I have a question about partitions of unity specifically in the book Calculus on Manifolds by Spivak. In case 1 for the proof of existence of partition of unity, why is there a need for the function $f$? The set $\Phi = \{\varphi_1, \dotsc, \varphi_n\}$ looks like is already the desired partition of unity. Following is the theorem and proof. Only Case 1 in the proof is relevant.

enter image description here enter image description here

  • $\begingroup$ Should I include pictures of the two pages with the theorem and proof? $\endgroup$ – Pratyush Sarkar Aug 24 '13 at 13:24
  • $\begingroup$ That would certainly help; or at least type in the relevant details (what are the $\phi_i, f$ and what is the proposed partition of unity). $\endgroup$ – Anthony Carapetis Aug 24 '13 at 14:00
  • $\begingroup$ I am reading calculus on manifolds too. And as I came across this proof today, I have exactly the same doubt. Glad that I find this post. $\endgroup$ – A. Chu Jul 10 '15 at 14:16
  • $\begingroup$ On what set is $\sigma$ define din the second case? Because I do not see how how defining $\phi / \sigma$ gives a partition of unity, or even how the division makes sense. $\endgroup$ – vaoy Apr 18 '18 at 10:51
  • $\begingroup$ @vaoy $\Phi_i$ is a partition of unity subordinate to $\mathcal O_i$, so we can assume $\varphi \in \Phi_i$ is defined on $A$ and zero outside $\operatorname{int}(A_{i + 1}) - A_{i - 2}$. And $\sigma$ is also defined on $A$ and as explained, the sum in the definition is at most finite for all $x \in A$ and necessarily positive, so division is possible. Note that $\sigma$ is just a scaling factor to make sure the new $\varphi'$ sum up to $1$ at each $x \in A$. This scaling is required because although $\Phi_i$ is a partition of unity, $\Phi_j$ for $j = i + 1, i - 1$ can interfere with $\Phi_i$. $\endgroup$ – Pratyush Sarkar Apr 24 '18 at 18:00

I belive that your assertion is correct. The functions $\varphi_{i}$ satisfy all of the conditions of Theorem 3-11. I don't see why Spivak used such an $f$. Particularly since the support of $f$ contains $A$. If at least the support of $f$ lied in $A$ then $f=\sum_{i=1}^{n}f\cdot\varphi_{i}$, thus giving a representation of $f$ as a sum of functions with small supports.

Since $A$ is compact we may assume WLOG that the $U_{i}$ are bounded. Therefore, by construction, the supports of the $\psi_{i}$ are compact. Hence, the word "closed" in item ($4$) of Theorem 3-11 can be changed to "compact". The proof remains unchanged. This helps clarify the first statement of the proof of Theorem 3-12.

Also, note as well that the functions $\varphi_{i}$ are $C^{\infty}$. This basically follows from Problem 2-26.

Posts related to the section:

  1. An application of partitions of unity: integrating over open sets and here

  2. Do we need additional assumptions for problem 3-37 (b) in Spivak´s calculus on Manifolds?

  3. Problem 3-38 in Spivak´s Calculus on Manifolds

  4. Extended integral in Spivak’s Calculus on Manifolds

  • $\begingroup$ Firstly, in property 4, isn't the closed set bounded? Which would mean it's compact? Secondly, in the proof it says "let $\psi_i$ be non-negative $C^\infty$ function positive on $D_i$ and $0$ outside some closed set contained in $U_i$". Denote the closed set as $S_i$. Then $\varphi_i$ has the same property, namely there is the set $U_i$ such that $\varphi_i = 0$ outside $S_i$ contained in $U_i$. So isn't property 4 already fulfilled? $\endgroup$ – Pratyush Sarkar Aug 24 '13 at 15:16
  • $\begingroup$ @Pratyush: I have looked at this more carefully and updated my answer. $\endgroup$ – John Aug 24 '13 at 21:10
  • $\begingroup$ Thanks for your help. I read the theorem and proof multiple times before and I couldn't figure out the purpose of the function $f$. I thought I was missing something really obvious. $\endgroup$ – Pratyush Sarkar Aug 24 '13 at 22:51

I was looking back at my question today for some reason and immediately saw why the function $f$ is required. Although $\psi_i$ is smooth with compact support in $U_i$, the functions $\varphi_i$ can only be defined on $U$ where $\sum_{i = 1}^n\psi_i > 0$. The problem is that $\varphi_i$ usually does not go to zero at the boundary of $\operatorname{supp}(\psi_i)$ (much less smoothly extend to zero outside the boundary). You can see this near the boundary of a $\operatorname{supp}(\psi_i)$ which is away from all other $\operatorname{supp}(\psi_j)$. Near this boundary $\varphi_i(x) = \frac{\psi_i(x)}{\psi_i(x)} = 1$. The solution is to use a cutoff function $f$ which forces everything to smoothly go to $0$ near the boundary of $U$.

  • $\begingroup$ I think you mean $\varphi_{i}$ (not $\psi_{i}$) does not usually go to zero at the boundary of $\text{supp}(\psi_{i})$. $\endgroup$ – fourierwho May 31 '16 at 16:37
  • $\begingroup$ @fourierwho Yes. Thanks for the correction. $\endgroup$ – Pratyush Sarkar Jul 25 '16 at 23:14
  • $\begingroup$ @PratyushSarkar, the fact of $\phi_i$ doesn't smoothly extend to zero outside the boundary of $\operatorname{supp} (\psi_i)$ doesn't contradicts the fact of $\psi_i$ are $C^{\infty}$ by the way was built? If your answer is not, so I think that I didn't understand the purpose of $f$. $\endgroup$ – George Jan 29 '17 at 18:28
  • $\begingroup$ @George No there is no contradiction. What I explained is, although $\psi_i$ obviously extends to $0$ smoothly outside $\operatorname{supp} (\psi_i)$, the biggest set where we can define $\varphi_i$ is only on $U = \bigcup_{i = 1}^n \operatorname{supp} (\psi_i)$ and it is possible that near the boundary of $U$ which, say, coincides with $\operatorname{supp} (\psi_i)$ and is away from all other $\operatorname{supp} (\psi_j)$, the value of $\varphi_i(x)$ tends to $1$. $\endgroup$ – Pratyush Sarkar Feb 17 '17 at 5:56
  • $\begingroup$ More explicitly, we can easily think of (and draw) an example where $B_i = \partial U \setminus \left(\bigcup_{j \neq i} \operatorname{supp} (\psi_j)\right)$ is nonempty and in that case we have $\lim_{x \to b} \varphi_i(x) = 1$ for any $b \in B_i$ since $\varphi_i(x) = \frac{\psi_i(x)}{\psi_i(x)} = 1$ for $x$ sufficiently close to $b$. $\endgroup$ – Pratyush Sarkar Feb 17 '17 at 6:03

Merry Christmas ! I am sorry to bother you.I revised my answer.Looking foward to your comments.

First of all, as Mr.John mentioned in his answer,the functions $\varphi_{i}$ have already satisfied all of the conditions of Th 3-11 but except for $\textit{(4)}$. A crucial question in Spivak's proof is: why the author "redundantly" required $f\cdot\varphi_{i}$ rather than $\varphi_{i}$ ? In my way of thinking , in accord with problem 2-26$^*$(d),the author wanted to extend $\varphi_{i}$ from its domain$-\ $an open subset in $\mathbf {R}^{n}$$-\ $to the whole of $\mathbf {R}^{n}$ smoothly and then guarantee such those extended functions $f\cdot \varphi_{i}$ satisfying all of four conditions.

The following is details:

Def. $\;\phi:X \rightarrow \mathbf {R}$ is a continuous real-valued function whose domain is an arbitrary set $X$ in $\mathbf {R}^{n}$, the support of $\phi$ is defined as the closure of the subset of $X$ where $\phi$ is non-zero $i.e.,\:$$supp(\phi):=$the closure of the set $\left\{\mathbf{x} \in X: \phi(\mathbf{x})\ne0 \right \}.$

Def. $\;\phi\in C^{\infty}(A,B)$ denotes $\phi:(\mathbf {R}^{n}\supset )A\rightarrow B(\subset\mathbf {R}^{m})$ is a $C^{\infty}$ function.

From the case 1,the compact sets $D_{i}(i=1,\cdots,n)$ whose interiors cover $A$.In order to explain why the author modified $\varphi_{i}$ with $f\cdot \varphi_{i}$ briefly,we can choose a specific open subset$-\ U:=\bigcup_{i=1}^{n}int D_{i}$$-\ $to set forth.If in the simplest case we need $\varphi_{i}$ multiplied by $f$,not to mention in most cases.


$\psi_i\in C^{\infty}(U_i,\mathbf{R}),$ which is positive on $D_i$ and $0$ outside of some closed set contained in $U_i$ .(problem 2-26$^*$(d))

Define $$\widetilde \psi_{i}:=\left\{\begin{matrix} \psi_{i}& x\in U_i\\ 0& x\in {U_i}^{c} \end{matrix}\right.\quad (i=1,2,\cdots,n),$$ then we have $\widetilde \psi_{i}\in C^{\infty}(\mathbf{R}^{n},\mathbf{R})$,$\;\widetilde \psi_{i}\bigg|_{U_{i}}=\widetilde \psi_{i}$ and $\psi_{i}$ (each domain is $U_{i}$) can be smoothly extended to $\widetilde \psi_{i}$(each domian is $\mathbf{R}^{n}$).

All functions $$\varphi_{i}=\frac{\psi_{i}}{\sum_{k=1}^{n}\psi_{k}}\;(i=1,\cdots,n)$$ are only constructed on $U$. $$ \forall \: x\in U,\: \sum_{k=1}^{n}\widetilde\psi_{k}\ne 0 ;\;\widetilde\psi_{k}\in C^{\infty}(\mathbf{R}^{n},\mathbf{R}).$$ $$\Longrightarrow \varphi_{i}=\frac{\psi_{i}}{\sum_{k=1}^{n}\psi_{k}}=\frac{\widetilde\psi_{i}}{\sum_{k=1}^{n}\widetilde\psi_{k}}\in C^{\infty}(U,\mathbf{R}) \quad (i=1,2,\cdots,n).$$


$f\in C^{\infty}(U,\mathbf{R}),$ which value is $1$ on $A$ and $0$ outside of some closed set contained in $U$.(problem 2-26$^*$(d))

$\\$Obviously,we have $f\cdot \varphi_{i}\in C^{\infty}(U,\mathbf{R})\;(i=1,2,\cdots,n).$ Note that each $f\cdot \varphi_{i}\in C^{\infty}(U,\mathbf{R})$ is only constructed on $U\:!$


Finally,we can extend $f\cdot \varphi_{i}$ from $U$ to the whole of $\mathbf {R}^{n}$ smoothly.

In fact,since $supp(f\cdot\varphi_{i})\subset U$, define $$\widetilde{f\cdot \varphi_{i}}:=\left\{\begin{matrix} f\cdot \varphi_{i}& x\in supp(f\cdot\varphi_{i})\\ 0 & x\notin supp(f\cdot\varphi_{i}) \end{matrix}\right.\quad (i=1,2,\cdots,n),$$ then we have $\widetilde {f\cdot \varphi_{i}}\in C^{\infty}(\mathbf{R}^{n},\mathbf{R})$,$\;\widetilde {f\cdot \varphi_{i}}\bigg|_{U}=f\cdot\varphi_{i}$ and $f\cdot \varphi_{i}$ (each domain is $U$) can be smoothly extended to $\widetilde {f\cdot \varphi_{i}}$ (each domian is $\mathbf{R}^{n}).\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\blacksquare$

A schematic diagram (A schematic diagram concretising above elaboration)

  • $\begingroup$ Thanks for elaborating, this makes sense. A few comments: 1. Actually the functions $\varphi_i$ do not satisfy property 4 of the theorem which is why we need $f$ to kill it off near some parts of the boundary of $U_i$. 2. In your method, $U$ is defined slightly differently but the result is the same, support of $\widetilde{f\cdot \varphi_{i}}$ is contained inside $U_i$ as required by property 4. I want to point out that this can't be made stronger to support contained in $D_i$. So I think I prefer the $U$ as defined by Spivak because it is easier to see the containment in $U_i$. $\endgroup$ – Pratyush Sarkar Dec 26 '18 at 23:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.