Time-Reversed Stochastic Inflation

• Time-reversed stochastic inflation is defined by evolving the dynamics backward from the end-of-inflation, yielding a conditioned diffusion process rather than a simple time reversal.
• It provides exactly solvable models for flat, bounded, and frictional potentials, highlighting contrasts between forward and reverse curvature distributions.
• The framework has practical implications for observables like primordial black hole formation, emphasizing how tail behaviors of curvature statistics differ in reverse formulations.

Time-reversed stochastic inflation is a formulation of stochastic inflation in which the stochastic dynamics is evolved backward from the end of the quantum-diffusion era, rather than forward from an initial field value deep in the stochastic region. In this approach the viewpoint of observers attached to the end-of-inflation hypersurface is enforced, the reverse time variable is counted from the exit event, and the resulting process is a conditioned diffusion rather than a naive replacement $N\to -N$ in the usual Langevin equation (Blachier et al., 24 Apr 2025). Explicit constructions were developed first for a flat semi-infinite potential, then for a bounded flat potential (“quantum well”), and later for a semi-infinite flat potential with a constant drift term (“friction”), where the reverse formulation could be compared directly with the conventional forward stochastic-$\delta N$ approach (Blachier et al., 24 Apr 2025, Animali et al., 29 May 2026, Blachier et al., 26 Nov 2025).

1. Definition and conceptual scope

In standard slow-roll stochastic inflation, the coarse-grained inflaton obeys the Langevin equation

$\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$

with $N=\ln a$ and $\xi$ a Gaussian white noise (Blachier et al., 24 Apr 2025). The forward problem starts from an initial value $\phi_0$ at $N_0$, evolves stochastically, and characterizes either the probability density of field values or the first-passage-time distribution to the end-of-inflation boundary.

Time-reversed stochastic inflation instead conditions on the fact that the quantum-diffusion era ended. If the forward process reaches the exit boundary at random time $N_+$, the reverse time variable is defined by

$\Delta N \equiv N_+ - N,$

or, in the notation of the friction paper,

$\bar N \equiv N_f-N.$

Thus $\delta N$0 at the exit hypersurface and increases as one traces the stochastic history backward toward earlier field values (Animali et al., 29 May 2026, Blachier et al., 26 Nov 2025).

The conceptual difference is not only a change of time variable. In the forward stochastic-$\delta N$1 picture, curvature perturbations are tied to fluctuations of the total lifetime until exit,

$\delta N$2

In the reverse picture, one first conditions on a given lifetime and defines

$\delta N$3

The reverse formalism therefore computes fluctuations inside fixed-lifetime subensembles before marginalizing over all lifetimes (Blachier et al., 24 Apr 2025, Animali et al., 29 May 2026).

This suggests that time-reversed stochastic inflation is best understood as a conditional diffusion adapted to cosmological histories that actually terminate, rather than as a microscopic time-reversal symmetry of the coarse-grained inflationary dynamics.

2. Reverse stochastic dynamics and the reverse Fokker–Planck equation

The reverse formalism is built from the forward Itô diffusion

$\delta N$4

together with the forward transition kernel $\delta N$5 (Blachier et al., 24 Apr 2025, Animali et al., 29 May 2026).

The reverse transition density $\delta N$6 obeys

$\delta N$7

with reverse drift

$\delta N$8

The reverse drift therefore depends explicitly on the forward transition probability (Blachier et al., 24 Apr 2025, Animali et al., 29 May 2026).

In the flat-potential examples the forward drift vanishes, $\delta N$9, and the reverse drift is generated entirely by conditioning on the endpoint. In the semi-infinite model with an absorbing wall, the reverse drift repels trajectories away from the wall near $\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$0 and attracts them toward the prescribed initial point near the end of the reverse evolution (Blachier et al., 24 Apr 2025). In the friction model with constant forward drift, the reverse conditioned drift simplifies to

$\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$1

and all dependence on the constant forward drift drops out (Blachier et al., 26 Nov 2025).

Boundary conditions remain essential. The exit wall is absorbing in the forward process, while the reverse process is initialized at the wall and conditioned to reach the earlier field value after a specified reverse lifetime (Blachier et al., 24 Apr 2025, Animali et al., 29 May 2026). This is the structural reason the reverse process differs from a simple sign flip of the forward drift.

3. Exactly solvable models

The first explicit solution was obtained for a flat semi-infinite potential with an absorbing boundary at the quantum wall. In that case the forward process is pure Brownian motion,

$\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$2

and the first-passage-time distribution is Lévy-like. The reverse transition density was solved exactly, and the resulting curvature distribution was shown to be normalisable while exhibiting tails slowly decaying as a Lévy distribution (Blachier et al., 24 Apr 2025).

The bounded flat potential, or quantum well, introduces a second boundary: $\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$3 is absorbing, $\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$4 is reflecting, and the well width is

$\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$5

The relevant control parameter is

$\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$6

When $\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$7, the reflecting wall is irrelevant and the reverse dynamics is indistinguishable from the semi-infinite flat potential. When $\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$8, the system enters “saturated quantum diffusion,” the reverse transition density loses dependence on both $\dv{\phi}{N} = -\frac{1}{3H^2}\dv{V}{\phi} + \frac{H}{2\pi}\,\xi(N),$9 and $N=\ln a$0, and memory of the initial state is erased (Animali et al., 29 May 2026).

The friction model adds a constant drift to the semi-infinite flat potential,

$N=\ln a$1

This regularizes the conventional forward first-passage problem, which is otherwise ill-defined in the driftless unbounded case, and permits a direct quantitative comparison between forward and reverse curvature distributions. The reverse curvature PDF can again be written exactly as a one-dimensional lifetime integral, and the forward and reverse answers remain inequivalent even though both acquire exponentially decaying tails (Blachier et al., 26 Nov 2025).

A further toy model uses negative drift. In that case inflation never ends for many forward trajectories, the forward approach becomes pathological, while the reverse formalism still gives a finite curvature distribution with exponential tails (Blachier et al., 26 Nov 2025). This suggests that the reverse construction can remain meaningful in settings where the forward stochastic-$N=\ln a$2 prescription is ambiguous because no preferred background lifetime exists.

4. Curvature perturbations and tail structure

The most striking differences between forward and reverse stochastic inflation appear in the statistics of rare fluctuations. In the semi-infinite flat model, the reverse curvature distribution is normalisable but has Lévy-like tails,

$N=\ln a$3

so its variance diverges even though the PDF itself is integrable (Blachier et al., 24 Apr 2025).

In the quantum well, the fixed-lifetime reverse curvature distribution becomes a rectangle function in the saturated regime,

$N=\ln a$4

with $N=\ln a$5. After marginalization over lifetimes, the final reverse curvature PDF acquires two-sided exponential tails,

$N=\ln a$6

valid for $N=\ln a$7 (Animali et al., 29 May 2026).

The forward curvature PDF in the same bounded model has a one-sided large-positive tail,

$N=\ln a$8

and is cut off on the negative side by the lower bound on the first-passage time (Animali et al., 29 May 2026). The reverse tails therefore decay exactly twice as fast as the forward positive tail. The quantum-well analysis also shows that small fluctuations reproduce the semi-infinite result, whereas large fluctuations probe the bounded-domain regime (Animali et al., 29 May 2026).

With friction, the reverse PDF still has exponential tails, but the asymptotics differ from the forward answer. In particular, in the classical-like limit of very large drift, the tails become Gaussian only in the time-reversed picture (Blachier et al., 26 Nov 2025).

These results matter directly for tail-sensitive observables such as primordial black hole formation. The relevant papers emphasize that the reverse and forward tails can imply quantitatively different abundances, although they also stress that the physically correct probability measure for PBH counting requires further clarification (Animali et al., 29 May 2026, Blachier et al., 26 Nov 2025).

5. Relation to broader stochastic-inflation formalisms

The recent reverse-time papers emerged against a broader literature that had already isolated many of the required forward ingredients. “Path Integral for Stochastic Inflation” formulated stochastic inflation as a history integral, treated trajectories connecting specified initial and final field values, identified uphill and overshooting saddle histories, and showed that volume weighting can lead to memory loss of initial conditions through complex loitering saddles; it did not, however, formulate a backward Fokker–Planck equation or a reversed Langevin process (Gratton, 2010).

“Stochastic Inflation at NNLO” derived the forward long-wavelength generator through next-to-next-to-leading order within Soft de Sitter Effective Theory. In canonical form the generator contains not only drift and diffusion but also the first higher-derivative correction,

$N=\ln a$9

so the forward process is no longer a pure second-order diffusion once NNLO non-Gaussian noise is included (Cohen et al., 2021). This suggests that reverse-time stochastic inflation beyond leading order must generally be formulated as the reverse of a Kramers–Moyal evolution rather than of an ordinary Fokker–Planck diffusion.

“Functional renormalization group in stochastic inflation” recast the overdamped forward process into a Hermitian Euclidean Schrödinger problem,

$\xi$0

with equilibrium measure

$\xi$1

That paper did not construct a reversed diffusion, but it provided the stationary distribution, spectral gap, and unequal-time decay law $\xi$2 that control late-time memory loss (Prokopec et al., 2017). This suggests that reverse reconstruction is simplest in regimes where the forward stochastic dynamics admits such a stationary representation.

Open-EFT analyses further clarified that the forward super-Hubble dynamics is an open-system master equation with IR-finite drift, noise, and stationary probability density once the transport coefficients are defined correctly (Burgess et al., 2015). That result is directly relevant because any formal reverse process requires well-defined forward transport data.

6. Limitations, validity, and open problems

Several limitations delimit the present status of time-reversed stochastic inflation. First, the explicit reverse-time constructions are one-dimensional and rely on the standard slow-roll-type coarse-grained diffusion. Beyond slow roll, stochastic inflation must generally be formulated in phase space. “Stochastic inflation beyond slow roll” showed that the formalism remains consistent away from the slow-roll attractor provided one works with $\xi$3, uses the separate-universe approximation on super-Hubble scales, and treats the gauge of the noise carefully; it also emphasized that the stochastic source is tied to forward coarse-graining as modes cross the horizon, which makes the effective description intrinsically time-asymmetric (Pattison et al., 2019). By contrast, “Failure of the stochastic approach to inflation beyond slow-roll” argued that the standard stochastic $\xi$4 formalism is generically reliable only at zeroth order in $\xi$5 if and only if

$\xi$6

with slow roll as a special case (Cruces et al., 2018).

Second, full-gravity formulations sharpen the source of the time asymmetry. “Stochastic Inflation in General Relativity” derived gauge-invariant stochastic ADM source terms from the moving coarse-graining window,

$\xi$7

and showed that the stochastic sources vanish when $\xi$8 (Launay et al., 2024). This ties the effective irreversibility directly to the time-dependent UV/IR split. “Stochastic Inflation in Numerical Relativity” then implemented the corresponding stochastic BSSN system and noted that smooth windows render the real-space forcing effectively non-Markovian (Launay et al., 16 Dec 2025).

Third, gradient interactions become essential near attractor/non-attractor transitions. “Stochastic inflation with gradient interactions” showed that the reduced $\xi$9 process becomes non-Markovian once gradient effects are retained, but can be embedded into a higher-dimensional Markov process with auxiliary stochastic variables. It also uncovered a “pullback” effect by which gradient interactions damp the tails of first-passage-time distributions (Briaud et al., 5 Sep 2025). A plausible implication is that a rigorous reverse-time treatment in such regimes must be defined on the enlarged Markov state space rather than on the reduced field variables alone.

Finally, the Lagrangian influence-functional approach remains unsettled in interacting theory. “Revisiting Lagrangian Formulation of Stochastic inflation” rederived the influence functional, emphasized non-orthogonality between long- and short-wavelength sectors, and highlighted the absence of a consistent prescription for handling general interaction terms in the imaginary part of the effective action (Panda et al., 3 Oct 2025). This suggests that reverse path measures beyond the simplest Gaussian Markovian regime remain formally ambiguous until the forward stochastic action is itself uniquely defined.

Time-reversed stochastic inflation is therefore an explicit and rapidly developing subfield rather than a finished framework. Its exact solvable models have established that reverse and forward stochastic cosmologies need not agree in quantum-diffusion-dominated regimes, especially in the tails of the curvature distribution, while the wider literature indicates that extending the reverse formalism to non-attractor phases, gradient interactions, and full general relativity will require a correspondingly richer forward theory (Blachier et al., 24 Apr 2025, Animali et al., 29 May 2026, Blachier et al., 26 Nov 2025).