Professional Documents
Culture Documents
Department of Mathematics
CASE WESTERN RESERVE UNIVERSITY
May 2012
Contents
1 Introduction
11
2.1
Definitions of Metrics . . . . . . . . . . . . . . . . . . . . 11
2.2
Incompleteness Issues . . . . . . . . . . . . . . . . . . . . 14
2.3
New Results . . . . . . . . . . . . . . . . . . . . . . . . . 17
27
3.1
Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2
3.3
SPDE Results . . . . . . . . . . . . . . . . . . . . . . . . 42
3.4
Extension by Continuity . . . . . . . . . . . . . . . . . . 45
3.5
Hahn-Jordan Decompositions . . . . . . . . . . . . . . . 50
3.6
New Results . . . . . . . . . . . . . . . . . . . . . . . . . 55
61
4.1
Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2
New Results . . . . . . . . . . . . . . . . . . . . . . . . . 68
5 Conclusion
72
6 Appendix
76
6.1
Fluid Dynamics . . . . . . . . . . . . . . . . . . . . . . . 76
6.2
Stochastics . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Bibliography
97
ii
Acknowledgements
There are several people whom I would like to thank for their support. Without
their help and guidance, I certainly would not have been successful at my time here
at Case. First and most importantly, I would like to thank to my research advisor,
Prof. Peter Kotelenez, for the endless support. His patience and guidance throughout
the process is something I never took for granted. I was very lucky to have him to
turn to for ideas and guidance. He also taught me a lot of life lessons, and he was
the reason I decided on the Case math program and my field of studies.
Prof. Daniela Calvetti has been a wonderful supporter of the graduate students
at Case, but she has been an especially strong advocate for me. I am indebted to her
for the innumerable ways that she has benefited my education and career. During
my time at Case, Prof. Calvetti has put lots of time and energy to create a great
environment for the graduate studies. I am very grateful for this, as it fostered very
strong friendships with my fellow students and our professors.
Prof. Elizabeth Meckes, Prof. Marshall Leitman, Prof. Manfred Denker, and
Prof. Peter Kotelenez put forth a lot of energy and time into serving on my thesis
committee. I would like to thank them for their effort and time to evaluate my thesis.
Chris Butler started me on this journey through mathematics. He gave me my
first exposure to the math department when he hired me as an S.I. in my second year
of undergrad. I am very grateful to him for giving me this opportunity.
Diane Robinson, Jeanne Jurkovich and Gaythreesa Lewis are the unsung heroines
of the math department. They have taken care of so many issues for me during the
last seven years. I am very appreciative of all their efforts.
The graduate students at Case are so much more than just students in the same
iii
program. We are united by the same difficult trials and stresses throughout our years
here. They have provided me with friendship and motivation, as well as distraction
when I needed it. Thank you for all the memories at Yost.
My family always took the time to listen to my excitement and my frustration. I
cannot thank them enough for the support that they provide me.
Finally, my biggest supporter is my girlfriend, Emily. She has been by my side
throughout my entire time in the Ph.D. program. She has seen my biggest successes
and hardest failures, and she has always provided support and an ear to let me vent.
If there is anyone who is more excited for me to be completing my Ph.D., it is she. I
will always be indebted to her for the patience she has shown me during the difficult
periods of the program. I dedicate this thesis to her.
iv
Abstract
Signed Measure Valued Stochastic Partial Differential Equations
BRADLEY T. SEADLER
Introduction
The behavior of a fluid in motion is a phenomenon that intrigues and fascinates
,
r1
r2
(1.1)
U 0
where is the kinematic viscosity of the fluid, is the Laplacian, is the gradient,
and is the inner product on R2 . For the case of an ideal or inviscid fluid, = 0,
and these equations represent the Euler equations in vorticity form. For > 0, these
1
For a full introduction to the field of mathematical fluid dynamics, refer to the classic texts of
[Lad69] or [CM93].
2
An incompressible fluid is a fluid which has a divergence free velocity field.
(1.2)
This follows as (1.1) represents the differential form of the continuity equation3 for
the vorticity, as remarked by [Lon88].
Another consequence of the incompressibility condition is that one can explicitly
express the velocity field in terms of the vorticity distribution as follows.4
Z
U (r, t) =
Z
Here
(1.3)
()dq denotes integration over R2 with respect to Lebesgue measure and K()
1
1
ln(|r|) =
(r2 , r1 )
2
2|r|2
(1.4)
,
) and |r|2 = r12 + r22 . Solving the Euler or Navier-Stokes
r2 r1
equations encounters several difficulties due in part to the singularity of the Biotwhere = (
Savart kernel at 0. Rather than explicitly solve the equations, many researchers
sought to use (1.1) as a way to simulate flows. To create a numerical method for the
Euler equations, Alexandre Chorin developed the regularized point-vortex method
3
4
(1.5)
The second technique assumes the initial vorticity, X (r, 0), has N N point vorticies
with nonzero intensities, ak R, and positions, r0k R2 . Thus, for k = 1, . . . , N
X (r, 0) =
N
X
(1.6)
ak r0k
k=1
dri X
ak K (ri rk ),
=
dt
k=1
ri (0) = r0i .
(1.7)
XN (t) :=
N
X
ak rk (t) ,
(1.8)
k=1
is a weak solution5 of the Euler-equation, (1.1), with the smoothed Biot-Savart ker5
In this article, weak solutions shall always be in the sense of partial differential equations (hereafter PDE).
nel.6 This means that if is a smooth function from R2 to R with compact support
and < , > is the duality between generalized functions and smooth functions, then
we have the following differential:
(1.9)
where
Z
U,N (r, t) :=
(1.10)
Further, Chorin devises and implements a sampling algorithm to approximate arbitary vortex distributions by discrete vortex distributions of the form (1.6). As
remarked in [Cho73], the algorithm is successful as (1.7) is a rectangular quadrature rule to (1.3). Also, the approximation of initial vortex distributions leads to
approximations in the general vortex distibution.
The techniques developed in [Cho73] generate significant progress in the case of
Eulers equation. However, the Navier-Stokes equation still provides difficulties due
to the viscosity term (the term involving in (1.1)). One approach is to examine
the random point vortex method and the Stochastic Navier-Stokes equations. The
random point vortex method modifies (1.7) to be a stochastic ordinary differential
equation (hereafter SODE) rather than a deterministic differential equation. [MP82],
[GHL90], [Lon88], [Kot95], [Ami00] all consider (1.7) with a stochastic driving term
of the following form. For i = 1, . . . , N ,
dr (t) =
N
X
(1.11)
k=1
6
Since this property is an essential part of the new analysis presented in this work, a proof is
provided in (3.40)-(3.42)
N
X
i=1
2
X
ai
2
kl
((ri (t)))d << mik (rN , t), mil (rN , t) >>
k,l=1
N
X
(1.12)
i=1
2
where kl
denotes the second derivative with respect to the spatial coordinates
7
8
rk , rl and << mik (rN , t), mil (rN , t) >> denotes the mutual quadratic variation process
of the one-dimensional components of mi ().9
In [MP82], [Lon88], and [GHL90] (as well as numerous other approaches), the
authors choose mi (rN (t), t) as independent 2-dimensional Brownian motions, i (t),
for i = 1, . . . , N . This reduces (1.12) to the following:
d < XN (t), > = < XN (t), (U,N ) > dt
+ < XN (t), > dt +
N
X
(1.13)
i
ai (r (t)) d (t).
i=1
Note that if the stochastic term did not appear in (1.13), the equation would represent
the weak form of the smoothed Navier-Stokes equations. Hence, (1.13) represents a
SNSE.
In both [Lon88] and [GHL90], the authors focus on the numerical aspects of the
algorithms following from [Cho73]. The authors in [GHL90] show that the point
vortex algorithm for the 2D Euler equation is consistent and stable. Further, the
authors derive a second order estimate on the error. Long examines the random point
vortex method as a numerical algorithm in [Lon88]. Here the author proves a nearly
optimal rate of convergence according to the Central Limit Theorem. Yet, in [MP82],
the authors establish an interesting advance in the general theory. The authors show
the following with some additional hypotheses. As the number of vortices, N , tends
to infinity, the empirical process, XN (t), will approximate weak solutions of (1.1).10
We call limits where N continuum-limits, and such limits play a central role in
the later analysis.
The results of [MP82], [Lon88], and [GHL90] provide important advances to the
9
10
theory of SNSE. Yet, the choice of mi (rN (t), t) = i (t) has several drawbacks. In
[Kot95], the author summarizes the limitations of using independent Brownian motions in (1.11) as follows:
In (1.13), the position of each particle is perturbed by its own fluctuation
force, i (t). This creates an identification that is preserved in the stochastic
term, yet disappears in the deterministic term. Consequently, the identification
prevents (1.13) from representing a smoothed signed-measure valued SNSE.
(1.13) represents a SNSE, but only in a formal way.
The choice of mi (rN (t), t) = i (t) yields fluctuation forces which are state independent (i.e. do not depend on rN (t)). From the physical interpretation, it is
more desirable that the fluctuation forces are state dependent.
Choosing mi (rN (t), t) as independent Brownian motions rather than spatially
correlated increases the high singularity to an associated SPDE on signedmeasures.
In order to address these limitations, [Kot95] introduces correlation functions,
(r, q), in [Kot95] to describe the fluctuation in the motion of the particles. Here
the removal of the taggings and fluctuation forces that are state dependent, spatially
correlated, and driven by Brownian sheets. In this method, the following choice for
the square-integrable martingales is made.
Z tZ
m (rN , t) :=
(1.14)
where w(dp, dt) = (w1 (dp, dt), w2 (dp, dt))T is standard space-time Gaussian white
noise differentials.11 With this choice of mi (rN , t), (1.12) becomes the following.
d < XN (t), > = < XN (t), (U ) > dt
2 XN (t)
Note that integrating (1.15) by parts in the sense of generalized functions shows
that XN (t) is a weak solution of the following smoothed, signed-measure valued
Stochastic Navier-Stokes Equation:
dX (t) = [X (U X )] dt
Z
(, p) w(dp, dt).
2 X
(1.16)
stochastic vortex theory. Here, Amirdjanova remarks on the difficulties that incompleteness imposes on the analysis. However, in [Ami07], the author can only show
the existence and uniqueness of the SODE for positive measures. As a result of the
work in [Kot95] and [Ami07], several questions naturally arise.
One of the standard choices for a metric in SPDE theory is not complete on
the space of signed-measures. Can one alter the definition of the metric to yield
completeness on this space, but without destroying the usefulness of this metric
in the analysis?
The continuum-limit in [Kot95] does not show conservation of vorticity. With
an alternative choice of metric, can one show conservation of vorticity in the
continuum-limit?
Seeking answers to these questions led to the original work derived in this work. The
analysis created new results in several areas, and we divide this work into chapters
based on these areas.
empirical process for an associated SODE-system will satisfy the SPDE. Work in
[Kot10] shows that the Hahn-Jordan decomposition for an initial signed-measure
is preserved provided that either the initial measure is discrete or that the
coefficients of the SODE are sufficiently smooth. Using the partially-complete
metric from Chapter 2, we generalize the Hahn-Jordan result to the case where
the coefficients of the SODE are only assumed to be Lipschitz.
12
12
13
13
10
most notably the Kantorovich-Rubinstein metric. It is shown that this metric fails
completeness on the signed measures despite being complete on positive measures.
Identifying a relationship between signed measures and diagonal sets inspires a new
metric similar to the quotient-space metric. With the goal of completeness in mind,
this metric fails convergence for arbitrary Cauchy sequences of signed measures. However, the metric satisfies a more useful partial-completeness result. A Cauchy sequence
of signed measures converges if and only if the following property holds. The HahnJordan decompositions of a subsequence converge to a limit which can be identified as
a signed measure. Such a result is very desirable in the subsequent analysis. One can
conclude a pair of measures represents the Hahn-Jordan decomposition of a signed
measure by the convergence properties of the metric.
2.1
Definitions of Metrics
Let (S, ) be a complete, separable metric space with countable, dense set T . One
typically takes S as Rd for d N and as the Euclidean metric; however, the following
results hold in more generality. Define %(r, q) := (r, q) 1 where r, q S and
denotes minimum. By [Mun00], % induces the same topology as . Consequently, it
follows that (S, %) is also a complete, separable metric space. Denote Mf (S) as the
space of finite, Borel measures on (S, %). Many of the later applications require a
metric on Mf (S) to derive certain a priori estimates. To introduce such metrics, the
Wasserstein distance must be defined. For , Mf (S), we define C(, ) as the set
of joint representations of and . This means that Q C(, ) is a measure on the
11
inf
(2.1)
QC(,)
kf kL :=
sS
sup
r,qS,r6=q
|f (r) f (q)|
%(r, q)
.
(2.2)
f (, ) :=
sup
kf kL, 1
Z
f (r)( )(dr)
then, f (, ) = W1 (, ).
12
(2.3)
:=
N
X
)
ai ti , N N, ti T and ai [0, ) Q, i = 1, . . . , N
i=1
(2.4)
By Theorem 6.7, Mf,d forms a countable, dense set in Mf (S) for f . Similarly,
N
X
restricting Mf,d to elements with
ai = 1 is a dense set for (P1 , W1 ).
i=1
(S, R). To
(S, R), and note that Mf (S) is a cone in CL,
Denote this dual by CL,
show that (Mf (S), f ) is complete, it suffices to show that it is closed in the norm
topology. Suppose n Mf (S) converges in the norm topology to a limit . As
CL,
(S, R) is a Banach space, convergence in the norm topology implies convergence
in the weak- topology. By the Riesz Representation Theorem, it follows that must
be a finite, Borel measure, and thus Mf (S). The other case follows by the
Kantorovich-Rubinstein Theorem.
13
2.2
Incompleteness Issues
Denote the space of finite, Borel signed measures on (S, %) by Mf,s (S). From
(2.3), the definition of f (, ) easily extends from Mf (S) to Mf,s (S). The extension,
(Mf,s (S), f ), is a normed vector space, and f is called the Kantorovich-Rubinstein
distance. Due to its convenient dual form, the Kantorovich-Rubinstein distance is
used throughout the analysis. One might expect that a theorem similar to Theorem
2.2 holds for (Mf,s (S), f ). However, the following result shows that the signed
measures under f do not have the same properties.
Theorem 2.3 Mf,s (S) is not complete with respect to f , but (Mf,s (S), f ) is separable.
Proof: Fix s S and k N choose tk T such that %(tk , s) <
1
. For all n N,
k2
We verify that {n }
n=1 is Cauchy with respect to f . For m, n N with m n and
f CL, (S, R) with kf kL, 1:
Z
f (r)(n (dr) m (dr))
m
n
X
X
(1)k f (tk )
= (1)k f (tk )
k=1
k=1
n
X
k
=
(1) f (tk )
k=m+1
n1
X
|f (tk ) f (tk+1 )|
k=m+1
14
n1
X
k=m+1
%(tk , tk+1 )
Thus, f (n m ) :=
n1
X
k=m+1
n1
X
k=m+1
n1
X
1
1
+
)
2
k
(k + 1)2
2
0 as n, m
2
k
k=m+1
sup
kf kL, 1
Z
f (r)(n m )(dr) 0 as n, m .
Consequently, {n }
n=1 is Cauchy with respect to f . Suppose there exists a finite
signed measure, , such that f (n ) 0 as n . Then, f CL, (S, R)
Z
with kf kL, 1,
f (r)(n )(dr) 0 as n . Yet, consider f 1
CL, (S, R). For this choice of f ,
Z
f (r)(n )(dr)
Z
n
X
= 1(r)(n )(dr) = (1)k (S)
k=1
|(S)|
if n = 2` for some ` N.
:=
N
X
)
ai ti , N N, ti T and ai Q i = 1, . . . , N
(2.5)
i=1
Theorem 6.7 shows that Mf,s,d is dense in Mf,s (S) under f . Since Mf,s,d is countable,
the claim follows.
The Kantorovich-Rubinstein distance, f , fails completeness on Mf,s (S), but its
restriction to Mf (S) is complete. From this remark, one might believe that using the
Hahn-Jordan decomposition of a signed measure may yield completeness on Mf,s (S).
15
(2.6)
f,s (S) := {
M
= (1 , 2 ) : 1 , 2 Mf,s (S)}.
f (S) (respectively, M
f,s (S)) the space of measure pairs (respectively, signed
Call M
f,s (S) is a real vector space with componentwise addition
measure pairs). Note that M
f (S) is a cone in M
f,s (S) as M
f (S) is closed
and scalar multiplication. Hence, M
under componentwise addition, but only scalar multiplication for positive scalars is
f,s (S):
well-defined. Define the product metric for
= (1 , 2 ), = (1 , 2 ) M
f (
, ) := f (1 , 1 ) + f (2 , 2 )
(2.7)
16
2.3
New Results
n converges in
n := (y + vn , un ). We claim that
this, it suffices to show that un x under f . The other terms follow analogously.
f (un , x ) =
sup
kf kL, 1
sup
Z
f (r)(un x )(dr)
|f (un ) f (x)| %(un , x) 0 as n
kf kL, 1
n (y + x , x ) as n whereas (y , 0)
is the Hahn-Jordan decomposition of the signed measure (y + x , x ), and 0 is the
measure which assigns 0 to all Borel sets of S.
As a result, (Mf,s (S), f ) is not a complete space. Indeed, one cannot identify
limits under f as signed measures due to the Hahn-Jordan representations. However,
the set of signed measures plays a special role in the product space. One can identify
f (S). Let
the signed measures with a quotient space under M
f,s (S) : = in Mf,s (S)}.
s := {(, ) M
D
17
f,s (S)
M
/Ds .
f,s (S)
its equivalence class, (
), by : M
f,s (S)
M
/Ds .
(2.8)
1 2 = 1 2 in Mf,s (S)
and define the same signed measure.
This important relationship between the signed measures and the quotient space
is crucial in the following analysis. Identifying a signed measure with its equivalence
class avoids the Hahn-Jordan preservation issues apparent in Example 2.5. Futhermore, the relationship provides the correct setting for a well-defined, KantorovichRubinstein-type metric on the space of signed measures. Prior to defining such a
metric, we make one more observation regarding the Hahn-Jordan decomposition in
the quotient-space setting. Let
= (+ , ) be the Hahn-Jordan decomposition of
f (S), then we
a signed measure . Note that if also = 1 2 where (1 , 2 ) M
have that for some Mf,s (S):
=
+ = 1 2
:= 2 = 1 +
2 = +
(2.9)
1 = + + .
The supports of + and are disjoint up to a -null set. Thus, Mf (S) since
otherwise 1 or 2 would be a signed measure. Extend the Hahn-Jordan decomposi-
18
tion for the case when 1 = 2 = by setting (, ) = (0, 0). This implies that for
f (S) that
all
M
+
(
) =
+ D
(2.10)
f,s (S)
M
/Ds .
f (S)
For
, M
((
), (
)) := inf [
f (
) f (
)].
+
D
(2.11)
((
), (
)) +
> f (
d1 ) f (
d1 )
2
+ > f (
d2 ) f (
d2 ).
((
), ())
2
+ . Since for nonnegative constants a, b, c, d, we have that:
Note that d1 + d2 D
(a + b) (c + d) (a c) + (b d),
19
it follows that:
f (
((
), ())
(d1 + d2 )) f ( (d1 + d2 ))
[
f (
d1 ) + f (
d2 )] [
f (
d1 ) + f (
d2 )]
[
f (
d1 ) f (
d1 )] + [
f (
d2 ) f (
d2 )]
+ .
< ((
), (
)) + ((
), ())
2
((
), (0)) = f (
)
(2.12)
and
f ((
) (
)) ((
), (
)) f (
) ((
), (
)) f (
).
(2.13)
20
sup
f (r)( + )(dr) +
kf kL, 1
sup
f (r)( + )(dr)
kf kL, 1
+
D
f (
) + 2(S) = f (
)
))(dx)
Z
+
= f (x) ( ) (dx)
Z
+
+
= f (x) ( ) (dx)
21
Z
Z
+
+
f (x)( )(dx) + f (x)( )(dx)
sup
kf kL, 1
Z
f (x)(+ + )(dx)
Z
+ f (x)( )(dx)
f (
)
where = (, ). Thus, taking the supremum over f CL, (S, R) with kf kL, 1
yields
f ((
) (
)) f (
).
Since Mf (Rd ) was arbitrary, it follows from the preceding inequality that
f ((
) (
)) inf f (
)
+
D
f ((
) (
)) inf [
f (
) f (
+ )] = ((
), (
)).
+
D
f ((
) (
)) = f (+ ( + )) f (+ ) + f ( + ) = f (
).
The second inequality follows immediately, and the third follows simply by choosing
= (0, 0) in (2.11).
22
The motivation for the work was to construct a metric that is complete on Mf,s (S).
We next examine whether is a complete metric on the space of signed measures.
Suppose {(
n )}nN is a Cauchy sequence for with a subsequence {(
nk )}. If
{(
nk )}nN converges, then it would follow that (Mf,s (S), ) is complete. However, the following analysis shows that this subsequence does not converge in general.
Rather, the following derivation yields something that is in a way more powerful than
completeness itself.
Theorem 2.8 A Cauchy sequence {(
n )nN } converges in iff a subsequence of
f ) whose limit,
, satisfies
measure pairs {
=
In particular, the limit,
, satisfies the identification of the Hahn-Jordan
decomposition of a signed measure.
Proof : For the reverse implication, note that by Proposition 2.7,
((
nk ), (
)) f (
) 0 as k .
nk ,
So, {(
n )nN } must converge as it is Cauchy.
For the forward implication, choose a subsequence {(
nk )}kN of {(
n )}nN such
that
((
nk ), (
nk+1 )) < 2k k N.
+ such that
Set 1 := 0 and by the definition of choose k D
f (
f (
nk
nk +1 + k ) < 2k k N.
nk
nk +1 k )
23
+ by k :=
Further, define another sequence in D
k1
X
i=0
f (
k N.
nk + k
nk +1 k+1 ) < 2
Setting k :=
nk + k Mf (S), we obtain
f (k k+1 ) < 2k .
D
f (S), f ) is complete and by (2.10), there is a unique = + ,
+
Since (M
such that
f
+ .
k =
nk + k
sup f (
kN
+.
i exists and D
i=1
Further, for m N,
X
X
X
= 2f (
f (
i )
i )
2i (S) 0 as m .
i>m
im
i>m
Thus,
X
0 as m .
f (
i )
im
24
:= +
nk + Mf (S). Denote
f (S) and D
+ such that
class of this limit and using (2.10), there are
M
Thus, we obtain
+ = + .
+ .
nk
(2.14)
To complete the proof of Theorem 2.5, it suffices to show the following lemma:
Lemma 2.9
0 iff = 0
+ ))
((
nk ), (
where 0 = (0, 0).
Proof : The reverse implication follows from Proposition 2.7 as
((
)) f (
)
nk ), (
nk
and
in f as k .
nk
> 0. Thus,
For the forward implication, it follows that if 6= 0, then f ()
((
)) inf f (
) inf f (
).
nk ), (
nk
nk
+
D
+
D
) = inf f (
+
nk
nk
+
D
+
D
inf f ( ) f (
+
nk )
+
D
inf f ( + )
+
D
25
(2.15)
nk ) 0. Hence,
> 0.
) inf f ( + ) = f ()
+
D
> 0.
lim inf ((
nk ), (
)) f ()
n
26
3.1
Definitions
Prior to stating the SODE-system or SPDE, we state some of the standard assumptions and notation for the stochastic analysis. Consider a probability space,
Z
(, F, P ), and a filtration {Ft }t0 contained in F. Denote E() and ()dP as the
mathematical expectation with respect to the probability measure. To avoid certain
pathological issues that may arise, we make the following standard assumptions on
the filtration:
F0 contains all the P -null sets (and hence so does Ft for all t 0).
The filtration is right continuous (i.e. Ft = Ft+ := st Fs ).
One reason for making such assumptions on the filtration is for convenience when
working with stopping times. A stopping time is a mapping, : [s, ), such
27
that the event { t} Ft for all t 0. Assuming that the filtration is right
continuous implies that the events { < t} Ft for a stopping time, .
For notation, we need to define the standard Lp spaces for p 1. Let K be
a metric space with metric . L0,Fs (K) is the space of Kvalued Fs -measurable
random variables . Lp,Fs (K) L0,Fs (K) is the set such that for Lp,Fs (K),
E p (, ) < where K is an arbitrary fixed element.
Denote, C([s, T ]; K) as the set of continuous functions from [s, T ] to (K, )
where s, T [0, ). L0,F (C([s, T ]; K)) is the space of random variables with values in C([s, T ]; K) which are adapted to the filtration Ft . Lp,F (C([s, T ]; K))
L0,F (C([s, T ]; K)) is the space of p-th integrable random variables with values in
C([s, T ]; K). Finally, Lloc,p,F (C([s, T ]; K)) is the space of processes () such that
there is a sequence of stopping times {n }nN with ( n ) Lp,F (C([s, T ]; K)) and
n . Such a sequence of stopping times is said to be localizing.
To aid in the notation, let f,s (, ) denote the partially-complete metric from
Chapter 2.14 Let wl (dp, dt) denote i.i.d. standard Gaussian space-time white noise
differentials15 where l = 1, . . . , d, and define w(dp, dt) = (w1 (dp, dt), . . . , wd (dp, dt))T .
Now, let F : Rd Mf,s (Rd ) [0, ) Rd , J : Rd Rd Mf,s (Rd ) [0, )
Mdd , where is the correlation length,16 and Mdd denotes the set of d d matrices
over R. We can now state the SODE system in differential form.17 For i = 1, . . . , N
N, and s t T with s, T [0, )
dri
Z
+
14
(3.1)
28
(3.2)
and Y(t) is a signed-measure valued, adapted process. One may think of the system
as representing N particles moving in a random medium with positions given by ri .
The signed measure input process is represented by Y(t) in Rd , and it provides the
mathematical interpretation of the random medium.
3.2
To address questions of existence and uniqueness to (3.1), one must specify more
conditions on the coefficients, F and J . Assume that F and J are jointly measurable
in all arguments. Also, assume that the one dimensional components of J , denoted
by J, ij for i, j = 1, . . . , d, are square integrable with respect to Lebesgue measure, dq,
on Rd . Also, assume the following Lipschitz and growth constants on the coefficients
in terms of f,s and %(r, q) := |r q| 1 where18 | | denotes the Euclidean metric on
Rd and r, q Rd .
f,s (Rd ), t s
There exists19 a cF,J (0, ) such that for all r, ri Rd ,
,
i M
and i = 1, 2:
|F (r1 , (
1 ), t) F (r2 , (
2 ), t)| cF,J {(
f (
1 ) f (
2 ))(r1 , r2 ) + f,s (1 , 2 )}
(3.3)
18
We write %(q) for %(q, 0) and %(r q) for %(r, q) without further mention.
Throughout this work, we use the notation that the subscripts on a constant indicate dependence
on those parameters.
19
29
d Z
X
(J,k` (r1 , p, (
1 ), t) J,k` (r2 , p, (
2 ), t))2 dp
k,`=1
2
2
c2F,J (
f (
1 ) f2 (
2 ))2 (r1 , r2 ) + f,s
(1 , 2 )
(3.4)
|F (r, (
), t)|2 +
d Z
X
2
J,k`
(r, p, (
), t)dp cF,J f2 (
).
(3.5)
k,`=1
The constant cF,J in (3.3)-(3.5) may also depend on the space dimension, d. Alternatively, if we relax the boundedness assumption on the coefficients in (3.5), we need
to impose a linear growth condition. This growth condition is assumed in addition
to the corresponding Lipschitz conditions from (3.3)-(3.4) in terms of the Euclidean
metric.
f,s (Rd ), t s
There exists a cF,J (0, ) such that for all r, ri Rd ,
,
i M
and i = 1, 2:
|F (r1 , (
1 ), t) F (r2 , (
2 ), t)| cF,J {(
f (
1 ) f (
2 ))|r1 r2 | + f,s (1 , 2 )}
(3.6)
d Z
X
(J,k` (r1 , p, (
1 ), t) J,k` (r2 , p, (
2 ), t))2 dp
k,`=1
2
2
c2F,J (
f (
1 ) f2 (
2 ))|r1 r2 |2 + f,s
(1 , 2 )
(3.7)
|F (r, (
), t)|2 +
d Z
X
2
J,k`
(r, p, (
), t)dp c2F,J (1 + |r|2 )
f2 (
).
(3.8)
k,`=1
With the appropriate Lipschitz assumptions on the coefficients, the following existence
and uniqueness result follows the standard techniques in the literature.20 However,
due to its importance in the later analysis, the complete proof of the result is given.
20
30
The case for positive measures is established in [Kot08]. The following argument
follows the work in [Kot08] and generalizes it to signed measures. In the following
propositions, we simply write the localizing stopping times {n }n1 as . Furthermore,
let % {%, | |}.
Proposition 3.1 Suppose either (3.3)-(3.5) or (3.6)-(3.8) holds:
i. Let the initial condition, qs = (qs1 , . . . , qsN ) have qsi L2,Fs (Rd ) for i = 1, . . . , N ,
and the signed valued input process have, Y Lloc,2,F (C((s, T ]; Mf,s (Rd ))). There
exists a unique, continuous, adapted solution of (3.1):
qs , s) = (r1 (t, Y,
qs1 , s), . . . , rN (t, Y,
qsN , s) Lloc,2,F (C([s, T ]; Rd )).
r (t, Y,
ii. For initial conditions, qji L2,Fs (Rd ), i = 1, . . . , N, j = 1, 2, signed measure valued processes Yj Lloc,2,F (C((s, T ]; Mf,s (Rd ))), j = 1, 2 and a localizing stopping
time, :
E
sup
stT
Z
i
2 i
cT,F,J , E (q1 , q2 )1{ s} + E
2
f,s
(Y1 (u), Y2 (u))1{ s} du.
(3.9)
iii. For any N N there is a RdN -valued map in the variables (t, , , r, s), 0 s
t < such that for any fixed s 0:
such that:
(a) For any t [s, T ], r (t, , . . . , , s) is Gs,t FMf,s (Rd ),s,t B dN B dN measurable. Here, Gs,t is the completed -algebra generated by w(dp, du) between s
and t, FMf,s (Rd ),s,t denotes the -algebra of cylinder sets on Mf,s (Rd ), and
B dN denotes the Borel sets on RdN .
(b) The ith component of r = (r1 , . . . , rN ) depends only on the ith initial condition, w(dq, dt), and Y L2,F (C([s, T ]; Mf,s (Rd ))):
qsi , s).
(r (t, , Y , qsi , s))i = ri (t, Y,
Proof: (i). Without loss of generality, we will prove the result on [0, T ]. For notational convenience, we will suppress the dependence on and the particle index in
the argument. Assume (3.3)-(3.5) as the other case will follow similarly. Since the
stopping times are localizing, we have that for all > 0 and for all b > 0, there is a
stopping time, , and an F0 measurable set b such that:
P
sup
0tT
f,s (Y1 (t, )) f,s (Y2 (t, ))1b () b 1 .
(3.11)
0tT
Similarly, the same holds with f replacing f,s . To show existence, the approach
is to use completeness and the Picard-Lindel
of approximation procedure. To use
this approach, we need to show that the approximations lie in the Banach Space
32
of L2,F (C([0, ); Rd )). To this end, assume that qj ( ) L2,F (C([0, ); Rd )) for
j = 1, 2, and define:
Z
qj (t) := qj (0) +
Z tZ
(3.12)
For j = 1, 2,
2
Z
E sup (
qj (t )) 9E (qj (0)) + 9E sup
0tT
0tT
Z
+ 9E sup
0tT
(3.13)
By Cauchy-Schwarz,21 (3.5), and (3.11), we have that there is a constant cF,J ,T
such that:
E sup
0tT
Z
Z
E sup T
0tT
1ds cF,J ,T .
0
(3.14)
Applying Doobs Inequality,22 (3.5), and (3.11) to the random term yields another
21
22
33
E sup
Z
0tT
d
X
d
X
k=1
!2
J,ki (qj (s), p, Yj (s), s)
dpds
i=1
(3.15)
2
4d
d
X
k,i=1
Z
E
2
J,ki
(qj (s), p, Yj (s), s)dpds
cF,J ,T .
where we have used (6.15) in the first inequality. Combining (3.13), (3.14), and (3.15)
yields that there is a finite constant, cF,J ,T,d, , such that:
E sup 2 (
qj (t )) 9E 2 (qj (0)) + cF,J ,T,d, .
(3.16)
0tT
Thus, we have shown that if the initial condition satisfies qj (0) L2,F0 (Rd ), then
qj (t ) defined by (3.12) is in L2,F (C([0, ); Rd )). Now, we state estimates to
compare terms of the type defined by (3.12). We may assume that cF,J 1, and note
that a c (a 1)c for all a 0 and c 1. Thus,
s) F (q2 (s), Y2 (s), s) cF,J F (q1 (s), Y1 (s), s), F (q2 (s), Y2 (s), s) .
F (q1 (s), Y(s),
(3.17)
34
Z
2
F (q2 (s), Y2 (s), s)ds
(3.18)
T
0
Z
cF,J ,Y
2
f,s
(Y1 (s), Y2 (s))ds
.
Z
2
d X
d Z t Z
X
E
(J,ki (q1 (s), p, Y1 (s), s) J,ki (q2 (s), p, Y2 (s), s))wi (dp, ds)
0
k=1
d
X
i=1
k=1
d E
2
Z X
d
(J,ki (q1 (s), p, Y1 (s), s) J,ki (q2 (s), p, Y2 (s), s)) dpds
i=1
d Z
X
k,i=1
|(J,ki (q1 (s), p, Y1 (s), s) J,ki (q2 (s), p, Y2 (s), s)|2 dpds
Z
cF,J ,Y,d E
t
2
t
2
f,s
(Y1 (s), Y2 (s))ds
(3.19)
.
It now follows from (3.18), (3.19) and Doobs Inequality that there exist a constant
35
(
q1 (t ), q2 (t )) cF,J ,T,,d E 2 (q1 (0), q2 (0))
2
0tT
(q1 (s ), q2 (s ))ds + E
+E
0
2
(Y1 (s
f,s
), Y2 (s ))ds .
(3.20)
From [Arn92], apply the Picard-Lindel
of procedure. Define iteratively for n N
:= Y1 () Y2 (),
with Y()
Z
qn+1 (t) := q(0) +
F (
qn (s), Y(s),
s)ds +
Z tZ
J (
qn (s), p, Y(s),
s)w(dp, ds)
q0 (t) := q(0).
(3.21)
By (3.20), it follows that with cN := cF,J ,T,,d :
E sup
(
qn+1 (t ), qn (t )) cN
0tT
E sup 2 (
qn+1 (s), qn (s))ds.
(3.22)
0tT
E sup
2 (
qn+1 (t ), qn (t ))
0tT
cnN T n
.
n!
(3.23)
From (3.16), qn lies in the complete space, L2,F (C([0, ); Rd )). So, by (3.23) it
follows that the sequence converges uniformly to a limit, q , in mean square on the
compact interval [0, T ]. By the construction, q is continuous as each qn is continuous.
36
Z
Z
|
q (s) qn (s)| ds lim sup E
2
T
2
|
qm (s) qn (s)| ds
0
(3.24)
F (
qn (t), Y(t), t)dt
0
F (
q (t), Y(t), t)dt
(3.25)
J (
q (t), p, Y(t), t)w(dp, dt)
(3.26)
J (
qn (t), p, Y(t), t)w(dp, dt)
0
2 (
q (t ), r (t ))
0tT
Z
cF,J ,T,,d E
(
q (s ), r (s ))ds + E
Z
f,s
(Y(s),
Y(s))ds
(
q (s ), r (s ))ds .
= cF,J ,T,,d E
0
(3.27)
Applying Gronwalls Inequality23 yields that:
E sup
2 (
q (t ), r (t )) = 0.
0tT
23
37
(3.28)
dri
Z
+
(3.29)
with
XN (t) :=
N
X
ai ri (t)
(3.30)
i=1
38
(3.31)
1iN
|XN |(0, , R ) :=
N
X
i=1
N
X
|ai | b a.s..
i=1
Following the argument for the decoupled case, define for j = 1, 2, i = 1, . . . , N and
nN
i
qj,n+1
(t)
:=
i
qj,n
(0)
Z
+
Z tZ
t
i
F (qj,n
(s), Xj,n (s), s)ds
i
J (qj,n
(s), p, Xj,n (s), s)w(dp, ds)
(3.32)
i
i
and qj,0
= qN
(0) where
Xj,n (t) :=
N
X
i (t) .
ai qj,n
(3.33)
i=1
Note that the arguments from (3.18)-(3.20) hold in the coupled case with Yj (t) =
Xj,n (t). However, to compare the qj,n (), one must derive the relationships between
the associated empirical measures, (3.33). As these empirical measures are signed
39
+
Xj,n
(t)
:=
N
X
i (t)
ai 1{ai >0} qj,n
(3.34)
i (t) .
ai 1{ai <0} qj,n
(3.35)
i=1
and
Xj,n
(t) :=
N
X
i=1
+
+
= f X1,n
(t), X2,n
(t) + f X1,n
(t), X2,n
(t)
Z
=
sup
kf kL, 1
!
+
+
f (s)d(X1,n
(s) X2,n
(s))
Z
+
sup
kf kL, 1
sup
N
X
kf kL, 1 i=1
f (s)d(X1,n
(s) X2,n
(s))
!
i
i
ai 1{ai >0} f (q1,n (t)) f (q2,n
(t))
sup
N
X
kf kL, 1 i=1
40
!
ai 1{ai <0}
i
i
f (q1,n
(t)) f (q2,n
(t))
N
X
i
i
(t)
(t), q2,n
ai 1{ai >0} % q1,n
i=1
N
X
i
i
ai 1{ai <0} % q1,n
(t), q2,n
(t)
i=1
N
X
i
i
|ai |% q1,n
(t), q2,n
(t)
i=1
(3.36)
b%N (q1,n (t), q2,n (t)) .
E sup
0tT
i
i
(t))
(t), q2,n+1
(q1,n+1
cF,J ,T, E%2 (q1i (0), q2i (0))
(3.37)
+E
where cF,J ,T, and cb are nonnegative constants depending on the associated parameters. From which it follows that
E sup
0tT
cF,J ,T E%2N (q1 (0), q2 (0))
(3.38)
Z
+E
Consequently, using the analogous arguments as (3.16)-(3.28), one can derive the existence of a unique, continuous solution by the contraction mapping principle.24
24
41
3.3
SPDE Results
With the existence and uniqueness of solutions to (3.1) and (3.29), we can now
examine a general type of SPDE. As we will soon derive, the empirical processes of
(3.1) and (3.29) automatically yield weak solutions of a SPDE. Prior to stating the
SPDE, one also needs to define the one and two particle diffusion matrices. These
matrices represent a measure of correlations in the stochastic term in (3.1) and (3.29).
Define the diffusion matrices as follows. For l, k = 1, . . . , d, i, j = 1, ..., N, ri , rj
R2 , Mf,s (Rd )
kl (ri , rj , , t) :=
D
d Z
X
q=1
(3.39)
and
i , ri , , t).
D(ri , , t) := D(r
42
!
N
N
N
X
X
X
i
i
dhXN (t), i = dh
ai r (t) , i = d
ai (r (t)) =
ai d(ri (t))
i=1
N
X
ai ()(ri (t))
i=1
d(ri )
i=1
i=1
N
d
1X X 2
kl (ri (t))d << ri,k (t), ri,l (t) >>
ai
+
2 i=1 k,l=1
N
X
i=1
N
X
ai ()(ri (t))
i=1
(3.40)
N
d
1X X 2
+
ai
(ri (t))d << ri,k (t), ri,l (t) >>
2 i=1 k,l=1 kl
(3.41)
as any terms with finite total variation in (3.29) do not contribute to (3.41). Combining (3.40) and (3.41) yields:
43
N
X
i=1
N
X
ai ()(ri (t))
i=1
d
N
1X X 2
ai
kl (ri (t))Dkl (ri , XN , t)dt
2 i=1 k,l=1
d
1X 2
+ XN (t),
( )()Dkl (, XN , t)dt
2 k,l=1 kl
d
X
k=1
d
X
Z
XN (t)
J,k (, p, XN (t), t)w(dp, dt), k
(3.42)
k=1
X
d
1
2
+
Dkl (, XN , t)XN (t)dt, kl .
2 k,l=1
44
(3.43)
Z
XN (t) J (, p, XN (t), t)w(dp, dt) , .
dY =
!
d
1X 2
YDkl (, Y, t) 5 (YF (, Y, t)) dt
2 k,l=1 kl
(3.44)
Z
5 Y J (, Y, p, t)w(dp, dt)
N
X
i=1
3.4
Extension by Continuity
Proposition 3.3 inspires the following method and representation. The empirical
process associated with (3.29) satisfies (3.44). Thus, one can construct a weak solution
of (3.44) provided that the initial signed measure, Ys , is discrete. One approach to
construct solutions for arbitrary signed measures is Extension by Continuity. This
method approximates arbitrary signed measures by discrete signed measures in an
45
appropriate metric. One then shows that the solutions associated with the discrete
signed measures converge. Finally, one shows that this limit is the solution of (3.44)
with the arbitrary initial signed measure.
The first step in the methods argument is justifying approximations of initial
signed measures by discrete signed measures. To show such a statement, it is necessary to derive certain a priori estimates on the empirical processes. These estimates
show that the empirical processes associated with solutions of (3.1) or (3.29) can be
estimated in terms of the initial signed measures. Again, to simplify notation, we
assume that s = 0.
Proposition 3.4 Suppose for i = 1, 2, Xi (0) L2,F0 (Mf,s (Rd )) are the initial signed
measures and Y Lloc,2,F (C((s, T ]; Mf,s (Rd ))) is the signed-measure input process.
Further, Yi () are the empirical processes, given by (3.30) associated with Xi (0) for
X1 (0) and X2 (0). Then
i = 1, 2. Finally, let be a localizing stopping time for Y,
there is a constant cT,F,J , such that
E
sup
0tT
2
f,s
(Y1 (t), Y2 (t))1{ >0}
cT,F,J , E f2 (X1 (0), X2 (0))1{ >0} .
(3.45)
Proof: The following proof is claimed in [KS12a] and follows the measure case established in [Kot08]. Suppose that the initial distributions are non-random. Once we
establish the result under this case, we can condition on the -algebra, F0 , to establish the general claim. Further, assume that Yi (t) has Hahn-Jordan decomposition
Yi (t) with measure Yi (0, Rd ) = m
i 0 for i = 1, 2.
46
By Proposition 2.7:
E
sup
0tT
2
(Y1 (t), Y2 (t))
f,s
sup
0tT
2E
sup
0tT
+ 2E
(3.46)
sup
0tT
We focus on the first term as the second will follow by a similar argument. The
following argument is based on a relationship between the Kantorovich-Rubinstein
distance and the Wasserstein distance. These details are given in Proposition 6.8. By
this proposition, we have the following.
E
sup
0tT
E
sup
0tT
(m+
1
Y1+ (t)
m+
)W
(
,
1
2
m+
1
2
Y2+ (t)
+
+
) + |m1 m2 |
m+
2
sup
0tT
(m+
1
(m+
1
Y1+ (t)
+
m2 )W1 ( + ,
m1
2
m+
2) E
2
Y2+ (t)
)
m+
2
Z Z
sup
0tT
inf
X + (0) X + (0)
Q+ C( 1 + , 2 + )
m1
m2
47
2
q), r(t, Y,
q))Q+ (dq, d
%(r(t, Y,
q)
Now, let Q+ C(
E
sup
0tT
(m+
1
(m+
1
(m+
1
be arbitrary.25
Y1+ (t)
m+
)W
(
,
1
2
m+
1
2
m+
2) E
2
m+
2) E
Z Z
sup
2
Y2+ (t)
)
m+
2
2
0tT
Z Z Z Z
sup
q), r(t, Y,
q))
%(r(t, Y,
0tT
p), r(t, Y,
p))Q+ (dp, d
%(r(t, Y,
p)Q+ (dq, d
q)
sup
q), r(t, Y,
q))%(r(t, Y,
p), r(t, Y,
p))Q+ (dp, d
%(r(t, Y,
p)Q+ (dq, d
q)
0tT
Z Z Z Z r
E
q), r(t, Y,
q))
sup %2 (r(t, Y,
0tT
r
E
p), r(t, Y,
p))Q+ (dp, d
sup %2 (r(t, Y,
p)Q+ (dq, d
q)
0tT
Z Z Z Z p
p
%2 (q, q) %2 (p, p)Q+ (dp, d
p)Q+ (dq, d
q ).
cT,F,J ,
48
that
E
sup
0tT
(m+
1
X1+ (0)
+
m2 )W1 (
,
m+
1
2
X2+ (0)
)
m+
2
+ 2
2
cT,F,J , (m+
1 m2 ) W1
+
Yet, by Proposition 6.8, adding |m+
1 m2 | to the right side represents a metric
sup
0tT
cT,F,J , E f2 (X1+ (0), X2+ (0)) .
(3.47)
The term for the negative components of the signed measures follows similarly. Combining these terms yields the claim.
The technique of extension by continuity is employed in [Kot95], [KX99], [MP82],
and [KK10]. We shall comment on [MP82] and [Kot95] in Chapter 4 as an application
of the general results. However, the other articles have generated general results
in signed measure valued SPDE. In [KX99], the authors focus on a more general
class of SPDE. In particular, the associated SODE (3.1) (or (3.29)) includes another
random term driven by independent Brownian motions as an external noise term.
Furthermore, the case allows for the particle weights, ai , to be time dependent and
not constant. However, there are several drawbacks from the approach in [KX99]:
The authors use that the initial conditions of the weights and the positions are
exchangeable in addition to square integrable and measurable.
The authors assume Lipschitz and growth conditions such as (3.3)-(3.5) (3.6)(3.8) for every possible representation of a signed measure as the difference
of two measures. This assumption is not only difficult to verify, but such a
strong assumption implies that the assumptions are not the ideal setting for the
49
analysis.
Despite these drawbacks, the authors derive existence and uniqueness result for their
SODE system, and similarly pass the solution to a solution of an SPDE by Itos
Formula.
Following the results from [KX99], [KK10] uses the results on exchangeable pairs
to derive limit results for deterministic PDE. In particular, the authors examine the
same class of SPDE as in this work, but with a different result in mind. The authors
derive macroscopic limit results for solutions of SPDE. That is, the correlation length,
, converges to 0, and consequently, the limit of the solutions of SPDE converge to a
solution of a deterministic PDE.
3.5
Hahn-Jordan Decompositions
Another important work intimately related to this thesis is [Kot10]. In this work,
the author examines the same SODE system and class of SPDE, and derives the
following interesting result.
Proposition 3.5 Assume that the coefficients of the SODE systems, (3.1) and (3.29),
satisfy either the Lipschitz conditions, (3.3)-(3.4) or (3.6)-(3.7). Then, the particles
satisfying (3.1) or (3.29) do not coalesce in finite time with probability one. That is:
(3.48)
With this proposition and some additional hypothesis, the author establishes that
a solution to the SPDE, (3.44), preserves the Hahn-Jordan decomposition of the initial
empirical process. An important part of the argument in [Kot10] relates solutions
50
(3.49)
(, q) 7 r(t,,q) , Rd 7 Mf (Rd ),
(3.50)
is Ft B d BMf (Rd ) measurable, where BMf (Rd ) is the Borel algebra on (Mf (Rd ), f ).
The following proposition from [Kot10] shows the definition and the importance of
the flow representation.
Proposition 3.6 Suppose X0 L2,F0 (Mf,s (Rd )) and Y(t, ) Mf,s (Rd ) is defined
by26
Z
r(t,,q) X0 (dq, ).
(3.51)
Then, Y(t, ) is a solution to the SPDE, (3.44), where Y replaces Y in the arguments
of D, F , and J .
Proof: The proof is given in [Kot10]. We reproduce it to fill in the details for
the argument in this thesis. To simplify notation, denote r(t, q) = r(t, , q) and
Z
The right side is by definition the image of the initial measure X0 under the flow q r(t, , q).
51
in Mdd . Set
Z tZ
(t) :=
n (p)w(dp, ds).
0
The n () are i.i.d. standard Rd -valued Brownian motions by [Kot08]. Also, note
that we have the following
Z
n=1
where
Z
J (r, p, , t)
n (p)dp.
n (r, , t) :=
Z
(r)Y(t, dr) =
Z
=
Z
(r)
Z Z
(q)X0 (dq) +
0
Z Z
+
0
d Z Z t
1X
2
(kl
)(r(s, q))Dkl (r(s, q), Y(s),
s)X0 (dq).
+
2 k,l=1
0
X
d
X
n=1 l=1
52
I3 (t) =
0
X
d Z Z
X
X
d Z tZ
X
n=1 k,l=1
r(s,q) X0 (dq)
X
d Z
X
n=1 k,l=1
n=1 k,l=1
X
d Z
X
n=1 k,l=1
=<
Z tX
d
0 k=1
Z
=<
Y(s)
X
d
X
n,kl (, Y(s),
s) ln (ds), () >
n=1 l=1
Z
53
<
Z
<
Z
d Z
1X t 2
+<
(Y(s)Dkl (, Y(s),
s)ds), () > .
2 k,l=1 0 kl
Yet, note that this is a weak form of (3.44) with Y replacing Y in D, F , and J .
The empirical process associated with the SODE system, (3.1), has such a flow
representation given by
N
X
Z
ai ri (t,,Y(t),X
=
0)
r(t,,Y(t),q)
X0 (dq, ).
i=1
(3.52)
Further, the empirical process has a Hahn-Jordan decomposition given by the following.
r(t,,Y(t),q)
X0 (dq, ).
(3.53)
3.6
New Results
With this approach in mind, we wish to generalize the results of [Kot10]. With
the apriori estimates given in Proposition 3.4, we are now ready to address the role
of the Hahn-Jordan decomposition. Proposition 3.5 implies that the Hahn-Jordan
decomposition of the empirical process is preserved through the flow representation,
(3.53), provided the initial empirical process, X0 , is discrete. If the particles driving
the SODE system never coalesce, then the weights, ai , in the definition of XN (t)
cannot change values. We now extend this result to arbitrary initial signed measures.
Corollary 3.7 The Hahn-Jordan decomposition of the initial condition is preserved
in the solution of (3.44) for all t > 0 for arbitrary F0 -measurable X0,N . That is, for
the flow representation,
Z
r(t,,q) X0,N (dq, )
(3.54)
the Hahn-Jordan decomposition can be found by the flow representation against the
Hahn-Jordan decomposition of the initial signed measures
r(t,,q) X0,N
(dq, ).
(3.55)
Proof : Let N N. As mentioned, (3.55) holds for discrete random initial signed
measures, X0,N .
E
f2 (X0,N , X0 ) 0.
55
Truncating the initial signed measure if necessary, we may assume that the initial
signed measures are square integrable with respect to f . By (3.45), for M N
2
Ef,s
(Y(t, X0,N ), Y(t, X0,M )) cT,F,J , E(
f2 (X0,N , X0,M )).
Hence, as N, M
2
(Y(t, X0,N ), Y(t, X0,M )) 0,
Ef,s
estimate shows how solutions depend on their input signed measure processes, Y().
Proposition 3.8 Suppose Y1 , Y2 Lloc,2,F (C((s, T ]; Mf,s (Rd ))). Let Y(t, Y1 ) and
Y(t, Y2 ) be two solutions of (3.44) with X (0, ) := Y(0, Y1 ) = Y(0, Y2 ) and flow
representations, (3.54). Then, T > 0 there is a positive cT < such that
E sup
0tT
2
f,s
Y(t, Y1 ), Y(t, Y2 ) E sup f2 Y (t, Y1 ), Y (t, Y2 )
0tT
Z
cT
T
2
Ef,s
Y1 (s), Y2 (s) ds cT
E
f2
(3.56)
56
Proof : Truncating the initial measure, X (0, ), if necessary, we may without loss of
generality assume that
ess sup
f (X (0, )) c < .
(3.57)
Hence,
2
f,s
Y(t, Y1 ), Y(t, Y2 )
E sup
0tT
E sup
0tT
= E sup
f2
sup
Y (t, Y1 ), Y (t, Y2 )
X Z
0tT kf kL, 1
2
2
X Z
cT
Z
E
(X0+ (Rd ) + X0 (Rd ))2 sup %2 (r(t, Y1 , q), r(t, Y2 , q))X0 (dq)
0tT
T
2
f,s
(Y1 (u), Y2 (u))du
cT E
0
Z
cT E
f2
Y1 (s), Y2 (s) ds
57
(3.58)
Further,
X (t) := X (t, X , X0 ) =
(3.59)
is the Hahn-Jordan decomposition X (t) of X (t) for all t > 0. Thus, the Hahn-Jordan
decomposition is preserved.
Y0 (t) X0 ,
Yn (t)
Z
:=
(3.60)
Without loss of generality, assume that the total variation of initial signed measure
is bounded uniformly in . Due to conservation of the intensities by Proposition 3.5,
58
E sup
0tT
(t)
f2 Yn (t), Ym
cT
E
f2 Yn1
(s), Ym1
(s) ds
Z
cT
E sup f2 Yn1
(s), Ym1
(s) ds.
0st
(3.61)
0tT
X (t) :=
Z
r(t,X ,q) X0 (dq),
X (t) =
(3.62)
E sup
0tT
f2
Z
Yn (t), X (t) cT
E
f2
59
60
4.1
Formulation
We recall the formulation of the SNSE with smoothed Biot-Savart kernel. The
smoothed SNSE are random perturbations of the smoothed Navier-Stokes equations:
(4.1)
where
Z
U (r, t) =
(4.2)
and K () is the smoothed form of the Biot-Savart kernel. That is, for > 0 and
1
< |r| <
1
K (r) = K(r) :=
(r2 , r1 )
2|r|2
where r = (r1 , r2 ) R2 and K C 2 (R2 , R) with bounded derivatives up to order
two. Without loss of generality, we can also assume that K (0) = 0.
61
There is extensive literature on random perturbations of the Navier-Stokes equations since the work of [Cho73]. The results of [Cho73], [Lon88], and [GHL90] are
more numerical analytic than our focus. Consequently, the discussion of these works
in the introduction is sufficient for our purposes. However, the works, [MP82], [Kot95],
[AX06] and [Ami07] form a background to the stochastic Navier-Stokes equations in
this work.
From [Cho73], the authors of [MP82] approach the SNSE by considering a system
of SODE of the form:
dr (t) =
N
X
2d i (t)
(4.3)
k=1
62
2
2
XX
(2a )2 X
= 2
{ (k )2 (rj (t))}dt,
2
N
j=1
k=1
1
, , T ).
N
(4.4)
63
initial conditions towards the initial condition in (4.9), it should follow that:
(4.5)
2 (r, p)dp = 1,
ii,
(4.6)
(4.7)
11, (r, p)
0
(r, p) :=
.
22, (r, p)
0
(4.8)
dX (t) = [X (U X )] dt
Z
(, p) w(dp, dt),
2 X
64
(4.9)
dr (t) =
N
X
ak K (r (t) r (t))dt +
Z
2
(4.10)
k=1
(4.11)
With these assumptions, the author in [Kot95] shows that the smoothed Stochastic
Navier-Stokes Equations have a solution for any adapted initial signed measure.
Finally, we discuss the works of [Ami00], [Ami07] and [AX06]. Recall that [Ami00]
generalized (4.10) so that the vorticity could include jump processes driven by Poisson random measures. The author reexamines the work in [Ami07], and presents the
results in the case of only positive measures due to completeness issues on the space
of signed measures. Finally, Amirdjanova derives a diffusion approximation to the
vorticity model with jump processes. In [AX06], the authors verify a form of exponential tightness for the stochastic vorticity. Such a result is then used to derive a
macroscopic limit theorem. That is, the authors show that as the magnitude of the
stochastic term in (4.10) tends to zero, the solutions of the stochastic equations tend
to a solution of a deterministic equation.
With these results in mind, we now state the problem we wish to analyze in this
work. Following [Kot95], we use the same the SODE system and conditions as in
(4.10) to analyze the SNSE. The issue of existence and uniqueness for the smoothed
stochastic Navier-Stokes equations follows from general results. The existence and
uniqueness results from Chapter 3 yield solutions to the SNSE. What is of greater
65
(4.12)
where a+ , a and a = a+ + a .
To establish the vorticity claim, we must establish existence and uniqueness for
(4.10). However, as we will soon show, (4.9) is a special form of the general SPDE
from Chapter 3. Consequently, we apply the results from the previous chapter to the
SNSE.
Theorem 4.3 To each F0 -adapted initial condition qN (0) R2N , (4.10) has a unique
Ft -adapted solution, qN () C([0, ); R2N ) a.s.
Proof: By Proposition 3.2, it suffices to verify that the coefficients satisfy the desired Lipschitz and boundedness conditions, (3.3)-(3.5) or (3.6)-(3.8). Clearly, the
boundedness properties follow from the definition of the correlation functionals and
the boundedness of K . For the Lipschitz conditions, if {n }nN is a complete orthonormal system in L2,F0 (R2 ) then, define
n 0
n :=
.
0 n
66
We have
Z tZ
Z
X
n=1
Z tZ
where (t) :=
tions by [Kot08]. It follows from the definition of the correlation function that:
2 Z
Z
2
X
Now, we note that the drift coefficient can be represented by F (XN (t), r) : Mf,s
R2 R2 where F (, r) := K (r) and denotes convolution. For 1 , 2
Mf,s , r1 , r2 R2 we have the following:
|F (1 , r1 ) F (2 , r2 )| |F (1 , r1 ) F (1 , r2 )| + |F (1 , r2 ) F (2 , r2 )|
Z
Z
K (r1 q)1 (dq) K (r2 q)1 (dq)
Z
Z
+ K (r2 q)1 (dq) K (r2 q)2 (dq)
Z
c
Z
%(r1 q, r2 q)1 (dq) + cK
cK, a%(r1 , r2 ) + f (
1 ,
2)
67
(1 (dq) 2 (dq))
where
= (+ , ) and + , is the Hahn-Jordan decomposition of and we
have used the assumption that K is bounded with bounded derivatives.
4.2
New Results
The next result is the first step in showing conservation of vorticity for solutions
of (4.9). It shows that the empirical process associated with (4.10) is a solution
of (4.9). To show conservation of vorticity, we provide a similar argument as for
extension by continuity. We show that the vorticity is conserved if the initial signed
measure is discrete. After establishing this result, the general case will follow by an
approximation result.
Theorem 4.4
Proof:
i. Note that the one particle diffusion matrix associated with (4.10) is given by the
following. For i, j {1, . . . , N } and k, l = 1, 2
i , ri , , t)kl = 2
D(r , , t)kl := D(r
i
2 Z
X
q=1
68
kq, (ri , p)
lq, (ri , p)dp.
(4.13)
D(r , , t)kk = 2
(4.14)
XN+ (t) =
XN (t) =
ai ri (t)
ai >0
ai ri (t) .
ai <0
69
(4.15)
Proof: In Theorem 4.3, it was shown that the coefficients of (4.10) satisfy the Lipschitz
and boundedness properties. By Theorem 4.4, it follows that for discrete initial signed
measures, (4.9) has a weak solution. Consequently, by Theorem 3.9, we have a weak
solution of (4.9) for arbitrary initial signed measures.
To show conservation of vorticity for arbitrary initial signed measures, we employ
the flow representation of the solution, X (), from Theorem 3.9.
Z
X (t) := X (t, X , X0 ) =
(4.16)
Further,
X (t) := X (t, X , X0 ) =
(4.17)
sup
0tT
2
f,s
(X (t, XM,0 ), X (t, XN,0 ))
cT,F,J , E
f2 (XN,0 , XM,0 ).
(4.18)
for the discrete case by Theorem 4.4, and we have conservation of vorticity, (4.15).
Thus, we have established that the vorticity is a conserved quantity for the smooth
stochastic Navier-Stokes equations. The arguments generated by showing this claim
also have important applications to understanding the Hahn-Jordan decompositions
for fluids. Such a conclusion represents an important advancement to fluid dynamics.
Furthermore, the results established in Theorem 4.5 also establish the existence and
uniqueness of solutions to the smoothed stochastic Navier-Stokes.
71
Conclusion
The new research presented in this thesis provide significant advancements to the
fields of fluid dynamics and stochastic differential equations. The advancements, while
in diverse areas, followed a specific trend of theoretical development and application.
We briefly recall the main results.
In chapter two, the analysis focuses on the incompleteness of certain metrics on
the signed measures. The first important result involves identifying a relationship between the Kantorovich-Rubinstein metric and quotient spaces. Recall
that one can identify the signed measures with a quotient space on the space of
product measures. With this identification, we define a quotient-type metric on
the signed measures. The main result involves completeness and this quotienttype metric. Although the metric does not satisfy completeness on the signed
measures, it satisfies a very useful partial-completeness result. One can use the
convergence properties of the metric to conclude a limit is in Hahn-Jordan form.
In chapter three, we introduce a general class of signed-measure valued SPDE
and their associated SODE system. The focus of this thesis was the role of the
Hahn-Jordan decomposition for solutions. The quotient-metric from the previous chapter yields a powerful new result for the Hahn-Jordan decomposition.
Applying the quotient-metric and the product metric in conjunction yields that
the Hahn-Jordan decomposition of the initial signed measure is preserved in solutions. This represents a significant advancement as previous results assumed
smoothness of the coefficients of the SODE. From a stochastic viewpoint, the
result is ideal as the conditions are the same as those required for existence and
uniqueness.
72
In chapter four, the general results from chapter three apply to the smoothed
SNSE. In the general literature, questions of conservation of certain quantities
related to fluid dynamics naturally arise. Of great importance is that dealing with the conservation of vorticity. With the analysis of the Hahn-Jordan
decomposition established in the previous chapter, we show that the vorticity
of a fluid is conserved. As this question is important from both the fluid dynamic and stochastic analytic approaches, this advancement in the fields is a
significant one.
Each of these results provide a significant advancement to the current research
in the fields of fluid dynamics and stochastic partial differential equations. However,
with these advancements, new questions arise that one would like to answer. Although
numerous paths follow, we discuss only the following, yet very important possible
future work.
In order to show that a solution of the SNSE exists, one must solve the singular
vortex SODE with unsmoothed kernel.
dr (t) =
N
X
Z
2
(5.1)
j=1
with ri (s) = rsi for i = 1, . . . , N . To show that solutions of (5.1) exist on [s, ),
one needs to show that the point vortices satisfying (5.1) do not coalesce. Several
approaches to (5.1) use a stochastic driver of independent Brownian motions instead
. The works of [Tak85], [Osa85], and [FM07] show
of the correlation functionals,
that point vortices satisfying (5.1) never coalesce under various conditions on the
vortex intensities, aj . Takanobu shows the result when all the intensities, aj , have
the same sign in [Tak85]. Osada shows the result for arbitrary intensities, aj in
73
[Osa85], but must employ abstract techniques from PDE theory that disguise the
physical interpretation. Yet, [FM07] employs a technique to preserve the physical
interpretation called path-clustering. To use this technique, [FM07] must have vortex
X
intensities that satisfy the following. For all I {1, . . . , N },
ai 6= 0.
iI
The results of [FM07] show potential for extension to the case where correlation
functionals drive (5.1). To generalize the arguments of [FM07], one must address two
issues.
The stochastic flow is the mapping that sends an initial condition, x R2N ,
to the solution of (5.1) with this initial condition. In the case of independent
Brownian motions, the stochastic flow is differentiable, and the Jacobian has
a determinant that is identically 1. In the case of correlation functionals, the
determinant is no longer identically 1, and must be controlled to use the results
of [FM07].
To generate solutions of (5.1), one uses solutions of (4.10) with smoothing
parameter and lets 0 in a certain sense. In the case of independent
Brownian motions as drivers, the diffusion term is independent of . Yet, in the
case of the correlation functionals, the diffusion term depends on . This fact
presents a difficulty in adapting the arguments from [FM07].
Suppose that one can address these issues and prove that a solution of (5.1) exists
on [s, ). Itos Formula can be applied to yield solutions to the stochastic NavierStokes equations for discrete initial signed measures. However, as in chapter three and
four, it is the continuum-limit that poses the most difficulty. A priori estimates must
be derived for the singular Biot-Savart kernel which provides a significant barrier to
the general solution of the stochastic Navier-Stokes equations. Consequently, if one
74
can establish these estimates, then one can solve the true stochastic Navier-Stokes
equations. This will be a significant advancement for the fields of fluid dynamics and
stochastic analysis.
75
Appendix
In this appendix, we collect many of the standard results and notations for the
development of the new results. We separate the statements into two categories:
Fluid Dynamics and Stochastics.
6.1
Fluid Dynamics
Lemma 6.1 Let r(t, r0 ) be a path of a fluid particle in R2 under the Euler equations,
,
X (r, t) = curl U (r, t) =
r1
r2
(6.1)
U 0,
with position r0 at t = 0. Here U (r, t) = (U1 (r, t), U2 (r, t))T is the velocity field,
X (r, t) is the vorticity, is the gradient and is the inner product on R2 . Then, for
all t 0, it follows that
(6.2)
=
X
+ (X ) U
t
r(t,r0 )
X
+ ( U )X + (X ) U
t
76
r(t,r0 )
=
X
+ (U X )
t
r(t,r0 )
=0
Thus, (6.2) follows immediately.
The following result establishes exactly how the Biot-Savart kernel enters into the
analysis of the Navier-Stokes equation.
Lemma 6.2 Assume that U (r, t) is the velocity field and X (r, t) is the vorticity associated with an incompressible fluid in R2 satisfying:
,
r1
r2
(6.3)
U 0.
Then, the velocity can be written explicitly in terms of the vorticity as follows:
Z
U (r, t) =
(6.4)
where K() is the Biot-Savart kernel which for r = (r1 , r2 ) R2 is given by:
K(r) =
and = (
1
1
ln(|r|) =
(r2 , r1 )
2
2|r|2
(6.5)
,
) and |r|2 = r12 + r22 .
r2 r1
Proof: Note that from the incompressibility condition, it follows that U (r, t) is a
divergence free vector field in two dimensions. Consequently, from [Bar11], there
77
,
r2 r1
.
X (r, t) =
U (r, t) =
6.2
(6.6)
Stochastics
Let (S, ) be a separable metric space with metric bounded by 1. Recall the
definition of the Wasserstein distance is the following.
W1 (, ) :=
Z
inf
%(r, q)Q(dr, dq)
(6.7)
QC(,)
where C(, ) is the set of joint distributions of the probability measures , P1 (S)
(the set of probability measures on S).
We begin with the following important theorem and its proof.
78
sup
kf kL, 1
Z
f (r)( )(dr)
(6.8)
then, f (, ) = W1 (, ).
Proof: The following arguments are those of [Dud02] and [Pan08]. The result is first
shown in the case where the supremum in the definition of f is taken over kf kL 1.
Use to denote (6.8) in this case. This is the original statement of the KantorovichRubinstein Theorem. The result is proved by a sequence of intermediate lemmas.
Note that the sets in the definitions of and f are invariant under the multiplication
of 1. Consequently, one can remove the absolute value in their definitions.
Lemma 6.4 Define the following term for , P1 (S):
Z
m (, ) := sup
Z
f (x)d(x) +
g(y)d(y) : f, g C(S, R), f (x) + g(y) < (x, y)
(6.9)
Then, m (, ) = (, )
Proof: Let f CL (S, R) satisfy kf kL 1, and pick > 0. Define g(y) = f (y) .
f (x) + g(y) = f (x) f (y) (x, y) < (x, y)
Thus, f and g satisfy the conditions for m (, ), and have the following relation.
Z
Z
f (x)(dx) +
Z
f (x)(dx)
g(y)(dy) =
79
f (y)(dy)
Z
sup
Z
f (x)(dx) +
g(y)(dy) : f (x) + g(y) < (x, y) +
and, so
(, ) m (, ).
Now, for f, g C(S, R) such that f (x) + g(y) < (x, y). Define
This implies
f (x) e(x) g(x)
and
Z
Z
f (x)(dx) +
Z
g(y)(dy)
Z
e(x)(dx)
e(y)(dy).
80
Note that U is convex and open because S S is compact. Define the subspace E of
V as
E = { V : (x, y) = f (x) + g(y) where f, g C(S; R)}.
This implies that
U E = {f (x) + g(y) < (x, y)}.
One can define a linear functional on E, r, by
Z
r() =
Z
f (x)(dx) +
So, by the Hahn-Banach Theorem, one can extend r to r : V R such that r|E = r
and
sup r() = sup r() = m (, )
U
U E
Note that if a(x, y) 0 and c 0, then (x, y) ca(x, y) < (x, y) U . Thus,
for arbitrary c 0:
r( ca ) = r() c
r(a) () sup r <
U
Yet, the above is true only if r(a) 0. So, is a positive linear functional on the
compact space S S. By the Riesz Representation Theorem, there exists a Borel
81
Z
(f (x) + g(y))(dx, dy) =
Z
f (x)(dx) +
g(y)(dy).
sup
Z
f (x, y)(dx, dy) =
f (x,y)<(x,y)
Z
f (x)(dx) +
Z
g(y)(dy) =
Z
(f (x) + g(y))(dx, dy)
82
as we can remove empty sets. Let xn,k Sn,k , and define the following.
fn,k (x) =
xn,j
xn,1
if
jk
otherwise
XZ
j1
Sn,j
1X
(Sn,j ) +
n jk
Z
(x, xn,1 )(dx)
S\(S1 Sk )
2
.
n
Define n P1 (S) as the measure concentrated on the sets {Sn,1 , . . . , Sn,k }, and let
n,k be the push-forward measure of under the map x (fn,k (x), x) so that
n,k C(n , ):
Z
W1 (n , )
Z
(x, y)n,k (dx, dy) =
2
.
n
83
= W1 (, n ) + (n , n ) + W1 (n , )
W1 (, n ) + (n , ) + W1 (n , ) + (, n ) + (, ).
Letting n shows that W1 (, ) (, ). The other inequality was established
previously, so we have equality in the case of .
To show the claim with f , we must show that
Z
f (, ) =
f (q)((dq) (dq))
sup
kf kL, 1
and
Z
(, ) = sup
f (q)((dq) (dq))
kf kL 1
(f (q))((dq) (dq)) c((S) (S))
kf kL 1
Z
= sup
f (q)((dq) (dq)).
kf kL 1
As both , are probability measures. Consequently, one can choose without loss of
generality that f (x0 ) = 0 in the definition of (, ) where x0 is a fixed origin in S.
84
Yet, a function with Lipschitz constant bounded by 1 and 0 at the origin, x0 , satisfies:
:=
N
X
)
ai ti , N N, ti T and ai [0, ) Q i = 1, . . . , N
i=1
(6.10)
is dense in (Mf (S), f ). Similarly, the set of discrete, finite, Borel, signed measures:
(
Mf,s,d :=
:=
N
X
)
ai ti , N N, ti T and ai Q i = 1, . . . , N
(6.11)
i=1
85
i ) R, we have that
under f . Consequently, by writing ai := (B
() = (f 1 ()) =
ai xi ().
i=1
i=1
C( , ). Consequently, by Fubinia
a
a a
Tonelli,
aW1 ( , ) a
a a
Z
%(x, y)(dx, dy) =
X
i=1
Z
ai
i
B
Since was arbitrary, it follows that one can approximate by infinite sums of point
masses with arbitrary weights.
Note that for > 0, one can pick a discrete measure such that f (, ) < .
M
X
X
Choose M N so that
ai < . Define Mf by =
ai x i
i=1
i=M +1
f (, ) =
sup
kgkL, 1
sup
Z
g(r)((dr) (dr))
kgkL, 1 i=M +1
ai g(xi )
ai < .
i=M +1
Finally, note that can be approximated elements in Mf,d . For > 0, pick i
M
X
i
i xi , one has the
Q [0, ) such that |ai i | < 2 . Then, by defining :=
i=1
86
following.
f (, ) =
sup
kgkL, 1
sup
Z
g(r)((dr) (dr))
M
X
(ai i )f (xi )
kgkL, 1 i=1
M
X
|ai i |
i=1
<
Proposition 6.8 Let , Mf (S) with total masses m and n respectively. The
following metric is equivalent to the Kantorovich-Rubinstein distance, f ,:
f (, ) := (n m)W1 ( , ) + |m n|
n m
(6.12)
f (, ) = mf (
, ) = mW1 ( , )
m m
m m
which is (6.12).
87
f (q)((dq) (dq))
sup
kf kL, 1
Z
=
sup
f (q)
kf kL, 1
sup
f (q)
kf kL, 1
n
= f ( , ) +
m
= nf (
n
m
n
m
(dq) (dq) +
(dq) (dq) +
Z
1(q)
mn
f (q)
(dq)
m
Z
sup
f (q)
kf kL, 1
mn
(dq)
m
mn
(dq)
m
, )+mn
m n
= nW1 (
, )+mn
m n
where the last equality follows from the previous case. Consequently, we have shown
that f (, ) f (, ). We need to establish the other inequality to show that these
metrics are equivalent.
Now, assume that
n
n
6= , so that c := f ( , ) > 0. Choose fi with kfi kL,
m
m
1 such that
Z
c = lim
n
fi (q)( (dq) (dq)) = lim
i
m
Z
n
fi (q) (dq)
m
fi (q)(dq) .
mmn
mmn
m
mn
(dq) = i
= ( 1) > 0.
n
n m
n m
n
n
f ( , ) = c = lim
i
m
Z
lim
fi (q)(
n
(dq) (dq))
m
n
fi (q)( (dq) (dq)) + lim
i
m
Z
fi (q)
mn
(dq)
m
Z
= lim
fi (q)((dq) (dq)) f (, ).
Now, assume that c. Since c > 0, we may assume that i < 0 for all i.
n
f ( , ) = c = lim
i
m
n
=
lim
m i
n
lim
=
m i
lim
m i
fi (q)(
n
(dq) (dq))
m
fi (q)((dq)
m
(dq))
n
mn
fi (q)((dq) (dq)) +
lim
n i
fi (q)((dq) (dq))
n
f (, )
m
89
Z
fi (q)(dq)
(n m)f (
n
, ) = f ( , ).
m m
m
The left side is symmetric with respect to m and n. Consequently, in both cases we
get that
(n m)f (
, ) f (, ).
m m
1((dq) (dq)) f (, ).
f (, ) (n m)f (
, ) + |m n| 2f (, )
m n
Z
kf kp :=
p1
|f | ()(d)
<
p
Theorem 6.9 H
olders Inequality
p
F,
), g Lq (,
F,
), then f g
Let p (1, ) and q :=
and f Lp (,
p1
90
F,
) with
L1 (,
kf gk1 kf kp kgkq
Proof For a proof of the inequality, we refer to [Fol99].
Of particular importance is the Cauchy-Schwarz Inequality which can be obtained
1
by picking p = q = in Holders Inequality. Typically, one uses Cauchy-Schwarz in
2
tandem with Doobs Martingale Inequality to obtain global estimates on behavior.
Before we state Doobs inequality, we provide the definition for a martingale in the
Hilbert space setting. Let (H, kkH ) be a separable, real Hilbert space with scalar
product < , >H and norm kkH . An H-valued martingale is a stochastic process (i.e.
a collection of H-valued random variables) {Xt }t0 such that the following hold:
{Xt } is adapted to the filtration {F}t0 . That is, Xt is Ft measurable for all
t0
Xt is integrable for all t 0 in the sense that E kXt kH < .
E(Xt |Fs ) = Xs a.s. for t s 0.
By [MP80] we note that an H-valued martingale always has a modification that is
a.s. and X()
has sample
cadlag (i.e. an H-valued martingale such that X() = X()
paths that are continuous from the right and limits from the left exist).
Theorem 6.10 Doobs Martingale Inequality
For any stopping time and H-valued martingale Xt
E sup kXt k2H 4E kX k2H
0t
Theorem 6.11 Gronwalls Lemma Suppose u, v, w are R-valued piecewise continuous functions on a t b. Suppose u(t) is nonnegative on [a, b] and the following
holds for a t b:
Z
v(t) w(t) +
u(s)v(s)ds
a
then,
Z
v(t) w(t) +
Z
u(s)w(s) exp
a
u(x)dx ds
(6.13)
92
where denotes minimum. We call << X >> (t)t0 a quadratic variation for the
process {X(t)}t0
Proposition 6.14 Suppose X() is an H-valued square-integrable martingale. Then,
there is a unique quadratic variation << X >> () such that (6.13) holds. If X() is
continuous, then so is its quadratic variation << X >> (t).
Proof: See [MP80], [Kot08].
Now, for two H-valued processes with finite quadratic variation, X1 () and X2 (),
we define the mutual quadratic variation of X1 () and X2 () as follows:
<< X1 , X2 >> () :=
1
(<< X1 + X2 >> () << X1 X2 >> ())
4
(6.14)
Note that this expression is well defined as the existence of the quadratic variation
processes, << X1 >> and << X2 >>, imply the existence of both << X1 + X2 >>
and << X1 X2 >>. Furthermore, we note that the mutual quadratic variation
defines a bilinear form by [MP80].
We recall the definition for the Ito integral driven by a continuous square integrable
martingale, m() = (m1 (), . . . , md ())T , where d N and T denotes transpose. Define
Mdd to be the set of d d matrices over R and define L2,F ,loc ([0, ) : Mdd )
to be the set of (, ) that are Mdd valued, Ft -adapted, jointly measurable in (t, )
with respect to dt P (where dt is the Lebesgue measure on [0, )). Also, such
must satisfy the following.
d
X
k,l=1
Z
E
93
n ) Ftn -adapted Mdd Given a sequence of {0 = tn0 < . . . < tnk < . . .}, and (t
k
k
valued random variables for i = 0, 1, . . ., define
n) +
n (t, ) = (t
0
k=1
Such processes are called simple. We define the stochastic integral by:
t
n (s)m(ds) := 0 +
k=1
By [IW89], the simple processes are dense in L2,F ,loc ([0, ) : Mdd ) in the
L2 metric. Consequently, the definition of the stochastic integral can be extended to
arbitrary . It can be shown that if L2,F ,loc ([0, ) : Mdd ), and n is a
sequence of simple functions converging to , then the stochastic integrals converge
in probability by [IW89]:
Z
(s)m(ds) = lim
Z
We note that
n (s)m(ds).
0
(s)m(ds) >>=
0
Z
d
X
i,j,k=1
(6.15)
where << , >> denotes the mutual quadratic variation of the components.
Generalization of Brownian Motion differentials
Typically, one of the main choices is m(t) = (t) where (t) is d-dimensional Brownian
motion. However, such a choice will not be sufficient for our analysis. Denote R+ :=
94
[0, ).
Definition 6.15 Following [Wal84] and [Kot08], a process exists called Standard
Gaussian Space-Time White Noise, w(dq, dt, ). It is a finitely additive random
signed measure on the Borel sets of Rd R+ of finite Lebesgue measure |A| such that
the following holds:
Z
Z
1A (q, t)w(dq, dt, ) is a normally distributed random variable with mean
independent
Let wi (dq, dt, ) be independent standard Gaussian White-Noise for i = 1, . . . d.
We define d-dimensional standard Gaussian White-Noise as
.
Finally, we provide a fundamental theorem from stochastic integration theory,
Itos Formula. We state the theorem as in [Kot08] and [IW89]
Theorem 6.16 It
os Formula Let (r, t) : Rd+1 R be twice continuously differentiable in space and once continuously differentiable in t such that all partial
derivatives are bounded. Further, we let m() be a continuous square-integrable Rd valued martingale and b() a continuous process of bounded variation. Set
95
Z
+
0
(a(s), s)ds
s
d Z
1X t 2
()(a(s), s) (b(ds) + m(ds)) +
( )(a(s), s) << mi , mj >> (ds)
2 i,j=1 0 i,j
(6.16)
where << mi , mj >> () are the mutual quadratic variations of the one-dimensional
components of m().
96
References
[Ami00]
A. Amirdjanova.
[Arn92]
[AX06]
A. Amirdjanova and J. Xiong. Large deviation principle for a stochastic navier-stokes equation in its vorticity form for a two-dimensional incompressible flow. Discrete and Continuous Dynamical Systems series B,
6(4):651, 2006.
[Bar11]
[Cho73]
[CM93]
[Dud02]
R.M. Dudley. Real Analysis and Probability. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 2002.
[EK05]
[Eva10]
L.C. Evans. Partial Differential Equations. Graduate Studies in Mathematics. American Mathematical Society, 2010.
[FM07]
Joaquin Fontbona and Miguel Martinez. Paths clustering and an existence result for stochastic vortex systems. Journal of Statistical Physics,
128(3):699719, 2007.
[Fol99]
[GG99]
[Kat67]
[KK10]
P.M. Kotelenez and T.G. Kurtz. Macroscopic limits for stochastic partial differential equations of mckeanvlasov type. Probability Theory and
Related Fields, 146(1):189222, 2010.
98
[Kot95]
[Kot08]
P. Kotelenez.
Equations: Transition from Microscopic to Macroscopic Equations. Applications of Mathematics. Springer Science+Business Media, 2008.
[Kot10]
[KS12a]
Peter M. Kotelenez and Bradley T. Seadler. Conservation of total vorticity for a 2d stochastic navier stokes equation. Advances in Mathematical
Physics, 2012, 2012.
[KS12b]
Peter M. Kotelenez and Bradley T. Seadler. On the hahn-jordan decomposition for signed measure valued stochastic partial differential equations.
Stochastics and Dynamics, 0(2), 2012.
[KX99]
[Lad69]
[Lon88]
Ding-Gwo Long. Convergence of the random vortex method in two dimensions. Journal of the American Mathematical Society, 1(4):pp. 779804,
1988.
99
[MP80]
[MP82]
1982. 10.1007/BF01209630.
[Mun00] J.R. Munkres. Topology. Prentice Hall, 2000.
[Osa85]
H. Osada. A stochastic differential equation arising from the vortex problem. Proceedings of the Japan Academy, Series A, Mathematical Sciences,
61(10):333336, 1985.
[Pan08]
[Pro04]
P.E. Protter. Stochastic Integration and Differential Equations. Applications of mathematics. Springer, 2004.
[Ros31]
[Tak85]
S. Takanobu. On the existence and uniqueness of sde describing an nparticle system interacting via a singular potential. Proceedings of the
Japan Academy, Series A, Mathematical Sciences, 61(9):287290, 1985.
[Wal84]
J.B. Walsh. Introduction to Stochastic Partial Differential Equations. University of British Columbia, 1984.
100