II E↵ective “Large-Scale” Equations

and the Turbulent Energy Cascade

(A) Coarse-Graining/Filtering/Mollifying

We have seen that the non-vanishing of turbulent energy dissipation as Re ! 1 requires

that velocity-gradients |ru| ! 1 in this limit. But, in that case, the usual formulation

of the fluid equations as PDE’s with smooth solutions no longer makes sense! To obtain a

dynamical description, we must regularize those equations. We shall use an approach based on

coarse-graining/filtering/mollifying with a smooth kernel G that satisfies

G(r) 0

G(r) ! 0 rapidly for |r| ! 1

Z
dd r G(r) = 1

It is also understood that G is centered at r = 0:

Z
dd r r G(r) = 0,

R
and that dd r|r|2 G(r) ⇡ 1. Other specific requirements shall be introduced as needed. Set

d
G` (r) ⌘ ` G(r/`)

R
so that all of the above properties hold, except that now dd r|r|2 G` (r) ⇡ `2 .

Using this kernel, now define a coarse-grained velocity at length-scale ` by

Z
v̄` (x) = dd r G` (r)v(x + r)

This represents the average velocity of a fluid “parcel” of size ` at position x. It can also

be called a low-pass filtered velocity, containing only length-scales > `, or a mollified (i.e.

smoothed) velocity. The corresponding small-scale / high-pass filtered velocity is given by

Z
v`0 (x) ⌘ v(x) v̄` (x) = dd r G` (r) v(r; x) (1)

1
where

v(r; x) = v(x + r) v(x)

is the velocity-increment across a separation vector r at point x

Comments:

? This coarse-graining is similar to that used to derive hydrodynamics from MD. However,
⇠ |v| whereas here we have in mind `
in that case ` ⌧ Lr = gradient length = Lr .
|rv|

? In principle, only coarse-grained fields v̄` (x) are experimentally measurable. Every ex-

periment has some spatial resolution `, such that only averaged properties for length-scales

` are obtained. The fine-grained/bare field v(x) are unobservable objects corresponding to

a mathematical idealization

v̄` (x) ! v(x) as ` ! 0

This idealization is physically unachievable, in the strictest sense, since the hydrodynamic

equations are not valid for ` ⇡ , the mean-free path. In general, v̄` (x) is the more physical

object and v(x) is an “ideal” object which is useful if Lr ` .

? In physics, the coarse-grained field is similar to an “e↵ective block-spin” that appears

in the method of real-space renormalization group (RG). It removes the ultraviolet divergence

associated with blow-up of velocity gradients |rv| ! 1 as ⌫ ! 0, since necessarily

Z
1
rv̄` (x) = dd r (rG)` (r) v(x + r)
`

remains finite. This “regularization” introduces an arbitrary length-scale `, on which no ob-

jective physical fact can depend. Note that coarse-graining is a purely passive operation—

“removing one’s spectacles”—which changes no physical process.

?The method of filtering is also employed as part of the large-eddy simulation (LES) mod-

elling technique for turbulent flow. Here are some key works:

M. Germano, “Turbulence: the filtering approach,” J. Fluid Mech. 238, 325–336 (1992).

S. B. Pope, “Ten questions concerning the large-eddy simulation of turbulent flows,” New J.

Phys. 6 35 (2004)

2
3
(B) E↵ective Large-Scale Equations

See also T&L , Section 2.1. Starting with the incompressible Navier-Stokes equations

@t v + (v · r)v = rp + ⌫4v + f , r · v = 0

for the bare/fine-grained velocity field, we can derive an equation for v̄` . Note that

(rf )` = rf̄` ,

i.e. space-derivatives commute with filtering. Thus,

@t v̄` + r · (vv)` = rp̄` + ⌫4v̄` + f̄` , r · v̄` = 0

Define the turbulent (or Reynolds) stress tensor

⌧ ` = (vv)` v̄` v̄`

so that

@t v̄` + (v̄` · r)v̄` + r · ⌧ ` = rp̄` + ⌫4v̄` + f̄`

This is the “e↵ective equation” for the large-scale velocity. Note that it is not closed, i.e. ⌧

is not given (in a simple way) as a function of v̄` .

We now wish to estimate the viscous term ⌫4v̄` as small, i.e. to show that it can be

neglected relative to the other terms when ⌫ is small or when ` is large. To measure “size”, we

need the notion of a norm of a function f : Rd ! Rd , i.e. a mapping f 7 ! kf k such that:

(i) kf k 0 and kf k = 0 i↵ f = 0

(ii) k↵f k = |↵| · kf k for any real scalar ↵.

(iii) kf + gk  kf k + kgk (triangle inequality)

4
Some common norms are the Lp norms for p 1
 Z 1/p
1
kf kp ⌘ dd x |f (x)|p
V V

and for p = +1

kf k1 ⌘ sup |f (x)|
x2V

Note that these satisfy

kf kp  kf kp0 for p0 p

and

kf k1 = lim kf kp .
p!1

Some of these have simple meaning

Z
1
kf k1 = h|f |i where hgi = d3 x g(x)
V V
p
kf k2 = h|f |2 i = frms when hf i = 0

kf k1 = |f |max

For more details, see A. N. Kolmogorov & S. U. Fomin, Introductory Real Analysis, Dover, 1975

Now we estimate
Z
⌫
⌫4v̄` (x) = 2 dd r(4G)` (r)v(x + r) (integration by parts!)
`

Then1
Z
⌫
k⌫4v̄` kp  dd r |(4G)` (r)| · kv(· + r)kp (by triangle inequality)
`2
⌫
= (const.)kvkp
`2
R
using kv(· + r)kp = kvkp and assuming that dd ⇢ |4G(⇢)| < +1.
1
The “continuous” version of the triangle inequality that we use below and repeatedly in these notes is usually
R R
called the Minkowski integral inequality. It states that k dd r F (r, ·)kp  dd r kF (r, ·)kp . See G. H. Hardy, J.

E. Littlewood, and G. Pólya, “Inequalities” (Cambridge University Press, 1952), Theorem 202, or E. Stein,

“Singular Integrals and Di↵erentiability Properties of Functions” (Princeton University Press, 1970), §A.1.

5
 If kvkp stays finite in the limit that ⌫ ! 0, then

lim k⌫4v̄` kp = 0 !!!

⌫!0

Note that energy is given by

1
E(t) = kv(t)k22
2
Since dE/dt = ⌫krvk22  0, kv(t)k2 can only decrease in time. Thus, kv(t)k2  kv(t0 )k2 =

initial energy. Thus for decaying turbulence and p = 2, it is true that kv(t)k2 must stay finite

in the limit as ⌫ ! 0. Experimental evidence is that kv(t)kp stays finite for all p 1.

Heuristically, we may say that

⌫U
⌫4v̄` ⇠
`2
We shall see later that
U2
(v̄` · r)v̄` ⇠ r · ⌧ ` ⇠ rp̄` ⇠
`
⌫ 1
Thus, the viscous term is smaller by a factor U` = Re` .

One of the important terms in the large-scale equation is from the turbulent (or subscale) force

f`s = r · ⌧`

This can be thought of as an e↵ective “body-force” on the large-scales produced by the elimi-

nated subscales. Note that

Prop: The stress (⌧ ` )ij = (vi vj )` v̄`i v̄`j is a symmetric, nonnegative-definite matrix at

each space-time point (x, t).

Proof: Omit ` for simplicity of notation. Clearly,

⌧ij = vi vj v̄i v̄j

6
is symmetric in i, j. Note also that
Z
⌧ij = dd r G` (r)vi (x + r)vj (x + r) v̄i (x)v̄j (x)
Z
= dd r G` (r)[vi (x + r) v̄i (x)][vj (x + r) v̄j (x)]

so that
X Z X
2
c⇤i cj ⌧ij = d
d r G` (r) cj [vj (x + r) v̄j (x)] 0
ij j

In fact, ⌧ij is just the covariance of the random variable vi (x + r) with r distributed according

to the “probability density” G` (r). Thus, ⌧ij is the velocity covariance of the fluid parcel of size

` at spacetime point (x, t).

This result is due to Vreman, Geurts & Kuerten, “Realizability conditions for the turbulent

stress tensor in Large Eddy Simulation,” J. Fluid Mech., 278, 351-362, 1994.

(C) Energy Balance

See T & L, Sections 3, 1-2; Frisch, Section 2.4.

We have seen that the viscous term ⌫4v̄` is negligible is the large-scale e↵ective equation.

Since kv̄` (t)k22  kv(t)k22 by convexity, the kinetic energy decays even with “spectacles o↵”!

How? To analyze this question, we must consider energy balance in detail.

Large-Scale Energy Balance

From the equation for v̄` it is easy to derive an evolution equation for the large-scale energy

(per unit mass)

Z
1
Ē` (t) = dd x |v̄` (x, t)|2
2

and its density

1
ē` (x, t) = |v̄` (x, t)|2 .
2

One finds by straightforward calculus that

7
 @t ē` + r · [(e¯` + p̄` )v̄` + ⌧ ` · v̄` ⌫rē` ] = rv̄` : ⌧ ` ⌫|rv̄` |2 + v̄` · f̄`

with

J̄E
` = (ē` + p̄` )v̄` + ⌧ ` · v̄` ⌫rē`

= space transport (flux) of large-scale energy,

8
>
> (ē` + p̄` )v̄` = transport by large-scale advection
>
>
<
(or in detail: ) ⌫rē` = viscous di↵usion of large-scale energy
>
>
>
>
: ⌧ ` · v̄` = turbulent di↵usion of large-scale energy

Q̄` = v̄` · f` = power input from external force into large-scales (per mass)

"¯` = ⌫|rv̄` |2 = large-scale energy dissipation rate (per unit mass)

Z
Ē` (t) = dd x "¯` (x, t) = total energy dissipation (per mass) in the large-scales

= ⌫krv̄` (t)k22
⌫ R
 (const.)kv(t)k22 by Minkowski estimate of rv̄` = 1
` dd r (rG)` (r)v(x + r)
`2
! 0 as ⌫ ! 0 !

We have assumed here that kv(t)k22 stays bounded as ⌫ ! 0.

The important term in the large-scale balance is

⇧` = rv̄` : ⌧ `

= deformation work of the large-scale strain against the small-scale stress

⇧` > 0 =) large-scale sink; ⇧` < 0 =) large-scale source

Alternative forms:

⇧` = S̄` : ⌧ `

8
with S̄` = 12 [(rv̄` ) + (rv̄` )T ] = large-scale strain (by symmetry of ⌧ ` ); or

⇧` = rv̄` : ˚
⌧`

1
with ˚
⌧` = ⌧` d tr(⌧ ` )I = deviatoric/traceless part of the stress (by r · v̄` = 0); or

⇧` = S̄` : ˚
⌧ ` (by both together).

Where does the energy go? To the small scales! Define

1
k` (x, t) = tr(⌧ ` ) = small-scale kinetic energy (per mass)
2
1 1
= (|v|2 )` |v̄` |2 0 by positive-definiteness of stress ⌧ `
2 2

Note that ē` = 12 |v̄` |2 so that

1
ē` + k` = (|v|2 )`
2
R R
and thus (since dd xf¯(x) = dd xf (x))
Z Z
d 1
d x [ē` + k` ] = dd x |v(x, t)|2
2
= E(t) = total kinetic energy (per mass)

An evolution equation of the following form holds for k` :

0
@ t k` + r · JE
` = rv̄ : ⌧ ` "0` + Q0`

Recall that ⇧` = rv̄` : ⌧ ` . Also

"0` = ⌫[(|rv|2 )` |rv̄` |2 ] = viscous energy dissipation in the small-scales

so that "¯` +"0` = ⌫(|rv|2 )` gives the total dissipation averaged over the region of radius ` around

x. Note that ⇧` appears with opposite signs in the equations for ē` and k` : it tends to act as a

9
sink for ē` and a source for k` . Thus, it transfers energy from large-scales to small-scales. For

this reason, ⇧` is often called (scale-to-scale) energy flux.

We now derive the equation for k` , following Germano (1992). In fact, we derive a more

general evolution equation for ⌧ ` . Starting with @t vi = (vk @k )vi @i p + ⌫@k2 vi + fi ,

@t vi vj = (@t vi )vj + vi (@t vj )

= vk (@k vi )vj (@i p)vj + ⌫(@k2 vi )vj + fi vj + (i $ j)

= @ k vi vj vk @i pvj @j pvi + p(@i vj + @j vi )

+ ⌫@k2 vi vj 2⌫@k vi @k vj + fi vj + fj vi

Similarly,

@t (v̄i v̄j ) = (@t v̄i )v̄j + v̄i (@t v̄j )

= [ (v̄k @k )v̄i @k ⌧ik @i p̄ + ⌫4v̄i + f¯i ]v̄j + (i $ j)

= @k (v̄i v̄j v̄k ) @k [⌧ik v̄j + ⌧jk v̄i ] + ⌧ik (@k v̄j ) + ⌧jk (@k v̄i )

@i (p̄v̄j ) @j (p̄v̄i ) + p̄(@i v̄j + @j v̄i )

+ ⌫@k2 (v̄i v̄j ) 2⌫@k v̄i @k v̄j + (f¯i v̄j + f¯j v̄i )

Subtracting the two equations yields an equation for ⌧ij = vi vj v̄i v̄j .

To express the various terms that appear, we must introduce the generalized central moments

of Germano (usually called cumulants in probability theory, or connected correlation functions

in statistical physics). The nth-order generalized central moment ⌧ (f1 , . . . , fn ) is defined as

follows:

f¯1 = ⌧ (f1 )

f1 f2 = ⌧ (f1 , f2 ) + f¯1 f¯2

f1 f2 f3 = ⌧ (f1 , f2 , f3 ) + f¯1 ⌧ (f2 , f3 ) + f¯2 ⌧ (f1 , f3 ) + f¯3 ⌧ (f1 , f2 ) + f¯1 f¯2 f¯3

and, iteratively,
p
XY
f1 . . . fn = ⌧ (fi(j) , . . . , fi(j) )
1 nj
I2P j=1

10
 (1) (1) (p) (p)
where the sum is over the set P of all partitions I = {i1 , . . . , in1 }, . . . , {i1 , . . . , inp } of the set
P
{1, 2, ..., n} with pj=1 nj = n. We thus see that

f1 . . . fn = ⌧ (f1 , . . . , fn ) + terms defined by lower-order cumulant functions

so that we may solve successively to obtain

⌧ (f1 ) = f¯1

⌧ (f1 , f2 ) = f1 f2 f¯1 f¯2

⌧ (f1 , f2 , f3 ) = f1 f2 f3 f¯1 ⌧ (f2 , f3 ) f¯2 ⌧ (f1 , f2 ) f¯3 ⌧ (f1 , f2 ) f¯1 f¯2 f¯3

= f1 f2 f3 f¯1 f2 f3 f¯2 f1 f3 f¯3 f1 f2 + 2f¯1 f¯2 f¯3

and etc.! Note: The “generalized central moments” ⌧ (f1 , . . . , fn ) are the cumulants of the

random variables f1 (x + r), . . . , fn (x + r), distributed according to the density G` (r) on r.

The final equation obtained for ⌧ij has the form

@t ⌧ij + @k Jijk = [v̄i,k ⌧kj + ⌧ik v̄j,k ] production of stress by large-scale straining

+ 2⌧ (p, Sij ) pressure-strain correlation

2⌫⌧ (vi,k , vj,k ) viscous destruction of stress

+ [⌧ (vi , fj ) + ⌧ (vj , fi )] production of stress by forcing

@vi
with vi,k = @xk , etc. and

Jijk = ⌧ij v̄k + ⌧ (p, vi ) jk + ⌧ (p, vj ) ik + ⌧ (vi , vj , vk ) ⌫⌧ij,k

where

⌧ij v̄k = advective transport of stress

⌫⌧ij,k = viscous transport of stress

1
Taking 2 of the trace of the equation for ⌧ij gives the equation for k, with

0 1
JiE = kv̄i + ⌧ (p, vi ) + ⌧ (vk , vk , vi )
2

11
 Q0 = ⌧ (vi , fi )

Remark: There is a tempting analogy

TURBULENCE : MOLECULAR DYNAMICS

ē` = 12 ⇢|v̄` |2 large-scale kinetic energy $ 1

2 ⇢|v|
2 kinetic energy

⇢k` small-scale kinetic energy $ u = ⇢cP T internal energy

For this reason, ⇧` = S̄` : ⌧ ` is sometimes called subscale dissipation (or “subgrid-scale

dissipation” in LES). Note, however, that 12 ⇢|v|2 + u is conserved, while 12 ⇢|v̄` |2 + ⇢k` is NOT
1 2
and, in fact, has the same space integral as total kinetic energy 2 ⇢|v| ! (For this reason, a

better correspondence is u for molecular dynamics and ū⇤` = ū` + ⇢k` for turbulence, so that

the total 12 ⇢|v̄` |2 + ū⇤` is conserved.) Furthermore, there is a big separation in scale between the

length Lr of variation of v and the mean-free-path mf p of molecules, whose energies (kinetic

+ potential) constitute u. As we discuss in more detail later, this is not true for ē` , k` .

There are some important alternative forms for the energy balances that we now discuss.

Note that

rv̄` : ⌧ ` = v̄` · f`s + r · (⌧ ` · v̄` )

Where f`s = r · ⌧ ` . Thus, we may rewrite the energy balance as

@t ē` + r · [(ē` + p̄` )v̄` ⌫rē` ] = v̄` · f`s ⌫|rv̄` |2 + v̄` · f¯` .

Where v̄` · f`s is the (negative) power input by the subscale force f`s . Note, however, that this

term is not Galilei invariant — an observer at rest and an observer moving with respect to a

turbulent fluid would disagree about the “dissipation” due to such a term!

Another form of the balance can be written using the turbulent vortex-force

f`v = (v ⇥ !)` ¯ ` , f`iv = ✏ijk ⌧` (vj , !k ).

v̄` ⇥ !

It is not hard to show using r · (vv) = v ⇥ ! r( 12 |v|2 ) that

f`s = f`v rk`

12
so that

@t v̄` + (v̄` · r)v̄` = f`v rh̄` + ⌫4v̄` + f̄`

with h̄` ⌘ p̄` + k` = “turbulent enthalphy”. Then,

@t ē` + r · [(ē` + h̄` )v̄` ⌫rē` ] = v̄` · f`v ⌫|rv̄` |2 + v̄` · f`

Estimation of Energy Flux

We have seen that the viscous dissipation in large-scale ⌫|rv̄` |2 is negligible and that the

energy flux

⇧` = rv̄` : ⌧ `

must therefore be the main “sink” term in the large-scale energy balance. We will now estimate

this term. Note that

Z
1
rv̄` (x) = dd r (rG)` (r)v(x + r)
`Z
1
= dd r (rG)` (r)[v(x + r) v(x)],
`
R
since dd r rG(r) = 0.

To get simple estimates, let us assume for the moment that G is C 1 with compact support

in the unit ball, e.g. 8

>
< N exp[ 1
(1 r2 )
] for |r|  1
G(r) =
>
: 0 for other r
.
where N = 0.8822 is a normalization factor for dimension d = 3.

13
 Smooth, Compactly-Supp orted Filter Kernel
0.45

0.4

0.35

0.3

0.25
G(r)

0.2

0.15

0.1

0.05

0
−1.5 −1 −0.5 0 0.5 1 1.5
r

We shall remove this restriction later on!

qP
Then, with |A| = 2
ij |aij | ,
Z
C
|rv̄` (x)|  sup | v(r; x)| with C = dd ⇢|rG(⇢)|
` r<`

or
v(`; x)
rv̄` (x) = O( )
`

with v(`; x) ⌘ supr<` | v(r; x)|. Now we must develop similar estimates for the stress ⌧ ` . For

this purpose, the following formula is crucial

Z Z Z
d d
⌧` (f, g) = d r G` (r) f (r) g(r) d r G` (r) f (r) dd r G` (r) g(r)

or

⌧` (f, g) = h f gi` h f i` h gi`

where h·.i` denotes average over r with respect to G` (r) and f (r; x) = f (x + r) f (x), etc.

The above formula will turn out to be absolutely essential for much of our further analysis. It

14
is one of the most useful formulas in the course, which we shall use many times. Amazingly,

it was not discovered in this context until the 1990’s [P. Constantin, et. al. Commum. Math.

Phys., 165, 207 (1994); G. L. Eyink, J. [Link]., 78, 335 (1995).]

It is easy to verify the formula by substituting the definitions of f , g and integrating.

Later on, we shall prove a more general formula for all generalized central moments. Applying

this formula to ⌧`ij = ⌧` (vi , vj ), we get

Z Z 2
|⌧ ` (x)|  dd r G` (r)| v(r; x)|2 + dd r G` (r)| v(r; x)|
Z Z 2
d 2
 d r G` (r) v (`; x) + dd r G` (r) v 2 (`; x)

= 2 v 2 (`; x),
R
with dd r G` (r) = 1, or,

⌧ ` (x) = O( v 2 (`; x)).

Putting the two estimates together gives

v 3 (`; x)
|⇧` (x)|  (const.) .
`
An estimate of this type was first derived by Lars Onsager around 1945 and closely related

results were obtained by A. N. Kolomogorov in his famous papers in 1941, using probabilistic

assumptions. The estimate has some important implications that we now discuss.

If v is continuously di↵erentiable at point x, then Taylor expansion in r gives

v(x + r) = v(x) + (r · r)v(x? )

=) v(r; x) = (r · r)v(x? )

=) v(`; x)  ` · sup |rv(x)| = O(`)

x

In that case,
v 3 (`)
⇧` = O( ) = O(`2 ) ! 0 as ` ! 0
`
p
Thus, ⇧` is too small for `  "/|rv|3 to account for a non-vanishing energy dissipation "!

More generally, suppose that

v(`; x) ⇠ `h

15
for some 0 < h < 1. Then,

v 3 (`)
⇧` = O( ) = O(`3h 1
) as ` ! 0 if h > 1/3
`

This was first pointed out by Onsager(1945, 1949). A “minimal assumption” is that

v(`) ⇠ ("`)1/3 ,

which is the famous prediction of Kolmogorov & Obukhov(1941), Onsager(1945, 1949), and

Heisenberg & von Weizsäcker(1948)

This scaling is often explained as the result of “dimensional analysis”, but it has a deeper

dynamic basis. We have already seen that v cannot remain di↵erentiable in the limit that

⌫ ! 0, if the experiments are correct that

" = ⌫h|rv|2 i 9 0 as ⌫ ! 0.

There is a more refined result. We say that v is Hölder continuous at point x with exponent

h, 0 < h < 1, if

| v(r; x)|  C|r|h (?)

for all |r| < r0 and some constant C. If it holds, then v(`; x) = O(`)h . We thus see that

(?) cannot hold with h > 1/3 for all x, as ⌫ ! 0, if the experiments are correct that " 9 0

in that limit (Onsager, 1949)

To make this argument a bit more general and detailed, we consider the total flux
Z
dd x ⇧` (x) ⌘ ⇧` ,
V

or, equivalently, the mean flux over the flow domain

Z
1
h⇧` i = d3 x ⇧` (x).
|V | V

16
Then
Z
1
|h⇧` i|  h|⇧` |i = d3 x |⇧` (x)| = k⇧` k1 ,
|V | V

and

k⇧` k1  k⇧` kr for r 1.

To get further estimates, we must recall some basic results for the Lp -norms, the

1 1
Hölder inequality: kf gk1  kf kp kgkq , + =1
p q
n
Y n
Y Xn
1 1
generalized Hölder inequality: k fi kr  kfi kpi , = , r 1
pi r
i=1 i=1 i=1

For example, see Kolmogorov & Fomin(1975), or any good textbook on real analysis. Since

⇧` = rv̄` : ⌧ ` ,

k⇧` kr  krv̄` k3r k⌧ ` k3r/2

To simplify notation, set p = 3r or r = p/3, with p 3. We must bound the terms krv̄` kp ,

k⌧ ` kp/2 . Since
Z
1
rv̄` (x) = dd r (rG)` (r) v(r; x),
`

the triangle inequality gives

Z
1
krv̄` kp  dd r |(rG)` (r)|k v(r)kp .
`

Now, let us assume that for some p, 0< p < 1,

k v(r)kp  C|r| p
(?)

for |r|  r0 and some constant C > 0. Then,

Z
C
krv̄` kp  dd r |(rG)` (r)| · |r| p
`
r
= C 0 ` p 1 (substitute ⇢ = )
`
R
with C 0 = C dd ⇢ |rG(⇢)| · |⇢| p , which is assumed to be finite. Notice that we have not had to

assume that G is compactly supported, but only that it decays rapidly enough for large |⇢| so

17
that the integral converges. This same approach could have been used earlier for the pointwise

estimates. Thus,
1
krv̄` kp = O(` p
)

To estimate k⌧ ` kp/2 we use the Hölder inequality again with

Z Z Z
⌧ ` = d3 rG` (r) v(r) v(r) dd rG` (r) v(r) dd rG` (r) v(r)

to get, for p 2,
Z Z
k⌧ ` kp/2  dd r G` (r)k v(r) v(r)kp/2 + k dd r G` (r) v(r)k2p
Z Z 2
d 2
 d rG` (r)k v(r)kp + dd rG` (r)k v(r)kp
(Z Z )
2
 C2 dd rG` (r)|r|2 p + dd rG` (r)|r| p

= Cp `2 p

R R
where Cp ⌘ C 2 [ dd ⇢G(⇢)|⇢|2 p + ( dd ⇢G(⇢)|⇢|2 p )], which is assumed to be finite (a very

modest requirement on G). Thus,

k⌧ ` kp/2 = O(`2 p ), p 2

and, finally,

1
k⇧` kp = O(`3 p 1 ), p 3, so that h⇧` i ! 0 as ` ! 0 unless p  3 for p 3.

We note that a function v with

kvkp < +1

and

k v(r)kp  C|r| p

is called Besov regular with pth-order Besov exponent p. This is an “Lp -version” of Hölder

continuity. Note that

lim k v(r)kp = k v(r)k1

p!1

18
 = sup | v(r; x)|
x2V

Thus, the p ! 1 limit of Besov regularity corresponds to uniform Hölder continuity, i.e.

| v(r; x)|  C|r| 1

for all x 2 V

for |r|  r0 . Our previous result thus says that non-vanishing energy dissipation requires
1
a velocity field which is not too regular in the limit ⌫ ! 0, i.e. v may not have p > 3 for any

p 3.

It is more traditional to consider so-called (absolute) structure functions for order p:

Sp (r) ⌘ h| v(r)|p i

= k v(r)kpp

with assumed scaling exponent ⇣p

|r| ⇣p
Sp (r) ⇠ Ap uprms ( ) (??)
L

We have written this in a dimensionally correct form with urms = h|v hvi|2 i1/2 the root-square

velocity and L a length-scale characteristic of the large-scale production mechanism. Note that

f (z) ⇠ g(z) for z ⌧ 1

means that limz!0 f (z)/g(z) = 1. Then (??) implies the previous estimate (?) with p = ⇣p /p.

Our previous result then implies that

p
h⇧` i ! 0 for ` ⌧ L, unless ⇣p  3 for p 3

The classical Kolmogorov(1941) theory assumes that

p
⇣p = 3 for all p.

Then, using

19
 u3rms
"⇠ L ,

one gets

Sp (r) ⇠ Cp ("|r|)p/3 for all p, and |r| ⌧ L.

Of course, only the inequality ⇣p  p/3, p 3 is rigorously implied. We shall discuss later the

physical meaning of assuming that ⇣p = p/3, but essentially, it is a “uniformity” assumption on

velocity increments | v(r; x)| which rules out large fluctuations in values for di↵erent x.

Note that the estimate v(`) ⇠ ("`)1/3 is consistent with

v 3 (`)
h⇧` i ⇠ ` ⇠"

for all ` ⌧ L. Thus, one can explain the observed rate of energy dissipation " — independent

of viscosity ⌫ — by the efficient transfer of energy down to small-scales where viscosity is

e↵ective. This length-scale is the so-called Kolmogorov (micro) scale ⌘. It can be obtained as

the length-scale at which

⇧` ⇠ ⌫|rv̄` |2

We previously estimated the viscous dissipation by the rms velocity, but now we have an

improved estimate in terms of velocity increments, as

v 2 (`)
⌫|rv̄` |2 = O( ⌫ `2
)

Using ⇧` = O( v 3 (`))/`, we get an estimate for ⌘ as the solution of

v 3 (⌘) ⇠ v 2 (⌘)
=⌫ =) v(⌘)⌘ ⇠
=⌫
⌘ ⌘2

Note that this implies that the “turbulent Reynolds number” v(`)`
⌫ is approximately ⇠
= 1 for

`⇠
= ⌘. If we now use the K41 (i.e. Kolmogorov 1941) conjecture that v(⌘) ⇠ ("⌘)1/3 , then

"1/3 ⌘ 4/3 ⇠
= ⌫ =) ⌘ ⇠
= ⌫ 3/4 " 1/4

This is dimensionally correct, since ["] = L2 /T 3 , [⌫] = L2 /T . The K41 scaling v(`) ⇠ ("`)1/3

is thus expected for a range of length-scales ⌘ ⌧ ` ⌧ L, the so-called inertial (sub)range. This

scaling prediction was one of the great early successes of turbulence theory, usually described

in terms of energy spectra E(k) in Fourier space. We do not discuss Fourier spectra here, but

note only the rough equivalence kE(k) ⇠ ( v(`))2 with ` ⇠ 1/k.

20
2
256 H . L. Grant, R. W . Stewart and A . Moilliet
0

The K41 prediction E(k) ⇠ "2/3 k 5/3 was finally verified in a rather convincing way by
6
C :DD C

H.L. Grant, R.W. Stewart & A. Moilliet, “Turbulent spectra from a tidal channel,”

J. Fluid Mech., 12, 241-268 (1962).

7 D7

Their data was presented at a famous meeting in Marseille in 1961, confirming the predictions
4 6 7

of Kolmogorov twenty years earlier (1941). However, Kolmogorov himself was in attendance ...
3

and he pointed out difficulties with his previous theory and proposed a “refinement” !!!
C 8 C7 3 3 34 7 3D :DD C

FIGURE
11. Calculated values of the quantity h" as a function
of the rate of dissipation of energy.
7 D7
4 6 7
C 4 7 D D D:7 3
3D
/ D
4 3
7 17 : .76 3

log 2
4 6 7

FIGURE 12. Seventeen spectra compared to the theories of Kolmogoroff, Heisenberg and
Kovasznay. The straight line has a slope of - 9, the curved solid line is Heisenberg's theory
3

and the dashed line is Kovasznay's theory. Within the square, the observations are too
crowded to display on this scale and they are shown in figure 13.
:DD C
3676 8

21
,
The original K41 theory thus works reasonably well for Sp (r) with p not too large (p ⇡ 1 3),

although even here there are deviations. See data from the 81923 DNS of

K. Iyer et al., “Scaling exponents saturate in three-dimensional isotropic turbu-

lence,” Phys. Rev. Fluids 5 054605 (2020)

We also show their results for the local slope ⇣(r) = d ln S2 (r)/d(ln r) :

22
 We now make a preliminary attempt to answer the question: Is the fluid approximation valid?

If ⌘ is indeed the smallest length-scale in a turbulent flow, then

|v| ⇠
L r = |r v|
= ⌘.

The condition for validity of the hydrodynamic description is that (microscale) Knudsen number
mf
be small: ⌘ ⌧1

Using

⇠
= ⌫/c with c the sound speed
mf
⌫ 3/4 1/4 u3
⌘ ⇠
= ⌫ 3/4 " 1/4 ⇠
=( ) L with " ⇠
= rms
urms L
mf ⇠ ⌫ u rms
=) = ( )1/4 ( ) = M a(Re) 1/4
⌘ urms L c

Thus, the fluid approximation is valid for

M a ⌧ 1 and/or Re 1.

This argument goes back to S. Corrsin, “Outline of some topics in homogeneous turbulent

flow,” J. Geophys. Res.,64, 2134-2150(1959). It is important to note that validity of the

Landau-Lifschitz fluctuating hydrodynamics requires the same condition.

Note that an important distinction is usually made between turbulence and molecular dynamics,

due to a presumed separation of scales (both length and time) in the latter case, but NOT in

the former. In derivations of hydrodynamics from deterministic Boltzmann equation, the sig-

nature of separation of length-scales is that a smooth “hydrodynamic profile” ⇢a (x, t) exists,

so that ⇢¯a,` (x, t) ⇡ ⇢a (x, t) for all ` in the range mf ⌧ ` ⌧ Lr . The same is true for

gradients r⇢a (x, t), time-derivatives @t ⇢a (x, t), etc. In the case of Boltzmann equation the

coarse-grained stress tensor has the form T` = ⇢vv + P I ⌘[(rv) + (rv)> ] for all ` in the

range mf ⌧ ` ⌧ Lr . The dynamics is the same for all ` in this range.

In turbulence, however, all of the quantities such as v̄` , v`0 , rv̄` , etc. change significantly

throughout the whole range (Lr =)⌘ ⌧ ` ⌧ L. For example, in K41 theory v`0 ⇠ (h"i`)1/3

and keeps growing as ` increases. We say that turbulent flows have a continuous spectrum of

23
excitations in the range ⌘ ⌧ ` ⌧ L. Among other things, this means that there is no law of

large numbers in turbulence and v̄` is fluctuating, with dynamics that depends stochastically

on unknown subscale modes v`0 . The turbulent subscale stress for ⌘ ⌧ ` ⌧ L scales as

⌧ ` ⇠ ( v(`))2 ⇠ (h"i`)2/3 (for K41)

and is thus a monotonically increasing function of length-scale `. Moreover, the turbulent stress

is much more essentially stochastic. To see this, note that the turbulent stress at length-scale

` is not uniquely determined from the stress at scale `0 < ` but instead (Germano, 1992):

.
⌧ ` (v, v) = (⌧ `0 (v, v))` + ⌧ ` (v`0 , v`0 )

To obtain the stress ⌧ ` requires knowledge not only of ⌧ `0 but also of the velocity field v`0

resolved down to length-scale `0 , which is unknown given only v` .

In fact, however, we shall see that scale-separation is vitiated even in laminar flows by

thermal fluctuations, so that gradients r¯

⇢a,` (x, t) become `-dependent! In that case, the coarse-

grained stress tensor has the form T` = ⇢vv +P I ⌘` [(rv` )+(rv` )> ]+⌧ 0 with ⌧ 0 the thermal

noise arising as a central limit theorem correction to the law of large numbers and viscosity ⌘`

becomes `-dependent. The distinction between turbulent and laminar is thus not so clear-cut.

24