Complex and asymptotic methods repair

Contour deformation, analytic continuation, and saddle expansion are powerful because they replace a difficult representation with a more useful one. They are also easy to overstate. This focused review teaches you to name the domain, singularities, contour, limiting parameter, and error statement that make each move valid.

Two distinctions guide the page:

$$\begin{aligned} \text{analytic continuation} &\ne \text{continued convergence of the original integral},\\ \text{asymptotic expansion} &\ne \text{convergent equality at finite parameter}. \end{aligned}$$

Required background. You should be able to integrate elementary complex functions, compute a simple residue, find stationary points, and manipulate a quadratic form. Review linear and tensor methods if the Hessian or change-of-variables step is unstable.

Contours may move only through analytic regions

If $f$ is holomorphic in the region swept between two oriented contours with the same endpoints, and the connecting pieces contribute nothing, Cauchy’s theorem makes the integrals equal. The useful checklist is concrete:

• mark every pole, branch point, and cut;
• record the contour orientation and endpoints;
• identify the region swept during the deformation;
• include residues of any singularities crossed; and
• justify arcs at infinity or endpoint limits.

For a closed counterclockwise contour $C$,

$$\oint_C f(z)\,\mathrm dz =2\pi i\sum_{z_k\ \mathrm{inside}\ C} \operatorname{Res}_{z=z_k}f(z).$$

Changing the orientation changes the sign. Moving a pole from outside to inside changes the answer by its residue contribution. When poles approach the integration contour from opposite sides, the contour can become pinched; then the naive deformation fails. These are topological facts about the singularities and contour, not algebraic properties of a substitution. See Ahlfors 1979, ch. 4 for Cauchy theory and the residue theorem.

Continuation extends a function, not every representation

Consider

$$F(a)=\int_0^\infty e^{-ax}\,\mathrm dx.$$

The integral converges when $\operatorname{Re}a>0$ and equals $1/a$ there. The function $1/a$ is analytic on $\mathbb C\setminus\{0\}$, so it uniquely continues $F$ to that larger connected domain. But the original real-axis integral still diverges when $\operatorname{Re}a<0$. Thus these are three separate pieces of information:

Object	Domain
Original integral	$\operatorname{Re}a>0$
Continued function $1/a$	$a\ne0$
Singularity	$a=0$

Multivalued functions require more data. For $\log z$, choose a base point, branch, cut, and continuation path. A circuit around $z=0$ changes the value by $2\pi i$; no globally single-valued logarithm exists on $\mathbb C\setminus\{0\}$. Analytic continuation is unique along a fixed path when it exists, but different homotopy classes can land on different sheets.

Saddles produce approximations with a declared limit

Let $\lambda\to+\infty$ and suppose $f$ has a unique interior minimum at $x_0$, with $f''(x_0)>0$, while the amplitude $g$ is smooth near $x_0$. Laplace’s method gives

$$\int e^{-\lambda f(x)}g(x)\,\mathrm dx \sim e^{-\lambda f(x_0)}g(x_0) \sqrt{\frac{2\pi}{\lambda f''(x_0)}}$$

under hypotheses ensuring that the remainder of the contour is suppressed. The symbol $\sim$ means that the ratio of the two sides tends to one in the declared limit. It does not claim equality for finite $\lambda$.

For the concrete integral

$$I(\lambda,g)=\int_{-\infty}^{\infty} \exp\left[-\lambda\left(\frac{x^2}{2} +\frac{gx^4}{4}\right)\right]\mathrm dx, \qquad g\ge0,$$

set $x=y/\sqrt\lambda$. Expanding near the only minimum gives

$$\begin{aligned} I(\lambda,g) &=\frac1{\sqrt\lambda}\int_{-\infty}^{\infty} e^{-y^2/2} \exp\left(-\frac{gy^4}{4\lambda}\right)\mathrm dy\\ &\sim\sqrt{\frac{2\pi}{\lambda}} \left(1-\frac{3g}{4\lambda} +\mathcal O(\lambda^{-2})\right). \end{aligned}$$

Here $\langle y^4\rangle$ under the standard Gaussian is $3$, which fixes the first correction. The result assumes fixed nonnegative $g$ and real positive $\lambda$. Complexifying either parameter changes the convergence sectors and accessible steepest-descent contours. A systematic account of remainders, multiple saddles, and Stokes transitions appears in Olver 1997, chs. 3–4.

A QFT bridge: the pole prescription is part of the answer

The elementary energy integral

$$J(t)=\int_{-\infty}^{\infty}\frac{\mathrm dp^0}{2\pi} \frac{i\,e^{-ip^0t}} {(p^0)^2-E_{\mathbf p}^2+i0}$$

illustrates why a denominator alone does not define a propagator. The Feynman prescription places the positive-energy pole just below the real axis and the negative-energy pole just above it. For $t>0$, $e^{-ip^0t}$ decays in the lower half-plane, whose closure is clockwise. The residue and orientation give

$$J(t)=\frac{e^{-iE_{\mathbf p}t}}{2E_{\mathbf p}}.$$

For $t<0$, close counterclockwise above and obtain

$$J(t)=\frac{e^{iE_{\mathbf p}t}}{2E_{\mathbf p}}.$$

Together these are the time-ordered boundary conditions. Retarded, advanced, and principal-value prescriptions move the poles differently and produce different distributions even though the polynomial denominator is the same. The relation between contour choice and causal propagators is developed in Schwartz 2014, §§ 6.1–6.2.

Exercises

1. A contour that crosses a pole

Let $C_R$ be the real interval from $-R$ to $R$ closed by the upper semicircle, with $R>2$. Evaluate

$$\oint_{C_R}\frac{\mathrm dz}{(z-i)(z+2i)}.$$

Then move the closing semicircle to the lower half-plane and explain the change, keeping the real segment oriented from $-R$ to $R$.

Solution

The upper contour is counterclockwise and encloses only $z=i$. Its residue is

$$\operatorname{Res}_{z=i}\frac1{(z-i)(z+2i)} =\frac1{3i},$$

so the integral is $2\pi/3$. The lower closure, while retaining the real segment from left to right, is clockwise and encloses $z=-2i$. Its residue is $1/(-3i)=i/3$, hence

$$\oint_{C_R^-}f(z)\,\mathrm dz =-2\pi i\left(\frac{i}{3}\right)=\frac{2\pi}{3}.$$

The two closed-contour integrals happen to agree, but their arc integrals and enclosed poles differ. In the $R\to\infty$ limit both arcs vanish as $1/R$, which makes the shared real-axis integral equal to $2\pi/3$. The orientation minus sign is essential in the lower closure.

2. Check the first omitted order

For $g=1$, define

$$I_1(\lambda)=\sqrt{\frac{2\pi}{\lambda}} \left(1-\frac{3}{4\lambda}\right).$$

What scaling should $I(\lambda,1)-I_1(\lambda)$ have as $\lambda\to\infty$? Give a numerical test that distinguishes this claim from the leading approximation alone.

Solution

The bracket omits terms of order $\lambda^{-2}$, while the prefactor is $\lambda^{-1/2}$. Therefore

$$I(\lambda,1)-I_1(\lambda)=\mathcal O(\lambda^{-5/2}).$$

Numerically evaluate the integral at a sequence such as $\lambda=10,20,40,80$ using sufficient precision, and inspect

$$\lambda^{5/2}\bigl[I(\lambda,1)-I_1(\lambda)\bigr].$$

A bounded sequence approaching a constant supports the stated order. By contrast, multiplying the error of the leading approximation only by $\lambda^{3/2}$ tests the first correction but not the remainder after that correction has been included.

3. Continue without extending convergence

Use $F(a)$ above to compare $a=1$, $a=-1$, and $a=i$. State separately the value of the continued function and whether the original improper integral converges.

Solution

At $a=1$, the integral converges and the value is $F(1)=1$. At $a=-1$, the continued function has value $-1$, while the integral of $e^x$ diverges. At $a=i$, the continued function has value $1/i=-i$, but the undamped integral $\int_0^\infty e^{-ix}\,\mathrm dx$ does not converge as an ordinary improper Riemann or Lebesgue integral. A regulator or distributional boundary value can assign a related object, but that is a new prescription and must be named.

Re-check and return

Choose one parameter-dependent contour integral and one Laplace-type integral. For the contour problem, state the convergence domain, all singularities and branch data, the deformation, and any crossed residues. For the saddle problem, state the limit, fixed parameters, relevant saddles and endpoints, the leading term, first correction, and the order of the residual.

The repair is complete when the continued function is distinguished from its original representation, every contour obstruction is visible, and the residual scales like the first omitted term. If a branch, arc, or remainder condition is still implicit, annotate it and repeat with a different parameter value.

Then retry the mathematics diagnostic or return to Readiness. The next Core use is in Loops and regularization.

References

• Lars V. Ahlfors, Complex Analysis, third edition, McGraw–Hill, 1979; reprint, American Mathematical Society, 2021, doi:10.1090/chel/385.
• Frank W. J. Olver, Asymptotics and Special Functions, A K Peters, 1997, doi:10.1201/9781439864548.
• Matthew D. Schwartz, Quantum Field Theory and the Standard Model, Cambridge University Press, 2014, doi:10.1017/9781139540940.