8

I am trying to prove that for a given Hölder parameter $\alpha \in (0, 1)$ and a distribution $f \in \mathcal{D}'(\mathbb{R}^d)$ the following are equivalent:

  • $f \in C^{\alpha}$
  • For any $x$ there exists a polynomial $P_x$ such that $| \langle f - P_x, \phi_x^{\lambda} \rangle | \le C \lambda^{\alpha.}$

Where the latter estimate holds uniformly over all$x$ and $\phi \in \mathcal{D}$ with compact support in the unit ball, and: $$\phi_x^{\lambda}(\cdot) = \lambda^{-d} \phi\left( \frac{\cdot \ - \ x}{\lambda} \right)$$

Proving that the first implies the second is fairly easy. The other way around gives me some problems.

  1. The first step I took is to realize that we care only about the order zero term of $P_x,$ since all other terms vanish at a order higher than $\alpha.$ Here we assume that the polynomial is centered in $x.$
  2. Then I would like to prove that $P_x(x) = g(x)$ defines a $\alpha$ - Hölder function.
  3. Thus we would get that $g \in \mathcal{D}'.$
  4. Eventually I would like to prove that $g = f$ in $\mathcal{D}'.$

My problem is that I can't prove any implications between 2,3,4 nor any of 2,3,4 starting from 1. Any hints, help, suggestions?

Kore-N
  • 4,275

1 Answers1

4

Your steps are the right ones, in this answer I'll fill in the details.

Firstly, let's check that $g$ defines an $\alpha$-Holder function. Let $\rho$ be a smooth probability density with support in $B(0,1)$.

As you observed, we can assume that $P_x$ is constant for each $x$ (and I write $P_x$ now to denote that constant so that $g(x) = P_x$). This means that since $\int \rho(x) dx = 1$ we have that $P_x = \langle P_x, \rho_y^\lambda \rangle$ for any $x,y$.

I then write $|g(x) - g(y)| = \langle P_x - P_y, \rho_x^\lambda \rangle$ for $\lambda = |x-y|$. Notice that this means that there exists $\psi$ such that $\rho_x^\lambda = c \psi_y^{2\lambda}$ where $c$ is bounded independently of $x,y$ and $\lambda$.

Therefore \begin{align*} |g(x) - g(y)| \le |\langle f - P_x, \rho_x^\lambda \rangle| + c|\langle f - P_y, \psi_y^\lambda \rangle| \lesssim |x-y|^\alpha. \end{align*}

Finally, we must show that $f = g$ in $\mathcal{D}'$. This will follow from the fact that any distribution that vanishes at order $\alpha > 0$ must be the zero distribution.


Lemma: Suppose $F \in \mathcal{D}'(\mathbb{R}^d)$ and for every compact set $k \subseteq \mathbb{R}^d$, $$|\langle F, \phi_x^\lambda \rangle| \leq C \lambda^\alpha$$ for $\alpha>0$ where $C$ is uniform over $x \in k$, $\gamma \in [0,1)$ and $\phi \in \mathcal{D}(\mathbb{R}^d)$ with $\operatorname{supp} \phi \subseteq B(0,1)$. Then $F = 0$.

Proof of Lemma: Fix a smooth probability density $\rho$ with compact support in $B(0,1)$. Then for any test function $\varphi \in \mathcal{D}'$, by standard facts about mollification, $$\langle \varphi, \rho_x^\lambda \rangle = \int_{\mathbb{R}^d} \varphi(y) \rho_x^\lambda(y) dy \to \varphi(x)$$ as $\lambda \to 0$ and in fact, the convergence takes place in $\mathcal{D}$ where $\Phi^\lambda(x) := \langle \varphi, \rho_x^\lambda \rangle$ is considered as a function of $x$. Hence \begin{align*} |\langle F, \varphi \rangle| =& \lim_{\lambda \to 0} |\langle F, \Phi^\lambda \rangle| \\ =& \lim_{\lambda \to 0} \bigg |\int_{\mathbb{R}^d} \langle F, \rho_x^\lambda \rangle \varphi(y) dy \bigg|\\ \leq& \lim_{\lambda \to 0} C\lambda^\alpha \int_{\mathbb{R}^d} |\varphi(y)| dy = 0 \end{align*}


In our particular case, we have \begin{align*} |\langle f-g, \phi_x^\lambda \rangle \leq& |\langle f - P_x, \phi_x^\lambda \rangle| + |\langle P_x - g(x), \phi_x^\lambda \rangle| + |\langle g - g(x), \phi_x^\lambda \rangle | \\ \leq& C_3 \lambda^\alpha \end{align*} by our assumption on $f$ and $\alpha$-Holder continuity of $g$ and $P_x$ (since $P_x - g(x) = P_x-P_x(x)$). Hence, by the lemma, $f=g$ as desired.

Rhys Steele
  • 20,326
  • Why the function $\Phi_{x,y}^\lambda$ is a test function at scale $\lambda + \frac{|x-y|}{2}$ centered at $(x+y)/2$? Could you provide more explanation? – Chival Apr 26 '21 at 02:31