---
title: Lagrangian Spontaneous Stochasticity
url: https://www.emergentmind.com/topics/lagrangian-spontaneous-stochasticity
type: topic
---

# Lagrangian Spontaneous Stochasticity

Searching arXiv for recent and foundational papers on Lagrangian spontaneous stochasticity.

Searching arXiv for the 2016 fluctuation–dissipation paper and related recent work.

Lagrangian spontaneous stochasticity is the persistent non-determinism of Lagrangian particle trajectories in the joint limit of vanishing viscosity and diffusivity. In the standard scalar-transport setting, a divergence-free velocity field $\mathbf{u}^\nu$ advects a scalar $\theta(\mathbf{x},t)$ while molecular diffusion is represented by Brownian forcing in a backward Itô SDE; spontaneous stochasticity occurs when the resulting transition probabilities converge, as $\nu,\kappa\to 0$, to a nontrivial limiting density rather than to a delta distribution concentrated on a unique trajectory [1606.00729]. In that sense, the phenomenon is not ordinary sensitivity to initial conditions but a singular loss of Lagrangian determinism in a fixed velocity realization, and it has become a central mechanism for understanding anomalous scalar dissipation, turbulent transport, breakdown of classical flux-freezing, and finite-time intrinsic randomness in rough flows [1103.1882].

## 1. Definition and stochastic Lagrangian formulation

For scalars in a compact domain without walls, the advection–diffusion equation is
$$
\partial_t \theta + \mathbf{u}^\nu\!\cdot\nabla\theta = \kappa \Delta \theta + S(\mathbf{x},t),
$$
with $\nabla\cdot\mathbf{u}^\nu=0$, $\kappa>0$, and source $S$. Diffusion is represented Lagrangianly by a backward stochastic flow
$$
d_s \widetilde{\mathbf{x}_{t,s}^{\nu,\kappa}(\mathbf{x})}
= \mathbf{u}^\nu\big(\widetilde{\mathbf{x}_{t,s}^{\nu,\kappa}(\mathbf{x}),s\big)\,ds
+ \sqrt{2\kappa}\,d\mathbf{W}_s,
\qquad
\widetilde{\mathbf{x}_{t,t}^{\nu,\kappa}(\mathbf{x})=\mathbf{x},
$$
and the scalar admits the stochastic representation
$$
\theta(\mathbf{x},t)
=
\mathbb{E}\Big[
\theta_0\big(\widetilde{\mathbf{x}_{t,0}^{\nu,\kappa}(\mathbf{x})\big)
+
\int_0^t S\big(\widetilde{\mathbf{x}_{t,s}^{\nu,\kappa}(\mathbf{x}),s\big)\,ds
\Big].
$$
The associated backward transition probability is
$$
p^{\nu,\kappa}(\mathbf{x}_0,0\mid\mathbf{x},t)
=
\mathbb{E}\Big[
\delta^d\big(\mathbf{x}_0-\widetilde{\mathbf{x}_{t,0}^{\nu,\kappa}(\mathbf{x})\big)
\Big].
$$
For fixed $\nu>0$, the limit $\kappa\to 0$ yields deterministic trajectories and hence a delta distribution. Lagrangian spontaneous stochasticity is the statement that, in the joint limit $\nu\to 0$, $\kappa\to 0$, possibly at fixed Prandtl number $Pr=\nu/\kappa$, the transition probabilities converge to
$$
p^*(\mathbf{x}_0,0\mid\mathbf{x},t)
=
\lim_{\nu,\kappa\to 0}
p^{\nu,\kappa}(\mathbf{x}_0,0\mid\mathbf{x},t),
$$
where $p^*$ is not a delta distribution [1606.00729].

The defining randomness is therefore attached to Lagrangian trajectories at fixed velocity realization. It does not arise from ensemble-averaging over turbulent velocity fields; it is the Brownian regularization that becomes dynamically amplified and survives the vanishing-noise limit. The standard physical interpretation is that explosive pair dispersion erases memory of the initial diffusive perturbation, so the limiting stochastic process selects a probability measure over non-unique generalized trajectories rather than a single path. In Kraichnan-type settings, this is tied to Hölder exponents smaller than $1$, for which the deterministic ODE $d\mathbf{x}/dt=\mathbf{u}(\mathbf{x},t)$ ceases to have a unique solution set [1606.00729].

A closely related construction appears in Burgers dynamics, where viscous or noisy backward trajectories converge to stochastic generalized characteristics at shocks. There, spontaneous stochasticity is explicitly backward in time: a point on a shock has a nontrivial distribution over earlier Lagrangian pre-images, and the limiting backward flow remains random even as viscosity and noise vanish [1401.5541].

## 2. Lagrangian fluctuation–dissipation relation

The principal structural result for scalar turbulence is the exact Lagrangian fluctuation–dissipation relation. Defining the stochastic scalar sampled along the backward flow by
$$
\tilde{\theta}(\mathbf{x},t)
=
\theta_0\big(\widetilde{\mathbf{x}_{t,0}^{\nu,\kappa}(\mathbf{x})\big)
+
\int_0^t S\big(\widetilde{\mathbf{x}_{t,s}^{\nu,\kappa}(\mathbf{x}),s\big)\,ds,
$$
one has $\theta(\mathbf{x},t)=\mathbb{E}[\tilde{\theta}(\mathbf{x},t)]$ and the backward Itô identity
$$
\tilde{\theta}(\mathbf{x},t)-\mathbb{E}[\tilde{\theta}(\mathbf{x},t)]
=
-\sqrt{2\kappa}\int_0^t d\mathbf{W}_s\cdot
\nabla\theta\big(\widetilde{\mathbf{x}_{t,s}^{\nu,\kappa}(\mathbf{x}),s\big).
$$
Itô isometry yields the local FDR
$$
\mathrm{Var}\big[\tilde{\theta}(\mathbf{x},t)\big]
=
2\kappa\int_0^t ds\,
\mathbb{E}\Big[
|\nabla\theta\big(\widetilde{\mathbf{x}_{t,s}^{\nu,\kappa}(\mathbf{x}),s\big)|^2
\Big],
$$
and incompressibility upgrades this to the global relation
$$
\frac{1}{2}\big\langle \mathrm{Var}[\tilde{\theta}(t)]\big\rangle_\Omega
=
\kappa\int_0^t ds\,\langle |\nabla\theta(s)|^2\rangle_\Omega.
$$
The right-hand side is the cumulative scalar dissipation due to molecular diffusivity, while the left-hand side is the variance of scalar inputs sampled backward in time by stochastic trajectories [1606.00729].

This relation is exact for fixed realization of $\mathbf{u}^\nu$, $\theta_0$, and $S$. A common misconception is that the variance on the left-hand side reflects only turbulent ensemble statistics; in fact, the randomness comes solely from the Brownian motion in the stochastic flow. The FDR thus identifies dissipation with Lagrangian sampling variance rather than with an independent closure ansatz.

For wall-bounded flows, the same structure survives but requires additional probabilistic objects. With Neumann data, one introduces boundary local-time density along reflected backward trajectories; with Dirichlet data, one introduces the backward first hitting-time of the space-time exit surface. In the no-flux case, the global FDR remains formally identical to the wall-free relation. With prescribed boundary values, the stopped representation yields a lower-bound inequality for dissipation, while an unstopped representation involving boundary fluxes restores an exact equality [1703.08133].

## 3. Equivalence with anomalous scalar dissipation

Anomalous scalar dissipation is the persistence of nonzero scalar dissipation in the limit $\kappa\to 0$, namely
$$
\lim_{\kappa\to 0}
\kappa \int_0^t ds\,\langle |\nabla\theta^{\nu,\kappa}(s)|^2\rangle_\Omega >0
$$
for some fixed $t>0$ along some sequence $(\nu_k,\kappa_k)\to(0,0)$. In the passive, source-free case, the FDR becomes
$$
\frac{1}{2}
\big\langle
\mathrm{Var}[\theta_0(\widetilde{\mathbf{x}_{t,0}^{\nu,\kappa}(\mathbf{x})})]
\big\rangle_\Omega
=
\kappa \int_0^t ds\,\langle |\nabla\theta^{\nu,\kappa}(s)|^2\rangle_\Omega.
$$
Using one- and two-time transition probabilities, the variance can be written as a covariance kernel involving
$$
p_2^{\nu,\kappa}(\mathbf{x},s;\mathbf{x}',s'\mid\mathbf{x},t)
=
\mathbb{E}\Big[
\delta^d(\mathbf{x}-\widetilde{\mathbf{x}_{t,s}^{\nu,\kappa}(\mathbf{x})})
\delta^d(\mathbf{x}'-\widetilde{\mathbf{x}_{t,s'}^{\nu,\kappa}(\mathbf{x})})
\Big].
$$
If the limiting trajectories were deterministic, then one would have factorization $p_2^*=p^*p^*$, the covariance kernel would vanish, and anomalous dissipation would be impossible. Therefore anomalous dissipation requires spontaneous stochasticity, expressed precisely as non-factorization of limiting transition probabilities [1606.00729].

For passive scalars, the converse also holds: if spontaneous stochasticity occurs, then one can choose suitable smooth initial scalar data, and if necessary suitable sources, so that the limiting Lagrangian variance is strictly positive. Hence passive-scalar spontaneous stochasticity and anomalous scalar dissipation are equivalent phenomena in flows without walls. For active scalars, the necessity direction survives, but sufficiency is more delicate because the scalar field influences the velocity and initial scalar data cannot be varied independently.

This equivalence extends with modifications to wall-bounded domains. With no scalar flux through walls, the wall-free logic remains valid: spontaneous stochasticity is still the only possible mechanism for anomalous dissipation of passive or active scalars. With scalar fluxes or prescribed wall values, thin scalar boundary layers provide a distinct mechanism for non-vanishing dissipation, so spontaneous stochasticity ceases to be the unique explanation, although it remains another possible mechanism [1703.08133].

The Burgers equation offers a complementary one-dimensional paradigm. There, shock solutions exhibit backward spontaneous stochasticity for all Prandtl numbers, and the backward stochastic flow makes the Burgers velocity a backward martingale. This martingale property guarantees dissipativity of conservation-law anomalies for general convex functions of the velocity, providing a direct Lagrangian mechanism for anomalous dissipation [1401.5541].

## 4. Physical mechanism and distinction from classical chaos

The phenomenological mechanism is turbulent explosive separation. In high-Reynolds and high-Péclet turbulence, particle pairs undergo Richardson dispersion,
$$
\langle r^2(\tau)\rangle \propto \varepsilon \tau^3,
$$
so arbitrarily small perturbations in initial separation or diffusive noise become dynamically irrelevant after a short transient. In the homogeneous, isotropic DNS examined in the fluctuation–dissipation study, backward stochastic trajectories exhibit initial diffusive growth $\propto 12\kappa s$, followed by rapid crossover to super-ballistic behavior close to Richardson’s $\varepsilon s^3$ law, with approximate independence on the Prandtl number in the inertial range [1606.00729]. This supports the interpretation that the limiting transition probability becomes independent of the microscopic diffusion amplitude.

Spontaneous stochasticity is therefore distinct from the butterfly effect. Classical chaos implies exponential divergence of nearby deterministic trajectories, but for any fixed finite time the separation can be made arbitrarily small by shrinking the initial perturbation. By contrast, intrinsic randomness in Lorenz’s sense is finite-time explosive separation that does not disappear as the initial perturbation tends to zero. In the Kelvin–Helmholtz instability of a singular vortex sheet, the separation energy between initially $\varepsilon$-close solutions enters a universal algebraic regime
$$
\mathcal{E}(t)\approx 0.14\,t,
$$
independent of initial separation and viscosity in the vanishing-regularization limit; this is presented as an Eulerian counterpart of Richardson’s law for particles [2004.08908].

Recent work sharpens the distinction further by proposing Eulerian spontaneous stochasticity: the zero-noise, infinite-Reynolds-number limit yields a probability distribution over velocity fields rather than a unique deterministic Euler solution. In shell-model simulations of fluctuating hydrodynamics, microscopic noise is amplified by an inverse error cascade into a stochastic front that randomizes inertial-range shells in a few local turnover times, producing nontrivial universal PDFs at finite times [2401.13881]. This suggests that Lagrangian spontaneous stochasticity can be viewed as one manifestation of a broader singular limit in which both trajectories and fields become statistically well-defined but individually non-predictive.

## 5. Generalizations, model systems, and modifications by intermittency

The concept originated in synthetic scalar-advection models and has since been developed across a wide range of deterministic and stochastic systems. In Kraichnan’s white-in-time Gaussian ensemble, spontaneous stochasticity and anomalous scalar dissipation were established as linked phenomena; the 2016 fluctuation–dissipation relation generalized this to any divergence-free velocity field, including Navier–Stokes solutions away from walls, thereby resolving the objection that spontaneous stochasticity might be a peculiarity of white-in-time models [1606.00729].

The phenomenon also appears in deterministic active systems. In viscous surface quasi-geostrophic turbulence from smooth deterministic initial data, numerics reveal three anomalies in the turbulent regime: dissipative anomaly, multifractal scaling, and super-diffusive separation of fluid particles both backward and forward in time. The interpretation advanced there is that anomalous dissipation, broken scale invariance, and loss of uniqueness of the Lagrangian flow are intertwined, although the resulting spontaneous stochasticity is described as tempered and non-universal in the deterministic SQG setting [2210.12366].

Intermittency can modify the picture in nontrivial ways. In a one-dimensional multifractal random flow built from a rough Markovian Gaussian field decorated by frozen Gaussian multiplicative chaos, the effective exponent controlling pair-separation dynamics becomes
$$
\xi_\gamma = \xi + 4\gamma^2,
$$
leading to a transition at
$$
\gamma_c = \frac{1}{2}\sqrt{1-\xi}.
$$
Below $\gamma_c$ the model is spontaneously stochastic, while above $\gamma_c$ it becomes deterministic. This is a counterexample to the naive expectation that greater roughness always favors spontaneous stochasticity; in that setting, intermittency can suppress it [2305.09839].

More geometric deterministic realizations have also been constructed. In the 3D Weierstrass-ABC flow,
$$
\bm u_W(\bm x) = \sum_{i=1}^{\infty} \frac{\omega_i}{k_i}\,\bm U(k_i\bm x),
\qquad
k_i=\lambda^i,\quad \omega_i=\lambda^{(1-h)i},
$$
the velocity is $h$-Hölder and the ideal Lagrangian problem is ill-posed for $h<1$. Monte Carlo simulations with both Langevin and randomized-initial-data regularizations yield convergent non-Dirac marginals at $h=1/3$, and the limiting statistics are insensitive to the type of stochastic regularization, which is presented as numerical evidence for spontaneous stochasticity in a rough 3D flow [2502.07581].

A different deterministic construction is the Armstrong–Vicol passive scalar, where a divergence-free multiscale vector field arbitrarily close to a weak Euler solution drives anomalous scalar dissipation and the absence of a selection principle in the vanishing-diffusivity limit. In that model, the passive scalar exhibits both Lagrangian and Eulerian spontaneous stochasticity, and the lack of a classical selection principle is reinterpreted as a measure selection principle over weak inviscid solutions [2504.15795].

## 6. Broader implications, wall effects, and unresolved directions

One major implication concerns magnetic flux-freezing and reconnection in MHD. In turbulent plasmas with inertial-range power-law spectra, Lagrangian trajectories become spontaneously stochastic, so infinitely many magnetic field-lines are advected to each point and the magnetic field must be represented by an average over stochastic histories. This stochastic flux-freezing framework yields the Lazarian–Vishniac scaling for the reconnection rate of large-scale magnetic structures and explains why fast reconnection can persist when microscopic non-ideal plasma terms are negligible at inertial-range scales [1103.1882]. A plausible implication is that spontaneous stochasticity is not confined to passive transport but is a generic route by which singular turbulent dynamics invalidates classical frozen-in laws.

Another implication concerns the inviscid limit of deterministic PDEs more broadly. RG-based studies on scale-invariant lattices and shell models interpret spontaneous stochasticity as the emergence of a stochastic fixed-point kernel for the renormalization-group dynamics of flow maps or Markov kernels. In that formulation, universality of the limiting stochastic process follows from attraction to an RG fixed point, while diversity of limiting behavior can arise through period-doubling bifurcations and stochastic attractors [2410.14903]. This suggests a route toward classifying universality classes of spontaneous stochasticity beyond specific transport equations.

Wall-bounded flows complicate the narrative because walls permit scalar boundary layers that can sustain non-vanishing dissipation independently of bulk trajectory randomness. Pure conduction with constant imposed flux or oscillatory Dirichlet boundary values gives explicit examples in which dissipation remains finite or diverges as $\kappa\to 0$ even when the advecting velocity is zero; in such cases, the mechanism is not spontaneous stochasticity but thin wall layers with gradients scaling like powers of $\kappa^{-1}$ [1703.08133]. The distinction is conceptually important: spontaneous stochasticity is a Lagrangian bulk mechanism tied to non-factorization of transition probabilities, whereas boundary-layer anomaly is an Eulerian near-wall mechanism tied to wall-imposed scalar structure.

A long-standing controversy asked whether spontaneous stochasticity was merely an artifact of synthetic models such as Kraichnan’s ensemble. The fluctuation–dissipation relation for scalar turbulence effectively closed that debate for flows without walls: for any divergence-free velocity field, including incompressible Navier–Stokes solutions, anomalous scalar dissipation requires spontaneous stochasticity, and for passive scalars the two are equivalent [1606.00729]. What remains less settled is the precise universality of the phenomenon in realistic intermittent, active, and wall-bounded settings, and the extent to which Eulerian and Lagrangian versions of spontaneous stochasticity can be unified in a single theory of singular turbulent limits.

Source: https://www.emergentmind.com/topics/lagrangian-spontaneous-stochasticity