Duality maps for constructible inverse and direct images

Dualizing a sheaf twice comes with a map from the original sheaf. Whether that map is invertible determines which duality comparisons can be reversed. We will construct the comparison maps, factor the reversed maps through evaluation, and then identify the geometric hypotheses that make those evaluations invertible. An infinite locally constant coefficient requires a different argument: move it through an internal Hom, rather than evaluate it twice.

Scope. The operation, orientation and evaluation arguments are proved below. The constructible applications additionally require the geometric results identified in the prerequisite section; their full proof chains are not claimed here.

AI-written exposition and exercises: GPT-6 Astra (OpenAI), Ultra. The modern weak-coefficient results are due to Andreas Hohl and Pierre Schapira and are credited at their point of use. Original programme expression is dedicated to the public domain under CC0.

Coefficients, pairings and the operation contracts

Let kk be a commutative ring of finite global dimension. Spaces in this lesson are finite-dimensional real analytic manifolds, Hausdorff and countable at infinity, with a uniform dimension bound. Complexes are globally bounded. All tensors and internal Homs are derived. We use

ωX=aX!k,DXE=RℋomX(E,ωX),Hj(E[s])=Hj+s(E).(1) \omega_X=a_X^!k,\qquad D_XE=R\mathcal Hom_X(E,\omega_X), \qquad H^j(E[s])=H^{j+s}(E). \qquad\text{(1)}

For a dd-manifold, ωX=orX[d]\omega_X=\mathrm{or}_X[d]. Keep the orientation line in this formula unless an orientation has been chosen.

We first work formally with a map f:Y→Xf:Y\to X. The required operation contracts are ordinary adjunction f−1⊣Rf*f^{-1}\dashv Rf_*, proper-support adjunction Rf!⊣f!Rf_!\dashv f^!, tensor–Hom adjunction, the projection isomorphism

Rf!(B⊗f−1A)≃Rf!B⊗A,(2) Rf_!(B\otimes f^{-1}A)\simeq Rf_!B\otimes A, \qquad\text{(2)}

and composition f!ωX≃ωYf^!\omega_X\simeq\omega_Y, with its compatible trace Rf!ωY→ωXRf_!\omega_Y\to\omega_X. These contracts include their units, counits, naturality and boundedness. The sections below prove compact-support extension and lifting, the proper-image fibre formulas, base change and composition, the uniform manifold dimension bound, and arbitrary-coefficient projection. The adjunction section then constructs both adjunctions on bounded-below categories and fixes their trace-compatible exceptional composition. The following bounds section proves globally bounded preservation for ordinary direct image, exceptional inverse image and duality. The tensor–Hom section then constructs the remaining algebraic contract, including the actual evaluation map and its signs. The orientation section proves the constant local-support and orientation inputs. Constructible and weakly constructible geometric inputs remain prerequisites. In particular, (2) permits arbitrary bounded AA; it is not a projection theorem restricted to perfect coefficients. The formal arguments also apply outside the analytic setting whenever these same bounded operation contracts hold.

Evaluation is the pairing ev⁡E:DXE⊗E→ωX\operatorname{ev}_E:D_XE\otimes E\to\omega_X. By symmetry and currying it gives

ηE:E⟶DXDXE.(3) \eta_E:E\longrightarrow D_XD_XE. \qquad\text{(3)}

We call EE reflexive for this duality when this particular map is invertible. An abstract isomorphism between EE and DXDXED_XD_XE would not be sufficient.

All maps below use the usual symmetry on complexes. On homogeneous tensors that symmetry is v⊗w↦(−1)|v||w|w⊗vv\otimes w\mapsto(-1)^{|v||w|}w\otimes v. Currying and evaluation use the same convention, so reversing a pairing introduces no unrecorded sign.

Compact support: extension, lifting and acyclicity

The projection proof needs precise control of support when a section is extended or lifted. We supply that part of its foundation here. Throughout this section, spaces are locally compact Hausdorff and coefficients are arbitrary modules over the fixed commutative ring kk. Finite global dimension is unnecessary for these support arguments. For a compact subset K⊂XK\subset X, Γ(K;F)\Gamma(K;F) means sections of the inverse-image sheaf on KK, not sections on XX whose support lies in KK.

The human source for these results is Pierre Schapira’s freely accessible An Introduction to Sheaves on Grothendieck Topologies, 1 August 2026, §4.3. His word “soft” in that section means what we call c-soft: every section on every compact subset extends to a global section. We organise the proof around two concrete operations—extending with prescribed support and lifting through a quotient—and include the compact-neighbourhood details they require. The following sections supply the proper-image fibre formula and the uniform manifold dimension bound.

Compact subsets remember neighbourhood germs

We first justify the comparison

colim⁡K⊂V openΓ(V;F)→∼Γ(K;F).(C1) \mathop{\mathrm{colim}}_{K\subset V\text{ open}}\Gamma(V;F) \xrightarrow{\sim}\Gamma(K;F). \qquad\text{(C1)}

The elementary topological facts needed below follow from compactness and Hausdorff separation. Disjoint compact sets have disjoint open neighbourhoods: separate each pair of points, take a finite intersection on the neighbourhood of a fixed point and a finite union on the opposite compact set, then repeat over a finite cover of the first compact set. Thus a compact Hausdorff space is normal. Inside a compact neighbourhood, this separation shrinks a neighbourhood of a point so that its closure is compact and lies in any specified open neighbourhood. Applying these pointwise shrinkings and then taking a finite subcover gives compact sets subordinate to any finite open cover of a compact set.

Surjectivity of (C1) is not merely the definition of inverse image, since that definition also sheafifies. A section on KK has local representatives sis_i on ambient open sets NiN_i. Choose finitely many compact sets Ki⊂K∩NiK_i\subset K\cap N_i covering KK, and relatively compact open neighbourhoods OiO_i of KiK_i with Oi¯⊂Ni\overline{O_i}\subset N_i. For each pair i,ji,j, let EijE_{ij} be the open subset of Ni∩NjN_i\cap N_j where the germs of si,sjs_i,s_j agree. It contains Ki∩KjK_i\cap K_j. The compact set Dij=(Oi¯∩Oj¯)\EijD_{ij}=(\overline{O_i}\cap\overline{O_j})\setminus E_{ij} therefore meets KiK_i and KjK_j in disjoint closed subsets. Separate those two subsets by disjoint relatively open subsets of DijD_{ij}. Their complements in DijD_{ij} are compact, so taking their complements in XX gives open neighbourhoods of Ki,KjK_i,K_j whose intersection misses DijD_{ij}. Intersect these neighbourhoods over the finitely many pairs, also intersecting with OiO_i. The resulting opens WiW_i still contain KiK_i, and the representatives agree on every Wi∩WjW_i\cap W_j. The sheaf axiom glues them to a section on ⋃iWi\bigcup_iW_i, a neighbourhood of KK. This proves surjectivity.

For injectivity, two representatives with equal restriction to KK have equal germs at every point of KK. Their equality locus is open and contains KK. They are therefore equal after shrinking their common neighbourhood, exactly the equivalence relation in the colimit.

We also use gluing over a finite closed cover of a compact space. If sections on closed sets K1,…,KmK_1,\ldots,K_m agree on their intersections, their germs give a section on their union. To check this locally at a point, discard the finitely many closed sets not containing that point. Representatives from the remaining sets have the same germ there, so they agree on a common smaller neighbourhood. This supplies a local representative of the proposed glued section; these representatives agree wherever both are defined. Hence the sheaf axiom proves existence and uniqueness of the glued section. This argument concerns sections of restricted sheaves, so equality on an intersection includes equality of the corresponding ambient germs.

Extending a section while controlling its support

Support extension lemma. Suppose FF is c-soft, K⊂XK\subset X is compact, and s∈Γ(K;F)s\in\Gamma(K;F). If the support of ss in KK lies in an open set V⊂XV\subset X, there is t∈Γc(X;F)t\in\Gamma_c(X;F) such that

t|K=s,supp⁡(t)⊂V.(C2) t|_K=s,\qquad \operatorname{supp}(t)\subset V. \qquad\text{(C2)}

Proof. The support of a section is closed in its domain because its zero-germ locus is open. Thus A=supp⁡(s)A=\operatorname{supp}(s) is compact. Choose a compact neighbourhood LL of AA contained in VV; if AA is empty, take t=0t=0. On the compact set (K∩L)∪∂L(K\cap L)\cup\partial L, prescribe ss on K∩LK\cap L and zero on ∂L\partial L. These prescriptions agree on the intersection because A⊂int⁡LA\subset\operatorname{int}L. Finite closed gluing supplies a section there. By c-softness it extends to a global section uu of FF.

The germs of uu vanish on ∂L\partial L, so uu vanishes on an open neighbourhood of that boundary. Glue uu on int⁡L\operatorname{int}L to zero on the union of that neighbourhood with X\LX\setminus L. These opens cover XX, and the prescribed sections agree on the overlap. The resulting tt has support in L⊂VL\subset V. On K∩LK\cap L it agrees with ss; on K\LK\setminus L both vanish. This proves (C2). ▫\square

Several stability properties now have direct proofs.

Finally, injective sheaves are c-soft. For opens U⊂VU\subset V, the map kU→kVk_U\to k_V is a monomorphism, as its stalks show. The identification Hom⁡(kU,I)=Γ(U;I)\operatorname{Hom}(k_U,I)=\Gamma(U;I) and injectivity of II make every restriction Γ(V;I)→Γ(U;I)\Gamma(V;I)\to\Gamma(U;I) surjective. Thus II is flabby. A section on a compact set is represented on a neighbourhood by (C1), and flabbiness extends that representative to XX. The same argument proves directly that every flabby sheaf is c-soft.

Lifting compactly supported sections

Lifting lemma. If 0→F′→F→F”→00\to F'\to F\to F\text{”}\to0 is exact and F′F' is c-soft, then

0⟶Γc(X;F′)⟶Γc(X;F)⟶Γc(X;F”)⟶0(C3) 0\longrightarrow\Gamma_c(X;F')\longrightarrow\Gamma_c(X;F) \longrightarrow\Gamma_c(X;F\text{”})\longrightarrow0 \qquad\text{(C3)}

is exact.

Proof. Left exactness follows from kernels and the definition of support. We prove surjectivity first when XX is compact. A section s”s\text{”} has lifts on an open cover because the sheaf map F→F”F\to F\text{”} is surjective on stalks. Choose a finite compact closed cover KiK_i subordinate to that cover, and use the corresponding local lifts on KiK_i.

Suppose lifts have already been glued on K1∪⋯∪Ki−1K_1\cup\cdots\cup K_{i-1}. On its intersection with KiK_i, the difference between the existing lift and the new lift is a section of F′F': restriction is exact and sections preserve kernels. The intersection is compact, so c-softness extends that difference to a global section of F′F'. Add this extension to the new lift. The adjusted lifts agree on the intersection, and finite closed gluing combines them. Finite induction gives a lift on all of XX.

For general XX, let s”s\text{”} have compact support AA, and choose a relatively compact open UU containing AA. Write FU=j!(F|U)F_U=j_!(F|_U), and similarly for the other two sheaves. Open extension by zero is exact by its stalk formula, so these sheaves give a short exact sequence. Its first term is c-soft by the restriction and extension properties just proved. Each term has zero stalks outside the compact set U¯\overline U.

A sheaf supported on a closed set is the direct image of its restriction to that set: the natural map to that direct image is an isomorphism on every stalk. Restrict our sequence to U¯\overline U, apply the compact case there, and push the resulting section back to XX. All its support is compact. The section s”s\text{”} belongs to Γc(X;F”U)\Gamma_c(X;F\text{”}_U), since it vanishes near every point outside its compact support A⊂UA\subset U. The compact-case lift is therefore a compactly supported lift of the original s”s\text{”}. This proves (C3). ▫\square

Quotient consequence. In a short exact sequence, if F′F' and FF are c-soft, then so is F”F\text{”}. Given a section of F”F\text{”} on a compact KK, restrict the sequence to KK. Its first term is c-soft there, so the compact case of (C3) lifts that section to Γ(K;F)\Gamma(K;F). C-softness of FF extends the lift globally. Its image is the desired extension of the original section.

The compact-cohomology criterion

We now prove, rather than assume, the equivalence

F is c-soft⟺Hcq(U;F|U)=0for every open U⊂X and q>0.(C4) F\text{ is c-soft}\quad\Longleftrightarrow\quad H_c^q(U;F|_U)=0 \quad\text{for every open }U\subset X\text{ and }q>0. \qquad\text{(C4)}

Use the usual injective definition of the right derived functor of Γc\Gamma_c. If FF is c-soft, embed it into an injective and continue to an injective resolution. Every successive cokernel is c-soft by the quotient consequence. The lifting lemma makes each of the resulting short exact sequences exact after Γc\Gamma_c. Hence the complex of compactly supported sections of the injective resolution is exact in positive degrees. This proves Γc\Gamma_c-acyclicity. Apply the same argument to the c-soft restriction on each open UU to obtain the forward implication in (C4).

We will also use the following derived extension identity, including for a sheaf GG which is not c-soft:

Hcq(X;j!G)=Hcq(U;G),j:U↪X.(C5) H_c^q(X;j_!G)=H_c^q(U;G),\qquad j:U\hookrightarrow X. \qquad\text{(C5)}

Take an injective resolution of GG on UU. Its terms are c-soft, so their open extensions are c-soft on XX and therefore Γc\Gamma_c-acyclic by the preceding paragraph. Exactness of j!j_! makes this an acyclic resolution of j!Gj_!G. Termwise, compactly supported sections on UU and on XX are identical under extension by zero: a compact support in UU is compact in XX, and a compact support of a section of j!Gj_!G lies inside UU. The acyclic-resolution comparison, applied to these degreewise identical complexes, proves (C5).

Here is the comparison needed in that last step. For an acyclic resolution 0→M→A0→A1→⋯0\to M\to A^0\to A^1\to\cdots, put Z1=A0/MZ^1=A^0/M. The first derived-functor exact sequence identifies R1T(M)R^1T(M) with coker⁡(T(A0)→T(Z1))\operatorname{coker}(T(A^0)\to T(Z^1)), which is H1(T(A•))H^1(T(A^\bullet)) because TT preserves kernels. For q≥2q\geq2, the same exact sequence gives RqT(M)=Rq−1T(Z1)R^qT(M)=R^{q-1}T(Z^1). Repeat using the tail resolution of Z1Z^1; after finitely many repetitions this is the same degree-one calculation. Degree zero is left exactness. The connecting maps come from the given short exact sequences, so the identifications are natural. Apply this with T=ΓcT=\Gamma_c.

For the reverse implication in (C4), fix a compact K⊂XK\subset X, put U=X\KU=X\setminus K, and let i:K↪Xi:K\hookrightarrow X, j:U↪Xj:U\hookrightarrow X. Stalks give an exact sequence

0⟶j!(F|U)⟶F⟶i*(F|K)⟶0.(C6) 0\longrightarrow j_!(F|_U)\longrightarrow F \longrightarrow i_*(F|_K)\longrightarrow0. \qquad\text{(C6)}

The long exact sequence for compactly supported cohomology and (C5) show that Γc(X;F)→Γc(X;i*(F|K))\Gamma_c(X;F)\to\Gamma_c(X;i_*(F|_K)) is surjective, because its obstruction lies in Hc1(U;F|U)=0H_c^1(U;F|_U)=0. The group on the right is Γ(K;F)\Gamma(K;F), since KK is compact. Every section on KK thus has a compactly supported global extension, proving c-softness. This completes (C4).

What this proves for proper direct image

The support arguments above also supply proper-image acyclicity from the underived proper-image fibre formula, proved in the next section:

(f!Q)x≃Γc(f−1(x);Q|f−1(x)).(C7) (f_!Q)_x\simeq\Gamma_c(f^{-1}(x);Q|_{f^{-1}(x)}). \qquad\text{(C7)}

Indeed, restrict an injective resolution of a c-soft QQ to the closed fibre. Restriction is exact, all its terms are c-soft there, and all the successive cokernels are c-soft. The lifting lemma makes its compact-section complex exact in positive degrees. Applying (C7) termwise identifies that complex with the stalk of f!I•f_!I^\bullet. Thus Rqf!Q=0R^qf_!Q=0 for q>0q>0. This proves the acyclicity implication without assuming the derived fibre formula as an extra theorem.

The next section proves (C7) and the coproduct and open-extension compatibilities of f!f_!, thereby establishing the acyclicity conclusion. The following composition and dimension section proves the uniform compact-cohomology bound. The subsequent bounded c-soft construction shows exactly how it supplies finite cohomological dimension and arbitrary-coefficient projection.

Proper support and the fibre calculation

We now supply the proper-image facts used in (C7) and in the projection argument. Let f:Y→Xf:Y\to X be continuous between locally compact Hausdorff spaces. The results in this section hold over an arbitrary fixed commutative coefficient ring kk; finite generation and finite global dimension are unnecessary. The source for the construction is Pierre Schapira’s freely accessible An Introduction to Sheaves on Grothendieck Topologies, 1 August 2026, §§4.1–4.2. We use compactly supported pieces to make the stalk calculation explicit, including its agreement with the ordinary proper-support definition.

A proper map shrinks neighbourhoods of its fibre

Suppose first that p:Z→Xp:Z\to X is proper, meaning that inverse images of compact subsets are compact. Such a map is closed. Indeed, if A⊂ZA\subset Z is closed and x∉p(A)x\notin p(A), choose a compact neighbourhood NN of xx. The set A∩p−1NA\cap p^{-1}N is compact, so its image is compact and hence closed in XX. Removing that image from int⁡N\operatorname{int}N gives a neighbourhood of xx disjoint from p(A)p(A).

Let WW be any open neighbourhood of the compact fibre Zx=p−1(x)Z_x=p^{-1}(x). Closedness of pp shows that V=X\p(Z\W)V=X\setminus p(Z\setminus W) is an open neighbourhood of xx, and p−1V⊂Wp^{-1}V\subset W. Consequently the sets p−1Vp^{-1}V, for VV ranging over neighbourhoods of xx, are cofinal among neighbourhoods of ZxZ_x. The compact-germ comparison (C1) gives

(p*H)x=colim⁡x∈VΓ(p−1V;H)≃Γ(Zx;H|Zx).(F1) (p_*H)_x =\mathop{\mathrm{colim}}_{x\in V}\Gamma(p^{-1}V;H) \simeq\Gamma(Z_x;H|_{Z_x}). \qquad\text{(F1)}

This is the restriction map on germs, and is natural in HH. If the fibre is empty, the same closedness argument supplies a neighbourhood with empty inverse image, so both sides are zero. No surjectivity or local triviality of pp was assumed.

The same formula applies when pp is proper only on a closed set supporting HH: write HH as the direct image of its restriction to that closed set and apply the proper case there. In particular it applies to every sheaf supported on a fixed compact set.

Constructing the proper-support subsheaf

For an open U⊂YU\subset Y, let QUQ_U denote the restriction of QQ to UU, extended by zero. Each QU→QQ_U\to Q is monic on stalks. Relatively compact open subsets of YY form a directed family under finite unions. Define the sheaf

f!Q=colim⁡U⋐Yf*(QU),U⋐Y means U¯ is compact.(F2) f_!Q=\mathop{\mathrm{colim}}_{U\Subset Y}f_*(Q_U), \qquad U\Subset Y\text{ means }\overline U\text{ is compact}. \qquad\text{(F2)}

The colimit here is a sheaf colimit. The inclusions into f*Qf_*Q identify it with a subsheaf: taking stalks gives a directed union of submodules. Thus a section of f!Qf_!Q over an open V⊂XV\subset X is locally represented by sections of QUQ_U for relatively compact UU. We prove that this is exactly the familiar condition that its support be proper over VV.

Take such a section, viewed as s∈Γ(f−1V;Q)s\in\Gamma(f^{-1}V;Q), and write S=supp⁡(s)S=\operatorname{supp}(s), a closed subset of f−1Vf^{-1}V. Locally on VV, the representative through QUQ_U implies S⊂U¯S\subset\overline U. Over a compact subset CC of that local neighbourhood, the set S∩f−1CS\cap f^{-1}C is closed in YY and contained in the compact set U¯\overline U; it is therefore compact. For a general compact C⊂VC\subset V, choose finitely many compact subsets covering CC, each subordinate to one of those local neighbourhoods. This finite shrinking is available by the compact Hausdorff argument preceding (C1). Their inverse images in SS are compact and cover the inverse image of CC. Thus S→VS\to V is proper.

Conversely, suppose S→VS\to V is proper. For x∈Vx\in V, choose a compact neighbourhood N⊂VN\subset V of xx. Then S∩f−1NS\cap f^{-1}N is compact, so it lies in some relatively compact open U⊂YU\subset Y. On f−1(int⁡N)f^{-1}(\operatorname{int}N), the section has zero germs outside UU. The stalk description of the subsheaf QU⊂QQ_U\subset Q shows that it factors through that subsheaf there. Hence ss locally belongs to a term in (F2), so it is a section of f!Qf_!Q. This proves the claimed identification with proper support.

In particular, when ff itself is proper, every section of f*Qf_*Q satisfies this condition: a closed support over an open of the target remains proper over that open. The natural inclusion is then an isomorphism f!Q=f*Qf_!Q=f_*Q.

The underived fibre formula

Fix x∈Xx\in X and put Yx=f−1(x)Y_x=f^{-1}(x). For every relatively compact open U⊂YU\subset Y, the sheaf QUQ_U is supported on the compact set U¯\overline U. Applying (F1) to that support gives

(f!Q)x≃colim⁡U⋐YΓ(Yx;(QU)|Yx)≃Γc(Yx;Q|Yx).(F3) (f_!Q)_x \simeq\mathop{\mathrm{colim}}_{U\Subset Y} \Gamma(Y_x;(Q_U)|_{Y_x}) \simeq\Gamma_c(Y_x;Q|_{Y_x}). \qquad\text{(F3)}

Here are the details of the last isomorphism. By the stalk formula for open extension, (QU)|Yx(Q_U)|_{Y_x} is Q|YxQ|_{Y_x} restricted to U∩YxU\cap Y_x and extended by zero within the fibre. A global section of this sheaf has support closed in YxY_x and contained in U¯∩Yx\overline U\cap Y_x, a compact set. It therefore gives a compactly supported section of Q|YxQ|_{Y_x}. All transition maps are the inclusions of these sections into the same section module.

Conversely, if tt has compact support in YxY_x, that support is compact in YY. Choose a relatively compact open UU containing it. The section has zero germs outside U∩YxU\cap Y_x, so it belongs to Γ(Yx;(QU)|Yx)\Gamma(Y_x;(Q_U)|_{Y_x}). The directed union therefore contains exactly all compactly supported sections. This proves (F3), including injectivity, and proves the previously stated formula (C7). No commutation of unrestricted sections with a colimit was used.

Restriction to the fibre is exact, and compactly supported sections preserve kernels: a section in a kernel has the same support whether regarded in the kernel sheaf or in the containing sheaf. Testing on stalks in (F3) consequently proves that f!f_! is left exact. It also confirms that, for the map to a point, (F2) is exactly Γc\Gamma_c.

Coproducts and open extensions

Compactly supported sections commute with arbitrary coproducts. A section with compact support KK locally uses only finitely many summands; a finite cover of KK gives one finite set of indices which suffices everywhere on KK. Outside KK all its germs are zero. Stalkwise factorization therefore puts the whole section in that finite subsum, and its finitely many components have compact support. Conversely a finite family of compactly supported sections gives such a section of the coproduct. These inverse constructions prove the claim. Inverse image to a fibre preserves coproducts, so (F3) shows that the canonical map

⨁λf!Qλ→∼f!(⨁λQλ)(F4) \bigoplus_\lambda f_!Q_\lambda \xrightarrow{\sim}f_!\left(\bigoplus_\lambda Q_\lambda\right) \qquad\text{(F4)}

is an isomorphism on every stalk. This does not assert that arbitrary sections commute with coproducts.

Let V⊂XV\subset X be open, U=f−1VU=f^{-1}V, and fV:U→Vf_V:U\to V the restricted map. The proper-support description of sections immediately gives (f!Q)|V=(fV)!(Q|U)(f_!Q)|_V=(f_V)_!(Q|_U): over every open subset of VV, the section modules and proper-support conditions are identical.

There is also the extension comparison

(f!Q)V→∼f!(QU),U=f−1V.(F5) (f_!Q)_V\xrightarrow{\sim}f_!(Q_U), \qquad U=f^{-1}V. \qquad\text{(F5)}

Both sides are subsheaves of f!Qf_!Q. The right side is a subsheaf by left exactness. On a stalk inside VV, formula (F3) identifies each inclusion with the identity of Γc(Yx;Q|Yx)\Gamma_c(Y_x;Q|_{Y_x}). On a stalk outside VV, the fibre misses UU, so both sides are zero. Their images in f!Qf_!Q are thus the same subsheaf, proving (F5) canonically.

The stalk identifications Q⊗kU=QUQ\otimes k_U=Q_U and f!Q⊗kV=(f!Q)Vf_!Q\otimes k_V=(f_!Q)_V turn (F5) into the open-generator projection map (P7). It is the map which multiplies a properly supported section by the pulled-back coefficient section. Hence the isomorphism just proved is the specific map used in the projection proof, not merely an unspecified isomorphism of its endpoints.

The derived fibre formula and c-soft acyclicity

Combining (F3) with the compact-support proofs gives, for B∈D+(kY)B\in D^+(k_Y), the natural identity

(Rf!B)x≃RΓc(Yx;B|Yx).(F6) (Rf_!B)_x\simeq R\Gamma_c(Y_x;B|_{Y_x}). \qquad\text{(F6)}

To prove it, choose a bounded-below injective representative II of BB. Stalks are exact, so (Rf!B)x(Rf_!B)_x is represented by (f!I)x(f_!I)_x, which (F3) identifies termwise with Γc(Yx;I|Yx)\Gamma_c(Y_x;I|_{Y_x}). Restriction is exact, hence I|YxI|_{Y_x} represents B|YxB|_{Y_x}. Its terms are c-soft, by injective c-softness and restriction stability, and thus are Γc\Gamma_c-acyclic.

For clarity, a bounded-below complex of acyclic terms computes a right derived functor. In any fixed degree qq, cut the complex off brutally above degree q+1q+1. The kernel starts in degree q+2q+2; applying the termwise functor or its right derived functor gives no cohomology below that degree. For the derived statement, use a bounded-below injective replacement starting in the same degree, which exists by the usual degreewise resolution construction. The cutoff therefore changes neither computation in degree qq. The cutoff complex is bounded, and the finite acyclic-complex argument proved after (P4) applies to it. These natural comparisons prove the assertion in each degree. This argument does not use an unbounded totalization or a uniform cohomological-dimension bound.

Apply that comparison to I|YxI|_{Y_x} to prove (F6). The maps arise from restriction and resolution comparisons, so the identity is natural in BB. For an ordinary sheaf QQ whose restriction to every fibre is c-soft, (C4) and (F6) imply Rqf!Q=0R^qf_!Q=0 for q>0q>0. In particular every c-soft sheaf on YY is f!f_!-acyclic. Thus the conditional acyclicity argument after (C7) is now established without an imported fibre theorem.

No bounded-output assertion follows merely from (F6). The next section supplies a uniform compact-cohomology bound and a finite c-soft resolution on the standing manifolds, thereby keeping Rf!Rf_! in a globally bounded range. The later adjunction section constructs the bounded-below exceptional adjoint. The subsequent bounds section proves its globally bounded output estimate. The geometric duality inputs remain separate obligations.

Composition and the uniform dimension bound

The fibre calculation alone does not bound cohomology. We now prove the missing bound, using proper-image composition to reduce Euclidean space one coordinate at a time to the interval. We then turn the local bounds into finite c-soft resolutions on manifolds. These arguments work for arbitrary sheaves over the fixed commutative ring kk; neither finite generation nor finite global dimension is needed in this section.

The human source is Schapira’s free An Introduction to Sheaves on Grothendieck Topologies, 1 August 2026: §4.2 for proper-support composition, §§4.3 and 4.5 for c-soft images and base change, Lemma 3.5.1 for the interval, and Lemma 5.1.1 and Proposition 5.1.2 for the dimension bound. We supply the compact-neighbourhood continuity, closed-interval gluing and dimension-shifting details used in the proof. The dimension argument does not use the projection formula it is about to justify.

Pulling back a proper support

Consider a Cartesian square with f:Y→Xf:Y\to X, g:X′→Xg:X'\to X, Y′=Y×XX′Y'=Y\times_X X', and projections f′:Y′→X′f':Y'\to X', g′:Y′→Yg':Y'\to Y. These spaces are locally compact Hausdorff: the fibre product is closed in Y×X′Y\times X', since the diagonal of XX is closed. Pulling back sections defines

g−1f!Q→∼f!′g′−1Q.(D1) g^{-1}f_!Q\xrightarrow{\sim}f'_!g'^{-1}Q. \qquad\text{(D1)}

We verify both the map and its invertibility. A section of f!Qf_!Q over an open V⊂XV\subset X is a section of QQ on f−1Vf^{-1}V with a support SS proper over VV. Its pullback has support g′−1Sg'^{-1}S. For a compact C⊂g−1VC\subset g^{-1}V, the part of that support over CC is a closed subset of (S∩f−1g(C))×C(S\cap f^{-1}g(C))\times C. Both factors are compact, and the equality f(y)=g(x′)f(y)=g(x') is closed. Thus the pulled-back support is proper. Pullback of local sections, followed by sheafification of the inverse-image presheaf, therefore gives the displayed map.

At a point x′∈X′x'\in X', (F3) identifies its two stalks with compactly supported sections on Yg(x′)Y_{g(x')} and on Yx′′Y'_{x'}. The projection g′g' identifies these fibres homeomorphically, with the same restricted sheaf, and the map is the resulting pullback of sections. It is an isomorphism. Stalkwise detection proves (D1). This proof also shows that the comparison respects restriction and consecutive changes of base: in each case it pulls back the same section.

Composing proper images and preserving c-softness

Let f:Y→Xf:Y\to X and g:X→Zg:X\to Z. Then there is a natural identification

g!f!Q≃(gf)!Q.(D2) g_!f_!Q\simeq(gf)_!Q. \qquad\text{(D2)}

To check it on an open V⊂ZV\subset Z, compare both sides as subsheaves of (gf)*Q(gf)_*Q. For a section ss of QQ on (gf)−1V(gf)^{-1}V, put S=supp⁡(s)S=\operatorname{supp}(s). If S→VS\to V is proper, then S→g−1VS\to g^{-1}V through ff is proper: the inverse image of a compact C⊂g−1VC\subset g^{-1}V is closed in the compact set S∩(gf)−1g(C)S\cap(gf)^{-1}g(C). The corresponding section tt of f!Qf_!Q has support exactly T=f(S)T=f(S). Indeed, (F3) says that its stalk at xx is the restricted section on the fibre, which is nonzero exactly when that fibre meets SS. The set TT is closed in g−1Vg^{-1}V, since f|Sf|_S is proper and hence closed. Its inverse image under gg of a compact C⊂VC\subset V is the compact image f(S∩(gf)−1C)f(S\cap(gf)^{-1}C). Thus tt has proper support over VV.

Conversely, if ss represents a section of g!f!Qg_!f_!Q, then S→TS\to T and T→VT\to V are proper by the same support description. The inverse image in SS of a compact subset of VV is compact by two successive properness applications. Hence S→VS\to V is proper. Both sheaves therefore consist of the same sections with the same restriction maps. This proves (D2), and also proves its identity and threefold-associativity compatibilities: every comparison retains the underlying section of QQ.

In particular, taking the last map to a point gives Γc(X;f!Q)=Γc(Y;Q)\Gamma_c(X;f_!Q)=\Gamma_c(Y;Q). We use this to prove that f!f_! preserves c-soft sheaves.

First, a c-soft sheaf QQ has surjective restriction Γc(Y;Q)→Γc(A;Q|A)\Gamma_c(Y;Q)\to\Gamma_c(A;Q|_A) for every closed A⊂YA\subset Y. Given a compactly supported section on AA, choose a compact neighbourhood LL of its support. Prescribe that section on A∩LA\cap L and zero on ∂L\partial L. The prescriptions agree, since the support lies in int⁡L\operatorname{int}L. Both sets are compact. The compact extension and cutoff argument proving (C2) gives a global compactly supported section with the required restriction to all of AA; off LL, both restrictions vanish.

Now let K⊂XK\subset X be compact. Base change (D1) for the closed inclusion of KK, followed by (D2) for the map from KK to a point, identifies Γ(K;f!Q)\Gamma(K;f_!Q) with Γc(f−1K;Q|f−1K)\Gamma_c(f^{-1}K;Q|_{f^{-1}K}). The restriction map from Γc(X;f!Q)=Γc(Y;Q)\Gamma_c(X;f_!Q)=\Gamma_c(Y;Q) becomes restriction to the closed set f−1Kf^{-1}K, which is surjective by the preceding paragraph. This proves c-softness of f!Qf_!Q.

Choose a bounded-below injective representative II of B∈D+(kY)B\in D^+(k_Y). Its terms are c-soft, so f!If_!I has c-soft terms and computes Rf!BRf_!B. Those terms are also g!g_!-acyclic by (F6). The bounded-below acyclic-complex comparison proved there therefore gives

Rg!Rf!B≃R(gf)!B,RΓc(X;Rf!B)≃RΓc(Y;B).(D3) Rg_!Rf_!B\simeq R(gf)_!B, \qquad R\Gamma_c(X;Rf_!B)\simeq R\Gamma_c(Y;B). \qquad\text{(D3)}

The termwise maps are (D2), so they are the natural composition maps and retain identity and threefold-associativity compatibility. This proof does not need a cohomological-dimension bound.

The same method derives (D1). Each term of g′−1Ig'^{-1}I is c-soft on every fibre of f′f': the fibre is identified with a fibre of ff, on which the original injective term restricts to a c-soft sheaf. Formula (F6) makes these terms f!′f'_!-acyclic. Since inverse image is exact, applying the bounded-below comparison and (D1) termwise proves

g−1Rf!B≃Rf!′g′−1B.(D4) g^{-1}Rf_!B\simeq Rf'_!g'^{-1}B. \qquad\text{(D4)}

The maps still pull back the original local sections before passing to resolutions; their base-change compatibilities follow from that construction. No properness of ff, beyond the proper supports defining f!f_!, is required.

Cohomology near a compact set and the interval bound

Let KK be a compact subset of a locally compact Hausdorff space. There is a natural continuity isomorphism

colim⁡K⊂V openHq(V;Q|V)≃Hq(K;Q|K).(D5) \mathop{\mathrm{colim}}_{K\subset V\text{ open}}H^q(V;Q|_V) \simeq H^q(K;Q|_K). \qquad\text{(D5)}

To prove it, take an injective resolution II of QQ on the ambient space. Restriction to an open preserves injectives: its left adjoint is the exact extension by zero, so its Hom test for injectivity remains exact. Thus Γ(V;I)\Gamma(V;I) computes the left-hand groups. The restricted complex I|KI|_K is exact over Q|KQ|_K, and its terms are c-soft on the compact space KK. There Γ=Γc\Gamma=\Gamma_c, so the same terms are acyclic and Γ(K;I)\Gamma(K;I) computes the right-hand groups. Comparison (C1) identifies these complexes after taking the filtered colimit. Filtered colimits of modules are exact: each equality and each witness for a preimage is witnessed at some common later index. Consequently they commute with kernels modulo images, hence with cohomology. This proves (D5).

The same proof works with compact neighbourhoods instead of open neighbourhoods. Every neighbourhood of a compact set contains a compact neighbourhood, and every compact neighbourhood contains an open one. These interleavings give the same colimit termwise. Restrictions of II to the compact neighbourhoods remain c-soft and acyclic. In particular, closed intervals shrinking down to a closed interval or to an endpoint have this cohomology continuity.

We claim that Hq([0,1];Q)=0H^q([0,1];Q)=0 for every q>1q>1 and every sheaf QQ. For 0<t<a≤10<t<a\leq1, write [0,a]=[0,t]∪[t,a][0,a]=[0,t]\cup[t,a]. Restrict an injective resolution on [0,a][0,a] to these compact sets. Finite closed gluing gives the kernel in the sequence of section complexes, and c-softness gives surjectivity onto the section complex at {t}\{t\}: every germ there extends globally. The difference of restrictions therefore gives a termwise short exact sequence. Its long exact cohomology sequence, and exactness of sections on a point, give

Hq([0,a];Q)≃Hq([0,t];Q)⊕Hq([t,a];Q)(q>1).(D6) H^q([0,a];Q)\simeq H^q([0,t];Q)\oplus H^q([t,a];Q) \quad(q>1). \qquad\text{(D6)}

Take s∈Hq([0,1];Q)s\in H^q([0,1];Q), q>1q>1, and let JJ be the set of aa for which s|[0,a]=0s|_{[0,a]}=0. It is downward closed. The restriction to {0}\{0\} is zero; continuity at that compact set makes the restriction zero on a nontrivial initial interval. Hence τ=sup⁡J\tau=\sup J is positive. Continuity at {τ}\{\tau\}, considered inside [0,τ][0,\tau], makes the restriction zero on some terminal interval [t,τ][t,\tau] with t<τt<\tau. The initial interval [0,t][0,t] also has zero restriction, by the definition of the supremum and downward closure. Formula (D6) shows that the restriction to [0,τ][0,\tau] is zero. If τ<1\tau<1, continuity around the compact set [0,τ][0,\tau] makes it zero on a strictly larger initial interval, a contradiction. Thus τ=1\tau=1, and s=0s=0. This proves the claim without a constructibility hypothesis.

An open interval is homeomorphic to ℝ\mathbb R. Extend a sheaf on (0,1)(0,1) by zero to [0,1][0,1]. Formula (C5), and compactness of the closed interval, identify its compactly supported cohomology with the ordinary cohomology of that extension. The proved interval bound therefore gives Hcq(ℝ;Q)=0H_c^q(\mathbb R;Q)=0 for q>1q>1.

Euclidean dimension by one-dimensional fibres

We prove by induction that

Hcq(ℝn;Q)=0(q>n)(D7) H_c^q(\mathbb R^n;Q)=0\qquad(q>n) \qquad\text{(D7)}

for every sheaf QQ. For n=0n=0, sections on a point are exact. The case n=1n=1 was just proved. For n>1n>1, use the projection p:ℝn→ℝn−1p:\mathbb R^n\to\mathbb R^{n-1}. Its fibres are lines. Formula (F6) and the interval result give Rjp!Q=0R^jp_!Q=0 for j∉{0,1}j\notin\{0,1\}. The cohomology truncation triangle is therefore R0p!Q→Rp!Q→(R1p!Q)[−1]→(R0p!Q)[1]R^0p_!Q\to Rp_!Q\to(R^1p_!Q)[-1]\to(R^0p_!Q)[1]. Applying RΓc(ℝn−1;−)R\Gamma_c(\mathbb R^{n-1};-), the induction hypothesis kills its first term in cohomological degrees greater than n−1n-1 and its third term in degrees greater than nn. The long exact sequence kills the middle term in degrees greater than nn. Composition (D3) identifies that middle term with RΓc(ℝn;Q)R\Gamma_c(\mathbb R^n;Q), proving (D7).

If V⊂ℝnV\subset\mathbb R^n is open and GG is any sheaf on VV, (C5) identifies Hcq(V;G)H_c^q(V;G) with Hcq(ℝn;j!G)H_c^q(\mathbb R^n;j_!G). The same bound therefore holds on every open subset, with the same integer nn.

C-softness is local

Suppose a sheaf QQ is c-soft on each member UiU_i of an open cover of a locally compact Hausdorff space. Given a section ss on a compact set KK, choose finitely many compact subsets K1,…,KmK_1,\ldots,K_m of KK covering KK, each contained in a member of the cover. Such a closed refinement follows by choosing, at each point of KK, an open neighbourhood whose closure lies in a cover member, and then taking a finite subcover and intersecting those closures with KK. We extend ss by successive corrections, keeping the already corrected pieces fixed.

Suppose a compactly supported global section already agrees with ss on A=K1∪⋯∪Kr−1A=K_1\cup\cdots\cup K_{r-1}. Its residual section on KK has zero germs on AA, and therefore factors through QX\A|KQ_{X\setminus A}|_K, by the stalk formula for extension by zero. On the cover member containing KrK_r, the sheaf QX\AQ_{X\setminus A} is c-soft: it is an open extension of a restriction of the locally c-soft sheaf QQ. Apply (C2) there to extend the residual on KrK_r to a compactly supported section of that sheaf. Extend the correction by zero to the whole space and add it. It changes no germs on AA, and corrects KrK_r.

Start with zero and repeat finitely many times. The resulting compactly supported section agrees with ss on all of KK, proving c-softness of QQ. Only a finite cover of the given compact set was used; no locally finite global partition was assumed.

The uniform manifold bound and finite c-soft resolutions

Let YY be a manifold of the standing kind, with a fixed finite upper bound NN on all its local dimensions. Then, for every open U⊂YU\subset Y and every sheaf QQ on UU,

Hcq(U;Q)=0(q>N).(D8) H_c^q(U;Q)=0\qquad(q>N). \qquad\text{(D8)}

Here is the finite resolution that proves this globally. For a sheaf QQ on YY, choose an injective resolution II in nonnegative degrees and put Zr=ker⁡(d:Ir→Ir+1)Z^r=\ker(d:I^r\to I^{r+1}), with Z0=QZ^0=Q. The sequences 0→Zr→Ir→Zr+1→00\to Z^r\to I^r\to Z^{r+1}\to0 are exact. On an open subset VV of a coordinate chart, the terms Ir|VI^r|_V are c-soft and compact-section acyclic. Their long exact sequences give, for q>0q>0,

Hcq(V;ZN|V)≃Hcq+N(V;Q|V)=0.(D9) H_c^q(V;Z^N|_V)\simeq H_c^{q+N}(V;Q|_V)=0. \qquad\text{(D9)}

The last equality is the Euclidean open-set bound, since the chart dimension is at most NN. If N=0N=0, it is that bound directly, with no dimension-shifting steps. Criterion (C4) makes ZNZ^N c-soft on each chart. Locality, just proved, makes it c-soft on YY. We thus have the exact finite c-soft resolution

0⟶Q⟶I0⟶⋯⟶IN−1⟶ZN⟶0.(D10) 0\longrightarrow Q\longrightarrow I^0\longrightarrow\cdots \longrightarrow I^{N-1}\longrightarrow Z^N\longrightarrow0. \qquad\text{(D10)}

For N=0N=0, this means that QQ itself is c-soft; there are no injective middle terms. Every term is Γc\Gamma_c-acyclic, so this resolution gives zero compact cohomology above degree NN. Applying the same argument to the open manifold UU, whose local dimensions have the same bound, proves (D8).

Every term of (D10) is also f!f_!-acyclic for any continuous map from YY to a locally compact Hausdorff space, by (F6). Thus the same bound gives Rqf!Q=0R^qf_!Q=0 for q>Nq>N. This proves both dimension inputs used by the bounded c-soft and left-tail constructions below. The bound is uniform because the standing manifold dimensions are uniformly bounded; no claim of a finite global bound is made for an unbounded union of dimensions.

The compact-support, fibre, composition and dimension arguments needed for projection have now been supplied. The later adjunction section supplies the bounded-below exceptional construction and its trace. The later bounds section supplies globally bounded operation estimates. Geometric duality and constructibility remain separate obligations.

Why projection permits arbitrary bounded coefficients

The projection isomorphism (2) is essential to both transport arguments. We prove it here for the standing manifold setting, without assuming finite generation of the coefficient complex. The resolution method is developed from Pierre Schapira’s freely accessible An Introduction to Sheaves on Grothendieck Topologies, 1 August 2026, §4.4. We give the resolution and truncation arguments explicitly; they are needed when a flat resolution has infinitely many terms to the left.

The compact-support section proves injective c-softness, restriction/open-extension/coproduct stability, and the compact-cohomology criterion. The proper-image section proves the fibre formula, c-soft acyclicity for proper direct image and its required underived support compatibilities. The composition and dimension section proves the uniform integer bound Hcj(U;Q)=0H_c^j(U;Q)=0 for j>nj>n, every open U⊂YU\subset Y and every sheaf QQ, where nn bounds the standing manifold dimensions. It also proves Rjf!Q=0R^jf_!Q=0 for j>nj>n. These are the topological inputs used below.

Resolving coefficients by sums of open constant sheaves

For an open subset U⊂XU\subset X, write kUk_U for its constant sheaf extended by zero to XX. A sum ⨁λkUλ\bigoplus_\lambda k_{U_\lambda} is flat: its stalk at a point is a direct sum of copies of kk, and tensor exactness is detected on stalks. Such sums surject onto every sheaf QQ. Indeed, a section s∈Q(U)s\in Q(U) gives a map kU→Qk_U\to Q; take the sum over all open sets and sections. Each germ of QQ lies in the image of one summand, proving stalkwise surjectivity.

The same generators give a bounded-above flat replacement of every bounded-above complex G•G^\bullet. We include the construction because the generators need not be projective as sheaves. Suppose PiP^i, its differential and a chain map ui:Pi→Giu^i:P^i\to G^i have been constructed for i>qi>q. Form the sheaf

Cq={(x,p)∈Gq⊕Pq+1:dGx=uq+1p,dPp=0}.(P1) C^q=\{(x,p)\in G^q\oplus P^{q+1}: d_Gx=u^{q+1}p,\ d_Pp=0\}. \qquad\text{(P1)}

Choose a sum of open constant sheaves PqP^q surjecting onto CqC^q. Compose that epimorphism with the two projections to define uqu^q and dPqd_P^q. Equation (P1) gives dP2=0d_P^2=0 and dGu=udPd_Gu=ud_P. Start above the upper bound of GG with zero terms and repeat in descending degrees.

To verify that u:P→Gu:P\to G is a quasi-isomorphism, use the convention

Cone⁡(u)q=Gq⊕Pq+1,d(x,p)=(dGx+u(p),−dPp).(P2) \operatorname{Cone}(u)^q=G^q\oplus P^{q+1},\qquad d(x,p)=(d_Gx+u(p),-d_Pp). \qquad\text{(P2)}

A cycle (x,p)(x,p) determines (x,−p)∈Cq(x,-p)\in C^q. Locally it lifts to some z∈Pqz\in P^q, by the epimorphism just chosen. Then d(0,z)=(x,p)d(0,z)=(x,p). The cone is therefore exact on stalks, which proves the claim. No lifting through a globally projective open-generator sheaf was assumed.

A bounded-above complex of flat sheaves is K-flat for the direct-sum tensor totalization used here. For completeness, its brutal truncations obtained by discarding all terms below a fixed degree are bounded subcomplexes. A bounded flat complex tensored with an exact complex is exact: filter it by its finitely many terms and use flatness and the long exact cohomology sequences of the resulting short exact sequences of complexes. The whole bounded-above flat complex is the directed union of these subcomplexes. Tensor products commute with that union, and directed colimits of sheaves of modules are exact because this can be checked on stalks. Tensoring the union with an exact complex is thus exact. This proves K-flatness and justifies using PP to compute a derived tensor product.

Removing an exact left tail before deriving

Finite-dimension lemma. Let TT be a left exact additive functor between abelian categories with enough injectives, and assume RjT=0R^jT=0 for j>rj>r. Let C•C^\bullet be bounded above, with TT-acyclic terms and bounded cohomology. Then the termwise complex T(C•)T(C^\bullet) computes RT(C•)RT(C^\bullet).

Proof. Choose aa such that Hi(C)=0H^i(C)=0 for i<ai<a, and write Zi=ker⁡dCiZ^i=\ker d_C^i. For i<ai<a, the exact left tail gives short exact sequences

0⟶Zi−1⟶Ci−1⟶Zi⟶0.(P3) 0\longrightarrow Z^{i-1}\longrightarrow C^{i-1} \longrightarrow Z^i\longrightarrow0. \qquad\text{(P3)}

Since the middle terms are TT-acyclic, the long exact derived-functor sequences give, for j>0j>0,

RjT(Zi)≃Rj+rT(Zi−r)=0.(P4) R^jT(Z^i)\simeq R^{j+r}T(Z^{i-r})=0. \qquad\text{(P4)}

For r=0r=0, vanishing is immediate from the hypothesis; no connecting maps are required. Thus every ZiZ^i with i<ai<a is TT-acyclic. Let Ba=im⁡dCa−1B^a=\operatorname{im}d_C^{a-1}. The sequence 0→Za−1→Ca−1→Ba→00\to Z^{a-1}\to C^{a-1}\to B^a\to0 also makes BaB^a acyclic for TT.

Replace the entire left tail by BaB^a in degree a−1a-1, with its inclusion into CaC^a. Call the resulting bounded complex C′C'. The map C→C′C\to C', given by the quotient in degree a−1a-1 and the identity above it, is a quasi-isomorphism. Applying TT preserves that quasi-isomorphism: (P3) and the last short exact sequence remain exact after applying TT, since their left terms are acyclic. The bounded complex C′C' has acyclic terms, so the usual bounded acyclic-resolution argument computes RT(C′)RT(C') by T(C′)T(C'). Hence T(C)T(C) represents the same derived object. All comparisons are induced by the specified truncation maps, so the result identifies the natural computation map. ▫\square

The bounded acyclic-resolution argument used in the last sentence follows by filtering C′C' by its finitely many brutal truncations. For a single acyclic term it is the definition of acyclicity. At each additional term the short exact sequence of complexes and its derived triangle identify the two computations by the long exact cohomology sequence. Induction over the finite number of terms proves the assertion.

A bounded c-soft model for the source input

Let B∈Db(kY)B\in D^b(k_Y), with cohomology in degrees [a,b][a,b]. Choose a bounded-below injective representative I•I^\bullet starting in degree aa. Its terms are c-soft. Put Zi=ker⁡dIiZ^i=\ker d_I^i. Since the cohomology above bb is zero, for every i≥bi\geq b there is an exact sequence

0⟶Zi⟶Ii⟶Zi+1⟶0.(P5) 0\longrightarrow Z^i\longrightarrow I^i\longrightarrow Z^{i+1}\longrightarrow0. \qquad\text{(P5)}

Restriction to an open U⊂YU\subset Y preserves these exact sequences and c-softness. Dimension shifting gives, for j>0j>0,

Hcj(U;Zb+n|U)≃Hcj+n(U;Zb|U)=0.(P6) H_c^j(U;Z^{b+n}|_U) \simeq H_c^{j+n}(U;Z^b|_U)=0. \qquad\text{(P6)}

For n=0n=0, the vanishing follows directly from the uniform compact-cohomology bound. The c-soft criterion therefore makes Zb+nZ^{b+n} c-soft. Replace Ib+nI^{b+n} by this kernel and discard higher degrees. This upper truncation A•A^\bullet is quasi-isomorphic to BB, has finitely many nonzero terms, and every term is c-soft. Consequently its terms are f!f_!-acyclic and f!Af_!A computes Rf!BRf_!B.

For an ordinary sheaf BB in degree zero, this construction produces AA in degrees 0,…,n0,\ldots,n. Its terms are f!f_!-acyclic by the proved fibre and c-soft results. Hence Rjf!B=0R^jf_!B=0 for j>nj>n. Thus the same uniform compact-cohomology bound implies finite cohomological dimension of f!f_!; no additional dimension hypothesis on the map is needed for the left-tail lemma.

The projection proof

Choose the bounded c-soft model AA of BB, and an open-generator replacement P→EP\to E for E∈Db(kX)E\in D^b(k_X). Inverse image is exact and sends kUk_U to kf−1Uk_{f^{-1}U}, so f−1Pf^{-1}P is again a bounded-above K-flat complex. It follows that A⊗f−1PA\otimes f^{-1}P represents B⊗Lf−1EB\otimes^Lf^{-1}E.

Each term of that tensor total complex is a finite sum, over the degrees of AA, of coproducts of open extensions of c-soft sheaves. It is c-soft and hence f!f_!-acyclic. The complex is bounded above. Its cohomology is bounded below as well: stalkwise, the two inputs are bounded complexes of kk-modules and finite global dimension bounds all Tor degrees. The finite-dimension lemma therefore proves that applying f!f_! termwise calculates the desired derived image even though PP may have an infinite left tail.

For an individual open generator the underived projection map is

f!Ap⊗kU→∼f!(Ap⊗kf−1U).(P7) f_!A^p\otimes k_U\xrightarrow{\sim} f_!(A^p\otimes k_{f^{-1}U}). \qquad\text{(P7)}

It is the compatibility of proper-support image with open restriction and extension by zero. Since f!f_! commutes with coproducts, (P7) holds with any PqP^q in place of kUk_U. Totalization uses only finitely many terms in each degree, because AA is bounded. Thus the termwise maps assemble to an isomorphism of complexes

(f!A)⊗P→∼f!(A⊗f−1P).(P8) (f_!A)\otimes P\xrightarrow{\sim}f_!(A\otimes f^{-1}P). \qquad\text{(P8)}

The left side represents Rf!B⊗LERf_!B\otimes^LE, since PP is K-flat. The right side represents Rf!(B⊗Lf−1E)Rf_!(B\otimes^Lf^{-1}E), by the preceding paragraph. This proves (2) for arbitrary bounded coefficients.

Finally, these are the canonical projection maps. Before deriving, (P7) multiplies a properly supported section by a pulled-back coefficient section. Multiplying by two coefficients consecutively is the same as multiplying by their tensor product; multiplication by the unit coefficient is the identity. Swapping homogeneous coefficients gives exactly the Koszul sign. These equalities hold termwise in (P8), commute with differentials and pass to the derived category. Hence this proof supplies the unit, associativity and symmetry compatibilities needed by the pairings (5)–(11).

Constructing the adjunction and fixing its trace

We now construct the adjunction used in (1). For this construction let f:Y→Xf:Y\to X, with YY a manifold whose local dimensions are at most NN, and XX locally compact Hausdorff. The coefficients may be any fixed commutative ring kk. The construction gives Rf!⊣f!Rf_!\dashv f^! on D+D^+ and a lower cohomological bound. The next section proves the additional upper bound needed to keep its output globally bounded in the standing manifold setting.

A human source for the finite-resolution method is Akhil Mathew’s freely accessible Verdier duality, 29 July 2011, Corollary 3.6, Lemma 3.7 and §§4.1–4.3. That treatment assumes Noetherian coefficients. Here we work first over ℤ\mathbb Z, prove the needed tensor and generator statements, and then carry the arbitrary kk-action through the construction. In particular, we do not assume that products of flat modules over an arbitrary ring are flat. The compact-support and dimension inputs are the proofs already given in this lesson.

A finite resolution that works with every coefficient sheaf

The bound (D8), with integral coefficients, has the following consequence. Suppose an abelian sheaf EE admits an exact sequence 0→A→SN−1→⋯→S0→E→00\to A\to S_{N-1}\to\cdots\to S_0\to E\to0, where the NN middle terms are c-soft. Restrict to any open V⊂YV\subset Y. Repeated connecting morphisms identify Hcq(V;E)H_c^q(V;E) with Hcq+N(V;A)H_c^{q+N}(V;A) for q>0q>0. The latter vanishes. Criterion (C4) makes EE c-soft. If N=0N=0, every abelian sheaf is c-soft directly by the same bound and criterion.

Consequently, if LL is a flat c-soft abelian sheaf, then M⊗ℤLM\otimes_{\mathbb Z}L is c-soft for every abelian sheaf MM. Indeed, resolve MM to the left by coproducts of open generators ℤU\mathbb Z_U, choosing generators for local sections and repeating on the kernel. Flatness preserves exactness after tensoring. The resulting terms are coproducts of LUL_U, hence c-soft by the proved support stability. Use the last NN terms and the preceding dimension shift. This proof applies to the underlying abelian sheaf of a kk-module, without assuming that kk is flat over ℤ\mathbb Z.

We construct a finite resolution of the constant integral sheaf. For an abelian sheaf EE, put C(E)(V)=∏y∈VEyC(E)(V)=\prod_{y\in V}E_y, with restrictions deleting coordinates. This is a sheaf and is flabby, because arbitrary coordinate families extend by zero. The germ map E→C(E)E\to C(E) is injective. On the stalk at yy, evaluating the yy-coordinate is a retraction: an element of C(E)yC(E)_y is represented on a neighbourhood containing yy, so that coordinate is well defined and sends the germ of a genuine section back to itself.

If the stalks of EE are torsion-free, so are the products defining C(E)C(E), their filtered stalk colimits, and the direct summands (C(E)/E)y(C(E)/E)_y. Torsion-free abelian groups are flat: each is the filtered union of its finitely generated subgroups, which are free by integer row reduction, and filtered colimits of tensor functors preserve exactness. Thus C(E)C(E) and C(E)/EC(E)/E are flat whenever EE is flat over ℤ\mathbb Z.

Starting from E0=ℤYE^0=\mathbb Z_Y, form Lp=C(Ep)L^p=C(E^p) and Ep+1=Lp/EpE^{p+1}=L^p/E^p for 0≤p<N0\le p<N, and put LN=ENL^N=E^N. The preceding dimension shift makes the final term c-soft; the other terms are flabby. We obtain

0⟶ℤY⟶L0⟶⋯⟶LN⟶0,Lp flat and c-soft.(A1) 0\longrightarrow\mathbb Z_Y\longrightarrow L^0\longrightarrow\cdots \longrightarrow L^N\longrightarrow0, \qquad L^p\text{ flat and c-soft}. \qquad\text{(A1)}

For N=0N=0, take L0=ℤYL^0=\mathbb Z_Y. Each constituent short exact sequence of (A1) splits on stalks by the coordinate retraction, so tensoring it with any sheaf preserves exactness. Hence M→M⊗ℤL•M\to M\otimes_{\mathbb Z}L^\bullet is a quasi-isomorphism for every sheaf MM. It remains so for a bounded-below complex: the augmented double complex has exact finite rows in the LL-direction, and in each total degree only finitely many entries occur. Successive elimination along these rows proves its total complex acyclic. Every term of the unaugmented total complex is c-soft by the tensor result. Formula (F6) and its acyclic-complex argument therefore give

Rf!G≃f!Tot⁡(G⊗ℤL•)(G∈D+(kY)).(A2) Rf_!G\simeq f_!\operatorname{Tot}(G\otimes_{\mathbb Z}L^\bullet) \qquad(G\in D^+(k_Y)). \qquad\text{(A2)}

Representing the supported pairing on open sets

For a flat c-soft abelian sheaf LL, write TL(M)=f!(M⊗ℤL)T_L(M)=f_!(M\otimes_{\mathbb Z}L) for kk-module sheaves MM. This functor is exact. Tensoring a short exact sequence is exact, and its c-soft terms are f!f_!-acyclic, so the proper-image long exact sequence proves the assertion. It also preserves coproducts, by (F4) and distributivity of tensor.

For an injective kXk_X-module sheaf II, define

JL(I)(U)=Hom⁡kX(TL(kU),I)(U⊂Y open).(A3) J_L(I)(U)=\operatorname{Hom}_{k_X}\bigl(T_L(k_U),I\bigr) \qquad(U\subset Y\text{ open}). \qquad\text{(A3)}

The inclusion kV→kUk_V\to k_U for V⊂UV\subset U defines restriction. This presheaf is a sheaf: if U=⋃iUiU=\bigcup_iU_i, the sequence ⨁i,jkUi∩Uj→⨁ikUi→kU→0\bigoplus_{i,j}k_{U_i\cap U_j}\to\bigoplus_i k_{U_i}\to k_U\to0 is right exact, with the first map the difference of the two inclusions. On a stalk in UU, this is the presentation of one copy of kk by generators indexed by cover members containing the point and relations identifying any two; outside UU all terms vanish. Applying the exact coproduct-preserving TLT_L, then Hom⁡(−,I)\operatorname{Hom}(-,I), gives precisely the sheaf equalizer for (A3). The presentation uses coproducts of sheaves; the resulting equalizer uses products of section modules.

There is a natural bijection

Hom⁡kX(TL(M),I)≃Hom⁡kY(M,JL(I)).(A4) \operatorname{Hom}_{k_X}(T_L(M),I) \simeq\operatorname{Hom}_{k_Y}(M,J_L(I)). \qquad\text{(A4)}

Given a morphism on the left, precompose with TL(kU)→TL(M)T_L(k_U)\to T_L(M) for each local section of MM. The resulting compatible linear maps on sections define the morphism on the right. For M=kUM=k_U this is the identity in (A3), since Hom⁡(kU,JL(I))=JL(I)(U)\operatorname{Hom}(k_U,J_L(I))=J_L(I)(U). It is therefore a bijection for coproducts of open generators. Every MM has a presentation P1→P0→M→0P_1\to P_0\to M\to0 by such coproducts: use all local sections to construct the first surjection and repeat for its kernel. Both sides of (A4) send that presentation to the kernel of the corresponding map from the value on P0P_0 to the value on P1P_1. The bijections on the two generators identify the kernels. This proves (A4) and its naturality. No projectivity of kUk_U, or colimit formula omitting additive relations, has been used.

The left side of (A4) is exact as a contravariant functor of MM. Thus JL(I)J_L(I) is injective. This construction is covariant in II and contravariant in LL, as is already visible in (A3).

The complex and the adjunction

Represent F∈D+(kX)F\in D^+(k_X) by a bounded-below injective complex II, and use the fixed finite resolution (A1). Define a complex J(I)J(I) by

J(I)n=⨁p=0NJLp(In+p),(dh)p=dIhp−(−1)nhp+1dLp.(A5) J(I)^n=\bigoplus_{p=0}^{N}J_{L^p}(I^{n+p}), \qquad (dh)_p=d_Ih_p-(-1)^n h_{p+1}d_L^p. \qquad\text{(A5)}

The last term is zero when p=Np=N. The notation means postcomposition in II and precomposition in LL. The two mixed terms cancel when this differential is squared, while dI2=dL2=0d_I^2=d_L^2=0. Every term is a finite sum of injectives, hence injective. If II starts in degree aa, then J(I)J(I) starts in degree a−Na-N.

Apply (A4) in each bidegree of (A2). Reindexing the Hom products is legitimate because there are only N+1N+1 possible pp’s. It gives an isomorphism of Hom complexes

Hom⁡•(f!Tot(G⊗ℤL),I)≃Hom⁡•(G,J(I)).(A6) \operatorname{Hom}^{\bullet}\!\left( f_!\operatorname{Tot}(G\otimes_{\mathbb Z}L),I\right) \simeq\operatorname{Hom}^{\bullet}(G,J(I)). \qquad\text{(A6)}

Here is the sign check. A degree-nn map has components ϕi,p:f!(Gi⊗Lp)→Ii+p+n\phi_{i,p}:f_!(G^i\otimes L^p)\to I^{i+p+n}. Its Hom differential has three terms: postcomposition by dId_I, minus (−1)n(-1)^n times precomposition by dGd_G, and minus (−1)n+i(-1)^{n+i} times precomposition by dLd_L. After currying, ϕi\phi_i takes values in J(I)i+nJ(I)^{i+n}. Formula (A5), followed by the Hom differential for GG, gives these same three signs. Thus (A6) is a chain isomorphism, not only a graded bijection.

Both target complexes in (A6) are bounded-below injective complexes. Their Hom-complex degree-zero cohomology computes morphisms in D+D^+. Together with (A2), this proves

Hom⁡D+(kX)(Rf!G,F)≃Hom⁡D+(kY)(G,f!F),f!F=J(I),f!D≥a⊂D≥a−N.(A7) \operatorname{Hom}_{D^+(k_X)}(Rf_!G,F) \simeq\operatorname{Hom}_{D^+(k_Y)}(G,f^!F), \qquad f^!F=J(I), \quad f^!D^{\ge a}\subset D^{\ge a-N}. \qquad\text{(A7)}

Maps and homotopies of injective complexes induce maps and homotopies of (A5). Bounded-below injective replacements therefore make this a functor on D+D^+; the Hom differential also gives its shift and cone compatibility. The Hom complexes themselves need not be bounded below when both inputs extend arbitrarily far to the right. The argument only uses their degree-zero cohomology and makes no such extra bound claim.

For completeness, ordinary adjunction comes from the sheaf inverse-image construction. The inverse-image presheaf is the colimit of sections over open sets containing the image, and sheafification does not change maps into a sheaf. This identifies maps f−1A→Bf^{-1}A\to B with maps A→f*BA\to f_*B. Inverse image is exact, since its stalk at yy is the stalk at f(y)f(y). Therefore f*f_* takes injectives to injectives, by its Hom test. Applying the same degreewise adjunction to a bounded-below injective replacement gives

Hom⁡D+(kY)(f−1A,B)≃Hom⁡D+(kX)(A,Rf*B).(A8) \operatorname{Hom}_{D^+(k_Y)}(f^{-1}A,B) \simeq\operatorname{Hom}_{D^+(k_X)}(A,Rf_*B). \qquad\text{(A8)}

Units, traces and exceptional composition

Let ηG:G→f!Rf!G\eta_G:G\to f^!Rf_!G correspond to 1Rf!G1_{Rf_!G} under (A7), and let ϵF:Rf!f!F→F\epsilon_F:Rf_!f^!F\to F correspond to 1f!F1_{f^!F}. The latter is the trace. Naturality of the bijection says that the transpose of u:Rf!G→Fu:Rf_!G\to F is f!u∘ηGf^!u\circ\eta_G, and the inverse transpose of v:G→f!Fv:G\to f^!F is ϵF∘Rf!v\epsilon_F\circ Rf_!v. Apply these two inverse operations to the identity morphisms. They give

ϵRf!G∘Rf!ηG=1Rf!G,f!ϵF∘ηf!F=1f!F.(A9) \epsilon_{Rf_!G}\circ Rf_!\eta_G=1_{Rf_!G}, \qquad f^!\epsilon_F\circ\eta_{f^!F}=1_{f^!F}. \qquad\text{(A9)}

If another choice of (A1) or injective representative produces a second adjoint, its counit transposes to a unique comparison with this one. Exchanging the choices produces the inverse, by (A9). The comparison preserves the trace, is natural, and satisfies the cocycle identity for three choices, since its transpose is fixed. The chain isomorphism (A6) makes this normalization compatible with shifts; there is no free choice of a sign for the shifted trace.

For Z→gY→fXZ\xrightarrow{g}Y\xrightarrow{f}X, successive adjunctions and (D3) identify maps from GG to g!f!Fg^!f^!F with maps from R(fg)!GR(fg)_!G to FF. Consequently there is a unique adjoint comparison

(fg)!F≃g!f!F,ϵfg,F=ϵf,F∘Rf!(ϵg,f!F),(A10) (fg)^!F\simeq g^!f^!F, \qquad \epsilon_{fg,F}=\epsilon_{f,F}\circ Rf_!(\epsilon_{g,f^!F}), \qquad\text{(A10)}

where the source of the trace is identified by (D3) and this comparison. Indeed, the composite on the right is the inverse transpose of the identity under the successive adjunctions, which characterizes the counit on the left. With three maps, either parenthesization has this same ordered composite of traces; associativity of (D3) and uniqueness of a trace-preserving adjoint comparison make the two identifications equal. For the identity map, proper image is the identity and its adjunction has identity trace, proving the unit compatibility. The same argument with the last map aX:X→{*}a_X:X\to\{*\} gives f!ωX≃ωYf^!\omega_X\simeq\omega_Y for ωX=aX!k\omega_X=a_X^!k, with exactly the composite trace required in the opening contracts.

These constructions supply both adjunctions and trace-compatible exceptional composition on D+D^+. The finite model gives the lower bound in (A7); the next section proves the remaining globally bounded estimates. The tensor–Hom section below supplies explicit resolutions, currying and evaluation. The later orientation section supplies orientation identification and constant relative-ball calculations. Constructible biduality and weak-constructibility results remain explicit proof obligations. They are not inferred from existence of the adjoint.

Why the operations remain globally bounded

Write NN and MM for finite upper bounds on the local dimensions of YY and XX, respectively. We now prove the bounds needed in the opening contracts for a continuous map f:Y→Xf:Y\to X between the standing manifolds. Besides the lower bound in (A7), the missing ingredients are ordinary cohomology and an upper bound for exceptional inverse image.

The freely accessible human reference is Schapira’s An Introduction to Sheaves on Grothendieck Topologies, 1 August 2026, Proposition 5.1.2 and Proposition 5.1.9. We supply the ordinary-section lifting argument, the constant-interval computation and the graph factorization explicitly. For the projection bound we use supported rectangular opens and the adjunction just constructed; an orientation identification is not needed for this bound.

From compact extension to ordinary cohomology

Every standing manifold has a compact exhaustion K1⊂int⁡K2⊂⋯K_1\subset\operatorname{int}K_2\subset\cdots with union the whole manifold. To construct one, start with its countable compact cover. At each stage cover the preceding compact set and the next member of that cover by finitely many relatively compact open neighbourhoods, and take the union of their compact closures. Each compact subset of a manifold is covered by finitely many coordinate charts, so the countable compact cover also gives a countable union of countable chart bases. Thus the manifold is second countable. Every open subset inherits a countable base and has a countable cover by relatively compact neighbourhoods; the same exhaustion construction applies there.

Consider a short exact sequence of sheaves 0→S→E→Q→00\to S\to E\to Q\to0, where SS is c-soft, and let s∈Γ(Y;Q)s\in\Gamma(Y;Q). On every compact KnK_n, the compact lifting result (C3), applied to the restricted sequence, lifts s|Kns|_{K_n} to a section of E|KnE|_{K_n}. Make these lifts compatible inductively. If a lift ene_n is already chosen, take any lift en+1′e'_{n+1} on Kn+1K_{n+1}. The difference en−en+1′|Kne_n-e'_{n+1}|_{K_n} lies in S|KnS|_{K_n}. C-softness extends that difference to a global section of SS; add its restriction to en+1′e'_{n+1}. The corrected lift agrees with ene_n. The resulting sections agree on the nested interiors and glue to a global lift of ss. Hence

Γ(Y;E)⟶Γ(Y;Q)⟶0is exact when the kernel sheaf is c-soft.(B1) \Gamma(Y;E)\longrightarrow\Gamma(Y;Q)\longrightarrow0 \quad\text{is exact when the kernel sheaf is c-soft}. \qquad\text{(B1)}

A c-soft sheaf is therefore acyclic for ordinary sections on these spaces. Indeed, embed it into an injective sheaf. The quotient is c-soft by the quotient result preceding (C4); repeat. Formula (B1) makes the resulting injective resolution exact after taking ordinary sections in positive degrees. Apply this to the finite c-soft resolution (D10), and then to any open subset. For every sheaf QQ on every open U⊂YU\subset Y, we obtain

Hq(U;Q)=0(q>N).(B2) H^q(U;Q)=0\qquad(q>N). \qquad\text{(B2)}

This argument concerns ordinary sections and uses the countable exhaustion. Compact-section acyclicity alone was not silently substituted for it.

For an injective resolution II of a sheaf QQ on YY, the stalk at xx of the cohomology of f*If_*I is colim⁡x∈VHq(f−1V;Q)\mathop{\mathrm{colim}}_{x\in V}H^q(f^{-1}V;Q). Open restriction preserves injectives, and filtered colimits of modules are exact, so this is the stalk of Rqf*QR^qf_*Q. Formula (B2) makes it zero for q>Nq>N. The proper-image bound was proved in (D10). Applying the finite cohomology truncation triangles of a bounded complex now yields

Rf*D[a,b](kY)⊂D[a,b+N](kX),Rf!D[a,b](kY)⊂D[a,b+N](kX).(B3) Rf_*D^{[a,b]}(k_Y)\subset D^{[a,b+N]}(k_X), \qquad Rf_!D^{[a,b]}(k_Y)\subset D^{[a,b+N]}(k_X). \qquad\text{(B3)}

Here each successive triangle attaches a single cohomology sheaf in its original degree; the two sheaf bounds and the long exact sequence give exactly the displayed interval. Inverse image remains exact and preserves [a,b][a,b].

Closed support and its bound

Let i:Z↪Wi:Z\hookrightarrow W be closed, and let j:W\Z↪Wj:W\setminus Z\hookrightarrow W be its open complement. The sheaf ΓZE\Gamma_Z E is the kernel of E→j*j−1EE\to j_*j^{-1}E. A map from i*Ai_*A to EE has zero stalks off ZZ, so lands in ΓZE\Gamma_ZE, and is uniquely determined by its restriction to ZZ. Thus i−1ΓZi^{-1}\Gamma_Z is the sheaf right adjoint to the exact functor i*i_*. It preserves injectives. For an injective II, flabbiness makes I(V)→I(V\Z)I(V)\to I(V\setminus Z) surjective for each open VV. Therefore 0→ΓZI→I→j*j−1I→00\to\Gamma_ZI\to I\to j_*j^{-1}I\to0 is exact.

For a bounded-below injective representative of FF, open restriction is injective and computes the derived open image. The preceding short exact sequence of complexes gives the localization triangle

i*i!F⟶F⟶Rj*j−1F⟶(i*i!F)[1].(B4) i_*i^!F\longrightarrow F\longrightarrow Rj_*j^{-1}F \longrightarrow(i_*i^!F)[1]. \qquad\text{(B4)}

The left term uses the derived sheaf adjunction just proved. When the exceptional adjoint (A7) applies to ii, uniqueness of the adjoint identifies the two constructions and their counits, since i!=i*i_!=i_*. This also fixes the first arrow as the support-inclusion map.

If WW is a standing manifold of local dimension at most DD, (B3) applies to jj. The long exact sequence of (B4) and left exactness of the sheaf support functor give

i!D[a,b](kW)⊂D[a,b+D+1](kZ).(B5) i^!D^{[a,b]}(k_W)\subset D^{[a,b+D+1]}(k_Z). \qquad\text{(B5)}

Indeed, the two middle terms in (B4) vanish above b+Db+D, so the first vanishes above b+D+1b+D+1; its lower bound is aa. Exact closed direct image and restriction detect these same bounds on ZZ. This estimate is sufficient here; it is not asserted optimal.

We also need the open restriction of an exceptional image. For an open inclusion u:U↪Yu:U\hookrightarrow Y, exact extension by zero has exact right adjoint u−1u^{-1}: the bijection follows by restricting a map out of a section extended by zero, and extending it back. It gives u!=u−1u^!=u^{-1}. Composition (A10) consequently identifies (fu)!F(fu)^!F with u−1f!Fu^{-1}f^!F, compatibly with the trace. The boundedness question is therefore local on the source.

Compact cohomology of constant coefficients on boxes

First, a constant sheaf on a closed interval has no positive cohomology. The case q>1q>1 was proved in (D6). For q=1q=1, its same closed-interval Mayer–Vietoris sequence has preceding map k⊕k→kk\oplus k\to k, the difference of the values on the two intervals at their common endpoint. This map is surjective. Thus (D6) is also an isomorphism in degree one for constant coefficients. The proof following (D6) now works unchanged: a class vanishes on a small initial interval by compact-set continuity, the supremum of the zero initial intervals belongs to that set by two-interval gluing, and continuity extends beyond it unless it is the right endpoint. Hence the class is zero. Degree-zero sections are kk, since an interval is connected. We have proved

RΓ([0,1];k)≃k.(B6) R\Gamma([0,1];k)\simeq k. \qquad\text{(B6)}

Let v:(0,1)↪[0,1]v:(0,1)\hookrightarrow[0,1] and let e:{0,1}↪[0,1]e:\{0,1\}\hookrightarrow[0,1]. The sequence 0→v!k→k[0,1]→e*k{0,1}→00\to v_!k\to k_{[0,1]}\to e_*k_{\{0,1\}}\to0 is exact by its stalks. Compact sections on the closed interval are ordinary sections. Applying cohomology, using (B6) and the two-point calculation, gives

RΓc((0,1);k)≃k[−1],0⟶k→t↦(t,t)k2→(s,t)↦t−sk⟶0.(B7) R\Gamma_c((0,1);k)\simeq k[-1], \qquad 0\longrightarrow k\xrightarrow{t\mapsto(t,t)}k^2 \xrightarrow{(s,t)\mapsto t-s}k\longrightarrow0. \qquad\text{(B7)}

The displayed cokernel identifies the degree-one group; the other degrees vanish. Equivalently, the compact-support complex is the shifted cone of the diagonal map, with this cokernel identification. Every open interval has the same conclusion by a homeomorphism.

For finite-dimensional manifolds A,BA,B, proper base change identifies the proper image under A×B→BA\times B\to B of the inverse image of E∈Db(kA)E\in D^b(k_A) with the constant complex on BB associated to RΓc(A;E)R\Gamma_c(A;E). Apply the projection formula with F∈Db(kB)F\in D^b(k_B), then compose with the map from BB to a point. A second projection formula gives

RΓc(A×B;E⊠LF)≃RΓc(A;E)⊗kLRΓc(B;F).(B8) R\Gamma_c(A\times B;E\boxtimes^L F) \simeq R\Gamma_c(A;E)\otimes_k^L R\Gamma_c(B;F). \qquad\text{(B8)}

All operations here are the already constructed ones; (D3), (D4) and (P8) supply these maps. The standing finite global dimension keeps the tensor products bounded. For constant coefficients on a product of nn open intervals, induction using (B7) gives

RΓc(U;k)≃k[−n]for an open coordinate box U⊂ℝn.(B9) R\Gamma_c(U;k)\simeq k[-n] \qquad\text{for an open coordinate box }U\subset\mathbb R^n. \qquad\text{(B9)}

For n=0n=0, the box is a point. For higher nn, tensoring the free rank-one complexes introduces no Tor groups. Only the existence of these identifications is needed for the bound below. The later orientation section identifies their transition maps, including the determinant signs.

Projection bounds from the actual adjunction

Let p:T×X→Xp:T\times X\to X, where TT is a manifold of constant dimension nn. For a coordinate box U⊂TU\subset T and an open V⊂XV\subset X, proper-image composition for the open inclusions, base change and (B9) give

Rp!kU×V≃kV[−n].(B10) Rp_!k_{U\times V}\simeq k_V[-n]. \qquad\text{(B10)}

For fixed UU, choose its identification in (B9) once. The resulting comparison (B10) is then natural as VV shrinks; it is extension of the same constant compact-cohomology complex from VV.

Let F∈D[a,b](kX)F\in D^{[a,b]}(k_X). Derived sections on U×VU\times V, the adjunction (A7) and (B10) give

Hq(U×V;p!F)≃Hom⁡(Rp!kU×V,F[q])≃Hq+n(V;F).(B11) H^q(U\times V;p^!F) \simeq\operatorname{Hom}(Rp_!k_{U\times V},F[q]) \simeq H^{q+n}(V;F). \qquad\text{(B11)}

These equalities are natural in VV for fixed UU. To pass to a stalk at (t,x)(t,x), represent a class on a rectangle U×VU\times V. For any bounded-below complex CC, exactness of filtered colimits and the definition of a stalk give colim⁡x∈VHj(V;C)=ℋj(C)x\mathop{\mathrm{colim}}_{x\in V}H^j(V;C)=\mathcal H^j(C)_x: use a bounded-below injective representative and commute the colimit with cohomology. If q+n∉[a,b]q+n\notin[a,b], the class corresponding under (B11) therefore becomes zero after shrinking VV. Naturality makes the original class zero on that smaller rectangle. Rectangles form a neighbourhood basis, so every such stalk class vanishes. This proves the sharp degree interval

p!D[a,b](kX)⊂D[a−n,b−n](kT×X).(B12) p^!D^{[a,b]}(k_X)\subset D^{[a-n,b-n]}(k_{T\times X}). \qquad\text{(B12)}

No compatibility between orientation choices on different boxes was required for this vanishing argument. The later orientation section proves that compatibility and the global orientation-sheaf identification.

A closed graph gives a bound for every continuous map

Restrict f:Y→Xf:Y\to X to a coordinate neighbourhood U⊂YU\subset Y of dimension nn. Factor it as the graph embedding i:U→U×Xi:U\to U\times X followed by the projection pp. The graph is closed because XX is Hausdorff: two unequal points of XX have disjoint neighbourhoods, so the complement of the graph is open. The ambient manifold has local dimension at most n+Mn+M.

For F∈D[a,b](kX)F\in D^{[a,b]}(k_X), (B12) places p!Fp^!F in degrees [a−n,b−n][a-n,b-n]. Apply (B5) to the closed graph, and use exceptional composition and open restriction. On UU, the resulting f!Ff^!F has cohomology only in degrees [a−n,b+M+1][a-n,b+M+1]. All chart dimensions satisfy n≤Nn\le N. Hence globally

f!D[a,b](kX)⊂D[a−N,b+M+1](kY).(B13) f^!D^{[a,b]}(k_X)\subset D^{[a-N,b+M+1]}(k_Y). \qquad\text{(B13)}

The upper estimate is a uniform sufficient bound, not a proposed optimal amplitude for every map. It proves the globally bounded preservation required by the opening contracts, with no constructibility or properness restriction. Together with (B3), it closes the previously missing ordinary and exceptional operation bounds.

A bounded injective model for the dualizing object

Let dd be the finite global dimension of kk. The module kk has an injective resolution in degrees 0,…,d0,\ldots,d. To see the bound, take any injective resolution and let ZdZ^d be its dd-th kernel. Dimension shifting gives Ext⁡k1(M,Zd)=Ext⁡kd+1(M,k)=0\operatorname{Ext}^1_k(M,Z^d)=\operatorname{Ext}^{d+1}_k(M,k)=0 for every module MM, the last vanishing following from a projective resolution of length at most dd. Thus ZdZ^d is injective: vanishing of these extension groups makes every extension with kernel ZdZ^d split, which, by pushing out along a map from a submodule, is exactly the extension property for injectivity. For d=0d=0, the same argument applies to kk itself.

Apply the explicit construction (A5) to aY:Y→{*}a_Y:Y\to\{*\} and this finite injective module resolution. It supplies a representative WYW_Y of ωY\omega_Y with

WYq=0(q∉[−N,d]),WYq injective as a kY-module sheaf.(B14) W_Y^q=0\quad(q\notin[-N,d]), \qquad W_Y^q\text{ injective as a }k_Y\text{-module sheaf}. \qquad\text{(B14)}

For F∈D[a,b](kY)F\in D^{[a,b]}(k_Y), choose its good-truncated representative in degrees [a,b][a,b]. The internal Hom complex ℋom•(F,WY)\mathcal Hom^\bullet(F,W_Y) computes DYFD_YF. In detail, restriction of an injective sheaf to any open remains injective, so ℋom(−,WYq)\mathcal Hom(-,W_Y^q) is exact as a contravariant sheaf functor. On each open its sections are the usual Hom into that injective restriction. The finite Hom totalization consequently sends bounded acyclic source complexes to acyclic complexes and computes the derived internal Hom into the chosen injective target. Its term in degree rr is a finite product of ℋom(Fi,WYi+r)\mathcal Hom(F^i,W_Y^{i+r}); all such terms vanish outside the interval

DYD[a,b](kY)⊂D[−N−b,d−a](kY).(B15) D_YD^{[a,b]}(k_Y)\subset D^{[-N-b,d-a]}(k_Y). \qquad\text{(B15)}

This proves boundedness of duality for arbitrary bounded sheaves, without assuming biduality, finite rank or constructibility. Identifying ωY\omega_Y with the shifted orientation sheaf, and proving when evaluation is an isomorphism, remain separate matters. The next section proves the tensor–Hom and evaluation statements on explicit models. The following orientation section proves the orientation and constant relative-ball statements. Biduality and constructibility retain their separate proof obligations.

Tensor–Hom adjunction and the actual evaluation map

We now supply the complex models behind the remaining algebraic operation contract. Put d=gldim⁡kd=\operatorname{gldim}k. This section uses bounded complexes for the lesson’s inputs and bounded-below injective resolutions for their targets. No perfectness or finite generation is imposed. Schapira’s freely accessible An Introduction to Sheaves on Grothendieck Topologies, §§2.1–2.3, pp. 43–45 and 49–53 supplies the tensor and internal-Hom statements used for comparison. The constructions below include the injective replacement, the finite flat truncation and the evaluation signs needed here.

Constructing enough injectives

We first justify the injective resolutions used above. The abelian group ℚ/ℤ\mathbb Q/\mathbb Z is divisible and injective. Here is the extension argument. For a homomorphism defined on a subgroup A⊂BA\subset B, choose a maximal extension by the union-of-chains form of Zorn’s lemma. If its domain omits b∈Bb\in B, the set of integers nn with nbnb in that domain is an ideal mℤm\mathbb Z. If m=0m=0, assign any value to bb. If m>0m>0, divisibility supplies a value whose mm-fold multiple is the prescribed value on mbmb. In either case the rule extends consistently to the subgroup generated by the domain and bb, a contradiction. This proves injectivity. It is also a cogenerator: a nonzero element generates a finite or infinite cyclic subgroup admitting a character into ℚ/ℤ\mathbb Q/\mathbb Z nonzero on that element, and injectivity extends the character to the whole group.

Give E=Hom⁡ℤ(k,ℚ/ℤ)E=\operatorname{Hom}_{\mathbb Z}(k,\mathbb Q/\mathbb Z) its kk-action (sφ)(r)=φ(sr)(s\varphi)(r)=\varphi(sr). The mutually inverse operations of evaluation at 11 and χ↦[m↦(r↦χ(rm))]\chi\mapsto[m\mapsto(r\mapsto\chi(rm))] give

Hom⁡k(M,E)≃Hom⁡ℤ(M,ℚ/ℤ),M↪EHom⁡ℤ(M,ℚ/ℤ).(H1) \operatorname{Hom}_k(M,E) \simeq\operatorname{Hom}_{\mathbb Z}(M,\mathbb Q/\mathbb Z), \qquad M\hookrightarrow E^{\operatorname{Hom}_{\mathbb Z}(M,\mathbb Q/\mathbb Z)}. \qquad\text{(H1)}

The first functor is exact, so EE is injective. The second map has component m↦(r↦χ(rm))m\mapsto(r\mapsto\chi(rm)); the cogenerator property makes it injective. Products of injectives are injective, since a map into a product extends component by component. Thus (H1) constructs an injective module J(M)J(M) containing each MM.

For a point inclusion ix:{x}→Xi_x:\{x\}\to X, the skyscraper ix*Ji_{x*}J is injective when JJ is: maps into it are maps from the exact stalk functor into JJ. Consequently every sheaf FF has a monomorphism

F⟶∏x∈Xix*J(Fx).(H2) F\longrightarrow\prod_{x\in X}i_{x*}J(F_x). \qquad\text{(H2)}

The component at xx, followed by taking its stalk at xx, is the chosen injection of FxF_x. This proves monicity without assuming that stalks commute with infinite products. The product on the right is injective by the same componentwise extension argument. Repeating on cokernels gives an injective resolution of a sheaf. This construction also supplies the injective modules invoked in (B14).

There is an explicit replacement for a bounded-below complex CC, say Cn=0C^n=0 for n<an<a. Set In=0I^n=0 below aa and embed CaC^a into an injective IaI^a, defining faf^a. Suppose ff and dId_I have been defined through degree nn. Form the quotient sheaf

Qn+1=In⊕Cn+1im⁡[(u,c)↦(dIu+fnc,−dCc)],(u,c)∈In−1⊕Cn.(H3) Q^{n+1}= \frac{I^n\oplus C^{n+1}} {\operatorname{im}\bigl[(u,c)\mapsto (d_Iu+f^nc,-d_Cc)\bigr]}, \quad (u,c)\in I^{n-1}\oplus C^n. \qquad\text{(H3)}

Embed Qn+1Q^{n+1} in an injective In+1I^{n+1}. The classes of (u,0)(u,0) and (0,c)(0,c) define dInud_I^nu and fn+1cf^{n+1}c. The relations in (H3) give dI2=0d_I^2=0 and dIf=fdCd_If=fd_C. To prove that f:C→If:C\to I is a quasi-isomorphism, use the cone with differential (u,c)↦(dIu+fc,−dCc)(u,c)\mapsto(d_Iu+fc,-d_Cc). In degree nn, its first output component is the composite of the quotient to Qn+1Q^{n+1} and an injection. A cone cycle therefore lies in the preceding image; conversely that image consists of cycles because the differential squares to zero. At degree a−1a-1 exactness is the initial injection. This proves exactness of the whole cone on stalks and hence the claim.

We also need the homotopy property, not just existence. Every chain map t:A→It:A\to I from an acyclic complex to a bounded-below complex of injectives is null-homotopic. Construct hn:An→In−1h^n:A^n\to I^{n-1} in ascending degrees. Below the lower bound of II take zero. Once the previous step is fixed, the residual tn−dIhnt^n-d_Ih^n vanishes on ker⁡dAn=im⁡dAn−1\ker d_A^n=\operatorname{im}d_A^{n-1}, by the chain-map equation and the preceding homotopy equation. It therefore defines a map from im⁡dAn\operatorname{im}d_A^n into InI^n. Injectivity extends it to An+1A^{n+1}; take that extension as hn+1h^{n+1}. Induction gives

t=dIh+hdA.(H4) t=d_Ih+hd_A. \qquad\text{(H4)}

Applying this to shifts shows that the global Hom complex Hom⁡•(A,I)\operatorname{Hom}^\bullet(A,I) is acyclic. Thus precomposition with a quasi-isomorphism induces an isomorphism on homotopy classes of maps into II: its cone is acyclic, and the Hom cone has zero cohomology. A roof in the derived category with target II consequently has a unique representing homotopy class from its left endpoint. This proves that maps into II in the homotopy and derived categories agree. In particular, two such resolutions of the same object have comparison maps inverse up to homotopy, with their classes uniquely determined by the resolution maps.

Flat models of finite length

For sheaves F,GF,G, define F⊗GF\otimes G by sheafifying the presheaf tensor product. Define ℋom(F,G)(U)=Hom⁡(F|U,G|U)\mathcal Hom(F,G)(U)=\operatorname{Hom}(F|_U,G|_U). Compatible local maps glue uniquely, so this is a sheaf. A map out of a tensor is a locally bilinear map; currying that map and gluing gives

Hom⁡(A⊗B,C)≃Hom⁡(A,ℋom(B,C)).(H5) \operatorname{Hom}(A\otimes B,C) \simeq\operatorname{Hom}(A,\mathcal Hom(B,C)). \qquad\text{(H5)}

The same argument on every open gives the internal version. Tensor exactness is stalkwise exactness. In particular, the open-generator resolutions (P1)–(P2) have flat terms with free stalks and are K-flat by the proof there.

Let BB have cohomology in [a,b][a,b]. The elementary good truncations give a representative zero outside that interval: at the lower edge quotient by incoming boundaries, and at the upper edge take outgoing cycles. Resolve it by (P1), obtaining P→BP\to B with Pi=0P^i=0 for i>bi>b. Write Ci=coker⁡(Pi−1→Pi)C^i=\operatorname{coker}(P^{i-1}\to P^i). Exactness below aa gives 0→Ci→Pi+1→Ci+1→00\to C^i\to P^{i+1}\to C^{i+1}\to0 for i<ai<a. The stalk Cxa−dC^{a-d}_x is therefore a dd-th syzygy of CxaC^a_x through projective modules. It is projective, because every kk-module has projective dimension at most dd.

For clarity, the last assertion does not require these sheaves themselves to be projective. Given two projective presentations of a module, the fibre product of their epimorphisms projects onto each projective term. Splitting these two projections identifies it both with K⊕P′K\oplus P' and with K′⊕PK'\oplus P, where K,K′K,K' are the two kernels. Repeat this comparison through dd stages against a projective resolution of length dd. It identifies the proposed dd-th syzygy, after adding projective summands, with a projective module; hence that syzygy is a direct summand of a projective module and is projective. For d=0d=0, all modules are already projective.

Thus Ca−dC^{a-d} is a flat sheaf. Quotient the left end of PP to obtain

Pfin=[Ca−d⟶Pa−d+1⟶⋯⟶Pb],P⟶Pfin≃B.(H6) P_{\mathrm{fin}}= [\,C^{a-d}\longrightarrow P^{a-d+1}\longrightarrow\cdots \longrightarrow P^b\,], \qquad P\longrightarrow P_{\mathrm{fin}}\simeq B. \qquad\text{(H6)}

The displayed quotient map is a quasi-isomorphism, since the removed degrees are exact and its boundary quotient preserves the remaining cohomology. The final identification in (H6) is in the derived category; a chain map from the quotient to the originally chosen representative is not being presumed. All terms are flat. This gives finite flat models for every bounded coefficient complex, including arbitrary infinite coefficients. In particular, tensoring inputs in [a,b][a,b] and [c,e][c,e] produces cohomology in [a+c−d,b+e][a+c-d,b+e], by using (H6) for the first and a bounded representative for the second.

Currying on complexes and passage to the derived category

Use the cochain conventions

d(a⊗b)=dAa⊗b+(−1)|a|a⊗dBb,ℋomn(A,C)=∏pℋom(Ap,Cp+n),(dh)(a)=dCh(a)−(−1)|h|h(dAa).(H7) \begin{aligned} d(a\otimes b)&=d_Aa\otimes b+(-1)^{|a|}a\otimes d_Bb,\\ \mathcal Hom^n(A,C)&=\prod_p\mathcal Hom(A^p,C^{p+n}),\\ (dh)(a)&=d_Ch(a)-(-1)^{|h|}h(d_Aa). \end{aligned} \qquad\text{(H7)}

For bounded-above inputs and a bounded-below target all products contributing to any one degree here are finite. The curried map α:A→ℋom(B,C)\alpha:A\to\mathcal Hom(B,C) and its uncurried map are related by α̃(a⊗b)=α(a)(b)\widetilde\alpha(a\otimes b)=\alpha(a)(b), with no sign for this order of factors. For |α|=n|\alpha|=n and |a|=p|a|=p, both differentials evaluate to

dCα(a)(b)−(−1)nα(dAa)(b)−(−1)n+pα(a)(dBb).(H8) d_C\alpha(a)(b)-(-1)^n\alpha(d_Aa)(b) -(-1)^{n+p}\alpha(a)(d_Bb). \qquad\text{(H8)}

This proves the isomorphism of Hom complexes, including its signs, and the identical calculation on each open proves the internal isomorphism. The tensor associator and unit act by their ordinary formulas. The symmetry is a⊗b↦(−1)|a||b|b⊗aa\otimes b\mapsto(-1)^{|a||b|}b\otimes a; substituting (H7) checks that it commutes with the differential. The associativity and symmetry coherence identities follow on a pure tensor: regrouping preserves the order, and swapping a homogeneous entry past two others has exponent p(q+r)=pq+prp(q+r)=pq+pr. These identities hold on local sections and hence on sheaves.

Two elementary injectivity observations allow these formulas to compute derived functors. An injective sheaf remains injective on an open subset: extension by zero is its exact left adjoint, as is seen on stalks. Thus ℋom(−,I)\mathcal Hom(-,I) is exact when II is injective, by applying injectivity on every open. Also ℋom(F,I)\mathcal Hom(F,I) is injective for a flat FF, since (H5) expresses maps into it as the composite of exact tensoring with FF and exact contravariant Hom into II.

It follows that ℋom(P,I)\mathcal Hom(P,I), for bounded-above flat PP and bounded-below injective II, is a bounded-below complex of injective sheaves. In each degree only finitely many terms occur. For any acyclic AA, (H4) applied on every open shows that ℋom(A,I)\mathcal Hom(A,I) is acyclic: restriction preserves acyclicity of AA and the bounded-below complex of injectives II. Consequently ℋom(B,I)\mathcal Hom(B,I) computes Rℋom(B,C)R\mathcal Hom(B,C) when II resolves CC, and replacing a bounded-above BB by PP gives a quasi-isomorphic injective model.

These constructions are independent of the choices. For injective targets this follows from the unique homotopy classes proved after (H4). For tensor, resolve roofs by (P1) and use K-flatness: tensoring a quasi-isomorphism with a K-flat complex preserves its acyclic cone. Resolving the middle object of a roof gives a roof between flat models, so the localization on flat models computes the same derived category of bounded-above objects. All tensor comparison maps and their composites are therefore the ordinary ones on these roofs. Finite models (H6) can be substituted through their explicit quasi-isomorphisms.

Take bounded-above flat models PA,PBP_A,P_B and a bounded-below injective ICI_C. Currying (H8) identifies the two Hom complexes; its inner Hom is a bounded-below injective model. Passing to homotopy classes and then using (H4) proves

Hom⁡D(X)(A⊗LB,C)≃Hom⁡D(X)(A,Rℋom(B,C)),Rℋom(A⊗LB,C)≃Rℋom(A,Rℋom(B,C)).(H9) \begin{aligned} \operatorname{Hom}_{D(X)}(A\otimes^L B,C) &\simeq\operatorname{Hom}_{D(X)}(A,R\mathcal Hom(B,C)),\\ R\mathcal Hom(A\otimes^L B,C) &\simeq R\mathcal Hom(A,R\mathcal Hom(B,C)). \end{aligned} \qquad\text{(H9)}

The first line is computed by global sections of the corresponding Hom complexes; the second uses their equality as sheaf complexes. In the lesson all inputs are bounded, and (H6) supplies bounded tensor products. The target of an internal Hom in (H9) is allowed to be bounded below; no assertion that arbitrary internal Hom preserves boundedness is needed. For duality its stronger boundedness was proved using the special model (B14).

The same injective Hom model proves RHom⁡(A,C)=RΓ(X;Rℋom(A,C))R\operatorname{Hom}(A,C)=R\Gamma(X;R\mathcal Hom(A,C)): global sections of ℋom(PA,IC)\mathcal Hom(P_A,I_C) compute both sides. Here RHomR\operatorname{Hom} means the complex of global derived morphisms. Naturality in all the variables follows already from evaluation on local sections, so the isomorphisms identify actual adjunction maps. Inverse image is exact and sends the stalk of a tensor to the tensor of the stalks. The resulting natural monoidal isomorphism preserves flat models and the symmetry, associator and unit just checked. Thus pullbacks use these same conventions.

Evaluation, units and the double-dual identity

For the order ℋom(B,C)⊗B\mathcal Hom(B,C)\otimes B, evaluation is h⊗b↦h(b)h\otimes b\mapsto h(b). Formula (H7) verifies that it is a chain map. For the adjunction −⊗B⊣ℋom(B,−)-\otimes B\dashv\mathcal Hom(B,-), its unit sends aa to the map b↦a⊗bb\mapsto a\otimes b. Evaluating this unit returns a⊗ba\otimes b; applying the unit to a Hom and then postcomposing with evaluation returns hh. These are the two triangle identities on complexes. Replacing the inputs as above proves them for the derived adjunction as well. Likewise composition of homogeneous maps is (g,h)↦g∘h(g,h)\mapsto g\circ h, with differential d(gh)=(dg)h+(−1)|g|g(dh)d(gh)=(dg)h+(-1)^{|g|}g(dh). Evaluation of its transpose is g(h(a))g(h(a)), so it is associative and natural, with the identity maps as units.

Fix the bounded injective model WW for ωX\omega_X from (B14). If AA is a bounded representative, write DWA=ℋom(A,W)D_WA=\mathcal Hom(A,W). This complex is bounded and represents DXAD_XA. It can therefore be used again as the input of the same Hom model. The map (3) is represented by

ηA(a)(h)=(−1)|a||h|h(a),ηA:A⟶DWDWA.(H10) \eta_A(a)(h)=(-1)^{|a||h|}h(a), \qquad \eta_A:A\longrightarrow D_WD_WA. \qquad\text{(H10)}

The sign comes from moving aa past hh before evaluation. Explicitly, for |a|=p|a|=p, |h|=q|h|=q, the differential of this double-Hom expression is

(dηA(a))(h)=(−1)pqdWh(a)−(−1)p(−1)p(q+1)(dh)(a)=(−1)(p+1)qh(dAa)=ηA(dAa)(h).(H11) \begin{aligned} (d\eta_A(a))(h) &=(-1)^{pq}d_Wh(a) -(-1)^p(-1)^{p(q+1)}(dh)(a)\\ &=(-1)^{(p+1)q}h(d_Aa) =\eta_A(d_Aa)(h). \end{aligned} \qquad\text{(H11)}

Thus (H10) is a chain map. Although the ordinary tensor product with AA need not compute a derived tensor product, its evaluation gives the derived evaluation after flat replacement. Currying that pairing gives exactly (H10), by naturality of the replacement and (H8). No flatness of AA or of DWAD_WA is required here.

For a degree-zero chain map u:A→A′u:A\to A', evaluation on a,ha,h gives DWDW(u)ηA=ηA′uD_WD_W(u)\eta_A=\eta_{A'}u. Since DWD_W sends quasi-isomorphisms between bounded complexes to quasi-isomorphisms, this equality descends to all derived morphisms represented by roofs. Finally, for homogeneous h∈DWAh\in D_WA, precomposition with ηA\eta_A gives

(DWηA∘ηDWA)(h)(a)=(−1)|h||a|ηA(a)(h)=h(a).DXηA∘ηDXA=1DXA.(H12) \bigl(D_W\eta_A\circ\eta_{D_WA}\bigr)(h)(a) =(-1)^{|h||a|}\eta_A(a)(h) =h(a). \qquad D_X\eta_A\circ\eta_{D_XA}=1_{D_XA}. \qquad\text{(H12)}

Both signs have the same exponent and cancel. This is the double-evaluation identity used below. It holds without assuming that either evaluation map is an isomorphism. The comparison and evaluation-obstruction arguments therefore have the tensor–Hom adjunction, naturality and triangle identities they require. The next section identifies orientation transition maps and proves the constant relative-ball comparisons. Constructible biduality and weak-constructibility statements retain their separate prerequisite chains.

From local compact classes to orientation and duality

We identify the orientation object and its actual transition maps before using constructible biduality. The free reference is Schapira’s An Introduction to Sheaves on Grothendieck Topologies, Lemma 5.1.3 and Proposition 5.1.5, pp. 106–107. That treatment states the orientation identification and refers elsewhere for the differentiable-coordinate comparison. Here we prove the local-support comparison, the signs of coordinate changes and the normalization by trace. These arguments use the compact-support and algebraic constructions already supplied above.

Homotopy invariance from a proper interval

Let TT be a locally compact Hausdorff space occurring below, and put q:T×[0,1]→Tq:T\times[0,1]\to T. The map qq is proper. For a bounded complex of coefficient modules MM, the actual unit is an isomorphism

MT⟶Rq*MT×[0,1].(O1) M_T\longrightarrow Rq_*M_{T\times[0,1]}. \qquad\text{(O1)}

Here the auxiliary spaces include closed balls and closed interval cylinders. The compact-neighbourhood and fibre arguments (C1), (F1)–(F3) and (D5) apply to these spaces: their proofs require local compactness and the Hausdorff property, not the absence of boundary. One can check (O1) directly without an additional manifold theorem. Represent the sheaf input by injectives. The stalk of its direct-image complex is the filtered colimit of sections on inverse images of neighbourhoods of tt. Properness makes these neighbourhoods cofinal among neighbourhoods of the compact fibre. Compact-germ continuity identifies that complex with sections of the restricted injective resolution on the fibre. Injectives are flabby and c-soft, their restrictions to the compact fibre are c-soft, and hence these restricted terms are acyclic for its sections, by (C3). Thus the stalk computes the fibre cohomology. For MM a module, (B6), whose proof applies to any constant module, identifies the stalk of the unit with M→RΓ([0,1];M)M\to R\Gamma([0,1];M), an isomorphism. Finite truncation triangles give (O1) for bounded MM.

The two endpoint restrictions are inverses to (O1), because their composites with the unit are the identity. A continuous map g:S→Tg:S\to T gives the usual pullback of constant-coefficient cohomology by the unit for g−1⊣Rg*g^{-1}\dashv Rg_*; these maps compose by ordinary adjunction. Applying this to a homotopy g:S×[0,1]→Tg:S\times[0,1]\to T proves

g0*=g1*:RΓ(T;MT)⟶RΓ(S;MS).(O2) g_0^*=g_1^*:R\Gamma(T;M_T)\longrightarrow R\Gamma(S;M_S). \qquad\text{(O2)}

If gg is proper, the same equality holds for compactly supported cohomology: proper pullback preserves compact support, q*=q!q_*=q_!, and applying RΓcR\Gamma_c to (O1) and proper-image composition makes both endpoint maps inverse to the same isomorphism. For a homotopy of pairs (S,A)→(T,B)(S,A)\to(T,B), apply the ordinary argument simultaneously to the two restriction maps and take their fibres. It proves homotopy invariance of relative sheaf cohomology as well. We use this for open complements of a point; no properness of a contraction is needed for the ordinary or relative statement.

The actual point-to-compact comparison

Let V=ℝnV=\mathbb R^n, let x∈Vx\in V, and let M∈Db(k)M\in D^b(k). Contraction of VV to xx, using (O2), proves that the constant-section map and evaluation at xx are inverse isomorphisms. The compact calculation (B7)–(B9), with the coefficient projection formula, gives the other end of

RΓ(V;MV)≃M,RΓ{x}(V;MV)⟶RΓc(V;MV).(O3) R\Gamma(V;M_V)\simeq M, \qquad R\Gamma_{\{x\}}(V;M_V) \longrightarrow R\Gamma_c(V;M_V). \qquad\text{(O3)}

We prove that the displayed support-forgetting map is an isomorphism. Translate xx to zero. For a closed ball KrK_r of radius r>0r>0, inclusion V\Kr→V\{0}V\setminus K_r\to V\setminus\{0\} is a homotopy equivalence: choose a radius R>rR>r, and deform both spaces radially onto the sphere of radius RR. The formula sends v≠0v\ne0 at time tt to ((1−t)+tR/‖v‖)v((1-t)+tR/\lVert v\rVert)v; if ‖v‖>r\lVert v\rVert>r, its radius stays greater than rr. These retractions and inclusions supply inverse homotopy classes. Equation (O2) therefore makes restriction between the two complement cohomologies an isomorphism.

The closed-support localization triangles have the same middle term RΓ(V;MV)R\Gamma(V;M_V). The complement isomorphism just proved gives

RΓ{0}(V;MV)→∼RΓKr(V;MV)⟶RΓc(V;MV).(O4) R\Gamma_{\{0\}}(V;M_V) \xrightarrow{\sim} R\Gamma_{K_r}(V;M_V) \longrightarrow R\Gamma_c(V;M_V). \qquad\text{(O4)}

Closed balls are cofinal among compact subsets of VV. Compact sections of an injective resolution are the filtered union of its sections with compact closed support. Filtered colimits are exact, so passage to that union proves that the second map in (O3) is an isomorphism too. This establishes the actual map, naturally in MM, rather than choosing an abstract identification of the groups. For n=0n=0, all spaces are a point and the assertion is immediate.

Excision for point supports follows directly by restriction and extension by zero: a section supported at a point inside an open neighbourhood is zero near its boundary, so these operations are inverse; the same calculation on injective restrictions proves the derived assertion. Consequently (O3) holds on every open coordinate ball, for any point in it. If two such balls satisfy U′⊂UU'\subset U and contain xx, the maps from the common point-supported complex to their compact-section complexes commute with extension from U′U' to UU. All three comparisons are isomorphisms. In particular, extension between the compact-cohomology groups of nested coordinate balls is an isomorphism with a specified geometric normalization.

The integral sign line

Work first on a component of dimension nn. Define oXℤo_X^{\mathbb Z} as the sheafification of the presheaf

U⟼Hom⁡ℤ(Hcn(U;ℤ),ℤ),(O5) U\longmapsto \operatorname{Hom}_{\mathbb Z}(H_c^n(U;\mathbb Z),\mathbb Z), \qquad\text{(O5)}

whose restrictions are dual to extension of compact supports. On a coordinate ball the compact group is free of rank one, by (B9) over ℤ\mathbb Z. The preceding point-support comparison makes all restrictions to smaller coordinate balls isomorphisms. A chart itself can be taken to be a ball; its chosen generator therefore trivializes (O5) on a basis inside that chart. Hence the sheaf is locally constant of rank one. On a connected overlap, any two integral bases differ by +1+1 or −1-1, since these are the units of ℤ\mathbb Z; on a disconnected overlap this assertion holds locally.

Set oX=kX⊗ℤoXℤo_X=k_X\otimes_{\mathbb Z}o_X^{\mathbb Z}. The interval boundary calculation is natural in its coefficient module: changing ℤ\mathbb Z to kk sends the endpoint difference generator to its kk-valued counterpart. Iteration in ordered coordinates gives the same assertion on boxes. Excision and (O3) then give it on coordinate balls with their restriction maps. Thus sheafifying U↦Hom⁡k(Hcn(U;k),k)U\mapsto\operatorname{Hom}_k(H_c^n(U;k),k) gives precisely oXo_X, including rings of positive characteristic. No claim that dualization commutes with arbitrary change of coefficients is used; the local groups in this comparison are free of rank one.

There is a canonical square pairing on the integral line: for either generator ee, send e⊗ee\otimes e to 11. Changing both signs preserves that rule, so it glues. After change of coefficients it gives

oX⊗koX≃kX,oX∨:=ℋomk(oX,kX)≃oX.(O6) o_X\otimes_k o_X\simeq k_X, \qquad o_X^\vee:=\mathcal Hom_k(o_X,k_X)\simeq o_X. \qquad\text{(O6)}

These tensors and Homs are already exact locally because the line is free. The self-duality uses the integral sign structure; it is not asserted for every rank-one kk-local system. The dual line (oX)x∨(o_X)_x^\vee identifies canonically with H{x}n(X;k)H^n_{\{x\}}(X;k), by (O3) and the evaluation pairing of that free local group with its dual.

Coordinate changes and their signs

The ordered coordinate generator of Hcn(ℝn;ℤ)H_c^n(\mathbb R^n;\mathbb Z) is the external product of the increasing-interval classes from (B7). In one dimension a positive rescaling preserves the endpoint order, so it preserves the generator. Reflection exchanges the endpoints and sends their difference to its negative. A swap of two coordinates exchanges two degree-one factors in (B8); the tensor symmetry (H7) therefore contributes −1-1. These assertions concern the actual pullback maps and the same ordered product generator.

An elementary shear xi↦xi+txjx_i\mapsto x_i+t x_j, i≠ji\ne j, is properly homotopic to the identity as tt varies over a compact interval. Indeed its inverse is xi↦xi−txjx_i\mapsto x_i-tx_j, which sends a bounded set into a uniformly bounded set; the inverse image of a compact set is also closed and is therefore compact in ℝn×[0,1]\mathbb R^n\times[0,1]. Proper homotopy invariance from (O2) makes its action the identity. Gaussian elimination expresses an invertible real matrix using these shears, swaps and nonzero rescalings: choose a nonzero pivot in each remaining column, swap it into place, rescale it and clear its column, then continue on the remaining square block. Multiplying the actions of these elementary operations proves

A*=sgn⁡(det⁡A)on Hcn(ℝn;ℤ),A∈GL⁡n(ℝ).(O7) A^*=\operatorname{sgn}(\det A) \quad\text{on }H_c^n(\mathbb R^n;\mathbb Z), \qquad A\in\operatorname{GL}_n(\mathbb R). \qquad\text{(O7)}

Point-support comparison gives the same formula on local cohomology at zero.

Now let hh be a C1C^1 change of coordinates fixing zero, with A=Dh(0)A=Dh(0) invertible. Shrink a ball about zero so that differentiability gives ‖A−1h(v)−v‖<12‖v‖\lVert A^{-1}h(v)-v\rVert<\tfrac12\lVert v\rVert for v≠0v\ne0 in it. For

ht(v)=(1−t)Av+th(v),‖A−1ht(v)‖>12‖v‖(v≠0),(O8) h_t(v)=(1-t)Av+t h(v), \qquad \lVert A^{-1}h_t(v)\rVert>\tfrac12\lVert v\rVert \quad(v\ne0), \qquad\text{(O8)}

we have a homotopy of maps of pairs into (ℝn,ℝn\{0})(\mathbb R^n,\mathbb R^n\setminus\{0\}). Intermediate maps need only avoid zero away from zero; their invertibility is unnecessary. Relative homotopy invariance and excision identify the local-cohomology actions of hh and AA. Thus the transition in (O5), and its dual, is multiplication by sgn⁡det⁡Dh(0)\operatorname{sgn}\det Dh(0). Repeat at every point of an overlap. The determinant is continuous and never zero, so its sign is locally constant. This proves that the sign line just constructed is the usual orientation sheaf for the real analytic manifolds of this lesson.

Identifying the dualizing object and its trace

The projection estimate (B12), applied locally to the map to a point, shows that ωX\omega_X has only degree −n-n on an nn-dimensional component. Put L=ℋ−n(ωX)L=\mathcal H^{-n}(\omega_X). Good truncation gives the canonical isomorphism ωX≃L[n]\omega_X\simeq L[n]. On any coordinate ball UU, adjunction and (B9) identify

Γ(U;L)=H−n(U;ωX)=Hom⁡D(k)(RΓc(U;k),k[−n])=Hom⁡k(Hcn(U;k),k).ωX≃oX[n].(O9) \begin{aligned} \Gamma(U;L) &=H^{-n}(U;\omega_X)\\ &=\operatorname{Hom}_{D(k)}(R\Gamma_c(U;k),k[-n])\\ &=\operatorname{Hom}_k(H_c^n(U;k),k). \end{aligned} \qquad \omega_X\simeq o_X[n]. \qquad\text{(O9)}

The maps as UU shrinks are dual to compact-support extension, because the adjunction was constructed from that extension on open generators. Thus (O9) identifies the actual sheaf LL with (O5) after change of coefficients, not merely their stalk ranks. It also identifies the actual maps on overlaps with the signs proved in (O8). On manifolds with components of different dimensions the formula is read componentwise, with the standing uniform bound.

The adjunction specifies the trace. Let cUc_U be an integral compact generator, extended to kk, and let eUe_U be its dual local orientation section. In the order with the dualizing section first, the pairing obtained from compact-support projection and the trace is

RΓ(U;ωU)⊗kLRΓc(U;k)⟶k,eU⊗cU⟼1.(O10) R\Gamma(U;\omega_U)\otimes_k^L R\Gamma_c(U;k) \longrightarrow k, \qquad e_U\otimes c_U\longmapsto1. \qquad\text{(O10)}

Indeed (O9) is the adjunction map taking that section to its functional on compact cohomology. Its inverse reconstructs exactly the same pairing by the counit. This proves the normalization, including compatibility with smaller balls and extension of compact supports. Both generators change sign when the orientation changes, so the pairing is unchanged. Their cohomological degrees are −n-n and nn; reversing their order uses the Koszul sign (−1)n(-1)^n. Formula (O10) fixes the order and avoids concealing that sign in an unnamed identification of shifted lines.

Constant coefficients on relative balls

Let ix:{x}→Xi_x:\{x\}\to X, and suppose a bounded coefficient complex is constant, MUM_U, on a coordinate ball around xx. Closed-support adjunction (B4), excision and (O3) give the natural identification

ix!MU≃RΓ{x}(U;MU)→∼RΓc(U;MU)≃M⊗k(oX)x∨[−n].(O11) i_x^!M_U \simeq R\Gamma_{\{x\}}(U;M_U) \xrightarrow{\sim} R\Gamma_c(U;M_U) \simeq M\otimes_k (o_X)_x^\vee[-n]. \qquad\text{(O11)}

The final expression puts the coefficient first. It uses the projection formula and the symmetry fixed in (H7); moving a degree-pp coefficient past the degree-nn compact class has sign (−1)np(-1)^{np}. The line is free, so no Tor correction or finite-rank assumption on MM is needed. All maps are natural in MM and in shrinking balls, by the actual support maps in (O3)–(O4).

There is also a closed-ball description with its boundary map retained. For a closed coordinate ball B¯\overline B about xx, let BB be its interior and SS its boundary. Write relative sheaf cohomology as the fibre of restriction. The exact sequence of constant sheaves and their extensions gives

RΓ(B¯,S;M):=Cone⁡(RΓ(B¯;M)⟶RΓ(S;M))[−1]≃RΓc(B;M)≃M⊗k(oX)x∨[−n].(O12) \begin{aligned} R\Gamma(\overline B,S;M) &:=\operatorname{Cone}\bigl(R\Gamma(\overline B;M) \longrightarrow R\Gamma(S;M)\bigr)[-1]\\ &\simeq R\Gamma_c(B;M) \simeq M\otimes_k(o_X)_x^\vee[-n]. \end{aligned} \qquad\text{(O12)}

To verify the first isomorphism, apply sections on the compact B¯\overline B to 0→j!MB→MB¯→i*MS→00\to j_!M_B\to M_{\overline B}\to i_*M_S\to0, degree by degree and then to its bounded total complex. Stalks prove this sequence exact. Open proper-image composition identifies its first derived-section term with RΓc(B;M)R\Gamma_c(B;M). Thus the displayed relative map is the localization boundary with the cone convention already fixed in (P2). For n=1n=1 it is the endpoint difference (s,t)↦t−s(s,t)\mapsto t-s of (B7). For n=0n=0 the boundary is empty and the result is MM in degree zero. Formula (O12) is a statement about relative sheaf cohomology; it does not silently import a comparison with a chosen singular or cellular cochain model.

These results establish the orientation and constant local-support inputs, including arbitrary bounded constant coefficients, transition signs and trace normalization. The companion small-ball lesson uses these foundations to prove the scalar endpoint, closed-support and compact-support comparisons. That lesson also proves the proper-image microsupport estimate and closed-cutoff sheaf argument. The companion cotangent lesson now supplies the transport argument and a direct analytic-curve proof of critical-value finiteness. Its subanalytic foundations and the constructibility/microsupport criterion with its microlocal proofs remain unfinished. Compact finiteness, local dual pairings, constructible biduality and the full geometric prerequisite chains still require their exact proofs.

Transporting a pairing through a map

There are two comparisons which require no constructibility:

uE:f!DXE→∼DYf−1E,bB:Rf*DYB→∼DXRf!B.(4) u_E:f^!D_XE\xrightarrow{\sim}D_Yf^{-1}E, \qquad b_B:Rf_*D_YB\xrightarrow{\sim}D_XRf_!B. \qquad\text{(4)}

Here is a proof of both from the operation contracts. It also specifies the actual maps.

For a bounded test object TT on YY, compute

Hom⁡Y(T,f!DXE)≃Hom⁡X(Rf!T,DXE)≃Hom⁡X(Rf!T⊗E,ωX)≃Hom⁡X(Rf!(T⊗f−1E),ωX)≃Hom⁡Y(T⊗f−1E,ωY)≃Hom⁡Y(T,DYf−1E).(5) \begin{aligned} \operatorname{Hom}_Y(T,f^!D_XE) &\simeq\operatorname{Hom}_X(Rf_!T,D_XE)\\ &\simeq\operatorname{Hom}_X(Rf_!T\otimes E,\omega_X)\\ &\simeq\operatorname{Hom}_X(Rf_!(T\otimes f^{-1}E),\omega_X)\\ &\simeq\operatorname{Hom}_Y(T\otimes f^{-1}E,\omega_Y)\\ &\simeq\operatorname{Hom}_Y(T,D_Yf^{-1}E). \end{aligned} \qquad\text{(5)}

Every step is natural in TT and EE. The first and fourth use proper-support adjunction, the middle step uses (2), and the remaining steps curry a pairing. Yoneda therefore gives the first isomorphism in (4). Its uncurried pairing is

f!DXE⊗f−1E⟶f!(DXE⊗E)⟶f!ωX=ωY.(6) f^!D_XE\otimes f^{-1}E \longrightarrow f^!(D_XE\otimes E) \longrightarrow f^!\omega_X=\omega_Y. \qquad\text{(6)}

The first arrow in (6) is obtained by taking the right-adjunction mate of projection followed by the counit Rf!f!DXE→DXERf_!f^!D_XE\to D_XE. This describes it without assuming an internal exceptional-Hom theorem. Tracing the identity map through (5) yields precisely (6), so (5) proves invertibility of the designated comparison.

For a bounded test object AA on XX, a second computation gives

Hom⁡X(A,Rf*DYB)≃Hom⁡Y(f−1A,DYB)≃Hom⁡Y(f−1A⊗B,ωY)≃Hom⁡X(Rf!(f−1A⊗B),ωX)≃Hom⁡X(A⊗Rf!B,ωX)≃Hom⁡X(A,DXRf!B).(7) \begin{aligned} \operatorname{Hom}_X(A,Rf_*D_YB) &\simeq\operatorname{Hom}_Y(f^{-1}A,D_YB)\\ &\simeq\operatorname{Hom}_Y(f^{-1}A\otimes B,\omega_Y)\\ &\simeq\operatorname{Hom}_X(Rf_!(f^{-1}A\otimes B),\omega_X)\\ &\simeq\operatorname{Hom}_X(A\otimes Rf_!B,\omega_X)\\ &\simeq\operatorname{Hom}_X(A,D_XRf_!B). \end{aligned} \qquad\text{(7)}

Ordinary adjunction enters in the first step; proper-support adjunction enters in the third. Naturality and Yoneda again give an isomorphism. In this case its pairing is

Rf*DYB⊗Rf!B≃Rf!(f−1Rf*DYB⊗B)⟶Rf!(DYB⊗B)⟶Rf!ωY⟶ωX.(8) \begin{aligned} Rf_*D_YB\otimes Rf_!B &\simeq Rf_!(f^{-1}Rf_*D_YB\otimes B)\\ &\longrightarrow Rf_!(D_YB\otimes B) \longrightarrow Rf_!\omega_Y\longrightarrow\omega_X. \end{aligned} \qquad\text{(8)}

The first counit in (8) is for ordinary inverse/direct image; the last arrow is the proper-support trace. This checks that the isomorphism is bBb_B, with those counits and that trace. Finite global dimension keeps the tensors bounded, and the standing operation bounds keep the remaining objects in the bounded categories. Neither proof uses perfection or properness of ff.

Where a reversed comparison can fail

We now define the two reversed maps for arbitrary bounded inputs. Let

AE=f−1DXE,δE=uDXE∘f!ηE:f!E⟶DYAE,HB=Rf!DYB,γB=bDYB∘Rf*ηB:Rf*B⟶DXHB.(9) \begin{aligned} A_E&=f^{-1}D_XE,& \delta_E&=u_{D_XE}\circ f^!\eta_E:f^!E\longrightarrow D_YA_E,\\ H_B&=Rf_!D_YB,& \gamma_B&=b_{D_YB}\circ Rf_*\eta_B:Rf_*B\longrightarrow D_XH_B. \end{aligned} \qquad\text{(9)}

These are defined even when evaluation is not invertible. Dualization reverses their directions. Thus the composites

cE:AE→ηAEDYDYAE→DYδEDYf!E,eB:HB→ηHBDXDXHB→DXγBDXRf*B(10) \begin{aligned} c_E&:A_E\xrightarrow{\eta_{A_E}}D_YD_YA_E \xrightarrow{D_Y\delta_E}D_Yf^!E,\\ e_B&:H_B\xrightarrow{\eta_{H_B}}D_XD_XH_B \xrightarrow{D_X\gamma_B}D_XRf_*B \end{aligned} \qquad\text{(10)}

have the advertised inverse- and direct-image types. We use (10) as the explicit definition of these dual comparisons. Equivalently, uncurry them to the evaluation pairings

f−1DXE⊗f!E⟶f!(DXE⊗E)⟶ωY,Rf!DYB⊗Rf*B⟶Rf!(DYB⊗f−1Rf*B)⟶Rf!ωY⟶ωX.(11) \begin{aligned} f^{-1}D_XE\otimes f^!E&\longrightarrow f^!(D_XE\otimes E)\longrightarrow\omega_Y,\\ Rf_!D_YB\otimes Rf_*B&\longrightarrow Rf_!(D_YB\otimes f^{-1}Rf_*B) \longrightarrow Rf_!\omega_Y\longrightarrow\omega_X. \end{aligned} \qquad\text{(11)}

To check this equivalence, expand δE\delta_E using (6), or γB\gamma_B using (8). The composite of double evaluation with evaluation is evaluation again: explicitly, DηQ∘ηDQ=1DQD\eta_Q\circ\eta_{DQ}=1_{DQ}. The chain calculation (H12) proves this identity with its two cancelling Koszul signs; the tensor–Hom unit and counit were checked immediately before (H10). In the second line of (11), ordinary adjunction additionally cancels f−1⊣Rf*f^{-1}\dashv Rf_*’s unit–counit pair. In the first line, the mate defining (6) cancels the corresponding Rf!⊣f!Rf_!\dashv f^! pair. The remaining pairings are exactly (11). Thus the definitions retain the canonical evaluation and trace, not just the isomorphism classes of their endpoints.

Evaluation criterion. If ηE\eta_E is invertible, then cEc_E is invertible if and only if ηAE\eta_{A_E} is invertible. If ηB\eta_B is invertible, then eBe_B is invertible if and only if ηHB\eta_{H_B} is invertible.

Proof. Under the first assumption, both factors defining δE\delta_E in (9) are isomorphisms. Therefore DYδED_Y\delta_E is an isomorphism, and the first factorization in (10) proves the first assertion. Under the second assumption γB\gamma_B, hence DXγBD_X\gamma_B, is an isomorphism. The second factorization proves the other assertion. No conservativity of dualization is used. ▫\square

This criterion isolates the obstruction at a specific object: the pulled-back dual for inverse image, and the proper-support image of the dual for direct image. It does not attempt to dualize a noninvertible map and infer invertibility of that map.

Applying the criterion to constructible coefficients

An ℝ\mathbb R-constructible complex here is weakly subanalytic constructible with perfect stalk complexes. A perfect complex has a bounded representative by finitely generated projective kk-modules. We retain the full ring kk; no Noetherian hypothesis is added.

The geometric inputs are constructible biduality and dual stability, analytic inverse-image stability, and proper direct-image stability on the closed coefficient support. Their current in-course arguments are identified in the prerequisite ledger below. They are precisely the inputs needed to apply the formal criterion.

If ff is analytic and EE is constructible, then DXED_XE and AE=f−1DXEA_E=f^{-1}D_XE are constructible. Constructible biduality makes both ηE\eta_E and ηAE\eta_{A_E} invertible. Consequently

f−1DXE→cE,∼DYf!E.(12) f^{-1}D_XE\xrightarrow{c_E,\ \sim}D_Yf^!E. \qquad\text{(12)}

For direct image, assume explicitly that

B∈Dℝ-cb(kY),HB=Rf!DYB∈Dℝ-cb(kX).(13) B\in D^b_{\mathbb R\text{-}c}(k_Y),\qquad H_B=Rf_!D_YB\in D^b_{\mathbb R\text{-}c}(k_X). \qquad\text{(13)}

Evaluation of BB makes γB:Rf*B→DXHB\gamma_B:Rf_*B\to D_XH_B invertible. Since HBH_B is constructible, its dual is constructible. This proves constructibility of Rf*BRf_*B before invoking any biduality for that image. Evaluation of HBH_B now gives

Rf!DYB→eB,∼DXRf*B,Rf*B∈Dℝ-cb(kX).(14) Rf_!D_YB\xrightarrow{e_B,\ \sim}D_XRf_*B, \qquad Rf_*B\in D^b_{\mathbb R\text{-}c}(k_X). \qquad\text{(14)}

Properness on the closed support of BB is one sufficient condition for (13). The closed supports of a constructible object and its dual agree: each vanishes on an open set if and only if its dual does, by local duality and biduality. Thus proper direct-image stability applies to DYBD_YB. A nonproper map can also satisfy (13); the actual condition concerns its coefficient image.

Infinite twists: applying Hom instead of biduality

Let M∈Db(kX)M\in D^b(k_X) be locally constant. Its stalks may be arbitrary modules. Locally, the whole derived object is isomorphic to LUL_U for some bounded coefficient complex LL. Do not replace this assumption by global constancy or by a chosen extension to an ambient boundary.

The free paper by Andreas Hohl and Pierre Schapira, Unusual functorialities for weakly constructible sheaves, version 2, proves the comparisons used here in Theorems 4.3 and 4.8 and the resulting duality statements in Corollaries 4.10 and 4.11. The following arguments use those results through their current in-course comparison proofs. The free paper’s references to earlier foundations do not themselves supply the missing programme proofs.

Write ℋM(Q)=Rℋom(M,Q)\mathcal H_M(Q)=R\mathcal Hom(M,Q). Tensor–Hom currying gives a natural identity

DX(Q⊗M)=ℋM(DXQ).(15) D_X(Q\otimes M)=\mathcal H_M(D_XQ). \qquad\text{(15)}

This is valid for arbitrary bounded MM. It does not identify ℋM\mathcal H_M with tensoring by DMD M.

Inverse-image twist. Let f:Y→Xf:Y\to X be analytic. For constructible EE, the canonical comparison is invertible:

f−1DX(E⊗M)→∼DYf!(E⊗M).(16) f^{-1}D_X(E\otimes M)\xrightarrow{\sim}D_Yf^!(E\otimes M). \qquad\text{(16)}

Proof. Start with f−1ℋM(DXE)f^{-1}\mathcal H_M(D_XE). The inverse-Hom comparison for a locally constant first argument and a weakly constructible second argument identifies it with ℋf−1M(f−1DXE)\mathcal H_{f^{-1}M}(f^{-1}D_XE). Apply this Hom functor to (12); its value is ℋf−1M(DYf!E)=DY(f!E⊗f−1M)\mathcal H_{f^{-1}M}(D_Yf^!E)=D_Y(f^!E\otimes f^{-1}M). Finally the exceptional-tensor comparison f!E⊗f−1M→∼f!(E⊗M)(17) f^!E\otimes f^{-1}M\xrightarrow{\sim}f^!(E\otimes M) \qquad\text{(17)} identifies this with the right side of (16), after dualizing (17) and using its inverse. All three steps are isomorphisms. The inverse-Hom step applies because DXED_XE is constructible; (17) applies because EE is weakly constructible and MM is locally constant. Currying (11) and then evaluating MM identifies their composite with the canonical comparison. The only finite biduality used was (12) for EE, before adding MM. ▫\square

Direct-image twist. Let Y∞=(Y,Ŷ)Y_\infty=(Y,\widehat Y) and X∞=(X,X̂)X_\infty=(X,\widehat X) be b-analytic pairs: their open subsets are relatively compact and subanalytic in the ambient analytic manifolds. Let f:Y→Xf:Y\to X be analytic with graph subanalytic in Ŷ×X̂\widehat Y\times\widehat X, so that it is a morphism of these b-analytic pairs. Suppose the zero extension jY!Bj_{Y!}B is constructible with perfect stalks. Then

Rf!DY(B⊗f−1M)→∼DX(Rf*B⊗M).(18) Rf_!D_Y(B\otimes f^{-1}M) \xrightarrow{\sim}D_X(Rf_*B\otimes M). \qquad\text{(18)}

Proof. First check the untwisted input to (14). Put B̂=jY!B\widehat B=j_{Y!}B. Open restriction of duality gives DYB=jY−1DŶB̂D_YB=j_Y^{-1}D_{\widehat Y}\widehat B. Its zero extension is therefore kY⊗DŶB̂k_Y\otimes D_{\widehat Y}\widehat B, where kYk_Y is the ambient zero extension of the constant sheaf. Dual stability and the perfect cutoff make this constructible with perfect stalks. Thus DYBD_YB is constructible up to infinity as well.

The b-analytic graph argument for image stability now makes HB=Rf!DYBH_B=Rf_!D_YB constructible with perfect stalks. Compactness here concerns the closed supports of the ambient graph coefficients: they lie in the compact closure of the graph. Projection is proper on those actual supports. It does not assert that ff is proper on YY. We have verified both conditions in (13), so (14) is available for the untwisted BB.

Use (15) on the source. The compact-graph Hom comparison then gives Rf!DY(B⊗f−1M)=Rf!ℋf−1M(DYB)→∼ℋM(HB)→ℋM(eB),∼ℋM(DXRf*B)=DX(Rf*B⊗M).(19) \begin{aligned} Rf_!D_Y(B\otimes f^{-1}M) &=Rf_!\mathcal H_{f^{-1}M}(D_YB)\\ &\xrightarrow{\sim}\mathcal H_M(H_B) \xrightarrow{\mathcal H_M(e_B),\ \sim}\mathcal H_M(D_XRf_*B)\\ &=D_X(Rf_*B\otimes M). \end{aligned} \qquad\text{(19)}

The Hom comparison requires weak constructibility up to infinity of DYBD_YB, already checked. It is local on XX, so local constant models for MM suffice; no ambient extension of MM is imposed. Its map is defined by evaluation and the proper-support trace, and composing it with the pairing defining eBe_B gives (18). Neither B⊗f−1MB\otimes f^{-1}M nor Rf*B⊗MRf_*B\otimes M was assumed reflexive. ▫\square

Eight calculations and tests

1. Finite pairings and the object to evaluate

Difficulty: Intermediate. Let YY be a finite discrete space with coefficients PyP_y that are perfect complexes, and map it to a point. Describe bBb_B, eBe_B, and the evaluations that justify them. Include the empty space.

Solution. Both images are the finite direct sum of the PyP_y. The source dual has coefficient Py∨P_y^\vee. A finite direct sum commutes with coefficient duality, so the pairing is ∑yev⁡Py\sum_y\operatorname{ev}_{P_y}, with no cross term between different points. The map bBb_B already follows from (7). For eBe_B, evaluate BB, obtaining γB\gamma_B, and then evaluate HB=⨁yPy∨H_B=\bigoplus_yP_y^\vee. Finite projective evaluation makes both isomorphisms. Formula (10) gives the displayed sum pairing. For the empty space every object is zero and all comparison maps are the unique isomorphism of zero objects. This also shows why evaluating an ordinary image before proving its finiteness is unnecessary.

2. A discrete source whose ordinary image is too large

Difficulty: Advanced. Work over a field. Embed S={3n:n∈ℤ}S=\{3n:n\in\mathbb Z\} as a closed subset of the real line, let B=i*kSB=i_*k_S, and map the line to a point. Decide which maps in (4) and (10) remain invertible.

Solution. The points of SS and the intervening open intervals give a locally finite subanalytic stratification; the restrictions have coefficients kk or zero. Thus BB has perfect stalks and is constructible. Closed-embedding duality gives DℝB=i*kSD_{\mathbb R}B=i_*k_S, because the intrinsic dualizing object of each isolated point is kk in degree zero. Ordinary sections on SS form a product and compact sections a direct sum. Both are exact functors on a discrete space, so Rf*B=kℤ,HB=k(ℤ)=:V.(20) Rf_*B=k^{\mathbb Z},\qquad H_B=k^{(\mathbb Z)}=:V. \qquad\text{(20)} Here γB\gamma_B identifies the product with V*V^*, while eBe_B becomes the canonical evaluation V→V**V\to V^{**}. This map is injective because coordinate functionals separate finite families. It is not surjective: the constant sequence 11 gives a nonzero class in kℤ/k(ℤ)k^{\mathbb Z}/k^{(\mathbb Z)}. Extend that class to a basis and choose a linear functional nonzero on it. Composing with the quotient gives a nonzero functional on the product vanishing on every finite family. A finite coordinate functional with that vanishing property must be zero, so this functional is not an evaluation at an element of VV. The general map bBb_B remains an isomorphism; both its sides are kℤk^{\mathbb Z}. The failed evaluation is exactly ηHB\eta_{H_B}. Neither vector space in (20) is finite dimensional.

3. The orientation line at a point

Difficulty: Intermediate. For i:{p}↪Xdi:\{p\}\hookrightarrow X^d, let PXP_X be a constant perfect complex and write L=orX,pL=\mathrm{or}_{X,p}. Calculate (12) without choosing an orientation.

Solution. Finite projective duality gives i−1DXPX=P∨⊗L[d]i^{-1}D_XP_X=P^\vee\otimes L[d]. The relative-ball calculation (O11)–(O12) gives i!PX=P⊗L∨[−d]i^!P_X=P\otimes L^\vee[-d]: the relative constant-coefficient cohomology has its generator in degree dd, and change of orientation acts by the dual orientation line. Dualizing on the point yields P∨⊗L[d]P^\vee\otimes L[d], the same object. Its pairing with P⊗L∨[−d]P\otimes L^\vee[-d] is finite-projective evaluation times L⊗L∨→kL\otimes L^\vee\to k, with the complex symmetry convention of (1)–(3). For P=kP=k, both sides of (12) have their nonzero cohomology in degree −d-d. Replacing i!i^! by ordinary restriction would lose the codimension shift.

4. Torsion tests the derived degree

Difficulty: Intermediate. In the preceding calculation take k=ℤk=\mathbb Z, P=ℤ/mP=\mathbb Z/m with m>1m>1, and trivialize the local orientation. Determine the nonzero degree of (12).

Solution. A representative of PP is [ℤ→mℤ][\mathbb Z\xrightarrow{m}\mathbb Z] in degrees −1,0-1,0. Its dual has cokernel ℤ/m\mathbb Z/m in degree 11, hence P∨=P[−1]P^\vee=P[-1]. Both sides of (12) are P[d−1]P[d-1], with nonzero cohomology in degree 1−d1-d. Indeed, i!PX=P[−d]i^!P_X=P[-d], so its point dual is P∨[d]P^\vee[d]. Ordinary degree-zero Hom would give Hom⁡ℤ(ℤ/m,ℤ)=0\operatorname{Hom}_{\mathbb Z}(\mathbb Z/m,\mathbb Z)=0 and lose the actual dual object. The finite free resolution verifies perfection and retains its Ext degree.

5. Nonproper integration and its shift

Difficulty: Introductory. Give ℝ\mathbb R its increasing orientation and let a:ℝ→{pt}a:\mathbb R\to\{\mathrm{pt}\}. For a perfect constant complex PℝP_{\mathbb R}, calculate the two direct-image comparisons.

Solution. Contractibility, with the homotopy comparison (O1)–(O3), gives Ra*Pℝ=PRa_*P_{\mathbb R}=P. The endpoint-difference calculation (B7), followed by coefficient projection, gives Ra!Pℝ=P[−1]Ra_!P_{\mathbb R}=P[-1] for every bounded PP. Since DℝPℝ=(P∨)ℝ[1]D_{\mathbb R}P_{\mathbb R}=(P^\vee)_{\mathbb R}[1], the two maps have endpoints Ra*DℝPℝ=P∨[1]=DptRa!Pℝ,Ra!DℝPℝ=P∨=DptRa*Pℝ.(21) Ra_*D_{\mathbb R}P_{\mathbb R}=P^\vee[1]=D_{\mathrm{pt}}Ra_!P_{\mathbb R}, \quad Ra_!D_{\mathbb R}P_{\mathbb R}=P^\vee=D_{\mathrm{pt}}Ra_*P_{\mathbb R}. \qquad\text{(21)} The oriented trace sends the compact fundamental cohomology generator to 11; the remaining pairing is evaluation on PP. The two constructibility inputs of (13) hold, although aa is nonproper.

6. Endpoints distinguish the two images

Difficulty: Intermediate. For j:U=(−2,2)↪ℝj:U=(-2,2)\hookrightarrow\mathbb R, compare ordinary and proper-support images of kUk_U, and calculate their duality maps at an endpoint.

Solution. On a small neighborhood of an endpoint, intersection with UU is a contractible half-interval. Thus Rj*kU=k[−2,2]Rj_*k_U=k_{[-2,2]} with endpoint stalk kk, while j!kU=k(−2,2)j_!k_U=k_{(-2,2)} has endpoint stalk zero. On the oriented source, DUkU=kU[1]D_Uk_U=k_U[1]. Formula (4) gives Dℝk(−2,2)=k[−2,2][1]D_{\mathbb R}k_{(-2,2)}=k_{[-2,2]}[1], and (14) gives Dℝk[−2,2]=k(−2,2)[1]D_{\mathbb R}k_{[-2,2]}=k_{(-2,2)}[1]. The first comparison has endpoint stalk k[1]k[1]; the reversed comparison has zero endpoint stalk. The inputs in (13) are constructible in both cases. The difference concerns boundary stalks, not invertibility of either comparison.

7. A twist can preserve a comparison without becoming reflexive

Difficulty: Intermediate. Let kk be a field and VV any vector space. Take E=kXE=k_X, M=VXM=V_X, and the inclusion of a point in an oriented dd-manifold. Calculate (16). Test biduality when V=k(ℕ)V=k^{(\mathbb N)}.

Solution. Constant-sheaf Hom on a sufficiently small ball gives DXVX=(V*)X[d]D_XV_X=(V^*)_X[d]. Formula (O11) applies to arbitrary bounded constant coefficients and gives i!VX=V[−d]i^!V_X=V[-d], including infinite VV. Both sides of (16) are consequently V*[d]V^*[d], and their map is the evaluation comparison from these identifications. For V=k(ℕ)V=k^{(\mathbb N)}, the quotient-functional argument in calculation 2 gives a functional on kℕk^{\mathbb N} which is not a finite coordinate functional. Thus V→V**V\to V^{**} is not surjective. The comparison used finite biduality of kXk_X, not biduality of the twist.

8. A controlled boundary permits an infinite twist

Difficulty: Intermediate. Regard j:(−2,2)↪(−3,3)j:(-2,2)\hookrightarrow(-3,3) as a morphism of pairs ((−2,2),ℝ)→((−3,3),ℝ)((-2,2),\mathbb R)\to((-3,3),\mathbb R). Use B=kUB=k_U and M=V(−3,3)M=V_{(-3,3)}, with arbitrary VV over a field. Check (18) including both endpoints.

Solution. The ambient zero extension of BB is constructible with perfect stalks. The ambient graph is subanalytic with compact closure. All boundary hypotheses are therefore satisfied. The left side of (18) is (V*)(−2,2)[1](V^*)_{(-2,2)}[1], with zero endpoint stalks. The ordinary untwisted image is k[−2,2]k_{[-2,2]}; after tensoring it becomes V[−2,2]V_{[-2,2]}. At an endpoint its costalk is the fibre of the identity V→VV\to V, hence zero. In the interior its costalk is V[−1]V[-1]. Stalk–costalk duality identifies its dual with (V*)(−2,2)[1](V^*)_{(-2,2)}[1]. The oriented trace matches the interior evaluation, and the endpoint maps are the unique maps of zero objects. This verifies (18) without requiring finite dimension of VV or an ambient extension as part of the theorem’s hypotheses.

Prerequisite ledger and reading

The compact-support arguments (C1)–(C6) prove the extension, lifting, stability and c-soft criterion needed by the projection proof (P1)–(P8). The proper-image arguments (F1)–(F6) supply the fibre and support-operation inputs. The composition and dimension arguments (D1)–(D10) supply base change, proper-image composition and the uniform bound. Projection therefore supplies contract (2) with its compact-support foundations proved here. The construction (A1)–(A10) supplies ordinary and exceptional adjunction and trace-compatible exceptional composition on bounded-below categories. The bounds (B1)–(B15) establish ordinary and exceptional boundedness, closed localization and a finite injective dualizing model. The constructions (H1)–(H12) supply enough injectives, finite flat models, tensor–Hom adjunction, natural evaluation and the double-dual triangle identity. These provide the algebraic operation contracts used by (5), (7), (9)–(10) and the evaluation criterion. The local arguments (O1)–(O12) prove homotopy invariance through a proper interval, point-to-compact comparison, orientation signs, the dualizing identification with its trace and the constant relative-ball formulas. The full chain of prerequisite proofs is not claimed to be complete: the underlying localization/truncation formalism and the geometric inputs listed below are not proved in this lesson. The following additional inputs are needed for their geometric applications. These links identify current arguments; they do not certify that those arguments’ own source and proof chains have finished reconstruction.

Input Current in-course argument Use here
Actual evaluation is constructible biduality over the stated ring Constructible costalks and Verdier duality, “The evaluation map is biduality” Evaluations in (12)–(14)
Perfect constructible inverse and proper-support image stability Perfect operations and finite microlocal coefficients Verifying constructibility, including the untwisted image in (18)
Stabilized stalk and costalk comparisons Small balls, central fibres and supported cohomology Local duality and arbitrary coefficients
Inverse-Hom and exceptional-tensor comparisons Weak constructibility under sheaf operations, “Analytic inverse images commute with these coefficients” (16)–(17)
Compact graph image and Hom comparisons Weak constructibility under sheaf operations, “Nonproper maps with controlled behavior at infinity” (18)–(19)
General weakly constructible small-ball stabilization and local duality The bundled scalar small-ball proof supplies all three comparisons once proper-image weak constructibility is known. The proper-image microsupport estimate and closed-cutoff sheaf argument are now supplied there; the constructibility criterion, geometric cotangent transport, compact finiteness and local dual pairings retain their proof obligations. Calculations 2–8 and local duality

The general duality discussion in David B. Massey’s Notes on Perverse Sheaves and Vanishing Cycles, version 13 provides further reading. Its stated coefficient assumptions are different from ours; it is not used to supply the full-ring biduality proof here. Andreas Hohl and Pierre Schapira’s Unusual functorialities for weakly constructible sheaves, version 2 supplies the infinite-twist results treated above.