0
$\begingroup$

I am looking for a dense sub-algebra $B$ in $C_{b}((0,1))$ in uniform topology such that it satisfy following requirements:

  1. $B\cap C^{\infty}_{b}((0,1))=\mathbb{R}$ (No polynomial, no bump function).

$(0,1)$ can be replaced with other connected open set. The only thing I can come up with is nowhere differentiable function union with $\mathbb{R}$, but I don't know if they form a sub-algebra.

$\endgroup$
5
  • 2
    $\begingroup$ Since $0$ is not nowhere differentiable, the nowhere differentiable functions do not form a subalgebra. What does "other connected open set is fine" mean? $\endgroup$
    – LSpice
    Feb 12, 2021 at 18:43
  • 1
    $\begingroup$ If $f$ and $g$ are positive with $f$ nowhere differentiable and $g$ smooth, then $g/f$ is nowhere differentiable but $f\cdot g/f$ is smooth, so what you describe is not an algebra. $\endgroup$
    – Pierre PC
    Feb 12, 2021 at 19:15
  • 1
    $\begingroup$ I would think of something like starting with a single function $f$ which is strictly increasing and not differentiable, and consider the algebra it generates, which is the set of all polynomials in $f$. By Stone-Weierstrass it will be dense. The question is then in effect whether $p(f)$ could be smooth and non-constant for some polynomial $p$. It seems unlikely. $\endgroup$ Feb 12, 2021 at 20:55
  • 1
    $\begingroup$ @NateEldredge I think the argument in my answer is enough to show that such $p$ does not exist. However, I am not sure $\mathbb R[f]$ in dense, since the underlying space is not compact, and we do not assume the functions are zero at infinity. $\endgroup$
    – Pierre PC
    Feb 12, 2021 at 21:35
  • 1
    $\begingroup$ Oh, good point. It would be dense in $C([0,1])$ but of course $C_b((0,1))$ contains functions which do not continuously extend to $[0,1]$, so it cannot approximate such functions. $\endgroup$ Feb 12, 2021 at 21:37

1 Answer 1

2
$\begingroup$

Such a subalgebra can be defined. See also the end of this answer for a more general result.

Let $I=(0,1)$. In the following, an open interval means a nonempty open interval of $I$. Fix a function $f\in C_b(I)$ with values in $[1,2]$ such that $f$ is not smooth over any open interval; for instance, $f:t\mapsto\alpha W(t)+\beta$, for $W$ a realisation of Brownian motion. Set $B$ the subset of $C_b(I)$ consisting of functions of the form $$a_nf^n+\cdots+a_1f+a_0,$$ with $a_1,\ldots,a_n$ smooth and $a_0$ constant. $B$ is clearly a subalbegra of $C_b(I)$. We need to show that it is dense, and that its smooth functions have to be constants.

The following fact follows from considering a smooth partition of unity over $I$.

Fact. Given $h,\delta\in C_b(I)$ such that $\delta>0$ over $I$, there exists $\tilde h\in C^\infty(I)$ such that $|h-\tilde h|<\delta$ over $I$.

In particular, it will be enough to show that the closure of $B$ in $C_b(I)$ contains all smooth functions. Let $g$ be a smooth function of $C_b(I)$ and $\varepsilon>0$. We are looking for a function in $B$ within distance $\varepsilon$ of $g$. Using the above fact, let $\tilde f\in C^\infty(I)$ such that $2|f-\tilde f|<\min(1,\varepsilon/|g|)$. Of course $\tilde f\geq f-|f-\tilde f|\geq 1/2$. Then $$ \left|g-\frac g{\tilde f}f\right|\leq\frac{|g|}{|\tilde f|}\cdot|\tilde f-f|<\frac{|g|}{1/2}\cdot\frac\varepsilon{2|g|}=\varepsilon,$$ so $g$ is at distance at most $\varepsilon$ of $(g/\tilde f)f\in B$, which concludes the density argument.

To show that smooth functions of $B$ are constant, here is a lemma that we'll prove later.

Lemma. If $h\in C_b(I)$, $b_n,\ldots,b_0\in C^\infty(I)$ and $h(x)$ is a root of $b_n(x)z^n+\cdots+b_0(x)$ for all $x\in I$, then either $h$ is smooth over an open interval of $I$ or the $b_i$ are all zero everywhere.

Choose some $g\in B$. We can write it as $a_nf^n+\cdots+a_1f+a_0$ for $a_1,\ldots,a_n$ smooth and $a_0$ constant. Now if $g$ is smooth, then $f(x)$ is a root of $a_n(x)z^n+\cdots+a_1(x)z+(a_0-g)(x)$ for all $x$, with all coefficients smooth. Since $f$ is not smooth over any open interval of $I$, the above lemma shows that all coefficients of this polynomial are zero, and in particular $g=a_0$ is a constant.

It remains to prove the lemma. I will work by induction over $n$. If $n=0$, the result is obvious since $b_0$ has to be zero everywhere. Suppose the result is true for a given $n$, and let us show it is true for $n+1$.

Let $h$ and $b_{n+1},\ldots b_0$ be as assumed in the lemma.

  1. If there exists an $x_0$ such that $h(x_0)$ is a simple root of $b_{n+1}(x_0)z^{n+1}+\cdots+b_0(x_0)$, a classical application of the inverse function theorem shows that $h(x)$ depends smoothly on the vector $(b_0(x),\ldots,b_{n+1}(x))$, which itself depends smoothly on $x$, so $h$ is smooth over a neighbourhood of $x_0$.

  2. If such an $x_0$ does not exist, it means that for all $x$, $h(x)$ is a root of the derivative $(n+1)b_{n+1}(x)z^n+\cdots+2b_2(x)z+b_1(x)$. By the induction hypothesis, either $h$ is smooth on an open interval and we are done, or all functions $b_1$ to $b_{n+1}$ are zero everywhere. Since $h(x)$ is now a root of the constant polynomial $b_0(x)$ for all $x$, $b_0$ is zero as well, which concludes.


Let us rewrite that result in an abstract-ish context. Let $X$ be a topological space, and $\mathcal S\subset C(X)$ a certain notion of regularity; in the above case, $X=(0,1)$ and $\mathcal S=C^\infty(X)$. Suppose the following three conditions are met.

  1. $\mathcal S$ is a local notion of regularity, in the sense that a function $f\in C(X)$ is of regularity $\mathcal S$ if $f_{|U}$ is of regularity $\mathcal S$ for all $U$ in an open cover of $X$ (for instance, not globally Lipschitz but rather locally Lipschitz). In technical terms, $\mathcal S$ should come from a subsheaf of $U\mapsto C(U)$. Moreover, for all $s_1,\ldots,s_k\in\mathcal S$ and $g\in C^\infty(\mathbb R^k)$, one should have $g\circ (s_1,\ldots,s_k)\in\mathcal S$. In particular, $\mathcal S$ is an algebra, and $g/h\in\mathcal S$ if $g,h\in\mathcal S$, $h>0$.

  2. The above fact should still hold: given $h,\delta\in C_b(X)$ such that $\delta>0$ over $I$, there exists $\tilde h\in \mathcal S$ such that $|h-\tilde h|<\delta$ over $X$. In particular, $\tilde h$ is bounded. This holds if $\mathcal S$ admits partitions of unity (a property sometimes known as $\mathcal S$-paracompactness); as described in Theorem 16.2 of The Convenient Setting of Global Analysis by Kriegl and Michor, it is enough to have $X$ paracompact and $\mathcal S$ separating for closed sets (given $F_0,F_1\subset X$ disjont closed sets, there exists $f\in\mathcal S$ such that $f_{|F_i}=i$).

  3. There should exist $f\in C_b(X)$ such that $f_{|U}$ is never of regularity $\mathcal S$, for $U\subset X$ a nonempty open subset.

Then the reasoning above shows that there exists a dense subalgebra $B\subset C_b(X)$ such that $B\cap\mathcal S$ contains only constants.

For instance, for any second countable metric space $(X,d)$, there exists a dense subalgebra $B\subset C_b(X)$ such that the only locally Hölder functions of $B$ ($\forall x,\exists\varepsilon,a,M>0,\forall y,z\in B_x(\varepsilon),|h(z)-h(y)|\leq Md(y,z)^a$) are constant. Indeed, setting $\mathcal S$ the local Hölder regularity, the first point is not difficult, and the second one is done using $\mathcal S$-paracompactness (consider the locally Lipschitz function $x\mapsto d(x,F_1)/(d(x,F_0)+d(x,F_1))$). We can deal with the third point using a Baire-type argument, noting that the set $$ A(x,\varepsilon,M) = \lbrace h\in C_b(X):\forall y,d(x,y)<\varepsilon\Rightarrow|h(y)-h(x)|\leq-1/\ln d(x,y) \rbrace $$ is closed with empty interior, and that a function that is locally Hölder over a nonempty open subset of $X$ has to belong to some $A(x_k,2^{-\ell})$, where $\lbrace x_0,x_1,\ldots\rbrace$ is a dense subset of $X$.

$\endgroup$
7
  • $\begingroup$ @NateEldredge The $a_i$ are smooth functions, not constants. $\endgroup$
    – Pierre PC
    Feb 12, 2021 at 21:28
  • $\begingroup$ Oh, thanks. Now it makes more sense. $\endgroup$ Feb 12, 2021 at 21:29
  • $\begingroup$ @NateEldredge No problem, thank you for your interest. :) $\endgroup$
    – Pierre PC
    Feb 12, 2021 at 21:30
  • 1
    $\begingroup$ @Ken.Wong I don't think so? If $g$ were precisely equal to $(g/\tilde f)f$, then we could apply the lemma, but that is not the case. The difference between these two functions is not smooth. $\endgroup$
    – Pierre PC
    Feb 13, 2021 at 15:56
  • 1
    $\begingroup$ $B$ is not an ideal, except in $\mathbb R[f]$ or something like that. The density argument relies on $C^\infty(I)$ being dense in $C_b(I)$, which is more or less the fact above. The fact that $(g/\tilde f)f$ is in $B$ comes from the definition of $B$, with $a_1=g/\tilde f$. $\endgroup$
    – Pierre PC
    Feb 13, 2021 at 16:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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