Let $f\in L^{\infty}(\mathbb{R}^{n})$ be supported in the unit ball and have mean zero. Let $\phi\in L^{1}(\mathbb{R}^{n})\cap C^{\alpha}(\mathbb{R}^{n})$ be a Holder continuous function with exponent $\alpha>0$. Define an approximation to the identity $\left\{\phi_{t}\right\}_{t>0}$ by $\phi_{t}:=t^{-n}\phi(\cdot/t)$, and consider the maximal convolution operator $M_{\phi}$ defined by $$M_{\phi}(f)(x)=\sup_{t>0}\left|(f\ast\phi_{t})(x)\right|, \qquad\forall x\in\mathbb{R}^{n} \tag{1}$$

If $\phi$ is compactly supported in a ball about the origin $B_{A}(0)$ of radius $A>0$, then it is not hard to show that $M_{\phi}f\in L^{1}(\mathbb{R}^{n})$. Indeed, for $\left|x\right|\leq 2$, we have from Young's inequality

$$M_{\phi}(f)(x)\leq\sup_{t>0}\left\|f\right\|_{L^{\infty}}\left\|\phi_{t}\right\|_{L^{1}}=\left\|f\right\|_{L^{\infty}}\left\|\phi\right\|_{L^{1}} \tag{2}$$

Now consider case $\left|x\right|>2$. Using $\int f=0$, we may write \begin{align*} (f\ast\phi_{t})(x)=t^{-n}\int_{\mathbb{R}^{n}}f(y)\left[\phi\left(\dfrac{x-y}{t}\right)-\phi\left(\dfrac{x}{t}\right)\right]\mathrm{d}y \tag{3} \end{align*} for any $t>0$. Since $\left|x\right|>2$ and $\left|y\right|\leq 1$ (by the support hypothesis for $f$), $\left|x-y\right|\geq\left|x\right|/2$ by the reverse triangle inequality. Since $\phi((x-y)/t)=0$, unless $\left|x-y\right|/t\leq A$, we must have $t\geq\left|x\right|/(2A)$. Whence by Holder continuity, $$t^{-n}\left|\phi\left(\dfrac{x-y}{t}\right)-\phi\left(\dfrac{x}{t}\right)\right|\leq\left\|\phi\right\|_{C^{\alpha}}t^{-n-\alpha}\leq\dfrac{\left\|\phi\right\|_{C^{\alpha}}}{(2A)^{\alpha}}\left|x\right|^{-n-\alpha}\cdot\left|y\right|^{\alpha}$$

Substituting this result into (3), we obtain that \begin{align*} \left|(f\ast\phi_{t})(x)\right|\leq \left|x\right|^{-n-\alpha}\dfrac{\left\|\phi\right\|_{C^{\alpha}}}{(2A)^{\alpha}}\int_{\mathbb{R}^{n}}\left|f(y)\right|\left|y\right|^{\alpha}\mathrm{d}y\lesssim_{\phi}\left\|f\right\|_{L^{1}}\left|x\right|^{-n-\alpha}, \qquad\forall t>0 \tag{4} \end{align*} which is integrable on $\mathbb{R}^{n}$ away from the origin.

I am interested in how much we can relax the decay hypothesis on $\phi$ (i.e. compact support) and still obtain $M_{\phi}f\in L^{1}(\mathbb{R}^{n})$. Motivated by this related question here, where a commenter suggests such an $f$ above will work for an arbitrary Holder continuous, integrable function $\phi$, I ask:

Question: If $\phi\in C^{\alpha}(\mathbb{R}^{n})\cap L^{1}(\mathbb{R}^{n})$, then for $f$ as above, does it hold that $M_{\phi}(f)\in L^{1}(\mathbb{R}^{n})$? If not, does there exist a function $g\in L^{1}(\mathbb{R}^{n})$ such that $M_{\phi}(g)\in L^{1}(\mathbb{R}^{n})$?

For example, if we have that $\phi\in C^{1}(\mathbb{R}^{n})\cap L^{1}(\mathbb{R}^{n})$ and $\left|\nabla\phi\right|$ satisfies the decay conditions

$$\left|\nabla\phi(x)\right|\leq C(1+\left|x\right|)^{-n-\delta},\qquad\forall x\in\mathbb{R}^{n} \tag{5}$$

for some $\delta>0$ fixed, then I believe the answer is affirmative.

Indeed, for notational convenience, assume $\delta=1$. Fix $\left|x\right|>2$, and partition the set $\left\{t>0, \left|y\right|<1\right\}$ into disjoint annuli $$\mathcal{A}_{0}:=\left\{0<\left|\dfrac{x-y}{t}\right|\leq A\right\}\cup\mathcal{A}_{j}:=\left\{2^{j-1}A<\left|\dfrac{x-y}{t}\right|\leq 2^{j}A\right\}, \qquad j\geq 1$$ We have already dealt with the case of the first annulus. If $(y,t)\in\mathcal{A}_{j}$, then $$\dfrac{\left|x\right|}{2t}\leq 2^{j}A, \quad 2^{j-1}A<\dfrac{2\left|x\right|}{t}$$ By the mean value theorem, we have \begin{align*} t^{-n}\left|\phi\left(\dfrac{x-y}{t}\right)-\phi\left(\dfrac{x}{t}\right)\right|\leq t^{-n}\left|\nabla\phi(\bar{x})\right|\dfrac{\left|y\right|}{t}\leq \left|\nabla\phi(\bar{x})\right|\dfrac{(2^{j+1}A)^{n+1}}{\left|x\right|^{n+1}}, \tag{6} \end{align*} where $\bar{x}=\theta\frac{x-y}{t}+(1-\theta)\frac{x}{t}$ for some $\theta\in (0,1)$. Since $\left|\bar{x}\right|\geq\left|x\right|/2t>2^{j-2}A$, the decay condition for $\left|\nabla\phi\right|$ implies $$\left|\nabla\phi(\bar{x})\right|\leq C\dfrac{1}{(1+(\left|x\right|/2t))^{n+1}}\leq \dfrac{C}{(2^{j-2}A)^{n+1}} \tag{7}$$

Putting these results together, we see that for $(y,t)\in\mathcal{A}_{j}$

\begin{align*} \left|(f\ast\phi_{t})(x)\right|\leq\int_{\mathbb{R}^{n}}\left|f(y)\right|\dfrac{C(2^{j+1}A)^{n+1}}{(2^{j-2}A)^{n+1}\left|x\right|^{n+1}}\mathrm{d}y=\dfrac{C(8A)^{n+1}\left\|f\right\|_{L^{1}(\mathbb{R}^{n})}}{\left|x\right|^{n+1}} \tag{8} \end{align*}

The constants are independent of $x$ and $\mathcal{A}_{j}$ and only depend on $\phi$. Since $\left|x\right|^{-n-1}$ is integrable on $\mathbb{R}^{n}$, we conclude that $M_{\phi}(f)\in L^{1}(\mathbb{R}^{n})$.

Without some integrability condition for $\nabla\phi$, I don't know how to control cases where $\left|x\right|/t$ is very large, so the above argument doesn't seem to suffice to answer my question, as it is written, affirmatively.

  • $\begingroup$ You're no doubt aware that your $f$ is precisely an $H^1$ atom. And the whole point to $H^1$ is you get integrable maximal functions. But I doubt that just assuming $\phi\in C^\alpha\cap L^1$ is enough. No actual example quite yet... $\endgroup$ Jul 19, 2015 at 18:28
  • $\begingroup$ @DavidC.Ullrich: I am aware. I omitted that detail because I didn't think it added anything to solving the problem as formulated. The maximal function characterizations of $H^{1}(\mathbb{R}^{n})$ that I've seen are with Schwartz functions of nonzero mean. As you have been very helpful with other questions of mine, I would very much appreciate your assistance. $\endgroup$ Jul 19, 2015 at 22:47
  • $\begingroup$ Well like I said, I'll be in touch if I come up with anything... $\endgroup$ Jul 19, 2015 at 23:46
  • $\begingroup$ Counterexample exists. New activity in thread... $\endgroup$ Jul 20, 2015 at 16:44

2 Answers 2


Three examples.

Example 1

No, $M_\phi f$ need not be in $L^1$. In fact if $n=1$ it can die arbitrarily slowly at infinity; it need not be in weak $L^1$ either. Even if $\phi$ is infinitely differentiable with all derivatives vanishing at infinity. (We can throw in $\int\phi=0$ too if you want.)

I'm going to take $n=1$ and $$f=\chi_{[0,1]}-\chi_{[-1,0)}.$$

True Fact: If $h(x)\to0$ as $x\to\infty$ there exists $\phi\in C^\infty(\mathbb R)$ with all derivatives vanishing at infinity and $\int\phi=0$ such that $M_\phi f(x)\ge h(x)$ ($x\ge 10$).

Proof, leaving details to you: Let $\psi$ be infinitely differentiable with support in $[-1,1]$. Say $\psi$ is odd, and assume $\psi\ge0$ in $[-1,0]$.

Let $$\phi(x)=\sum_{n=1}^\infty c_n\psi(x-r_n),$$where $c_n\ge0$, $\sum c_n<\infty$, and $(r_n)$ is a sequence of positive numbers increasing to infinity. We'll be specifying the $r_n$ at the end; for now note that one constraint is $$r_{n+1}\ge 2r_n.$$

Now suppose $$10\le x\le r_n$$and set $$\delta_n=x/r_n.$$It follows that $$f*\phi_{\delta_n}(x)=c_n||\psi||_1.$$You can work that out as well as I can, unless it's not quite true, in which case there's an off by one error or a dropped minus sign in my scribblings; the picture is clear enough to me that I know any errors can be fixed. Two things that will be relevant in the calculation: First, since $\delta_n\le1$, $$(-\delta_n,\delta_n)\subset[-1,1].$$ Second, the fact that $r_{k+1}\ge 2r_k$ shows that $$[-1,1]\cap\left( x\left(1-\frac{r_m}{r_n}\right)-\delta_n,\,x\left(1-\frac{r_m}{r_n}\right)+\delta_n\right)=\emptyset\quad(m\ne n).$$

So we have $$M_\phi f\ge||\psi||_1\sup_n c_n\chi_{[10,r_n]}.$$ Given $(c_n)$ it's easy to see that there exist $r_n$ such that this is larger than $h$ on $[10,\infty)$.

Or, if $\alpha>0$ and we set $r_n=2^n$, $c_n=r_n^{-\alpha}$ we obtain $|\phi(x)|\le c(1+|x|)^{-\alpha}$ and $M_\phi f(x)\ge c(1+|x|)^{-\alpha}$ for $x>10$. (Modifying $\phi$ to be even would give the last inequality for all $x$.)

Now what about higher dimensions? I'm not going to give an example in $\mathbb R^n$; the letter $n$ is too useful to waste. Instead I'll give examples in $\mathbb R^d$. (It's not hard to show that $\mathbb R^n$ and $\mathbb R^d$ are isomorphic if $d=n$.)

Example 2

A construction like $$\phi(x)=\sum_{j=1}^\infty c_j\psi(|x|-r_j)$$ does everything you want in $\mathbb R^d$. Things come out particularly nice if you allow $\mathbb R^d$-valued functions, taking $f(t)=t/|t|$ for $|t|<1$.

I'm not going to prove this. Instead I'm going to give another example, which is just as simple to work out as Example 1 was, but which I claim has the additional property that if you "understand" how Example 3 works you will then "see" that Example 2 also works. (Comments on the connection at the very end.)

Example 3

Points in $\mathbb R^d$ will often be written $x=(x_1,x')$, where $x_1\in\mathbb R$ and $x'\in\mathbb R^{d-1}$. Similarly $B'(x',r)$ denotes a ball in $\mathbb R^{d-1}$, etc.

Let $K$ be the set $|x_1|\le 1$, $|x'|\le 1$: $$K=[-1,1]\times B'(0,1).$$ Take $f$ vanishing outside $K$; for $x\in K$ let $f(x)=1$ for $x_1>0$, $f(x)=-1$ for $x_1<0$.

Let $\psi\ge0$ be a smooth function supported in $[-1,0]$.

Note this is not quite the same as in Example 1.

(Comment that will make no sense to readers who haven't worked out the details for Example 1: In the first example we took $f$ and $\psi$ both odd, and then the estimate on the integral used the facts that $(-1)^2>0$ and $1^2>0$. That was cute but a little silly. We could set Example 3 up that way, but then my claim that Example 2 works the same as Example 3 would be problematic. Instead in the relevant estimate in Example 3 we will use $(-1)(0)\ge0$ and $1^2>0$.)

Let $\chi$ be a smooth function with $$\chi_{[-1,1]}\le\chi\le\chi_{[-2,2]}.$$

Given sequences $c_j$ and $r_j>0$ with $r_{j+1}\ge2r_j$, set $$\phi(x)=\sum_{j=1}^\infty c_j\psi(x_1-r_j)\chi\left(\frac{|x'|}{r_j}\right).$$

(For your immediate purposes there's no reason not to take $r_j=2^j$; I'm leaving the more general version for future applications.)

Define a truncated sector $S$ by $$S=\{x\in\mathbb R^d\,:\,x_1>10,\,\,|x'|\le x_1/2\}.$$

Now suppose $x\in S$ and $x_1\le r_n$. Let $\delta=x_1/r_n\le1$. As in Example 1 the condition $r_{j+1}\ge2r_j$ shows that only one term in the definition of $\phi$ matters, and we get $$f*\phi_\delta(x)=\delta^{-d}c_n\int_Kf(y)\,\psi\left(\frac{-y_1}\delta\right)\chi\left(\frac{|x'-y'|}{r_n\delta}\right)\,dy.$$Now the conditions we put on $\psi$ show that the integrand vanishes except at points where $f$ and $\chi$ equal $1$, so$$f*\phi_\delta(x)=\delta^{-d}c_n\int_K\psi\left(\frac{-y_1}\delta\right)\,dy=C\delta^{-d+1}c_n\ge C\,c_n.$$

And there you are.

Now about Example 2 being really the same. In Example 3 we had $\delta\le1$, and we essentially ended up with the integral of $c_n$ over the set $[0,\delta]\times B'(0,1)$. We could have set it up slightly differently, getting $0<\delta\le1/4$, and getting an integral over the set $$R=[1/2-\delta,1/2+\delta]\times B'(0,1).$$Imagine a picture of $R$ in the $x_1,|x'|$ plane: $R$ is a "vertical" rectangle. Take that rectangle and curve it just a little bit, so instead of a rectangle it becomes a part of an annulus with center at a point $(x_1,0)$ where $x_1$ is large.

The new $R$ looks the same as the old $R$ if you take your glasses off. Nothing essential has changed, we get the same sort of estimate. And this shows that Example 2 works. (Ok, there are other differences, in the definition of $\phi$. But none of the differences really matter. QED.)

  • $\begingroup$ Very nice, David. Thank you! I believe that I also came up with a decay at $\infty$ condition for $\left|\phi\right|$, as opposed to $\left|\nabla\phi\right|$, which your counterexample shows is sharp in dimension $n=1$. $\endgroup$ Jul 21, 2015 at 16:49
  • $\begingroup$ Excellent and very insightful! In particular, you've shown that the maximal convolution operator $M_{\phi}$ need not be bounded $H^{1}(\mathbb{R}^{d})\rightarrow L^{1}(\mathbb{R}^{d})$ without a reasonable decay at $\infty$ condition on $\phi$...in addition to a Holder condition or some smoothness. Perhaps if you'd like now, you can put to your mind to this problem. $\endgroup$ Jul 25, 2015 at 17:17
  • $\begingroup$ I looked at it. Got as far as you did pretty quickly. Haven't got any farther in a few days - no idea at all. In fact I have no idea what sort of argument he has in mind, that would allow us to derive the bound $CN$ from the bound $CN^2$. There's some trick we haven't seen. $\endgroup$ Jul 27, 2015 at 16:26

David's very nice counterexample shows that in dimension $n=1$ the following result is sharp. We cannot only require that $\phi\in L^{1}(\mathbb{R})\cap C^{0,\alpha}(\mathbb{R})$ and $\left|\phi(x)\right|\lesssim(1+\left|x\right|)^{a}$, for any $0<a\leq n$, and still expect $M_{\phi}(f)\in L^{1}(\mathbb{R})$.

Proposition. If $\phi\in L^{1}(\mathbb{R}^{n})\cap C^{0,\alpha}(\mathbb{R}^{n})$ and satisfies the decay condition \begin{align*} \left|\phi(x)\right|\lesssim (1+\left|x\right|)^{-n-\delta},\qquad\forall x\in\mathbb{R}^{n} \end{align*} for some fixed $\delta>0$, then $M_{\phi}f\in L^{1}(\mathbb{R}^{n})$.

Proof. Let $0<\beta<1$ denote a parameter to be chosen at the end. On $\left|x\right|\leq 2$, then we can easily bound $M_{\phi}(f)$ by a constant. Suppose that $\left|x\right|>2$, and consider two cases.

First, suppose $\left|x-y\right|/t\leq\left|x\right|^{\beta}$, so that $\left|x\right|/2t\leq\left|x\right|^{\beta}$ for all $\left|y\right|\leq 1$. Using the zero mean of $f$, we may write \begin{align*} (f\ast\phi_{t})(x)=\int_{\mathbb{R}^{n}}f(y)t^{-n}\left[\phi\left(\dfrac{x-y}{t}\right)-\phi\left(\dfrac{x}{t}\right)\right]\mathrm{d}y \tag{1} \end{align*} Whence by Holder continuity, the RHS is bounded by \begin{align*} \int_{\mathbb{R}^{n}}\left|f(y)\right|t^{-n}\left\|\phi\right\|_{C^{0,\alpha}}\left(\dfrac{\left|y\right|}{t}\right)^{\alpha}\mathrm{d}y&\lesssim\dfrac{\left\|\phi\right\|_{C^{0,\alpha}}}{\left|x\right|^{(1-\beta)(n+\alpha)}}\int_{\mathbb{R}^{n}}\left|f(y)\right|\left|y\right|^{\alpha}\mathrm{d}y \tag{2} \end{align*}

Now suppose that $\left|x-y\right|/t > \left|x\right|^{\beta}$, so that $2\left|x\right|/t>\left|x\right|^{\beta}$ for all $\left|y\right|\leq 1$. Whence by the decay at $\infty$ hypothesis for $\phi$, we have that \begin{align*} \left|(f\ast\phi_{t})(x)\right|&\lesssim\int_{\mathbb{R}^{n}}\left|f(y)\right|\dfrac{t^{-n}}{(1+(\left|x-y\right|/t))^{n+\delta}}\mathrm{d}y\\ &\lesssim\int_{\mathbb{R}^{n}}\left|f(y)\right|\dfrac{\left|x\right|^{-n(1-\beta)}}{\left|x\right|^{(n+\delta)\beta}}\mathrm{d}y\\ &=\left|x\right|^{-n-\delta\beta}\int_{\mathbb{R}^{n}}\left|f(y)\right|\mathrm{d}y \tag{3} \end{align*} Note that $n+\delta\beta>0$ for any choice of $0<\beta<1$.

So choosing $\beta$ sufficiently close to $1$ so that $(1-\beta)(n+\alpha)>n$, we conclude that for $\left|x\right|\leq 2$, \begin{align*} \left|(f\ast\phi_{t})(x)\right|\lesssim_{\phi}\max\left\{\left|x\right|^{(1-\beta)(n+\alpha)},\left|x\right|^{-n-\delta\beta}\right\},\qquad\forall t>0 \end{align*} Taking the supremum over $t>0$, we conclude that \begin{align*} M_{\phi}(f)(x)\lesssim_{\phi}\chi_{\left|x\right|\leq 2}(x)+\max\left\{\left|x\right|^{(1-\beta)(n+\alpha)},\left|x\right|^{-n-\delta\beta}\right\}\chi_{\left|x\right|>2}(x), \tag{4} \end{align*} where the RHS is integrable on $\mathbb{R}^{n}$. $\Box$

  • $\begingroup$ Heh. So now we have to show that that's sharp for $n>1$. Raising the question of whether one can use something like $\phi=\sum c_j\psi(|x|-r_j)$. Hmm. One trick was used for $n=1$: we made things line up exacty right so that the fact that $\int f=0$ was irrelevant. Probably can't hit things on the head like when $n=1$. But, for example, if $f$ vanishes outside the unit disk in $\mathbb C$ and $f(re^{it})=e^{it}$ in the disk then $\int_H f\ne0$ if $H$ is any half-disk. Can't get quite that for $n>2$... OH. You can get something like that for $n>2$, except that $f$ is $\mathbb R^n$-valued! $\endgroup$ Jul 21, 2015 at 17:08
  • $\begingroup$ Woo-hoo. Seems to me essentially the same example, modified as outlined in my previous comment, actually does work to show that $\phi\le(1+|x|)^{-n}$ is not enough. I'll let you go first writing it down... $\endgroup$ Jul 21, 2015 at 17:27
  • $\begingroup$ @DavidC.Ullrich: I don't follow what modification your proposing for $f$ in dimensions $n\geq 2$. Would you please elaborate? $\endgroup$ Jul 22, 2015 at 18:14
  • $\begingroup$ What I was claiming works was what I said in my first comment here: $\phi=\sum c_j\psi(|x|-r_j)$. With, say, $f$ supported in the unit ball, $f(x)=x/|x|$. Have you tried to get this to work? $\endgroup$ Jul 22, 2015 at 18:21
  • $\begingroup$ @DavidC.Ullrich So I'm playing around with it now. I believe I get $(f\ast\phi_{\delta_{n}})(x)=(c_{n}/\delta_{n})\int_{B(0,1)}(y/\left|y\right|) \psi((\left|x-y\right|-\left|x\right|)/\delta_{n})dy$ for $10\leq\left|x\right|\leq r_{n}$. This looks like it behaves nicely under orthogonal transformation applied to $x$; but I don't think we get something nice like $c_{n}\alpha\frac{x}{\left|x\right|}$, where $\alpha\neq 0$. $\endgroup$ Jul 22, 2015 at 21:46

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.