SH02-SUP-EXH-001 — Recovering sheaf cohomology from closed pieces

Support item: SH02-SUP-EXH-001.

This note supplies the geometric comparison needed before applying a Mittag-Leffler argument. Its proof does not require the closed pieces to be compact. In particular it applies to strips with a noncompact horizontal base.

SH02-EXH-COMPARISON — The comparison

Let T be a topological space and k a fixed unital coefficient ring. Let

T0⊂T1⊂⋯⊂T,Tn closed,Tn⊂Int⁡T(Tn+1),⋃nTn=T. T_0\subset T_1\subset\cdots\subset T, \qquad T_n\text{ closed},\qquad T_n\subset\operatorname{Int}_T(T_{n+1}),\qquad \bigcup_nT_n=T.

Write in:Tn↪Ti_n:T_n\hookrightarrow T. For KK in the bounded-below derived category of sheaves of left kk-modules on TT, put Kn=in*in−1KK_n=i_{n*}i_n^{-1}K. Restriction to the smaller closed piece gives the transition Kn+1→KnK_{n+1}\to K_n. Then the restriction maps induce an equivalence

K≃holimnKn.(1) K\simeq\operatorname*{holim}_n K_n. \qquad\text{(1)}

Here holim means the homotopy inverse limit, or derived inverse limit. It is not an ordinary categorical inverse limit in the triangulated derived category. One may work in the enhanced derived category. Concretely, represent KK by a bounded-below complex CC and use the strict tower in*in−1Ci_{n*}i_n^{-1}C. For hn:Tn↪Tn+1h_n:T_n\hookrightarrow T_{n+1}, the transition is in+1*i_{n+1*} applied to the restriction unit into hn*hn−1in+1−1Ch_{n*}h_n^{-1}i_{n+1}^{-1}C. The resulting strict map from the constant tower defines the map in (1) after derived inverse limit.

For an explicit resolution, use the Grothendieck category of inverse sequences of sheaves. Evaluation at index n preserves injectives: its left adjoint places the given sheaf at indices at most n, zero at later indices, and has identity transitions where applicable; it is exact. A bounded-below injective resolution of the tower consequently evaluates to K-injective complexes InI_n. The cone of 1−shift1-\mathrm{shift} on their product, shifted by minus one, computes the homotopy limit. Strict compatibility provides the zero null-homotopy needed for the comparison map. This agrees with the enhanced-derived-category construction, without asserting uniqueness of a cone in a bare triangulated category.

Proof. Fix an open inclusion j:U↪Tj:U\hookrightarrow T. Restriction j−1j^{-1} is exact and has the exact extension-by-zero functor j!j_! as a left adjoint. The adjunction therefore passes to derived categories, and j−1j^{-1} preserves derived products. It is exact on distinguished triangles as well. Apply it to the derived-product triangle defining the countable homotopy limit,

holimnKn⟶∏nDKn→1−shift∏nDKn. \operatorname*{holim}_n K_n\longrightarrow \prod_n^{\mathrm D} K_n \xrightarrow{\,1-\mathrm{shift}\,}\prod_n^{\mathrm D}K_n.

It follows that restriction commutes with this homotopy limit. The products here are derived products; termwise products of arbitrary complexes must not be substituted without justification.

The open sets UN=Int⁡T(TN)U_N=\operatorname{Int}_T(T_N) cover TT: a point of TnT_n lies in Un+1U_{n+1}. On UNU_N, every KnK_n with n≥Nn\geq N restricts to K|UNK|_{U_N}, with identity transition maps under the restriction identifications. Discarding a finite initial segment does not change a homotopy inverse limit. The homotopy limit of this remaining constant tower is K|UNK|_{U_N}, and the map in (1) restricts to that equivalence. For clarity, the latter constant-tower computation is elementary: the map x↦(x0,(1−shift)x)x\mapsto(x_0,(1-\mathrm{shift})x) identifies ∏L\prod L with L⊕∏LL\oplus\prod L. Its inverse has nth component a−∑0≤r<nyra-\sum_{0\leq r<n}y_r, a finite sum. Thus the homotopy fibre of the second projection is LL, with the diagonal comparison. Finite initial factors are removed by the same finite triangular elimination. Equivalences of complexes of sheaves can be checked on an open cover, by their cohomology sheaves. This proves (1). No compactness or Hausdorff assumption was used. ▫\square

SH02-EXH-MILNOR — The precise cohomological obstruction

Closed pushforward in*i_{n*} is exact and is right adjoint to the exact inverse-image functor. Thus it preserves injectives, and

RΓ(T,Kn)≃RΓ(Tn,in−1K). R\Gamma(T,K_n)\simeq R\Gamma(T_n,i_n^{-1}K).

The derived-sections functor commutes with homotopy inverse limits. Applying the Milnor exact sequence to (1) gives, for every integer q,

0⟶lim⁡n1Hq−1(Tn,in−1K)⟶Hq(T,K)⟶limnHq(Tn,in−1K)⟶0.(2) 0\longrightarrow \lim{}^1_n H^{q-1}(T_n,i_n^{-1}K) \longrightarrow H^q(T,K) \longrightarrow \lim_n H^q(T_n,i_n^{-1}K) \longrightarrow0. \qquad\text{(2)}

Consequently, the inverse-limit description of Hq(T,K)H^q(T,K) holds if the tower in degree q−1q-1 is Mittag-Leffler: for each n, the images in its nth term from all sufficiently late terms stabilize. Surjective transition maps in that degree are one sufficient condition. The relevant condition is on degree q−1q-1; Mittag-Leffler in degree qq alone does not remove the left term in (2). For all-degree continuity, check the required condition in every degree.

If every transition is an isomorphism in every degree, global cohomology is identified with that on every piece. If transitions are eventually isomorphisms in every degree, this conclusion holds in each degree on its stable tail. The stabilization index may depend on the degree; a common tail requires a uniform index. In degree q the obstruction also uses degree q-1, so take a tail beyond both stabilization indices. No identification with arbitrary early pieces follows.

SH02-EXH-TOWER-PROOFS — Tower resolutions, sections and the Milnor sequence

Here are proofs of the inverse-limit facts used above. They apply to the same bounded-below complexes and to left modules over any unital ring. Neither exactness of products of sheaves nor compactness of the closed pieces is assumed.

Resolving the whole tower

Let 𝒜\mathcal A be the category of sheaves of left kk-modules on TT, and let Tow⁡(𝒜)\operatorname{Tow}(\mathcal A) consist of sequences Fn+1→unFnF_{n+1}\xrightarrow{u_n}F_n, indexed by the nonnegative integers. Its morphisms commute with all transitions. Kernels, cokernels and colimits are computed at each index, so it is abelian and has exact filtered colimits. These sheaf-category facts are proved in Sheaves of modules on a ringed space, Theorems 2.1 and 3.1.

For an object MM of 𝒜\mathcal A, define LnML_nM to have value MM at indices 0,…,n0,\ldots,n, zero at later indices, and identity transitions between its nonzero terms. A tower map LnM→FL_nM\to F is determined by its component M→FnM\to F_n: earlier components are its composites with the transitions. Thus LnL_n is an exact left adjoint of evaluation at nn. If UU is a generator of 𝒜\mathcal A, then ⨁n≥0LnU\bigoplus_{n\geq0}L_nU is a generator of the tower category: a nonzero tower morphism has a nonzero component, detected by a map from the corresponding LnUL_nU. The sheaf generator and its detection property are proved in Proposition 1.1 of K-injective resolutions in Grothendieck abelian categories. Hom collections in the tower category are subsets of products of Hom sets, hence sets. Therefore it is a locally small Grothendieck abelian category.

The injective embedding construction, Theorem 2.4, applies to this category. The degree-by-degree resolution construction, Theorem 4.1, consequently resolves any bounded-below complex of towers by a bounded-below complex J•J^\bullet of injective towers. In particular it applies to the strict tower in*in−1Ci_{n*}i_n^{-1}C above. Evaluation preserves injectives, because its left adjoint LnL_n is exact: extending a map into an evaluated injective is the same as extending the adjoint map from the corresponding monomorphism of towers. Hence each evaluated complex Jn•J_n^\bullet is a bounded-below injective resolution of KnK_n. All these resolutions have the same lower bound.

An injective tower JJ has split-surjective transitions. Indeed, Ln(Jn)→Ln+1(Jn)L_n(J_n)\to L_{n+1}(J_n) is a monomorphism. The tower map from its source to JJ corresponding to 1Jn1_{J_n} extends by injectivity. At index n+1n+1, the extension is a map σn:Jn→Jn+1\sigma_n:J_n\to J_{n+1} satisfying unσn=1Jnu_n\sigma_n=1_{J_n}. For any family of morphisms yn:M→Jny_n:M\to J_n, set x0=0x_0=0 and recursively set xn+1=σn(xn−yn)x_{n+1}=\sigma_n(x_n-y_n). This gives xn−unxn+1=ynx_n-u_nx_{n+1}=y_n. Applied to the projections from the product, the construction gives a right inverse of

δ:∏n≥0Jn⟶∏n≥0Jn,δ(x)n=xn−unxn+1.(E3) \delta:\prod_{n\geq0}J_n\longrightarrow\prod_{n\geq0}J_n, \qquad \delta(x)_n=x_n-u_nx_{n+1}. \qquad\text{(E3)}

Each coordinate of that right inverse uses only finitely many projections and splittings. It is a genuine morphism of sheaves, not an assumed ability to lift infinitely many local sections on a common neighbourhood. Its kernel is lim⁡nJn\lim_n J_n, by the defining compatibility equations. The limit is also injective: limit is right adjoint to the exact constant-tower functor, so the same extension argument proves preservation of injectives.

Apply these facts separately in every degree of J•J^\bullet. The splittings need not commute with its differential; they establish degreewise split exactness of

0⟶limnJn•⟶∏nJn•→δ∏nJn•⟶0.(E4) 0\longrightarrow\lim_n J_n^\bullet \longrightarrow\prod_n J_n^\bullet \xrightarrow{\delta}\prod_n J_n^\bullet\longrightarrow0. \qquad\text{(E4)}

The kernel-to-fibre comparison is a quasi-isomorphism, by the cohomology sequence of this short exact sequence of complexes. In the cone convention d(y,x)=(dy+δx,−dx)d(y,x)=(dy+\delta x,-dx), the fibre is Cone⁡(δ)[−1]\operatorname{Cone}(\delta)[-1], its differential is (y,x)↦(−dy−δx,dx)(y,x)\mapsto(-dy-\delta x,dx), and the comparison sends a kernel element xx to (0,x)(0,x). This checks the actual comparison map and its signs.

The termwise limit of J•J^\bullet computes the right derived limit by the cited bounded-below derived-functor construction. Each product in (E4) represents the derived product of the KnK_n: products of K-injective complexes are K-injective and represent derived products, as proved in Proposition 5.4 of the K-injective lesson. Thus (E4) proves the homotopy-limit triangle used in the geometric comparison, not merely a formula for an ordinary inverse limit. Comparison maps between tower resolutions are unique up to homotopy by the same bounded-below comparison theorem. The construction is therefore independent of the chosen resolution and natural in maps of strict towers.

Restriction, finite initial segments and sections

For an open inclusion j:V↪Tj:V\hookrightarrow T, restriction is exact and has the exact left adjoint j!j_!, proved in Lemma 5.1 of the module-sheaf lesson. It therefore preserves injectives and K-injectives by the adjunction argument. Restriction also commutes with ordinary products: both products have sections on an open of VV equal to the product of the same sections on that open in TT. Apply this to the resolved products in (E4). It proves commutation of restriction with the derived-product fibre, even though arbitrary products of sheaves need not be exact.

To remove the first NN terms of a tower, split its resolved product into the first NN factors and the tail. In these coordinates δ\delta is block upper triangular. Its head block is upper triangular with identity diagonal; backward substitution gives its inverse using only finite sums of transition maps. Eliminating this block identifies its cone with the cone of the tail block plus the contractible cone of an identity. Thus deleting the head preserves the homotopy fibre and its comparison maps. For a constant tail use one K-injective model II of its value. The coordinate transformation

∏n≥0I⟶I⊕∏n≥0I,x⟼(x0,(xn−xn+1)n)(E5) \prod_{n\geq0} I\longrightarrow I\oplus\prod_{n\geq0}I, \qquad x\longmapsto\bigl(x_0,(x_n-x_{n+1})_n\bigr) \qquad\text{(E5)}

is a chain isomorphism. Its inverse has coordinate a−∑r<nyra-\sum_{r<n}y_r. It changes δ\delta into projection onto the second factor, whose homotopy fibre is II. The diagonal map gives this identification. These computations justify precisely the local constant-tail step in (1).

Sections commute with products by the module-sheaf limit construction. The complexes in (E4), and their fibre, are K-injective; alternatively its kernel is a bounded-below complex of injectives by the limit argument above. Applying ordinary sections to these models therefore computes derived sections. Applying sections also commutes with the cone, since it commutes with finite sums and the displayed differentials. Consequently

RΓ(T,holim⁡nKn)≃Cone⁡(∏nRΓ(T,Kn)→1−shift∏nRΓ(T,Kn))[−1].(E6) R\Gamma\bigl(T,\operatorname{holim}_nK_n\bigr) \simeq\operatorname{Cone}\!\left( \prod_n R\Gamma(T,K_n)\xrightarrow{1-\mathrm{shift}} \prod_n R\Gamma(T,K_n)\right)[-1]. \qquad\text{(E6)}

Here the products on the right are represented by complexes of kk-modules. Products of modules are exact: kernels are coordinatewise, and a product of surjections is surjective by choosing a lift in each coordinate. Hence cohomology of these products is the product of their cohomologies. The long exact sequence of the fibre (E6) gives a natural short exact sequence with left term the cokernel of 1−shift1-\mathrm{shift} on the degree-(q−1)(q-1) product and right term its kernel on the degree-qq product. For the present closed embeddings, exactness of in*i_{n*} follows on stalks: its stalks are the original stalks on TnT_n, and zero off the closed set. Its exact left adjoint in−1i_n^{-1} shows that it preserves injectives. The equality of ordinary sections then identifies RΓ(T,Kn)R\Gamma(T,K_n) with RΓ(Tn,in−1K)R\Gamma(T_n,i_n^{-1}K), as used in (2).

For completeness, the cokernel just obtained is the first right derived limit of a module tower, not a new notation for a different obstruction. Resolve a module tower by injective towers and use (E4) in the module category. Exactness of products says that the products of the resolutions resolve the products of the original modules. Passing to their cones therefore identifies derived limit with the two-term complex ∏Mn→1−shift∏Mn\prod M_n\xrightarrow{1-\mathrm{shift}}\prod M_n in degrees zero and one. Its degree-zero cohomology is lim⁡Mn\lim M_n, its degree-one cohomology is lim⁡1Mn\lim^1 M_n, and its higher cohomology is zero. Substitution into the preceding fibre sequence proves (2), with exactly the degree-(q−1)(q-1) obstruction asserted there.

Why the Mittag–Leffler hypothesis kills that obstruction

Let M=(Mn,un)M=(M_n,u_n) be a countable inverse sequence of left kk-modules with stabilizing images. Let Sn⊂MnS_n\subset M_n be the eventual image in MnM_n. The transitions carry Sn+1S_{n+1} onto SnS_n: choose an index beyond stabilization at both positions; a lift from that index of an element of SnS_n maps into Sn+1S_{n+1}. Put Qn=Mn/SnQ_n=M_n/S_n. For every fixed nn, the transition Qm→QnQ_m\to Q_n is zero for all sufficiently large mm, since the corresponding image in MnM_n has become SnS_n.

The difference map δS\delta_S on ∏Sn\prod S_n is onto. For a prescribed yy, start with x0=0x_0=0 and choose xn+1x_{n+1} mapping to xn−ynx_n-y_n, using the surjective transitions. The difference map on ∏Qn\prod Q_n is invertible: if vn,m:Qm→Qnv_{n,m}:Q_m\to Q_n denotes the composite transition and vn,n=1v_{n,n}=1, its inverse sends yy to

xn=∑m≥nvn,m(ym).(E7) x_n=\sum_{m\geq n}v_{n,m}(y_m). \qquad\text{(E7)}

For each nn only finitely many summands can be nonzero, by the preceding eventual-zero property. Subtracting the shifted expression leaves yny_n. Conversely a compatible family in QQ is zero because each coordinate is the image of an arbitrarily late coordinate and that transition eventually vanishes. This proves both inverse identities.

Products of the short exact sequences 0→Sn→Mn→Qn→00\to S_n\to M_n\to Q_n\to0 remain exact. Given y∈∏Mny\in\prod M_n, solve its image under δQ\delta_Q, lift that solution to x′∈∏Mnx'\in\prod M_n, and observe that y−δMx′y-\delta_Mx' lies in ∏Sn\prod S_n. Surjectivity of δS\delta_S supplies zz with this difference, and x′+zx'+z solves δMx=y\delta_Mx=y. Thus lim⁡1Mn=coker⁡δM=0\lim^1 M_n=\operatorname{coker}\delta_M=0. This proof uses the same classical choice convention as the injective constructions. It does not require finite generation, commutativity, a field, or a uniform stabilization index. Applied to the degree-(q−1)(q-1) tower it gives exactly the continuity criterion following (2), including all noncompact-strip restrictions.

SH02-EXH-STRIPS — Noncompact-strip application

For an arbitrary topological space VV, take T=V×ℝT=V\times\mathbb R and Tn=V×[−n,n]T_n=V\times[-n,n] for positive integers n. These are closed, satisfy the interior condition, and cover T. Thus (1) and (2) apply without requiring V to be compact. Any claimed vanishing of the lim-one term still needs a separate proof for the actual sheaf or complex under consideration. The geometry of the exhaustion alone does not prove that condition.

SH02-EXH-COUNTEREXAMPLE — Why the interior condition is substantive

A closed increasing cover alone does not suffice. Let T=[0,1]T=[0,1], Tn={0}∪[1/n,1]T_n=\{0\}\cup[1/n,1] for n≥2n\geq2, and K=ℤTK=\mathbb Z_T. Each piece has two components, hence degree-zero sections ℤ2\mathbb Z^2, and its restriction transitions are identities. Global sections on the connected interval are ℤ\mathbb Z, with diagonal restriction. The degree-zero limit comparison is therefore not an isomorphism. The interiors fail to cover zero, exactly where the local proof cannot apply.

SH02-EXH-DEPENDENCIES — Dependencies and attribution

This is a short independent bridge built from standard derived-sheaf facts, not a claim of a new research theorem. The exact Stacks inputs are 0D60 (derived sections and the Milnor sequence) and 02UV (closed pushforward and cohomology). The internal tower proof supplies the derived extension-by-zero/restriction comparison, the countable homotopy-limit triangle, invariance under removing a finite initial segment, and the countable Mittag-Leffler vanishing theorem. Their precise supporting tags are 00A7, 01AK, 01AX, 08TC, 07D9, 0BK7 and 07KW. The nonunique-cone caution is 0H9J.

For noncommutative kk, the same additive arguments apply to left modules: no tensor-product exactness, commutativity or finite global dimension is used. For these coefficients, commutation of derived sections with derived inverse limits follows from 08U1. Exact products of modules then give the Milnor sequence by the same product-triangle argument underlying 0D60. The internal tower and module proofs above establish these facts with this coefficient interpretation; the links also provide source credit and further reading.

Original text is dedicated under CC0 1.0 Universal.