Let $f:\mathbb R\to \mathbb R$ be a function differentiable on $\mathbb R\backslash\{0\}$. Define the jump of the derivatives of $f$ by $$ \sigma_k =\lim_{\epsilon\to 0} (f^{(k)}(\epsilon)-f^{(k)}(-\epsilon)). $$ $f$ defines a distribution in $\mathcal D'(\mathbb R)$. Let $f'$ be the derivative of $f$ in the sense of distributions, and let $\{f'\}$ be the distribution given by the derivative $\frac{df}{dx}$ in the sense of functions. (This is awkward - please suggest better notation if possible.)

Now, by direct computation, $$ f'=\{f'\}+\sigma_0\delta. $$

This question concerns the generalisation of this to higher dimensions. Let's now work in $\mathbb R^n$. Let $S$ be a surface in $\mathbb R^n$ defined by $S=\{\mathbf x\in\mathbb R^n: F(\mathbf x)=0\}$, where $\nabla F\neq 0$. Let $f$ be a function defined on $\mathbb R^n$ and differentiable on $\mathbb R^n\backslash S$. For $\mathbf x\in S$ and a multi-index $\alpha$, define $$ \sigma_\alpha(x)=\lim_{\mathbf y\to \mathbf x\\ F(\mathbf x)>0} \partial^\alpha f(\mathbf y)-\lim_{\mathbf y\to \mathbf x\\ F(\mathbf x)<0} \partial^\alpha f(\mathbf y) $$

Now, consider the derivative $\frac{\partial f}{\partial x_1}$ in the sense of distributions. Let $\phi\in \mathcal D$. Under the assumption that on $S$, $x_1=x_1(x_2,...,x_n)$, $$ \langle\frac{\partial f}{\partial x_1},\phi\rangle =-\int \mathbb d x_2...\mathbb d x_n\int f(x_1,..., x_n) \frac{\partial \phi}{\partial x_1} \mathbb dx_1\\ =\int \mathbb d x_2...\mathbb d x_n\int \left(\frac{\partial f(x_1,..., x_n)}{\partial x_1}+\delta_{x_1(x_2,...,x_n)}\sigma_0(x_1(x_2,...,x_n))\right) \phi(x_1,\ldots, x_n) \mathbb dx_1\\ =\int \mathbb d x_2\ldots\mathbb d x_n\left[\int \left(\frac{\partial f(x_1,..., x_n)}{\partial x_1}\phi(x_1,\ldots, x_n)\right) \mathbb dx_1+\phi(x_1(x_2,...,x_n),\ldots, x_n)\sigma_0(x_1(x_2,...,x_n))\right]\\ =\langle \{\frac{\partial f}{\partial x_1}\},\phi \rangle+\int \phi(x_1(x_2,...,x_n),\ldots, x_n)\sigma_0(x_1(x_2,...,x_n)) \mathbb d x_2\ldots\mathbb d x_n $$ Note that the second term of the last expression is $\int_S \sigma_0 \phi \cos \theta_1 ds$, where $\theta_1$ is the angle between $x_1$-axis and the normal of the surface, because the surface element $ds=\cos \theta_1 \mathbb d x_2\ldots\mathbb d x_n$.

Now, here is my question: can I establish the result $$ \langle\frac{\partial f}{\partial x_1},\phi\rangle =\langle \{\frac{\partial f}{\partial x_1}\},\phi \rangle+\int_S \sigma_0 \phi \cos \theta_1 ds $$ rigorously and formally, without the assumption that $x_1$ is a function of other $x_i$?

Also: $\int_S \sigma_0 \phi \cos \theta_1 ds$ also gives a distribution. Is there a special notation for this distribution?


1 Answer 1


With $g,h$ smooth $\Bbb{R^n\to R}$ $$f(x) = g(x) 1_{h(x) > 0}$$ is a distribution, $$\partial_{x_1} f =(\partial_{x_1} g) 1_{h > 0}+g\ (\partial_{x_1}1_{h>0})=(\partial_{x_1} g) 1_{h > 0}+g\ (\partial_{x_1}h)\delta(h)$$ where $\partial_{x_1}1_{h>0}= (\partial_{x_1}h)\delta(h)$ is always a distribution but for $$\delta(h)= \underset{\text{in the sense of distributions}}{\lim_{t\to 0}} \frac{1_{h \in [0,t]}}{t}$$ to be a distribution on its own we need that $\|\nabla h\|$ doesn't vanish on $h=0$ and $h$ has locally finitely many vanishing hypersurfaces. If so the second term of

$$\langle \partial_{x_1} f,\phi\rangle= \langle (\partial_{x_1} g) 1_{h > 0},\phi\rangle-\langle 1_{h>0} , \partial_{x_1} (g\phi)\rangle$$

can be evaluated as $$-\langle 1_{h>0} , \partial_{x_1} (g\phi)\rangle =\langle \delta(h) , (\partial_{x_1} h)g\phi\rangle=\int_{h = 0} (\partial_{x_1} h)g\phi d\nu $$ $\nu$ is the measure such that for $h(x)=0$ as $r\to 0$ it approximates $\frac{1_{h\in [0,r]}}{r}$ ie. $$\nu(B_r(x)\cap (h=0))\sim \frac{Vol(B_r(x)\cap h\in [0,r])}{r}\sim \frac{S_{n-1}r^{n-1}}{\|\nabla h(x)\|}$$

  • $\begingroup$ Are you assuming that $f$ is zero when $h(x)<0$? What about the general case (the $f$ described in my question)? Thank you. $\endgroup$
    – Ma Joad
    Dec 31, 2019 at 5:06
  • $\begingroup$ Also, what is $1_{h\in [0,r]}$? I think it is an indicator function, but in the last line, it seems to be a set. $\endgroup$
    – Ma Joad
    Dec 31, 2019 at 5:08
  • $\begingroup$ With $g(x) 1_{h(x) > 0}+g_2(x) 1_{h(x) < 0}$ (if $\| \nabla h\|$ doesn't vanish) you have all the functions with a jump discontinuity at $h=0$ $\endgroup$
    – reuns
    Dec 31, 2019 at 5:09
  • $\begingroup$ Thanks. What about the measure $\nu$? Why the second expression in the last line (with $B_r(x)$ in it) is the surface element? $\endgroup$
    – Ma Joad
    Dec 31, 2019 at 5:14
  • $\begingroup$ Also, $g$ doesn't need to be $\mathbb R^n\to \mathbb R$, does it? $g$ just need to be defined in the region where $h>0$. $\endgroup$
    – Ma Joad
    Dec 31, 2019 at 7:11

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.