---
title: "A trained physical network's function is readable from the soft modes of its Hessian, tying conserved protein regions to slow collective modes"
type: "claim"
status: "seedling"
writer_model: "claude-opus-4-8"
source_url: "https://arxiv.org/abs/2406.09689"
source_title: "Physical networks become what they learn"
source_author: "Menachem Stern, Marcelo Guzman, Felipe Martins, Andrea J. Liu, Vijay Balasubramanian"
source_date: 2024
source_venue: "'Physical networks become what they learn' (arXiv:2406.09689)"
source_quote: "conserved regions in proteins were associated with slow collective modes"
source_tier: 1
audit_status: "capture-verified but NOT verbatim-grounded for the mechanism — the hop capture (2026-07-11) read arXiv:2406.09689 as a lead and preserved only the phrase 'conserved regions in proteins were associated with slow collective modes' verbatim. The load-bearing mechanism (function recoverable from the physical Hessian's soft modes; the physical-Hessian ↔ cost-Hessian correspondence) is recorded as a hop paraphrase, not an exact quote, so it is held [unverified-mechanism] pending a primary re-read. Routed to [[question-verify-physical-networks-hessian-cost-formalism]]. Freely fetchable on arXiv. | 2026-07-12 (audit, big-opus-8): primary re-read of arXiv:2406.09689v2 confirms the verbatim quote in the Discussion ('Conserved regions in proteins were associated with slow collective modes') and grounds the mechanism — the paper states the physical Hessian 'gives insight into an adaptable network's function without knowing the task' and that 'the satisfied constraint can be discovered by examining the physical responses of the network around equilibrium without any prior knowledge of the task'; the physical↔cost Hessian link is derived in Eqs. 8–9. [unverified-mechanism] resolved to verified-against-primary. Author list recovered from the primary and filled into source_author."
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-physical-learning-allostery.md, 2026-07-12"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-11-hop-physical-learning-allostery.md"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["physical-learning","structure-function","soft-modes","allostery","protein-biophysics","energy-based-models","coupled-learning"]
audits: ["2026-07-12 claude-opus-4-8"]
drafted_in: ["2026-07-13-cut-one-percent-of-the-bonds","cut-one-percent-of-the-bonds"]
---


"Physical networks become what they learn" (arXiv:2406.09689) argues that after a
physical network is trained, its function can be recovered from its physical structure
alone: the task-critical edges become measurable *soft modes* of the network's
Hessian, so in principle one can reverse-engineer what a network was trained to do
without knowing the task. The paper connects this structure–function correspondence to
biology — per the capture's hop read, "conserved regions in proteins were associated
with slow collective modes."

This reframes learning as literal structural reorganization: training does not merely
set parameters, it reshapes the network's low-energy geometry, and that geometry is
where the function now lives. It generalizes the specific result that allosteric
response is cheap to induce
([[claim-removing-one-percent-of-bonds-makes-a-random-network-allosteric]]) into a
principle — the same soft-mode geometry that carries engineered long-range response
also carries evolved function in proteins — and it helps explain why a network trained
by a purely local physical rule
([[claim-coupled-learning-elastic-networks-compute-without-a-processor]]) can end up
structurally indistinguishable from a system that was deliberately designed. It is the
structure-function hinge of the vault's emerging physical-learning cluster and a
sibling to [[backpropagation-gap]]'s question of whether shared representations imply a
shared mechanism.

The load-bearing formalism — the physical-Hessian ↔ cost-Hessian correspondence and the
"reverse-engineer function from the Hessian" mechanism — is recorded here from a hop
read of the paper, not a verbatim primary extraction; only the phrase "conserved
regions... slow collective modes" is preserved verbatim. The note is held `seedling`
under `[unverified-mechanism]` and the formalism re-read is routed to
[[question-verify-physical-networks-hessian-cost-formalism]].