(D) Estimating General Terms in the E↵ective Equations

We now return to the e↵ective large-scale equations

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

It is important to be able to estimate the magnitude of the various terms that appear here. We

start with f` . Note that f̄` = f f`0 and

Z
f`0 = dd r G` (r) f (r)

Z
=) |f`0 |  dd r G` (r)| f (r)|
Z
 (sup |rf (x)|) dd r G` (r)|r|, assuming f is smooth and Taylor expanding!
x Z
= C`krf k1 with C = dd ⇢G(⇢)|⇢|

Thus, with a smooth (large-scale) external force, we see that

|f |
f̄` ⇠ f
= f when ` ⌧ Lr = |rf | = gradient-length of the force

We have already considered ⌫4v̄` , but we now obtain an improved estimate:

R
⌫4v̄` (x) = `12 dd r (4G)` (r) v(r; x)

=) (with compactly-supported filter)

⌫4v̄` = O( ⌫ v(`)
`2
)

Before we had ⌫4v̄` = O(⌫kvk2 /`2 ) but the above estimate is much smaller for ⌘ ⌧ ` ⌧ L.

Now consider the subscale force f`s = r · ⌧ ` , using the identity

⇢Z Z Z
1
f`is = d
d r (@j G)` vi (r) vj (r) d
d r (@j G)` (r) vi (r) dd r0 G` (r0 ) vj (r0 )
`

which is easily verified (using r · v = 0). From this, it follows that

v 2 (`)
f`s = O( ` ).

It is useful for later purposes to note a more general technique to derive such an identity

for @k ⌧ij , using the fact that

25
 Z
⌧` (vi , vj )(x + a, t) = dd r Ga,` (r) vi (r, t) vj (r, t)
Z Z
d r Ga,` (r) vi (r, t) dd r0 Ga,` (r0 ) vj (r0 , t)
d
(⇤)

with kernel Ga,` centered at point a:

Ga,` (r) = ` d G( r a ).
`

By expanding to 1st-order in a, we derive

⇢Z
1
@k ⌧` (vi , vj ) = dd r(@k G)` (r) vi (r) vj (r)
` Z Z
d r(@k G)` (r) vi (r) dd r0 G` (r0 ) vj (r0 )
d
Z Z
dd rG` (r) vi (r) dd r0 (@k G)` (r0 ) vj (r0 ) .

Summing over k = j and using incompressibility r · v = 0, then gives the identity on the

previous page for f`s .

We have only to justify the identity (*). We first note that by its definition
Z
⌧` (vi , vj )(x + a) = dd r G` (r)[vi (x + a + r) vi (x + a)][vj (x + a + r) vj (x + a)]
Z Z
d r G` (r)[vi (x + a + r) vi (x + a)] dd r G` (r)[vj (x + a + r)
d
vj (x + a)]

Making the change of variables r + a ! r gives

Z
⌧` (vi , vj )(x + a) = dd r Ga,` (r)[ vi (r; x) vi (a; x)][ vj (r; x) vj (a; x)]
Z Z
dd r Ga,` (r)[ vi (r; x) vi (a; x)] dd r Ga,` (r)[ vj (r; x) vj (a; x)],

since

v(x + r) v(x + a) = [v(x + r) v(x)] [v(x + a) v(x)]

= v(r; x) v(a; x).

However, v(a; x) does not depend upon r and is thus a “constant” with respect to the average

over r with density Ga,` (r). Since cumulants are invariant to shifts of the random variables by

26
constants, this yields the formula (*). We shall use this “shift-trick” again later for higher-order

terms.

We now see indeed that the viscous term is negligible with respect to the subscale force

term, for ` such that

v(`)`
Re` ⌘ ⌫ 1.

Within K41 theory, this holds whenever ` ⌘.

Now consider the large-scale pressure gradient rp̄` . It is very easy to derive
R
rp̄` = 1` (rG)` (r) p(r)dd r

from which
p(`)
rp̄` = O( ` )

with p(`; x) = sup|r|` | p(r; x)|. However, this is not so useful, since we do not know the

magnitude of p(`)! It is more instructive to use the relation

R
4p̄` = `12 dd r (@i @j G)` (r) vi (r) vj (r)

which follows by filtering the Poisson equation 4p = @i @j (vi vj ). From this, we get
v 2 (`)
4p̄` = O( `2
).

This suggests that

p(`) = O( v 2 (`)) (F)

which would imply that

v 2 (`)
rp̄` = O( ` ),

and thus has the same order of magnitude as the subscale force term. In fact, it is possible to

prove various rigorous forms of (F). This will involve a bit of mathematics more sophisticated

than any that we have used up to now (or will later!) This is not surprising — the pressure is

one of the most mysterious and challenging objects in the study of incompressible turbulence!

As an important example of a rigorous form of (F), we show that

k v(r)k2q = O(|r|s ) =) k p(r)kq = O(|r|2s )

i.e. if v is Besov regular of order (2q) with exponent s, then p is Besov regular of order q with

27
exponent (2s). To show this, we need to use an equivalent formulation of Besov regularity in

terms of band-pass filtered fields, a so-called “Paley-Littlewood criterion”. It is known that for

p 1 and 0 < s < 1 that

k f (r)kp = O(|r|s ), 8|r|  r0 () kf [n] kp = O(2 ns

), 8n n0

where

f [n] = f¯`n f¯`n 1 , `n = 2 n

`0

is the band-pass filtered function. Notice that we have already proved the =) direction, since

f [n] = f`0n 1
f`0n ! For a discussion of Besov regularity from this point of view, see

M. Frazier, B. Jawerth & G. Weiss, Littlewood-Paley Theory and the Study of

Function Spaces (Amer. Math. Soc., Providence, RI, 1991), Chapter 5.

We shall use this result to establish first the Besov regularity of

⌘ 4p.

We have seen that

Z
1
¯` = dd r (@i @j G)` (r) vi (r) vj (r)
`2 Z
1
=) k¯` kq  dd r |(@i @j G)` (r)| k v(r)k22q by Hölder inequality
`2

Then, if k v(r)k2q = O(|r|s ), it easily follows that

k¯` kq = O(`2s 2)

But then

k [n] k = k¯`n ¯`n 1 kq  k¯`n kq + k¯`n 1 kq = O(`2s 2)

q n

By the Paley-Littlewood criterion, this implies that is Besov regular of order q with exponent

2s 2. But now notice that

4p =

where has the stated Besov regularity. It then follows by elliptic regularization that the

solution p is Besov regular of order q with exponent

28
 (2s 2) + 2 = 2s,

i.e. inverting 4 adds two orders of derivatives! E.g. see H. Triebel, Theory of Function Spaces

(Birkhäuser Verlag, Basel, 1983), Section 4.2.4.

If we define pressure structure functions

Sq(p) (r) = h| p(r)|q i

= k p(r)kqq

and corresponding scaling exponents

(p) |r| ⇣q (p)
Sq (r) ⇠ u2q
rms ( L )
(u)
analogous to those defined before for the velocity, ⇣q , then the previous result implies that
(p) (u)
⇣q ⇣2q .

In K41 theory one would expect that

(p) (u)
⇣q = ⇣2q = 23 q.

Experimental Data on the Energy & Pressure Spectra from Tsuji & Ishihara (2003). The data were

taken on the center line of a free jet with 200  Re  1200. The measurement of pressure fluctuation

in the flow field was accomplished with a small piezoresistive transducer (low-k) and a quarter-inch

condenser microphone (high-k).

29
The K41 prediction was by

A. M. Obukhov, “Pressure fluctuations in a turbulent flow,” Dokl. Akad. Nauk, SSSR, 66

(1), 17-20(1949).

A. M. Yaglom, “Acceleration field in a turbulent flow,” Dokl. Akad. Nauk, SSSR, 67 (5),

795-798(1949).

G. K. Batchelor, “Pressure fluctuations in isotropic turbulence,” Prov. Camb. Phil. Soc.

47 (2) 359-374(1951).

Numerical simulations give some evidence of an intermittency correction:

Plots of Spectrum of the Pressure Laplacian 4p.

More precisely, normalized spectra k 4 P (k)/(⌫ 5

h"i4 )1/4 versus k⌘ in 2563 ,5123 , 10243 ,20483 DNS. The
5/3
inset shows the compensated spectra multiplied by (k⌘) , for 0.001 < k⌘ < 0.1; the straight line

shows the slope / k 1.81 5/3

. From T. Ishihara et al., J. Phys. Soc. Japan, vol. 72, pp. 983-986 (2003).

We now consider the last two terms in the large-scale equation, namely, @t v̄` and (v̄` · r)v̄` .

Note that

v̄` = v v`0 = v + O( v(`))

and v(`) ! 0 for ` ⌧ L. For example, in K41 theory, v(`) ⇠ ("`)1/3 . Thus,

v̄` ⇡ v, ` ⌧ L

30
and
v(`)
v̄` · rv̄` = O(umax · ` ).

This is much larger than any of the terms in the equation that we have previously examined,

except possibly f̄` . Hence, if f̄` = 0 (as in most natural flows!), then the only term that can

balance it is @t v̄` , so that @t v̄` ⇡ (v̄` · r)v̄` and

v(`)
@t v̄` = O(umax · ` ).

The combined term, or (large-scale) Lagrangian time-derivative

D̄t v̄` = @t v̄` + (v̄` · r)v̄`

must have the magnitude of the remaining terms

v 2 (`)
D̄t v̄` = O( ` ).

This implies that the Lagrangian time-scale at length ` must be of the order of t` = `/ v(`),

the so-called local eddy-turnover time.

A summary of all these estimates is

v(`)
@t v̄` = O(umax ` )
v(`)
(v̄` · r)v̄` = O(umax ` )
v 2 (`)
D̄t v̄` = O( ` )
v 2 (`)
f`s = r · ⌧ ` = O( ` )
p(`) v 2 (`)
rp̄` = O( ` ) = O( ` )
v(`) v 2 (`)
⌫4v̄` = O(⌫ `2
) = O( ` · Re` 1 )

f̄` = O(fmax ) [if present]

These may be compared with the estimates developed by Tennekes & Lumley, Ch.2. Note that

their ` always corresponds to the integral length-scale or L, while we have derived estimates for

all `  L. Also,

` ⇡ L =) v ⇡ umax ⇡ urms

31
 Estimation of Terms in Energy Balances

We now estimate the various terms in both the large-scale and small-scale energy balances.

(i) Large-scale Energy Balance

See T & L, Section 3.1

We may write the large-scale balance as

D̄t ē` + (v̄` · r)p̄` + r · (⌧ ` · v̄` ) ⌫4ē` = ⇧` ⌫|rv̄` |2 + v̄` · f̄`

We have already estimated all of the terms on the righthand side, except the forcing term

v̄ · f̄` = O(umax fmax )

The terms on the lefthand side can be estimated from our previous work. It is not hard to

check that the largest contributions to the four lefthand terms (from the left to right) are from

v̄` · (D̄t v̄` ), v̄` · (rp̄` ), (r · ⌧ ` ) · v̄` , (⌫4v̄` ) · v̄` ,

respectively. Of course, these four contributions sum up to give v̄` · f̄` exactly, since D̄t v̄` +

rp̄` + r · ⌧ ` ⌫4v̄` = f̄` . We see then that the following estimates hold:
v)2
D̄t ē` = O(umax ( ` )
v)2
(v̄` · r)p̄` = O(umax ( ` )
v)2
r · (⌧ ` · v̄` ) = O(umax ( ` )
v)2
⌫4ē` = O(umax ⌫`2v ) = O(umax ( ` · Re` 1 )
v)3
⇧` = O( ( ` )
2 v)3
⌫|rv̄` |2 = O(⌫ ( `v)
2 ) = O(
(
` · Re` 1 )

v̄` · f̄` = O(umax fmax ).

Integrating over space and assuming that there is no transport of energy across the boundary,

the space transport terms must all vanish (including (v̄` · r)ē` ) so that the dominant terms

must be in the set

@t ē` ⇠
= @t e, ⇧` , v̄` · f̄` ⇠
= v · f.

Which terms give the dominant balance depends upon the precise situation:

32
 In steady-state forced turbulence, taking time-averages

h@t eitime = 0, h⇧` itime ⇠

= hv · f itime (independent of `!)

In homogenous decaying turbulence, taking global space average

h@t eispace ⇠
= h⇧` ispace (again independent of `!)

In other cases, the balance may be di↵erent. For example, in steady-state pipe flow, there is

space transport of energy into each cross-section of the pipe, by the pressure head. Taking

an average over time and also over space from a cross-section at point 1 to a cross-section at

another point 2 downstream gives

⌫h|rv̄` |2 ispace time + h⇧` ispace time

⇠
= [hpuicross section 1 & time hpuicross section 2 & time ] /L12

where L12 is the axial distance between points 1 and 2. Again one finds a range of length-scales

L ` ⌘ where the energy flux is constant (note that L depends on the distance to the pipe

wall!) Of course, for ` ⇡ ⌘ the term ⌫|rv̄` |2 which we have neglected becomes significant and

the mean flux is no longer constant. In the case of pipe flow, there is always some region close

the side wall where this term cannot be neglected. E.g. see Figure 5.5, in T & L, Section 5.2.

(ii) Small-scale energy balance & stress production

See T & L, Section 3.2

We now estimate the terms in the small-scale energy balance, which may be written as

1
D̄t k` + @i ⌧` (p, vi ) + @i ⌧` (vj , vj , vi )
2
= +⇧` ⌫⌧` (vi,j , vi,j ) + ⌧` (vi , fi )
| {z } | {z }
"0` Q0`

In fact, we shall do something more general and estimate the terms in the small-scale stress

production equation

D̄t ⌧` (vi , vj ) + [@i ⌧` (p, vj ) + @j ⌧` (p, vi )] + @k ⌧` (vi , vj , vk )

= [v̄i,k ⌧` (vk , vj ) + ⌧` (vi , vk )v̄j,k ] ; stress production by large-scale strain

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

33
 2⌫⌧` (vi,k , vj,k ) ; viscous destruction of stress

+[⌧` (vi , fj ) + ⌧` (vj , fi )] ; stress production by large-scale force

Some of the terms are easy to estimate by our previous techniques:

p v ( v)3
[@i ⌧` (p, vj ) + @j ⌧` (p, vi )] = O( ) = O( )
` `
v ( v)3
[v̄i,k ⌧` (vk , vj ) + v̄j,k ⌧` (vk , vi )] = O( · ( v)2 ) = O( )
` `
[⌧` (vi , fj ) + ⌧` (vj , fi )] = O( v f )

If f is a smooth (large-scale) force, then f (`) = O(`krf k1 ). This term is negligible compared

to the others if ` is small enough that

` 1/2
t` = v(`) ⌧ krf k1 = Tf = time-scale of large-scale force

The next term that we consider is the space-transport of stress by triple-correlation @k ⌧` (vi , vj , vk ).

Here it is useful to note the identity

⌧` (vi , vj , vk ) = h vi vj vk i` h vi vj i ` h vk i `

h vi vk i ` h vj i ` h vj vk i ` h vi i `

+2h vi i` h vj i` h vk i`

where h.i` is the average over the seperation vector r with respect to the filter kernel G` (r).

This identity can be verified by a direct computation. However, it is useful to give a more

general derivation. Suppose that {fi |i 2 I} are any set of fields. Note that
Z
(fi1 . . . fip )` (x) = dd r G` (r)fi1 (x + r) . . . fip (x + r)

= h( fi1 ) . . . ( fip )i` (x)

where

( fi )(x) = fi (x + r)

is the shift operator. We thus see that fi1 . . . fip is a correlation function of the “random

variables” fi1 ,..., fip . Likewise, the generalized central moments defined by Germano are the

connected correlation functions of fi1 ,..., fip :

34
 ⌧` (fi1 , . . . , fip ) = h( fi1 ) . . . ( fip )ic` .

There is a powerful method in statistical mechanics for computing these objects using so-called

generating functions. The generating function for the correlation function is

X
Z` (↵) = hexp( ↵i fi )i` .
i2I

It is easy to check that

@p
(fi1 . . . fip )` = Z (↵)|↵=0
@↵i1 . . . @↵ip `
The corresponding cumulants or connected correlation functions are generated by the function

W` (↵) = ln Z` (↵),

i.e.
@p
⌧` (fi1 . . . fip ) = @↵i1 ...@↵ip W` (↵)|↵=0 .

This is the so-called linked cluster-theorem. See

K. Huang, Statistical Mechanics, 2nd Ed. John Wiley & Sons, NY, 1987, Section 10.1.

On the other hand, rather than the shift fields, one can consider instead the increment fields

r fi (x) = r fi (x) fi (x).

The correlation functions of the increments h fi1 . . . fip i` are generated by the function

X
Z` (↵) = hexp( ↵i fi )i` ,
i2I

and the connected correlation functions by the function

W` (↵) = ln Z` (↵),

i.e.
@p
h( fi1 ) . . . ( fip )ic` = @↵i1 ...@↵ip W` (↵)|↵=0 ,

again by the linked-cluster theorem. Now comes the key observation: since fi (x) does not

depend on r, it can be taken outside the average h.i` . Thus, using fi = fi fi ,

X
Z` (↵) = hexp( ↵i fi )i` ,
i2I
X X
= hexp( ↵i fi )i` exp( ↵i fi )
i2I i2I

35
 X
= Z` (↵) exp( ↵i fi )
i2I

Taking the logarithm of both sides then gives

X
W` (↵) = W` (↵) ↵i fi
i2I

Since they di↵er only by a term linear in ↵, we can draw our main conclusion:

Proposition: The connected correlation functions of fi and fi are related for

p = 1 by

h fi i ` = h fi i ` fi

and for p > 1 are equal

h fi1 . . . fip ic` = h fi1 . . . fip ic

In terms of the quantities defined by Germano (1992), this means that

(fi )0` = fi (fi )` = h fi i `

and, for p > 1,

⌧` (fi1 , . . . , fip ) = h fi1 . . . fip ic` .

For example, for p = 2

⌧` (fi , fj ) = h fi fj i` h fi i ` h fj i `

and for p = 3

⌧` (fi , fj , fk ) = h fi fj fk i ` h fi fj i ` h fk i `

h fi fk i ` h fj i ` h fj fk i ` h fi i `

+2h fi i` h fj i` h fk i` , etc.!!!

After this somewhat lengthy interlude, we conclude that

⌧ (vi , vj , vk ) = O(( v)3 ).

36
Furthermore, using the same “shift trick” as for the 2nd-order term
⇢Z
1
@m ⌧` (vi , vj , vk ) = dd r(@m G)` (r) vi (r) vj (r) vk (r)
Z ` Z
d r(@m G)` (r) vi (r) vj (r) dd r0 G` (r0 ) vk (r0 )
d
Z Z
+ dd rG` (r) vi (r) vj (r) dd r0 (@m G)` (r0 ) vk (r0 )

+ cyclic permutations among i, j, k

Z Z Z
d 0 0 0
2 d
d r(@m G)` (r) vi (r) d r G` (r ) vj (r ) dd r00 G` (r00 ) vk (r00 )

+ cyclic permutations among i, j, k

From this one obtains

v)3
@k ⌧ (vi , vj , vk ) = O( ( ` ) ...finally!

The next term that we examine in the stress balance is the viscous destruction term:

"0`ij = 2⌫⌧` (vi,k , vj,k )

This satisfies matrix positivity:

"0` 0

The Cauchy-Schwarz inequality also implies that

q
"0`ij  2⌫ ⌧` (vi,k , vi,k )⌧` (vj,k , vj,k
q
= "0`ii "0`jj

P
Note that Tr("0` ) = i "0`ii = "0` is the subscale dissipation. Thus, in particular,

"0`ij  "0` for all i, j.

Aside from these rigorous bounds, it is difficult to develop exact estimates for "0` , because rv is

a dissipation-range variable which does not remain Hölder continuous/Besov regular ( or even

Lp !) in the limit as ⌫ ! 0. We can get some additional insight, however, by decomposing the

viscous destruction into two parts:

"0`,ij = 2⌫(vi,k vj,k )` 2⌫(v̄i,k )` (v̄j,k )`

Since rv̄` = O( v(`)/`), we can see that the second term can be estimated as

37
 v 2 (`) v 3 (`)
2⌫rv̄` (rv̄` )> = O(⌫ `2
) = O( ` · Re` 1 )
v(`)`
with Re` = ⌫ . This is small compared with the other terms in the stress production

equation. However, the first term is expected to be bigger and, in fact, it is expected that
v 3 (`)
2⌫(vi,k vj,k )` ⇠ (const.) `

For the trace, i.e. for the viscous dissipation, this is a famous conjecture of [Link]

(1962) the refined similarity hypothesis (RSH)

R v 3 (`)
"` (x) = (2⌫|rv|2 )` = dr G` (r)"(x + r) ⇠ (const.) `

This estimate is certainly consistent with K41, since then |rv|2 ⇠ (h"i/⌫) everywhere, but

also v(`) ⇠ (h"i`)1/3 , so that both sides are proportional to h"i. But RSH is also consistent

with intermittancy, which we discuss later. It is reasonable to extend these ideas also to the

o↵-diagonal terms. At least, it is reasonable to guess that there is an upper bound:

v 3 (`)
"0` = O( ` )

The last — and most difficult! — term to estimate is the pressure-strain correlation

`,ij = 2⌧` (p, Sij )

This term is especially tricky, because it is a mixed quantity, with p inertial-range and Sij

dissipation range. Thus,

` = O( p(`)|S|)
"
= O( v 2 (`) · ( )1/2 )
⌫

However, this upper bound is expected to be a big overestimate. Because p and Sij “live” on

di↵erent length-scales, they are presumably very poorly correlated in space. Thus, the local

average over the region of radius ⇠ ` that defines ` should have substantial cancellations!

Following T & L, section 3.2, we may estimate the correlation coefficient between these term

by the ratio of time-scales

t⌘ v(`)/`
⇢( p(`), S) ⇠ t` ⇠ ("/⌫)1/2
⌧ 1.

We can therefore expect that, in fact,

v 3 (`)
` = O( v 2 (`) · ( ⌫" )1/2 · ⇢( p(`), S)) = O( ` )!

38
 v 3 (`)
Lastly, we note that D̄t ⌧` (vi , vj ) = O( ` ), since it is equal to the sum of all the other terms

and these are, at most, of that magnitude. This is a reasonable estimate of that term since

⌧ij = O( v 2 (`)) and we can expect that D̄`,t = O( v(`)/`). We thus obtain, finally,

v 3 (`) ?
D̄t ⌧` (vi , vj ) = O( )
`

v 3 (`)
@i ⌧` (p, vj ) + @j ⌧` (p, vi ) = O( )
`

v 3 (`)
@k ⌧` (vi , vj , vk ) = O( )
`

v 3 (`)
v̄i,k ⌧` (vk , vj ) + v̄j,k ⌧` (vi,k ) = O( )
`

v 3 (`) ?
2⌧` (p, Sij ) = O( )
`

v 3 (`) ?
2⌫⌧` (vi,k , vj,k ) = O( )
`

⌧` (vi , fj ) + ⌧` (vj , fi ) = O( v(`) f (`))

1
Note that the terms marked with (?) are not completely rigorous upper estimates but only heuristic.

39