Since the cokernels of (2.29) and (2.30) have equal dimension the image of (2.29) is precisely the orthogonal complement of Hm+ in Cδk,α(Λm), i.e.
Cδk,α(Λm) = Hm+ ⊕ △
Cδk+2,α(Λm)⊕ρHm∞⊕ρtHm∞ . Since Hm+ is finite-dimensional the projections are bounded.
Now let G be a Fredholm inverse for (2.29) and forβ ∈Cδk,α(Λm) define PEβ =dd∗Gβ,
PE∗β =d∗dGβ.
PE andPE∗ map intoCδk,α[dΛm−1] andCδk,α[d∗Λm+1] respectively, andβ =PE(β) +PE∗(β) for any β ∈ △
Cδk,α(Λm)⊕ρHm∞⊕ρtHm∞
. Therefore
△
Cδk+2,α(Λm)⊕ρHm∞⊕ρtHm∞
=Cδk,α[dΛm−1]⊕Cδk,α[d∗Λm+1] with projections PE and PE∗. The result follows.
Definition 2.3.29. Let Hcpt∗ (M) be the compactly supported cohomology of M, i.e. the cohomology of the cochain complex
Ω0cpt(M)−→d Ω1cpt(M)−→d Ω2cpt(M)−→ · · ·d of compactly supported smooth forms on M.
Recall that whenM has cylindrical ends then we can compactifyM by ‘adding a copy of X at infinity’, i.e. by including it in M = M0 ∪(X×[0,∞]). The cohomology (with coefficients R) of M relative to its boundary can be identified with Hcpt∗ (M). The long exact sequence for relative cohomology of M can be written as
· · · −→Hm−1(X)−→∂ Hcptm(M)−→e Hm(M) j
∗
−→Hm(X)−→ · · · (2.32) e : Hcptm(M) →Hm(M) is induced by the inclusion Ω∗cpt(M) ֒→Ω∗(M). The image of e is the subspace of de Rham cohomology classes with compact representatives.
Definition 2.3.30. Let H0m(M) = im e:Hcptm(M)→Hm(M) .
In the asymptotically cylindrical setting the map j∗ : Hm(M) → Hm(X) can be de-scribed as follows: For s ∈ R+ let js : X ֒→ M be the inclusion x 7→ (x, s) ∈ M∞. The maps js are homotopic, so they all give the same map j∗ :Hm(M)→Hm(X).
If α is an asymptotically translation-invariant m-form let B(α) denote its asymptotic limit, which is a translation-invariant form on the cylinder X×R. B(α) can be written as Ba(α) +dt∧Be(α), where Ba(α),Be(α) are forms onXof degreemandm−1 respectively.
If α is a closed asymptotically translation-invariant m-form then for any m-cycle C in X
Z
C
js∗([α])→ Z
C
[Ba(α)]
as s→ ∞, so
j∗([α]) = [Ba(α)]. (2.33)
By proposition 2.3.24 any α ∈ Hm0 is exponentially asymptotically translation-invariant, and B(α)∈ Hm∞. By remark 2.3.22
Hm∞∼=HXm⊕dt∧ Hm−1X , so we get maps Ba :Hm0 → HmX, Be :Hm0 → Hm−1X .
Definition 2.3.31. Let
Hmabs= kerBe⊆ Hm0 , Hrelm = kerBa ⊆ Hm0 .
Further let HmE,HEm∗ ⊆ Hm0 denote the spaces of bounded exact and coexact harmonic forms respectively.
It follows immediately from (2.33) that HEm ⊆ Hmrel. It follows by applying the Hodge star that HmE∗ ⊆ Hmabs (using remark 2.3.28 ifM is not orientable).
Let Ω∗−δ(M) be the cochain complex of forms β such that e−δtβ is uniformly bounded with all derivatives, and let H−δ∗ (M) be its cohomology.
Lemma 2.3.32. For δ >0 the natural map H−δ∗ (M)→H∗(M) is an isomorphism.
Proof. The inclusion i: Ω∗−δ(M) ֒→ Ω∗(M) is a chain map, so induces a well-defined map i:H−δ∗ (M)→H∗(M).
Let ρ : R → [0,1] be a smooth function with ρ(t) = 0 for t ≤ 0 and ρ(t) = 1 for t ≥ 1. Define a map c : M → M which is the identity on the compact piece M0 and (x, t) 7→ (x,(1−ρ(t))t+ρ(t) arctant) on the cylindrical part. c is smooth, and c∗ maps Ω∗(M) to Ω∗−δ(M). It is a chain map, so induces a mapc∗ :H∗(M)→H−δ∗ (M). We deduce that i and c∗ are inverses from the fact that cis homotopic to the identity.
The next theorem is part of [42, Theorem 6.18].
Theorem 2.3.33. Let M be an EAC manifold. The natural map Hmabs → Hm(M) is an isomorphism.
Proof. Note that by integration by parts the elements of Hm0 are closed, so represent cohomology classes. Pick some 0< δ < ǫ1. By the previous lemma it suffices to check that Habsm →H−δm(M) is an isomorphism.
We first prove injectivity. Ifα∈ Hmabs and α =dβ for some β ∈Ωm−δ(M) thenα ∈ Hm+, since HmE ⊆ Hmrel and Hmabs∩ Hmrel=Hm+. As α is exponentially decaying with rateδ we can integrate by parts to deduce α = 0. ThusHabsm →H−δm(M) is an injection.
It remains to show that if γ ∈Ωm−δ(M) is closed then γ is cohomologous to an element of Habsm . C−δ0,α(Λm) =Hm+ ⊕ △C−δ2,α(Λm) by proposition 2.3.16. Thusγ can be written as
γ =φ+△ψ
with φ ∈ Hm+ and ψ ∈C−δ2,α(Λm). Let χ=dψ. Then
dγ = 0⇒dd∗χ= 0 ⇒ △χ= 0.
By proposition 2.3.23,χ∈Cδ2,α(Λm+1)⊕ρHm+1∞ ⊕ρtHm+1∞ sod∗χ∈ HmE∗ ⊆ Hmabs. Therefore γ −dd∗ψ =φ+d∗χ∈ Hmabs (2.34) represents the same cohomology class as γ. Hence Hmabs →H−δm(M) is surjective.
Proposition 2.3.34. Hmabs =Hm+ ⊕ HmE∗, Hmrel =Hm+ ⊕ HmE. Furthermore HmE =dHm−1− , HEm∗ =d∗Hm+1− .
Proof. Theorem 2.3.33 implies that HmE ∩ Hm+ = 0. Similarly HmE∗∩ H+m = 0, so the sums are direct. If γ ∈ Hm0 then, since γ is closed, (2.34) gives
γ =φ+d∗χ+dd∗ψ,
whereφ∈ Hm+,d∗χ∈d∗Hm+1− . Analogously, sinceγ is also coclosed,dd∗ψ ∈dH−m−1. Hence Hm0 =Hm+ ⊕d∗H−m+1⊕dHm−1− .
Since dH−m−1 ⊆ HmE ⊆ Hmrel and d∗Hm+1− ⊆ HmE∗ ⊆ Hmabs the result follows.
As a corollary of theorem 2.3.33 we can determine that the image of the space Hm+ of L2 harmonic forms in the de Rham cohomologyHm(M) is precisely the subspaceH0m(M) of compactly supported classes. Similar results (with more general hypotheses) appears in e.g. [1, Proposition 4.9] and [38, Theorems 7.6 and 7.9].
Theorem 2.3.35. Let M be an EAC manifold. The natural map Hm+ → Hm(M) is an isomorphism onto H0m(M).
Proof. H+mis kerBainHmabs, and it follows from theorem 2.3.33 it is mapped isomorphically to H0m(M) = kerj∗ ⊆Hm(M).
The fact that the image of H+m is contained in H0m(M) could be seen more explicitly by applying the following lemma, which is useful for other purposes.
Ifα is a closed exponentially asymptotically translation-invariantm-form onM we can write it as α∞ +βt +dt∧γt on the cylindrical part X ×R+, where α∞ is translation-invariant, and βt, γt are sections of ΛmT∗X, Λm−1T∗X respectively, and are exponentially
decaying in t. Let
η(α) =ρ Z ∞
t
γsds, (2.35)
where ρ is a smooth cut-off function for the cylindrical end of M, equal to 1 for t > t0. η(α) is a well-defined exponentially decaying (m−1)-form on M.
Lemma 2.3.36. Let M be a manifold with cylindrical ends, and α a closed exponentially asymptotically translation-invariant m-form onM. Thenα+dη(α)is translation-invariant on {y ∈M :t > t0}.
Proof. Closure of α means that ∂t∂βt+dXγt = 0 where dX is the exterior derivative on X.
Thus βt =−R∞
t dXγsds. Since X is compact the dominated convergence theorem ensures that for t > t0
dη =dXη+dt∧ ∂t∂η=−βt−dt∧γt.
Definition 2.3.37. LetAm =Ba(Hm0 )⊆ HmX, Em =Be(Hm+10 )⊆ HmX, and letAm, Em be the subspaces of Hm(X) that they represent (Am is of course justj∗(Hm(M))⊆Hm(X)).
When Mn is oriented the Hodge star on M identifies Hmabs and Hm−nrel . Ifβ ∈ Hm0 then Be(∗β) =∗Ba(β), so the Hodge star onX identifiesAm with En−m−1 (and henceAm with En−m−1).
By proposition 2.3.23 anyψ ∈ H−m can be written as χ+α+tβ+dt∧γ+tdt∧δ with χ exponentially decaying, α, β ∈ HmX and γ, δ ∈ Hm−1X . Thus we can define a ‘boundary data’ map
BD :Hm− →(HmX)2⊕(Hm−1X )2, ψ 7→ β δ α γ
! .
Let BDa, BDe be the composition of BD with the projection to (HXm)2 and (HXm−1)2 respectively. Let ˜Am = BDa(Hm−) and ˜Em−1 = BDe(H−m). The following proposition is a refinement of [42, Lemma 6.15].
Proposition 2.3.38. Let M be an EAC manifold.
(i) If ψ1, ψ2 ∈ Hm− with BD(ψi) = βi δi
αi γi
!
then in the L2 inner product on X
<α1, β2>L2 =<α2, β1>L2, (2.36a)
<γ1, δ2>L2 =<γ2, δ1>L2 . (2.36b)
(ii) BD(H−m) = ˜Am ⊕E˜m, and A˜m ⊂ (HmX)2 and E˜m−1 ⊂ (Hm−1X )2 are Lagrangian subspaces.
Proof. First assume that M is oriented. Then ∂ : Hm(X) → Hcptm+1(M) is the Poincar´e dual of j∗ :Hn−m+1(M)→Hn−m(X), so
<αi, βj>L2 = ([αi],[∗βj])X = ([αi], j∗[∗dψj])X = (∂[αi],[∗dψj])M. Note that ∂[αi] = [d(ψi+η(dψi)−ρtβi)]∈Hcptm+1(M). Hence
<α1, β2>L2 −<α2, β1>L2 = Z
M
(dψ1 −d(ρtβ1))∧ ∗dψ2−(dψ2−d(ρtβ2))∧ ∗dψ1
= Z
M
d(ρt(β2∧ ∗dψ1−β1∧ ∗dψ2)). (2.37) Since
β2∧B(∗dψ1)−β1∧B(∗dψ2) = β2∧ ∗β1−β1 ∧ ∗β2 = 0
the integrand in the RHS of (2.37) is the exterior derivative of an exponentially decaying form. The vanishing of the integral proves (2.36a), and (2.36b) follows by applying ∗.
This proves (i) in the oriented case. WhenM is not orientable we use remark 2.3.28.
(i) implies that ˜Am ⊂(HmX)2 and ˜Em−1 ⊂(Hm−1X )2 are null spaces. In particular dim ˜Am ≤bm(X), dim ˜Em−1 ≤bm−1(X).
Since BD(Hm−)⊆A˜m⊕E˜m−1 and has dimensionbm−1(X) +bm(X) equality must hold, so A˜m and ˜Em−1 are Lagrangian.
Proposition 2.3.39. If Mn is an EAC manifold with cross-section X then HmX =Am⊕ Em
is an orthogonal direct sum.
Proof. Ifψ ∈ Hm− with BDa(ψ) = β1
α1
!
∈A˜m thenβ1 =Be(dψ)∈ Em. Thus the second projection ˜Am → HmX has image Em, and kernel Am, so
dimAm+ dimEm = dim ˜Am = dimHmX
Furthermore if α2 ∈ Am then 0 α2
!
∈ A˜m, and (2.36b) implies that <β1, α2> = 0, so Am and Em are orthogonal.
It follows from proposition 2.3.39 that we can define an isomorphism Hmrel →Hcptm(M), α7→
( [α+dη(α)] for α∈ Hm+
∂([Be(α)]) for α∈ HmE .
Corollary 2.3.40. Let Mn be an asymptotically cylindrical manifold which has a single end (i.e. the cross-section X is connected). Then e:Hcpt1 (M)→H1(M) is injective.
In particular H1E = 0, and H10 →H1(M) is an isomorphism.
Proof. Consider the start of the long exact sequence for relative cohomology Hcpt0 (M)→H0(M)→H0(X)→∂ Hcpt1 (M)→e H1(M).
The dimensions of the first three terms are 0,1, and 1, so ∂ = 0, and thus e is injective.
Hrel1 ↔Hcpt1 (M) identifies H1E with kere, so the result follows.
Finally we make two simple observations for the case whenM is Ricci-flat, as it is when the holonomy is G2 orSpin(7).
Proposition 2.3.41. If M is a Ricci-flat EAC manifold then H10 is the space of parallel 1-forms on M. In particular H1+= 0, and j∗ :H1(M)→H1(X) is injective.
Proof. This is proved by the same standard ‘Bochner argument’ as corollary 2.1.9, adapted to the EAC setting. If φ is a 1-form then by proposition 2.1.8
△φ =∇∗∇φ. (2.38)
It follows that any parallel 1-form φis harmonic, and parallel forms are of course bounded.
To show that any bounded harmonic form is parallel we use (2.38) together with an integration by parts argument, justified by lemma 2.3.15.
H1+ = 0 since it consists of parallel decaying forms. By theorem 2.3.35, the kernel H01(M) of j∗ :H1(M)→H1(X) is represented by H+1.
Corollary 2.3.42. If M is a Ricci-flat EAC manifold with a single end thenHcpt1 (M) = 0.
Proof. Follows from proposition 2.3.41 and corollary 2.3.40.