Saddles and the Semiclassical Expansion

At a fixed regulator, a stationary configuration organizes a local asymptotic expansion by separating a field into a background and small fluctuations. If the contributing Euclidean saddle is isolated and its Hessian is positive on the allowed fluctuation space, the leading fluctuation integral is Gaussian: its determinant is the first correction to the classical exponential, and higher derivatives generate successive powers of the declared small parameter. Stationarity alone is not enough. The integration cycle, boundary data, zero and negative modes, determinant branch, competing saddles, and remainder all decide whether that formula applies.

Required background. Regulated Bosonic Field Integrals supplies the finite variables, domain or cycle, reference measure, and order of limits. The Action Principle and Field Equations supplies stationarity relative to the allowed variations and the boundary terms that determine the fluctuation space.

Helpful background. Second Variation, Hessians, and Jacobi Operators develops the Hessian language used here. Laplace Method and Steepest Descent and Stationary Phase, Coalescing Saddles, and Stokes Geometry treat the corresponding asymptotic theorems and contour geometry in greater generality.

A stationary configuration defines the fluctuation problem

Throughout this page, $R$ denotes a fixed regulator and $N_R<\infty$ the number of retained real variables. We restore $\hbar$ as a bookkeeping parameter and take $\hbar\to0^+$ before considering removal of $R$. Every action is scaled so that $S/\hbar$ is dimensionless.

The regulated Euclidean object has the form

$$\mathcal I_R(\hbar) = \int_{\Gamma_R} \mathrm d\mu_R(x)\,A_R(x) \exp\!\left[-\frac{S_{E,R}(x)}{\hbar}\right].$$

Its inputs are not just the action. They include the oriented $N_R$-real dimensional cycle $\Gamma_R$, the measure $\mathrm d\mu_R$, the amplitude or insertion $A_R$, the boundary conditions, the regulator, and the small parameter. A saddle $x_\star$ is a critical point with respect to tangent vectors allowed by those data:

$$\delta S_{E,R}[x_\star;\eta]=0 \qquad \text{for every allowed }\eta .$$

This definition does not say that $x_\star$ is a minimum. The quadratic form on the same tangent space is

$$H_{\star,ab} = \left. \frac{\partial^2 S_{E,R}} {\partial x^a\partial x^b} \right|_{x_\star}.$$

Boundary conditions matter twice: they help make the first variation vanish, and they select the vectors on which $H_\star$ acts. A mode forbidden by the boundary data is not an eigenvector of the fluctuation problem.

In local coordinates with $\mathrm d\mu_R=\rho_R(x)\,\mathrm d^{N_R}x$, set $a_R=\rho_R A_R$ and write $x=x_\star+\sqrt{\hbar}\,\eta$. Taylor expansion gives

$$\begin{aligned} \frac{S_{E,R}(x)}{\hbar} ={}& \frac{S_\star}{\hbar} +\frac12\eta^{\mathsf T}H_\star\eta \\ &+ \sum_{n\ge 3} \hbar^{n/2-1} \frac{S_\star^{(n)}[\eta^n]}{n!}. \end{aligned}$$

The $\sqrt{\hbar}$ rescaling is the source of both the Gaussian determinant and the power counting below.

A positive Hessian gives the Euclidean Gaussian term

Suppose $x_\star$ is an isolated interior saddle on a real local chart, $H_\star$ is positive definite, $a_R$ is smooth, and the real part of the action grows away from the saddle on the localized cycle. For any fixed truncation order $m$, Laplace expansion then has the form

$$\begin{aligned} \mathcal I_{\star,R}(\hbar) ={}& e^{-S_\star/\hbar} \frac{(2\pi\hbar)^{N_R/2}} {\sqrt{\det H_\star}} \\ &\times \left[ \sum_{k=0}^{m}c_{k,R}\hbar^k +O_R\!\left(\hbar^{m+1}\right) \right], \qquad c_{0,R}=a_R(x_\star). \end{aligned}$$

The square root here is the positive one. The subscript on $O_R$ is essential: its constant may depend on the cutoff, volume, and boundary conditions. For an exactly quadratic action with constant $a_R$ and the full real cycle, the leading Gaussian expression is exact. For a general action, the coefficients also contain derivatives of the measure density and insertion, not just interaction vertices. This finite-dimensional derivation and its Gaussian-moment expansion are given in Zinn-Justin 2021, § 1.3, pp. 4–5.

This is a local contribution. It becomes an approximation to the full integral only after one shows that the original cycle reaches this saddle and that its complement, including other contributing saddles and endpoints, is smaller at the claimed order. One must not sum every critical point merely because it solves the stationary equation.

A regulated quartic scalar checks the first correction

Consider $N_R$ retained scalar coordinates with the normalized measure

$$\begin{aligned} \mathcal Z_R(\hbar,g) = \int_{\mathbb R^{N_R}} \prod_{n=1}^{N_R}\frac{\mathrm d\phi_n}{\sqrt{2\pi}} \exp\!\Bigg[ -\frac1\hbar \Bigg( \frac12\phi^{\mathsf T}K_R\phi +\frac{g}{4!}\sum_{n=1}^{N_R}\phi_n^4 \Bigg) \Bigg]. \end{aligned}$$

Let $K_R=K_R^{\mathsf T}>0$ and hold $g\ge0$ fixed as $\hbar\to0^+$. The integral is then convergent, the saddle $\phi_\star=0$ is nondegenerate, and its Hessian is $K_R$. Put $C_R=K_R^{-1}$ and rescale $\phi=\sqrt{\hbar}\,\eta$. Wick’s rule for the resulting finite Gaussian gives

$$\left\langle\eta_n^4\right\rangle_0 = 3(C_R)_{nn}^2.$$

Expanding the quartic exponential therefore yields

$$\begin{aligned} \mathcal Z_R(\hbar,g) ={}& \hbar^{N_R/2}(\det K_R)^{-1/2} \\ &\times \left[ 1-\frac{g\hbar}{8} \sum_{n=1}^{N_R}(C_R)_{nn}^2 +O_R(\hbar^2) \right]. \end{aligned}$$

The remainder is genuine at fixed $R$: for $t\ge0$, $\lvert e^{-t}-1+t\rvert\le t^2/2$, and the required Gaussian eighth moments are finite. Nothing here says that the bound is uniform as $N_R\to\infty$.

There is an independent source check. Differentiating at $g=0$ gives

$$\left. \frac{\partial\log\mathcal Z_R}{\partial g} \right|_{g=0} = -\frac{1}{4!\hbar} \sum_n\left\langle\phi_n^4\right\rangle_0 = -\frac{\hbar}{8} \sum_n(C_R)_{nn}^2,$$

because the exact Gaussian covariance is $\langle\phi_m\phi_n\rangle_0=\hbar(C_R)_{mn}$. This reproduces the coefficient using the source derivatives developed on Gaussian Fields and Sources.

For a more general regulated scalar action

$$S_{E,R}(\phi) = \frac12\phi^{\mathsf T}K_R\phi +\sum_n w_n V(\phi_n),$$

a classical configuration $\bar\phi$ obeys

$$(K_R\bar\phi)_n+w_nV'(\bar\phi_n)=0,$$

and its fluctuation matrix is

$$(H_{\bar\phi})_{mn} = (K_R)_{mn} +\delta_{mn}w_nV''(\bar\phi_n).$$

If this matrix is positive on the allowed finite-dimensional fluctuation space and the saddle contributes on the chosen cycle, its local contribution starts as

$$\mathcal Z_{\bar\phi,R}^{\mathrm{loc}} = e^{-S_{E,R}(\bar\phi)/\hbar} \frac{\hbar^{N_R/2}} {\sqrt{\det H_{\bar\phi}}} \left[1+O_R(\hbar)\right].$$

The continuum-looking symbol $-\partial^2+V''(\bar\phi)$ is only a mnemonic for a limit of such regulated matrices. Its determinant is not a defined continuum number until a normalization, regulator removal, and any required renormalization have been supplied.

Powers of ℏ orient loop counting

The rescaled action assigns a factor $\hbar^{n/2-1}$ to an $n$-leg interaction vertex, while contractions of the $\eta$ variables carry no further power of $\hbar$. For a connected vacuum graph with $I$ internal lines and $V$ vertices,

$$\hbar^{\sum_v(n_v/2-1)} = \hbar^{I-V} = \hbar^{L-1}, \qquad L=I-V+1,$$

where $2I=\sum_v n_v$. First factor out the explicitly displayed Gaussian normalization by defining $\widehat{\mathcal Z}_R=\hbar^{-N_R/2}\mathcal Z_R$, or use a ratio with the same $N_R$ in which that factor cancels. Assuming the action parameters and Hessian do not themselves depend on $\hbar$, the classical action in $\log\widehat{\mathcal Z}_R$ is order $\hbar^{-1}$, the saddle-dependent finite determinant is the one-loop term of order $\hbar^0$, and an $L$-loop connected vacuum contribution is order $\hbar^{L-1}$. Equivalently, in $-\hbar\log\widehat{\mathcal Z}_R$, tree, one-loop, and two-loop terms scale as $\hbar^0,\hbar^1,\hbar^2$. Without this normalization, $(N_R/2)\log\hbar$ is also present in $\log\mathcal Z_R$.

The single quartic vertex in the worked example contracts into two internal lines. It has $I=2$, $V=1$, and $L=2$, so its order-$\hbar$ term in $\log\mathcal Z_R$ is a two-loop connected vacuum contribution. Small $g$ and small $\hbar$ are distinct expansions; here $g$ is fixed. The regulated field expansion and this loop organization are developed in Zinn-Justin 2021, § 7.9, pp. 146–150.

Zero modes and negative directions change the local model

The determinant formula is a diagnostic, not a prescription that survives every Hessian spectrum.

Hessian or cycle situation	What must change
Positive, nondegenerate Euclidean Hessian	Use the positive-root Gaussian determinant on the allowed fluctuation space.
Symmetry-generated zero mode	Replace the orbit direction by a collective coordinate, include its induced Jacobian, and take a determinant only on the transverse space.
Other zero or critical mode	Retain the first nonzero higher term and rescale again; fractional powers of $\hbar$ can result.
Gauge zero mode	Fix the redundancy and include the corresponding measure factor; it is not automatically a physical modulus.
Negative Euclidean direction on the real cycle	The quadratic Gaussian is not damping; an admissible contour and determinant phase must be specified.
Complex critical point	Existence of the point does not show that the original cycle has a contribution from it.
Several saddles	Local series do not determine which saddles contribute or dominate.

If a continuous family $\bar\phi(a)$ consists of saddles, differentiating the stationary equation gives $H_{\bar\phi}\,\partial_a\bar\phi=0$. The tangent coefficient is not an ordinary Gaussian variable. One projects it out, normalizes the zero mode, and replaces it by integration over $a$ with the induced Jacobian. Writing $\det' H$ records only the transverse determinant; it does not perform the replacement or determine its normalization. These zero-mode steps are illustrated explicitly in Coleman 1985, ch. 7, § 2.2, pp. 276–277. A gauge orbit instead requires gauge fixing; it is a redundancy, not automatically a physical modulus.

An accidental degeneracy has different scaling. For example,

$$\int_{-\infty}^{\infty} \mathrm dx\, e^{-x^4/(4\hbar)} = \hbar^{1/4} \int_{-\infty}^{\infty} \mathrm dy\, e^{-y^4/4},$$

so the leading power is $\hbar^{1/4}$, not the $\hbar^{1/2}$ predicted by a nonexistent quadratic term.

A negative eigenvalue is equally decisive. Along that direction the real Euclidean Gaussian grows, so the positive-root formula is invalid even when the full nonlinear integral happens to converge. A contour deformation may define a local contribution and assign it a phase, but that phase belongs to the oriented contour and its analytic continuation—not to the eigenvalue alone. In particular, a negative mode does not by itself imply an imaginary physical observable or a decay rate. See Coleman 1985, ch. 7, § 2.4, pp. 278–282 for a specific contour-defined unstable saddle.

For comparison, with the Lorentzian convention $e^{iS/\hbar}$, a real nonsingular symmetric matrix $H$, the standard orientation of $\mathbb R^{N_R}$, and the usual damping continuation, diagonalization gives

$$\int_{\mathbb R^{N_R}} \frac{\mathrm d^{N_R}\eta}{(2\pi)^{N_R/2}} \exp\!\left[ \frac{i}{2\hbar}\eta^{\mathsf T}H\eta \right] = \frac{\hbar^{N_R/2}} {\sqrt{\lvert\det H\rvert}} \exp\!\left[ \frac{i\pi}{4}\operatorname{sig}(H) \right],$$

where $\operatorname{sig}(H)=n_+-n_-$. Each diagonal factor is the one-dimensional stationary-phase model described in NIST DLMF 2026, § 2.3(iv); multiplying those factors gives the displayed signature phase under the declared hypotheses. A differently oriented real or complex cycle can select a different square-root branch. This is why a Lorentzian saddle requires a local $i0$ or contour prescription before its determinant phase is quoted.

The expansion is asymptotic at fixed regulator

A useful saddle calculation must state what its remainder means. For the positive-Hessian formula above, $O_R(\hbar^{m+1})$ is a local fixed-$R$ claim under the stated smoothness and descent assumptions. The quartic example has an explicit fixed-$R$ bound. In a formal diagrammatic calculation with no estimate, the honest statement is only that terms have been organized by powers of $\hbar$. Asymptotic smallness cannot be inferred from a few coefficients without controlling the remainder or comparing with an independent calculation; see NIST DLMF 2026, § 2.11(i).

The correct order of questions is:

1. At fixed $R$, does the chosen cycle reach an isolated saddle with a controlled transverse expansion?
2. Are omitted regions, endpoints, or other saddles smaller at the claimed order?
3. Do the coefficients and remainders remain controlled under the desired cutoff, volume, $i0$, or long-time limit?

Do not apply the nondegenerate Gaussian formula when the transverse Hessian is singular or its quadratic form is not damping on the declared local cycle. Do not call a local series a controlled approximation to the full integral when the small parameter or contributing cycle is unknown, or when no remainder and suppression statement is available. A formal power series may still organize terms in some of those cases, but it carries no demonstrated error bound or saddle-dominance claim. Likewise, fixed-$R$ control does not authorize replacing a finite determinant by a continuum functional determinant. Contour deformation through a simple saddle and local branch selection are treated in NIST DLMF 2026, § 2.4(iv). Determining a global cycle decomposition is a separate problem, handed off to Complex Saddles, Lefschetz Thimbles, and Integration Cycles.

Common pitfalls

Stationary is not synonymous with stable. The first variation tests criticality. Positivity, zero modes, and negative directions are properties of the second variation on the allowed fluctuation space.

A determinant is not the entire semiclassical answer. The action, measure density, insertion, integration cycle, boundary data, determinant branch, and remainder all contribute. With several saddles, the local determinants do not determine the global combination.

A primed determinant does not cure a zero mode. It is meaningful only after the omitted subspace has been identified and its integration replaced by a collective coordinate, gauge-fixing factor, or a different degenerate local model.

Small coupling is not automatically the semiclassical limit. In the quartic example $g$ is held fixed while $\hbar\to0$. A weak-coupling expansion can have different scaling and different control conditions.

A regulated determinant is not yet a continuum determinant. Its cutoff dependence may require ratios, local counterterms, and renormalization. Those operations must be established before the regulator is removed.

Check your understanding

Use the solutions to compare each finite-dimensional calculation with the stated saddle-point assumptions.

1. One retained scalar coordinate. Set $N_R=1$ and $K_R=k>0$ in the quartic example. Find the expansion through order $\hbar$ relative to the Gaussian result.

Solution

Here $C_R=k^{-1}$, so the general formula gives

$$\mathcal Z_R = \sqrt{\frac{\hbar}{k}} \left[ 1-\frac{g\hbar}{8k^2} +O(\hbar^2) \right].$$

Equivalently, expand the interaction and use $\langle\phi^4\rangle_0=3(\hbar/k)^2$.

2. A degenerate saddle. Determine the exact $\hbar$-scaling of $\int_{\mathbb R}\mathrm dx\,e^{-x^4/(4\hbar)}$.

Solution

Set $x=\hbar^{1/4}y$. Then

$$\int_{\mathbb R}\mathrm dx\,e^{-x^4/(4\hbar)} = \frac12\,4^{1/4}\Gamma\!\left(\frac14\right) \hbar^{1/4}.$$

The vanishing Hessian makes the ordinary Gaussian $\hbar^{1/2}$ scaling inapplicable.

3. Loop power. Show that a connected vacuum graph built from rescaled $n_v$-leg vertices carries $\hbar^{L-1}$.

Solution

Multiplying the vertex factors gives

$$\hbar^{\sum_v(n_v/2-1)} = \hbar^{\frac12\sum_v n_v-V}.$$

Every internal line has two ends, so $\sum_vn_v=2I$. The exponent is $I-V=L-1$ for a connected graph.

4. A declared Fresnel phase. With the damping prescription understood, evaluate $\int_{\mathbb R}\mathrm dx\,e^{i\lambda x^2/(2\hbar)}/\sqrt{2\pi}$ for real $\lambda\ne0$.

Solution

Analytically continue the convergent Gaussian from a positive real quadratic coefficient. The result is

$$\sqrt{\frac{\hbar}{\lvert\lambda\rvert}} \exp\!\left[ \frac{i\pi}{4}\operatorname{sgn}(\lambda) \right].$$

The sign fixes the phase only because the real orientation, the $e^{iS/\hbar}$ convention, and the damping continuation were declared.

Where to continue

• Changes of Variables and Regulated Jacobians develops finite coordinate transformations and their measure factors.
• Saddles, Control Parameters, and Loop Counting develops physical saddle sectors, competing exponentials, and their actual control parameters.
• Fluctuation Operators and Determinant Ratios treats nontrivial determinant ratios and their renormalization interface.
• Zero Modes, Collective Coordinates, and Moduli Measures develops normalized zero modes and collective-coordinate measures.
• Negative Modes and Instability Indices and Complex Saddles, Lefschetz Thimbles, and Integration Cycles develop the contour-dependent extensions that were stop conditions here.
• The Faddeev–Popov Construction treats gauge zero modes as redundancies rather than physical moduli.

References

• Coleman, Sidney. Aspects of Symmetry: Selected Erice Lectures. Cambridge University Press, 1985. doi:10.1017/CBO9780511565045.
• NIST Digital Library of Mathematical Functions. F. W. J. Olver et al., editors. Release 1.2.7, 15 June 2026. National Institute of Standards and Technology. DLMF.
• Zinn-Justin, Jean. Quantum Field Theory and Critical Phenomena. 5th ed. Oxford University Press, 2021. doi:10.1093/oso/9780198834625.001.0001.