overview

Planned work, by section, and the ledger of open sorrys. A row moves to Done with the commit that completes it. Rows were proposed in the planning phase (2026-11-19); the maintainer approves sections for implementation (so far: J0, the nimber fields, J2a); Status.md says which.

Sections: V vault content, T tooling and site, L Lean, J Julia. Characteristic 2 comes first in every section.

V — Vault content

  • V0 — Restructure the vault. Folders by subject (proposal: Fields/, Characteristic 2/, Prime Fields/, Extension Fields/, Linear Algebra/, Analysis/, Category Theory/, Comparisons/, Algorithms/), a Map of Content.md, a concept-note template in vault/templates/, by-name wikilinks.
  • V1 — Foundations. One note each: field; characteristic; prime field; as ZMod p; finite field; order is a prime power; existence and uniqueness of (GaloisField); multiplicative group is cyclic; primitive element; Frobenius; subfield lattice ↔ divisibility of 𝑛; Galois group is cyclic generated by Frobenius; trace and norm; minimal polynomial; irreducible polynomials and counting them; Conway polynomials.
  • V2 — Characteristic 2 (priority). and Boolean arithmetic; −1=1 and addition is XOR; polynomial basis of ; every element is a square and the square-root map; Artin–Schreier equation 𝑥2+𝑥=𝑐 and the trace criterion; normal bases; tower fields; quadratic forms in characteristic 2 (symmetric ≠ alternating, Arf invariant); standard fields (AES), (GHASH); examples worked by hand.
  • V3 — Categorical view. Fields have no products and no initial object; the category of finite fields of characteristic 𝑝 and embeddings (hom-sets nonempty iff , then of size 𝑚); as a colimit; absolute Galois group ; finite étale algebras ↔ finite continuous -sets (Galois categories, Mathlib/CategoryTheory/Galois/); Frobenius as a natural transformation; ModuleCat over .
  • V4 — Comparisons with and . No ordering; only the trivial absolute value; discrete topology and no limits; scalar products: no InnerProductSpace, but bilinear/hermitian forms, isotropic vectors, failure of Gram–Schmidt, orthogonal complements that meet the space; differentiability: formal derivative, (𝑥𝑝)′=0, Hasse derivatives and Taylor expansion in characteristic 𝑝; Fourier analysis via additive characters (Walsh–Hadamard in characteristic 2); what “analysis over ” means (scope of FF-004).
  • V6 — Nimbers and Conway polynomials: nim-addition and nim-multiplication; the nimber fields and the quadratic closure; Conway polynomials and their compatibility as a functor on the divisibility poset; pseudo-Conway polynomials; changes of representation as linear maps. Lean: CombinatorialGames’ Nimber (check what it proves before relying on it).
  • V7 — Learners over finite fields (the maintainer’s experiment): what replaces gradients (discrete optimisation, interpolation, Reed–Muller, learning parity with noise), once the maintainer has written the ideas down.
  • V5 — Linear algebra over (was FF-003): counting subspaces, order, rank, solving; specification before code.

T — Tooling and site

  • T1 — Snippet sync tool. Takes a declaration name, asks Lean (against the pinned mathlib) for signature, docstring and module, and writes a > [!mathlib] callout between markers (…) with a link to the mathlib docs and an attribution line (mathlib is Apache-2.0). Plain Markdown, so Obsidian and Quartz both render it; a re-run refreshes it; a renamed declaration fails loudly. The same tool syncs our own Lean and Julia snippets from anchors in the sources. Caveat: the hosted mathlib docs follow master, not our pin. A Quartz plugin would not render in Obsidian, hence the generator.
  • T2 — Lean API docs with doc-gen4 (the tool behind mathlib’s docs), published under a subpath (e.g. <site>/lean/) beside the vault, as Lenticulum.jl publishes Documenter under api/; Julia API docs (Documenter) likewise under <site>/julia/. First check whether doc-gen4 can be limited to our modules: by default it documents every import, i.e. much of mathlib.
  • T3 — Note checker (Lenticulum’s check_notes.py as model): dangling wikilinks, missing tag lines, missing sources, unlisted sorrys.
  • T4 — CI, remaining: a benchmark job (on demand, not per push) once bench/ exists; a job that fails on an unlisted sorry (with T3).
  • T7 — Register FiniteFieldArithmetic.jl in the General registry (subdir=julia/FiniteFieldArithmetic.jl) once it exists and has a first release.

L — Lean

  • L0 — Library layout under lean/FiniteFields/ mirroring the vault subjects; one file per subject, each linked from its notes.
  • L1 — Characteristic 2: statements for V2 (squares, Artin–Schreier trace criterion, CharTwo lemmas), reusing Mathlib/Algebra/CharP/Two.lean and FieldTheory/Finite/*.
  • L2 — Categorical statements for V3, beside the classical ones.
  • L3 — Generalisation survey: list mathlib results stated only for , or RCLike that hold over any field, then restate them generically.
  • L4 — Proof sweep: discharge the sorry ledger.

J — Julia (FiniteFieldArithmetic.jl)

  • J1a — Representations: nimbers for 𝑛=2𝑘 (subfields nested as integers), Conway polynomials for every 𝑛 (pseudo-Conway beyond the tables), standard polynomials (AES, GHASH, …) for cryptanalysis and interoperability; selected by a type parameter; conversion maps between them. Decided (2026-11-19): nimbers default, Conway secondary, the others alternatives. Printing always hex; the REPL adds the field when it does not fill its storage.
  • J1b — Arbitrary 𝑛: a function (or macro, if a function cannot give type stability) that constructs for any 𝑛: picks the smallest fitting UInt for , multi-word storage above; finds an irreducible polynomial (low-weight trinomial or pentanomial search with an irreducibility test) or takes a Conway polynomial from a table.
  • J2 — Multiplication strategies, selected by dispatch and compared by benchmark; both table and computed variants where possible (maintainer, 2026-11-19), small fields probably tables only: shift-and-add, log/antilog tables (small 𝑛), carry-less multiplication (x86 PCLMULQDQ, ARM PMULL via LLVM intrinsics), reduction for sparse polynomials; profiled.
  • J3 — Inversion and division: Fermat, extended Euclid, Itoh–Tsujii; benchmarked.
  • J4 — in the smallest fitting UInt; Barrett/Montgomery as options.
  • J5 — generic.
  • J6 — Constant time: not a goal. Document which strategies happen to be constant-time; do not engineer for it.
  • J8 — Bulk and array operations for learners: broadcasting, SIMD/bitsliced bulk multiplication, matrix products over , random sampling; GPU-compatible element types. Benchmarked like single-element arithmetic.
  • J7 — (""): every element lives in some finite , so elements carry their degree and an operation embeds both operands into through Conway polynomials (pseudo-Conway beyond the tables). Exact, not fast; sizes grow with the lcm. The nimber quadratic closure (all 2{2𝑘}) is the cheap special case. After J1–J3. Lean counterpart: AlgebraicClosure (ZMod 2).

Sorry ledger

No sorry exists yet. Each row: file, declaration, note, date added.

filedeclarationnoteadded

Done

  • J0 — FiniteFieldArithmetic.jl skeleton (2026-11-19): tests first, Aqua, JET, DispatchDoctor, bench/ suite and profile script, implementation notes; CI tests Julia 1 and min.
  • J1 — nimber fields GF(2) … GF(2^128) (2026-11-19): Nimber{0..7} in the smallest fitting UInt; strategies Recursive, MulTable (default), LogTable sharing one recursion; measurements in src/nimber.md.
  • J2a — faster large nimber levels (2026-11-19): LogTable16 (GF(2^16) log/exp tables as the recursion base) and CLMul (polynomial basis with PCLMULQDQ, change of basis by byte tables); CLMul is the default for GF(2^128) (about 1.5 times faster than MulTable); details and measurements in src/nimber_fast.md, open question FF-P1.
  • T4a — CI (2026-11-19): path-filtered lean.yml and julia.yml, and full-ci.yml (all components plus a Quartz build) on the full-ci branch. Not yet run on GitHub.
  • T6 — Move to MathStruct (2026-11-19): repository transferred by the maintainer; baseUrl mathstruct.org/FiniteFields, footer link updated.
  • T5 — flake.lock and lean/lake-manifest.json committed (2026-11-19).
  • FF-001 — Prime-field starter (2026-11): Lean smoke test of mathlib’s ZMod field instance. Its Julia half was removed on 2026-11-19 (implemented without the maintainer’s go).