(E) Locality of Energy Transfer

See T & L, Section 8.2; U. Frisch, Section 7.3

The Essence of the Matter

We have seen that energy is transferred from scales > ` to scales < ` by the deformation work

⇧` (x) = S̄` (x) : ⌧ ` (x)

As we now discuss, this transfer is scale-local under conditions that are realistic for turbulent

flow. Suppose that the velocity v is Hölder continuous at point x with exponent 0 < h < 1, i.e.

v(r; x) = O(|r|h ).

For example, in K41 theory, h = 1/3 at every point x in the flow. Then, as follows from our

earlier discussion,

S̄` (x) = O(`h 1 )?

and

⌧ ` (x) = O(`2h )??

so that

⇧` = O(`3h 1)

at the point x. We have proved these only as upper bounds, but let us assume, for the sake of

argument, that = O(`↵ ) here means in fact ⇠ (const.)`↵ . (We’ll return to this issue later!)

Where does most of the strain come from? We can consider a larger length-scale , with

`⌧ ⌧ L and write

S̄` (x) = S̄ (x) + S[`, ] (x)

where

S̄ = strain from scales >

S[`, ] = band-pass filtered strain from scales between ` and

But the previous estimates apply to S̄ , so that

R R
? requires dd ⇢ |⇢|h |rG(⇢)| < +1. ?? requires dd ⇢ |⇢|2h G(⇢) < +1

40
 S̄ (x) = O( h 1 ).

In that case,
|S̄ (x)| h 1
|S̄` (x)|
= O( `h 1 ) = O(( ` )1 h)

which is ⌧ 1 whenever h < 1. Thus, we see that most of the strain S̄` (x) comes from scales

near ` and very little comes from scales `, whenever h < 1.

Now, what about the stress from small-scales? We can likewise consider a smaller length ,

with ⌘ ⌧ ⌧ ` and write

⌧ ` (x) = ⌧ (x) + ⌧ [ ,`] (x)

where

⌧ (x) = ⌧ (v, v) = stress from scales <

⌧[ ,`] (x) = stress from scales between and `

The above equation defines ⌧ [ ,`] . (NOTE: A better way to do this is by using the so-called

Germano identity. This will be explored in the homework!) But, again, the previous estimates

apply to ⌧ , so that

⌧ (x) = O( 2h ).

In that case
|⌧ (x)| 2h
|⌧ ` (x)| = O( `2h ) = O(( ` )2h )

which is ⌧ 1 whenever h > 0. Thus, we see that most of the stress ⌧ ` (x) comes from scales near

` and very little comes from scales ⌧ `, whenever h > 0.

The conclusion is that the energy transfer from length-scales > ` to length-scales < ` is domi-

nated by interactions of modes at scales ⇡ `. In fact, two modes with length scales > ` can, by

quadratic nonlinearities, interact only with modes down to length scales > `/2. This is a basic

result of Fourier analysis, since, if two modes have only wavenumbers k, k0 such that
2⇡
|k|, |k0 | < `

then their product can contain wavenumbers k00 with at most

41
 |k00 | = |k + k0 |  |k| + |k0 |
2⇡ 2⇡ 4⇡ 2⇡
 + = =
` ` ` `/2

The picture that emerges is of energy transfer across the length-scale ` by interaction of modes

with scale near `, to a length-scale & 2` . The energy at this scale is then, in turn, transferred

by similar scale-local interactions to a length-scale & `

4, and so on. This stepwise process is

called a local cascade (in this case, of energy). Here it is crucial that only modes with scale ⇡ `

participate in the transfer of excitation across the length-scale `.

If the individual steps in the cascade process are also chaotic nonlinear process, then it is rea-

sonable to expect that the small-scales will “forget” about the detailed geometry and statistics

of the large-scale flow modes. In particular, there is no “direct communication” with large-scale

modes in the process which creates and maintains the small-scale motions.

These considerations motivate the idea of universality of the small-scales, that is, the notion

that the statistics of the small-scale modes shall be the same for all flows and independent of the

details of the large-scale geometry, generation mechanisms, etc. In particular, the symmetries

of the dynamics — space homogeneity, temporal invariance, rotational isotropy, scale invari-

ance,etc. — should be restored on a statistical level.

This university — and thus also scale-locality — is quite important for the physical foundation

of large-eddy simulation (LES) modelling of turbulent flow. It raises the hope that generally

applicable (universal) models of small-scale stress ⌧ ` may be possible!

42
 A More Precise Statement of Scale Locality

The filtered velocity gradient

D̄` (v) = rv̄`

is a linear functional D` of the velocity field v. Likewise,

⌧ ` (v, v) = (vv)` v̄` v̄`

is a quadratic functional of v. The concept of infrared (IR) or large-scale locality is that re-

placing any v with v̄ will lead to a much smaller contribution, for `. Conversely, this

means that most of the contribution will be obtained by replacing v with v0 , for a sufficiently

large `. The concept of ultraviolet (UV) or small-scale locality is that replacing any v

with v0 will lead to a much smaller contribution, for ⌧ `. Again, this means that most of

the contribution will be obtained by replacing v with v̄ , for a sufficiently small ⌧ `.

These results hold if the velocity field v is Hölder continuous at space point x with exponent

0 < h < 1, as the following precise estimates show:

IR Locality: For > `,

|D̄` (v̄ | = |D̄` (v)| · O(( ` )1 h ), or D̄` (v0 ) = D̄` (v) · [1 + O(( ` )1 h )]

|⌧ ` (v̄ , v)| = |⌧ ` (v, v)| · O(( ` )1 h ), or ⌧ ` (v0 , v) = ⌧ ` (v, v) · [1 + O(( ` )1 h )].

Note, BTW, that replacing both v0 s with v̄`0 s in ⌧ ` leads to an even smaller result:

|⌧ ` (v̄ , v̄ )| = |⌧ ` (v, v)| · O(( ` )2(1 h) )

UV Locality: For < `,

|D̄` (v0 )| = |D̄` (v)| · O(( ` )h ), or D̄` (v̄ ) = D̄` (v) · [1 + O(( ` )h )]

|⌧ ` (v0 , v)| = |⌧ ` (v, v)| · O(( ` )h ), or ⌧ ` (v̄ , v) = ⌧ ` (v, v) · [1 + O(( ` )h )].

and, of course,

|⌧ ` (v0 , v0 )| = |⌧ ` (v, v)| · O(( )2h ).

`

For a complete discussion and proof, see

G. L. Eyink, “Locality of turbulent cascades,” Physica D, 207, 91-116 (2005)

However, the basic point is that D̄` (v), ⌧ ` (v, v) are — as we have seen long ago — given by

43
integrals of velocity increments v(r) for |r|  ` (essentially). Furthermore, it is a consequence

of the following lemma that the velocity increments themselves are scale-local:

Lemma: If v is Hölder continuous at point x with exponent 0 < h < 1, then for `

(1) v̄ (`; x) = O(` h 1 ), (2) v0 (`; x) = v(`; x) + O(` h 1) = v(`; x)[1 + O(( ` )1 h )]

and for `

(3) v0 (`; x) = O( h ), (4) v̄ (`; x) = v(`; x) + O( h ) = v(`; x)[1 + O(( ` )h )],

if v is Hölder continuous with exponent 0 < h < 1 in a neighbourhood of the point x.

Proof: Note that (1)&(2) are, in fact, equivalent, since v0 (`; x) = v(`; x) v̄ (`; x). Like-

wise, (3)&(4) are also equivalent.

The proof of (3) is simple, because

v0 (`; x) = v0 (x + `) v0 (x).

However, if ` is close enough to zero, then both x + ` and x are in the neighbourhood with

exponent h and thus

R
v0 (x) = dr G (r) v(r; x) = O( h )

and likewise for v0 (x + `).

The proof of (1) is just a bit more complex:

Z
v̄ (`; x) = dr G (r) v(`; x + r)
| {z }
[v(x+r+`) v(x+r)]
Z
= dr[G (r `) G (r)]v(x + r)
Z
= dr[G (r `) G (r)] [v(x + r) v(x)]
| {z }
v(r;x)
Z 1 Z
1
= d✓ dr ` · (rG) (r ✓`) v(r; x)
0
h 1
= O(` ) QED!

Remark: The idea behind the last estimate is very simple: it is just a consequence of the fact

that v̄ is smooth. Thus,

v̄ (`) ⇠
= ` · rv̄ = O(` h 1)

44
The first estimate is even simpler: v0 (`) is small, because the small-scale fluctuation field

v0 = O( h ) is small! This is the essence of scale-locality. Note that these facts about velocity-

increments were stated by Kolmogorov in the first of his 1941 papers on turbulence and also

by Onsager in his 1945 letter to Lin and von Kármán.

Some important comments:

? There are some shortcomings in the above locality results, which we shall discuss later in

the chapter on intermittency & scaling. In particular, it is not reasonable to assume pointwise

scaling v(`; x) ⇠ (const)`h and thus some of the estimates above are based on unrealistic

assumptions. E.g. in the lemma above, (2a) & (4a) are OK, but (2b) & (4b) are not. For most

purposes, (2a),(4a) suffice.

? We have been deriving upper bounds on nonlocal contributions. However, we have used only

very simple estimates and the true contributions could be considerably smaller. For example,

our estimate on the small-scale stress contribution to energy flux is

⇧stress<
` = S̄` : ⌧ ` (v0 , v0 )

= ⇧` · O(( )2h )
`

1
Thus, the relative contribution of ⇧stress<
` is at most O(( ` )2/3 ) for the K41 value h = 3.

However, this argument assumes that S̄` and ⌧ <

` = ⌧ ` (v0 , v0 ) are well-correlated, when, in

fact, they are not. Notice that S̄` is associated to scales > `, while ⌧ <
` is associated to scales

< . In integrals over space ( or over ensemble), there shall be large cancellations that are not

considered in our rather crude estimates. Thus, the averages will be smaller by a factor of the

correlation coefficient ⇢(S̄` , ⌧ <

` ), which can be estimated as in T&L, Section 8.2, p.261:
t
⇢(S̄` , ⌧ <
` )⇠ t` = O(( ` )1 h)

where t` = `/ v(`) is the turnover time at length `. Thus, in K41 theory one can argue that

h⇧stress<
` i = h⇧` i · O(( ` )2/3 · ( ` )2/3 ) = h⇧` i · O(( ` )4/3 )

which is much smaller.

45
 On the other hand, even in this strengthened form, the nonlocal contributions decay only

as a small power-law of the scale-ratios /` and `/ . This fact led Kraichnan to remark, for

the similar wave-number ratio, that

“However, the dependence on k/k 0 is not particularly strong, and thus the cascade

is rather di↵use.” –R. H. Kraichnan, J. Fluid Mech. 5 497 (1959)

Similarly, T&L remarked

“..we should not expect too much from the cascade model. After all, it is a very

leaky cascade if half the water crossing a given level comes directly from all other

pools uphill.” – Tennekes & Lumley (1972), p. 261.

As these quotations reflect, a very large number of cascade steps is required before local in-

teractions really dominate. In practice, a substantial fraction of the the energy transfer often

comes from very non-local interactions.

Local Time-Scales

Now is a good time to discuss in more detail the time-scales t` at length-scale `, or the

local eddy-turnover time

`
t` = v(`) .

We have already argued in several places earlier that this is the time-scale for Lagrangian evolution

at length-scale `, i.e. that D̄`,t ⇠ 1/t` . It is usually regarded as the time-scale for any “eddy”

of size ` to change by O(1) or to “turnover”. It is also the time-scale set by the strain-rate at

length-scale `
v(`)
S̄` = O( ` ) = O( t1` )

Thus, it is the time-scale in which structures of size ` are deformed by the fluid shears.

If the velocity field is Hölder continuous with exponent h, 0 < h < 1, then

v(`) ⇠ urms ( L` )h

and

46
 t` ⇠ tL ( L` )1 h

with tL = L/urms . We see that, as long as h < 1, then

t` ! 0 for ` ⌧ L.

Thus, the small-scale “eddies” evolve faster as ` decreases! For example, in K41 theory,

t` ⇠ tL ( L` )2/3 ⇠ h"i 1/3 `2/3

The energy cascade is thus an accelerated cascade, with each successive step sped up. Taking

`n = 2 nL for the length-scale of the nth step, then

tn = tL ( `Ln )1 h = tL 2 (1 h)n .

This has the remarkable property that

1
X 1
X
(1 h)n
T = tn = tL 2 < +1
n=0 n=0

for any 0 < h < 1. Thus, the time T that it takes to make an infinite number of cascade steps

is finite! This observation, first made by [Link] (1945,1949), has important implications.

First, t` is the time that it takes to transfer an O(1) amount of energy across the length-scale

`. Indeed,

⇧` = O( t1` · v 2 (`)).

Thus, the acceleration of the cascade is important to explain the observed dissipation of energy.

In reality, there are only a finite number N of cascade steps to reach the dissipation range. For

example, in K41 theory

3
N = log2 (L/⌘) = 4 log2 (Re).

The time for an O(1) amount of energy to reach the dissipation range is then
N
X
TN = tn
n=0

and Tn ! T < +1 in the limit Re ! 1. Thus, turbulence can dissipate an O(1) amount of

energy in a time which is independent of Re, for Re 1!

47
Another important consequence of “acceleration” of the cascade, is that it makes more plausible

the hypothesis of universality. As the small-scales are evolving so quickly, the very large scales —

which are non-universal — appear essentially “frozen”. Furthermore, because of scale locality,

there is no direct contact or communication between the largest scales and the small scales.

Excitation is transferred only by a chain of chaotic intermediate scales. Thus, the small scales

have plenty of time to approach an invariant distribution for fixed input (energy flux) from large-

scales. Except for the conserved fluxes, which may vary slowly in time, all other information of

the large scales is lost in the course of the cascade.

The chaotic dynamics of the turbulent flow at length-scale ` can be measured by the

Lyapunov exponent `, which has units of inverse time and which gives the exponential rate of

perturbation growth at that scale. Put another way, 1/ ` is the “e-folding time” of perturbations

at length-scale ` and may be expected to be approximately the same as the eddy-turnover time

t` inside the inertial range L ` ⌘. At length-scales ` ⌧ ⌘, the Kolmogorov length-scale,

viscosity damps out perturbations so that one expects a negative exponent ` ⇡ ⌫/`2 and

non-chaotic dynamics. These expectations are hard to prove mathematically for Navier-Stokes,

although there has been some important progress for stochastically forced models:

J. Bedrossian and S. Punshon-Smith. “Chaos in stochastic 2D Galerkin-Navier-

Stokes,” Comm. Math. Phys. 405 107 (2024).

For some toy “shell models” of the Navier-Stokes equation, these expected behaviors of Lya-

punov exponents have been verified by numerical simulations:

M. Yamada and Y. Saiki, “Chaotic properties of a fully developed model turbu-

lence,” Nonlin. Proc. Geophys. 14 631-640 (2007).

Motivated by such considerations, the mathematical physicist David Ruelle in the following

paper,

48
 D. Ruelle, “Microscopic fluctuations and turbulence,” Phys. Lett. 72A 81–82

(1979),

asked the interesting question: how long will it take for thermal fluctuations at length-scale

` to be amplified to a macroscopic size? The largest positive Lyapunov should occur for ` ⇠

⌘, the Kolmogorov length, with magnitude 1/t⌘ for t⌘ = ⌘ 2/3 /"1/3 = (⌫/")1/2 the so-called

Kolmogorov time. Thus, Ruelle asked what time t will it take for an initial thermal velocity

fluctuation v⌘0 at scale ⌘ to grow exponentially according to the equation

et/t⌘ v⌘0 ' v⌘

to a magnitude of order the Kolmogorov velocity v⌘ = ("⌘)1/3 = ("⌫)1/4 , which characterizes

the magnitude of the turbulent velocity fluctuations at length-scale ` ' ⌘. To estimate the size

of the initial thermal velocity fluctuation v`0 at length-scale `, Ruelle appealed to the Central

Limit Theorem, which gives

vth
v`0 ' p ,
n`d
with vth = (kB T /m)1/2 the thermal velocity (which is of order the speed of sound cs ) and n the

particle density, so that N = n`d represents the total number of particles in the region of radius

`. Here we recall that the N molecules of the fluid in a small spatial region with temperature

T have their velocities distributed according to independent Maxwellian’s:

N ✓
Y ◆d/2 ✓ ◆
m m|vn u|2
P (v1 , ..., vN ) = exp
2⇡kB T 2kB T
n=1

If v(`) is the characteristic turbulent velocity at length-scale `, then notice that

✓ ◆1/2
v`0 kB T 1/2
' := ✓`
v(`) ⇢( v(`))2 `d

and the term on the right is the small parameter which appears as the amplitude of the thermal

noise term in the fluctuating Navier-Stokes equation at length-scale `. As expected, the thermal

fluctuations are negligible for large ` but grow as ` decreases.

49
 Putting together these various estimates, Ruelle found that for ` ' ⌘, the time t for the

thermal perturbation v⌘0 to grow to order the Kolmogorov velocity v⌘ is, for d = 3,

kB T
t ' t⌘ ln(1/✓⌘ ), ✓⌘ = .
⇢v⌘2 ⌘ 3

Although one expects a large value of 1/✓⌘ 1, this quantity appears only inside a logarithm

so that the growth time t is only a modest multiple of t⌘ ! For example, consider some values

of physical constants typical of the turbulent atmospheric boundary layer

T = 300 K, ⇢ = 1.2 gm/cm3 , ⌫ = 0.15 cm2 /sec, " = 400 cm3 /sec3 .

Then

⌘ = 0.54 mm, t⌘ = 19.4 msec, v⌘ = 2.78 cm/sec

. 16
and with kB = 1.38 ⇥ erg/K,

8 .
✓⌘ = 2.83 ⇥ 10 , t = (14.4)t⌘ .

More generally, one finds for typical terrestrial turbulent flows that the growth time for thermal

perturbations at the Kolmogorov scale ⌘ to reach macroscopic magnitude is only t ' 5 15t⌘ ,

i.e. just a few Kolmogorov times.

Once the perturbations have grown to macroscopic size at scale ` ' ⌘, they will infect

the dynamics at the next larger scale ` ' 2⌘ and produce errors of the size of the turbulent

fluctuations at that twice larger scale, then ` ' 4⌘, and so forth. Thus, one can expect an

inverse cascade of errors. The expected time to double the length-scale of the error from `n 1

to `n is just the turnover-time at that scale, or tn = t`n . The total time Tn that it takes for a

thermal perturbation to grow from the Kolmogorov scale ⌘ to length-scale `n is just

N
X
Tn = tm ' Atn ,
m=n

3
with A a constant of order unity at high Re with N = 4 log2 (Re) 1. This basic picture and

the above formula for the growth time Tn were obtained in a fundamental paper:

50
 E. N. Lorenz, “The predictability of a flow which possesses many scales of motion,”’

Tellus 21, 289-307 (1969)

which we shall discuss in more depth later. These predictions were verified using a spectral

closure (the “test-field model”) in the following paper:

C. Leith and R. H. Kraichnan, “Predictability of turbulent flows,” J. Atmos. Sci.

29 1041–1058 (1972)

for both 2D and 3D space. Note that there is considerable di↵erence in the physics of incom-

pressible fluid turbulence in 2D and 3D, with energy cascading to large scales in 2D! However,

the above paper found the above formula for growth time of errors to hold in the energy cascade
. .
range in both cases, with A = 10 for 3D and A = 2.5 for 2D. The papers

G. Bo↵etta and S. Musacchio, “Predictability of the inverse energy cascade in 2D

turbulence,” Phys. Fluids 13 1060–1062 (2001)

G. Bo↵etta and S. Musacchio, “Chaos and predictability of homogeneous-isotropic

turbulence,”’ PRL 119, 054102 (2017)

verified the closure predictions for 2D and 3D turbulence via numerical simulation. See also

D. Bandak et al., “Spontaneous Stochasticity Amplifies Even Thermal Noise to the

Largest Scales of Turbulence in a Few Eddy Turnover Times,” PRL 132, 104002

(2024)

for very high Reynolds number simulations in a shell model.

All of the above considerations imply, remarkably, that two di↵erent flows with precisely

the same macroscopic initial velocity but with di↵erent realizations of thermal noise will lead

to completely di↵erent velocity fields at all length-scales in the inertial-range within about one

large-eddy turnover time! This suggests an intrinsic unpredictability of turbulent flows, with

radically di↵erent solutions for the same macroscopic initial data.

51
 Thermal Energy Spectrum
p
This is a suitable point to remark on the spectral interpretation of the result v`0 = vth / n`3

for the thermal velocity fluctuation in space dimension d = 3. This formula corresponds to an

energy spectrum of velocity fluctuations

3 kB T 4⇡k 2
E(k) = .
2 ⇢ (2⇡)3
Apart from the specific numerical prefactor, the result is obviously the correct one to yield
Z 1 Z 1/`
kB T
h|v0` |2 i = |Ĝ` (k)|2 E(k) dk ' (const.) E(k) dk ' (const.) 3
0 0 ⇢`
with v0` := v` hv` i. The specific prefactor is found by using the equipartition of modal

energies 12 ⇢|b
v(k)|2 with thermal equilibrium value 12 kB T for each Fourier mode k and velocity

component i = 1, 2, 3 and then counting the number of modes. This spectrum occurs for any

fluid in local thermodynamic equilibrium and it has been observed in fluids at rest by light-

scattering experiments. For a good review, see:

José M. Ortiz de Zárate & Jan V. Sengers, Hydrodynamic Fluctuations in Fluids

and Fluid Mixtures (Elsevier, 2006)

The fluctuating hydrodynamics of Landau & Lifschitz (1959), discussed in Chapter I, was

proposed precisely to account for such thermal fluctuations. As we saw, this equation for an

incompressible fluid takes the form

@t u + (u · r)u = rp + ⌫⇤ 4u + f

where p is chosen to enforce r · u = 0 and fi (x, t) is a Gaussian space-time white-noise random

process with mean zero. The covariance prescribed by the fluctuation-dissipation relation
2⌫⇤ kB T
hfi (x, t)fj (x0 , t0 )i = d
ij 4x ⇤ (x x0 ) (t t0 )
⇢
was chosen precisely to recover the equipartition energy spectrum
kB T 4⇡k 2
E(k) = (⇤)
⇢ (2⇡)3
b (k) = 0. For a careful derivation, see:
of the incompressible fluid modes, satisfying k · v

52
 D. Bandak et al., “Dissipation-range fluid turbulence and thermal noise,” PRE 105,

065113 (2022).

3
The derived energy spectrum (*) di↵ers from the previous result by the prefactor 2 ! 1,

because the incompressibility constraint eliminates one degree of freedom per wavenumber. The

equipartition spectrum of the longitudinal modes (sound waves) must be physically present, but

these modes are expected to be dynamically decoupled from the transverse/solenoidal modes:

A. Donev et al., “Low Mach number fluctuating hydrodynamics of di↵usively mixing

fluids,” Comm. App. Math. Comp. Sci. 9, 47–105 (2014)

Nothing that is stated above is special to turbulent flows and holds equally for laminar flows.

There are, however, important implications for turbulent flows, where thermal and turbulent

velocity fluctuations can compete. It has generally been expected that the turbulent spectrum

in the dissipation range at scales (sufficiently) below the Kolmogorov length ⌘ should exhibit

an exponential decay:

E(k) ⇠ v⌘2 ⌘ exp( k⌘), k⌘ & 1,

valid down to nearly k mf p ' 1. For example, see

S. Khurshid, D. A. Donzis, and K. R. Sreenivasan, “Energy spectrum in the dissi-

pation range,” Phys. Rev. Fluids 3, 082601R (2018)

and many references therein. However, this conventional expectation is inconsistent with the

thermal spectrum (*)! In fact, equating the two spectra allows us to identify the crossover

wavenumber kc as

kc ⌘ = (2/ )W ( /2✓⌘1/2 ) ' (2/ ) ln( /2✓⌘1/2 )

in terms of the Lambert W -function. Because of the slow logarithmic dependence, kc ⌘ ' 1 even

when ✓⌘ ⌧ 1. For example, in the ABL with ✓⌘ = 2.83 ⇥ 10 8 and taking ' 7 (in agreement
.
with Khurshid et al., 2018), one finds that kc ⌘ = 2.25. The first person, apparently, to point out

the relevance of thermal fluctuations to the turbulent dissipation range was Robert Betchov:

53
 R. Betchov, “ On the fine structure of turbulent flows,” J. Fluid Mech. 3, 205216

(1957); R. Betchov, “Thermal agitation and turbulence,” in: Rarefied Gas Dynam-

ics, Proceedings of the Second International Symposium, University of California,

Berkeley, CA, 1960. L. Talbot, Ed. (New York, Academic Press, 1961), pp. 307-321.

However, only recently were Betchov’s ideas given empirical support, first by Eyink et al. (2021)

using a stochastic shell-model simulation, and subsequently by

J. B. Bell et al. “Thermal fluctuations in the dissipation range of homogeneous

isotropic turbulence,” J. Fluid Mech. 939: A12 (2022),

[Link]

who numerically simulated homogeneous forced turbulence in a liquid water-glycerol mixture

using incompressible fluctuating hydrodynamics. Their simulation was in periodic domain of

size 5 cm3 at room temperature (T = 300 K), with a mean dissipation per mass " = 17.7 cm2 /s3 ,

Kolmogorov length ⌘ = 0.025 cm, and Reynolds number Re=554 (Re = 61). See below:

Turbulent Energy Spectrum, from Bell et al. (2022).

Thermal fluctuations thus become relevant near ⌘, generally of order millimeters.

54
 A well-known e↵ect of such thermal fluctuations is renormalization of transport coefficients,

such as viscosity, so that they become dependent on the wavenumber cut-o↵ ⇤ (and also the

frequency cuto↵ ⌦). For example, see

D. Forster, D. R. Nelson, and M. J. Stephen, “Large-distance and long-time prop-

erties of a randomly stirred fluid,” Phys. Rev. A 16 732–749 (1977)

who discussed this phenomenon by means of renormalization group methods. Note that we

already indicated above in writing out the fluctuating Navier-Stokes equations that the viscosity

⌫⇤ is, in fact, ⇤-dependent, just as the turbulent stress tensor ⌧ ` is `-dependent! Thus, the

usual picture that there is “scale-separation” in molecular dynamics is not true, because of the

e↵ects of thermal fluctuations, and there is more similarity between “molecular viscosity” and

“turbulent viscosity” than is generally understood.

Nevertheless, essential di↵erences remain. Indeed, the description of the turbulent stress

⌧ ` by means of an “eddy viscosity” is not accurate (as we shall discuss in more detail later).

Furthermore, the thermal fluctuations are always weakly nonlinearly coupled, or “low Reynolds

number”. This may be seen by forming the “thermal Reynolds number” at every scale ` by
0
`v✓,` ` vth
Re✓` := ⇠ p
⌫` ⌫` n`d

If the “bare viscosity” ⌫0 at scales near ` ⇠ mf p , before renormalization by thermal fluctuations,

can be estimated by ⌫0 ⇠ vth 1/d

mf p , then we find using intp := n that
!1/2
d
intp
Re✓` ⇠ 2 d 2
.1 (⇤⇤)
mf p `

for ` & ( d / 2
intp mf p )
1/(d 2) . Recall mf p intp for a low-density gas, whereas mf p ' intp for

a liquid. Thus, (**) holds down to ` ' mf p , where the hydrodynamic approximation breaks

down. In particular, the “UV strong coupling regime” for d > 2 discussed by Forster et al.

(1977) is never achieved in molecular fluids.

55
 These considerations do, however, raise an important issue about our earlier discussion of

kinetic energy balance for turbulent fluid flows, which was based entirely on the deterministic

Navier-Stokes equations. How does the existence of thermal noise alter that calculation? It is

straightforward to derive the kinetic energy balance for 3D fluctuating Navier-Stokes equation:
✓ ◆ ✓ ◆ ✓ ◆
1 2 1 2 1 2 2⌫⇤ kB T 1 X
@t |v| + r · |v| + p v ⌫⇤ r |v| = ⌫⇤ |rv|2 + v · f + · 3 |k|2 ,
2 2 2 ⇢ L
|k|<⇤

where L is the side-length of the periodic flow domain. Two new terms arise from the stochastic

forcing, the obvious term v · f which has vanishing mean and the final constant term which

arises from the Itō chain rule. Note that in the statistical steady-state (fluid entirely at rest),

one obtains from this equation the mean energy balance

Z ⇤
⌫⇤ h|rv⇤ |2 i = 2⌫⇤ dk k 2 E(k), (?)
0

where E(k) is the thermal equipartition spectrum (*) and where now v⇤ denotes the velocity

in the model with UV cuto↵ ⇤. It follows that in thermal equilibrium, the gradients become

strongly ⇤-dependent with h|rv⇤ |2 i / ⇤5 . As we can see from the simulation of Bell et al.

(2021), however, the same equipartition spectrum should occur in a turbulent flow, e.g. decaying

turbulence behind a grid. Thus the same divergence with ⇤ must occur in turbulent flows! How

is it possible then to use deterministic Navier-Stokes to analyze the kinetic energy balance?

To answer this question, we will study the magnitude of the gradients rv⇤ by means of a

model turbulent energy spectrum

L5 k 4 kB T 4⇡ 2
E(k) = CK ("L)2/3 exp( bk⌘) + k
(1 + (kL) )2 17/6 ⇢ (2⇡)3
that was proposed by

T. von Kármán, “Progress in the statistical theory of turbulence,” Proc. Nat. Acad.

Sci. 34, 530 (1948)

to which we have added the thermal equipartition spectrum. In the following figure we plot

the enstrophy spectrum k 2 E(k) obtained from this model and also the mean-square gradients

h|rv⇤ |2 i obtained from (?) by integration over k:

56
The parameters here were chosen to correspond to the atmospheric boundary layer, as discussed

earlier. The vertical red line marks the position of the crossover wavenumber kc ⌘, where it can be

seen that k 2 E(k) has a deep local minimum, after the exponential decay around the Kolmogorov

scale but before the k 2 growth of the equipartition spectrum. Because of this deep minimum,

the mean-square gradient h|rv⇤ |2 i obtained by the integration (?) is almost independent of ⇤

within a 3-decade interval around kc . For any ⇤ chosen in this range it follows that

2kB T 1 X
|rv⇤ |2 · 3 |k|2 ,
⇢ L
|k|<⇤

and thus the kinetic energy balance for fluctuating Navier-Stokes becomes
✓ ◆ ✓ ◆ ✓ ◆
1 2 1 2 1 2 .
@t |v⇤ | + r · |v⇤ | + p⇤ v⇤ ⌫⇤ r |v⇤ | = ⌫⇤ |rv⇤ |2 ,
2 2 2

where we have neglected also the term v · f , which is small and has mean zero. It follows

that the energy balance reduces approximately to that for deterministic Navier-Stokes in this

range of ⇤. We may refer to the velocity-gradients rv⇤ in this interval around ⇤ ' kc as

e↵ective gradients, because these are the quantities which would be measured with any space-

resolution 1/⇤ selected in this range. Likewise, ⌫⇤ is the viscosity which would be measured

57
experimentally in a laminar flow, slightly renormalized from its “bare value” ⌫0 by thermal

fluctuations. Thus, the validity of the energy balance for deterministic Navier-Stokes is justified

to a good approximation.

Note, however, that not all of the predictions of deterministic Navier-Stokes are correct! For

example, the following plot shows the mean-square of the 6th-order space-derivative h|r6 v⇤ |2 i =
R⇤
2 0 dk k 12 E(k) calculated from the von Kármán model spectrum:

As can be seen, there is no range of ⇤ for which h|r6 v⇤ |2 i is ⇤-independent and the predictions

of deterministic Navier-Stokes are not accurate for any choice of ⇤.

Helicity Cascade

Recall that kinetic energy is not the only ideal invariant of 3D Euler! There is also the helicity
R
H = d3 x v(x) · !(x).

It was conjectured by

A. Brissand, U. Frisch, J. Leorat, M. Lesieur, and A. Mazure, “Helicity cascades in

fully developed isotropic turbulence,” Phys. Fluids 16, 1366-1368 (1973)

58
that flows with large-scale helicity (either by forcing or initial) shall have a joint cascade of

energy & helicity to small-scales, i.e. a helicity cascade coexisting with the energy cascade.

The large-scale helicity balance can be derived from the coarse-grained Navier-Stokes equation

in the form

@t v̄` = v̄` ⇥ !
¯` r(p̄` + 12 |v̄` |2 ) + f`s + ⌫4v̄`
1
with f`s = r · ⌧ ` the subscale force; ē` = 2
2 |v̄` | . Taking the curl of both sides gives the

coarse-grained vorticity equation

@t ! ¯ ` + f`s ) + ⌫4!
¯ ` = r ⇥ (v̄` ⇥ ! ¯`

From this it is easy to derive that the large-scale helicity density h̄` = v̄` · !
¯ ` satisfies

@t h̄` + r · J̄H
` = ⇤` 2⌫rv̄` : r!
¯`

where

J̄H
` = h̄` v̄` + (p̄` ¯ ` + v̄` ⇥ f`s
ē` )! ⌫rh̄`

= space transport of large-scale helicity

2⌫rv̄` : r!
¯` = viscous dissipation of helicity

⇤` = ¯ ` · f`s = helicity flux

2!

The latter quantity transfers helicity between scales. It is easy to see how this term does so if

one recalls the topological meaning of large-scale helicity:

R
H` = d3 x v̄` (x) · ! ¯ ` (x),

¯ ` , i.e. the flux of large-scale vorticity

which gives the asymptotic linking number of the lines of !

through the closed lines of the large-scale vorticity itself. This was proved by [Link], Sel.

[Link]. 5, 327 (1986); see also Arnold & Khesin (1998). Thus, we can understand that it

is the parallel component of f`s , along the lines of !

¯ ` , which modifies the large-scale helicity.

The component of the turbulent force f`s parallel to !

¯ ` accelerated fluid about closed loops of

¯ ` -lines, generating circulation around them. Vorticity flux is this created/destroyed though
!

the vortex-loop:

59
 FIGURE: A VORTEX LOOP.

Note that
v 2 (`)
2⌫rv̄` : r!
¯ ` = O(⌫ `3
)

whereas
v 3 (`)
⇤` = O( `2
).

Thus the viscous destruction of helicity is certainly negligible in the limit as ⌫ ! 0 with ` fixed.

If v is Hölder continuous with exponent 0 < h < 1, then

⇤` = O(`3h 2)

2
as an upper bound. Thus, a non-vanishing ⇤` for ` ! 0 is possible with any h  3 and, in

particular, for the K41 value h = 13 . It is quite possible to have co-existing cascade of energy

& helicity! Note that, a priori

⇤` = ⇧` · O(1/`)

However, these are only upper bounds and both ⇤` (and ⇧` ) take on both positive and negative

values and can have significant cancellations in averages over space or time. These cancellations

are disccussed in more detail by

60
 Q. Chen, [Link] & G. L. Eyink, “The joint cascade of energy and helicity in three-

dimensional turbulence,”, Phys. Fluids 15, 361-374(2003)

Space-time averages of energy flux and helicity flux versus filter length. = helicity input, " = energy

input. Data from the 5123 DNS of Chen et al. (2003) of forced homogeneous, isotropic turbulence.

The role of helicity in turbulent flow is still rather mysterious. Note that H is a pseudoscalar

(which changes sign under space-reflection) so that it can only be present for reflection-non-

symmetric forcing and/or initial conditions, on average. Of course, there can still be local

helicity — positive in some regions, negative in others — that vanishes on average. It has

been suggested that high local helicity (of either sign) may be correlated with low local energy

dissipation:

A. Tsinober & E. Levich, “On the helical nature of three-dimensional coherent struc-

tures in turbulent flows,” Phys. Lett. A 99 321-323 (1983)

61
 H. K. Mo↵at & A. Tsinober, “Helicity in laminar and turbulent flow,” Annu. Rev.

Fluid Mech. 24, 281 (1992)

but this remains very controversial. For a recent discussion, see:

O. L. Pham & D. V. Papavassiliou, “Helicity and dissipation correlation in anisotropic

turbulent flow fields,” Phys. Fluids 35, 105135 (2023)

It is less controversial that turbulence with strong helical forcing can have suppressed energy

cascade or even an “inverse cascade” of energy to large scales rather than to small scales:

F. Plunian et al., “Inverse cascade of energy in helical turbulence,” J. Fluid Mech.

895 A13 (2020).

We shall return to helicity and related issues when we consider later in depth the vorticity

dynamics in turbulent flows.

Entropy Cascade

If we consider the complete Navier-Stokes-Fourier system of equations governing the dynamics

of an incompressible fluid

@t v + (v · r)v = rp + ⌫4v, r·v =0

@t T + (v · r)T = T 4T + "/cP

with " = 2⌫|S|2 , then there is yet another ideal invariant, the total thermodynamic entropy:
R
S = d3 x s with

s = ⇢cP ln T.

It is easy to check that, with thermal conductivity defined as  = ⇢cP T,

|rT |2
@t s + r · (sv T rs) = + ⇢" 0,
T2

62
which is the local form of the second law of thermodynamics for an incompressible fluid. For

smooth solutions of ideal Euler equations with ⌫ =  = 0, entropy is thus conserved.

We have discussed extensive evidence that, in a turbulent flow, " 9 0 as ⌫ ! 0. This already

suggests that entropy may, in fact, not be conserved for the ideal limit. In addition, it is

possible that |rT |2 /T 2 9 0 as ⌫,  ! 0 together, because of development of large temperature

gradients. This possibility was suggested (in a special case of strong temperature forcing) by

A. M. Obukhov, Temperature Field Structure in a Turbulent Flow, Izvestiia Akademii

Nauk SSSR, Ser. Geogr. i Geofiz 13, 58 (1949).

and more generally by

G. L. Eyink and T. D. Drivas, Cascades and Dissipative Anomalies in Compressible

Fluid Turbulence, Phys. Rev. X, 8, 011022 (2018)

Both of these papers suggested that such anomalous entropy production may indeed occur,

and be associated to a cascade of entropy from the small-scales where entropy is produced to

the large-scales where entropy accumulates, in the form of a more spatially uniform and/or

rising large-scale temperature. A statistical steady-state is possible as well, if excess entropy is

removed by large-scale cooling or by a source of temperature inhomogeneities (e.g. di↵erential

heating/cooling) that sustains large-scale thermal structure.

A large-scale entropy balance can be introduced in an obvious way by considering

s` = ⇢cP ln T `

as the measure of the resolved/large-scale entropy. In particular, note that s` s` , because

entropy is a concave function of temperature. Thus, the total entropy increases under coarse-

graining, so that “myopic” observations at space resolution ` cannot miss anomalous entropy

production. It is straightforward to show using the thermofluid equations that

|rT ` |2
@t s` + r · [s` v` + `
⌧` (u, v) rT ` ] = r `
· ⌧` (u, v) + 2 + ⇢ ` "` ,
T`

63
where u = ⇢cP T is the internal energy per volume of the fluid and `
= 1/T ` . The terms

in this entropy balance proportional to  vanish in the limit as , ⌫ ! 0 under reasonable

assumptions (see Homework). However, the quantity ⇢"` representing the viscous dissipation

of kinetic energy is clearly not vanishing! This is the same term that appears in the balance

equation of the unresolved/small-scale kinetic energy, which in the limit ⌫ ! 0 becomes


1
@t ⇢k` + r · ⇢k` v` + ⌧` (P, v) + ⇢⌧` (vi , vi , v) = ⇢⇧` ⇢"` .
2

As discussed earlier, the coarse-grained viscous dissipation rate does not vanish as ⌫ ! 0. Thus,

the balance equation for s as , ⌫ ! 0 does not involve only ideal dynamical terms, unlike the

large-scale balances for kinetic energy and helicity.

There is, however, an alternative definition of “resolved/large-scale entropy” whose balance

involves only ideal dynamics, which is given simply by

s⇤` = s` + `
⇢k` .

Because k` 0, it is again true that s⇤` s` and furthermore lim`!0 s⇤` = s. Thus, the quantity

s⇤` is a reasonable choice as a “resolved/large-scale entropy” (and it is furthermore shown in

Eyink & Drivas (2016) that s⇤` is the entropy obtained from standard thermodynamic relations

if the “resolved internal energy” is obtained from coarse-grained observations of the conserved

energy and momentum). It is straightforward to show that the term ⇢ ` "` cancels in the balance

equation for s⇤` , which in the limit , ⌫ ! 0 becomes:

@t s⇤` + r · [s⇤` v` + q]
` `
= ⇢ ` ⇧` + (Dt `
)⇢k` + r `
· q` := ⌃`

where we have defined a “turbulent heat-transport vector”

1
q` := ⌧` (h, v) + ⇢⌧` (vi , vi , v)
2

with h = u + P = ⇢(cp T + p) the thermodynamic enthalpy per volume. The expression ⌃` on

the righthand side of this resolved entropy balance does not depend upon  or ⌫ explicitly and

64
represents an entropy flux from unresolved length scales < ` to resolved scales > `. Each of the

three terms appearing in ⌃` has a transparent physical interpretation:

⇢ ` ⇧` = entropy production from turbulent energy cascade

(Dt `
)⇢k` = entropy production from large-scale temperature change

anti-correlated with subscale kinetic energy density

r `
· q` = entropy production from turbulent heat-transport

down the gradient of large-scale temperature

When the entropy production is non-vanishing in the limit , ⌫ ! 0 (anomalous), then now-

standard arguments show that there must be a cascade of entropy. Increase of entropy at

large-scales seen by a “myopic” observer cannot be accounted for by thermal-conductive or

viscous entropy production, and require non-vanishing entropy flux. For example, if an initial

large-scale temperature distribution is created in a turbulent flow, one expects the temperature

field to become nearly homogenous at large-scales due to turbulent heat transport. This is

the type of situation considered by Obukhov (1949), where decay of an initial temperature

inhomogeneity is associated to cascade of entropy from small-scales up to the scale of the

inhomogeneity. Another example is driven turbulence with a nearly homogeneous temperature,

where the large-scale temperature must slowly increase due to viscous heating. Here also entropy

must cascade from small-scales to account for the gradual increase in large-scale temperature.

Just as for cascades of kinetic energy and helicity, a non-vanishing flux of entropy requires

“rough” or “non-smooth” fields of both velocity and temperature. By exploiting the above

explicit expression ⌃` for entropy flux, Eyink & Drivas (2018) derive constraints on the scaling

exponents ⇣pv of velocity and ⇣pT of temperature for all p 3 of the form

2⇣pT + ⇣pv  p, ⇣pT + 2⇣pv  p, 3⇣pv  p,

in order that a turbulent entropy cascade can be sustained.

65