Spontaneous Stochasticity in Deterministic Systems

• Spontaneous stochasticity is the persistence of inherent randomness in systems where vanishing regularization yields non-unique deterministic outcomes.
• It arises in various contexts such as turbulence, passive-scalar transport, and quantum semiclassical limits, highlighting failures in classical uniqueness.
• The phenomenon challenges traditional deterministic chaos by demonstrating that non-commuting limits and rough dynamics lead to universal stochastic behaviors.

Spontaneous stochasticity is the persistence of intrinsic randomness in a singular limit of an otherwise deterministic dynamical system, typically when viscosity, diffusivity, noise amplitude, or a microscopic cutoff vanish while the limiting ideal dynamics ceases to be uniquely solvable. In the literature it appears in Lagrangian form, as non-unique particle trajectories in rough velocity fields, and in Eulerian form, as non-Dirac limiting laws for fields or weak solutions of ideal equations; in both cases the limiting evolution remains stochastic even though the governing equations and data are deterministic (Eyink et al., 2015, Mailybaev et al., 27 Feb 2026, Ruffenach et al., 19 Sep 2025). Across turbulence, passive-scalar transport, Burgers shocks, singular shear layers, shell models, logarithmic-lattice Navier–Stokes, and quantum semiclassical limits, the phenomenon is tied to Hölder roughness, anomalous dissipation, and finite-time loss of predictability, and it is explicitly distinguished from both ordinary noise-induced randomness and deterministic chaos (Thalabard et al., 2020, Mailybaev, 2015).

1. Core concept and taxonomy

The basic criterion is that a vanishing-regularization limit fails to collapse onto a single deterministic outcome. In Lagrangian spontaneous stochasticity, transition probabilities for particle trajectories remain nontrivial as molecular diffusivity or auxiliary noise tends to zero. In Eulerian spontaneous stochasticity, the limiting object is a stochastic process for the field itself, or a non-Dirac probability measure on weak solutions of the inviscid problem (Mailybaev, 2024, Ruffenach et al., 19 Sep 2025).

A recent general formulation distinguishes two regimes. In the “strong” regime, there exists an ambient measure family on the regularization parameter such that the pushforward converges weakly to a non-Dirac limit measure. In the “weak” regime, only subsequential weak limits exist, so the limiting statistics is not uniquely selected. In finite-dimensional systems, after excluding purely Dirac limits, lack of selection principle, observable lack of selection principle, and spontaneous stochasticity are equivalent (Ruffenach et al., 19 Sep 2025).

This framework also clarifies a frequent misconception. Deterministic chaos concerns unique trajectories with sensitive dependence on initial data, usually diagnosed by Lyapunov growth. Spontaneous stochasticity instead concerns non-uniqueness of the ideal limit itself. The real GOY shell model analyzed in the vanishing-viscosity limit is a particularly sharp illustration: before blowup it is deterministic, and after blowup its renormalized dynamics is periodic or quasi-periodic rather than chaotic, yet the inviscid limit becomes non-unique and, under randomization of viscosity, spontaneously stochastic (Mailybaev, 2015).

A related structural marker is non-commutation of limits. In stochastic-field-theoretic language, one encounters

$$\lim_{D\to 0}\lim_{t\to\infty} x(t;D)\neq \lim_{t\to\infty}\lim_{D\to 0}x(t;D),$$

which is interpreted as persistence of randomness in the zero-noise limit because the effective deterministic flow is ill-posed (Kikuchi, 2024).

2. Mathematical mechanism: roughness, non-uniqueness, and fluctuation–dissipation

The canonical classical picture is Richardson dispersion. In the turbulence model

$\frac{dr}{dt}=C(\epsilon r)^{1/3},$

the solution is

$$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$

Hence $r^2(t)\propto \epsilon t^3$, and as $r_0\to 0$,

$$r(t)\to \Big[\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}>0.$$

Moreover, one may construct infinitely many solutions that remain at $r=0$ up to an arbitrary waiting time $\tau$ and then separate. This is the standard non-uniqueness of ODEs driven by non-Lipschitz fields: if $|u(x)-u(y)|\le C|x-y|^\alpha$ with $0<\alpha<1$, Picard–Lindelöf uniqueness may fail (Eyink et al., 2015).

The same mechanism appears in the minimal one-dimensional singular model

$\frac{dr}{dt}=C(\epsilon r)^{1/3},$0

For $\frac{dr}{dt}=C(\epsilon r)^{1/3},$1, this ODE admits an uncountable family of solutions parametrized by a waiting time and a sign, with extremal solutions

$\frac{dr}{dt}=C(\epsilon r)^{1/3},$2

A renormalization-group analysis shows that, in the near-singular weak-noise limit, the statistics flows to fixed points supported on these extremal branches, while the approach inside the extremal envelope is controlled by a singular large-deviation scaling rather than the standard saddle-point approximation to the Onsager–Machlup functional (Eyink et al., 2020).

For passive scalars, the fluctuation–dissipation relation makes the connection to dissipation exact. In a two-dimensional autonomous rough divergence-free field, backward stochastic characteristics $\frac{dr}{dt}=C(\epsilon r)^{1/3},$3 satisfy

$\frac{dr}{dt}=C(\epsilon r)^{1/3},$4

and the scalar obeys the exact identity

$\frac{dr}{dt}=C(\epsilon r)^{1/3},$5

A strictly positive zero-noise limit of the trajectory variance therefore forces anomalous dissipation, and conversely persistent dissipation implies that the limiting backward flow cannot be deterministic (Johansson et al., 2024). This equivalence is also central in the Armstrong–Vicol passive scalar and in the Kraichnan-type literature summarized in later work (Ruffenach et al., 19 Sep 2025).

3. Fluid and transport realizations

Burgers dynamics provides a paradigmatic compressible example. For entropy solutions with shocks, forward characteristics coalesce, but backward characteristics spontaneously branch in the zero-noise limit. The backward stochastic flow constructed from the Constantin–Iyer representation yields transition measures supported on the Hopf–Lax minimizers; at a shock these are convex combinations of the two extreme preimages, with equal weights $\frac{dr}{dt}=C(\epsilon r)^{1/3},$6 on the viscous shock surface. The Lagrangian velocity is then a backward martingale, and this property enforces entropy inequalities and anomalous dissipation for convex entropies (Eyink et al., 2014).

In two-dimensional Kelvin–Helmholtz evolution of a vortex sheet, spontaneous stochasticity appears in Eulerian form. The singular shear layer is ill-posed for Euler, and vanishingly small noise or regularization produces a universal stochastic process rather than a deterministic limit. The paper measuring this effect defines a macroscopic separation energy

$\frac{dr}{dt}=C(\epsilon r)^{1/3},$7

and finds

$\frac{dr}{dt}=C(\epsilon r)^{1/3},$8

in the simultaneous zero-noise, zero-viscosity limit. The mixing-layer thickness obeys

$\frac{dr}{dt}=C(\epsilon r)^{1/3},$9

with one- and two-point vorticity statistics collapsing across Navier–Stokes and Birkhoff–Rott regularizations (Thalabard et al., 2020).

Surface quasi-geostrophic turbulence displays a related but subtler picture. Starting from smooth deterministic data, viscous SQG develops a turbulent regime with anomalous dissipation of scalar variance, multifractal scaling, and super-diffusive pair separation approximately proportional to $$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$0. However, the stochasticity is “tempered”: the limiting pair-dispersion statistics and Eulerian predictability diagnostics are non-universal and sensitive to the vanishing-viscosity protocol, unlike the more rigid universal behavior of the Kraichnan model (Valade et al., 2022).

Two recent deterministic passive-scalar constructions sharpen the PDE side. First, a two-dimensional autonomous divergence-free field in $$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$1, $$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$2, was constructed so that backward stochastic trajectories retain positive variance as $$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$3, which by the fluctuation–dissipation relation yields anomalous scalar dissipation; the same construction also gives $$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$4-dimensional forced Navier–Stokes sequences with anomalous dissipation and an Euler limit whose energy is Lipschitz in time (Johansson et al., 2024). Second, in the Armstrong–Vicol passive scalar, the inviscid limit is not merely non-unique: it selects a non-Dirac measure in the space of weak solutions, giving strong Eulerian spontaneous stochasticity. In the accompanying general finite-dimensional framework, the set of selected measures is compact and equals the closed convex hull of Dirac measures on the inviscid solution set, and every non-Dirac measure supported there can be produced by a suitable regularization (Ruffenach et al., 19 Sep 2025).

4. Eulerian spontaneous stochasticity, renormalization group, and universality

Multiscale shell and lattice models have become the main arena for explicit Eulerian spontaneous stochasticity. In the real GOY shell model, finite-time blowup separates a deterministic pre-blowup regime from a post-blowup regime where the vanishing-viscosity limit depends on a phase $$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$5. When viscosity is treated as a small random parameter, the inviscid limit becomes a singular probability measure supported on this one-parameter family. The resulting renormalized dynamics is periodic before blowup and quasi-periodic after blowup, making clear that spontaneous stochasticity is not synonymous with chaos (Mailybaev, 2015).

A rigorous Markovian realization appears in the multiscale Arnold’s cat model. There, deterministic initial data generate uncountably many ideal solutions, deterministic cutoffs yield non-unique inviscid limits, and injecting a single random perturbation at the smallest active scale produces a universal stochastic limit as the cutoff is removed. The limit is Markovian and assigns equal probability to the admissible deterministic continuations, providing an explicit Eulerian selection mechanism by infinitesimal noise (Mailybaev et al., 2021).

This picture has been abstracted into RG-on-kernels formalisms. On self-similar fractal lattices, the inviscid limit is represented as an attractor of an RG operator acting on flow maps or Markov kernels; deterministic fixed points yield universal deterministic limits, while nontrivial kernel attractors yield spontaneous stochasticity (Mailybaev et al., 2022). Subsequent work analyzed stability and bifurcations of these attractors, including deterministic and stochastic period-doubling transitions and a universal linearized RG eigenmode governing convergence (Mailybaev, 2024). A later synthesis recast spontaneous stochasticity as a universality phenomenon on the same footing as the Feigenbaum equation, the central limit theorem, and hierarchical spin models, with a fixed point in the space of Markov kernels as the central object (Mailybaev et al., 27 Feb 2026).

For the fluctuating Sabra shell model, this RG program becomes quantitative. The ideal-limit stochastic process is identified with an RG fixed point, and the leading correction is controlled by a universal complex eigenvalue

$$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$6

which explains the slow, oscillatory convergence in cutoff or Reynolds number. Numerical data confirm that this $$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$7 is independent of the specific dissipation and noise terms within a broad canonical class (Mailybaev, 15 Sep 2025).

The same finite-time universality is seen in noisy shell and lattice Navier–Stokes analogues. In the Sabra model with thermal noise, the randomization time at scale $$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$8 obeys

$$r(t)=\Big[r_0^{2/3}+\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}.$$9

and the discussion argues that any noise whose amplitude vanishes more slowly than $r^2(t)\propto \epsilon t^3$0 triggers spontaneous stochasticity at high Reynolds number; thermal noise alone is therefore sufficient in that model (Bandak et al., 2024). On a logarithmic lattice that preserves the exact algebraic form, symmetries, and inviscid invariants of the 3D Navier–Stokes/Euler equations, PDFs of large-scale vorticity modes converge to Dirac masses before blowup but to non-degenerate limiting laws after blowup or from rough initial data, which the authors describe as the first numerical probing of spontaneous stochasticity in a model retaining the exact NS/Euler structure (Ortiz et al., 3 Jul 2025).

5. Quantum and cross-domain extensions

The quantum counterpart is formulated in “Quantum Spontaneous Stochasticity.” There a massive particle evolves in a one-dimensional repulsive nearly rough potential

$r^2(t)\propto \epsilon t^3$1

regularized at short distance $r^2(t)\propto \epsilon t^3$2, with an initial minimum-uncertainty Gaussian centered at the origin. In the order of limits $r^2(t)\propto \epsilon t^3$3 followed by $r^2(t)\propto \epsilon t^3$4, the WKB density converges to the non-deterministic mixture

$r^2(t)\propto \epsilon t^3$5

for symmetric initial data, where $r^2(t)\propto \epsilon t^3$6 are the two extremal non-unique classical solutions. The second moment exhibits Richardson-like super-ballistic growth,

$r^2(t)\propto \epsilon t^3$7

and the splitting occurs on a short timescale $r^2(t)\propto \epsilon t^3$8, with $r^2(t)\propto \epsilon t^3$9 for $r_0\to 0$0 and $r_0\to 0$1 for $r_0\to 0$2. Full Schrödinger simulations further show momentum-space splitting and effective decoherence of the two branches in the classical limit (Eyink et al., 2015).

A complementary deterministic rough-flow realization is the three-dimensional Weierstrass–ABC model. This stationary incompressible field is built as a Weierstrass superposition of ABC modes with Hölder exponent $r_0\to 0$3, and for $r_0\to 0$4 it exhibits structure functions $r_0\to 0$5 and an Onsager-type anomalous dissipation plateau in the Duchon–Robert flux. Two distinct regularizations—additive white noise in the dynamics and randomized initial conditions under truncation—produce the same non-degenerate limiting coordinate distributions, providing a controlled three-dimensional demonstration of spontaneous stochasticity independent of the stochastic regularization used (Barlet et al., 11 Feb 2025).

Magnetohydrodynamics supplies another major extension. In turbulent MHD with inertial-range power-law spectra, spontaneous stochasticity of Lagrangian trajectories destroys deterministic Alfvén flux freezing and replaces it with stochastic flux freezing: the magnetic field at a point is the average over infinitely many backward stochastic trajectories. Using Goldreich–Sridhar and weak-MHD scaling, this yields the Lazarian–Vishniac reconnection width

$r_0\to 0$6

and reconnection rate

$r_0\to 0$7

while Hall effects remain negligible for the inertial-range dynamics whenever $r_0\to 0$8 and $r_0\to 0$9 (Eyink et al., 2011).

6. Diagnostics, controversies, and open directions

The operational diagnostics vary by setting but follow a common pattern: a non-degenerate weak limit for PDFs or path measures, finite variance of backward or forward trajectories as regularization vanishes, or asymptotically nonzero dissipation tied by an exact fluctuation–dissipation identity to residual trajectory randomness. In the AV passive scalar this appears as a non-Dirac limiting measure on weak solutions (Ruffenach et al., 19 Sep 2025); in the autonomous passive-scalar construction it appears as a uniform lower bound on backward-trajectory variance on a positive-measure set (Johansson et al., 2024); in the logarithmic-lattice Navier–Stokes model it appears as convergence of large-scale mode PDFs to finite-width laws after blowup (Ortiz et al., 3 Jul 2025).

Two controversies recur. The first is whether spontaneous stochasticity is simply another name for chaos. The answer in the cited literature is negative: the real GOY model is explicitly non-chaotic while exhibiting spontaneous stochasticity after blowup (Mailybaev, 2015), and the shell-model analysis of thermal triggering emphasizes finite-time universal statistics rather than infinite-time attractor sensitivity (Bandak et al., 2024). The second concerns universality. Many RG-based models and several rough-flow constructions display regularization-independent limits, but not all systems do. Deterministic SQG exhibits a tempered form of spontaneous stochasticity with non-universal statistics (Valade et al., 2022), and in a one-dimensional Gaussian-multiplicative-chaos-decorated Kraichnan model, increasing intermittency can suppress spontaneous stochasticity altogether. There the mean-field effective Hurst parameter is

$$r(t)\to \Big[\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}>0.$$0

and the transition from spontaneous stochasticity to deterministic sticking occurs near

$$r(t)\to \Big[\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}>0.$$1

despite the fact that pathwise roughness increases with $$r(t)\to \Big[\frac{2}{3}C\epsilon^{1/3}t\Big]^{3/2}>0.$$2 (Considera et al., 2023). This suggests that roughness is necessary in many settings but not, by itself, a complete criterion.

Several open problems remain explicit in the literature. Rigorous continuum PDE theories for Eulerian spontaneous stochasticity in 3D Navier–Stokes/Euler are still lacking, even though shell, lattice, passive-scalar, and Arnold-cat models now furnish detailed mechanisms (Ortiz et al., 3 Jul 2025, Mailybaev et al., 27 Feb 2026). The full temporal structure of the limiting process—Markovianity, self-similarity, and basin dependence—remains unresolved in many models (Ortiz et al., 3 Jul 2025). Higher-dimensional quantum analogues, where radially symmetric cusp potentials would produce a continuum of random outgoing directions rather than two branches, are proposed but not yet analyzed in comparable detail (Eyink et al., 2015). More broadly, the RG perspective suggests a classification of universality classes by fixed points, periodic attractors, or stochastic invariant kernels, but extending this program from discrete multiscale models to continuum turbulence is still an open program (Mailybaev, 2024, Mailybaev, 15 Sep 2025).

Taken together, these results establish spontaneous stochasticity as a precise mechanism by which deterministic multiscale systems generate intrinsic randomness in singular limits. The common structure is the loss of uniqueness caused by rough or nearly singular dynamics, the survival of nontrivial measures as regularization vanishes, and the emergence—sometimes universal, sometimes only subsequential—of statistical laws that replace deterministic selection.