Imaging Algorithms vs. Perfect Data

2012 March 15

Fun fact: when I’m not rooting my phone or Linuxing my computer, I do some astronomy. I’ve been doing some data simulations and used the opportunity to look into the effect that different approaches to interferometric imaging have. The image below summarizes what I’ve found:

Using real ATA observations as a template, I generated some perfect fake data of a single point source — no noise, calibration errors, etc. I’m interested in wide-field imaging so I put the source at 0.63 deg from phase center. I then imaged the data using the casarest lwimager fork of the CASA imager with various options.

The top-left panel shows the imaging residuals using traditional grid-and-FFT and 200 iterations of Högbom (1974) clean. The S/N is just 1001 — for ideal data with no noise!

The top-right panel turns on the w-projection algorithm with 128 planes. The S/N jumps to 2987. I’m too lazy to measure it, but the speed penalty is nontrivial. Annoyingly, a bunch of time is spent in startup calculating convolution kernels that are constant from one imaging run to another in most cases. In the imaging that I do, a lot of time could be saved by caching those kernels on disk.

The bottom-left panel turns on the “wide- field” Högbom clean which subtracts the modeled CLEAN components from the visibilities. This also adds a significant time penalty. The S/N increases to 8713.

Finally, and most interestingly to me, in the bottom-right panel I’ve moved the source slightly to land it precisely in the center of the nearest image pixel. The S/N jumps to 34542. This effect is investigated in Cotton & Uson (2008).

Not shown: using Cotton-Schwab clean with w-projection gives results nearly identical to those of the wide-field Högbom clean. Wide-field Högbom should theoretically give more accurate results since it uses the full dirty beam in the image domain. In a quick test, CS clean is about 30% faster than wide-field Högbom.