Skip to content

Roadmap Numerical K Theory

Chris Dare edited this page Nov 15, 2026 · 1 revision

Numerical K Theory

schema: roadmap/1
slug: numerical-k-theory
project: derived-alg-geo-lean
title: "The numerical lane — one Grothendieck group, and dimension as a parameter"
status: active
phase: sequenced          # init -> refined -> decomposed -> sequenced -> complete
brief: |
  This track exists because the lane it covers had no track.

  As of 2026-08-27 four roadmaps are active — contract-v1, dg-enhancements,
  stability-families, theorem22-2-dependencies — and none owns K-theory,
  Riemann--Roch, or numerical geometry. That is 261 files under
  `AlgebraicGeometry/` with no epic, no milestone, and no review pressure.

  It is also, precisely, where the duplication is. An adversarial architecture
  audit run on 2026-08-26 found a second complete hand-rolled Grothendieck
  group, two surviving copies of the axiom triple that #760 announced it had
  eliminated, an eleven-declaration verbatim fork between a Surface file and a
  HigherDimension file, a documented-and-kept fork of the real Mukai extension,
  and three parallel directories in which dimension is encoded as tuple arity.
  Every one of those is in this lane. None is in the stability or dg lanes,
  which are tracked.

  The claim this track rests on is that the correlation is causal: untracked
  work accretes duplicates because nothing is watching the shape. The
  deletions below are the repair; the track itself is the fix.

  What this is NOT: a rewrite of the numerical mathematics. Every epic here
  either deletes a copy in favour of an existing definition or replaces an
  enumeration with a parameter. No new mathematical content is claimed, and the
  Riemann--Roch and intersection-theory statements are untouched except where
  they name a type that moves.
goal:
  hmw: "How might we make the numerical lane carry the same abstraction tower the stability and dg lanes already do — one Grothendieck group, invariants as bundled homs, dimension as an index — without restating a single theorem of its mathematics?"
  objective: "Collapse every hand-rolled Grothendieck-group construction and axiom-triple carrier in AlgebraicGeometry onto GrothendieckPresentation and K₀Ab, and replace the three dimension-indexed directories with one parameterised statement, lowering the AlgebraicGeometry unaudited ceiling in the process."
  key_results:
    - "Exactly one Grothendieck group construction exists in the repository; grep for a second returns nothing. `CoherentGrothendieckGroup` and its bespoke universal property are gone, replaced by `K₀Ab (Coh X)`."
    - "No structure in AlgebraicGeometry carries a function together with `_iso` and `_shortExact` (or `_zero`) fields. Every additive invariant is an `AddMonoidHom` out of `K₀Ab`, and the isomorphism and additivity laws are theorems obtained from it rather than hypotheses supplied to it."
    - "`ReconstructionSystem` exists once. The Surface case is obtained from the higher-dimensional one at n = 2 rather than by a second copy of eleven declarations."
    - "`RealMukai` and `RealExtension` are one type with one bundled pairing; the cross-reference that names one in the other's docstring is replaced by an import."
    - "Dimension is a parameter. `SurfaceNum`, `ThreefoldNum` and `FourfoldNum` are replaced by a single indexed type, and the hard-coded dimension names in CI move with it."
    - "The AlgebraicGeometry ceiling in `scripts/check_audit_complete.py` is strictly lower at the end of this track than the 1059 it starts at, and every drop is recorded in the PR that causes it."
  assumptions:
    - tier: must
      text: "`Coh X` is abelian for every `Variety k`, so `K₀Ab (Coh X)` is defined and is the same object `CoherentGrothendieckGroup` hand-rolls."
      validation: "CONFIRMED 2026-08-26 against the tree: the `Abelian (Coh X)` instance is in `CoherentSheaf/Abelian/Basic.lean` and applies to every `Variety k`."
    - tier: must
      text: "The `iso` generator in `CoherentGrothendieckRelation` is redundant, so dropping it loses nothing."
      validation: "CONFIRMED: `K₀Ab.of_iso` is a theorem in `CategoryTheory/GrothendieckGroup/Abelian.lean`, derived from the short exact sequence `X ≅ Y ⟶ 0`. The hand-rolled relation carries it as a constructor instead."
    - tier: must
      text: "The positivity fields of an invariant — `rank_nonneg`, `degree_pos_of_rank_zero` and their analogues — do NOT descend to K₀ and must survive as fields."
      validation: "CONFIRMED by inspection: they are statements about objects, and a difference of classes can have negative rank. They are rephrased through the hom, never deleted. This is the single most likely way to get these epics wrong."
    - tier: should
      text: "The fourfold numerical formulas admit the same uniform shape as the surface and threefold ones, so one indexed type covers all three."
      validation: "FALSIFIED 2026-08-27. The Chern-character coefficients are not uniform in dimension, and the repository already said so: `Fourfold/LinearSection.lean:54` records that the `7/12` and `-3/2` in entries 3 and 4 are 'the terms that do not follow the threefold pattern', coming from the expansion of `(1 - e^{-H})^k`. Concretely, threefold entry 3 is `E.2.1/6 - E.2.2.1 + E.2.2.2/d` and fourfold entry 3 is `E.2.1/6 - E.2.2.1 + E.2.2.2.1`: the `/d` degree normalisation sits in the top slot, so 'entry 3' denotes different things in dimension 3 and 4. A second disproof: `SurfaceNum` is shared by `ProjectivePlane.lean` and `Abelian.lean` with DIFFERENT chCoeff formulas, so these are not one-formula-per-dimension types at all -- they are coordinate types, and the formulas are per-variety data. e5 is parked; e6 carries the part that survives."
  wont:
    - "No new numerical mathematics. Every epic deletes a copy or parameterises an enumeration; none proves a new theorem about a variety."
    - "No merge of the Mukai lattice theory above the real quadratic extension. Mukai lattice structure is special to K3 and abelian surfaces and does not generalise to arbitrary dimension; only `ℝ × V × ℝ` with its pairing unifies."
    - "No isomorphism `K₀Ab 𝒜 ≅ K₀ C`. That needs a bounded t-structure and is not available at the pin. The comparison stays the one-way `K₀Ab.toAmbient` that #760 shipped."
    - "No `sorry`, including as a refactoring convenience while a copy is being removed."
    - "No stability-lane or dg-lane work. `ClassDatum`, `heartDatum` and the enhancement tower belong to stability-families and dg-enhancements respectively."
  evidence:
    - "The 2026-08-26 architecture audit: four adversarial critics over the tree at 3079d8a, each required to run a mathematical soundness gate and drop unifications that would force a `sorry`. Findings 02, 03, 05, 06 and 08 are this track's epics."
    - "AlgebraicGeometry/Cohomology/EulerCharacteristic/Additivity.lean:360-378 — `CoherentGrothendieckGroup` rebuilds `FreeAbelianGroup ⧸ AddSubgroup.closure {relations}`, its own class map and its own lift, and imports nothing from `GrothendieckGroup/`."
    - "AlgebraicGeometry/RiemannRoch/Grothendieck.lean:31 and AlgebraicGeometry/Variety/Numerical.lean:63 — the two axiom triples that survived #760, which claimed in `GrothendieckGroup/Abelian.lean` that those fields had stopped being data. #766 collapsed `SlopeData`; these two were not in it."
    - "AlgebraicGeometry/RiemannRoch/Surface/NumericalVariety.lean:102-206 vs AlgebraicGeometry/RiemannRoch/HigherDimension/Hirzebruch.lean:124-222 — eleven declarations, ~80 code lines, differing only in line wrapping and two docstring phrasings. Diffed with comments stripped."
    - "LinearAlgebra/Lattice/Mukai/RealForm.lean:76 — names `Walls/Spherical/Basic.lean`'s `RealMukai` in prose as 'the same underlying type carrying a bare pairing', rather than importing it. The fork was seen and kept."
    - "AlgebraicGeometry/Numerical/Examples/{Surface/RankOne.lean:192, Threefold/LinearSection.lean:61, Fourfold/LinearSection.lean:50} — `ℤ×ℤ×ℤ`, `ℤ×ℤ×ℤ×ℤ`, `ℤ×ℤ×ℤ×ℤ×ℤ`. Dimension is the arity and the arity is typed out."
    - ".claude/roadmap/{contract-v1,dg-enhancements,stability-families,theorem22-2-dependencies}.yaml — searched 2026-08-27 for K-theory, Grothendieck, Riemann--Roch and numerical ownership. Four incidental mentions, zero epics. This lane was unowned."
generated_by: { agent: claude-opus-5, at: "2026-08-27" }
generations:
  - { rev: 1, at: "2026-08-27" }
items:
  - id: numerical-k-theory-m1
    gh_milestone: null
    kind: milestone
    title: "NK1 — Absorb: one Grothendieck group, and invariants as homs"
    status: done
    lane: now
    acceptance:
      - "`CoherentGrothendieckGroup`, `coherentGrothendieckRelations`, `CoherentGrothendieckRelation` and `coherentGrothendieckClass` no longer exist; every consumer names `K₀Ab (Coh X)`."
      - "`CoherentAdditiveInvariant` and the `classOf` triple in `NumericalVarietyData` each carry one `AddMonoidHom` field where they carried a function and two laws."
      - "Every positivity field that does not descend to K₀ is preserved and rephrased through the hom, and each such field is named in its PR description."
      - "`ReconstructionSystem` is declared once; the Surface lane obtains it by instantiation."
      - "The AlgebraicGeometry ceiling is lower than 1059 and the new figure is recorded, per CONTRIBUTING."
      - "Zero sorries; all gates green, including `check_roadmap.py --require-api`."
  - id: numerical-k-theory-e1
    gh_issue: null
    gh_sub_issues: []
    kind: epic
    parent: numerical-k-theory-m1
    title: "NumericalHypotheses's class map becomes a hom out of K₀Ab"
    status: done
    summary: "`NumericalHypotheses` (`Variety/Numerical.lean:63`) carried `classOf`, `classOf_iso`, `classOf_shortExact`. The two laws say the function factors through the Grothendieck group, so they were never data: three fields become `classOfHom : K₀Ab (Coh X) →+ N`, with `classOf` surviving as an object-level abbreviation and both laws as theorems of unchanged name and shape. This is the second of the three axiom triples #760 announced; #766 collapsed `SlopeData`. Scoped to `NumericalHypotheses` alone because it targets an abstract `N` and is therefore separable — see e2 for why `CoherentAdditiveInvariant` is not."
    priority: must
    size: M
    tags: [lean, k-theory, refactor]
    labels: [area:geometry, type:feature, type:finding]
  - id: numerical-k-theory-e2
    gh_issue: null
    gh_sub_issues: []
    kind: epic
    parent: numerical-k-theory-m1
    title: "Delete CoherentGrothendieckGroup, and with it CoherentAdditiveInvariant"
    status: done
    summary: "A second, complete Grothendieck group: `Additivity.lean:360-378` rebuilds the quotient, the class map and the universal property, with the property re-proved at `RiemannRoch/Grothendieck.lean` and the defining relation re-derived by hand at `Surface/Devissage.lean`. ~176 lines and 6 restated proofs. The `iso` relation generator is redundant because `K₀Ab.of_iso` is a theorem. **`CoherentAdditiveInvariant` is part of this epic, not e1.** Its only purpose is `grothendieckHom : CoherentGrothendieckGroup X →+ M`, so collapsing it onto `K₀Ab` IS deleting `CoherentGrothendieckGroup`; separating them would mean building a bridge between two Grothendieck groups in order to delete one. Discovered on contact with the code while implementing e1, and recorded here rather than left as a split that was abandoned."
    priority: must
    size: L
    tags: [lean, k-theory, deletion]
    labels: [area:geometry, type:feature, type:finding]
    blocked_by: [numerical-k-theory-e1]
  - id: numerical-k-theory-e3
    gh_issue: null
    gh_sub_issues: []
    kind: epic
    parent: numerical-k-theory-m1
    title: "One ReconstructionSystem, obtained at n = 2 for surfaces"
    status: done
    summary: "Eleven declarations exist verbatim in `Surface/NumericalVariety.lean:102-206` and `HigherDimension/Hirzebruch.lean:124-222`; diffed with comments stripped, the only differences across ~80 code lines are line wrapping and two docstring phrasings. Both consume `CoherentGrothendieckGroup`, so the second K-group has already propagated into two copies of an API. Surface is the special case of higher-dimension. Blocked on e2 so the merge happens once, against the surviving carrier."
    priority: must
    size: M
    tags: [lean, deduplication]
    labels: [area:geometry, type:chore, type:finding]
    blocked_by: [numerical-k-theory-e2]
  - id: numerical-k-theory-m2
    gh_milestone: null
    kind: milestone
    title: "NK2 — Unify: parameters, not enumerations"
    status: planned
    lane: next
    acceptance:
      - "One real quadratic extension type with one bundled pairing; the prose cross-reference in `RealForm.lean:76` is replaced by an import."
      - "The Mukai lattice theory above the real extension is untouched, and the module docstring says why it does not generalise."
      - "Dimension is an index. Either all three of surface, threefold and fourfold are covered by one type, or the fourfold case is parked with its obstruction stated — both are acceptable exits."
      - "The hard-coded dimension names in CI move with the types, in the same change."
      - "Zero sorries; all gates green."
  - id: numerical-k-theory-e4
    gh_issue: null
    gh_sub_issues: []
    kind: epic
    parent: numerical-k-theory-m2
    title: "One RealExtension, bundled"
    status: in_progress
    summary: "`RealMukai V := ℝ × V × ℝ` (`Walls/Spherical/Basic.lean:109`) and `RealExtension V := ℝ × V × ℝ` (`LinearAlgebra/Lattice/Mukai/RealForm.lean:79`) are the same type with the same pairing `q c c' - r·s' - r'·s`, in two different top-level subjects. One is a bare `def`, the other a bundled `BilinForm` plus `QuadraticForm`. `RealForm.lean:76` names the other in a docstring rather than importing it: the shared abstraction was seen, described in prose, and not written. Unify at the bundled form. The Mukai lattice structure above it does NOT generalise and is out of scope."
    priority: should
    size: M
    tags: [lean, deduplication, lattice]
    labels: [area:geometry, area:stability, type:chore, type:finding]
  - id: numerical-k-theory-e5
    gh_issue: null
    gh_sub_issues: []
    kind: epic
    parent: numerical-k-theory-m2
    title: "Dimension as an index, not as tuple arity"
    status: parked
    summary: "PARKED 2026-08-27, assumption falsified. The plan was to replace `SurfaceNum`, `ThreefoldNum` and `FourfoldNum` with `Fin (n+1) → ℤ` and share one set of Chern-character formulas. The formulas are not shareable: they differ per dimension by terms the repository itself documents as not following the pattern, and `SurfaceNum` carries different formulas for different surfaces besides. The type-level part of the idea survives and moves to e6; nothing of the formula-level part does. Parking rather than forcing it is the exit this epic was written to allow -- it was tagged `falsifiable` from the start."
    priority: should
    size: L
    tags: [lean, polymorphism, falsifiable]
    labels: [area:geometry, type:feature, type:finding]

  - id: numerical-k-theory-e6
    gh_issue: null
    gh_sub_issues: []
    kind: epic
    parent: numerical-k-theory-m2
    title: "Numerical coordinates as an indexed family, for the boilerplate alone"
    status: planned
    summary: "What survives e5's falsification, scoped on its own evidence rather than inherited from it. `Fin (n+1) → ℤ` replaces the three nested-tuple coordinate types. The claim is NOT that the Chern-character formulas unify -- they demonstrably do not -- but that the additivity proofs about them are boilerplate produced by the encoding: `threefoldChCoeff_add` is 63 lines and `fourfoldChCoeff_add` is 78, both case-splitting to show that projection from a nested tuple is additive, which `Pi.add_apply` gives for free. Rank also becomes `E 0` uniformly. Cost: every construction site across the eight Examples files, plus twelve hard-coded dimension names in CI. Worth measuring against that 141-line saving before starting."
    priority: could
    size: L
    tags: [lean, polymorphism, boilerplate]
    labels: [area:geometry, type:chore]

Exported from .claude/roadmap/numerical-k-theory.yaml at the commit that removed it from the tree.

Home

Architecture

CI and runners

Decision records

Roadmaps

Clone this wiki locally