2
$\begingroup$

Let $f\colon [0, \infty)\to\mathbb{R}$ be nonnegative, nondecreasing, and concave. Prove the following claim or give a counter example: There is a sequence of functions $f_n\colon [0, \infty)\to\mathbb{R}$ with following properties:

  1. $f_n$ is nonnegative, nondecreasing, and concave for all $n\in\mathbb N$.
  2. $f_n$ is smooth, i.e., infinitely differentiable on $(0,\infty)$, for all $n\in\mathbb N$.
  3. $\lim_{n\to\infty} f_n(x) = f(x)$ for all $x\in[0,\infty)$.

My first approach was to use mollifiers. But I get problems close to $x=0$:

Let $\varphi\colon\mathbb{R} \to \mathbb{R}$ be a positive mollifier, e.g., $$ \varphi(x) := \begin{cases} C \exp(-\frac{1}{1-x^2}) & \text{ if } x^2 < 1,\\ 0 & \text{ if } x^2 \geq 1, \end{cases} $$ where $C\in(0,\infty)$ is a normalization constant such that $\int_{-\infty}^\infty \varphi(x) d x = 1$. Define $$ f_n(x) := \int_{-\infty}^\infty n \varphi(ny) \tilde f(x-y) d y, $$ where $\tilde f\colon\mathbb{R} \to \mathbb{R}$ is an extension of $f$ to $\mathbb R$. If we could find an extension $\tilde f$ that is nonnegative, nondecreasing, and concave on $[-1/n, \infty)$, we would be fine. But if $f(0) = 0$ and $f$ is strictly increasing, this is not possible. If we use $\tilde f(x) = f(\max(0, x))$, then we can only guarantee concavity on $[1/n, \infty)$ but not in $(0, 1/n)$.


Any ideas?

$\endgroup$

1 Answer 1

2
$\begingroup$

Replace the convolution on $\mathbb{R}$ by a convolution on the group $\mathbb{R}_+^*$, endowed with the invariant measure $dx/x$, namely set $$f_n(x) := \int_0^\infty \varphi_n(y) f(x/y) \frac{dy}{y},$$ where $(\varphi_n)_{n \ge 1}$ is a sequence of non-negative $\mathcal{C}^\infty$ functions on $\mathbb{R}_+^*$ with support contained in $[e^{-1/n},e^{1/n}]$ such that for every $n \ge 1$, $$\int_0^\infty \varphi_n(y) \frac{dy}{y} = 1.$$

$\endgroup$
2
  • $\begingroup$ Thanks a lot! Is this something one can find in a textbook? $\endgroup$ Commented May 28, 2023 at 13:59
  • 1
    $\begingroup$ Fourier analysis on groups by Walter Rudin for the general theory. Here, on can note that $$f_n(e^t) = \int_\mathbb{R} \varphi_n(e^s) f(e^{t-s}) ds$$ and recognise a usual convolution product of $s \mapsto \varphi(e^s)$ and $r \mapsto f(e^r)$. $\endgroup$ Commented May 28, 2023 at 16:05

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .