Stability of finite graded simplicial homology

Let \(C\) be a finite simplicial complex whose faces are nonempty finite sets of vertices. Assign a vector \(g(s)\in\mathbb R^n\) to each face, with

\[ s\subseteq t \quad\Longrightarrow\quad g(s)\le g(t) \]

in the coordinatewise order. At a parameter \(a\in\mathbb R^n\), the sublevel

\[ C_g(a)=\{s\in C:g(s)\le a\} \]

is a subcomplex. Its homology in degree \(q\in\mathbb N\), over an arbitrary field \(k\), forms a persistence module \(M_g(a)=H_q(C_g(a);k)\); the structure maps are induced by inclusions of subcomplexes. These are ordinary simplicial homology groups, so the empty complex has zero homology in every degree.

Filtration.FiniteGraded C n stores exactly the face grades and their monotonicity, with C : Simplicial.FiniteComplex m on vertices Fin m. finiteSimplicialHomology k g q constructs this module through the existing simplicialHomologyPModule.

Finite presentation from the input

The module \(M_g\) satisfies the original GradedFree.FinitelyPresented predicate. A finite presentation consists of finitely many graded generators and relations whose cokernel is naturally isomorphic to \(M_g\). finiteSimplicialHomology_finitelyPresented derives this data from \(C,g\).

For each coordinate, take the finite set of values occurring among the face grades, include zero in each coordinate set, and let \(G\) be their product grid. Whenever \(a\) has a grid point below it, the coordinatewise greatest such point \(\lfloor a\rfloor_G\) satisfies

\[ g(s)\le a \quad\Longleftrightarrow\quad g(s)\le\lfloor a\rfloor_G. \]

Thus the inclusion from the floor sublevel to the original sublevel is an equality of complexes. When no lower grid point exists, no face can occur: the grade of any occurring face would itself be a lower grid point.

Restrict \(M_g\) to \(G\) and take its left Kan extension along the grid inclusion. The canonical comparison to \(M_g\) is an isomorphism at every parameter: at a floor it is the inclusion-induced homology isomorphism, and outside the lower-grid region both fibres are zero. Since the comparison is a natural transformation, it preserves all structure maps. The restricted module has finite-dimensional fibres on a finite poset. The existing finite-poset presentation theorem and finite-grid Kan extension theorem therefore give the required graded-free presentation of \(M_g\).

This construction permits repeated grades and empty complexes. In zero coordinates the grid product is a singleton and every sublevel is \(C\); the same argument gives finite presentation of the resulting constant module.

Perturbing the grades

For two monotone grade assignments \(g,h\) on the same complex, let

\[ \delta=\max_{s\in C,\,i\in\{1,\ldots,n\}} |g(s)_i-h(s)_i|, \]

with value zero when the indexing set is empty. The declaration FiniteGraded.supGradeError uses this finite maximum. More generally, any nonnegative upper bound \(\varepsilon\) for these coordinate errors gives

\[ C_g(a)\subseteq C_h(a+\varepsilon\mathbf1),\qquad C_h(a)\subseteq C_g(a+\varepsilon\mathbf1). \]

Functoriality of inclusion-induced homology makes the two composites the structure maps shifted by \(2\varepsilon\). This constructs an actual coordinate interleaving, finiteSimplicialHomology_interleaving, and proves

\[ d_I(M_g,M_h)\le\delta. \]

For \(n\ge1\), the derived finite presentations define the canonical signed hook barcodes \(B_g,B_h\). Applying the existing signed hook stability theorem gives

\[ d_{\mathrm{signed}}(B_g,B_h)\le (2n-1)^2\delta. \]

finiteSimplicial_signedHookBottleneckDissimilarity_le accepts any nonnegative error bound, without a caller-supplied presentation, resolution, or interleaving. In one coordinate the constant is one. The reused signed bound follows Botnan–Oppermann–Oudot–Scoccola, Theorem 6.1. Its signed cross-sum comparison is a dissimilarity; it has no triangle inequality in general, as explained in their Remark 6.3 and the library's two-coordinate example.

The constructions are mathematical APIs and may be noncomputable. They do not provide an executable multiparameter barcode algorithm.