---
title: Stochastic Constant-Roll Inflation
url: https://www.emergentmind.com/topics/stochastic-constant-roll-inflation
type: topic
---

# Stochastic Constant-Roll Inflation

Searching arXiv for recent papers on stochastic constant-roll inflation and closely related stochastic-inflation formalisms.
Stochastic constant-roll inflation is the stochastic treatment of inflationary dynamics in backgrounds for which the second slow-roll parameter is constant, $\epsilon_2 \equiv d\ln \epsilon_1/dN = \mathrm{const}$. In this setting, the long-wavelength inflaton evolves under classical drift and quantum diffusion, and the resulting perturbation statistics can be analyzed with Langevin and Fokker–Planck methods, or, in pure constant roll, with a closed-form nonlinear $\Delta N$ map. For a quadratic hilltop potential,
$$
V(\phi)=V_0-\tfrac12 m^2\phi^2,
$$
the growing attractor obeys
$$
\phi_{\rm cl}(N)=\phi_0 e^{A_+N},\qquad
A_+=\tfrac32\Bigl(\sqrt{1-\tfrac43\eta_V}-1\Bigr),
$$
so that $\epsilon_2=2A_+$ is constant. Recent work has shown that, once one allows trajectories to cross the hilltop and encounter a reflecting boundary on the far side, rare but very long-lived histories qualitatively alter first-passage statistics and the tail structure of the coarse-grained $\Delta N$ distribution; earlier work had already derived an exact non-Gaussian curvature-perturbation distribution in pure constant roll and connected it to primordial black-hole production, while Pattison et al. established the validity of the stochastic formalism beyond slow roll and quantified gauge corrections to the noise amplitude [2606.09690][2304.10903][1905.06300].

## 1. Foundational stochastic framework beyond slow roll

Pattison et al. formulated the stochastic dynamics of a single inflaton field beyond the slow-roll approximation in the uniform-$N$ gauge, where the number of e-folds $N\equiv \ln a$ is unperturbed. The exact phase-space Langevin system is
$$
\frac{d\phi}{dN}=\pi+\xi_\phi(N),\qquad
\frac{d\pi}{dN}=-[3-\epsilon_1(\phi,\pi)]\,\pi-\frac{V_{,\phi}(\phi)}{H^2(\phi,\pi)}+\xi_\pi(N),
$$
with $\epsilon_1\equiv \pi^2/(2M_{\rm p}^2)$ and Gaussian white noises satisfying
$$
\langle \xi_i(N_1)\xi_j(N_2)\rangle=D_{ij}(\phi,\pi)\,\delta(N_1-N_2).
$$
In many practical cases one reduces this system to a one-field Langevin equation,
$$
\frac{d\phi}{dN}=A(\phi)+\sqrt{2D(\phi)}\,\eta(N),
$$
where $A(\phi)$ is the classical drift and $D(\phi)$ is the field-only diffusion coefficient [1905.06300].

A key technical issue is that the noise amplitude is usually computed from the mode function of the Mukhanov–Sasaki variable in the spatially-flat gauge, whereas the Langevin equation is written in uniform-$N$ gauge. Pattison et al. showed that the corresponding gauge correction can be written as
$$
\widetilde{D}_{\phi\phi}=D_{\phi\phi}^{\rm flat}\,[1+\Delta D(\phi)],
\qquad
\Delta D(\phi)=2\,\mathrm{Re}\{\phi'\alpha_k/Q_k\}\big|_{k=k_\sigma},
$$
and found that for attractor backgrounds the correction satisfies $\Delta D\sim O(\epsilon,k_\sigma^2)\ll1$. They also demonstrated that the separate-universe approach remains valid on super-Hubble scales without any slow-roll assumption: each patch obeys the homogeneous Klein–Gordon equation up to $O(\nabla^2)$ corrections. This establishes the formal basis for applying stochastic inflationary methods to constant-roll regimes rather than only to slow-roll backgrounds [1905.06300].

## 2. Constant-roll realization near a quadratic hilltop

In the hilltop model emphasized by Tomberg, the local potential maximum at $\phi=0$ is characterized by
$$
\eta_V\equiv \frac{V''(0)}{V(0)}=-\frac{m^2}{V_0}<0.
$$
In an almost constant-Hubble regime, $V\simeq V_0$ and $H^2\simeq V_0/3$, the homogeneous inflaton equation becomes
$$
\ddot\phi+3H\dot\phi+V'(\phi)=0
\;\Rightarrow\;
\frac{d^2\phi}{dN^2}+3\frac{d\phi}{dN}+3\eta_V\phi=0.
$$
The growing attractor solution,
$$
\phi_{\rm cl}(N)=\phi_0 e^{A_+N},
$$
implies
$$
\epsilon_2\equiv \frac{d\ln\epsilon_1}{dN}=2A_+=\mathrm{const},
$$
which is precisely the constant-roll condition. Along this attractor, the stochastic long mode obeys
$$
\frac{d\phi}{dN}=\mu(\phi)+\sigma\,\xi(N),\qquad
\langle \xi(N)\xi(N')\rangle=\delta(N-N'),
$$
with linear drift
$$
\mu(\phi)=b\,\phi,\qquad b\equiv \tfrac12\epsilon_2=\mathrm{const}>0,
$$
and constant diffusion amplitude $\sigma^2=(H/2\pi)^2$ times a constant-roll correction factor [2606.09690].

This linear-drift, constant-diffusion specialization is the technically decisive simplification behind the later spectral analysis. It permits an exact characterization of the relevant Fokker–Planck operator in a bounded domain with one absorbing and one reflecting boundary, and it sharply separates the classical outward drift from the diffusion-driven possibility of hilltop crossing. The latter possibility is the source of the qualitative differences between beyond-hilltop stochastic solutions and treatments that restrict trajectories to a single side of the hilltop [2606.09690].

## 3. Fokker–Planck operator and spectral solution

For the one-field constant-roll system, the probability density $P(\phi,N)$ satisfies
$$
\partial_N P
=
\partial_\phi\Bigl[\tfrac12\sigma^2\,\partial_\phi P-\mu(\phi)\,P\Bigr].
$$
Tomberg imposed an absorbing boundary at $\phi=\phi_a$, interpreted as the end-of-inflation field value, and a reflecting boundary at $\phi=\phi_r$, mimicking a steep classical slope or second minimum. The reflecting condition is a vanishing probability current,
$$
j(\phi_r,N)=0
\quad\Rightarrow\quad
\Bigl[\tfrac12\sigma^2\partial_\phi P-\mu P\Bigr]_{\phi_r}=0,
$$
while the absorbing condition is
$$
P(\phi_a,N)=0.
$$
The solution is sought in separated form,
$$
P(\phi,N)=\sum_n c_n e^{-\lambda_n N}\psi_n(\phi),
$$
where the eigenfunctions satisfy
$$
L\psi_n(\phi)=-\lambda_n\psi_n(\phi),
\qquad
L=\partial_\phi\Bigl[\tfrac12\sigma^2\partial_\phi-\mu(\phi)\Bigr].
$$
Introducing the weight function
$$
w(\phi)=\sigma^2\exp\Bigl[-\int^\phi \frac{2\mu(\phi')}{\sigma^2}d\phi'\Bigr]
=\sigma^2 e^{-b\phi^2/\sigma^2},
$$
and setting $A=\sqrt{w}$, the operator can be rewritten in self-adjoint form $L_S=A^{-1}LA$. Sturm–Liouville theory then guarantees a discrete spectrum
$$
0\le \lambda_1<\lambda_2<\cdots,
$$
with orthonormal eigenfunctions $v_n(\phi)=\sqrt{w(\phi)}\,\psi_n(\phi)$ and completeness relation
$$
\sum_n v_n(\phi)v_n(\phi')=\delta(\phi-\phi').
$$
In practice, Tomberg solved the eigenvalue problem by a shooting method, requiring exactly $n-1$ nodes in $(\phi_a,\phi_r)$ while enforcing the two boundary conditions [2606.09690].

The methodological significance is that the spectral method yields the full late-time probability density and, through the same eigensystem, the full first-passage-time statistics. This contrasts with approximations that focus only on low moments or assume that the large-$\Delta N$ tail remains purely exponential all the way to arbitrarily large $\Delta N$ [2606.09690].

## 4. First-passage statistics and the failure of the mean background

The first-passage-time distribution through the absorbing boundary is given by the boundary current,
$$
\mathrm{FPT}(N,\phi_0)=\pm j(\phi_a,N|\phi_0),
$$
with spectral decomposition
$$
\mathrm{FPT}(N,\phi_0)
=
\sum_{n=1}^{\infty}\bar u_n(\phi_0)\,j_n(\phi_a)\,e^{-\lambda_n N},
$$
where
$$
j_n(\phi)\equiv \Bigl[-\tfrac12\sigma^2\partial_\phi+\mu\Bigr]\psi_n(\phi).
$$
The mean first-passage time is
$$
\langle N\rangle_{\phi_0}
=
\int_0^\infty dN\,N\,\mathrm{FPT}(N,\phi_0)
=
\sum_n \frac{\bar u_n(\phi_0)\,j_n(\phi_a)}{\lambda_n^2}.
$$
In the wide-boundary limit, $\phi_r\gg1$ and $|\phi_a|\gg1$ in suitable units, the smallest eigenvalue satisfies $\lambda_1\to0$, so the mean diverges as $\sim 1/\lambda_1^2$. Tomberg showed that this divergence is dominated by extremely rare trajectories that cross the hilltop, get stuck near the reflecting boundary on the other side, and then tunnel out slowly in a way dictated by the lowest eigensolution. Although rare, these trajectories dominate the global average, so the mean does not properly describe the inflationary background [2606.09690].

Tomberg therefore proposed using the median rather than the mean. The median $N_{\rm med}$ satisfies
$$
\int_{N_{\rm med}}^\infty dN\,\mathrm{FPT}(N,\phi_0)=\tfrac12,
$$
or equivalently
$$
\sum_n \frac{\bar u_n(\phi_0)\,j_n(\phi_a)}{\lambda_n}\,
e^{-\lambda_n N_{\rm med}}
=
\tfrac12.
$$
At moderate barrier heights this relation is dominated by the first few eigenterms, and for initial $\phi_0$ on the inflationary slope, $\phi_0<0$, the median tracks the classical value
$$
N_{\rm cl}=\ln(\phi_a/\phi_0)
$$
almost exactly. A common simplification is to identify the background e-fold number with the mean; in the beyond-hilltop stochastic problem, Tomberg argued that the median is the more robust background measure for the $\Delta N$ formalism because it reflects the bulk of realizations rather than rare outliers [2606.09690].

## 5. Structure of the coarse-grained $\Delta N$ distribution

In the coarse-grained $\Delta N_c$ formalism, one selects a coarse-graining time $N_c$, reads off the stochastic field value $\phi_c$, and adds the remaining first-passage time $N_{\rm av}(\phi_c)$ to the absorbing boundary. Defining a single background by the median prescription,
$$
\Delta N_c=N_c+N_{\rm av}(\phi_c)-N_{\rm bg},
$$
one obtains a one-to-one map $\phi_c\mapsto \Delta N_c$. For $\phi_c<0$, the average remaining time is approximately classical,
$$
N_{\rm av}\simeq N_{\rm cl}(\phi_c)=\ln(\phi_a/\phi_c),
$$
and the probability density $P(\phi_c,N_c|\phi_0)$ is nearly Gaussian of width $\Omega(N_c)$. The associated small-$\Delta N_c$ distribution is the modified Gaussian
$$
P_G(\Delta N_c)
=
\frac{1}{\sqrt{2\pi}\,\Omega}\,
\exp\!\Bigl[-\tfrac{(1-e^{-\Delta N_c})^2}{2\Omega^2}-\Delta N_c\Bigr]
\;\xrightarrow{\Delta N_c\gg1}\;
\exp[-\Delta N_c].
$$
This gives a nearly Gaussian core and an exponential tail [2606.09690].

For $\phi_c>0$, the behavior changes qualitatively. The median remaining time saturates at a constant,
$$
N_{\max}\simeq (\ln 2)/\lambda_1<\infty,
$$
which is the largest allowed $\Delta N_c$. The distribution then piles up into a skewed plateau and finally a sharp peak at $\Delta N_c\simeq N_{\max}$. Quantitatively, Tomberg found a plateau height of order
$$
P_{\rm trans}\sim P_G(\phi_c=0)\,\bigl|d\phi_c/dN_{\rm av}\bigr|,
$$
and ultimately a $\delta$-function-like spike at $\Delta N_c=N_{\max}$. On a log plot, the full $P(\Delta N_c)$ displays a Gaussian peak around small $\Delta N_c$, an intermediate approximate exponential tail $\sim e^{-\lambda_2\Delta N_c}$ and then $e^{-\lambda_1\Delta N_c}$, a flattening to a plateau once the first tunneling mode overtakes the classical modes, and a final spike near $\Delta N_{c,\max}=\ln 2/\lambda_1$ [2606.09690].

These results directly modify the standard expectation that stochastic inflation near a hilltop produces only a Gaussian core plus a single exponential tail. Tomberg argued that similar intricacies should arise in primordial-black-hole models with a shallow secondary minimum, because patches that cross into a diffusion-dominated well can inflate for almost exactly a maximal duration before escaping [2606.09690].

## 6. Curvature perturbations, deterministic $\Delta N$, and primordial black holes

Tomberg’s earlier treatment of stochastic constant-roll inflation and primordial black holes solved the system along the classical constant-roll trajectory by drawing stochastic kicks from a numerically computed power spectrum, beyond the usual de Sitter approximation. The integrated curvature power spectrum is
$$
\sigma_k^2\equiv \int_{N_{\rm ini}}^{N(k)} P(N',k_\sigma)\,dN'
=
\int_{k_{\rm ini}}^{k} P(k')\,d\ln k',
$$
and in pure constant roll, with $\epsilon_2=\mathrm{const}$ and $\epsilon_1\ll1$,
$$
\phi_{\rm cl}(N)=\phi_0\,e^{\frac{\epsilon_2}{2}N}.
$$
Writing the stochastic solution as
$$
\phi(N)=\phi_{\rm cl}(N)\Bigl[1-\tfrac{\epsilon_2}{2}X(N)\Bigr],
$$
one finds that $X$ is Gaussian with
$$
p_X(X)=\frac{1}{\sqrt{2\pi}\,\sigma_k}
\exp\!\Bigl[-\tfrac{X^2}{2\sigma_k^2}\Bigr].
$$
The nonlinear $\Delta N$ formula gives the coarse-grained curvature perturbation
$$
\zeta\equiv \Delta N
=
-\frac{2}{\epsilon_2}\,
\ln\!\Bigl[1-\tfrac{\epsilon_2}{2}X\Bigr],
$$
which implies
$$
X(\zeta)=\frac{2}{\epsilon_2}\Bigl[1-e^{-\tfrac{\epsilon_2}{2}\zeta}\Bigr]
$$
and therefore
$$
p(\zeta)=
\frac{\epsilon_2}{2\sqrt{2\pi}\,\sigma_k}\,
e^{-\tfrac{\epsilon_2}{2}\zeta}\,
\exp\!\Bigl[-\tfrac{2}{\sigma_k^2\epsilon_2^2}
\bigl(1-e^{-\tfrac{\epsilon_2}{2}\zeta}\bigr)^2\Bigr].
$$
For $\zeta\ll1$ this is approximately Gaussian with variance $\sigma_k^2$, whereas for $\zeta\gg1$ it develops an exponential tail,
$$
p(\zeta)\sim e^{-\tfrac{\epsilon_2}{2}\zeta}.
$$
In this regime, the perturbation distribution is controlled by the two parameters $\{\sigma_k^2,\epsilon_2\}$ [2304.10903].

The same work connected the non-Gaussian tail to primordial-black-hole formation. With collapse threshold $\zeta_c\approx1$, the PBH mass at scale $k$ is
$$
M_k\approx 6.8\times10^{49}\,{\rm g}\,
\Bigl(\tfrac{k}{0.05\,{\rm Mpc}^{-1}}\Bigr)^{-2},
$$
and the collapsing fraction is
$$
\beta_k
=
2\int_{\zeta_c}^{\infty} p(\zeta)\,d\zeta
\approx
\frac{\sigma_k\epsilon_2}{\sqrt{2\pi}\bigl(1-e^{-\tfrac{\epsilon_2}{2}\zeta_c}\bigr)}
\exp\!\Bigl[
-\frac{2}{\sigma_k^2\epsilon_2^2}
\bigl(1-e^{-\tfrac{\epsilon_2}{2}\zeta_c}\bigr)^2
\Bigr].
$$
Numerically, for asteroid-mass PBHs with $M\sim 10^{19}\,{\rm g}$, $\beta\sim10^{-17}$, and $\zeta_c\approx1$, pure slow roll requires $P_{\rm peak}\sim10^{-2}$, whereas for $\epsilon_2\simeq1\ldots4$ one can reduce this to
$$
P_{\rm peak}\sim(1.6\ldots5)\times10^{-3},
$$
an ${\cal O}(10)$-fold reduction. Tomberg also showed that a non-stochastic $\Delta N$ treatment with a single initial Gaussian field kick reproduces precisely the same $p(\zeta)$ in the pure constant-roll regime, but that the stochastic framework clarifies that the kicks actually arrive continuously over the constant-roll phase and that outside pure constant roll the deterministic shortcut no longer applies [2304.10903].

Source: https://www.emergentmind.com/topics/stochastic-constant-roll-inflation