2026-11-21: Energy is the right noise model, and morphology routes information
Question. Is the energy metric the best way to model a body’s resolution for measuring mechanical intelligence, or only a reasonable one?
Findings.
- Maximum entropy fixes the shape of the noise once energy is chosen as the constrained quantity, but it does not choose energy. Constraining mean action instead gives the quantum zero-point shape.
- Energy is the only noise metric that is coordinate-free, assembled element by element, free of outside scales, symmetric in time and invariant under free motion (Theorem 7.1).
- Invariance under free motion is equivalent to “free motion never creates information” (Proposition 7.1). Under per-coordinate noise, free motion raised the stored information from 2 to 3.58 bits with no input at all.
- Of the alternatives, only action weighting passes both the stiff-spring limit and the mesh-refinement limit. Locality and the classical regime exclude it. The frequency-weighted family converges in the stiff limit exactly when \(p<2\), as predicted (Table 7.1).
- Whitening by the data gives every body the same bits (Proposition 7.2).
- In a Langevin body, with noise injected through the dampers, every design gives the same 37.742 bits when the full state is observed. The limit is set by the damping at the contact, \(JD^+J^\top=1/b_i\) (Proposition 8.2). Through the motor encoder, the rigid transmission captures 99.9% of that limit, and the SEA climbs from 6% to 99.5% with stiffness (Table 8.1).
- Readout-only noise lets a resonant SEA beat the rigid leg through the encoder by up to 5.9 bits. Under the fluctuation-dissipation theorem the resonance amplifies the thermal noise by the same factor, so that advantage is an artifact.
Open from this note. The constant 1/2 in Proposition 8.2 is confirmed numerically but proved only in continuous time (roadmap thread T3). The routing efficiency in point 6 needs a definition and properties of its own (thread T2).