12
$\begingroup$

The standard mollifier function is defined as follows

$$f(x)=\begin{cases} 0 & \text{if } |x| \ge 1\\ \exp \left(-\cfrac{1}{1-x^2}\right) & \text{if } |x|<1.\end{cases}$$

It is well known that $f$ is $C^\infty$, and $f^{(n)}(x)=0$ for $|x| \ge 1$. On the interval $x\in (-1,1)$, the derivative

$$\displaystyle f^{(n)}(x)=\frac{P_n(x)}{(1-x^2)^{2n}}\cdot f(x)$$

where $P_n$ is a polynomial function of $x$ defined inductively by

$$P_0(x) \equiv 1, \qquad P_1(x)=-2x, \qquad P_{n+1}(x)=P_n'(x)(1-x^2)^2+4nx(1-x^2) P_n(x)-2xP_n(x)$$

Note that $\displaystyle \sup_{|x|<1} f^{(n)}(x)<+\infty$, since $f^{(n)}(\pm 1)=0$. So $\displaystyle |f|_n:=\max_{|x|<1} f^{(n)}(x)$ is well defined.

Are there some rough/good estimates on the size $|f|_n$ of the derivatives?

Thanks!

$\endgroup$
6
  • 3
    $\begingroup$ The local extrema of $f^{(n)}(x)$ can be found by computing the critical points of $f^{(n+1)}(x)$. These are the same as the zeroes of $P_{n+1}(x)$. Once you have those, you just evaluate $f^{(n)}(x)$ at them to find your global extrema. This, of course, needs you to find a decent formula for $P_{n+1}(x)$ and its zeros. That could be a good deal more challenging, as finding roots of polynomials is not such a simple task. Playing with wolfram alpha suggests within (-1,1) that there is one root at zero for $n+1$ odd, and no roots when $n+1$ even. I'm not sure of a proof of that, though. $\endgroup$ Jun 8, 2015 at 1:11
  • $\begingroup$ I modified the formulation a little bit. The above comment is for the initial formulation I used: $f^{(n)}(x)=P_n(x)\cdot f(x)$. $\endgroup$
    – Pengfei
    Jun 8, 2015 at 4:07
  • 1
    $\begingroup$ My comment holds in the new formulation, as well, actually. The inductive formula may be useful for proving the apparent pattern I suggested. Incidentally, you need to replace the $\max$ with a $\sup$: it is not immediately clear that maximum is attained in the open interval. Using $\max |f^{(n)}(x)|$ would also work, but may give you something other than what you were looking for. $\endgroup$ Jun 8, 2015 at 4:28
  • 1
    $\begingroup$ @Pengfei Have you tried $||f^{(n)}||_\infty \le ||\widehat{f^{(n)}}||_1 = ||\xi^n \hat{f}(\xi)||_1 = \int_{-\infty}^\infty |\xi|^n|\hat{f}(\xi)|d\xi$? $\endgroup$ Jul 21, 2019 at 12:55
  • 1
    $\begingroup$ @Pengfei Or just consider your function in the complex plane and use Cauchy bounds choosing radius appropriately. $\endgroup$
    – fedja
    Jul 21, 2019 at 19:01

2 Answers 2

5
+100
$\begingroup$

First we want to study $$h_\alpha:x\mapsto \frac{\exp\left(-\frac{\alpha}{1-x^2}\right)}{1-x^2}=Xe^{-\alpha X}$$ for $x \in [0,1]$ i.e. $X \in [1,+\infty[$.

The second expression $H_\alpha(X)=Xe^{-\alpha X}$ is positive over $\mathbb R_+$, and zero for $X=0$ and $X\rightarrow +\infty$, so there must be a maximum on which $H_\alpha'(X^*)=0$.

But there is only one zero of the differential, for $X=\frac1\alpha$.

So $h_\alpha$ is bounded by $\frac1{\alpha e}$.

Rewriting:

$$ f^{(n)}(x)=\frac{P_n(x)}{(1-x^2)^{2n}} \exp\left(-\frac{1}{1-x^2}\right)=P_n(x) h_{1/2n}(x)^{2n} $$

we get that: $$|f^{(n)}|_\infty\leq \left(\frac{2n}e\right)^{2n}\sup_{[-1,1]}|P_n|$$

Now we want to bound $\sup_{[-1,1]}|P_n|$ by $|P_n|_1$ the sum of the absolute values of the coefficients of $P_n$, and use the inductive formula to find a probably give a very rough bound:

$$P_{n+1}(x)=P_n'(x)(1-x^2)^2+4nx(1-x^2) P_n(x)-2xP_n(x)$$

By triangular inequality:

$$ |P_{n+1}|_1\leq|(1-x^2)^2P_n'|_1 + 4n|(1-x^2) P_n|_1 + 2|P_n|_1 $$

Because $P_n$ is of degree less than $3n$:

$$ |P_{n+1}|_1\leq 12n|P_n|_1 + 6n|P_n|_1 + 2|P_n|_1 $$

$$ |P_{n+1}|_1\leq 18(n+1)|P_n|_1 $$

Thus:

$$ |P_n|_1\leq 18^n n! $$

And finally:

$$|f^{(n)}|_\infty \leq \left(\frac{2n\sqrt{18}}e\right)^{2n}n!$$

$\endgroup$
3
  • $\begingroup$ In fact, numerical calculations appear to show that $|f^{(n)}|_{\infty}=(2ne^{-1}+o(1))^{2n}$. However, it is also interesting to see that your method can not do better than $(4ne^{-1})^{2n}$, since $|P_n(\pm 1)|=2^n$. $\endgroup$
    – pre-kidney
    Jul 22, 2019 at 6:16
  • $\begingroup$ @pre-kidney Yes I have little doubt that this bound is far from the actual maximum by several orders of magnitude! $\endgroup$
    – FXV
    Jul 22, 2019 at 17:54
  • $\begingroup$ It seems that this method cannot be tightened much because it is indeed the case that $\sup_{[-1,1]} P_n \gg |P_n(\pm 1)|$. $\endgroup$
    – user58955
    Oct 11, 2019 at 8:24
4
$\begingroup$

As mathworker21 suggested in a comment, $$ \|f^{(n)}\|_\infty \leq \|\widehat{f^{(n)}}\|_1 = \|(i\xi)^n\hat f(\xi)\|_1 = \int |\xi|^n |\hat f(\xi)| d\xi. $$ This paper (https://arxiv.org/abs/1508.04376) shows that $|\hat f(\xi)|$ decays as $|\xi|^{-3/4}e^{-\sqrt{|\xi|}}$. Hence, when $|\xi| \gtrsim (Cn\log n)^2$, we have $|\xi|^n |\hat f(\xi)|\leq |\xi|^{-2}$. On the other hand, $|\hat f(\xi)| \leq \|f\|_1 = 1$. This gives $$ \int |\xi|^k |\hat f(\xi)| d\xi \lesssim (C n\log n)^{2n+2}. $$ This improves FXV's bound with an $n!$ factor in it but still seems suboptimal by an $n^2(\log n)^{2n+2}$ factor.

$\endgroup$
1
  • $\begingroup$ This is a very interesting approach! $\endgroup$
    – Pengfei
    Oct 14, 2019 at 14:56

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.