Multiscale / Renormalization

Definition

This motif appears when a system can be described at multiple resolutions and the challenge is preserving the variables that matter across those resolutions.

This motif becomes actionable when the coarse-graining operation is explicit. ISOM asks which variables survive scale changes and which details can be safely discarded without damaging downstream decisions.

Mathematical Structure

The core structure is a hierarchy of coarse-graining operations, effective descriptions, and scale-dependent parameters.

Physics Side

Renormalization gives physics a disciplined way to move from microscopic detail to macroscopic law without pretending all details remain equally relevant.

AI Side

In AI, this suggests representation stacks, temporal abstraction, hierarchical planning, and scale-aware routing that preserve only the effective degrees of freedom.

Potential AI transfers include hierarchical planning, multi-resolution perception, token compression, and representation routing. The key metric is whether information remains useful after repeated abstraction, not whether the hierarchy looks elegant.

Failure Modes

A multiscale story breaks when the chosen abstraction discards variables that later become decision-critical. The wrong coarse-graining can produce confidence without control.

A poor abstraction can erase rare but decisive variables. ISOM treats multiscale transfer as risky unless the experiment measures what is lost at each scale and whether that loss changes the final decision.

Open Questions

Which latent variables remain predictive under repeated coarse-graining, and how should we learn them without hand-designed scale boundaries?