1 of 37

Separation of Point Source Contaminants at the Visibility Level

�Shreyam Parth Krishna

��26th August 2025, Swiss SKA Days 2025, Winterthur

1

2 of 37

Radio Interferometric Imaging

2

Electric Field

with Brightness

Baseline

Phase Difference :

Antenna P

Antenna Q

Voltage P

Voltage Q

X

Visibilities

3 of 37

Radio Interferometric Imaging

3

Electric Field

with Brightness

Baseline

Phase Difference :

Antenna P

Antenna Q

Voltage P

Voltage Q

X

4 of 37

Radio Interferometric Imaging

4

Electric Field

with Brightness

Baseline

Phase Difference :

Antenna P

Antenna Q

Voltage P

Voltage Q

X

Visibilities

5 of 37

CLEAN family of algorithms

5

1

1: [Corda 2022]

(Typical) Dirty image uses backprojection:

​

6 of 37

Bluebild Algorithm

6

​

  • Bluebild uses the Psuedoinverse as a solution to the least �squares problem:��
  • G often ill-conditioned and difficult to invert, instead we want an optimal solution of the form:

��

  • Solves for I(r) in by framing a generalised eigenvalue problem and ��decomposing visibilities into different eigenvectors, via fPCA.

Eigenvector + sampling operator gives eigenvisibilities. These can be recombined and are sorted by energy.

​

​

​

  • Flexible continuous spherical imager for interferometric applications; 3D nuFFT Type-3

Normalised Eigenvisibilities

7 of 37

Bluebild Imaging Plus Plus (BIPP)

  • C++ ported version with python wrapper for wider community release
  • HPC implementation with support for CUDA (NVIDIA) and HIP (AMD)
  • MPI Parallelism and Domain partitioning inbuilt
  • No deconvolution - only outputs dirty images or .ms files
  • Github: https://github.com/epfl-radio-astro/bipp
  • Arxiv: https://arxiv.org/pdf/2310.09200

7

8 of 37

Bluebild Imaging Plus Plus (BIPP)

  • C++ ported version with python wrapper for wider community release
  • HPC implementation with support for CUDA (NVIDIA) and HIP (AMD)
  • MPI Parallelism and Domain partitioning inbuilt
  • No deconvolution - only outputs dirty images or .ms files
  • Github: https://github.com/epfl-radio-astro/bipp
  • Arxiv: https://arxiv.org/pdf/2310.09200

8

9 of 37

Bluebild Toy Example

9

10 of 37

Bluebild Toy Example

10

11 of 37

Bluebild Toy Example

11

12 of 37

Bluebild Toy Example

12

13 of 37

Bluebild Toy Example

13

14 of 37

Bluebild Toy Example

14

15 of 37

Bright Point Source Contaminants are Ubiquitous in Radio Astronomy

Typically solved by peeling or direction dependeny calibration followed by subtracting source model in visibility space Eg: Obit4, Ionpeel5 etc.

15

1

1: Williams et al 2019, 2: Credit: Watson, Harrison et al. 2024 3: Obit Development Memo Series No. 54, Cotton 4: Cotton 2013 5: Offringa et. al. 2016

3

2

16 of 37

Bright Point Source Contaminants are Ubiquitous in Radio Astronomy

But perfect peeling is not always possible due to:

  • Varying observing conditions (Radio Frequency Interference (RFI), ionospheric activity)
  • Direction Dependent Effects, Incomplete sky models and Calibration Errors

16

1

1: Williams et al 2019, 2: Credit: Watson, Harrison, 3: Obit Development Memo Series No. 54, Cotton

3

2

17 of 37

Separating Contaminating Point Sources at the Visibility Level

  • Model: T-RECS1 x 1000�
  • Visibilities Generated using SKA Low Configuration on OSKAR2 using Karabo Pipeline3 @200 MHz�
  • Added Observational Effects4�
  • Separated measurement set using BIPP�
  • Imaged using WSClean

17

1: Bonaldi et. al. 2018, 2023, 2: Mort et. al. 2010 , 3: Sharma et. al. 2025 4: https://github.com/micbia/mirto, 5:skao.int

18 of 37

18

1: Price et. al. (2015)

1

(Simplified) Radio Interferometric Measurement Equation

Where

Erqt = Direction Dependent gain

Gqt = Direction Independent gain

N(t) - additive system noise

19 of 37

Additive (Noise) and Multiplicative (Calibration) Errors

19

Spectral Frequency

Time Steps

Gain Deviation

Power Spectral Density (dB)

20 of 37

3 types of output:

  • “Complete” Images: Unseparated Visibilities run through WSClean�
  • “Interval” Images: Visibilities separated via BIPP run through WSClean�
  • “Summed” Images: Visibilities separated via BIPP run through WSClean then summed

Fit 1-D Gaussians to pixel value distribution to determine noise in images after cleaning.

20

Separating Contaminating Point Sources at the Visibility Level

21 of 37

Eigenvalue Histogram

21

22 of 37

200 MHz Observation; No Noise; No Calibration Errors

22

23 of 37

200 MHz Observation; 8h Additive Noise;

Realistic Calibration Errors

23

24 of 37

Image Noise Comparison for all Cases

24

1,2

1:Van Weeren et. al. 2016, 2: de Gasperin et. al. 2019,

25 of 37

Real Data

Real Data

25

  • MWA Phase I Observation of Centaurus A
  • 11.37 deg FoV
  • 175.35 MHz Observation
  • MWA Phase I - sensitivity to extended structure

26 of 37

Real Data

26

  • MWA Phase I Observation of Centaurus A
  • Closest AGN, A-team source
  • 11.37 deg FoV, 175.35 MHz Observation
  • MWA Phase I - sensitivity to extended structure

27 of 37

Conclusions

  • BIPP’s eigendecomposition can help to probe a lower noise floor in the presence of realistic calibration errors and noise. �
  • BIPP’s eigendecomposition can help to avoid peeling errors while still effectively removing bright contaminants.

​

Contact me at shreyam.krishna@epfl.ch for any more questions!

27

28 of 37

Conclusions

  • BIPP’s eigendecomposition can help to probe a lower noise floor in the presence of realistic calibration errors and noise. �
  • BIPP’s eigendecomposition can help to avoid peeling errors while still effectively removing bright contaminants.

​

Contact me at shreyam.krishna@epfl.ch for any more questions!

28

29 of 37

Backup Slides start here

29

30 of 37

Future Scope

  • MeerKAT Exploration of Relics, Giant Halos, and Extragalactic Radio Sources (MERGHERs) follow up using BIPP + WSClean methodology
  • Broader extension of BIPP + WSClean methodology to real data

30

31 of 37

Bluebild Toy Example

31

32 of 37

Parameter Estimator ~ 10% of Imaging Time with I/O Overhead

33 of 37

33

34 of 37

200 MHz Observation

8h Additive Noise

No Calibration Errors

34

35 of 37

200 MHz Observation; 8h Additive Noise; No Calibration Errors

35

36 of 37

200 MHz Observation; 8h Additive Noise; Low Calibration Errors

36

37 of 37

200 MHz Observation; 8h Additive Noise; very very Low Calibration Errors

37