Method of Steepest Descent

Derivation

Method of Steepest Descent

D-193 Home PU-205 Threads chance · energy Depends on Laplace's Method for Exponential Integrals, Real Integrals by Residues & Jordan's Lemma, Cauchy-Goursat Theorem

Statement

For an integral \( I(\lambda)=\int_{C} g(z)\,e^{\lambda f(z)}\,dz \) with \( f,g \) analytic and \( \lambda\to+\infty \), deform the contour \( C \) through a saddle point \( z_0 \) (where \( f'(z_0)=0 \)) along the path of steepest descent, on which \( \operatorname{Im}f \) is constant and \( \operatorname{Re}f \) falls most rapidly, to obtain the leading asymptotics \( I(\lambda)\sim g(z_0)\,e^{\lambda f(z_0)}\,e^{i\theta}\sqrt{\dfrac{2\pi}{\lambda\,|f''(z_0)|}} \).

Why it matters

Laplace's method handles real exponents where the integrand simply peaks; but an enormous class of physical integrals — partition functions in the microcanonical ensemble, inverse Laplace transforms, semiclassical (WKB) wavefunctions, diffraction fields, the density of states — carry a large complex exponent that oscillates rather than merely peaking. Naive peaking fails because the phase spins. Steepest descent rescues the estimate by using Cauchy's theorem to slide the contour onto a route where the oscillation is switched off.

It is the analytic backbone of the saddle-point approximation: it converts "count the overwhelmingly most probable microstate" (chance) and "the configuration of stationary action" (energy) into the same Gaussian-around-a-critical-point calculation, and it tells you not just the magnitude but the crucial phase \( e^{i\theta} \) that stationary-phase and Fresnel diffraction results turn on.

Assumptions

Assumption.\( f \) and \( g \) are analytic (holomorphic) in a simply connected region containing both \( C \) and the deformed contour. Without analyticity Cauchy–Goursat does not apply and the contour cannot be moved freely; the whole strategy collapses.

Assumption.The large parameter \( \lambda \) is real and \( \lambda\to+\infty \), so that \( e^{\lambda\operatorname{Re}f} \) concentrates. If \( \lambda \) stays finite the "leading term" is not dominant and the neglected corrections are not small.

Assumption.The saddle is simple and isolated: \( f'(z_0)=0 \) but \( f''(z_0)\neq 0 \). If \( f''(z_0)=0 \) the quadratic term vanishes, the Gaussian collapses, and one needs the higher-order (Airy / cubic) generalisation instead.

Assumption.The contour can be deformed onto the steepest-descent path with no singularities of \( g \) crossed and no residual contribution from the arcs at infinity (the tails of \( e^{\lambda f} \) decay). If a pole is swept across, a residue term must be added (Stokes phenomenon); if the arcs do not decay the endpoint contributions survive and the estimate is incomplete.

Assumption.A single saddle dominates. If two saddles have equal \( \operatorname{Re}f(z_0) \) their contributions must be summed coherently, and interference (an anti-Stokes line) governs the answer.

Derivation

\[ I(\lambda)=\int_{C} g(z)\,e^{\lambda f(z)}\,dz,\qquad \lambda\to+\infty \]

Starting integral. The modulus of the integrand is \( |g|\,e^{\lambda\operatorname{Re}f} \), so for large \( \lambda \) the value is controlled entirely by where \( \operatorname{Re}f \) is largest on the contour. A

\[ \operatorname{Re}f \text{ is harmonic} \;\Rightarrow\; \nabla^2(\operatorname{Re}f)=0 \;\Rightarrow\; \text{no interior maxima} \]

The real part of an analytic function is harmonic, so by the maximum principle it has no interior local maximum. Its critical points are therefore saddles, not peaks — the largest value on a freely deformable contour is achieved by routing through a saddle. B

\[ f'(z_0)=0 \]

Define the saddle point \( z_0 \) as a critical point of \( f \). By the Cauchy–Riemann equations, \( \nabla(\operatorname{Re}f)=0 \) exactly when \( f'(z_0)=0 \), so this single complex condition locates the saddle of the real relief. B

\[ \int_{C}=\int_{C'} \quad\text{whenever } C\to C' \text{ sweeps no singularity} \]

Cauchy–Goursat: for \( f,g \) analytic, the integral is unchanged under continuous deformation of the contour between fixed endpoints, provided no pole or branch cut is crossed and the connecting arcs contribute nothing. This lets us replace \( C \) by a contour \( C' \) through \( z_0 \). A

\[ f(z)=f(z_0)+\tfrac12 f''(z_0)\,(z-z_0)^2+O\!\big((z-z_0)^3\big) \]

Taylor-expand \( f \) about the saddle. The linear term drops because \( f'(z_0)=0 \), leaving a quadratic as the leading variation. For large \( \lambda \) only a neighbourhood of \( z_0 \) contributes, so cubic and higher terms are subdominant. A

\[ f''(z_0)=|f''(z_0)|\,e^{i\alpha},\qquad z-z_0=t\,e^{i\theta},\ t\in\mathbb{R} \]

Write the (complex) curvature in polar form and parametrise the local contour as a straight ray through \( z_0 \) at angle \( \theta \). The angle \( \theta \) is the direction we are still free to choose. B

\[ \tfrac12 f''(z_0)(z-z_0)^2=\tfrac12\,|f''(z_0)|\,t^2\,e^{i(\alpha+2\theta)} \]

Substitute step 6 into the quadratic term. Its phase is \( \alpha+2\theta \); to make \( \operatorname{Re}f \) decrease as fast as possible (steepest descent) and to kill the oscillation, we force this term to be real and negative. B

\[ \alpha+2\theta=\pi \;\Rightarrow\; \boxed{\ \theta=\frac{\pi-\alpha}{2}\ } \]

Choosing this angle makes \( \tfrac12 f''(z_0)(z-z_0)^2=-\tfrac12|f''(z_0)|\,t^2 \). Along this ray \( \operatorname{Im}f \) is stationary (constant to quadratic order) so the integrand does not oscillate, and \( \operatorname{Re}f \) falls like a Gaussian — this is the steepest-descent direction. C

\[ I(\lambda)\approx g(z_0)\,e^{\lambda f(z_0)}\,e^{i\theta}\!\int_{-\infty}^{\infty} e^{-\frac12\lambda|f''(z_0)|\,t^2}\,dt \]

Insert the local expansion, use \( dz=e^{i\theta}\,dt \), and pull the slowly varying factors \( g(z_0)\,e^{\lambda f(z_0)} \) out at the saddle. Because the Gaussian confines \( t \) to \( O(\lambda^{-1/2}) \), the \( t \)-limits are extended to \( \pm\infty \) with exponentially small error. B

\[ \int_{-\infty}^{\infty} e^{-a t^2}\,dt=\sqrt{\frac{\pi}{a}},\qquad a=\tfrac12\lambda|f''(z_0)| \]

Evaluate the standard Gaussian integral (a real, convergent integral because \( a>0 \) by construction in step 8). This yields \( \sqrt{2\pi/(\lambda|f''(z_0)|)} \). A

Result

\[ I(\lambda)\;\sim\; g(z_0)\,e^{\lambda f(z_0)}\;e^{i\theta}\sqrt{\frac{2\pi}{\lambda\,|f''(z_0)|}}\,,\qquad \theta=\frac{\pi-\alpha}{2},\ \ \alpha=\arg f''(z_0) \]

Reading. The answer factorises into three transparent pieces: the amplitude of the smooth part at the saddle, \( g(z_0) \); the exponential set by the saddle height, \( e^{\lambda f(z_0)} \); and a Gaussian width factor \( \sqrt{2\pi/(\lambda|f''|)} \) that shrinks like \( \lambda^{-1/2} \) because the peak grows sharper with \( \lambda \). The phase \( e^{i\theta} \) records the tilt of the steepest-descent line; it is the term Laplace's real-axis method never sees, and it is exactly the \( \pi/4 \) Fresnel phase in diffraction. Equivalently one writes \( I\sim g(z_0)\,e^{\lambda f(z_0)}\sqrt{-2\pi/\big(\lambda f''(z_0)\big)} \) with the branch of the root fixed by \( \theta \).

Units check. The exponent \( \lambda f \) is dimensionless, so \( [\lambda][f]=1 \). Then \( \lambda f''=\lambda f/[z]^2 \) has units \( [z]^{-2} \), so \( \sqrt{2\pi/(\lambda|f''|)} \) carries units of \( [z] \). The result therefore has units \( [g]\,[z] \), matching \( \int g\,e^{\lambda f}\,dz \) since the exponential is dimensionless. Consistent.

Limiting cases

Real saddle on the real axis with \( f''(z_0)<0 \): then \( \alpha=\pi \), \( \theta=0 \), \( e^{i\theta}=1 \), and the formula collapses to Laplace's method \( I\sim g(z_0)e^{\lambda f(z_0)}\sqrt{2\pi/(\lambda|f''|)} \). Steepest descent contains Laplace as its no-phase special case.
Purely imaginary exponent, \( f=i\phi \) with real \( \phi \): the saddle becomes a stationary-phase point \( \phi'(x_0)=0 \), \( \alpha=\pm\pi/2 \), \( \theta=\pm\pi/4 \), recovering the method of stationary phase and its signature \( e^{\pm i\pi/4} \).
Saddle at a contour endpoint (half-line): only half the Gaussian is captured, so the result is halved, \( I\sim\tfrac12 g(z_0)e^{\lambda f(z_0)}e^{i\theta}\sqrt{2\pi/(\lambda|f''|)} \).
Very large \( \lambda \): relative correction is \( O(1/\lambda) \), so accuracy improves monotonically as the parameter grows — the approximation is asymptotic, not convergent.

Breaks when

Degenerate saddle, \( f''(z_0)=0 \). The quadratic term vanishes; the Gaussian width diverges and the leading coefficient blows up. One must keep the cubic term, giving an Airy-function scaling \( \lambda^{-1/3} \) instead of \( \lambda^{-1/2} \) — the coalescence of two saddles.
A singularity of \( g \) sits at or near the saddle, or a pole is crossed during deformation. The residue is comparable to the saddle contribution; the Cauchy–Goursat step is no longer clean and a residue term must be added (Stokes phenomenon), so the bare saddle formula is wrong.
Two or more saddles with equal \( \operatorname{Re}f(z_0) \). On an anti-Stokes line their contributions interfere; using one saddle alone misses the oscillatory interference that dominates the answer.
\( \lambda \) not large, or complex \( \lambda \) with wandering phase. The neglected \( O(1/\lambda) \) terms are not small and the "leading" term is not leading; the asymptotic series may even be worse than a few terms of a convergent expansion.

Failure modes

Picking the wrong steepest direction. The condition \( \alpha+2\theta=\pi \) selects descent; the other root \( \alpha+2\theta=0 \) gives steepest ascent. Students who solve \( \operatorname{Im}f=\text{const} \) without checking the sign of \( \operatorname{Re}f'' \) along the ray get an exponentially growing (nonsense) answer.
Dropping the phase \( e^{i\theta} \). Treating \( |f''| \) under the root but forgetting the tilt loses the \( \pi/4 \) factor — fatal in diffraction and stationary-phase problems where the magnitude is right but the phase is observable.
Using \( f''(z_0) \) with the wrong branch of the square root. Writing \( \sqrt{-2\pi/(\lambda f'')} \) and choosing the principal branch blindly can flip the sign of the whole amplitude; the branch must be tied to \( \theta \).
Forgetting to halve at an endpoint saddle. When the critical point lands at the limit of integration (e.g. Bessel \( K_0 \) from \( 0 \) to \( \infty \)) students keep the full Gaussian and overcount by a factor of two.
Extending the \( t \)-integral to \( \pm\infty \) when a nearby singularity forbids it. The Gaussian tail assumption fails if \( g \) has a pole a distance \( O(\lambda^{-1/2}) \) away — the neighbourhood is not "empty".
Solving \( f'(z_0)=0 \) but keeping a saddle that lies off the deformable region. Not every root of \( f' \) is reachable; only saddles the contour can actually be pushed onto contribute.

Discussion

The geometric picture is a mountain relief: \( \operatorname{Re}f \) is a landscape with no summits (harmonicity forbids interior maxima), only passes. A saddle point is a mountain pass; the steepest-descent path is the valley floor running through the pass, and the steepest-ascent path is the ridge crossing it perpendicularly. Because \( \operatorname{Im}f \) is the harmonic conjugate, its level lines are orthogonal to those of \( \operatorname{Re}f \) — which is precisely why the steepest-descent direction is also the constant-phase direction. Killing the oscillation and maximising the descent are the same choice, and that coincidence is what makes the Gaussian integral legitimate.

Physically the method is the mathematical face of "the most probable state dominates". In statistical mechanics the microcanonical count \( \Omega=\oint e^{N s(E)}\,dz \) is evaluated at the saddle where the entropy is stationary, delivering the thermodynamic limit and the equivalence of ensembles (the chance thread). In mechanics and optics the same integral with \( \lambda=1/\hbar \) or \( \lambda=k \) is stationary at the classical trajectory or the geometric ray, so the saddle is the point of stationary action or stationary phase (the energy thread). Steepest descent is therefore the analytic engine behind both the second law's dominance of typical states and the emergence of classical paths from quantum amplitudes.

More deeply, the subleading terms of the expansion form an asymptotic (generally divergent) series in \( 1/\lambda \) whose optimal truncation error is exponentially small, of order \( e^{-c\lambda} \). That exponentially small remainder is not noise: it is the contribution of a rival saddle. As a parameter is varied across a Stokes line, this hidden exponential switches on discontinuously in the asymptotics even though the exact function is analytic — the Stokes phenomenon. Resurgence theory (Écalle, Berry–Howls) makes this precise, showing the late terms of one saddle's series encode the early terms of another's, so the full trans-series reconstructs the exact integral from its saddles.

Common misconceptions. The method does not require the integrand to be large or even bounded on the original contour — it only requires that a deformed contour exist on which \( \operatorname{Re}f \) peaks. It is also not "just Laplace with complex numbers": the essential new content is the freedom to move the contour (Cauchy's theorem) and the phase \( e^{i\theta} \) that this freedom generates. Finally, "steepest descent" refers to the descent of \( \operatorname{Re}f \) in the complex plane, not to any minimisation being performed — the saddle is a stationary point, never a minimum.

Worked examples

Example 1 — Stirling's formula from the Gamma integral (real saddle).

\[ n!=\Gamma(n+1)=\int_0^{\infty} t^{n}e^{-t}\,dt=\int_0^{\infty} e^{\,n\ln t-t}\,dt \]

Write the factorial as an integral and combine into a single exponential. The exponent is not yet in the form \( \lambda f \) with a fixed profile because \( t \) ranges over all scales. A

\[ t=n s:\qquad n!=n^{\,n+1}\!\int_0^{\infty} e^{\,n(\ln s-s)}\,ds,\qquad f(s)=\ln s-s,\ \lambda=n \]

Rescale so the large parameter \( n \) multiplies a fixed profile \( f(s) \). Now the method applies with \( \lambda=n \), \( g=1 \). B

\[ f'(s)=\frac1s-1=0\ \Rightarrow\ s_0=1,\qquad f(1)=-1,\qquad f''(s)=-\frac1{s^2},\ f''(1)=-1 \]

Locate the saddle and its curvature. Since \( f''(1)=-1<0 \) it is a real saddle: \( \alpha=\pi \), \( \theta=0 \), \( e^{i\theta}=1 \). A

\[ n!\sim n^{\,n+1}\,e^{\,n f(1)}\sqrt{\frac{2\pi}{n\,|f''(1)|}}=n^{\,n+1}e^{-n}\sqrt{\frac{2\pi}{n}} \]

Apply the boxed result with \( g(z_0)=1 \). Simplify \( n^{\,n+1}\sqrt{1/n}=n^{\,n}\sqrt{n} \). B

\[ n!\;\sim\;\sqrt{2\pi n}\;\Big(\frac{n}{e}\Big)^{n} \]

Reading. The celebrated Stirling formula, obtained purely from a real saddle. For \( n=10 \) it gives \( 3\,598\,696 \) against the exact \( 3\,628\,800 \) — a \( 0.83\% \) error, already \( \approx 1/(12n) \) as the next correction predicts. Units. Dimensionless throughout (\( t \) and \( n \) pure numbers), as a counting quantity must be.

Example 2 — the Fresnel integral (complex saddle, exact).

\[ I(\lambda)=\int_{-\infty}^{\infty} e^{\,i\lambda x^{2}}\,dx,\qquad f(x)=i x^{2},\ g=1,\ \lambda>0 \]

A purely oscillatory integrand: \( |e^{i\lambda x^2}|=1 \) everywhere, so Laplace's real-axis peaking gives nothing. The exponent is analytic, so we may deform. A

\[ f'(x)=2ix=0\ \Rightarrow\ x_0=0,\qquad f(0)=0,\qquad f''(x)=2i \]

The saddle sits at the origin. The curvature is imaginary: \( f''=2i=2\,e^{i\pi/2} \), so \( |f''|=2 \) and \( \alpha=\pi/2 \). A

\[ \theta=\frac{\pi-\alpha}{2}=\frac{\pi-\pi/2}{2}=\frac{\pi}{4} \]

The steepest-descent line is tilted at \( 45^\circ \) to the real axis. Rotating the real-axis contour onto this ray sweeps no singularity (entire integrand) and the corner arcs vanish, so Cauchy–Goursat licenses the move. C

\[ I(\lambda)\sim g(0)\,e^{\lambda f(0)}\,e^{i\theta}\sqrt{\frac{2\pi}{\lambda|f''|}}=e^{\,i\pi/4}\sqrt{\frac{2\pi}{2\lambda}}=e^{\,i\pi/4}\sqrt{\frac{\pi}{\lambda}} \]

Apply the boxed result. Because \( f \) is exactly quadratic there are no higher terms, so the asymptotic estimate is in fact the exact value. B

\[ \int_{-\infty}^{\infty} e^{\,i\lambda x^{2}}\,dx\;=\;\sqrt{\frac{\pi}{\lambda}}\;e^{\,i\pi/4} \]

Reading. The full Fresnel result, phase and all. The magnitude \( \sqrt{\pi/\lambda} \) is what a real-saddle argument would give; the \( e^{i\pi/4} \) is the pure gift of steepest descent and is exactly the phase advance seen at a caustic in wave optics. Splitting real and imaginary parts reproduces \( \int\cos(\lambda x^2)dx=\int\sin(\lambda x^2)dx=\tfrac12\sqrt{\pi/\lambda} \). Units. With \( x \) a length and \( \lambda \) an inverse-area (so \( \lambda x^2 \) is a dimensionless phase), \( \sqrt{\pi/\lambda} \) has units of length, matching \( \int dx \).

Problems

Verify that steepest descent reproduces the exact Gaussian \( I(\lambda)=\int_{-\infty}^{\infty} e^{-\lambda x^{2}/2}\,dx \).
Solution
Here \( f(x)=-x^2/2 \), \( g=1 \). Then \( f'(x)=-x=0\Rightarrow x_0=0 \), \( f(0)=0 \), \( f''(0)=-1 \). Since \( f''<0 \) real, \( \alpha=\pi \), \( \theta=0 \), \( |f''|=1 \). The boxed result gives \( I\sim 1\cdot e^{0}\cdot 1\cdot\sqrt{2\pi/(\lambda\cdot 1)}=\sqrt{2\pi/\lambda} \). Because \( f \) is exactly quadratic this is exact: \( \int_{-\infty}^{\infty}e^{-\lambda x^2/2}dx=\sqrt{2\pi/\lambda} \). For \( \lambda=2 \) this is \( \sqrt{\pi}\approx1.7725 \).
Find the leading large-\( \lambda \) behaviour of the modified Bessel function \( K_0(\lambda)=\int_{0}^{\infty} e^{-\lambda\cosh t}\,dt \). (Note the endpoint.)
Solution
With \( f(t)=-\cosh t \), \( g=1 \): \( f'(t)=-\sinh t=0\Rightarrow t_0=0 \), which is the endpoint of integration. There \( f(0)=-1 \), \( f''(t)=-\cosh t \), \( f''(0)=-1 \) (real, \( \theta=0 \)). Because the saddle sits at the boundary, only half the Gaussian is captured, so multiply the full-saddle value by \( \tfrac12 \): \( K_0(\lambda)\sim\tfrac12 e^{-\lambda}\sqrt{2\pi/(\lambda\cdot1)}=\sqrt{\pi/(2\lambda)}\;e^{-\lambda} \). This is the standard asymptotic \( K_0(\lambda)\sim\sqrt{\pi/(2\lambda)}\,e^{-\lambda} \). At \( \lambda=5 \): \( \sqrt{\pi/10}\,e^{-5}=0.5605\times0.006738=3.78\times10^{-3} \), against the true \( K_0(5)=3.69\times10^{-3} \) — about \( 2.5\% \) high, consistent with the \( O(1/\lambda) \) correction.
Evaluate the leading asymptotics of \( I(\lambda)=\int_{-\infty}^{\infty} e^{-\lambda x^{2}/2}\,e^{\,i\lambda x^{2}/2}\,dx \) and identify the phase.
Solution
Combine exponents: \( f(x)=-\tfrac12(1-i)x^2 \), \( g=1 \). Saddle \( x_0=0 \), \( f(0)=0 \), \( f''(0)=-(1-i)=-1+i \). In polar form \( f''=\sqrt{2}\,e^{\,i(3\pi/4)} \), so \( |f''|=\sqrt2 \), \( \alpha=3\pi/4 \), and \( \theta=(\pi-3\pi/4)/2=\pi/8 \). Then \( I\sim e^{\,i\pi/8}\sqrt{2\pi/(\lambda\sqrt2)}=e^{\,i\pi/8}\,(2\pi)^{1/2}2^{-1/4}\lambda^{-1/2} \). Check against the exact Gaussian \( \int e^{-\tfrac12(1-i)\lambda x^2}dx=\sqrt{2\pi/[(1-i)\lambda]} \); since \( 1-i=\sqrt2\,e^{-i\pi/4} \), the root is \( \sqrt{2\pi/(\sqrt2\lambda)}\,e^{\,i\pi/8} \) — identical, confirming both magnitude and the \( \pi/8 \) phase.
Estimate the central binomial coefficient \( \binom{2n}{n} \) for large \( n \) using Stirling (Example 1), and give the numeric value at \( n=10 \).
Solution
From \( m!\sim\sqrt{2\pi m}\,(m/e)^m \): \( \binom{2n}{n}=\dfrac{(2n)!}{(n!)^2}\sim\dfrac{\sqrt{4\pi n}\,(2n/e)^{2n}}{\big[\sqrt{2\pi n}\,(n/e)^{n}\big]^2}=\dfrac{\sqrt{4\pi n}\,(2n)^{2n}e^{-2n}}{2\pi n\,n^{2n}e^{-2n}} \). The \( e^{-2n} \) cancels and \( (2n)^{2n}/n^{2n}=4^{n} \), giving \( \binom{2n}{n}\sim\dfrac{\sqrt{4\pi n}}{2\pi n}\,4^{n}=\dfrac{4^{n}}{\sqrt{\pi n}} \). At \( n=10 \): \( 4^{10}/\sqrt{10\pi}=1\,048\,576/5.6050=187\,079 \), against the exact \( \binom{20}{10}=184\,756 \) — a \( 1.3\% \) overestimate.
The function \( I(\lambda)=\int_{C} e^{\lambda(z-z^{3}/3)}\,dz \) has two saddles. Locate them, compute \( \operatorname{Re}f \) at each, and state which dominates when the contour can be routed through either.
Solution
With \( f(z)=z-z^3/3 \), \( f'(z)=1-z^2=0\Rightarrow z_0=\pm1 \). Values: \( f(1)=1-\tfrac13=\tfrac23 \), \( f(-1)=-1+\tfrac13=-\tfrac23 \). Curvatures: \( f''(z)=-2z \), so \( f''(1)=-2 \) (\( \alpha=\pi,\ \theta=0 \)) and \( f''(-1)=+2 \) (\( \alpha=0,\ \theta=\pi/2 \)). The exponential weight is \( e^{\lambda\operatorname{Re}f} \); since \( \operatorname{Re}f(1)=+\tfrac23>\operatorname{Re}f(-1)=-\tfrac23 \), the saddle at \( z_0=+1 \) dominates by a factor \( e^{\,4\lambda/3} \). The leading estimate through it is \( I\sim e^{\,2\lambda/3}\sqrt{2\pi/(2\lambda)}=e^{\,2\lambda/3}\sqrt{\pi/\lambda} \) (with \( \theta=0 \)). The \( z_0=-1 \) saddle contributes an exponentially smaller \( \sim e^{-2\lambda/3} \) term that only matters on a Stokes line where the two are comparable.