Non-Physical Positive Spectral Index from DDFacet Subband Imaging

[BANELE-MTHEMBU]

I'm encountering non-physical positive spectral indices when performing in-band spectral analysis of extended radio galaxies imaged with DDFacet. This contradicts the expected synchrotron power-law behavior (negative spectral index) and persists across multiple sources and subbands. The issue appears to be isolated to DDFacet and does not occur with WSCLEAN using the same data and calibration.

Test Setup

• Dataset: MeerKAT field centered on J0438-5419

• Number of Sources: 5 extended radio galaxies

• Subband Imaging:

  • 3 subbands extracted with central frequencies:

•        Subband 1: 1.00 GHz
  

•       Subband 2: 1.38 GHz
  

•       Subband 3: 1.56 GHz
  

• Processing Notes:

All images were convolved to a common beam size

Flux densities were measured at fixed RA/Dec positions for each source

Identical measurement method was used for both DDFacet and WSCLEAN

Observed Behavior

• DDFacet shows either flat or increasing flux density with frequency across subbands.

• WSCLEAN shows the expected decreasing flux density consistent with synchrotron emission.

• This behavior is consistent across all 5 sources tested.

Impact
This behavior results in positive spectral indices, which are not physically expected for the types of sources observed (i.e., extended synchrotron-emitting radio galaxies). I got the In-band spectral index map measurements that are falling on the positive side with the DDFacet subband images.

Attached Plot

The attached plot compares DDFacet and WSCLEAN flux density trends across subbands for five sources:

• Solid lines with filled markers: DDFacet

• Dashed lines with open markers: WSCLEAN

• Each pair of lines corresponds to one source. Note how WSCLEAN follows the expected power-law trend, while DDFacet does not.

Request
I request to please confirm if this is a known issue with subband imaging in DDFacet, provide insight into what might be causing this discrepancy, suggest any recommended configuration tweaks or processing steps, and let me know if sub-band FITS files are needed I’m happy to share.

Thanks again for your work on DDFacet. I really appreciate the tool and your support.

[bennahugo]

Need a bit more information here. Are you applying kMS solutions in the DDF run or not - I see this was applied in the previous ticket. If so this may just be a case of flux absorption due to inadequite modelling of the extended emission. Have you crossmatched your fluxscales before and after calibration? Obviously the wsclean version will not have kms solutions applied so it is a bit hard to tell if it is an apples to apples comparison of calibrated products.

[BANELE-MTHEMBU]

Thanks for the response. After splitting the full bandwidth into sub-bands, I ran killMS independently on each sub-band to generate fresh calibration solutions. I did not reuse the killMS solutions from the final 3GC image generated with oxkat.

As for your question on crossmatching flux scales before and after calibration, I haven’t explicitly done that yet. However, during the 1GC to 2GC step, I did check the flux consistency across the 8 multifrequency images, and the fluxes followed a clean power-law trend, so the spectral behavior appeared reasonable.

Do you mean I should compare the total flux (or flux scale) before and after applying the sub-band killMS + DDFacet imaging to assess if there is flux absorption? I can try that next.

[bennahugo]

Do you mean I should compare the total flux (or flux scale) before and after applying the sub-band killMS + DDFacet imaging to assess if there is flux absorption? I can try that next.

Yes - for each of the subbands. Because the lower band will typically contain more extended diffuse flux due to the steep spectra there is more chance for absorption. Depending on the number of DD solutions derived with KillMS and for fine cadences there is a good chance of this.

Alternatively do not apply solutions and reimage to give a more true comparison to wsclean?

[cyriltasse]

Hi @BANELE-MTHEMBU thanks for your feedback! What are you using to deconvolve (SSD/SSD2/HMP/Hogbom)? How many taylor terms? Are you using a beam model as input to kMS/DDF?

[BANELE-MTHEMBU]

@cyriltasse Thanks for your question. For deconvolution, I’m using the Högbom CLEAN algorithm. My current setup includes: Taylor term fitting: Polynomial order = 4 (via PolyFitOrder 4) Beam model: None (no primary beam correction applied).

[BANELE-MTHEMBU]

Hi @bennahugo,

Thanks again for your input.

Since I applied the KMS solutions, you mentioned that the positive spectral index values might be due to flux absorption caused by poor modelling of extended emission. I wanted to ask, does using a mask help retain the flux from extended emission in this case?

To test things further, I tried running DDFacet without applying KMS, just for the first subband image. I also made sure that all images were convolved to the same beam size before comparing the fluxes. I selected a few extended sources and compared their flux across three versions:

1. Subband image imaged with WSCLEAN
2. Subband image imaged with DDFacet (no KMS)
3. Final 3GC image

In all cases, the sources imaged with WSCLEAN had the highest flux, followed by the 3GC image, and lastly the DDFacet (no KMS) image, which showed the lowest flux for the same sources.

[bennahugo]

My thought is it is probably a difference in weighting for the images where solutions are not applied. Convolving the image post imaging is not enough to recover flux filtered out by the weighting applied to the interferometer. You would need to apply equal weighting to the imaging weights to do a 1:1 comparison. One possible discrepency can be your briggs weights - the implementations are different between the two imagers, so you may need to play around to get the synthesized beam to the same size for a comparison.

[BANELE-MTHEMBU]

Hi @bennahugo,

I conducted a new test by imaging each of the three subbands using DDFacet without applying any KillMS solutions. After imaging, I convolved the outputs to a common beam size and compared the flux values. For both extended and point sources, I observed that the flux at lower frequencies was consistently lower, while the flux increased toward the higher-frequency bands. This trend does not follow the expected power-law behavior.

In contrast, when I imaged the same three subbands using WSClean (again convolving the outputs to a common beam size), I observed the opposite trend: the flux at lower frequencies was higher, and it decreased with increasing frequency, as expected for typical synchrotron sources. I ensured that I used consistent imaging parameters between DDFacet and WSClean for a fair comparison.

Below are the exact imaging commands I used:

DDFacet:

singularity exec /idia/software/containers/ddfacet_v0.7.2.sif DDF.py \ --Data-MS cfp2020_1_subband1.ms \ --Data-ColName DATA \ --Data-ChunkHours 1 \ --Data-Sort 1 \ --Predict-InitDicoModel /scratch3/users/banele/oxkat_r/IMAGES/img_cfp2020_1_full_1024ch_J0438-5419.ms_DDFpcal.DicoModel \ --Output-Name /scratch3/users/banele/killms/no_killms_sol/img_cfp2020_1_full_1024ch_J0438-5419.ms_DDFkMS_subband1 \ --Output-Mode Clean \ --Output-Also oenNS \ --Output-Images DdPMmRrIikz \ --Output-Cubes MmRi \ --Image-NPix 10125 \ --Image-Cell 1.1 \ --Facets-DiamMax 0.25 \ --Facets-DiamMin 0.05 \ --Facets-NFacets 4 \ --Facets-PSFOversize 1.5 \ --Facets-Padding 3.0 \ --Weight-Mode Briggs \ --Weight-Robust 0.0 \ --Comp-GridDecorr 0.01 \ --Comp-DegridDecorr 0.01 \ --Comp-Sparsification 0 \ --Parallel-NCPU 16 \ --Cache-Reset 0 \ --Cache-Dir . \ --Cache-HMP 1 \ --Beam-Model None \ --Freq-NBand 4 \ --Freq-NDegridBand 4 \ --Deconv-Gain 0.15 \ --Deconv-FluxThreshold 3e-06 \ --Deconv-CycleFactor 0 \ --Deconv-RMSFactor 3.0 \ --Deconv-Mode Hogbom \ --Deconv-PeakFactor 0.1 \ --Deconv-MaxMajorIter 0 \ --Deconv-MaxMinorIter 1000 \ --Hogbom-PolyFitOrder 4 \ --Mask-Auto 0 \ --Mask-External /scratch3/users/banele/oxkat_r/IMAGES/img_cfp2020_1_full_1024ch_J0438-5419.ms_pcalmask-MFS-image.mask1.zoom10125.fits \ --Misc-ConserveMemory 1 \ --Log-Memory 1 \ --Log-Boring 1 \ --Selection-ChanStart 0 --Selection-ChanEnd 119 --Selection-ChanStep 1

WSCLEAN:

singularity exec /idia/software/containers/wsclean-v3.3-no-cuda.sif wsclean \ -name /scratch3/users/banele/killms/wsclean_images/cfp2020_1_subband1_wsclean \ -log-time \ -abs-mem 225 \ -parallel-reordering 8 \ -save-source-list \ -data-column DATA \ -field 0 \ -size 10240 10240 \ -scale 1.1asec \ -gridder wgridder \ -weight briggs 0.0 \ -parallel-deconvolution 2560 \ -no-update-model-required \ -niter 80000 \ -gain 0.15 \ -mgain 0.9 \ -channels-out 4 \ -fit-spectral-pol 4 \ -join-channels \ -circular-beam \ -fits-mask /scratch3/users/banele/oxkat_r/IMAGES/img_cfp2020_1_full_1024ch_J0438-5419.ms_datablind-MFS-image.mask0.fits \ -threshold 3e-06 \ cfp2020_1_subband1.ms

Let me know if you think the discrepancy is still due to differences in weighting schemes, or if I should test additional configurations.