0
$\begingroup$

How can I build (i.e. find an explicit formula) a smooth function $f_\epsilon: \mathbb R \to \mathbb R$ depending on a parameter $\epsilon >0$ which is "almost the identity" but constant around a (fixed) value $\bar x \in \mathbb R$ in the following sense:

  • $f_\epsilon \in C^2(\mathbb R)$;
  • $L_\epsilon > f'_\epsilon \ge 0$;
  • $|f''_\epsilon| \le C_\epsilon $;
  • $f_\epsilon(\cdot) = \bar x$ in $(\bar x - \epsilon/2, \bar x + \epsilon/2)$;
  • $\|f_\epsilon(\cdot) - \mathrm{Id}\|_{L^p([a,b])} \lesssim \epsilon $ for any $p \in [1,\infty]$ (here $Id(x) = x$ is the identity and $[a,b]$ is any compact interval of $\mathbb R$)
$\endgroup$
3
  • 1
    $\begingroup$ I am pretty sure if you apply Gronwall's Lemma to the fourth inequality, that $|x f''_\epsilon(x)| \leq f'(\epsilon)(x)$, and the fifth condition on $f_{\epsilon}$ being constant on an interval, together these will require $f_{\epsilon}$ to be identically constant. This makes the "approximation of identity" property impossible. $\endgroup$ Commented Dec 7, 2021 at 19:44
  • $\begingroup$ @WillieWong What if I drop the 4th condition? $\endgroup$
    – user139844
    Commented Dec 7, 2021 at 20:07
  • $\begingroup$ In what is now the final condition, the implicit constant in the $\lesssim \epsilon$, will you allow it to depend on the interval $[a,b]$? If yes, then Iosif's answer does the job. It not (if you want the bound to be uniform over different intervals) I think it is probably doable but will require more work. $\endgroup$ Commented Dec 8, 2021 at 0:41

1 Answer 1

0
$\begingroup$

$\newcommand\R{\mathbb R}\newcommand{\ep}{\epsilon}$This is achieved simply by convolution. Indeed, for real $x$ let \begin{equation*} F(x):=\begin{cases} x+2 &\text{ if } x\leq -2, \\ 0 &\text{ if } -2<x\leq 2, \\ x-2 &\text{ if } x>2, \end{cases} \end{equation*} \begin{equation*} g(x):=\max(0,1-|x|), \end{equation*} \begin{equation*} h(x):=(F*g)(x)=\int_{\R}F(x-y)g(y)\,dy. \end{equation*} It is straightforward to obtain an explicit piecewise-polynomial expression for $h(x)$ (now given at the end of the answer, on a request by the OP). Note that $g\ge0$, $\int_{\R}g(y)\,dy=1$, and $0\le F'\le1$, where $F'$ denotes (say) the right derivative of $F'$.

It is easy to see that \begin{equation*} \text{$h(x)=0$ if $|x|\le1$},\tag{1} \end{equation*} \begin{equation*} h'(x)=\int_{\R}F'(x-y)g(y)\,dy\in[0,1],\tag{2} \end{equation*} \begin{equation*} h''(x)=g(x-2)-g(x+2)\in[-1,1],\tag{3} \end{equation*} \begin{equation*} h\in C^2(\R). \tag{3a} \end{equation*}

Also, $|F(x)-x|\le2$ and $\int_{\R}(x-y)g(y)\,dy=x$, and hence \begin{equation*} |h(x)-x|\le\int_{\R}|F'(x-y)-(x-y)|g(y)\,dy\le\int_{\R}2g(y)\,dy=2. \tag{4} \end{equation*}

Let now $a:=\bar x$ and then \begin{equation*} f_\ep(x):=f_{a;\ep}(x):=a+\ep h\Big(\frac{x-a}\ep\Big). \end{equation*} Then $f_\ep\in C^2(\R)$ (by (3a)), $f'_\ep(x)=h'\big(\frac{x-a}\ep\big)\in[0,1]$ (by (2)), $f''_\ep(x)=\frac1\ep\,h''\big(\frac{x-a}\ep\big)\in[-\frac1\ep,\frac1\ep]$ (by (3)), $f_\ep=a$ on $[-\ep,\ep]$ (by (1)), and $\|f_\ep-\mathrm{Id}\|_{L^\infty}\le2\ep$ (by (4)). So, $f_\ep$ satisfies all the desired conditions.


Here are the graphs $\{(x,f_{a;\ep}(x))\colon-1\le x\le2\}$ for $a=1$ and $\ep=2/10$ (black), $\{(x,x)\colon-1\le x\le2\}$ (blue, dashed), and $\{(x,1)\colon-1\le x\le2\}$ (green, dashed):

enter image description here


Here is the mentioned explicit piecewise-polynomial expression for $h(x)$:

\begin{equation} h(x)= \begin{cases} x-2 &\text{ if } x>3 , \\ x+2 &\text{ if } x\leq -3 , \\ \frac{1}{6} \left(-x^3-9 x^2-21 x-15\right) &\text{ if } -3<x\leq -2 , \\ \frac{1}{6} \left(-x^3+9 x^2-21 x+15\right) &\text{ if } 2<x\leq 3 , \\ \frac{1}{6} \left(x^3-3 x^2+3 x-1\right) &\text{ if } 1<x\leq 2 , \\ \frac{1}{6} \left(x^3+3 x^2+3 x+1\right) &\text{ if } -2<x<-1. \end{cases} \end{equation}

$\endgroup$
5
  • $\begingroup$ Thank you so much. I have a couple of questions: (1) how can one compute $h$ explicitly? (2) Is it possible for the function $f_\epsilon$ to be built also to be convex? $\endgroup$
    – user139844
    Commented Dec 8, 2021 at 7:38
  • $\begingroup$ Other two questions: (3) what is the dependence of $f'_\epsilon$ on $\epsilon$? Is it true that $f''_\epsilon \le C_\epsilon f'_\epsilon$? (4) Also, can the construction be generalized to get a $C^3$ function? If yes, on this $C^3$ function, do we have a better control on the second derivative, e.g. something like $f''_\epsilon \le C_\epsilon f'_\epsilon$? $\endgroup$
    – user139844
    Commented Dec 8, 2021 at 10:35
  • $\begingroup$ @Dal : (1) I have added the explicit expression for $h$. (2) No, because a convex function which equals a constant $c$ on a nonzero-length interval must be $\ge c$ everywhere and therefore cannot approximate the identity function. $\endgroup$ Commented Dec 9, 2021 at 21:19
  • $\begingroup$ @Dal : No to (3) and (4) either. Indeed, if for $F:=f_\epsilon$ we have $F''\le CF'$ and $F'>0$ on an interval $(a,b]$, then on $(a,b]$ we have $(\ln F')'\le C$ and hence $\ln F'(x)-\ln F'(a)\le C(x-a)$ and $F'(x)\le F'(a)e^ {C(x-a)}$ for $x\in[a,b]$, which implies $F'\ge0$ on $[a,b]$ if $F'(a)=0$. So, we get a contradiction with the condition $F'>0$ on an interval $(a,b]$. Now take $a:=\max\{z\colon F(z)=\bar x\}$. $\endgroup$ Commented Dec 9, 2021 at 21:37
  • 1
    $\begingroup$ @Dal : Your original question has been fully answered, and I even answered all your extra questions. Also, multiple questions for one post are discouraged on MathOverflow, anyway. So, let us have a closure here. $\endgroup$ Commented Dec 9, 2021 at 21:37

You must log in to answer this question.