Stochastic Resetting: Fundamentals & Applications

• Stochastic resetting is the process of interrupting and reinitializing a stochastic dynamic, transforming unbounded diffusion into a localized nonequilibrium stationary state.
• It optimizes first-passage times by introducing an optimal reset rate that minimizes mean completion times under varying reset-time distributions.
• The concept extends to complex systems—from non-Poissonian and state-dependent resets to quantum and many-body scenarios—while balancing reset costs and protocol tradeoffs.

Stochastic resetting is the interruption of a stochastic process at random times and its reinitialization to a prescribed reference state, usually the initial condition. The reset times are drawn from a waiting-time law $\psi(\tau)$, so the dynamics acquires a renewal structure. In its canonical form, resetting converts unbounded diffusion into a localized nonequilibrium stationary state and can render a previously divergent mean first-passage time finite, with a nontrivial optimal reset rate. Subsequent work has extended the concept to non-Poissonian protocols, state-dependent and partial resets, interacting many-body systems, multiplicative processes, non-Markovian dynamics, and quantum channels, while also clarifying its limitations, costs, and application-specific tradeoffs (Evans et al., 2019, Gupta et al., 2021).

1. Canonical formulation and renewal structure

For overdamped Brownian motion in one dimension, stochastic resetting is defined by

$$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$

with Gaussian white noise satisfying $\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$. The corresponding Fokker–Planck equation is

$$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$

and the Poissonian renewal representation reads

$$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$

where $G_0$ is the bare propagator (Evans et al., 2019).

This renewal structure is the defining analytic feature of stochastic resetting. It separates trajectories with no reset from those whose last reset occurred $\tau$ time units earlier, and it generalizes beyond ordinary diffusion to arbitrary propagators and higher-dimensional settings. For diffusion on the line, the stationary state is

$$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$

a Laplace-shaped nonequilibrium steady state localized around the reset point. The cusp at the reset point is a robust signature of resetting; for Gaussian diffusion with a reset-time distribution possessing finite mean and variance, the leading stationary profile remains Laplace and the cusp is independent of the specific details of the reset-time law (Evans et al., 2019, Singh et al., 2022).

The generalization from Poissonian to arbitrary waiting-time distributions changes the existence conditions for a stationary state. For diffusion, the mean squared displacement saturates to a constant only if the reset-time distribution has both finite mean and finite variance; equivalently, broad reset-time laws with divergent variance do not produce a true stationary diffusive state. For subdiffusion, the criterion is weaker: a steady state may exist even when the reset-time distribution possesses only finite mean (Singh et al., 2022). This protocol sensitivity is one of the central distinctions between resetting in normal and anomalous transport.

2. First-passage optimization and protocol dependence

Resetting is especially consequential for first-passage and completion problems. For a target at the origin and reset to the starting point $x_0>0$, the mean first-passage time under Poissonian resetting becomes

$$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$

which diverges as $$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$0 and as $$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$1, and therefore has a unique minimum at an optimal resetting rate $$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$2. Writing $$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$3, the optimality condition is

$$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$4

More generally, for a bare completion-time distribution $$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$5, Poissonian resetting gives

$$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$6

and at the optimum the coefficient of variation satisfies

$$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$7

a widely used universality criterion (Evans et al., 2019, Gupta et al., 2021).

A related sufficient condition for the utility of small-rate resetting is that the coefficient of variation of the unbiased first-passage-time distribution exceed unity:

$$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$8

This criterion has been used explicitly in molecular and escape problems to diagnose when restarting can reduce the mean completion time (Blumer et al., 2022).

Protocol dependence becomes particularly sharp when the target location is itself random. For a target at random distance $$x(t+dt)= \begin{cases} x_0 & \text{with prob. } r\,dt,\ x(t)+\eta(t)\,dt & \text{with prob. } 1-r\,dt, \end{cases}$$9 with distribution $\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$0 and reset intervals drawn from a general waiting-time distribution $\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$1, the averaged mean first-passage time can be functionally optimized with respect to $\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$2. The associated conjugate target distribution for a given protocol is

$\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$3

For one-dimensional diffusion, Poissonian resetting has an exponential conjugate target distribution,

$\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$4

whereas sharp restart $\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$5 has

$\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$6

with a Gaussian-like tail. Accordingly, sharp restart is optimal for a fixed or narrowly distributed target, but stochastic resetting prevails over sharp restart for target distributions with exponential or heavier tails (Evans et al., 2024). A common misconception is therefore that deterministic restart is uniformly superior; the comparison depends on the spatial statistics of the target.

3. State-dependent, partial, and generalized resetting

Classical resetting to a fixed point at externally prescribed times is only one corner of the theory. A recent extension is energy-based stochastic resetting in the escape dynamics of the noisy Hénon–Heiles Hamiltonian. In that system, the escape threshold is $\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$7, the initial energy is $\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$8, and noise-enhanced stability occurs at $\langle \eta(t)\eta(t')\rangle=2D\delta(t-t')$9, where the mean escape time peaks at $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$0. Standard sharp restart at fixed time $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$1 becomes beneficial only when $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$2 exceeds about $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$3, with optimum $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$4 and reduced mean escape time $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$5, about $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$6 below the no-reset value. The proposed energy-triggered protocol resets when $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$7; its optimum is $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$8, giving $$\frac{\partial P_r(x,t)}{\partial t} = D\frac{\partial^2 P_r(x,t)}{\partial x^2} -rP_r(x,t)+r\delta(x-x_0),$$9, about $$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$0 below the no-reset value and slightly below the noiseless baseline $$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$1. It also truncates the survival tail much more strongly, with $$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$2 under energy-based resetting versus $$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$3 without resetting. The mechanism is selective removal of trajectories whose energy falls rapidly and that would otherwise linger near KAM structures or the chaotic saddle; in this sense, chaotic dynamics catalyzes the usefulness of resetting (Cantisán et al., 2024).

Resetting need not return the system to a single global point. In overdamped diffusion on a mirror-asymmetric periodic potential, a particle can instead be reset to the bottom of the potential well it currently occupies. This degenerate resetting point produces a genuine ratchet current: the long-time drift velocity becomes nonzero, reaches a maximum at an optimal average resetting time $$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$4, and vanishes both for very slow and very fast resetting. The mechanism is not minimization of escape time from a well, but rectification through asymmetric splitting probabilities $$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$5 and $$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$6 that are biased by local resetting (Ghosh et al., 2023).

Reset amplitudes can also be random. In random-amplitude stochastic resetting, one distinguishes an independent additive-jump version,

$$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$7

and a dependent multiplicative version,

$$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$8

The dependent case preserves positivity and contains classical resetting as the limit $$p(x,t|x_0)=e^{-rt}G_0(x,t|x_0)+r\int_0^t d\tau\,e^{-r\tau}G_0(x,\tau|X_r),$$9. For ballistic motion with Poissonian multiplicative resets and uniform $G_0$0 on $G_0$1, the mean and variance saturate and the stationary density is

$G_0$2

whereas the independent additive-jump case remains nonstationary under Poisson resets (Dahlenburg et al., 2021). These constructions show that partial resetting and overshooting are intrinsic generalizations rather than perturbations of the classical model.

4. Multiplicative, interacting, and networked systems

In multiplicative processes, resetting can stabilize distributions without restoring ergodicity. Geometric Brownian motion with Poissonian resetting is the canonical example:

$G_0$3

Its $G_0$4-th moment changes behavior at the threshold

$G_0$5

For $G_0$6, the $G_0$7-th moment converges; for $G_0$8, it grows linearly; for $G_0$9, it diverges exponentially. The stationary density has a power-law right tail,

$\tau$0

and the long-time behavior separates into three regimes: quenched or frozen for $\tau$1, unstable annealed for $\tau$2, and stable annealed for $\tau$3. The process is stationary yet non-ergodic, and the critical self-averaging time is minimized at $\tau$4 (Stojkoski et al., 2021).

Resetting also stabilizes the reallocating geometric Brownian motion, which for negative reallocation rate $\tau$5 is non-stationary, non-ergodic, and mean-repulsive. With Poissonian resetting to $\tau$6,

$\tau$7

the mean becomes stationary for

$\tau$8

while full stationarity of the first two moments requires

$\tau$9

The critical resetting rate

$$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$0

marks the regime in which the self-averaging time effectively diverges, $$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$1, and the process mimics ergodic behavior at the level of its first two moments (Jolakoski et al., 2024).

Many-body resetting can be genuinely correlated rather than factorized over degrees of freedom. In a network growth model with node deletion, a growth event at rate $$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$2 adds a link between two uniformly chosen nodes, while a resetting event at rate $$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$3 removes a uniformly chosen node and inserts a new isolated node. The degree distribution obeys

$$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$4

which has an exact full-time solution and a stationary state for any nonzero $$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$5. The model exhibits a time-dependent percolation-like phase transition, and above

$$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$6

a giant component never appears (Artime, 2022). Reviews of interacting particle systems further distinguish global resetting, in which all particles reset collectively, from local resetting, in which particles reset independently; the interplay between interaction timescales and resetting produces nonequilibrium stationary states, altered currents, and reset-induced phase structure (Nagar et al., 2023).

5. Memory, non-Markovianity, and quantum resetting

The effect of resetting on memory is two-sided. In classical non-Markovian escape from a harmonic well described by a generalized Langevin equation with power-law memory kernel

$$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$7

Poissonian resetting to a fixed point $$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$8 yields a renewal expression for the propagator and survival probability. In that setting, resetting lowers the survival probability and often reduces the mean first-passage time, with an optimal reset rate analogous to the Markovian case. More distinctively, repeated resets disrupt temporal correlations: the first-passage distribution narrows and tends toward near-exponential decay at moderate and high reset rates, which the authors interpret as effective memory loss or induced Markovianity. The effect is not instantaneous; early-time statistics remain non-exponential because the process has not yet experienced enough resets (Saha et al., 15 Sep 2025).

In quantum systems, stochastic resetting can instead generate non-Markovianity. For discrete-time open-system evolution

$$p^*(x)=\frac{\alpha_0}{2}e^{-\alpha_0|x-X_r|},\qquad \alpha_0=\sqrt{\frac{r}{D}},$$9

with resets to $x_0>0$0 after $x_0>0$1 steps with probability $x_0>0$2, the averaged dynamics is renewal-like:

$x_0>0$3

For constant reset probability, the map reduces to repeated application of a single CPTP channel,

$x_0>0$4

and is Markovian. For time-dependent reset schedules, the resulting maps are generically non-divisible. In non-classical scenarios where the effective reset probability can become negative, the survival factor $x_0>0$5 may increase, producing revivals in trace-distance distinguishability and thereby a witness of quantum non-Markovianity. Carefully chosen time-dependent reset sequences can also stabilize stronger stationary quantum correlations than any constant-rate protocol in the two-qubit example studied (Carollo et al., 19 Jan 2026). Thus classical resetting can erase memory in non-Markovian escape, whereas quantum time-dependent resetting can itself be a source of memory effects.

6. Applications, algorithmic uses, and limitations

Resetting has become a methodological tool in molecular simulation. In molecular dynamics, restarting trajectories from the original initial condition at random times can accelerate rare-event sampling by up to an order of magnitude in examples ranging from simple model systems to alanine dipeptide, and a key exact relation allows the mean first-passage time at any larger reset rate $x_0>0$6 to be inferred from simulations at a single reset rate $x_0>0$7:

$x_0>0$8

The method is collective-variable-free, easy to implement, highly parallelizable, and can be combined with Metadynamics or OPES; adaptive resetting protocols of the form

$x_0>0$9

further restrict resets to undesirable regions of configuration space (Blumer et al., 2022, Blumer et al., 8 Apr 2025).

The same logic has been imported into optimization and learning. For noisy-label deep learning, parameters are reset to a checkpoint $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$0 with probability $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$1,

$$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$2

which repeatedly returns training to a parameter region that has learned general patterns but not yet memorized corrupted labels. Reported gains include CIFAR-10 with cross-entropy and noise rate $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$3, where test accuracy improves from $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$4 to $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$5, and CIFAR-100 under the same setting, where it improves from $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$6 to $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$7 (Bae et al., 2024). In reinforcement learning, resetting returns the agent to the start state with probability $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$8 at each training step while preserving learned parameters. In tabular grids and MountainCar with DQN, this accelerates policy convergence even when it does not reduce the search time of a purely diffusive agent, because it truncates long, uninformative trajectories and speeds value propagation, unlike temporal discounting, which changes the optimal policy (Zhou et al., 17 Mar 2026).

The limitations of resetting are equally explicit in the literature. Resetting is not free: if each reset carries a cost $$\langle t_f\rangle_r=\frac{e^{x_0\sqrt{r/D}}-1}{r},$$9, then in the limit of zero reset rate the mean total cost is finite for linear cost, vanishes for sub-linear cost, and diverges for super-linear cost. For exponential cost, the mean total cost diverges at a finite resetting rate, reflecting a power-law cost distribution with a continuously varying exponent (Sunil et al., 2023). Nor does resetting always accelerate completion. In unstable potentials, very frequent resetting can trivially pin the system near the starting point, and in the weak-noise regime resetting can increase rather than decrease the lifetime by interrupting nearly deterministic motion (Capała et al., 2022). More generally, sharp restart can outperform stochastic reset timing for narrow target distributions, whereas stochastic reset timing is superior for broad target distributions (Evans et al., 2024). The modern theory therefore treats stochastic resetting not as a universally beneficial intervention, but as a tunable renewal mechanism whose efficacy depends on the tail structure of completion times, the statistics of targets, the cost model, and the availability of informative state variables such as energy.