The thickening addition map

Setting

This digest continues reeb-geometric-thickening.md and uses its thickening without redefinition. For ε : ℝ≥0, Band ε = [-ε, ε], the thickening of (X, f) is X × Band ε with height height f ε (x, t) = f x + t, map g ε (x, t) = (g x, t), and functor ε is the thickening on the real slice Over (TopCat.of ℝ). The cosheaf side is the smoothing of reeb-smoothing.md, with its composition isomorphism smoothingCompIso ε δ : smoothing δ ⋙ smoothing ε ≅ smoothing (ε + δ).

For product thickening, T_ε T_δ ≅ T_(ε+δ) is false. Over a point at height 0 with ε = δ = 1, the height-0 fibre of T_1 T_1 is the segment {((*, s), -s)}, while that of T_2 is the single point (*, 0). The correct statement is a height-preserving addition map with lax coherence.

The addition map

addMap X ε δ : C((X × Band δ) × Band ε, X × Band (ε + δ)),   ((x, s), t) ↦ (x, s + t)

The outer parameter comes first: the source is T_ε (T_δ X), matching smoothingCompIso ε δ and smoothing δ ⋙ smoothing ε. The band coordinate lies in Band (ε + δ) because |s| ≤ δ and |t| ≤ ε (add_mem_band). The parameter ε + δ is never rewritten to δ + ε. Equalities of parameters are bridged by bandInclusion, never by casts on Band.

The laws, each an equation of continuous maps proved on coordinates:

  • Height. height f (ε + δ) (addMap X ε δ z) = height (height f δ) ε z, that is f x + (s + t) = (f x + s) + t (height_addMap).
  • Naturality in every continuous g, with no height hypothesis (addMap_natural).
  • Band enlargement in both parameters, ε ≤ ε' and δ ≤ δ' (addMap_comp_bandInclusion).
  • Associativity on ((X × Band γ) × Band δ) × Band ε, the two composites agree after (ε + δ) + γ ≤ ε + (δ + γ) (addMap_assoc).
  • Units. After 0 + δ ≤ δ, μ_(0,δ) is zeroHomeomorph; after δ + 0 ≤ δ, μ_(δ,0) is T_δ of zeroHomeomorph (addMap_zero_left, addMap_zero_right).
  • Zero sections. μ_(ε,δ) after the outer zero section is the enlargement δ ≤ ε + δ, and after T_ε of the inner zero section it is the enlargement ε ≤ ε + δ (addMap_comp_zeroSection, addMap_comp_map_zeroSection).

The addition map is surjective: (x, u) is the image of ((x, s), u - s) with s the projection of u onto [-δ, δ], and |u| ≤ ε + δ puts u - s in [-ε, ε] (addMap_surjective). For nonempty X and ε, δ > 0 it is not injective: with m = min ε δ > 0, the points ((x, 0), 0) and ((x, m), -m) differ and have the same image (addMap_not_injective).

The real slice

bandInclusionNatTrans (h : ε ≤ δ) : functor ε ⟶ functor δ
addNatTrans ε δ : functor δ ⋙ functor ε ⟶ functor (ε + δ)

The components are bandInclusion and addMap of the underlying space, by rfl (bandInclusionNatTrans_app_left_hom, addNatTrans_app_left_hom). The laws above hold as equations of natural transformations, with the associator and unitors definitional:

  • associativity, whiskerLeft (functor γ) (addNatTrans ε δ) ≫ addNatTrans (ε + δ) γ ≫ bandInclusionNatTrans _ = whiskerRight (addNatTrans δ γ) (functor ε) ≫ addNatTrans ε (δ + γ) (addNatTrans_assoc);
  • units against zeroIso, addNatTrans 0 δ ≫ bandInclusionNatTrans _ = whiskerLeft (functor δ) zeroIso.hom and addNatTrans δ 0 ≫ bandInclusionNatTrans _ = whiskerRight zeroIso.hom (functor δ) (addNatTrans_zero_left, addNatTrans_zero_right);
  • zero sections, whiskerLeft (functor δ) (zeroSectionNatTrans ε) ≫ addNatTrans ε δ and whiskerRight (zeroSectionNatTrans δ) (functor ε) ≫ addNatTrans ε δ are the enlargements δ ≤ ε + δ and ε ≤ ε + δ;
  • band enlargement in both parameters (addNatTrans_comp_bandInclusionNatTrans).

So (functor, zeroSectionNatTrans, addNatTrans) is a lax composition law over the additive monoid ℝ≥0, with unit and associativity holding up to band enlargement.

Path components

With ρ_ε = componentComparison f ε : C(T_ε f) ⟶ S_ε C(f),

intervalComponentNatTrans (addMap X ε δ) _ ≫ componentComparison f (ε + δ) =
  componentComparison (height f δ) ε ≫ (smoothing ε).map (componentComparison f δ) ≫
    (smoothingCompIso ε δ).hom.app (intervalComponents f)

(addMap_comp_componentComparison). On a representative ((x, s), t) both sides give the class of x in f ⁻¹ (thicken (ε + δ) I): the left side through (x, s + t), the right side through (x, s), then x in f ⁻¹ (thicken δ (thicken ε I)), then the transport of thickenCompIso ε δ, which is the identity on points. The equation uses the smoothingCompIso. On the slice it holds objectwise (map_addNatTrans_comp_componentComparisonNatIso) and as an equation of natural transformations (functor δ ⋙ functor ε) ⋙ 𝒞 ⟶ 𝒞 ⋙ smoothing (ε + δ) (whiskerRight_addNatTrans_comp_componentComparisonNatIso).

Since every other factor is an isomorphism, π₀ of the addition map is an isomorphism (isIso_intervalComponentNatTrans_addMap, isIso_map_addNatTrans_app), bijective at every interval (intervalComponentNatTrans_addMap_app_bijective), although the addition map is not a homeomorphism.

The one-point counterexample

onePoint is PUnit over height 0. The map fiberSegment : s ↦ ((*, s), -s) from Band 1 to T_1 T_1 of the point is injective and lands at height 0 (fiber_zero_injective). Every height-0 fibre of T_ε of the point is a subsingleton, since 0 + t = 0 forces t = 0 (fiber_zero_subsingleton). A slice isomorphism is injective and preserves height, so it would send fiberSegment 0 and fiberSegment 1 to one point. Hence

not_iso_one_one : IsEmpty ((functor 1).obj ((functor 1).obj onePoint) ≅ (functor (1 + 1)).obj onePoint)

and (addNatTrans 1 1).app onePoint is not an isomorphism (not_isIso_addNatTrans_app_one_one). The proof compares fibres; non-injectivity of μ alone would not exclude other isomorphisms.

Fixtures

TDA.Reeb.GeometricThickeningAdditionExamples:

  • One point. The non-isomorphism, the explicit pair ((*, 0), 0) ≠ ((*, 1), -1) with one image under μ_(1,1), the boundary point (*, 2) as the image of ((*, 1), 1), and bijective π₀ of μ_(1,1) at ℝ on the same object.
  • Boundary control. On the point, μ_(0,δ) and μ_(ε,0) are injective by the unit laws, so non-injectivity needs both parameters positive.
  • One edge. On the realization of singleEdge over 0 < 1, both sides of height_addMap at ((q, 1/4), 1/4) with q at height 1/2 are 1, and associativity at ε = δ = γ = 1/4 gives band coordinate 3/4.
  • Two merging edges. On parallelData, the naturality squares of addMap and addNatTrans for the edge swap send the point over edge true at thickened height 1/2 to the point over edge false, which differs from it. The collapse onto the single edge is checked the same way; its one-class target alone could not detect a wrong map.
  • Distinct classes. On the two disjoint segments, π₀ of μ_(1,1) at ℝ sends the classes over the labels false and true to classes separated by twoSegLabel of the first coordinate.

Source correspondence

de Silva, Munch and Patel, Categorified Reeb Graphs, §4.3 Definition 4.22, supplies the thickening X_ε = X × [-ε, ε], f_ε(x, t) = f(x) + t and T_ε[α] used here. Its Proposition 4.16, the cosheaf-side composition law, is the merged smoothingCompIso. §4.4 Observation 4.33 asserts semigroup isomorphisms U_(ε₁+ε₂) ≃ U_(ε₁) U_(ε₂) for the Reeb smoothing U_ε = ℛ T_ε.

The addition map, its coherence laws, the π₀ compatibility and the one-point non-isomorphism are library statements; none is attributed to the source.

Not claimed here: the Reeb quotient comparison ℛ T_ε T_δ ≅ ℛ T_(ε+δ), which is proved in reeb-quotient-thickening-addition.md; the absorption ℛ T_ε ℛ ≅ ℛ T_ε; the semigroup isomorphisms of Observation 4.33 for geometric smoothing; the geometric interleavings of Definition 4.36 and any distance inequality; a first projection or any other non-height-preserving map packaged as a slice morphism. The one-point witness shows that the analogue of Observation 4.33 for T itself fails.

Source: de Silva, Munch and Patel, Categorified Reeb Graphs, §4.3, Definition 4.22, and §4.4, Observation 4.33.