaging 400 TB of visibility data to produce high-fidelity sub-arcsecond resolution images over a 2.5 × 2.5 deg2^{2} field at metre wavelengths is now feasible within an acceptable computational time below 1 million CPU core hours.

  • We achieve a highest central point-source sensitivity of 5.3 µJy beam−1^{-1} at a resolution of 0.33″×0.39″ and an area-weighted sensitivity of 7.4 µJy beam−1^{-1} at 0.36″×0.45″, representing the deepest radio images below 1.28 GHz and among the overall deepest radio images to date. In the centre of the field, it is nearly a factor of two improvement over the previous deepest LOFAR wide-field image of the

same field at lower angular resolution [114] and a factor of ~2.5 improvement compared to the previous data release of ELAIS-N1 at sub-arcsecond resolution [41]. After calibration, the astrometry yields an accuracy of about ΔRA=22±71\Delta\mathrm{RA} = 22 \pm71 mas and ΔDEC=−17±41\Delta\mathrm{DEC} = -17 \pm41 mas, which, using 95 Gaia quasars, we further improved to ΔRA=−3±27\Delta\mathrm{RA} = -3 \pm27 mas and ΔDEC=2±17\Delta\mathrm{DEC} = 2 \pm17 mas for 22 of our 24 facets. We also applied facet-to-facet flux density scale corrections to align our measurements with those of [114] at 6′′ resolution.

  • We produced a multi-resolution catalogue by combining source detections from our 0.3′′, 0.6′′, and 1.2′′ images with those identified in the 6′′ image of [114] within our FoV. The resulting catalogue contains 62 916 unique sources. This is one of the largest numbers of radio source detections from a single telescope pointing, highlighting LOFAR’s capability for ultra-deep imaging across large FoVs. The combination of high source density and sub-arcsecond resolution makes this catalogue unique compared to any other radio catalogue.

  • We show that time and bandwidth smearing remain a limitation at the highest angular resolution, particularly towards the edges of the FoV, where peak intensities are reduced by up to 80% of their intrinsic values at 0.3′′. This has a disproportionate effect on the completeness of source detections at 0.3′′ compared to the other lower resolutions. A significant contribution to this effect arises from additional time averaging in roughly half of the ELAIS-N1 observations. Smearing effects introduced by ionospheric conditions and by SVA to combine observations in visibility space are comparatively small.

  • Our high-resolution image is not limited by confusion noise and will not become confusion limited with increasing observing time, in contrast to lower-resolution imaging with LOFAR’s Dutch-only baselines and future SKA-Low observations. This provides a unique capability to explore increasingly faint radio source populations while having the resolution required to study structures such as jets, lobes, and compact components in distant galaxies.

  • After leakage calibration, we measure linear polarisation leakages of 0.12% and circular polarisation leakages of 0.02% on the leakage calibrator, which is about a factor of two better than the typical leakage in LoTSS at 6′′ resolution. We also recover more linear polarisation across the lobe from the polarisation angle alignment calibrator. This demonstrates the potential for future deep polarisation studies at sub-arcsecond resolution, enabling the detection of fainter polarised emission with increasing total integration time.

Our results establish a new benchmark for low-frequency radio imaging in terms of the combination of depth, resolution, and survey area. This work lays the foundation for future ultra-deep surveys with LOFAR2.0, such as the planned LUDO survey, which will extend these techniques to the imaging of several thousand hours of LOFAR observations. Overall, this work demonstrates that ultra-deep high-resolution imaging at metre wavelengths with LOFAR is now reaching a mature and scalable regime, opening new windows for studying the deep radio sky at the lowest frequencies.

Data availability

The input visibility data that were processed in this work are publicly available from the LOFAR LTA (https://lta.lofar.eu). The images and catalogues produced in this work have been deposited, with data provenance, in the SURF Data Repository (SDR) and can be accessed without restrictions or registration at https://repository.surf.nl/records/xq:qr-cdf21. The SDR guarantees a retention period of at least 5 years, after which the retention period will be extended or the data will be moved to another public repository. The catalogues and images will also be made available through the ASTRON Virtual Observatory service, https://vo.astron.nl. The data have been enriched with metadata to improve reproducibility. These efforts are part of the OSCARS project [36] and ensure that the resulting data products comply with the FAIR principles [158, 31].

Acknowledgements

The authors thank C. Schreiber for useful conversations about cataloguing. The authors acknowledge the OSCARS project, which has received funding from the European Commission’s Horizon Europe Research and Innovation programme under grant agreement No. 101127951. RT is grateful for support from the UKRI Future Leaders Fellowship (grant MR/Y020405/1). This work was supported by the STFC [grants ST/T000244/1, ST/W002406/1]. FS appreciates the support of STFC [ST/Y004159/1]. RK acknowledges support from the Leverhulme Trust through a Leverhulme Early Career Fellowship. CLH acknowledges support from the Science and Technology Facilities Council (STFC) through grant ST/Y000951/1. This work is part of the project CORTEX (NWA.1160.18.316) of the research programme NWA-ORC which is (partly) financed by the Dutch Research Council (NWO). This work is co-funded by the EGI-ACE project (Horizon 2020) under Grant number 101017567. This publication is part of the project Deep high-resolution LOFAR imaging with file number 203.040 of the research programme Computing Time on National Computing Facilities which is (partly) financed by the Dutch Research Council (NWO). This publication is part of the project Uncovering the wide high-resolution sky with LOFAR with file number 2025.043 of the research programme Computing Time on National Computer Facilities which is (partly) financed by the Dutch Research Council (NWO) under the grant https://doi.org/10.61686/JACAQ20438. This publication is part of the project Probing the Formation of Galaxy Clusters and their Progenitors with High-Resolution Radio Imaging with file number VI.C.252.058 of the NWO Talent Programme which is (partly) financed by the Dutch Research Council (NWO) under the grant https://doi.org/10.61686/AZV0I80686. The public release of the data associated with this work was made possible by funding from the Astrophysics Centre for Multi-messenger studies (ACME) project. ACME project has received funding from the European Union’s Horizon Europe Research and Innovation programme under Grant Agreement No 101131928. LOFAR is the Low Frequency Array designed and constructed by ASTRON [142]. It has observing, data processing, and data storage facilities in several countries, which are owned by various parties (each with their own funding sources), which are collectively operated by the LOFAR ERIC under a joint scientific policy. The LOFAR resources have benefited from the following recent major funding sources: CNRS-INSU, Observatoire de Paris and Université d’Orléans, France; BMBF, MIWF-NRW, MPG, Germany; Science Foundation Ireland (SFI), Department of Business, Enterprise and Innovation (DBEI), Ireland; NWO, The Netherlands; The Science and Technology Facilities Council, UK; Ministry of Science and Higher Education, Poland; The Istituto Nazionale di Astrofisica (INAF), Italy. This research made use of the Dutch national e-infrastructure with support of the SURF Cooperative (e-infra 180169) and the LOFAR e-infra group. The Jülich LOFAR Long Term Archive and the German LOFAR network are both coordinated and operated by the Jülich Supercomputing Centre (JSC), and computing resources on the supercomputer JUWELS at JSC were provided by the Gauss Centre for Supercomputing e.V. (grant CHTB00) through the John von Neumann Institute for Computing (NIC). This research made use of the University of Hertfordshire high-performance computing facility and the LOFAR-UK computing facility located at the University of Hertfordshire and supported by STFC [ST/P000096/1], and of the Italian LOFAR IT computing infrastructure supported and operated by INAF, and by the Physics Department of Turin University (under an agreement with Consorzio Interuniversitario per la Fisica Spaziale) at the C3S Supercomputing Centre, Italy. This research is part of the project LOFAR Data Valorization (LDV) project numbers 2020.031, 2022.033, and 2024.047 of the research programme Computing Time on National Facilities using SPIDER that is (co-)funded by the Dutch Research Council (NWO), hosted by SURF through the call for proposals of Computing Time on National Facilities. This work made use of Flocs (https://github.com/FLOCSoftware) for (part of) its data reduction.

Flux density scale across wavelengths

In Figure A.1 we plot the spectra of several compact sources across the ELAIS-N1 FoV to demonstrate the accuracy of our final flux density scale. We used data from the Pi GHz Sky Survey (PiGSS Croft et al. [28]), NRAO VLA Sky Survey at 1.4 GHz (NVSS; Condon et al. [24]), the GMRT at 610 MHz (Garn et al. [55]), the Westerbork Northern Sky Survey at 325 MHz (WENSS; Rengelink et al. [109]), and an image obtained after reducing data from an LBA observation at 54 MHz from the ELAIS-N1 field with the LiLF16 calibration pipeline (de Gasperin et al. [35, 33, 32]). We also adopt a conservative uncertainty of 20% on the integrated flux density scale, following LoTSS (Shimwell et al. [115]).

Dutch versus European LOFAR

To demonstrate the increased resolving power of the full European LOFAR at different resolutions, Figure B.1 presents selected regions imaged for the 0.3″, 1.2″, and 6″ images. These comparisons highlight how radio galaxies that appear as compact, blob-like sources at 6″ are fully resolved into complex morphologies at sub-arcsecond resolution. With the depth achieved from 200 hours of LOFAR observations, most of the faint extended emission seen at 6″ and 1.2″ resolution is also recovered in the 0.3″ images. This demonstrates that much of the low surface brightness emission is at these depths also captured at high resolutions.

Smearing reductions

Time and bandwidth smearing remain significant issues (see § 4.2), particularly for science cases where the recovery of accurate morphologies is essential, such as weak lensing (e.g. Liu et al. [74]) or precise source size measurements (e.g. Sweijen et al. [125]). One straightforward strategy for reducing smearing would be to increase the time and frequency resolution of the data in the LTA by a factor of two. This would improve the maximum peak-intensity recovery at 0.3″ from 50% to 85% (Bridle & Schwab [18]). However, this change would also increase the number of visibilities by a factor of four. Therefore, adopting baseline-dependent averaging (BDA; e.g. Cotton [27, 26]; Wijnholds et al. [157]), in which shorter baselines are averaged more than longer ones, throughout the data-processing pipeline would be in this case essential (e.g. Atemkeng et al. [7]; de Jong et al. [42]). Enabling BDA in combination with Dynamical Statistical Compression (Dysco; Offringa [96]) or Simulated Signal Compression (Sisco; Offringa & van Weeren [99]) is also important to keep data volumes manageable.

To reduce the remaining smearing, the imaging software could be extended to apply smearing corrections by modelling the smeared PSF. Such corrections are already implemented in several existing imagers (Noordam & Smirnov [90]; Tasse et al. [129]). Applying such corrections on a per-facet basis could mitigate the smearing at the boundaries of each facet, observed in the left panel of Figure 9. More comprehensive PSF modelling that accounts for both the intrinsic smearing present in the input data and the additional averaging introduced during the construction of the facet datasets could further reduce smearing across the full image.

Fitted spectra of bright compact sources across the field after flux density scaling

Figure A.1. Fitted spectra of bright compact sources across the field after flux density scaling (see § 2.4). The red star corresponds to the ELAIS-N1 data presented in this work.

Comparison between 0.3″, 1.2″, and 6″ resolution images for different sky areas containing various resolved radio galaxies

Figure B.1. Comparison between 0.3′′0.3^{\prime\prime}, 1.2′′1.2^{\prime\prime}, and 6′′6^{\prime\prime} resolution images for different sky areas containing various resolved radio galaxies.

References

  1. [1]Adams, E. A. K., Adebahr, B., de Blok, W. J. G., et al. 2022, A&A, 667, A38
  2. [2]Aihara, H., Armstrong, R., Bickerton, S., et al. 2018, PASJ, 70, S8
  3. [3]Algera, H. S. B., van der Vlugt, D., Hodge, J. A., et al. 2020, ApJ, 903, 139
  4. [6]Arras, P., Reinecke, M., Westermann, R., & Enßlin, T. A. 2021, A&A, 646, A58
  5. [7]Atemkeng, M., Smirnov, O., Tasse, C., et al. 2018, MNRAS, 477, 4511
  6. [8]Bertin, E., Mellier, Y., Radovich, M., et al. 2002, in Astronomical Society of the Pacific Conference Series, Vol. 281, Astronomical Data Analysis Software and Systems XI, ed. D. A. Bohlender, D. Durand, & T. H. Handley, 228
  7. [9]Best, P. N., Kondapally, R., Williams, W. L., et al. 2023, MNRAS, 523, 1729
  8. [10]Bisigello, L., Giulietti, M., Prandoni, I., et al. 2025, The Open Journal of Astrophysics, 8, 101
  9. [12]Bock, D. C.-J., Large, M. I., & Sadler, E. M. 1999, AJ, 117, 1578
  10. [13]Böhme, L., Schwarz, D. J., de Gasperin, F., Röttgering, H. J. A., & Williams, W. L. 2023, A&A, 674, A189
  11. [14]Bondi, M., Scar​​amella, R., Zamorani, G., et al. 2024, A&A, 683, A179
  12. [15]Botteon, A., Shimwell, T. W., Cassano, R., et al. 2022, A&A, 660, A78
  13. [16]Botteon, A., van Weeren, R. J., Brunetti, G., et al. 2020, MNRAS, 499, L11
  14. [17]Brentjens, M. A. & de Bruyn, A. G. 2005, A&A, 441, 1217
  15. [18]Bridle, A. H. & Schwab, F. R. 1999, in Astronomical Society of the Pacific Conference Series, Vol. 180, Synthesis Imaging in Radio Astronomy II, ed. G. B. Taylor, C. L. Carilli, & R. A. Perley, 371
  16. [19]Briggs, D. S. 1995, PhD thesis, New Mexico Institute of Mining and Technology
  17. [20]Chaturvedi, P., Sinha, A., Datta, A., & Saha, K. 2026, MNRAS, 550, stag1291
  18. [21]Cochrane, R. K., Kondapally, R., Best, P. N., et al. 2023, MNRAS, 523, 6082
  19. [22]Condon, J. J. 1974, ApJ, 188, 279
  20. [23]Condon, J. J., Cotton, W. D., Fomalont, E. B., et al. 2012, ApJ, 758, 23
  21. [24]Condon, J. J., Cotton, W. D., Greisen, E. W., et al. 1998, AJ, 115, 1693
  22. [25]Cornwell, T. J. 2008, IEEE Journal of Selected Topics in Signal Processing, 2, 793
  23. [26]Cotton, W. 2009, Effects of baseline dependent time averaging of uv data, Tech. rep., National Radio Astronomy Observatory
  24. [27]Cotton, W. D. 1986, in Synthesis Imaging, ed. R. A. Perley, F. R. Schwab, & A. H. Bridle, 123–136
  25. [28]Croft, S., Bower, G. C., & Whysong, D. 2013, ApJ, 762, 93
  26. [29]Crusoe, M. R., Abeln, S., Iosup, A., et al. 2022, Communications of the ACM, 65, 54
  27. [30]Dabhade, P., Röttgering, H. J. A., Bagchi, J., et al. 2020, A&A, 635, A5
  28. [31]David, R., Schmidt-Tremmel, F., Carboni, N., et al. 2025, Building Capacity for FAIR and Open Science: Insights from the World Café sessions at the OSCARS 1st AGM, version v1, published Dec 8, 2025
  29. [32]de Gasperin, F., Brunetti, G., Brüggen, M., et al. 2020, A&A, 642, A85
  30. [33]de Gasperin, F., Dijkema, T. J., Drabent, A., et al. 2019, A&A, 622, A5
  31. [34]de Gasperin, F., Edler, H. W., Williams, W. L., et al. 2023, A&A, 673, A165
  32. [35]de Gasperin, F., Mevius, M., Rafferty, D. A., Intema, H. T., & Fallows, R. A. 2018, A&A, 615, A179
  33. [36]de Jong, J. 2026, COPLI poster - OSCARS 2nd AGM 2026DOI
  34. [37]de Jong, J., Schlimbach, R. J., Veenk​​ind, L., Sweijen, F., & Botteon, A. 2026, j​​urgen93/lofar_helpers: lofar_helpers, Zenodo, version v2.0.0
  35. [38]de Jong, J. M. G. H. J., Röttgering, H. J. A., Kondapally, R., et al. 2024a, A&A, 683, A23
  36. [39]de Jong, J. M. G. H. J., van Weeren, R. J., Botteon, A., et al. 2022, A&A, 668, A107
  37. [40]de Jong, J. M. G. H. J., van Weeren, R. J., Dijkema, T. J., et al. 2025a, A&A, 694, A98
  38. [41]de Jong, J. M. G. H. J., van Weeren, R. J., Sweijen, F., et al. 2024b, A&A, 689, 542
  39. [42]de Jong, J. M. G. H. J., Veefkind, L., van Weeren, R. J., et al. 2025b, MNRAS, 542, 3253
  40. [43]De Rubeis, E., Lacopo, G., Gheller, C., Tornatore, L., & Taffoni, G. 2025, Astronomy and Computing, 50, 100895
  41. [45]Dewdney, P. E., Hall, P. J., Schilizzi, R. T., & Lazio, T. J. L. W. 2009, IEEE Proceedings, 97, 1482
  42. [46]Dijkema, T. J., Nijhuis, M., van Diepen, G., et al. 2023, DP3: Streaming processing pipeline for radio interferometric data, Astrophysics Source Code Library, record ascl:2305.014
  43. [47]Duchesne, S., Ross, K., Thomson, A. J. M., et al. 2025, PASA, 42, 38
  44. [48]Duchesne, S. W., Thomson, A. J. M., Pritchard, J., et al. 2023, PASA, 40, e034
  45. [49]Duncan, K. J., Kondapally, R., Brown, M. J. I., et al. 2021, A&A, 648, A4
  46. [50]Edler, H. W., de Gasperin, F., & Rafferty, D. 2021, A&A, 652, A37
  47. [51]Escott, E. L., Morabito, L. K., Sweijen, F., et al. 2026, MNRAS, 547, stag274
  48. [52]Euclid Collaboration, Scaramella, R., Amiaux, J., et al. 2022, A&A, 662, A112
  49. [53]Flesch, E. W. 2023, The Open Journal of Astrophysics, 6, 49
  50. [54]Gaia Collaboration, Vallenari, A., Brown, A. G. A., Prusti, T., et al. 2023, A&A, 674, A1
  51. [55]Garn, T., Green, D. A., Riley, J. M., & Alexander, P. 2008, MNRAS, 383, 75
  52. [56]Gendre, M. A., Best, P. N., & Wall, J. V. 2010, MNRAS, 404, 1719
  53. [57]Gloudemans, A. J., Sweijen, F., Morabito, L. K., et al. 2025, ApJ, 980, L8
  54. [58]Govoni, F., Orrù, E., Bonafede, A., et al. 2019, Science, 364, 981
  55. [60]Hale, C. L., Heywood, I., Jarvis, M. J., et al. 2025, MNRAS, 536, 2187
  56. [61]Helfand, D. J., White, R. L., & Becker, R. H. 2015, ApJ, 801, 26
  57. [62]Herrera Ruiz, N., O’Sullivan, S. P., Vacca, V., et al. 2021, A&A, 648, A12
  58. [63]Hurley-Walker, N., Galvin, T. J., Duchesne, S. W., et al. 2022, PASA, 39, e035
  59. [64]Intema, H. T., Jagannathan, P., Mooley, K. P., & Frail, D. A. 2017, A&A, 598, A78
  60. [65]Jackson, N., Badole, S., Morgan, J., et al. 2022, A&A, 658, A2
  61. [66]Jackson, N., Tagore, A., Deller, A., et al. 2016, A&A, 595, A86
  62. [67]Kassim, N. E., Lazio, T. J. W., Erickson, W. C., et al. 2007, ApJS, 172, 686
  63. [68]Kondapally, R., Best, P. N., Cochrane, R. K., et al. 2022, MNRAS, 513, 3742
  64. [69]Kondapally, R., Best, P. N., Hardcastle, M. J., et al. 2021, A&A, 648, A3
  65. [70]Kurtzer, G. M., Sochat, V., & Bauer, M. W. 2017, PLOS ONE, 12, e0177459
  66. [71]Lacy, M., Baum, S. A., Chandler, C. J., et al. 2020, PASP, 132, 035001
  67. [72]Lane, W. M., Cotton, W. D., van Velzen, S., et al. 2014, MNRAS, 440, 327
  68. [73]Lawrence, A., Warren, S. J., Almaini, O., et al. 2007, MNRAS, 379, 1599
  69. [74]Liu, J., van Weeren, R., Röttgering, H., & Kuijken, K. 2025, A&A, 700, A250
  70. [75]Loi, S. T., Murphy, T., Bell, M. E., et al. 2015, MNRAS, 453, 2731
  71. [76]Magliocchetti, M., La Marca, A., Bisigello, L., et al. 2026, A&A, 711, A40
  72. [77]Mahatma, V. H., Basu, A., Hardcastle, M. J., Morabito, L. K., & van Weeren, R. J. 2023, MNRAS, 520, 4427
  73. [78]Manners, J. C., Johnson, O., Almaini, O., et al. 2003, MNRAS, 343, 293
  74. [79]Martin, D. C., Fanson, J., Schiminovich, D., et al. 2005, ApJ, 619, L1
  75. [80]Mauch, T., Cotton, W. D., Condon, J. J., et al. 2020, ApJ, 888, 61
  76. [81]Mauduit, J. C., Lacy, M., Farrah, D., et al. 2012, PASP, 124, 714
  77. [82]McMahon, R. G., Walton, N. A., Irwin, M. J., et al. 2001, New A Rev., 45, 97
  78. [83]Mohan, N. & Rafferty, D. 2015, PyBDSF: Python Blob Detection and Source Finder, Astrophysics Source Code Library, record ascl:1502.007
  79. [84]Morabito, L. K., Jackson, N., de Jong, J., et al. 2025a, Ap&SS, 370, 19
  80. [85]Morabito, L. K., Jackson, N. J., Mooney, S., et al. 2022a, A&A, 658, A1
  81. [86]Morabito, L. K., Kondapally, R., Best, P. N., et al. 2025b, MNRAS, 536, L32
  82. [87]Morabito, L. K., Sweijen, F., Radcliffe, J. F., et al. 2022b, MNRAS, 515, 5758
  83. [88]Mostert, R. I. J., Oei, M. S. S. L., Barkus, B., et al. 2024, A&A, 691, A185
  84. [89]Murphy, P. C., Callanan, P., McCauley, J., et al. 2021, A&A, 655, A16
  85. [90]Noordam, J. E. & Smirnov, O. M. 2010, A&A, 524, A61
  86. [91]Norris, R. P., Marvil, J., Collier, J. D., et al. 2021, PASA, 38, e046
  87. [92]Novak, M., Smolčić, V., Delhaize, J., et al. 2017, A&A, 602, A5
  88. [93]Novak, M., Smolčić, V., Schinnerer, E., et al. 2018, A&A, 614, A47
  89. [94]Ocran, E. F., Taylor, A. R., Vaccari, M., Ishwara-Chandra, C. H., & Prandoni, I. 2020, MNRAS, 491, 1127
  90. [95]Oei, M. S. S. L., Hardcastle, M. J., Timmerman, R., et al. 2024, Nature, 633, 537
  91. [96]Offringa, A. R. 2016, A&A, 595, A99
  92. [97]Offringa, A. R., McKinley, B., Hurley-Walker, N., et al. 2014, MNRAS, 444, 606
  93. [98]Offringa, A. R. & Smirnov, O. 2017, MNRAS, 471, 301
  94. [99]Offringa, A. R. & van Weeren, R. J. 2026, A&A, 706, A68
  95. [100]Oliver, S., Rowan-Robinson, M., Alexander, D. M., et al. 2000, MNRAS, 316, 749
  96. [101]Oliver, S. J., Bock, J., Altieri, B., et al. 2012, MNRAS, 424, 1614
  97. [102]Orrú, E., Norden, M. J., Iacobelli, M., ter Veen, S., & Ahmadi, A. 2024, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 13098, Observatory Operations: Strategies, Processes, and Systems X, ed. C. R. Benn, A. Chrysostomou, & L. J. Storrie-Lombardi, 130980R
  98. [103]Piras, S., Horellou, C., Conway, J. E., et al. 2024, A&A, 687, A267
  99. [104]Piras, S., Horellou, C., Conway, J. E., et al. 2025, A&A, 693, A100
  100. [105]Prandoni, I. & Seymour, N. 2015, in Advancing Astrophysics with the Square Kilometre Array (AASKA14), 67
  101. [106]Purcell, C. R., Van Eck, C. L., West, J., Sun, X. H., & Gaensler, B. M. 2020, RM-Tools: Rotation measure (RM) synthesis and Stokes QU-fitting, Astrophysics Source Code Library, record ascl:2005.003
  102. [107]Ramírez-Olivencia, N., Varenius, E., Pérez-Torres, M., et al. 2022, A&A, 658, A4
  103. [108]Ramírez-Olivencia, N., Varenius, E., Pérez-Torres, M., et al. 2018, A&A, 610, L18
  104. [109]Rengeling, R. B., Tang, Y., de Bruyn, A. G., et al. 1997, A&AS, 124, 259
  105. [110]Renting, G. A. & Holties, H. A. 2012, in Astronomical Society of the Pacific Conference Series, Vol. 461, Astronomical Data Analysis Software and Systems XXI, ed. P. Ballester, D. Egret, & N. P. F. Lorente, 689
  106. [111]Sabater, J., Best, P. N., Tasse, C., et al. 2021, A&A, 648, A2
  107. [112]Schinnerer, E., Smolčić, V., Carilli, C. L., et al. 2007, ApJS, 172, 46
  108. [113]Schwab, F. R. 1984, AJ, 89, 1076
  109. [114]Shimwell, T. W., Hale, C. L., Best, P. N., et al. 2025, A&A, 695, A80
  110. [115]Shimwell, T. W., Hardcastle, M. J., Tasse, C., et al. 2022, A&A, 659, A1
  111. [116]Shimwell, T. W., Hardcastle, M. J., Tasse, C., et al. 2026, A&A, 707, A198
  112. [117]Shimwell, T. W., Röttgering, H. J. A., Best, P. N., et al. 2017, A&A, 598, A104
  113. [118]Sinha, A., Basu, A., Datta, A., & Chakraborty, A. 2022, MNRAS, 514, 4343
  114. [119]Sirothia, S. K., Dennefeld, M., Saikia, D. J., et al. 2009, in Astronomical Society of the Pacific Conference Series, Vol. 407, The Low-Frequency Radio Universe, ed. D. J. Saikia, D. A. Green, Y. Gupta, & T. Venturi, 27
  115. [120]Skrutskie, M. F., Cutri, R. M., Stiening, R., et al. 2006, AJ, 131, 1163
  116. [122]Smolčić, V., Novak, M., Delvecchio, I., et al. 2017, A&A, 602, A6
  117. [123]Starck, J.-L. & Murtagh, F. 1994, A&A, 288, 342
  118. [124]Sweijen, F., Kurek, A., van Weeren, R., & van der Wild, M. 2025a, FLoCS: Frits’ LOFAR Containers, Zenodo, version v6.0.0
  119. [125]Sweijen, F., Pierce, J. C. S., Hardcastle, M. J., et al. 2025b, MNRAS, 540, 416
  120. [126]Sweijen, F., van Weeren, R. J., Röttgering, H. J. A., et al. 2022, Nature Astronomy, 6, 350
  121. [129]Tasse, C., Hugo, B., Mirmont, M., et al. 2018, A&A, 611, A87
  122. [130]Tasse, C., Shimwell, T., Hardcastle, M. J., et al. 2021, A&A, 648, A1
  123. [132]Timmerman, R., Escott, E. L., Vernstrom, T., Morabito, L. K., & Reynolds, C. 2026, arXiv e-prints, arXiv:2606.25017arxiv.org/abs/2606.25017
  124. [133]Timmerman, R., van Weeren, R. J., Callingham, J. R., et al. 2022, A&A, 658, A5
  125. [134]van den Haak, L. B., Wijs, A., Huisman, M., & van den Brand, M. 2024, in Formal Methods for Industrial Critical Systems, ed. A. E. Haxthausen & W. Serwe (Cham: Springer Nature Switzerland), 152–169
  126. [135]van der Tol, S., Veenboer, B., & Offringa, A. R. 2018, A&A, 616, A27
  127. [136]van der Tol, S., Yatawatta, S., Veenboer, B., & Rafferty, D. 2026, arXiv e-prints, arXiv:2602.06002arxiv.org/abs/2602.06002
  128. [137]van der Velden, E. 2020, The Journal of Open Source Software, 5, 2004
  129. [138]van der Vlugt, D., Algera, H. S. B., Hodge, J. A., et al. 2021, ApJ, 907, 5
  130. [139]van der Wel, A., Chang, Y.-Y., Bell, E. F., et al. 2014, ApJ, 792, L6
  131. [140]van Diepen, G., Dijkema, T. J., & Offringa, A. 2018, DPPCP: Default Pre-Processing Pipeline, Astrophysics Source Code Library, record ascl:1804.003
  132. [141]Van Eck, C. L., Purcell, C. R., Baidoo, L., et al. 2026, arXiv e-prints, arXiv:2601.20092arxiv.org/abs/2601.20092
  133. [142]van Haarlem, M. P., Wise, M. W., Gunst, A. W., et al. 2013, A&A, 556, A2
  134. [143]van Weeren, R. J., de Gasperin, F., Akamatsu, H., et al. 2019, Space Sci. Rev., 215, 16
  135. [144]van Weeren, R. J., de Jong, J. M. G. H. J., Le Saux, X. K., et al. 2026a, MNRAS, 549, stag992
  136. [145]van Weeren, R. J., Osinga, E., Brunetti, G., et al. 2026b, MNRAS, 546, stag054
  137. [146]van Weeren, R. J., Shimwell, T. W., Botteon, A., et al. 2021, A&A, 651, A115
  138. [147]van Weeren, R. J., Timmerman, R., Vaidya, V., et al. 2024, A&A, 692, A12
  139. [148]van Weeren, R. J., Williams, W. L., Hardcastle, M. J., et al. 2016, ApJS, 223, 2
  140. [149]Varenius, E., Conway, J. E., Martí-Vidal, I., et al. 2016, A&A, 593, A86
  141. [150]Veenboer, B. & Romein, J. W. 2020, Astronomy and Computing, 32, 100386
  142. [151]Vivian, J., Rao, A., Nothaft, F., et al. 2017, Nature Biotechnology, 35, 314
  143. [153]Wang, J., Norden, M., & Donker, P. 2025, Journal of Astronomical Telescopes, Instruments, and Systems, 11, 017002
  144. [154]Wang, W., Yuan, Z., Yu, H., & Mao, J. 2024, A&A, 683, A174
  145. [155]Ward, E., de la Vega, A., Mobasher, B., et al. 2024, ApJ, 962, 176
  146. [156]Wayth, R. B., Lenc, E., Bell, M. E., et al. 2015, PASA, 32, e025
  147. [157]Wijnholds, S. J., Willis, A. G., & Salvini, S. 2018, MNRAS, 476, 2029
  148. [158]Wilkinson, M. D., Dumontier, M., Aalbersberg, I. J., et al. 2016, Scientific data, 3, 160018
  149. [161]Yoo, A. B., Jette, M. A., & Grondona, M. 2003, in Job Scheduling Strategies for Parallel Processing, ed. D. Feitelson, L. Rudolph, & U. Schwiegelshohn (Berlin, Heidelberg: Springer Berlin Heidelberg), 44–60
  150. [162]Zwart, J. T. L., Santos, M., & Jarvis, M. J. 2015, MNRAS, 453, 1740

Paper details

vlnc:W_zuatvqkw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents