Skip to content

[FORM-S9] Predictive-coding energy as a negative log joint, with precision inferred #92

Closed

Description

@docxology

Scope ID: FORM-S9. Size: Minor. Scoped on 2026-10-08 in SCOPE-2026-10-08.md; acceptance remains open.

Owner: predictive-coding and Gaussian information-geometry foundations.

Dependencies: FORM-D3

Work

hierarchicalEnergy sums weighted squared errors over arbitrary variables; fep-091's primary theorem is monotonicity in a fixed weight. Prove (a) the linear-Gaussian chain energy equals the negative log native density up to a constant, and (b) ½πε² − ½ log π has the unique minimiser π* = 1/ε².

Acceptance

Both statements, plus a two-level witness tied to fep087_twoLevel_decomposition_nonvacuity.

Failure controls and claim boundary

No disposition promotion from compilation, import or numerical agreement; the maturity record changes only after independent semantic review. No duplicate topics and no automatic sealed-roster expansion.

Canonical references

Closure requires actual exact-source commands, results and retained receipts, with independent review where specified. Remove its open TODO row only after acceptance; retain delivery facts in changelog/history. Size describes implementation scope, not a semantic-version promise.

Activity

  1. docxology commented on Oct 9, 2026

    @docxology
    CollaboratorAuthor

    Status update (2026-10-09). Delivered in #160 (via #164), merged to main: Precision as inference: the unique minimiser π* = 1/ε², the derivative, strict convexity, and the ε = 0 unboundedness boundary. Also the linear-Gaussian chain energy as exactly −log ∏ N up to card/2·log(2π) via Mathlib's gaussianPDFReal, and a two-level witness.

    Evidence: lake --wfail build FepSketches passes at the pin (Lean v4.34.1, Mathlib d13f23b7), every new declaration depends only on propext/Classical.choice/Quot.sound, and all generator --checks plus hosted CI pass.

    Remaining for closure: independent semantic review of the delivered statements against the topic's theorem proxy. Per this issue's claim boundary, no disposition in config/theorem_maturity.yaml was changed, and none may change without that review.

  2. docxology commented on Oct 11, 2026

    @docxology
    CollaboratorAuthor

    Accepted on main via #171. In FEP.PredictiveCoding: gaussianChainEnergy_eq_neg_log_prod (energy is a negative log joint), precisionFreeEnergy_eq_iff (unique optimum 1/ε²) and twoLevel_chain_energy_witness, tied to the fep-087 two-level data. Any maturity change still needs independent semantic review. Row retired.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions