Skip to content

Prove sharp uniform bounds for triangular parity macro cuts - #4722

Closed
lensora wants to merge 1 commit into
CrocSwap:mainfrom
lensora:brian/triangular-parity-prefix-audit
Closed

lensora wants to merge 1 commit into
CrocSwap:mainfrom
lensora:brian/triangular-parity-prefix-audit

Conversation

@lensora

@lensora lensora commented Oct 22, 2026 •

Copy link
Copy Markdown

Superseded by PR #4731, reopened under the intended contributor account, JoyboyBrian. The tested commit and research contents are unchanged.

The triangular Toeplitz parity compiler's one-level bounds do not directly control an unfinished child composed with its ancestor shear. This package proves a sharp, depth-independent common infinity-norm bound of 1/(1−η) for every prefix, suffix and their inverses at complete-shear boundaries, for every dyadic dimension and every legal interleaving of the disjoint child programs. At tail l1 ≤ 1/2, all four bounds are at most 2. This includes the source's parallel child schedule.

The proof simultaneously inducts on all four cut operators while retaining the shared even/odd coefficient budget. The geometric inverse family proves the common constant is sharp. An exact N=4 suffix of norm 7/4 shows why a one-level 3/2 cap cannot cover all composed cuts. The sharper forward-prefix-only 3/2 question remains open.

This sharpens macro-boundary conditioning within the polynomial stability claim already present in SHALLOW-TOEPLITZ.md; it does not replace the internal dirty-shear gate, rounding, full Gaussian, normalization or supplier-admission obligations. No new κ or selected-frontier change.

Validation:

  • Python 3.12 registered metadata/source-pin validation and package replay pass.
  • Seven exact rational fixtures at N=1,2,4,8,16: 116 distinct cut matrices, including all pairs of child cuts for each fixture; all four operators checked.
  • Exact suffix witness and geometric inverse checks; four failure controls; direct replay also passes under Python -O.
  • Committed-byte and declared-source-history publication preflight pass.
  • A second Codex agent independently reviewed the proof and checker and used a separate sparse row-action implementation for three interleaving schedules. Root agent reviewed the four block identities and inequalities. These are AI checks, not independent human or maintainer acceptance.

Credit: Vlado Vojdanovski/#1673 for the shallow compiler and Cinna Davis/#4100 for its later caller; Jose Manuel/#4198 for the separate rational-generator work; Ryanson Jonathan's one-factor-of-t question and unpaidmoon's #4371; James Reed's #4699 one-level budget and #4701/#4702/#4704/#4705/#4706 finite examples. Exact predecessor heads and source hashes are retained in the package. No unmerged executable dependency is required.

Canonical investigation: #4715 . Base: b58ad9ef4d4e861b0eaa5dba41742bdc072dca51, priority revision 7. Prepared for Brian with substantial OpenAI Codex assistance; Apache-2.0.

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants