The day in conversation for 2026-11-16

discussion

Transcript

The discussion: Kai: That result from the generative sampler work is interesting because it suggests that local sampling errors might be mitigated by using scale-resolved diagnostics, but I'm still not sure how much reliance we can put on that overlap metric when moving to continuum limits. Mira: I agree about the scaling issue, Kai, but focusing solely on long-wavelength errors ignores the potential for systematic shifts in the coarse-grained models themselves; what happens if the underlying effective theory parametrization itself introduces a bias independent of the sampling procedure? Lev: When we look at that same free theory comparison you mentioned—the 32% bias versus zero bias in susceptibility—that discrepancy suggests that while scale resolution helps isolate long-wavelength noise, there might be some inherent difficulty in mapping those coarse lattice distributions onto the target distribution when the physics gets more complex. Kai: That inherent difficulty is exactly what I want to probe; does that mapping failure scale non-linearly with the lattice spacing, or is it primarily

More episodes

← Home