Batchelor's formula and infrared renormalization for sedimentation
arXiv:2607.09995
2026
Architecture
1 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The paper develops a constructive method for making infinite-volume observables well-defined when interactions have non-integrable long-range tails: isolate an explicit singular contribution, subtract a counterterm representing divergent mean backflow, and control the remaining regular interaction using decay estimates. The transferable asset is not the Stokes-flow application, but the mode-separation principle and cluster-by-cluster cancellation of large-scale contributions. A promising neural analogue is a long-context or point-cloud aggregation layer that separates a global low-frequency interaction from a centered residual, preventing context-size-dependent activation drift while retaining long-range information. The extracted estimates provide a principled infrared scale and falsifiable predictions for how residual magnitudes should change with context length.
Ideas from this paper
Unverified
2026
Add a coordinate-aware long-range aggregation branch whose singular low-frequency component is explicitly centered before it is mixed into token representations. The centering acts as a neural counterterm: constant or slowly varying value fields cannot accumulate an activation contribution that grows with context size, while local and higher-frequency interactions remain available through an ordinary attention residual branch.
Useful6/10
Difficulty5/10
Novelty7/10