Mellin Transforms and Scaling Asymptotics

The Mellin transform is Fourier analysis on the logarithm of a positive variable. Because a dilation becomes multiplication by a character $\lambda^{-s}$, the transform separates scaling behavior in much the same way that the Fourier transform separates oscillations. Its singularities then have a direct asymptotic meaning: a pole at $s=s_0$ produces a power $x^{-s_0}$, and a pole of order $m$ produces that power times a polynomial in $\log x$ of degree $m-1$.

That slogan becomes a method only after four qualifications. One must state the fundamental strip in which the defining integral converges, specify the branch of $x^{-s}$, justify Mellin inversion, and control the new vertical contour after it is moved. Subject to those conditions, contour displacement turns meromorphic information into a power-and-logarithm expansion with an explicit remainder integral. A finite Euclidean one-loop integral below will check the pole order, contour direction, normalization, and logarithmic coefficient independently.

Required background. Lebesgue Integration and Convergence Theorems supplies the convergence and interchange results needed for the defining integral, inversion, and parameter-dependent examples; Holomorphic Functions and Cauchy Theory supplies holomorphy, Cauchy’s theorem, and the contour-deformation logic behind the rectangular shifts below. The residue formulas needed here are derived where they are used.

Mellin variables, branches, and scaling scope

In particular, WKB and Eikonal Methods and Turning-Point Matching studies local differential-equation approximations and matching, whereas this page extracts scaling information from a transform in the complex $s$-plane.

Unless stated otherwise, $x>0$, $s=\sigma+i\tau$, and $\log x$ is the real logarithm. For complex $x$, the expression

$$x^{-s}=\exp[-s\operatorname{Log}x]$$

is meaningful only after choosing a branch of $\operatorname{Log}x$ and a sector that avoids its cut. Any complex-$x$ contour shift must remain uniform in the claimed closed subsector.

Definition and the fundamental strip

For a locally integrable measurable function $f:(0,\infty)\to\mathbb C$, use the convention

$$F(s)=\mathcal M[f](s) = \int_0^\infty x^{s-1}f(x)\,\mathrm dx.$$

The set of real parts $\sigma$ for which this integral converges absolutely is normally an open vertical strip. If, for example,

$$f(x)=O(x^u)\quad(x\to0^+), \qquad f(x)=O(x^v)\quad(x\to\infty),$$

with $u>v$, then absolute convergence holds at least in

$$-u<\Re s<-v.$$

The two inequalities come from different endpoints: $x^{\sigma+u-1}$ must be integrable at zero, while $x^{\sigma+v-1}$ must be integrable at infinity. The fundamental strip is the maximal open strip obtained from the actual endpoint behavior; the bound above need not determine it sharply.

This distinction matters. The defining integral, its meromorphic continuation, and a regularized value are three different objects. A formula continued beyond the fundamental strip may remain useful, but it is no longer represented there by the original absolutely convergent integral.

The conventions, inversion formula, convolution law, and contour method are collected in NIST DLMF 2026, §2.5(i)–(ii) and developed with explicit mapping theorems in Flajolet, Gourdon, and Dumas 1995, , Part I §§1–2, printed pp. 9–21, PDF. The formulas here incorporate the corrections to the reflection identity on the original pp. 11–12 and to the logarithmic power in Theorem 4 on the original p. 20, as listed in the Flajolet, Gourdon, and Dumas 1995, official corrigenda, unnumbered p. 1, PDF.

Fourier analysis in logarithmic coordinates

Set $x=e^t$, so that $\mathrm dx/x=\mathrm dt$. On a vertical line $s=c+i\tau$ inside the fundamental strip,

$$F(c+i\tau) = \int_{-\infty}^{\infty} e^{ct}f(e^t)e^{i\tau t}\,\mathrm dt.$$

Thus $F(c+i\tau)$ is the Fourier transform, with the volume’s positive-sign forward convention, of

$$g_c(t)=e^{ct}f(e^t).$$

Mellin inversion is consequently Fourier inversion in $t$. Under standard Fourier inversion hypotheses—for example, when $g_c$ and its Fourier transform are integrable—the inverse at a continuity point is

$$f(x) = \frac{1}{2\pi i} \int_{c-i\infty}^{c+i\infty} x^{-s}F(s)\,\mathrm ds,$$

where the vertical contour is oriented upward. Weaker $L^2$, almost-everywhere, or distributional inversions require their corresponding hypotheses; the pointwise formula should not be silently transferred between these settings.

Two elementary identities explain why the transform is adapted to scaling. For $\lambda>0$,

$$\mathcal M[f(\lambda\,\cdot)](s) = \lambda^{-s}F(s),$$

and, for the Mellin convolution

$$(f*_\mathrm M g)(x) = \int_0^\infty f\!\left(\frac{x}{t}\right)g(t)\,\frac{\mathrm dt}{t},$$

one has

$$\mathcal M[f*_\mathrm M g](s)=F(s)G(s)$$

whenever the interchange of integrals is justified. Dilation is therefore diagonalized, while scale convolution becomes multiplication.

Endpoint terms become Mellin poles

The basic correspondence can be read directly from elementary endpoint integrals. A term near zero of the form

$$f(x)\sim a\,x^\alpha(\log x)^k, \qquad k\in\mathbb Z_{\ge0},$$

contributes

$$\int_0^1 x^{s-1}a\,x^\alpha(\log x)^k\,\mathrm dx = \frac{a(-1)^k k!}{(s+\alpha)^{k+1}}.$$

It therefore gives a pole at $s=-\alpha$ of order $k+1$. By contrast, a term at infinity written as

$$f(x)\sim b\,x^{-\beta}(\log x)^k$$

contributes

$$\int_1^\infty x^{s-1}b\,x^{-\beta}(\log x)^k\,\mathrm dx = -\frac{b(-1)^k k!}{(s-\beta)^{k+1}}.$$

The minus sign is not cosmetic: it combines with the direction of the large-$x$ contour shift to recover the positive coefficient $b$.

These calculations establish the direct mapping from a controlled endpoint expansion to the principal parts of a meromorphic continuation. The converse is stronger. Inferring an endpoint expansion from poles requires an inversion formula, meromorphic continuation into the displaced strip, enough decay on vertical lines, and a bound on the remainder contour. A pole diagram alone is not an asymptotic theorem.

For later use, suppose the local principal part at $s=s_0$ is

$$F(s) = \frac{A}{(s-s_0)^2} + \frac{B}{s-s_0} +O(1).$$

Since

$$x^{-s} = x^{-s_0} \left[ 1-(s-s_0)\log x+O((s-s_0)^2) \right],$$

the residue of the inverse-Mellin integrand is

$$\operatorname*{Res}_{s=s_0} \left[x^{-s}F(s)\right] = x^{-s_0}(B-A\log x).$$

More generally, a pole of order $m$ gives $x^{-s_0}$ times a polynomial in $\log x$ of degree $m-1$. Higher-order poles—often formed when simple pole families coincide—are therefore the transform-space source of logarithms.

Which way should the contour move?

Let the inversion line be $\Re s=c$. Suppose $F$ continues meromorphically to a wider vertical region, and choose horizontal segments on which the rectangular contour contributions vanish. Also assume that the residues crossed are finite in number or converge in the same limiting sequence of rectangles; otherwise the displayed residue sum has not been defined.

For $x\to0^+$, move the line to $\Re s=d<c$. With upward orientation on both vertical lines,

$$f(x) = \sum_{d<\Re s_j<c} \operatorname*{Res}_{s=s_j} \left[x^{-s}F(s)\right] + \frac{1}{2\pi i} \int_{d-i\infty}^{d+i\infty} x^{-s}F(s)\,\mathrm ds.$$

Thus the left-hand poles generate the small-$x$ expansion and enter with a plus sign.

For $x\to\infty$, move the line to $\Re s=e>c$. Then

$$f(x) = -\sum_{c<\Re s_j<e} \operatorname*{Res}_{s=s_j} \left[x^{-s}F(s)\right] + \frac{1}{2\pi i} \int_{e-i\infty}^{e+i\infty} x^{-s}F(s)\,\mathrm ds.$$

The right-hand poles generate the large-$x$ expansion and enter with a minus sign. This is the easiest sign to lose in a Mellin calculation.

If, on the new line,

$$\int_{-\infty}^{\infty} |F(d+i\tau)|\,\mathrm d\tau<\infty,$$

then the first remainder is $O(x^{-d})$; the analogous condition at $e$ gives an $O(x^{-e})$ large-$x$ remainder. Other vertical-growth assumptions lead to other remainder meanings, and uniformity in auxiliary parameters must be proved rather than inferred. Moving a contour without controlling its horizontal edges and final vertical line is only a formal residue calculation.

A sign check: the geometric kernel

For

$$f(x)=\frac{1}{1+x},$$

Euler’s beta integral gives

$$F(s) = \int_0^\infty\frac{x^{s-1}}{1+x}\,\mathrm dx = \frac{\pi}{\sin\pi s}, \qquad 0<\Re s<1.$$

For small $x$, moving left crosses $s=0,-1,-2,\ldots$ and gives

$$\frac{1}{1+x} \sim 1-x+x^2-x^3+\cdots.$$

For large $x$, moving right crosses $s=1,2,3,\ldots$. The residue of $\pi/\sin\pi s$ at $s=n$ is $(-1)^n$, and the overall minus sign from the rightward shift gives

$$\frac{1}{1+x} \sim x^{-1}-x^{-2}+x^{-3}-\cdots.$$

Both are familiar geometric expansions. They independently verify the contour-direction rule.

Scaled sums and colliding poles

Consider a scaled sum

$$K(t)=\sum_n a_n h(t\lambda_n), \qquad t>0,$$

with $\lambda_n>0$. If the sum and Mellin integral can be interchanged in a common strip, the dilation identity gives

$$\mathcal M[K](s) = H(s)Z_a(s), \qquad Z_a(s)=\sum_n a_n\lambda_n^{-s}.$$

The poles of the kernel transform $H$ and of the Dirichlet series $Z_a$ now jointly determine the scaling expansion. If a simple pole of $H$ coincides with a simple pole of $Z_a$, their product has a double pole and inversion produces a logarithm. This factorization is powerful precisely because it separates the analytic behavior of one kernel from the distribution of the scales $\lambda_n$.

The formula is conditional. It can fail when there is no common convergence strip, when the sum-integral interchange is unjustified, or when the continued product grows too quickly on vertical lines. The systematic use of this observation for heat traces and spectral zeta functions belongs to Heat Kernels, Zeta Functions, and Spectral Determinants.

Controlled QFT example: a finite Euclidean bubble

Consider the equal-mass scalar integral in two Euclidean dimensions,

$$B(Q^2,m^2) = \int_{\mathbb R^2}\frac{\mathrm d^2k}{(2\pi)^2} \frac{1} {(k^2+m^2)((k+Q)^2+m^2)},$$

with $Q^2>0$ and $m^2>0$. It is ultraviolet and infrared finite under these assumptions, so no regulator or subtraction is being hidden. Feynman parameterization followed by a translation of the loop momentum gives

$$B(Q^2,m^2) = \int_0^1\mathrm dx \int_{\mathbb R^2}\frac{\mathrm d^2\ell}{(2\pi)^2} \frac{1} {[\ell^2+m^2+x(1-x)Q^2]^2}.$$

The radial integral is elementary:

$$\int_{\mathbb R^2}\frac{\mathrm d^2\ell}{(2\pi)^2} \frac{1}{(\ell^2+\Delta)^2} = \frac{1}{4\pi\Delta}.$$

With the dimensionless ratio $r=Q^2/m^2$,

$$B(Q^2,m^2) = \frac{1}{4\pi m^2}J(r), \qquad J(r)=\int_0^1\frac{\mathrm dx}{1+r x(1-x)}.$$

The inverse Mellin formula for the geometric kernel is

$$\frac{1}{1+z} = \frac{1}{2\pi i} \int_{c-i\infty}^{c+i\infty} \frac{\pi}{\sin\pi s}z^{-s}\,\mathrm ds, \qquad 0<c<1.$$

Insert $z=r x(1-x)$ and perform the parameter integral. Absolute convergence in $0<c<1$ justifies the interchange and yields the Mellin–Barnes representation

$$J(r) = \frac{1}{2\pi i} \int_{c-i\infty}^{c+i\infty} \frac{\pi}{\sin\pi s} \frac{\Gamma(1-s)^2}{\Gamma(2-2s)} r^{-s}\,\mathrm ds.$$

On closed substrips avoiding poles, Stirling’s formula gives at most polynomial vertical growth for the gamma quotient, while $\pi/\sin\pi s$ decays exponentially. The inversion integral and the rectangular contour shifts used below are therefore absolutely controlled.

Here a Mellin–Barnes integral is simply inverse Mellin applied to a sum in a denominator. The basic identity and an introductory parametric Feynman-integral application are given in Dubovyk, Gluza, and Somogyi 2022, Chapter 1 §§1.3–1.4, printed pp. 12–20.

Small momentum ratio

For $r\to0^+$, move the contour left. At $s=-n$, $n\ge0$, the residue is

$$(-1)^n \frac{(n!)^2}{(2n+1)!} r^n.$$

Consequently, for each fixed $N\in\mathbb Z_{\ge0}$,

$$J(r) = \sum_{n=0}^N (-1)^n\frac{(n!)^2}{(2n+1)!}r^n +O(r^{N+1}), \qquad r\to0^+.$$

In particular,

$$J(r)=1-\frac r6+\frac{r^2}{30}+O(r^3).$$

This is also obtained by expanding the original integrand geometrically and using

$$\int_0^1[x(1-x)]^n\,\mathrm dx = \frac{(n!)^2}{(2n+1)!}.$$

The parameter integral is therefore an independent coefficient check. In fact, the full power series converges for $|r|<4$, because $|r x(1-x)|<1$ uniformly on $0\le x\le1$ in that disk.

Large momentum ratio and a double pole

For $r\to\infty$, move the contour right. The first crossed singularity is at $s=1$. Write $s=1+t$. Near $t=0$,

$$\frac{\pi}{\sin\pi s} = -\frac1t+O(t), \qquad \frac{\Gamma(1-s)^2}{\Gamma(2-2s)} = -\frac2t+O(t).$$

Their product has a double pole with no simple-pole term:

$$\frac{\pi}{\sin\pi s} \frac{\Gamma(1-s)^2}{\Gamma(2-2s)} = \frac{2}{t^2}+O(1).$$

Since

$$r^{-s} = r^{-1}(1-t\log r+O(t^2)),$$

the residue of the full integrand is $-2r^{-1}\log r$. A rightward contour shift subtracts this residue. The next right-hand singularity is a double pole at $s=2$, so the leading result with its remainder is

$$J(r) = \frac{2\log r}{r} +O\!\left(\frac{\log r}{r^2}\right) \qquad(r\to\infty).$$

The logarithm is not guessed from dimensional analysis; it is forced by the coincident simple poles that form the double pole at $s=1$.

For an independent check, the parameter integral can be evaluated exactly:

$$J(r) = \frac{4}{\sqrt{r(r+4)}} \operatorname{artanh} \sqrt{\frac{r}{r+4}}.$$

Its small- and large-$r$ expansions reproduce both residue calculations. The example establishes only the asymptotics of this finite Euclidean integral. The Mellin poles here are singularities in an auxiliary transform variable; they are neither particle poles nor dimensional-regularization poles. When a large logarithm is tied to a running coupling and repeated logarithms must be resummed, the relevant continuation is Large Logarithms and RG Improvement.

A reproducible Mellin workflow

For a positive scaling variable $x$, a reliable calculation proceeds as follows.

1. Nondimensionalize. Identify the limit $x\to0$, $x\to\infty$, or a sectorial complex limit. Do not Mellin-transform a dimensional variable without first choosing a reference scale.
2. Find the fundamental strip. Analyze zero and infinity separately and record which endpoint supplies each boundary.
3. Compute inside the strip. Establish the transform there before using analytic continuation. Justify any parameter, sum, or integral interchange in a common domain.
4. Continue meromorphically. Locate poles, their orders, and their principal parts. Record branch cuts or non-polar singularities rather than forcing them into a pole table.
5. Choose the contour direction. Move left for $x\to0$ and right for $x\to\infty$ with the present $x^{-s}$ inversion convention.
6. Bound what remains. Show that the horizontal pieces vanish and state a bound, norm, or distributional meaning for the new vertical integral.
7. Check independently. Compare with an exact expression, direct endpoint expansion, differential equation, numerical quadrature, or known scaling law.

Residue extraction is often cheap once the transform is known. The difficult parts are finding a useful continuation, controlling vertical growth, and maintaining uniformity when parameters move singularities. Stop the method and reformulate when no common strip exists, when pole families pinch the contour, or when the physical integral needs a regulator whose removal has not been specified.

For numerical inversion, truncating the vertical line at $|\tau|\le T$ adds a second approximation. Increase $T$, vary the line within the same analytic strip, and compare against direct quadrature. Large cancellations between residues and the remainder contour are a warning that the chosen representation is poorly conditioned.

Limitations and common pitfalls

Reading poles without a converse theorem. Meromorphic continuation identifies candidate powers and logarithms. It does not by itself show that the inverse contour exists, that its displacement is legal, or that the remaining integral is smaller than the retained terms.

Using the wrong contour direction. With $x^{-s}$ in the inverse transform, small $x$ uses poles to the left and large $x$ uses poles to the right. The rightward shift also carries an overall minus sign.

Forgetting the fundamental strip. An analytically continued formula can be finite where the defining Mellin integral diverges. State both objects and do not use continuation as an unannounced convergence prescription.

Ignoring moving poles. If an auxiliary parameter makes poles collide or pinch the contour, a termwise expansion may cease to be uniform and new logarithms can appear. The parameter range and order of limits are part of the result.

Treating every singularity as a pole. Branch points and cuts can produce non-power behavior, and distributions require a different inversion framework. A finite list of residues cannot capture contributions that are smaller than every power, such as $e^{-1/x}$ as $x\to0^+$.

Confusing transform poles with physical poles. The Mellin variable is an auxiliary scaling variable. Its poles organize asymptotics; their physical interpretation depends on the original observable and cannot be assigned by the transform alone.

Check your understanding

1. Retrieval: conventions and contour direction

State the transform, inversion formula, dilation law, and contour direction for each endpoint limit under this page’s convention.

Solution

The formulas are

$$F(s)=\int_0^\infty x^{s-1}f(x)\,\mathrm dx, \qquad f(x)=\frac{1}{2\pi i}\int_{c-i\infty}^{c+i\infty} x^{-s}F(s)\,\mathrm ds,$$

and

$$\mathcal M[f(\lambda\,\cdot)](s)=\lambda^{-s}F(s).$$

For $x\to0^+$, shift left and add the crossed residues. For $x\to\infty$, shift right and subtract the crossed residues. In both cases the final vertical contour is the remainder and must be bounded.

2. Hypothesis check: a function with no classical strip

Show that $f(x)=1$ has no fundamental strip for the ordinary Mellin integral on $(0,\infty)$. Explain why assigning a regularized value does not repair Mellin inversion automatically.

Solution

At zero,

$$\int_0^1x^{s-1}\,\mathrm dx$$

converges only for $\Re s>0$. At infinity,

$$\int_1^\infty x^{s-1}\,\mathrm dx$$

converges only for $\Re s<0$. The two half-planes do not overlap, so there is no vertical line on which the original transform converges at both endpoints. A regularization may define another analytic object, but ordinary inversion cannot be invoked until a compatible distributional or regularized theorem is stated and its hypotheses are checked.

3. Calculation: invert a double pole

Suppose that near $s=s_0$,

$$F(s)=\frac{A}{(s-s_0)^2}+\frac{B}{s-s_0}+O(1).$$

Find the residue of $x^{-s}F(s)$ and state how it enters a small-$x$ expansion when the pole is crossed to the left.

Solution

Expanding

$$x^{-s}=x^{-s_0} \left[1-(s-s_0)\log x+O((s-s_0)^2)\right]$$

shows that

$$\operatorname*{Res}_{s=s_0}[x^{-s}F(s)] = x^{-s_0}(B-A\log x).$$

A leftward shift adds this residue. A rightward shift would subtract it.

4. QFT transfer: the bubble logarithm

In the Mellin–Barnes representation for $J(r)$, use the local behavior

$$\frac{\pi}{\sin\pi s} \frac{\Gamma(1-s)^2}{\Gamma(2-2s)} = \frac{2}{(s-1)^2}+O(1)$$

to find the leading term as $r\to\infty$. Why is this not, by itself, an RG improvement result?

Solution

Writing $t=s-1$ gives

$$r^{-s}=r^{-1}(1-t\log r+O(t^2)).$$

The residue at $s=1$ is therefore $-2r^{-1}\log r$. The contour moves right for large $r$, so this residue is subtracted and

$$J(r)\sim\frac{2\log r}{r}.$$

This calculation extracts a kinematic asymptotic logarithm from one finite integral. RG improvement additionally identifies scale dependence governed by renormalization, derives the running parameters, and resums the relevant logarithmic tower. None of those steps follows from the Mellin pole alone.

Synthesis and continuations

The Mellin transform converts dilation into multiplication. Once its defining strip and inversion line are fixed, meromorphic continuation exposes the scaling data: pole position fixes a power, pole order fixes the degree of its logarithm, and contour direction distinguishes the two endpoint limits. The new vertical contour supplies the remainder, so an honest result always pairs its residue sum with a decay or remainder statement.

The next page, Special Functions from Equations and Boundary Data, develops a different organizing principle: differential equations, singular points, and boundary conditions select special functions and their branches. Later, Heat Kernels, Zeta Functions, and Spectral Determinants applies Mellin factorization to spectral data. In QFT, Large Logarithms and RG Improvement explains when logarithmic scaling is controlled and resumable by the renormalization group.

References

• Ievgen Dubovyk, Janusz Gluza, and Gábor Somogyi, Mellin–Barnes Integrals: A Primer on Particle Physics Applications, Lecture Notes in Physics 1008 (Springer, 2022), Chapter 1 §§1.3–1.4, printed pp. 12–20. Mellin–Barnes decompositions and their use in parametric Feynman integrals.

• Philippe Flajolet, Xavier Gourdon, and Philippe Dumas, “Mellin Transforms and Asymptotics: Harmonic Sums,” Theoretical Computer Science 144 (1995), 3–58, Part I §§1–2, printed pp. 9–21, PDF. Fundamental strips, inversion, and direct and converse mapping theorems. Read with the Flajolet, Gourdon, and Dumas official corrigenda, unnumbered p. 1, PDF, which corrects the original pp. 11–12 and p. 20.

• NIST Digital Library of Mathematical Functions, version 1.2.7, released June 15, 2026, National Institute of Standards and Technology, §2.5, “Mellin Transform Methods,” especially §2.5(i)–(ii). Definition, inversion, convolution, contour translation, and residue asymptotics.