WorksPhysics and Astronomy
Intermittent fluctuations determine the nature of chaos in turbulence
Abstract
We adapt recent ideas for many-body chaos in nonlinear, Hamiltonian fluids [Murugan et al., Phys. Rev. Lett. 127, 124501 (2021)] to revisit the question of the Reynolds number Re dependence of the Lyapunov exponent in fully developed turbulence. The use of such decorrelators allow us to investigate the interplay of the competing effects of viscous dissipation and nonlinearity. We obtain a precise value of and show that departure from the Kolmogorov mean field result is a consequence of the intermittent fluctuations in the velocity-gradient tensor. The robustness of our results are further confirmed in a local, dynamical systems model for turbulence.
Fully developed, incompressible turbulence is perhaps the most celebrated example of a chaotic system. In contrast to other examples of classical, many-body systems showing chaotic behaviour, turbulent flows have the distinction of being central across natural world. Unsurprisingly, therefore, physicists working in problems ranging from astrophysics, atmospheric sciences, oceanography, and of course fluid dynamics have to factor in the underlying chaotic nature of such flows [1]. This ubiquity is not surprising: After all the underlying Navier-Stokes equation, in all such systems, lead to solutions which are turbulent in the limit of small viscosities commonly seen in most fluids [2].
While mathematical and engineering tools remain indispensable, the ideas of statistical physics provide the basis for much of our understanding of turbulence, especially in its most general form where the flow is homogeneous and isotropic. In particular, chaos along with the other fingerprint of fully developed turbulence, namely intermittency, and their dependence inter alia the Reynolds number , forms a major challenge in a complete understanding of turbulence.
The degree of chaos is usually quantified by the (largest) Lyapunov exponent of the flow. By using arguments tracing back to Kolmogorov’s seminal work from 1941, Ruelle showed that [3]. This result is a consequence of associating the (largest) Lyapunov exponent with the inverse of the smallest time-scale of the flow , obtained most simply from phenomenological arguments, and thence the -scaling. More generally, assuming a Hölder exponent characterising the flow, it is easy to show [4]; indeed taking the monofractal, Kolmogorov limit [5] of a unique recovers the Ruelle scaling.
The result implicitly assumes a unique small length and time-scale characteristic of a monofractal flow.
However, the observational, experimental, and numerical evidence against this is overwhelming. Indeed, a modern rationalisation of turbulence rests on the multifractal approach [6] developed by Frisch and Parisi [7]. Adapting this model, which allows for a spread of Hölder exponents in the flow, leads to the revised scaling obtained by Crisanti et al. , with [4].
This scaling exponent has been challenged in recent years with data from different experiments and direct numerical simulations. For example, Berera and Ho [8] have recently reported . This lack of consensus between the theoretical estimate and measurements — as well as the lack of agreement between different simulations [4, 8, 9, 10, 11, 12, 13] — suggests that the origins of chaos in turbulence is far from settled.
But is there a way to address the microscopic origin of from the Navier-Stokes equation itself and connect it to the intermittent aspect of fully developed turbulence? And what role does the non-locality (in length scale) of interactions — a defining feature of the Navier-Stokes equation — play?
In this paper we show that this is indeed possible through recently constructed ideas of decorrelators [14] — largely limited to Hamiltonian [15, 16, 17] systems in the context ergodicity, thermalisation and chaos — and adapt them for driven-dissipative, non-equilibrium systems such as turbulence. Consequently, by using a suitable mix of theory and numerical simulations, we uncover the competing effects of the viscous and nonlinear terms to show exactly the chaos and the scaling emerges along with what determining the value of and why this is so. We also confirm the robustness of this result by using a dynamical systems approach to turbulence with the additional advantage of exploiting its nearest and next-nearest neighbour interactions to investigate short-time effects of locality which are absent in the Navier-Stokes equation.
We begin with the three-dimensional, incompressible, unit-density Navier-Stokes equation, driven by an external force to a non-equilibrium statistically steady state. The force varies to ensure a constant energy injection [18] , where the mean energy dissipation rate is the time and space average of the spatio-temporally fluctuating dissipation characteristic of such non-equilibrium steady states of fully developed turbulence with a statistically steady velocity field . This follows from the familiar, instantaneous (spatially-averaged) kinetic energy balance:
with a mean kinetic energy .
We now construct two, nearly identical, velocity fields and , where the form of ensures that system B remains incompressible. We choose the amplitude of perturbation sufficiently small (see Appendix A) and localise it in space at ; the precise choice of is irrelevant and can be chosen to be the center of our domain.
We choose these fields — and — as initial conditions for the Navier-Stokes equation, evolve them independently in time, and calculate the velocity difference for . From such velocity differences, we are able to construct the decorrelator [17, 19]. It is easy to show, starting from the Navier-Stokes equation written separately for systems A and B, that the decorrelator follows an evolution (written in component form and sum over repeated indices assumed)
with the strain-rate tensor , , where is the Green’s function obtained from the Poisson equation for pressure, and the coefficient of kinematic viscosity (see Appendix B).
A further simplification exploits the statistical isotropy and homogeneity of fully developed turbulence to construct, over the volume , the spatially integrated decorrelator . This leads to a simple cancellation of the divergence term in Eq. (1) and a resulting evolution equation for the integrated decorrelator:
where and as the contributions from the strain and dissipative terms, respectively. There is one additional remark to make. In the derivations of Eqs. (1)–(2) we have neglected the forcing terms on systems A and B. This is perfectly valid at short times but in the long time limit it plays a critical role as we shall see below.
At short times we conjecture an exponential growth of the decorrelator: . We confirm this numerically by integrating Eq. (2) with an initial perturbation field and measurements of the strain-rate tensor drawn from direct numerical simulations (DNSs) of the Navier-Stokes equation (Appendix A).
In Figure 1 we show a representative plot of vs time for displaying a clear plateau demarcated

Figure 1. Representative plots of the normalised spatially averaged decorrelator , and vs time for . The pair of vertical dashed lines indicate the short-time exponential-growth phase. (Inset) Probability density function of the direction cosines of (at short times) showing a preferential growth of the difference field along the compressional direction.
by a pair of vertical dashed lines. While this is not very surprising, given that turbulence is chaotic, it is far from obvious how to trace its origins in the structure of Eq. (2) and in particular how the competing effects of the strain and viscous terms conspire to give such exponential-growth phases. In Figure 1 we plot and (compensated by ) and find strong evidence of their short-time exponential growth — the plateau between the vertical dashed lines — corresponding to Lyapunov exponents and , respectively. Furthermore, the dissipative effect of the viscous term is compensated by the effect of the strain term , leading to an overall positive Lyapunov exponent .
How do we understand the origins of which leads to a compensation of the negative and hence a positive ? In particular, how does the strain field ensure the exponential amplification of at short times and thence ? The answer lies in decomposing the strain term in the eigenbasis of the strain tensor , with direction cosines of (along the eigendirections) and the eigenvalues. The relative distributions for the extensional (; ), intermediate (; ) and compressional directions (; ), shown in the inset of Figure 1, underlines a bias for the compressional direction. Thus, clearly the dominant contribution in this expansion must come from the statistics of the compressional eigenvalue.
At such short times, a careful analysis of shows (see Appendix B) that . And hence, the Reynolds number dependence of must be dominated by the competing effects of the mean eigenvalue and their intermittency-induced fluctuations .
In the long time limit, , all eigenvalues are sampled equally (see Appendix B, Figure 5(c)) leading to because of incompressibility. We see from Figure 1 that this is indeed the case as the compensated term rapidly falls to 0 for .

Figure 5. (a) Plots of , at late times, for different Reynolds numbers, showing a saturation to (black dotted, horizontal line). (b) A plot of the probability distribution of components of (for ), for the Navier-Stokes equation, along the three eigen vectors of the strain rate tensor at late times once the decorrelator saturates. (c) The distribution of the compressional eigenvalues for different values of . Both the peak and the spread of the distribution scales with . (d) The evolution of vs , for the GOY shell model, on a loglog scale at different times; the late time shows an asymptotic scaling corresponding to indicated by the dashed line.
Does the viscous term also go to zero similarly at late times leading to ? At such late times it is easy to show that (see Appendix B) . Indeed, in Figure 1 we do see a clear saturation of the normalised viscous term to a value consistent with (corresponding to the simulations from which the rate of strain matrix is drawn).
A trivial consequence of this is that within the framework of Eq. (2) (clearly seen in Figure 1) and suggesting that the decorrelator never saturates. However this is a contradiction: By definition, the decorrelator is defined as the spatial average of the velocity difference of systems A and B, and hence at long times since the cross-correlator vanishes as the fields decorrelate as .
This contradiction is resolved by recalling that the derivation of Eq. (2) neglects the effective energy injection. While at short times, this is zero, at long times this contribution is no longer irrelevant. In fact we can show (see Appendix B) that as and thus compensates the viscous contribution (see Appendix B, Figure 5(a)) leading to .
How consistent are these ideas in actual measurements of the decorrelator in DNSs of the Navier–Stokes equation? We check for this by solving the incompressible Navier–Stokes equation on a triply-periodic domain by using a pseudospectral method with a large-scale constant energy [18] injection scheme to drive
the flow and maintain it in non equilibrium statistically steady state with a mean energy ; see Appendix A for a detailed summary of the numerical scheme and the parameters used. By using the same protocol described earlier, we calculate the decorrelator for several different values of the Reynolds number . In Figure 2 we show representative snapshots at (left inset) and (right inset), corresponding to the initial and the full exponential-growth phases, respectively. (Ref. [20] shows the full evolution of .)

Figure 2. Semilog plots of vs from DNSs for different showing an exponential regime followed by saturation at a value $2E_0$ (dashed horizontal line). The insets show two snapshots at (left) and (right) of the square of the difference field for . (See Ref. [20] for the full evolution of $\lvert\delta\mathbf{u}\rvert^2.)
These difference fields allow us to construct the spatially averaged decorrelator which we plot, on a semilog scale, in the main panel of Figure 2 as a function of time. We see a convincing exponential regime from which we can extract the Lyapunov exponent , followed by a saturation at a value approximately equal to , indicated by the dashed horizontal line, as our theory suggests. (We have confirmed from our data that the onset of the saturation of the decorrelators (see Figure 1 and Figure 2) is of the order of the inverse of the Lyapunov exponent.)
It is important to observe that even at very short times, before the exponential growth phase, our results from the Navier–Stokes equation summarised in Figure 2 do not suggest an initial power-law growth which has been seen in many local models of chaotic systems[21, 22]. Could this — as well as the somewhat self-similar evolution of the difference field (see Figure 2, insets) — be a consequence of the essential non-local nature of the equation?
To test this as well as the robustness of our conclusions of an exponentially growing decorrelator, we use a simpler, phenomenological (cascade) model for turbulence, inspired from dynamical systems, namely the Gledzer–
Ohkitani-Yamada (GOY) shell model [22, 23, 24]. In such shell models (Appendix A), the velocities , associated with a scalar, exponentially growing, wave-number , are treated as complex, dynamical variables, with the shells numbers being integers (see Appendix A for further details).
In the statistically steady state (Appendix A), let the velocity field be denoted as . We now construct a second field and . Our choice of ensures that the tiny perturbation (with ) are always added at wavenumbers just before the onset of the dissipation scales for all the Reynolds number being considered. (We have checked that our results are robust to the specific choice of or variations in .)
We now use the nearly-identical fields and as initial conditions and evolve the GOY shell model independently for systems A and B and hence construct the shell model decorrelator . The denote an ensemble averaging over independent realisations of the initial field . While it is straightforward to write the evolution equation for (Appendix C), we find this is neither revealing nor analytically tractable. Instead, we choose to work with the numerically constructed and assess its behaviour from our simulations.
Given the local nature of interactions in the shell model[21], it is reasonable to expect that decorrelators such as may well have a self-similar short time growth
followed by an eventual exponential rise with a saturation as . In the inset of Figure 3 we show a loglog plot of the temporal evolution of for a small initial time window with a clear power-law range for . This power-law eventually () gives way to the exponential-growth phase of decorrelators. The Lyapunov exponent characterising such shell-by-shell decorrelators are of course -dependent. Hence we find it useful to construct the -independent Lyapunov exponent via the integrated decorrelator . In the main panel of Figure 3 we show a semilog plot of the temporal evolution of for different values of Re showing, yet again, a clear exponential growth, with a shell model Lyapunov exponent , followed by a saturation to values roughly corresponding to , with the statistically steady state energy in our shell model.

Figure 3. Semilog plot of the shell model decorrelator vs for different showing an exponential growth. (Inset) Loglog plot of the shell-specific decorrelator for different shells (the arrow indicates the direction to shells) at very early times shows a power-law growth.
We can now summarise all of these ideas in answering the question of how consistent are the various (largest) Lyapunov exponent — , , and — with each other given the different ways and indeed the different models from which they are derived? In Figure 4 we show a composite plot of , and versus the Reynolds number Re. For measurements from the DNSs and the GOY shell model simulations, we estimate the Lyapunov exponents from a linear fit in the exponential-growth phase of the decorrelators (as shown in Figure 2 and Figure 3); the errors on such fits yield the errorbars on and . We find our errorbars smaller than the symbols sizes in Figure 4. The errorbars for the exponent are extracted likewise and are of the same order as those on and . The exponent is obtained differently. We use plots of , such as the one shown in Figure 1, to extract and its errorbar as the mean and standard deviation of the plateau, respectively.

Figure 4. A loglog plot of , , and , obtained from various DNSs versus the Reynolds number Re (DNS) (bottom X axis), as well as the Lyapunov exponent obtained from the shell model (rescaled by a constant factor to overlap with the DNS data since the Reynolds number ranges are different) with the corresponding shell model Reynolds numbers (top X axis). (Inset) A loglog plot of the probability density function for showing a power-law tail.
We find remarkable consistency (within error bars) between the different measurements of the Lyapunov exponents and indeed across the different models, as well as the range of Reynolds number between the DNSs (bottom axis) and shell models (top axis), that we use. All the Lyapunov exponents seem to follow a universal scaling: , with . Remarkably, combining the shell models and DNS data we show in Figure 4 that our reported scaling holds for nearly 7 decades in Reynolds number.
The theory developed earlier suggests a connection between the Lyapunov exponent and the statistics of with (Appendix B). The departure of the measured Lyapunov exponents (Figure 4) from a Kolmogorov-like mean field result , is a clear indication that the dynamics must stem from intermittent fluctuations and the dominant scaling in the statistics of is due to . In Figure 4 we show a plot of from our DNSs and find a near perfect agreement with the Lyapunov exponent confirming nearly 7 decades in Reynolds number, the robustness of this theory: . Furthermore, we confirm that the mean eigenvalue (Figure 4) consistent with the simpler scaling analysis, not accounting for fluctuations, and at odds with the measured Lyapunov exponent.
In this work we have shown how recent ideas of decorrelators [14, 15, 16, 17, 25] provide a microscopic way to understand the origins of chaos in fully developed turbulence. Remarkably, these ideas are just as robust for simpler, cascade models of turbulence where locality [21] ensures an initial scale-invariant growth of the decorrelator. We find signatures of these fluctuations in one further aspect of the strain and viscous terms in Eq. (2) which connects with the idea of intermittency. The theory developed by us involve the mean exponents: and (and hence ) are obtained from the spatial integrals over the corresponding terms. However, it is possible to measure the local exponents and in the exponential-growth phase and thence their probability distribution functions . These distribution are found to have exponential (for ) or power-law (for ; see inset of Figure 4) tails strongly suggestive of intermittency and its connections with the question of chaos in turbulence. We do not explore these ideas, and in particular the origins of the power-law tails, any further in this work as well as its implications for Lagrangian chaos [26, 27], but leave it for a more detailed study of the statistics of and in the future.
Acknowledgments
We acknowledge Anupam Kundu, Siddhartha Mukherjee, Sthitadhi Roy, Sibaram Ruidas and Akash Sarkar for discussions on related issues. A.B. acknowledges the Long Term Visiting Students’ Program (LTSVP) of the ICTS-TIFR which enabled this collaboration. SB acknowledges funding from the Swarna Jayanti fellowship of SERB-DST (India) Grant No. SB/SJF/2021-22/12 and DST, Government of India (Nano mission), under Project No. DST/NM/TUE/QM-10/2019 (C)/7. S.S.R. acknowledges SERB-DST (India) projects STR/2021/000023 and CRG/2021/002766 for financial support. This research was supported in part by the International Centre for Theoretical Sciences (ICTS) for the programs - Field Theory and Turbulence ( (code:ICTS/ftt2023/12), Indo-French workshop on Classical and quantum dynamics in out of equilibrium systems (code: ICTS/ifwcqm2024/12) and 10th Indian Statistical Physics Community Meeting (code: ICTS/10thISPCM2025/04). The simulations were performed on the ICTS clusters Mario, Tetris, and Contra. The authors acknowledge the support of the DAE, Government of India, under projects nos. 12-R&D-TFR-5.10-1100 and RTI4001.
Direct Numerical Simulations
The Navier-Stokes Equation
We perform direct numerical simulations (DNSs), by using a pseudospectral method, of the three-dimensional (3D), incompressible, unit density Navier Stokes equation, with the pressure field
on a periodic box of size with collocation points. We choose and (to check for numerical convergence) and vary the coefficient viscosity to obtain Reynolds numbers . We use a second-order Adams-Bashforth for time-marching with a time step (for ) and (for ).
We initialise the flow with a random initial condition such that the initial energy spectrum spectrum with . The system is driven to a non-equilibrium steady state (NESS) through a large-scale forcing with a constant energy injection at large scales corresponding to wavenumber .
This statistically steady velocity field is taken as the initial field for system A: and further define the initial condition for system B via . We use a Gaussian perturbation with , and at the center of the domain and which also satisfies the incompressibility constraint.
The nearly identical fields and are then used as initial conditions for the Navier-Stokes equation and evolved as before. The decorrelator is calculated from the evolution of such twin simulations as and thence . The decorrelators are then used to estimate the Lyapunov exponents (obtained by fitting the initial exponential growth phase) and its dependence on the Reynolds number.
The Gledzer-Ohkitani-Yamada (GOY) Model
The Gledzer-Ohkitani-Yamada or the GOY shell model
is one of several cascade models which mimic Navier-Stokes turbulence and are designed to achieve very high Reynolds number. In such models, the velocities (with boundary conditions ) associated with a scalar, exponentially growing, wave-number , are complex, dynamical variables. The shell numbers range from ; in our simulations we choose and the coefficient of viscosity varies as yielding Reynolds numbers in the range ; the shell model, Reynolds number is simply defined as . The equation is integrated numerically by using a second-order Runge-Kutta scheme with a time step . The constant amplitude forcing , applied on the shell and drives the system to a statistically steady state [28].
As with the strategy for the Navier-Stokes equation, we first obtain a non-equilibrium statistically steady state with an energy spectrum over an inertial range of shells whose extent is determined by Re. We use this steady state to define the initial field for system A — and construct the initial condition for system B via . In all our simulations we have used and . Numerically it has been seen that the shell wise decorrelators first show a power-law rise later followed by the exponential-growth regime. The initial power law largely depends on the distance from the perturbed shell. Therefore, we add the perturbation in a fixed shell for all values of Re.
Derivation of the decorrelator from the Navier-Stokes equation and its limits
Evolution equation for and
In this section we provide the derivation of the equation of motion of the decorrelator for the Navier-Stokes equation and in particular estimate the long and short time limits of and .
We begin by recalling that the pressure term in the Navier–Stokes equation can be rewritten as a Poisson equation
by exploiting the incompressibility constraint. This Poisson equation underlines the non-local nature of turbulence.
It is useful, in what follows, to have a Green’s function representation
where the denotes the quantity evaluated at .
This allows us to construct the evolution equation of the solenoidal velocity perturbation in component form
and hence, via a dot product of with Eq. (B-3), the equation for the decorrelator :
Here, is the strain-rate tensor and .
A spatial integration with now leads to the cancellation of the divergence term and hence
the evolution equation for our spatially integrated decorrelator as discussed in the main text of the manuscript.
Short and long-time asymptotics of
At early times, (since the forcing terms are nearly identical), the pre saturation exponential growth of the decorrelator with the Lyapunov exponent is controlled by the short time asymptotics of . Hence, in what follows, we drop the contribution from , to simplify notation.
We now expand in the eigenbasis of the strain-rate tensor, with eigenvalues and direction cosines , to obtain
with dominated by the compressional eigendirection as discussed in the main text.
We choose an integrating factor , such that , to obtain the form of the spatially-resolved decorrelator at time
We are now able to spatially integrate Eq. (B-7) and, keeping in mind that the initially localised perturbation is independent of , we obtain
where the angular brackets .
Since we are interested in the short-time asymptotics of , it is possible to Taylor expand the exponential and, keeping in mind the mapping of to , we obtain
where . Considering terms up to , to account for corrections stemming from intermittent fluctuations [8],
Now all that remains is the evaluation of and . From the choice of it is easy to see (by interchanging time and space integrals suitably) that (see Figure 4 of the main text). The determination of requires a further Ansatz of an approximate delta-correlation in time, at least over time-scales of the order associated with strain-rate, to yield (see Figure 4 of the main text as well as Figure 5(c) which shows that indeed the distribution of changes with ) which, in the large Reynolds number limit dominates over the of the mean exponent. Hence with .
This then completes the proof of how the intermittency induced fluctuations determines the Lyapunov exponent for fully-developed turbulence.
At long times, beyond the exponential-growth phase when the systems have started to completely decorrelate the forcing term can not be neglected. Thus as , we estimate . As we shall see below, this exactly compensates the contribution of the viscous term at long times.
Short and long-time asymptotics of
The viscous term can be expanded as
We find at short times, in the exponential-growth phase, this terms scales as , where corresponds to the smallest Kolmogorov time scales. However, at long times , systems A and B decorrelate leading to and hence (see Figure 5(a) as well as the main text). This is easily understandable because, on average, dissipation is balanced by the energy injection derived above.
Decorrelator in the shell model for turbulence
It is possible to construct an evolution equation for the decorrelator in our shell model but, as we shall see, its structure is less transparent than what we have seen for the Navier-Stokes equation.
By subtracting the GOY model equations for systems A and B we obtain the evolution of the shell-by-shell velocity difference .
From this by suitably multiplying with the complex conjugate field , the decorrelator obeys
A summation over all shells leads to simplifications in the evolution of the GOY model decorrelator
Further progress is only possible through a numerical solution of the decorrelator equation as we report in the main text of the manuscript. The shell-wise decorrelator at very short times show a power-law growth with a wavenumber-dependent scaling exponent. In particular at long times when the decorrelator saturates, we find, unsurprisingly, as shown in Figure 5(d).
References
- [1]R. Pandit, P. Perlekar, and S. S. Ray, Pramana 73, 157 (2009).
- [2]U. Frisch, Turbulence: The Legacy of A. N. Kolmogorov (Cambridge University Press, 1995).
- [3]D. Ruelle, Physics Letters A 72, 81 (1979).
- [4]A. Crisanti, M. H. Jensen, A. Vulpiani, and G. Paladin, Phys. Rev. Lett. 70, 166 (1993).
- [5]A. N. Kolmogorov, Dokl. Akad. Nauk SSSR 30, 299–303 (1941).
- [6]R. Benzi, G. Paladin, G. Parisi, and A. Vulpiani, Journal of Physics A: Mathematical and General 17, 3521 (1984).
- [7]U. Frisch and G. Parisi, Proceedings of the International School of Physics Enrico Fermi, Course LXXXVIII, Varenna, 1985 (1985).
- [8]A. Berera and R. D. J. G. Ho, Phys. Rev. Lett. 120, 024101 (2018).
- [9]M. Yamada and K. Ohkitani, Journal of the Physical Society of Japan 56, 4210 (1987).
- [10]S. Mukherjee, J. Schalkwijk, and H. J. Jonker, Journal of the Atmospheric Sciences 73, 2715 (2016).
- [11]G. Boffetta and S. Musacchio, Phys. Rev. Lett. 119, 054102 (2017).
- [12]P. Mohan, N. Fitzsimmons, and R. D. Moser, Phys. Rev. Fluids 2, 114606 (2017).
- [13]J. Ge, J. Rolland, and J. C. Vassilicos, Journal of Fluid Mechanics 977, A17 (2023).
- [14]A. Das, S. Chakrabarty, A. Dhar, A. Kundu, D. A. Huse, R. Moessner, S. S. Ray, and S. Bhattacharjee, Phys. Rev. Lett. 121, 024101 (2018).
- [15]T. Bilitewski, S. Bhattacharjee, and R. Moessner, Phys. Rev. B 103, 174302 (2021).
- [16]T. Bilitewski, S. Bhattacharjee, and R. Moessner, Phys. Rev. Lett. 121, 250602 (2018).
- [17]S. D. Murugan, D. Kumar, S. Bhattacharjee, and S. S. Ray, Phys. Rev. Lett. 127, 124501 (2021).
- [18]A. G. Lamorgese, D. A. Caughey, and S. B. Pope, Physics of Fluids 17, 015106 (2004).
- [19]S. Mukherjee, R. K. Singh, M. James, and S. S. Ray, Nature Physics 19, 891 (2023).
- [20]Supplemental video (2025), spatio-temporal evolution of an initially localised perturbation.
- [21]L. Biferale, Annual Review of Fluid Mechanics 35, 441 (2003).
- [22]P. D. Ditlevsen, Turbulence and Shell Models (Cambridge University Press, Cambridge, UK, 2010).DOI
- [23]E. B. Gledzer, Soviet Physics Doklady 18, 216 (1973).
- [24]K. Ohkitani and M. Yamada, Progress of Theoretical Physics 81, 329 (1989).
- [25]S. H. Shenker and D. Stanford, Journal of High Energy Physics 2014 (2014).
- [26]J. Bec, L. Biferale, G. Boffetta, M. Cencini, S. Musacchio, and F. Toschi, Physics of Fluids 18, 091702 (2006).
- [27]S. S. Ray, Phys. Rev. Fluids 3, 072601 (2018).
- [28]S. S. Ray, D. Mitra, and R. Pandit, New Journal of Physics 10, 033003 (2008).