Laplace Method and Steepest Descent

For an integral

$$I(\Lambda) = \int_\Gamma a(z)e^{-\Lambda S(z)}\,\mathrm dz, \qquad \Lambda\to+\infty,$$

a critical point contributes only if the original oriented contour can be deformed legally onto a descent path through it. Among those accessible saddles, the smallest relevant value of $\operatorname{Re}S$ sets the exponential scale. The local Hessian, amplitude, contour orientation, and square-root branch then set the power of $\Lambda$, coefficient, and phase. Solving $S'(z)=0$ and ranking every solution is therefore not the method.

This page treats isolated nondegenerate saddles in a fixed parameter region. Endpoints are included because they can dominate without being stationary. Oscillatory limits, coalescing saddles, and changes of contributing cycles require a different uniform analysis.

Required background. Asymptotic Scales, Remainders, Uniformity, and Optimal Truncation supplies the fixed-order meaning of the expansions and explains why a fixed-order remainder estimate does not justify a truncation order that grows with $\Lambda$.

Helpful background. Holomorphic Functions and Cauchy Theory supplies the deformation conditions that make a complex contour replacement an equality rather than a picture.

Laplace’s method and steepest-descent data

The input is not just the formula for $S$. One needs:

• the limiting direction of $\Lambda$ and any auxiliary-parameter region;
• the amplitude $a$, phase or action $S$, and their singularities;
• the oriented contour $\Gamma$, including its endpoints or decay sectors;
• the sheet and branches on which the integrand is defined;
• and the desired number of fixed asymptotic orders.

A reproducible calculation then has six stages:

1. List interior critical points, endpoints, singularities, and other candidate dominant regions.
2. Determine which candidates are reachable by a deformation that preserves the integral.
3. At each retained candidate, choose a local coordinate that makes the leading exponent Gaussian or endpoint-linear.
4. Expand the transformed amplitude, integrate its moments, and record the orientation and branch.
5. Sum all contributions of the same exponential size before simplifying phases or declaring dominance.
6. State the parameter region, remainder meaning, checks, and the condition that would stop the calculation.

The local algebra is usually inexpensive. The difficult part is global: finding all relevant critical points and proving a contour deformation. In several variables, locating saddles and factoring the Hessian may dominate the work. A nearly singular or badly conditioned Hessian is both a numerical warning and evidence that the nondegenerate expansion may be losing uniformity.

A real interior minimum

Consider

$$I(\Lambda) = \int_a^b g(x)e^{-\Lambda f(x)}\,\mathrm dx,$$

where $f\in C^\infty([a,b];\mathbb R)$ and $g\in C^\infty([a,b];\mathbb C)$. Suppose $f$ has a unique global minimum at an interior point $x_0$ and

$$f'(x_0)=0, \qquad f_2\equiv f''(x_0)>0.$$

Write $f_k=f^{(k)}(x_0)$, $g_k=g^{(k)}(x_0)$, and $f_0=f(x_0)$. For every fixed $N$,

$$\begin{aligned} I(\Lambda) &= e^{-\Lambda f_0} \sqrt{\frac{2\pi}{\Lambda f_2}} \left[ \sum_{n=0}^{N-1}\frac{A_n}{\Lambda^n} +O\!\left(\Lambda^{-N}\right) \right],\\ A_0&=g_0. \end{aligned}$$

The first correction is

$$\begin{aligned} A_1={}& \frac{g_2}{2f_2} -\frac{g_1f_3}{2f_2^2} -\frac{g_0f_4}{8f_2^2}\\ &+\frac{5g_0f_3^2}{24f_2^3}. \end{aligned}$$

The hidden constant is allowed to depend on $N$, $f$, $g$, and the interval. If $g_0\neq0$, the leading statement can be written

$$I(\Lambda) \sim g_0e^{-\Lambda f_0} \sqrt{\frac{2\pi}{\Lambda f_2}}.$$

If $g_0=0$, that equivalence is false: continue to the first nonzero transformed even coefficient. If every such coefficient vanishes, the local contribution is beyond all algebraic orders and may vanish identically.

This is the standard nondegenerate interior form of NIST DLMF 2026, Laplace’s method in §2.3(iii). Hunter 2004, §§3.5–3.6 treatment, PDF gives the corresponding localization and Gaussian-reduction argument on pp. 43–47.

Why the Gaussian appears

Choose a small neighborhood $U$ of $x_0$. Uniqueness of the minimum on a compact interval gives an action gap $\eta>0$ outside $U$:

$$f(x)\geq f_0+\eta \qquad (x\in[a,b]\setminus U).$$

That part of the integral is exponentially smaller than $e^{-\Lambda f_0}\Lambda^{-m}$ for every fixed $m$. Inside $U$, define

$$u = \operatorname{sgn}(x-x_0) \sqrt{2\bigl(f(x)-f_0\bigr)}.$$

Nondegeneracy makes this a smooth local coordinate with

$$f(x)=f_0+\frac{u^2}{2}, \qquad h(u)=g(x(u))\frac{\mathrm dx}{\mathrm du}, \qquad h(0)=\frac{g_0}{\sqrt{f_2}}.$$

Taylor expansion of $h$ and the Gaussian moments give

$$I(\Lambda) \sim e^{-\Lambda f_0} \sqrt{\frac{2\pi}{\Lambda}} \sum_{n=0}^{\infty} \frac{h^{(2n)}(0)} {2^n n!\,\Lambda^n}.$$

Odd powers integrate to zero. This argument separates the two ingredients of Laplace’s method: a global action gap localizes the integral, and a local quadratic normal form computes it.

Several minima

If there are finitely many separated nondegenerate local minima, repeat the localization around every candidate not excluded by a larger-scale argument. Those above the global minimum are exponentially suppressed, while all global minima contribute at the same order. Their amplitudes must be summed; symmetry factors, relative signs, or complex phases can matter.

Endpoints and other changes of scale

A dominant point need not satisfy $f'=0$. Suppose the unique minimum is the left endpoint $a$, with

$$f_1\equiv f'(a)>0.$$

Writing $g_k=g^{(k)}(a)$ and $f_2=f''(a)$, the substitution $u=f(x)-f(a)$ gives

$$\begin{aligned} I(\Lambda) =e^{-\Lambda f(a)} \biggl[ &\frac{g_0}{\Lambda f_1}\\ &+\frac{1}{\Lambda^2} \left( \frac{g_1}{f_1^2} -\frac{g_0f_2}{f_1^3} \right) +O\!\left(\Lambda^{-3}\right) \biggr]. \end{aligned}$$

The contributing width is now $O(\Lambda^{-1})$, not $O(\Lambda^{-1/2})$. Thus blindly attaching a Gaussian to every dominant region gives the wrong power.

Other failures of the quadratic model are equally diagnostic. A zero Hessian eigenvalue changes the scaling; a saddle approaching an endpoint or singularity destroys the separated-neighborhood argument; and a singular or vanishing amplitude can change the leading power. These are not small corrections to the same formula.

Complex descent geometry

Return to

$$I(\Lambda) = \int_\Gamma a(z)e^{-\Lambda S(z)}\,\mathrm dz, \qquad \Lambda>0.$$

Assume $S$ and $a$ are holomorphic on a specified sheet throughout the deformation region. The contour is oriented, its finite endpoints are fixed or its infinite ends remain in decay sectors, and added arcs at infinity vanish. Retained saddles $z_\sigma$ are isolated and satisfy

$$S'(z_\sigma)=0, \qquad S''(z_\sigma)\neq0.$$

These analyticity, endpoint, sector, and branch requirements are the structural conditions in NIST DLMF 2026, §2.4(iii)–(iv).

For the sign convention $e^{-\Lambda S}$, the outward gradient-flow equation

$$\frac{\mathrm dz}{\mathrm d\tau} = \overline{S'(z)}$$

obeys

$$\frac{\mathrm d}{\mathrm d\tau}\operatorname{Re}S =|S'(z)|^2, \qquad \frac{\mathrm d}{\mathrm d\tau}\operatorname{Im}S =0.$$

A descent path therefore leaves a saddle along constant $\operatorname{Im}S$ while $\operatorname{Re}S$ increases. The integrand decays away from the saddle. Near $z_\sigma$,

$$S(z)-S(z_\sigma) = \frac12 S''(z_\sigma)(z-z_\sigma)^2 +O\!\left((z-z_\sigma)^3\right).$$

If the outgoing tangent has angle $\theta$, its two decay directions satisfy

$$\arg S''(z_\sigma)+2\theta \equiv0\pmod{2\pi};$$

the ascent directions are rotated by $\pi/2$. Reversing the sign in the exponential reverses which directions descend, so this sign check should precede any contour sketch.

Legal deformation and the saddle coefficient

Within a region where the integrand is holomorphic and the endpoint conditions remain valid, an allowed deformation may be represented schematically as

$$\Gamma \simeq \sum_\sigma n_\sigma\mathcal J_\sigma,$$

where $\mathcal J_\sigma$ is an oriented descent cycle and $n_\sigma$ records whether and with what orientation it occurs. This statement must be established globally. Crossing a pole, a branch cut, or a pinch is not an innocuous deformation; nor is moving an endpoint out of its decay sector.

On an oriented descent cycle, choose a real coordinate $u$ increasing with the contour such that

$$S(z)-S(z_\sigma)=\frac{u^2}{2}.$$

Then

$$h_\sigma(u) = a(z(u))\frac{\mathrm dz}{\mathrm du}$$

is the transformed amplitude, and the local contribution is

$$\begin{aligned} I_\sigma(\Lambda) =n_\sigma e^{-\Lambda S(z_\sigma)} \sqrt{\frac{2\pi}{\Lambda}} \left[ h_\sigma(0) +O\!\left(\Lambda^{-1}\right) \right]. \end{aligned}$$

Because

$$\left(\frac{\mathrm du}{\mathrm dz}\right)^2 =S''(z_\sigma),$$

one often writes

$$h_\sigma(0) = \frac{a(z_\sigma)} {\sqrt{S''(z_\sigma)}}.$$

The square root here is not automatically the principal root. Its sign and phase are fixed by the oriented local coordinate. Reversing the contour reverses the contribution.

Only after finding $n_\sigma\neq0$ is it meaningful to rank saddles by $\operatorname{Re}[\Lambda S(z_\sigma)]$. If several retained saddles have the same real exponential weight, sum their complex contributions before taking a magnitude. Relative phases or a zero of the amplitude may cancel a term that appears dominant saddle by saddle.

Exact check: a rotated Gaussian

The family

$$G_\theta(\Lambda) = \int_{e^{i\theta}\mathbb R} e^{-\Lambda z^2/2}\,\mathrm dz, \qquad |\theta|<\frac{\pi}{4},$$

is an exact test of orientation and branch bookkeeping. Orient the line by $z=e^{i\theta}x$ with $x$ increasing from $-\infty$ to $+\infty$. Since $\operatorname{Re}(e^{2i\theta})>0$,

$$\begin{aligned} G_\theta(\Lambda) &= e^{i\theta} \int_{\mathbb R} \exp\!\left( -\frac{\Lambda e^{2i\theta}x^2}{2} \right)\mathrm dx\\ &= e^{i\theta} \sqrt{\frac{2\pi} {\Lambda e^{2i\theta}}} = \sqrt{\frac{2\pi}{\Lambda}}. \end{aligned}$$

The square root is continued from $\theta=0$, so $\sqrt{e^{2i\theta}}=e^{i\theta}$ in this sector. Its phase cancels the tangent factor. Choosing an unrelated principal value could give a spurious sign. At $|\theta|=\pi/4$, absolute Gaussian decay is lost; the calculation has reached an oscillatory boundary rather than a uniform continuation of this result.

Uniformity and stop rules

If the saddles depend on an auxiliary parameter $\alpha$ in a compact set $K$, a uniform expansion needs uniform versions of the assumptions. In particular, one needs:

• a lower bound $|S''(z_\sigma(\alpha),\alpha)|\geq c>0$;
• separation of retained saddles from one another, endpoints, singularities, and branch cuts;
• a positive real-action gap outside their neighborhoods;
• on an unbounded contour, a uniform tail estimate, such as coercive growth of $\operatorname{Re}S$ together with controlled growth of the amplitude;
• bounded derivatives needed for the chosen fixed order;
• and one stable contour deformation, orientation, and branch choice on $K$.

Stop the nondegenerate calculation when a Hessian eigenvalue approaches zero, saddles coalesce, a saddle meets an endpoint or singularity, a contour becomes pinched, a decay sector closes, or the contributing-cycle data changes. Stationary Phase, Coalescing Saddles, and Stokes Geometry develops the oscillatory, coalescing, and transition analysis needed at those boundaries.

A finite-dimensional Euclidean field integral

Revisit the normalized zero-dimensional Euclidean scalar integral from the prerequisite:

$$Z(\Lambda) = \sqrt{\frac{\Lambda}{2\pi}} \int_{\mathbb R} \exp\!\left[ -\Lambda \left( \frac{\phi^2}{2} +\frac{\phi^4}{24} \right) \right]\mathrm d\phi, \qquad \Lambda\to+\infty.$$

It is a finite-dimensional analogue of a regulated field integral, with

$$S(z)=\frac{z^2}{2}+\frac{z^4}{24}.$$

The candidate saddles and their local quadratic data are:

• Real saddle $z_0=0$. Here $S(z_0)=0$ and $S''(z_0)=1$.
• Upper saddle $z_+=+i\sqrt6$. Here $S(z_+)=-3/2$ and $S''(z_+)=-2$.
• Lower saddle $z_-=-i\sqrt6$. Here $S(z_-)=-3/2$ and $S''(z_-)=-2$.

The complex saddles appear exponentially preferable if one ranks action values alone. That conclusion is wrong for the stated integral. The real axis is already the oriented descent path through $0$: $S(x)$ is real and increases away from its unique real minimum. No legal deformation used here introduces the descent paths through $\pm i\sqrt6$.

The saddle actions are phase-aligned at this parameter value, so alternative thimble bases can be ambiguous on the Stokes boundary. The direct real-contour representation is the unambiguous statement needed here; the next page treats how saddle descriptions change under continuation. The elementary bound

$$0<Z(\Lambda)\leq1$$

also rules out any uncancelled contribution proportional to $e^{3\Lambda/2}$.

At $z=0$, the local scale is $x=\sqrt{\Lambda}\phi$. With $X$ a standard normal variable,

$$Z(\Lambda) = \mathbb E\!\left[ e^{-X^4/(24\Lambda)} \right].$$

Gaussian moments give

$$Z(\Lambda) = 1-\frac{1}{8\Lambda} +\frac{35}{384\Lambda^2} +O\!\left(\Lambda^{-3}\right).$$

This agrees with the general coefficient formula: $f_2=1$, $f_3=0$, $f_4=1$, and $g=1$ give $A_1=-1/8$. More strongly, the sign-definite remainder proved on the prerequisite page supplies the independent check

$$0 \leq \left( 1-\frac{1}{8\Lambda} +\frac{35}{384\Lambda^2} \right) -Z(\Lambda) \leq \frac{385}{3072\Lambda^3}$$

for every $\Lambda>0$. Here the contour analysis selects the saddle, while the Gaussian-moment calculation produces its fixed-order coefficients.

After nondimensionalization, the several-variable analogue is

$$Z_\hbar = \int_\Gamma \frac{\mathrm d^d\phi}{(2\pi\hbar)^{d/2}}\, a(\phi)e^{-S_E(\phi)/\hbar}.$$

For one contributing real minimum $\phi_\star$ with positive-definite Hessian $H$,

$$Z_\hbar = e^{-S_E(\phi_\star)/\hbar} \frac{1}{\sqrt{\det H}} \left[ a(\phi_\star)+O(\hbar) \right].$$

This is the finite-dimensional Gaussian fluctuation determinant. It is not yet an infinite-dimensional path integral: regularization, gauge fixing, zero modes, renormalized determinants, and the physical choice of integration cycle require additional arguments. In particular, a zero or negative eigenvalue must not be hidden by replacing $\det H$ with $|\det H|$.

Zinn-Justin 2021, §1.3 derives this finite-dimensional Hessian factor, while Mariño 2015, §1.3 shows in a zero-dimensional quartic model how decay sectors and the deformed contour determine which saddle contributions occur.

Common pitfalls

Ranking before checking accessibility. A saddle with smaller $\operatorname{Re}S$ is irrelevant when its descent cycle has zero coefficient for the original contour.

Ignoring endpoints and singularities. The leading region may be a nonstationary endpoint, and singularities can obstruct a proposed deformation or generate separate contributions.

Using an untracked square root. The Hessian square root inherits its phase from the oriented contour. A principal-root convention chosen after the calculation can change the answer’s sign or phase.

Turning a local result into a global one. The Gaussian normal form computes a neighborhood. A separate action-gap or contour argument must show that all omitted regions are smaller.

Taking an absolute determinant. Negative and zero Hessian eigenvalues signal instability, a collective coordinate, or a different integration cycle; they are not ordinary positive Gaussians.

Claiming uniformity through a transition. Coalescence, a pinched contour, a closing decay sector, or a changing contributing set invalidates fixed separated-saddle estimates.

Choosing a growing truncation order from a fixed-order proof. The expansion here is Poincaré asymptotic at each fixed order. Optimal truncation needs the order-dependent control developed on the prerequisite page.

Check your understanding

1. Retrieve the selection rule

For $\int_\Gamma a(z)e^{-\Lambda S(z)}\,\mathrm dz$, what must be established before comparing the real parts of saddle actions?

Solution

One must first fix the limiting sector, sheet, branches, and oriented contour; locate endpoints and singularities; and prove which descent cycles occur in a legal deformation of $\Gamma$. Only saddles with nonzero contour coefficients enter the comparison. The smallest real action among that retained set controls the exponential size, subject to amplitude zeros and cancellations.

2. Remove a hypothesis

Why does

$$\int_0^\infty e^{-\Lambda x}\,\mathrm dx = \frac{1}{\Lambda}$$

contradict a rule that every leading contribution has Gaussian width $\Lambda^{-1/2}$?

Solution

The dominant point is the endpoint $x=0$, where the phase has nonzero derivative. Setting $u=\Lambda x$ shows that the contributing width is $O(\Lambda^{-1})$. The quadratic interior-saddle hypothesis has been removed, so its Gaussian scaling does not apply.

3. Calculate a local correction

Evaluate

$$I_\alpha(\Lambda) = \int_{-\infty}^{\infty} (1+\alpha x^2) e^{-\Lambda(1+x^2/2)}\,\mathrm dx$$

and identify the saddle coefficient.

Solution

The unique saddle is $x_0=0$, with $f(0)=1$ and $f''(0)=1$. The ordinary Gaussian and its second moment give the exact result

$$I_\alpha(\Lambda) = e^{-\Lambda} \sqrt{\frac{2\pi}{\Lambda}} \left( 1+\frac{\alpha}{\Lambda} \right).$$

Thus $A_0=1$ and $A_1=\alpha$. The general formula agrees because $g''(0)=2\alpha$ and all derivatives $f_k$ with $k\geq3$ vanish.

4. Transfer to two regulated modes

For

$$S_E(\phi_1,\phi_2) = S_\star +\frac12 \left( m^2\phi_1^2+\omega^2\phi_2^2 \right),$$

evaluate the normalized real integral and diagnose the limits $\omega\to0$ and $\omega^2<0$.

Solution

If $m>0$ and $\omega>0$, then

$$\int_{\mathbb R^2} \frac{\mathrm d\phi_1\,\mathrm d\phi_2}{2\pi\hbar} e^{-S_E/\hbar} = \frac{e^{-S_\star/\hbar}}{m\omega}.$$

As $\omega\to0$, the Hessian loses its spectral gap and the nondegenerate formula ceases to be uniform. If $\omega^2<0$, the stated real Gaussian diverges. A collective-coordinate treatment, stabilizing higher-order terms, or a justified different cycle is required; taking an absolute determinant is not a repair.

What the calculation establishes

For isolated nondegenerate saddles in a fixed sector, the method now has a definite order: establish global contour accessibility, compare the real actions of the accessible candidates, and only then compute oriented local Gaussian coefficients. The same analysis distinguishes interior, endpoint, and several-saddle contributions, supplies exact branch and remainder checks, and identifies when a uniform transition or physical cycle analysis must replace it.

Where the method continues

Stationary Phase, Coalescing Saddles, and Stokes Geometry treats oscillatory cancellation, degenerate or merging saddles, uniform transition approximations, and changes across Stokes geometry. WKB and Eikonal Methods and Turning-Point Matching carries related local models into differential equations.

For physical semiclassical expansions, Saddles, Control Parameters, and Loop Counting develops the field-theory interpretation. Negative Modes and Instability Indices treats negative modes, imaginary contributions, and stability, while Complex Saddles, Lefschetz Thimbles, and Integration Cycles develops physical cycle selection and intersection data. Those later pages require regulated functional-integral arguments beyond the finite-dimensional method proved here.

References

• John K. Hunter, Asymptotic Analysis and Singular Perturbation Theory, PDF, University of California, Davis (2004), §§3.5–3.6, pp. 43–47. Real Laplace localization, Gaussian reduction, and introductory contour geometry.

• Marcos Mariño, Instantons and Large N: An Introduction to Non-Perturbative Methods in Quantum Field Theory, Cambridge University Press (2015), §1.3, pp. 12–15. A zero-dimensional quartic integral, decay sectors, rotated contours, accessible saddles, and Gaussian contributions.

• NIST Digital Library of Mathematical Functions, version 1.2.7, released June 15, 2026, National Institute of Standards and Technology, §2.3(iii), “Laplace’s Method”, §2.4(iii), “Laplace’s Method” for contour integrals, and §2.4(iv), “Saddle Points”. Structural hypotheses, sectors, contour deformation, saddle paths, branches, and fixed-order expansions.

• Jean Zinn-Justin, Quantum Field Theory and Critical Phenomena, 5th ed., Oxford University Press (2021), §1.3, pp. 4–5. Finite-dimensional steepest descent and the Gaussian Hessian factor.