Physica D 179 (2003) 161182

Invariance principle for inertial-scale behavior of scalar

elds in Kolmogorov-type turbulence
Albert C. Fannjiang

Department of Mathematics, University of California at Davis, Davis, CA 95616, USA

Received 24 August 2002; received in revised form 11 December 2002; accepted 21 January 2003
Communicated by M. Vergassola
Dedicated to George Papanicolaou on the occasion of his 60th birthday
Abstract
We prove limit theorems for small-scale pair dispersion in synthetic velocity elds with power-law spatial spectra and
wavenumber dependent correlation times. These limit theorems are related to a family of generalized Richardsons laws with
a limiting case corresponding to Richardsons t
3
- and 4/3-laws. We also characterize a regime of positive dissipation of passive
scalars.
2003 Elsevier Science B.V. All rights reserved.
Keywords: Invariance principle; Inertial-scale behavior; Kolmogorov-type turbulence
1. Introduction
The celebrated Richardsons t
3
-law [36] states that a pair of particles located at (x
(0)
(t), x
(1)
(t)) R
2d
being
transported in the incompressible turbulence satises
E[x
(1)
(t) x
(0)
(t)[
2
C
R
t
3
for
1
< [x
(1)
(t) x
(0)
(t)[ <
0
, (1)
where is the energy dissipation rate, C
R
the Richardson constant and
0
and
1
are respectively the integral
and viscous scales. Here and below E stands for the expectations w.r.t. the ensemble of the velocity elds. This
law has been conrmed experimentally [22,31,39] and numerically [4,11,18,43]. A stronger statement is that the
relative diffusivity of the tracer particles is proportional to the 4/3 power of their momentary separation, and this
is called Richardsons 4/3-law [36], see also [1,7,29,32]. This paper presents several small-scale limit theorems
(Theorems 13) related to the Richardsons laws for a family of colored-noise-in-time velocity elds that have
Kolmogorov-type spatial spectra and wavenumber dependent correlation times. The other aspect of the scaling limit
concerns the dissipation of the scalar eld in the limit of vanishing molecular diffusion (Corollaries 1 and 2).

Tel.: 1-530-752-2216; fax: 1-530-752-6635.

E-mail address: fannjian@math.ucdavis.edu (A.C. Fannjiang).
0167-2789/03/$ see front matter 2003 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0167-2789(03)00027-7
162 A.C. Fannjiang / Physica D 179 (2003) 161182
The nature of time correlation in fully developed turbulences in the inertial range is not entirely clear (see [30]
and the references therein). But it seems reasonable to assume that, to the leading order, the temporal correlation
structure of the Eulerian velocity eld u(t, x) is determined by the energy-containing velocity components above the
integral scale, consistent with Taylors hypothesis commonly used in the uid ow measurements in the presence
of a mean ow or the random sweeping hypothesis in the absence of a mean ow (see [35,40]). In both cases the
temporal correlation function on the small scales is anisotropic and depends on external forcing. The more robust
features of small-scale turbulence can be revealed by considering the relative velocity eld U(t, x) = u(t, x
x
(0)
(t)) u(t, x
(0)
(t)), with respect to a reference uid particle x
(0)
(t), which tends to preserve invariance properties
of the uid equations. The velocity eld u(t, x x
(0)
(t)) as viewed from a uid particle, which is a useful tool for
turbulence modeling [2,23], is called the quasi-Lagrangian velocity eld in the physics literature and is an example
of the general notion of the Lagrangian environment process [15,33,34].
We assume [13,16] that the two-time structure function of U(t, x) has the power-law form
E[U(t, x) U(t, y)] [U(s, x) U(s, y)]
=

R
d
2[1 cos (k (x y))] exp(a[k[
2
[t s[)E
(
1
,
0
)
(, k)[k[
1d
dk,
(1, 2), > 0, a > 0 (2)
with the energy spectrum
E
(
1
,
0
)
(, k) =

E
0
(I k k[k[
2
)[k[
12
for [k[ (
1
0
,
1
1
),
0 for [k[ / (
1
0
,
1
1
),

0
< ,
1
> 0, E
0
> 0, (3)
where
1
and
0
are respectively the viscous and integral scales. The assumed temporally stationary vector eld
U(t, x) has homogeneous spatial increments and its expectation E
s
[U(t, x)], conditioning on the events up to time
s < t, is assumed to admit the spectral representation
E
s
[U(t, x) U(t, y)] =

R
[1 exp(ik (x y))] exp(a[k[
2
[t s[)

U(s, dk), s < t, (4)

where

U(t, k) is a time-stationary process with uncorrelated increments over k such that
E[

U(t, dk)

(t, dk
/
)] = E
(
1
,
0
)
(, k)(k k
/
) dk dk
/
t, k, k
/
. (5)
The exponential form of the temporal correlation in (2) and (4) is not important for us; it can be replaced by a more
general one like
(a[k[
2
[t s[)
with an integrable function () decaying to zero as . Since the exponential form seems to agree well with
the Lagrangian measurements (see [37] for the Reynolds number around 100 and [41] for high Reynolds numbers)
we will use it for the sake of simplicity.
Set the rescaled velocity
U

(t, x)
1
U(
2
t, x). (6)
Then U

(t, x) has the energy spectrum

E
(
1

1
,
0

1
)
(, k) =

E
0
(I k k[k[
2
)[k[
12
for [k[ (
1
0
,
1
1
),
0 else.
(7)
A.C. Fannjiang / Physica D 179 (2003) 161182 163
However, we do not assume in this paper the full scale-invariance, namely,
U

(t, x)
d
=U(t, x) for
1
= 0,
0
= , (8)
where =
d
means the identity of the distributions. Instead, we assume the weaker assumption of the 4th order
scale-invariance, i.e. that up to the 4th moments of the velocity eld can be estimated in term of the energy spectrum
as in the case of Gaussian elds.
The viscous and integral scales
1
and
0
can be related to each other via the Reynolds number Re as

1
Re
1/(42)
by using the positivity of kinetic energy dissipation of uid in the limit Re . The correlation time a
1
[k[
2
decreases as the wavenumber k increases. The spatial Hurst exponent of the velocity equals 1 in the inertial range
(
1
,
0
). It should be noted that because of the temporal stationarity of the Lagrangian eld u(t, xx
(0)
(t)) [15,42],
U(t, x) has the same one-time statistics as the Eulerian velocity u(t, x); in particular they share the same energy
spectrum, but their multiple-time statistics are usually different. We could work with the modied von Karman
spectrum but it is irrelevant for our purpose since we are concerned with transport in the inertial-convective range.
It is convenient to express the coefcients E
0
, a in terms of U
0
, the root mean-square longitudinal velocity
increment over the integral length
0
, as
E
0
C

U
2
0

22
0
, a c
0

21
0
U
0
as

0

1
(9)
with dimensionless constants c
0
and
C

=
(4)
d/2
2
23
(2 2)( d/2)
(d 1)(2 )
, (10)
where (r) is the Gamma function.
Assuming that the lifetime (i.e. correlation time (k) = a
1
[k[
2
) of eddy of size [k[
1
is same as its turnover
time one gets the relation
2 = 2. (11)
Assuming that the energy ux given by E
(
1
,
0
)
[k[/(k) is constant across the scales in the inertial range one gets
the relation
= 1. (12)
The values of parameters satisfying both Eqs. (11) and (12) correspond to the Kolmogorov spectrum with = 4/3,
= 1/3. For the Kolmogorov spectrum, one has the expression, by estimating by U
3
0

1
0
,
E
0
C

2/3
, a c
0

1/3
. (13)
Writing x(t) = x
(1)
(t)x
(0)
(t) and adding the molecular diffusivity we have the following It os stochastic equation
for the pair separation x(t)
dx(t) = [u(t, x
(0)
(t) x(t)) u(t, x
(0)
(t))] dt

dw(t) = U(t, x(t)) dt

dw(t),
where w(t) is the standard Brownian motion in R
d
. It is also useful to consider the associated backward stochastic
ow which is the solution of the backward stochastic differential equation
d
t
s
(x) = U(s,
t
s
(x)) ds

dw(t), 0 s t, (14)
164 A.C. Fannjiang / Physica D 179 (2003) 161182

t
t
(x) = x. (15)
Denote by Mthe expectation with respect to the molecular diffusion and consider the scalar eld T(t, x)
T(t, x) M[T
0
(
t
0
(x))], (16)
which satises the advectiondiffusion equation
T(t, x)
t
= U(t, x) T(t, x)

2
T(t, x), T(0, x) = T
0
(x). (17)
We interpret Eq. (17) in the weak sense
!T(t, ), ) !T
0
, ) =

2

t
0
!T(s, ), ) ds

t
0
!T(s, ), V(s, ) ) ds (18)
for any test function C

c
(R
d
), the space of smooth functions with compact supports.
To study the small-scale behavior we introduce the following scaling limit. First we assume that the integral and
viscous scales of the eld U are
0
= L,
1
= /K with L, K tending to in a way to be specied later. Then
we re-scale the variables x x, t
2q
t amounting to consider the re-scaled pair separation
x

(t) =
1
x(
2q
t).
The scaling parameter will tend to zero, indicating that we are considering the emergent inertial range of scales

1
< [x[ <
0
(since K, L ) as a result of a large Reynolds number. We also set
=
22q
with = (). (19)
After re-scaling, the advectiondiffusion equation becomes
T

t
=
2q1
U(
2q
t, x) T

2
T

. (20)
We take the initial data T

(0, x) = T
0
(x) L

(R
d
) L
2
(R
d
). Let
V(t, x) =
1
U(
2
t, x).
As before (cf. (7)) the energy spectrum of the rescaled eld V is given by
E
K,L
(, k) =

E
0
(I k k[k[
2
)[k[
12
for [k[ (L
1
, K),
0 else.
We rewrite Eq. (20) in terms of V as
T

t
=
2q2
V(
2(q)
t, x) T

2
T

, T

(0, x) = T
0
(x). (21)
A simple, non-trivial scaling limit is the white-noise limit when
q < (22)
and
q = 2 (23)
resulting from equating 2q 2 and q . Inequalities (22) and (23) then gives the condition
2 > 2. (24)
A.C. Fannjiang / Physica D 179 (2003) 161182 165
Note that for
< 2 (25)
and thus q > 0 we have a short-time limit; otherwise, it is a long time (but small spatial scale) limit.
The paper is organized as follows. In Section 2 we state the main results and discuss their implications. In Section
3 we discuss the meaning of solutions for the colored-noise and white-noise models and prove the uniqueness for
the latter. In Section 4, we prove Theorem 1: we prove the tightness of the measures in Section 4.1 and, in Section
4.2, identify the limiting measure by the martingale formulation. In Section 5, we prove Theorem 2. The method
of proof is the same as that in [14] (see also [5]). We refer the reader to [26] for the full exposition of the perturbed
test function method used here. We note that the method of Kunita [24] requires sub-Gaussian behavior and spatial
regularity of the velocity eld and is not applicable here.
2. Main theorems and interpretation
Let us begin by briey recalling the Kraichnan model. The model has a white-noise-in-time incompressible
velocity eld which can be described as the time derivative of a zero mean, isotropic Brownian vector eld B
t
with
the two-time structure function
E[B
t
(x) B
t
(y)] [B
s
(x) B
s
(y)] = min(t, s)

2[1 cos (k (x y))]a

1

E
L
( 1, k)[k[
1d
dk,
(0, 1) (26)
with

E
L
( 1, k) = lim
K
E
K,L
( 1, k).
In this paper, we interpret the corresponding advectiondiffusion equation for the Kraichnan model in the sense of
Stratonovichs integral
dT
t
(x) = [T
t
(x)]

[dB
t
(x) dB
t
(0)]

0
2
T
t
(x) dt,
0
0, T(0, x) = T
0
(x), (27)
which can be rewritten as an It os SDE
dT
t
=

0
2

1
a

T
t
dt

2a
1/2
T
t
d

W
(1)
t
, (28)
where

W
(1)
t
(x) is the Brownian vector eld with the spatial covariance

(1)
(x, y) =

[exp(ik x) 1][exp(ik y) 1]

E
L
( 1, k)[k[
1d
dk, = 1 (29)
and the operator

B is given by

B(x) =

i,j

(1)
ij
(x, x)

2
(x)
x
i
x
j
, C

(R
d
). (30)
We will discuss the meaning of solutions for the Kraichnan model and prove the uniqueness property in Section
3. The Kraichnan model for passive scalar has been widely studied to understand turbulent transport in the inertial
range because of its tractability (see, e.g., [6,10,12,19,20,28,30,38] and the references therein). The tractability of
this model lies in the Gaussian and white-noise nature of the velocity eld.
166 A.C. Fannjiang / Physica D 179 (2003) 161182
Theorem1. Suppose 2 > 2. Let L < be xed and let K = K() such that lim
0
K = . Let = () > 0
such that lim
0
=
0
< . Let T
0
L

(R
d
) L
2
(R
d
). If, additionally, any one of the following conditions is
satised:
(i) 2 > 4;
(ii) 2 = 4, lim
0

2

log K = 0;
(iii) 3 < 2 < 4, lim
0

2
K
42
= 0;
(iv) 2 = 3, lim
0

2
K = lim
0

log K = 0;
(v) 2 < 2 < 3, lim
0

2
K
42
= lim
0
K
32
= 0.
Then for the exponent q given in (23) the solution T

t
of (21) converges in distribution, as 0, in the space
D([0, ): L

(R
d
)) to the scalar eld T
t
for pair dispersion in the Kraichnan model in the time interval [0, t
0
] t
0
<
. The limiting Kraichnan model has the spatial covariance given by (29). Here D([0, ): L

(R
d
)) is the space
of L

(R
d
)-valued right continuous processes with left limits endowed with the Skorohod metric [3] and L

(R
d
)
is the standard space L

(R
d
) endowed with the weak

topology.
Remark 1. In addition to the assumptions stated in Section 1 and in the theorem, we use in the proof of
Theorem 1 the assumption
sup
t<t
0

[x[M
[

t
(x)[ dx = o

, 0 0 < M < (31)

with a random constant possessing a nite moment where

t
(x) =
1

t
E
t
2 V

2
, x

ds.
For Gaussian velocity elds one has
M
d
sup
[x[M
tt
0

2
, x

CL
22
log

M
d
t
0

= o

, (32)
where the random constant C has a Gaussian-like tail by Chernoffs bound. Condition (31) allows certain degree
of intermittency in the velocity eld.
Note that, in Theorem1, when
0
> 0 and 2 < 2 < 3, lim
0

2
K
42
= 0 implies lim
0
K
32
=
0. Also, 2 < 3 contains the regime < 2 in which the limiting Brownian velocity eld is spatially Hlder
continuous and has a Hurst exponent = 1 (1/2, 1), i.e. the limiting velocity eld has a persistent spatial
correlation.
If we let L in the Kraichnan model, we see that it gives rise to a Brownian velocity eld

B
t
with the
structure function
E[

B
t
(x)

B
t
(y)] [

B
s
(x)

B
s
(y)] = min(t, s)

2[1 cos (k (x y))]a

1

E( , k)[k[
1d
dk, (33)
where

E( , k) = lim
L

E
L
( , k).
The spectral integral in (33) is convergent only for < 2. The convergence of the integral in (33) means that
the limiting Brownian velocity eld

B
t
has spatially homogeneous increments.
A.C. Fannjiang / Physica D 179 (2003) 161182 167
We can prove the convergence to the Kraichnan model with velocity eld

B
t
in the simultaneous limit of
0, K, L if additional conditions are satised.
Theorem 2. Suppose < 2 and all the assumptions of Theorem 1 (thus, only regime (v) is relevant) except
for the niteness of L. Instead, let L = L() such that
lim
0
L
2(22)
= 0. (34)
Then the same convergence holds as in Theorem1. The limiting Brownian velocity eld

B
t
has the structure function
given by (33).
Remark 2. In addition to the assumptions of Theorem 1 (cf. Remark 1) we use in the proof of Theorem 2 the
assumption
sup
t<t
0

[x[M
[

t
(x)[
2
dx CL
2(22)
1

, 0, L 0 < M < (35)

with a random constant C possessing a nite moment. For Gaussian velocity elds one has
sup
t<t
0

[x[M
[

t
(x)[
2
dx CL
2(22)

log
1

2
, 0, L 0 < M < .
One sees that condition (35) is in some sense more tolerant of intermittency than (31) is.
Due to the divergence-free property of the velocity eld, the pre-limit scalar eld satises the energy identity [29,
Chapter III, Theorem 7.2]

[T

t
(x)[
2
dx

t
0

[T

t
[
2
(x) dx ds =

[T
0
(x)[
2
dx (36)
provided that T
0
L
2
(R
d
). From (36) we have the estimates
|T

t
|
2
2
< |T
0
|
2
,

t
0
|T

s
|
2
H
1
ds

t
1

|T
0
|
2
2
, t > 0,
where ||
H
1 is the normof the standard Sobolev space H
1
(R
d
) of square-integrable functions with square-integrable
rst derivative. Thus the lawof T

is naturally supported by the space of continuous L

2
(R
d
)-valued processes which
are also in L
2
loc
([0, ): H
1
(R
d
)). Following [5] we consider the space
= D([0, ): L
2
w
(R
d
) L

w
(R
d
)) L
2
w,loc
([0, ): H
1
w
(R
d
)),
where the subscripts w and loc denote the weak and the local topologies, respectively.
In the case of > 0,
0
> 0 the above observation and the tightness argument for Theorems 1 and 2 then imply
the tightness of T

t
in the space . We have the following corollary.
Corollary 1. If
0
> 0 and T
0
L

(R
d
) L
2
(R
d
) then the convergence holds in the space in the following
regimes:
Case 1: Let L < be xed and K as 0.
(i) 2 > 4;
(ii) 2 = 4, lim
0

2

log K = 0;
168 A.C. Fannjiang / Physica D 179 (2003) 161182
(iii) 2 < 2 < 4, lim
0

2
K
42
= 0.
Case 2: Suppose < 2 < 2 and L, K as 0 such that
lim
0

2
K
42
= lim
0
L
2(22)
= 0.
In particular,
|T
0
|
2
2
limsup
0
E[|T

t
|
2
2
] = liminf
0

t
0
E[|T

s
|
2
2
] ds
0

t
0
E[|T
s
|
2
2
] ds > 0,
t > 0, unless T
s
0, 0 s t, (37)
where T
t
is the solution of the corresponding Kraichnan model.
In the case of > 0,
0
= 0 and T
0
L
2
L

, the limiting Kraichnan model conserves the L

2
-norm of T
t
. The
energy identity (36) then implies
|T

t
|
2
2
|T
0
|
2
2
= |T
t
|
2
2
> 0 t > 0,
which in turn implies lim
0
|T

t
|
2
= |T
t
|
2
. Hence the weak sense of convergence in Theorems 1 and 2 can be
strengthened to the strong L
2
convergence.
Corollary 2. If
0
= 0 and T
0
L

(R
d
) L
2
(R
d
) then the convergence holds in the space D([0, ): L
2
(R
d
)
L

(R
d
)) in the respective regimes listed in Theorems 1 and 2. In particular,
|T
0
|
2
2
lim
0
E[|T

t
|
2
2
] = 0, t > 0 a.e.
We see that in the context of Corollary 1 there is positive dissipation (37) while there is none in the context of
Corollary 2. The conditions of the limit theorems set a constraint for the presence of positive dissipation: on the
observation scale , if the molecular diffusion is of order
22q
, then there is always positive dissipation no matter
how slow
1
vanishes. On the other hand, if <
22q
(i.e.
0
= 0) and the dissipation is positive, then

1
= O(

), =
4 2
3 2
with (2, ) in the regime < 2 < 2 (cf. (41)). An open question is whether there is a positive
dissipation as , 0 with
1
= 0 at the outset. If there is, then the Kraichnan model (27) is unlikely to be the
governing equation of the scaling limit (if exists).
In the case of = 0, a still stronger sense of convergence holds since now Eq. (21) is of rst order and any
locally bounded measurable function (T

) of the scalar eld satises the same equation (18) with = 0. The same
argument for the proof of Theorems 1 and 2 will then yield the following theorem.
Theorem 3. Assume the conditions stated in Remarks 1 and 2. Let = 0, T
0
, (T
0
) L

(R
d
) L
2
(R
d
) and is
a locally bounded measurable function from R to R. Then T

t
, (T

t
) converge in the space D([0, ): L

(R
d
)
L
2
(R
d
)) to the corresponding Kraichnan model in the following regimes.
Case 1: Let L < be xed and K as 0.
(i) 2 > 3;
(ii) 2 = 3, lim
0

log K = 0;
A.C. Fannjiang / Physica D 179 (2003) 161182 169
(iii) 2 < 2 < 3, lim
0
K
32
= 0.
Case 2: Suppose < 2 < 2 and L, K as 0 such that
lim
0
K
32
= lim
0
L
2(22)
= 0.
Remark 3. The assertions of Theorems 13 and Corollaries 1 and 2 hold true for random as well as deterministic
initial data.
When the parameters are in the regime < 2 < 2, by taking the expectation in the It os equation with
the Brownian velocity eld

B
t
one sees readily that the longitudinal relative diffusion coefcient is given by

0
2

1
a
x
[x[

(1)
(x, x)
x
[x[

1
a
C
1

E
0
[x[
2
for
0
< 1, = 1 q = 1

1
2
, 1

(38)
with

(1)
(x, x) = lim
L

(1)
(x, x) = C
1

E
0
[x[
2

1
2( 1)
d 1

I
2( 1)
d 1
x x[x[
2

,
where C

is dened as in (10), except with replaced by . The exponent q is related to the exponent p in
the expression for the mean-square pair separation as follows:
E[x[
2
(t) a
p
E
p
0
t
p
, p =
1
q
=
1
2
(39)
uptoa dimensionless constant dependingonlyon. Expressions (38) and(39) canbe viewedas the generalization
of Richardsons t
3
- and 4/3-laws, respectively. In general, p (2, ), indicating super-ballistic (i.e. accelerating)
motion as a result of a scale-dependent relative diffusivity.
We now remark on the range of scales for which Theorem 2 is proved and Richardsons laws can be reasonably
interpreted. Let be the scale of dispersion. Then the limit theorem holds in the range
< min

1
,

2(22)/(524)

, =

3 2
4 2
if
0
= 0,
4 2
6 2
if
0
> 0.
(40)
In the usual situation with
0
= O(1) the range of scales covered by the limit theorem has an upper limit of

1
with

(0,
1
2
) if
0
= 0
(
1
3
,
1
2
) if
0
> 0
for 2 > 2 > , (41)
which is limited to the low end of the inertial range depending on , ,
0
. It is not clear whether this is physical
or a technical matter. Qualitatively similar restriction of Richardsons laws in synthetic ows has been observed in
numerical calculation (cf. [4,18]).
If we stretch the validity of (38) and (39) by taking the limit 4/3, 1/3 from within the valid regime,
the resulting exponents are p = 3, 2 = 4/3 in accordance with Richardsons laws. On the boundary 2 = 2
the scaling exponent q should be given by
q = = 1
1
2
, (42)
170 A.C. Fannjiang / Physica D 179 (2003) 161182
which also coincides with the limiting value of (23). With (42) and K, L , the solution of (21) converges
to that of the advectiondiffusion equation with the molecular diffusivity
0
= lim
0
and the time-stationary,
spatially Hlder continuous velocity eld

V whose two-time correlation function is
E[

V(t, x)

V(s, y)] =

R
d
[exp(ik x) 1][exp(ik y) 1] exp(a[k[
2
[t s[)

E(, k)[k[
1d
dk,
(1, 2),
which has the self-similar structure
E[

V(
2
t, x)

V(
2
s, y)] =
22
E[

V(t, x)

V(s, y)].
In view of the 4th order scale-invariance property it is reasonable to postulate the temporal self-similarity on the
mean-square relative dispersion as
0
0
E[x(t)[
2
= f(E
0
, a)t
1/
,
which has the same exponent as the limiting case of (39) as 2 2, where the unknown function f satises
the relation
f(E
0
, a)
1/
= f(
2
E
0
, a) > 0.
Dimensional analysis with (9) then leads to the relation
E[x(t)[
2
=

C
R
C
1/2

E
1/2
0
t
1/
,
where

C
R
is the generalized Richardson constant. For = 1/3 the exponent p is 1/3 as predicted by Richardsons
t
3
-law. However, since the limiting velocity eld is non-white-in-time, the notion of relative diffusivity is not strictly
well dened. Therefore the temporal memory persists on small or intermediate time scales and the notion of relative
diffusivity does not describe accurately the process of relative dispersion on the boundary 2 = 2 (cf., e.g.,
[18,21,30]).
Let us consider the regime 2 < 2. The correct scaling is to set
2q 2 = 0 or q = 1
1
2
. (43)
Then the exponent 2(q ) of the temporal scaling in (21) is positive due to 2 < 2, meaning the time variable
is slowed down as 0. It is easy to see by a regular perturbation argument that the solution T

t
converges in the
sense described in Theorem 1 to the solution

T
t
of the following equation:

T
t
t
= V(0, x)

T
t


0
2

T
t
,

T
0
= T
0
L

(R
d
)
if
0
> 0. If, however,
0
= 0, the above equation probably have multiple solutions for a given initial condition.
The relation (43) is consistent with the numerical simulation using two-dimensional frozen velocity elds with
Kolmogorov-type spectrum [11].
Unlike the previous regime, for either 2 = 2 or 2 < 2 there is no restriction on the vanishing rate
of
1
.
A.C. Fannjiang / Physica D 179 (2003) 161182 171
3. Formulation
From the general theory of parabolic partial differential equations [17], for any xed > 0, > 0, there is a
unique C
2
-solution T

t
(x), 0 < < 1. But the solutions T

t
may lose all the regularity as 0, 0.
So we consider the weak formulation of the equation:
!T

t
, ) !T
0
, ) =

2

t
0
!T

s
, ) ds
1

t
0

s
, V

2
,

ds (44)
for any test function C

c
(R
d
), the space of smooth functions with compact support. On the other hand, the
energy identity (36) implies T

t
L
2
([0, t
0
]: H
1
(R
d
)) if T
0
L
2
(R
d
). Hence for L
2
initial data the pre-limit
measure P

is supported in the space L

2
([0, t
0
]: H
1
(R
d
)) and, by the tightness result (Section 4.1), the limiting
measure P is supported in L
2
w
([0, t
0
]: H
1
w
(R
d
)).
As in (14) and (16) the solutions T

t
can be represented as
T

t
= M[T
0
(
t,
0
(x))], (45)
where
t,
s
(x) is the unique stochastic ow satisfying
d
t,
s
(x) =
1

2
,
t,
s
(x)

ds

dw(t), 0 s t, (46)

t,
t
(x) = x. (47)
In the case of = 0,
t,
0
(x) t, is almost surely a diffeomorphism of R
d
and T

t
= T
0
(
t,
0
(x)). Moreover, for any
locally bounded measurable function : R R, (T

t
(x)) = ( T
0
)(
t,
0
(x)).
In view of the averaging in the representation (45) we have the following proposition.
Proposition 1.
|T

t
|

|T
0
|

a.s.
Clearly, Proposition 1 holds for the case of = 0 as well.
For tightness as well as identication of the limit, the following innitesimal operator A

will play an important

role. Let V

t
V(t/
2
, ). Let F

t
be the -algebras generated by {V

s
, s t} and E

t
the corresponding conditional
expectation w.r.t. F

t
. Let M

be the space of measurable function adapted to {F

t
t} such that sup
t<t
0
E[f(t)[ < .
We say f() D(A

), the domain of A

, and A

f = g if f, g M

and for f

(t)
1
[E

t
f(t ) f(t)] we
have
sup
t,
E[f

(t)[ < , lim

0
E[f

(t) g(t)[ = 0 t.
For f(t) = (!T

t
, )), f
/
(t) =
/
(!T

t
, )) C

(R) we have the following expression from (44) and the chain
rule:
A

f(t) =

2
f
/
(t)!T

t
, )
1

f
/
(t)!T

t
, V

t
()), (48)
where
V

t
() V

t
. (49)
172 A.C. Fannjiang / Physica D 179 (2003) 161182
A main property of A

is that
f(t)

t
0
A

f(s) ds is an F

t
-martingale f D(A

). (50)
Also,
E

s
f(t) f(s) =

t
0
E

s
A

f() d s < t a.s. (51)

(see [25]).
Likewise we formulate the solutions for the Kraichnan model (28) as the solutions to the corresponding martingale
problem. Find a measure P (of T
t
) on the space D([0, ): L

(R
d
)) such that
f(!T
t
, ))

t
0

f
/
(!T
s
, ))

0
2
!T
s
, )
1
a
!T
s
,

B

1
a
f
//
(!T
s
, ))!,

K
(1)
T
s
)

ds
is a martingale w.r.t. the ltration of a cylindrical Wiener process, for each f C

(R), (52)
where

B

is the adjoint of

B and
!,

K
(1)
T
s
) =

T
s
(x)T
s
(y)(x)

(1)
(x, y) (y) dy (53)
with

(1)
(x, y) given, respectively, by (29) and

(1)
(x, y) =

[exp(ik x) 1][exp(ik y) 1]

E( 1, k)[k[
1d
dk, = 1 (54)
for L < and L = . To identify the limit for the proof of convergence one needs the uniqueness of solution to
the martingale problem (52) which can be easily obtained as follows.
Taking expectation of (52) with f(r) = r
n
, n N we get for the n-point correlation function
F
t
n
(x
1
, x
2
, x
3
, . . . , x
n
) E
T
0
[T
t
(x
1
)T
t
(x
2
) T
t
(x
n
)]
the equation
!F
t
n
,
n
) !F
0
n
,
n
)
=

t
0

0
2
!F
s
n
, (x
1
) (x
j
) (x
n
))

j
1
a
!F
s
n
, (x
1
)

B

(x
j
) (x
n
))

i<j
2
a
!F
s
n
,

(1)
(x
i
, x
j
) : (x
1
) (x
i
) (x
j
) (x
n
))

ds,
which induces a weakly continuous (hence strongly continuous) sub-Markovian semigroup on L
p
(R
nd
) p
(1, ). The sub-Markovianity property is inherited from the pre-limit process T

t
. The generator of the semigroup
is given formally as
L
n
(x
1
, . . . , x
n
)

0
2
n

j=1

x
j

1
a
n

i,j=1

(1)
(x
i
, x
j
) :
x
i

x
j
, C

c
(R
nd
),
0
0 (55)
with the spatial covariance tensor

(1)
(x
i
, x
j
) given by (29) and (54), respectively, for L < and L = . Note
that the symmetric operator L
n
(55) is an essentially self-adjoint positive operator on C

c
(R
N
), N = nd which then
A.C. Fannjiang / Physica D 179 (2003) 161182 173
induces a unique symmetric Markov semigroup of contractions on L
2
(R
N
). The essential self-adjointness is due to
the sub-Lipschitz growth of the square-root of

(1)
(x
1
, x
2
) at large [x
1
[, [x
2
[ (hence no escape to innity) [8].
By Theorem 1.4.1 of [9] this semigroup induces a sub-Markovian C
0
-semigroup on L
p
(R
N
), p [1, ). The
uniqueness holds for these semigroups in their respective space as well but we will not pursue it here.
4. Proof of Theorem 1
4.1. Tightness
In the sequel we will adopt the following notation
f(t) f(!T

t
, )), f
/
(t) f
/
(!T

t
, )), f
//
(t) f
//
(!T

t
, )) f C

(R).
Namely, the prime stands for the differentiation w.r.t. the original argument (not t) of f, f
/
, etc.
A family of processes {T

, 0 < < 1} D([0, ): L

(R
d
)) is tight if and only if the family of processes
{!T

, ), 0 < < 1} D([0, ): L

(R
d
)) is tight for all C

c
(R
d
). We use the tightness criterion of Kushner
[28, Chapter 3, Theorem 4], namely, we will prove: rstly,
lim
N
limsup
0
P{sup
t<t
0
[!T

, )[ N} = 0 t
0
< . (56)
Secondly, for each f C

(R) there is a sequence f

(t) D(A

) such that for each t

0
< {A

(t), 0 < <

1, 0 < t < t
0
} is uniformly integrable and
lim
0
P{sup
t<t
0
[f

(t) f(!T

, ))[ } = 0 > 0. (57)

Then it follows that the laws of {!T

, ), 0 < < 1} are tight in the space of D([0, ): L

(R
d
)).
Condition (56) is satised as a result of Proposition 1. Let
f

1
(t)
1

t
E

t
f
/
(t)!T

t
, V

s
()) ds
be the rst perturbation of f(t). We obtain
f

1
(t) =

a
f
/
(t)!T

t
,

V

t
()) (58)
with

t
() =

V

t
, (59)

t


V

2
,

t
E

t
V

s
ds, (60)
where

V has the power spectrum E
K,L
( 2, k) by the spectral representation
E

t
V

s
=

[e
ixk
1] e
a[k[
2
[st[
2

t
(dk) s t. (61)
Note that while V

t
loses differentiability as K ,

V

t
is almost surely a C
1,
-function in the limit with
0 < < 2 2
and has uniformly bounded local W
1,p
-norm, p 1.
174 A.C. Fannjiang / Physica D 179 (2003) 161182
Proposition 2.
lim
0
sup
t<t
0
E[f

1
(t)[ = 0, lim
0
sup
t<t
0
[f

1
(t)[ = 0 in probability.
Proof. By Proposition 1 we have
E[[f

1
(t)[]

a
|f
/
|

|T
0
|

||

[x[M
E[

t
[ dx (62)
and
sup
t<t
0
[f

1
(t)[

a
|f
/
|

|T
0
|

||

sup
t<t
0

[x[M
[

t
[ dx. (63)
By the temporal stationarity of

V

t
we can replace the terms E[

t
(x)[ in (62) by E[

V(0, x)[. By assumption (cf.

(31), Remark 1), we have the desired estimate. Proposition 2 now follows from (31), (62) and (63).
Set f

(t) = f(t) f

1
(t). A straightforward calculation yields
A

1
=

2a
f
//
(t)!T

t
, )!T

t
,

V

t
())

2a
f
/
(t)!T

t
,

t
())
1
a
f
//
(t)!T

t
, V

t
())!T

t
,

V

t
())

1
a
f
/
(t)!T

t
, V

t
(

t
()))
1

f
/
(t)!T

t
, V

t
())
and, hence
A

(t) =

2
f
/
(t)!T

t
, )
1
a
f
/
(t)!T

t
, V

t
(

t
()))
1
a
f
//
(t)!T

t
, V

t
())!T

t
,

V

t
())

2a
[f
//
(t)!T

t
, )!T

t
, V

t
()) f
/
(t)!T

t
,

t
())] = A

1
(t) A

2
(t) A

3
(t) A

4
(t), (64)
where A

2
(t) and A

3
(t) are the O(1) statistical coupling terms.
For the tightness criterion stated in the beginning of the section, it remains to show the folowing proposition.
Proposition 3. {A

} are uniformly integrable and

lim
0
sup
t<t
0
E[A

4
(t)[ = 0.
Proof. We show that {A

i
}, i = 1, 2, 3, 4 are uniformly integrable. To see this, we have the following estimates:
[A

1
(t)[ =

2
[f
/
(t)!T

t
, )[

2
|f
/
|

|T
0
|

||
1
.
Thus A

1
is uniformly integrable since it is uniformly bounded:
[A

2
(t)[ =
1
a
[f
/
(t)!T

t
, V

t
(

t
()))[
C
a
|f
/
|

|T
0
|

[x[<M
[V

t
[
2
dx

1/2

[x[<M
[

V

t
[
2
dx

1/2
.
Similarly,
[A

3
(t)[ =
1
a
[f
//
(t)!T

t
, V

t
())!T

t
,

V

t
())[
C
a
|f
/
|

|T
0
|
2

[x[<M
[V

t
[
2
dx

[x[<M
[

t
[
2
dx

.
A.C. Fannjiang / Physica D 179 (2003) 161182 175
Thus A

2
and A

3
are uniformly integrable in view of the uniform boundedness of the 4th moment of V

t
,

V

t
and

t
as L < is xed and K (the 4th order scale-invariance):
[A

4
[ =

2a
[f
//
(t)!T

t
, )!T

t
,

V

t
()) f
/
(t)!T

t
,

t
())

C
2a

|f
//
|

|T
0
|
2

[x[<M
[

t
[
2
dx

1/2
|f
/
|

|T
0
|

[x[<M
[

t
[
2
dx

[x[<M
[

V

t
[
2
dx

[x[<M
[

t
[
2
dx

1/2

. (65)
The most severe term in the above argument as a result of K is

2a
[f
/
(t)!T

t
,

t
())[,
whose second moment can be bounded as

2a

E[f
/
(t)!T

t
,

t
())[
2
C
1

2a
|f
/
|

|T
0
|

[x[<M
E[[

t
[
2
] dx

1/2
C
2

K
32
for 2 < 3,

log K for 2 = 3,
1 for 2 > 3,
(66)
and, thus, vanishes in the limit by the assumptions of the theorem. The 4th moment behaves the same way by the
4th order scale-invariance. Hence A

4
is uniformly integrable. Clearly
lim
0
sup
t<t
0
E[A

4
(t)[ = 0.
4.2. Identication of the limit
Once the tightness is established we can use another result in [28, Chapter 3, Theorem 2] to identify the limit. Let
Abe a diffusion or jump diffusion operator such that there is a unique solution
t
in the space D([0, ): L

(R
d
))
such that
f(
t
)

t
0
Af(
s
) ds (67)
is a martingale. We shall show that for each f C

(R) there exists f

D(A

) such that
sup
t<t
0
,
E[f

(t) f(!T

t
, ))[ < , (68)
lim
0
E[f

(t) f(!T

t
, ))[ = 0 t < t
0
, (69)
sup
t<t
0
,
E[A

(t) Af(!T

t
, ))[ < , (70)
lim
0
E[A

(t) Af(!T

t
, ))[ = 0 t < t
0
. (71)
Then the aforementioned theorem implies that any tight processes !T

t
, ) converge in law to the unique process
generated by A. As before we adopt the notation f(t) = f(!T

t
, )).
176 A.C. Fannjiang / Physica D 179 (2003) 161182
For this purpose, we introduce the next perturbations f

2
, f

3
. Let
A
(1)
2
() !, K
(1)

), (72)
A
(1)
3
() !, E[V

t
(

t
())]), (73)
where the positive-denite operator K
(1)

is dened as
K
(1)

(y)(x)
(1)
(x, y)(y) dy, (74)

(1)
(x, y) =

[exp(ik x) 1][exp(ik y) 1]E

K,L
( , k)[k[
1d
dk (75)
such that
!
1
, K
(1)
T
t

2
) =

(x)(y)G
(1)

1
,
2
(x, y) dx dy, (76)
G
(1)

1
,
2

i,j

2
x
i
y
j
[
1
(x)
2
(y)
(1)
ij
(x, y)] (77)
(cf. (53)).
It is easy to see that
A
(1)
2
() = E[!, V

t
())!,

V

t
())], (78)
A
(1)
3
() = !B, ), (79)
where the operator B is given by
B(x) =

i,j

(1)
ij
(x, x)

2
(x)
x
i
x
j
.
Dene
f

2
(t)
1
a
f
//
(t)

t
E

t
[!T

t
, V

s
())!T

t
,

V

s
()) A
(1)
2
(T

t
)] ds,
f

3
(t)
1
a
f
/
(t)

t
E

t
[!T

t
, V

s
(

s
())) A
(1)
3
(T

t
)] ds.
Let
G
(2)

1
,
2
(x, y)

i,j

(2)
ij
(x, y)

1
(x)
x
i

2
(y)
y
j
, !
1
, K
(2)

2
)

(x)(y)G
(2)

1
,
2
(x, y) dx dy,
where the covariance function
(2)
(x, y) E[

t
(x)

V

t
(y)] has the spectral density E
K,L
( 2, k). Let
A
(2)
2
() !, K
(2)

), A
(2)
3
() !, E[

t
(

t
())]).
Noting that
E

t
[V

s
(x)

V

s
(y)] =

[e
ixk
1][e
iyk
/
1] e
a[k[
2
[st[
2
e
a[k
/
[
2
[st[
2

t
(dk)

t
(dk
/
)

[e
ixk
1][e
iyk
1][1 e
2a[k[
2
[st[
2
]E
K,L
( , k) dk, (80)
A.C. Fannjiang / Physica D 179 (2003) 161182 177
we then have
f

2
(t) =

2
2a
2
f
//
(t)[!T

t
,

V

t
())
2
A
(2)
2
(T

t
)] (81)
and similarly
f

3
(t) =

2
2a
2
f
/
(t)[!T

t
,

V

t
(

t
())) A
(2)
3
(T

t
)]. (82)
Proposition 4.
lim
0
sup
t<t
0
E[f

2
(t)[ = 0, lim
0
sup
t<t
0
E[f

3
(t)[ = 0.
Proof. We have the bounds
sup
t<t
0
E[f

2
(t)[ sup
t<t
0

2
2a
2
|f
//
|

|T
0
|
2

||
2

[x[<M
E[

t
[
2
(x) dx

[x[<M
[
(2)
(x, x)[ dx

C
1

2
,
sup
t<t
0
E[f

3
(t)[ sup
t<t
0

2
2a
2
|f
/
|

|T
0
|

||

[x[<M
E[

t
[
2
(x) dx ||

[x[<M
E[

t
[
2
(x) dx

1/2

[x[<M
E[

V

t
[
2
(x) dx

1/2

C
2

2
K
22
both of which tend to zero.
We have
A

2
(t) =
1
a
f
//
(t)[!T

t
, V

t
())!T

t
,

V

t
()) A
(1)
2
(T

t
)] R

2
(t),
A

3
(t) =
1
a
f
/
(t)[!T

t
, V

t
(

t
())) A
(1)
3
(T

t
)] R

3
(t)
with
R

2
(t) =
f
///
(t)
2

2

2a
2
!T

t
, )

a
2
!T

t
, V

t
())

[!T

t
,

V

t
())
2
A
(2)
2
(T

t
)]
f
//
(t)!T

t
,

V

t
())

2
2a
2
!T

t
,

t
())

a
2
!T

t
, V

t
(

t
()))

f
//
(t)

2
4a
2
!T

t
, G
(2)

t
)

a
2
!T

t
, V

t
(G
(2)

t
))

, (83)
where G
(2)

denotes the operator

G
(2)

G
(2)
,
(x, y)(y) dy,
and similarly
R

3
(t) = f
//
(t)

2
4a
2
!T

t
, )

2a
2
!T

t
, V

t
())

[!T

t
,

V

t
(

t
())) A
(2)
3
(T

t
)]
f
/
(t)

2
4a
2
!T

t
,

t
(

t
()))

2a
2
!T

t
, V

t
(

t
(

t
())))

f
/
(t)

2
4a
2
!T

t
, E[

t
(

t
())])

2a
2
!T

t
, V

t
(E[

t
(

t
())]))

.
178 A.C. Fannjiang / Physica D 179 (2003) 161182
Proposition 5.
lim
0
sup
t<t
0
E[R

2
(t)[ = 0, lim
0
sup
t<t
0
E[R

3
(t)[ = 0.
Proof. The argument is entirely analogous to that for Proposition 4. The most severe term without the prefactor
occurs in the expression for R

3
(t) and can be bounded as
E[!T

t
, V

t
(

t
(

t
())))[ |T
0
|

E[V

t
(

t
(

t
()))[ C
1
|T
0
|

[x[<M
E[V

t
[
2
dx

1/2

[x[<M
{E[[

t
[
4
]E[[
2

t
[
4
]}
1/2
dx

[x[<M
E[[

V

t
[
4
] dx

1/2
(84)
by assumption. The right-hand side of the above tends to zero if either
2 > 3
or
2 = 3, lim
0

log K = 0 (85)
or
2 < 3, lim
0
K
32
= 0 (86)
is satised. The term involving !T

t
, V

t
(G
(2)

t
)) can be similarly estimated.
The most severe term involving the prefactor occurs in R

3
and can be bounded as

2
E[!T

t
,

t
(

t
()))[ C
2
|T
0
|

[x[<M
E[[
3

t
[
2
]

1/2

2
for 2 > 4,

2

log K for 2 = 4,

2
K
42
for 2 < 4,
(87)
the right-hand side of which tends to zero if either
2 > 4
or
2 = 4, lim
0

2

log K = 0
or
3 < 2 < 4, lim
0

2
K
42
= 0 (88)
or
2 < 2 < 3, lim
0

2
K
42
= lim
0
K
32
= 0.
Note that for 2 2 the condition (85) or (86) implies that
lim
0

2
K
42
= 0.
A.C. Fannjiang / Physica D 179 (2003) 161182 179
Set
R

(t) = A

4
(t) R

2
(t) R

3
(t).
It follows from Propositions 3 and 5 that:
lim
0
sup
t<t
0
E[R

(t)[ = 0.
Recall that
M

t
() = f

(t)

t
0
A

(s) ds = f(t) f

1
(t) f

2
(t) f

3
(t)

t
0

2
f
/
(t)!T

t
, ) ds

t
0
1
a
[f
//
(s)A
(1)
2
(T

s
) f
/
(s)A
(1)
3
(T

s
)] ds

t
0
R

(s) ds
is a martingale. Now that (68)(71) are satised we can identify the limiting martingale to be
M
t
() = f(t)

t
0

f
/
(s)

0
2
!T
s
, )
1
a

A
(1)
3
(T
s
)

1
a
f
//
(s)

A
(1)
2
(T
s
)

ds, (89)
where

A
(1)
2
() = lim
K
A
(1)
2
(),

A
(1)
3
() = lim
K
A
(1)
3
()
(cf. (72) and (79)).
Since !T

t
, ) is uniformly bounded
[!T

t
, )[ |T
0
|

||
1
,
we have the convergence of the second moment
lim
0
E{!T

t
, )
2
} = E{!T
t
, )
2
}.
Use f(r) = r and r
2
in (89)
M
(1)
t
() = !T
t
, )

t
0

0
2
!T
s
, )
1
a

A
(1)
3
(T
s
)

ds
is a martingale with the quadratic variation
[M
(1)
(), M
(1)
()]
t
=
2
a

t
0

A
(1)
2
(T
s
) ds =
2
a

t
0
!,

K
(1)
T
s
) ds,
where

K
(1)
T
t
is a positive-denite operator given formally as

K
(1)
T
t
=

(y)T
t
(x)

(1)
(x, y)T
t
(y) dy, (90)
(cf. (74)). Therefore,
M
(1)
t
=

2
a

t
0

K
(1)
T
s
dW
s
,
180 A.C. Fannjiang / Physica D 179 (2003) 161182
where W
s
is a cylindrical Wiener process (i.e. dW
t
(x) is a spacetime white-noise eld) and

K
(1)
T
s
the square-root
of the positive-denite operator given in (90). From (72) and (79) we see that the limiting process T
t
is the (assumed
unique) distributional solution to the martingale problem (52) of the It os equation
dT
t
=

0
2

1
a

T
t
dt

2a
1
K
(1)
T
t
dW
t
=

0
2

1
a

T
t
dt

2a
1/2
T
t
d

W
(1)
t
,
where the operator

B is given by (30) and

W
(1)
t
is the Brownian vector eld with the spatial covariance

(1)
(x, y).
5. Proof of Theorem 2
As we let L along with 0 the proof of the uniform integrability of A

[f(t) f

1
(t)] (the rst part of
Proposition 3) breaks down. In this case, we work with the perturbed test function
f

(t) = f(t) f

1
(t) f

2
(t) f

3
(t).
Proposition 6.
lim
0
sup
t<t
0
E[f

j
(t)[ = 0, lim
0
sup
t<t
0
[f

j
(t)[ = 0 in probability j = 1, 2, 3. (91)
Proof. The argument for the case of f

1
(t) is the same as Proposition 2. For f

2
(t) and f

3
(t) we have the bounds
sup
t<t
0
E[f

2
(t)[ sup
t<t
0

2
2a
2
|f
//
|

|T
0
|
2

||
2

[x[<M
E[

t
[
2
(x) dx

[x[<2M
[
(2)
(x, x)[ dx

C
1

2
L
2(2)4
,
sup
t<t
0
E[f

3
(t)[ sup
t<t
0

2
2a
2
|f
/
|

|T
0
|

||

[x[<M
E[

t
[
2
(x) dx ||

[x[<M
E[

t
[
2
(x) dx

1/2

[x[<M
E[

V

t
[
2
(x) dx

1/2

C
2

2
L
2(2)4
both of which vanish under the assumptions of the theorem. Here we have used the fact that

[x[<M
E[

t
[
2
(x) dx = O(L
2(2)4
), L .
As for estimating sup
t<t
0
[f

j
(t)[, j = 2, 3, we can use
M
d

[x[<M
[

t
[
2
(x) dx in place of

[x[<M
E[

t
[
2
(x) dx
in the above bounds and obtain by assumption (cf. (35), Remark 2) the desired estimate which have a similar order
of magnitude with an additional factor of 1/ and a random constant possessing a nite moment.
We have
A

(t) =

2
f
/
(t)!T

t
, )
1
a
f
//
(t)A
(1)
2
(T

t
)
1
a
f
/
(t)A
(1)
3
(T

t
) R

1
(t) R

2
(t) R

3
(t) (92)
A.C. Fannjiang / Physica D 179 (2003) 161182 181
with
R

1
(t) =

2a
[f
//
(t)!T

t
, )!T

t
, V

t
()) f
/
(t)!T

t
,

t
())] (93)
and R

2
(t), R

3
(t) as before.
Proposition 7.
lim
0
sup
t<t
0
E[R

j
(t)[ = 0, j = 1, 2, 3.
Proof. The proof is similar to that of Proposition 5 with the additional consideration due to L . These
additional terms can all be estimated by
C
1

[x[<M
E[[

t
(x)

t
(x)[] dx C
2
L
2(22)
,
which tends to zero under the assumptions of the theorem.
For the tightness it remains to show the following proposition.
Proposition 8. {A

} are uniformly integrable.

Proof. We shall prove that each term in the expression (92) is uniformly integrable.
The rst three terms are clearly bounded under the assumption of < 2. The last three terms can be estimated
as in Proposition 7 by
C
1
sup
t<t
0

[x[<M
[

t
(x)

t
(x)[
whose second moment behaves like
2
L
4(22)
, by the 4th order scale-invariance property, and tends to zero.
Now we have all the estimates needed to identify the limit as in the proof of Theorem 1.
Acknowledgements
I thank L. Biferale and K. Gawedzki for stimulating discussions on the nature of Lagrangian turbulent velocity
during the Developed Turbulence program, June 2002, at The Erwin Schrdinger International Institute for
Mathematical Physics, Vienna. I appreciate the nancial support and hospitality of ESI. The research is supported
in part by The Centennial Fellowship from American Mathematical Society and a grant from US National Science
Foundation, DMS-9971322.
References
[1] G.K. Batchelor, Diffusion in a eld of homogeneous turbulence. II. The relative motion of particles, Proc. Camb. Philos. Soc. 48 (1952)
345362.
[2] V.I. Belinicher, V.S. L vov, A scale-invariant theory of fully developed hydrodynamic turbulence, Sov. Phys. JETP 66 (2) (1987) 303313.
[3] P. Billingsley, Convergence of Probability Measures, Wiley, New York, 1968.
182 A.C. Fannjiang / Physica D 179 (2003) 161182
[4] G. Boffetta, A. Celani, A. Crisanti, A. Vulpiani, Pair dispersion in synthetic fully developed turbulence, Phys. Rev. E 60 (6) (1999)
67346741.
[5] R. Carmona, J.-P. Fouque, Diffusion-approximation for the advectiondiffusion of a passive scalar by a spacetime Gaussian velocity eld,
in: E. Bolthaunsen, M. Dozzi, F. Russo (Eds.), Seminar on Stochastic Analysis, Random Fields and Applications, Birkhauser, Basel, 1995,
pp. 3750.
[6] M. Chertkov, G. Falkovich, I. Kolokolov, V. Lebedev, Normal and anomalous scaling of the fourth-order correlation function of a randomly
advected scalar, Phys. Rev. E 51 (1995) 56095627.
[7] S. Corrsin, On the spectrum of isotropic temperature uctuations in an isotropic turbulence, J. Appl. Phys. 22 (1951) 469473.
[8] E.B. Davies, L
1
properties of second order elliptic operators, Bull. London Math. Soc. 17 (1985) 417436.
[9] E.B. Davies, Heat Kernel and Spectral Theory, Cambridge University Press, Cambridge, 1989.
[10] W. E, E. Vanden-Eijnden, Turbulent Prandtl number effect on passive scalar advection, Physica D 152153 (2001) 636645.
[11] F.W. Elliott Jr., A.J. Majda, Pair dispersion over an inertial range spanning many decades, Phys. Fluids 8 (4) (1996) 10521060.
[12] G. Falkovich, G. Gawedzki, M. Vergassola, Particles and elds in uid turbulence, Rev. Mod. Phys. 73 (2001) 913975.
[13] A. Fannjiang, Phase diagram for turbulent transport: sampling drift, eddy diffusivity and variational principles, Physica D 136 (12) (2000)
145174.
[14] A. Fannjiang, Convergence of passive scalars in OrnsteinUhlenbeck ows to Kraichnans model. arXiv:/abs/math-ph/0209011.
[15] A. Fannjiang, T. Komorowski, Turbulent diffusion in Markovian ows, Ann. Appl. Probab. 9 (3) (1999) 591610.
[16] A. Fannjiang, T. Komorowski, S. Peszat, Lagrangian dynamics for a passive tracer in a class of Gaussian Markovian ows, Stochast.
Process. Appl. 97 (2002) 171198.
[17] A. Friedman, Partial Differential Equations of Parabolic Type, Prentice-Hall, Englewood Cliffs, NJ, 1964.
[18] J.C.H. Fung, J.C. Vassilicos, Two-particle dispersion in turbulent ows, Phys. Rev. E 57 (2) (1998) 16771690.
[19] K. Gawedzki, A. Kupiainen, Anomalous scaling of the passive scalar, Phys. Rev. Lett. 75 (1995) 38343837.
[20] K. Gawedzki, M. Vergassola, Phase transition in the passive scalar advection, Physica D 138 (2000) 6390.
[21] H.G.E. Hentschel, I. Procaccia, Relative diffusion in turbulent media: the fractal dimension of clouds, Phys. Rev. A29 (3) (1984) 14611471.
[22] M.-C. Jullien, J. Paret, P. Tabeling, Richardson pair dispersion in two-dimensional turbulence, Phys. Rev. Lett. 82 (14) (1999) 28722875.
[23] V.S. L vov, E. Podivilov, I. Procaccia, Temporal multiscaling in hydrodynamic turbulence, Phys. Rev. E 55 (6) (1997) 70307035.
[24] H. Kunita, Stochastic Flows and Stochastic Differential Equations, Cambridge University Press, Cambridge, 1990.
[25] T.G. Kurtz, Semigroups of conditional shifts and approximations of Markov processes, Ann. Probab. 3 (4) (1975) 618642.
[26] H.J. Kushner, Approximation and Weak Convergence Methods for Random Processes, with Applications to Stochastic Systems Theory,
MIT Press, Cambridge, MA, 1984.
[27] O.A. Ladyhenskaya, V.A. Solonnikov, N.N. Uralceva, Linear and Quasilinear Equations of Parabolic Type, AMS Translation of
Mathematical Monographs, vol. 23, 1968.
[28] Y. Le Jan, O. Raimond, Integration of Brownian vector elds, 1999. math.PR/9909147.
[29] C.C. Lin, On a theory of dispersion by continuous movements, Proc. Natl. Acad. Sci. 46 (1960) 566569.
[30] A.J. Majda, P.R. Kramer, Simplied models for turbulent diffusion: theory, numerical modeling, and physical phenomena, Phys. Rep.
314 (45) (1999) 237574.
[31] A. Monin, A. Yaglom, Statistical Fluid Mechanics, vol. 2, MIT Press, Cambridge, MA, 1975.
[32] A.M. Obukhov, Structure of the temperature eld in a turbulent ow, Izv. Akad. Nauk SSSU, Geogr. Geoz. 13 (1949) 5869.
[33] H. Osada, Homogenization of diffusion processes with random stationary coefcients, in: Proceedings of the Fourth JapanUSSR
Symposium on Probability Theory, Lecture Notes in Mathematics, vol. 1021, Springer, Berlin, 1982, pp. 507517.
[34] G.C. Papanicolaou, S.R.S. Varadhan, Diffusion with random coefcients, in: G. Kallianpur, P.R. Krishnaiah, J.K. Ghosh (Eds.), Statistics
and Probability: Essays in Honor of C.R. Rao, North-Holland, Amsterdam, 1982, pp. 547552.
[35] A.A. Praskovsky, E.B. Gledzer, M.Yu. Karyakin, Y. Zhou, The sweeping decorrelation hypothesis and energy-inertial scale interaction in
high Reynolds number ows, J. Fluid Mech. 248 (1993) 493511.
[36] L.F. Richardson, Atmospheric diffusion shown on a distance-neighbor graph, Proc. R. Soc. London A 110 (1926) 709737.
[37] Y. Sato, K. Yamamoto, Lagrangian measurement of uid-particle motion in an isotropic turbulent eld, J. Fluid Mech. 175 (1987) 183199.
[38] B. Shraiman, E. Siggia, Anomalous scaling of a passive scalar in turbulent ow, C.R. Scand. Sci. 321 (1995) 279284.
[39] V.I. Tartarski, Izv. Vyssh. Uchebn. Zaved. Radioz. 4 (1960) 551.
[40] H. Tennekes, Eulerian and Lagrangian time microscales in isotropic turbulence, J. Fluid Mech. 67 (1975) 561567.
[41] M. Virant, T. Dracos, 3D PTV and its application on Lagrangian motion, Measur. Sci. Technol. 8 (1997) 15391552.
[42] C.L. Zirbel, Lagrangian observations of homogeneous random environments, Adv. Appl. Probab. 33 (2001) 810835.
[43] N. Zouari, A. Babiano, Derivation of the relative law in the inverse energy cascade of two-dimensional turbulence, Physica D 76 (1994)
318328.