Čech complexes, refinements and acyclic covers

Draft lesson.

Written by GPT-6.1 Sol (OpenAI), October 2026. Public domain (CC0).

For finite closed covers, use the separate closed-restriction resolution and its cohomology comparison. Sections on a cover satisfy an overlap equation. Higher Čech cochains continue that equation through multiple intersections. To use them for sheaf cohomology, one must account for the higher cohomology already present on those intersections. A double complex makes that distinction precise and also explains which covers really compute the answer.

Let \((X,\mathcal O)\) be a topological space with a sheaf of unital rings, possibly noncommutative, and let \(\mathcal F\) be a left module sheaf. Fix universes and use choice. Let \(\mathcal U=(U_i)_{i\in I}\) be any set-indexed open cover of \(X\). Repetitions and empty members are allowed. Products below are products of modules, and sections on the empty open are zero. All cohomology groups are ordinary right-derived section groups.

We use Godement resolutions and flabby sheaf cohomology, Sections 1–6, for enough injectives, exact open extension by zero, restriction of injectives and their flabbiness; Injective replacements and right derived functors, Sections 2–5, for comparison homotopies and long exact sequences; and Filtered complexes and Cartan–Eilenberg resolutions, Sections 1–3, for the proved spectral pages, signs, finite convergence and total-complex comparison. The exercises also use the sheaf–diagram equivalence of Finite-poset sheaves and canonical homological resolutions, Sections 1–3 and 6.

1. Cochains on all ordered intersections

For a tuple \(\mathbf i=(i_0,\ldots,i_p)\), put \(U_{\mathbf i}=\bigcap_{a=0}^p U_{i_a}\). Use all ordered tuples, including repeated indices, and set

C^p(\mathcal U,\mathcal F)=\prod_{\mathbf i\in I^{p+1}}\mathcal F(U_{\mathbf i}),\qquad (\delta c)_{i_0\ldots i_{p+1}} =\sum_{a=0}^{p+1}(-1)^a c_{i_0\ldots\widehat{i_a}\ldots i_{p+1}}|_{U_{i_0\ldots i_{p+1}}}. \tag{1.1}

Each sum is finite even for an infinite cover. Deleting positions \(a<b\) in either order gives the same restriction; the corresponding coefficients have opposite signs because the second deletion shifts its position by one. Pairing these terms proves \(\delta^2=0\). The augmentation sends a section on \(X\) to its restrictions on the \(U_i\). Its composite with \(\delta\) is zero, and it identifies \(\mathcal F(X)\) with \(\ker\delta^0\): a degree-zero cocycle is exactly a compatible family, so the sheaf gluing axiom applies. This also holds for an empty cover of the empty space.

Write \(\check H^p(\mathcal U,\mathcal F)=H^p(C^\bullet(\mathcal U,\mathcal F))\). A module sheaf map induces cochain maps by applying it to every section. The same formula defines \(C^\bullet(\mathcal U,P)\) for a presheaf \(P\); its degree-zero kernel need not be \(P(X)\) unless the relevant gluing holds.

A refinement consists of a cover \(\mathcal V=(V_j)_{j\in J}\) and a function \(r:J\to I\) with \(V_j\subseteq U_{r(j)}\). It defines

(r^*c)_{j_0\ldots j_p} =c_{r(j_0)\ldots r(j_p)}|_{V_{j_0\ldots j_p}}. \tag{2.1}

Restrictions commute with deletion, so \(r^*\delta=\delta r^*\). Identity refinements give identity maps, and successive refinement functions compose exactly in (2.1). No order on either index set is needed.

If \(r,s:J\to I\) are two refinement functions for the same covers, define \(H^0=0\) and, for \(p\ge1\),

(H^p c)_{j_0\ldots j_{p-1}} =\sum_{a=0}^{p-1}(-1)^a c_{r(j_0)\ldots r(j_a)\,s(j_a)\ldots s(j_{p-1})} |_{V_{j_0\ldots j_{p-1}}}. \tag{2.2}

Every restriction is defined: the smaller intersection lies in both chosen opens for each of its indices. These maps satisfy

\delta H+H\delta=s^*-r^*. \tag{2.3}

Here is the cancellation proving the signs. On a formal tuple \([j_0,\ldots,j_n]\), form the sum of \([rj_0,\ldots,rj_a,sj_a,\ldots,sj_n]\) with coefficient \((-1)^a\). In its boundary, deleting an entry strictly before the switch pairs with the corresponding summand of the prism applied to the boundary tuple; the same holds strictly after the switch. The coefficients are opposite after moving that deletion past the switch. Deleting one of the two entries at the switch pairs with the adjacent switch, leaving just the all-\(s\) tuple minus the all-\(r\) tuple. This proves \(\partial P+P\partial=s_*-r_*\). Evaluate cochains on these chains and restrict to the common intersection to obtain (2.3). In degree zero it reads \((H^1\delta c)_j=c_{s(j)}|_{V_j}-c_{r(j)}|_{V_j}\), so it includes the bottom degree. Thus refinement maps on Čech cohomology are independent of the chosen refinement function, and respect composition.

3. A free open-extension chain resolution

For an open \(W\), write \(j_W:W\hookrightarrow X\). The first prerequisite proves exactness of \(j_{W!}\) and the adjunction identifying \(\operatorname{Hom}_{\mathcal O}(j_{W!}\mathcal O|_W,\mathcal G)\) with \(\mathcal G(W)\), by evaluation at one. Define

P_p=\bigoplus_{\mathbf i\in I^{p+1}}j_{U_{\mathbf i}!}\mathcal O|_{U_{\mathbf i}} \quad(p\ge0),\qquad P_{-1}=\mathcal O. \tag{3.1}

If \(W\subseteq V\), extension of the unit section gives a canonical map \(j_{W!}\mathcal O|_W\to j_{V!}\mathcal O|_V\); it is the identity on stalks in \(W\) and zero on all other source stalks. These maps compose under inclusions. On each tuple summand define \(\partial_p:P_p\to P_{p-1}\) as the sum of the deletion maps with coefficients \((-1)^a\). For \(p=0\), use the canonical inclusions into \(\mathcal O\). The same deletion pairing as in Section 1 proves \(\partial^2=0\).

This augmented chain complex is exact. Indeed its stalk at \(x\) is the free \(\mathcal O_x\)-module on tuples of the nonempty set \(I_x=\{i:x\in U_i\}\), augmented to \(\mathcal O_x\). Stalks commute with the direct sum, since a germ of a section of the sheafified sum is represented locally by finitely many summands. Choose \(a\in I_x\). Define \(s_{-1}(b)=b[a]\) and \(s_p(b[i_0,\ldots,i_p])=b[a,i_0,\ldots,i_p]\). Deleting the initial \(a\) gives the original tuple; the other deletions have the negative of the signs in \(s\partial\). Consequently \(\partial s+s\partial=1\) in every augmented degree. Exactness on stalks proves exactness of the sheaf complex. The chosen insertion may depend on \(x\); a global sheaf contraction is not asserted. If \(X\) is empty, the sheaf complex is zero and the conclusion still holds.

4. Why injectives have no positive Čech cohomology

Let \(J\) be an injective module sheaf. The direct-sum universal property and the open adjunction give actual identifications

\operatorname{Hom}_{\mathcal O}(P_p,J) \simeq\prod_{\mathbf i}J(U_{\mathbf i})=C^p(\mathcal U,J). \tag{4.1}

The forward map restricts a morphism to each summand and evaluates at one. Its inverse takes each section to its multiplication map on that open, extends by the adjunction, and assembles the maps on the direct sum. Both composites are identities. Precomposition with \(\partial\) becomes precisely (1.1), including its signs and restriction maps.

For completeness, exactness after applying this Hom functor can be checked directly. A cocycle \(f:P_p\to J\) vanishes on \(\operatorname{im}\partial_{p+1}=\ker\partial_p\). It therefore factors through \(\operatorname{im}\partial_p\subseteq P_{p-1}\). For \(p>0\), injectivity extends that factor to \(P_{p-1}\to J\), making \(f\) a coboundary. For \(p=0\), the same factor is a map \(\mathcal O\to J\), since the augmentation is onto. Hence the augmented Čech complex of \(J\) is exact, and

\check H^0(\mathcal U,J)=J(X),\qquad \check H^p(\mathcal U,J)=0\quad(p>0). \tag{4.2}

The sheaves \(P_p\) need not be projective. Injectivity of the target is the extension property used in this proof.

5. The signed double complex and its augmentation

Choose an ordinary injective resolution \(\mathcal F\to I^\bullet\), starting in degree zero. Set

B^{p,q}=C^p(\mathcal U,I^q),\quad \operatorname{Tot}^n B=\bigoplus_{p+q=n}B^{p,q},\quad D|_{B^{p,q}}=\delta+(-1)^p d_I. \tag{5.1}

The two unsigned differentials commute, since the resolution maps commute with restrictions. The mixed terms in \(D^2\) cancel: the horizontal map raises \(p\) by one, reversing the vertical sign. The pure terms vanish. There are only \(n+1\) summands in total degree \(n\), regardless of the size of the cover.

Restriction defines a chain map

a:\Gamma(X,I^\bullet)\longrightarrow\operatorname{Tot}B \tag{5.2}

into column zero. The horizontal differential of a restricted section is zero, and the vertical sign there is positive. Section 4 says that, for each \(q\), the unaugmented horizontal row has cohomology \(I^q(X)\) in horizontal degree zero and zero elsewhere. Regarding the source of (5.2) as a double complex supported in column zero, \(a\) is thus a quasi-isomorphism on every row. The row comparison of the third prerequisite applies: both complexes are first quadrant, and the row filtration has finite length in each total degree. It proves that \(a\) is a quasi-isomorphism. Its abutment is therefore \(H^n(X,\mathcal F)\).

There is also an actual chain map \(b:C^\bullet(\mathcal U,\mathcal F)\to\operatorname{Tot}B\), from \(\mathcal F\to I^0\) in row zero. Its image has zero vertical differential. Define the Čech comparison map by

\sigma^n_{\mathcal U}=H^n(a)^{-1}H^n(b): \check H^n(\mathcal U,\mathcal F)\longrightarrow H^n(X,\mathcal F). \tag{5.3}

In degree zero it is the sheaf gluing identification. In higher degrees it need not be an isomorphism.

6. The comparison spectral sequence and its exact hypotheses

Use the column filtration \(F^r\operatorname{Tot}^nB=\bigoplus_{p\ge r,\ p+q=n}B^{p,q}\). It is the whole degree for \(r\le0\) and zero for \(r>n\). The filtered-complex theorem therefore applies with its proved finite convergence. Restricting an injective resolution to any \(U_{\mathbf i}\) remains an injective resolution. Products of module sequences are exact by coordinatewise lifting, so vertical kernels, images and their quotient give

E_1^{p,q}=\prod_{\mathbf i\in I^{p+1}}H^q(U_{\mathbf i},\mathcal F),\qquad d_1=\delta,\qquad E_2^{p,q}=\check H^p(\mathcal U,\mathcal H^q_{\mathcal F}), \tag{6.1}

where \(\mathcal H^q_{\mathcal F}(W)=H^q(W,\mathcal F)\) with the usual open restriction maps. This is a presheaf; no sheafification is used. The signed vertical complex has the same cycles and boundaries as the unsigned one, which fixes the direct identification in (6.1) and gives the stated sign of \(d_1\).

All higher differentials and the abutment filtration are those of the third prerequisite: \(d_r\) has bidegree \((r,1-r)\), and

E_\infty^{p,n-p}\simeq F^pH^n(X,\mathcal F)/F^{p+1}H^n(X,\mathcal F),\qquad F^0H^n=H^n,\quad F^{n+1}H^n=0. \tag{6.2}

The filtration is transported along \(H(a)\). Every position stabilizes after finitely many pages by the degreewise finite theorem. Thus (6.1) converges even for an infinite cover. It specifies associated pieces, not a chosen splitting of the resulting extensions.

Theorem 6.1 (acyclic-cover comparison). Suppose \(H^q(U_{i_0\ldots i_p},\mathcal F)=0\) for every finite nonempty tuple and every \(q>0\). Then (5.3) is an isomorphism for every \(n\ge0\).

Proof. Regard \(C^\bullet(\mathcal U,\mathcal F)\) as a double complex in row zero. In each column the map \(b\) is a quasi-isomorphism: degree-zero cohomology of the target is the product of the section kernels, hence \(C^p(\mathcal U,\mathcal F)\), and all its positive cohomology vanishes by the assumed intersection groups and exact products. Column comparison with the same finite filtrations proves that \(b\) is a quasi-isomorphism on totals. Since \(a\) is already one, (5.3) is an isomorphism. This proof identifies the actual comparison rather than only an abstract group isomorphism. Empty intersections have zero section complex and cause no exception. No finite-cover or separation hypothesis was used. \(\square\)

7. Naturality, choices and restriction to an open

A sheaf map lifts to a map of injective resolutions, unique up to the comparison homotopy proved in the second prerequisite. If \(h:I^q\to I'^{q-1}\) is a homotopy, the corresponding total homotopy on \(B^{p,q}\) is \((-1)^p C^p(\mathcal U,h)\). Its horizontal terms cancel, and its vertical terms give \(d'h+hd\). It preserves the column filtration, so homotopic comparisons induce the same maps on \(E_1\), all later pages and the filtered abutment. Identity and composition comparisons follow by that same uniqueness. This proves independence of the injective resolution, naturality in \(\mathcal F\), and all composition compatibilities for (5.3) and (6.1).

A refinement induces a double-complex map by (2.1) in each resolution degree, commuting with \(a\) on its source and with \(b\). Its map on the abutment, identified with \(H^*(X,\mathcal F)\), is the identity. For two refinement functions the prism (2.2), applied in every vertical degree, is a total homotopy without an extra sign: its vertical mixed terms have coefficients \((-1)^{p-1}\) and \((-1)^p\) and cancel. On vertical cohomology it gives a homotopy of the horizontal \(E_1\) complexes. Consequently the two maps agree on \(E_2\) and all later pages, and on the abutment. They need not agree as maps of \(E_1\) groups before taking horizontal cohomology. The cochain formulas still compose exactly when refinement functions are specified.

For an open \(W\subseteq X\), use the cover \((U_i\cap W)_i\) of \(W\). Restricting cochains gives the map on all tuple intersections. Restriction of \(I^\bullet\) is an injective resolution on \(W\) by the first prerequisite, so these maps commute with (5.2), (5.3), every page and the usual \(H^*(X,\mathcal F)\to H^*(W,\mathcal F|_W)\). Thus all comparisons are compatible with actual open restrictions. These conclusions require no assertion that arbitrary covers satisfy Theorem 6.1.

8. Four exercises with full solutions

8.1 Infinitely many copies of one open

Exercise 1 (basic). On a one-point space with ring \(A\), let a module \(M\) be covered by copies of that point indexed by any nonempty set \(I\). Compute the augmented Čech complex and its cohomology, including infinite \(I\).

Solution. In degree \(p\) the group is \(\prod_{I^{p+1}}M\), and every restriction in (1.1) is an identity. Choose \(a\in I\). Define \(h^0(c)=c_a\in M\) and \((h^pc)_{i_0\ldots i_{p-1}}=c_{a,i_0,\ldots,i_{p-1}}\) for \(p\ge1\). With the augmentation as degree \(-1\), the formula \(h^{p+1}\delta+\delta h^p=1\) follows by deleting the initial \(a\) and pairing all the other deletions with opposite signs. Also \(h^0\eta=1_M\). These are maps on full products: they select coordinates and take finite sums, and never require finite support. Hence \(\check H^0=M\), \(\check H^p=0\) for \(p>0\). This agrees with sheaf cohomology, since sections on a point are the exact identity functor, even if \(M\) is not injective as an \(A\)-module.

8.2 The full Mayer–Vietoris boundary

Exercise 2 (intermediate). For \(X=U\cup V\), derive the Mayer–Vietoris long exact sequence for any module sheaf. Specify its difference map and connecting homomorphism without assuming any of \(U,V,U\cap V\) acyclic.

Solution. Choose \(\mathcal F\to I^\bullet\). Injectives are flabby, so in every degree there is a short exact sequence

0\to I^q(X)\longrightarrow I^q(U)\oplus I^q(V) \xrightarrow{\ (u,v)\mapsto v|_{U\cap V}-u|_{U\cap V}\ }I^q(U\cap V)\to0. \tag{8.1}

The first map is restriction. Its kernel and image assertions are exactly uniqueness and existence of gluing. For surjectivity, extend a section on \(U\cap V\) to \(V\) by flabbiness and use zero on \(U\). These maps commute with the resolution differential, so the long exact sequence of complexes from the prerequisites gives

\cdots\to H^q(X,\mathcal F)\to H^q(U,\mathcal F)\oplus H^q(V,\mathcal F) \xrightarrow{\ v|-u|\ }H^q(U\cap V,\mathcal F) \xrightarrow{\Delta}H^{q+1}(X,\mathcal F)\to\cdots. \tag{8.2}

For a cocycle \(c\in I^q(U\cap V)\), choose \((u,v)\) mapping to it in (8.1). Then \(dv-du=dc=0\) on the overlap, so \(du\) and \(dv\) glue to a cocycle \(z\in I^{q+1}(X)\). The boundary is \(\Delta[c]=[z]\), with this positive sign. Changing the chosen lift adds a restricted global cochain \(w\), so changes \(z\) by \(dw\). Replacing \(c\) by \(c+d\gamma\), lift \(\gamma\) one degree earlier and add its differential to \((u,v)\); this leaves \(z\) unchanged. Thus the formula is well-defined. The short-exact-complex proof supplies exactness and naturality of every arrow. Restriction preserves injectives, so all the displayed groups are the indicated sheaf cohomology groups on the smaller opens. No acyclicity assumption occurred.

8.3 A four-point circle

Exercise 3 (advanced). Give \(X=\{u,v,a,b\}\) the upper-set topology for the relations \(u<a,u<b,v<a,v<b\), and no others except reflexivity. For the constant diagram \(k\) with identity structure maps, use \(U=\{u,a,b\}\) and \(V=\{v,a,b\}\) to compute sheaf cohomology and exhibit an actual Čech generator.

Solution. The sheaf–diagram equivalence makes global sections compatible tuples. Their four entries must agree, so \(\mathcal F(X)=k\). On \(U\), sections are evaluation at the least vertex \(u\); this functor is exact on all diagrams. Thus \(H^0(U,\mathcal F)=k\) and its positive groups vanish. The same holds on \(V\). The intersection \(\{a,b\}\) is discrete; its section functor is the product of two evaluations, so its degree-zero group is \(k^2\) and its positive groups vanish. Every finite tuple intersection of this two-member cover is \(U,V\), or \(U\cap V\), so Theorem 6.1 applies.

In (8.2) the degree-zero difference map is \(k^2\to k^2\), \((r,s)\mapsto(s-r,s-r)\). Its kernel is the diagonal \(k\) and its cokernel is \(k\), identified by \((x,y)\mapsto y-x\). Consequently \(H^0(X,\mathcal F)=k\), \(H^1(X,\mathcal F)=k\), and all groups in degrees at least two vanish. An actual Čech cocycle has \(c_{UU}=c_{VV}=0\), \(c_{UV}=(0,1)\), and \(c_{VU}=(0,-1)\). Every triple has only two possible labels, and substitution in \(c_{jk}-c_{ik}+c_{ij}\) gives zero. A Čech coboundary on \(UV\) has the form \((s-r,s-r)\), so this cocycle is not a boundary and represents the displayed generator. With the conventions (5.1) and (8.2), its image under (5.3) is \(-\Delta[(0,1)]\). Indeed choose the local pair lifting \((0,1)\) as in Exercise 2, and let \(z\) be its glued differential. In the total complex its differential is \(b(c)+a(z)\). Hence \([b(c)]=-[a(z)]\), proving the claimed sign. Both classes are generators; the explicit convention determines which one is obtained.

8.4 A cover that loses the first cohomology

Exercise 4 (advanced). Use the same space and sheaf as Exercise 3, but take the singleton cover \((X)\). Compute its Čech cohomology and comparison spectral sequence. Compare with the refinement \((U,V)\to(X)\).

Solution. Every tuple intersection is \(X\), so every cochain module is \(k\). The differential is the sum of \(p+2\) alternating identities: it is zero for even \(p\) and identity for odd \(p\). Its Čech cohomology is therefore \(k\) in degree zero and zero in every positive degree. In particular it misses \(H^1(X,\mathcal F)=k\). The singleton intersection fails the positive-cohomology hypothesis of Theorem 6.1.

For the spectral sequence, \(E_1^{p,0}=k\) and \(E_1^{p,1}=k\) for every \(p\ge0\), with zero other rows, by Exercise 3. Each horizontal differential again alternates between zero and identity. Hence \(E_2^{0,0}=E_2^{0,1}=k\), with all other positions zero. Higher differentials vanish, giving the correct two abutment groups. The first group in total degree one comes from \((p,q)=(0,1)\), not from the bottom Čech row.

There is only one refinement function to the singleton index set. Formula (2.1) restricts a global section into every smaller tuple intersection. On first Čech cohomology it is the zero map \(0\to k\), whereas on the spectral-sequence abutment it is the identity of \(H^1(X,\mathcal F)\). This is consistent with (5.3) and Section 7. Independence of the refinement function does not say that arbitrary refinement maps on Čech cohomology are isomorphisms.