The Green's function for the Laplacian in the plane is $$G(\mathbf{x},\mathbf{x_0}) = \frac{1}{4\pi}\log(||\mathbf{x}-\mathbf{x_0}||^2)$$

That is to say, $G$ satisfies $$\triangle_{\mathbf{x}}G = \delta(\mathbf{x}-\mathbf{x_0})$$

where $\triangle$ is the Laplacian operator, and $\delta$ is the Dirac function.

In an effort to convince myself of this, I computed by hand the Laplacian of $G$, and obtained $$\triangle_{\mathbf{x}}G = \frac{-1}{2\pi}\frac{2r^2-2r^2}{r^4}$$ where $r= ||\mathbf{x}-\mathbf{x_0}||^2$.

Two questions:

Is this really equal to the $\delta$ function? Certainly it's equal to $0$ for $r \neq 0$, but for $r=0$ can we really say that it blows up? I suppose since $r > 0$, the value tends to $\infty$.

The more important question: Why does the initial factor of $\frac{1}{4\pi}$ even matter? We could have started out with any arbitrary constant and obtained the same property for the Laplacian.

  • $\begingroup$ The prefactor $\require{cancel}\cancel{1 \over 4\pi}$ must be, indeed $\color{#f00}{-\,{1 \over 2\pi}}$. Usually, the Green function is defined as $\ln\left\vert\,\vec{r} - \vec{R}\,\right\vert$ such that $\nabla^{2}\ln\left\vert\,\vec{r} - \vec{R}\,\right\vert = -2\pi\,\delta\left(\,\vec{r} - \vec{R}\,\right)$. $\endgroup$ – Felix Marin May 15 '17 at 3:18

We can redefine the problem to circumvent use of the Dirac Delta. So, rather than describing the problem by the expression $\nabla^2 G(\vec \rho|\vec \rho')=\delta(\vec \rho-\vec \rho')$ we describe the function $G(\vec \rho|\vec \rho')$ to satisfy the equations

$$\begin{align} \nabla^2 G(\vec \rho|\vec \rho')&=0\,\,\ \dots\vec \rho\ne \vec \rho'\tag 1\\\\ \lim_{\epsilon\to 0}\oint_{|\vec \rho-\vec \rho'|=\epsilon}\nabla' G(\vec \rho|\vec \rho')\cdot \hat n'\,d\ell'&=1\tag2 \end{align}$$

As determined in the OP, $(1)$ is satisfied for $G(\vec \rho|\vec \rho')=\frac{1}{2\pi }\log(|\vec \rho-\vec \rho'|)$.

To see that this form of $G$ satisfies $(2)$ we see that on $|\vec \rho-\vec \rho'|=\epsilon$, we have $\color{blue}{\nabla' G(\vec \rho|\vec \rho')=\frac{1}{2\pi}\frac{\vec \rho'-\vec \rho}{|\vec \rho-\vec \rho'|^2}}$, $\color{red}{\hat n'=\frac{\vec \rho'-\vec \rho}{|\vec \rho-\vec \rho'|}}$, and $\color{green}{d\ell'=|\vec \rho-\vec \rho'|\,d\phi'}$, where $\phi'\in [0,2\pi]$. Putting everything together reveals

$$\oint_{|\vec \rho-\vec \rho'|=R} \color{blue}{\nabla'G(\vec \rho|\vec \rho')}\cdot \color{red}{\hat n'}\,\color{green}{d\ell'}=\int_0^{2\pi} \color{blue}{\left(\frac{1}{2\pi}\frac{\vec \rho'-\vec \rho}{|\vec \rho-\vec \rho'|^2}\right)}\cdot \color{red}{\left(\frac{\vec \rho'-\vec \rho}{|\vec \rho-\vec \rho'|}\right)}\,\color{green}{\left(|\vec \rho-\vec \rho'|\,d\phi'\right)}=1$$

as was to be shown! In fact, note that we didn't need to let $\epsilon\to 0$. We only require that $\epsilon$ is small enough so that the disk $|\vec \rho-\vec \rho'|=\epsilon$ is contained inside the open domain of interest.

Establishing that Solution to $(1)$ and $(2)$ is the Green (Green's) Function

We show now the motivation for describing the Green (or Green's) function to the problem $\nabla^2 \phi(\rho)=f(\rho)$, where $f$ is a smooth function with compact support. We write $\phi(\vec \rho)$ as the superposition integral

$$\phi(\vec \rho)=\int_S G(\vec\rho|\vec\rho')f(\vec \rho')\,dS'\tag 3$$

We have $\nabla^2 G(\vec \rho|\vec\rho')=0$ for $\vec\rho\ne\vec\rho'$. In addition, we require $\phi$ to satisfy $\nabla^2 \phi(\vec \rho)=f(\vec \rho)$.

Accordingly, we apply the Laplacian to $(3)$ and write

$$\begin{align} \nabla^2 \phi(\vec \rho)&=\nabla \cdot \int_S \nabla G(\vec\rho|\vec\rho')f(\vec \rho')\,dS'\\\\ &=-\nabla \cdot \int_S \nabla' G(\vec\rho|\vec\rho')f(\vec \rho')\,dS'\\\\ &=-\nabla \cdot \int_S \left(\nabla'\left(G(\vec\rho|\vec\rho')f(\vec \rho')\right)-G(\vec\rho|\vec\rho')\nabla'f(\vec \rho')\right)\,dS'\\\\ &=-\nabla \cdot \underbrace{\oint_{\partial S} G(\vec\rho|\vec\rho')f(\vec \rho') \,\hat n'\,d\ell'}_{=0\,\,\text{since}\,f\,\text{is of compact support}}+\nabla \cdot \int_S G(\vec\rho|\vec\rho')\nabla'f(\vec \rho')\,dS'\\\\ &=-\int_S \nabla' G(\vec\rho|\vec\rho')\cdot \nabla'f(\vec \rho')\,dS'\\\\ &=-\int_S \nabla' \cdot \left(\nabla' G(\vec\rho|\vec\rho')f(\vec \rho')-f(\vec \rho')\underbrace{\nabla'^2 G(\vec \rho|\vec \rho')}_{=0}\right)\,dS'\\\\ &=-\underbrace{\oint_{\partial S}\hat n'\cdot\nabla' G(\vec\rho|\vec\rho')f(\vec \rho')\,d\ell'}_{=0,\,\text{since}\,f\,\text{is of compact support}}+\oint_{|\vec \rho-\vec \rho'|=\epsilon}\hat n'\cdot \nabla' G(\vec\rho|\vec\rho')f(\vec \rho')\,d\ell'\\\\ &=\underbrace{\oint_{|\vec \rho-\vec \rho'|=\epsilon}\hat n'\cdot \nabla' G(\vec\rho|\vec\rho')(f(\vec \rho')-f(\vec \rho))\,d\ell'}_{\to 0\,\,\text{as}\,\,\epsilon \to 0}+f(\vec \rho)\oint_{|\vec \rho-\vec \rho'|=\epsilon}\nabla' G(\vec\rho|\vec\rho')\cdot \hat n'\,d\ell' \end{align}$$

Therefore, we see that $\nabla^2 \phi(\vec \rho)=f(\vec \rho)$ if $\nabla^2 G(\vec\rho|\vec\rho')=0$ for $\vec \rho\ne \vec \rho'$ and

$$\lim_{\epsilon \to 0}\oint_{|\vec \rho-\vec \rho'|=\epsilon}\nabla' G(\vec\rho|\vec\rho')\cdot \hat n'\,d\ell'=1$$

This is precisely the description given by $(1)$ and $(2)$ and establishes a way forward without using distributions.

  • $\begingroup$ There is still a bit of "formal integration by parts" going on in order to obtain your second identity, which is the whole idea of the distribution theory calculation, but still, nice. $\endgroup$ – Ian May 14 '17 at 22:14
  • $\begingroup$ @Ian No, there is not. I'm redefining the problem in a way that circumvents generalized functions. $\endgroup$ – Mark Viola May 14 '17 at 22:29
  • $\begingroup$ How do you arrive at (2) then? $\endgroup$ – Ian May 14 '17 at 22:40
  • $\begingroup$ @Ian I've edited to provide the answer to your question. ;-)) $\endgroup$ – Mark Viola May 14 '17 at 23:45
  • $\begingroup$ What a well-explained and thoughtful answer. Many thanks! $\endgroup$ – Alex May 15 '17 at 13:19

You need distribution theory to really know what's going on. In the sense of distribution theory, the equation $\Delta G(x) = \delta(x)$ means that whenever $\phi$ is a smooth function with compact support, you have the "formal" equation

$$\int_{\mathbb{R}^2} (\Delta G)(x) \phi(x) dx = \int_{\mathbb{R}^2} \phi(x) \delta(x) dx.$$

The right side is understood by the definition of the delta function to be $\phi(0)$. The left side is understood through the definition of the distributional derivative as meaning

$$\int_{\mathbb{R}^2} G(x) (\Delta \phi)(x) dx.$$

This has meaning because the singularity in $G$ at $x=0$ is integrable.

This also tells you why the constant factor matters: if you replace $G$ with $cG$ for a constant $c \neq 1$ then you get $c \phi(0)$ out of the integration, which is not what you should have for the Green's function.

  • 3
    $\begingroup$ Another way to understand what's going on is to consider instead some nascent delta function $\delta_t(\mathbf{x})$ which converges (in an appropriate weak sense) to $\delta(\mathbf{x})$ as $t\to 0^+$. In that case, the requirement that $\int_{\mathbb{R}^2}\delta(\mathbf{x})\,d\mathbf{x}=1$ again fixes the overall normalization constant. $\endgroup$ – Semiclassical May 14 '17 at 21:50

Here is a non-rigorous argument (along the lines of SemiClassical's comment), which I hope you find satisfying nonetheless:

Without loss of generality, set $\mathbf x_0 = 0$. So $$ G(\mathbf x, \mathbf 0) = \frac 1 {4\pi} \ln r^2 = \frac 1 {2\pi} \ln r,$$ where $r$ is the radial polar coordinate.

Let's consider the integral $$ \int_{B(0, \epsilon)} \nabla^2_{\mathbf x} G(\mathbf x, \mathbf 0) \ dA,$$ where $B(0, \epsilon)$ is the disk of radius $\epsilon$ centred at the origin. Since $$ \nabla^2_{\mathbf x} G(\mathbf x, \mathbf 0) = \delta (\mathbf x),$$ we expect the integral to be equal to one.

Now, let's check this, by "computing" the integral using the divergence theorem: \begin{multline} \int_{B(0, \epsilon)} \nabla^2_{\mathbf x} G(\mathbf x, \mathbf 0) \ dA = \int_{\partial B(0, \epsilon)} \nabla_{\mathbf x} G(\mathbf x, \mathbf 0) . \mathbf n \ dl \\ = \int_{\phi = 0}^{\phi = 2\pi} \frac{\partial }{\partial r} \left( \frac{1}{2\pi} \ln r \right) \big\rvert_{r = \epsilon} .\epsilon d \phi = \frac{1}{2\pi\epsilon}.(2\pi \epsilon) = 1,\end{multline} which is what we want.

The reason why this argument is non-rigorous is that the "function" we're integrating is not a genuine function: it takes "values" of zero and infinity. As Ian explained, we need to think about the delta function as a distribution in order to do this properly. (See also Section 2.2.4 of Evans's PDEs book.) Nonetheless, I hope my intuitive argument manages to convince you that we need the factor of $\frac 1 {4\pi}$ in the Green's function.


$\newcommand{\bbx}[1]{\,\bbox[8px,border:1px groove navy]{\displaystyle{#1}}\,} \newcommand{\braces}[1]{\left\lbrace\,{#1}\,\right\rbrace} \newcommand{\bracks}[1]{\left\lbrack\,{#1}\,\right\rbrack} \newcommand{\dd}{\mathrm{d}} \newcommand{\ds}[1]{\displaystyle{#1}} \newcommand{\expo}[1]{\,\mathrm{e}^{#1}\,} \newcommand{\ic}{\mathrm{i}} \newcommand{\mc}[1]{\mathcal{#1}} \newcommand{\mrm}[1]{\mathrm{#1}} \newcommand{\pars}[1]{\left(\,{#1}\,\right)} \newcommand{\partiald}[3][]{\frac{\partial^{#1} #2}{\partial #3^{#1}}} \newcommand{\root}[2][]{\,\sqrt[#1]{\,{#2}\,}\,} \newcommand{\totald}[3][]{\frac{\mathrm{d}^{#1} #2}{\mathrm{d} #3^{#1}}} \newcommand{\verts}[1]{\left\vert\,{#1}\,\right\vert}$

Integrate, for example, along a circle in the $\ds{xy}$ plane which encloses the origin. Namely,

\begin{align} \int\nabla^{2}\ln\pars{r}\,\dd^{2}\vec{r} & = \int\nabla\cdot\nabla\ln\pars{r}\,\dd^{2}\vec{r} = \int\braces{\partiald{}{x}\bracks{\partiald{\ln\pars{r}}{x}} - \partiald{}{y}\bracks{-\,\partiald{\ln\pars{r}}{y}}}\,\dd^{2}\vec{r} \\[5mm] & = \int\braces{\nabla \times \bracks{-\,\partiald{\ln\pars{r}}{y}\,\hat{x} + \partiald{\ln\pars{r}}{x}\,\hat{y}}}_{\ z}\,\dd^{2}\vec{r} \\[5mm] &= \int\nabla \times \bracks{-\,\partiald{\ln\pars{r}}{y}\,\hat{x} + \partiald{\ln\pars{r}}{x}\,\hat{y}}\cdot \dd\vec{S} \\[5mm] & = \int\bracks{-\,\partiald{\ln\pars{r}}{y}\,\hat{x} + \partiald{\ln\pars{r}}{x}\,\hat{y}}\cdot\dd\vec{r}\qquad\qquad\pars{~Stokes\ Theorem~} \\[5mm] & = \int\bracks{-\,{y \over r^{2}}\,\hat{x} + {x \over r^{2}}\,\hat{y}}\cdot\dd\vec{r} = -\int{ x\,\dd y - y\,\dd x \over r^{2}} = -\int{\pars{\vec{r} \times \,\dd\vec{r}}_{z} \over r^{2}} \\[5mm] & = -\int{r\verts{\dd\vec{r}} \over r^{2}} = -\,{1 \over r}\ \overbrace{\int\verts{\dd\vec{r}}}^{\ds{2\pi r}} = \bbx{-2\pi}\,,\qquad\quad \nabla^{2}\ln\pars{r} = -2\pi\,\delta\pars{\vec{r}} \end{align}


Another approach, similar in spirit to my comment on Ian's answer, is to modify the Green's function to $G_a(\mathbf{x}-\mathbf{x}_0)=\dfrac{1}{4\pi}\ln[a^2+(\mathbf{x}-\mathbf{x}_0)^2]$. Then $$\Delta G_a(\mathbf{x}-\mathbf{x}_0)=\left(\partial_x^2+\partial_y^2\right)\left(\frac{1}{4\pi}\ln[a^2+(x-x_0)^2+(y-y_0)^2]\right)=\frac{1}{\pi}\frac{a^2}{(a^2+(\mathbf{x}-\mathbf{x_0}^2)^2}.$$ One can then verify that $\int_{\mathbb{R}^2} \Delta G_a(\mathbf{x}-\mathbf{x}_0)\,dA=1$ for all $a>0$; if the constant factor wasn't there, this integral wouldn't equal 1.


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.