Is it possible to use brighter-fatter kernel to object detection in difference imaging analysis (DIA)?

The brighter-fatter effect modifies the statistical properties of imaging data, which would possibly affect the object detection and flux measurement in DIA. Is it possible to use the brighter-fatter kernel to object detection in DIA?

Because you mentioned “statistical properties”, I’m guessing that you’re assuming we’ve already applied some kind of brighter-fatter correction to the image itself (i.e. the signal), and you’re talking about the fact that brighter-fatter also affects the noise, and correcting the signal does not correct the noise?

Correcting the signal is certainly important for DIA, and something we try to do already. I think also correcting for changes to the noise during detection is probably straightforward from a theoretical standpoint (it’s easy to account for correlated noise if you know what it is) - but I’d also be surprised if it made big difference, as there are many larger effects (including correlated noise just from PSF matching) that are hard to fully correct for in practice, and I think we would have to be doing extremely well at image subtraction generally for brighter-fatter noise effects to be a major concern. That’s just intuition, though; I don’t have any numbers to back that statement up.


Thanks for the explanation. Actually my question of the brighter-fatter correction is related to both signal and noise. In the bright region, the pixel correlation is large and we usually have subtraction artifacts there. Would brighter-fatter effect make a large contribution to the pixel correlation in bright region? One related question is the photometry, if the bright-fatter does make large contribution to pixel correlation in bright region, then would we underestimate the photometry error a lot if we ignore that correlation?

I think we can identify the limit in which the brighter-fatter effect contribution to noise correlations in DIA is large compared to other contributions - some combination of very deep templates (so low noise) with PSFs very similar to the science image they’re subtracted from (so little correlation of that little noise by matching, and few other problems with subtraction artifacts). Whether that limit is a common one or a rare one depends on the distribution of observing conditions, how the PSF changes in practice, and how far along we are into the survey, and I think one would have to use simulations to get a real feel for how important this actually is.

1 Like