4

The Sokhotski–Plemelj theorem for the real line is stated at https://en.wikipedia.org/wiki/Sokhotski–Plemelj_theorem:

Sokhotski–Plemelj theorem. Let $f$ be a complex-valued function that is defined and continuous on the real line, and let $a$, $b$, and $x_0$ be real constants with ${\displaystyle a<x_0<b}$. Then $${\displaystyle \lim _{\varepsilon \to 0^{+}}\int _{a}^{b}{\frac {f(x)}{x-x_0\pm i\varepsilon }}\,dx=\mp i\pi f(x_0)+{\mathcal {P}}\int _{a}^{b}{\frac {f(x)}{x-x_0}}\,dx.}$$

I also gather from Confusion concerning the Sokhotski–Plemelj theorem: two different values for the same real integral that there is a more general version of the theorem that is valid over the whole real line, so that, under some mild hypotheses on $f(x)$, one has $${\displaystyle \lim _{\varepsilon \to 0^{+}}\int _{-\infty}^{\infty}{\frac {f(x)}{x-x_0\pm i\varepsilon }}\,dx=\mp i\pi f(x_0)+{\mathcal {P}}\int _{-\infty}^{\infty}{\frac {f(x)}{x-x_0}}\,dx.}$$

My question is: what (mild) hypotheses on $f(x)$ are sufficient for the more general version stated above to hold? I'm also looking for a proof or appropriate reference.

  • I'm looking for the real-line version that includes the entire real line, for integrals from $-\infty$ to $\infty$, not integrals from $a$ to $b$. – Jesse Elliott May 18 '21 at 22:01
  • 1
    As you comment below, there is no need for $f$ to be holomorphic... More later, if no one else gives the answer I would give. :) – paul garrett Dec 03 '21 at 21:17

2 Answers2

3

The first version also holds for $a,b\in\overline{\mathbb{R}}=\mathbb R\cup\{+\infty\}\cup\{-\infty\}$. The important part of the Sokhotski–Plemelj theorem is integrating around $x_0$ and not about the limits of the integral. It is not enough to say that $f$ must be continuous. The function also cannot have singularities neighborhood around $[a,b]$. Keep this in mind at the end result.

Proof: Define the contour $C=[a,x_0-\delta]\cup C_{x_0,\delta}\cup [x_0+\delta,b]$, where $C_{x_0,\delta}$ is a semicircular path of radius $\delta$ centered at $x_0$ with $\delta>0$. Then for $\delta\to0^+$ $$\int_C\frac{f(x)}{x-x_0}\,\mathrm{d}x=\mathcal P\int_a^b \frac{f(x)}{x-x_0}\,\mathrm{d}x+\int_{C_{x_0,\delta}}\frac{f(x)}{x-x_0}\,\mathrm{d}x.$$ Note that the contour $C_{x_0,\delta}$ can be parameterized as $x=x_0+\delta e^{i\varphi}$ for $0\leqslant\varphi\leqslant\pi$ or $-\pi\leqslant\varphi\leqslant0$. Taking $\delta\to0^+$, $$\lim_{\delta\to0^+}\int_{\pm\pi}^0\frac{f(x_0+\delta e^{i\varphi})}{\delta e^{i\varphi}}i\delta e^{i\varphi}\,\mathrm{d}\varphi=f(x_0)\lim_{\delta\to0^+}\int_{\pm\pi}^0\frac{i\delta e^{i\varphi}}{\delta e^{i\varphi}}\,\mathrm{d}\varphi=\mp i\pi f(x_0).$$
We were able to pull the limit inside $f$, since it is by definition continuous on the real line. Assuming that $f$ has no singularities neighborhood around the interval $[a,b]$, we can deform the contour $C$ into a straight line, which runs from $a +i\varepsilon$ to $b+i\varepsilon$, so $$\int_C\frac{f(x)}{x-x_0}\,\mathrm{d}x=\lim_{\varepsilon\to0^+}\int_{a\pm i\varepsilon}^{b\pm i\varepsilon}\frac{f(x)}{x-x_0}\,\mathrm{d}x=\lim_{\varepsilon\to0^+}\int_{a}^{b}\frac{f(x\pm i\varepsilon)}{x-x_0\pm i\varepsilon}\,\mathrm{d}x.$$
We may pull the limit inside $f$ again since the function is continuous on the real line. The end result is $${\displaystyle \lim _{\varepsilon \to 0^{+}}\int _{a}^{b}{\frac {f(x)}{x-x_0\pm i\varepsilon }}\,\mathrm{d}x=\mp i\pi f(x_0)+{\mathcal {P}}\int _{a}^{b}{\frac {f(x)}{x-x_0}}\,\mathrm{d}x},$$ where $a<x_0<b$ and $a,b \in\overline{\mathbb{R}}$.

vitamin d
  • 5,783
  • 1
    Three questions: (1) In the version of the theorem stated on Wikipedia, $f(x)$ is assumed to be function of a real variable, not of a complex variable. Shouldn't there be a proof that doesn't require $f$ to be a function of a complex variable? (2) The version on Wikipedia also explicitly states: "Note that this version makes no use of analyticity." It claims that continuity is all that is needed. Is the Wikipedia page wrong? (3) Don't you need some decay conditions on $f(x)$ so that the integral on an infinite interval exists? – Jesse Elliott May 19 '21 at 23:28
  • (1) I can't see where it is written that $f$ must be a function of a real variable. (2) If the bounds are infinite, then of course $f$ must satisfy certain conditions, which are not needed in the finite version. At the end it breaks down to if the integrals conv. or not, which is a trivial condition. (3) Indeed, analyticity is not needed. The condition that $f$ cannot have singularities in the neighborhood around $[a,b]$ is finicky and is almost never violated, which is probably the reason why Wikipedia didn't include this information - but is still needed as shown in this answer. – vitamin d May 19 '21 at 23:48
  • 1
    (1) I can't see where it is written that $f$ must be a function of a complex variable, which is a much stronger condition than being a function of a real variable. "Let $f$ be a complex-valued function which is defined and continuous on the real line" means to me that it is a function of a real variable. Nothing in the statement of the theorem implies that it must be defined for complex values of the variable $x$. (3) The lack of singularities is needed for your proof, but that doesn't mean it is needed in the statement of the theorem---unless, of course, you have a counterexample. – Jesse Elliott May 20 '21 at 00:22
  • 1
    Also, in the examples the theorem is often applied to, $f(x)$ is a density function of some measure on $\mathbb R$, and therefore the assumption that $f(x)$ is a function of a complex variable is unwarranted. – Jesse Elliott May 20 '21 at 00:28
  • (1) The proof holds for all complex functions (+mild conditions.) Note that if you have a strictly real function, you can apply analytic continuation to work with complex contour integration. Functions are usually denoted the same. (3) What does "my proof" mean? This is a variation (adding $a,b,x_0$) of a well known theorem. This is a rigorous proof with which the theorem is usually proved. Wikipedia is not known for absolute rigorosity. – vitamin d May 20 '21 at 10:56
  • OK, thank you for the information! Is there a reference I can find this proof in? I'd like to cite it in a paper I'm writing. (BTW, wikipedia also provides a "proof" of the theorem as it is stated there, but I don't follow the proof. I guess maybe you could find the error there.) – Jesse Elliott May 20 '21 at 19:50
  • The references in the Wikipedia article are useless. After searching for quite some time I found this arcticle which cites: Vladimirov, V. S. Equations of Mathematical Physics. New York: Dekker, 1971. – vitamin d May 20 '21 at 20:48
  • @JesseElliott If my answer has helped you, I'd appreciate a checkmark. – vitamin d May 21 '21 at 22:39
  • It helped, and I just gave it a ^, but I'm not ready to accept it as the answer because I'm not convinced that there's not a more general version with milder conditions on $f(x)$ and that the "theorem" as stated on Wikidepia is false. – Jesse Elliott May 22 '21 at 05:43
1

To make clear that the function $f$ need not be holomorphic: whether the integrals are over finite or infinite intervals, taking $x_0=0$ for simplicity of notation, $$ \lim_\epsilon\int {f(x)\over x+i\epsilon}dx \;=\; \lim_\epsilon\int {f(x)-f(0)e^{-x^2}\over x+i\epsilon}dx + f(0)\lim_\epsilon\int {e^{-x^2}\over x+i\epsilon}dx $$ $$ \;=\; \int {f(x)-f(0)e^{-x^2}\over x}dx + f(0)\lim_\epsilon\int {e^{-x^2}\over x+i\epsilon}dx $$ whenever ${f(x)-f(0)e^{-x^2}\over x}$ is locally integrable at $0$. The integral involving the Gaussian can be evaluated explicitly as a function of $\epsilon$. :)

paul garrett
  • 52,465