Absorption of the Reeb quotient and the geometric smoothing semigroup

Setting

This digest continues reeb-quotient-thickening-addition.md and uses its objects without redefinition. On the real slice Over (TopCat.of ℝ), T_ε = functor ε is the product thickening, ℛ = ReebQuotient.quotientFunctor ℝ is the Reeb functor and q_X = (projNatTrans ℝ).app X is the Reeb projection X ⟶ ℛ X. On graph objects, U_ε = geometricSmoothing ε = ℛ T_ε (reeb-rgraph-smoothing.md), and the merged geometricSmoothingAddIso ε δ identifies ℛ T_ε T_δ A with U_(ε+δ) A. Write Φ = intervalComponentsOver for interval components and κ_X = intervalComponentComparison (overBaseMap X) : Φ X ⟶ Φ (ℛ X) for the comparison of reeb-constructible-comparison.md. The outer parameter comes first: U_ε U_δ is geometricSmoothing δ ⋙ geometricSmoothing ε. Everything is in universe 0.

The absorption morphism and its domain

Composing two geometric smoothings needs the absorption morphism

ℛ (T_ε q_X) : ℛ T_ε X ⟶ ℛ T_ε ℛ X,     [(x, t)] ↦ [([x], t)]

to be an isomorphism at X = T_δ A for a graph object A. That object is a constructible ℝ-space but need not be a graph object (at δ = 0, zeroIso identifies it with A), and its Reeb projection q_X need not be an isomorphism: on the single edge, T_(1/4) contains two distinct points at height 1/2 joined by a constant-height path (fixture below). So the merged projNatIso, which inverts q_A on graph objects, does not apply there. On a graph object A itself absorption is immediate from projNatIso, and that case is only a control.

Why no point-set inverse

A direct inverse would send [([x], t)] to [(x, t)]. That this is well defined requires lifting constant-height paths of T_ε ℛ X, which are paths in ℛ X of varying height together with a compensating band coordinate, through q_X. The library has no such path lifting, and a compact-to-Hausdorff argument would give continuity only once bijectivity is known. The proof therefore detects the isomorphism on interval components and reflects it.

The π₀ argument

  • Φ q_X = κ_X by definition (intervalComponentsOver_map_projNatTrans_app), and κ_X is an isomorphism for constructible X (DMP Theorem 3.17, merged). The merged first-projection comparison componentComparisonNatIso ε : T_ε ⋙ Φ ≅ Φ ⋙ S_ε is hypothesis-free, so Φ (T_ε q_X) is conjugate to S_ε (Φ q_X) and is an isomorphism (isIso_map_functor_map_projNatTrans_app).
  • The naturality square of κ at g = T_ε q_X reads Φ g ≫ κ_(T_ε ℛ X) = κ_(T_ε X) ≫ Φ (ℛ g). Since ℛ X is a graph object, T_ε ℛ X is constructible and κ_(T_ε ℛ X) is an isomorphism, so the left side is one. Hence κ_(T_ε X) is injective at every interval. It is surjective at every interval for every space (merged), so κ_(T_ε X) is an isomorphism (isIso_intervalComponentComparison_thickening), without using constructibility of T_ε X. Cancelling it, Φ (ℛ g) is an isomorphism (isIso_map_quotientFunctor_map_functor_map_projNatTrans_app).

Reflection and the hypotheses

Under

  • (H1) IsConstructibleSpace X, and
  • (H2) IsRGraph (ℛ (T_ε X)),

both ends of ℛ g are graph objects: the source by (H2), the target because T_ε ℛ X is constructible. As a morphism of RGraphCat, its image under the merged cosheafFunctor is Φ (ℛ g) in the full subcategory of constructible cosheaves. The inclusion of that full subcategory and cosheafFunctor are fully faithful, so both reflect isomorphisms, and ℛ g is an isomorphism (isIso_quotientFunctor_map_functor_map_projNatTrans_app). Its inverse sends [([x], t)] to [(x, t)] (absorption_inv_mk), so the underlying map is a height-preserving homeomorphism. Continuity of the inverse comes from the reflection, not from a point-set argument.

Only full faithfulness of cosheafFunctor is used, not its essential surjectivity or an inverse equivalence, so no critical-value or witness scope of cosheafEquivalence is inherited beyond universe 0. No compactness, Hausdorff or local path-connectedness hypothesis is added.

Discharge on graph objects. For A : RGraphCat and X = T_δ A, (H1) is the merged isConstructibleSpace_thickening and (H2) is the merged isRGraph_quotient_thickening_thickening; both are proved, not assumed (isIso_quotientFunctor_map_functor_map_projNatTrans_app_thickening). For X = A itself, (H1) is isConstructibleSpace_of_isRGraph and (H2) follows from isConstructibleSpace_thickening; there the inverse is ℛ T_ε of the inverse of projNatIso (inv_quotientFunctor_map_functor_map_projNatTrans_app_of_isRGraph).

Absorption on all constructible ℝ-spaces

For every constructible X and every ε ≥ 0, (H2) follows from (H1). For a closed window [a, b], exists_clampData_Icc supplies δ > 0 and clamp data with lo = a, hi = b on the open window (a − δ, b + δ). If two points of X|[a, b] have Reeb classes joined in ℛ X|[a, b], then, because κ is injective on open intervals, they are joined in X|(a − δ, b + δ), and the clamp retracts that path into X|[a, b] (joinedIn_Icc_of_joinedIn, joinedIn_Icc_of_isConstructibleSpace). By reebRel_height_iff, the level sets of T_ε X are precisely the closed windows X|[t−ε, t+ε]. Consequently, bijective_quotientFunctor_map_functor_map_projNatTrans_app makes ℛ (T_ε q_X) a continuous bijective map; its source is compact because the constructible pieces are compact, and its target is Hausdorff because it is a graph object, so it is a homeomorphism over ℝ (absorptionOverIso). No Hausdorffness of the pieces is used. Transport along this homeomorphism gives (H2) via isRGraph_quotient_thickening_of_isConstructibleSpace. This absorption argument does not use DMP15-v1 Theorem 4.27, and it needs no Hausdorffness assumption on X. Separately, the general form of that theorem says that T_ε X is constructible for every constructible ℝ-space X and every ε : ℝ≥0, with no T₂ hypothesis. Applying it first at δ to a constructible A and then at ε to T_δ A shows that T_ε T_δ A is constructible; see Theorem 4.27. Following the Reeb quotient, its lift thickeningFunctor ε is thickeningQuotient ε.

The result is packaged on the full subcategory of constructible spaces as

thickeningQuotient ε : ConstructibleSpaceCat ⥤ RGraphCat
thickeningQuotientAbsorptionIso ε : thickeningQuotient ε ≅
  quotientRGraphFunctor ⋙ geometricSmoothing ε
thickeningQuotientCompIso ε δ : thickeningQuotient δ ⋙ geometricSmoothing ε ≅
  thickeningQuotient (ε + δ)

The last isomorphism applies U_ε to thickeningQuotientAbsorptionIso δ, then applies geometricSmoothingCompIso ε δ on ℛ X, and then applies the inverse of thickeningQuotientAbsorptionIso (ε + δ), with only an associator between these composites. The absorption instance isIso_quotientFunctor_map_functor_map_projNatTrans_app_constructible removes (H2) from the merged absorption theorem.

Interleavings of constructible ℝ-spaces built on thickeningQuotient and thickeningQuotientCompIso are described in reeb-geometric-interleaving-equivalence.md, section Interleavings of constructible ℝ-spaces.

The semigroup

On graph objects absorption assembles into the natural isomorphism

geometricSmoothingAbsorptionIso ε δ : doubleThickeningQuotient ε δ ≅ geometricSmoothing δ ⋙ geometricSmoothing ε

with components ℛ (T_ε q_(T_δ A)) and inverse [([z], t)] ↦ [(z, t)]. Composing its inverse with the merged geometricSmoothingAddIso ε δ gives

geometricSmoothingCompIso ε δ : geometricSmoothing δ ⋙ geometricSmoothing ε ≅ geometricSmoothing (ε + δ)

sending [([(x, s)], t)] to [(x, s + t)], with inverse [w] ↦ [([σ w].1, (σ w).2)] for the merged section σ = addSection. Naturality in every graph map is part of the natural isomorphism.

Parameter equalities ε = δ are bridged by geometricSmoothingEqIso, the Reeb quotient of the band inclusions bandEqIso, never by eqToIso or casts. With it:

  • Associativity (geometricSmoothingCompIso_assoc): the two composites U_ε U_δ U_γ ≅ U_(ε+(δ+γ)) agree, with (ε + δ) + γ = ε + (δ + γ) bridged on the left.
  • Unit laws (geometricSmoothingCompIso_zero_left, geometricSmoothingCompIso_zero_right): against the merged geometricSmoothingZeroIso : U_0 ≅ 𝟭, which sends [(b, s)] to b.

All three are checked on representatives [([([(x, s)], t)], u)], where both sides give the class of (x, s + t + u).

Cosheaf compatibility (cosheafFunctor_map_geometricSmoothingCompIso_hom_app). Under the merged comparisons β_ε : C(U_ε A) ≅ S_ε C(A) of Proposition 4.32, the image of the semigroup on underlying precosheaves is the precosheaf semigroup smoothingCompIso ε δ: C(U_ε U_δ A) ⟶ C(U_(ε+δ) A) ⟶ S_(ε+δ) C(A) equals C(U_ε U_δ A) ⟶ S_ε C(U_δ A) ⟶ S_ε S_δ C(A) ⟶ S_(ε+δ) C(A). Both send the class of [([(x, s)], t)] to the class of x.

Fixtures

TDA.Reeb.QuotientThickeningAbsorptionExamples checks the theory on the merged fixture objects.

  • Negative control. The Reeb projection of T_(1/4) of the single edge is not an isomorphism: (q_(1/2), 0) and (q_(1/4), 1/4) are distinct, both at height 1/2, and joined by the constant-height path r ↦ (q_(1/2 - r/4), r/4). Absorption holds at that object.
  • Graph-object control. On the parallel edges with ε = 1/4, the inverse of absorption is ℛ T_(1/4) of the inverse of projNatIso, and it sends [([q_true], 1/4)] to [(q_true, 1/4)], computed through that description and through absorption_inv_mk.
  • Boundary control. ℛ q_X = q_(ℛ X) for every X, and on the point the left unit law sends [([(*, s)], 0)] to [(*, s)].
  • One edge. At ε = δ = 1/4, [([(q, 1/4)], 1/4)] ↦ [(q, 1/2)], the inverse returns [([(q, 1/4)], 1/4)] (the clamp of 1/2 onto [-1/4, 1/4] is 1/4), associativity gives [(q, 3/4)] on both sides, and the cosheaf square at ⊤ gives the class of q on both sides.
  • Two merging edges. The height-1/2 fibre of U_(1/4) U_(1/4) of the parallel edges is a subsingleton, transported from U_(1/4 + 1/4) through the height-preserving semigroup isomorphism, and the classes over the two edges agree there, while the height-1/2 fibre of the single smoothing U_(1/4) is not a subsingleton. The height-1/2 fibre of U_(1/8) U_(1/8) is not a subsingleton, by injectivity of absorption from the merged statement for ℛ T_(1/8) T_(1/8). Together these pin the threshold ε + δ ≥ 1/2.
  • Naturality. On the edge swap both composites of the square send [([(q_true, 1/4)], -1/4)] to [(q_false, 0)], through the merged swap square of geometricSmoothingAddIso; the representative differs from that of the identity image. The classes themselves coincide, since every fibre of U_(1/4 + 1/4) of the parallel edges is a subsingleton. At 1/8 + 1/8, where the loop survives, the square separates classes: both composites send [([(q_true, 0)], 0)] to the image of [([(q_false, 0)], 0)], which differs from the identity image, as geometricSmoothingEqIso carries the two classes to the distinct height-1/2 classes over the two edges in U_(1/4).
  • One point. (geometricSmoothingCompIso 1 1).app pointGraph is an explicit isomorphism U_1 U_1 P ≅ U_(1+1) P in RGraphCat, whereas no slice isomorphism T_1 T_1 P ≅ T_(1+1) P exists (not_iso_one_one).
  • Flat interval. interval_isConstructibleSpace, interval_isRGraph_quotient_thickening_quarter, and interval_isIso_absorption_quarter show that the constant-height interval is constructible but not a graph-object input for the merged route; its quarter-thickening has a graph quotient with an isomorphic absorption map.
  • Thickened pinch. pinch_isRGraph_quotient_thickening_quarter and pinch_thickeningQuotientAbsorptionIso_hom show that the quarter-thickened pinch has a graph quotient and an isomorphic absorption component despite the non-injective attaching map.
  • Point at ε = 0 and ε = 1. point_isRGraph_quotient_thickening_zero, point_isRGraph_quotient_thickening_one, and point_thickeningQuotient_one_fiber_subsingleton show that the one-point graph satisfies (H2) at ε = 0 and ε = 1, and every fibre of its unit thickening quotient is a subsingleton.
  • Negative control. id_thickening_quotient_not_isRGraph shows that the quotient of the unit thickening of id ℝ is not a graph object: the height of the unbounded input is surjective onto ℝ, while graph-object heights lie in [T.value 0, T.value (Fin.last m)], so the constructibility hypothesis cannot be dropped.

Source correspondence

de Silva, Munch and Patel, Categorified Reeb Graphs, §4.4 Definition 4.30 is U_ε = ℛ T_ε, merged as geometricSmoothing. The first bullet of Observation 4.33 asserts semigroup isomorphisms U_(ε₁+ε₂) ≃ U_(ε₁) U_(ε₂). Only its statement shape is attributed, on RGraphCat in the order U_ε U_δ ≅ U_(ε+δ), and it is proved here independently through absorption and the merged geometricSmoothingAddIso. Theorem 3.17 is used as the merged isIso_intervalComponentComparison_of_constructible, and Theorem 3.22 only through the full faithfulness cosheafFullyFaithful.

The absorption theorem, its hypotheses (H1) and (H2), the reflection proof, the coherence laws and the cosheaf compatibility square are library statements; none is attributed to the source.

The all-constructible absorption results and the thickeningQuotient packaging are library statements, not attributions to the source.

The geometric interleavings of Definition 4.36, their distance and the inequality d_cosheaf ≤ d_geo are in reeb-geometric-interleaving.md.

Not claimed here: the reverse inequality d_geo ≤ d_cosheaf; absorption on arbitrary objects of the real slice; a lift of smoothingCompIso to constructible cosheaves, so the compatibility square is stated on underlying precosheaves.

Source: de Silva, Munch and Patel, Categorified Reeb Graphs, §4.4, Definition 4.30 and Observation 4.33.