nLab partition of unity




topology (point-set topology, point-free topology)

see also differential topology, algebraic topology, functional analysis and topological homotopy theory


Basic concepts

Universal constructions

Extra stuff, structure, properties


Basic statements


Analysis Theorems

topological homotopy theory



A partition of unity is a partition of the unit function on a topological space into a sum of continuous functions that are each non-zero only on small parts of the space.


Let XX be a topological space. A partition of unity on XX is a collection {u j} jJ\{u_j\}_{j \in J} of continuous functions u j:X[0,1]u_j \colon X \to [0,1], jJj\in J to the closed interval with its Euclidean metric topology such that jJu j(x)=1\sum_{j \in J} u_j(x) = 1 for all xXx\in X.

A partition of unity defines an open cover of XX, consisting of the open sets u j 1(0,1]u_j^{-1}(0,1]. Call this the induced cover.

Given a cover 𝒰={U j} jJ\mathcal{U} = \{U_j\}_{j\in J} of a topological space (open cover or closed or neither), the partition of unity {u j} J\{u_j\}_J is subordinate to 𝒰\mathcal{U} if for all jJj\in J,

u j 1(0,1]¯U j.\overline{u_j^{-1}(0,1]} \subset U_j.

What this means is that the open sets u j 1(0,1]u_j^{-1}(0,1] form an open cover refining the cover 𝒰\mathcal{U}.

A partition of unity is point finite if for every xXx\in X there is only a finite number of jJj\in J such that u j(x)0u_j(x) \neq 0.

A partition of unity is locally finite if for every xXx\in X there is an open neighborhood UU of xx such that for only a finite number of jJj\in J there is xUx\in U such that u j(x)0u_j(x) \neq 0.

Often, the property of local finiteness is included in the definition of a partition of unity. This is harmless, since a result due to Michael R. Mather (Prop. below) says that for any partition of unity we can find a locally finite partition of unity with the same indexing set and whose induced cover refines the original induced cover.



Consider \mathbb{R} with its Euclidean metric topology.

Let ϵ(0,)\epsilon \in (0,\infty) and consider the open cover

{(n1ϵ,n+1+ϵ)} n. \{ (n-1-\epsilon , n+1 + \epsilon) \subset \mathbb{R} \}_{n \in \mathbb{Z} \subset \mathbb{R} }.

Then a partition of unity {f n:[0,1]} n\{ f_n \colon \mathbb{R} \to [0,1] \}_{n \in \mathbb{N}} subordinate to this cover is given by

f n(x){x(n1) | n1xn 1(xn) | nxn+1 0 | otherwise}. f_n(x) \coloneqq \left\{ \array{ x - (n - 1) &\vert& n - 1 \leq x \leq n \\ 1- (x-n) &\vert& n \leq x \leq n+1 \\ 0 &\vert& \text{otherwise} } \right\}.


Existence on paracompact topological spaces


(paracompact Hausdorff spaces equivalently admit subordinate partitions of unity)

Assuming the axiom of choice then:

Let (X,τ)(X,\tau) be a topological space. Then the following are equivalent:

  1. (X,τ)(X,\tau) is a paracompact Hausdorff space.

  2. Every open cover of (X,τ)(X,\tau) admits a subordinate partition of unity.

Similarly normal spaces are equivalently those such that every locally finite cover has a subordinate partition of unity (reference Bourbaki, Topology Generale - find this!)

The case of non-Hausdorff spaces

Slightly more generally, a topological space (not necessarily Hausdorff) is fully normal if and only every open cover admits a subordinate partition of unity.

A T1-space is fully normal if and only if it is paracompact, in which case it is also Hausdorff.

For topological spaces that are not T1-spaces, the condition of being fully normal is strictly stronger than paracompactness.

The case of locales

A regular locale is fully normal if and only if it is paracompact.

The usual proof of the existence of partitions of unity goes through for such locales since it does not make any use of points.

Existence on smooth manifolds

Paracompact smooth manifolds admit locally finite smooth partitions of unity subordinate to any open cover (this follows from the existence of a smooth bump function on [1,1][-1,1]). It is not true, however, that analytic manifolds have analytic partitions of unity - the aforementioned bump function is smooth but not analytic:


(open cover of smooth manifold admits locally finite refinement by closed balls)

Let XX be a smooth manifold and let {U iX} iI\{U_i \subset X\}_{i \in I} be an open cover. Then there exists cover

{B 0(ϵ j)ψ jV jX} iJ \left\{ B_0(\epsilon_j) \underoverset{\simeq}{\psi_j}{\to} V_j \subset X \right\}_{i \in J}

which is a locally finite refinement of {U iX} iI\{U_i \subset X\}_{i \in I} with each patch diffeomorphic to a closed ball in Euclidean space.


First consider the special case that XX is compact topological space.


{ nϕ jV jX} \left\{ \mathbb{R}^n \underoverset{\simeq}{\phi_j}{\longrightarrow} V_j \subset X \right\}

be a smooth atlas representing the smooth structure on XX. The intersections

{U iV j} iI,jJ \left\{ U_i \cap V_j \right\}_{i \in I, j \in J}

still form an open cover of XX. Hence for each point xXx \in X there is iIi \in I and jJj \in J with xU iV jx \in U_i \cap V_j. By the nature of the Euclidean topology, there exists a closed ball B xB_x around ϕ j 1(x)\phi_j^{-1}(x) in ϕ j 1(U iV j) n\phi_j^{-1}(U_i \cap V_j) \subset \mathbb{R}^n. Its image ϕ j(B x)X\phi_j(B_x) \subset X is a neighbourhood of xXx \in X diffeomorphic to a closed ball.

The interiors of these balls form an open cover

{Int(B x)X} xX \left\{ Int(B_x) \subset X \right\}_{x \in X}

of XX which, by construction, is a refinement of {U iX} iI\{U_i \subset X\}_{i \in I}. By the assumption that XX is compact, this has a finite subcover

{Int(B l)X} lL \left\{ Int(B_l) \subset X \right\}_{l \in L}

for LL a finite set. Hence

{B lX} lL \left\{ B_l \subset X \right\}_{l \in L}

is a finite cover by closed balls, hence in particular locally finite, and by construction it is still a refinement of the orignal cover. This shows the statement for XX compact.

Now for general XX, notice that without restriction we may assume that XX is connected, for if it is not, then we obtain the required refinement on all of XX by finding one on each connected component.

But if a locally Euclidean paracompact Hausdorff space XX is connected, then it is sigma-compact and in fact admits a countable increasing exhaustion

V 0V 1V 2 V_0 \subset V_1 \subset V_2 \subset \cdots

by open subsets whose topological closures

K 0K 1K 2 K_0 \subset K_1 \subset K_2 \subset \cdots

exhaust XX by compact subspaces K nK_n (by the proof of this prop.).

For nn \in \mathbb{N}, consider the open subspace

V n+2K n1X V_{n+2} \setminus K_{n-1} \;\subset\; X

which canonically inherits the structure of a smooth manifold (this prop.). As above we find a refinement of the restriction of {U iX} iI\{U_i \subset X\}_{i \in I} to this open subset by closed balls and since the further subspace K n+1K nK_{n+1}\setminus K_n is still compact (by this lemma) there is a finite set L nL_n such that

{B l nV n+2K n1X} l nL n \{B_{l_n} \subset V_{n+2} \setminus K_{n-1} \subset X \}_{l_n \in L_n}

is a finite cover of K n+1K nK_{n+1} \setminus K_n by closed balls refining the original cover.

It follows that the union of all these

{B l nX} n,l nL n \left\{ B_{l_n} \subset X \right\}_{n \in \mathbb{N}, l_n \in L_n}

is a refinement by closed balls as required. Its local finiteness follows by the fact that each B l nB_{l_n} is contained in the “strip” V n+2K n1V_{n+2} \setminus K_{n-1}, each strip contains only a finite set of B l nB_{l_n}-s and each strip intersects only a finite number of other strips. (Hence an open subset around a point xx which intersects only a finite number of elements of the refined cover is given by any one of the balls B l nB_{l_n} that contain xx.)


(smooth manifolds admit locally finite smooth partitions of unity)

Let XX be a paracompact smooth manifold. Then every open cover {U iX} iI\{U_i \subset X\}_{i \in I} has a subordinate partition of unity by functions {f i:U i} iI\{f_i \colon U_i \to \mathbb{R}\}_{i \in I} which are smooth functions.


By lemma the given cover has a locally finite refinement by closed subsets diffeomorphic to closed balls:

{B 0(ϵ j)ψ jV jX} jJ. \left\{ B_0(\epsilon_j) \underoverset{\simeq}{\psi_j}{\to} V_j \subset X \right\}_{j \in J} \,.

Given this, let

h j:X h_j \;\colon\; X \longrightarrow \mathbb{R}

be the function which on V jV_j is given by a smooth bump function

b j: b_j \;\colon\; \mathbb{R} \longrightarrow \mathbb{R}

with support supp(b j)=B 0(ϵ j)supp(b_j) = B_0(\epsilon_j):

h j:x{b j(ψ j 1(x)) | xV j 0 | otherwise. h_j \;\colon\; x \mapsto \left\{ \array{ b_j(\psi_j^{-1}(x)) &\vert& x \in V_j \\ 0 &\vert& \text{otherwise} } \right. \,.

By the nature of bump functions this is indeed a smooth function on all of XX. By local finiteness of the cover by closed balls, the function

h:X h \;\colon\; X \longrightarrow \mathbb{R}

given by

h(x)jJh j(x) h(x) \coloneqq \underset{j \in J}{\sum} h_j(x)

is well defined (the sum involves only a finite number of non-vanishing contributions) and is smooth. Therefore setting

f jh jh f_j \;\coloneqq\; \frac{h_j}{h}


{f j} jJ \left\{ f_j \right\}_{j \in J}

is a subordinate partition of unity by smooth functions as required.

From a partition of unity to a locally finite partition of unity


A collection of functions 𝒰={u i:X[0,1]}\mathcal{U} = \{u_i : X \to [0,1]\} such that every xXx\in X is in the support of some u iu_i. Then 𝒰\mathcal{U} is called locally finite if the cover u i 1(0,1]u_i^{-1}(0,1] (i.e. the induced cover) is locally finite.


(Mather, 1965)

Let {u i} J\{u_i\}_J be a partition of unity. Then there is a locally finite partition of unity {v i} iJ\{v_i\}_{i\in J} such that the induced cover of the latter is a refinement of the induced cover of the former.

The proof is Proposition A.2.8 in Dold 95 and Lemma 5.1.8 on page 301 of Engelking 89.

This implies that (locally finite) numerable covers are cofinal in induced covers arising from collections of functions as in the definition. In particular, given the Milnor classifying space MG\mathcal{B}^M G of a topological group GG, which comes with a countable family of ‘coordinate functions’ MG[0,1]\mathcal{B}^M G \to [0,1], has a numerable cover. This is shown by Dold to be a trivialising cover for the universal bundle constructed by Milnor, and so the universal bundle is numerable.


Maps to geometric realizations

Partitions of unity can be used in constructing maps from spaces to geometric realizations of simplicial spaces (incl. simplicial sets) - for example a classifying map for a GG-bundle where GG is a Lie group.

Coboundaries for Cech cocycles

Partitions of unity can be used to give explicit coboundaries for the cocycles of the complex of functions on a cover.

Let {U iX}\{U_i \to X\} be a open cover and {ρ iC(X,)}\{\rho_i \in C(X,\mathbb{R})\} a collection of functions with

  • (xU i)ρ i(x)=0(x \notin U_i) \Rightarrow \rho_i(x) = 0

  • iρ i=const 1\sum_i \rho_i = const_1.

Write C({U i}):Δ opTopC(\{U_i\}) : \Delta^{op} \to Top for the Cech nerve of the cover and C(C({U i}),)C(C(\{U_i\}), \mathbb{R}) for the cosimplicial ring of functions on this simplicial topological space; and (C (C({U i}),),δ)(C_\bullet(C(\{U_i\}), \mathbb{R}), \delta) for the corresponding (normalized) cochain complex: its differential is the alternating sum of the pullbacks of functions along the face maps, i.e. along the restriction maps

δ= k(1) kδ k *. \delta = \sum_k (-1)^k \delta_{k}^* \,.

For instance for f={f i 1,i 2,,i nC(U i 1U i n+1)}f = \{f_{i_1, i_2, \cdots, i_n} \in C(U_{i_1} \cap \cdots \cap U_{i_{n+1}})\} a collection of functions in degree nn, we have

(δf) i 0i ni n+1= k=0 n+1(1) kf i 0i k1i k+1i n+1. (\delta f)_{i_0 \cdots i_n i_{n+1}} = \sum_{k = 0}^{n+1} (-1)^k f_{i_0 \cdots i_{k-1} i_{k+1} \cdots i_{n+1}} \,.

This cochain complex has vanishing cochain cohomology in positive degree. We can explicitly construct corresponding coboundaries using the partion of unity:

assume that with the above notation ff is a cocycle in positive degree, in that δf=0\delta f = 0. Then define the (n1)(n-1)-cochain

λ i 1i n:= i 0ρ i 0f i 0i 1i n. \lambda_{i_1 \cdots i_n} := \sum_{i_0} \rho_{i_0} f_{i_0 i_1 \cdots i_n} \,.

Here in the summands on the right the product is defined on U i 0U i 1U i nU_{i_0} \cap U_{i_1} \cap \cdots \cap U_{i_n} and extended as 0 to all of U i 1U i nU_{i_1} \cap \cdots \cap U_{i_n}.

With this definition we have

δλ=f. \delta \lambda = f \,.

To see this we compute

(δλ) i 1i n+1 := i 0ρ i 0 k=1 n(1) kf i 0i 1i k1i k+1i n+1 =± i 0ρ i 0f i 1i n+1 =f i 1i n+1, \begin{aligned} (\delta \lambda)_{i_1 \cdots i_{n+1}} & := \sum_{i_0} \rho_{i_0} \sum_{k=1}^n (-1)^k f_{i_0 i_1 \cdots i_{k-1} i_{k+1} \cdots i_{n+1}} \\ & = \pm \sum_{i_0} \rho_{i_0} f_{i_1 \cdots i_{n+1}} \\ & = f_{i_1 \cdots i_{n+1}} \end{aligned} \,,

where in the second step we used the condition δf=0\delta f = 0 and in the last step we used the property of the partition of unity.

This construction is used a lot in Cech cohomology. For instance it can be used to show in Chech cocycles that every principal bundle admits a connection on a bundle (see there for the details).


Discussion of partitions of unity in constructive mathematics is in

Last revised on May 20, 2022 at 02:53:58. See the history of this page for a list of all contributions to it.