Kinetic, Secular, and Resummation Regimes

A perturbative correction can be small in coupling yet grow with elapsed time, number of e-folds, or another secular variable until it becomes order one. That diagnoses the failure of a fixed-order truncation, but it does not by itself select a unique late-time description. Kinetic equations, dynamical renormalization, large-N or two-particle-irreducible approximations, and stochastic coarse graining resum different structures under different assumptions.

Required background. Flat-Spacetime Quantum Field Theory supplies loop expansions and scattering-based kinetic intuition; Partial Differential Equations and Green Functions provides initial-value evolution and memory kernels.

Helpful background. Causality and Global Structure establishes the causal domains over which correlations accumulate; Quantum Fields in Curved Spacetime supplies state dependence, Hadamard ultraviolet control, and prescribed-background limitations.

The minimum local bridge is to distinguish two times. A microscopic correlation time controls how rapidly a memory kernel decays, while a macroscopic relaxation or expansion time controls the observable. Replacing an integral over the past by an instantaneous collision term is justified only when these scales are parametrically separated.

When a small loop becomes order one

Suppose an observable has the formal expansion

$$F(t)=F_0(t)\left[1+\lambda L(t)+O(\lambda^2L(t)^2,\lambda)\right],$$

with $\lambda\ll1$ but $L(t)$ growing. Fixed order is controlled by $|\lambda L(t)|\ll1$, not merely by $\lambda\ll1$. A secular time $t_{\mathrm{sec}}$ is defined parametrically by

$$|\lambda L(t_{\mathrm{sec}})|\sim1.$$

The expression does not show that the exact observable diverges. It shows that terms nominally higher order may be as important as the retained term. A resummation is credible only if it reproduces the known early-time coefficients and controls the terms it neglects.

In a quartically interacting massless scalar model on exact de Sitter space, a one-loop two-point function contains a term proportional to $\lambda\ln(\mu L)\ln(-k\tau)$. The time-dependent logarithm counts e-folds after horizon exit. Burgess and collaborators identify the two infrared logarithms and the fixed-order limitation in Burgess et al. 2010, §§2.1–2.4. This is a concrete model result, not a theorem that every curved-spacetime secular term exponentiates.

Kinetic closure and its hierarchy

Nonequilibrium two-point functions can be written in statistical and spectral combinations. In a convex normal neighborhood, with the necessary parallel transport to one tangent space left implicit, introduce a center coordinate $X$ and separation $s$. A local Wigner transform takes the schematic form

$$F(X,p)=\int_{T_XM}d^4s\,e^{ip_as^a} F\!\left(\exp_X\frac{s}{2},\exp_X\!\left(-\frac{s}{2}\right)\right).$$

A gradient expansion of the exact Kadanoff–Baym equations can reduce them to a kinetic equation such as

$$p^\mu\nabla_\mu f(X,p)=C[f](X,p).$$

Here $p$ is a momentum covector at $X$ in the Wigner transform; the exponential-map construction is restricted to a normal neighborhood, with a smooth localization when needed. In the on-shell kinetic equation, $\nabla_\mu$ means the horizontal phase-space derivative, not the derivative of a spacetime scalar with its momentum components held fixed. In coordinates using contravariant spatial momenta, free streaming is

$$\mathcal Lf=p^\mu\partial_{x^\mu}f -\Gamma^i{}_{\alpha\beta}p^\alpha p^\beta \frac{\partial f}{\partial p^i}.$$

The second term changes the momentum along a geodesic. For a homogeneous isotropic FLRW distribution written in terms of the locally measured momentum magnitude $p$, this becomes $\partial_t f-Hp\,\partial_p f=C/E$. A collisionless distribution $f(t,p)=F(ap)$ solves it exactly. Omitting the momentum derivative would erase cosmological redshift.

This compact equation hides its assumptions. A quasiparticle description requires a narrow spectral width, schematically $\Gamma/E\ll1$. A first-gradient truncation needs the background and distribution to vary slowly across microscopic correlation lengths. A Markov collision term requires the memory time to be much shorter than the relaxation time. Weak coupling alone does not guarantee any of these conditions, especially near thresholds, critical points, horizons, or strongly infrared-enhanced states.

Conserving approximations add another constraint. A truncation derived consistently from a symmetry-respecting two-particle-irreducible functional can preserve specified conservation laws, whereas inserting a width in propagators while leaving vertices unmatched can violate Ward identities or energy balance. On a nonstationary prescribed geometry, “energy conservation” must itself be stated carefully; local covariant stress-energy conservation is the relevant check when the renormalized equations permit it.

Dynamical RG and stochastic resummation

Dynamical renormalization introduces an arbitrary intermediate time, absorbs the secular part into running initial data, and requires the prediction to be independent of that time. If the resulting leading evolution is multiplicative, $d\log F/dt=\epsilon f'(t)$, then a correction $F=F_0[1+\epsilon f(t)+O(\epsilon^2)]$ can be promoted to

$$F_{\mathrm{DRG}}(t)=F_0\exp[\epsilon f(t)]$$

with the reference-time value of $f$ absorbed in $F_0$. Exponentiation is a consequence of that evolution equation, not a consequence of seeing one growing term: $e^x$ and $(1-x)^{-1}$ both start with $1+x$ but already disagree at order $x^2$. The de Sitter application is developed in Burgess et al. 2010, §§3.1–3.3. Its controlled statement is leading-logarithmic, with subleading terms requiring further information.

Stochastic inflation instead splits long and short wavelength sectors and models the transfer of modes across a coarse-graining scale as noise. It can resum a different class of infrared effects when expansion, state, field content, and observable support that split. Agreement between a stochastic equilibrium distribution, a large-N gap equation, and a dynamical-RG exponent is strong evidence in an overlap regime; disagreement need not mean one method is algebraically wrong, because the coarse graining and observables may differ.

Burgess et al. 2010, §§4.1–4.3 and §6 compare the resummed scalar behavior with massive and large-N calculations: the quartic examples soften the infrared behavior in a mass-like way, while the cubic example does not support the same self-regulation. That adverse comparison is precisely why the conclusion must remain field- and approximation-dependent.

Results and limitations through 2026-08-19

The literature assessment below is current through 2026-08-19. Each row states the assumptions needed to interpret the result; none licenses a universal late-time claim.

Evidence or position	Assumptions	Supported result	Scope, limitation, or adverse case
Calzetta–Hu closed-time-path examples, 1987	Prescribed cosmological geometries, specified initial state, loop truncation, renormalized in-in functional	Real and causal expectation equations can describe dissipation and backreaction observables within the chosen semiclassical setup	Does not establish a universal Markovian kinetic equation, all-order control, or a unique resummation
Burgess et al. fixed-order de Sitter calculation, §§2.1–2.4	Weakly coupled massless scalar, exact de Sitter, selected state, specified infrared scale and two-point observable	Exhibits an explicit loop-induced secular logarithm and the time at which fixed order loses uniform validity	Infrared completion is model-dependent; a secular term in this correlator need not be a secular term in every gauge-invariant observable
Burgess et al. dynamical RG, §§3.1–3.3	Small leading-log coefficient, identifiable running integration constant, time-independent infrared scale in the worked application	Resums the leading secular combination and matches the known early-time expansion	Does not control unenhanced higher orders or automatically extend to de Sitter-breaking initial data and memory-dominated evolution
Massive and large-N comparisons, §§4.1–4.3	Particular scalar interactions and controlled comparison limits	Quartic resummation shows mass-like infrared softening and agrees qualitatively with independent large-N structure	The cubic adverse case does not generate the same mass-like regulation, so the mechanism is not generic

Calzetta and Hu’s curved-spacetime treatment explicitly contrasts in-in causal equations with in-out boundary data Calzetta–Hu 1987, pp. 495–509. It supports the formal setting for nonequilibrium evolution, but it should not be read as proving every later kinetic or stochastic closure.

Competing positions and decisive discriminators

One position treats leading secular logarithms as the central data and uses dynamical RG to extend the perturbative answer. A second position derives a kinetic equation after quasiparticle, gradient, and Markov approximations. A third uses stochastic coarse graining for long-wavelength fields in an expanding geometry. These are credible competing accounts only after the field content, state class, geometry, boundary condition, regulator, approximation, algebra or representation, and observable are aligned.

Useful discriminators are concrete. Competing approximations should reproduce the same fixed-order coefficient in their overlap region, preserve the applicable Ward identities and conservation equations, show regulator independence after allowed counterterms, and agree on a state- and observable-specific late-time limit. A kinetic closure fails decisively if the spectral peak is broad or the memory kernel decays on the relaxation timescale. A stochastic description fails if its coarse-graining predictions depend uncontrollably on the arbitrary split. A resummed gauge-dependent quantity does not establish growth of a physical observable.

The unresolved questions are correspondingly conditional. It remains open how broadly the major resummations agree for interacting fields on generic nonstationary curved spacetimes, which state classes admit uniform late-time control, and when self-consistent metric backreaction changes the secular variable before fixed-background resummation becomes valid. Evidence does not establish one universal late-time attractor across these regimes.

Memory and infrared failure modes

The strongest adverse case for a local kinetic equation is long memory. If a self-energy kernel $\Sigma_R(t,t')$ has substantial support over the entire evolution, replacing

$$\int_{t_0}^{t}dt'\,\Sigma_R(t,t')F(t')$$

by an instantaneous rate times $F(t)$ erases phase information at leading order. Infrared enhancement can produce precisely this behavior. Boundaries and non-Hadamard initial states may add persistent image or initial-surface terms, while strong expansion can prevent a stationary quasiparticle mass shell from forming.

Resummation also cannot repair a poor observable. A coordinate-dependent correlator may grow while a relational or gauge-invariant quantity remains bounded. Conversely, a bounded equal-time two-point function does not prove that stress-energy, entropy production, or detector response is controlled. The observable must be declared before deciding what “secular” means.

Solved diagnostic: early-time matching and conservation

Consider two occupations obeying $\dot n_1=-\epsilon\gamma(n_1-n_2)$ and $\dot n_2=+\epsilon\gamma(n_1-n_2)$, with $\epsilon\gamma>0$. Find the exact solution, identify the secular breakdown of first order, and check the conserved quantity.

Solution

The sum $N=n_1+n_2$ has zero derivative, while the difference $d=n_1-n_2$ satisfies $\dot d=-2\epsilon\gamma d$. Therefore $d(t)=d_0e^{-2\epsilon\gamma(t-t_0)}$ and $n_{1,2}=(N\pm d)/2$. Expanding the exponential gives $d=d_0[1-2\epsilon\gamma(t-t_0)+O(\epsilon^2\gamma^2(t-t_0)^2)]$, so fixed order fails when $2\epsilon\gamma(t-t_0)$ is order one. The exponential resummation matches the early coefficient, preserves $N$ exactly, and relaxes the difference. A proposed resummation that changed $N$ would fail an independent conservation check even if it removed the secular term.

What could change these conclusions

The conclusions could change if a theorem proves uniform late-time control for a broader state class or geometry, if an independent calculation or simulation reproduces or overturns a leading secular coefficient, if a counterexample exposes persistent memory or gauge dependence in a claimed resummation, or if a boundary, regulator, field-content, or backreaction analysis changes the overlap conditions in the comparison table. Any such event must be assessed for the same observable and approximation regime before changing the conclusions.

References

• Burgess, C. P., R. Holman, L. Leblond, and S. Shandera. “Breakdown of Semiclassical Methods in de Sitter Space.” Journal of Cosmology and Astroparticle Physics 2010, no. 03 (2010): 033. DOI.
• Calzetta, Esteban, and B. L. Hu. “Closed-Time-Path Functional Formalism in Curved Spacetime: Application to Cosmological Back-Reaction Problems.” Physical Review D 35 (1987): 495–509. DOI.