Skip to content

P104: node-partition refinement of the frontier-word ceiling, κ ≤ .0116750 (Lean, conditional, no κ claim) - #7678

Open
DreamingOfClouds wants to merge 1 commit into
CrocSwap:mainfrom
DreamingOfClouds:research/p104-tree-ceiling
Open

DreamingOfClouds wants to merge 1 commit into
CrocSwap:mainfrom
DreamingOfClouds:research/p104-tree-ceiling

Conversation

@DreamingOfClouds

Copy link
Copy Markdown
Contributor

Lean-only package, builds on #7668 (its 46 import-closure modules are included unchanged; 148 new modules). No κ claim and no word claim in Lean: all objects in the two theorems are abstract, and nothing in Lean ties a word to the hypotheses.

Results. Two theorems in Nov15/CongestionTree.lean, each proving, for t ∈ [t0, 1] (that is, t0 ≤ t ≤ 1), Finv (1 − t) (invoice (caller 6 7) n 17) < 0 (caller caller 6 7, L = 17) under the hypotheses listed in PROVED.md (copied from the Lean text):

  • ceiling_tree_CG: t0 = .0117694 (a0CG = 58847/5000000). Node-partition tree bound: the I2 paths of each source contain an out-tree of charts over the 230-shape DAG (cgR, root cgr0), each nested move is charged c2/n(A) per source tree, with the minimum over source counts (|J| ≥ 4962, ≤ 48 partners per source, ≤ 104 sources).
  • ceiling_tree_GD: t0 = .0116750 (a0GD = 467/40000). The same with |S| ≤ 104, Bset ⊆ S, 48 ≤ |Bset|: sources in Bset use gbR/gbr0 (the shape edges E48 realised by concrete nested charts), sources in S \ Bset use gaR/gar0.

Both sit on top of ceiling_native_J (t0 = .0118103, #7668): on [.0118103, 1] they apply it with its own hypotheses, and on [t0, .0118103] they use two kernel-certified tree pieces. Kernel-checked here: row soundness of three 230-row DP tables (one decide +kernel per row, no float, no shrinking), the tree cost bound for any out-tree, the double-counting reduction, the Lagrange line bound, a certificate that the integer move weights lower-bound c2 on the window, and the per-piece numeric certificates. All theorems print only propext, Classical.choice, Quot.sound; no sorry, no native_decide; every new module was compiled fresh (CHECK.txt).

Scope. One 104-port set. The 230 shapes, E48 and the 48-port class were computed for that port set only. Five word files share it (st1, #7637, x1t2a, #7660, and t7669, our reproduction of the word of #7669, not that file); the word-level numbers in NATIVES.md are for these five.

Hypotheses. All new hypotheses are in PROVED.md. In short:

  • Binders: the 32 explicit binders shared with ceiling_native_J (after the implicit {ι : Type*} [DecidableEq ι]) are textually identical to that theorem; new are σ ρ μ, T : TreeData, and hcstMonoT, hcst0, hcst1, hP1, hP2.
  • cstF premises: hcstMonoT (monotone on [a0, 1/10]), hcst0, hcst1 (two brackets at the new breakpoints).
  • TreeData fields: hJS, hScard, hfib, hJcard, hB, hBcard (48 ≤ |Bset| in GD), hcell, hxk, htreeA, htreeB, hnd, hMv, hcap, hMsub (Mv ⊆ Mreg ∪ Mun), hdim (1 ≤ dA < dB ≤ 8).
  • Tree premises Tr out-trees (htreeA, htreeB), hcap, hnd, hP1, hP2, hcell: open hypotheses, not verified on any word (no script was run for them). They are new statements, not consequences of hcong, although they rest on the same families (nested moves, I2 capacity, I2 charge). T, T.J and J are separate abstract objects.

Caveats, stated plainly.

  • Earlier (non-Lean) computations gave .0117604 and .0116675 using measured count profiles. The digits certified here are weaker by 9.0e-6 (.0117694 − .0117604) and 7.5e-6 (.0116750 − .0116675), because Lean takes only the stated counts. The certificate generator reported negative margins one digit lower (.0117693, .0116749); that is a generator output, not a Lean statement.
  • The count behind E48 (the (3,4,12) and (5,8,6) charts cover only 1,344 of the 4,992 even cells) is neither in Lean nor script-checked for this package. NATIVES.md recomputes the realised edge sets with an independent program (e48.c: all 29,212 subspaces, 2,350,136 nested pairs, all 4,992 even cells): three port classes of 24, 32 and 48 ports, and the 48-port class equals the gbR mask exactly (cmp.py); cgR has the same shapes and masks as gaR (cgcmp.py).
  • Word-level premises are computed by programs that are not in the package (NATIVES.md: premise × word table). aud.py replays modulo two primes (2^61 − 1 and 2^64 − 59); the others are exact integer or rational computations. aud2d.py was reused from P104: tiered ceilings for words meeting the native-class hypotheses (incl. the frontier words), Lean only #7668, as were the aud2a.py entries for st1, w7637, x1t2a and the hc12.py entries for all words but t7669. The owned layers agree in mass 56/56 but not in shape: t7669 has 23 rises / 23 descents and len ks = 36, the other four 23 / 24 and 37. Everything inherited from P104: tiered ceilings for words meeting the native-class hypotheses (incl. the frontier words), Lean only #7668 keeps its status there (hcong, hdist, hspMono, hcstMono, the brackets, the 8-row decoder constants; the bound |J| ≥ 4962 is a conjecture outside the known words).
  • The hypotheses lists in README.md and PROVED.md are generated from the Lean binders (gendocs.py, hyps.py).
  • CHECK.txt: the 148 new modules were compiled fresh, one at a time under a 3.5 GB guard. The 46 modules inherited from P104: tiered ceilings for words meeting the native-class hypotheses (incl. the frontier words), Lean only #7668 were not rebuilt for this package; their oleans are those of P104: tiered ceilings for words meeting the native-class hypotheses (incl. the frontier words), Lean only #7668's build of the same sources (prebuilt oleans).

The premises under study are SYLforge's (#7616, #7620); the caller is rohanarun's (#7113). DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code.

🤖 Generated with Claude Code

…= .0116750 (Lean, conditional, no kappa claim)

ceiling_tree_CG (t0 .0117694) and ceiling_tree_GD (t0 .0116750); kernel-only; builds on CrocSwap#7668. Tree premises are open hypotheses (not verified on any word); hypotheses copied from the Lean in PROVED.md.
DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.

1 participant