Skip to content

Latest commit

 

History

History
55 lines (44 loc) · 4.32 KB

File metadata and controls

55 lines (44 loc) · 4.32 KB

Proof complexity — the P-vs-NP certificate ladder

The repository's third research arc (alongside the Brualdi–Goldwasser campaign in proof/ and the Telperion engine): a program of kernel-checked lower bounds in proof complexity, climbing a ladder of progressively harder systems with the same discipline as the BG campaign — exact-arithmetic validation first, then Lean theorems whose only trusted component is the kernel. p_vs_np_proved = False, permanently implied; each rung is a certified formalization of a known lower bound, and the novelty claims live in the formalization, not the mathematics.

This page is the front door and index. The artifacts currently live inside the Telperion tree (they grew out of it and share its pinned Lean toolchain); nothing here is a second copy — every row links to the single source.

The rungs

Rung Result Status Write-up / code Lean (kernel-checked)
Knapsack (Grigoriev 2001) Symbolic-n SOS degree lower bound: the pseudoexpectation E[x_S] = (r)_{|S|}/(n)_{|S|} is a valid dual witness for {xᵢ² = xᵢ, Σxᵢ = n/2} for every odd n and every degree simultaneously — via a rank-one collapse of the harmonic blocks, g_k = ∏(n−2j)/(2(n−2j−1)) Discharged — 51 theorems, axioms clean; scalar layer + the d=4 Gram bridge (35 emitted identities); the harmonic-completeness layer is Python-pinned, stated honestly WRITEUP.md (the pipeline paper) · README.md · knapsack_pseudoexpectation.py KnapsackSOS.lean, SumEqProd.lean, BridgeD4.lean
3XOR (Grigoriev / Schoenebeck) Per-instance certified SOS lower-bound machinery: closure-consistency ⟹ block-rank-one PSD (a GF(2) class-rank-1 consistency argument), with Tseitin on the Petersen graph (refutation width exactly 6) as the canonical fully-certified instance Structure theorem kernel-checked (abstract partially-multiplicative-kernel PSD theorem); Petersen instance fully kernel-certified; the asymptotic expansion layer is not yet formalized xor3_pseudoexpectation.py, gen_petersen_cert.py (same directory) Xor3Structure.lean, PetersenCertificate.lean
Next (planned) Duality layer; LRS; a generic 3XOR instance emitter; planted clique (W2) open

What fed back into the engine

The arc crystallized three reusable Telperion emitter shapes (all drift-gated production examples now):

  • finite_decide — guarded decidable facts over ℕ-bitmask/sign tables, closed by kernel decide (lesson learned the hard way: ℚ does not kernel-reduce; state tables in ℕ/ℤ and cast);
  • fwd_telescope — forward-difference telescoping (the W2 prover): a whole Lean induction certified by one polynomial contiguous relation;
  • rational_identity — rational function identities on a ray (the Gram-bridge shapes).

Why the Lean lives where it does

The proof-complexity theorems compile inside the g1_floors Lean package — originally a BG artifact — because it is the repository's standalone pinned Lean + Mathlib shell, and the CI compile gate already covers it. That is a pragmatic co-location, not a conceptual one; if the arc keeps growing it should get its own lake package (tracked as future cleanup, deliberately not done while the formalization sessions are active in those files).

Honest status

See PUBLICATION_LEDGER.md rows 8–9 for the conservative novelty positioning: the mathematics is Grigoriev / Laurent / Kurpisz–Leppänen–Mastrolilli (knapsack) and Grigoriev / Schoenebeck (3XOR); what is claimed is the certified pipeline and the machine-checked symbolic-n statement — believed to be the first kernel-checked asymptotic proof-complexity lower bound, pending a formalization-literature check.