5
$\begingroup$

for some positive integer $d \geq 1$ I have a globally Lipschitz continuous function $f \colon \mathbb{R}^d \to \mathbb{R}$ with Lipschitz constant $1$ and would like to approximate it by a sequence $f_\varepsilon$ with the following properties:

  1. For all $\varepsilon > 0$, the function $f_\varepsilon$ is $k$ times partially differentiable and all its partial derivatives up to order $k$ are bounded. Here, $k$ is some positive integer ($k=3$ suffices for my purposes but arbitrary $k$ would be nicer);
  2. For all $\varepsilon > 0$, it holds that $\| f_\varepsilon \|_{\text{Lip}} \leq \| f \|_{\text{Lip}}$, where $\|\cdot\|_\text{Lip}$ denotes the usual Lipschitz norm;
  3. For $\varepsilon \to 0$, it holds that $\| f_\varepsilon - f \|_{\infty} \to 0$.

For $d=1$ and $k=2$ I have found the following example in the literature which is stated without proof (it should continue to work for arbitrary $d$ if one replaces the one dimensional Gaussian distribution by a $d$-dimensional one) but don't understand why the second derivative exists (in the classical sense) and is bounded: $$f_\varepsilon(x) = \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty} g(x + \varepsilon y) e^{- y^2/2} \, \mathrm{d}y.$$ Can anybody give me a hint or provide an example of such a sequence?

Thanks in advance!

$\endgroup$
2
  • 1
    $\begingroup$ Do you want the (2nd and higher) derivatives to be bounded for each $\epsilon$ by a constant that may depend on $\epsilon$ or uniformly in $\epsilon$? If uniformly, this is not possible in general. $\endgroup$
    – PhoemueX
    Sep 19, 2014 at 21:09
  • $\begingroup$ @PhoemueX In my case, the derivatives can be bounded by a constant depending on $\varepsilon$ but that's a very interesting point. Is it already impossible in one dimension? What about the norms $\| \partial (f_{\varepsilon} - f)\|_{\infty}$, where $\partial$ stands for some partial derivatives, can these be bounded uniformly in $\varepsilon$? $\endgroup$
    – herrsimon
    Sep 20, 2014 at 7:36

2 Answers 2

4
$\begingroup$

Let $$ h(x)=\left\{\begin{array}{lll}\mathrm{e}^{-1/x^2} & \text{if} & x>0,\\ 0 & \text{if} & x\le 0.\end{array}\right. $$ Then $h\in C^\infty(\mathbb R)$. Then set $$ j(\boldsymbol{x})=c\,h\big(1-\|\boldsymbol{x}\|^2\big), $$ where $\boldsymbol{x}\in\mathbb R^n$, and $c>0$, so that $\int_{\mathbb R^n}j(\boldsymbol{x})\,d\boldsymbol{x}=1$. Clearly, $j\ge 0$, $j\in C^\infty(\mathbb R^n)$ and $\,\mathrm{supp}\,j\subset B(0,1)$ - the unit ball.

Next define $j_e(\boldsymbol{x})=\varepsilon^{-n}j(\varepsilon^{-1}\boldsymbol{x})$. Then $\int_{\mathbb R^n}j_\varepsilon(\boldsymbol{x})\,d\boldsymbol{x}=1$ and let the function $$ f_\varepsilon=f*j_\varepsilon, $$ i.e., $$ f_\varepsilon(\boldsymbol{x})=\int_{\mathbb R^n} f(\boldsymbol{y})\,j_\varepsilon(\boldsymbol{x}-\boldsymbol{y})\,d\boldsymbol{y}=\int_{\mathbb R^n} f(\boldsymbol{x}-\boldsymbol{y})\,j_\varepsilon(\boldsymbol{y})\,d\boldsymbol{y}= \frac{1}{\varepsilon^n}\int_{B(0,\varepsilon)} f(\boldsymbol{x}-\boldsymbol{y})\,j(\boldsymbol{y}/\varepsilon)\,d\boldsymbol{y}=\frac{1}{\varepsilon^n}\\=\int_{B(0,1)} f(\boldsymbol{x}-\varepsilon\boldsymbol{y})\,j(\boldsymbol{y})\,d\boldsymbol{y}. $$ Clearly $f_\varepsilon\in C^\infty(\mathbb R^n)$. Next $$ f_\varepsilon(\boldsymbol{x}_1)-f_\varepsilon(\boldsymbol{x}_2)= \int_{\mathbb R^n} \big(f(\boldsymbol{x}_1-\boldsymbol{y})-f(\boldsymbol{x}_2-\boldsymbol{y})\big)\,j_\varepsilon(\boldsymbol{y})\,d\boldsymbol{y} $$ and hence $$ \lvert\,f_\varepsilon(\boldsymbol{x}_1)-f_\varepsilon(\boldsymbol{x}_2)\rvert\le \int_{\mathbb R^n} \lvert\, f(\boldsymbol{x}_1-\boldsymbol{y})-f(\boldsymbol{x}_2-\boldsymbol{y})\rvert\,j_\varepsilon(\boldsymbol{y})\,d\boldsymbol{y}\le\|\boldsymbol{x}_1-\boldsymbol{x}_2\|\int_{\mathbb R^n}j_\varepsilon(\boldsymbol{y})\,d\boldsymbol{y}=\|\boldsymbol{x}_1-\boldsymbol{x}_2\|. $$ Finally $$ \lvert\,f_\varepsilon(\boldsymbol{x})-f(\boldsymbol{x})\rvert\le \left|\int_{B(0,1)} \big(f(\boldsymbol{x}-\varepsilon\boldsymbol{y})-f(\boldsymbol{x})\big)\,j(\boldsymbol{y})\,d\boldsymbol{y}.\,\right|\le \cdots\le \varepsilon. $$

$\endgroup$
2
  • $\begingroup$ Thank you, now I've understood the main arguments. Actually, you still have to show that all (partial) derivatives are bounded, right? Although this is heuristically clear, I think that taking $\exp{x^2}$ instead of $\exp{1/x^2}$ from the start as in my one-dimensional example would have simplified things (unless I'm missing something). $\endgroup$
    – herrsimon
    Sep 20, 2014 at 7:26
  • $\begingroup$ Is it true that the order $1$ partial derivatives of $f_{\epsilon}$ converge to those of $f$? (they exist by Rademacher theorem) $\endgroup$
    – Leonardo
    Feb 22, 2020 at 12:06
2
$\begingroup$

Rewrite $f_{\epsilon}(x)$ by substituting $y=y'-\frac{1}{\epsilon}x$: $$ \begin{align} f_{\epsilon}(x) & = \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}g(\epsilon y') e^{-(y'-\frac{1}{\epsilon}x)^{2}/2}\,dy' \\ & = \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}g(\epsilon y')e^{-(\epsilon y'-x)^{2}/2\epsilon^{2}}\,dy' \end{align} $$ Now you can see that the derivatives respect to $x$ fall onto the Gaussian, instead onto the function $g$. You can further substitute $y''=\epsilon y$ to make the integral a little more transparent.

$\endgroup$
1
  • $\begingroup$ And again it's a shame that one can only accept a single answer... thanks for the clarification, so $f_{\varepsilon}$ is indeed $\mathcal{C}^{\infty}$. $\endgroup$
    – herrsimon
    Sep 20, 2014 at 7:05

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.