Complex Saddles, Lefschetz Thimbles, and Integration Cycles

A complex critical point contributes only when its steepest-descent cycle occurs in the homology class of the original integration cycle. The coefficients are oriented intersection numbers with dual cycles, and the Gaussian phases follow from the orientation and Hessian branch on each thimble. Complexifying the field equations is therefore necessary but not sufficient.

Required background. Laplace’s method and steepest descent supplies convergent contour deformation in a regulated integral; negative modes and instability indices supplies the relation between local Gaussian directions and contour phases.

Helpful background. Stationary phase, coalescing saddles, and Stokes phenomena supplies uniform finite-dimensional asymptotics; Wick rotation and analytic continuation supplies the distinction between an analytic continuation and an assumed Euclidean contour.

Downward cycles and their duals

Begin with a finite-dimensional regulator,

$$Z_\Gamma(g)=\int_\Gamma \Omega(z)\, \exp[-h(z;g)], \qquad h(z;g)=\frac{S(z;g)}{\epsilon},$$

where $S$ is holomorphic, $\Omega$ is a holomorphic top form, and $\Gamma$ is a convergent middle-dimensional cycle in the complexified configuration space. Suppose the critical points $z_\sigma$ are isolated and nondegenerate.

Choose a Hermitian metric and define the upward-$\operatorname{Re}h$ flow

$$\frac{dz^i}{dt} =\overline{\frac{\partial h}{\partial z^i}}.$$

Along it,

$$\frac{d}{dt}\operatorname{Re}h =\sum_i\left|\frac{\partial h}{\partial z^i}\right|^2\geq0, \qquad \frac{d}{dt}\operatorname{Im}h=0.$$

The downward integration cycle $\mathcal J_\sigma$ consists of trajectories approaching $z_\sigma$ as $t\to-\infty$ and running toward regions with $\operatorname{Re}h\to+\infty$. The dual cycle $\mathcal K_\sigma$ is defined by the opposite flow. With compatible orientations,

$$\langle\mathcal J_\sigma,\mathcal K_\tau\rangle =\delta_{\sigma\tau}, \qquad \Gamma=\sum_\sigma n_\sigma\mathcal J_\sigma, \qquad n_\sigma=\langle\Gamma,\mathcal K_\sigma\rangle\in\mathbb Z.$$

This answers the coefficient question: $n_\sigma$ is fixed by the original cycle, not by comparing critical values. The local contribution is

$$Z_\sigma \sim e^{-h_\sigma} \frac{(2\pi)^{N/2}\,\Omega(z_\sigma)} {\sqrt{\det H_\sigma}}\, \bigl[1+O(\epsilon)\bigr], \qquad (H_\sigma)_{ij} =\frac{\partial^2 h}{\partial z^i\partial z^j}\bigg|_{z_\sigma}.$$

The square-root branch is continued along the oriented thimble. Choosing a principal square root independently at each saddle generally gives inconsistent phases. Witten 2011, §§2.1–3.1, preprint pp. 5–23 develops the relative-homology construction and its intersection pairing.

The construction requires modification at a degenerate critical point, at a singularity of $\Omega$, or when flow escapes to a nonconvergent end. In a gauge theory, gauge fixing or an appropriate quotient is required before the critical set can be treated as Morse. In continuum QFT, a lattice, mode cutoff, or other regulator must define both the complexified space and the cycle before formal infinite-dimensional flow equations acquire mathematical meaning.

A quartic integral

Consider the normalized zero-dimensional quartic integral

$$Z(g)=\frac1{\sqrt{2\pi}} \int_{\mathbb R}\mathrm dx\, \exp\!\left(-\frac{x^2}{2}-\frac{g x^4}{4}\right), \qquad \operatorname{Re}g>0.$$

The critical points of

$$S(z;g)=\frac{z^2}{2}+\frac{g z^4}{4}$$

are

$$z_0=0, \qquad z_\pm=\pm\frac{i}{\sqrt g},$$

with critical values

$$S_0=0, \qquad S_\pm=-\frac1{4g}.$$

All three saddles exist for $g\neq0$, but on the positive real contour only the thimble content selected by $\mathbb R$ contributes. In particular, the exponentially large factors $e^{+1/(4g)}$ associated with $z_\pm$ for $g>0$ do not appear merely because those critical points exist.

The integral can be evaluated exactly:

$$Z(g) =\frac{e^{1/(8g)}}{2\sqrt{\pi g}}\, K_{1/4}\!\left(\frac1{8g}\right), \qquad \operatorname{Re}g>0,$$

with branches fixed by continuation from positive $g$. The formula follows from the integral representation of the modified Bessel function; see NIST DLMF, §10.32(ii). Its small-$g$ asymptotic expansion begins with $Z(g)\sim1-3g/4+\cdots$, while the other critical values set the exponential scales that can enter after analytic continuation.

For a pair of saddles $\sigma,\tau$, phase alignment occurs when

$$\operatorname{Im}(S_\sigma-S_\tau)=0.$$

For $z_0$ and $z_\pm$ this occurs on $\arg g=0$ and $\arg g=\pi$. A thimble jump requires more than this equality: a connecting flow must exist. Across a lateral continuation near a Stokes ray, the quartic integral’s convergent cycle and the Bessel-function branch can be followed on both sides. The saddle basis can jump even though the analytically continued cycle has not.

The figure shows what must be tracked. Since positive real $g$ is itself a phase-alignment ray for these critical values, its left panel freezes the nearby upper-lateral representative $g=e^{0.15i}$. The endpoint sectors and opposite local tangents were checked by integrating the flow; their drawn shapes are schematic rather than quantitative. Inspect the distinction between a critical point, a downward thimble, a dual cycle, and the original contour.

Schematic thimble map for the regulated quartic integral at $g=e^{0.15i}$, not to scale. The solid-cycle topology and local tangents are checked flow data, while the drawn curve shapes are illustrative; dashed curves are dual cycles. Intersection numbers determine which saddles contribute, and at phase alignment a connecting flow can change the thimble basis and its coefficients in compensating ways.

Why the exact cycle does not jump

Suppose crossing a Stokes ray changes a two-thimble basis by

$$\mathcal J_\tau^{+} =\mathcal J_\tau^{-} +m\,\mathcal J_\sigma^{-}, \qquad \mathcal J_\sigma^{+}=\mathcal J_\sigma^{-}, \qquad m\in\mathbb Z.$$

The coefficients transform contragrediently:

$$n_\tau^{+}=n_\tau^{-}, \qquad n_\sigma^{+}=n_\sigma^{-}-m n_\tau^{-}.$$

Therefore

$$n_\sigma^{+}\mathcal J_\sigma^{+} +n_\tau^{+}\mathcal J_\tau^{+} =n_\sigma^{-}\mathcal J_\sigma^{-} +n_\tau^{-}\mathcal J_\tau^{-} =\Gamma.$$

This algebra is the local reason an exact integral can remain analytic while its saddle representation changes. The integer $m$ and its sign come from the oriented connecting flow; they cannot be inferred from a drawing alone. Direct evaluation of the quartic integral through its Bessel representation reproduces the same lateral analytic continuations as the corresponding cycle deformations.

It is useful to distinguish two conditions:

$$\begin{aligned} \text{phase alignment:}\quad& \operatorname{Im}\frac{S_\sigma-S_\tau}{\epsilon}=0,\\ \text{equal exponential magnitude:}\quad& \operatorname{Re}\frac{S_\sigma-S_\tau}{\epsilon}=0. \end{aligned}$$

This chapter calls the first condition a Stokes condition when a connecting flow exists, and the second an anti-Stokes condition. Some literature exchanges the names; the equations remove the ambiguity.

Boundaries of the method

A complex saddle calculation is controlled only when:

• the original Lorentzian or Euclidean prescription defines $\Gamma$;
• the regulator preserves the analytic structure needed for contour deformation;
• all singularities crossed by the deformation are included;
• critical points and flows are isolated or treated by a valid Morse–Bott generalization;
• the regulator can be removed without losing convergence or changing the claimed observable.

Real-time QFT, gauge orbit spaces, fermion determinants with zeros, and infinite-volume limits can violate several conditions at once. Formal complex solutions are still useful candidates, but their contribution and phase remain unproved until a cycle or an equivalent analytic-continuation prescription is supplied.

Shared comparison. The canonical saddle comparison places this contour requirement beside the action, spectrum, determinant, renormalization, and breakdown data required of any saddle calculation.

Common pitfalls

Counting every complex solution. Existence solves only the critical-point equation. The intersection number with the original cycle can vanish.

Assigning Gaussian phases by a principal square root. The correct branch is transported continuously with the oriented thimble. Independent local branch choices can violate analyticity.

Calling every phase-alignment ray a jump. A Stokes jump also requires a connecting flow with nonzero incidence. Global topology can make the jump coefficient zero.

Exercises

1. Verify the critical points and critical values of the quartic integral.

Solution

Since

$$S'(z)=z+gz^3=z(1+gz^2),$$

the critical points are $z_0=0$ and $z_\pm=\pm i/\sqrt g$. At $z_\pm$, $z^2=-1/g$ and $z^4=1/g^2$, so

$$S(z_\pm)=-\frac1{2g}+\frac1{4g} =-\frac1{4g}.$$

2. Prove that $\operatorname{Im}h$ is constant along the thimble flow.

Solution

Using $\dot z^i=\overline{\partial_i h}$,

$$\frac{dh}{dt} =\sum_i\partial_i h\,\overline{\partial_i h} =\sum_i|\partial_i h|^2,$$

which is real and nonnegative. Hence the imaginary part is constant and the real part increases away from the saddle.

3. Check explicitly that the basis and coefficient jumps displayed above leave $\Gamma$ invariant.

Solution

Substitution gives

$$(n_\sigma^{-}-mn_\tau^{-})\mathcal J_\sigma^{-} +n_\tau^{-}(\mathcal J_\tau^{-}+m\mathcal J_\sigma^{-}) =n_\sigma^{-}\mathcal J_\sigma^{-} +n_\tau^{-}\mathcal J_\tau^{-}.$$

The two terms proportional to $m$ cancel. A truncated saddle sum that changes only the basis or only the coefficients would create a spurious discontinuity.

References

• Olver, Frank W. J., et al., eds. NIST Digital Library of Mathematical Functions, §10.32, “Integral Representations” for modified Bessel functions. National Institute of Standards and Technology. DLMF.
• Witten, Edward. “Analytic Continuation of Chern–Simons Theory.” In Chern–Simons Gauge Theory: 20 Years After, AMS/IP Studies in Advanced Mathematics 50 (2011): 347–446. arXiv:1001.2933.