III Small-Scale Intermittency & Anomalous Scaling

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

⇣  p/3 for p 3.

The K41 theory assumes the “minimal singularity” sufficient to dissipate energy, or ⇣ = p/3

for all p. However, other possibilities are allowed by the above estimate! In this set of notes we

consider the subject of turbulent scaling laws and their relation to turbulent energy cascade.

(A) A Simple Model of Energy Dissipation: Burgers Equation

In this section we consider a simple 1-dimensional PDE model that has non-vanishing energy

dissipation for Re ! 1 but for which K41 theory fails. It is a useful counterexample! The

model is the 1-dimensional Burgers equation for a velocity field u(x, t):

@t u + u@x u = ⌫@x2 u.

It can also be written as

@t u + @x ( 12 u2 ) = ⌫@x2 u
R
so that it corresponds to a conservation of “momentum” u(x, t)dx. As a simple prototype of

turbulence, it was first proposed by J. M. Burgers, “ A mathematical model illustrating the

theory of turbulence,” Adv. Appl. Mech. 1, 171-199 (1948).

The “energy”
1
R
E(t) = 2 u2 (x, t)dx

is also conserved in the limit ⌫ ! 0, i.e. for “ideal” Burgers equation, since

@t ( 12 u2 ) + @x [ 13 u3 ⌫@x ( 12 u2 )] = ⌫(@x u)2 .

Formally, ⌫(@x u)2 ! 0 as ⌫ ! 0. However, this is NOT what occurs!

Consider a simple exact solution of 1-D burgers:

u(x, t) = 1t [x Lx
L tanh( 2⌫t )].

This seems to have been first written down in

1
 S. I. Soluyan & R. V. Khokhlov, Vestnik. Moscow State. Univ. Phys. Astron. 3

52-61(1961)

and is sometimes called the “Khokhlov saw-tooth solution”. The reason for the term “sawtooth”

is that

1
u⌫ (x, t) ⇠ (x + L) as x ! 1
t
1
u⌫ (x, t) ⇠ (x L) as x ! +1
t

with a discontinuity 4u = (2L)/t across the origin.

Plots of the Khokhlov sawtooth solution at fixed time for various viscosities

In the limit ⌫ ! 0, this becomes 8

>
< (x + L)/t Lx<0
u(x, t) =
>
: (x L)/t 0<xL

2
a function on the interval [ L, L] with u(±L, t) = 0 a sharp discontinuity of size 4u = (2L)/t

at x = 0. Such a discontinuity is called a shock or, in this case, a stationary shock, since it is

located at the same point x = 0 for all time t.

Now it is easy to see that there is nonvanishing mean energy dissipation in the limit that ⌫ ! 0.

For example, at ⌫ = 0, using the above explicit formula, it is easy to see that
Z L
1 2 1 1 2
hu (t)i = u (x, t)dx
2 2L L 2
Z L✓ ◆ ✓ ◆
1 x L 2 1 L 2
= dx =
2L 0 t 6 t

so that
d 1 2 1 L2 (4u)2
h"(t)i = dt 2 hu (t)i = 3 t3 = 12t > 0!!!

Alternatively, one can consider the viscous dissipation

"⌫ (x, t) = ⌫|@x u⌫ (x, t)|2

using
L2
@x u⌫ (x, t) = 1
t 2⌫t2
sech2 ( 2⌫t
Lx
)

so that ⌫ ⌧ L2 /t, with L and t fixed,

L4
"⌫ (x, t) ⇡ 4⌫t4
sech4 ( 2⌫t
Lx
)
L 4
The energy dissipation becomes very large ⇠ ⌫t 4 in a small region of size ⇠ ⌫t/L. Using the

R +1
simple integral 1 sech4 u du = 43 , we again find that
Z L
⌫ 1
h" (t)i = "⌫ (x, t)dx
2L L
⇠ 1 L4 2⌫t 4
= · · · for ⌫ ! 0
2L 4⌫t4 L 3
1L 2 (4u) 2
= 3
or
3t 12t

Again, the limit as ⌫ ! 0 is positive! This is exactly like the observations from experiments &

simulations for real fluids!

3
 Plots of energy dissipation in the Khokhlov sawtooth solution for various viscosities

All of our previous theory applies to this problem. E.g. we may consider the coarse-grained

equation

@t ū` + @x ( 12 ū2` + ⌧` ) = ⌫@x2 ū`

with

1 2
⌧` = [(u )` ū2` ]
2
1
= [h( u)2 i` h ui2` ] = O( u2 (`))
2

The large-scale energy balance is

@t ( 12 ū2` ) + @x [ 13 ū3` + ⌧` ū` ⌫@x ( 12 ū2` )] = ⌧` (@x ū` ) ⌫(@x ū` )2 .

2 u(`)
For fixed ` we see again that ⌫(@x ū` )2 = O( ⌫ `2
), which goes to zero as ⌫ ! 0. Large-scale

“dissipation” must come from

4
 u3 (`)
⌧` @x ū` = O( ` )

from which we can deduce, as before, that

⇣p  p/3 for p 3

However, K41 theory does not work here!8 For the limiting shock profile at ⌫ ! 0:
>
< (x + L)/t Lx<0
u(x, t) = ,
>
: (x L)/t 0 < x  L
periodically extended to R with period 2L,8we can see that (with ` > 0)
>
< `/t if 0 2/ [x, x + `]
u(x + `, t) u(x, t) =
>
: (2L + `)/t if 0 2 [x, x + `]
=)
Z L
p 1
h| u(`)| i = |u(x + `) u(x)|p dx
2L L
` ` ` 2L + ` p
= (1 ) · ( )p + ·( )
2L 8 t 2L t
>
< ( ` )p 0 < p < 1
p 2L
⇠ (4u) for ` ⌧ L
>
: `
2L p 1

Thus, 8
>
< p 0<p<1
⇣p =
>
: 1 p 1
Of course, 1  p/3 for p 3, so that our inequality is verified that ⇣p  p/3 but only for p = 3

is ⇣p = p/3!
1
The problem is that K41 theory assumes that h = 3 at every point of space and that is not

what happens here. Instead, there is one point (x = 0) where

u(`, x) ⇠ 4u ⇠ `0 for all ` < L

and, at every other point,

u(`, x) ⇠ `/t ⇠ `1

for sufficiently small `. This is an extreme example of small-scale intermittency, in which

velocity increments are “spotty” in space, big in some places and small in others.

5
Incidentally, it is known that the above features that we have seen in a simple specific solution

of Burgers equation are, in fact, generic for that dynamics. Except for fields in which @x u > 0

everywhere, the Burgers solutions always develop shocks that become exact discontinuities in

the limit ⌫ ! 0 and these dissipate a finite amount of energy that does not vanish as ⌫ ! 0.

The scaling exponents ⇣p that we determined above are also universal to a wide class of initial

data and forcing schemes.

For more information about Burgers equation and “Burgulence,” see :

J. M. Burgers, The Non-Linear Di↵usion Equation: Asymptotic Solutions &

Statistical Problems, (Springer, Boston, 1974)

W. E. K. Khanin, A. Mazel and Y. Sinai, “Invariant Measures for Burgers Equa-

tion with Stochastic Forcing,” Ann. Math. 151 (3): 817-960 (2000)

and a review article

J. Bec & K. Khanin, “Burgers Turbulence,” Physics Reports 447 1-66 (2007)

Before we leave the topic of Burgers equation, there is one last important remark we wish to

make. Consider again the limiting

8 shock profile
>
< (x + L)/t Lx<0
u(x, t) = periodic on [ L, L]
>
: (x L)/t 0 < x  L
These results are just special cases of a general theory for Burgers equation (and for a whole

class of scalar conservation laws in 1-dimension). It is known that the solution u⌫ (x, t) of

@t u⌫ (x, t) + u⌫ (x, t)@x u⌫ (x, t) = ⌫@x2 u⌫ (x, t)

converges in the sense of distributions u⌫ (x, t) ! u(x, t) as ⌫ ! 0 to a solution of the inviscid

equation

@t u(x, t) + @x ( 12 u2 (x, t)) = 0

which satisfies

@t ( 12 u2 (x, t)) + @x ( 13 u3 (x, t))  0,

6
both in the sense of distributions. It is quite easy to check that this profile satisfies the inviscid

Burgers equation

@t u(x, t) + u(x, t)@x u(x, t) = 0

at every spacetime point (x, t) with x 6= 0. The problem at x = 0 is that a derivative (@x u)

does not even exist in the classical sense. However, it is not hard to check that the limiting

shock profile satisfies everywhere in space and for all ` > 0 the coarse-grained equations.

@t ū` + @x ( 12 u2 )` = 0 (*)

or, with a mere change of notation,

@t ū` + @x ( 12 ū2` + ⌧` ) = 0
1
with ⌧` = 2
2 [(u )` ū2` ], where all derivatives of the coarse-grained variables do exist in a

classical sense. Here the initial data ū` (x, 0) for (*) are also coarse-grained. An equivalent,

more conventional formulation of (*) is obtained by noting that derivatives can be taken to

exist in the sense of distributions and it can easily be verified that

R1 RL 1 2
RL
0 dt L dx[u(x, t)@t '(x, t) + 2 u (x, t)@x '(x, t)] + L dx u(x, 0)'(x, 0) = 0

for all C 1 functions ' on [ L, L] ⇥ [0, 1] with compact support (in time). Thus, u(x, t) satifies

the inviscid Burgers equation in the sense of distributions, with initial data u(x, 0). However,

it is also easy to check that energy is not conserved and, in fact, that

( u)3 (**)
@t ( 12 u2 (x, t)) + @x ( 13 u3 (x, t)) = 12 (x)  0

in the sense of distributions.

Furthermore, for a given initial datum u(x, 0) this distributional solution is unique. Note

that, in general, such “weak” or distributional solutions are not unique and there can be more

than one solution even for the same initial data. A simple example is

u↵ (x, t) = ↵ sign(x)✓(t 2|x|/↵)

which is a weak solution of inviscid Burgers with initial data u↵ (x, 0) ⌘ 0 for any ↵ > 0. It does

not make sense to talk about the solution of the inviscid equation — it is an ill-defined concept!

It is only weak solutions with the energy inequality (**) that are unique. It is expected that

7
any hyperviscous regularized solution

@t u✏ (x, t) + @x ( 12 |u✏ (x, t)|2 ) = ✏( @x2 )p u✏ (x, t), ✏ > 0

with p 1 converges as ✏ ! 0 to the same u(x, t) as for p = 1. See:

E. Tadmor, “Burgers equation with vanishing hyperviscosity,” Comm. [Link].

2(2), 317-324(2004).

However, a dispersive regularization such as the famous Korteweg-de Vries (KDV) equation

@t u✏ (x, t) + @x ( 12 |u✏ (x, t)|2 ) = ✏@x3 u✏ (x, t), ✏ > 0

which describes weakly nonlinear shallow water waves, has a completely di↵erent class of solu-

tions, even for the same initial data u(x, 0). See

P. D. Lax and C. D. Levermore, “The small dispersion limit for the KDV equation,

I-III,” Comm. Pure Appl. Math. 36 253-290; 571-594; 889-829 (1983).

For more discussion of these matters see, e.g.

P. D. Lax, “Shock waves and entropy,” in: Contributions to Nonlinear Functional

Analysis, ed. E.H. Zarantonello (Academic Press, NY, 1971)

R. J. DiPerna, “Measure-valued solutions to conservation laws,” Arch. Rat. Mech.

Anal., 8 223-270(1985).

L. C. Evans, Partial Di↵erential Equations, Graduate Studies in Mathematics, vol.

19 (American Mathematical Society, Providence, RI, 1998)

The issues are especially interesting since it was conjectured by Lars Onsager in 1949 that real

turbulent fluids are described in the limit ⌫ ! 0 ( or Re ! 1) by singular solutions of 3D

incompressible Euler equtions that dissipate energy! See

L. Onsager, “Statistical hydrodynamics,” Nuovo Cimento, 6 279-287 (1949)

We shall discuss this in more detail a bit later...

8
(B) Scaling & Scaling Exponents

See UF, Section 8.4

We have been discussing scaling exponents p

r  L, k v(r)kp ⇠ urms (r/L) p , (?)

or equivalently, ⇣p = p p

r  L, Sp (r) ⇠ uprms (r/L)⇣p , (??)

with Sp (r) = h| v(r)|p i = k v(r)kpp , the pth-order structure functions. (To be precise, this is

the pth-order absolute structure function; see below). However, we have not yet given a precise

definition of p or ⇣p . In particular, there is no proof from first principles that scaling laws like

(?) or (??) even hold! We would like to make definitions of p and ⇣p that are guaranteed to

exist, at least when v 2 Lp .

A precise statement of (?) is that

ln k v(r)kp
lim = p
r!0 ln(r/L)

This means that a plot of ln k v(r)kp vs. ln(r/L) should better approximate a straight line for

r  L and getting smaller. Unfortunately, this limit is not guaranteed to exist! Instead, we can

consider
ln k v(r)kp
inf
r<` ln(r/L)

which is a function of ` that is increasing. Thus, its limit as ` ! 0 is guaranteed to exist

(possibly = +1), a quantity which in mathematics is called the limit-infimum and denoted by

limr!0 inf. We therefore define

ln k v(r)kp
p ⌘ lim inf .
r!0 ln(r/L)

[By the way, this is properly called the lower exponent of order p and denoted by p. There

is also an upper exponent ¯p obtained by replacing “inf” with “sup”. This limit also exists,

because now the function of ` is decreasing, and is called the limit-supremum. However, for

reasons discussed just below, we refer to use the limit infimum.]

9
What does this definition mean? Unwrapping it, we see that 8✏ > 0, 9 > 0, such that

ln k v(r)kp
`< =) p ✏  inf  p +✏
r<` ln(r/L)

In other words, for all ` < ,

ln k v(r)kp ln k v(r)kp
9r  `, ln(r/L)  p + ✏ and 8r  `, ln(r/L) p ✏

or, again, for all ` <

9r  `, k v(r)kp ( Lr ) p +✏ and 8r  `, k v(r)kp  ( Lr ) p ✏

Note that ln(r/L) < 0, since r < L. Now, the second inequality above means that v is Besov

regular of order p with exponent p ✏, or that

8
>
> space of functions in Lp
>
>
<
✏,1
v 2 Bp p = with k v(r)kp  (const.)(r/L) p ✏
>
>
>
>
: for some constant.

On the other hand, the first inequality above means that v is NOT Besov regular of order p

with exponent p + 2✏ (for example). Thus, we see that, because ✏ > 0 was arbitrary,

ln k v(r)kp
p ⌘ lim inf = maximal Besov exponent of order p.
r!0 ln(r/L)
In other words, v 2 Bp ,1 for any < p / Bp ,1 for any
and v 2 > p. We sometimes just call

this the Besov exponent of v of order p, since it is uniquely defined whenever v 2 Lp .

ln Sp (r)
In the same manner, we can define with ⇣p (`) := inf |r|<` ln(r/L)

ln Sp (r)
⇣p = lim ⇣p (`) = lim inf
`!0 r!0 ln(r/L)

so that ⇣p = p p. This definition is satisfactory if we take the limit ⌫ ! 0 first (assuming

that limit exists!) In that case, an infinitely long inertial range exists with singular velocity

fields. However, for finite ⌫ > 0, p = 1 and ⇣p = p for all p if — as seems plausible for the

Navier-Stokes solution — the velocity is smooth1 . In that case, k v(r)kp ⇠ (const.)(r/L) for

sufficiently small r.
1
In fact, it is only proved for the Navier-Stokes solution that p = 1 for p  2!

10
Instead, one can hope to get scaling with p < 1 or ⇣p < p only in an intermediate asymptotic

sense. That is, there should exist a small length-scale

⌘p = (const.)L(Re) ↵p , ↵p > 0

such that
ln k v(r)kp
inf r<` ln(r/L) ! p

better and better for decreasing ` in the range ⌘p  `  L. Here it is key that this intermediate

range should be able to be made longer and longer as Re ! 1.

Defining ⇣p and p as above, either by taking the limit Re ! 1 first or else by intermediate

asymptotics for Re 1, we can then establish some important general properties:

Proposition 1: ⇣p is a concave function of p 2 [0, 1), i.e. for all t, 0  t  1,

⇣tp+(1 t)p0 t⇣p + (1 t)⇣p0 for p, p0 0

Proof: For simplicity we will give the proof for Re ! 1 first. Note that

t)p0
Stp+(1 t)p0 (r) = h| v(r)|tp+(1 i
t)p0
= h| v(r)|tp · | v(r)|(1 i
tp· 1t t)p0 · 1 1 t 1 t
 h| v(r)| it · h| v(r)|(1 i by Hölder inequality
0
= h| v(r)|p it · h| v(r)|p i(1 t)

= [Sp (r)]t [Sp0 (r)](1 t)

Thus, for r < L,

ln Stp+(1 t)p0 (r) ln Sp (r) ln Sp0 (r)
ln(r/L) t· ln(r/L) + (1 t) · ln(r/l)

Taking the limit as r ! 0 of both sides gives the result ⇣tp+(1 t)p0 t⇣p + (1 t)⇣p0 . QED!

One important consequence is the following:

Corollary 1: The exponent p = ⇣p /p is non-increasing in p.

Proof: Note first that

⇣0 = 0

since S0 (r) = h| v(r)|0 i = h1i = 1! Thus, take any p0 p and write

11
 p p
p = (1 p0 ) ·0+ p0 · p0

so that
p
⇣p (1 )⇣0 + pp0 ⇣p0
p0
| {z }
=0
or
⇣p ⇣p0
p = p p0 = p0 QED

On the other hand, we have

Proposition 2: If v is bounded, then ⇣p is non-decreasing in p.

Proof: if v 2 L1 , then for all x 2 V ,

| v(r; x)|  2kvk1

Taking any p0 p,

0 0
h| v(r)|p i = h| v(r)|p p
· | v(r)|p i
0
 (2kvk1 )p p
· h| v(r)|p i

so that
0
lnh| v(r)|p i lnh| v(r)|p i ln(2kvk1 )
+ (p0 p)
ln(r/L) ln(r/L) ln(r/L)

The last term vanishes for r ! 0, so that taking the lim inf as r ! 0 of both sides gives

⇣p 0 ⇣p QED.

Again, there is a corollory for p:

Corollary 2: If v is bounded, then p 0 for all p > 0.

⇣p
Proof: For p > 0, ⇣p ⇣0 = 0. Thus, p = p 0. QED

It follows from Corollaries 1 & 2 that

lim p = 1 0
p!1

must exist, if v is bounded. It is more usual to denote

hmin = 1,

12
which is called the minimum Hölder exponent of v in V . It is the most singular behavior of

the flow velocity v that determines hmin8

. For example, for Burgers equation
>
< 1 0<p<1
p =
>
: 1/p p 1
so that
1
hmin = lim = 0
p!1 p

corresponding to shock discontinuities in u(x, t). The relation p ! hmin can also be stated as

⇣p ⇠ phmin , p ! 1,

at least if hmin > 0.

Note that p and ⇣p can be shown to be nonnegative under an even weaker assumption for each

specific p 1. That is,

8p 1, v 2 Lp =) p 0.

The proof is essentially the same, since the triangle inequality (only valid for p 1!) implies

that

k v(r)kp  2kvkp

and thus
ln k v(r)kp ln(2kvkp )
p = lim inf lim inf =0
r!0 ln(r/L) r!0 ln(r/L)

This means that, in order to get a negative p or ⇣p , it must be that v 2

/ Lp ! In that case,

even the definition of p and ⇣p that we gave above are invalid, since we always assumed that

v 2 Lp . It is possible to give a more general definition that allows negative exponents p or ⇣p ,

for example, by using wavelet coefficients in place of velocity increments. We shall discuss this

a little bit later.

If we go back to the proof that ⇣p is non-decreasing for bounded v, we derived

0 0
h| v(r)|p i  (2kvk1 )p p h| v(r)|p i.

If we assume scaling, so that Sp (r) ⇠ p

= urms (r/L)⇣p for ⌘p ⌧ r ⌧ L, then

0 0
uprms (r/L)⇣p0  (2umax )p p
· uprms (r/L)⇣p

13
 2umax p0
=) (r/L)⇣p0 ⇣p
 ( ) p
urms

If we take ⌘ = max ⌘p , ⌘p0 ! 0 as Re ! 1, then by taking r/L sufficiently small as Re ! 1

we get that
umax
lim = +1
Re!1 urms

whenever ⇣p0 < ⇣p for some p0 > p. Even if we assume that urms < +1 as Re ! 1, we see that

umax ! 1. This is the contrapositive of the statement that v bounded implies ⇣p increasing:

if ⇣p0 < ⇣p for even one value p0 > p, then umax ! 1 as Re ! 1. Uriel Frisch, Section 8.4

argues that incompressible Navier-Stokes equation will not be a physically valid model at very

large Reynolds number if it leads to

⇣p0 < ⇣p , for some p0 > p.

His reasoning is that

umax ! 1 as Re ! 1

in that case, so that the maximum Mach number

umax
M amax = c ! 1,

violating the assumption of small Mach number in the derivation of the incompressible Navier-

Stokes from either molecular dynamics or from compressible flow.

However, the work of Quastel & Yan (1998) cited earlier shows that this argument is incorrect.

They derived the incompressible Navier-Stokes equation rigorously from a stochastic lattice gas

in the limit where a bulk Mach number and Knudsen number are small. Their derivation is

valid even if the Navier-Stokes equation leads to singular Leray solutions which, it is known,

must have umax = +1. Thus, it is proved that incompressible Navier-Stokes (INS) may

be physically valid even if umax = +1. This is related to partial regularity results for the

Leray solutions of the Navier-Stokes equation, which imply that the fractal (actually, parabolic

Hausdor↵) dimension of the singular set is small, much less than the dimension of spacetime. In

particular, the set where umax = +1 is proved to have zero measure (and possibly is empty.)

Thus, the assumptions under which INS are derived — in a distribution sense! — are not

14
violated by singular Leray solutions.

It is worth emphasizing, however, that the proof of Quastel & Yan (1998) does not, in fact,

justify the validity of deterministic Navier-Stokes equations for turbulent flows, in practice.

Their argument is based on the scaling symmetry of the deterministic equations discussed in

Chapter II:

v ! v0 = v, x ! x0 = 1
x, t ! t0 = 2
t

which maps an incompressible Navier-Stokes solution v(x, t) to another solution v0 (x0 , t0 ) with

the same Reynolds number Re0 = Re and with molecular viscosity ⌫ also unchanged. This self-

similarity is broken, however, by thermal noise. For the case of turbulent flows where Taylor’s

relation holds:

" ⇠ U 3 /L ! "0 = 4
"

u⌘ = (⌫")1/4 ! u0⌘ = u⌘ ,

⌘ = ⌫ 3/4 " 1/4

! ⌘0 = 1
⌘

✓⌘ = kB T /⇢u2⌘ ⌘ 3 ! ✓⌘0 = ✓⌘

Because ✓⌘ decreases with , it is possible, in principle, to observe deterministic Navier-Stokes

predictions in the dissipation range of turbulent flows by taking v ! v0 = v, x ! x0 = 1x

with ⌧ 1. However, in practice, must be chosen unrealistically small. If relation x2 ex = 1/✓⌘

is solved for x = kc ⌘, then satisfying it for x0 = 2x, requires 1/ = ✓⌘ /✓⌘0 = 4ex . For example,

in the case of the ABL we argued in Chapter II that thermal noise would be relevant already

at a length-scale ` ⇠ ⌘/11. To make the deterministic predictions valid down to ` ⇠ ⌘/22 would

require that the integral length be made 4e11 = 240, 000 times larger and r.m.s. velocities

240,000 times weaker!

Lastly, let us comment on the special role of K41 among all possible scaling laws

Sp (r) ⇠ Cp uprms ( Lr )⇣p .

Using urms ⇠
= (h"iL)1/3 and setting ⇣p = p/3 + ⇣p , gives

15
 Sp (r) ⇠ Cp (h"ir)p/3 ( Lr ) ⇣p . (?)

Thus we see that K41 with ⇣p = p/3 or ⇣p = 0 for all p 0 is unique in that it is

the only possible scaling in which Sp (r) is independent of L, depending only upon h"i. K41

predictions are often “derived” by dimensional considerations, assuming that h"i is the only

relevant parameter.

More generally, Sp (r) should be expected to depend upon both h"i and L, even for r ⌧ L. This

means that the small-scales “remember” not only the energy flux from the large-scales but also

N = log2 ( Lr )

the number of “cascade steps” (by factors of 2) in going from the length L to the length r.

Intuitively, intermittency corresponds to fluctuations that build up and become larger and

larger as r decreases further below L.

In physics, scaling laws like (?) above are called anomalous because the dimensional scaling

based upon h"i breaks down. K41 scaling implicitly assumes that the limit L/r ! 1 ex-

ists, which need not be true (e.g. Burgers equation!) The exponent ⇣p that describes this

divergence is called an anomalous dimension. Unfortunately, ⇣p cannot be obtained by a

dimensional analysis. The renormalization group is a general tool developed in physics to cal-

culate such anomalous dimensions. See N. Goldenfeld, Lectures on Phase Transitions and the

Renormalization Group (Westview Press, 1992).

16