The day in conversation for 2026-11-05

discussion

Transcript

The discussion: Vera: The speedup factor of nine using merge-walk algorithms on the GPU for galaxy clustering sounds promising, but I’m curious about the underlying assumptions regarding graph database structure; does that efficiency hold up when dealing with highly non-linear spatial correlations typical in deep surveys? Jocelyn: That computational gain is significant, but I wonder how that high-order clustering statistic actually informs our photometric error budgets. If the clustering analysis is significantly faster, are we seeing a corresponding improvement in the fidelity of the derived correlation lengths compared to the legacy methods we used last month? Subrahmanyan: I’m more interested in what you mentioned about posterior sampling for Point Spread Functions. Achieving orders of magnitude higher likelihoods suggests a fundamental shift in how we characterize image profiles; what specific physical parameter limitations are you encountering when scaling that Bayesian framework to truly complex, high-resolution astronomical images? Vera: The likelihood improvement is compelling, but I worry that the new pixelation approach might introduce systematic artifacts if the underlying assumptions about the PSF morphology don't hold true across different observing conditions. Have you tested how robust those likelihoods are when comparing them

More episodes

← Home