Eulerian Spontaneous Stochasticity

• Eulerian spontaneous stochasticity is the persistence of intrinsic randomness in fluid fields, leading to a probability distribution over weak Euler solutions rather than a single deterministic outcome.
• Methodologies such as fluctuating hydrodynamics and shell-model simulations illustrate the transition from delta measures to non-Dirac distributions in the vanishing-noise, vanishing-viscosity limit.
• Key results reveal that vanishing small-scale perturbations trigger universal finite-time statistics, offering a probabilistic resolution to the ill-posed nature of deterministic Euler equations.

Eulerian spontaneous stochasticity is the persistence of intrinsic randomness in the Eulerian field itself—velocity, vorticity, or other field variables—in a singular limit where microscopic regularization and the noise that triggers the effect both vanish. In this formulation, the limiting object is not a single deterministic Euler solution but a nontrivial probability distribution over weak Euler solutions, or equivalently a stochastic process whose realizations satisfy the ideal equations in a weak sense (Thalabard et al., 2020). The concept is designed to capture a form of unpredictability stronger than the ordinary butterfly effect: not merely rapid amplification of small errors within a deterministic dynamics, but failure of deterministic selection in the ideal limit, with universal finite-time statistics arising from vanishing small-scale perturbations (Bandak et al., 2024).

1. Conceptual definition and mathematical form

A standard formulation begins from noisy, regularized fluid dynamics—such as Landau–Lifshitz fluctuating hydrodynamics or a shell-model analogue—and considers the transition probability

$$P_{Re,\Theta}(\mathbf u_f,t_f\mid \mathbf u_i,t_i),$$

which for fixed $Re$ collapses to a delta measure as $\Theta\to 0$. Eulerian spontaneous stochasticity is the statement that in a double singular limit

$$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$

the limit $P_\infty$ need not be a delta distribution. If $P_\infty$ is nontrivial, then the limiting Eulerian dynamics remain random even though the formal ideal equation is deterministic (Bandak et al., 2024).

This field-level viewpoint is explicitly distinct from formulations based on particle trajectories. In the Eulerian version, the random object is the velocity or vorticity field itself, or a law on weak solutions selected by vanishing-viscosity, vanishing-noise, and often vanishing-cutoff limits. Several papers describe this as a probabilistic resolution of nonuniqueness: the deterministic ideal problem is ill-posed, but the physically relevant asymptotic object is a probability measure supported on weak solutions (Thalabard et al., 2020).

Recent work on deterministic regularizations of inviscid limits pushes this formulation further. One abstract criterion defines spontaneous stochasticity for a regularized family $({\cal P}_\kappa)$ through an observable ${\cal A}(\kappa)$ satisfying

$$-\infty<\liminf_{\kappa\to 0}{\cal A}(\kappa) < \limsup_{\kappa\to 0}{\cal A}(\kappa) <+\infty,$$

thereby identifying Eulerian spontaneous stochasticity with lack of deterministic selection for some Eulerian observable (Ruffenach et al., 22 Apr 2025). In the Armstrong–Vicol passive scalar, the vanishing-diffusivity limit has been described as selecting a non-Dirac measure in the space of weak solutions, with a distinction between weak and strong regimes; the advection–diffusion system of Armstrong and Vicol is placed in the strong regime (Ruffenach et al., 19 Sep 2025).

2. Distinction from deterministic chaos and from Lagrangian spontaneous stochasticity

The distinction from chaos is foundational. In ordinary chaos, two trajectories with initial mismatch $\delta_0$ separate exponentially, but at fixed observation time the separation still goes to zero as $Re$0. The dependence on initial data is violent but continuous. In the stronger Lorenz-type scenario invoked in this literature, arbitrarily small errors evolve within a finite time into states “as different as random states,” and that predictability horizon does not improve when the triggering noise is reduced (Thalabard et al., 2020). In modern terms, the limiting deterministic problem is ill-posed rather than merely chaotic.

The distinction from Lagrangian spontaneous stochasticity is equally important. Lagrangian spontaneous stochasticity concerns particle trajectories in a rough velocity field; Eulerian spontaneous stochasticity concerns the field variables themselves. The 2020 vortex-sheet study explicitly interprets its phenomenon as an Eulerian counterpart of Richardson dispersion and of Lagrangian spontaneous stochasticity: the “separating objects” are whole flow realizations in field space rather than tracer pairs in physical space (Thalabard et al., 2020).

Several adjacent literatures clarify the boundary between the two notions. Studies of 1D multifractal random flows show that increasing intermittency can suppress Lagrangian spontaneous stochasticity even though the velocity field becomes rougher in a Kolmogorov/Hölder sense; this shows that roughness and spontaneous stochasticity are not equivalent notions once intermittency is present (Considera et al., 2023). Work on a rough 3D Weierstrass-ABC flow provides a numerical Lagrangian foundation—vanishing-regularization trajectory laws converge to non-Dirac limits—but does not establish an Eulerian nonuniqueness theorem for transported fields (Barlet et al., 11 Feb 2025). For Burgers, backward-in-time spontaneous stochasticity of generalized characteristics is rigorously tied to Eulerian conservation-law anomalies, but the stochasticity itself remains fundamentally Lagrangian in formulation (Eyink et al., 2014).

A related but more ambiguous case is deterministic SQG. There, numerics reveal anomalous dissipation, multifractality, and super-diffusive separation of fluid particles, and the work argues that Eulerian irreversibility and Lagrangian spontaneous stochasticity are intertwined; yet it also states that it is mathematically unclear whether SQG should display Eulerian spontaneous stochasticity, and describes the observed phenomenon as a tempered, non-universal version (Valade et al., 2022).

3. Canonical realizations and numerical evidence

The clearest direct realization is the singular Kelvin–Helmholtz vortex sheet. The unperturbed flow is a discontinuous shear layer with vorticity

$Re$1

and linear Kelvin–Helmholtz theory predicts growth of a perturbation of wavelength $Re$2 like

$Re$3

Because the growth rate diverges as $Re$4, the ideal evolution is violently ill-posed. When the initial sheet is modulated by infinitesimal random perturbations and regularized either by viscous/hyperviscous Navier–Stokes or by a nondissipative Birkhoff–Rott point-vortex approximation, the large-scale flow does not converge to a single deterministic solution. Instead, the separation energy between two independently perturbed realizations crosses over from an early nonuniversal exponential regime to a viscosity-independent algebraic regime

$Re$5

At the same time, the mixing-layer width follows

$Re$6

and both one-point and two-point vorticity statistics collapse to universal self-similar forms across the two microscopic regularizations (Thalabard et al., 2020). The paper interprets this as the Eulerian field-level analogue of Richardson-like explosive separation.

A second line of evidence comes from high-Reynolds-number turbulence models with thermal noise. In a stochastic Sabra shell model motivated by Landau–Lifshitz fluctuating hydrodynamics, the PDFs of $Re$7, energy fluxes, and multipliers converge at fixed finite times to broad non-delta distributions as $Re$8, while the same deterministic truncated model without thermal noise yields only delta functions (Bandak et al., 2024). The randomization time for shell $Re$9 is reported, for the atmospheric-boundary-layer parameterization and for scales above about $\Theta\to 0$0 cm, as

$\Theta\to 0$1

The direct thermal contribution to the shell velocity is estimated to be 4–5 orders of magnitude smaller than the observed spontaneous fluctuations, so the effect is interpreted as turbulence-mediated amplification of dissipation-range noise rather than direct thermal agitation (Bandak et al., 2024).

A more Eulerian 3D realization is provided by fluctuating incompressible Navier–Stokes on a logarithmic lattice in Fourier space. There the asymptotic regime is posed as

$\Theta\to 0$2

and the measured observables are PDFs of large-scale vorticity modes, especially $\Theta\to 0$3 with $\Theta\to 0$4. For rough initial data $\Theta\to 0$5, the PDFs converge with increasing $\Theta\to 0$6 to a non-delta distribution even though the noise amplitude itself shrinks. For more regular initial data $\Theta\to 0$7, the deterministic log-lattice Euler solution blows up at

$\Theta\to 0$8

and the large-scale PDFs sharpen toward a delta before $\Theta\to 0$9 but remain broad and nontrivial after $$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$0 (Ortiz et al., 3 Jul 2025). This gives a concrete separation between a deterministic strong-solution regime and a post-blowup stochastic weak-solution regime.

4. Probabilistic well-posedness, nonunique weak solutions, and measure selection

A recurrent claim in the literature is that the appropriate asymptotic object is not a deterministic weak solution but a probability law over weak solutions. The Kelvin–Helmholtz study states this directly: the ideal vortex-sheet evolution is ill-posed deterministically but becomes “well-posed in a probabilistic sense,” with the limiting object a stochastic process for the Eulerian field (Thalabard et al., 2020). The 2024 shell-model work uses the same language, describing spontaneous stochasticity as “stochastic behavior of a formally deterministic Euler system, which each realization of the limiting distribution satisfies in a weak sense” (Bandak et al., 2024).

This perspective becomes explicit in recent measure-selection formulations. In the Armstrong–Vicol passive scalar, the absence of a deterministic selection principle in the vanishing-diffusivity limit is reinterpreted as Eulerian spontaneous stochasticity: the inviscid limit selects a non-Dirac measure on the space of weak solutions (Ruffenach et al., 19 Sep 2025). A related framework proves that the set of selected measures is compact and equals the closed convex hull of Dirac measures, and that for any non-Dirac measure supported on the set of nonunique solutions of the inviscid system there exists a regularization that produces strong spontaneous stochasticity (Ruffenach et al., 19 Sep 2025). In the broader abstract theory, the set of limiting measures satisfies

$$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$1

where $$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$2 is the full probability simplex over the set of inviscid states, and Dirac measures represent classical deterministic selection while non-Dirac measures represent spontaneous stochasticity (Ruffenach et al., 22 Apr 2025).

Earlier shell-model work had already provided a mathematically sharp example of this structure. In the real GOY model, deterministic viscous dynamics is non-chaotic for every fixed $$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$3, but after finite-time blowup the inviscid limit depends on subsequences

$$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$4

producing a one-parameter family of inviscid limits $$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$5. Randomizing the logarithmic viscosity phase $$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$6 yields a probability measure on shell-velocity states, so the velocity variables themselves become nonunique/random in the vanishing-viscosity limit (Mailybaev, 2015). This is a model realization of Eulerian spontaneous stochasticity in the strict sense that the random object is the Eulerian velocity degree of freedom rather than a particle path.

A rigorous prototype of the same idea appears in the “Spontaneously stochastic Arnold’s cat” model. There, an infinite-dimensional multiscale deterministic system with deterministic initial conditions possesses uncountably many solutions, deterministic cutoff regularizations have many subsequential limits, and adding a single microscopic random seed at the ultraviolet cutoff yields a unique universal limiting probability measure with Markovian properties (Mailybaev et al., 2021). The paper presents this explicitly as a toy model of Eulerian spontaneous stochasticity.

5. Universality, renormalization group, and regularization dependence

A central claim of the subject is universality: the limiting stochastic law, when it exists, should be independent of the detailed microscopic regularization. In the vortex-sheet problem this is formulated as collapse of macroscopic statistics—$$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$7, $$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$8, $$P_{Re, \Theta}(\mathbf u_f, t_f|\mathbf u_i, t_i) \xrightarrow[Re\rightarrow \infty]{\Theta \rightarrow 0} P_\infty (\mathbf u_f, t_f|\mathbf u_i, t_i),$$9, and $P_\infty$0—across viscous/hyperviscous Navier–Stokes and a Birkhoff–Rott point-vortex approximation, above the regularization scale (Thalabard et al., 2020). In the shell-model turbulence setting, PDFs converge to the same limits for different microscopic noise scalings, notably $P_\infty$1 and $P_\infty$2, and for different regularization paths toward $P_\infty$3 (Bandak et al., 2024). In the logarithmic-lattice Navier–Stokes study, the phenomenon is observed under two distinct noise scalings, $P_\infty$4 and $P_\infty$5, again suggesting universality of the limiting law (Ortiz et al., 3 Jul 2025).

The most systematic attempts to explain this universality use renormalization-group language. One abstract theory defines the inviscid limit as an attractor of an RG operator acting on flow maps or probability kernels on a scale-invariant space-time lattice; if the attractor is a nontrivial probability kernel, the deterministic ideal system is spontaneously stochastic (Mailybaev et al., 2022). A later synthesis states the program more sharply: spontaneous stochasticity emerges as a universal fixed point of an RG transformation acting on Markov kernels, independent of microscopic regularization (Mailybaev et al., 27 Feb 2026). In that formulation, convergence, stochasticity, spontaneity, and universality are all encoded as properties of an RG fixed point in the space of transition kernels.

The fractal-lattice RG analysis extends this picture by allowing multiple attractor types. Fixed-point attractors explain deterministic universal inviscid limits; period-doubling bifurcations produce two universal parity-dependent branches; chaotic RG dynamics motivate stochastic RG operators and stochastic attractors, which the paper interprets as prototypes of Eulerian spontaneous stochasticity and of its “potential diversity” (Mailybaev, 2024). For the Sabra model, an RG theory is developed in which the dominant eigenmode of the linearized RG operator controls the leading correction to the ideal-limit law, with a universal complex eigenvalue

$P_\infty$6

explaining slow and oscillatory convergence (Mailybaev, 15 Sep 2025).

Universality, however, is not automatic. The real GOY shell model shows that replacing $P_\infty$7 by hyperviscosity can change the limiting support in $P_\infty$8-space, so the spontaneously stochastic solution there depends on the small-scale regularization mechanism (Mailybaev, 2015). A plausible implication is that “universality” is model- and class-dependent: it may hold robustly within a basin of RG attraction or a class of admissible regularizations, yet fail across qualitatively different microscopic mechanisms.

6. Limitations, controversies, and open directions

The principal limitation is that direct evidence for Eulerian spontaneous stochasticity in full 3D continuum Navier–Stokes or Euler remains absent. The 2020 vortex-sheet paper is purely numerical, confined to two dimensions, finite periodic domains, and times short enough to avoid layer interaction; it does not construct the limiting stochastic process rigorously or prove uniqueness of any associated statistical solution class (Thalabard et al., 2020). The 2024 thermal-noise work is based on the Sabra shell model and explicitly states that it does not prove Eulerian spontaneous stochasticity for 3D Navier–Stokes (Bandak et al., 2024). The 2025 logarithmic-lattice study emphasizes that it provides only numerical evidence in a Fourier-space surrogate rather than a theorem for the continuum PDE (Ortiz et al., 3 Jul 2025).

A second unresolved issue is the role of intermittency and regularization path. In 1D multifractal random flows, increasing intermittency can destroy spontaneous stochasticity even though sample paths become rougher, so roughness alone is not a sufficient criterion (Considera et al., 2023). Deterministic SQG suggests a tempered, non-universal version whose detailed behavior may depend sensitively on the joint limits $P_\infty$9 and perturbation amplitude $P_\infty$0 (Valade et al., 2022). The passive-scalar measure-selection papers argue that the correct language is not deterministic selection but measure selection, yet they also show that different ambient measures or regularization procedures can lead to different probabilistic limits unless one is in a strong regime (Ruffenach et al., 22 Apr 2025).

A common misconception is that Eulerian spontaneous stochasticity is merely “chaos with noise.” The literature rejects this. At fixed $P_\infty$1, taking the noise amplitude to zero returns a deterministic system. The phenomenon requires a singular multiscale limit in which smaller and smaller active scales enter the dynamics, often with divergent local Lyapunov exponents or post-blowup nonuniqueness. Another misconception is that Lagrangian spontaneous stochasticity automatically implies an Eulerian version. The passive-scalar literature supplies counterexamples: Lagrangian nonuniqueness of characteristics can coexist with deterministic Eulerian limits, whereas Eulerian spontaneous stochasticity requires nontrivial probabilistic selection at the level of the field or weak solution itself (Ruffenach et al., 22 Apr 2025).

The open program is therefore twofold. One branch seeks rigorous continuum theorems linking vanishing-noise, vanishing-viscosity Navier–Stokes limits to nontrivial probability measures on weak Euler solutions. The other seeks a structural explanation—often RG-based—for why any such measure, if it exists, should be universal. Current evidence strongly suggests that the right asymptotic object in singular turbulent regimes may be a stochastic Eulerian law rather than a single inviscid solution, but that claim remains established only in models, surrogates, and specially constructed transport problems, not yet in full 3D turbulence (Mailybaev et al., 27 Feb 2026).