Analytic Structure of Singular Integrals

A vast catalogue of physical quantities arrives as an integral whose kernel blows up exactly where you want to evaluate it: the potential of a charged wire on the wire itself, the forward scattering amplitude at physical energy, a Green function on its own light cone, the refractive index inside an absorption band. Strip away the physics and one object remains — the Cauchy-type integral $F(z)=\frac{1}{2\pi i}\int_\Gamma\frac{\phi(t)}{t-z}\,\mathrm{d}t$ , analytic in $z$ everywhere except on the contour $\Gamma$ that carries the density $\phi$ . This article studies its analytic structure: where it is holomorphic, how fast it dies at infinity, and above all how it jumps when you carry $z$ across $\Gamma$ . The crisis comes when $z$ lands on $\Gamma$ , where $1/(t-z)$ stops being integrable. Two devices rescue the integral and turn out to be two faces of one identity. The Cauchy principal value excises a symmetric sliver around the pole and lets it shrink, so the equal-and-opposite infinities cancel and a finite number survives. The $\pm i\varepsilon$ prescription instead nudges the pole off the axis. The Sokhotski–Plemelj formula reconciles them, and once you have it the Plemelj relations follow: the two boundary values differ by exactly the density and average to a principal-value integral of it. On the real line that average is the Hilbert transform, a rotation by 90° that maps the real part of an analytic boundary value onto its imaginary part — the operator content of causality. These few lines, due to Cauchy, Sokhotski, Plemelj, Privalov and Muskhelishvili, are the working grammar of Kramers–Kronig relations, Riemann–Hilbert problems, the Wiener–Hopf method and the analytic S-matrix.

Key takeaways

The Cauchy-type integral $F(z)=\frac{1}{2\pi i}\int_\Gamma \frac{\phi(t)}{t-z}\,\mathrm{d}t$ is analytic off the contour, but its limits from the two sides differ by exactly the density $\phi$ .
The Cauchy principal value $\mathrm{P}\!\int$ extracts a finite, symmetric limit from a non-integrable $1/(t-x)$ singularity sitting on the path.
The Sokhotski–Plemelj identity $\frac{1}{x\mp i0}=\mathrm{P}\frac{1}{x}\pm i\pi\,\delta(x)$ packs the principal value and a half-residue delta into one statement.
The Plemelj jump and average relations $F_+-F_-=\phi$ and $F_++F_-=\frac{1}{\pi i}\mathrm{P}\!\int\frac{\phi}{t-x}\,\mathrm{d}t$ are the backbone of dispersion relations and Riemann–Hilbert problems.
On the real line the average becomes the Hilbert transform $\mathcal{H}$ , the Fourier multiplier $-i\,\mathrm{sgn}\,k$ , an $L^2$ isometry with $\mathcal{H}^2=-\mathbb{1}$ .
Analyticity in a half-plane plus decay turns the reproducing formula into the Kramers–Kronig relations linking dispersion to absorption.
Key idea: A single density $\phi$ on a contour builds an analytic function whose two boundary values differ by $\phi$ and average to a principal-value (Hilbert) transform of $\phi$ ; the Sokhotski–Plemelj identity $\frac{1}{x\mp i0}=\mathrm{P}\frac{1}{x}\pm i\pi\delta(x)$ is what ties the off-contour limit to the on-contour principal value.

The core insight

A single density $\phi$ on a contour generates an analytic function whose two boundary values differ by $\phi$ and average to a principal-value (Hilbert) transform of $\phi$ ; the Sokhotski–Plemelj identity $\frac{1}{x\mp i0}=\mathrm{P}\frac{1}{x}\pm i\pi\delta(x)$ ties the off-contour limit to the on-contour principal value.

Where this lives today

The machinery is anything but a museum piece. In optical and condensed-matter spectroscopy, Kramers–Kronig analysis is the daily tool that turns a measured reflectivity or absorption spectrum into the full complex dielectric function — you cannot measure the phase, so you reconstruct it by Hilbert transform. In high-energy physics the modern S-matrix bootstrap revives the 1960s dream of pinning down scattering amplitudes from analyticity, unitarity and crossing alone: dispersion relations and the Plemelj discontinuity across physical cuts become numerical constraints that yield rigorous bounds on pion scattering and effective field theories. Riemann–Hilbert problems are the analytic engine of the inverse scattering transform for the KdV and nonlinear Schrödinger equations, and the Deift–Zhou steepest-descent method turns them into precise asymptotics for orthogonal polynomials, random-matrix statistics, and the Tracy–Widom distribution. Closer to engineering, the Wiener–Hopf factorisation — Plemelj's construction in another guise — solves half-plane diffraction and crack-tip stress fields with their signature $1/\sqrt{r}$ singularity. The same $i\pi\delta$ a student first meets as "the delta term you must not drop" is the on-shell prescription that makes loop integrals in quantum field theory finite and causal.

The physical picture

Smear the density $\phi$ along a thin wire bent into the shape of $\Gamma$ . The Cauchy integral $F(z)$ is the complex potential of that wire — smooth on either side, but stepping through the wire from one face to the other you pick up exactly the local source strength $\phi$ . The jump is not an accident of the kernel; the jump is the density. That one sentence is the whole theory of sectionally analytic functions, and the first figure is its portrait: an oriented contour splitting the plane into a plus and a minus side, with $F_+$ and $F_-$ differing by $\phi$ at every point.

Figure

A Cauchy integral is sectionally analytic — smooth on each side of the oriented contour Γ, but its boundary limits F₊ (left of the orientation) and F₋ (right) differ by exactly the local density φ. The jump relation F₊ − F₋ = φ holds pointwise for Hölder-continuous φ

The principal value is a tug-of-war that ends in a draw. March toward a first-order pole at $t=x$ and the kernel $1/(t-x)$ races to $+\infty$ on one side and $-\infty$ on the other; cut out a symmetric slice of width $2\varepsilon$ and the two runaway logarithms are equal and opposite, so they annihilate and leave a finite, fair average. Sokhotski–Plemelj is the quantitative version of stepping off the axis instead: displacing the pole by $\pm i\varepsilon$ keeps that average but adds half a residue, because the deformed path now slips entirely past the pole on one side. The delta is the bookkeeping of that half-encirclement, and the second figure shows it being born — the Lorentzian $\varepsilon/[\pi(x^2+\varepsilon^2)]$ , always of unit area, sharpening into a spike of height $1/(\pi\varepsilon)$ as $\varepsilon\to0$ .

Figure

The imaginary part of the ±iε prescription is a nascent delta — the Lorentzian δ_ε(x) = ε /(π(x² + ε²)) keeps unit area for every ε while its peak climbs as 1/(πε): 0.64 at ε = 0.5, up to 1.59 at ε = 0.2. In the limit it becomes π δ(x), the half-residue term of Sokhotski–Plemelj

The core equations

The founding object is the Cauchy-type integral: a density $\phi$ on an oriented contour $\Gamma$ , integrated against the Cauchy kernel $1/(t-z)$ . It defines a function of the free point $z$ that is holomorphic in every region not touching $\Gamma$ — sectionally (or piecewise) holomorphic.

F(z)=\frac{1}{2\pi i}\int_\Gamma\frac{\phi(t)}{t-z}\,\mathrm{d}t \qquad (1)

where: $\Gamma$ — oriented contour (line, arc, or closed curve); $\phi(t)$ — density, assumed Hölder continuous on $\Gamma$ ; $z$ — complex evaluation point with $z\notin\Gamma$ ; $F(z)$ — the resulting sectionally holomorphic function. Read it as a superposition of simple poles, one at each source point $t$ , weighted by $\phi(t)$ . Before that means anything on $\Gamma$ we need a finite value when the pole sits on the path. The Cauchy principal value supplies it by deleting a symmetric neighbourhood of the singular point and shrinking the radius to zero.

\mathrm{P}\!\int_a^b\frac{f(x)}{x-x_0}\,\mathrm{d}x=\lim_{\varepsilon\to0^+}\left(\int_a^{x_0-\varepsilon}+\int_{x_0+\varepsilon}^{b}\right)\frac{f(x)}{x-x_0}\,\mathrm{d}x \qquad (2)

where: $x_0$ — the first-order singularity, $a<x_0<b$ ; $\varepsilon$ — the symmetric exclusion radius, sent to $0^+$ ; $f$ — a regular (e.g. Hölder) numerator. Symmetry is the whole point: an asymmetric cut leaves an uncancelled $\ln$ and diverges. The bridge between deleting the pole and stepping around it is Sokhotski–Plemelj — one distributional identity, valid when both sides act on a smooth test function.

\frac{1}{x\mp i0^+}=\mathrm{P}\frac{1}{x}\pm i\pi\,\delta(x) \qquad (3)

where: $i0^+$ — shorthand for $\lim_{\varepsilon\to0^+}i\varepsilon$ ; $\mathrm{P}\frac{1}{x}$ — the principal-value distribution; $\delta(x)$ — the Dirac delta from the half-residue. The top (lower) sign goes with the pole displaced into the upper (lower) half-plane; the delta is what a physicist loses by "just taking the principal value." Applying this as $z$ approaches a point of $\Gamma$ gives both boundary values at once — a jump term $\pm\tfrac12\phi$ plus a principal-value average.

F_\pm(x)=\pm\tfrac12\phi(x)+\frac{1}{2\pi i}\,\mathrm{P}\!\int_\Gamma\frac{\phi(t)}{t-x}\,\mathrm{d}t \qquad (4)

where: $F_\pm(x)$ — the limit of $F(z)$ as $z\to x\in\Gamma$ from the plus (left of orientation) or minus (right) side; $x$ — a point on the contour. The $\pm\tfrac12\phi$ is the half-encirclement; the integral is common to both sides. Subtracting and adding the two limits strips out the jump and the average — the Plemelj formulas proper, the two facts everything downstream is built on.

F_+(x)-F_-(x)=\phi(x),\qquad F_+(x)+F_-(x)=\frac{1}{\pi i}\,\mathrm{P}\!\int_\Gamma\frac{\phi(t)}{t-x}\,\mathrm{d}t \qquad (5)

where: $F_+-F_-$ — the jump across $\Gamma$ , equal to the density; $F_++F_-$ — the average, a singular Hilbert-type integral. The jump reconstructs the source; the average is the constraint linking real and imaginary parts. On the real line the singular average is exactly the Hilbert transform, which turns the real part of an analytic boundary value into its imaginary part.

(\mathcal{H}f)(x)=\frac{1}{\pi}\,\mathrm{P}\!\int_{-\infty}^{\infty}\frac{f(t)}{x-t}\,\mathrm{d}t,\qquad \widehat{\mathcal{H}f}(k)=-i\,\mathrm{sgn}(k)\,\hat f(k) \qquad (6)

where: $\mathcal{H}$ — the Hilbert transform on $\mathbb{R}$ ; $\hat f$ — the Fourier transform of $f\in L^2$ . The multiplier $-i\,\mathrm{sgn}\,k$ is a 90° phase shift; hence $\mathcal{H}^2=-\mathbb{1}$ and $\|\mathcal{H}f\|_2=\|f\|_2$ . When $F$ is analytic in a half-plane and decays at infinity, that reproducing formula becomes the Kramers–Kronig relations: the dispersive (real) part is the Hilbert transform of the absorptive (imaginary) part.

\operatorname{Re}f(\omega)=\frac{1}{\pi}\,\mathrm{P}\!\int_{-\infty}^{\infty}\frac{\operatorname{Im}f(\omega')}{\omega'-\omega}\,\mathrm{d}\omega' \qquad (7)

where: $f(\omega)$ — a causal response, analytic and bounded in the upper half $\omega$ -plane; $\operatorname{Re}f,\operatorname{Im}f$ — the dispersive and absorptive parts. Measure one across the spectrum and you have computed the other.

Derivation APlemelj boundary values from an indented contour

The goal is to prove equation (4) — that the two boundary limits of the Cauchy integral (1) are a half-density jump plus a common principal-value average — by the most physical route available: deforming the contour with a tiny semicircle around the approach point and reading off where each piece comes from. We take $\Gamma$ smooth and $\phi$ Hölder continuous, approach a point $x_0\in\Gamma$ , isolate the one singular term, evaluate the half-residue from the arc, and reassemble; the Hölder condition is what keeps the leftover integral honest.

Step 1. Add and subtract the value at the target point, $F(z)=\frac{\phi(x_0)}{2\pi i}\int_\Gamma\frac{\mathrm{d}t}{t-z}+\frac{1}{2\pi i}\int_\Gamma\frac{\phi(t)-\phi(x_0)}{t-z}\,\mathrm{d}t$ . Why: the second integrand is bounded near $t=x_0$ since $|\phi(t)-\phi(x_0)|\le C|t-x_0|^\mu$ , so its $z\to x_0$ limit is an ordinary integral — all the singular behaviour is quarantined in the first term.

Step 2. Handle the singular term $\frac{1}{2\pi i}\int_\Gamma\frac{\mathrm{d}t}{t-z}$ by indenting $\Gamma$ with a semicircle of radius $\varepsilon$ around $x_0$ on the side away from $z$ . Why: the straight portions furnish precisely the principal value $\mathrm{P}\!\int_\Gamma\mathrm{d}t/(t-x_0)$ , while the small arc collects a computable fraction of a residue that survives the limit.

Step 3. Evaluate the arc: with $t-x_0=\varepsilon e^{i\theta}$ the semicircle integral is $\frac{1}{2\pi i}\int_0^{\pm\pi} i\,\mathrm{d}\theta=\pm\tfrac12$ . Why: a semicircle subtends angle $\pi$ , half of the $2\pi$ of a full encirclement, so a full residue $1$ becomes $\pm\tfrac12$ ; the sign records which side of $\Gamma$ the point $z$ sat on.

Step 4. Reassemble, $F_\pm(x_0)=\phi(x_0)\big(\pm\tfrac12+\frac{1}{2\pi i}\mathrm{P}\!\int_\Gamma\frac{\mathrm{d}t}{t-x_0}\big)+\frac{1}{2\pi i}\mathrm{P}\!\int_\Gamma\frac{\phi(t)-\phi(x_0)}{t-x_0}\,\mathrm{d}t$ . Why: the $\phi(x_0)$ principal value recombines with the regular remainder into a single principal value of $\phi(t)/(t-x_0)$ , leaving $F_\pm=\pm\tfrac12\phi(x_0)+\frac{1}{2\pi i}\mathrm{P}\!\int_\Gamma\phi(t)/(t-x_0)\,\mathrm{d}t$ .

Step 5. Subtract and add the two limits: $F_+-F_-=\phi$ and $F_++F_-=\frac{1}{\pi i}\mathrm{P}\!\int_\Gamma\frac{\phi(t)}{t-x_0}\,\mathrm{d}t$ . Why: these are the relations (5) — the jump term $\pm\tfrac12\phi$ cancels in the sum and doubles in the difference, separating the density from its singular transform. *Verified: reproduces Muskhelishvili, Singular Integral Equations §17, and the Stone & Goldbart boundary-value formulas for Cauchy integrals.*

Go deeper · Derivation B · 1 minThe Sokhotski–Plemelj identity by rationalising the kernel

Derivation BThe Sokhotski–Plemelj identity by rationalising the kernel

The goal is to prove equation (3) directly as a statement about distributions, showing that the two pieces a physicist keeps — a principal value and an $i\pi\delta$ — are literally the real and imaginary parts of the regularised kernel $1/(x\mp i\varepsilon)$ . We rationalise, integrate each part against a smooth test function, and take $\varepsilon\to0$ ; the real part converges to the principal value and the imaginary part to a delta of the right weight — the entire content of the identity.

Step 1. Multiply by the conjugate, $\frac{1}{x\mp i\varepsilon}=\frac{x\pm i\varepsilon}{x^2+\varepsilon^2}$ , splitting the kernel into a manifestly real and a manifestly imaginary piece. Why: on the real axis the $\varepsilon$ -regularised kernel is a genuine function, so its distributional limit can be studied part by part.

Step 2. For the real part, $\int\frac{x\,g(x)}{x^2+\varepsilon^2}\,\mathrm{d}x\xrightarrow{\varepsilon\to0}\mathrm{P}\!\int\frac{g(x)}{x}\,\mathrm{d}x$ . Why: the weight $x/(x^2+\varepsilon^2)$ is odd and, away from the origin, tends to $1/x$ ; oddness cancels the symmetric neighbourhood of $x=0$ — precisely the principal-value prescription, with no delta.

Step 3. For the imaginary part, $\frac{\varepsilon}{\pi(x^2+\varepsilon^2)}\xrightarrow{\varepsilon\to0}\delta(x)$ . Why: this Lorentzian has unit area for every $\varepsilon$ (the $\arctan$ integrates to $\pi$ , cancelling the $1/\pi$ ) and concentrates at the origin — a nascent delta, the curve drawn in the second figure.

Step 4. Collect real plus imaginary, $\frac{1}{x\mp i0^+}=\mathrm{P}\frac{1}{x}\pm i\pi\,\delta(x)$ . Why: reinstating the $\pm i$ on the imaginary piece produces the signed half-residue $\pm i\pi\delta$ ; dropping it is exactly the error that breaks causality and unitarity downstream. *Verified: the Lorentzian area is $\int_{-\infty}^{\infty}\frac{\varepsilon\,\mathrm{d}x}{\pi(x^2+\varepsilon^2)}=1$ ; matches the standard $i\varepsilon$ propagator prescription of quantum field theory.*

Go deeper · Derivation C · 1 minKramers–Kronig relations from causality

Derivation CKramers–Kronig relations from causality

The goal is to show that a causal linear response — one that cannot react before it is driven — is automatically analytic in the upper half of the complex frequency plane, and that this analyticity, fed through Sokhotski–Plemelj, forces its real and imaginary parts to be Hilbert transforms of each other. We start from strict causality in time, promote it to analyticity in $\omega$ , close a contour, and pick up the half-residue at the physical frequency, giving relation (7).

Step 1. Let $\chi(t)=0$ for $t<0$ (causality) and set $\tilde\chi(\omega)=\int_0^\infty\chi(t)e^{i\omega t}\,\mathrm{d}t$ ; for $\operatorname{Im}\omega>0$ the factor $e^{i\omega t}=e^{i(\operatorname{Re}\omega)t}e^{-(\operatorname{Im}\omega)t}$ decays. Why: with lower time limit $0$ , the damping $e^{-(\operatorname{Im}\omega)t}$ only helps convergence, so $\tilde\chi$ is analytic and bounded in the open upper half-plane — causality has become analyticity.

Step 2. Close the contour along the real axis and a large upper semicircle. Why: with $\tilde\chi$ analytic inside and decaying at infinity the big arc contributes nothing, so the entire content of the loop is the real-axis integral plus whatever the pole at $\omega'=\omega$ supplies.

Step 3. Push $\omega$ down onto the real axis. Indenting around the pole gives $\mathrm{P}\!\int_{-\infty}^{\infty}\frac{\tilde\chi(\omega')}{\omega'-\omega}\,\mathrm{d}\omega'-i\pi\,\tilde\chi(\omega)=0$ , i.e. $\tilde\chi(\omega)=\frac{1}{i\pi}\mathrm{P}\!\int_{-\infty}^{\infty}\frac{\tilde\chi(\omega')}{\omega'-\omega}\,\mathrm{d}\omega'$ . Why: the half-residue $-i\pi\tilde\chi(\omega)$ is Sokhotski–Plemelj again — the same $i\pi\delta$ from the kernel, now the on-shell contribution of the response at its own frequency.

Step 4. Separate real and imaginary parts of $\tilde\chi=\frac{1}{i\pi}\mathrm{P}\!\int\frac{\tilde\chi(\omega')}{\omega'-\omega}\,\mathrm{d}\omega'$ . Why: with $\tilde\chi=\operatorname{Re}\tilde\chi+i\operatorname{Im}\tilde\chi$ and $1/i=-i$ , the real part is $+\frac{1}{\pi}\mathrm{P}\!\int\operatorname{Im}\tilde\chi/(\omega'-\omega)$ and the imaginary part $-\frac{1}{\pi}\mathrm{P}\!\int\operatorname{Re}\tilde\chi/(\omega'-\omega)$ — the two Kramers–Kronig relations, each the Hilbert transform of the other. *Verified: matches the Arfken–Weber dispersion-relation derivation; equivalent to $\operatorname{Re}\tilde\chi=\mathcal{H}[\operatorname{Im}\tilde\chi]$ .*

Limiting cases

How the regularised kernel and the Cauchy integral behave in characteristic limits — each row is a sanity check the general formulas must reproduce.

Case	Result	Meaning
On-contour, distributional	$\lim_{\varepsilon\to0^+}\frac{1}{x\mp i\varepsilon}=\mathrm{P}\frac{1}{x}\pm i\pi\,\delta(x)$	Sokhotski–Plemelj: the $\pm i\varepsilon$ prescription is a principal value plus a signed half-residue delta.
Far field $z\to\infty$	$F(z)\sim -\frac{1}{2\pi i z}\int_\Gamma\phi\,\mathrm{d}t+O(z^{-2})$	decays like $1/z$ ; the coefficient is minus the total "charge", vanishing iff $\int_\Gamma\phi=0$ .
Hölder density on closed $\Gamma$	$F_+-F_-=\phi,\quad F_++F_-=\frac{1}{\pi i}\mathrm{P}\!\int_\Gamma\frac{\phi(t)}{t-x}\,\mathrm{d}t$	the Plemelj relations hold pointwise once $\phi$ is Hölder continuous (Privalov's theorem).
Open arc, endpoint $\phi\neq0$	$F(z)\sim \frac{\phi}{2\pi i}\ln(z-c)$ near endpoint $c$	the average diverges logarithmically at a bare arc endpoint — the origin of edge singularities.
Hilbert transform on $L^2$	$\mathcal{H}^2=-\mathbb{1},\ \\|\mathcal{H}f\\|_2=\\|f\\|_2$	a unitary involution-up-to-sign; the Fourier multiplier $-i\,\mathrm{sgn}\,k$ .

Common trapThe principal value is not "just ignoring the pole"

Two silent errors sink more calculations here than any algebra slip. First, the principal value demands a symmetric deletion, $\lim_{\varepsilon\to0}\big(\int^{x_0-\varepsilon}+\int_{x_0+\varepsilon}\big)$ with the same $\varepsilon$ on both sides. Cut asymmetrically — say $\int^{x_0-\varepsilon}+\int_{x_0+2\varepsilon}$ — and a leftover $\ln 2$ survives the limit, giving a prescription-dependent number (or $\infty$ for a genuine divergence). Second, $\mathrm{P}\frac{1}{x}$ and $\frac{1}{x\pm i0}$ are different distributions, related by $\frac{1}{x\pm i0}=\mathrm{P}\frac{1}{x}\mp i\pi\delta(x)$ . Quietly swapping one for the other drops the $i\pi\delta$ — and that term is the on-shell part of a propagator, the imaginary part of a self-energy, the absorptive channel of an amplitude. Losing it breaks causality and unitarity while looking perfectly reasonable.

Worked examples

1 · A principal-value integral by Sokhotski–Plemelj and residues (fully worked)

Evaluate $I=\mathrm{P}\!\int_{-\infty}^{\infty}\dfrac{\mathrm{d}x}{(x-1)(x^2+1)}$ , a real integral with a first-order pole on the path at $x=1$ .

Step 1. Lift the on-axis pole into the upper half-plane with the $-i0$ prescription. Writing $J=\int_{-\infty}^{\infty}\frac{\mathrm{d}x}{(x-1-i0)(x^2+1)}$ , equation (3) gives $\frac{1}{x-1-i0}=\mathrm{P}\frac{1}{x-1}+i\pi\delta(x-1)$ , so $J=I+i\pi\left.\frac{1}{x^2+1}\right|_{x=1}=I+\frac{i\pi}{2}$ .

Step 2. Compute $J$ by closing in the upper half-plane, where the integrand has simple poles at $x=i$ and the lifted $x=1+i0$ , with residues $\frac{1}{(i-1)(2i)}$ and $\frac12$ . Thus $J=2\pi i\left[\frac{1}{2i(i-1)}+\frac12\right]=\frac{\pi}{i-1}+i\pi$ .

Step 3. Rationalise $\frac{\pi}{i-1}=-\frac{\pi}{2}-\frac{i\pi}{2}$ , so $J=-\frac{\pi}{2}+\frac{i\pi}{2}$ . The imaginary part $+\frac{i\pi}{2}$ is exactly the half-residue predicted in Step 1 — a built-in consistency check.

Step 4. Extract the principal value, $I=\operatorname{Re}J=J-\frac{i\pi}{2}=-\frac{\pi}{2}$ . The answer is $I=-\dfrac{\pi}{2}\approx-1.5708$ . As a cross-check, partial fractions give $\frac{1}{(x-1)(x^2+1)}=\frac{1/2}{x-1}-\frac{(x+1)/2}{x^2+1}$ ; the first and the odd $x$ term vanish by symmetry, leaving $-\frac12\int\frac{\mathrm{d}x}{x^2+1}=-\frac{\pi}{2}$ — the same number, no complex analysis required.

2 · Dispersion from a single absorption line (Kramers–Kronig)

A medium's absorptive response is a single sharp line, $\operatorname{Im}\chi(\omega)=A\big[\delta(\omega-\omega_0)-\delta(\omega+\omega_0)\big]$ , $A>0$ . Use Kramers–Kronig (7) to find $\operatorname{Re}\chi(\omega)$ and evaluate it at $\omega=0$ .

Step 1. The odd combination of deltas is mandatory: a real response $\chi(t)$ forces $\operatorname{Im}\chi$ to be odd in $\omega$ , so absorption at $+\omega_0$ is mirrored at $-\omega_0$ . This is the physical input; the rest is bookkeeping.

Step 2. Insert into (7): $\operatorname{Re}\chi(\omega)=\frac{1}{\pi}\mathrm{P}\!\int_{-\infty}^{\infty}\frac{A[\delta(\omega'-\omega_0)-\delta(\omega'+\omega_0)]}{\omega'-\omega}\,\mathrm{d}\omega'=\frac{A}{\pi}\!\left[\frac{1}{\omega_0-\omega}-\frac{1}{-\omega_0-\omega}\right]$ . The deltas collapse the integral to two evaluations.

Step 3. Combine: $\frac{1}{\omega_0-\omega}+\frac{1}{\omega_0+\omega}=\frac{2\omega_0}{\omega_0^2-\omega^2}$ , so $\operatorname{Re}\chi(\omega)=\dfrac{2A\omega_0}{\pi(\omega_0^2-\omega^2)}$ . This is the textbook single-resonance dispersion curve, changing sign as $\omega$ crosses $\omega_0$ — the hallmark of anomalous dispersion, produced purely from causality plus one absorption line.

Step 4. Evaluate the static response: $\operatorname{Re}\chi(0)=\dfrac{2A}{\pi\omega_0}$ . For $A=1$ , $\omega_0=2$ (arbitrary units) this is $\frac{1}{\pi}\approx0.318$ ; at $\omega=1$ it gives $\frac{4}{3\pi}\approx0.424$ . The single line has fixed the entire dispersion curve.

Physics of the portrait

In this subject the physics — dispersion, causality, scattering — is entirely downstream of a handful of nineteenth-century theorems about boundary values. The name most tied to the punchline is Josip Plemelj, who wrote the jump-and-average formulas in the form physicists still use and, in the same breath, opened the Riemann–Hilbert problem.

Origins · 1 minJosip Plemelj and how this was discovered

Josip Plemelj

1873–1967, Slovenian mathematician · boundary values of Cauchy-type integrals and the Riemann–Hilbert problem

Josip Plemelj was the founding rector of the University of Ljubljana and, decades earlier, the young analyst who gave the theory of Cauchy integrals its working form. Building on Julian Sokhotski's 1873 St. Petersburg dissertation — where the jump formula first appeared, largely unnoticed in the West — Plemelj in 1908 derived the boundary-value relations $F_+-F_-=\phi$ and $F_++F_-=\frac{1}{\pi i}\mathrm{P}\!\int\phi/(t-x)\,\mathrm{d}t$ that now carry both men's names. His deeper aim was Hilbert's twenty-first problem: to build a linear differential system with a prescribed monodromy group, which he reduced to finding a sectionally analytic function with a given multiplicative jump $g$ across a contour — the scalar Riemann–Hilbert problem. His solution, an exponential of a Cauchy integral of $\log g$ , is the direct ancestor of the Wiener–Hopf method and of the matrix Riemann–Hilbert machinery that today powers inverse scattering, random-matrix asymptotics, and the analytic S-matrix. He also proved the sharp continuity results (later completed by Privalov) fixing exactly how smooth $\phi$ must be — Hölder, not merely continuous — for the formulas to hold pointwise, a bridge from Cauchy's era of contour deformation to the operator language of modern mathematical physics.

History

1814

— Cauchy launches the theory of integrals of functions of a complex variable

1873

— Sokhotski states the jump formula for Cauchy-type integrals in his dissertation

1908

— Plemelj derives the boundary formulas and attacks the Riemann–Hilbert problem

1927

— Kramers and Kronig link refraction and absorption by dispersion relations

1953

— Muskhelishvili's Singular Integral Equations systematizes the whole theory

Going further

Texts. Stone & Goldbart, Mathematics for Physics, Ch. 18–19 (Cauchy integrals, Plemelj, Hilbert transform) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 11 (principal values, dispersion relations) · Hassani, Mathematical Physics, Ch. 11–12 · Nakahara, Geometry, Topology and Physics, §1.4 (complex analysis background) · Byron & Fuller, Mathematics of Classical and Quantum Physics, Ch. 6–7 · Muskhelishvili, Singular Integral Equations · Gakhov, Boundary Value Problems · King, Hilbert Transforms, Vols. I–II.

Provenance. Every equation is a standard textbook result and each was re-verified for this article. A checker should: (1) recompute the half-residue $\pm\tfrac12$ from $\frac{1}{2\pi i}\int_0^{\pm\pi}i\,\mathrm{d}\theta$ ; (2) confirm the Lorentzian $\varepsilon/[\pi(x^2+\varepsilon^2)]$ has unit area and that $\mathcal{F}[1/(\pi x)]=-i\,\mathrm{sgn}\,k$ , hence $\mathcal{H}^2=-\mathbb{1}$ by Plancherel; (3) reproduce Worked Example 1, $\mathrm{P}\!\int_{-\infty}^{\infty}\mathrm{d}x/[(x-1)(x^2+1)]=-\pi/2\approx-1.5708$ , both by Sokhotski–Plemelj-plus-residues and by partial fractions; (4) reproduce Worked Example 2, $\operatorname{Re}\chi(\omega)=2A\omega_0/[\pi(\omega_0^2-\omega^2)]$ . Cross-check the Plemelj formulas against Muskhelishvili §17 and Stone & Goldbart Ch. 18–19, and the dispersion relations against Arfken–Weber Ch. 11.

Builds on: Complex Integration and Contour Integrals · Cauchy's Integral Formula and Derivatives · Analytic Functions and the Cauchy–Riemann Equations · Zeros, Poles, and Isolated Singularities · Harmonic Functions and Conjugates. Leads to: Kramers–Kronig and Dispersion Relations · Riemann–Hilbert Problems · Wiener–Hopf Method · Analytic S-matrix Theory.

Ask the tutor: Ask me why the $i\pi\delta$ term in Sokhotski–Plemelj is exactly the on-shell part of a propagator, or to reconstruct a full dispersion curve from a measured absorption spectrum step by step.

What is the difference between the principal value and the ±i0 prescription?

They answer two different questions about the same integral. The principal value $\mathrm{P}\!\int$ stays on the real axis and deletes a symmetric sliver around the pole, returning a real number. The $\pm i0$ prescription instead moves the pole infinitesimally off the axis, so the integral never hits a singularity and stays complex. Sokhotski–Plemelj relates them, $\frac{1}{x\mp i0}=\mathrm{P}\frac{1}{x}\pm i\pi\delta(x)$ : the $\pm i0$ answer is the principal value plus a purely imaginary half-residue $\pm i\pi f(x_0)$ . If your problem is genuinely real and symmetric you want the principal value; if it carries a causal or on-shell condition (a propagator, a response function) you want the prescription with its delta term intact.

Why does the jump of a Cauchy integral equal the density exactly?

Because the whole discontinuity is manufactured by a single pole that the free point $z$ crosses. As $z$ passes through the contour at $x$ , the kernel $1/(t-z)$ swings the contribution of the immediate neighbourhood of $t=x$ by a full residue, $2\pi i$ times the local weight $\phi(x)/(2\pi i)=\phi(x)$ . Everything else — the far parts of the contour, the smooth variation of $\phi$ — is continuous across the crossing and cancels in $F_+-F_-$ . What survives is precisely the local source strength. This is why a Cauchy integral is the natural way to encode a prescribed jump: hand it any Hölder density and it builds a sectionally analytic function with that exact discontinuity — the starting point for every Riemann–Hilbert construction.

What does "sectionally analytic" mean, and why insist on it?

A sectionally (or piecewise) holomorphic function is analytic in each open region into which $\Gamma$ divides the plane, but generally discontinuous across $\Gamma$ itself. The Cauchy integral (1) is the archetype: holomorphic on the plus side, holomorphic on the minus side, with a controlled jump on the seam. Insisting on this structure is what lets you pose problems like "find $F$ analytic off $\Gamma$ with $F_+=gF_-+h$ " — the Riemann–Hilbert problem. Ordinary analytic-continuation intuition, which forbids any jump, is too rigid to describe scattering data, spectral densities, or diffraction, all naturally encoded as discontinuities. Sectional analyticity is the precise home for "an analytic function with a prescribed cut."

How exactly does causality turn into analyticity in the upper half-plane?

Through the one-sided Fourier transform. A causal response has $\chi(t)=0$ for $t<0$ , so its transform $\tilde\chi(\omega)=\int_0^\infty\chi(t)e^{i\omega t}\,\mathrm{d}t$ runs only over $t\ge0$ . Give $\omega$ a positive imaginary part and the integrand gains a factor $e^{-(\operatorname{Im}\omega)t}$ that decays because $t$ is never negative. That extra convergence makes $\tilde\chi$ analytic and bounded throughout the upper half-plane. No poles up there means you can close a contour and derive a dispersion relation: the absence of a response before the stimulus is literally the absence of singularities above the real axis. Reverse the time convention and it is the lower half-plane, but the logic is identical.

When do I need a subtracted dispersion relation?

Whenever the response does not decay to zero at high frequency. The unsubtracted relation (7) assumes the big semicircle contributes nothing, which needs $\tilde\chi(\omega)\to0$ as $|\omega|\to\infty$ . If instead $\tilde\chi$ tends to a nonzero constant (or grows), that arc no longer vanishes and the naive integral diverges. The fix is to write the dispersion relation for $\tilde\chi(\omega)-\tilde\chi(\omega_0)$ at some reference $\omega_0$ ; the subtraction improves the large- $\omega$ falloff by an extra power, restoring convergence, at the cost of one experimentally fixed constant $\tilde\chi(\omega_0)$ . In particle physics these subtraction constants are physical inputs (scattering lengths, low-energy theorems); in optics a single subtraction handles the constant high-frequency limit of the dielectric response.

What is a Riemann–Hilbert problem, and how do the Plemelj formulas solve it?

A scalar Riemann–Hilbert problem asks for a sectionally analytic $F$ whose boundary values satisfy a prescribed multiplicative jump, $F_+(x)=g(x)F_-(x)$ on $\Gamma$ , with $F$ well-behaved at infinity. Take logarithms: $\log F_+-\log F_-=\log g$ , exactly the Plemelj jump relation with density $\phi=\log g$ . So $\log F(z)=\frac{1}{2\pi i}\int_\Gamma\frac{\log g(t)}{t-z}\,\mathrm{d}t$ — $F$ is the exponential of a Cauchy integral of $\log g$ . That single line is Plemelj's solution, ancestor of the Wiener–Hopf method (which splits $g$ into upper- and lower-analytic factors) and, in matrix form, the backbone of inverse scattering. The Plemelj formulas are quite literally the inversion formula reconstructing an analytic function from its jump.