Attainment of the multiparameter interleaving distance
Let R be a ring, and let M and N be finitely presented persistence modules on coordinatewise ordered ℝⁿ. A finite presentation means a cokernel of a homogeneous map between free modules on finite graded generator families, together with a natural isomorphism to the given module. The construction is described in graded free modules and finite presentations.
Write 1 for the all-ones vector. For a nonnegative finite parameter ε,
write M[ε](a) = M(a + ε·1). An
ε-interleaving consists of natural maps M → N[ε] and N → M[ε] whose two
composites are the structure maps to M[2ε] and N[2ε]. The extended distance
dᵢ(M,N) is the infimum of the parameters admitting such maps; an empty
parameter set has distance infinity.
The attained threshold
For every finite ε ≥ 0,
dᵢ(M,N) ≤ ε if and only if M and N admit an ε-interleaving.
Consequently, every finite distance is attained. Distance zero is equivalent to natural isomorphism. When the distance is finite, the admissible parameter set is the closed ray [dᵢ(M,N),∞); when it is infinite, the set is empty. The distance is invariant under natural isomorphisms, symmetric and satisfies the extended triangle inequality. It thus separates isomorphism classes.
GradedFree.finitelyPresented_distance_le_iff_interleaving is the threshold
theorem. The FinitePresentationClosure module gives attainment,
separation and closedness. CoordinateInterleavingDistance supplies the
infimum and distance laws, and CoordinateInterleaving.zeroEquivIso identifies
zero interleavings with natural isomorphisms.
GradedFree.FinitePresentationIsoClasses is the quotient of finitely presented
modules by natural isomorphism. Its EMetricSpace instance uses the descended
interleaving distance, and FinitePresentationIsoClasses.edist_classOf identifies
that distance on representatives. The separation proof uses finite-presentation
attainment at zero.
These statements include the zero ring, empty presentations and n = 0. Finite-dimensional fibers and interval decompositions are unnecessary. Over a field and for n ≥ 1, the result is Lesnick, Theorem 6.1 and Corollary 6.2. The ring and zero-coordinate cases are generalizations of that source's domain.
Removing a small increase of the parameter
The proof uses two consequences of finite presentation. First, generator and relation grades give a finite set S of test points: if a natural map X → Y is bijective at every point of S, postcomposition gives a bijection Hom(M,X) → Hom(M,Y). One lifts generator images through X → Y; injectivity at relation grades makes the lifted relations vanish. Generator injectivity gives uniqueness. This criterion is proved for the presented cokernel and then transferred to M through its presentation isomorphism.
Second, the target N has finite critical coordinate sets. At any fixed point there is a positive right-diagonal slack on which its structure maps are isomorphisms. For a fixed ε, apply this at the finitely many translated test points a + ε·1, a ∈ S. The minimum of their positive slacks makes
Hom(M,N[ε]) → Hom(M,N[ε+δ])
bijective for every sufficiently small δ ≥ 0. The empty test set causes no restriction. This argument requires no single positive slack valid at every point of ℝⁿ.
Apply the Hom argument to the four pairs and parameters (M,N,ε), (N,M,ε), (M,M,2ε) and (N,N,2ε). Choose δ smaller than the first two slacks and half the last two. Given an (ε+δ)-interleaving, the first two Hom bijections lift both natural maps to shift ε. After postcomposing their double-shift composites to parameter 2ε+2δ, naturality identifies the results with the original interleaving equations. Injectivity for the last two Hom maps therefore proves both equations at 2ε.
The construction uses actual natural transformations. Its intermediate
algebraic step, CoordinateInterleaving.of_postcomp_lifting, accepts the two
surjections and two injections; the finite-presentation proof establishes
all four conditions. If dᵢ(M,N) ≤ ε, the infimum property gives an
interleaving below ε + η for the chosen positive slack η. Monotonicity
extends it to ε + η, and the lowering construction gives one at ε.
Why the finite-presentation hypothesis matters
For a field k, the real interval modules k_[0,∞) and k_(0,∞) are interleaved at every
positive parameter. Their distance is zero, while their fibers at zero have
different dimensions, so they are not naturally isomorphic. The declarations
RealRay.distance_eq_zero and RealRay.not_nonempty_iso prove these facts.
Thus distance zero does not separate arbitrary real-indexed modules.
Informally, the usual generator at zero gives the closed ray a finite presentation while the open
ray cannot have one under the convention above; this ray-presentability remark is not formalized.
The one-dimensional finite-constructible theorem also obtains attainment
from barcode isometry and an optimal finite matching; see
finite-constructible closure.
The multiparameter proof uses finite presentations directly and needs no
barcode decomposition. CoordinateRealBridge.interleavingEquiv identifies the
real-indexed and one-coordinate witness types, and CoordinateRealBridge.distance_eq
identifies their distances. Thus the real examples and the coordinate theorem
use the same one-dimensional distance.