betaFeedback

WorksPhysics and Astronomy

Spontaneous stochasticity amplifies even thermal noise to the largest scales of turbulence in a few eddy turnover times

Abstract

How predictable are turbulent flows? Here we use theoretical estimates and shell model simulations to argue that Eulerian spontaneous stochasticity, a manifestation of the non-uniqueness of the solutions to the Euler equation that is conjectured to occur in Navier-Stokes turbulence at high Reynolds numbers, leads to universal statistics at finite times, not just at infinite time as for standard chaos. These universal statistics are predictable, even though individual flow realizations are not. Any small-scale noise vanishing slowly enough with increasing Reynolds number can trigger spontaneous stochasticity and here we show that thermal noise alone, in the absence of any larger disturbances, would suffice. If confirmed for Navier-Stokes turbulence, our findings would imply that intrinsic stochasticity of turbulent fluid motions at all scales can be triggered even by unavoidable molecular noise, with implications for modeling in engineering, climate, astrophysics and cosmology.

Spontaneous stochasticity is a recently discovered phenomenon [2, 11, 15] in turbulent flows [4, 5], whereby solutions of model fluid equations remain unpredictable and stochastic due to diverging Lyapunov exponents in the high Reynolds number limit, even though random perturbations to the flow are negligible asymptotically. It is an open question as to whether or not it occurs in real fluids or in the Navier-Stokes equations. Here we report that at large (but finite) Reynolds numbers, a stochastic wave propagating from small to large scales, as first postulated by Lorenz [23], rapidly randomizes the large-scale flow, even if the only sources of noise to trigger the wave are molecular fluctuations at small scales. Going beyond Lorenz, we show that flow fluctuations at large-scales exhibit universal statistics due to spontaneous stochasticity and not directly due to whatever small-scale noise triggers the stochastic wave. Spontaneous stochasticity is not inevitable; for it to be triggered, the small-scale noise must become negligible in the large Reynolds number limit sufficiently slowly. The surprise is that even thermal noise satisfies this condition. These new indirect effects at large scales are distinct from a growing body of work showing that thermal noise directly alters the turbulent dissipation range below the Kolmogorov scale [2, 7, 8, 9, 11, 13, 15, 16, 25, 30], and in fact other small-scale disturbances, typically much larger than thermal agitation, will produce indistinguishable large-scale stochasticity. To demonstrate our claim that the probability distributions of relevant flow quantities have a universal, non-delta-function form at large but finite Reynolds numbers, we are forced to employ a simplified but well-studied dynamical model [17] of turbulence, since the Navier-Stokes equations are computationally unfeasible at the large Reynolds numbers required for convincing convergence to universal statistics. These results suggest an essential indeterminism of turbulent flows at scales of practical interest, with potentially far-ranging implications for engineering, geophysics, and astrophysics.

Fluctuating Hydrodynamics. The fluctuating hydrodynamics of Landau-Lifshitz [18] describes the effect of thermal noise in fluid flows by including fluctuating stresses into the Navier-Stokes equation. It is expressed in a form non-dimensionalized by large-scale velocity UU and outer/integral length LL as

∂tu+(u⋅∇)u=−∇p+1ReΔu+Θ∇⋅ξ+Ff,(1)\partial_t \mathbf{u}+(\mathbf{u}\cdot\nabla)\mathbf{u}=-\nabla p+\frac{1}{Re}\Delta\mathbf{u}+\sqrt{\Theta}\nabla\cdot\boldsymbol{\xi}+F\mathbf{f}, \tag*{(1)}

where the fluctuating stress is modeled as a Gaussian random field ξ\boldsymbol{\xi} with mean zero and covariance

⟨ξij(x,t)ξkl(x′,t′)⟩=(δikδjl+δilδjk−23δijδkl)×δ3(x−x′)δ(t−t′).(2)\langle\xi_{ij}(\mathbf{x},t)\xi_{kl}(\mathbf{x}',t')\rangle=\left(\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk}-\frac{2}{3}\delta_{ij}\delta_{kl}\right)\times\delta^3(\mathbf{x}-\mathbf{x}')\delta(t-t'). \tag*{(2)}

Here, Reynolds number is defined as Re≡UL/νRe\equiv UL/\nu where UU is the large-scale velocity of the flow, LL is the forcing scale at which energy is injected, and ν\nu is kinematic viscosity. Θ≡2νkBT/ρL4U3\Theta\equiv2\nu k_B T/\rho L^4U^3 appears due to the presence of noise and characterizes its strength according to the fluctuation-dissipation relation. kBk_B is Boltzmann’s constant, TT is absolute temperature, and ρ\rho is mass density. To drive turbulence, we have added an external forcing f\mathbf{f} non-dimensionalized by r.m.s. value frmsf_{rms}, with prefactor F=Lfrms/U2F=Lf_{rms}/U^2 setting the magnitude. Strictly speaking, equation (6) describes a coarse-grained velocity field on lengthscale Λ−1\Lambda^{-1} larger than the mean free path and some care is needed to interpret this mathematically, see Supplemental Material (SM §I), for details [1, 3, 4, 5].

Using flow parameters characteristic of the atmospheric boundary layer (ABL) [24] T=300∘ KT=300^\circ\,\mathrm{K}, ν=1.5×10−5 m2/sec\nu=1.5\times10^{-5}\,\mathrm{m^2/sec}, ρ=1.2 kg/m3\rho=1.2\,\mathrm{kg/m^3}, ϵ=4×10−2 m2/sec3\epsilon=4\times10^{-2}\,\mathrm{m^2/sec^3}, L=103 mL = 10^{3}\,m, U=3.42 m/secU = 3.42\,m/\mathrm{sec}, where ε\varepsilon is the mean energy dissipation per mass, we find Θ≃2.59×10−39\Theta\simeq2.59 \times10^{-39}, naively justifying dropping the fluctuating stress term from the Landau-Lifschitz equations and supporting the conventional wisdom that physically relevant turbulent fluid flows are well modeled by the deterministic (cut-off) Navier-Stokes equations. This conclusion is consistent with numerical findings [2, 13, 25] that the turbulent steady-state statistics at scales larger than the Kolmogorov length η=ν3/4ε−1/4\eta= \nu^{3/4}\varepsilon^{-1/4} are unchanged by molecular fluctuations, but not addressing the issue of the flow predictability.

Fully-developed fluid turbulence in the ABL has Re≃2.28×108Re \simeq2.28 \times10^{8}, justifying dropping as well the viscous term proportional to Re−1Re^{-1} from the equations [25]. However, the limiting equations with no viscosity, no wavenumber cutoff, and no thermal noise are the continuum Euler equations, which do not have unique solutions and are formally ill-posed [16, 17, 18]. Such non-uniqueness or “flexibility” of solutions suggests an intrinsic unpredictability of turbulent fluid motions at high Reynolds numbers known as Eulerian spontaneous stochasticity [14, 15], and provides a possibility for tiny thermal noise to influence the predictability of all scales of the flow up to the largest.

Spontaneous Stochasticity. Spontaneous stochasticity can be given a precise meaning through a probability distribution on solutions of the governing stochastic differential equation. In the case of Eulerian spontaneous stochasticity triggered by thermal noise the corresponding equation is (6). Solution of this equation may be expressed in terms of a transition probability density PRe,Θ(uf,tf∣ui,ti)P_{Re,\Theta}(u_f,t_f\mid u_i,t_i) as follows

PRe,Θ(uf,tf)=∫Dui PRe,Θ(uf,tf∣ui,ti)PRe,Θ(ui,ti),(3)P_{Re,\Theta}(u_f,t_f) = \int\mathcal{D}u_i\,P_{Re,\Theta}(u_f,t_f\mid u_i,t_i)P_{Re,\Theta}(u_i,t_i), \tag*{(3)}

where PRe,Θ(ui/f,ti/f)P_{Re,\Theta}(u_{i/f},t_{i/f}) is a probability distribution of velocity fields at the initial/final time. The transition probability PRe,Θ(uf,tf∣ui,ti)P_{Re,\Theta}(u_f,t_f\mid u_i,t_i) satisfies the Fokker-Planck equation corresponding to Langevin equation (6), and it is parameterized by the non-dimensional numbers ReRe and Θ\Theta. In the limit of zero noise Θ→0\Theta\to0 with ReRe fixed the transition probability becomes deterministic, that is, expressed as a delta-distribution on the unique solution of the limiting deterministic problem. The issue of uniqueness of solutions plays the central role in emergence of spontaneous stochasticity: see SM §II [19] for more details. Note that the limiting equations for the fluctuating hydrodynamics of Landau-Lifschitz in the limit Θ→0\Theta\to0 is Navier-Stokes equations with the finite cut-off Λ\Lambda, so that uniqueness of solutions to the Cauchy initial-value problem is then elementary and well-known. This physically necessary cut-off is crucial, since the uniqueness of Leray solutions of the continuum Navier-Stokes equations is a major open problem in pure mathematics [30]. However, if the zero-noise limit is taken together with Re→∞Re \to\infty and Λ∝Re→∞\Lambda\propto Re \to\infty, it leads to singular Euler dynamics with no unique solutions. Then the limiting transition probability may cease to become deterministic

PRe,Θ(uf,tf∣ui,ti)→Re→∞Θ→0P∞(uf,tf∣ui,ti).(4)P_{Re,\Theta}(u_f,t_f\mid u_i,t_i) \xrightarrow[ Re\to\infty]{ \Theta\to0} P_{\infty}(u_f,t_f\mid u_i,t_i). \tag*{(4)}

If such a non-trivial limiting transition probability exists, the limit is called spontaneously stochastic and corresponds to stochastic behavior of a formally deterministic Euler system, which each realization of the limiting distribution satisfies in a weak sense [11, 12, 13, 14, 15].

Because the spontaneously stochastic limit is a double limit Re−1,Θ→0Re^{-1},\Theta\to0, there is no unique way to arrive at it. Furthermore, if the noise strength is taken to zero sufficiently fast, the limit becomes deterministic. In many cases of practical importance, such as turbulent flows past a grid or a cylinder, experimental evidence [33, 34] points at a non-vanishing energy dissipation with the limiting dissipation rate satisfying a relation first proposed by Taylor Taylor [35]: ε=CU3/L\varepsilon= CU^{3}/L, where CC is a dimensionless constant. In this scenario we can control the macroscopic flow parameters UU (or ε\varepsilon) and LL, while ν\nu, TT, and ρ\rho are fixed material parameters of the fluid. These considerations motivate as a “canonical limit” the one obtained by fixing the ratio U3/LU^{3}/L, along with ν\nu, TT, and ρ\rho, while taking Re→∞Re \to\infty. This leads us to define the second non-dimensional number as θη=2kBT/ρν1/4ε−1/4\theta_{\eta} = 2k_{\mathrm{B}}T/\rho\nu^{1/4}\varepsilon^{-1/4} (see [2]) and consider the limit Θ=Re−15/4θη→0\Theta= Re^{-15/4}\theta_{\eta} \to0 with θη\theta_{\eta} held constant. This is the physically relevant continuum limit in which also Λ∝Re→∞\Lambda\propto Re \to\infty, describing fully-developed 3D hydrodynamic turbulence.

Numerical Verification of Eulerian Spontaneous Stochasticity. In order to check that the limit (4) is indeed spontaneously stochastic we need to simulate Landau-Lifshitz equations at very high Reynolds numbers. State-of-the-art simulation can achieve only Re≈500Re \approx500 [25, 26] for incompressible flows, so we use here the Sabra model [17], a simplified dynamical model of turbulent cascade that preserves many key features of Navier-Stokes equations (6), but discretizes length scales as ℓn=2−nL\ell_n = 2^{-n}L and keeps only one complex mode unu_n to represent velocity increments δℓu∝∣un∣\delta_\ell u \propto|u_n| at scale ℓn\ell_n. We also include a stochastic term to model thermal noise, and a deterministic forcing fnf_n that acts only at large scales. In non-dimensionalized form this modified Sabra model is given by the stochastic ODEs:

dundt=i(kn+1un+1un+2∗−12knun+1un−1∗+12kn−1un−1un−2)−1Rekn2un+Θknξn(t)+Ffn,n=1,…,N(5)\frac{d u_n}{d t}=i\left(k_{n+1}u_{n+1}u_{n+2}^{*}-\frac{1}{2}k_nu_{n+1}u_{n-1}^{*}+\frac{1}{2}k_{n-1}u_{n-1}u_{n-2}\right)-\frac{1}{Re}k_n^2u_n+\sqrt{\Theta}k_n\xi_n(t)+Ff_n, \qquad n=1,\ldots,N \tag*{(5)}

where kn=1/ℓnk_n = 1/\ell_n, covariance of the white noise ⟨ξn∗(t)ξm(t′)⟩=2knαδnmδ(t−t′)\langle\xi_n^{*}(t)\xi_m(t')\rangle= 2k_n^{\alpha}\delta_{nm}\delta(t-t'), and the second non-dimensional number group is Θ=Re−βθη\Theta= Re^{-\beta}\theta_{\eta}, $\beta = 3(\alpha + 2)/42)/4. Here we take the number of shells N∝34log⁡2(Re)N \propto\frac{3}{4}\log_2(Re), sufficient to resolve a few shells above the Kolmogorov wavenumber kη=1/ηk_\eta= 1/\eta. The choice α=3\alpha= 3 in the noise covariance is dimensionally identical to 3D Landau-Lifshitz, with β=15/4\beta= 15/4, and it produces also an energy spectrum ∝k2\propto k^2 in the dissipation range, the same as for 3D fluids, but violates the shell-model fluctuation-dissipation relation. On the other hand the choice α=0\alpha= 0 preserves this relation although the equipartition energy spectrum in the dissipation range is ∝k−1\propto k^{-1} rather than ∝k2\propto k^2. Since it is impossible to match exactly all relevant properties of 3D Landau-Lifshitz equations with a single choice of α\alpha, we investigated both choices α=0\alpha= 0 and α=3\alpha= 3 and we find the overall results are insensitive to this choice. We would like to emphasize that for either choice of α\alpha the noise does not serve as a large scale forcing and in fact, together with the viscous damping, it becomes vanishingly small in the limit Re→∞Re \to\infty. For more details on the numerical simulations, including the forcing fnf_n used and the choice of α\alpha, see SM §III-VI [19].

Transition probability density functions for the self-similar initial state with noise scaling $\alpha = 0$

Figure 1a. Self-similar initial state with noise scaling α=0\alpha= 0.

Transition probability density functions for the self-similar initial state with noise scaling $\alpha = 3$

Figure 1b. Self-similar initial state with noise scaling α=3\alpha= 3.

Transition probability density functions for the “burst” initial state with noise scaling $\alpha = 0$

Figure 1c. “Burst” initial state with noise scaling α=0\alpha= 0.

We study the Cauchy problem for (9) with two different deterministic but “quasi-singular” initial data that are not smooth uniformly in Reynolds number. It is convenient to study spontaneous stochasticity with such quasi-singular initial data since with large-scale initial data independent of ReRe one would otherwise have to wait for singularities to form by finite-time blow-up [14]. The first is the Kolmogorov initial datum un=−iAε1/3kn−1/3u_n = -iA\varepsilon^{1/3}k_n^{-1/3}, which is an exact stationary (but unstable) solution of the inviscid, deterministic Sabra model if suitable deterministic forces fnf_n are added to the two lowest shells n=1,2n = 1, 2 [37]. The other initial datum is a “burst” state selected from the ensemble of turbulent steady states of the Sabra model at very high ReRe (see SM §VII [19]). This particular initial datum has approximately a power-law form un∝kn−hu_n \propto k_n^{-h} in the inertial range, with Hölder exponent h≃0.258h \simeq0.258; by construction, this is not intended to be the scaling of the statistical steady-state. Both of these initial data are quasi-singular with exponent h<1h < 1, regularized only at very high-wavenumber either by the cut-off NN or by viscosity ν\nu. The numerical details of how ReRe was varied differs for the two initial data: see SM, §IV for the K41 case, §VI for the “burst” case [19].

The key statistical quantities which we calculate are the probability density functions (PDF’s) of local-in-scale variables, such as the absolute values of velocities at a fixed shellnumbers nn, fixed time tft_f, at an increasing sequence of Reynolds numbers. These reduced PDF’s are integrals over the transition probability densities in (3). Without external noise, these are delta-distributions; see SM, §VII [19]. Presented in Figure 1 are plots of the PDF’s for shell n=18n = 18 and time tf=1t_f = 1 (a,b), tf=1.477×10−3t_f = 1.477 \times10^{-3} (c), where panels (a),(b) are for the K41 initial datum with noise exponents α=0\alpha= 0 and α=3\alpha= 3, respectively, and panel (c) is for the “burst” initial datum with α=0\alpha= 0. As seen clearly, the PDFs converge with increasing ReRe to non-delta distributions and therefore do not become deterministic. The direct effects of thermal fluctuations at this scale can be estimated from θn\theta_n and the resulting RMS velocity fluctuations are 4–5 orders of magnitude smaller than the ones shown in Figure 1. Thus, the universal statistics reflect spontaneous stochasticity not direct effects of thermal noise. We have obtained similar results for the PDFs of other scale-local variables, e.g. energy flux Πn\Pi_n (see SM §VIII [19]).

Transition probability density functions for the absolute values $|u_n|$

Figure 1. Transition probability density function for the absolute values ∣un∣|u_n| at a single fixed shellnumber n=18n = 18 and time tf=1t_f = 1 (a,b), tf=1.477×10−3t_f = 1.477 \times10^{-3} (c) in inertial units for Reynolds numbers spanning almost two decades. The bottom axis represents the inertial range units, while the top axis represents the SI units for the ABL parameters. All the errors are estimated as standard errors using the bootstrap method. [19] contains details how Reynolds number was varied, SM §IV for the K41 initial condition, SM §VI for the “burst” initial condition.

These observations constitute our crucial numerical evidence for Eulerian spontaneous stochasticity triggered by thermal noise in the Sabra model and, presumably, for the Landau-Lifshitz equations. The two cases in panels (a) and (b) correspond to the same initial datum and the same limiting equations when Re→∞Re \to\infty, but a different scale-by-scale approach towards it. Nevertheless, the limiting probability distributions are the same, and independent of the type of regularization and the type of noise that triggers random perturbations.

Inverse error cascade and stochastic wave. What causes this unpredictability if the direct effects of thermal noise are too small? The mechanism was first suggested by Lorenz: an inverse cascade of error [23] that has since been extensively studied [31, 32, 38, 40]. Perhaps the simplest way to illustrate this mechanism is to look at the time-dependent variances Var⁡[un]=E[∣un−E[un]∣2]\operatorname{Var}[u_n] = \mathbb{E}[|u_n - \mathbb{E}[u_n]|^2] calculated across an ensemble of noise realizations with fixed initial datum. These are shown in Figure 2 for the K41 datum. Initially, variances at all scales exhibit diffusive linear growth in time, with higher rate at larger knk_n. Next, modes become chaotic scale-by-scale, starting from high wavenumbers, and eventually the variance for a particular shell saturates when it reaches twice the average energy at that scale. In the early stage of development of the stochastic wave the total variance of the system Var⁡(u)=∑nVar⁡(un)\operatorname{Var}(u) = \sum_n \operatorname{Var}(u_n) grows exponentially (see SM §IX [19]), and this regime is fully consistent with work of Ruelle [30] on the effects of thermal noise in predictability of developed turbulence. However, when the stochastic front starts to propagate across the inertial range [14] [42] the system enters the spontaneously stochastic regime. In the case of a self-similar initial state u0,n∝Akn−hu_{0,n} \propto A k_n^{-h} the front is self-similar, located at length scale ℓ(t)=(At)1/(1−h)\ell(t) = (At)^{1/(1-h)} with amplitude u(t)=(Ath)1/(1−h)u(t) = (At^h)^{1/(1-h)} at time tt. Plotted as Var⁡(un)/u2(t)\operatorname{Var}(u_n)/u^2(t) versus knℓ(t)k_n\ell(t), the curves collapse for the three late times t=12.14,48.56,194.24 st = 12.14, 48.56, 194.24\ \mathrm{s}. For more details on the stages of the stochastic wave formation and propagation, see SM §IX [19]. Furthermore, after the stochastic front passes some scale the statistics of Kolmogorov multipliers [36] at that scale converges to the steady state distribution. Such super-universality has been observed before in [14]; see SM §X [19]. We draw attention to the striking resem-

Energy and velocity variances across the ensemble

Figure 2. Twice ensemble average energy E[ϵn]\mathbb{E}[\epsilon_n] (orange) for ϵn=12∣un∣2\epsilon_n = \frac{1}{2}|u_n|^2 and velocity variances (blue, red, green, black) across the ensemble as a function of wavenumber in SI units for 4 increasing times. The smallest time in the variances plots the initial non-self-similar transient and the subsequent three times show the self-similar propagation of the stochastic wave towards large scales. 2E[ϵn]2\mathbb{E}[\epsilon_n] is almost unchanged in time and forms the envelope of the propagating wave.

blance of our Figure 2 to Figure 2 of Lorenz [23], which he obtained for 2-dimensional Euler equations using a turbulence closure model. For the analogous plot with the “burst” initial state, see SM §XI [19], where the same picture holds qualitatively, although there is no exact self-similarity. The large spontaneous fluctuations illustrated in Figure 1(a) are thus due to effects of thermal noise in the dissipation range which are propagated up into the inertial range by nonlinear error cascade, and not due to the direct local effects of thermal noise.

An important feature of this “inverse error cascade” is that in the inertial range the universal statistical distributions are achieved at each length-scale ℓ\ell in a time which is a constant multiple of the eddy-turnover time τℓ=ℓ/uℓ\tau_\ell= \ell/u_\ell, indifferent to the noise magnitude. This should be contrasted with a predictability horizon in conventional chaotic systems, which is dependent on the noise strength [15, 23, 24] [45]. To illustrate this point, Figure 3 shows the randomization time tr(n)t_r(n), defined as the time when the nnth shell’s variance reaches its ensemble average energy, plotted versus index nn. As is clear from the figure, the randomization times above the lengthscale of 10 cm for the flow parameters of the ABL are given by tr(n)=3.4ϵ−1/3ℓn2/3t_r(n) = 3.4\epsilon^{-1/3}\ell_n^{2/3}. Therefore we conclude that the length scales of about the size of a coffee mug and above in 3D ABL turbulence behave in a spontaneously stochastic fashion. In the SM §XII [19] we provide a theoretical estimate on dimensional grounds of that length scale as a function of ReRe and Θ\Theta. Crucially, we observe that tr(n)t_r(n) approaches the asymptotic value ∝ϵ−1/3ℓn2/3\propto\epsilon^{-1/3}\ell_n^{2/3} for any shell nn in the limit Re→∞Re \to\infty: see inset of Figure 3. Thus all scales are spontaneously stochastic in that idealized limit.

Local randomization times

Figure 3. Local randomization times tr(n)t_r(n) as a function of length-scale ℓn=2−nL\ell_n = 2^{-n}L for the K41 initial datum. tr(n)t_r(n) is defined as the time in which the nnth shell’s variance reaches the ensemble average energy E[ϵn]\mathbb{E}[\epsilon_n]. The inset plot depicts tr(18)t_r(18) as a function of Reynolds number.

Discussion. It is important to emphasize our finding that the spontaneous large-scale statistics are universal with respect to the small-scale noise that triggers them, as long as the noise amplitude becomes negligible with respect to the deterministic equation more slowly than some ReRe-dependent threshold. On dimensional grounds we estimate this threshold to be ∼exp⁡(−Re)\sim\exp(-\sqrt{Re}) as Re→∞Re \to\infty (see SM §XII [19]). Even the inevitably present molecular noise satisfies this criterion and our simulations suggest that it is sufficient to trigger spontaneous stochasticity. In one turnover time of the largest 3D turbulent eddies the unknown molecular motions will impact the evolution, rendering only statistical predictions possible.

Our work has implications for turbulence across multiple scales. For climate models, even if the projected goal of 1 km horizontal resolution in the next decade is achieved [45], such refined resolution will not obviate the need for stochastic models [24] [45, 46, 47]. For the dynamics of galaxy formation, it has already been shown that microscale chaos and stochasticity lead to large variations in star-formation histories and distribution of stellar mass [48], and our results suggest that these effects may be even more severe than currently thought. At the large scale of hydrodynamic simulations of cosmological galaxy formation, the sensitivity of simulations to minute perturbations has also been examined with regard to chaotic dynamics [49], and would be expected to be amplified further by the results we have discussed [50]. Closer to home, there have been recent efforts to reconstruct best-fit individual solution trajectories of Navier-Stokes equations using variational data assimilation techniques [51, 52]. It is already recognized that these reconstruction problems are highly ill-conditioned due to chaotic dynamics. The inclusion of spontaneous stochasticity into this program poses even more severe limitations and implies that a well-posed problem is instead the reconstruction of statistical ensembles of solutions [53, 54]. These examples show that there are many potential ramifications of spontaneous stochasticity in turbulence and related phenomena. It will be important to determine if our findings, based admittedly on a shell model, are valid beyond the necessary simplifications entailed in our work.

This work was partially supported by the Simons Foundation through Targeted Grant “Revisiting the Turbulence Problem Using Statistical Mechanics” (Grant Nos. 663054 (G.E.) and 662985(N.G.)). A.A.M. was supported by the CNPq grant 308721/2021-7 and FAPERJ grant E-26/201.054/2022.

Landau-Lifschitz Fluctuating Hydrodynamics and the Role of the UV Cut-Off

Here we review briefly the equations of Landau-Lifschitz fluctuating hydrodynamics and the role of the wavenumber cut-off Λ\Lambda. We explain also why taking the limit Λ→∞\Lambda\to\infty at fixed ReRe, Θ\Theta to obtain a continuum SPDE is physically irrelevant to the understanding of 3D fluid turbulence.

The fluctuating hydrodynamics equations for an incompressible fluid are given in dimensional form by:

∂tu+(u⋅∇)u=−∇p+ν△u+2νkBTρ∇⋅ξ,∇⋅u=0(6)\partial_t\mathbf{u}+(\mathbf{u}\cdot\nabla)\mathbf{u}=-\nabla p+\nu\triangle\mathbf{u}+\sqrt{\frac{2\nu k_{\mathrm{B}}T}{\rho}}\nabla\cdot\boldsymbol{\xi},\qquad\nabla\cdot\mathbf{u}=0 \tag*{(6)}

where ν\nu is kinematic viscosity, TT is absolute temperature, kBk_{\mathrm{B}} is Boltzmann’s constant, ρ\rho is mass density, and the fluctuating stress is modeled as a Gaussian random matrix field ξ\boldsymbol{\xi} with mean zero and covariance

⟨ξij(x,t)ξkl(x′,t′)⟩=(δikδjl+δilδjk−23δijδkl)δΛ3(x−x′)δ(t−t′).(7)\left\langle\xi_{ij}(\mathbf{x},t)\xi_{kl}(\mathbf{x}',t')\right\rangle=\left(\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk}-\frac{2}{3}\delta_{ij}\delta_{kl}\right)\delta_{\Lambda}^{3}(\mathbf{x}-\mathbf{x}')\delta(t-t'). \tag*{(7)}

For example, see [1, 2]. Here we consider these equations for simplicity in a periodic box and note that δΛ3(r)\delta_{\Lambda}^{3}(\mathbf{r}) in (7) is a “cut-off delta-function”, defined as [3]

δΛ3(x−x′)=1V∑∣k∣<Λeik⋅r.(8)\delta_{\Lambda}^{3}(\mathbf{x}-\mathbf{x}')=\frac{1}{V}\sum_{\lvert\mathbf{k}\rvert<\Lambda}e^{i\mathbf{k}\cdot\mathbf{r}}. \tag*{(8)}

In the microscopic derivation of these equations using Zwanzig-Mori techniques, e.g. see [3, 4, 5], the viscosity is given by Green-Kubo-type formulas which show that ν=νΛ\nu=\nu_{\Lambda} must be cut-off dependent. This Λ\Lambda-dependence of the effective viscosity has been studied also by renormalization-group methods for a quiescent fluid in thermal equilibrium [1]. In practical computations the values of νΛ\nu_{\Lambda} are often fixed by comparison with molecular dynamics simulations, as in [6], Appendix C.

The choice of the high-wavenumber cut-off Λ\Lambda in the Landau-Lifschitz equations (6) is somewhat arbitrary, but subject to important constraints. The microscopic derivations [3, 4, 5] show that Λ\Lambda must be chosen somewhat smaller than the microscopic wavenumber 1/λmicr1/\lambda_{micr}, where λmicr=max⁡{λmfp,λintp}\lambda_{micr}=\max\{\lambda_{mfp},\lambda_{intp}\} with λmfp\lambda_{mfp} the mean-free-path length and λintp=n−1/3\lambda_{intp}=n^{-1/3} the interparticle spacing (where n=ρ/mn=\rho/m is the number of molecules of mass mm per volume). Note that λmicr=λmfp≫λintp\lambda_{micr}=\lambda_{mfp}\gg\lambda_{intp} in a low density gas and λmicr=λintp≳λmfp\lambda_{micr}=\lambda_{intp}\gtrsim\lambda_{mfp} in a liquid. Empirically, the model (6) is found to be accurate with Λ\Lambda just moderately smaller than 1/λmicr1/\lambda_{micr} (e.g. see [7]). On the other hand, Λ\Lambda must not be chosen too small, so that the only velocity fluctuations which are eliminated by coarse-graining correspond to local thermodynamic equilibrium fluctuations. In a turbulent flow, this implies that, at least, Λ≳1/η\Lambda\gtrsim1/\eta, with η\eta the Kolmogorov length, so that the eliminated wavenumber modes are not subject to turbulent fluctuations. However, as discussed at some length in our previous work [2], intermittency effects can lead to turbulent effects propagating to sub-Kolmogorov lengths much smaller than η\eta. In that case, it would be safer to choose Λ\Lambda just a factor of a few smaller than 1/λmicr1/\lambda_{micr}. Note that the integral length LL and η\eta are related through the scaling L/η≃Re3/4L/\eta\simeq Re^{3/4} using Taylor’s relation ε∝U3/L\varepsilon\propto U^3/L. LL and λmfp\lambda_{mfp} are related through L/λmfp≃Re/MaL/\lambda_{mfp}\simeq Re/Ma using the kinetic theory estimate ν=λmfpc\nu=\lambda_{mfp}c where cc is the sound speed. In dimensionless units in terms of the outer scales, it therefore follows that the cut-off wavenumber at the Kolmogorov-scale is Λ^:=ΛL∼Re3/4\hat{\Lambda}:=\Lambda L\sim Re^{3/4} whereas the cutoff at the molecular scale is Λ^∼Re/Ma\hat{\Lambda}\sim Re/Ma. It should be emphasized that even the latter cut-off may not be sufficient to guarantee validity of fluctuating hydrodynamics. As discussed in [2], extremely large, rare events could produce turbulent fluctuations at the length scale λmicr\lambda_{micr}. If such extreme events occur, then the Landau-Lifschitz equations (6) would break down, at least locally in space-time.

The well-known renormalization group analysis of Forster et al. [1] suggests that there is a UV strong-coupling regime in model (6) in 3D, because the “Reynolds number” Reθ=uΛθ/νΛΛRe_{\theta}=u^{\theta}_{\Lambda}/\nu_{\Lambda}\Lambda of thermal velocity fluctuations uΛθ=(kBTΛ3/ρ)1/2u^{\theta}_{\Lambda}=(k_{B}T\Lambda^3/\rho)^{1/2} increases with Λ\Lambda for fixed ReRe (even for Re≪1Re\ll1!) Although the equations (6) are not continuum stochastic partial-differential equations, because of the cut-off Λ\Lambda, the mathematical methods of stochastic partial differential equations (SPDE’s) developed to study the formal Λ→∞\Lambda\rightarrow\infty limit might be relevant to describe the UV strong-coupling regime at finite ReRe, if the latter occurred in a physical range of wavenumbers. However, simple estimates [2] show that the nonlinear coupling becomes strong in (6) only for Λ−1\Lambda^{-1} of order the mean-free-path length, or smaller, where any hydrodynamic description breaks down! This is in contrast to the situation for similar stochastic models in less than two dimensions, such as 1D KPZ, where the RG analysis of [1] showed that the strong-coupling regime occurs instead in a physical regime at low wavenumbers. Thus, the methods of stochastic PDE’s to study the limit Λ→∞\Lambda\rightarrow\infty for KPZ [8] are physically very relevant, since that limit corresponds to the nontrivial IR-attractive RG fixed point which governs KPZ scaling at low wavenumbers [9]. In other words, the infrared limit k≪Λk\ll\Lambda for a microscopic stochastic growth model with physical UV cutoff Λ\Lambda corresponds mathematically, by a simple rescaling, to holding wavenumber kk fixed and taking Λ→∞\Lambda\rightarrow\infty.

The latter limit should be clearly distinguished from that considered in our work. Here we keep the dimensional UV cutoff Λ\Lambda fixed at some value between 1/η1/\eta and 1/λmicr1/\lambda_{micr}, while taking Re=UL/ν→∞Re=UL/\nu\rightarrow\infty. In this limit, the outer-scaled cutoff Λ^\hat{\Lambda} lies between Re3/4Re^{3/4} and Re/MaRe/Ma and thus Λ^→∞\hat{\Lambda}\rightarrow\infty, similar to KPZ. Simultaneously, however, the dimensionless viscosity ν^=1/Re\hat{\nu}=1/Re and thermal noise Θ=Re−1/4θη\Theta=Re^{-1/4}\theta_{\eta} both vanish, so that the solution of the Cauchy problem for (6) with initial data u0\mathbf{u}_{0} is described in this limit by a statistical ensemble of solutions to deterministic Euler equations, each with initial datum u0\mathbf{u}_{0} [10]. The infinite ReRe limit is governed not by an SPDE but instead by a deterministic PDE, the Euler equations, with spontaneously stochastic solutions.

Spontaneous Stochasticity and Non-uniqueness of Solutions

The concept of spontaneous stochasticity was first introduced in the seminal paper [11] that studied the Kraichnan model of passive scalar advection (see also [12]). This work showed that trajectories of Lagrangian particles advected by a rough velocity field remain stochastic even in the limit of zero noise, a phenomenon now known as Lagrangian spontaneous stochasticity. The limiting distribution is over the set of non-unique particle trajectories and is responsible for the anomalous dissipation of scalars advected by turbulence [11, 13] and the emergence of universal particle statistics in finite time. This is in sharp contrast to chaos, where universality is associated with infinite times. Eulerian spontaneous stochasticity is a related phenomenon with the limiting distribution over the non-unique velocity fields solving the Euler equation [14, 15]. Recently, by PDE methods of convex integration it has been shown that there is at least a dense set of “rough” initial data for which infinitely many solutions of the Cauchy problem for the Euler equations exist, even if one imposes the conditions that the Euler solutions are locally dissipative and Hölder continuous in spacetime [16, 17, 18] as required for the energy cascade by Onsager’s theorem [19, 20]. Such non-uniqueness, as illustrated by the more tractable example of Lagrangian spontaneous stochasticity, may lead to Eulerian spontaneous stochasticity, as has been recently verified rigorously in some simple toy models [21, 22].

It is clear in retrospect that Lorenz in his seminal work [23] anticipated several key components of the modern concept of spontaneous stochasticity. Lorenz understood the essential difference between now-standard chaos and spontaneous stochasticity, stating that the predictability horizon of the latter would not increase with decreasing noise. Lorenz recognized also that this fundamental property would lead to essentially stochastic dynamics. In his own words from [23], “certain formally deterministic fluid systems which possess many scales of motion are observationally indistinguishable from indeterministic systems”. What is added here is the clear understanding that spontaneous stochasticity is a phenomenon that occurs in a singular limit such as Re→∞Re \to\infty [11]. In this limit, Lyapunov exponents are not merely positive, as in standard chaos, but in fact diverge to positive infinity. In Eulerian stochasticity the divergence happens at small scales, while in the Lagrangian case divergence happens at rough points of the fluid flow. Ultimately this is the root cause of the qualitative difference in predictability horizon of chaotic and spontaneously stochastic systems. We furthermore stress the importance of the insight of [11] that connects spontaneous stochasticity with non-unique solutions of a limiting singular initial-value problem. It is precisely the non-uniqueness of singular solutions of the limiting ideal dynamics which permits the randomness from vanishingly small stochastic perturbations to persist and to yield robust, universal statistics in a finite time, independent of the noise source. The “formally deterministic fluid system” envisioned by Lorenz to have stochastic solutions must be the ideal Euler equation, and not the viscous Navier-Stokes equation. The latter equation at finite Reynolds number exhibits spontaneous statistics only if it is subject to tiny random perturbations, either in the dynamics or in the initial data, which vanish in the limit Re→∞Re \to\infty. The stochastic behavior of solutions of the deterministic Euler equations requires not merely “many scales of motion”, but in fact infinitely many infinitesimally-small scales in the asymptotic high-ReRe limit. We believe that identification of the “real butterfly effect” [24] of Lorenz with Eulerian spontaneous stochasticity clarifies the mathematical foundations of this effect and should accelerate further progress.

Shell Model and Numerical Methods

The recent state-of-the-art simulations of Bell et al. [25] could attain only Re=554Re = 554, which is orders of magnitude below the value in the atmospheric boundary layer, for example. Likewise, Gallis et al. [26] using the molecular gas dynamics method of Direct Simulation Monte Carlo (DSMC) could achieve only Re=500Re = 500 for nearly incompressible flow and Re=2000Re = 2000 for supersonic flow. While this latter paper did observe effects of thermal noise on large scales, their Reynolds numbers were not high enough to verify the convergence of statistics which is the signature of spontaneous stochasticity. Therefore to verify our claims it was necessary to use the Sabra shell model supplemented with an appropriate thermal noise term.

Dynamics of velocities unu_n in the noisy Sabra model in its nondimensional form is governed by the following system of stochastic ODEs (equation [9] in the main text):

dundt=∇n[u,u∗]+Θκnξn(t)=Bn[u,u∗]+fn−1Rekn2un+Θκnξn(t)=i(kn+1un+2un+1∗−12knun+1un−1∗+12kn−1un−1un−2)−1Rekn2un+Θκnξn(t)(9)\frac{d u_n}{d t} = \nabla_n[u,u^*] + \sqrt{\Theta}\kappa_n\xi_n(t) = B_n[u,u^*] + f_n - \frac{1}{Re}k_n^2u_n + \sqrt{\Theta}\kappa_n\xi_n(t) = i\left(k_{n+1}u_{n+2}u^*_{n+1} - \frac{1}{2}k_nu_{n+1}u^*_{n-1} + \frac{1}{2}k_{n-1}u_{n-1}u_{n-2}\right) - \frac{1}{Re}k_n^2u_n + \sqrt{\Theta}\kappa_n\xi_n(t) \tag*{(9)}

where kn=2n/Lk_n = 2^n/L are wavenumbers, Re=UL/νRe = UL/\nu is Reynolds number, fnf_n represent a deterministic external forcing to drive turbulence, and ξn\xi_n are white noises that model thermal fluctuations, so that Θ=2νkBT/ρLα+3\Theta= 2\nu k_{\mathrm{B}}T/\rho L^{\alpha+3} is the second dimensionless number due to the presence of thermal noise term. The latter depends upon the scaling power α\alpha of the covariance of white noise ⟨ξn∗(t)ξm(t′)⟩=2δnmknαδ(t−t′)\langle\xi^*_n(t)\xi_m(t')\rangle= 2\delta_{nm}k_n^\alpha\delta(t-t'), which is an important degree of freedom at our disposal. As noted in the main text, we may write alternatively Θ=Re−βθη\Theta= Re^{-\beta}\theta_\eta with θη=2kBT/ρν(3α+2)/4ε(2−α)/4\theta_\eta= 2k_{\mathrm{B}}T/\rho\nu^{(3\alpha+2)/4}\varepsilon^{(2-\alpha)/4} and with ρ\rho assigned dimensions (mass)/(length)α^\alpha. In our work we studied two different values, α=0\alpha= 0 and α=3\alpha= 3. The noise spectrum given by α=0\alpha= 0 leads to high-wavenumber modes that are in equipartition and form a thermal bath. This choice implies that there is detailed balance in the stochastic dynamics of small-scale modes. This can be seen from the Fokker-Planck operator L\mathcal{L}

LP=∑n=0N(−∂∂un(Vn[u,u∗]P)−∂∂un∗(Vn∗[u,u∗]P)+2Θkn2+α∂2∂un∂un∗P)(10)\mathcal{L}P = \sum_{n=0}^{N}\left(-\frac{\partial}{\partial u_n}\left(V_n[u,u^*]P\right)-\frac{\partial}{\partial u^*_n}\left(V^*_n[u,u^*]P\right)+2\Theta k_n^{2+\alpha}\frac{\partial^2}{\partial u_n\partial u^*_n}P\right) \tag*{(10)}

of stochastic dynamics (9) with fn=0f_n = 0, which is self-adjoint with respect to the Gaussian equilibrium measure PG[u,u∗]=exp⁡{−1Re Θ∑n=0Nϵn}P_{\mathrm{G}}[u,u^*] = \exp\left\{-\frac{1}{Re\,\Theta}\sum_{n=0}^{N}\epsilon_n\right\}, ϵn=12∣un∣2\epsilon_n = \frac{1}{2}|u_n|^2 for α=0\alpha= 0. See [2], Appendix A. In molecular fluids, the validity of the hydrodynamic approximation is predicated on the assumption of local thermodynamic equilibrium. This is indeed equivalent to high-wavenumber modes remaining in thermal equilibrium, which is why the case α=0\alpha= 0 is a good choice to represent this feature of molecular fluids.

In contrast, the noise spectrum given by α=3\alpha= 3 leads to breaking of detailed balance and therefore the high-wavenumber modes are not in equipartition. However, in this case the energy spectrum En:=ϵn/knE_n := \epsilon_n/k_n scales as kn2k_n^2, as is expected in 3D fluids, and thus closer to 3D turbulence from this perspective. The kn2k_n^2-scaling can be expected from the balance of the viscous and noise term in (9), neglecting the nonlinear interactions. To numerically verify this expectation, we ran a simulation with the K41 initial data un=−iAϵ1/3kn−1/3u_n = -i A\epsilon^{1/3}k_n^{-1/3}. Figure 4 shows the modal energy ϵn\epsilon_n averaged over different noise realizations as a function of wavenumber at time t=1t = 1 in dissipation units. The modal energy at high-wavenumbers is fit to a power-law scaling, with fit depicted in green. The power-law fit has the power of 3.00, and thus the spectrum En∝kn2E_n \propto k_n^2, as expected. Since it is impossible to match exactly all of the relevant properties of 3D Landau-Lifschitz equations with a single choice of α\alpha, we investigated both choices α=0\alpha= 0 and α=3\alpha= 3 and the overall results are insensitive to this choice.

Energy ensemble average as a function of wavenumber

Figure 4. Energy ensemble average as a function of wavenumber at time t=1t = 1 in dissipation units for the K41 initial datum in a simulation with noise exponent α=3\alpha= 3. The modal energy at high-wavenumbers is fit to power-law scaling, with fit depicted in green. The power-law fit has the power of 3.00

Stochastic integration of (9) has been performed using a slaved Taylor-Ito scheme from [27] which is nominally 32\frac{3}{2}-order but, in practice, 2nd-order. The details of the derivation of the scheme for stochastic Sabra model may be found in reference [2], and here we just state the update rule for complex velocity unu_n

un(tk+1)=e−νkn2Δt{un(tk)+Δt[Bn(tk,u(tk))+fn(tk)]+12(Δt)2[νkn2[Bn(tk,u(tk))+fn(tk)]+f˙n(tk)]+12(Δt)2∑m[am∂Bn∂um(tk,u(tk))+am∗∂Bn∂um∗(tk,u(tk))+2bm2∂2Bn∂um∂um∗(tk,u(tk))]+∑mbm[ΔZm(tk)∂Bn∂um(tk,u(tk))+ΔZm∗(tk)∂Bn∂um∗(tk,u(tk))]+bn[(1+νkn2Δt)ΔWn(tk)−νkn2ΔZn(tk)]}(11)\begin{aligned} u_n(t_{k+1}) = e^{-\nu k_n^2\Delta t}\Bigl\{&u_n(t_k)+\Delta t[B_n(t_k,u(t_k))+f_n(t_k)] \\ &+\frac{1}{2}(\Delta t)^2\bigl[\nu k_n^2[B_n(t_k,u(t_k))+f_n(t_k)]+\dot{f}_n(t_k)\bigr] \\ &+\frac{1}{2}(\Delta t)^2\sum_m\Bigl[a_m\frac{\partial B_n}{\partial u_m}(t_k,u(t_k))+a_m^*\frac{\partial B_n}{\partial u_m^*}(t_k,u(t_k)) \\ &\qquad\qquad\qquad\qquad+2b_m^2\frac{\partial^2 B_n}{\partial u_m\partial u_m^*}(t_k,u(t_k))\Bigr] \\ &+\sum_m b_m\Bigl[\Delta Z_m(t_k)\frac{\partial B_n}{\partial u_m}(t_k,u(t_k))+\Delta Z_m^*(t_k)\frac{\partial B_n}{\partial u_m^*}(t_k,u(t_k))\Bigr] \\ &+b_n\bigl[(1+\nu k_n^2\Delta t)\Delta W_n(t_k)-\nu k_n^2\Delta Z_n(t_k)\bigr]\Bigr\} \tag*{(11)} \end{aligned}
Bn(u)=Vn[u,u∗],an=Bn+fn,bn=Θkn,(12)B_n(u)=V_n[u,u^*],\qquad a_n=B_n+f_n,\qquad b_n=\sqrt{\Theta}k_n, \tag*{(12)}

where fnf_n is forcing at shell nn.

Stationary Kolmogorov self-similar solution, Reynolds number and time-inertial units

A few words are required to explain how to define a “Reynolds number” for the K41 solution un=−iAε1/3kn−1/3u_n=-iA\varepsilon^{1/3}k_n^{-1/3}, which has no natural large scale LL. Note that we can truncate this self-similar state at any arbitrarily large scale, the dynamical effect of which is negligible at short enough times. As a suitable proxy for LL, we instead introduce an “observation time” tot_o, which we will employ for non-dimensionalization. To that end we state dimensions of all the basic dimensionful quantities

[t]=[time],[kn]=[length]−1,[u]=[length][time]−1,[f]=[length][time]−2,[t]=[time],\quad[k_n]=[length]^{-1},\quad[u]=[length][time]^{-1},\quad[f]=[length][time]^{-2},
[ν]=[length]2[time]−1,[kBTρ]=[length]2[time]−2,[\nu]=[length]^2[time]^{-1},\quad\left[\frac{k_{\mathrm{B}}T}{\rho}\right]=[length]^2[time]^{-2},
[ε]=[length]2[time]−3,[η]=[time]−1/2.[\varepsilon]=[length]^2[time]^{-3},\quad[\eta]=[time]^{-1/2}.

This leads us to introduce a length-scale ℓo=ε1/2to3/2\ell_o=\varepsilon^{1/2}t_o^{3/2}, which we note from Figure 4 in the main text is the length scale reached by the spontaneously stochastic wave at time tot_o. The corresponding velocity is uo=(εℓo)1/3=(εto)1/2u_o=(\varepsilon\ell_o)^{1/3}=(\varepsilon t_o)^{1/2}. We then non-dimensionalize the equations with these scales to obtain the Sabra model dynamics in “time inertial units” as

dundt=Bn[ui]−1Reokn2un+fn+Θoknξn(t),(13)\frac{du_n}{dt}=B_n[u_i]-\frac{1}{Re_o}k_n^2u_n+f_n+\sqrt{\Theta_o}k_n\xi_n(t), \tag*{(13)}

upon introduction of dimensionless parameters Reo=uoℓoν=εto2νRe_o=\frac{u_o\ell_o}{\nu}=\frac{\varepsilon t_o^2}{\nu} and Θo=2kBTρuo2ℓo2=Reo−3(2+α)/4θη\Theta_o=\frac{2k_{\mathrm{B}}T}{\rho u_o^2\ell_o^2}=Re_o^{-3(2+\alpha)/4}\theta_\eta. In this manner, the limit of large Reynolds number Reo≫1Re_o\gg1 over a unit time interval is achieved by simulating the model at Reo=1Re_o=1, Θo=θη\Theta_o=\theta_\eta for a long “observation time” tot_o and then non-dimensionalizing as discussed above. The Reynolds numbers Re=6.25×106Re=6.25\times10^6, 15.75×10615.75\times10^6, 39.68×10639.68\times10^6, 100×106100\times10^6 reported in Figure 1 were achieved with to=2500t_o=2500, 3968, 6299, 10000.

Although the K41 solution does not have a natural cutoff, we had to introduce one for numerical purposes, employing 40 shells in our computations. In order for the inviscid solution to remain stationary with a cutoff, forcing can be introduced at the two shells that correspond to the largest scales, which are the only shells that are directly affected by the introduction of the cutoff. Dynamics at these two shells is given by

∂tu0=0=B0[ui]+f0=ik1u2u1∗+f0(14)\partial_tu_0=0=B_0[u_i]+f_0=ik_1u_2u_1^*+f_0 \tag*{(14)}
∂tu1=0=B1[ui]+f1=i(k2u3u2∗−12k1u2u0∗)+f1(15)\partial_tu_1=0=B_1[u_i]+f_1=i\left(k_2u_3u_2^*-\frac{1}{2}k_1u_2u_0^*\right)+f_1 \tag*{(15)}

Hence the forcing that ensures stationarity is the following

f0=−ik1u2u1∗(16)f_0=-ik_1u_2u_1^* \tag*{(16)}
f1=−i(k2u3u2∗−12k1u2u0∗)(17)f_1=-i\left(k_2u_3u_2^*-\frac{1}{2}k_1u_2u_0^*\right) \tag*{(17)}

These values of forcing held constant were used in the whole simulation with K41 initial state. To illustrate that, beyond the initial transient, stochastic evolution from the K41 state is self-similar, we plot variances of shell velocities across the noise ensemble as a function of wavenumber for several Reynolds numbers in time-inertial units (see Figure 5). Outside the dissipation range and the region close to the numerical large-scale cutoff the data points collapse to a very good approximation, which provides evidence of self-similarity of stochastic dynamics. An even more refined evidence in terms of probability distributions of the absolute values of shell velocities is given in the main text.

Variances of shell velocities averaged over the noise ensemble

Figure 5. Variances of shell velocities averaged over the noise ensemble for the K41 initial datum, plotted as a function of wavenumber in the time-inertial units scaled with ℓo=ε1/2to3/2\ell_o=\varepsilon^{1/2}t_o^{3/2} and uo=ε1/2to1/2u_o=\varepsilon^{1/2}t_o^{1/2}, for several Reynolds numbers.

Burst initial datum

The “burst” state was generated to represent a quasi-singular state that naturally arises in the dynamics of Sabra model. Because of the very high Reynolds number required, the state was not generated via direct dynamics, but rather by a sequence of three steps. Each step corresponds to evolution of 40 shells with decreasing viscosity ν=106\nu=10^6, 10310^3, 1. At the first step the dynamics was carried out for 10 large eddy turnover times L/UL/U. The second and third steps were carried out for 10410^4 and 10310^3 viscous times respectively. This was done to ensure that the small scales are representative of dynamics with the corresponding viscosity, while keeping overall simulation time manageable. At ν=1\nu=1 the simulation had only 5 shells in the dissipation range, which is marginally adequate for calculating statistical steady states but suffices to construct a single realization. In our subsequent simulation with thermal noise the randomness washes out the fine details of the deterministic evolution in the dissipation range and reduces the resolution requirements [2]. Furthermore, the effects we are examining and measuring occur in the inertial range, which is well-resolved.

Using this procedure we identified the following “burst” state:

u=[2.70338267×103−4.29501775×103i, −5.99644021×102−1.08362024×103i, −1.73526401×102−8.70031976×102i,−6.62990674×102+4.43382806×102i, 3.50125557×102+1.94712011×102i, −2.90199477×102−1.47957982×102i,1.99604442×101+3.17275261×102i, −3.87134201×101−6.38137074×101i, −6.81213766×101+2.00246718×102i,−3.05298025×102+9.49052891×101i, 4.02510712×102+8.48877540×101i, −2.93971753×102−1.75452647×102i,−1.65022863×102−2.13115578×102i, −2.33841272×102+1.26798430×102i, −6.08830647×101+1.50048655×102i,−1.28391586×101−1.20865013×102i, 1.98177167×101+1.51201621×102i, −1.47897620×101+1.33319974×102i,−2.35708969×10−1−1.15215796×102i, −6.96084174×101+6.92793032×101i, −7.05189999×101−5.46971275×101i,−4.89715996×101+3.74218157×101i, −3.41976233×101+4.18048345×101i, 6.54699804×100+1.35739046×101i,1.50962455×101+2.39918367×101i, −2.04000328×101−3.26894560×101i, 1.65019396×101−3.49312633×101i,−2.94633411×10−1−7.39808658×100i, 2.77261064×100+1.71067071×100i, 4.92888927×100+1.19176886×101i,−1.09264397×100−2.89297482×100i, 1.92793947×100+3.61432832×100i, −2.40679385×100+2.32440175×100i,2.66376041×100+1.05559511×100i, 4.77644022×10−1−2.06105535×100i, 7.31776213×10−1+6.51759355×10−1i,8.37075637×10−2+1.93103368×10−1i, −1.07909657×10−2−4.72826940×10−3i, 7.46353392×10−5+3.30852429×10−6i,6.18596728×10−9−1.20561074×10−8i].\begin{aligned} u={}&[2.70338267\times10^3-4.29501775\times10^3i,\,-5.99644021\times10^2-1.08362024\times10^3i,\,-1.73526401\times10^2-8.70031976\times10^2i,\\ &-6.62990674\times10^2+4.43382806\times10^2i,\ 3.50125557\times10^2+1.94712011\times10^2i,\,-2.90199477\times10^2-1.47957982\times10^2i,\\ &1.99604442\times10^1+3.17275261\times10^2i,\,-3.87134201\times10^1-6.38137074\times10^1i,\,-6.81213766\times10^1+2.00246718\times10^2i,\\ &-3.05298025\times10^2+9.49052891\times10^1i,\ 4.02510712\times10^2+8.48877540\times10^1i,\,-2.93971753\times10^2-1.75452647\times10^2i,\\ &-1.65022863\times10^2-2.13115578\times10^2i,\,-2.33841272\times10^2+1.26798430\times10^2i,\,-6.08830647\times10^1+1.50048655\times10^2i,\\ &-1.28391586\times10^1-1.20865013\times10^2i,\ 1.98177167\times10^1+1.51201621\times10^2i,\,-1.47897620\times10^1+1.33319974\times10^2i,\\ &-2.35708969\times10^{-1}-1.15215796\times10^2i,\,-6.96084174\times10^1+6.92793032\times10^1i,\,-7.05189999\times10^1-5.46971275\times10^1i,\\ &-4.89715996\times10^1+3.74218157\times10^1i,\,-3.41976233\times10^1+4.18048345\times10^1i,\ 6.54699804\times10^0+1.35739046\times10^1i,\\ &1.50962455\times10^1+2.39918367\times10^1i,\,-2.04000328\times10^1-3.26894560\times10^1i,\ 1.65019396\times10^1-3.49312633\times10^1i,\\ &-2.94633411\times10^{-1}-7.39808658\times10^0i,\ 2.77261064\times10^0+1.71067071\times10^0i,\ 4.92888927\times10^0+1.19176886\times10^1i,\\ &-1.09264397\times10^0-2.89297482\times10^0i,\ 1.92793947\times10^0+3.61432832\times10^0i,\,-2.40679385\times10^0+2.32440175\times10^0i,\\ &2.66376041\times10^0+1.05559511\times10^0i,\ 4.77644022\times10^{-1}-2.06105535\times10^0i,\ 7.31776213\times10^{-1}+6.51759355\times10^{-1}i,\\ &8.37075637\times10^{-2}+1.93103368\times10^{-1}i,\,-1.07909657\times10^{-2}-4.72826940\times10^{-3}i,\ 7.46353392\times10^{-5}+3.30852429\times10^{-6}i,\\ &6.18596728\times10^{-9}-1.20561074\times10^{-8}i]. \end{aligned}

In Figure 6 we depict the absolute values of the burst state as a function of shell index. The scaling of the absolute values is roughly described by the Hölder exponent h≃1/4h\simeq1/4. The values of u0u_0, u1u_1, u2u_2 given above were used to calculate the constant forcing in the simulation of the dynamics of the burst state from (16) and (17).

Absolute values of shell velocities of a burst initial state as a function of shell index

Figure 6. Absolute values of shell velocities of a burst initial state as a function of shell index n=log⁡2kn+log⁡2Ln = \log_{2} k_n + \log_{2} L, and a power-law fit with exponent −0.258-0.258.

Supplement to Figure 1(c): Varying Reynolds number for the ``burst'' initial condition.

We explain here briefly how we performed the numerical simulations for the “burst” initial datum at different Reynolds numbers Re=UL/νRe=UL/\nu obtained by varying ν\nu. The main computational issue in performing these integrations was the choice of time-step Δt\Delta t. The scaling of the viscous wavenumber is given by

kν∝(ϵν−3)1/4.(18)k_{\nu} \propto(\epsilon\nu^{-3})^{1/4}. \tag*{(18)}

Therefore increasing of viscosity by a factor of 252^{5} decreases the viscous wavenumber by a factor of 215/42^{15/4}, which is equivalent to a shift of index nn by −3.75-3.75. Given that our slaved integration scheme requires νkn2Δt\nu k_n^2 \Delta t to be small, decreasing our truncation index by 4 allowed us to increase the timestep in our simulation as follows

Δt′∝1(25ν)kn−42=23Δt.(19)\Delta t' \propto\frac{1}{(2^{5}\nu)k_{n-4}^{2}} = 2^{3}\Delta t. \tag*{(19)}

The nominal Reynolds numbers achieved by this approach are extremely large, up to 101410^{14}. However, for the study of spontaneous statistics at shell n=18n = 18 in Figure 1, panel (c) these large Reynolds numbers are unrealistically inflated, because the modes with shell numbers n<18n < 18 are essentially frozen on the time-scales of interest and play no role. We have therefore defined Reynolds number for the purpose of Figure 1(c) by the length-scale ℓn\ell_n and root-mean-square velocity UnU_n that correspond to the shell of interest, that is n=18n = 18. The final-time r.m.s. velocity is calculated by an average over noise realizations.

Supplement to Figure 1: Control experiment without thermal noise

As a simple control experiment, we have recalculated the transition PDF’s plotted in Figure 1(a), but now with thermal noise set identically to zero. The results are plotted in Figure 7, where it can be seen that the PDF’s become Dirac delta functions. This is expected because, in the absence of external noise, we are solving deterministic ODE’s in finitely many variables with deterministic initial data. There is thus exactly one solution of the initial-value problem for each Reynolds number and the corresponding unique value of ∣u18∣|u_{18}| lies in a single bin of the histogram plotted below for the control experiment. There is no discernible limit as Re→∞Re \to\infty. In fact, we observe periodicity in ReRe, explaining why the result for Re=6.25×106Re = 6.25 \times10^{6} and Re=108Re = 10^{8} coincide. Our observations are consistent with rigorous results of [28, 29], who proved that solutions of the viscous Sabra shell model will in general converge as Re→∞Re \to\infty to a weak solution of the inviscid Sabra model with infinitely many shells. However, such convergence is only obtained along a suitable subsequence of Reynolds numbers, and the limiting solution may depend on the subsequence. Figure 7 illustrates how essential is the thermal noise (or some external noise) for the results in Figure 1 of the main text.

Transition probability density function for the absolute values $|u_n|$

Figure 7. Transition probability density function for the absolute values ∣un∣|u_n| at a single fixed shellnumber n=18n = 18 and time tf=1t_f = 1, exactly as in Figure 1(a) in the main text, but with thermal noise turned off. The PDF’s for Re=6.25×106Re = 6.25 \times10^6 and Re=108Re = 10^8 are identical and lie on top of each other. For comparison, the original PDF’s from Figure 1(a) in the main text are replotted here as well.

Although the randomness appearing in Figure 1 of the main text is due to the thermal noise, it is not a direct effect of that noise. As discussed in the main text of the letter, the observed randomness is due instead to the “inverse error cascade” of Lorenz, which propagates the effect of thermal noise from the dissipation range up to any inertial-range scale in a few local turnover times. To explain why the direct effects of thermal noise are negligible, let us consider the parameter values α=0\alpha= 0, n=18n = 18, and tf=1t_f = 1 used in Figure 1(a). Note that the direct thermal noise effects are even smaller for the other cases illustrated in Figure 1(b,Figure 1c). The r.m.s. velocity due to the direct effect of thermal noise is urmsθ=(2Θkn2αtf)1/2u_{\mathrm{rms}}^{\theta} = (2\Theta k_n^{2\alpha} t_f)^{1/2} with Θ=Re−βθn\Theta= Re^{-\beta}\theta_n, where for all cases considered we took the value θn=2.83×10−8\theta_n = 2.83 \times10^{-8} typical of the atmospheric boundary layer. Thus, for n=18n = 18 at the lowest considered Reynolds number, Re=6.25×106Re = 6.25 \times10^{6}, we have urmsθ=4.99×10−4u_{\mathrm{rms}}^\theta= 4.99 \times10^{-4} and at the highest Reynolds number, Re=108Re = 10^8, we have urmsθ=6.24×10−5u_{\mathrm{rms}}^\theta= 6.24 \times10^{-5}. These values are obviously negligible compared with the r.m.s velocity of the spontaneous statistics exhibited in Figure 1(a). These large spontaneous values originate from the stochastic wave shown in Figure 2 of the main text, which arrives to shellnumber n=18n = 18 at a dimensionless time t=3.4(218)−2/3=8.3×10−4t = 3.4(2^{18})^{-2/3} = 8.3 \times10^{-4}, asymptotically independent of ReRe. The large spontaneous fluctuations seen in Figure 1(a) are thus due to thermal noise effects in the dissipation range propagated up into the inertial range by nonlinear error cascade, and not due to direct local effects of thermal noise.

Supplement to Figure 1: Energy flux probability density functions

In Figure 1 of the main text we plotted probability density functions of ∣un∣|u_n| for multiple Reynolds numbers as evidence of spontaneous stochasticity. To provide more evidence, Figure 8 plots probability density functions of energy flux Πn=12Tn−1+Tn\Pi_n = \frac{1}{2}T_{n-1} + T_n with Tn=knIm⁡{un∗un+1∗un+2}T_n = k_n\operatorname{Im}\{u_n^*u_{n+1}^*u_{n+2}\}. These appear as well to converge to limiting distributions with increasing ReRe.

Probability density functions of the energy flux $\Pi_{18}$ for a fixed wavenumber and time

Figure 8. (a) Self-similar initial state with noise scaling α=0\alpha= 0. (b) Self-similar initial state with noise scaling α=3\alpha= 3. Probability density functions of the energy flux Π18\Pi_{18} for a fixed wavenumber and time in inertial units for several Reynolds numbers. All the errors are estimated as standard errors using the bootstrap method.

Dimensional analysis of the self-similar stochastic wave

The self-similarity of the stochastic wave propagation is best visible in the evolution of the total variance Var⁡(u)=∑nVar⁡(un)\operatorname{Var}(\mathbf{u}) = \sum_n \operatorname{Var}(u_n) summed over all shells, with

Var⁡(un):=⟨∣un−⟨un⟩∣2⟩.(20)\operatorname{Var}(u_n) := \left\langle\lvert u_n-\langle u_n\rangle\rvert^2\right\rangle. \tag*{(20)}

Here the average is over the ensemble of noise realizations. This quantity is plotted in Figure 10 for both the K41 (panel a) and the “burst” (panel b) initial conditions, at several values of ReRe. Note that the thermal noise in (9) by itself would produce diffusive variance growth, linear in time for each shell. This linear growth is given by Var⁡(un)∼(2θn/Re3/2)kn2t\operatorname{Var}(u_n) \sim(2\theta_n/Re^{3/2})k_n^2t, or Var⁡(un)∼(4νkBT/ϱ)kn2t\operatorname{Var}(u_n) \sim(4\nu k_BT/\varrho)k_n^2t before non-dimensionalization, which is what we observe for each shell separately at early time. This can be seen in Figure 9 for n=10n=10 and n=20n=20, but similar behavior is found for all nn. Note the slight deviation from linear growth near t=0t=0 due to round-off error of the double precision arithmetic.

Growth of variance for $n=10$ and $n=20$

Figure 9. Growth of variance Var⁡[un]\operatorname{Var}[u_n] (blue line) as a function of time. For comparison we also plot the function 4θnkn2t4\theta_n k_n^2 t (orange line), which clearly coincides with Var⁡[un]\operatorname{Var}[u_n] for early times once the variance supersedes the roundoff error.

Total variance growth for stationary K41 and “burst” initial states

Figure 10. (a) Stationary K41 initial state. (b) “Burst” initial state. Total variance growth as a function of time in inertial range units for the two initial states. The inset plots show the total variance growth as a function of time in SI units for the largest Reynolds number. The green line on the inset plots shows predicted power-law growth in the self-similar regime, and the red line on the inset plots shows exponential fit in the chaotic regime.

Exponential growth, indicated by the red curve in the insets in Figure 10, commences when the nonlinearity starts to dominate, which is the chaotic phase predicted by Ruelle [30]. Finally, however, power-law growth ensues, to a very good approximation for the K41 initial datum and to rough approximation for the “burst” datum. For any exactly self-similar initial condition scaling as un=Akn−hu_n = Ak_n^{-h}, the variance could be expected to grow as a power law Var⁡(u)∝A21−ht2h1−h\operatorname{Var}(\mathbf{u}) \propto A^{\frac{2}{1-h}}t^{\frac{2h}{1-h}} on dimensional grounds. This is observed in Figure 10 for the K41 solution, where h=1/3h=1/3 gives Var⁡(u)∝εt\operatorname{Var}(\mathbf{u}) \propto\varepsilon t, in good agreement with the numerical results. Notice that this is the same power-law observed in previous studies of predictability of turbulence, both by analytical closure [31] and by direct numerical simulation [32, 33]. For the “burst” solution the power-law Var⁡(u)∝t2/3\operatorname{Var}(\mathbf{u}) \propto t^{2/3} predicted for h=1/4h=1/4 is only crudely consistent with the data. In the spirit of the multifractal model [34], we conjecture that power-law growth would be recovered if the variance were averaged over many such “burst” initial data all selected from the statistical steady-state with the same Hölder exponent hh. Averaging further over a distribution of hh-values should reproduce the anomalous growth ⟨∣δu(t)∣p⟩∼tξp\langle\lvert\delta u(t)\rvert^p\rangle\sim t^{\xi_p}, with ξp\xi_p nonlinear in pp, previously observed for growth of differences δu(t)=u′(t)−u(t)\delta u(t)=u'(t)-u(t) between solutions in shell models [35].

Statistics of Kolmogorov multipliers and super-universality

Once the stochastic wave has passed a given shell nn, further changes at that scale are due only to the intermittent nature of the turbulent dynamics behind the front, which we find to be similar to that in the infinite-time statistical steady-state. To expose this deep similarity, we consider the Kolmogorov multipliers [36] for the shell amplitudes ωn=∣un∣/∣un−1∣\omega_n=\lvert u_n\rvert/\lvert u_{n-1}\rvert and the phase multipliers [37] Δn=arg⁡(un−2un−1un∗)\Delta_n=\arg(u_n-2u_{n-1}u_n^*), which have been used previously to characterize the build-up of intermittency in the turbulent cascade. The PDFs of angle multipliers are shown in Figure 11 for three fixed shells, n=16,18,20n=16,18,20, as a function of time. They are invariant under the shift n↦n−1n\mapsto n-1, t=τn↦t=τn−1t=\tau_n\mapsto t=\tau_{n-1}, which is a consequence of the self-similar growth. Furthermore, unlike the PDFs of modal energy and energy flux, which keep evolving in time for each shell and only coincide in time-inertial units, the PDFs of multipliers saturate and become time-invariant past the front. This is indicative of the scale-invariance of the multiplier statistics, a well-known feature in the turbulent steady-state which has been explained there by a “hidden symmetry” [38, 39]. In fact, the multiplier PDFs in the stochastic wake of the traveling wave are indistinguishable from those in the infinite-time steady state. This is what we call super-universality of the multiplier statistics, distinct from the usual universality of the spontaneous statistics, that is, its independence from the precise regularization and the small-scale noise that triggers it [14]. As evidence of super-universality in the Sabra model, we compare the probability distribution functions of magnitude and angle multipliers in the wake of the stochastic front with those same probability densities observed in a turbulent steady-state from our previous study [2]. As seen in Figure 12, there is excellent agreement. As a consequence, we expect that structure functions ⟨∣un∣p⟩\langle|u_n|^p \rangle calculated in the stochastic wake will exhibit anomalous scaling ∼(ε/kn)p/3(knℓ(t))−δζp\sim(\varepsilon/k_n)^{p/3}(k_n\ell(t))^{-\delta\zeta_p}, exactly as in the steady-state cascade but with usual outer length LL replaced by ℓ(t)\ell(t). Super-universality thus implies that characteristic turbulent statistics are reached in finite time, unlike the universal steady-state distributions in low-dimensional chaotic systems which require an infinite time limit.

Probability density functions of angle multipliers

Figure 11. (a) n=16n = 16 (b) n=18n = 18 (c) n=20n = 20. Probability density functions of angle multipliers as a function of time in SI units for several shell indices in the inertial range. The three subfigures illustrate the self-similarity of the stochastic wave, which follows from identical probability density functions for different shells up to a time shift.

Comparison of the probability distributions of angle and magnitude multipliers

Figure 12. Comparison of the probability distributions of angle and magnitude multipliers in the steady-state with the probability distribution in the stochastic wake past the front.

Supplement to Figure 2: Lorenz plot for ``burst'' initial datum

In Figure 2 of the main text we presented a plot analogous to that of Lorenz (1969) [23], in his Figure 2, for our shell model with K41 initial data. Figure 13 shows the similar plot for the “burst” initial datum, exhibiting a similar spread of stochasticity across scales.

Twice ensemble average energy and velocity variances across the ensemble

Figure 13. Twice ensemble average energy E[ϵn]\mathbb{E}[\epsilon_n] (∙\bullet) for ϵn=12∣un∣2\epsilon_n = \frac{1}{2}|u_n|^2 and velocity variances (∙\bullet,∙\bullet,∙\bullet,∙\bullet) across the ensemble as a function of wavenumber in SI units for 4 increasing times. The smallest time in the variances plots the initial transient and the subsequent three times show the propagation of the stochastic front across the inertial range towards large scales. 2E[ϵn]2\mathbb{E}[\epsilon_n] of the initial state is provided for comparison with the Fig. 2 of the main text.

Theoretical Estimate of the Condition for Eulerian Spontaneous Stochasticity.

Building on the earlier work of Ruelle [30] and Lorenz [23], we can provide a theoretical estimate of the randomization times tr(n)t_r(n), which will also allow us to estimate how we can take the double limit Θ\Theta, Re⁡−1→0\operatorname{Re}^{-1} \to0 to arrive at the spontaneously stochastic state. According to Ruelle’s estimate [30] the time it takes for a thermally-triggered disturbance Δ\Delta of velocity to become significant at the Kolmogorov scale η\eta of a turbulent flow is tR∼(νϵ)1/2log⁡(θη−1/2)t_R \sim\left(\frac{\nu}{\epsilon}\right)^{1/2}\log(\theta_\eta^{-1/2}). The growth of a disturbance in Ruelle’s regime is exponential with Lyapunov exponent that corresponds to the Kolmogorov scale, but after it reaches the inertial range it enters the Lorenz regime, and grows in a self-similar way across the cascade. If we assume complete self-similarity, then by dimensional argument the disturbance grows as Δ(t)∼ϵ1/2(t−to)1/2\Delta(t) \sim\epsilon^{1/2}(t-t_o)^{1/2}, and its length-scale propagated up the cascade is ℓ(t)∼ϵ1/2(t−to)3/2\ell(t) \sim\epsilon^{1/2}(t-t_o)^{3/2}. Both of these scalings have been verified in our numerical simulations as can be seen in Figure 10 and in Figure 2 of the main text. Here tot_o is a constant to be determined from matching Ruelle’s and Lorenz’ regimes. From the condition that Kolmogorov-scale velocity perturbations are of order Δ(tR)∼(ϵν)1/4\Delta(t_R) \sim(\epsilon\nu)^{1/4} and the expressions for Δ(t)\Delta(t) and ℓ(t)\ell(t), we deduce to∼tR−(ν/ϵ)1/2t_o \sim t_R-(\nu/\epsilon)^{1/2}. This implies that the randomization time tr(ℓ)t_r(\ell) required for the stochastic wave to reach the scale ℓ\ell is the sum of two terms: one proportional to the eddy-turnover time (ℓ/L)2/3(L/U)∼ϵ−1/3ℓ2/3(\ell/L)^{2/3}(L/U) \sim\epsilon^{-1/3}\ell^{2/3} and the other dependent on the magnitude of thermal noise and proportional to Re⁡−1/2(log⁡(θη−1/2)−1)(L/U)\operatorname{Re}^{-1/2}\left(\log(\theta_\eta^{-1/2})-1\right)(L/U). If the limit Re⁡−1\operatorname{Re}^{-1}, Θ→0\Theta\to0 is taken in any manner such that

Re−1/2log⁡θη→0(21)Re^{-1/2}\log\theta_{\eta} \to0 \tag*{(21)}

we arrive at the self-similar randomization time tr(ℓ)→Cε−1/3ℓ2/3t_r(\ell) \to C\varepsilon^{-1/3}\ell^{2/3}, which is independent of the noise strength. The existence of a non-zero limit for tr(η)t_r(\eta) is the precise statement of spontaneous stochasticity. This result is consistent with earlier investigations of toy models of spontaneous stochasticity [21, 40], which suggest it is sufficient that the double limit be taken with Θ\Theta not decaying faster than e−γRee^{-\gamma Re}, where γ\gamma is a constant. Such extreme robustness of spontaneous stochasticity is one of its characteristic features.

The above heuristic arguments based on “eddy-turnover times” at each length-scale are closely related to the spectrum of Lyapunov exponents observed in chaotic shell models [41], where local exponents λn∼±ε1/3kn2/3\lambda_n \sim\pm\varepsilon^{1/3}k_n^{2/3} are observed by numerical calculations at each wavenumber knk_n in the inertial-range of scales. The existence of such a Lyapunov spectrum implies that the leading (largest positive) Lyapunov exponent must diverge as Re→∞Re \to\infty. Such a spectrum of Lyapunov exponents thus constitutes a more refined criterion for existence of the “inverse error cascade” conjectured by Lorenz.

References

  1. [1]D. Forster, D. R. Nelson, and M. J. Stephen, Large-distance and long-time properties of a randomly stirred fluid, Physical Review A 16, 732 (1977).DOI
  2. [2]D. Bandak, N. Goldenfeld, A. A. Mailybaev, and G. Eyink, Dissipation-range fluid turbulence and thermal noise, Physical Review E 105, 065113 (2022).DOI
  3. [3]D. Zubarev and V. Morozov, Statistical mechanics of nonlinear hydrodynamic fluctuations, Physica A: Statistical Mechanics and its Applications 120, 411 (1983).DOI
  4. [4]V. Morozov, On the langevin formalism for nonlinear and nonequilibrium hydrodynamic fluctuations, Physica A: Statistical Mechanics and its Applications 126, 443 (1984).DOI
  5. [5]P. Español, J. G. Anero, and I. Zúñiga, Microscopic derivation of discrete hydrodynamics, The Journal of chemical physics 131, 244117 (2009).
  6. [6]A. Donev, A. Nonaka, Y. Sun, T. Fai, A. Garcia, and J. Bell, Low mach number fluctuating hydrodynamics of diffusively mixing fluids, Communications in Applied Mathematics and Computational Science 9, 47 (2014).
  7. [7]A. Donev, J. B. Bell, A. De La Fuente, and A. L. Garcia, Enhancement of diffusive transport by non-equilibrium thermal fluctuations, Journal of Statistical Mechanics: Theory and Experiment 2011, P06014 (2011).DOI
  8. [8]M. Hairer, Solving the KPZ equation, Annals of mathematics 178, 559 (2013).
  9. [9]I. Corwin, J. Quastel, and D. Remenik, Renormalization fixed point of the KPZ universality class, Journal of Statistical Physics 160, 815 (2015).
  10. [10]G. L. Eyink and L. Peng, Infinite Reynolds number limit of Landau-Lifschitz fluctuating hydrodynamics, to appear (2023).
  11. [11]D. Bernard, K. Gawȩdzki, and A. Kupiainen, Slow modes in passive advection, Journal of Statistical Physics 90, 519 (1998).
  12. [12]G. Falkovich, K. Gawȩdzki, and M. Vergassola, Particles and fields in fluid turbulence, Reviews of modern Physics 73, 913 (2001).DOI
  13. [13]T. D. Drivas and G. L. Eyink, A lagrangian fluctuation–dissipation relation for scalar turbulence. part i. flows with no bounding walls, Journal of Fluid Mechanics 829, 153 (2017).DOI
  14. [14]A. A. Mailybaev, Spontaneously stochastic solutions in one-dimensional inviscid systems, Nonlinearity 29, 2238 (2016).DOI
  15. [15]S. Thalabard, J. Bec, and A. A. Mailybaev, From the butterfly effect to spontaneous stochasticity in singular shear flows, Communications Physics 3, 1 (2020).DOI
  16. [16]C. De Lellis and L. Székelyhidi, On admissibility criteria for weak solutions of the euler equations, Archive for rational mechanics and analysis 195, 225 (2010).
  17. [17]C. De Lellis and H. Kwon, On non-uniqueness of hölder continuous globally dissipative euler flows, arXiv preprint arXiv:2006.06482 (2020).DOI
  18. [18]S. Daneri, E. Runa, and L. Székelyhidi, Non-uniqueness for the euler equations up to onsager’s critical exponent, Annals of PDE 7, 1 (2021).arxiv.org/abs/2004.00391
  19. [19]L. Onsager, Statistical hydrodynamics, II Nuovo Cimento (1943-1954) 6, 279 (1949).
  20. [20]G. L. Eyink and K. R. Sreenivasan, Onsager and the theory of hydrodynamic turbulence, Reviews of modern physics 78, 87 (2006).DOI
  21. [21]A. A. Mailybaev and A. Raibekas, Spontaneously stochastic arnold’s cat, Arnold Mathematical Journal , 1 (2022).arxiv.org/abs/2111.03666
  22. [22]A. A. Mailybaev and A. Raibekas, Spontaneous stochasticity and renormalization group in discrete multi-scale dynamics, arXiv e-prints , arXiv (2022).DOI
  23. [23]E. N. Lorenz, The predictability of a flow which possesses many scales of motion, Tellus 21, 289 (1969).DOI
  24. [24]T. Palmer, A. Döring, and G. Seregin, The real butterfly effect, Nonlinearity 27, R123 (2014).DOI
  25. [25]J. B. Bell, A. Nonaka, A. L. Garcia, and G. Eyink, Thermal fluctuations in the dissipation range of homogeneous isotropic turbulence, Journal of Fluid Mechanics 939, A12 (2022).DOI
  26. [26]M. Gallis, J. Torczynski, M. Krygier, N. Bitter, and S. Plimpton, Turbulence at the edge of continuum, Physical Review Fluids 6, 013401 (2021).DOI
  27. [27]P. E. Kloeden, G. J. Lord, A. Neuenkirch, and T. Shardlow, The exponential integrator scheme for stochastic partial differential equations: Pathwise error bounds, Journal of Computational and Applied Mathematics 235, 1245 (2011).DOI
  28. [28]P. Constantin, B. Levant, and E. S. Titi, Regularity of inviscid shell models of turbulence, Physical Review E 75, 016304 (2007).pubmed.ncbi.nlm.nih.gov/17358250
  29. [29]A. A. Mailybaev, Spontaneous stochasticity of velocity in turbulence models, Multiscale Modeling & Simulation 14, 96 (2016).DOI
  30. [30]D. Ruelle, Microscopic fluctuations and turbulence, Physics Letters A 72, 81 (1979).
  31. [31]C. Leith and R. Kraichnan, Predictability of turbulent flows, Journal of Atmospheric Sciences 29, 1041 (1972).doi.org/10.1175/1520-0469(1972)029<1041:potf>2.0.co;2
  32. [32]G. Boffetta and S. Musacchio, Predictability of the inverse energy cascade in 2d turbulence, Physics of Fluids 13, 1060 (2001).DOI
  33. [33]G. Boffetta and S. Musacchio, Chaos and predictability of homogeneous-isotropic turbulence, Physical review letters 119, 054102 (2017).DOI
  34. [34]U. Frisch and G. Parisi, On the singularity structure of fully developed turbulence, in Turbulence and Predictability in Geophysical Fluid Dynamics and Climate Dynamics, Enrico Fermi International School of Physics Series, Vol. 88, edited by M. Ghil, R. Benzi, and G. Parisi (North-Holland, 1985) pp. 84–88.
  35. [35]E. Aurell, G. Boffetta, A. Crisanti, G. Paladin, and A. Vulpiani, Predictability in systems with many characteristic times: The case of turbulence, Physical Review E 53, 2337 (1996).pubmed.ncbi.nlm.nih.gov/9964518
  36. [36]A. N. Kolmogorov, A refinement of previous hypotheses concerning the local structure of turbulence in a viscous incompressible fluid at high reynolds number, Journal of Fluid Mechanics 13, 82 (1962).DOI
  37. [37]G. L. Eyink, S. Chen, and Q. Chen, Gibbsian hypothesis in turbulence, Journal of statistical physics 113, 719 (2003).DOI
  38. [38]A. A. Mailybaev, Hidden spatiotemporal symmetries and intermittency in turbulence, Nonlinearity 35, 3630 (2022).DOI
  39. [39]A. A. Mailybaev and S. Thalabard, Hidden scale invariance in Navier–Stokes intermittency, Philosophical Transactions of the Royal Society A 380, 20210098 (2022).DOI
  40. [40]G. L. Eyink and D. Bandak, Renormalization group approach to spontaneous stochasticity, Physical Review Research 2, 043161 (2020).DOI
  41. [41]M. Yamada and K. Ohkitani, Asymptotic formulas for the lyapunov spectrum of fully developed shell model turbulence, Physical Review E 57, R6257 (1998).

Paper details

vlnc:W_7645nnzw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents