betaFeedback

WorksPhysics and Astronomy

Contrasting Eulerian and Lagrangian organization of turbulent dissipation

Abstract

The dissipation of turbulent kinetic energy is conventionally attributed to a scale-invariant spatial multiplicative cascade, with Lagrangian statistics inferred through bridge relations. Direct numerical simulations with tracers reveal instead that Eulerian coarse-grained dissipation has a squared-logarithmic cascade depth and scale-dependent multipliers, whereas time-averaged dissipation along trajectories displays multifractal scaling and an approximately scale-invariant temporal cascade. The classical bridge fails at the level of individual trajectories. Eulerian multipliers nevertheless possess a multifractal organization whose accumulation accounts for the squared-logarithmic spatial statistics. This contrast motivates a material-inheritance picture in which Eulerian dissipation statistics reflect the spatial assembly, deformation, and mixing of distinct Lagrangian histories. The results therefore distinguish the temporal cascade organization revealed along fluid trajectories from the non-stationary spatial refinement observed in the Eulerian frame.

Introduction

Turbulence thrives on paradox. Time and again, elegant theoretical frameworks built on universal laws or scaling principles have seemed to capture the essence of turbulent motion, only to be challenged by new experiments or simulations revealing subtle yet profound deviations from the expected behavior. Kolmogorov’s 1941 theory provides perhaps its most celebrated example: assuming that inertial-range statistics depend only on the mean energy-transfer rate and the observation scale, it predicts universal self-similar scaling, with velocity structure-function exponents ζp=p/3\zeta_p = p/3 [1]. This phenomenology has extraordinary explanatory power, yet experiments and numerical simulations consistently reveal intense rare fluctuations, increasingly non-Gaussian statistics toward small scales, and scaling exponents that depart from the dimensional prediction [2, 3]. Intermittency is therefore not merely a small correction to K41, but signals that the energy transfer becomes progressively more heterogeneous along the cascade [4]. The refined similarity hypotheses of Kolmogorov and Obukhov incorporate this heterogeneity through the locally coarse-grained dissipation εℓ\varepsilon_\ell, relating velocity increments to the fluctuating local transfer rate through δv∼(εℓℓ)1/3\delta v \sim(\varepsilon_\ell\ell)^{1/3} [5, 6]. Anomalous velocity scaling is thereby connected to the statistics of εℓ\varepsilon_\ell, placing the Eulerian dissipation field at the center of the intermittency phenomenology.

This connection has encouraged two notions that are closely related but not equivalent. Multifractal scaling is first of all a statement about moments: the normalized moments of the coarse-grained dissipation behave as ⟨(εℓ/εˉ)q⟩∝(ℓ/L)χq\langle(\varepsilon_\ell/\bar{\varepsilon})^q\rangle\propto(\ell/L)^{\chi_q}, with a nonlinear dependence of χq\chi_q on qq reflecting intermittency [1, 7, 8]. A scale-stationary multiplicative cascade makes a stronger assertion about the process across scales: passing from one scale to another amounts statistically to multiplication by a random factor whose law depends only on the scale ratio, so that equal intervals of logarithmic scale are statistically equivalent. In its log-divisible form, the logarithmic increments accumulated over disjoint scale intervals are also independent. Such a process naturally produces multifractal moment scaling, but the converse need not hold: observing power-law moments over a finite range does not, by itself, establish an underlying scale-stationary, log-divisible process. This distinction will be important when comparing the spatial and temporal organizations of dissipation fluctuations.

The multifractal description of Eulerian dissipation has a substantial empirical basis. Early experiments, relying largely on one-dimensional surrogates reconstructed from hot-wire velocity measurements, found nonlinear moment exponents and consistent multifractal spectra across different flows and Reynolds numbers [7, 8]. Direct numerical simulations later made it possible to evaluate the full tensorial dissipation and to perform genuinely three-dimensional local averages. These studies provided direct support for refined similarity [9] and reported anomalous scaling of the moments of gradients [10, 11, 12] and locally averaged dissipation at Reynolds numbers up to Rλ≃1300R_\lambda\simeq1300 [13, 14]. In particular, the second-order intermittency exponent was found to approach the commonly reported value μ≃0.23\mu\simeq0.23 at the highest Reynolds numbers. Multifractal moment scaling of the Eulerian dissipation field is therefore supported by several generations of measurements, although the measured observable, coarse-graining geometry, and accessible scale range have varied among studies.

Evidence bearing on the stronger multiplicative-cascade interpretation has come from additional diagnostics. Hot-wire measurements of surrogate-dissipation multipliers found probability distributions that were approximately independent of the cascade level, supporting the idea that successive intervals of logarithmic scale are statistically equivalent [15]. Subsequent analyses nevertheless detected correlations between multipliers at successive levels [16], showing that scale-invariant multiplier distributions alone do not establish a log-divisible process with independent increments. A complementary line of investigation has examined correlations of the logarithm of dissipation, whose logarithmic dependence on spatial separation is another natural signature of multiplicative cascades [17]. Recent analysis of three-dimensional DNS by Ruffenach and Chevillard reported spatial and Eulerian temporal statistics compatible with such log-correlated fields and developed their representation in terms of Gaussian multiplicative chaos [18]. Together, these results motivate examining moment scaling, multipliers, and logarithmic correlations as related but logically distinct signatures of a cascade.

The Lagrangian viewpoint, which follows individual fluid particles rather than observing the flow at fixed spatial positions, provides a complementary temporal description of intermittency [19]. Particle-tracking experiments and direct numerical simulations have shown that Lagrangian velocity increments and accelerations undergo extreme fluctuations, often displaying stronger intermittency than their Eulerian counterparts [20, 21, 22, 23, 24]. Dissipation sampled along trajectories has received less attention, but the study of Huang and Schmitt [25] provided several indications of a temporal cascade. Using DNS at Rλ=400R_\lambda=400, they found approximately log-normal statistics for both the instantaneous and time-averaged Lagrangian dissipation. The moments of the time-averaged dissipation exhibited power-law scaling

over 10<τ/τη<10010<\tau/\tau_\eta<100, with exponents consistent with those inferred from Lagrangian velocity statistics. Moreover, both the autocorrelation of ln⁡[ε~(t)/εˉ]\ln[\widetilde{\varepsilon}(t)/\bar{\varepsilon}] and the variance of ln⁡(ε~τ/εˉ)\ln(\widetilde{\varepsilon}_\tau/\bar{\varepsilon}) followed logarithmic laws characterized by a common intermittency coefficient μ≃0.30\mu\simeq0.30. The concurrence of these signatures led to the proposal that dissipation along fluid trajectories can be described by a multiplicative cascade evolving in time.

Despite these parallel developments, the relation between the spatial and temporal organizations of dissipation remains unclear. Eulerian and Lagrangian intermittency are traditionally connected through bridge relations, which associate a spatial scale with a characteristic eddy-turnover time and thereby infer temporal scaling from spatial statistics [26, 27, 28, 29]. Although this approach has been successful for velocity fluctuations, its extension to dissipation is less direct. The Eulerian quantity εℓ(x,t)\varepsilon_\ell(\mathbf{x},t) averages the dissipation over a spatial region at a fixed time, whereas its Lagrangian counterpart averages the values accumulated by a moving particle through an evolving flow. A correspondence between their coarse-graining scales, and still more between their multiplier statistics, does not therefore follow from dimensional arguments alone. As illustrated in Figure 1, the same dissipation field appears as highly localized spatial structures in the Eulerian frame [2, 30, 31, 32] and as intermittent temporal bursts along trajectories. Whether these two manifestations are governed by a common scale-stationary cascade mechanism is the central question addressed here.

Snapshots of the local Eulerian energy dissipation field and dissipation sampled along Lagrangian tracer trajectories

Figure 1. (a) Snapshots of the local Eulerian energy dissipation field ε(x,t)=(ν/2)∥∇v+∇vT∥2\varepsilon(\mathbf{x},t)=(\nu/2)\lVert\nabla\mathbf{v}+\nabla\mathbf{v}^{\mathsf{T}}\rVert^2 in a two-dimensional slice of the simulation at Rλ=730R_\lambda=730, shown at increasing magnification in regions of linear size (b) 1024 η1024\,\eta and (c) 128 η128\,\eta. Intense dissipation events appear highly localized and filamentary in space. (d) Dissipation sampled along Lagrangian tracer trajectories, ε~(t)=ε(X(t),t)\widetilde{\varepsilon}(t)=\varepsilon(\mathbf{X}(t),t), shown for the same flow. Individual tracers experience strongly intermittent temporal fluctuations.

To address this question, high-resolution direct numerical simulations of homogeneous isotropic turbulence seeded with Lagrangian tracers are analyzed. The incompressible Navier–Stokes equations are integrated using a pseudo-spectral method for three datasets, with resolutions up to 409634096^3 and Taylor-scale Reynolds numbers Rλ=315R_\lambda= 315, 420, and 730. The numerical parameters are summarized in Table I.

N3N^3η\etaλ\lambdaLLτη\tau_\etaτL\tau_LRλR_\lambda
102431024^30.00200.0731.530.0221.76315
204832048^30.00270.113.030.0212.25420
409634096^30.000740.0391.860.0519.68730

Table I. DNS parameters. η=ν3/4/εˉ1/4\eta= \nu^{3/4}/\bar{\varepsilon}^{1/4} is the Kolmogorov dissipative length scale, with εˉ\bar{\varepsilon} the average dissipation rate, λ=15νurms2/εˉ\lambda= \sqrt{15\nu u_{\mathrm{rms}}^2/\bar{\varepsilon}} the Taylor microscale, and L=urms3/εˉL = u_{\mathrm{rms}}^3/\bar{\varepsilon} the integral length scale, with urms=⟨∣v∣2⟩/3u_{\mathrm{rms}} = \sqrt{\langle|\mathbf{v}|^2\rangle/3}. The corresponding time scales are the Kolmogorov time τη=ν/εˉ\tau_\eta= \sqrt{\nu/\bar{\varepsilon}} and the large-eddy turnover time τL=urms2/εˉ\tau_L = u_{\mathrm{rms}}^2/\bar{\varepsilon}. The Taylor-scale Reynolds number is Rλ=urmsλ/νR_\lambda= u_{\mathrm{rms}}\lambda/\nu.

The velocity field is denoted by v(x,t)\mathbf{v}(\mathbf{x},t), and the local kinetic-energy dissipation rate is ε(x,t)=2νSijSij\varepsilon(\mathbf{x},t) = 2\nu S_{ij}S_{ij}, where Sij=(∂jvi+∂ivj)/2S_{ij} = (\partial_j v_i + \partial_i v_j)/2 and ν\nu is the kinematic viscosity. Its Eulerian coarse-grained form εℓ(x,t)\varepsilon_\ell(\mathbf{x},t) is obtained by averaging ε\varepsilon over a ball of size ℓ\ell centered at x\mathbf{x}. Lagrangian tracers satisfy dX/dt=v(X,t)d\mathbf{X}/dt = \mathbf{v}(\mathbf{X},t), sample the instantaneous dissipation ε~(t)=ε(X(t),t)\widetilde{\varepsilon}(t) = \varepsilon(\mathbf{X}(t),t), and define ε~τ\widetilde{\varepsilon}_\tau by averaging along a trajectory over an interval τ\tau. For uniformly distributed tracers in a statistically stationary, homogeneous, incompressible flow, all these quantities share the mean εˉ=⟨ε⟩\bar{\varepsilon} = \langle\varepsilon\rangle.

These observables are examined using the same hierarchy of statistical diagnostics in the two frames. Moment scaling tests for multifractality, while logarithmic cumulants and correlations, together with multiplier statistics, probe the stronger requirements of scale stationarity and log-divisibility. Large-deviation spectra provide a complementary characterization of rare fluctuations and identify the effective cascade depth governing their accumulation across scales [1, 33]. Rather than assuming an Eulerian–Lagrangian bridge from the outset, this comparison asks which scale variable renders the statistics stationary in each frame and how the resulting cascade organizations are related.

Three principal findings emerge. First, over the resolved range of Reynolds numbers and scales, Eulerian coarse-grained dissipation does not display the organization of a conventional scale-stationary multifractal cascade. Its spatial covariance, logarithmic cumulants, large-deviation statistics, and moment slopes are instead consistently governed by an effective depth proportional to [log⁡(ℓ/ℓc)]2[\log(\ell/\ell_c)]^2. Second, dissipation averaged along Lagrangian trajectories exhibits the standard signatures of a temporal cascade: power-law moment scaling, logarithmic temporal correlations, a large-deviation principle with depth proportional to log⁡(τc/τ)\log(\tau_c/\tau), and approximately scale-stationary multipliers. The associated rate function is nevertheless non-quadratic, revealing a multifractal but non-log-normal temporal process. Third, the classical bridge between spatial and temporal dissipation averages fails as a trajectory-wise identification. Eule-

rian multipliers are not scale invariant, but their scale dependence has a multifractal organization of its own; the accumulation of these increasingly intermittent refinements accounts for the squared-logarithmic Eulerian depth. This structure motivates a dynamical-inheritance picture in which spatial dissipation patterns result from the deformation and mixing of distinct Lagrangian histories.

The remainder of the paper is organized as follows. Section II introduces the statistical diagnostics used throughout. Sections III and IV present the Eulerian and Lagrangian analyses, respectively. Section V revisits bridge relations and proposes a dynamical-inheritance picture, before the conclusions are presented in Section VI.

Cascades, large deviations, and multifractal scaling

Multiplicative cascades provide the standard phenomenological framework for describing intermittency in turbulence. Their statistical structure is naturally formulated in terms of a cascade depth nn. When this depth is proportional to the logarithmic separation between physical scales, the cascade gives rise to logarithmic cumulants and correlations, together with multifractal power-law scaling. As these properties are examined in the following sections, we briefly summarize their connections.

Following the phenomenology introduced by Kolmogorov and Obukhov [5, 6] (see also [1]), a positive quantity, such as the coarse-grained dissipation, is successively refined according to

εn=ε0∏i=1nMi(1)\varepsilon_n = \varepsilon_0 \prod_{i=1}^{n} \mathcal{M}_i \tag*{(1)}

where the random multipliers Mi\mathcal{M}_i describe the transfer of fluctuations between successive cascade levels. The cascade depth nn need not be identified a priori with a particular physical scale. In the standard scale-invariant construction with a fixed cascade ratio rc>1r_c > 1, the scale after nn refinement steps is sn=S/rcns_n = S/r_c^n, where SS denotes a reference scale. We then get n(s)=log⁡(S/s)/log⁡rcn(s) = \log(S/s)/\log r_c.

and εs=εn(s)\varepsilon_s = \varepsilon_{n(s)}. Figure 2 illustrates the dyadic case rc=2r_c = 2. More generally, an effective cascade depth n(s)n(s) may depend differently on the physical scale, for instance as an arbitrary function of its logarithm.

Schematic diagram of a multiplicative cascade

Figure 2. Schematic diagram of a multiplicative cascade.

Taking logarithms transforms the multiplicative cascade into an additive process,

Yn:=log⁡(εn/ε0)=∑i=1nXi,Xi:=log⁡Mi.(2)Y_n := \log(\varepsilon_n/\varepsilon_0) = \sum_{i=1}^{n} \mathcal{X}_i,\qquad\mathcal{X}_i := \log\mathcal{M}_i. \tag*{(2)}

If the logarithmic increments Xi\mathcal{X}_i are independent and stationary, then YnY_n has independent and stationary increments in the cascade depth. In a continuous-depth description, it forms a Lévy process whose cumulant generating function takes the form

log⁡⟨eqYn⟩=nΨ(q),(3)\log\langle e^{qY_n}\rangle= n\Psi(q), \tag*{(3)}

where the per-unit-depth exponent Ψ\Psi admits the Lévy–Khintchine representation [34]

Ψ(q)=μq+σ22q2+∫R(eqx−1−qx1∣x∣<1)ρ(dx).(4)\Psi(q) = \mu q + \frac{\sigma^2}{2}q^2 + \int_{\mathbb{R}}\left(e^{qx}-1-qx\mathbf{1}_{|x|<1}\right)\rho(\mathrm{d}x). \tag*{(4)}

Here μ\mu is the mean logarithmic increment per unit cascade depth, σ2\sigma^2 is the variance of its Gaussian component, and ρ\rho is the Lévy measure describing jump contributions. The classical log-normal and log-Poisson models correspond respectively to purely Gaussian increments and to jump processes with discrete amplitudes [35, 36].

Provided the corresponding derivatives of Ψ\Psi exist, all cumulants of YnY_n grow linearly with the cascade depth,

Km[Yn]=nΨ(m)(0).(5)\mathcal{K}_m[Y_n] = n\Psi^{(m)}(0). \tag*{(5)}

In particular, ⟨Yn⟩=nμ\langle Y_n\rangle= n\mu and Var⁡[Yn]=nσtot2\operatorname{Var}[Y_n] = n\sigma_{\mathrm{tot}}^2, where σtot2=Ψ′′(0)\sigma_{\mathrm{tot}}^2 = \Psi''(0) includes both Gaussian and jump contributions. For coarse-grained dissipation at scale ss, these cumulants are therefore proportional to the effective depth n(s)n(s). They depend linearly on log⁡(S/s)\log(S/s) only in the standard case where n(s)=log⁡(S/s)/log⁡rcn(s) = \log(S/s)/\log r_c; another dependence of n(s)n(s) produces a corresponding departure from logarithmic behaviour.

The same observation applies to spatial correlations. Two points separated by a distance Δ\Delta share a common cascade history over a depth n(Δ)n(\Delta) before their branches separate. The additive cascade structure therefore gives

Cov⁡[log⁡ε(0),log⁡ε(Δ)]∝n(Δ).(6)\operatorname{Cov}[\log\varepsilon(0),\log\varepsilon(\Delta)] \propto n(\Delta). \tag*{(6)}

The familiar logarithmic correlation [17] Cov⁡[log⁡ε(0),log⁡ε(Δ)]∝log⁡(S/Δ)/log⁡rc\operatorname{Cov}[\log\varepsilon(0),\log\varepsilon(\Delta)] \propto\log(S/\Delta)/\log r_c is recovered when the common depth is logarithmic in the separation. Thus, the fundamental signature of log-divisibility is linearity with cascade depth, rather than necessarily with the logarithm of the physical scale.

The cascade also admits a large-deviation description. The depth-normalized exponent

αn=−1nlog⁡(εn/ε0)=−1n∑i=1nXi.(7)\alpha_n = -\frac{1}{n}\log(\varepsilon_n/\varepsilon_0) = -\frac{1}{n}\sum_{i=1}^{n}\mathcal{X}_i. \tag*{(7)}

is an empirical average of independent and identically distributed random variables. It therefore satisfies a large-deviation principle [37]

p(α;s)≍exp⁡[−n(s)I(α)].(8)p(\alpha;s) \asymp\exp[-n(s)I(\alpha)]. \tag*{(8)}

The rate function can then be written

I(α)=sup⁡q[qα−Ψ(q)].(9)I(\alpha) = \sup_q [q\alpha-\Psi(q)]. \tag*{(9)}

In the standard scale-invariant case n(s)=log⁡(S/s)/log⁡rcn(s) = \log(S/s)/\log r_c, the exponent αn\alpha_n coincides with the usual scale-dependent exponent αs=[log⁡(εs/εS)]/[log⁡(s/S)]\alpha_s = [\log(\varepsilon_s/\varepsilon_S)]/[\log(s/S)] and the large-deviation form (8) reduces to the scale-invariant form

p(α;s)≍(s/S)I(α).(10)p(\alpha;s) \asymp(s/S)^{I(\alpha)}. \tag*{(10)}

In that case, the same function I(α)I(\alpha) is directly the multifractal co-dimension spectrum: along a one-dimensional cut through the flow, the set of points characterized by exponent α\alpha has fractal dimension 1−I(α)1-I(\alpha) [1, 7, 8]. The corresponding dissipation moments behave as

⟨(εs/εS)q⟩∝(s/S)χq,χq=−Ψ(q)/log⁡rc,(11)\left\langle(\varepsilon_s/\varepsilon_S)^q\right\rangle\propto(s/S)^{\chi_q},\qquad\chi_q=-\Psi(q)/\log r_c, \tag*{(11)}

where intermittency manifests itself through the nonlinear, concave dependence of χq\chi_q on the order qq.

It is therefore important to distinguish between log-divisibility, large deviations, and multifractality. Log-divisibility is a property of the stochastic evolution in cascade depth and implies that logarithmic cumulants and correlations are proportional to the corresponding depth. A large-deviation principle describes the fluctuations accumulated over large depths, independently of how that depth is related to physical scale. Multifractal power-law scaling is obtained only when the effective depth is asymptotically proportional to log⁡(S/s)\log(S/s). For example, a depth proportional to [log⁡(s/sc)]2[\log(s/s_c)]^2 may still yield a well-defined large-deviation function I(α)I(\alpha), but not conventional multifractal scaling in ss. Conversely, multifractal moment scaling does not by itself establish that the underlying increments are independent and stationary, and therefore does not imply log-divisibility.

The following sections examine the effective cascade depth governing Eulerian and Lagrangian coarse-grained dissipation and test whether their statistics are consistent with an underlying scale-invariant multiplicative cascade.

Eulerian dissipation: beyond the scale-invariant spatial cascade

As discussed in the previous section, two properties commonly associated with multiplicative cascades should be distinguished. A scale-invariant log-divisible cascade requires the logarithm of the coarse-grained dissipation to have stationary increments in logarithmic scale. This entails logarithmic spatial correlations, cumulants growing linearly with the cascade depth, and scale-independent multiplier statistics. Multifractality is a separate property: it requires a large-deviation principle with speed proportional to log⁡(L/ℓ)\log(L/\ell), or equivalently power-law scaling of the moments of the coarse-grained dissipation. We examine these two structures successively and show that neither is observed. Instead, the data reveal a different large-deviation structure governed by a squared-logarithmic scale variable.

We first test the log-divisible cascade picture. In a spatial multiplicative cascade, two points separated by a distance r=∣r∣r=|\mathbf{r}| share the same cascade history down to a scale of order rr. The covariance of the logarithmic dissipation should therefore be proportional to the number of common cascade steps and hence to log⁡(L/r)\log(L/r). Introducing Λ(x,t)=log⁡[ε(x,t)/εˉ]\Lambda(x,t)=\log[\varepsilon(x,t)/\bar{\varepsilon}], we measure

CΛ(r)=⟨Λ(x,t)Λ(x+r,t)⟩−⟨Λ⟩2,r=∣r∣(12)C_\Lambda(r)=\langle\Lambda(x,t)\Lambda(x+r,t)\rangle-\langle\Lambda\rangle^2,\qquad r=|\mathbf{r}| \tag*{(12)}

As shown in Figure 3(a), CΛ(r)C_\Lambda(r) does not display an extended range of linear dependence on log⁡r\log r. At most, such a behavior could be inferred over a short interval below the Taylor microscale, which broadens only weakly as the Reynolds number increases. Over the resolved range, the covariance is instead much better described by

Figure 3. (a) Spatial covariance of Λ(x,t)=log⁡[ε(x,t)/εˉ]\Lambda(x,t)=\log[\varepsilon(x,t)/\bar{\varepsilon}] as a function of the separation rr. The dashed line is a[log⁡(2r/ℓc)]2a[\log(2r/\ell_c)]^2, with a=0.023a=0.023 and ℓc=5.5L\ell_c=5.5L. (b) Mean μℓ\mu_\ell and variance σℓ2\sigma_\ell^2 of Λℓ=log⁡(εℓ/εˉ)\Lambda_\ell=\log(\varepsilon_\ell/\bar{\varepsilon}) as functions of the coarse-graining scale ℓ\ell. are quadratic functions of log⁡(ℓ/ℓc)\log(\ell/\ell_c) with the same aa and ℓc\ell_c as in (a). Vertical dotted lines in panels (a,b) mark the Taylor microscales. (c) Finite-scale rate functions of γℓ=Λℓ/[log⁡(ℓ/ℓc)]2\gamma_\ell=\Lambda_\ell/[\log(\ell/\ell_c)]^2 for Rλ=730R_\lambda=730. The dashed line is J(γ)=(γ+a/2)2/(2a)J(\gamma)=(\gamma+a/2)^2/(2a).

CΛ(r)≈a [log⁡(2r/ℓc)]2,(13)C_\Lambda(r)\approx a\,[\log(2r/\ell_c)]^2, \tag*{(13)}

with a≈0.023a\approx0.023 and ℓc≈5.5L\ell_c\approx5.5L. This behavior is incompatible with the logarithmic correlations expected from a scale-invariant spatial cascade [38].

These observations should be compared with the recent analysis of Ruffenach and Chevillard [18], who reported an approximately logarithmic spatial covariance of log⁡ε\log\varepsilon, with μ≈0.21\mu\approx0.21, for simulations with much higher Reynolds numbers compared to ours. Their logarithmic form μlog⁡(L⋆/r)\mu\log(L^\star/r) introduces an effective length L⋆L^\star that absorbs an additive finite-scale correction, while the slope is fixed to the reported value μ≈0.21\mu\approx0.21. Over a finite range, the two forms are therefore more sharply distinguished by their local slope than by the quality of the fit to CΛC_\Lambda alone. Equation (13) implies a scale-dependent effective coefficient

μeff(r)=−dCΛdlog⁡r=2alog⁡(ℓc2r),(14)\mu_{\mathrm{eff}}(r)=-\frac{\mathrm{d}C_\Lambda}{\mathrm{d}\log r}=2a\log\left(\frac{\ell_c}{2r}\right), \tag*{(14)}

whereas a logarithmic covariance requires a constant slope. With the values of aa and ℓc\ell_c given above, μeff\mu_{\mathrm{eff}} decreases systematically across the resolved range, from approximately 0.220.22 near the Taylor microscale to about 0.050.05 near LL for Rλ=730R_\lambda=730. Thus, over a restricted interval concentrated around and below λ\lambda, a logarithmic fit can return a value close to the conventional μ≈0.21\mu\approx0.21, even though the local slope continues to vary across the broader range. The present analysis differs in that the same pair (a,ℓc)(a,\ell_c) also constrains the one-point cumulants, the rate function, and the moments of εℓ\varepsilon_\ell, providing a joint constraint that a covariance fit alone does not test. Whether the broader logarithmic range reported in Ref. [18] at higher Reynolds numbers represents an asymptotic logarithmic regime or instead a slowly varying effective slope can therefore be tested directly through the local-slope behavior of CΛC_\Lambda.

A complementary test is provided by the cumulants of the coarse-grained logarithmic dissipation

Λℓ=log⁡(εℓ/εˉ),μℓ=⟨Λℓ⟩,σℓ2=Var⁡(Λℓ).(15)\Lambda_\ell=\log(\varepsilon_\ell/\bar{\varepsilon}),\qquad\mu_\ell=\langle\Lambda_\ell\rangle,\qquad\sigma_\ell^2=\operatorname{Var}(\Lambda_\ell). \tag*{(15)}

Let n=log⁡2(L/ℓ)n=\log_2(L/\ell) denote the number of dyadic scale steps between LL and ℓ\ell. For a log-divisible cascade with stationary increments in nn, every cumulant of Λℓ\Lambda_\ell must be proportional to nn. In particular, both μℓ\mu_\ell and σℓ2\sigma_\ell^2 should depend linearly on log⁡(L/ℓ)\log(L/\ell). Figure 3(b) shows a markedly different behavior. For all three Reynolds numbers, the mean and variance are well approximated by quadratic functions,

μℓ≈−(a/2)[log⁡(ℓ/ℓc)]2,σℓ2≈a[log⁡(ℓ/ℓc)]2,(16)\mu_\ell\approx-(a/2)[\log(\ell/\ell_c)]^2,\qquad\sigma_\ell^2\approx a[\log(\ell/\ell_c)]^2, \tag*{(16)}

with the same values of aa and ℓc\ell_c as in (13). The logarithm of the Eulerian coarse-grained dissipation therefore cannot be represented as a process with stationary additive increments in the usual logarithmic scale variable.

The agreement between Figure 3(a) and (Figure 3b) is, however, noteworthy. Upon identifying the coarse-graining scale with ℓ≃2r\ell\simeq2r, the one-point cumulants and the two-point covariance are governed by the same function

n(ℓ):=[log⁡(ℓ/ℓc)]2.(17)n(\ell):=[\log(\ell/\ell_c)]^2. \tag*{(17)}

This suggests that n(ℓ)n(\ell) acts as an effective, nonlinear cascade depth. The data thus retain a coherent organization across scales, but its clock is quadratic rather than linear in the logarithmic scale separation. This observation motivates introducing the scale-averaged logarithmic dissipation

γℓ:=Λℓn(ℓ)=log⁡(εℓ/εˉ)[log⁡(ℓ/ℓc)]2.(18)\gamma_\ell:=\frac{\Lambda_\ell}{n(\ell)}=\frac{\log(\varepsilon_\ell/\bar{\varepsilon})}{[\log(\ell/\ell_c)]^2}. \tag*{(18)}

Figure 3(c) shows finite-scale estimates of the associated rate function,

Jℓ(γ)=−1n(ℓ)log⁡[pγℓ(γ;ℓ)pγℓ(γℓ∗;ℓ)],(19)J_\ell(\gamma)=-\frac{1}{n(\ell)}\log\left[\frac{p_{\gamma_\ell}(\gamma;\ell)}{p_{\gamma_\ell}(\gamma_\ell^*;\ell)}\right], \tag*{(19)}

where γℓ∗\gamma_\ell^* denotes the most probable value, so that min⁡Jℓ=0\min J_\ell=0. The curves obtained at different scales collapse closely over more than one decade in ℓ\ell. This supports the large-deviation form

pγℓ(γ;ℓ)≍exp⁡[−n(ℓ)J(γ)],η≪ℓ≪L,(20)p_{\gamma_\ell}(\gamma;\ell)\asymp\exp[-n(\ell)J(\gamma)],\qquad\eta\ll\ell\ll L, \tag*{(20)}

where equality is understood at logarithmic accuracy. The measured rate function is remarkably close to the parabola

J(γ)≈(γ+a/2)22a.(21)J(\gamma)\approx\frac{(\gamma+a/2)^2}{2a}. \tag*{(21)}

The Eulerian dissipation therefore displays an approximately log-normal large-deviation structure, but with the nonstandard speed n(ℓ)=[log⁡(ℓ/ℓc)]2n(\ell)=[\log(\ell/\ell_c)]^2. This differs fundamentally from both a log-normal multiplicative cascade and the conventional multifractal formalism, for which the large-deviation speed is proportional to log⁡(L/ℓ)\log(L/\ell).

We next examine the implications of this nonstandard large-deviation principle for the moments of the coarse-grained dissipation. Define the normalized moments Mq(ℓ)=⟨(εℓ/εˉ)q⟩M_q(\ell)=\langle(\varepsilon_\ell/\bar{\varepsilon})^q\rangle and their logarithmic derivatives χqloc(ℓ)=dlog⁡Mq(ℓ)/dlog⁡ℓ\chi_q^{\mathrm{loc}}(\ell)=\mathrm{d}\log M_q(\ell)/\mathrm{d}\log\ell. If the dissipation were multifractal over an inertial range, Mq(ℓ)∝(ℓ/L)χqM_q(\ell)\propto(\ell/L)^{\chi_q}, then χqloc(ℓ)\chi_q^{\mathrm{loc}}(\ell) would develop a scale-independent plateau. As seen in Figure 4(a), no extended plateau is observed for q=2q=2, 33, or 44, except possibly over the restricted interval η≲ℓ≲λ\eta\lesssim\ell\lesssim\lambda below the Taylor microscale, where the local slopes can be approximated by effective constants. At Rλ=730R_\lambda=730, this gives χ2≃−0.25\chi_2\simeq-0.25, χ3≃−0.69\chi_3\simeq-0.69, and χ4≃−1.29\chi_4\simeq-1.29, values that can be made compatible with familiar log-normal or log-Poisson estimates. These effective exponents are not, however, maintained over a substantial range of scales. Above the Taylor microscale, the local slopes drift approximately linearly with log⁡ℓ\log\ell, and comparison among the three Reynolds numbers reveals no systematic tendency for scale-independent plateaux to emerge.

Four-panel plot of local slopes, cumulant-generating functions, PDFs, and multiplier variance

Figure 4. (a) Local slopes χqloc(ℓ)=dlog⁡Mq/dlog⁡ℓ\chi_q^{\mathrm{loc}}(\ell)=\mathrm{d}\log M_q/\mathrm{d}\log\ell of Mq(ℓ)=⟨(εℓ/εˉ)q⟩M_q(\ell)=\langle(\varepsilon_\ell/\bar{\varepsilon})^q\rangle. Horizontal lines and shaded bands show estimates of the scaling exponents χq\chi_q over η≲ℓ≲λ\eta\lesssim\ell\lesssim\lambda for the three Reynolds numbers; dashed lines follow from (27). (b) Centered cumulant-generating function plotted against q∣log⁡(ℓ/ℓc)∣q|\log(\ell/\ell_c)| for ℓ∈[2η,2000η]\ell\in[2\eta,2000\eta] and q≤4q\leq4. The dotted line is the quadratic form ax2/2ax^2/2. (c) PDFs of the Eulerian multipliers Mℓ=εℓ/ε2ℓM_\ell=\varepsilon_\ell/\varepsilon_{2\ell} for Rλ=730R_\lambda=730. (d) Variance of MℓM_\ell for the three Reynolds numbers.

This drift follows directly from the large-deviation structure identified above. By the Laplace principle,

1n(ℓ)log⁡Mq(ℓ)⟶ψ(q):=sup⁡γ{qγ−J(γ)},(22)\frac{1}{n(\ell)}\log M_q(\ell)\longrightarrow\psi(q):=\sup_\gamma\{q\gamma-J(\gamma)\}, \tag*{(22)}

where ψ\psi is the scaled cumulant-generating function. For the parabolic rate function (21),

ψ(q)=a2(q2−q).(23)\psi(q)=\frac{a}{2}(q^2-q). \tag*{(23)}

The linear term ensures ψ(1)=0\psi(1)=0, consistently with M1(ℓ)=1M_1(\ell)=1. The moments therefore obey

log⁡Mq(ℓ)≈(a/2)(q2−q)[log⁡(ℓ/ℓc)]2.(24)\log M_q(\ell) \approx(a/2)(q^2-q)[\log(\ell/\ell_c)]^2. \tag*{(24)}

This is a squared-logarithmic scaling law, rather than a power law in ℓ\ell.

Figure 4(b) tests this prediction directly. It displays the centered cumulant-generating function

K(q;ℓ):=log⁡Mq(ℓ)−qμℓ=log⁡⟨exp⁡[q(Λℓ−μℓ)]⟩.(25)K(q;\ell):=\log M_q(\ell)-q\mu_\ell=\log\langle\exp[q(\Lambda_\ell-\mu_\ell)]\rangle. \tag*{(25)}

Equations (16) and (21) predict

K(q;ℓ)≈(a/2)q2n(ℓ)=(a/2)[qlog⁡(ℓ/ℓc)]2.(26)K(q;\ell)\approx(a/2)q^2n(\ell)=(a/2)[q\log(\ell/\ell_c)]^2. \tag*{(26)}

The collapse of the data onto the parabola (a/2)x2(a/2)x^2 confirms this prediction over the resolved scales and for moments up to q=4q=4.

Differentiating (24) gives

χqloc(ℓ)≈a(q2−q)log⁡(ℓ/ℓc).(27)\chi_q^{\mathrm{loc}}(\ell)\approx a(q^2-q)\log(\ell/\ell_c). \tag*{(27)}

The dashed lines in Figure 4(a) show this prediction. They account for the approximately linear evolution of the local slopes throughout the upper part of the inertial range. Their drift is therefore not merely due to the absence of a sufficiently broad fitting interval: it is the moment-space signature of the squared-logarithmic large-deviation speed.

The failure of a self-similar Eulerian cascade is also apparent directly at the level of the multipliers. For each reference scale ℓ\ell, define

Mℓ=εℓ/ε2ℓ,log⁡Mℓ=Λℓ−Λ2ℓ.(28)\mathcal{M}_\ell=\varepsilon_\ell/\varepsilon_{2\ell},\qquad\log\mathcal{M}_\ell=\Lambda_\ell-\Lambda_{2\ell}. \tag*{(28)}

Stationarity of the increments in logarithmic scale would require the distribution of Mℓ\mathcal{M}_\ell to be independent of ℓ\ell. Figure 4(c) instead shows a systematic evolution of the multiplier PDFs with the reference scale. Their widths first increase toward the dissipative range and then decrease continuously as ℓ\ell approaches the integral scale. The variance shown in Figure 4(d) confirms this strong scale dependence for all three Reynolds numbers. The curves become close in the upper part of the inertial range, but retain a Reynolds-number dependence at the smallest resolved scales, where viscous effects are expected. Most importantly, increasing RλR_\lambda does not produce a scale-independent plateau. The multiplier increments therefore remain nonstationary across scales. The observed scale dependence does not exclude more general multiplicative descriptions involving nonstationary or scale-dependent multiplier statistics [17, 39]. Such constructions, however, depart from the scale-invariant cascade phenomenology summarized in Sec. II. More specifically, the present results point to an effective cascade depth n(ℓ)=[log⁡(ℓ/ℓc)]2n(\ell)=[\log(\ell/\ell_c)]^2, whose physical origin is not accounted for by conventional cascade models.

These observations do not constitute a statement about the formal limit Rλ→∞R_\lambda\to\infty. They do, however, show that no conventional multifractal scaling regime is resolved in the present simulations. At the largest Reynolds number, the separation between the integral and Taylor scales reaches approximately L/λ≈50L/\lambda\approx50, yet no extended plateau develops. Moreover, neither the local slopes nor the scale dependence of the logarithmic cumulants or multiplier distributions shows a systematic evolution toward the expectations of a scale-invariant cascade.

Taken together, Figure 3 and Figure 4 establish a coherent but nonclassical organization of Eulerian dissipation. Its logarithm does not have stationary increments in log⁡ℓ\log\ell, and its moments do not obey multifractal power laws. Instead, the one-point cumulants, spatial covariance, rate function, and moments are all governed by the effective depth n(ℓ)=[log⁡(ℓ/ℓc)]2n(\ell)=[\log(\ell/\ell_c)]^2, with an approximately parabolic rate function. Eulerian intermittency thus retains a large-deviation structure, but one whose speed is incompatible with the scale-invariant spatial cascade phenomenology usually invoked in turbulence.

The emergence of this squared-logarithmic depth is not a minor correction to the conventional picture. It implies that the effective amount of statistical accumulation between two scales is not proportional to their logarithmic separation, as it would be for a self-similar hierarchy of scale-local cascade steps. Its physical origin therefore remains to be identified. It reflects non-stationary transfer statistics, correlations extending across scales, or the projection onto space of dynamical events whose natural organization is temporal rather than spatial. Since dissipation is experienced along fluid trajectories, this observation motivates examining whether a more conventional cascade structure emerges in a Lagrangian description.

Multifractality and temporal cascade of Lagrangian dissipation

We now turn to the dissipation sampled along fluid trajectories. Tracer particles evolve according to dX/dt=v(X,t)\mathrm{d}\boldsymbol{X}/\mathrm{d}t=\boldsymbol{v}(\boldsymbol{X},t), and the instantaneous Lagrangian dissipation is defined as ε~(t)=ε(X(t),t)\widetilde{\varepsilon}(t)=\varepsilon(\boldsymbol{X}(t),t). For tracers uniformly distributed in an incompressible flow, the one-time probability distribution of ε~\widetilde{\varepsilon} coincides with the single-point Eulerian distribution of ε\varepsilon. Any difference between the two descriptions must therefore originate from the correlations and scale organization of the fluctuations, rather than from their instantaneous statistics. To probe this temporal organization, we introduce the dissipation averaged along a trajectory over a time interval τ\tau,

ε~τ(t)=1τ∫tt+τε~(t′) dt′.(29)\widetilde{\varepsilon}_\tau(t)=\frac{1}{\tau}\int_t^{t+\tau}\widetilde{\varepsilon}(t')\,\mathrm{d}t'. \tag*{(29)}

We use the large-time cutoff τc=3.2τL\tau_c=3.2\tau_L, as determined below from the logarithmic scaling of the mean and variance of log⁡(ε~τ/εˉ)\log(\widetilde{\varepsilon}_\tau/\bar{\varepsilon}) [Figure 6(b)].

Figure 6. (a) Temporal covariance of Λ~(t)=log⁡[ε~(t)/εˉ]\widetilde{\Lambda}(t)=\log[\widetilde{\varepsilon}(t)/\bar{\varepsilon}]. The dashed line shows the logarithmic behavior 0.33log⁡(τ∗/τ)0.33\log(\tau_\ast/\tau). (b) Mean μ~τ\widetilde{\mu}_\tau and variance σ~τ2\widetilde{\sigma}_\tau^2 of Λ~τ=log⁡(ε~τ/εˉ)\widetilde{\Lambda}_\tau=\log(\widetilde{\varepsilon}_\tau/\bar{\varepsilon}). Dotted lines indicate linear dependences on log⁡(τ/τc)\log(\tau/\tau_c). (c) PDFs of the temporal multipliers Nτ=ε~τ/ε~2τN_\tau=\widetilde{\varepsilon}_\tau/\widetilde{\varepsilon}_{2\tau}, for Rλ=730R_\lambda=730 and the indicated coarse-graining times. (d) Correlation coefficients Corr⁡(log⁡Nτ,log⁡Nτ′)\operatorname{Corr}(\log N_\tau,\log N_{\tau^\prime}) between logarithmic multipliers, as functions of the timescale ratio τ′/τ\tau^\prime/\tau, for the same reference times τ\tau as in panel (c).

We first examine whether ε~τ\widetilde{\varepsilon}_\tau displays multifractal scaling. Its normalized moments are M‾q(τ)=⟨(ε~τ/εˉ)q⟩\overline{M}_q(\tau)=\langle(\widetilde{\varepsilon}_\tau/\bar{\varepsilon})^q\rangle. Figure 5(a) shows their local logarithmic slopes χqloc(τ)=\chi_q^{\mathrm{loc}}(\tau)= (dlog⁡M~q(τ))/(dlog⁡τ)(\mathrm{d}\log\tilde{M}_q(\tau))/(\mathrm{d}\log\tau). For q=2,3,4q=2,3,4, the slopes develop broad plateaux over the interval 8τη<τ<0.4τL8\tau_\eta<\tau<0.4\tau_L. These plateaux are observed at both Reynolds numbers and become more clearly defined as RλR_\lambda increases. This behavior sharply contrasts with the Eulerian results of Sec. III, where the local slopes retain a pronounced scale dependence, and provides direct evidence for the inertial-range power laws M~q(τ)∝(τ/τc)χ~q\tilde{M}_q(\tau)\propto(\tau/\tau_c)^{\tilde{\chi}_q}.

Three panels showing local slopes, scaling exponents, and the temporal rate function

Figure 5. (a) Local slopes χ~qloc(τ)=dlog⁡M~q(τ)/dlog⁡τ\tilde{\chi}^{\mathrm{loc}}_q(\tau)=\mathrm{d}\log\tilde{M}_q(\tau)/\mathrm{d}\log\tau of the moments M~q(τ)=⟨(ε~τ/εˉ)q⟩\tilde{M}_q(\tau)=\langle(\tilde{\varepsilon}_\tau/\bar{\varepsilon})^q\rangle. The vertical dotted lines delimit the interval 8τη<τ<0.4τL8\tau_\eta<\tau<0.4\tau_L used to estimate the scaling exponents, shown as horizontal lines. The shaded regions indicate the corresponding uncertainties, determined from the extrema of the local slopes over this interval. (b) Measured scaling exponents χ~q\tilde{\chi}_q, with associated error bars. The dashed curve is obtained from the Legendre transform of the quartic fit to the rate function shown in (c), while the dotted curve is the quadratic log-normal prediction χ~q=0.14q(1−q)\tilde{\chi}_q=0.14q(1-q). (c) Finite-scale estimates of the temporal rate function I~(α~)\tilde{I}(\tilde{\alpha}), with α~=log⁡(ε~τ/εˉ)/log⁡(τ/τc)\tilde{\alpha}=\log(\tilde{\varepsilon}_\tau/\bar{\varepsilon})/\log(\tau/\tau_c), for the indicated coarse-graining times and Reynolds numbers. The dashed curve is a quartic fit to the collapsed data, while the dotted curve is the quadratic rate function associated with the log-normal prediction in panel (b).

The resulting exponents are reported in Figure 5(b). The measurements at the two Reynolds numbers agree within their uncertainties and define a nonlinear concave function of qq. The constraints χ~0=0\tilde{\chi}_0=0 and χ~1=0\tilde{\chi}_1=0 follow, respectively, from normalization and from ⟨ε~τ⟩=εˉ\langle\tilde{\varepsilon}_\tau\rangle=\bar{\varepsilon}. The nonlinear dependence on qq establishes the multifractal character of the temporally coarse-grained dissipation. A quadratic approximation, χ~qLN=0.14q(1−q)\tilde{\chi}^{\mathrm{LN}}_q=0.14q(1-q), provides a reasonable description at low orders, but predicts increasingly negative exponents as qq increases and thus systematically overestimates their magnitude. The Lagrangian dissipation is therefore multifractal, but its statistics are not accurately described by a log-normal model.

The same conclusions follows from the complementary large-deviation formulation. Multifractal scaling implies

p(α~;τ)≍(τ/τc)I~(α~),α~=log⁡(ε~τ/εˉ)log⁡(τ/τc),(30)p(\tilde{\alpha};\tau)\asymp(\tau/\tau_c)^{\tilde{I}(\tilde{\alpha})},\qquad\tilde{\alpha}=\frac{\log(\tilde{\varepsilon}_\tau/\bar{\varepsilon})}{\log(\tau/\tau_c)}, \tag*{(30)}

where I~(α~)\tilde{I}(\tilde{\alpha}) is a convex temporal rate function normalized by inf⁡α~I~(α~)=0\inf_{\tilde{\alpha}}\tilde{I}(\tilde{\alpha})=0. The moment exponents are related to it through χ~q=inf⁡α~[qα~+I~(α~)]\tilde{\chi}_q=\inf_{\tilde{\alpha}}[q\tilde{\alpha}+\tilde{I}(\tilde{\alpha})]. Figure 5(c) shows the finite-scale estimates of I~(α~)\tilde{I}(\tilde{\alpha}). Results obtained at different coarse-graining times collapse closely onto a common curve, with only weak dependence on Reynolds number. The measured rate function is clearly non-quadratic. Its Legendre transform, obtained from the quartic fit shown by the dashed curve in Figure 5(c), is reported as the dashed curve in panel (b) and reproduces the measured exponents, including their departure from the quadratic log-normal prediction. The local-slope plateaux, the nonlinear exponents, and the collapse of the rate functions thus provide mutually consistent evidence for non-log-normal multifractality.

Similar inertial-range scaling and logarithmic statistics of the Lagrangian dissipation were previously reported at Rλ=400R_\lambda=400 by Huang & Schmitt [25]. Their results were found to be approximately compatible with a log-normal model. The present data, obtained at higher Reynolds numbers and with improved statistical resolution, confirm the multifractal scaling of the Lagrangian coarse-grained dissipation while revealing systematic departures from log-normality.

We next ask whether this multifractal scaling is accompanied by the statistical organization expected from a multiplicative process. We first consider the logarithm of the instantaneous dissipation, Λ^(t)=log⁡[ε~(t)/εˉ]\hat{\Lambda}(t)=\log[\tilde{\varepsilon}(t)/\bar{\varepsilon}]. Its connected temporal covariance,

CΛ^(τ)=⟨Λ^(t)Λ^(t+τ)⟩−⟨Λ^⟩2,(31)C_{\hat{\Lambda}}(\tau)=\langle\hat{\Lambda}(t)\hat{\Lambda}(t+\tau)\rangle-\langle\hat{\Lambda}\rangle^2, \tag*{(31)}

is shown in Figure 6(a). Over most of the inertial range, it follows

CΛ^(τ)≃c2log⁡(τ∗/τ),c2≃0.33,τ∗≃0.55τL.(32)C_{\hat{\Lambda}}(\tau)\simeq c_2\log(\tau_*/\tau),\qquad c_2\simeq0.33,\quad\tau_*\simeq0.55\tau_L. \tag*{(32)}

This logarithmic decay is the temporal counterpart of the logarithmic spatial correlations predicted by scale-invariant multiplicative cascades. Its clarity is particularly striking in view of the absence of an analogous regime in the Eulerian dissipation field.

Further evidence is provided by the logarithm of the temporally averaged dissipation, Λ^τ=log⁡(ε~τ/εˉ)\hat{\Lambda}_\tau=\log(\tilde{\varepsilon}_\tau/\bar{\varepsilon}). As shown in Figure 6(b), both its mean μ~τ=⟨Λ~τ⟩\widetilde{\mu}_\tau=\langle\widetilde{\Lambda}_\tau\rangle and its variance σ~τ2=Var⁡(Λ~τ)\widetilde{\sigma}_\tau^2=\operatorname{Var}(\widetilde{\Lambda}_\tau) vary linearly with the logarithm of the coarse-graining time,

μ~τ≃c1log⁡(τ/τc),σ~τ2≃c2log⁡(τc/τ).(33)\widetilde{\mu}_\tau\simeq c_1\log(\tau/\tau_c),\qquad\widetilde{\sigma}_\tau^2 \simeq c_2\log(\tau_c/\tau). \tag*{(33)}

The two linear dependences consistently extrapolate to zero at τc≃3.2τL\tau_c\simeq3.2\tau_L, thereby fixing the large-time cutoff used above in the definition of α~\widetilde{\alpha}. In particular, τc\tau_c is determined independently of the collapse of the rate functions in Figure 5(c). The variance has the same logarithmic slope c2≃0.33c_2\simeq0.33 as the covariance of the instantaneous signal. For a scale-invariant log-divisible process, the cumulants of Λ~τ\widetilde{\Lambda}_\tau are expected to depend linearly on log⁡(τ/τc)\log(\tau/\tau_c). Equation (33) tests this prediction for the first two cumulants. This behavior again differs markedly from the quadratic dependence found under Eulerian spatial coarse-graining.

A direct scale-by-scale test is provided from the temporal multipliers

Nτ=ε~τε~2τ.(34)N_\tau=\frac{\widetilde{\varepsilon}_\tau}{\widetilde{\varepsilon}_{2\tau}}. \tag*{(34)}

Their logarithms are the increments of Λ~τ\widetilde{\Lambda}_\tau between timescales separated by a fixed ratio of two. As shown in Figure 6(c), the probability distributions of NτN_\tau collapse closely for the various values of τ\tau within the inertial range. The multiplier statistics are therefore approximately stationary with respect to translation of the logarithmic timescale.

Scale invariance of the multiplier distributions does not imply that successive scale increments are independent. We therefore consider ρ(τ,τ′)=Corr⁡(log⁡Nτ,log⁡Nτ′)\rho(\tau,\tau^\prime)=\operatorname{Corr}(\log N_\tau,\log N_{\tau^\prime}). As shown in Figure 6(d), ρ(τ,τ′)\rho(\tau,\tau^\prime) depends only weakly on the reference scale τ\tau and decays primarily with the ratio τ′/τ\tau^\prime/\tau. The correlations remain measurable over scale ratios extending to several tens. Because the multipliers are constructed from nested temporal averages, part of this correlation may reflect the overlap of the averaging intervals. The measurements of Figure 6(d) nevertheless show that the logarithmic scale increments cannot be treated as independent. They are approximately stationary in logarithmic timescale, but retain correlations over several scales.

Taken together, the logarithmic covariance, the linear dependence of the first two logarithmic cumulants, and the scale invariance of the multiplier distributions provide strong evidence for an approximately scale-stationary multiplicative process in time. These observations support a log-divisible description of the single-scale distributions. The correlations in Figure 6(d), however, rule out the simplest cascade of independent multipliers and instead reveal a correlated multiplicative organization with memory extending over several temporal scales.

The particularly clear multiplicative organization observed in the Lagrangian frame may be related to the dynamics of material deformation [40, 41, 42]. Let Jij(t,t0)=∂Xi(t)/∂Xj(t0)J_{ij}(t,t_0)=\partial X_i(t)/\partial X_j(t_0) denote the deformation gradient along a trajectory. It satisfies ∂tJij=AikJkj\partial_tJ_{ij}=A_{ik}J_{kj} with Aij=∂jviA_{ij}=\partial_jv_i. This leads to rewriting the velocity gradient as A=∂tJJ−1A=\partial_tJ J^{-1}, so that the dissipation can be expressed exactly as

ε~(t)=ν2[(∂tJJ−1)+(∂tJJ−1)T]2(35)\widetilde{\varepsilon}(t)=\frac{\nu}{2}\left[\left(\partial_tJ J^{-1}\right)+\left(\partial_tJ J^{-1}\right)^{\mathrm{T}}\right]^2 \tag*{(35)}

It therefore measures the squared instantaneous rate of material deformation.

Finite-time deformations compose multiplicatively, J(t2,t0)=J(t2,t1)J(t1,t0)J(t_2,t_0)=J(t_2,t_1)J(t_1,t_0), so that a fluid trajectory retains the causal succession of stretching and compression events experienced by the same material element. (35) does not imply that the dissipation is itself a multiplicative functional of JJ; rather, it shows that its instantaneous fluctuations are generated within a material history whose finite-time deformations compose multiplicatively.

The inter-timescale correlations observed in Figure 6(d) may reflect the fluctuating lifetime of Lagrangian gradient-amplification events. Indeed, the quadratic self-amplification term in the velocity-gradient dynamics introduces a local evolution time of order τloc∼∣A∣−1\tau_{\mathrm{loc}} \sim|\mathbf{A}|^{-1}. At dissipative scales, this reduces, up to factors of order unity, to the local Kolmogorov time τloc∼ν/ε~\tau_{\mathrm{loc}} \sim\sqrt{\nu/\tilde{\varepsilon}}, while for coarse-grained gradients it is of the order of the corresponding local eddy-turnover time. A coherent amplification event may therefore contribute to several neighbouring temporal averages and correlate their associated multipliers. The material history sampled along a Lagrangian trajectory can thus retain memory across several temporal scales, rather than reducing to a sequence of independent cascade steps.

This material genealogy is lost under Eulerian spatial coarse-graining. Nested Eulerian regions contain fluid elements with different deformation histories and therefore mix distinct sequences of strain amplification. The preservation of a single causal history along each trajectory provides a plausible explanation for why the Lagrangian statistics recover the standard organization of a scale-invariant temporal cascade, with a cascade depth proportional to log⁡(τ/τc)\log(\tau/\tau_c). This interpretation is consistent with causal multifractal processes and with Lagrangian velocity-gradient models based on recent material deformation [40, 43, 44, 45]. It should not be viewed as a derivation of the observed temporal cascade from the Navier–Stokes equations, but rather as a dynamical rationale for its emergence.

The contrast with the Eulerian results of Sec. III is therefore not simply between multifractal and non-multifractal statistics. Both descriptions exhibit a large-deviation structure, but with fundamentally different scale depths. In the Lagrangian frame, the rate function is organized by the usual logarithmic depth n(τ)∝log⁡(τ/τc)n(\tau) \propto\log(\tau/\tau_c), consistently with logarithmic correlations and approximately scale-stationary multipliers. In the Eulerian frame, by contrast, the large-deviation collapse requires the non-standard depth n(ℓ)∝[log⁡(ℓ/ℓc)]2n(\ell) \propto[\log(\ell/\ell_c)]^2, while logarithmic spatial correlations and scale-independent multipliers are absent. The bridge between the two descriptions must therefore account not only for the conversion of spatial scales into temporal ones, but also for this change in the effective cascade depth. Classical bridge relations [26, 28] are re-examined from this perspective in the following section.

Revisiting bridge relations: from spatial sampling to temporal inheritance

Bridge relations seek to connect the spatial organization of turbulence to the temporal fluctuations experienced along particle trajectories. For velocity increments, dimensional conversions between spatial separations and eddy-turnover times have proved remarkably successful, yielding Eulerian and Lagrangian scaling exponents that are broadly consistent with one another [23]. Their application to dissipation is more delicate. Indeed, εℓ\varepsilon_\ell is an instantaneous spatial average, whereas ε~τ\tilde{\varepsilon}_{\tau} records the history of dissipation along a moving particle. Relating these two quantities therefore requires assumptions not only about the conversion between ℓ\ell and τ\tau, but also about how dissipation is sampled, transported, and inherited by material trajectories.

In the classical construction, illustrated in Figure 7a, a spatial scale ℓ\ell is first associated with its local eddy-turnover time. Using the refined-similarity estimate δℓu3/ℓ∼εℓ\delta_\ell u^3/\ell\sim\varepsilon_\ell gives τℓ∼εℓ−1/3ℓ2/3\tau_\ell\sim\varepsilon_\ell^{-1/3}\ell^{2/3}. The bridge then makes a second, stronger assumption: during this interval, a representative tracer samples the dissipation carried by the corresponding region sufficiently well that its temporal average can be identified with the instantaneous spatial average. These two steps lead to

Schematic representations of spatial sampling and temporal inheritance

Figure 7. Schematic representations of (a) the classical spatial-sampling bridge, ε~τℓ∼εℓ\tilde{\varepsilon}_{\tau_\ell} \sim\varepsilon_\ell for τℓ∼ℓ1/3ℓ2/3\tau_\ell\sim\ell^{1/3}\ell^{2/3}, and (b) dynamical inheritance. After the material doubling time τℓ\tau_\ell, the pre-image B~ℓ\widetilde{\mathcal{B}}_\ell of Bℓ\mathcal{B}_\ell lies within the parent region B~2ℓ\widetilde{\mathcal{B}}_{2\ell}. The resulting dynamical multiplier M~ℓ\widetilde{M}_\ell is statistically related to the Eulerian multiplier MℓM_\ell.

ε~τℓ∼εℓ,τℓ=cεℓ−1/3ℓ2/3,(36)\widetilde{\varepsilon}_{\tau_\ell} \sim\varepsilon_\ell,\qquad\tau_\ell= c\varepsilon_\ell^{-1/3}\ell^{2/3}, \tag*{(36)}

where the order-one constant cc accounts for the ambiguity in defining a turnover time. This is the spatial-sampling interpretation of the classical bridge.

Taken literally, Eq. (36) identifies not only the characteristic scales of the two descriptions, but also their local dissipation fluctuations. This is difficult to reconcile with the results of the previous sections. Lagrangian time averages display approximately scale-stationary multipliers and the usual logarithmic cascade depth ∝log⁡(τ/τc)\propto\log(\tau/\tau_c), whereas Eulerian spatial averages have scale-dependent multipliers and require a depth ∝[log⁡(ℓ/ℓc)]2\propto[\log(\ell/\ell_c)]^2. This difference in scale organization makes a trajectory-wise interpretation of Eq. (36) questionable. It does not, however, by itself imply that the two descriptions are statistically unrelated. We therefore test the stronger assumption of a one-to-one trajectory-wise correspondence directly.

For each tracer, we identify the Eulerian cell of size ℓ\ell containing its position at the reference time tt. The corresponding Eulerian dissipation εℓ\varepsilon_\ell determines the local time τℓ=cεℓ−1/3ℓ2/3\tau_\ell=c\varepsilon_\ell^{-1/3}\ell^{2/3}. It is compared with the backward Lagrangian average ε~τℓ=(1/τℓ)∫t−τℓtε(X(s),s) ds\widetilde{\varepsilon}_{\tau_\ell}=(1/\tau_\ell)\int_{t-\tau_\ell}^{t}\varepsilon(\boldsymbol{X}(s),s)\,\mathrm{d}s. Figure 8(a) conditions Λ~τℓ−=log⁡(ε~τℓ/εˉ)\widetilde{\Lambda}_{\tau_\ell}^{-}=\log(\widetilde{\varepsilon}_{\tau_\ell}/\bar{\varepsilon}) on Λℓ=log⁡(εℓ/εˉ)\Lambda_\ell=\log(\varepsilon_\ell/\bar{\varepsilon}) for c=1c=1, grouping cells into equal-population bins of the latter quantity. Figure 8(b) measures their normalized root-mean-square difference Eℓ(c)=⟨(Λ~τℓ−−Λℓ)2⟩1/2/[Var⁡(Λℓ)]1/2\mathcal{E}_\ell(c)=\langle(\widetilde{\Lambda}_{\tau_\ell}^{-}-\Lambda_\ell)^2\rangle^{1/2}/[\operatorname{Var}(\Lambda_\ell)]^{1/2} as a function of the parameter cc.

Direct test of the classical dissipation bridge

Figure 8. Direct test of the classical dissipation bridge at Rλ=730R_{\lambda}=730. (a) Conditional average of the backward Lagrangian log-dissipation Λ~τℓ−=log⁡(ε~τℓ/εˉ)\widetilde{\Lambda}_{\tau_\ell}^{-}=\log(\widetilde{\varepsilon}_{\tau_\ell}/\bar{\varepsilon}) given the Eulerian one Λℓ=log⁡(εℓ/εˉ)\Lambda_\ell=\log(\varepsilon_\ell/\bar{\varepsilon}), for c=1c=1 and the indicated scales. The shaded area shows the conditional standard deviation, and the dashed line represents the direct bridge. (b) Normalized rms logarithmic difference between both quantities as a function of cc. Dotted lines mark c=1c=1 and a unit error.

For c=1c=1, the conditional averages in Figure 8a increase with εℓ\varepsilon_\ell, confirming appreciable Eulerian–Lagrangian coupling. Nevertheless, the fluctuations among trajectories remain broad and the conditional means depart from the predicted diagonal. Correspondingly, Eℓ(c)\mathcal{E}_\ell(c) remains of order unity, and no single choice of cc reduces it simultaneously at all scales. The classical bridge therefore fails as a one-to-one relation for a representative trajectory, rather than through a complete absence of coupling between the Eulerian and Lagrangian descriptions.

This failure has a simple physical interpretation. The Eulerian quantity εℓ\varepsilon_\ell averages over the many fluid elements occupying a coarse-graining region Bℓ\mathcal{B}_\ell of size ℓ\ell at time tt, whereas ε~τℓ\widetilde{\varepsilon}_{\tau_\ell} records the history of one such element. Their connection must therefore be considered at the level of the ensemble of trajectories ending in Bℓ\mathcal{B}_\ell, rather than through a representative trajectory. The classical bridge replaces the resulting distribution by a representative trajectory, amounting to a mean-field approximation that Figure 8 rules out. Moreover, when this ensemble is followed backward in time, it is transported and deformed and can no longer be represented by the same spatial region. A refined bridge must consequently account for both trajectory-to-trajectory fluctuations and material deformation.

Once the one-to-one bridge is abandoned, the natural quantity to consider is the refinement between two successive spatial scales, measured by the Eulerian multiplier Mℓ=εℓ/ε2ℓ\mathcal{M}_\ell=\varepsilon_\ell/\varepsilon_{2\ell}. A conventional spatial cascade would associate this refinement with an approximately scale-independent multiplier. Section III showed that this is not the case. Its scale dependence need not, however, be devoid of structure and we next examine how multiplier statistics themselves evolve across scales.

Figure 9 establishes one of the central results of this work: the Eulerian multipliers possess a multifractal scale organization of their own. Their lack of scale invariance is not an arbitrary drift of their distributions. The mean and variance of Xℓ=log⁡Mℓ\mathcal{X}_\ell=\log\mathcal{M}_\ell vary linearly with the logarithmic distance from the cutoff scale, with compatible behaviors at both Reynolds numbers (Figure 9a). Correspondingly, the rescaled variable β=Xℓ/log⁡(ℓ/ℓc)\beta=\mathcal{X}_\ell/\log(\ell/\ell_c) obeys a finite-scale large-deviation principle, while the multiplier moments follow power laws,

Figure 9. Multifractal statistics of Eulerian dissipation multipliers. (a) Mean and variance of Xℓ=log⁡Mℓ=log⁡(εℓ/ε2ℓ)X_{\ell} = \log\mathcal{M}_{\ell} = \log(\varepsilon_{\ell}/\varepsilon_{2\ell}) at Rλ=420R_{\lambda} = 420 and 730730. Lines show the logarithmic fits indicated in the legend with b=0.013b = 0.013 and q⋆≃2.15q_{\star} \simeq2.15; vertical dotted lines mark the Taylor microscales. (b) Finite-scale rate functions of β=Xℓ/log⁡(ℓ/ℓc)\beta= X_{\ell}/\log(\ell/\ell_{c}) at the indicated scales. The dotted curve is the parabolic (log-normal) fit K(β)=(β−bq⋆/2)2/(2b)K(\beta) = (\beta- bq_{\star}/2)^2/(2b), with the same values of bb and q⋆q_{\star} as in panel (a). (c) Multiplier moment exponents ψℓ(q)=log⁡⟨Mℓq⟩/log⁡(ℓ/ℓc)\psi_{\ell}(q) = \log\langle\mathcal{M}_{\ell}^{q}\rangle/\log(\ell/\ell_{c}); error bars indicate their scale-to-scale variation. The dotted curve is ψ(q)=(b/2)q(q⋆−q)\psi(q) = (b/2)q(q_{\star} - q). (d) Correlation coefficients Corr⁡(Xℓ,Xℓ′)\operatorname{Corr}(X_{\ell}, X_{\ell^{\prime}}) between the logarithms of concentric Eulerian multipliers at Rλ=730R_{\lambda} = 730, plotted against the scale ratio ℓ′/ℓ\ell^{\prime}/\ell for the indicated reference scales.

pℓ(β)≍(ℓ/ℓc)K(β),⟨Mℓq⟩∼(ℓ/ℓc)ψ(q).(37)p_\ell(\beta)\asymp(\ell/\ell_c)^{K(\beta)},\qquad\langle\mathcal{M}_\ell^q\rangle\sim(\ell/\ell_c)^{\psi(q)}. \tag*{(37)}

These scaling forms are understood over the resolved scale interval. Indeed, nested averaging regions imply 0<Mℓ≤80<\mathcal{M}_\ell\leq8, so positive-order moments remain bounded and branches with ψ(q)<0\psi(q)<0 cannot persist as ℓ/ℓc→0\ell/\ell_c\to0. Within this finite-range interpretation, the collapse of the rate functions in Figure 9b and the nonlinear, approximately scale-independent function ψ(q)\psi(q) in Figure 9c provide equivalent representations of this multifractality. The conventional logarithmic cascade depth that is absent from the coarse-grained Eulerian dissipation therefore reappears one level lower in the hierarchy, in the increments log⁡εℓ−log⁡ε2ℓ\log\varepsilon_\ell-\log\varepsilon_{2\ell}.

The collapsed rate functions and moment exponents further show that this multiplier multifractality is approximately, but not exactly, log-normal. Over the central range of fluctuations, K(β)K(\beta) is close to the parabolic form shown in Figure 9b, while ψ(q)\psi(q) is reasonably described by the corresponding quadratic approximation in Figure 9c. Equivalently, Xℓ\mathcal{X}_\ell is approximately Gaussian, although systematic departures in the tails show that the log-normal description is not complete. A notable feature of the quadratic approximation is its nontrivial zero at q=q⋆≃2.15q=q_\star\simeq2.15: ψ(q⋆)=0\psi(q_\star)=0, ⟨Mℓq⋆⟩≃const⁡\langle\mathcal{M}_\ell^{q_\star}\rangle\simeq\operatorname{const}. Within the log-normal approximation, the same condition follows from the scale-independent relation ⟨Xℓ⟩≃−(q⋆/2)Var⁡(Xℓ)\langle\mathcal{X}_\ell\rangle\simeq-(q_\star/2)\operatorname{Var}(\mathcal{X}_\ell) observed in Figure 9a. The positive root q⋆q_\star, unlike the trivial normalization condition at q=0q=0, expresses a balance between the typical refinements with Mℓ<1\mathcal{M}_\ell<1 and the rarer intense events with Mℓ>1\mathcal{M}_\ell>1. If it persists over a broader inertial range, this approximately conserved moment would constitute a generalized constraint on the spatial refinement process, and hence on any dynamical bridge relating it to Lagrangian multipliers.

The observed intermittency of multipliers differs fundamentally from a conventional spatial cascade where a dyadic refinement would have the same statistics at every level. Here, instead, the fluctuations of a single refinement become increasingly broad as ℓ\ell moves away from ℓc\ell_{c}: the multiplier retains a memory of how far the cascade has progressed. The equal-time spatial refinement is therefore not an autonomous, scale-local step. The scale dependence of the Eulerian multiplier is instead a constraint that any connection between spatial and temporal descriptions must account for. A possible interpretation is that it reflects the material histories accumulated before fluid elements are assembled into the child region. Transport, deformation, and mixing provide natural mechanisms for such a reorganization, but the present measurements do not directly establish that they transform the scale-stationary temporal multipliers into the observed spatial multipliers. The measured multiplier multifractality therefore provides a quantitative constraint on, rather than a direct demonstration of, such a dynamical bridge.

The multifractality of Eulerian dissipation multipliers provides a statistical interpretation of the squared-logarithmic large-deviation depth found in Sec. III. Along a dyadic hierarchy ℓj=ℓc/2j\ell_j = \ell_c/2^j with ℓn=ℓ\ell_n = \ell, the identity Λℓ=∑j=1nXℓj\Lambda_{\ell} = \sum_{j=1}^{n} X_{\ell_j} gives ⟨Λℓ⟩=∑j=1n⟨Xℓj⟩\langle\Lambda_{\ell}\rangle= \sum_{j=1}^{n}\langle X_{\ell_j}\rangle and Var⁡(Λℓ)=∑j=1nVar⁡(Xℓj)+2∑i<jCov⁡(Xℓi,Xℓj)\operatorname{Var}(\Lambda_{\ell}) = \sum_{j=1}^{n}\operatorname{Var}(X_{\ell_j}) + 2\sum_{i<j}\operatorname{Cov}(X_{\ell_i}, X_{\ell_j}). The number of multiplier increments grows as n∼∣log⁡(ℓ/ℓc)∣n \sim|\log(\ell/\ell_c)|, while the mean and variance of each increment grow linearly with its logarithmic distance from ℓc\ell_c. Summing these marginal contributions therefore produces a quadratic dependence ∼[log⁡(ℓ/ℓc)]2\sim[\log(\ell/\ell_c)]^2. For the variance, however, the measured covariance contributions are comparable to the sum of marginal variances and cannot be neglected. Figure 9d shows that the correlation coefficients depend primarily on the scale ratio and decrease rapidly over the available separations. If this behavior extends to a summable correlation kernel in logarithmic scale, the accumulated interscale covariances can also contribute at order n2n^2, modifying the coefficient while preserving

the squared-logarithmic dependence. The measurements support this interpretation over the resolved hierarchy, although the limited range of scale pairs above the Taylor microscale precludes establishing its asymptotic validity.

The Taylor microscale λ\lambda approximately marks the lower boundary of this multiplier regime. In Figure 9a, the logarithmic behavior of the multiplier cumulants is observed mainly for ℓ≳λ\ell\gtrsim\lambda, while smaller scales display a Reynolds-number-dependent departure. Consistently, the rate functions in Figure 9b are approximately compatible from 64η64\eta upward, whereas the result at 32η32\eta, below or close to λ\lambda, remains distinct. The relevance of λ\lambda has a natural space-time interpretation: the Taylor microscale is, up to a conventional factor, the distance over which which dissipative structures are swept during one Kolmogorov time. The observed crossover is therefore consistent with a transition from nearly frozen spatial sampling at smaller scales to a regime in which temporal evolution becomes important. Uniform sweeping alone, however, cannot alter equal-time multiplier statistics. Their nontrivial scale dependence must involve the evolution and deformation of the dissipation field during transport. This motivates the material-inheritance construction illustrated in Figure 7b, which provides a way to make explicit the distinction between an instantaneous Eulerian parent and the material history contributing to the present child.

The fluid elements occupying Bℓ\mathcal{B}_{\ell} at time tt carry dissipation fluctuations acquired during their previous evolution. Denote by B~ℓ(t−τ)\widetilde{\mathcal{B}}_{\ell}(t-\tau) the pre-image of Bℓ(t)\mathcal{B}_{\ell}(t) under the flow. As τ\tau increases, this material region is transported, stretched, and folded. We define τℓ\tau_{\ell} as the first backward time at which the characteristic size of the pre-image reaches 2ℓ2\ell, and denote by B~2ℓ(t−τℓ)\widetilde{\mathcal{B}}_{2\ell}(t-\tau_{\ell}) a region of size 2ℓ2\ell containing this pre-image. We henceforth denote by εℓ(t)\varepsilon_\ell(t) the dissipation averaged over the child region Bℓ(t)\mathcal{B}_\ell(t), and by ε2ℓ(t−τℓ)\varepsilon_{2\ell}(t-\tau_\ell) that averaged over its dynamically selected parent B~2ℓ(t−τℓ)\widetilde{\mathcal{B}}_{2\ell}(t-\tau_\ell) (see Figure 7b). With this convention, the dynamical refinement reads

M~ℓ(t)=εℓ(t)ε2ℓ(t−τℓ).(38)\widetilde{\mathcal{M}}_\ell(t)=\frac{\varepsilon_\ell(t)}{\varepsilon_{2\ell}(t-\tau_\ell)}. \tag*{(38)}

Unlike the Eulerian multiplier Mℓ(t)\mathcal{M}_\ell(t), whose parent and child are sampled simultaneously, M~ℓ\widetilde{\mathcal{M}}_\ell compares the present child with its dynamically selected parent in the past. It therefore represents temporal inheritance rather than instantaneous spatial refinement.

To connect this dynamical refinement to the Lagrangian cascade, we use the correspondence between space and time only at the level of filtering. For an inertial-range scale ℓ\ell, averaging the instantaneous dissipation over all tracers in a region of size ℓ\ell suppresses fluctuations associated with times shorter than the corresponding turnover time τℓ\tau_\ell. We therefore use the following statistical representation: after averaging over the tracer ensemble occupying a spatial region, the instantaneous spatial average can be related to time-coarse-grained dissipation along the contributing trajectories. For uniformly distributed tracers, we further hypothesize the following approximate statistical correspondences:

εℓ(t)∼stat1Nℓ∑xi(t)∈Bℓ(t)ε~τℓ,i(t),\varepsilon_\ell(t)\mathrel{\stackrel{\mathrm{stat}}{\sim}}\frac{1}{N_\ell}\sum_{\mathbf{x}_i(t)\in\mathcal{B}_\ell(t)}\widetilde{\varepsilon}_{\tau_\ell,i}(t),
ε2ℓ(t−τℓ)∼stat1N2ℓ∑xj(t−τℓ)∈B~2ℓ(t−τℓ)ε~τ2ℓ,j(t−τℓ).(39)\varepsilon_{2\ell}(t-\tau_\ell)\mathrel{\stackrel{\mathrm{stat}}{\sim}}\frac{1}{N_{2\ell}}\sum_{\mathbf{x}_j(t-\tau_\ell)\in\widetilde{\mathcal{B}}_{2\ell}(t-\tau_\ell)}\widetilde{\varepsilon}_{\tau_{2\ell},j}(t-\tau_\ell). \tag*{(39)}

where NℓN_\ell and N2ℓN_{2\ell} are the numbers of tracers in the child region Bℓ(t)\mathcal{B}_\ell(t) and in the enclosing parent region B~2ℓ(t−τℓ)\widetilde{\mathcal{B}}_{2\ell}(t-\tau_\ell), respectively. The latter contains the material pre-image of the child as well as additional fluid elements, so the two tracer counts generally differ. The proposed statistical correspondences concern averages over these ensembles and do not assert that the temporal average along an individual tracer equals the corresponding spatial average.

Along each trajectory ending in the child region, the two temporal averages are related by the scale-adapted Lagrangian multiplier

Nℓ,i(t−τℓ)=ε~τℓ,i(t−τℓ)ε~τ2ℓ,i(t−τℓ).(40)\mathcal{N}_{\ell,i}(t-\tau_\ell)=\frac{\widetilde{\varepsilon}_{\tau_\ell,i}(t-\tau_\ell)}{\widetilde{\varepsilon}_{\tau_{2\ell},i}(t-\tau_\ell)}. \tag*{(40)}

This is the same temporal refinement as Nτ\mathcal{N}_\tau in Eq. (34), except that the two averaging times are now selected by the spatial scales ℓ\ell and 2ℓ2\ell; it reduces exactly to Nτℓ\mathcal{N}_{\tau_\ell} when τ2ℓ=2τℓ\tau_{2\ell}=2\tau_\ell. Equations (38) and (39) then give

M~ℓ(t)∼statN2ℓNℓ∑xi(t)∈Bℓ(t)Nℓ,iε~τ2ℓ,i∑xj(t−τℓ)∈B~2ℓ(t−τℓ)ε~τ2ℓ,j,(41)\widetilde{\mathcal{M}}_\ell(t)\mathrel{\stackrel{\mathrm{stat}}{\sim}}\frac{N_{2\ell}}{N_\ell}\frac{\sum_{\mathbf{x}_i(t)\in\mathcal{B}_\ell(t)}\mathcal{N}_{\ell,i}\widetilde{\varepsilon}_{\tau_{2\ell},i}}{\sum_{\mathbf{x}_j(t-\tau_\ell)\in\widetilde{\mathcal{B}}_{2\ell}(t-\tau_\ell)}\widetilde{\varepsilon}_{\tau_{2\ell},j}}, \tag*{(41)}

where the common time argument t−τℓt-\tau_\ell has been omitted on the right-hand side. Thus, M~ℓ\widetilde{\mathcal{M}}_\ell is not identified with a single Lagrangian multiplier. Rather, it involves the temporal refinements collectively carried by the fluid elements with their relative weights set by the parent-scale dissipation. Material deformation and mixing determine which histories contribute to the present spatial region and how they are assembled.

The final step is to relate this dynamical refinement to the Eulerian multiplier. Both quantities contain the same present child but refer to different parents:

Mℓ(t)=M~ℓ(t)ε2ℓ(t−τℓ)ε2ℓ(t).(42)\mathcal{M}_\ell(t)=\widetilde{\mathcal{M}}_\ell(t)\frac{\varepsilon_{2\ell}(t-\tau_\ell)}{\varepsilon_{2\ell}(t)}. \tag*{(42)}

Here ε2ℓ(t)\varepsilon_{2\ell}(t) denotes the dissipation in the instantaneous concentric parent B2ℓ(t)\mathcal{B}_{2\ell}(t), whereas ε2ℓ(t−τℓ)\varepsilon_{2\ell}(t-\tau_\ell) refers to the dynamically selected parent defined above. The second factor measures the difference between the dynamically selected material parent and the instantaneous Eulerian parent. It therefore contains the effects of transport, deformation, mixing, and temporal evolution of the dissipation field. Because this factor fluctuates and may correlate with Mℓ\mathcal{M}_\ell, the dynamical and Eulerian multipliers need not coincide. Equation (42) consequently provides a statistical decomposition of the Eulerian refinement into a material contribution and a parent-evolution contribution. It does not require, and does not establish, a one-to-one correspondence between Eulerian and Lagrangian multipliers.

The decomposition (42) turns the material-inheritance picture into a falsifiable statement. If the scale dependence of Mℓ\mathcal{M}_\ell arises from the assembly of material histories, the dynamical multiplier M~ℓ\widetilde{\mathcal{M}}_\ell should be approximately scale-stationary, with statistics close to those of the temporal multipliers Nτℓ\mathcal{N}_{\tau_\ell}, while the logarithmic drift of Figure 9(a) should be carried by the parent-evolution factor ε2ℓ(t−τℓ)/ε2ℓ(t)\varepsilon_{2\ell}(t-\tau_\ell)/\varepsilon_{2\ell}(t) and its correlation with M~ℓ\widetilde{\mathcal{M}}_\ell. Conversely, a dynamical multiplier as scale-dependent as Mℓ\mathcal{M}_\ell would show that the Eulerian refinement is not explained by material transport. Measuring M~ℓ\widetilde{\mathcal{M}}_\ell requires following the pre-images of coarse-graining regions with tracer densities beyond those used here.

Conclusions

These results show that the intermittency of turbulent dissipation has distinct statistical organizations under spatial and temporal coarse-graining. A central organizing idea in turbulence theory is that intermittency arises from a cascade process that can be expressed through multifractal statistics of energy dissipation. Our results show that this picture cannot be represented by a single, frame-independent cascade structure. Lagrangian time averages display an approximately scale-stationary temporal cascade, whereas Eulerian coarse-grained dissipation exhibits a distinct spatial organization with scale-dependent multipliers and a squared-logarithmic cascade depth. The commonly assumed equivalence between Eulerian and Lagrangian dissipation statistics, mediated by bridge relations, therefore cannot be interpreted as a trajectory-wise identification. Rather, the results indicate that the temporal organization revealed along Lagrangian trajectories provides an important description of dissipation intermittency, while its Eulerian manifestation reflects a different spatial organization.

To reach this conclusion, we performed a systematic comparison of Eulerian and Lagrangian statistics of energy dissipation in homogeneous isotropic turbulence. By combining analyses of single-point distributions, correlation functions, coarse-grained fields, large-deviation properties, and multiplier statistics, we showed that the statistical organization of dissipation depends fundamentally on the manner in which it is sampled and cannot be captured by a single universal cascade description.

In the Eulerian frame, the spatial correlations and coarse-grained properties of the instantaneous dissipation field do not support the existence of a conventional scale-stationary multiplicative cascade. In particular, the absence of logarithmic spatial correlations, the nonlinear dependence of cumulants on scale, and the pronounced scale dependence of Eulerian multipliers rule out a conventional log-divisible cascade interpretation of the spatially coarse-grained Eulerian dissipation in the inertial range. At the same time, this measure still shows a large deviation form with squared-logarithmic depth. Furthermore, the Eulerian multipliers possess a nontrivial multifractal organization. This surprising, yet self-consistent picture, forces us to reimagine the physically appealing, but certainly incomplete, cascade phenomenology and go beyond the scale-stationary processes which underlies much of conventional turbulence theory.

In contrast, dissipation sampled along Lagrangian trajectories displays the defining signatures of a multiplicative cascade in time. These include logarithmic temporal correlations, linear scaling of the mean and variance of the logarithm of the coarse-grained dissipation with the logarithm of the time scale, and approximately scale-invariant statistics of Lagrangian multipliers. The temporally coarse-grained Lagrangian dissipation is not log-normal but still follows a well-defined large-deviation principle. Thus the Lagrangian dissipation is consistent with a log-divisible stochastic process organized by a temporal multiplicative cascade.

Building on these findings, we revisited classical bridge relations that aim to connect Eulerian and Lagrangian dissipation statistics through dimensional arguments. We showed that such relations implicitly rely on a mean-field assumption, namely that dissipation fluctuations experienced by different Lagrangian trajectories within the same spatial region are effectively uniform. Our results demonstrate that this assumption is violated for dissipation, leading to a breakdown of the classical bridge relation as a trajectory-wise identification. This breakdown does not imply an absence of Eulerian–Lagrangian coupling: the two descriptions remain appreciably correlated, but their fluctuations cannot be represented by a single trajectory or by a single choice of turnover-time prefactor.

We therefore formulated a material-inheritance framework in which the present Eulerian coarse-graining region is related to the material histories of the fluid elements contributing to it. The corresponding dynamical refinement compares the present child region with a dynamically selected material parent and can be expressed statistically in terms of the temporal refinements carried by the trajectories that contribute to the child. The Eulerian multiplier then contains, in addition, a factor describing the evolution between this dynamically selected parent and the instantaneous Eulerian parent. This construction provides a statistical framework for connecting spatial and temporal descriptions without requiring a one-to-one mapping between Eulerian and Lagrangian multipliers.

More broadly, these results indicate that the distinction between Eulerian and Lagrangian intermittency is not simply a distinction between two measurements of the same cascade. The Lagrangian description exposes the scale organization of dissipation histories along material trajectories, while the Eulerian description captures the organization that results when these histories are spatially assembled, transported, and deformed by the flow. The present analysis does not establish a direct causal transformation from the temporal cascade into the observed Eulerian multiplier statistics. Instead, it identifies material inheritance as a possible framework for relating the two descriptions and shows that any such connection must be consistent with the distinct multiplier statistics and cascade depths observed in the two sampling procedures.

Future work may test this framework more directly by following material pre-images of Eulerian coarse-graining regions and examining how their dissipation histories are transported, deformed, and assembled across scales. While it may be tempting to connect our results on dissipation to the velocity increment statistics directly — since εℓ\varepsilon_\ell is treated as a proxy for the energy flux ∼δvℓ3/ℓ\sim\delta v_\ell^3/\ell — the absence of a clear multifractal statistics in the former does not preclude multifractality in the latter. In any case, the present results suggest that models of intermittency may benefit from distinguishing the temporal organization of fluctuation histories from the spatial statistics that emerge when those histories are sampled and assembled by the flow. A particularly well-adapted framework to factor this in could be quasi-Lagrangian approaches and models [46, 47, 48] and its generalisations [49] including those involving dynamical time-rescaling as suggested by hidden symmetry arguments [50].

Acknowledgments

This research has been partly supported by the Agence Nationale de la Recherche (Grant No. ANR-25-CE30-4603-01). RM acknowledges the Infosys-TIFR Leading Edge Travel Grant 2025 which facilitated part of this research. SM acknowledges the IITK Initiation Grant project IITK/ME/2024316 and support from the Govt. of India Grant ANRF/ECRG/2024/002467/ENS. JB and SSR acknowledges the Indo–French Centre for the Promotion of Advanced Scientific Research (IFC-PAR/CEFIPRA, project no. 6704-1) for support. The simulations were performed on the ICTS clusters Mario, Tetris, and Contra, using OPAL infrastructure from Université Côte d’Azur, and HPC resources from GENCI-TGCC (Grant t2016-2as027). RM and SSR acknowledge the support of the Department of Atomic Energy, Government of India, under project no. RTI4019.

References

References

  1. [1]U. Frisch, Turbulence: The Legacy of A. N. Kolmogorov (Cambridge University Press, 1995).
  2. [2]P. K. Yeung, X. M. Zhai, and K. R. Sreenivasan, Extreme events in computational turbulence, Proceedings of the National Academy of Sciences 112, 12633 (2015).DOI
  3. [3]D. Buaria, A. Pumir, E. Bodenschatz, and P.-K. Yeung, Extreme velocity gradients in turbulent flows, New J. Phys. 21, 043004 (2019).
  4. [4]A. Alexakis and L. Biferale, Cascades and transitions in turbulent flows, Phys. Rep. 767-769, 1 (2018), cascades and transitions in turbulent flows.
  5. [5]A. N. Kolmogorov, A refinement of previous hypotheses concerning the local structure of turbulence in a viscous incompressible fluid at high Reynolds number, J. Fluid Mech. 13, 82 (1962).DOI
  6. [6]A. M. Obukhov, Some specific features of atmospheric turbulence, J. Geophys. Res. 67, 3011 (1962).
  7. [7]C. Meneveau and K. R. Sreenivasan, The multifractal spectrum of the dissipation field in turbulent flows, Nucl. Phys. B 2, 49 (1987).
  8. [8]C. Meneveau and K. R. Sreenivasan, The multifractal nature of turbulent energy dissipation, J. Fluid Mech. 224, 429 (1991).
  9. [9]K. P. Iyer, K. R. Sreenivasan, and P. K. Yeung, Refined similarity hypothesis using 3D local averages, Phys. Rev. E 92, 063024 (2015).DOI
  10. [10]J. Schumacher, K. R. Sreenivasan, and V. Yakhot, Asymptotic exponents from low-Reynolds-number flows, New J. Phys. 9, 89 (2007).DOI
  11. [11]D. Donzis, P. Yeung, and K. Sreenivasan, Dissipation and enstrophy in isotropic turbulence: resolution effects and scaling in direct numerical simulations, Phys. Fluids 20, 10.1063/1.2907227 (2008).DOI
  12. [12]A. Dey, R. Mukherjee, A. Banerjee, and S. S. Ray, Parity-dependent scaling of velocity-gradient correlations in turbulence (2026), arXiv:2605.19921 [physics.flu-dyn].arxiv.org/abs/2605.19921
  13. [13]K. P. Iyer, J. Schumacher, K. R. Sreenivasan, and P. K. Yeung, Scaling of locally averaged energy dissipation and enstrophy density in isotropic turbulence, New J. Phys. 21, 033016 (2019).DOI
  14. [14]D. Buaria and K. R. Sreenivasan, Intermittency of turbulent velocity and scalar fields using three-dimensional local averaging, Phys. Rev. Fluids 7, L072601 (2022).DOI
  15. [15]A. B. Chhabra and K. R. Sreenivasan, Scale-invariant multiplier distributions in turbulence, Phys. Rev. Lett. 68, 2762 (1992).
  16. [16]K. R. Sreenivasan and G. Stolovitzky, Turbulent cascades, J. Stat. Phys. 78, 311 (1995).
  17. [17]A. Arnéodo, E. Bacry, S. Manneville, and J. Muzy, Analysis of random cascades using phase-scale correlation functions, Phys. Rev. Lett. 80, 708 (1998).
  18. [18][18*] W. Ruffenach and L. Chevillard, The spatio-temporal statistical structure of the turbulent dissipation field and its stochastic representation as a Gaussian multiplicative chaos, Comptes Rendus Physique 27, 275 (2026).
  19. [19]F. Toschi and E. Bodenschatz, Lagrangian properties of particles in turbulence, Annu. Rev. Fluid Mech. 41, 375 (2009).
  20. [20]B. L. Sawford and P.-K. Yeung, Kolmogorov similarity scaling for one-particle lagrangian statistics, Phys. Fluids 23, 10.1063/1.3643852 (2011).DOI
  21. [21]L. Biferale, G. Boffetta, A. Celani, B. J. Devenish, A. Lanotte, and F. Toschi, Lagrangian statistics of particle pairs in homogeneous isotropic turbulence, Phys. Fluids 17, 115101 (2005).DOI
  22. [22]B. L. Sawford and P. K. Yeung, Direct numerical simulation studies of Lagrangian intermittency in turbulence, Phys. Fluids 27, 065109 (2015).DOI
  23. [23]A. Arnéodo, R. Benzi, J. Berg, L. Biferale, E. Bodenschatz, A. Busse, E. Calzavarini, B. Castaing, M. Cencini, L. Chevillard, et al., Universal intermittent properties of particle trajectories in highly turbulent flows, Phys. Rev. Lett. 100, 254504 (2008).
  24. [24]L. Bentkamp, C. C. Lalescu, and M. Wilczek, Persistent accelerations disentangle lagrangian turbulence, Nat. Comm. 10, 3550 (2019).DOI
  25. [25]Y. Huang and F. G. Schmitt, Lagrangian cascade in three-dimensional homogeneous and isotropic turbulence, J. Fluid Mech. 741, R2 (2014).
  26. [26]M. Borgas, The multifractal lagrangian nature of turbulence, Phil. Trans. R. Soc. London A 342, 379 (1993).
  27. [27]F. G. Schmitt, Linking eulerian and lagrangian structure functions’ scaling exponents in turbulence, Physica A 368, 377 (2006).
  28. [28]L. Chevillard, B. Castaing, A. Arneodo, E. Lévêque, J.-F. Finton, and S. G. Roux, A phenomenological theory of Eulerian and Lagrangian velocity fluctuations in turbulent flows, C.R. Physique 13, 899 (2012).DOI
  29. [29]C. C. Lalescu and M. Wilczek, How tracer particles sample the complexity of turbulence, New J. Phys. 20, 013001 (2018).DOI
  30. [30]S. Mukherjee, S. D. Murugan, R. Mukherjee, and S. S. Ray, Turbulent flows are not uniformly multifractal, Phys. Rev. Lett. 132, 184002 (2024).
  31. [31]J. C. Vassilicos, Beyond scale-by-scale equilibrium, Atmosphere 14, 736 (2023).
  32. [32]R. Mukherjee and S. Mukherjee, Disentangling intermittent flow structure contributions to anomalous scaling and multifractality in turbulence (2026), arXiv:2607.26896 [physics.flu-dyn].arxiv.org/abs/2607.26896
  33. [33]A. Vulpiani, F. Cecconi, M. Cencini, A. Puglisi, and D. Vergni, Large deviations in physics, The Legacy of the Law of Large Numbers (Berlin: Springer) 10, 978 (2014).
  34. [34]J. Bertoin, Lévy processes, Vol. 199 (Cambridge university press Cambridge, UK, 1996).
  35. [35]Z.-S. She and E. Leveque, Universal scaling laws in fully developed turbulence, Phys. Rev. Lett. 72, 336 (1994).
  36. [36]B. Dubrulle, Intermittency in fully developed turbulence: Log-poisson statistics and generalized scale covariance, Phys. Rev. Lett. 73, 959 (1994).
  37. [37]H. Touchette, The large deviation approach to statistical mechanics, Phys. Rep. 478, 1 (2009).
  38. [38]J. O’Neil and C. Meneveau, Spatial correlations in turbulence: Predictions from the multifractal formalism and comparison with experiments, Phys. Fluids A 5, 158 (1993).DOI
  39. [39]B. Castaing and B. Dubrulle, Fully developed turbulence: a unifying point of view, J. Phys. II (France) 5, 895 (1995).DOI
  40. [40]C. Meneveau, Lagrangian dynamics and models of the velocity gradient tensor in turbulent flows, Annu. Rev. Fluid Mech. 43, 219 (2011).
  41. [41]H. Yu and C. Meneveau, Lagrangian refined kolmogorov similarity hypothesis for gradient time evolution and correlation in turbulent flows, Phys. Rev. Lett. 104, 084502 (2010).
  42. [42]P. L. Johnson and M. Wilczek, Multiscale velocity gradients in turbulence, Annu. Rev. Fluid Mech. 56, 463 (2024).
  43. [43]F. G. Schmitt, A causal multifractal stochastic equation and its statistical properties, Euro. Phys. J. B 34, 85 (2003).
  44. [44]L. Chevillard and C. Meneveau, Lagrangian dynamics and statistical geometric structure of turbulence, Phys. Rev. Lett. 97, 174501 (2006).
  45. [45]R. M. Pereira, L. Moriconi, and L. Chevillard, A multifractal model for the velocity gradient dynamics in turbulent flows, J. Fluid Mech. 839, 430 (2018).DOI
  46. [46]V. Belinicher and V. L’vov, A scale-invariant theory of fully developed hydrodynamic turbulence, Sov. Phys. JETP 66, 303 (1987).
  47. [47]G. Boffetta, F. De Lillo, and S. Musacchio, Lagrangian statistics and temporal intermittency in a shell model of turbulence, Phys. Rev. E 66, 066307 (2002).
  48. [48]D. Mitra and R. Pandit, Varieties of dynamic multi-scaling in fluid turbulence, Phys. Rev. Lett. 93, 024501 (2004).
  49. [49]S. S. Ray, D. Mitra, P. Perlekar, and R. Pandit, Dynamic multiscaling in two-dimensional fluid turbulence, Phys. Rev. Lett. 107, 184503 (2011).
  50. [50]A. A. Maillybaev and S. Thalabard, Hidden scale invariance in Navier–Stokes intermittency, Phil. Trans. R. Soc. London A 380, 1 (2022).DOI

Paper details

vlnc:W_vs5dm9uv
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