Weak-pair rank erosion

Let \(M,N\) be persistence modules on \(\mathbb R^n\) over a field \(k\), with finite-dimensional fibers. The current API permits every \(n\in\mathbb N\), including \(n=0\). Write \(r_M(a,b)\) for the dimension of the image of the actual structure map \(M(a)\to M(b)\), where \(a\le b\) coordinatewise.

The predicate and distance

For a finite nonnegative radius \(\varepsilon\), RankErosion ε M N means

\[r_M(a-\varepsilon\mathbf1,b+\varepsilon\mathbf1)\le r_N(a,b), \qquad r_N(a-\varepsilon\mathbf1,b+\varepsilon\mathbf1)\le r_M(a,b) \quad\text{for every }a\le b.\]

These are weak pairs: \(a=b\) is included, and \(r_M(a,a)=\dim_k M(a)\). The diagonal is significant, not an omitted endpoint convention. When \(n=0\), the grade set has one element and these diagonal tests still compare the fibers; a convention using only strict pairs would have no tests.

erosionLower and erosionUpper implement the two shifts. rankOfStructureMap_four_point proves that widening an interval decreases its rank. This yields RankErosion.refl and RankErosion.mono; symmetry and composition are RankErosion.symm and RankErosion.trans.

rankErosionDistance is the extended infimum of the admissible finite radii:

\[d_E(M,N)=\inf\{\varepsilon\ge0:\operatorname{RankErosion}(\varepsilon,M,N)\}.\]

An empty admissible set gives infinity. rankErosionDistance_self, rankErosionDistance_comm and rankErosionDistance_triangle give the extended pseudometric laws in the stated finite-dimensional scope. A supplied predicate witness gives an upper bound through rankErosionDistance_le_of_erosion. The definition uses Module.finrank, which is zero on infinite-dimensional spaces; it must not be read as a cardinal-rank theory without finiteness.

Stability from actual interleavings

An actual coordinate \(\varepsilon\)-interleaving factors the widened structure map through the other module's unshifted map:

\[M(a-\varepsilon\mathbf1)\longrightarrow N(a) \longrightarrow N(b)\longrightarrow M(b+\varepsilon\mathbf1).\]

The outside arrows are interleaving components. Naturality and the double-shift equations identify the composite with the original structure map, rather than assuming a rank inequality. The reverse interleaving gives the other inequality. CoordinateInterleaving.factor_structureMap supplies this factorization; CoordinateInterleaving.rank_erosion_le and CoordinateInterleaving.rankErosion extract the rank comparisons for pointwise finite-dimensional modules. Passing to infima proves

\[d_E(M,N)\le d_I(M,N)\]

as rankErosionDistance_le_coordinateInterleavingDistance. This includes zero radius, \(n=0\) and infinite distances; no minimizing interleaving is assumed.

Using a failed rank test

Choose a weak pair \(a\le b\) and prove

\[r_N(a,b)<r_M(a-\varepsilon\mathbf1,b+\varepsilon\mathbf1).\]
  • no_coordinateInterleaving_of_rank_violation excludes an actual \(\varepsilon\)-interleaving.
  • radius_le_coordinateInterleavingDistance_of_rank_violation gives \(\varepsilon\le d_I(M,N)\) without attainment.
  • radius_lt_coordinateInterleavingDistance_of_rank_violation gives \(\varepsilon<d_I(M,N)\) when both modules have original finite graded-free presentations (GradedFree.FinitelyPresented), in addition to the finite-dimensional instances in its signature. This uses the existing interleaving attainment theorem, not an erosion attainment theorem.

Interchange \(M,N\) for a violation in the other direction. A failed test alone does not justify a strict lower bound for arbitrary pointwise finite-dimensional inputs: an infimum need not be attained.

RankErosionWitness.originHook_diagonal_rank computes a nonzero diagonal rank for the principal upset over \(\mathbb Q\). originHook_no_zero_interleaving applies it against the zero module, including \(n=0\). originHook_positive_radius_no_interleaving uses the genuinely off-diagonal pair \(a=1,b=2\) at radius \(1\) in one coordinate. These are actual hook-module consumers, not an executable finite-test algorithm.

Zero radius, zero distance and future extensions

rankErosion_zero_iff identifies the predicate RankErosion 0 M N with equality of all weak-pair ranks. Consequently equal rank invariants give distance zero. This theorem does not identify the infimum equation \(d_E(M,N)=0\) with that predicate, nor does it classify modules up to natural isomorphism. In multiple parameters, ranks do not record how images of maps from incomparable grades meet, so rank equality cannot in general recover the module.

The next-release extension proposal separates the following targets: a precise comparison with strict-pair conventions; erosion attainment and strict erosion lower certificates for original finite presentations; image rank beyond finite-dimensional fibers; and certified finite grid tests with actual persistent-homology consumers. Executable tests additionally need effective grades, coefficients and a matrix-rank correctness bridge. These are proposed extensions, not declarations supplied by this note. The proposal also records endpoint and nonseparation counterexamples that a broader API must respect, rather than promising to remove necessary hypotheses.

Imports and source convention

Import TDA.Persistence for the public area API. Narrow imports are TDA.Persistence.RankErosion, TDA.Persistence.RankErosionInterleaving and TDA.Persistence.RankErosionWitness. Their declarations are in TDA.Persistence; the concrete hook witnesses are in its RankErosionWitness namespace. Coordinate notation uses open scoped TDA.Persistence.

Puuska, §2.2 defines the rank invariant on strict pairs and proves its erosion stability using the interleaving factorization (Theorem 22). The current library uses weak pairs and a finite-dimensional specialization; it does not yet prove equivalence with that strict-pair convention. The proposed convention comparisons and finite-presentation erosion attainment are additional library targets, not results attributed to the paper. See the source record.