The Mellin Transform

The Mellin transform is the integral transform built for the multiplicative world. Where the Fourier transform diagonalizes translation and the Laplace transform tames one-sided decay, the Mellin transform diagonalizes dilation — the stretching $x\mapsto\lambda x$ on the positive half-line $(0,\infty)$ . Its kernel $x^{s-1}$ is exactly the character of the scaling group: rescale the input and the transform merely picks up a power $\lambda^{-s}$ . That one algebraic fact organizes the whole theory, because $(0,\infty)$ under multiplication is the same group as $(\mathbb{R},+)$ under addition, viewed through the logarithm. This kinship makes the Mellin transform a Fourier transform in disguise, reached by $x=e^{-t}$ . Convergence, once a maze of endpoint estimates, becomes a single geometric object: the fundamental strip $a<\mathrm{Re}\,s<b$ , the exact image of the Fourier axis. But the transform earns its place in mathematical physics through something Fourier analysis cannot do so cleanly — it turns asymptotics into algebra. Slide the inversion contour across a pole of $\tilde f(s)$ and you peel off one term of the expansion of $f(x)$ near an endpoint: the residue is the coefficient, the order of the pole is the power of the logarithm. The question "how does $f$ behave as $x\to0$ ?" becomes "where are the poles of $\tilde f$ ?". This is the machinery behind the Gamma and Riemann-zeta functions, Mellin–Barnes contour integrals, the heat-kernel expansions of quantum field theory, and the analytic continuation that defines dimensionally regularized Feynman integrals.

Key takeaways

The Mellin transform $\tilde f(s)=\int_0^\infty f(x)\,x^{s-1}\,\mathrm{d}x$ maps a function on $(0,\infty)$ to a function of a complex variable $s$ , with the scaling group of positive reals as its natural symmetry.
It is the Fourier transform pulled back along the logarithm: setting $x=e^{-t}$ turns scaling on $(0,\infty)$ into translation on $\mathbb{R}$ , so every Mellin fact is a Fourier fact in new coordinates.
Convergence lives on a vertical strip $a<\mathrm{Re}\,s<b$ — the fundamental strip — fixed by the endpoint power laws of $f$ ; inversion is a Bromwich contour integral $\tfrac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}$ with $a<c<b$ .
Poles of $\tilde f(s)$ encode asymptotics: residues to the left of the strip give the small- $x$ expansion, residues to the right give the large- $x$ expansion, and a pole of order $m$ makes a $\log^{m-1}x$ .
The canonical output is $\Gamma(s)=\int_0^\infty e^{-x}x^{s-1}\mathrm{d}x$ (poles at $s=-n$ , residues $(-1)^n/n!$ ); the Bose–Einstein integral $\int_0^\infty x^{s-1}/(e^x-1)\,\mathrm{d}x=\Gamma(s)\zeta(s)$ gives the Stefan–Boltzmann number $\int_0^\infty x^3/(e^x-1)\,\mathrm{d}x=\pi^4/15=6.4939$ .
Key idea: set $x=e^{-t}$ and the Mellin transform becomes the two-sided Laplace/Fourier transform of $f(e^{-t})$ ; scaling is translation, the fundamental strip is a strip of convergence, and the poles of $\tilde f(s)$ are the residue-bookkeeping of $f$ 's endpoint asymptotics.

The core insight

Set $x=e^{-t}$ : the Mellin transform $\tilde f(s)=\int_0^\infty f(x)x^{s-1}\mathrm{d}x$ becomes the two-sided Laplace/Fourier transform of $f(e^{-t})$ in $t$ . Scaling in $x$ is translation in $t$ , the fundamental strip is a strip of convergence, and the poles of $\tilde f(s)$ are the residue-bookkeeping of $f$ 's endpoint asymptotics.

Where this lives today

The Mellin transform is the day-to-day engine of perturbative quantum field theory. Multi-loop Feynman integrals are evaluated by the Mellin–Barnes method: each propagator denominator $(A+B)^{-\alpha}$ is split with the representation of Derivation 3, the loop momenta integrated to yield products of Gamma functions, and the surviving contour integrals resolved into nested sums or evaluated numerically. The poles in $\varepsilon=(4-d)/2$ that encode ultraviolet and infrared divergences are literally residues of these Gamma factors, so dimensional regularization is Mellin analysis. Elsewhere, the heat-kernel expansion $\mathrm{Tr}\,e^{-t\hat H}\sim\sum a_k t^{(k-d)/2}$ behind spectral zeta functions and one-loop effective actions is a textbook Mellin-asymptotics calculation, and the analytic number theory around the Riemann hypothesis still runs on Riemann's theta-Mellin identity. Whenever a problem has a scale symmetry, the Mellin transform diagonalizes it.

The physical picture

The Mellin variable $s$ is a scale-frequency. A Fourier mode $e^{i\omega t}$ is the eigenfunction of translation — shift $t$ and it only changes phase. In the same way $x^{s-1}$ is the eigenfunction of scaling: stretch $x\mapsto\lambda x$ and it changes only by $\lambda^{-s}$ . The transform never mixes different values of $s$ , so the whole apparatus of "decompose into modes, act mode by mode, reassemble" carries over from frequency to scale. The first figure makes the dictionary explicit: the multiplicative half-line is the additive line seen through the logarithm, and the kernel $x^{s-1}\,\mathrm{d}x$ is the character $e^{-st}\,\mathrm{d}t$ .

Figure

The logarithmic dictionary: the substitution x = e⁻ᵗ turns the multiplicative half-line (0,∞) under scaling x → λx into the additive line ℝ under translation t → t + ln λ, and the Mellin kernel xˢ⁻¹ dx becomes the Laplace/Fourier kernel e⁻ˢᵗ dt

The fundamental strip is a tug-of-war between the two ends of the half-line. Near $x=0$ , if $f\sim x^{-a}$ then the integrand $\sim x^{s-1-a}$ is integrable only when $\mathrm{Re}\,s>a$ ; near $x=\infty$ , if $f\sim x^{-b}$ then integrability needs $\mathrm{Re}\,s<b$ . The transform lives in the overlap $a<\mathrm{Re}\,s<b$ , and the bounding powers are the leading endpoint behaviours of $f$ . Beyond the strip $\tilde f(s)$ continues meromorphically, and its poles are the higher terms of the endpoint expansions. The second figure shows this for $f(x)=e^{-x}$ , whose transform is $\Gamma(s)$ : closing the contour left and summing the residues at $s=0,-1,-2,\dots$ reconstructs the Taylor series exactly.

Figure

Asymptotics by residues for f(x) = e⁻ˣ, whose Mellin transform is Γ(s) with fundamental strip Re s > 0: closing the inversion contour to the left past the simple poles at s = 0, −1, −2, −3 (residues 1, −1, ½, −⅙) reconstructs the series e⁻ˣ = 1 − x + x²/2 − x³/6 = Σ (−1)ⁿ xⁿ/n!

The core equations

The definition is a single integral against the scaling character, pairing a function on the half-line with the kernel $x^{s-1}$ . The only subtlety — the one everything downstream depends on — is that it converges not for all $s$ but on a vertical strip.

\tilde f(s)=\mathcal{M}[f](s)=\int_0^{\infty} f(x)\,x^{s-1}\,\mathrm{d}x,\qquad a<\mathrm{Re}\,s<b \qquad (1)

where: $f(x)$ — a function on $(0,\infty)$ ; $s$ — the complex Mellin (scale-frequency) variable; $x^{s-1}$ — the kernel, character of the dilation group; $(a,b)$ — the fundamental strip of analyticity. Reading it: $\tilde f$ is analytic throughout the strip, and the edges $a,b$ are the endpoint power laws of $f$ at $0$ and $\infty$ . Inversion recovers $f$ from an integral up a vertical line inside the strip — the Fourier inversion theorem wearing the logarithm's clothes, which is why the contour is a straight line rather than a closed loop.

f(x)=\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\tilde f(s)\,x^{-s}\,\mathrm{d}s,\qquad a<c<b \qquad (2)

where: $c$ — the real abscissa of the contour, any value strictly inside the strip; $x^{-s}$ — the inverse kernel. Reading it: moving $c$ across a pole — which requires leaving the strip — is exactly what generates an asymptotic expansion. Scaling and multiplication by a power are the transform's two elementary symmetries. A dilation of the argument multiplies $\tilde f$ by a power; multiplying $f$ by $x^\mu$ shifts the strip.

\mathcal{M}[f(\lambda x)](s)=\lambda^{-s}\,\tilde f(s),\qquad \mathcal{M}[x^{\mu}f(x)](s)=\tilde f(s+\mu),\qquad \mathcal{M}[f(x^p)](s)=\tfrac1p\,\tilde f(s/p) \qquad (3)

where: $\lambda>0$ — a scaling factor; $\mu$ — a real power multiplying $f$ ; $p>0$ — a power of the argument. Reading it: scaling is diagonal because $x^{s-1}$ is the scaling eigenfunction — the defining property of the transform. The operator $x\,\mathrm{d}/\mathrm{d}x$ — the generator of scaling — is diagonalized to multiplication by $-s$ , the analogue of $\mathrm{d}/\mathrm{d}t\mapsto i\omega$ for Fourier, turning Euler differential equations into algebra.

\mathcal{M}\!\left[x\frac{\mathrm{d}f}{\mathrm{d}x}\right](s)=-s\,\tilde f(s),\qquad \mathcal{M}\!\left[\frac{\mathrm{d}f}{\mathrm{d}x}\right](s)=-(s-1)\,\tilde f(s-1) \qquad (4)

where: $x\,\mathrm{d}/\mathrm{d}x$ — the Euler (scaling) operator. Reading it: an ordinary derivative shifts the argument of $\tilde f$ , but the scale-covariant derivative $x\,\mathrm{d}/\mathrm{d}x$ stays at the same $s$ and simply multiplies by $-s$ . Multiplicative convolution — with respect to the scale-invariant measure $\mathrm{d}y/y$ — factorizes under the transform, exactly as ordinary convolution does under Fourier.

\mathcal{M}\big[f\star g\big](s)=\tilde f(s)\,\tilde g(s),\qquad (f\star g)(x)=\int_0^{\infty} f\!\left(\frac{x}{y}\right)g(y)\,\frac{\mathrm{d}y}{y} \qquad (5)

where: $f\star g$ — Mellin (multiplicative) convolution with invariant measure $\mathrm{d}y/y$ . Reading it: any integral of the harmonic-sum form $\int_0^\infty h(xt)g(t)\,\mathrm{d}t$ is such a convolution, so its transform is a product — the key to extracting asymptotics by Parseval. The Parseval, or product, relation converts a real integral over the half-line into a contour integral of transforms reflected about $s=\tfrac12$ — the workhorse for evaluating definite integrals and Dirichlet sums in closed form.

\int_0^{\infty} f(x)\,g(x)\,\mathrm{d}x=\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\tilde f(s)\,\tilde g(1-s)\,\mathrm{d}s \qquad (6)

where: $\tilde g(1-s)$ — the transform of $g$ at the reflected point $1-s$ . Reading it: the real integral equals a sum over the poles of $\tilde f(s)\tilde g(1-s)$ , usually a product of Gamma functions. Finally, the most important explicit transform: the Gamma function is the Mellin transform of the decaying exponential, and summing $\sum e^{-nx}$ against it produces the Gamma–zeta product that governs blackbody physics.

\Gamma(s)=\int_0^{\infty} e^{-x}\,x^{s-1}\,\mathrm{d}x,\qquad \int_0^{\infty}\frac{x^{s-1}}{e^{x}-1}\,\mathrm{d}x=\Gamma(s)\,\zeta(s),\qquad \mathrm{Re}\,s>1 \qquad (7)

where: $\Gamma(s)$ — the Euler Gamma function, meromorphic with poles at $s=0,-1,-2,\dots$ ; $\zeta(s)$ — the Riemann zeta function. Reading it: $\Gamma(s)$ has strip $\mathrm{Re}\,s>0$ (upper edge at $+\infty$ , since $e^{-x}$ decays faster than any power), and the Bose–Einstein integral just sums $\Gamma(s)/n^s$ over $n$ .

Derivation AInversion as a Fourier transform along the logarithm

The goal is to prove that Mellin inversion is a vertical-line contour integral, not a closed loop, by showing it is literally the Fourier inversion theorem after $x=e^{-t}$ . We turn the half-line into the whole line, recognise the Mellin transform on $\mathrm{Re}\,s=c$ as an ordinary Fourier transform, invert, and undo the substitution to land on the Bromwich contour — each move exact and reversible inside the fundamental strip.

Step 1. Substitute $x=e^{-t}$ , mapping $x\in(0,\infty)$ to $t\in(-\infty,\infty)$ with $x^{s-1}\mathrm{d}x=e^{-st}\mathrm{d}t$ ; on the line $s=c+i\tau$ define $g(t)=f(e^{-t})e^{-ct}$ . Then $\tilde f(c+i\tau)=\int_{-\infty}^{\infty}g(t)\,e^{-i\tau t}\,\mathrm{d}t=\hat g(\tau)$ . Why: the logarithm converts the multiplicative measure $\mathrm{d}x/x$ into $\mathrm{d}t$ , so the scaling kernel becomes the Fourier kernel $e^{-i\tau t}$ dressed by the weight $e^{-ct}$ .

Step 2. Apply the ordinary Fourier inversion theorem in the variable $\tau$ : $g(t)=\tfrac{1}{2\pi}\int_{-\infty}^{\infty}\hat g(\tau)\,e^{i\tau t}\,\mathrm{d}\tau=\tfrac{1}{2\pi}\int_{-\infty}^{\infty}\tilde f(c+i\tau)\,e^{i\tau t}\,\mathrm{d}\tau$ . Why: because $\tilde f(c+i\tau)$ is exactly the Fourier transform of $g$ , standard Fourier inversion applies verbatim — the fundamental strip guarantees $g\in L^1\cap L^2$ for admissible $c$ .

Step 3. Undo the substitution with $s=c+i\tau$ , $\mathrm{d}s=i\,\mathrm{d}\tau$ , $e^{-t}=x$ , so $g(t)=f(x)\,x^{-c}$ and $e^{i\tau t}=x^{-(s-c)}$ . Multiplying by $x^{c}$ collapses the weight and gives $f(x)=\tfrac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\tilde f(s)\,x^{-s}\,\mathrm{d}s$ . Why: the weight $x^{c}$ that made $g$ integrable cancels the $x^{-c}$ inside $g$ , leaving the Bromwich integral; the freedom to place $c$ inside the strip is exactly what Fourier convergence demanded.

Go deeper · Derivation B · 1 minAsymptotics from the poles of the transform

Derivation BAsymptotics from the poles of the transform

The goal is the converse mapping theorem: the location and order of the poles of $\tilde f(s)$ to the left of the fundamental strip determine, term by term, the expansion of $f(x)$ as $x\to0^+$ . We start from inversion, slide the contour leftward across the poles, and read off each residue as one term of the series, with higher-order poles supplying logarithms — the technique that makes Mellin the natural tool for asymptotics of integrals, heat kernels, and lattice sums.

Step 1. Begin with inversion, $f(x)=\tfrac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\tilde f(s)\,x^{-s}\,\mathrm{d}s$ , and push the contour left to $\mathrm{Re}\,s=c'<c$ . The residue theorem gives $\tfrac{1}{2\pi i}\int_{c}-\tfrac{1}{2\pi i}\int_{c'}=\sum_{-p\in(c',c)}\mathrm{Res}_{s=-p}\big[\tilde f(s)\,x^{-s}\big]$ . Why: deforming the contour changes the integral only by the residues of the poles it crosses, so those poles become an explicit finite sum.

Step 2. Evaluate a simple pole at $s=-p$ with residue $r_{-p}$ : since $x^{-s}=x^{p}$ there, the contribution is $\mathrm{Res}_{s=-p}[\tilde f(s)x^{-s}]=r_{-p}\,x^{p}$ — a pure power. Why: a first-order pole factors as a residue times an analytic remainder, and $x^{-s}$ at the pole is just $x^{p}$ , so each simple pole is one clean power.

Step 3. Handle a pole of order $m$ by expanding the kernel, $x^{-s}=x^{p}\sum_{k\ge0}\tfrac{(-\log x)^{k}}{k!}(s+p)^{k}$ , and multiplying by the Laurent tail of $\tilde f$ ; the product yields terms $x^{p}(\log x)^{j}$ for $j\le m-1$ . Why: the expansion of $x^{-s}$ supplies powers of $\log x$ , and only the first $m$ survive against the order- $m$ singularity, so a pole of order $m$ makes a logarithm of degree $m-1$ .

Step 4. Collect residues and bound the remainder: $f(x)\sim\sum_{p}(\text{residue terms})$ as $x\to0^+$ , the leftover integral on $\mathrm{Re}\,s=c'$ being $O(x^{-c'})$ , uniformly smaller than every term kept. Sliding right instead reproduces the $x\to\infty$ expansion. Why: because $|x^{-s}|=x^{-c'}$ on that line the shifted integral is a genuine remainder, so each pole crossed refines the expansion by one order. Verification: for $\tilde f=\Gamma(s)$ , poles at $s=-n$ with residues $(-1)^n/n!$ give $e^{-x}=\sum_n(-1)^n x^n/n!$ .

Go deeper · Derivation C · 1 minThe Beta integral and the Mellin–Barnes representation

Derivation CThe Beta integral and the Mellin–Barnes representation

The goal is to compute the Mellin transform of $(1+x)^{-\alpha}$ in closed form as a ratio of Gamma functions, then invert it into a Mellin–Barnes contour integral whose two ways of closing produce the two convergent expansions of the same function. This example is the seed of the Mellin–Barnes representations that define hypergeometric and Meijer $G$ -functions and split propagator denominators in dimensional regularization. We evaluate by a rational substitution and recognise the Euler Beta function.

Step 1. Write the transform $\mathcal{M}[(1+x)^{-\alpha}](s)=\int_0^{\infty}x^{s-1}(1+x)^{-\alpha}\,\mathrm{d}x$ , convergent for $0<\mathrm{Re}\,s<\mathrm{Re}\,\alpha$ . Why: at $x\to0$ the integrand $\sim x^{s-1}$ needs $\mathrm{Re}\,s>0$ ; at $x\to\infty$ it $\sim x^{s-1-\alpha}$ needs $\mathrm{Re}\,s<\mathrm{Re}\,\alpha$ — the endpoint powers set the strip.

Step 2. Substitute $t=x/(1+x)$ , so $\mathrm{d}x=\mathrm{d}t/(1-t)^2$ , mapping $(0,\infty)$ to $(0,1)$ ; the integrand becomes $t^{s-1}(1-t)^{\alpha-s-1}$ . Why: the map sends the denominator $(1+x)$ to the finite endpoint $1-t$ , converting the half-line integral into a compact one on the unit interval — the home of the Beta function.

Step 3. Recognise the Euler Beta integral: $\int_0^1 t^{s-1}(1-t)^{\alpha-s-1}\mathrm{d}t=B(s,\alpha-s)=\dfrac{\Gamma(s)\,\Gamma(\alpha-s)}{\Gamma(\alpha)}$ . Why: rewriting $B$ through Gamma functions exposes the pole structure — $\Gamma(s)$ at $s=0,-1,\dots$ on the left, $\Gamma(\alpha-s)$ at $s=\alpha,\alpha+1,\dots$ on the right.

Step 4. Invert to the Mellin–Barnes integral $(1+x)^{-\alpha}=\tfrac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty}\dfrac{\Gamma(s)\Gamma(\alpha-s)}{\Gamma(\alpha)}\,x^{-s}\,\mathrm{d}s$ . Closing left over the poles of $\Gamma(s)$ gives the small- $x$ binomial series $\sum_n\binom{-\alpha}{n}x^n$ ; closing right over the poles of $\Gamma(\alpha-s)$ gives the large- $x$ series in $x^{-\alpha-n}$ . Why: the two closures are the two convergent expansions of one function, decided entirely by which family of Gamma poles the contour sweeps. Verification: at $\alpha=1$ , $\Gamma(s)\Gamma(1-s)=\pi/\sin\pi s$ , the reflection formula, with strip $0<\mathrm{Re}\,s<1$ .

Limiting cases

The Mellin transform is the hinge connecting three neighbouring transforms; each limit below is the same object read in different coordinates.

Case	Result	Meaning
Fourier (log substitution)	$x=e^{-t}\Rightarrow\tilde f(c+i\tau)=\int_{-\infty}^{\infty}f(e^{-t})e^{-ct}e^{-i\tau t}\mathrm{d}t$	on the line $\mathrm{Re}\,s=c$ the Mellin transform is the Fourier transform of $g(t)=f(e^{-t})e^{-ct}$
Two-sided Laplace	$\tilde f(s)=\int_{-\infty}^{\infty}f(e^{-t})\,e^{-st}\,\mathrm{d}t$	identically the bilateral Laplace transform of $f(e^{-t})$ ; the fundamental strip is its abscissa-of-convergence strip
Strip edges from power laws	$f\sim x^{-a}\,(x\to0),\ f\sim x^{-b}\,(x\to\infty)\Rightarrow a<\mathrm{Re}\,s<b$	endpoint decay sets the strip; meromorphic continuation past an edge generates the asymptotic series
Euler operator eigenvalue	$\mathcal{M}[(x\,\mathrm{d}/\mathrm{d}x)f](s)=-s\,\tilde f(s)$	equidimensional operators become algebra, the scaling analogue of $\mathrm{d}/\mathrm{d}t\to i\omega$
Rapid decay (e.g. $e^{-x}$ )	upper edge $b\to\infty$ , transform is entire-ratio $\Gamma$ -type	when $f$ decays faster than any power, the strip is a half-plane and the poles are all on one side

Common trapQuoting the transform without tracking its fundamental strip

The commonest and most damaging error is to write down $\tilde f(s)$ as a formula and forget that it is only defined by the integral inside a strip. The same closed form — say $\Gamma(s)\Gamma(\alpha-s)/\Gamma(\alpha)$ — represents two different functions of $x$ depending on which side of a pole the inversion contour $c$ sits, because sliding $c$ past a pole adds or removes a whole asymptotic branch. Invert an analytically continued $\tilde f$ without asking which poles you enclose and you will produce the small- $x$ series when you wanted the large- $x$ one, or a divergent answer where a finite one exists. Always name the strip, place $c$ inside it, and decide before closing which family of poles you intend to sum.

Worked examples

1 · The Stefan–Boltzmann integral from the Gamma–zeta product (fully worked)

Evaluate $I_n=\int_0^\infty \dfrac{x^{n}}{e^{x}-1}\,\mathrm{d}x$ for $n=1$ and $n=3$ , the integrals behind the heat capacity of a photon gas and the total blackbody power. Use the Mellin machinery, not a table.

Step 1. Recognise the integral as a Mellin transform at $s=n+1$ : $I_n=\mathcal{M}[(e^x-1)^{-1}](n+1)$ , valid for $\mathrm{Re}\,s>1$ (near $x=0$ the integrand $\sim x^{s-2}$ needs $\mathrm{Re}\,s>1$ ).

Step 2. Expand the Bose factor as a geometric series, $\dfrac{1}{e^x-1}=\sum_{k=1}^{\infty}e^{-kx}$ , and transform term by term. Each term gives $\int_0^\infty x^{s-1}e^{-kx}\mathrm{d}x=\Gamma(s)/k^{s}$ by the scaling rule, so summing over $k$ gives $\mathcal{M}[(e^x-1)^{-1}](s)=\Gamma(s)\sum_{k\ge1}k^{-s}=\Gamma(s)\,\zeta(s)$ — equation (7).

Step 3. Put $n=1$ , so $s=2$ : $I_1=\Gamma(2)\,\zeta(2)=1\cdot\dfrac{\pi^2}{6}=\mathbf{\pi^2/6\approx1.6449}$ — the integral behind the $T^2$ heat capacity of a one-dimensional phonon gas.

Step 4. Put $n=3$ , so $s=4$ : $I_3=\Gamma(4)\,\zeta(4)=6\cdot\dfrac{\pi^4}{90}=\mathbf{\pi^4/15\approx6.4939}$ . This is the Stefan–Boltzmann integral — the dimensionless number in $u=\dfrac{8\pi (k_BT)^4}{(hc)^3}\int_0^\infty\dfrac{x^3\,\mathrm{d}x}{e^x-1}$ that makes the blackbody energy density scale as $T^4$ . One Mellin identity, $\Gamma(s)\zeta(s)$ , delivers both $\pi^2/6$ and $\pi^4/15$ at $s=2$ and $s=4$ .

2 · A half-line integral by the fundamental strip and the reflection formula

Evaluate $J(s)=\int_0^\infty \dfrac{x^{s-1}}{1+x}\,\mathrm{d}x$ and read off the numbers at $s=\tfrac12$ and $s=\tfrac13$ .

Step 1. This is the Beta integral of Derivation 3 at $\alpha=1$ : $J(s)=\dfrac{\Gamma(s)\Gamma(1-s)}{\Gamma(1)}=\Gamma(s)\Gamma(1-s)$ , strip $0<\mathrm{Re}\,s<1$ . Outside it the integral diverges — at $s\le0$ from the $x=0$ end, at $s\ge1$ from the $x=\infty$ end.

Step 2. Apply the reflection formula $\Gamma(s)\Gamma(1-s)=\pi/\sin\pi s$ , so $J(s)=\pi/\sin\pi s$ throughout the strip, symmetric about $s=\tfrac12$ where it takes its minimum $\pi$ .

Step 3. At $s=\tfrac12$ : $J=\pi/\sin(\pi/2)=\mathbf{\pi\approx3.1416}$ . Check by $x=u^2$ : $\int_0^\infty\frac{u^{-1}}{1+u^2}2u\,\mathrm{d}u=2\cdot\frac{\pi}{2}=\pi$ , an independent confirmation.

Step 4. At $s=\tfrac13$ : $J=\pi/\sin(\pi/3)=\mathbf{2\pi/\sqrt3\approx3.6276}$ . The value grows as $s$ moves off centre and blows up at the edges $s\to0,1$ where the integral stops converging. The strip $0<\mathrm{Re}\,s<1$ is not a technicality — it is the exact range where $\pi/\sin\pi s$ is finite and correct.

Physics of the portrait

The transform carries the name of the Finnish analyst who assembled its scattered pieces — the Gamma integral, Riemann's theta trick, Barnes's contour integrals — into one coherent theory.

Origins · 1 minHjalmar Mellin and how this was discovered

Hjalmar Mellin

1854–1933 · Finnish analyst who systematized the scaling transform

Robert Hjalmar Mellin studied under Karl Weierstrass in Berlin and Gösta Mittag-Leffler in Stockholm before settling in Helsinki. The integral that bears his name was, in pieces, older than he was: Euler's Gamma integral dates to the 1720s, and in 1859 Bernhard Riemann had transformed the Jacobi theta function to derive the functional equation of the zeta function — the deepest single application of the transform, though Riemann never named it as such. Mellin's contribution, around the turn of the century, was to isolate the transform as an object in its own right: to give the inversion formula as a vertical-line contour integral, to identify the fundamental strip of convergence, and to build a systematic calculus of its properties. He used it to analytically continue the hypergeometric functions and to study the asymptotics of linear differential equations, in parallel with the contour methods of his contemporary E. W. Barnes. The transform found its rigorous modern footing in Titchmarsh's 1937 treatise on Fourier integrals, and its role as the master tool of asymptotic analysis was codified much later by Flajolet and collaborators. What Mellin gave physics is the habit of thinking of scale the way Fourier taught us to think of frequency — and with it the Mellin–Barnes integrals that today evaluate multi-loop Feynman diagrams.

History

1859

— Riemann transforms the theta function to derive the zeta functional equation

1896

— Mellin publishes his systematic theory of the transform and its inversion

1908

— Barnes develops contour integrals for the hypergeometric function

1937

— Titchmarsh gives rigorous inversion conditions in his Fourier-integrals treatise

1995

— Flajolet, Gourdon and Dumas codify the Mellin asymptotics dictionary

Going further

Texts. Stone & Goldbart, Mathematics for Physics, integral transforms · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed. · Hassani, Mathematical Physics, contour integrals and special functions · Byron & Fuller, Mathematics of Classical and Quantum Physics · Bleistein & Handelsman, Asymptotic Expansions of Integrals · Flajolet, Gourdon & Dumas, "Mellin transforms and asymptotics", Theoretical Computer Science 144 (1995). Key anchor results to keep in mind: $\int_0^\infty e^{-x}x^{s-1}\mathrm{d}x=\Gamma(s)$ with $\mathrm{Res}_{s=-n}\Gamma=(-1)^n/n!$ ; $\int_0^\infty x^{s-1}/(e^x-1)\mathrm{d}x=\Gamma(s)\zeta(s)$ , giving $\int_0^\infty x^3/(e^x-1)\mathrm{d}x=\pi^4/15=6.4939$ and $\int_0^\infty x/(e^x-1)\mathrm{d}x=\pi^2/6=1.6449$ ; the reflection formula $\Gamma(s)\Gamma(1-s)=\pi/\sin\pi s$ , yielding $J(\tfrac12)=\pi$ and $J(\tfrac13)=2\pi/\sqrt3=3.6276$ ; and the Euler-operator rule follows by integration by parts inside the strip.

Prerequisites. The Fourier Transform · The Laplace Transform · Laplace Transform Inversion · The Convolution Theorem. Leads to. Parseval and Plancherel · Differential Equations by Transform Methods · Theory of Distributions.

Ask the tutor: ask me to slide the inversion contour on any $\tilde f(s)$ you like and watch the small- $x$ and large- $x$ expansions fall out of its poles.

How is the Mellin transform related to the Fourier and Laplace transforms?

They are the same machine in different coordinates. Substitute $x=e^{-t}$ in the Mellin integral: the multiplicative half-line becomes the whole real line, the measure $\mathrm{d}x/x$ becomes $\mathrm{d}t$ , and the kernel $x^{s-1}\mathrm{d}x$ becomes $e^{-st}\mathrm{d}t$ . So $\tilde f(s)$ is exactly the two-sided (bilateral) Laplace transform of $f(e^{-t})$ , and restricting to the vertical line $s=c+i\tau$ makes it the Fourier transform of $f(e^{-t})e^{-ct}$ . Everything Mellin — the inversion contour, the convolution theorem, Parseval — is the corresponding Fourier or Laplace statement pulled back through the logarithm. The practical upshot is that you never need to re-prove convergence or inversion; you translate.

What exactly is the fundamental strip and why can't I ignore it?

The defining integral $\int_0^\infty f(x)x^{s-1}\mathrm{d}x$ converges only for $s$ in a vertical strip $a<\mathrm{Re}\,s<b$ , where $a$ and $b$ are set by how $f$ decays at the two ends of the half-line. Inside that strip $\tilde f(s)$ is analytic; outside it the integral diverges, though the formula for $\tilde f$ usually continues meromorphically. You cannot ignore the strip because the inversion contour must lie inside it, and because the same closed-form $\tilde f(s)$ can represent different functions of $x$ depending on which poles the contour encloses. Naming the strip is how you keep the small- $x$ expansion from being confused with the large- $x$ one.

How do poles of the transform give asymptotics of the function?

Start from the inversion integral and slide the contour off to one side. Every pole of $\tilde f(s)$ you cross contributes its residue, and because the kernel is $x^{-s}$ , a pole at $s=-p$ contributes a term proportional to $x^{p}$ . Sliding left (to more negative $\mathrm{Re}\,s$ ) collects the poles that govern small $x$ , so you build the $x\to0$ expansion; sliding right builds the $x\to\infty$ expansion. A simple pole gives a pure power; a pole of order $m$ gives $x^{p}$ times a polynomial of degree $m-1$ in $\log x$ . The leftover integral on the shifted line is the rigorous remainder. This turns a hard analytic question into the algebraic one of locating poles and computing residues.

Why does the operator x d/dx become multiplication by −s?

Integrate by parts: $\int_0^\infty x f'(x)\,x^{s-1}\mathrm{d}x=[x^{s}f(x)]_0^\infty-s\int_0^\infty f(x)x^{s-1}\mathrm{d}x$ . Inside the fundamental strip the boundary term $x^{s}f(x)$ vanishes at both endpoints — that is exactly what the endpoint power laws defining the strip guarantee — so what remains is $-s\,\tilde f(s)$ . Conceptually, $x\,\mathrm{d}/\mathrm{d}x$ is the generator of scaling, and $x^{s-1}$ is its eigenfunction, so the transform diagonalizes it just as $\mathrm{d}/\mathrm{d}t$ is diagonalized to $i\omega$ by Fourier modes. This is why an Euler (equidimensional) differential equation turns into a polynomial in $-s$ , solvable by algebra and then inverted.

What is a Mellin–Barnes integral and where does it matter?

A Mellin–Barnes integral is an inversion integral whose transform is a ratio of Gamma functions — for instance $(1+x)^{-\alpha}=\tfrac{1}{2\pi i}\int\tfrac{\Gamma(s)\Gamma(\alpha-s)}{\Gamma(\alpha)}x^{-s}\mathrm{d}s$ . Closing the contour to one side sums one family of Gamma poles and gives the small-argument series; closing to the other gives the large-argument series. These integrals are the standard representations of the hypergeometric and Meijer $G$ -functions, and they are central to high-energy physics: every propagator in a Feynman diagram can be written this way, and after loop integration the divergences of the diagram appear as poles of the surviving Gamma functions in the regularization parameter $\varepsilon=(4-d)/2$ . Whole software packages exist to resolve such integrals.

Does the Mellin transform actually compute definite integrals I care about?

Constantly. The Bose–Einstein integrals of statistical mechanics are Mellin transforms: $\int_0^\infty x^{s-1}/(e^x-1)\mathrm{d}x=\Gamma(s)\zeta(s)$ delivers the Stefan–Boltzmann value $\pi^4/15$ at $s=4$ and the $\pi^2/6$ of low-temperature heat capacities at $s=2$ . The Parseval relation converts many other real integrals into contour integrals of Gamma-function products evaluated by residues, and the converse-mapping theorem gives the asymptotics of integrals with no closed form at all. If an integral has a scaling structure — a product $h(xt)g(t)$ , a power-law tail, an exponential — the Mellin transform is usually the most systematic way to attack it.