Covering dimension and locally finite refinements

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

A cohomological dimension bound concerns all open covers, whereas covering dimension is often defined using only finite covers. On a noncompact space an arbitrary cover need not have a finite subcover. The passage between these two formulations therefore needs a theorem. We prove it for every paracompact Hausdorff space, without a countability or metrizability assumption.

The construction starts with a locally finite cover and removes its unwanted intersections one at a time. An infinite succession of open shrinkings need not have an open limit. We arrange all modifications inside a fixed locally finite family of neighborhoods. Every point then sees only finitely many modifications, which makes the limit an open cover. The resulting refinement has bounded multiplicity even after taking closures.

We use normal shrinking from Fine module sheaves and central partitions of unity, closed locally finite refinements from Soft sheaves on paracompact Hausdorff spaces, and the finite shrinking and swelling lemma, Lemma 1.2, of Covering dimension and finite trivializing covers (Claude Opus 5.5, CC0). The final cohomological consequences use Cech comparison and cohomological bounds on paracompact spaces. Well-ordering and transfinite recursion are as in Well-orders, cardinal bounds and regular successors.

1. Finite covers and finite nerves

Throughout, \(X\) is a paracompact Hausdorff space and \(d\) is a nonnegative integer. A family has multiplicity at most \(d+1\) if no \(d+2\) distinctly indexed members have a common point. A refinement includes a choice of an original member containing each refining member. Empty members may be retained or discarded. A family is locally finite if each point has a neighborhood meeting only finitely many members. The closures of a locally finite family are locally finite: an open neighborhood meeting a closure also meets the original member.

We write \(\dim X\leq d\) if every finite open cover has an open refinement of multiplicity at most \(d+1\). One may require this refinement to be finite. Indeed, assign each of its members to one of the finitely many original members and take their unions label by label. The resulting finite family still covers, and multiplicity cannot increase: membership in \(r\) different groups requires membership in \(r\) distinct original refining members. These observations also apply to a finite indexed cover with repetitions.

Finite shrinking and swelling, recalled. In a normal space, a finite open cover \((A_i)\) has an open shrinking \((W_i)\) with \(\overline{W_i}\subseteq A_i\). If finitely many closed sets \(F_i\subseteq A_i\) have empty common intersection, there are open \(G_i\) with

These are precisely the two assertions of the linked Lemma 1.2, whose proofs use only normality. Paracompact Hausdorff spaces are normal by Lemma 1.1 of the fine-module lesson. In particular, if \(F\) is closed and \(F\subseteq A\) with \(A\) open, there is an open \(W\) with \(F\subseteq W\subseteq\overline W\subseteq A\).

Lemma 1.1 (Preserving a finite nerve). Let \(F_1,\ldots,F_m\) be closed and \(F_i\subseteq A_i\), with \(A_i\) open. There are open \(G_i\) containing \(F_i\), with \(\overline{G_i}\subseteq A_i\), such that every empty intersection of the \(F_i\) remains empty for the corresponding \(\overline{G_i}\). In particular, a multiplicity bound for the \(F_i\) holds for the \(\overline{G_i}\).

Proof. First choose open \(H_i\supseteq F_i\) with \(\overline{H_i}\subseteq A_i\). For each nonempty subset \(S\subseteq\{1,\ldots,m\}\) for which \(\bigcap_{i\in S}F_i=\varnothing\), apply finite swelling to the subfamily indexed by \(S\). It gives open \(H_i^S\supseteq F_i\), \(i\in S\), whose closures have empty common intersection. Set

There are only finitely many subsets \(S\), so \(G_i\) is open and contains \(F_i\). Its closure is contained in \(\overline{H_i}\subseteq A_i\). For a forbidden \(S\), every \(\overline{G_i}\), \(i\in S\), is contained in \(\overline{H_i^S}\), so their intersection is empty. Singleton subsets handle an empty \(F_i\). If the \(F_i\) cover, the \(G_i\) cover as well. \(\square\)

Lemma 1.2 (Finite covers with bounded closure multiplicity). If \(\dim X\leq d\), every finite open cover \((A_i)_{i=1}^m\) has an open shrinking \((G_i)_{i=1}^m\) whose closures lie in their assigned \(A_i\) and have multiplicity at most \(d+1\).

Proof. Choose a finite open refinement \((Q_j)\) of multiplicity at most \(d+1\), and an assignment \(a(j)\) with \(Q_j\subseteq A_{a(j)}\). Finite shrinking gives open \(W_j\) covering \(X\) with \(\overline{W_j}\subseteq Q_j\). Put

Each \(F_i\) is closed and contained in \(A_i\); they cover \(X\). If a point lies in \(r\) distinct \(F_i\), choose one witnessing \(j\) for each group. Those \(j\) are distinct, and the point lies in their \(Q_j\), so \(r\leq d+1\). Lemma 1.1 gives the required \(G_i\). \(\square\)

2. Locally finite neighborhoods of closed families

Lemma 2.1. Let \((C_s)_{s\in S}\) be a locally finite family of closed subsets of a paracompact Hausdorff space. For assigned open sets \(A_s\supseteq C_s\), there are locally finite open sets \(O_s\) with \(C_s\subseteq O_s\subseteq A_s\).

Proof. At \(x\), take an open neighborhood \(N_x\) meeting only finitely many \(C_s\). Remove from \(N_x\) those finitely many \(C_s\) which do not contain \(x\), and intersect with the finitely many \(A_s\) for which \(x\in C_s\). The result is an open neighborhood \(P_x\) of \(x\). If \(P_x\) meets \(C_s\), then \(x\in C_s\) and \(P_x\subseteq A_s\).

Take a locally finite open refinement \((V_t)\) of the cover \((P_x)\). Every \(V_t\) meets only finitely many \(C_s\), and lies in \(A_s\) whenever it meets \(C_s\). Define

The \(V_t\) cover, so \(C_s\subseteq O_s\); the preceding containment gives \(O_s\subseteq A_s\). Near a point only finitely many \(V_t\) occur, and each of those meets only finitely many \(C_s\). Only finitely many \(O_s\) can meet that neighborhood. Thus \((O_s)\) is locally finite. Empty \(C_s\) give empty \(O_s\). \(\square\)

The hypothesis concerns the entire closed family. Separating its members individually would give no local finiteness for the chosen neighborhoods. This distinction will control limit stages.

3. From finite covers to arbitrary covers

Theorem 3.1. The following conditions on a paracompact Hausdorff space \(X\) are equivalent.

1. \(\dim X\leq d\), using the finite-cover definition.
2. Every open cover has an open refinement of multiplicity at most \(d+1\).
3. Every open cover \(\mathcal U\) has a locally finite open refinement \((V_i)\), with assigned \(U_i\in\mathcal U\), such that \(\overline{V_i}\subseteq U_i\) and the family \((\overline{V_i})\) has multiplicity at most \(d+1\).

Proof. Condition 3 implies 2. Condition 2 implies 1, with finiteness of the resulting refinement supplied by the grouping argument in Section 1. We prove 1 implies 3.

Initial cover and fixed supports. The closed-refinement lemma in the soft-sheaf lesson gives a locally finite open cover \((V_i^0)_{i\in I}\) with \(K_i=\overline{V_i^0}\subseteq U_i\) for assigned original cover members \(U_i\). If \(I\) has fewer than \(d+2\) elements, there is nothing to prove. Otherwise let \(\mathcal S\) consist of the \((d+2)\)-element subsets \(S\subseteq I\) for which

is nonempty. This is a locally finite family of closed sets. In a neighborhood meeting only finitely many \(K_i\), a \(C_S\) can meet the neighborhood only if all its indices belong to that finite list, which leaves only finitely many subsets \(S\).

Lemma 2.1 supplies a fixed locally finite family of open sets \(O_S\supseteq C_S\). Well-order \(\mathcal S\). We shall shrink only the members whose indices belong to the subset being processed, and make the modification only inside \(O_S\).

One shrinking step. Suppose \((V_i)\) is the current open cover, with \(V_i\subseteq V_i^0\). For the current \(S\), put \(W=\bigcup_{j\notin S}V_j\) and consider the finite indexed open cover

Lemma 1.2 gives a finite open shrinking

whose closures have multiplicity at most \(d+1\). In particular, the \(d+2\) sets \(\overline{G_i}\), \(i\in S\), have empty common intersection. The set \(O_S\setminus\overline{G_*}\) is open and contains the closed set \(C_S\). By normal shrinking choose a closed \(D_S\) with

For \(i\in S\) replace \(V_i\) by

leave the other members unchanged. These are open subsets of the previous members. Outside \(D_S\) the cover is unchanged. Inside \(D_S\), the shrinking finite cover cannot use \(G_*\), so a point belongs to some \(G_i\) or to \(G_W\subseteq W\), and is still covered. The modification is supported in \(O_S\).

This step kills the intersection of the selected closures. Indeed, any point in \(\bigcap_{i\in S}\overline{V_i'}\) belongs to \(C_S\), because \(V_i'\subseteq V_i^0\). On the other hand,

Since \(C_S\subseteq\operatorname{int}D_S\), such a point would belong to every \(\overline{G_i}\), which is impossible.

Limit stages. Define a member at a limit ordinal by intersecting its earlier versions. This operation is open here for a specific reason. Around \(x\), choose a fixed open neighborhood \(N\) meeting only finitely many of the sets \(O_S\). Among the stages preceding the given limit, only finitely many can change any member on \(N\). After the last such stage, the entire cover restricted to \(N\) is constant. The intersection at the limit agrees on \(N\) with that earlier open cover. Consequently every limiting member is open near \(x\), and the limiting family still covers \(N\). This holds at every point. The same argument applies at the final ordinal if it is a limit.

Final bound. All covers in the recursion refine the initial locally finite family, so the final family is locally finite. Once an intersection of closures is empty, later shrinking preserves that fact. A \((d+2)\)-element subset outside \(\mathcal S\) already had empty intersection of its initial \(K_i\); a subset in \(\mathcal S\) has its intersection killed at its own stage. Thus no \(d+2\) final closures meet. They remain inside their initial \(K_i\subseteq U_i\), proving 3. \(\square\)

Notice that local finiteness is imposed on the modification supports before recursion starts. Local finiteness of the cover alone would not justify openness of an infinite intersection of its successive versions.

4. Closed subspaces and cohomological bounds

Corollary 4.1. If \(A\subseteq X\) is closed and \(\dim X\leq d\), then \(A\) is paracompact Hausdorff, \(\dim A\leq d\), and every open cover of \(A\) has the closure-multiplicity refinement in Theorem 3.1.

Proof. A closed subspace is paracompact: extend a relative open cover to ambient open sets and add \(X\setminus A\), refine locally finitely in \(X\), and restrict to \(A\). Hausdorffness is inherited. For the dimension bound, write a finite relative open cover as \((A\cap U_1,\ldots,A\cap U_m)\). Apply the finite-cover dimension condition in \(X\) to

Restrict its finite bounded-multiplicity refinement to \(A\), discarding the members assigned to \(X\setminus A\). This gives the required finite refinement of the original relative cover. Theorem 3.1 now applies to \(A\). \(\square\)

Corollary 4.2 (Sheaf cohomological dimension). Let \(\mathscr R\) be any sheaf of unital rings on \(X\), and let \(F\) be a sheaf of left \(\mathscr R\)-modules. If \(\dim X\leq d\), then

More generally, if \(K\in D^+(\mathscr R)\) and \(\mathcal H^q(K)=0\) for \(q>b\), then

The same assertions hold on every closed subspace with its own sheaf of rings and modules.

Proof. Theorem 3.1 supplies, cofinally among all open covers, covers of multiplicity at most \(d+1\). The complete comparison and finite-multiplicity theorem, Theorem 5.1 of the linked Cech-comparison lesson, then gives the first assertion: increasing-index Cech cochains are zero above degree \(d\), and their all-cover cohomology is sheaf cohomology.

For the second assertion use the hypercohomology spectral sequence proved there,

Its terms vanish for \(p>d\), for \(p<0\), and for \(q>b\). Since \(K\) is bounded below, each total degree has the finite filtration used in that comparison. If \(m>b+d\), every term with \(p+q=m\) is zero already on the second page, so every associated graded piece of \(\mathbb H^m(X,K)\) is zero. A finite filtration with zero pieces has zero underlying group. Apply Corollary 4.1 for a closed subspace. \(\square\)

For example, the case \(d=0\) makes global sections exact for all module sheaves: its higher derived functors vanish, and the long exact sequence makes every epimorphism surjective on global sections. This is a consequence of the topological refinement theorem, with no fine or soft hypothesis on the individual sheaf.

5. Exercises with complete solutions

Exercise 5.1 (Foundational: dimension zero and clopen separation). Prove that \(\dim X\leq0\) if and only if every pair of disjoint closed subsets \(A,B\) admits a clopen set \(D\) with \(A\subseteq D\subseteq X\setminus B\).

Solution. Suppose first that \(\dim X\leq0\). Refine the finite open cover \((X\setminus A,X\setminus B)\) by an open cover of multiplicity at most one. Its distinct members are disjoint. Group them by their assigned original member. Let \(D\) be the union of the members assigned to \(X\setminus B\). A point of \(A\) cannot belong to a member assigned to \(X\setminus A\), so it belongs to \(D\); and \(D\cap B=\varnothing\). The other group is exactly \(X\setminus D\), which is open. Thus \(D\) is clopen.

Conversely, take a finite open cover \((U_i)_{i=1}^m\). Finite shrinking gives closed sets \(F_i\subseteq U_i\) covering \(X\). The assumed separation property applied to \(F_i\) and \(X\setminus U_i\) gives a clopen \(D_i\) with \(F_i\subseteq D_i\subseteq U_i\). The \(D_i\) cover. Set

These are clopen, pairwise disjoint, subordinate to \(U_i\), and cover \(X\): each point belongs to the first \(D_i\) containing it. They are a refinement of multiplicity at most one, so \(\dim X\leq0\). This proof uses separation of arbitrary closed pairs; it does not replace that property by an unproved assertion about a clopen base. \(\square\)

Exercise 5.2 (Intermediate: why limit stages need control). Show that the closed family \(C_n=\{1/n\}\) in \(\mathbb R\), \(n\geq1\), has no locally finite family of neighborhoods \(O_n\supseteq C_n\). Give a decreasing sequence of two-member open covers whose memberwise limit contains a nonopen member, and identify the obstruction to using that recursion in Theorem 3.1.

Solution. Every neighborhood of zero contains \(1/n\) for infinitely many \(n\). It therefore meets infinitely many \(C_n\), and hence infinitely many \(O_n\). Thus neither family is locally finite at zero.

For the second assertion use the covers

Each is an open cover; the second member is fixed, and the first decreases. Their memberwise intersection is the covering family \((\{0\},\mathbb R\setminus\{0\})\), whose first member is not open. A finite indexed cover is locally finite at each stage, so this also shows that stagewise local finiteness is insufficient. Around zero the first member changes at arbitrarily late stages; there is no locally finite family of modification supports yielding local stabilization. The limit argument of Theorem 3.1 supplies exactly that additional property. \(\square\)

Exercise 5.3 (Advanced: an arbitrary disjoint union of circles). Let \(J\) be a nonempty set and \(X=\coprod_{j\in J}S^1\). Prove \(\dim X=1\), and for any abelian group \(A\) compute \(H^q(X,A_X)\), where \(A_X\) is the constant sheaf with fiber \(A\). Do not assume \(J\) finite or countable.

Solution. Each circle is compact Hausdorff. A finite subcover of any cover of a circle is a locally finite refinement, so the circle is paracompact. Refining a cover separately on every open component shows that \(X\) is paracompact: near a point only its own component occurs. It is Hausdorff because distinct components are disjoint open sets and each component is Hausdorff.

The circle-cover construction in Exercise 3 of the Cech-comparison lesson refines every open cover of \(S^1\) by a finite open cover of multiplicity at most two. Apply it on each component. Their union is locally finite in \(X\), refines the original cover, and has multiplicity at most two. Theorem 3.1 gives \(\dim X\leq1\).

For completeness, sheaves on a disjoint union are exactly families of sheaves on its components. Restriction gives one direction; in the other direction define sections over an open \(U\) to be the product of the sections over \(U\cap S^1_j\). The sheaf gluing axiom follows component by component, and these constructions are inverse. Kernels, cokernels and exactness are componentwise. A family of injectives is injective in this product category: extend a map from a subobject separately in each component. Thus componentwise injective resolutions compute

Here products of abelian groups are exact: kernels are componentwise, and a product of surjections is surjective by choosing a lift in each coordinate. Hence taking a product of the resolution complexes commutes with cohomology. Exercise 3 of the Cech-comparison lesson computes the circle groups, giving

Take \(A=\mathbb Z\). Since \(J\neq\varnothing\), the degree-one product is nonzero. Corollary 4.2 would make it zero if \(\dim X\leq0\). Therefore \(\dim X=1\). The refinement argument used no finite subcover of the whole disjoint union. \(\square\)

Exercise 5.4 (Advanced: swelling with variable radii). In \(\mathbb R\) put \(a_n=n+1/3\), \(\delta_n=1/(10n)\), \(b_n=a_n+\delta_n\), and \(A=\{a_n:n\geq1\}\), \(B=\{b_n:n\geq1\}\). Show that \(A,B\) are disjoint closed sets with distance zero. Explain why every pair of uniform positive-radius neighborhoods intersects. Construct open neighborhoods whose closures are disjoint.

Solution. Every bounded interval meets only finitely many of the listed points, since both sequences tend to \(+\infty\). Thus \(A\) and \(B\) are closed: near any point outside either set there are only finitely many candidate points to avoid. Each pair \(a_n,b_n\) consists of distinct points in \((n,n+1)\), and different indices lie in different such intervals, so \(A\cap B=\varnothing\). Their distance is zero because \(|a_n-b_n|=\delta_n\to0\).

If the two uniform radii are \(\epsilon_A,\epsilon_B>0\), choose \(n\) with \(\delta_n<2\min(\epsilon_A,\epsilon_B)\). The midpoint of \(a_n,b_n\) belongs to both neighborhoods. In particular, choosing a single sufficiently small positive radius cannot separate the sets.

Instead set

These are open neighborhoods. The two closed intervals belonging to one index are disjoint, because the sum of their radii is \(2\delta_n/3<\delta_n\). All intervals for index \(n\), including their endpoints, lie in \((n,n+1)\); indeed the smallest endpoint is at least \(n+3/10\), and the largest is at most \(n+7/15<n+1\). Thus intervals for different indices cannot meet. Both interval families are locally finite, so their unions of closed intervals are closed and equal to \(\overline{G_A}\) and \(\overline{G_B}\). These closures are disjoint. The finite swelling principle allows neighborhoods varying with location, exactly as this example requires. \(\square\)

6. Further reading

C. H. Dowker, Lectures on Sheaf Theory, Lecture 11, printed page 57, connects the finite-cover formulation of dimension with refinements of locally finite covers and sheaf-cohomological vanishing. The finite-cover dimension convention and the elementary finite shrinking and swelling used above also appear in the linked lesson on Covering dimension and finite trivializing covers.