Join-indexed resolutions of generated upsets

Let k be a field, P a lattice, and f : Fin m → P a finite indexed family. Its generated upset is

U = {p : P | ∃ i, f i ≤ p}.

For each nonempty set of occurrences A ⊆ Fin m, write j_A for the join of the values f i with i ∈ A. The corresponding summand is the principal-upset module born at j_A. In degree r, TDA.Persistence.GeneratedUpsetResolution.term f r is the finite biproduct of these modules over the faces with r+1 occurrences. The differential deletes each occurrence in increasing Fin m order, with the alternating sign, and maps the face module into the module of the smaller face. The augmentation maps degree zero into the support-interval module on U.

The construction preserves occurrences: if two values of f are equal, their singleton faces and the higher faces containing them remain distinct summands. Each term is projective because it is isomorphic to the graded-free module on its face generators, and each principal-upset module is projective.

At an index p, the active occurrences are i with f i ≤ p. The summands visible at p are exactly the nonempty faces of the active-occurrence simplex. The equivalence termEvaluationEquiv identifies the evaluated terms with its simplicial chains, sends the actual differential to the signed simplicial boundary, and identifies augmentation with the sum of degree-zero coefficients. If the active set is nonempty, coning by any active vertex contracts the simplex chain complex, including its augmentation. If the active set is empty, the evaluated face terms and the support-module fiber are all zero. Thus resolution_pointwise_exact_at gives surjectivity and exactness at every term, while differential_comp_differential and differential_comp_augmentation establish the chain laws.

The categorical chain complex is chainComplex f. Its object and differential equations are chainComplex_X and chainComplex_d. The canonical Mathlib projectiveResolution f has this complex, projective terms, and the original augmentation in degree zero, as recorded by projectiveResolution_complex and projectiveResolution_augmentation. Pointwise exactness and surjectivity make that augmentation a quasi-isomorphism to the complex concentrated in degree zero at the generated-upset module.

For every lattice and finite indexed family, generatedSupport_hasProjectiveDimensionLE proves Mathlib's ordinary projective-dimension bound m−1 for the actual support module; generatedSupport_projectiveDimension_le states its numerical form. No finiteness or decidability condition on the lattice is needed. The proof uses term_isZero_of_degree_ge: the degree-r term is zero for m ≤ r, since there are no faces with r+1 occurrences. Every Ext class in such a degree is represented by a morphism from that zero term of the canonical resolution, so the class vanishes. This computes ordinary dimension in the persistence-functor category.

The bound uses truncated natural subtraction. For m = 0, the generated support is empty and generatedSupport_fin_zero_module_isZero identifies the support module as zero; its numerical projective dimension is ⊥. For m = 1, the bound gives projectivity. Two equal labels still contribute two occurrences to the resolution, but their support equals that of a singleton family, so their module is already projective. The occurrence-count bound need not be sharp.

A sharp two-generator example is the upset generated by (true, false) and (false, true) in the lattice Bool × Bool. Its support consists of those two atoms and their common top. The module has one copy of the coefficient field at each supported point. It has ordinary projective dimension exactly one over every field. To prove nonprojectivity, suppose the free augmentation had a section. Naturality from each atom to the top would identify the section's top coefficient vector with vectors supported on each atom's available generators. These generator sets are disjoint, so the top vector would vanish, contradicting the section identity on the unit coefficient. The general bound supplies the upper bound of one.

The dimension theorem is an upper bound; it does not assert that the canonical resolution is shortest or minimal, or give a global-dimension bound. Proposition 5.1's separate strict-join condition belongs to its shortestness conclusion and is unnecessary for the upper bound formalized here.

The construction formalizes Botnan, Oppermann, Oudot and Scoccola, On the bottleneck stability of rank decompositions of multi-parameter persistence modules, arXiv:2208.00300v3, Proposition 5.1. The Lean interface is in GeneratedUpsetResolution.lean and GeneratedUpsetProjectiveDimension.lean; focused fixtures are in GeneratedUpsetResolutionExamples.lean.