Probabilistic Methods in Fluids Proceedings of The Swansea 2002 Workshop, Wale, UK, 14-19 April 2002 by I. M. Davies, PROBABILISTIC METHODS in FLUIDS WORKSHOP, N. Jacob, A. Truman, O. Hassan, K. Morga
Probabilistic Methods in Fluids Proceedings of The Swansea 2002 Workshop, Wale, UK, 14-19 April 2002 by I. M. Davies, PROBABILISTIC METHODS in FLUIDS WORKSHOP, N. Jacob, A. Truman, O. Hassan, K. Morga
Probabilistic Methods in Fluids Proceedings of The Swansea 2002 Workshop, Wale, UK, 14-19 April 2002 by I. M. Davies, PROBABILISTIC METHODS in FLUIDS WORKSHOP, N. Jacob, A. Truman, O. Hassan, K. Morga
Methods in
Fluids
This page intentionally left blank
editors:
I M Davies
N Jacob
A Truman
Department of Mathematics
University of Wales Swansea
UK
0 Hassan
K Morgan
N P Weatherill
School of Engineering
University of Wales Swansea
Probabilistic
Methods in
Fluids
Wales, UK 14 - 19 April 2002
ye L
World Scientific
NewJersey London Singapore Hong Kong
Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA ofice: Suite 202, 1060 Main Street, River Edge, NJ 07661
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN 981-238-226-7
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
IRIMA . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Participants . . . . . . . . . . . . . . . . . . . . . . . . . xi
Sergio Albeverio and Yana Belopolskaya . . . . . . . . . . . . . 1
Probabilistic Approach to Hydrodynamic Equations
Hakima Bessaih and Franco Flandoli . . . . . . . . . . . . . . 22
A Mean Field Result for 3D Vortex Filaments
Bjorn Bottcher and Niels Jacob . . . . . . . . . . . . . . . . . 35
Remarks on Meixner-type Processes
Zdzistaw Brzeinaak . . . . . . . . . . . . . . . . . . . . . . 48
Some Remarks on It6 and Stratonovich Integration in 2-smooth
Banach Spaces
Tomas Caraballo . . . . . . . . . . . . . . . . . . . . . . . 70
The Long-time Behaviour of Stochastic 2D-Navier-Stokes Equations
Pao-Liu Chow . . . . . . . . . . . . . . . . . . . . . . . . 84
Semilinear Stochastic Wave Equations
Nigel J. Cutland . . . . . . . . . . . . . . . . . . . . . . . 97
Stochastic Navier-Stokes Equations: Loeb Space
Techniques & Attractors
Arnaud Debussche . . . . . . . . . . . . . . . . . . . . . 115
The 2D-Navier-Stokes Equations Perturbed by a Delta
Correlated Noise
Sergio Albeverio and Benedetta Ferrario . . . . . . . . . . . . 130
Invariant Measures of Lkvy-Khinchine Type for 2D Fluids
Franco Flandoli . . . . . . . . . . . . . . . . . . . . . . . 144
Some Remarks on a Statistical Theory of Turbulent Flows
Christophe Giraud . . . . . . . . . . . . . . . . . . . . . 161
Some Properties of Burgers Turbulence with White Noise
Initial Conditions
V
vi
The aim of the meeting, the first IRIMA workshop, was to bring together
internationally reknowned researchers from the areas of Pure Mathemat-
ics, Applied Mathematics and Engineering t o participate in a workshop,
on probabilistic methods for fluids, and through collaboration further the
mathematical understanding of the fundamental problems in this field.
This international workshop successfully allowed leading researchers to
present, reflect upon and discuss their recent work in the probabilistic
modelling of fluids. This field stretches across Pure Mathematics, Applied
Mathematics and Engineering and consequently is ideally placed to benefit
from regularly arranged workshops for collaborative purposes. The Work-
shop mainly concentrated on the understanding of turbulence in stochastic
fluid dynamics, a problem which has numerous applications in science and
engineering and has defied many attempts to success full^ model it. The
workshop bridged a gap between the recent year of activity at the Univer-
sity of Warwick and the year of emphasis at Princeton, which started in
Autumn 2002. As such the workshop ensured that the research momentum
in Britain, in this subject, was maintained.
In this volume probabilistic approaches to hydrodynamic equations are re-
viewed and deterministic viscous hydrodynamics is discussed in terms of
stochastic processes on groups of diffeomorphisms. At the Workshop sig-
nificant progress was made in understanding the intermittence of stochastic
turbulence for Burgers equation and the application of L6vy processes t o
the Mathematics of Finance, both of which are represented in the pro-
ceedings. Other noteworthy developments concerned the Strong Law of
Large Numbers and ergodicity of the Gaussian invariant measures for 2-
dimensional Navier-Stokes equations with space-time white noise and pe-
riodic boundary conditions and mean field results for 3-dimensional vortex
filaments. Also, new results are presented on Loeb space techniques and
attractors for stochastic Navier-Stokes equations. The long time behaviour
of stochastic 2-dimensional Navier-Stokes equations is investigated as are
vii
...
Vlll
I M Davies 0 Hassan
N Jacob K Morgan
A Truman N P Weatherill
University of Wales Swansea, December 2002
International Research Institute in Mathematics and its
Applications
ix
X
xi
xii
S. ALBEVERIO
Institut fur Angewandte Mathematik, Universitat Bonn,
Wegelerstr. 6, D-53115 Bonn,
Germany
SFB 256,
Bonn, BiBoS, Bielefeld - Bonn, CERFIM, Locarno and USI (Switzerland)
YA. BELOPOLSKAYA
St. Petersburg State University for Architecture and Civil Engineering,
Russia, 198005, St. Petersburg, 2-ja Krasnoarmejskaja 4
1
2
Notice that among recent works the papers by Busnello and Flandoli,
Busnello, are mostly close to our approach. In these papers the authors
deal with the equation for the circulation of the velocity field and the Biot-
Savart law (instead of the original Navier-Stokes system) and study it by
probabilistic methods.
In the present paper we construct diffusion processes that can be used
for the probabilistic representation of the velocity field and the pressure
itself.
To this end we consider the Cauchy problem for the Navier-Stokes sys-
tem
divu =0 (2)
where u(t,z) E R3,z E R3,t E [0, co),P is a positive constant and p ( t , x) E
R1 and change ( 2 ) for the Poisson equation
- A p ( t , x) = y ( t ,x), y = Tr[VuI2 (3)
connecting the velocity field and the pressure. The topic of main interest
for us in the present paper is the construction of diffusion processes in R3
associated with (1),(3). Let (0,F,P ) be a probability space and w ( t ) E
R3,B ( t ) E R3 be a couple of independent Wiener processes defined on it.
Denote by E the expectation with respect to P and by EE the conditional
expectation with respect to a stochastic process [ ( t ) .Consider the Cauchy
problem for the stochastic differential equation
d<(T) = -U(t - 7 ,< ( T ) ) d 7+ fJdw('T), <(o) = 5 (4)
and assume the unknown drift coefficient u to be determined by
and
m l
V d t ,). = - -SE [ y ( t , + B(s))B(s)lds (10)
belongs to Cb(R3)and
IPk(t, .)I100 5 Ilr(t,.)l142,4. (13)
If in addition q 2 > $, then Fk(t,x)= 2&p(t,z) .
Lemma 1.2. Let y ( t ) E Lq2 nL4(R3)for some 1 < q 2 < 2 < 3 < q , t E
[0,TI. Then for each r > 2 the functions x H Fk(t,x) belong to L' and
IIFk(t)llr I Kllr(t)1142 , 4 ( 1 + Ilr(t)1142 A). (14)
Let us construct the solution for (4),(5), (7) by a successive approxi-
mation method.
To this end we consider
uO(t,x) = uo(x), p ( t ) = x, (15)
w
po(t,x) = ETr [Vu0l2(t,
x B(s))ds, +
uk+l(t,x) = E[UO(<"(t))- / 0
t
v p y t - 7,<'"(7))d7]. (18)
To prove the convergence of the successive approximations we have to add
-&-
to (15)-(18) the successive approximations for the derivatives q(t) = at (t)
and put
where
l l ~ l I ~ , r= [IP(t)llo + [V4t)lal.
x S~PtG[O,T]
with the smooth drift coefficient v ( t , x ) E R3,x E R 3 , t E [0, co) and its
derivatives with respect t o the initial data V"(t)given by
V v ( t - 7, &(.)) 0 VZ(T)dT,
I K,(t),
s~P"Ilv(t,z)1I2 5 K,(t),
s~PZIIVv(t,s)1I2
and
2
IIW1x) - Vv(t,Y)1l2I L:(t)Ilz - YII
where K, ( t ) ,K t ( t ) LA
, ( t ) are positive time dependent scalar functions
bounded on bounded interval [0,TI. Then
7
and
21 21 S O t [ K t ( t - - 7 ) + L t ( t - r ) ] d ~
I
EIlr12(t)- rlY(t)112112 - YII e
1 (31)
hold for t E [O, TI.
Proof. I t results from Ito's formula and Gronwall's lemma that
Elltdt) - t Y ( t ) l 1 2 1
t
5 1l2-YIl2'+211 El(4-7&(7)) -"(t-7ltY(7)),&(4 -Jy(7))1
Ilrz(.) - EY(7))/12(1-1)d7i 1
15 - YII
21 21
e
lot
K,'(t-r)dr
Ellrl(t)l12
< 1-+ 2
- it E(Vv(t - 7 ,t(7)>17(.>1rl(T))d7
I 1t 2 1t
K t ( t - 7)E11rl(7)/)2d7
The solution &(t)of (27) gives rise to the stochastic flow x H & ( t ) . In
addition
dt J ( t , x) = J ( t , x)div v(& ( t ) ) .
JVlv' + J V 2 v 2 + J V 3 v 3 = (div U ) J. 0
Remark If the drift vector field v(t, x) in (27) possesses the property
d i v v = 0 , we deduce d t J = 0 and J ( t ) = J ( 0 ) = I . 0
Assume that v is a smooth divergence free vector field and consider
functions p , and u given by
II~oII, < tor, I I v w I I I ~ < c&, SUP, JJv~o(z)IJ I K,, SUP, Il~o(x)IlI KO,
9
llvu(t)llT< y ( t ) (35)
hold f o r the function u ( t , x )given by (32), (33), 0 I t < T I and < r < 2 9
or r > 3.
Proof. From the heuristic expressions (23)-(26) rewritten for the func-
tions determined by (32), (33) by Jensen and Holder inequalities we deduce
that
( ( V u ( t - s , x ) ( Il m 1 ( t - s , x ) + m z ( t - s 1 s ) (36)
where
To derive the estimate for mz we apply the Holder inequality to the right
hand side of (37). Taking into account (30) this yields
- ~ , < z ( e ) ) I I z ~ l I r 1 z ( ~ ) 1 1 2 )5~ d e
10
for 1
43
I+
44
= 1. Choose 43 < 2 and q 4 > 2 to obtain
t
where c6 = do]& depends on t , s. If r > 4 and q4 are chosen so
that & 2 1 then we can apply Jensen’s inequality to get
Finally, we derive the required estimate for v in terms of the Lr- norm of
V p , using the properties of the Jacobian J proved in Lemma 2.2. In this
way we obtain
I c7 L k3
t
IIVP(t - 8, z)IITdzdQ.
Using the notation Cur@)= IIVu(t)ll. we deduce from the above estimates
that
d y = Py
-
ds
+ KY2[1+ y2], y(0) = c;,, (44)
dP
-
ds
= p2 + Ky2, P(0) = KO. (45)
By the general theory of ODE systems we know that there exists a unique
bounded solution to this system over an interval [O,Tl] depending on
KJ,CJr.Finally we notice that if K,(t) 5 p(t) and C&(t) 5 y(t) then
IIVull, 5 y ( t ) and sup,IIVu(t,x)II5 p(t) on the interval [O,Tl]as well. 0
Lemma 2.4 Under the conditions of Lemmas 2.1 - 2.3 there exist func-
tions M i @ ) ,M,(t), K u ( t )and Z,(t) (bounded o n [O,Tl)for above 7'1 ) such
12
that the vector fields u given by (33) and V p for p given b y (32) obey the
estimates
ll~P(t)llql,T,q2 -5
where 1 < r < $ or r = 4 and < q 1 < 2 , q 2 > 3.
Proof. The proof of these estimates can be derived from (32), (33) using
the results of Lemma 1.2 and Lemmas 2.1 - 2.3. 0
Lemma 2.5.Assume that conditions of Lemma 2.1 - Lemma 2.3 hold.
Let in addition u0 E C 2 ( R 3 )and llu0llcz 5 Kz. Then there exists an
interval [0,T I ]with TI < T depending on KZ, KOand the function Lk(t - s )
bounded on this interval such that the estimate
I1Vu(t - s , x ) - V u ( t - s , y ) l l 2 < LL(t - s)11x - (46)
holds.
Proof. It is easy to check that
C l ( 4 Z,Y) ) ~~0(Jy(t))rly(t)1I2,
= E / l [ ~ ~ o ( € z ( t ) ) 7 7 z ( t-
To derive the estimate for (2(t) we notice first that y(t, x) = Tr[V2w(t,
x)]
satisfies the estimate
pUTTING
N2(t) 5 JT" +
:2K;(t)(EI(V.u(t,z B ( s ) )- Vu(t, Y + B(s))Ilm)&
for q' =5. Finally, choosing q < 2 and T < z q we prove that the integrals
in the latter expression converge that leads t o the estimate
N ( t ) L CIK,(t)L:(t)Jlz - Yll.
14
+ u - ~ I C0 ~ ~
$(t,5 , Y) I IIx - YII’[C~
st C4[KLv(t-T)+K]dT
From the above estimates we get that there exist absolute positive constants
C3, C4 such that
L i ( t )5 [c3+ u-1]c4eJot
C~[KL:(~-T)+KI~T
+
where A4 = C4[C3 a-1]. Choose It - sI i TI to ensure that Ki(i-1) <
K < 03. As a result we deduce that n(t - s) solves the Cauchy problem
dK(t - S )
ds
+ )M
= C ~ [ KK ] K , ~ ( 0=
where C = 1 + K
c41c3+c-11. Hence, if 0 < t < T3 where
T3 = min(Tl,T2) (53)
and
The assertion of this lemma can be deduced from the estimates derived
in section 1, Lemma 1.2 and Lemma 2.5. 0
Lemma 2.7. Under the conditions of Lemma 2.1 the estimates
E l l l 2 ( t ) - E,""t)ll2 (54)
(55)
and Gronwall's lemma yields (54). We prove the second estimate applying
Gronwall's lemma t o (28) that yields
q r l z v ' ( t ) - 77""2(t)1I2 I
where
vipyt, x) = .I m l
-E[y'"(t,x
S
+ B(S))Bi(S)]dS
3 auk auf
where y'"(t, x) = dF we deduce that V p k ( t ,x) is uniformly (in
k) bounded in L' norm as well. Moreover given rk(t)E Lq n C1icy(R3), Q E
(0,1),1 5 q 5 4we know by Schauder's theorem that IIV2p(t)Ilc; 5
KIIYII Lqncg.
Let u k , Vuk be successive approximations of tensor fields u,Vu defined
by (15)-(18) and (55)-(58). Now we can prove our main result stated in
Theorem 1.1.
Proof of Theorem 1.1 Let us prove that u k ( t ) ,Vuk(t) given by (15)
- (18), (55) - (58) converge in L' norm. Set
17
We start our considerations with general remark that to estimate all above
functions we apply the Fatou-Fubini theorem to change the order of inte-
gration in t and z variables as well as Holder's and Jensen's inequalities.
We denote further by K,, i = 1 ' 2 , . . . , constants which depend only on
t and r and assume that s is choosen so that f +
= 1. Many of our
computations use the Fubini theorem and the Holder inequality and the
properties of the Jacobian J that allows to change the variables under the
integral sign. Since we have used already this reasoning before in proving
previous Lemmas we do not give below the detailed description of them.
To estimate aL(t) we apply (34)' the Gronwall lemma and the Holder
inequality to deduce
I K l [ es," KZPT(t-T)dT + 1
t
,qt - T)dT] (60)
where
1
;' ( t )= El/VUo(E,k-l ( t ) I/)'11 p k( t )- T p - ( t )11 ' d z
L 3
and the last three terms are derived by Bismut-Elworthy's formula that
along with Fubini's theorem and the Holder inequality gives
In addition
Ak(t) 5
t
L K5[/ r;(t - T ) ( Q ~ ( T ) ) ~ ~ T ] .
0
and by estimates from Lemma 2.1 and the Holder inequality applied in the
0 variable we deduce for ml = r m < 2
tO DERIVE THE ESTIMATE FOR (T0 WE USE THE ESTIMATES FROM lEMMA
20
where
and -rl1
<I.
2
Let us combine the above estimates (58)- (70) t o derive the following
inequality
t
IIGk(t)lIT I M ( t - ~)IIGk-l(~)IITd~.
(71)
t
Ilu(t - T ) - w ( t - T)IITdT
1
+ c1 Ilu(t - ). - v ( t - T)II.dTde
where the positive constants C , C1 depend on the interval [0,T ) and es-
timates for functions L h ( t ) , Kk(t) derived in the above Lemmas. Finally
(71) yields that IJu(t)- v(t)lJ. = 0.
21
Acknowledgements
We are very grateful1 to Professor Aubrey Truman for the kind invitation
to an interesting and stimulating conference and to the University of Wales
for the hospitality. The financial support by DFG Project 436 RUS 113/593
and by Grant RFBR 02-01-00483 are gratefully acknowledged.
References
1. Nelson E. Les e‘coulements incompressibles d’energie finie. Colloques intern.
d u CNRS” 117,159, (1962).
2. Belopolskaya Ya., Dalecky Yu. Investigation of the Cauchy problem for quasi-
linear parabolic systems with the help of Markov random processes. Izu, VUZ.
Matematika, N 12, 5 (1978).
3. Belopolskaya Ya. I., Dalecky, Yu. L. Stochastic equations and differential ge-
ometry, Kluwer Acad. Publ., (1990).
4. Busnello B. A probabilistic approach t o the two-dimensional Nauier-Stokes
equations. The Annals of Prob. 27, N 4, 1750,(1999).
5. Busnello B., Flandoli F., Romito M. A probabilistic representation f o r the
vorticity of a 3D viscous fluid and f o r general systems of parabolic equataons.
Preprint (2002).
6. Belopolskaya Ya. Probabilistic representation of solutions to boundary-value
problems for hydrodynamic equations Zap. nauchn.sem. P O M I , 249, 77,
(1997).
7. Belopolskaya Ya. Burgers equation o n a Hilbert manifold and the motion of
incompressible fluid,Methods of Functional Analysis and Topology, 5 , N4, 15
(1999).
8. Elworthy K.D., X-M.Li. Formulae for the derivatives of heat semigroup. JFA
125,252, (1994).
A MEAN F I E L D R E S U L T F O R 3D VORTEX F I L A M E N T S
A mean field result is proved for a n abstract model, under a class of conditions on
t h e rescaling of the energy. Propagation of chaos, variational characterization of
t h e limit Gibbs density h and a n equation for h are proved. The general results are
applied t o a model of 3D vortex filaments described by stochastic processes, includ-
ing Brownian motion and Brownian Bridge, other semimartingales, and fractional
Brownian Motion.
1. I n t r o d u c t i o n
The importance of thin vortex structures in 3D turbulence has been dis-
cussed intensively in the last ten years, see 4 , Some mathematical models
of vortex filaments, based on stochastic processes, have been proposed by
Chorin 4, Gallavotti Lions-Majda 14, Flandoli 5, Flandoli-Gubinelli ‘
and Flandoli-Minelli The importance of these models for the statistics
of turbulence or for the understanding of 3D Euler equations is under in-
vestigation.
The limit properties (mean field) of a collection of many interacting vor-
tices has been investigated by P. L. Lions and A. Majda l 4 for a particular
model of “nearly parallel” vortices.
The aim of our work is to investigate a similar limit for the model intro-
duced in 5 , ‘. Here the expression for the kinetic energy is not approximated
and filaments may fold, so some features are more realistic. However, the
filament structures have a fractal cross section (as observed numerically) to
eliminate a divergence in the energy.
The structure of the paper is the following one. In the next section we
present the abstract frameworks and state a mean field result for them.
Section 3 is devoted to the proofs. Then in the final section we apply the
general result to some models of vortex filament.
22
23
i=l i#j
(here and below, the second summation is extended from 1 to N ) with the
understanding that for N = 1 it reduces to
H s ( w ) 2 0 po-a.s.
(5)
and that
2
HI(u,u’) is symmetric in w and w‘, po -a.s. (7)
Let H N : ON 4 R, for any positive integer N , be the random variable
defined as
The variable H N has the meaning of a rescaled energy, where only the
interaction energy is reduced as N grows. A physical motivation for this
rescaling has to be found in each particular case.
Denote the product measure of N copies of po by p t . Let
hN :RN + R
be the probability density defined as
hN = (N)-’ e - p H N , Z ( N )= lNePPHNdpf
24
dpN = hNdp,N.
dpNh = hN,k k
PO
hN’k(W1l...lwk)=
s,r hN(W1,...,WN)dpo(#k+1) ...dPo(Wiy)
Under the assumptions (2), (5), (3), (6) and (7), we have the following
result.
dp = hdpo
3. Proofs
We introduce new notations to shorten the formulas. We set
H ( q J l , . . . , U p J ) = HS(W,),
and
H("qU1,. . ., U N ) = HI(Wn,Um).
In the sequel, we simply write H(") and H("1") without their arguments.
Proof. The proof of this lemma will be done in three steps. Let us define
In particular Z ( N ,N , 1) = ZN.
Step 1 Given k 2 1, there exists a constant No(k) such that for all
N 2 No(k)
2k
hNYk 5 ( Z ~ J ) -Z' ( N ,N - k , 1 - -)
N
Step 2 For every p > 0 and for every N 2 k 2 1,
k
Z", N,P ) 2 (CZ(IL)YZ ( N ,N - k, P + ;v)
where
2k
hN'k 5 (ZN)-' Z ( N ,N - k , 1 - -)
N
c k
= (ZN)-' Z ( N , N - k , 1 - - $- -)
N N
5 ( Z N ) - ' ( C Z ) -Z~( N ,N , 1 - z)
c k
5 ( Z N ) - ' c 3 ( k ) ( C Z ) -Z~( N ,N , 1)
= C3(k)(Cz)-"
The proof is complete.
{
P= f 2o p t - as., fEL~(oN;~:),
Lemma 3.2. The density hN is the unique minimum of F N .
F(4 = s, F(f)dT(f),
where F is the functional given in theorem 2.
Let us prove the following basic fact:
?i = 6h (28)
and h is the unique minimum of F over P .
Due to (28), this implies that hm3k factorizes] i.e. the associated mea-
sure is a product measure. This proves the first claim of Theorem 2. More-
over, i t proves that the functional F has a unique minimum. So, let us
prove the lemma.
F(%)= s, F ( f ) % ( d f )2 s, r;’?i(df) = F
It follows that
of a velocity field u(x) can be written in terms of the vorticity field [(z) =
curlu(s) as
29
E(.) = r 1
1
b(x - y(t))?(t)dt,
where the parameter r is the circulation and the energy takes the form
For regular curves, as well as for many examples of curves given by paths of
stochastic processes, this expression (with suitable interpretation for pro-
cesses) is divergent. Physical vort,ex structures, although very thin, have
a cross section. Re-introducing the cross section increases the degrees of
freedom and makes the model less intrinsic, but helps t o eliminate the di-
vergence of the energy. To keep a closer relation with the vortex structures
observed in fluids, it is better to consider fractal cross sections instead of
simply a tubular mollification. In the previous papers ', vorticity fields
of such kind with finite energy have been constructed. They are formally
expressed as
represent the self energies of the single filaments] while the terms
(44)
The presentation until now have been rather informal] but we give below
rigorous definitions.
R = c x M-1,
where C = C ([0, TI;R3)and M-1 is the space of probability measures p in
R3 defined below. The interpretation is that the filament has a core and
a cross-section. The core is a 3-D curve, i.e. an element of C([O,TI;R3).
The cross-section is a probability measure p, the support of the measure
represents the geometric cross-section, while the measure weights the inten-
sities of the different lines of vorticity. Thus R is the space of configurations
of a fluid when the vorticity field is made of a single vortex filament with
cross-section.
The space of configurations of a collection of N vortex filaments is O N l
the product of N copies of R.
where b ( k ) = Jw3 ei"'"p(dz) and let us denote by M-1 the set of all p such
that 11 p llT1< 00. We recall from classical potential theory, see l 2 that a
probability measure p E M-1 is called a measure with finite energy. Given
a set A in R3, there exists a probability measure p supported by A with
finite energy if and only if the capacity of A is strictly positive. Finally, by
Theorem 3.13 of 12, every compact set with Hausdorff dimension d > 1 has
positive capacity. Therefore, it supports a probability measure p satisfying
(46).
On the space M of all probability measures p on (R3, 2 3 (R')) there is a
metric d such that the convergence with respect to d is the weak convergence
of probability measures:
for all bounded continuous functions f . We endow the subset M-1 with
the metric d . Notice that ( M , d ) is complete, while ( M - 1 , d ) is not, but
this fact has no importance in the sequel. Let us denote by BM the Borel
c-algebra of (M-1, d ) .
Let PM be a probability measure on ( M - 1 , B M ) . Our vortex structures
will have a cross-section measure p choosen a t random with probability law
PM. In the sequel, we denote by B the product a-algebra Bc @ BM on R
a=&@BM
where BC is the Borel a-algebra on C = C ( [ 0 , T ] ; R 3 )Moreover,
. if PC
denotes the Wiener measure on C, we set
PO = PC @ P M .
then, we have
where (Xt)tE[o,ll
is a Brownian semimartingale, i.e. a process of the form
X t = Wt + 1
t
b,ds
33
References
1. P. Blanchard., E. Briining. (1992). Variational Methods in Mathematical
Physics. A unified Approach. Springer-Verlag.
2. E. Caglioti., P. L. Lions., C. Marchioro., M. Pulvirenti. (1992). A Special
Class of Stationary Flows for Two-Dimensional Euler Equations: A Statistical
Mechanics Description. Comm. Math. Phys 143,no 3, 501-525.
3. E. Caglioti., P. L. Lions., C. Marchioro., M. Pulvirenti. (1995). A Special
Class of Stationary Flows for Two-Dimensional Euler Equations: A Statistical
Mechanics Description 11. Comm. Math. Phys 174,no 2, 229-260.
4. A. Chorin. (1994). Vorticity and Turbulence. Springer-Verlag, New York.
5. F. Flandoli. A probabilistic description of small scale structures in 3D fluids.
To appear on Annales Inst. Henri Poincark, Probab. & Stat.
6. F. Flandoli, M. Gubinelli. Gibbs ensembles of Vortex filaments. To appear on
Prob. Theory and Related Fields.
7. F. Flandoli, I. Minelli. Probabilistics models of vortex filaments. To appear on
Czechoslovak Mathematical Journal.
8. U . Frisch. (1998). Turbulence, Cambridge Univ. Press, Cambridge.
9. G. Gallavotti. (1996). Meccanica dei jluidi. Quaderni CNR- GNAFA n. 52,
Roma.
10. E. Hewitt, L. J. Savage. (1955). Symmetric measures on Cartesian products.
Trans. A m e r . Math. SOC40, pp470-501.
11. H. Kunita. (1984). Stochastic Differential Equations and Stochastic Flows
of Diffeomorphisms, Ecole d’6t6 de Saint-Flour XII, 1982, LNM 1097, P.L.
Hennequin Ed., Springer-Verlag, Berlin.
12. N. S. Landkof. (1972) Foundations of Modern Potential Theory, Springer-
Verlag, New York.
13. P.L. Lions. (1997). O n Euler Equations and Statistical Physics, Scuola Nor-
male Superiore.
14. P.L. Lions, A . Majda. (2000). Equilibrium Statistical Theory for Nearly Par-
allel Vortex Filaments. C. P. A. M, Vol. LIII, pp 0076-0142.
15. C. Marchioro, M. Pulvirenti. (1994) Mathematical Theory of Incompressible
Noviscous Fluids, Springer- Verlag, Berlin.
16. D. Nualart, C. Rovira, S. Tindel. Probabilstic models for vortex filaments
based on fractional Brownian motion. In preparation
17. D. Ruelle. (1969). Statistical mechanics: rigorous results, W. A. Benjamin,
New York- Amsterdam.
REMARKS ON MEIXNER-TYPE PROCESSES
1. Introduction
L6vy processes are becoming more and more important in modeling, com-
pare the collection of surveys recently edited by 0. Barndorff-Nielsen, Th.
Mikosch and S.I. Resnick '. Several applications are related to the mathe-
matics of finance, see in particular the contributions of 0. Barndorff-Nielsen
and N. Shephard 4 , and E. Eberlein 7 , in '.
By definition a L6vy process has stationary and independent increments.
This is reflected by the fact that its transition probabilities form a con-
volution semigroup and that the generator A of the corresponding Feller
semigroup is translation invariant. On smooth functions it is given by
35
36
such a situation will not any longer be a Levy process, but a Lkvy-type
process. To our knowledge it had been 0. Barndorff-Nielsen and S. Z. Lev-
endorski; who first handled such a situation in case of normal inverse
Gaussian (type) processes. Their work had been stimulated by problems of
modeling in finance. Their approach via pseudo-differential operators had
partly been influenced by the survey 1 3 .
In this paper we take up the ideas from of treating L&y(-type) pro-
cesses with state space dependent parameters by using the theory of pseudo-
differential operators. The family under consideration consists of Meixner-
type processes, i.e. we start with the characteristic exponent (1) of a
Meixner process and make the parameters state space dependent. The
choice of these processes is due to more recent work of W. Schoutens 21
and W. Schoutens and J. Teugels 22 on modelling in finance with Meixner
processes.
In a first section we recall basic facts of Meixner processes and in the fol-
lowing section we introduce symbols of Meixner-type, i.e. symbols q(z,<)
<
where for fixed 20 the function H q ( q , <) is a characteristic exponent of
a Meixner process. It is proved that under reasonable restrictions on the
2-dependence of the parameter functions these symbols are elliptic sym-
bols in the classical class S1(R).Finally in Section 3 we prove that to every
Meixner-type process corresponds a unique Feller process, and some short
time asymptotics of the corresponding transition functions is discussed. An
asymptotic of the transition function with respect to the state space vari-
able will go along the lines of the considerations in and is not discussed
here.
A remark t o the style of the paper: We aim to give more, and in a certain
sense new tools into the hands of those who are modeling with LBvy pro-
cesses or more general jump processes. More details (on a technical level)
of the impact of pseudo-differential operators in the theory of Markov pro-
cesses are given in 13, the more recent survey '' and in the monographs l4
and 15.
2. Meixner Processes
In this section we summarize various results on Meixner processes, i.e. real-
valued LBvy processes whose characteristic exponent is a continuous nega-
tive definite function of type
a< - ib
$ m , 6 , a , b ( < ) := -Zm< + 26 2
where m E R, 6 > 0, a > 0 and -7r < b < 7r.
For the corresponding process (Xy'6'"'b)t20we find
and
Im < ) -m<
~ L + 5 , ~ , b (= + 26tan-1 (9)
Moreover we find
a6
Re ‘$m,6,a,b(<) 2 yl<1
for all < E R, 161 2 y ,as well as with some co > 0
IIm ‘$m,fi,a,b(()I 5 CO(1 + Re ‘$m,6,a,b(<)). (11)
In addition it follows that
E(u,w ) = y
s, u’(z)v(x)d x
Since for the study of symmetric Dirichlet forms much more (analytic)
techniques are available, compare the monograph of M. F’ukushima, Y. Os-
hima and M. Takeda let us have a short look at the symmetric part
of the Meixner process ( X r 1 6 1 a 1 b ) t 2 0 i.e.
, the Lkvy process ( q s 1 a 3 b ) t 2 0
with characteristic exponent being the continuous negative definite func-
tion Re ‘$m,fi,a,b. w e find now
as well as
39
and
For the general case a longer calculation, see B. Bottcher 6, leads to the
following series representation of pf'a'b :
where
3. Symbols of Meixner-type
As mentioned in the introduction we want to make the parameters m, 6,
a and b in ( 1 ) state space, i.e. x-dependent and then identify (under some
conditions) this function of x and ( as a symbol of a pseudo-differential
operator generating a Markov process. In the following we denote by S(IR)
the Schwartz space of rapidly decreasing functions.
Definition 3.1. A. An arbitrary often differentiable function q : IR x R 4
C is said to belong t o the symbol class Sk(IR),k E IR, if for all a,P E No
there are constants cap 2 0 such that
sin2 y
, x E R and y E (-- -),
2’ 2 (29)
41
)( 2 )
+ %(z) tan(-)--b(z)
2
b”(z)
2
+ b(z) b’’(z)
6(z)sec2(-)-
2 2
+ 2S(z) tanh( 2 2 1 (33)
ib(x) a’(.)[ - ib‘(x)
+ 2S(z) sech2(a(..)[ 2
-
)( 2 1
e),
we may use (31) to reduce the estimates for @q(z, p 2 3, t o the estimate
for a,”q(z,[) and the estimates for z H (sech’ z ) z k ,and then (29) and (30)
give the result. Next observe that
a ( z ) <- ib(z)
L$q(x, <) = -im(z) + 6 ( z ) a ( ztank(
)
2 ) (34)
which implies the desired estimates (21) for Q = 1,,B = 0. Further we have
which again by (29) and (30) leads to (21) for a = 2, /3 = 0. But now (32)-
( 3 5 ) together with (31) as well as (29)-(30) imply q E S'(R). The ellipticity
condition (28) follows from the restrictions for the parameters and (10).
4. Meixner-type Processes
The aim of this section is to show that every pseudo-differential operator
-q(x, D ) with q being a Meixner Symbol has an extension, in fact a unique
extension, generating a Feller semigroup, hence gives rise to a Feller process,
or equivalently, for every q E M S ( R ) there is a stochastic process (Xt)t>o
with state space R such that
and (Tt)t?o,where
principle, compare Theorem 2.6.1 in 15. The serious problem is the solv-
+
ability of the equation Xu q ( x , D)u = f . This problem is overcome in
several steps:
1. Show that for every f E L2(R) and X 2 0 sufficiently large there is a
weak solution u E H i , i.e. u satisfies
~ x ( u $1 +
, := ~ ( u4 ), ~ ( u 0 = (f,
4 1, $10 for all 4EH~(R),
where B(.,.) is the continuous extension of (u,
u) I+ ( q ( x ,D)u,u)o from
H y R ) to Hi(R).
4x1 D ) ( H 3 ( R )c cco(R)l
The proof of Theorem (4.1), more precisely working out step 1-3, yields
more, namely that there is A0 > 0 such that for X 2 A0 the operator
-qm,6,a,b(x, D ) - Aid extends also to a generator of an L2-sub-Markovian
semigroup which we denote by (T,(2)’x)t20. Clearly on Cco(R) n L2(R) we
have
e-~t~,(c=
o )Tu, ( ~ ) J U a,e. (39)
44
Ttu(z)= ( 2 ~ ) - 7
1 s, e
ixc - t q m J A b
e ( E ) Q ( < ) d< (43)
which holds (at least) for all u E S(R).Thus we should long for
O t ) ( 3 : , < )= e
-qm,6,a,b
+
(zL)t r ( t , 2 , <), (44)
where r ( t ,z, E ) satisfies certain smallness conditions.
To proceed further we need some preparations. The class Sk (R) as defined
by (21) is a Frkchet space if topologized with the serninorms
~ q ( +(2)1=
z, (2n)-+ IJ, e i x ~ z<)&(<)
, d< 5 cm(u)plc,o(q) (46)
l i m a ( T t ) ( z , J )=
t-0 1 weakly in S'(IW), (48)
a ( T t ) ( zJ) +
, = e-q(z>E)t ro(t,z,[) (49)
where ro(t,., .) E S-l(R) and
lim ro(t,z, <) =
t-io o weakly in s-'(R.), (50)
and
eizEro(t,z, [ ) G ( [ ) dJ = 0.
lii
Finally, let us consider the result in a heuristic way. Using the semigroup
- we arrive for small t > 0 at
property of (Tt),>0
~,+,u(xM
) (27r1-4 S, eizEe-q(z>E)t(TsuT(5) dJ1
In particular, if
s is small and therefore p,(y,A) can be substituted by
~ ~ p ~ ~ y ~ ' " y ~d z~ we
u ~find
y ~now
~ bfor~ Iz
y- ( ~as) well as s and t small
~ zoI
that
References
1. Adams, D. and L. I. Hedberg, Function spaces and potential theory. Vol. 314
of Grundlehren der math. Wissenschaften. Springer Verlag, Berlin 1996.
2. Barndorff-Nielsen, 0. and S. Z. Levendorski:, Feller processes of normal in-
verse Gaussian type. Quantitative Finance 1 (2001), pp. 318-331.
3. Barndorff-Nielsen, O., T. Mikosch, and S. I. Resnick (eds.), Le'vy Processes -
Theory and Applications. Birkhauser Verlag, Boston 2001.
4. Barndorff-Nielsen, 0. and N. Shephard, Modelling b y Le'vy processes f o r fi-
nancial econometrics. In 3 , pp. 283-318.
5. Berg, C. and G. Forst, Non-symmetric translation invariant Dirichlet forms.
Inventiones Math. 21 (1973), pp. 199-212.
6. Bottcher, B., PhD-thesis, University of Wales Swansea. (In preparation).
7. Eberlein, E., Application of generalized hyperbolic Le'vy motions to finance.
In ', pp. 319-336.
8. Fraenkel, L. E., Formulae for,higher derivatives of composite functions. Math.
Proc. Cambridge Phil. SOC.83 (1978), pp. 159-165.
9. Fukushima, M., Y . Oshima, and M. Takeda, Dirichlet forms and symmetric
Markov processes, Vol. 19 of d e Gruyter Studies in Mathematics. Walter de
Gruyter Verlag, Berlin 1994.
10. Grigelionis, B., Processes of Meixner type. Lithuanian Math. J . 39 (1999),
pp. 33-41.
11. Hoh, W., Pseudo differential operators generating Markov processes. Habili-
tationsschrift. Universitat Bielefeld, Bielefeld, 1998.
12. Hoh, W., A symbolic calculus for pseudo differential operators generating
Feller semigroups. Osaka 3. Math. 35 (1998), pp. 789-820.
13. Jacob, N., Pseudo-differential operators and Markov processes. Vol. 94 of
Mathematical Research. Akademie Verlag, Berlin 1996.
14. Jacob, N., Pseudo-Differentia1 Operators and Markov Processes, Vol. I:
Fourier Analysis and Semigroups. Imperial College Press, London 2001.
15. Jacob, N., Pseudo-Differential Operators and Markov Processes, Vol. 11: Gen-
erators and Their Potential Theory. Imperial College Press, London 2002.
16. Jacob, N. and R. L. Schilling, Estimates for Feller semigroups generated b y
pseudo differential operators. In: Rakosnik, J. (ed.), Function Spaces, Dif-
ferential Operators and Nonlinear Analysis. Prometheus Publishing House,
Praha 1996, pp. 27-49.
17. Jacob, N. and R. L. Schilling, Le'vy-type processes and pseudo differential
operators. In 3, pp. 139-168.
47
ZDZISlAW BRZEZNIAK
Department of Mathematics
University of Hull
Hull HU6 7RX, U.K.
E-mail: z. brzezniakQmaths.hu11.ac.uk
In this paper we study It8 integral in 2-smooth Banach spaces. Burkholder inequal-
ity is proved using It6 formula in certain subclass of such spaces. Relationship with
an integral introduced recently by Mikulevicius and Rozovskii is discussed. Finally,
Wong-Zakai type approximation for such integrals is proved.
1. Introduction
This paper has its origin in the author’s attempt to understand an impor-
tant work by Mikulevicius and Rozovskii 28. In order to study stochastic
Navier-Stokes equations in Rd for d = 2,3 in Sobolev space HS9P, the au-
thors introduce a new type of It6 integral for some Banach space valued
processes. One of the aims of the current presentation is to show that the
Mikulevicius-Rozovskii integral is a special case of an integral in 2-smooth
Banach spaces first introduced by Neidhardt in 31 and then extensively
studied and used by the present author and his collaborators. The main
object however is to present a concise and detailed exposition of the sub-
ject. The paper is organised as follows. In the section 2 we recall the basic
definitions, i.e. of 2-smooth Banach space and of It6 integral with valued
in 2-smooth Banach space. Section 3 is devoted to statement and proof of
the Burkholder inequality for It6 integrals taking values in certain class of
Banach spaces. Let us note here that Burkholder inequality is valid in 2-
smooth Banach spaces, see l6 and 32. The class of Banach spaces considered
in this section is big enough as it contains important examples of L P , p 2 2
and Besov and Sobolev-Slobodetski spaces. In section 4 we show how the
theory on It6 integration in 2-smooth Banach spaces can be used to solved
certain nonlocal stochastic differential equations. In the section 5 we inves-
tigate the relationship of the integral introduced by Mikulevicius-Rozovskii
with the one in 2-smooth Banach spaces. We show that the former is a
48
49
special case of the latter. For this we use a result of the author and Peszat
on identification of y-radonifying operators with values in LP-spaces with
certain class of integral operators.
We conlude the paper with a discussion of dependence of the It6 integral
on the Wiener process. We prove that the Stratovich integral is equal to
limits of the Riemann sums with the mid-point approximations is repalced
by interval averages. Our result should be seen as in conjunction with
the authour's paper with A Carroll on Wong-Zakai approximation for
stochastic differential equations in 2-smooth Banach spaces.
(X, I . I ) satisfies
PX(t) 5 kt2, t E (0,1].
The notion of a 2-smooth Banach space was introduced by Pisier in 34.
Pisier proved there that X is a 2-smooth Banach space iff one of the fol-
lowing two conditions is satisfied
(i) There exists a constant A > 0 such that
In fact, the implication X is 2-smooth -----r. (i) had been earlier proven by
Figiel & Pisier in 19, see also l4 p. 144. The proof of converse implication,
only alluded to in 34, is rather straightforward.
A Banach space X satisfying property (2) is usually called an M-
type 2 Banach space. Although an It6 type integral for 2-smooth Banach
spaces was first introduced by Hoffmann-Jorgensen and Pisier l8 only for
1-dimensional square integrable martingales, a complete construction was
50
carried out by Neidhardt in 31, see also Belopolskaya and Daletskii 2 , Det-
tweiler 16, Brzeiniak and references therein. In order to introduce this
integral we need one more new notion, i.e. of a y-radonifying operator. If
H and X are separable real Hilbert and resp. Banach spaces, a bounded
linear operator L : H -+ X is called y-radonifying iff L ( ~ H is) a-additive,
where Y H is the canonical Gaussian distribution on H . If this is the case,
L ( ~ Hhas) a unique extension to a a-additive Bore1 probability measure V L
on X. One can then also show that VL is a centered Gaussian measure on
X with Reproducing Kernel Hilbert Space (RKHS) (i.e. the Cameron Mar-
tin space) equal to H . In particular, in the spirit of L Gross 17, the triple
( H ,X , VL) is a Abstract Wiener Space (AWS). The set of all y-radonifying
operators from H to X we will denote by R ( H , X ) . Note that in 31 and
earlier papers this set is often denoted by R ( H ,X ) . For L E R ( H , X ) one
puts
that
One should mention here that in order to prove (6) both Neidhardt 31
and Dettweiler l5 used the property (i), while the author in and above
has used the M-type 2 property (ii). The last inequality (6) shows that
I is a bounded (obviously) linear map from I : Mstep(a,b;&,n(H,X))to
L2(R;X ) . Since the former is dense in M 2 ( a ,b; R(H,X)), I has a unique
extension t o a bounded linear map from the whole of M 2 ( a ,b; R ( H ,X ) )
with values in L2(R,IF,X). Moreover, this extension, also denoted by I
satisfies
b
W(t)I2I C 2 ( X ) E / Ilt(t)ll&H,X) dt. (7)
I'S(s) dW(s)=w[o,T)r)?
where I = la$.
One can prove that for 0 I r 5 t !E s," c(s) dW(s)lFT)=
Jit ( s ) d W ( s ) . We also have, see 4, the following
ai.e. (i)qlnt : Rt + X is Ft measurable, for any t 2 0; (ii) for almost all w E R, the
function [O,.(LO)) 3 t H q ( t ,w ) E X is continuous.
53
j
The series is absolutely convergent and its sum is independent of the choice
of the ONB { e j } . If x
= Y and A = B we write trAA instead of trA,AA.
z ( t )= ~ ( c )+ I " +L
a(.) ds ((s) d W ( s ) , t E [c,4. (8)
Then for all t E [c,d ) , a.s,
3
~ ~ ~~
bSimply, g,
appropriate space.
and exist and are continuous on [ c , d ) x X with values in the
54
3. Burkholder inequality
In this section we assume that our
( H ) X is a real separable Banach space such that there exists p E [ 2 , c o )
for which the function ' p p : X 3 x H l x l p E IR is of C2 class and there
are k 1 , kz > 0 such that for every zE X ,lp'(x)I 5 klIz(p-' and ('p''(x)(5
2k2 1 2 1 p - 2 .
Note that the Sobolev Hsip-spaces with p E [2,co) and s E IR satisfy the
condition (H). Moreover, a Banach space X satisfying (H) is 2-smooth7
see l4
If q 2 p , the following is a special case of Theorem 1.1 from 12.
Theorem 3.1. Assume that X is a Banach space satisfying the condition
(HI. A s s u m e that 5 E M ~ , ( O , c o , R ( H , X ) .Let x ( t ) = J,"C(s)dW(s),
t 2 0. Suppose q E ( 1 , a). T h e n there exists a constant Kq > 0 depending
only o n q, H , X , and the constants kl, kz appearing in (H), such that for
every T > 0 ,
Remark 3.1. With a slight modification of the proof below one can show
that in fact the Burkholder inequality above is also valid for any accessible
(and hence any bounded) stopping time.
55
Proof. The first step is to prove this result for q = p . We follow the above
mentioned paper l2 where a more general result is studied. Suppose first
that A is a bounded dissipative linear operator. Since ‘p(x)= 1xIP is of C2
class we can use It6’s formula of Neidhardt, see Theorem 2.1 above, and
obtain
Y(4 := 1 t
9’( 4 s ) ) ( E ( 4 ) dW(s), t 2 0.
4(t)= 1 (4t
Now, we shall deal with the second term on the RHS of (13). Since for L
from R ( H , X ) and a bilinear mapping A : X x X 4 R, Itr A o ( L ,L ) ( I
]A1 . JJL))2,
we have, again by ( H p ) ,
We shall study each term on the RHS of (13) separately. Let E > 0. First
we have
where we have used Holder and Young inequalities. Similarly for the second
term we have
which proves (12) for q = p. The proof in the case q > p follows the same
lines. It is enough to observe that if the Banach space satisfies the condition
(H) with p 2 2 then it also satisfies this condition with q 2 2. In order to
complete the proof we need to consider the case q < p. The proof in this
case is motivated by the proof of the Burkholder inequality given by Revuz
and Yor in 36. It is based on the following (see Proposition IV.4.7 therein)
57
4. An Example
This Example is motivated by a question raised by Terry Lyons. Let S1
be the unit circle (with normalized Haar measure) and let H = H19"(S1)
and B = L2(S1). Let E = H"ip(S1) with < a < 3.
Let us recall that
H"ip(S1) = [LP(S1);HIJ'(S1)la, the complex interpolation space. Then it
is well known that E is a Banach algebra. Note also that E is 2-smooth
Banach space.
Consider three maps
A : E 3 u H { H 3 H u . y E E } E R ( H ,E ) , (15)
A :E 3 uH { B 3 y H U * Y E E } E R(R,E), (16)
B : E 3 u H { E 3 y H u.yE E } E L ( E , E ) . (17)
Since E is a Banach algebra, B is a (well defined) bounded linear operator.
Since A(u) = B ( u ) o i, where i : H E is the natural embedding, the
map A : is well defined and bounded as well. Here we use the fact that
i E R(H,E).
Since for f E L'(S1) the map A, : u H f * u is bounded from LP
into L P (by the Young inequality) and from HIJ'(S1) into H1+'(S1) (by the
former fact and equality D ( f * u)= f * (Du)) we infer, by means of the
interpolation theory, that As is a bounded linear map from E = H"J'(SI)
into itself. Therefore, is a bounded linear map from E into L ( L 1 , E ) ,
hence into C ( f i , E ) . To prove that is a bonded linear map from E into
R(I?,E) we argue as follows. Let u E E . Then, Dau E LP(S'), where D"
58
is the fractional power of the derivative operator D . We will show first that
if u E El then
-
linear operator
K : L2 3 y v * y E Lp
is finite, as v E LP C L2, the result follows, see l1 and Theorem 5.1 in the
next section. In fact we have proven that the map A : E -+ M ( H ,E ) is
well defined, linear and bounded. Hence the following result follows directly
from 31 and 8 .
d u ( t ) = u(t)d W ( t ) + u ( t )* d W ( t )
u ( 0 ) =uo.
Remark 4.1. The reason we used the H"1p spaces and not the Sobolev-
Slobodetski W"tP was that for u E Wa,p,D"u may not be an element of
L2. Thus Theorem 4.1 may be not true in this case. However, in Theorem
4.2 we can use E = WaJ'(S1).
59
(1) K is y-radonifying;
(2) There exists a u-measurable function K. : 0 --+ Y with
( K ( Y ) ) ( X )= ( K ( X ) , Y ) , Y E Y.
Moreover, there exists a constant C > 0 such that for all IC E LP(0, Y),
& A x ) ( Y , e j ) = C ( 9 ( 4 , e j ) ( Y , e j )= ( S ( X ) , Y )
j j
l g ( r ) d W ( r )=
I’ i j ( r ) d W ( r ) , t 2 0.
This integral, being just a special case of the integral introduced earlier in
section 2 satisfies all its properties. In particular, it satisfies the Burkholder
60
inequality (12), a special case of which in the present situation takes the
following form. If p 2 2, then
Let us now show its another property whose a byproduct is that it coin-
cides with the It6 type integral of Mikulevicius-Rozovskii,see 28 (subsection
5.1 in the Appendix) and 27.
Proof. This result is in fact a special case of the It6 formula, see Theo-
rem 2.1. Indeed, cp can be identified with a bounded linear map on LP(0).
Since then for e E LP(0), = cp and cp”(J) = 0, we get that a.s.
(s,”g(r)d W ( r ) ,9) = J;(g(~)cp)dW(+ Denoting by E(r) = M r M E
L(Y,R) = Y* E Y and observing that L(Y,R) S R(Y,R) the integral
S,”c(r)dW(r)is again a special case of the It6 integral from section 2.
Hence, S,”S(r)dW(r) = C,”=, tj(r)dWj(r)= C&(gj(r),cp)dWj(r) what
concludes the proof of the Theorem. 0
Remark 5.1. The above can be generalised to any Banach space X which
is isomorphic with the space LP(O),in particular for the Bessel spaces
He>p(Rd). Indeed, the isomorphism between the latter space and LP(Rd)is
given by f H (1 - A)e/2f.
trA = C A ( e j , e j )
j
where 0 = tt < t? < . . . < t"Nn, 5 T < t",n)+l < 00 is a partition of the
interval [O,T].
Recall that L ( E ,X ) is the space of bounded linear maps from E t o X
and that the imbedding C ( E , X ) 3 A H A o E Z R ( H , X ) is bounded. Here
i : H 4 E is the canonical imbedding. Our main result is the following
theorem.
Then
62
Remark 6.1. We state and prove this theorem for the second moment.
A generalization to any p 2 2 is possible and will be discussed in ‘. For
simplicity and clarity of exposition we only give a proof in a special case of
one-dimensional Wiener process, i.e. H = E = R and (identifying C(W,X)
with X ) F ( t , x ) = x , t E [O,T],x E X . We will also assume that a(t) = 0
for t E [O,T].The proof in the general case will also be discussed in ‘.
m,(t) = Sup{k : tk 5 t }
Proof. From the uniform continuity (on interval [0,TI) of paths of both
processes W ( t )and <(t)we have
sup II;(t)l2 -+ o a.e.
o<t<T
Moreover,
From (24) and the Doob inequality, see Theorem IV.8.2 in 26 and Proposi-
tion 2.1 in this paper, we infer that
Lemma 6.2.
64
Lemma 6.3.
Therefore
65
I z ( t )- 1 t
+
b(s)d s = I i 1 ( t ) Iz2(t), (36)
m(t)-1
k+l
s)b(s)d s .
k=O n
Sk
Since n -
n
*(w n
- s)d s = n so&ds = & we obtain
s
cn l * ( D
m(t)-1
k=O n
k+l - s)b(s)ds=
cm(t)-1
k=O
nJ**(*
n
n - s) (6(s) - b ( k ) ) ds
+ cm(t)-1
'
n L * ( - - - rk) b+(1- ) d s k
k=O K
n n
m(t)-1 m(t1-1
1 1 k
k=O n
lctl S)
(
b(s)-b(-) ds+?
k-0 ,!I(--)
and thus in order t o prove (37) it is enough to prove
However (38) and (39) easily follow from continuity of paths of the
process b(s) and assumption (24) by applying the Lebesgue Dominated
Convergence Theorem. 0
66
Lemma 6.4.
Proof. Since 122(t)is constant on each time interval (k, y]we have
where
m(T)-1
Now we are going t o show that for fixed n, the sequence (Y?,)k=Ois a
m(T)-1
martingale with respect to a filtration ( C k ) + O , where c k = F w .For
this it is enough to show
IE(Xp-1) =0 (43)
Let us observe that each term on the right hand side of (44) can be estimated
in the same way. Thus we may take i = 0 and get
From the last twoinequalities, (44), (45) and the fact that m(T)= m,(T) 5
cn we get (40). This concludes the proof. 0
Acknowledgments
The authour would like to thank Marek Capihski, David Elworthy, Terry
Lyons, Jan van Neerve, Martin Ondrejat, Szymon Peszat and Boris Ro-
zovskii for their helpful discussion on various topics related to this paper.
References
1. Baxendale, P., Gaussian measures o n Function Spaces, Amer. J. Math.
98, 891-952 (1976).
2. Belopolskaya, Ya.L. and Daletskii, Yu.L., Diffusion processes in smooth
Banach manifolds. I, Trans. Moscow Math. SOC.1, 113-150 (1980). Rus-
sian original: Trudy Moskov. Mat. Obsh. 37, 107-141 (1978)
3. Belopolskaya, Ya.L. and Daletskii, Yu.L., STOCHASTIC EQUATIONS AND
DIFFERENTIALGEOMETRY,Mathematics and Its Applications vol. 30,
Kluwer Academic Publishers, Dortrecht Boston London 1990.
4. Brzeiniak, Z., Stochastic PDE in M-type 2 Banach Spaces, BiBoS
preprint (1991).
5. Brzeiniak, Z., Stochastic Convolution in Banach spaces, Stochastics and
Stochastics Reports 61, p.245-295, 1997.
68
6. Brzeiniak, Z. and Capinski, M., Wong Zakai Theorem for stochastic in-
tegrals in Banach spaces, in preparation.
7. Brzeiniak, Z. and Carroll, A,, Wong Zakai Theorem on Loop Manifolds,
40 pages, to appear in SQminairede probabilitis XXXVII, edt. M. Ledoux.
8. Brzeiniak, Z. and Elworthy, K.D., Stochastic differential equations o n
Banach manifolds; applications to diffusions on loop spaces, MFAT (a
special volume dedicated to the memory of Professor Yuri Daletski), 6,
no.1, 43-84 (2000).
9. Brzeiniak, Z. and van Neerven, J., Stochastic convolution in separable
Banach spaces and the stochastic linear Cauchy problem, Studia Math.
143, no. 1, 43-74 (2000)
10. Brzeiniak, Z. and van Neerven, J., Space-time Regularity for linear
stochastic evolution equations driven b y spatially homogeneous noise, to
appear in J. Math. Kyoto Univ.
11. Brzeiniak, Z. and Peszat, S., Space-time continuous solutions t o SPDEs
driven b y a homogeneous Wiener process, Studia Mathematica 137, 261-
299 (1999),
12. Brzeiniak, Z. and Peszat, S., Maximal inequalities and exponential tail
estimates for Stochastic Convolutions in Banach Spaces, pp. 55-64 in
STOCHASTIC PROCESSES, PHYSICS A N D GEOMETRY: NEW INTERPLAYS.
I A Volume in Honour of Sergio Albeverio, CMS Conference Proceedings,
v. 28,Providence, Rhode Island (2000).
13. Brzeiniak, Z. and Peszat, S., Stochastic two dimensional Euler Equations,
Annals of Probability 29, 1796-1832 (2001)
14. R. Deville, G. Godefroy and V. Zizler, SMOOTHNESS AND RENORMING IN
BANACH SPACES,Pitman Monographs and Surveys in Pure and Applied
Mathematics 64, Longman Scientific and Technical, 1993.
15. Dettweiler, E. Stochastic Integration of Banach Space Valued Functions,
in STOCHASTIC SPACE-TIME MODELSA N D LIMITTHEOREM, pp. 33-79,
Arnold, L. and Kotelenez, P. edts., D. Reidel Publ. Comp. 1985
16. Dettweiler, E., Stochastic Integration Relative to Brownian Motion o n a
General Banach Space, Boga - Tr. J. of Mathematics, 15,6-44 (1991).
17. Gross, L., Measurable functions o n Hilbert space Trans. Am. Math. SOC.
105 372-390 (1962).
18. Hoffman-Jorgensen J. and Pisier, G., The Law of Large Numbers and
the Central Limit Theorem in Banach Spaces, Annals of Probabilityl,
587-599 (1976)
19. Figiel, T. and Pisier, G., Series aleatoires duns les espaces uniformement
convexes ou uniformement lisses, C. R. Acad. Sci., Paris, Sir. A 279,
611-614 (1974).
20. N. Ikeda and s. Watanabe, STOCHASTIC DIFFERENTIAL EQUATIONS A N D
DIFFUSION PROCESSES, North-Holland, Amsterdam - Oxford - New York
1981.
21. Karatzas, I., Shreve, S.E. BROWNIAN MOTIONAND STOCHASTIC CALCU-
LUS, Springer Verlag, New York Berlin Heidelberg, 1988.
22. Kunita, H., STOCHASTIC FLOWS AND STOCHASTIC DIFFERENTIAL EQUA-
TIONS, Cambridge University Press, Cambridge, 1990.
69
T. CARABALLO
Dpto. Ecuaciones Diferenciales y A n d i s i s Nume'rico, Universidad de Sevilla,
Apdo. de Correos 1160, 41 080-SE VILLA, Spain,
E-mail: [email protected]
1. Introduction
The long-time behaviour of flows is a very interesting and important prob-
lem in the theory of fluid dynamics, as the vast literature shows (see
Temam 2 6 , Hale 18, Ladyzhenskaya 19, among others, and the references
therein), and has been receiving very much attention over the last three
decades.
One of the most studied models is the Navier-Stokes one (and its vari-
ants) since it provides a suitable model which covers several important
fluids (see Temam 24,26 and the references inside these).
On the other hand, another interesting question is to analyze the effects
produced on a deterministic system by some stochastic or random distur-
bances appeared in the problem. These facts motivated the analysis done
in Caraballo et al. l 1 and the one in the present work. Therefore, our main
objective is t o show some aspects of the effects produced in the long-time
behaviour of the solution to a two dimensional Navier-Stokes equation un-
der the presence of stochastic perturbations, since it is very interesting to
investigate if a fluid subjected to random influences is asymptotically more
or less stable than the deterministic unperturbed one.
There exists a controversy concerning the different interpretations which
can be given to the stochastic terms used to model our problem. Two for-
mulations are the most commonly used for the noise in the literature: It6's
formulation and Stratonovich's one. Each interpretation gives a different
solution of the stochastic equation, so they provide different answers t o the
70
71
same problem. There exist several reasons which make reasonable both
possibilities and there exists a rule which permits us to pass from one kind
of equation to the other (see Arnold Oksendal 2 1 , Kunita 2 0 , among
others). However, when one is analyzing the long-time behaviour of the
solutions, special care should be paid to the choice of the model since the
solutions of both stochastic equations can have totally different behaviour.
We will comment again about this in the final section.
In this work, we will first recall some results from Caraballo et al. l 1
concerning the exponential behaviour of the solutions to our stochastic 2D-
Navier-Stokes model. Then we will improve those results by giving some
information concerning the general decay rate of solutions. To this end, we
will consider the following stochastic 2D-Navier-Stokes equation:
+ + Vpldt +g(t,X ) d W ( t )
i
d X = [vAX - ( X ,0) X f(X)
divX = 0 in [O, co) x D ,
X = 0 on [O, co) x r,
X ( 0 , z ) = X o ( z ) , J: E D ,
where D is a regular open bounded domain of R2 with boundary I?, u is
the velocity field of the fluid, p the pressure, v > 0 the kinematic viscosity,
Xo the initial velocity field, f the external force field and g(t,z)dW(t)
the random field where W ( t )is an infinite dimensional Wiener process,
i.e., if (Q, P, 3) is a probability space on which an increasing and right
continuous family {3t}ZE~0,00)of complete sub-o-algebra of 3 is defined,
and &(t) ( n = 1 , 2 , 3 , ...) is a sequence of real valued one-dimensional
standard Brownian motions mutually independent on (0,P, S),then
M
n=l
where A; 2 0 ( n = 1 , 2 , 3 . . . ) are nonnegative real numbers such that
C,"==, A; < +co, and {en} ( n = 1 , 2 , 3 , ...) is a complete orthonormal
basis in the real and separable Hilbert space K . Let Q E L ( K , K ) be the
operator defined by Qe, = xien. The above K-valued stochastic process
W ( t )is called a Q-Wiener process.
Our problem can be set in the usual abstract framework by considering
the following Hilbert spaces:
H = the closure of the set {u E C p ( D ,R2) : divu = 0} in L 2 ( D ,R2)
with the norm (u(= ( u ,u ) ; , where for u,u E L 2 ( D ,R2),
72
Then, it follows that H and V are separable Hilbert spaces with associated
inner products (., .) and ((., .)) and the following is safisfied:
V c H = H' c V',
where injections are dense, continuous and compact. Now, we can set
A = -PA where P denotes here the orthogonal projector from L 2 ( D ,R2)
onto H, and define the trilinear form b by
As we shall need some properties on this trilinear form b, we list here the
ones we will use later on (see Temam 26):
d X ( t ) = [ - v A X ( t ) - B ( X ( t ) )+ f ( X ( t ) ) ld t + S ( t , X ( t ) ) d W ( t ) , (2)
where f : V + V', g : [ O , c o ) x V + L ( K ,H) are continuous functions
satisfying some additional assumptions (see conditions below). Also we
consider the deterministic version of this equation, namely,
+
d X ( t ) = [ - v A X ( t ) - B ( X ( t ) ) f ( X ( t ) ) ]dt. (3)
First, we give the definition of the weak solutions to stochastic 2D-Navier-
Stokes equation (2).
73
( l a ) X ( t ) is Qt-adapted,
( l b ) X ( t ) E L”(0, T ;H ) n L2(0,T ;V ) almost surely for all T > 0,
(lc) the following identity in V’ holds almost surely, for t E [0,a)
X(t)=X(0) + s,” [ - - Y A X ( S ) B ( X ( s ) )+ f(X(s))]ds
-
+ s,”ds, X(s))dW(s).
As we are mainly interested in the analysis of the asymptotic behaviour of
the weak solutions t o the problem (2), we will assume the existence of such
weak solutions (see, for instance, Bensoussan or Capinski and Gatarek
for some results on the existence and uniqueness of solutions).
We also recall some definitions from Caraballo et al. l 1
Theorem 2.2. Suppose that all the conditions in Theorem 2.1 are satisfied.
Then, any weak solution X ( t ) to (2) converges to the stationary solution
u, of (4) almost surely exponentially.
( a ) g ( t , urn) -- 0, t 2 0,
(b) II S ( t ,). - d t ,). ; ,1 5 cg II 21 - 21 II, cg > 0, u,v E v.
If 2v > 2p c i + + &
11 u , 11, then any weak solution X ( t ) to (2)
converges to u, exponentially an mean square and so u, is exponentially
stable in mean square. That is, there exists a real number y > 0 such that
d z ( t )= az(t)dt+ bz(t)dW(t),
z ( t )= z(0)exp ( a { - T) + t bW(t)}
76
cz
where y := ~ ( X I U- c - + + 9). 2
In particular, exponential stability of
sample paths with probability one holds i f y > 0.
We omit the proof since this result is a particular case of Theorem 4.1.
77
First we will prove a general theorem which extends Theorem 3.1 and
then we will comment about its consequences.
Let us apply Ito's formula for our soution X(t) satisfying the
assumptions mentioned in the theorem. Then, it follows
78
and, applying once again ItG’s formula to the function log lX(t)I2, and
taking into account the hypotheses, it follows
+q t
Ix;s)lz (X(S)Ig(s1 X ( s ) ) d W ( s ) )- 2l t
p(s)ds.
*
law of iterated logarithm,
and, taking into account the super-exponential growth of X(t), the result
follows immediately.
Acknowledgments
I would like t o thank Aubrey Truman and Ian Davies for the kind invita-
tion t o take part in this Conference on Probabilistic Methods in Fluids. I
finished this work during my stay in the Mathematics Institute (University
of Warwick, June-August 2002). I would like to thank the Royal Society of
London for their generosity, and especially, to James Robinson and Tania
Styles for the hospitality and friendship they offered me, what made me
feel as if I would have been at home.
This paper has been partially supported by Secretaria de Estado de
Universidades e Investigacibn (Spain).
References
1. L. Arnold, Stabilization by noise revisited, Z. angew. Math. Mech. 70 (1990),
235-246.
2. L. Arnold, Stochastic Differential Equations: Theory and Applications, J.
Wiley and Sons, New York, (1974).
3. A. Bensoussan, Stochastic Navier-Stokes equations, Acta Applicandae Math.,
38 (1995), 267-304.
4. Z. Brzezniak, M. Capinski and F. Flandoli, Pathwise global attractors for
stationary random dynamical systems, Prob. Th. Rel. Fields, 95 (1993), 87-
102.
5. M. Capinski and D. Gatarek, Stochastic equations in Hilbert spaces with
application to Navier-Stokes equations in any dimension, J. of Funct. Anal.
126(1994), 26-35.
6. M. Capinski and N. Cutland, Measure attractors for stochastic Navier-
Stockes Equations, Electronic J. of Prob., 3(1998), 1-15.
7. M. Capinski and N.J. Cutland, Existence of global stochastic flow and at-
tractors for Navier-Stokes equations, Prob. Th. and Rel. Fields 115(1999),
121-151.
8. T. Caraballo, M. J. Garrido-Atienza and J. Real, Asymptotic stability of non-
linear stochastic evolution equations, Stoch. Anal. Appl., to appear.
9. T. Caraballo, M.J. Garrido-Atienza and J. Real, Stochastic stabilization of
differential systems with general decay rate, Systems and Control Letters, to
appear.
10. T. Caraballo and J.A. Langa, Comparison of the long-time behaviour of
linear It6 and Stratonovich partial differential equations, Stoch. Anal. Appl.
19(2001), 183-195.
11. T. Caraballo, J.A. Langa and T. Taniguchi, The exponential behaviour
and stabilizability of stochastic 2D-Navier-Stokes equations, J . Diff. Eqns.
179(2002), 714-737.
12. T. Caraballo and K. Liu, On exponential stability criteriaof stochastic partial
differential equations, Stochastic Processes and their Applications 83 (1999),
289-301.
13. P. Constantin and C. Foias, "Navier-Stokes Equations", The University of
83
P.L. CHOW
Wayne State University
Detroit, Michigan 48202, USA
E-mail: [email protected]
1. Introduction
Consider the stochastic wave equation:
8,". =v2u + o(u)atW(x,t ) (1)
where at denotes the partial derivative in t , V2 the Laplacian; W ( . ,t ) is a
Wiener random field. For d = 1 or 2, Mueller proved that the equation
Eq. (1) has a unique long-time global solution pointwise in (5, t ) E Rd x
[0, m), provided that o(u)grows no faster then lul(logIu1)' with T E ( 0 , a ) .
In the case d > 1, the Weiner field W ( x , t )must be smooth in z, because
nonlinear equations such as Eq. (1) and Eq. (2) below are not well defined
if a t W ( x ,t ) is a space-time white noise (see Walsh '). In view of Mueller's
result, the following question arises naturally. That is, if o(u)grows like
' for a sufficiently large T > 1, whether a solution to Eq. (1) may blow
u
up in a finite time. This question is still open. Here we consider a related
problem as follows:
a,zu = v 2 u + +
f ( u ) .(u)atw(x, t ) , z E Rd, t > 0
{ u(x, 0) = g(z), dtu(z,0) = h ( x ) ,
(2)
where nonlinear terms f ( u ) and a(u) are assumed to grow like polyno-
mials in u,and the initial data g and h are given functions. In general
84
85
Let W ( x , t )be a continuous Wiener random field with mean zero and
covariance function ~ ( z9,) defined by
Now we consider the Cauchy problem for the linear hyperbolic equation
with a random perturbation:
+
i [a; - A(x,D ) ] u ( x t, ) = f (z, t ) & M ( x ,t ) , 0 < t < T ,
u ( z ,0) = uo(x), &u(x,0 ) = vo(2), 2 E Rd,
where A ( x ,D ) is a strongly elliptic operator of second order of the form:
(4)
A(& D)cp(x)= c d
i,j=l
p (b(z)cp(x),
~ z , [ ~ i j ( ~ ) ~ z , c- 41 (5)
where the coefficients aij = u j z and b are smooth functions such that
ao(1 + m2)5 c d
i,j=l
U%)&<j I + ls1)2), t , x E Rd,
for some constants a1 2 a0 2 0. This condition implies that ( - A ) is a
self-adjoint, strictly positive linear operator on H = L 2 ( R d )with domain
D(A)= H z ( R d )and its square root B = a
is also a self-adjoint, strictly
positive operator with domain D ( B ) , which is a Hilbert space under the
inner product (9, h ) := ~ (Bg,Bh). Since the norms 11 . I I B and 11 . 111 are
equivalent, we have D ( B ) 2 HI.
rt
87
or equivalently,
rt
(7)
where we set
and
with I being the identity operator on H . Introduce the Hilbert space ‘FI =
( H I x H ) . As a linear operator in ‘FI, A generates a strongly continuous
semigroup etA on ‘H. Now regarding Eq. (7) as a stochastic evolution
equation in ‘FI in a distributional sense, we have the following lemma:
Then the equation Eq. (7), or Eq. (6) has a unique (mild) solution $t =
(ut,ut) which is a continuous predictable ‘FI-valued process. Moreover at
satisfies the energy equation:
rt rt
Notice that, due t o the lack of required smoothness of solutions, the general
It6 formula does not hold here. The energy equation Eq. ( 9 ) can be proved
by a smoothing technique, such as the Yosida approximation (Yosida 5 , as
done in (Chap. 5, Da Prato and Zabczyk 6 ) , and then taking a proper limit.
88
The energy inequality Eq. (11) can be shown to hold by applying the It6
formula to the energy equation and by invoking Burkholder's submartingale
inequality. The proof is similar to the special case given in Chow and will
be omitted.
+
i (8,"- A)ut = f t ( J u t ) & M t ( J u ) , t > O
uo = 9, &uo = h. (12)
Again we rewrite the equation Eq. (12) as a stochastic system in the Hilbert
space IH:
ut = uo + J, usds
t
= YO + s,' Ausds + F t ( J u ) ,
(15)
Vt
$t = $0 + 1t
A$& + 3t(4),
where dt and A are defined as before and
89
j=O
T r [ C t ( J u ) R C ; ( J 7 4I C3(1 + llull:),
and
T r { [ C t ( J u )- C t ( J u ’ ) ] R [ C t ( J u-
) Ct(Ju’)]*
I C4llu - U’III: a s . ,
f o r any u , u‘ E H I ,t E [0,TI, where * denotes the adjoint.
(3) W, is a H-valued process with covariance operator R such that
+2
e ( u t , vt) = e(uo,uo)
Under the above conditions (1)-(3), it is easy to check that the coeffi-
cients of the evolution equation (3.5) in IFI satisfies the usual global Lip-
schitz continuity and linear growth conditions. Theorem 3.1 follows from
a standard existence theorem (Theorem 7.4, Da Prato and Zabczyk 6 , for
stochastic evolution equations in a Hilbert space.
and
and
T r { [ C , ( J u )- C t ( J u I ) ] R [ C t ( J u-) C t ( J u I ) ] * )
and
(N4) Suppose that, for any u.E C([0,TI,HI)n C'( [0,TI,H) with &u =
1
v, there exist constants c l , c2 > 0 and IE < - such that, for any
2
t E [O, TI,
t
< c1+ c2
- J, e(us,v,)ds + IEe(ut,vt) as..
4. Applications
Let us consider the following initial-value problem in R3:
(a; - c 2 v 2+ y2)ut = ft(.t) + at(J..t)atwt, t > 0 ,
210 = g, dtuo = h,
92
where c and y are positive constants, while ft and ut are nonlinear (deter-
ministic) functions of polynomial type. In comparison with Eq. (12), we
have A = (c2V2- y2),f t ( J u )= ft(u)and M t ( J u ) is defined by Eq. (13).
In particular, we assume that the following conditions hold:
j =O
j=1
and
and
Tr R =
s r(x,x)dx < 00
To = < m.
sup ?-(x,x)
z€Rd
Lemma 4.1. For u,v E C,W and 1 5 k 5 m, there exist positive constants
~ 1~2, such that
and
With the aid of this lemma and conditions (P1)-(P3),we can apply The-
orem 3.2 to give a local existence theorem for Eq. (20) with polynomial
nonlinearities.
Theorem 4.1. Suppose that conditions (Pl)-(P3) given above hold true.
Then, for g E H I and h E H , the Cauchy problem Eq. (20) in Rd,for
d _< 3, has a unique continuous local solution ut E H I with &u, E H ,
provided that m 5 3 for d = 3, and m 2 1 for d = 1 or 2.
The proof of this theorem under the stated assumptions is to verify the
conditions (Nl)-(N3) in Theorem 3.2 are satisfied for d 5 3 . We will sketch
the proof in steps:
1 j=1
Ilft(u)1125 b l ( l l ~ l l ) l l ~ l l f l (21)
where
m
bl(r) = (~OC~)~(C ~ j - ' ) ~ .
j=1
Step 2) : Similar to Step 1, it can be shown that, for u,v E H I and t E [0,TI,
F b4(11UII1,
TT{[gt(JU)-gt(JV)IR[gt(JU)-gt(J21)1*} 11~111)11~-~11~,(29)
+
with b 4 ( ~s), = Kz[l T ~ ( ~ - ~s ~ +
] . view of Eq. (27) and
) ( ~ - ' )In
Eq. (29), the condition (N2) is valid.
95
“ 2
f ( x ,r, t)dr = - -aj(x, t)uj+l
j=1 3 + 1
Then the next theorem holds true, and its proof can be found in Chow 3.
Theorem 4.2. Suppose that all conditions in Theorem 4.1 are fulfilled. In
addition to ( P l ) and (P2), assume that
+
( l a ) Given m = (2n 1) for a positive integer n, there exist constants
a 2 0 and ,L? 2 0 such that
G ( x ,r , t ) 2 ( a + PrZn)r2
+
(8,”- c 2 v 2 y2)u = xu3 + atMt(Ju) t > 0,
(31)
0) = g,
u(., a,.(., 0 ) = h, x E 723,
96
where c, y and X are real parameters, the functions g, h are given as before,
and M t ( J u ) is assumed to be of the bilinear form:
M t ( J u ) ( z ):= M ( z , J u , t ) = c/
3
j=1 0
t
[aZju(z,s ) ] W j ( zds).
,
+ r j ( x , z )< m.
~ j ( z , z ) ] d z sup
.€R3
j=1
Acknowledgments
This work was supported in part by the NSF Grant DMS-9971608.
References
1. C. Mueller, A n n . Probab. 25, 133 (1997).
2. J.B. Walsh, Lect. Notes in Math., Springer-Verlag, Berlin, Heidelberg, New
York, 1180, 265 (1984).
3. P.L. Chow, A n n . Appl. Probab. 12,361 (2002).
4. H. Kunita, Stochastic Flows and Stochastic Differential Equations, Cambridge
Univ. Press, Cambridge, England, 1990.
5 . K. Yosida, Functional Analysis, Springer-Verlag, New York, 1968.
6. G. Da Prato and J. Zabczyk, Stochastic Equations in Infinite Dimensions,
Cambridge Univ. Press, Cambridge, England, 1992.
7. R.A. Adams, Sobelev Spaces, Academic Press, New York, 1975.
8. M. Reed, Abstract Non-linear Wave Equations, Springer-Verlag,Berlin, 1976.
STOCHASTIC NAVIER-STOKES EQUATIONS: LOEB
SPACE TECHNIQUES & ATTRACTORS
NIGEL J. CUTLAND
Department of Mathematics,
University of Hull
Hull, HU6 7RX, UK
E-mail: n.j. [email protected]
We survey the use of Loeb space methods in stochastic fluid mechanics, with
particular emphasis on recent results concerning the existence of attractors for the
stochastic Navier-Stokes equations.
1. Introduction
A general version of the stochastic Navier-Stokes (sNS) equations in a
bounded domain D c EXd takes form:
+ +
V ) U f ( t , u)- V p ] d t g ( t , u ) d ~ t
du = [VAU- (u,
{ divu = 0 (1)
u : [O, m) x D xR ---f Rd
where R is the domain of an underlying probability space. The initial
condition u(0) = u g is prescribed (and may be random); the boundary
condition is either u(t,x) = 0 for x E 8D or, occasionally, when d = 2 we
assume periodic boundary conditions.
These equations have been the subject of considerable study since they
were first solved in [6], for d 5 4, using Loeb space methods. Some time
after the publication of [6] a number of alternative proofs of existence ap-
peared (see below) so that now there is considerable interest in more delicate
issues such as the existence of a stochastic flow and attractors for the sNS
equations.
Loeb space methods have continued to prove powerful in this field, in
combination with the well-developed techniques of LLc1assica177 infinite di-
mensional stochastic analysis. The purpose of this paper is to survey what
97
98
H and V are Hilbert spaces with scalar products (., .) and ((., .)) respectively,
and 1 . I 5 cII . 11 for some constant c.
By A we denote the self adjoint extension of the projection of -A in
H; A has an orthonormal basis { e k } of eigenfunctions with corresponding
eigenvalues X k r X k > 0 , X k m. For u E H we write U k = ( u , e k ) ,
and write Pr, for the projection of H on the subspace H, spanned by
{ e l , . . . , em}. Since each ek E V then H, C V.
The trilinear form b defined by
d
avi
(whenever the integrals make sense) has the well-known and crucial prop-
erty b(u,w , w) = --b(u,w, w) so that b(u,w,w) = 0.
In this framework, the stochastic Navier-Stokes equations (1) may be
formulated as a stochastic differential equation in H as follows:
+ +
du = l-vA.1~- B ( u ) f ( t ,u ) ] d t g ( t , u)dwt (2)
where B ( u ) = b(u,u,.). This is initially regarded as an equation in V’ (the
dual of V) although it turns out that the solution lives in H (and in fact
in V for almost all times). Compared to ( l ) ,note that the pressure has
disappeared, because V p = 0 in V’ (using divu = 0 in V and an integration
by parts).
The equation (2) is really an integral equation, with the first integral
being the Bochner integral and the second an extension of the It6 integral
t o Hilbert spaces, due t o Ichikawa [25]. The noise is given by a Wiener
process w : [0, m) x R t H with trace class covariance, and so the noise
coefficient g belongs t o L(H, H). It is assumed that
g : [ O , o o ) x V + L(H,H)
while
f : [O,m) x v -+ V’.
(The restriction to V in the domains is sufficient because we will have the
solution in V for almost all times.)
Definition 2.1. Suppose that u g E H and f,g as above are given, together
with a probability space R carrying an H-valued Wiener process w. A weak
solution of the stochastic Navier-Stokes equations is a stochastic process
u : [0,co) x R 4 H such that for 8.8. w
(i) u E L 2 ( 0 ,T ;V) n L"(0, T ;H) n C(0,T ;Hweak)for all T < co ,
(ii) for all t 2 0
rt rt
for all T
3.1. M e a s u r e attractors
This approach is currently applicable only to d = 2 since it is necessary
that the equation (4) has a unique solution. Thus it is assumed that f , g
satisfy an appropriate Lipschitz condition, to ensure that for each initial
condition u E H there is a unique solution u(t) = v(t,u)with u(0)= u (so
w(0,u) = u).A semigroup St is now defined on Ml(H),the set of Bore1
probability measures on H, by putting Stp = pt where
s, d(u)dCLt(u) = s,
IE 29(v(t1U))dP(UZL)
The methods in [12] do not make essential use of Loeb spaces although a t
some points they can be employed to assist the construction.
For a stochastic system such as (4) the idea of a stochastic attractor devel-
oped by Crauel & Flandoli [19] takes into account the fact that a t all times
new noise is introduced into the evolution of each path of any solution t o
(4). A stochastic attractor is defined to be a random set A ( w ) that, a t time
0, attracts trajectories “starting a t -m” (compared to the usual idea of
an attractor being a set “at time m” that attracts trajectories starting at
time 0).
This idea is spelled out below, and involves the introduction of a one
parameter group Bt : R + R of measure preserving maps, which should be
thought of as a shift of the noise to the left by t. In proving the existence of
a stochastic attractor for the system (4) the nonstandard framework makes
it particularly easy to consider -co.
Making this precise, suppose that cp is a stochastic flow of solutions to
(4). That is, cp is a measurable function
cp: [O,co) x H x R +H
such that cp(.,.,w) is continuous for 8.8. w, and for each fixed initial
condition uo the process u ( t , w ) = c p ( t , u 0 , W ) is a solution to (4) with
u(0,w)= uo.
The notion of a semigroup in the usual definition of a deterministic
attractor, along with the notion of an attractor itself, is now replaced by
the following.
Definition 3.1.
(i) The flow cp is a crude cocycle if for each s E R+ there is a full set R,
such that for all w E R,
cp(4 A ( w ) ,w ) = A(Qtw), t L 0,
lim dist(cp(t, B , Q P t w ) ,A ( w ) ) = 0
t+m
for each bounded set B c H.
Note that the existence of a perfect cocycle is necessary for the pos-
sibility of having a stochastic attractor. Constructing a perfect cocycle
is difficult for infinite dimensional systems, particularly for those that are
truly stochastic (as compared to random dynamical systems in which paths
may be treated individually).
The proof of this result is quite long and complicated, and uses heav-
ily the fact that solutions to (4) may be obtained as standard parts of
Galerkin approximations of dimension N , infinite. A delicate extension of
the Kolmogorov continuity theorem as adapted to a nonstandard setting
by Lindstrom [l]is at the heart of the construction of the perfect cocycle.
An outline of the main steps and ideas of the proof is given in Chapter 2
of [16].
that is sufficient for the concepts to make sense. For this an adapted Loeb
space is needed.
A precise formulation of the notion of a process attractor and the main
result of [18]is as follows.
On an arbitrary space R carrying a 1-dimensional Wiener process
(wt)t20 suppose that a class X of solutions to the sNS equations (4)is
defined. Suppose further that R is equipped with a family of measure
preserving maps Ot : R -+ R for t 2 0 with the following properties:
(el) B0 =identity and 8, 0 BS = 8t+s;
for all s , t 2 0, where ( F t ) is the filtration on R;
(82) 8 t 3 s = Ft+s
+
(83) w ( t s, &w) - w ( t , Otw) = W ( S , w) for all s 2 0.
Note that the property (83) tells us that for a fixed t the increments of
the process w ( t + s , 8,w) are the same as those of the process w ( s , w ) . Thus
Ot can be thought of as a shift of the noise to the right by t.
The family (8,) allows the following definition of a semiflow S, of
stochastic processes.
Definition 3.2. (Semiflow of Processes) Suppose that u = u ( t , w ) is
a stochastic process defined for t > 0. Then for any r 2 0 the process
u = S,u is defined by
v(t, w)= u ( r + t ,8,w)
It is clear that S, is a semigroup, and if u is adapted so is Stu.
Suppose now that X is closed under St. Then a process attractor for the
class X can now be defined. In the following, if u is a stochastic process
then Law(u) is defined to be the probability law (on path space) of the
coupled process ( u ,w ) .
Definition 3.3.
(a) A set of laws A c Law(X) is a Law-attractor if
(i) (Invariance) & A = A for all t 2 0, where St is the mapping of
laws induced by the semigroup St.
(ii) ( A t t r a c t i o n ) For any open set 0 2 A and bounded 2 c Law(X),
$2 c0
eventually (i.e. this holds for all t 2 to(O,2)).
(iii) ( C o m p a c t n e s s ) A is compactg
108
limt+,d(StZ, K ) 2 d(A, K )
(iv) A is closedh.
Remarks on Definition 3.3. (1) Since existence results for the stochastic
Navier-Stokes equations require a rather large probability space, it is to be
expected that any space carrying a whole class of solutions X as above
will be too big to allow an attractor A c X that is compact in the usual
sense. However, the attractor A of the following theorem is neo-compact,
the key notion developed in [23] It is a consequence of neo-compactness
that Law(A) is compact.
(2) The attraction property 3.3(b)(iii) is equivalent to the following:
stz 5 0 (5)
eventually for any bounded Z and any open 0 3 A of the form 0 =
L2(R,M)\KsE, with K compact. Property 3.3(b)(i) means that in addition
(5) holds eventually for any open set 0 of the form 0 = LawP1(0’) where
0’ is an open set of laws with Law(A) 2 0’.The usual attraction property
for attractors, namely that StZ C 0 eventually for any bounded 2 and
any open 0 2 A is probably too much to expect. However, the attractor
in the following theorem has property (5) for a smaller class of open sets -
namely those that are neo-open, a further key notion of [23]. Sets 0 of the
form L2(R,M ) \ K s Eor Law-l(O’) as above are neo-open.
where do is the Prohorov metric and pi (i = 1,2) is the projection of X i onto the first
coordinate- that is, path space for the solutions of (4).
lu12 = IEsom
hHere and in (iii) the topology is the L2 norm topology on processes in H given by
lu(t)12exp(-t)dt.
109
00; H) n &(o,
L:~(O,00; H) n L?~,[O, 00; V)n ~ ( 000;, Hweak)
4tl) = u(to)+lo tl
[--vAu(t)-B(u(t))+f(u(t))ldt+
1: 9(u(t))dwt
(X6) IE J i < co
Iu(t)12dt
(ii) Denote by xk the set of u E x with
(X6k) E J i Iu(t)I2dt i k
Remarks 1. The above conditions tell us nothing about u ( t , w ) at t = 0
and there may be a singularity there. In this sense the class X is a class
of generalized weak solutions to the stochastic Navier-Stokes equations (cf.
[28] p.12).
2. It follows from (X6) that IE(Iu(t))I2)< 00 for 8.8. t E ( 0 , l ) . Thus,
from (X3) we see that IE(Iu(t))I2) is bounded on [A,
m) for all n.
3. In condition (X5), the function cpn(u)is an explicit smooth approx-
imation to the function \~1~1{1..12,).The inequalities (X5) follow heuristi-
cally from the equation (4) as a particular instance of the Foias equation
corresponding to (4). The choice of the functions qn makes (X5) a kind
110
Appendix
Here we give a concise but mathematically complete construction of the
Loeb space used in the paper [18] and discussed in the previous section.
This is to take some of the mystery away from the notion of a Loeb space,
and to show that its construction is entirely algebraic. What we are not
able to do here is to expand on the properties of Loeb spaces that make
them so useful. For this see any of the introductions such as [1,11,16].
111
A . l . The hyperreals
To define the extension of the reals known as the hyperreals *R first fix
a nonprincapal ultrafilter U on N. That is, U is a collection of subsets of
N that is closed under intersections and supersets, does not contain any
finite sets, and is maximal with this property. This means that for every
set E C N either E E U or N \ E E U (but not both).
The hyperreals *Rare defined by
*R= RN/U
meaning the quotient of RN by the equivalence relation
It is easy to check that " ( a + b ) = ' a + " b and the same for products.
A . 2 . Construction of a Loeb space
Let W be two-sided Wiener measure on Co(R) = {x : R 4 R;x ( 0 ) = 0).
The set R is defined by
R = CO(R)"/IA
A = r I i e ~ A i / U= [(Ai)]
where Ai C: Co(R). That is, for x = [(xi)]E R we define
xEA @ xi E Ai a.c
0 # Anfl c A,
which means that if A, = [(A,,i)]we have 0 # A,+l,i C A,,i a.c. For
n = 1 , 2 , 3 , .. . in turn, systematically modify A,,i on a smallm set of indices
i , so that for each n
0 # A,+I,, 5 A,,% for all i E N
This does not alter the sets A, themselves. Now pick xi E Ai,i for each i
and note that zi E A,,i for n 5 i. Consider the element x = [ ( x i ) ]Then
.
xE nnEN
A, because for each n
{i : xi E A,,i} 2 {i : i 2 n } E ZA.
Thus A # 0 , the required contradiction.
(b) Carathkodory's extension theorem shows that the finitely additive
probability POon the algebra 4 extends uniquely to a a-additive probability
on a(G) provided that whenever n n E N A n= 0 for a decreasing sequence
of sets from G then Po(A,) 4 0 with n. In our case this follows trivially
from (a).
The above construction gives a probability space
( 0 , 4 4 ) P, )
kFor example, [ ( A i )U ] [ ( B i )=
] [(Ai U B i ) ] .
' I n fact for disjoint A , B we have Po(A U B ) = Po([(Ai)] ) Po([(AiU Bi)])
U [ ( B i ) ]= =
+
O[(W(AiU B i ) ) ]= " [ ( W ( A i ) W ( B i ) )= +
] O [ ( W ( A i ) ) ]O[W((Bi))] +
= Po(A) Po(B).
mThat is, a set not in the ultrafilter U .
113
The Loeb space resulting from the above construction is now the completion
of this space with respect t o the measure P (that is, adding in the P-null
sets) which is still denoted P , giving the space (0,F,P ) say.
The a-algebra F is a Loeb algebra, and t o indicate its origin i t is often
denoted F = L(G). Similarly we often write P = Q L , the Loeb measure
constructed from Q, where Q is the *R-valued function defined on 4 by
Q(A)= [(W(Ai))], so that Po = " Q .
The key to the use of Loeb spaces hinges on two main facts. T h e first
is due t o Loeb [26] and shows t h a t L(G) = G modulo null sets: for any
B E L ( 4 ) there is A E 6 with P ( B A A )= 0.
T h e second is t h a t the sets in 4 and their measures inherit (in a way
made precise by the Transfer Principle) the properties of the measurable
subsets of Co(R) and their Wiener measure. This makes the algebra 6
tractable, as expounded in any of the references cited above.
References
[l].S.Albeverio, J.-E.Fenstad, R.H@egh-Krohn,and T.Lindstr@m,Nonstandard
Methods in Stochastic Analysis and Mathematical Physics, Academic Press,
New York 1986.
[2]. Z.Brzeiniak, M.Capinski, and F.Flandoli, Stochastic Navier-Stokes equa-
tions with multiplicative noise, Stochastic Analysis and Applications 105
(1992), 523-532.
[3]. A.Bensoussan and R.Temam, Equations stochastiques du type Navier-
Stokes, J. Functional Analysis 13 (1973), 195-222.
[4]. A. Bensoussan, A model of stochastic differential equation in Hilbert space
applicable to Navier-Stokes equation in dimension 2, in: Stochastic Analy-
sis, Liber Amicorum for Moshe Zakai, eds. E.Mayer-Wolf, E.Merzbach, and
A.Schwartz, Academic Press 1991, pp.51-73.
[5]. A. Bensoussan, Stochastic Navier-Stokes equations, Acta Applicanda Math-
ematicae 38 (1995), 267-304.
[6]. M.Capiriski and N.J.Cutland, Stochastic Navier-Stokes equations, Acta Ap-
plicanda Mathematicae 25 (1991), 59-85.
[7]. M.Capifiski and N.J.Cutland, A simple proof of existence of weak and sta-
tistical solutions of Navier-Stokes equations, Proceedings of the Royal Society,
London, Ser.A, 436 (1992), 1-11.
181. M.Capiriski and N.J.Cutland, Navier-Stokes equations with multiplicative
noise, Nonlinearity 6 (1993), 71-77.
[9]. M.Capifiski and N.J.Cutland, Foias and Hopf statistical solutions of Navier-
Stokes equations, Stochastics and Stochastic Reports 5 2 ( 1995), 193-205.
[lo]. M.Capihski and N.J.Cutland, Statistical solutions of stochastic Navier-
Stokes equations, Indiana University Mathematics Journal 43(1994), 927-
940.
[ll]. M.Capiriski & N.J.Cutland, Nonstandard Methods for Stochastic Fluid Me-
chanics, World Scientific, Singapore, London, 1995.
114
[la]. M. Capiriski & N.J. Cutland, Measure attractors for stochastic Navier-
Stokes equations, Electronic J. Prob. 3(1998), Paper 8, 1-15.
[13]. M. Capinski & N.J. Cutland, Stochastic Euler equations on the torus, The
Annals of Applied Probability, 9(1999), 688-705.
[14]. M. Capiriski & N.J. Cutland, Existence of global stochastic flow and
atractors for Navier-Stokes equations, Probability Theory & Related Fields
115(1999), 121-151.
[15]. M.Capinski and D.Gqtarek, Stochastic equations in Hilbert space with ap-
plication to Navier-Stokes equation in any dimension, Journal of Functional
Analysis 126(1994), 26-35.
[16]. N.J. Cutland, Loeb Measures in Practice - Recent Advances, Springer Lec-
ture Notes in Mathematics 1751(2000), Springer, Berlin.
[17]. N.J. Cutland & H.J. Keisler, Neocompact sets and stochastic Navier-Stokes
equations, in Stochastic Partial Differential Equations, (Ed. A. Etheridge),
LMS Lecture Notes Series 216, CUP, 1995, 31-54.
[18]. N.J. Cutland & H.J. Keisler, Global attractors for 3-dimensional stochastic
Navier-Stokes equations, in preparation.
[19]. H.Craue1 and F.Flandoli, Attractors for random dynamical systems, Prob-
ability Theory and Related Fields 100 (1994), 365-393.
[20]. F. Flandoli and B. Schmalfui.3, Random attractors for the 3D stochastic
Navier-Stokes equation with multiplicative white noise, Stochastic & Stochas-
tics Reports, 59(1996), 21-45.
[21]. F. Flandoli and B. Schmalfui.3, Weak solutions and attractors for three-
dimensioanl Navier-Stokes equations with nonregular force, J . Dynamics and
Differential Equations, 11(1999), 355-398.
[22]. S. Fajardo and H.J. Keisler, Existence theorems in probability theory, Ad-
vances in Mathematics, 120(1996), 191-257.
[23]. S. Fajardo and H. J.Keisler, Neometric spaces, Advances in Mathematics,
118(1996), 134-175
[24]. F. Flandoli and D. Gatqrek, Martingale and stationary solutions for stochas-
tic Navier-Stokes equations, Probability Theory 63 Related Fields 102(1995),
367-391.
[25]. AJchikawa, Stability of semilinear stochastic evolution equations, Journal
of Mathematical Analysis and Applications 90 (1982), 12-44.
[26]. P.A.Loeb, Conversion from nonstandard to standard measure spaces and
applications in probability theory, Trans. Amer. Math. SOC.211(1975), 113-
122.
[27]. B. Schmallfui.3, Measure attractors of the stochastic Navier-Stokes equation,
Bremen Report No.258, 1991.
128). G.R. Sell, Global attractors for the three-dimensional Navier-Stokes equa-
tions, J. Dynamics and Diff. Equations 8 ( 1996), 1-33.
[29]. M.Viot, Solutions Faibles d 'Equations aux Derivees Partielles non Lin-
eaires, Thesis, Universite Paris VI (1976).
[30]. M.I.Vishik and A.V.Fursikov, Mathematical Problems of Statistical Hy-
dromechanics, Kluwer, Dordrecht - London 1988.
THE 2D-NAVIER-STOKES EQUATIONS PERTURBED BY A
DELTA CORRELATED NOISE.
A. DEBUSSCHE
ENS d e Cachan, a n t e n n e d e Bretayne
Campus de K e r L a n n
351 70 Bruz cedex
France
E-mail: arnaud. debussche@bretayne. ens-cachan.fr
1. Introduction
We consider the two dimensional incompressible Navies-Stokes equations
in a periodic domain driven by a space time white noise:
-
du = (vAu - (u. V ) u- V p ) d t + dW, in [0,TI x 0 ,
{ div u = 0, in [0,T ]x 0 ,
4 0 , t) = uo(t), 5' E 0,
u is periodic with period 27r,
where 0 = [0,27rI2. The unknown are random processes: the velocity field
(1)
u(t,6) = ( u l ( t ,<), uz(t,6)) and the pressure field p ( t , 6); these are defined
for (5'1, (2) E 0 and t 2 0. The kinematic viscosity v has no importance in
this work and we will take it equal to 1.
The equations are forced by a space time white noise. It is delta corre-
lated in time and in space, i.e. we formally have
and correlated in space. Recently much progress has been obtained in the
study of the associated invariant measures for noises which are very smooth
in space. Uniqueness and ergodicity properties have been proved 4117,20,21.
Also in 18, the singularities of the solutions in the three dimensional case
are studied.
In the work 16, a space-time white noise is also considered and equation
(1) has been studied through the associated Kolmogorov equation. They
prove directly the existence of a solution to this latter equation but are
unable t o connect it t o the original equation. The main difficulty is that,
as is well known, with such a rough noise it is expected that a solution of
(1) is not regular.
In this work, we first observe that, using ideas borrowed from the theory
of the Wick product the nonlinear term can be defined for random
11,22124,
2. Notations
We introduce standard notations used for the Navier-Stokes equations (see
for instance 2 5 ) . The subspace of ( L 2 ( 0 ) ) consisting
2 of periodic divergence
117
' k ik.E
k . E = klEl + k2E2, ek(E) = - -e
27T IN
E 7 = ( E l , E2) E 0.
AX =PAX, z E D ( A ) = ( H ; ( O ) ) n~ H .
We have
Aek = -Ikl2ek, k E Z,.2
Here and in the following H&(O) is the subspace of the Sobolev space
H T ( 0 )consisting of all periodic functions. For T E R, we use the fractional
power ( - A ) T on the domain D ( ( - A ) T ) .It is classical that D ( ( - A ) T ) is the
closure in ( H 2 ' ( 0 ) ) 2of the space spanned by (ek)&Z,2. Moreover .I
is a norm on D ( ( - A ) T ) equivalent to the usual norm on (H2'(0))2. For
any T E R, P can be defined on ( H z ( 0 ) ) 2and its image is D ( ( - A ) T ) .
We set
W =PW. (2)
It is not difficult to see that W is a cylindrical Wiener process on H thus,
for any complete orthonormal system (ek)kcz; in H , we can write
W= Pkek
k€Z;
118
We will use this form which is better suited to the case of non smooth
velocities. Whenever it makes sense, we set
b(z,y) =P div (z 18
y), b(z)= b(z,z). (3)
When projecting equations (1) on HI we get
+
du = (Au b(u))dt dW, +
{ u(0)= uo.
We wish to solve (4) and t o find a solution which is a D((-A)') valued
(4)
process. Implicitly this means that we restrict our attention t o zero average
initial data. This is no loss of generality since we can change the unknown
in (1) and consider only such initial data.
3. Preliminaries
It is not expected that (4)has a solution in D ( ( - A ) T )for T 2 0. This is
not even true for the linear equation
dz = Azdt + dW,
(5)
4 0 ) = 201
whose solution is given by
~ ( t=)etAzo + 1 t
e(t-s)AdW(s).
The second hrn in the right hand side is a continuous process with values
in D ( ( - A ) T )for any T < 0 but does not take its values in D((-A)') for
any T 2 0. This follows from
119
for r 2 0, see 12. We have denoted by L H S(K1, K2) the space of all Hilbert-
Schmidt operators from a Hilbert space K1 on a Hilbert space K2.
It follows that we have to work with non smooth processes and this
creates difficulties when working with the nonlinear term. Here we pro-
ceed as is usual when dealing with parabolic equations in negative Sobolev
spaces. We use Littlewood-Paley decomposition and paraproduct to define
the nonlinear terms, see 718123.
For cr E R , p 2 1, p 2 1 we define
Iulr3;,p = (~2~q~16qul;.;.))
The following result is crucial in our argument and is the main motivation
for working in Besov spaces, see 7,8.
+
Proposition 3.1. Let p , p 2 1, cy p > 0 , a < 2 / p , p < 2 / p . Then zf
we have uv E Bz,P where y = a P - p,
u E B& and v E i3t,,p 2
and +
14B;,p i C14Bpqpl~lBg,p. (6)
Let us also recall that the nonlinear term verifies the following identities,
see 1i26
x )0, ( b ( z ) ,A X )= 0.
(~(x),= (7)
120
These are true for any x such that the quantities on the left hand side make
sense.
Let us denote by p the product measure on 'H,
P= I-I"0,
kEZ:
1/(2142)).
X N = PNX = C (x,el)et, 1 2
XN = (zN,zN),
IlllN
and
b N ( x )= b ( P N z ) .
We also define for x E If, and N E N:
where
and
121
dz = Azdt +dW(t)
with invariant law C ( z ( t ) ) = p.
Proof : It is not difficult to prove that for any S < 0 the trajectories of
(-A)&. are continuous with respect to t , II: on R x 0. This uses for instance
the Kolmogorov criterion of continuity, see 12.
Since for a n y p 2 1, (C(a)>'c 13:,m, it follows that z has trajectories in
C(R; a;,,) for any 0 < 0. Moreover, it is easy to see that z has trajectories
in C([O,T];D((-A)"/')) = C([O,T];Bz,,), < 0, so that by interpolation
BE,,),0 < 0,p 2 p 2 2, P - a s . .
z E C([O,T]; 0
By Corollary 3.1, b(a(t)) can be defined for each t E R so that we have
a well defined process (b(z(t))tEw.
Proposition 3.3. Let X and 2 be random variables such that tke law of
Z is p and Y = X - Z E Lb(R;a),: where
2
b>2, ->a>O,
P
then the sequence of random variables ( b N ( X ) ) N E N converges in
; for any a < a - 1 - ;.2
L ~ / ~ (B:R,,)
If moreover, the law of X , ux, is absolutely continuous with respect to
p and % E L'(7-i;p ) with 1 > 1 then the limit coincides with b ( X ) defined
in Corollary 3.1 and
+ +
b ( X ) = b(Y) 2 b ( X , Y ) b ( 2 ) . (9)
Proof : Let a < a - 1 - 2 and set (T = a - a - 1 -
P
5.
Clearly, 0 < 0
and 2 E Lb(R;a,",,).Thanks t o Proposition 3.1, b(Y,2 ) is well defined in
Lb12((R;a:,,) and
, in L ~ / ~ ( (B;,,).
Z ) -+ b ( ~Y),
bN(y, R;
123
Remark 4.1. By Remark 3.1, the solution has the required properties to
ensure that b(u) is well defined. Indeed, u ( t )- z ( t ) E B;,, P a s . w E 0 for
for almost every t E [O, TI.
Remark 4.2. Note the condition - $ < 5 implies that p > 2. This is the
reason for working in Besov spaces. Recall that Sobolev spaces correspond
to Besov spaces with p = 2.
Proof : We split the proof in two parts. We first prove local existence
on a random time interval depending on the initial data. Then an a priori
estimate enables us to get global solutions.
First step : Let uo E B;,,. We fix w and solve (4) pathwise, w is taken
in a set of probability one such that the various properties on z and b ( z )
proved above are true.
124
we deduce that
since, by assumption
Similarly, we have
since, by assumption
S9imilarly, we have
Furthermore,
125
1
1 eA(t-s)b(z(s))dslETo 5 C(Q1 PI T)lb(z)Ilk(o,T;a~,,')'
thus the local in time construction of the first step can be iterated leading
to a global solution. Thus Theorem 4.1 is proved.
Let us prove that (11) holds. We use a formal argument which can
be easily justified by a Galerkin approximation. Indeed, it is not difficult
to prove that the local solution constructed above is the limit of Galerkin
solutions.
We have
therefore
u(tl uo)= e"(u0 - z ( 0 ) ) + 1" e(t-S)Ab(u(s,
u0))ds +z(t),
Then, integrating with respect to p, using again Holder inequality and the
invariance of p, we obtain
1/3
+C(P, p, a)W2(Jx/b(uo)l;;,;l)~P(~O)) .
By Lemma 3.2 and Proposition 3.2, we know that the right hand side is
finite. This proves our claim (11).
The last statement (10) is proved in the same way.
5 . Ergodicity
As already mentionned, using a galerkin approximation, it is not difficult
to prove that the Gaussian measure p is invariant for (4). We now study
some of its properties.
Given a functional cp defined on 'Ti, we denote by (p its average with
respect to p
r
I t is well known that the Gaussian measure satisfies the spectral gap in-
equality
References
1. Albeverio S., Cruzeiro A. B. (1990) Global flows with invariant (Gibbs) mea-
sures for Euler and Navier-Stokes two dimensional fluids, Commun. Math.
Phys. 129,431-444.
2. Bensoussan A., Temam R. (1973) kquations stochastiques du type Navier-
Stokes, J. Funct. Anal., 13,195-222.
3. Bourgain J. (1999) Nonlinear Schrodinger equations, in Hyperbolic equations
and frequency interactions, Providence, RI, edited by Caffarelli et a]., AMS.
Park City Math. Ser. 5,3-157.
4. Bricmont J., Kupiainen A., Lefevere R. (2000) Exponential mixing f o r the 2D
Nauier-Stokes dynamics, Preprint.
128
S. ALBEVERIO
Institut f u r Ang. Mathematik, Universitat Bonn, Wegelerstr. 6, 0 - 5 3 1 15 Bonn;
S F B 61 1, Bonn; BiBoS, Bielefeld; CERFIM, Locarno;
Accademia d i Architettura, USI, CH-6850 Mendrisio;
Dipartimento d i Matematica, Universitci di Trento, I-38050 Povo;
E-mail: [email protected]
B. FERRARIO
Institut fur Ang. Mathematik, Universitat Bonn, Wegelerstr. 6, 0 - 5 3 115 Bonn;
Dipartimento di Matematica, Universitci d i Pavia, via Ferrata 1, I-271 00 Pavia;
E-mail: [email protected];[email protected]
1. Introduction
Let us begin considering the classical motion of an ideal incompressible
fluid, that is the Euler equations
130
131
and applying the V' = (-&,&-) operator to the first equation in (l),
we obtain
{ ~ + u . V w = 0
(t1x) E ( O I T I x JD (2)
w=vL.u
with the tangential boundary conditions for the velocity. In (2) there is no
pressure, but this has t o be recovered from the velocity field u: applying V
to the first equation (1) we get -Ap = V . [(u. V)u].
Since w evolves according to a transport equation, the solution is w ( t , x) =
w ( 0 , E t x ) where Et is the flow of material points in the fluid ( & x ( t ) =
v ( t , x ( t ) ) ) .Since the vector field w is divergence free, any solution to sys-
tem (2) corresponds to a volume preserving flow Et (i.e. the Lebesgue
measure on ID is preserved in time). Moreover there are two other con-
served quantities
energy ,
E = L2 J I +)I2da:
enstrophy S= J , w(z)'dz
This has t o be understood as follows: if an Euler flow (2) with finite energy
is defined, then the energy is indeed constant. The same holds for the
enstrophy in a two dimensional spatial domain. The computations showing
this invariance in time are easily checked in these cases, e.g.
dS
- --
dt
J, w ( t ,). a t 4 t 1). dx
=- J, ~ ( X)
t ,~ ( 2)t , . V w ( t ,X) dx
= J
, V U ( X)~ ,. U ( t , X) w ( t , X) dx + J
, ~ ( x)V
t , . u ( t ,X) w ( t , X) dx
Since V . w(t, x) = 0, then 2
= 0. Notice that all the quantities have been
assumed t o be well defined, i.e. the solution w is regular enough.
By means of these conserved quantities, heuristic expressions of invariant
measures can be given. In the next section, we will deal with probability
measures m of Lkvy-Khinchine type. They are supported on distribution
spaces. Therefore the Euler dynamics with initial data in the support of
the measure rn (if it exists) is not a classical one. An overview on the
study of a deterministic dynamics having m as invariant measure will be
presented in section 3. According to the Koopman-von Neumann theory,
as soon as a (candidate) invariant measure m is known, any flow S t ,t E R,
in the space of distributions S' a gives rise to a flow in C2(rn),represented
aBy S' we denote the vector space of continuous linear functionals on C,"(ID) or, when
the spatial domain is the torus, CpMer(T).
132
Assume now that the random variables are identically distributed and
functional
p('Hb(D))= 1 V b < -1
that is, the support of the Gaussian measure p is given by nb<-lRb(IO).
Similarly when D is the torus.
Compound Poisson measure
The support of the compound Poisson measure IT is the space r of config-
urations. More precisely, for any n = 1 , 2 , . . . l let
= {((m
;i(", Z l ) , . . ., (vn,4)E. (Rlx D ) n : 51 # 21,for 1 # I C )
The space of n point configurations is defined as
n
r(n) = { w = C qszl : vl E R O , z1 E D~z1 + Xk for 1 # I C )
1=1
where S(") denotes the permutation group over (1,.. . ,n). Consider on the
Bore1 a-algebra of subsets of i ( n ) / S ( n the
) measure o @ =
' ~ (dO(v)dz)@'n,
where for simplicity we assume the LBvy measure 6 to be finite. The image
134
r = u=;W
,
is defined as disjoint union of topological spaces, with the corresponding
Bore1 a-algebra B ( r ) .
The compound Poisson measure II is defined by
Remark. Notice that the above measures are not the only invariant measures
known for the Euler equation (2). For instance, Albeverio et al.'l5 consid-
ered more general Gaussian white noises p ~(y ,> 0, ~ p > -y), expressed by
means of the enstrophy and of the renormalized energy. For other types of
invariant measures related to the Gaussian ones, see Capiriski&Cutland",
Ciprianoll. Anyway we consider here only white noise distributions for
the vorticity w , in order to have a unified approach ( p and II as particular
cases of a Lkvy-Khinchine measure). Infinitesimal invariance of measures of
Gaussian and Poisson type for the Euler equation has also been discussed
in Boldrighini&F'rigiog. 0
3. Deterministic dynamics
From now on, we choose the spatial domain D to be the torus T = [0,27rI2;
hence periodic boundary conditions are assumed. In this section this choice
is done for mathematical convenience; in the next one it will anyway appear
necessary for a right physical interpretation.
Let w be a periodic distribution; it can be developed in Fourier series
with respect to the complete orthonormal basis { & e i k ' z } k E Z z in the (com-
plex) L2(T). Let denote for short by P k the k-th element in this basis. Then
1
W = 2;; x k E Z 2 w k p k with w k := s ' ( W , p - k ) S . The Coefficients w k E and
w k = W - k , because w is real. Adding a constant to the velocity, solving
(l), yields again a solution of (1). We select that one of zero mean value.
Hence also the mean value of the vorticity is assumed to vanish: wo = O.b
bThe starting problem indeed is formulated in t h e real framework. Now the complex
structure arises in a somewhat artificial but practical way - via Fourier transform. Ac-
tually t h e relevant variables are {%~k,Swk}~.~~,~>~ or { w k } k E Z ~ , k > where
O, k > 0
means either !q > 0 or kl = 0 , Icz > 0. Anyway, whenever t he whole sequence { ~ k } ~ ~ ~
appears, t h e condition wl, = w-k is assumed.
135
dwk(t) -
-- ChkWh(t)wk-h(t)
dt h#k,h#O (4)
(The equivalence of infinitesimal and full invariance is due to the fact that
1E D(B).)
Finally, let us notice that even if any Gaussian measure can be approx-
imated by a sequence of Poisson measures, the II-a.s. well posed dynamics
(in S') is not helpful to define in the limit a p a s . dynamics. Indeed,
p and II are singular measures: supp II c supp p and II(F) = 1,
p ( r ) = 0 (see the proof by Colella and Lanford, in the modified version
in Albeverio&Ferrario4).
4. Stochastic dynamics
because -2-
auk = l2( a pi&), for k > o .
This form is closable, as easily seen by integration by parts, rewrit-
ing & as the positive symmetric sesquilinear form associated with a
densely defined positive symmetric operator (see, e.g., Albeverio&Rockner7,
Ma&Rockner" for this technique). The closure is a classical Dirichlet
form, quasi-regular and local; moreover, the minimal and the maximal ex-
tension coincide (we refer, e.g., t o Ma&Rockner" for results of this type
in the Gaussian case). Its associated classical Dirichlet operator is the
139
in which b / 2 represents the viscosity of the fluid ( b > 0). This corresponds
to the following equation for the velocity vector fields
b
dv(t,x) = ,A v(tl x)dt + V p ( t ,x)dt
v . v ( t , 2 )= 0
Therefore, the noise is defined by means of a Brownian motion, cylindrical
in Lz(T) for the velocity. Interpretation of (10) as an equation of motion of
a viscous fluid is possible only in the frame of periodic boundary conditions.
Indeed, the boundary conditions in a bounded domain ID are v . n 1 a D = 0
for an ideal fluid (viscosity b = 0) and ~ 1 =8 0 ~for a viscous fluid ( b >
0). Hence the torus is the only case in which the boundary conditions
for the two different fluids coincide, and therefore the functional spaces
introduced for the Euler problem fit also for the Stokes problem (and in
the next section for the Navier-Stokes problem). It is worth a t this point
to say that all the results of section 3 require the spatial domain ID to be
bounded. (For a formulation of the Euler problem as an infinite system of
nonlinear equations (4) when ID is a bounded domain in IR2 with piecewise
C1 boundary 8D1see Albeverio&H@egh-Krohn5.)
Compound Poisson measure
Similarly as before, we introduce the classical pre-Dirichlet form given by
140
X€W(O)
5 . Final remarks
Given the Liouville operator B and the diffusion operator Q, it is possible
to merge them together in the following sense. Since both the operators B
and Q are well defined on the dense subset FCF of C 2 ( m ) ,we can consider
the sum operator (Kolmogorov operator)
K =Q + iB, D ( K )= 3 C r (12)
which corresponds to a non-symmetric sesquilinear form. The operator
(Q, FC?) is negative definite, the operator ( B ,F C T ) is skew-symmetric;
141
J K f d m =0 Vf E F C ~
since it is so separately for Q and for B. Our analysis is based on the follow-
ing: if the closure generates a strongly continuous Markov semigroup in
C z ( m ) , then there would exist a unique Markov process (a diffusion) solving
the associated stochastic nonlinear equation. This property is called C2(rn)-
uniqueness (or strong uniqueness) of the Kolmogorov operator ( K ,FCF).
For the Gaussian case, this Markov process would be the (unique) weak
solution to the stochastic Navier-Stokes equation
b
d W k ( t ) = [ - -Ik12wk(t)
2
+ B k ( ~ ( t )d)t]+ -
lkl
Jz d P k ( t ) ,
k E Z2, k > 0 (13)
then this same technique would give C1-uniqueness for the Kolmogorov
operator K defined in (12). Unfortunately, the best estimates are (see
Albeverio&Ferrario4)
v j d z j ( t ) = V$ 2
l#j,l,j=l
v j . l g ( z j ( t ) - z l ( t ) ) c t t + d ~ ~ " ' ( t )j, = 1 , .. . , n(14)
We write the Kolmogorov operator with respect to the scalar product given by the
symmetric part as
Then, in our setting the LP-uniqueness result of Eberle15 (Th. 5 . 2 ) holds true if the
components asatisfy the integrability condition
Ikl
for 1 5 p < 2.
143
Acknowledgments
We would like t o thank the organizers of the Conference on Probabilistic
Methods in Fluids for the interesting meeting and for arranging a very pleas-
ant stay in Swansea. The second author gratefully acknowledges financial
support from the Alexander von Humboldt Stiftung.
References
1. S. Albeverio and A.B. Cruzeiro, Comm. Math. Phys. 129,431 (1990).
2. S. Albeverio, M. Ribeiro de Faria and R. H0egh-Krohn, J . Statist. Phys. 20
No. 6, 585 (1979).
3. S. Albeverio and B. Ferrario, J . Funct. Anal. 193 No. 1, 77 (2002).
4. S. Albeverio and B. Ferrario, Infin. Dimens. Anal. Quantum Probab. Relat.
Top. (2002) in press.
5 . S. Albeverio and R. H0egh-Krohn, Stochastic Process. Appl. 31,1 (1989).
6. S. Albeverio, Yu.G. Kondratiev and M. Rockner, J . Funct. Anal. 154,444
(1998) and 157,242 (1998).
7. S. Albeverio and M. Rockner, J . Funct. Anal. 88,395 (1990).
8. S. Albeverio, M. Rockner and T.S. Zhang, Markov uniqueness for a class
of infinite dimensional Dirichlet operators, in Stochastic processes and
optimal control Stochastics Monogr. 7 (eds. H.J. Engelbert, I. Karatzas
and M. Rockner) Gordon and Breach, Montreux, pp. 1-26 (1993).
9. C. Boldrighini and S. F'rigio, Comm. Math. Phys. 72 , 55 (1980); Errata:
ibid. 78,303 (1980).
10. M. Capiriski and N.J. Cutland, Nonstandard methods for stochastic fluid
mechanics, World Scientific Series on Advances in Mathematics for Applied
Sciences, Vol. 27 (1995).
11. F. Cipriano, Comm. Math. Phys. 201,139 (1999).
12. A.B. Cruzeiro, Expo. Math. 7, 73 (1989).
13. G. Da Prato and A. Debussche, J. Funct. Anal. (2002). To appear.
14. D. Diirr and M. Pulvirenti, Comm. Math. Phys. 85,265 (1982).
15. A. Eberle, Uniqueness and non-uniqueness of semigroups generated b y sin-
gular diffusion operators LNM 1718,Springer, Berlin (1999).
16. I.M. Gel'fand and N.Ya. Vilenkin, Generalized Functions Vol. 4, Academic
Press (1964).
17. R. Goodrich, K. Gustafson and B. Misra, Physica 102A,379 (1980).
18. Z.M. Ma and M. Rockner, Introduction to the theory of (non-symmetric)
Dirichlet forms, Springer, Berlin (1992).
19. C. Marchioro and M. Pulvirenti, Vortex methods in two-dimensional fluid
mechanics, L N P 203,Springer (1984).
20. W. Stannat, Ann. Scuola Norm. Sup. Pisa C1. Sci. (4) 28,99 (1999).
21. W. Stannat, Preprint Bielefeld (2002).
SOME REMARKS ON A STATISTICAL THEORY OF
TURBULENT FLOWS
FRANC0 FLANDOLI
Dipartimento d i Matematica Applicata, Universitci d i Pisa
Via Bonanno 25b, 56126 Pisa E-mail: [email protected]
Some recent notions and results, like invariant memures for the Navier-Stokes
equations, random attractors, random invariant measures and vortex filaments are
reviewed. Some conjectures about their relation are expressed.
1. Introduction
The statistical theory of turbulent fluids contains a number of scaling laws
derived on the basis of phenomenological arguments and experimental re-
sults, like the Kolmogorov K41 scaling law for the energy spectrum which
asserts that E ( k ) behaves as k - 3 for wave numbers in the inertial range
(between the integral scale and the dissipation scale); here E ( k ) is the
mean value of J',,! 1C(k)l2dk,where S ( k ) is the sphere of wave numbers k
of modulus k and u(k) is the Fourier transform of the velocity of the fluid.
Moreover, in some cases the experiments and certain pieces of the theory
have some discrepancies, like the scaling of the pmoments of velocity in-
crements, $ ~ ~ (=r (lu(z
) + r ) - u(z)Ip), that are not correctly described by
Kolmogorov theory and seem to require proper intermittency corrections.
The Kolmogorov theory would predict for the structure function 4 p ( r )a
scaling of the form r c ( p ) with < ( p ) = f (for small r in a suitable range), but
the experiments clearly show different exponents q ( p ) for p > 2. A number
of models have been proposed to recover exponents close t o the experimen-
tal ones but a final model is not known. See the review of F'rischZ0for an
extensive discussion of these topics.
The most rational approach to the analysis of fluids is by means of
the Navier-Stokes equations, but the previous facts and theories on the
statistical properties of fluids have not been explained on such a basis.
Of course a number of attempts to fill the gap between the Navier-Stokes
equations and the statistical theory of turbulence have been performed, but
the present understanding of this subject is very incomplete.
144
145
In the last ten years there has been some interest in the concept of
statistics of vortex filaments. It is quite clear from numerical simulations
and experiments that the vorticity field of a turbulent fluid presents some
degree of geometrical organization and the concept of coherent structure
has been introduced. Particularly interesting seem to be structures having
the shape of filaments] therefore called vortex filaments. The importance of
these structures for the statistics of turbulent fluids is not clarified yet, but
the question whether a relation exists between them must be considered.
In addition, the 3D geometric concreteness of these objects with respect
to the vague concepts of eddies (K41 theory and many others) or fractal
sets of singularities (multifractal models) and others, usually advocated in
phenomenological studies of turbulence, may open the door to a more rig-
orous connection with the Navier-Stokes equations. In other words, there is
some hope that vortex filaments (and maybe other structures not identified
yet) may constitute the bridge between the Navier-Stokes equations and
the phenomenological laws of turbulence.
Whether the typical scalings of turbulence can be derived from statis-
tical models of vortex filaments is still an open problem, with some pre-
liminary indications in the works of Chorin4, and some work in progress.
See also She et a1 26 and Boyer et a1 We devote this note t o the other
question] namely the possible connection between the Navier-Stokes equa-
tions and the ensembles of vortex filaments. We describe just a few rigorous
results that could build up such a bridge with the addition of.many other
still unclear ingredients.
In a sense, we meet in turbulence the same situation as in statistical
mechanics] as described by R. Feynman. The theory of statistical mechanics
is like a mountain: the ascent is the path from the Hamiltonian dynamics of
particles (or other miscroscopic models) to the Gibbs measures, the descent
goes from Gibbs measures to macroscopic predictions and laws. In fluid
mechanics we see the ascent from the Navier-Stokes equations to statistical
ensembles of vortex structures (filaments or others) and the descent from
the latter to the laws of turbulence.
This note is restricted to a few fragments of a possible path of the ascent.
The main tools will be SPDEs, random dynamical systems and stochastic
analysis.
2. Vortex filaments
We first describe the concept of vortex filaments following Flandoli et a1
12,15,161 which is a generalization to continuous processes of the ideas of
Chorin4. In the next sections we re-start from the Navier-Stokes equations]
146
I (cp) = Jlo 1
(cp ( X t ),d X t ) I s (9)= Jlo 1
('p ( X t ) I OdXt)
Definition 2.2. We say that the random 1-current cp H S (cp) has a path-
wise realization if there exists a measurable mapping
w I-+ S(w)
from (0,A, P ) to the space V1 of deterministic currents (endowed with the
natural topology of distributions) , such that
[ S (p)](w)= [S(w)]
(cp) for P-a.e. w E R. (3)
for every p E V'.
A general theorem of Minlos in nuclear spaces implies that the usual It6
and Stratonovich integrals have a pathwise realization. A direct spectral
argument provides (presumably optimal) Sobolev regularity properties of
the pathwise realization, see Flandoli et a1 14. We state here only the result
for the 3D Brownian motion, to minimize the digression.
defined over all cp E D1, with values in Lo ( 0 ) .With the same arguments
that yield the previous theorem we have:
Theorem 2.2. Assume that the measure p has finite energy, in the follow-
ing sense:
Then the random current cp H 5 (cp) just defined has a pathwise realization
< ( w ) , with
< E L2 (0,H-1 (R3,R3)) .
<
Remark 2.2. It is possible to show that the H-'-norm of is given (up
to a multiplicative constant) by the following double stochastic integral in
the Stratonovich sense:
case, and several works in the stochastic case, among which we just quote
* and references therein, Ferrariog, Flandoli et a1 17,
Kuksin et a1 21 and
subsequent works, Vishik et a1 27.
We complete this section with the precise statement of a rigorous result
on point 3 above.
In a sufficiently regular domain D c R3, consider the SPDE of Navier-
Stokes type
H =
{Q : D 3
3
R31Q E [L2 ( D ) ] , divQ = 0 , Q .nlaD = 0
1
where n is the outer normal to i3D
f o r all T 2 0. If in addition
2
IIG ( m 2 ( H ) 5 A0 1.;lI +c
f o r all x E V and f o r a suficiently small A0 2 0 , t h e n there exists a
stationary process u with the previous properties.
p = law of u ( t )
(t)[lt]
An inequality of the form E S U ~ ~ ~IIu[ ~ , ~ <I 00 would imply the well
[
posedness of the Navier-Stokes equations, but we do not have such a strong
estimate. Anyway, the weaker estimate (22) is sufficient to prove interesting
improvements of the theory of singularities, see Flandoli & Romito".
bounds on the Hausdorff dimension of the attractor that show that the di-
mension diverges with the Reynolds number, see Liu22). Therefore we do
not expect p to have a simple form, like a Gibbs measure. The paradigm
arising from the RSB theory, however, open the door to a quantitative anal-
ysis, even if very difficult. The picture that emerges in the RSB theory is
that a typical trajectory on the attractor crosses continuously local unsta-
ble manifolds Wp (such unstable manifolds are sets of points close to each
other and approaching each other exponentially in the reversed motion).
The measure p conditioned to W z is a Gibbs measure with energy propor-
tional to a certain logarithmic determinant of the flow, so a quantity that
in principle one can try to analyze to get quantitative informations. Hence
the statistics of plwp reflect into statistics of the flow. The latter sentences
require careful analysis since one has to mix up in a rather complex way the
measures p[wp for different manifolds Wp to get statistical properties of a
trajectory. Rigorously speaking, the point is still unclear. However, local-
ization on the attractor (which is related to conditioning to W:) seems t o
be compatible with scaling properties of p : one can presume that universal
scaling properties do not depend so much on the local piece of attractor we
observe, while more large scale properties (depending for instance on the
particular geometry of the boundary) may vary in essential way over the
attractor.
Fokker-Planck equation. In the stochastic case, again 2D, we think
that the unique invariant measure p is a sort of diffused regularization
of the invariant measure p d e t of the deterministic system. Again for the
models rigorously covered by the RSB theory, the measures ps of suitable
random perturbations of order E converge to p d e t as E -+ 0. The additional
regularity of the measure p has the good consequence that it satisfies certain
elliptic infinite dimensional equations of Fokker-Planck type, so in principle
there could be a way to obtain quantitative results from these equations.
However, at present, really promising results in this direction are not known,
especially as far as scaling properties of local quantities are concerned. This
approach seems to be more promising to get large scale informations by
suitable finite dimensional or large eddies approximations. An argument in
favor of the SRB approach instead of the Fokker-Planck one is the following.
In the section on vortex filaments we underlined the transient features of
such vortex structures. The restrictions plw2 may capture features that
change in time, while it looks less reasonable to see them directly from p
using global tools like the Fokker-Planck equation.
155
(&w) (s) = w (t + s) - iJ ( t ) .
cp(t,w) : H + P ( H )
Remark 4.1. One can also associate a random dynamical system by lifting
the dynamics in the path space L 2 (0,oo;H ) . This approach, introduced
by Sell in the deterministic case, has been developed also in the stochastic
one, see Cutland8, Flandoli et a1 19.
Remark 4.2. The map cp ( t ,w ) does not have good continuity properties
(similarly t o the lack of uniqueness). Just the following very weak form
The following theorem is proved in the series of papers and the part on
the weakly compact attractor in H by standard analysis is still a work in
progress. See Cutland', Flandoli et a1 l 9 and references therein. One of the
157
claims require the following assumption, which represents one of the main
open problems in the theory of the Navier-Stokes equations:
In 2-dimensions the compact global attractor exists and (at least for
certain noise) has finite Hausdorff dimension.
P (f)= (Ptf)
where Ptf (x)= E [f('p ( t ,.) x)]. Denote by F<o
- the a-algebra generated
by the mappings w H cp (t,&,w) x,with 0 5 t 5 s. It describes the past.
for some constant C > 0 depending only on g . Then there exists a (unique)
random invariant measure p,, supported by A ( w ) , such that for every u E
C, every f E C b ( H ) , every A' < X and P-a.e. w E R one has
159
5 . Conclusions
We arised the question whether the physical invariant measure p of the
Navier-Stokes equations, conditioned to the unstable manifolds W:, is ‘ re-
lated to the statistics of vortex structures, like the Gibbs measure of vortex
filaments. We believe that a statistical analysis of some typical instability
of fluid flows could throw some light.
Up t o now, just a few objects of such a story are known, like invariant
measures for the Navier-Stokes equations, random attractors, random in-
variant measures possibly with some RSB properties, and some ensembles
of vortex filaments.
References
1. L. Arnold, Random Dynamical Systems, Springer, Berlin 1998.
2. D. Bonn, Y. Couder, P. H. J. van Damm, S. Douady, Phys. Rev. E. 47, 28
(1993).
3. D. Boyer, J. C. Elicer-Cortbs, J. Phys. A : Math. Gen. 33,6859 (2000).
4. A. Chorin, Vorticity and Turbulence, Springer-Verlag, 1994.
5. P. Constantin, C. Foias, R. Temam, Attractors Representing Turbulent Flows,
Memoirs Amer. Math. SOC.314, Providence 1985.
6. H. Crauel, Random Probability Measures on Polish Spaces, Habilitations-
schrift, Universitat Bremen, 1995.
7. H. Crauel, F. Flandoli, Prvbab. Theory Rel. Fields 100, 365 (1994).
8. N. J. Cutland, these proceedings.
160
CHRISTOPHE GIRAUD
Laboratoire J.A.Dieudonne U M R CNRS 6621
Universite d e Nice Sophia-Antipolis
Parc Valrose
06108 Nice Cedex 2, F R A N C E
E-mail: [email protected]
1. Introduction
This text aims at surveying some key properties of the solutions of the
one-dimensional (inviscid) Burgers equation
&u+ud,u=O (1)
with initial condition of white noise” type. Burgers introduced this equation
+
around 1940 in its multidimensional form, &u u . Du = 0, as a toy model
for hydrodynamic turbulence. It is known nowadays this far from accurate;
see Kraichnan’l for a discussion on the similarities and the differences with
Navier-Stokes equation. Yet, Burgers equation appears in many fields of
mathematical physics, such as the formation of the large scale structures of
the universe, or the dynamics of growing surfaces, see e.g. Woyc~ynski’~.
The study of the solution of Burgers equation (1)with white noise initial
condition takes place in the field of the analysis of solutions of PDE’s with
random initial data. If we think to the phenomenon of turbulence, it seems
interesting to exhibit the statistical properties of the solutions of some
PDE of fluid mechanics, with random and chaotic initial conditions. Such
studies also appear in astrophysics, when one considers the formation of
the structures of the universe. Solutions of Burgers equation with random
161
162
2 HLZ + 2t
u(5,O)d z -2’.
1
In the case of a white noise initial condition u(.,0), the analysis of u thus
requires a deep analysis of the convex hull of a Brownian motion with
parabolic drift, which is mainly based on the work of Groeneboomlg. See
Section 3 for a sketch of this analysis. There are also some interesting
connections with the phenomenon of coalescence and fragmentation] see
Bertoin5.
The rest of the paper intends t o review the main properties of the so-
lutions of Burgers equation (1) with initial condition of white noise type.
Section 2 recalls necessary background on Burgers equation (with deter-
ministic initial condition). In Section 3, various results on the convex hull
of a Brownian motion with parabolic drift are collected. Even if at first
sight they seem to have little to do with Burgers turbulence] they are the
key for the understanding of the proofs of the next sections. In Section 4
the main properties of the solution of (1) with white noise initial condition
are depicted. A special attention is given to its time-evolution. Some other
163
Even for very smooth initial conditions, solutions may develop shocks
(discontinuities) at finite time. We then lose the existence of strong solu-
tions, as well as the uniqueness of weak solutions. We will focus henceforth
on a special (weak) solution of (l),the so-called entropy solution, since it
is the physically meaningful1 solution of (l),see g. This special solution
can be obtained in adding a vanishing viscosity term to equation (1). More
precisely, the viscid equation
dtu+udd33u=&d~ZU
has a unique solution u, which converges as E + 0, except maybe on a set
of Lebesgue measure 0, to the entropy solution u of (1).
Provided that the so-called initial potential W ( z ):= u(z,0) dz fulfills
the condition
z I--+ W ( z )+ -1( z - x)
2
.
2t
Then, on the one hand a(x,t) coincides with the right-continuous inverse of
t times the derivative of the convex hull 'lit. On the other hand, a versionb
of the entropy solution u of (1) is given by the Hopf-Cole formula
x - a ( z ,t)
t)=
u(2,
t '
see 11$20, Notice already that the discontinuities of x ++ u(x,t ) come from
the discontinuities of z H u ( z ,t ) . Since z H a(z,t)is right-continuous and
bA weak solution is only defined up t o a set of Lebesgue measure 0, we can thus only
speak of a version of it.
164
increasing, they are of the first kind and always negative (this is precisely
the entropy condition).
As mentioned before, we can interpret the entropy solution in terms
of a system of ballistic aggregation. Consider a t time t = 0, infinitesimal
particles spread on the real line according to the uniform density p(dz, 0) =
d z , with velocities given by the velocity field u,(., 0). Then, let the system
evolve according to the dynamics of free sticky particles described in the
introduction. At time t , the velocity field of the system fits with (a version
of) the entropy solution u(., t ) with initial condition u(.,
0). Moreover, the
function a ( z ,t ) defined above represents the right-most initial location of
the particles lying in ] - 00, z] at time t. In other words, the density of
mass in the system is given at time t by the Stieltjes measure
d l z , Yl, t ) = 4 Y , t ) - 4 x 1 t ) .
Therefore, the jumps of z H a(z,t ) ,which correspond to the shocks of z H
u(z,t ) , also correspond to the macroscopic clusters of particles (clusters of
positive mass) present in the system at time t . Actually, a jump of a(., t ) at
z corresponds exactly to a macroscopic cluster located in z, whose mass is
given by a ( z ,t ) - a ( z - , t ) ,where the notation a ( z - , t ) refers t o the left limit
t ) a t z. The velocity V of this cluster is enforced by the conservation
of a(.,
of momentum
a(x,t) 2z - a ( z ,t ) - a ( z - , t )
u ( z ,0) dz =
V = 451 t ) -I4 z - >t ) aS( z - , t ) 2t
In the special case where z H a ( z ,t ) is a step function, we say that the
shock structure is discrete at time t. The path z H u ( z , t )is then shaped
as a toothpath made of pieces of line of slope l / t separated by negative
jumps (shocks). In terms of ballistic aggregation, a discrete shock structure
corresponds to a state of the system where all particles have clumped into
macroscopic clusters, whose locations form a discrete sequence on the real
line. From a geometrical point of view, the shock structure is discrete if
+
and only if the convex hull ?it of z H W ( z ) &z2 is piecewise linear. It
is convenient in this case to introduce the so-called &-parabolic hull Pt of
the initial potential W , defined by
When the convex hull ?it is piecewise linear, the parabolic hull Pt is made
of pieces of parabola. Indeed, to a linear piece of X t with slope X / t , say
+
( z H 5 X z k ; a I z 5 b ) , corresponds a piece of parabola of Pt
165
Finally, we emphasize that the above analysis still makes sense when the
initial condition u(., 0) is not a real function, but is only the derivative in
the sense of Schwartz of an initial potential W fulfilling condition (2). The
solution u(., t ) is then a real function at any time t > 0 and when t -+ 0+,
it converges in the sense of Schwartz to u(., 0), which is still said to be the
initial condition. The white noise initial condition is to be understood in
this sense.
tion of Girsanov Theorem shows that this property still holds for the convex
hull of a Brownian motion with parabolic drift, see Groeneboomlg and also
Avellaneda & E4.
Theorem 3.1. The convex hull 7tt of a (two-sided) Brownian motion with
+
parabolic drift (Wz &t2;t E R) is piecewise linear with probability one.
Recall that when the convex hull 7 t t is piecewise linear, the &-parabolic
hull of W is made of pieces of parabola. We can index these pieces of
parabola by Z, with indices increasing from left to right and the convention
that parabola number 1 is the first parabola whose vertex is located at the
right of 0. We write X, for the abscissa of the vertex of the piece of nt h
parabola and also Mn-l and M, for the abscissae of its end-points; see
Figure 1. One may notice that, in the notation of the previous section,
M, = .(X,,t).
The parabolic hull Pt is fully determined by the sequence (X,, Mn)nE~.
A characterization of the distribution of this sequence can be easily derived
from the work of Groeneboomlg on Brownian motions with parabolic drift.
It involves the Laplace transform C(X) of the integral of a Brownian excur-
sion e of duration 1. According to Groeneboom's formula (see l9 Lemma
4.2.(iii))
n=l
for X > 0, where 0 > -w1 > -w2 > . . . denotes the zeros of the Airy
function A i (see on p 446). We also introduce, following Groeneboom's
notations, the function g : R + Rf defined by its Fourier transform
Theorem 3.2.
The sequences ((0, Mo), ( X n ,Mn)n>l} and ((0, Mo), (X-n+l, M-n)n>l}
are two Marlcov chains, independent conditionally on Mo , with transitions
given by
167
(an-i - z),
3
- (an-l - zn-i)
6t2
') dz,dm,. (5)
Moreover, the law of MO is given by
1
1
P(Mo E da) = -92 5 / 3 t 2 / 3 ( - ( 2 t ) - 2 / 3 a ) g ((2t)-2/3a da.
)
where m, = Mn - M,-1.
Remark: A straightforward application of Girsanov Theorem shows that
the law v ( m ,t ) is absolutely continuous with respect t o the law P[rn]of dml.
Actually,
exp (- $ STeiml d z )
d v ( m ,t ) = dPIrn]
IE (exp (-$ Sr eiml d z ) )
The law of the variables u(m) and q ( m ) plays a key role in the analysis of
Burgers turbulence with white noise initial data. It is specified in the next
theorem, in terms of the function C defined above. See l6 for proof.
168
kW( m - 3 / 2 a ( 1 ) ,
( a ( m )rl(m))
, .
For any a > 0 and 0 < x < 1, the probability density function of
( a ( l ) , ~ ( lis) )given by
e-a2/24
P(a(1)E da, ~ ( 1E) dx) = C (ax3/') C ( a ( 1 - x ) ~ /dadx.
~ )
diG$Tj
( u ( x t, ) ;x E R) 'EW(t-'/3u xt-2/3 1 R
( 1 ) ;
xE
).
169
2->-
the others. Moreover, since the conditional law of & given Pt only depends
on time t and the mass m of the cluster, the fragmentation of the cluster
only depends on m and t , and not on its velocity or location.
According to the previous theorem, we can focus on a single cluster of
mass m at time t. We now turn our attention to its first splitting.
Theorem 4.2. With probability one a cluster splits into exactly two clus-
ters at its first splitting. The law of the time t* of the splitting of a cluster
of mass m at time t and of the mass m* of the left-most cluster arising
from this splitting is given b y
P(t* E ds,m* E d m l )
for ( s ,m i ) E ] O , t [ x ] O , m [with
, the notation m2 = m - ml and C defined
bY (3).
Moreover, we have for 0 < s < t
The previous result depicts the first splitting. Combined with a Markov
property at the times of fragmentation (see 1 5 ) , it yields a complete descrip-
tion of the fragmentation of a cluster. This description can be formulated
as follows. We write ml, . . . , m k for the masses of the clusters resulting at
time s = t - r of the fragmentation of a cluster of mass m at time t. The
mass ml refers to the mass of the left-most cluster, the mass mk to the one
of the right-most cluster. We write also
1
M ( m ) t ) (+r h) = ( m l , .. . , mi,l,mi,2,. . . , m k ) M(mit)(r)=
( m l ,. . .,mi,.
. . ,m k )
We refer t o l5 for the proof of the Markov property and l6 for the compu-
tation of the rate of jump.
We end this section with a remark about the dynamics of fragmen-
tation. The property stated in Theorem 4.1 bears the same flavor as
the so-called fragmentation property considered by Aldous', PitmanZ4and
Bertoin7. Nevertheless, the fragmentation process r ++ M("lt)(r)we study
here is not homogeneous in time and therefore differs from those considered
by Aldous et al. Besides, a cluster of mass m at time t statistically breaks
into pieces in the same way as a cluster of mass mt-'l3 at time 1. This per-
mits us to associate a time homogeneous Markov process to r H M ( m , t ) ( r ) .
Indeed, the process
fi(Wt)(') := t-2/3e2"/3M(m,t)(te-s), s E Rf
Figure 4. Shape of x H u ( x ,t ) .
there are infinitesimal particles a t rest. On its right, all particles have
clumped into macroscopic clusters, whose locations and masses are given
by (XnlM n ) n E ~The
. location and the mass of the front cluster correspond
t o xt and Mt.
The first property to mention about the shock front is the time-scaling
identity in law
(xt,M t ) ’aw(t2l3x1,t2l3M1).
This property originates from the scaling property of the white noise and
permits t o focus on time t = 1. The second property to be noticed, is that
the shock front is completely described a t time t = 1 by the variables z1
and M I . Indeed, according to the conservation of mass and momentum the
velocity & of the shock front is given by V1 = - ; M I . This equality can
be extended at any time t > 0 by
Ai’ (-tun)
exp (-21/”zWn)
n=l
where, as before, 0 > -w1 > -w2 > ... represent the zeros of the Airy
function A i ranked in decreasing order. See l7 for proof and also the law
of x1 alone.
for M ,x > 0.
175
so that the evolution of the shock front can be fully expressed in terms of
the process t H Mt,which is characterized in the following theorem.
Some other aspects of the solution u(., t ) have also been investigated. The
main contributions are perhaps the description of the flux of particles cross-
ing a given point and the study of the different scaling regimes of the solu-
tion by Frachebourg, Jacquemet and Martin13, see also '. Besides, it can be
noticed that the genealogy of a macroscopic cluster present in the system,
is statistically the same as the genealogy considered in Section 4. Finally,
we mention the work of Tribe & ZaboronskiZ5 and also of Frachebourg et
176
References
1. M. Abramowitz, I.A. Stegun: Handbook of mathematical functions. Washing-
ton: Nat. Bur. Stand. 1964
2. D. Aldous: Deterministic and stochastic models for coalescence (aggregation,
coagulation): review of the mean-field theory f o r probabilists. Bernouilli 5
(1999), pp 3-48.
3. M. Avellaneda: Statistical properties of shocks in Burgers turbulence II: tail
probabilities for velocities, shock-strengths and rarefaction intervals. Comm.
Math. Phys. 169 (1995), pp 45-59.
4. M. Avellaneda and W. E: Statistical properties of shocks in Burgers turbu-
lence. Comm. Math. Phys. 172 (1995) pp 13-38
5. J. Bertoin, Clustering statistics for sticky particles with Brownian initial ve-
locity. J. Math. Pures Appl. 79 no 2 (2000), pp 173-194.
6. J. Bertoin, Some properties of Burgers turbulence with white or stable noise
initial data. In LBvy Processes : Theory and Applications. Eds Barndorff-
Nielsen, Mikosh et Resnick. Birkhuser (2001).
7. J. Bertoin Homogeneous fragmentation processes. Probab. Theory Related
Fields 121 (2001), no. 3, pp 301-318
8. J. Bertoin, C. Giraud, Y. Isozaki: Statistics of a flux in Burgers turbulence
with one-sided Brownian initial data. Cornrn. Math. Phys. 224 (200l), pp
551-564
9. Y. Brenier, E. Grenier: Sticky particles and scalar conservation laws SIAM J.
Numer. Anal. 35 No 6 (1998), pp 2317-2328.
10. J.M. Burgers: The nonlinear diffusion equation. Dordrecht, Reidel 1974
11. J.D. Cole: On a quasi linear parabolic equation occuring in aerodynamics.
Quart. Appl. Math. 9 (1951), pp 225-236
12. W. E, Ya.G. Rykov, Ya.G. Sinai: Generalized variational principles, global
weak solutions and behavior with random initial data for systems of conser-
vation laws arising in adhesion particle dynamics. Comm. Math. Phys. 177
(1996), pp 349-380
178
Y. E. GLIKLIKH
Mathematics Faculty
Voronezh State University
Universitetskaya p l . , 1
394006 Voronezh, Russia
E-mail: [email protected]
179
180
in R" and flat torus Inas well as natural analogues of such processes on
groups (infinite-dimensional manifulds) of diffeornorphisms. In (1) w ( t )is
a Wiener process, adapted to ( ( t ) ,a ( t ,x) is a vector field and 0 > 0 is a
real constant.
Following Nelson (see, e.g., - 11) we give the next:
where (as well as in (i)) the limit is supposed to exist in L1(R, F, P ) and
At +O means the same as in (i).
---f
Lemma 1.2. For process ( 1 ) an R" the following formulae take place:
DZ =
a
-2 + (YO.0)Z+ -v2z,
o2
at 2 (9)
D*Z = - Z +
a o2
(Y*".V ) 2 - - P Z ,
at 2
where V = (A,
..., &), V2 i s the Laplacian, the dot denotes the scalar
, an,d Y f ( t ,x) are introduced in
product in Rn and the vector fields Y o ( t x)
(4).
The main idea of description of viscous hydrodynamics in the language
of mean derivatives is as follows.
For the sake of convenience we deal with fluids moving in a flat n-
dimensional torus I".It is the quotient space of €2" with respect the
integral lattice where the Riemannian metric is inherited from Rn. Consider
182
the vector space V e c t ( s )of all Sobolev HS-vector fields (s > + 1) on I".
Introduce the L2-scalar product in V e c t ( s )by the formula
(X,
Y )= 1 I n
< X(X), Y ( x )> CL(dx) (11)
where < ., . > is the Riemannian metric on 7"and p is the form of Rie-
mannian volume (here it is the ordinary Lebesgue measure on 7").
Denote by p the subspace of V e c d s ) consisting of all divergence-free
vector fields. Then consider the projector
P : v e c d s ) -+ p (12)
orthogonal with respect to (11). Notice that from Hodge decomposition it
follows that the kernel of P is the subspace consisting of all gradients. Thus
for any Y E Vect(s)we have
P ( Y )= Y - gradp (13)
where p is a certain HS+l function on I" that is unique to within the
constants for given Y .
Let a random flow [ ( t ) be given on a flat n-dimensional torus 7".Sup-
pose that it is a general solution of a stochastic differential equation of the
type
d J ( t )= a ( s , J ( s ) ) d s + udw(t) (14)
where u > 0 is a real constant. Let o , c ( t ) = u ( t , E ( t ) ) where
, u ( t , z )is
a divergence-free vector field on I", C1-smooth in t and C2-smooth in
m E 7".Suppose that [ ( t ,x) satisfies the relation
= -u
a + (ulV)u -
U2
-V2u - gradp.
at 2
Thus (15) means that the divergence-free vector field u ( t l x )satisfies the
relation
a + (u,0 ) u
-u -
U2
-V2u - gradp = F,
at 2
that is the Navier-Stokes equation with viscosity $ and external force
F ( t ,XI.
183
(X,
Y ) f= 1 I"
< X ( z ) ,Y(,) >f(.) ll(dz). (18)
The family of these scalar products form the weak Riemannian metric on
D s ( I n ) (it generates the topology, weaker than H ' ) .
The right-hand translation R f : D"(I") + D"(I"), Rf o 0 = 0 o f ,
8 , f E V s ( I n )is , Cw-smooth and thus one may consider right-invariant
vector fields on D s ( I n ) . Note that the tangent to right translation takes
the form: T R f X = X o f for X E T D " ( I n ) .
A right-invariant vector field X on D i ( 7 " ) generated by a vector X E
T e D " ( l n )is C'-smooth iff the vector field X on 7"is Hs+'--smooth This
fact is a consequence of the so-called w-lemma (see 6 , and it is valid also
for more complicated fields. For example, if a tensor (or any other) field
on 7"is Cw-smooth, the corresponding right-invariant field on D s ( 7 " ) is
Cw-smooth as well.
One can easily check that the second tangent bundle T T D ' ( 7 " ) con-
sists of H" maps from 7"to TT7" with additional properties that they
are projected into maps from D S ( 7 " ) . Thus we can apply the connector
K : TT7" -+ T I " of introduced above to obtain the connector on
TTV"(7")by the formula
K : TTD"(7") 4 TD"(7"). (19)
The family of its kernels in second tangent spaces form the connection on
D " ( I " ) ,denoted by 7-1.
The geodesic spray 2 of is described as follows:
2(X)=2oX (20)
for X E T V S ( I n )where
, 2 is the geodesic spray of the connection 'Ft on
I" (see above). One can easily obtain from (20) the following statement:
185
S ( X ) = T B ( 2 o X ) , X E p. (21)
Since P and 2 are P ( P ) - r i g h t - i n v a r i a n t and C"-smooth on T D S ( 7 " ) ,
it evidently follows from (21) that so is S.
Introduce the operators:
B :TI" --f R",
the projection onto the second factor in 7"x R";
A ( z ) : R" 4 Tm7", (22)
the converse to B linear isomorphism from Rn onto the tangent space to
I"at m E I", and
Qg(z) = A ( g ( z ) )0 B (23)
where g E D s ( I " ) , m E 7".
For a vector Y E T f D s ( I " ) we get QgY = A ( g ( z ) )0 B ( Y ( z ) )E
T,DS(7") for any f E Ds(In). In particular, Q,Y E V e c t ( " ) . Notice
that for Y E pf the vector QeY may not belong to Be. The operation Q ,
is a formalization for D s ( I " ) of the usual finite-dimensional operation that
allows one to consider the composition X o f of a vector X E V e c d s ) and
diffeomorphism f E D s ( l n ) as a vector in V e c t ( s ) .It denotes the shift of
a vector, applied at the point f(z),to the point z with respect t o global
parallelism of the tangent bundle to torus.
The map A has the following property. For the natural orthonormal
frame b in R" we have an orthonormal frame A,(b) in T , P , the field of
frames A(b) on T7" consists of frames inherited from the constant frame
b. Thus for a fixed vector X E R" the vector field A ( X ) on I nis constant
(i.e., it is obtained from the constant vector field X on R" and has constant
coordinates with respect to A(b))and in particular A ( X ) is Coo-smooth and
186
divergent-free since such is the constant vector field X on R". So, A may
be considered as a map A : R" p = 0, c T e D s ( l n ) .
-+
d l ( t ) = q t ,l ( t ) ) d t + cA(l(t))dw(t) (25)
on D s(7").
Definition 3.1. If<(t) satisfies an equation of (25) type with some (maybe
random) initial condition, we say that it is a process with diffusion term
aA.
v ( t ) = [(T - t ) with inverse time direction has the same diffusion term aA
but, generally speaking, different drift.
The definition of mean derivatives for processes on V s(7") is analogous
to that on R" and on 7".In order to distinguish the derivatives on Ds('Tn)
and on Tnwe denote the former by D and D* while D and D , remain valid
for I".
The mechanical meaning of the subbundle fl is a constraint. According
to the ideology of geometric description of constraints suggested by Vershik
and Faddeev, we give the following
Theorem 3.2. If the process u(T - t ,q ( t ) )on p satisfies the It; equation
in Belopols~aya-Daletskiif o r m
T
d u ( T - t , q ( t ) )= exp,(T-t,l)(t))(-S(iZ(T-t,q ( t ) ) ) d t - F ' ( t , q(t)))dt
189
-T -
+aA - t , rl(t)))dw(t)), (29)
the process q ( t ) and the right-invariant admissible vector field ii on V s( I " )
satisfy (27) and (28) and so ( ( t )= q ( T - t ) satisfies (26) and the divergence-
free vector field u(t,x) on I" is a solution of (17).
Theorem 3.2 follows from a statement of Lemma 1.1 type for equations
in Belopolskaya-Daletskii form (see, e.g., 7, 8 ) .
The next finite-dimensional interpretation makes the construction more
clear. Notice that the process q ( t ) with initial condition q ( 0 ) = e on
V8(In), that satisfies (27), is a random flow on I". Denote this flow
by q ( t , x ) with q ( 0 , x ) = x. It is the general solution of It6 stochastic
differential equation on I"
dq(t,X) = -u(T - t , q ( t ,x ) ) d t -tadw(t) (30)
with divu(t, x) = 0, the finite-dimensional version of (27). By direct calcu-
lation of forward mean derivatives for the finite dimensional process q ( t ,x)
we show that
Drl(t1.) = - 4 T - t177(4 x)),
P D D q ( t ,z) =
a
--v(T
at - t ,q(t,.)) f (Gt , 77(t1z)), V)U(T- t , d t , x))-
-
U2
-V2u(T - t , q ( t ,x)) - gradp.
2
The latter equality is turned into (16) under the change of variables q ( t ,x) =
[ ( T - t ) . Thus equation (29) guarantees that for the process q ( t ) ,satisfying
(30), the relation P D D q ( t , x ) = F ( t , x ) holds. The same relation can be
achieved also by another way.
For a stochastic differential equation with respect t o a process ( ( t ) on
Vs(In) denote by & ( s ) its solution with initial condition & ( t )= e. Con-
sider the following system on V s(I"):
dq(t) = -G(T - +
t , q(t))dt a A ( q ( t ) ) d w ( t )
P t
Acknowledgments
T h e research is supported in part by Grant 99-00559 from INTAS, Grant
UR.04.01.008 of the program Universities of Russia and by U.S. CRDF -
RF Ministry of Education Award VZ-010-0.
References
1. Arnol’d V. Sur la gkomktrie diffkrentielle des groupes de Lie de dimen-
sion infinie et ses applications a l’hydrodynamique des fluides parfaits.
Ann.Inst.FourierT.16, N 1,319-361 (1966).
2. Belopolskaya Ya.1. Probabilistic presentation for solutions of boundary-value
problems for hydrodynamical equations. Trudy POMI, V. 249, 71-102 (1997).
3. Belopolskaya Ya.1. and Dalecky, Yu.L. Stochastic processes and differential
geometry. Kluwer Academic Publishers, Dordrecht 1989
4. Belopolskaya Ya.I., Gliklikh Yu.E. Diffusion processes on groups of diffeomor-
phisms and hydrodynamics of viscous incompressible fluid. Transactions of
RANS, ser. MMMIC, V. 3, N. 2 , 27-35 (1999).
5. Busnello B. A probabilistic approach to the two-dimensional Navier-Stokes
equation. The Annals of Probability, V. 27, No. 4, 1750-1780 (1999).
6. Ebin D.G. and Marsden J. Groups of diffeomorphisms and the motion of an
incompressible fluid Annals of Math.,V.92, N 1, 102-163 (1970).
7. Gliklikh Yu.E. Ordinary and Stochastic Differential Geometry as a Tool for
Mathematical Physics.- Dordrecht: Kluwer, 1996.
8. Gliklikh Yu.E. Global Analysis in Mathematical Physics. Geometric and
Stochastic Methods.- N .Y.: Springer-Verlag , 1997.
9. Nelson, E. Derivation of the Schrodinger equation from Newtonian mechanics.
Phys. Reviews, 150 (4), 1079-1085 (1966)
10. Nelson, E. Dynamical theory of Brownian motion.-Princeton: Princeton Uni-
versity Press, 1967
11. Nelson E. Quantum Fluctuations.-Princeton: Princeton University Press,
1985.
FURTHER CLASSES OF PSEUDO-DIFFERENTIAL
OPERATORS APPLICABLE TO MODELLING IN FINANCE
AND TURBULENCE
1. Introduction
Since the pioneering work of E. EberleinlO on comparing solutions for fi-
nance market models obtained from models driven by diffusions with real-
world data it is clear that jump processes yield much better models. The
most widely used class of jump processes for modelling are L&y processes.
We refer to the surveys of E. Eberleing and 0. Barndorff-Nielsen and N.
Shephard4, respectively, and the references given therein. The fact that
L6vy processes have stationary and independent increments implies a cer-
tain “translation-invariance” of the distribution corresponding to the under-
lying process. This fact excludes for example that a change of parameters
occurs when a certain threshold (for prices for example) is reached.
In the very original paper, 0. Barndorff-Nielsen and S. Levendorskii3
therefore started to model finance markets by using distributions in-
volving parameters depending on the price, or by an abuse of the lan-
guage of physics: price-homogeneous distributions were replaced by price-
inhomogeneous distributions. Tracking back to the generator of the un-
191
192
derlying Markov process, this change necessitates the switch from con-
stant coefficient operators to variable coefficient operators. In their paper,
Barndorff-Nielsen and Levendorskii used the representation of the generator
as a pseudo-differential operator following l6 where it was emphasised that
pseudo-differential operators are canonical tools in the theory of Markov
processes. They chose to work with classical pseudo-differential operators,
i.e. with symbols q(x,<)in the class Sz6 with the additional assumption
that for all z E R",q(z, .) : Rn -+ C is a continuous negative definite func-
tion to ensure that the generated semigroup is a Feller semigroup. Their
model symbol is:
which gives for frozen coefficients, i.e. z = ZO, just a normal inverse Gaus-
sian distribution with parameters ~ ( z o )~, ( z o )~, ( z oand
) P ( s 0 ) . The fact
that modelling with normal inverse Gaussian distribution is rather suc-
cessful and that Eq. (1) belongs to the Hormander class allowed them to
emphasise the need for having smooth symbols in modelling finance mar-
kets.
The purpose of this note is to show that there are large and rather
general classes of smooth but non-classical symbols leading to pseudo-
differential operators generating Feller semigroups. After we have discussed
some basic facts on pseudo-differential operators and Markov processes, we
will first discuss W. Hoh's l 1 > l 2symbol class and then the Weyl calculus
approach due to F. Baldus '.Further we will have a short look a t subor-
dination in the sense of Bochner as well as t o pseudo-differential operators
of variable order of differentiation.
Many results in finance have a counterpart in turbulence, compare for
example Barndorff-Nielsen2 and Barndorff-Nielsen and Shephard4. In par-
ticular experimental observations show that the time derivative of a fluid's
velocity field is not log normally distributed. It has instead a hyperbolic or
normally inverse Gaussian type of distribution such as those arising for the
above jump processes. Thus, the classes of pseudo-differential operators
introduced here may also be helpful in modelling turbulence problems, e.g.
by considering classical models of fluid dynamics in a random environment
where the driving noise is a jump process like those described in what fol-
lows. Extensive results are already known for models involving Burgers
equation when the driving noise is white noise. Here we have exact solu-
tions for the Burgers velocity field in terms of a stochastic mechanics with
additive white noise. The challenge is to replace this noise with a jump
193
l,k=l
al(z)EkEl
The basic observation is that (for reasonably nice processes) the genera-
tor of the semigroup Ttu(z)= Ex(u( X t ) )is given by the pseudo-differential
operator
-q(x, D ) u ( x )= - (27r-"
"S
Wn
eiz'cq(z,<)ii(<)d<, (5)
and
$(<) = 1
Wn\{o)
(1 -cosY+(dY) . (8)
195
aa$(E) = 1 a;
a-\{o)
(1 - cosy. E ) v (dy)
= ca 1
Wn\to)
ya cos) (Y . E ) v (dy) 7
ca being f l .
Thus, if s
wn\{o}
lyal v (dy) < 00, then a"$ exists by the dominated con-
Lemma 3.1. If $ has t h e representation Eq. (8) and if Mla1 exists for
2 5 J Q J5 m.t h e n E C" (Rn) and
<
Note that the continuous negative definite function H 1-cos a.<, a E R",
shows that Eq. (11) is optimal. Given any continuous negative definite
function with representation Eq. (8) we may always construct a continuous
negative definite function $R E C" (R")satisfying Eq. (11) for all m E No.
We just have to consider
IaE"a:P(x, 0
1Ica,&m"-P(IaI) . (15)
As worked out by W. Hoh11,12,it is possible to develop a complete symbolic
calculus including the parametrix construction for "elliptic" elements for
pseudo-differential operators p(x,D)with symbol p E S~~'. In particular,
if in addition $ satisfies a growth condition from below,
Theorem 3.1. (W. Hoh) A s s u m e Eq. (16), p E 5':)' such that p(x,.) :
R" R i s a negative definite function, and
P(Z, 02 fix2 (El (17)
for large 1
1
1 and some 6 > 0. T h e n (-p(z, D ) , C r (EXn)) is closable in
C, (R")and the closure generates a Feller semigroup.
Corollary 3.1. In case of Theorem 3.1 we find that p(x,5) i s the symbol
of a Feller process.
197
P(Z,O= .i'
an\{o)
(1- C O S Y . 5 ) N ( G d Y ) . (18)
P(X,C) = .i'
BR(O)\{O)
(1-cosY.<)N(x,dY)+
s
ak(o)
(1-cos(Y.t))"z,dY)
In this section we will briefly discuss results due to F. Baldus' who used
the Weyl calculus to construct Feller semigroups. Unfortunately we need
quite a lot of special notions to state the result. For a detailed discussion
we refer to F. Baldus' and also to the original paper and monograph by L.
HOrmander'sl4>15.
Denote by 0 the standard symplectic form on Rn x R",i.e.
we have:
hTM0 5 CM . (28)
If M and 6
are both sub-y-admissible we call M an invertible sub-y-
admissible weight function.
where Van is the gradient in R" x R" and (., .) is the scalar product. Further
we set for a metric y on R" x R" and Ic E No
-
m
S ( M , y ) = SE6 . (34)
Note that the Weyl-pseudo-differential operators q w ( z ,0)we are interested
in can always be transformed into the "usual" form
~ s) () i-, 'y ) .
Q(I,~)~B(L~(w= (36)
Further let M be a n invertible sub-y-admissible weight function and m 5 1
a n arbitrary sub-y-admissible weight function satisfying with some k E N
and C M > 0
Example 4.2. A. Elliptic elements p E Sz6 such that for all z E R"
the function 5 H p(x,<) is negative definite are included in Theorem 4.1.
20 1
Note however that certain extensions of Hoh's results, i.e. the perturba-
tion theory, is not covered. C . Symbols of mixed homogeneity are partly
included.
f ( z )= a + bz +
i: (1 - e-zt) p (dt) (42)
and q(xl<) depends on x , then the subordinated process X,f := Xs, and
the process yt with symbol f ( q (zl6)) are clearly distinct! The symbolic
calculi introduced in Sections 2 and 3 may be used to relate X s , and Yt. We
refer to Jacob and Schilling18 for a first simple approach and t o F. Baldus',
Section 6.5. What we may expect (and what holds true in many situations)
is that the generator -f (-4 (x,D))of X [ and the operator - (f o q) (x,D)
differ only by a "low order" term which will follow a reasonable asymptotics
of the transition function of ( X ! ) in terms of that of (K)t>O -
and vice
t2o
versa.
For modelling purposes maybe a different] but very related concept is
more important. Stable, especially symmetric stable processes, are very
often used for modelling] in finance and turbulence, but also in other
problems. We may interpret the symmetric stable process with index 2a,
0 < Q < 1, as the process obtained from Brownian motion by subordinat-
ing with the one-sided L6vy process associated with the Bernstein function
fa(s) = s". Indeed the characteristic exponent of the symmetric stable
process of index 2a is given by $a (I =) (
1 < 1 2 " which is just fa 1<12) and
<
of course H [<I2 is the characteristic exponent of Brownian motion. As
indicated in the introduction, often in modelling it is useful to make pa-
rameters z - (state space) dependent. Thus we may consider the function
c)
(x, H I<12a(z)l or more generally] if q(x16) is the symbol of a Feller pro-
cess we may have a look to the symbol
<
This function has the property that if H q(xl E ) is a continuous negative
<
definite function, so is H q(zl<)"(").Hence, it may lead t o a stochastic
process with generator
In case of q(xl <) = 1<12 (or more generally q(xl <) = ~ i , ak,l(x)<k<l)
~ = ~ we
enter the field of stable-like processes. For these processes a lot of results
(probabilistic as well as analytic) are known and we refer to the work of A.
Negoro20)21222 and coworkers and the references given therein.
W. Hoh13 was able to make his symbolic calculus also work for the
"stable-like" situation, an extension of F. Baldus' approach seems to be
possible. It should be mentioned that some work in this direction had been
stimulated by the paper Jacob and Schilling18.
203
References
1. Baldus, F., S ( M ,g)-pseudo-differential calculus with spectral invariance on
Rn and manifolds for Banach function spaces. Dissertation Universitat Mainz
2000, Logos Verlag, Berlin 2001.
2. Barndorff-Nielsen, 0. E., Probability and statistics: Selfdecomposability fi-
nance and turbulence. In : Probability Towards 2000 (eds. L. Accardi and
C. C. Heyde), Springer Verlag, Berlin 1998, 47 - 57.
3. Barndorff-Nielsen, 0. E., and Levendorskii, S. Z., Feller processes of normal
inverse Gaussian type. Quantitative Finance (to appear).
4. Barndorff-Nielsen, 0. E., and Shephard, N., Modelling by LBvy processes for
financial econometrics. In: L6vy processes: Theory and applications (eds. 0.
E. Barndorff-Nielsen, T. Mikosch, S. J . Resnick), Birkhauser Verlag, Boston
2001, 283 - 318.
5 . Bogarchenko, S. J., and Levendorskii, S. Z., Option pricing for truncated
Le'vy processes. Intern. J. Theor. Appl. Finance (to appear).
6. Bogarchenko, S. J., and Levendorskii, S. Z., Perpetual American options
under LBvy processes. Preprint 2000.
7. Bogarchenko, S. J., and Levendorskii, S. Z., Barrier options and touch-and-
out options under regular LBvy processes of exponential type. Preprint 2000.
8. Davies, I. M., Truman, A . , and Zhao, H. Z., Stochastic heat andSurgers
equations and their singularities - geometrical and analytical properties (The
fish and the butterfly, and why). Preprint 2001.
9. Eberlein, E., Application of generalised hyperbolic Le'vy motion to finance.
In: LBvy processes: Theory and applications (eds. 0. E. Barndorff-Nielsen,
T. Mikosch, S. J. Resnick), Birkhauser Verlag, Boston 2001, 319 - 336.
10. Eberlein, E., and Keller, U., Hyperbolic distributions in finance. Bernoulli 1
(1995), 281 - 299.
11. Hoh, W., Pseudo differential operators generating Markov processes. Habili-
tatsionschrift Universitait Bielefeld, 1998.
12. Hoh, W., A symbolic calculus for pseudo differential operators generating
Feller semigroups. Osaka J. Math. 35 (1998), 789-820.
13. Hoh, W., Pseudo differential operators with negative definite symbols of vari-
able order. Revista Mat. Iberoamericana 16 (2000), 219-241.
14. Hormander, L., The Weyl calculus of pseudo-differential operators. Comm.
Pure Appl. Math. 32 (1979), 359 - 443.
15. Hormander, L., The analysis of linear partial differential operators, vol. 3,
Springer Verlag, Berlin, 1985.
16. Jacob, N., Pseudo-differential operators and Markov processes. Akademie
Verlag, Berlin, 1996,
17. Jacob, N., and Leopold, H. G., Pseudo-differential operators with variable
204
1. Introduction
The rheology of non-newtonian fluids is a very lively field of modern fluid
mechanics. The challenge is to find a good relation linking within the fluid
the stress tensor to the velocity field in order to reproduce the behavior of
the fluid in some classical situations (shear flow, elongational flow) and to
simulate it in some more complex cases. This relation may be complicated
since the stress generally depends on the whole history of the velocity field.
Many approaches consist in deriving this relation from the microscopic
structure of the fluid. In some cases, it is possible to directly attack the
full system coupling the evolution of these microscopic structures to the
macroscopic quantities (such as velocity or pressure) : this is the so-called
micro-macro approach.
We are here interested in the modeling of polymeric fluids. More pre-
cisely, we consider dilute solutions of polymers, so that the chains of poly-
mers (the ‘hicroscopic structures”) do not interact with each other. In
order to describe the microscopic structure of this fluid, one can model a
polymer by a chain of beads and rods (this is the Kramers model) or more
simply by some beads linked by springs (see Figure 1). We consider here
the simplest model consisting in two beads linked by one spring : this is
the dumbbell model. In this model, the evolution of the end-to-end vector
(which joins the two beads) is described by a SDE. We refer the interested
205
206
reader to Refs 10)1,2,6 for the general physical background of these mod-
els. This SDE is actually coupled to the Navier-Stokes equation through
the expression of the stress tensor as an expectation value built from the
end-to-end vector.
u=o
Let us now introduce the equations we deal with. They read, in a non-
dimensional form :
where the parameter b > 0 measures the finite extensibility of the poly-
mer. The space variable y varies in c? = ( 0 , l ) and t varies in the
whole of R+. The random variables are defined on a filtered proba-
bility space (R,.FIFtllP). The random process (&, Wt) is a (.Ft)-two-
dimensional Brownian motion. We take Dirichlet boundary conditions on
, is a FO-
the velocity. The initial velocity is u(t = 0, .) = U O , and ( X O YO)
measurable random variable. We will suppose that ( X o ,Yo) is either such
+ +
that P ( X i Y: > b) = 0 (Section 2) or such that P ( X i Y; 2 b) = 0
(Sections 3 and 4).
We fix y in 0 , set g ( t ) = a,u(y,t) and suppose throughout this paper
that we have at least the following regularity on g :
system :
{ dXf = [-i1-
(5)
dY,g = -1.2 1- qg
(X?P+(Y,9)2
b
) dt+dWt,
with initial condition ( X O YO).
,
Let us begin by recalling from Ref. the precise mathematical meaning
we give to (5).
Definition 1.1. Let Xo = ( X o ,Yo)and Wt = (K,Wt). We shall say that
a (Ft)-adapted process Xi = ( X a ,yt”) is a solution to (5) when : for lP-a.e.
w , lft 2 0 ,
solution to reach the boundary of B (see Section 3). When b < 2 and
I P ( ( X O (<~ b) = 1, this probability is equal to one. This -
enables us
to construct (for g = 0) a solution to (6) that leaves a.s. B. Hence, if
b < 2 , uniqueness of solutions does not hold for solutions to (6) without
the additional requirement to take values in B. When b 2 2 and again
lP(IX0l2 < b) = 1, the probability to reach the boundary is equal to zero
and trajectorial uniqueness holds. We exhibit the unique invariant proba-
bility measure of the SDE (6) with g = 0 (see Section 4). All these results
on the SDE have an impact on the analysis and the understanding of the
coupled SDE-PDE system (for which we refer to Ref. They show that
the assumption b 2 2 adopted in Ref. to prove existence and uniqueness
of solution t o the coupled system is in some sense “optimal”.
209
Proposition 2.1. Under assumption (4), for any b > 0 and for any initial
, such that IP (Xi+ Y$ > b) = 0 , there exists a unique
condition ( X O YO)
solution to (6) with values in B.
We first prove the uniqueness statement (Section 2.1), then turn to the
existence first when g E L y (Section 2.2) and finally when g E L:,c (Section
2.3). In the following, the point is to notice that the singular term in the
drift derives from a convex potential II : R2 +]- 00, +m] :
(
Z t = X: - X,
- g, . By ItG's formula, we have :
dlZl: = 2 2 t . d Z t ,
= -2(VrI(X,s) - on(x;)).z,dt + 2g(t)(X,g- x,g)(qg - P)dt,
where x.y denotes the scalar product of x and y E lR2.
Using the fact that, since II is convex, for a,ny x and 2 E B , (VrI(x) -
VrI(%)).(x- 2 ) 2 0, we obtain, for any t 2 0 :
210
Using Gronwall Lemma and the fact that g E L~o,(R+),we have thus shown
- 9
that P-a.s, 'dt 2 0, XfAT= X t A T .Therefore, on {T < m}, IX$12= b. We
deduce that in case P(3t 2 0, IX:lz = b) = 0, T = 03 P-a.s. . 0
l t
I I ( X t )du 5 l t
II(a,) du + l t
( X t - a,).dKt. (10)
We are now going to recover a solution to (6) from the solution of (9).
More precisely, we follow the method of E. C6pa and D. L6pingle5 (see
Lemmas 3.3, 3.4 and 3.6) in order to identify the process K f .
We can thus show that for all 0 < t < co,we have :
where, for any a E [0, b ] , Lg denotes the local time in a of Rg. Using now
the occupation times formula (see Ref. l1 p. 215), we know (using ( 1 1 ) )
that, for any fixed t > 0 :
Since a + Lg is a.s. cadlag (see" p. 216), we deduce that for any t > 0,
IP-a.s., L: = 0. Using this in ( 1 5 ) , we obtain
dRf = 1Rf>O d R f .
Using this equality in ( 1 4 ) , we have : V t 2 0,
1
-
2& 0
s" 1Rj=O (-z+
b2
+ + +
ds 2g(s)X,SY,S d s ( 2 13)d s 2 X : . d W s
0
for any 0 < t < co,IP-as-.,
d X f = -VII(X:) d t
h 1-w
t
+ (g(t)Kg,O)dt
+
1
du
dWt.
< co,
Using the fact that, since II is convex, for any x and y E B , (VII(x) -
VII(y)).(z - y) 2 0, we obtain : V t E [O,T],
so that : V t E [O,T],
Ixin- x:" I 2
5 10
t
lgn(s)Y;" - p(s)Y;" I 1X.g" - x:" I ds
5 J' (lgn(s)l jY;"
0
- Y:* 1 + lY;m1 Ign(s) - g m ( s ) l ) 1X:" - X:" 1 ds
5 i1 t
lgl(s) IX:" - X;m d s I2 + 2b I' l g n ( s ) - g"(s)l d s .
Using Gronwall Lemma, we then obtain :
214
From this inequality and the fact that g E Li ([0, TI),we deduce that there
exists a continuous adapted process X : with values in B such that X:" --+
X : in L,"(L,oO([O,T I ) ) .
One has the following estimate on the total variation of VII(XE")du
on [O,T]:
Ix:"12
= IXOl2 - 1 t
Ix:" IXpd s + 2 t
g"(s) x;"Y,"" d s + 2 t + 2 1t
x:".dWs,
Ji
It is obvious that s,' X;" . d W , + X , . d W , in L:(Lp([O,TI))-norm. Up
to the extraction of a subsequence, we can suppose that this convergence
holds for almost every w . Using this property together with (18) and ( 1 9 ) ,
we deduce that for a.e. w , the measure VII(X:")dt on [O,T]is such that
Jz IVlr(Xig")ld t < C ( T , w ) where C(T,w) is a constant only depending
on T and w . 'One can thus extract a weakly converging subsequence of
the other hand, taking the limit n --+ co in ( 1 7 ) ,
1 du uniformly converges on [0, TI to K: satisfying :
K: = 1 t
(g(u)Y,, 0) du + Wt - (Xf- XO).
By identification of the limit, we have VII(X:") dt 2 dK: weakly.
By Definition 2.1, the processes X:" are such that for any continuous
(.Ft)-adapted process a t with values in R', for P-a.e. w , VO 5 s 5 t < co,
[ II(X5")d u 5
t
rI(a,)du + 1 t
(Xt"- a u ) . V r I ( X ~du.
") (20)
215
One can pass to the limit n -+ co in (20), using the fact that Il(Xt")+
II(Xt) pointwise in u and that II(Xt")is uniformly integrable. Indeed,
for any A 2 $, if we set Mu = , we have (since x ++
so that 1 T
du + 0 uniformly in n when A
11n(xzn,12AII(Xtn) + co. We
have thus obtained a continuous process Xf on [O,T]and a continuous
+
process with finite variation Kf = Ji(g(u)Y,lv,0 ) d u Wt - (Xi - XO)on
[0,TI such that for any continuous (Ft)-adapted process at with values in
lR2, for IP-a.e. w , VO 5 s 5 t < TI
It II(Xt)d u 5 1 t
II(a,) du +
t
(Xt- a,).dKt.
This shows that we have built a solution to the rnultivalued SDE (9) on
the time interval [0, TI. Since T is arbitrary, using Proposition 2.2, we have
built a solution on lR+.Following again the arguments of the last section
it is easy to show that :
Proposition 3.1. A s s u m e
9E Jm+), (21)
and that P(IX0J2< b) = 1. Let u s consider the process X : solution t o (6)
built above. We have :
if b _> 2, t h e n P ( 3> 0 , IX:12 = b) = 0 ,
0 if b < 2 , then P ( 3> 0 , IX:12 = b) = 1.
In view of Proposition 2.2, we deduce immediatly :
+
the process (Gg,WE) = (& s," g(s)Y,S d s , Wt) is a Brownian motion and
therefore ( X f ,y,",Kg,
WE,lPg)tER+ is a weak solution of the SDE ( 5 ) with
g = 0. Since this solution is with values in B,it is also a weak solution of
the multivalued SDE (9), with g = 0, for which uniqueness in law holds.
Since IPg and P are equivalent on F,we can then deduce the properties of
Proposition 3.1 in case g E L2(R+) from the properties of Proposition 3.1
in case g = 0.
In the following, we focus on the solution to (9) with g = .O, which we
denote by X t = (Xt,Y,). We fix x E B and the superscript x means that
we consider the solution to (9) with g = 0 such that Xo = x.
Let us first suppose that 1x1 > 0. Let us consider the process R," =
b - lXT12. We know that :
b2
dR," = -dt - ( 2 + b) d t - 2X,".dWt.
R,"
Let us introduce the stopping time
Let:'FI be defined by :
= 0. (23)
One can therefore show that IXrl > 0 on [O,T"),where
T" = lim T," = inf { t 2 0, 1XF12 = b } = inf {t 2 0, R," = O}.
n+o3
b2
dR: = -d t - (2 + b) d t + 2 d q d , & ,
R,"
where Pt is a Ft-adapted 1-dimensional Brownian motion.
Let us now introduce the stopping time
which leads to :
p' ( r ) = C(b - r ) - l ~ - - ~ / ~ ,
In case b < 2, we can deduce from the second item that 5'" = T". We now
want to know whether S" = +m or not in this case. Let us introduce the
speed measure m on (0, b) defined by
2 dr -
I- b / 2 dr
m(dr)= -
4(b-r)pt(r) 2C '
and the function v such that, for any r E (O,b),
= 0.
In case b < 2, we use the fact that, due to the proof of (23),
IP(IX~,,,I = 0) = 0. By the strong Markov property and
since IP-a.s., S U ~ ~IXt~0 I2 [ < ~ b, , we~ have
~ ~IP (
~ 3> 0, lX:12 = b ) =
E (P (3> 0, Ix:lz = b ) lx=X1/J = 1.
In case of a non-deterministic initial condition Xo with law po, we can
deduce the properties of Proposition 3.1 from the fact that (by uniqueness
of the solution) :
IP (3> 0, lXt12 = b) =
J IP ( 3> 0, IX:12 = b) dpo(x). 0
-b21n(r - b) if T > b,
otherwise.
so that af is a simple-valued maximal monotone operator with domain
I = (b,co) (for all T > b, d f ( r ) = { V f ( r ) }= {&}). By E. C6pa4, there
exists a unique process rt solution to (27). Following exactly the arguments
of Lemma 2.1, one can show that this process rt is such that :
for any 0
1 -
< t < co,P-a.s., 11
5t
1 1
du < M, with convention
- +W,
x t = (fi
cos(dt), fisin(&))
220
1 x t
d X t = --
21-1x,12 d t + (-sin(&), cos(8t))dyt+ (cos(&),sin(Ot))dpt.
b
Using Paul L6vy characterisation, one can show that
(- sin(&),cos(Ot))dyt + (cos(&),sin(Ot))d,Bt = d B t
where Bt is a two-dimensional Brownian motion, independent of W t.
b < 2. b 2 2.
Existence. Existence.
IP(IXo12 = b) = 0.
IP (32 0, lXt12 = b) IP ( 3 2 0, lXt12 = b)
= 1. = 0.
I Non-uniaueness. I Uniaueness
IP(IXo12= b) > 0.
Existence.
Non-uniqueness. I Existence.
Non-uniqueness
Table 1.: Properties of solutions to (6) when g = 0. We suppose
IP(I_XO~~5 b) = 1. In any case, uniqueness holds for solutions with values
in B according to Proposition 2.2. The terminology uniqueness and non
uniqueness relates to a solution that is not enforced to take values in B.
221
where X i = ( X i , & ) denotes (as in last section) the solution with values
in to (6) with g = 0 (see Ref. This expression of the stress yields the
following estimate (using Holder inequality) : for almost all y and t ,
w h e r e p = L.
9-1
Indeed, by (i), one easily checks that po(x)d z is invariant. By (ii), any
invariant probability measure is equivalent to the Lebesgue measure on B
which implies uniqueness (see Proposition 6.1.9 p. 188 of M. DuAo 7).
With Proposition 4.1, it is then straightforward to prove that, if X Ohas
the density po(x), then we have :
1zI2< b ( l - i)
and S 1 A n dx > 0 where A , = A n B
Girsanov Theorem, under IP: defined by :
Acknowledgments
This work has partly been motivated by some remarks of Claude Le Bris.
Bibliography
1. R.B. Bird, R.C. Armstrong, and 0. Hassager. Dynamics of polymeric liquids,
volume 1. Wiley Interscience, 1987.
2. R.B. Bird, C.F. Curtiss, R.C. Armstrong, and 0. Hassager. Dynamics of
polymeric liquids, volume 2. Wiley Interscience, 1987.
3. M. BOSSY,B. Jourdain, T. Leli$vre, C. Le Bris, and D. Talay. Existence of
solution for a micro-macro model of polymeric fluid : the FENE model. In
preparation.
4. E. CCpa. Equations diffkrentielles stochastiques multivoques. ThGse, Univer-
sit6 d’orlkans, 1994.
5. E. CCpa and D. Lepingle. Diffusing particles with electrostatic repulsion.
Probab. Theory Relat. Fields, 107:429-449, 1997.
6. M. Doi and S.F. Edwards. The Theory of Polymer Dynamics. International
Series of Monographs on Physics. Clarendon Press, 1988.
7. M. Duflo. Random iterative models. Springer, 1997.
8. B. Jourdain, T. LeliBvre, and C. Le Bris. Numerical analysis of micro-macro
simulations of polymeric fluid flows : a simple case. to appear in Math.
Models and Methods in Applied Sciences.
9. I. Karatzas and S.E. Shreve. Brownian motion and stochastic calculus.
Springer-Verlag, 1988.
10. H.C. Ottinger. Stochastic Processes in Polymeric Fluids. Springer, 1995.
11. D. Revuz and M. Yor. Continuous martingales and Brownian motion.
Springer-Verlag, 1994.
12. L.C.G. Rogers. Smooth transition densities for one-dimensional diffusions.
Bull. London Math. SOC.,17:157-161, 1985.
ON THE DISPERSION OF SETS UNDER THE ACTION OF
AN ISOTROPIC BROWNIAN FLOW*
H. LISEI
Faculty of Mathematics a n d Computer Science,
Babeg-Bolyai University,
Str. Koga"1niceanu Nr. 1,
RO - 3400 Cluj-Napoca, Romania
E-mail: [email protected]
M. SCHEUTZOW
Institut fur Mathematik, MA 7-5,
Technische Universitat Berlin,
Straj3e des 17. J u n i 136,
10623 Berlin, Germany
E-mail: [email protected]. de
We give a survey on results about the growth of the diameter of the image of a
bounded subset X of Rd under the action of a stochastic flow. We provide a new
proof of the fact that, under reasonable assumptions, the diameter of this image set
will almost surely grow a t most linearly in time, and we establish an explicit upper
bound for the linear growth rate which is both simpler and better than previous
bounds. Our main tool is the Garsia-Rodemich-Rumsey Lemma.
1. Introduction
Imagine that at time t = 0 an oil slick on the surface of the ocean covers
the set X and that each oil particle moves randomly according to a random
differential equation or a stochastic differential equation. Let &(z) be the
location of the particle at time t 2 0 which started at z E X at time 0. It is
of considerable practical importance to predict some characteristics of the
random set &(X) := { &(z), z E X}. We regard the particles as passive
tracers, which means that we assume they are being carried by the fluid
without interacting with the fluid or with other particles. This assumption
is rather unrealistic for oil particles but is in good agreement with reality
224
225
for light pollutants like dust. It has been conjectured by R. Carmona and
Y. Sinai3 that under reasonable assumptions, the diameter of the set &(X)
will grow linearly in t . Proving the conjecture consists in showing that
the set will grow at most linearly, i. e. in giving an upper bound for the
linear growth rate, and that it grows at least linearly, i. e. that it has a non
trivial linear lower bound. A linear upper bound was proved for a certain
class of stochastic flows by Cranston, Scheutzow and Steinsaltz' and by
the authors" using somewhat different methods. In section 3 we will use
yet another method - namely the Garsia-Rodemich-Rumsey Lemma (in
short: GRR) - t o prove an upper linear bound which in fact happens t o be
better than the previous ones. In addition, our proof seems to be shorter
and more transparent. We state the GRR-Lemma in the appendix. Lower
linear bounds have been proved under various assumptions by Cranston,
Scheutzow and Steinsaltz5, Scheutzow and Steinsaltz12 and Cranston and
Scheutzow4. We state a corresponding result for isotropic flows in section
4 but only provide an idea of the proof. The reader is referred to the
references for more general results and detailed proofs. Finally we state
some open problems.
time 0 time T
BL(r) = bii(rei), r 20
BN(r) = bii(rej), r 2 0, j # i,
where e k , Ic = 1 , .. . , d denotes the standard basis of Rd. Due to isotropy
the functions BL and BN do not depend on the choice of i and j . For later
reference, we introduce the strictly positive parameters
PL := -Bg(O),
PN := -BZ(O).
d X ( t ) = M ( d t ,X ( t ) ) , (1)
where M is an isotropic Brownian field. It wits shown by Kunitag, Theorem
4.5.1, that this equation does not only have a unique solution for every
initial condition X ( 0 ) = x E Rd but that it even generates a stochastic
flow of homeomorphisms, i. e. that there exists a family ( @ s l ) ~ j s , t < o oof
random homeomorphisms of Rd such that
condition X ( s ) = x.
0 The map ( s , t , x )H $,t(x) is continuous for all w E R.
with generator
A g ( z ) = (1 - BL(z))g”(z) + (d - 1) (1 - B N ( z )) S’(z),
where g E C.; Therefore pt satisfies the sde
where w
is a suitable standard Brownian motion
0 For each x E Rd, v E Rd\{O}
1 1
X := lim -logII(D&)(x)vII = - ( ( d - l),Ojv - P L ) a. s. (3)
t’cc t 2
X is called top Lyapunov exponent of the flow.
where
where A0 = 2c2d d- A ) .
For an isotropic Brownian flow with top Lyapunov exponent X 2 0 we get
the result above with
+
Proof. Choose E > 0 and ro > 0 such that logN(X, r ) 5 (A E ) log for 5
all 0 < r _< ro. Further, let y,T > 0 satisfy e-yT _< rg. Then N ( X , e - T T ) 5
exp{yT(A+e)}. Let Xi,i = 1 , . . . , N ( X ,e-TT) be compact sets of diameter
at most e-YT which cover X and choose arbitrary points xi E Xi.Define
-
x := {Xi, i = 1,.. . , N ( X ,C ' T ) } .
For K > 0 we have
+
P{ sup ll$t(x) - zll 2 KT b for some x E X } 5 S1
O<t<T
+S2,
where
S1 := exp{yT(A + E ) } maxP{ sup ll&(x) - 511 2 KT - eCYT}
sEX Wt<T
230
and
S 2 := exp{yT(A + E)} maxP{ sup diam($t(Xi)) 2 b } .
O<t<T
Our aim is to identify the infimum k over all r; for which there exists some
y > 0 and E > 0 such that the upper bounds of both S1 and S, above decay
to zero exponentially fast as T --+ 00. A simple Borel-Cantelli argument
will then show that k is indeed an upper bound for the linear growth
rate. Observing (6) we get k = B +
A m , where YO is the infimum
of all y > 0 for which there exists some E > 0 such that 5’2 decays to 0
exponentially fast as T .+ 00. Rather than identifying 70, we will instead
provide some yo 2 70.Then
K : = B + A ~
will turn out to be an upper bound for the linear growth rate.
We will estimate SZusing the Lemma of Garsia-Rodemich-Rumsey (see
Lemma 5.1).
Define
c := -T(1
U2
2
+ 6).
We have
and
Therefore
where
I := "Texp {/ s } d t .
0
232
We have
Define
U = U ( y , 6 ) :=
y-2fiJ72’(CGJ if y 2 a2ci(1 6) +
-a%( 1 6) + otherwise.
Then
1 5 2 exp{-UT) (1 + 6 ~ 2 0 2 4 1 6+) T )
Assuming betT 2 121 we have
P(ZT 2 b) 5 P ( e x p { {-} 7T Cd 2 g)
<P (v > -
Using Chebyshev’s inequality, we obtain
P ( ~ 2Tb) < E V Cd
8pexp{ - 121
‘(i.g(E))2}.
4C
Using we have
1
lim sup - log S2 5 y(A + E) - 2dy + N2
+ (t2026 - (5 + w2
T-oo 2a2(1 + 6) <
for some E > 0 provided that
(i) t + U > 0;
(ii) 5 + A 2 0;
(iii) Ay 2dy + A>2 - (5 + w2
+ (E2026 2+(1 + 6) < O.
-
If
2a2d(d - A)
A 2 A ’
then it is easy to check that
70:= A + a 2 A + Jo4A2 + 2A02A,
60 := (70- A)2d
- 1’
a27oA2
Eo := -U(yo, 60)
233
satisfy (i)-(iii) above, provided “>” and ‘(<” are replaced by “,” and “5”
in (i) and (iii) respectively, and that 60 > 0. Further, it is easy to see that
yo is greater or equal than the infimum of all y’sfor which there exist 6 > 0
and ‘$ such that (i)-(iii) are satisfied. If on the other hand -
~
2a2d(d - A)
A< A ’
then it is again easy to check that
A
so := -
a2d
+ 1,
‘$0 := -U(Yo, 60)
satisfy (i)-(iii) above, provided “>” and “<” are replaced by “2”and “5”
in (i) and (iii) respectively. Further, it is easy to see that yo is greater or
equal than the infimum of all y’s for which there exist 6 > 0 and E such
that (i)-(iii) are satisfied.
Therefore, for each €0 > 0 we have
Since the two events in (10) are disjoint Theorem 4.1 says that for any
subset X as above one of the following two cases will occur almost surely:
either the diameter will grow to infinity with a t least linear speed c* or
the diameter will shrink to zero. Even if the top Lyapunov exponent X is
negative, linear growth will occur with strictly positive probability.
Remark 4.1. It is easy to see that Theorem 4.1 will no longer hold if
we either allow the set X to be finite or if d = 1. In the first case the
diameter of 4 t ( X ) equals the maximum of the distance of a finite number
of (correlated) Brownian motions in Rdwhich grows at most like a constant
times (t log log t)'/2.In the second case the compact set X c R is contained
in a compact interval [a,b ] , and hence diam (& ( 2 ) 5
) &(b) - & ( a ) , which
again grows at most like a constant times (t loglogt)1/2.
* *
St St
t=O 1 coordinate t=l 1 coordinate
= x1 = x1 V yl.
E (+:(~)lFo)
Further it is plausible (and true) that there exists some p > 0 (not depend-
ing on the particular choice of z and y) such that
E( 4 m v 4ZY)IFo) =E ( 4 x 4 + (4XY) - 4 : ( 4 ) + I F o )
2 x1v y' +p a. s.
5. Open Problems
In this section we assume that (q&)t?o is an isotropic Brownian flow which
has a nonnegative top Lyapunov exponent A. We list some open problems.
Is it true that for any (reasonable) nontrivial compact subset XC
Rd the limit
1
lim - diam((bT(X))
T-cc T
exists almost surely? If so, is it deterministic?
If the answer to both questions is yes, does this limit depend on the
set X (e.g. on its dimension A)? Since our upper bound depends
on A we conjecture that the linear growth rate will depend on A in
general.
Let X be a curve in Rd of finite length LO > 0, and let LT be the
length of the curve ~ T ( X )How
. does LT grow as T + oo? I t seems
reasonable to conjecture that l i m $ l o g L ~ = X almost surely but
we conjecture that LT will grow faster, namely that
1
lirn - log LT = X
T-oo T
+-
PL
2
almost surely.
T-too T
1 log LT 5 lim sup -
X 5 lim inf -
T-m T
1 log LT 5 X -
+
PL
2
a. s.
t-,
X)n B
lim P {w : &(u, # 0} = 1
If such a function I exists, then it will take the value +00 for suf-
ficiently large values of y by Theorem 3.1. The following simple
observation shows that if such a function I exists, then I ( y ) 2 y2/2
for all y 2 0: using Chebychev's inequality and Fubini's theorem,
we get for any E > 0
Since
Appendix
We state the following lemma which is originally due to Garsia, Rodemich
and Rumsey and which we briefly refer to as the GRR-Lemma. A proof
(of a more general version) can be found in Arnold and Imkeller'.
then we have
References
1. L. Arnold and P. Imkeller, Stratonovich calculus with spatial parameters and
anticipative problems in multiplicative ergodic theory, Stoch. Proc. Appl. 62,
19-54 (1996).
2. P. Baxendale and T. Harris, Isotropic stochastic flows, Ann. Probab. 14,1155-
1179 (1986).
3. R. Carmona and F. Cerou, Transport b y incompressible random velocityfields:
simulations and mathematical conjectures, in: Stochastic partial differential
equations: six perspectives, eds. R. Carmona and B. Rozovskii, AMS, 1999.
4. M. Cranston and M. Scheutzow, Dispersion rates under finite mode Kol-
mogorov flows, Ann. Appl. Probab., 12,511-532 (2002).
5. M. Cranston, M. Scheutzow and D. Steinsaltz, Linear expansion of isotropic
Brownian flows, Electron. Commun. Probab. 4, 91-101 (1999).
6. M. Cranston, M. Scheutzow and D. Steinsaltz, Linear bounds for stochastic
dispersion, Ann. Probab. 28, 1852-1869 (2000).
7. G. Dimitroff, forthcoming Ph. D. thesis, Technische Universitiit Berlin.
8. J. Hoffmann-J~rgensen, Probability with a view toward statistics, Vol. II,
Chapman & Hall, 1994.
9. H. Kunita, Stochastic flows and stochastic differential equations, Cambridge
University Press, 1990.
10. M. Ledoux and M. Talagrand, Probability in Banach spaces, Springer, 1991.
11. H. Lisei and M. Scheutzow, Linear bounds and Gaussian tails in a stochastic
dispersion model, Stochastics and Dynamics 1,389-403 (2001).
12. M. Scheutzow and D. Steinsaltz, Chasing balls through martingale fields,
Ann. Probab. 30, 2046-2080 (2002).
STOCHASTIC BURGERS EQUATION IN D-DIMENSIONS -
A ONE-DIMENSIONAL ANALYSIS: HOT AND COOL
CAUSTICS AND INTERMITTENCE OF STOCHASTIC
TURBULENCE
1. I n t r o d u c t i o n
Burgers equation has been used to model the large scale structure of space-
time (Shandarin and Zeldovich and in a noisy environment in studies of
turbulence ( E, Khanin, Maze1 and Sinai Here we develop some related
results.
We begin by giving a brief account of the results of Davies, Truman and
> ~Wt
Z h a ~ ~Let . be a B M ( R ) process on the probability space 3,P ) (a,
with
lE[WtW,] = (s A t ) .
Consider the stochastic viscous Burgers equation for up = u p ( z , t ) , IC E Rd,
t > 0.
239
240
VP(Z,0) = VSO(Z) ,
Wt being white noise, pz coefficient of viscosity. We are interested in
Burgulence, that is the advent of discontinuities in
v’(z, t) = lim v ~ ( zt), .
P V
where
A ( X ( O ) ,Z, t ) = inf A [ X ],
X(S)
X(t)=x
A [ X ]= 2-1
t
X 2 ( s ) ds - 1 t
c ( X ( s ) )ds -E 1 t
k , ( X ( s ) )dW,
This gives the minimal entropy solution of the Burgers equation 12. Set
+
d [ X ]:= A [ X ] S o ( X ( 0 ) ).
Then necessary conditions for X to be an extremiser of A are :
dSt + (2-1 +
(VStj2 c(z)) dt + &(x) dWt =0 ,
St=o(.)= So(.) .
Definition 1.1. We define the stochastic wavefront Wt in x by {x :
S ( Z , t) = O}.
24 1
For small p , the heat equation solution u p switches continuously from being
exponentially large to small as we cross the wavefront. up can also switch
discontinuously.
Define the classical flow m a p Q S : Rd 4 Rd by
X(s)=@,q1X ]
UO(5, t ) = d&@;lz )
Definition 1.2.
Detax(t)
8x0
1 8x0
so=zo(z,t)
=0 Caustic in x, Ct.
where for i = 1 , 2 , . . . , n
t ) = So(x9(z1
SA("Cl +
t ) ) A(sA(z, t ) ,z, t ) I
and Oi is an asymptotic series in p2.
When zo(x,t)is unique, t < T ( w ) , the asymptotic series for up =
vp(x, t ) is given for each integer m 2 0 by
242
w p (x,t ) = cm
j =O
p2jvj (x,t ) - p2V lnE{ exp ( -
t
V . um(y;, t - s) ds
Here
Uj(X,t) = V S j ( 2 , t ),
Sj satisfying
j =O
yo” = x .
Here, we see as p - 0, the leading term
V P ( Z ,t ) - VSO(Z, t ) +O ( 2 ),
where SOis the solution of the Hamilton-Jacobi equation, which minimises
the action A.
When @T’{x} = {xA(x,t ) ,xg(x,t ) ,. . . , $(x, t ) } , there is a similar
asymptotic series 64 for the ith term in the series. Since
S(x,t)
= min Sh(x,t)
z=1,2, ...,n
,
we define the zero level surface by
H,“ = {x : ~ h ( r ct, ) = o , for some i} ,
where H: includes the wavefront. The dominant term for wo(x,t)comes
from the minimising 20(x,t)(assumed unique) and we obtain the corre-
sponding Burgers velocity field
VO(X, t ) = & q ’ z = &tS,(x,t)
Two x;(x,t)’scan coalesce a.nd disappear as we cross the caustic. When
this corresponds to the minimiser jumping, up=o has a jump discontinuity
and we say that this part of the caustic is cool.
243
cusp
3t 2 1
-zo - - .
2 t
Ht” : z ( z 0 , t )= s(ii
Y(X0,t)=
Figure 1. Multiplicity of
I(x,t) =
.I
R
G(zo)eiF(xO, a ? t ) / P zdzo ,G E C,-(R) 7
where i = G.
Consider the graph of the phase function F,,t(so) = F ( z o , z , t )as z
crosses the caustic. (D,F” > 0 in neighbourhood of (l).)
244
2 coalesce at the
Begin moving x in di- point of inflection.
New ZO(Z,t ) here.
rection n. Zo(x,t) is the global
minimiser of Fz,t(.) .
i). Zo(x,t ) jumps from position (1) to position (2). This causes u p and
up to change discontinuously as we cross the caustic.
ii). This only happens when a point of inflexion is the global minimiser
of &(.I.
iii). Some parts of the caustic (the cool parts) will be jump discontinu-
ities in wo and uo.Coalescing 2x0’sis associated with level surfaces of
Hamilton’s function having cusps on the caustic. I t is therefore im-
portant to know when Ht has a cusp on the caustic. We investigate
this in the next section.
I I4
\
\
,, ,,
I
, I
where
X ( s ) = -Vc(X(s)) ,
D@t(zo)= ( I + tV2So(zo)) .
The following elementary identity is the key to the free case
# 0 , so that xo i s
P r o p o s i t i o n 2.1. A s s u m e that [(I+tV2So(xo))VSo(xo)I
not a singular point of @;'Ht. T h e n @t(xo)can only be a generalised cusp,
if @t(xo) E Ct, the caustic. Moreover, if z = @txoE @t(@;lCt n @;'Ht),
x will indeed be a generalised cusp of the level surface.
Q = 1 , 2 , . . ., d ,
-(xo,po,t)
dA = X(t).% - X,(O) 7 a! = 1 , 2 , . . ., d
ax,. ax,.
Now let
Lemma 2.3. The random classical flow map has Frechet derivative a s .
Theorem 2.2.
Let x E cusp(^^) = { x E at , ( x o )# o } .
(@;lCt n @ ; ' H ~ ) , x = a t x O n
Then in 3 dimensions in the stochastic case, T,, the tangent space t o the
level surface at x , is at most one-dimensional.
The above explains the geometry of level surfaces of the Hamilton Jacobi
function. We know that u p changes dramatically as we cross Cusp(Ct n H t )
in the cool region. What about discontinuities in up as p O? Let us now N
y& = y k ( y , y o1 , y o2 , . . ., yyo'-',t) , r = d , d - l , d - 2 , . . . , 2 .
Yy," = Y~(Y,Y~,Y~,...,Yod-l,t)
250
Then
,
a). T(yA,af y, t ) = 0 and Equations (2) * y = atyo
8Yo
af
ii). Equations (2) and 7 ( y A 1y1t ) = - a2f 2(YILYlt) = 0 *
dY0 (aY; 1
y = a t y o is such that the number of solutions yo of this equation
changes.
Lemma 2.4.
where the last term is f”(yh, y, t ) and the first ( d - 1) terms are non-zero
as above.
25 1
Here the picture de- Here 2 a:'s coalesce, say Here the point of inflec-
forms as we move in di- UA-~ and uA. tion at (1) has disap-
rection n. a n1 P l ( y , t ) = uA(g,t) = peared.
a : ( y , t ) , a repeated
root.
Because the value f (a:(y, t ) ,y, t ) < mini=1,2,...,,-2 f (ai(y, t ) ,y, t ) the
disappearing root = a; is the minimising one. So the minimiser
jumps from (1) to (2). Hence wo is discontinuous and uo is exponentially
discontinuous. Hence the function fy,t(yi) = f (yi , y, t ) gives a complete
analysis of the discontinuities. A similar analysis may be given if you only
have local reducibility. This explains how to analyse hot and cool parts of
the caustic.
x = = o and xg
@txo s f{x,t)(xA) =x
i(., 1 t) .
xo,
W h e n x E Ct, the random caustic, let f[x,t,(xh)= 0 have the repeated root
xi = xg(x,t), Let X H xt(X) be a parameterisation of Ct, X E R, such
that X = A0 corresponds t o a cusp o n the caustic, or a point o n the caustic
where the Burgers velocity field is zero or orthogonal t o the caustic. T h e n
<
the processes f o r stochastic turbulence at x t ( X 0 ) are given by
<C(t) = f(z,(x,,,t)(.;;(.t(XO),t)) - c,
for c E R.
Remark 3.1. The ( 0 processes are just the stochastic action evaluated at
the relevant points on the caustic and their inverse images. Similar results
hold in &dimensions and for more general noise.
#{A E R : f(xc,(x,,t,(.2;(xt(X),t)) = c } .
Differentiating our last equation with respect to X gives X = Xo and
< c ( t ) = 0.
The above suggests the nomenclature for the three kinds of turbulence
- cusp, zero and orthogonal turbulence. We expect orthogonal turbulence
to be the most important. Similar results hold in higher dimensions.
dv + (u.V) dt = -VC(Z) dt - ) ,
~ V k ( 2dWt
253
O , dW9
dXi"(zo, S) = -VC(X'(ZO,s ) )ds - E V ~ ( X ' ( Z s)) ,
with X'(z0,O) = zo and Xi"(xo,O)= VSo(zo),0 < s < t . Let X o ( z ,s) =
@:xo satisfy the deterministic ( E = 0) version of the above equation and let
4 be given by Bi, = { X , " ( u ) , X ! ( s ) }8(s - u),the product of the Poisson
bracket { } and the Heaviside function 8.
Let
for s E [O, t]. This is the first term in, the perturbation expansion for X ' .
and
In particular,
and
as E \ 0 in probability.
254
It is not difficult to prove from the above that the pre-caustic surface of
the stochastic mechanics converges to the pre-caustic surface of classical
mechanics as E \ 0 in probability. Caustic surfaces are stable in probability.
What about the stability of level surfaces of the Hamilton-Jacobi function?
We can prove:
2
Theorem 4.2. Let @, be the minimiser of 2-1 Ji l&sx~l ds+So(@txo) -
s," c(@:xo)d s - E J i
k(@zxo)dW,, satisfying @zxo = x for almost all w E R,
with corresponding minimum S E ( x , t ) .Then we have for almost all w ER
So(x,
t)- EL t
k(@;xo)dW, I SE(x,
t ) 5 So(x,
t )- E
Finally, if we assume there exists a unique z o for fixed t and x such that
Qtxo = x, then the first approximation is
rt
S'(Z,t ) = S 0 ( z t, ) - E
lo k ( @ . , ~ odW,
) + O(E) ,
where So(x, t ) is Hamilton's principal function for the path X o ( x o ,s). Sim-
ilar results hold for xh(z,t ) and corresponding S'.
5. Some Applications
We give two elementary results illustrating the kind of applications now
accessible.
where
As expected
(;)
Moreover, additionally
= Qt (;:) - f’(X0) = 0 and yo
2
= y - txo
( 5 ,y ) E ct * f ” ( Z 0 ) =0 .
Analysis of the graph o f f ( . ) yields the hot and cool parts of swallowtail, as
shown in Figure 6, where
= ( -t5(3+ 8v@)
18000
-EL t
Wsds,--+
2t
1 t3 ( 9 - 4)
450
and
K = (-&-EL t W s d r , - - ~ + ~. )
2t 50
f(xo)= Z; -
t 4
-xi + -(1+ 2 t y ) - -+ -
502 x2
2 2t t 2t
We consider the roots of f’(x0)for the following two cases.
Case 1 : y < -2t1 or y > --%l &.
3
+
Since (x,y ) E Ct we know f’(x0) = 0 has only one solution, namely the
repeated solution x2;(( z, y ) , t ) . Thus f (zo) has only one stationary point
which is a point of inflection and so one side of this part of CL is cool.
Case 2 : -%1 5 y 5 -& + &.
We adopt the labelling scheme for the caustic shown in Figure 6. On branch
(A), f’(zo) = 0 will have one solution which is repeated and as in Case 1
one side of this part of Ct is cool.
256
= ( -t5(3 +8&)
18000
1
, --
2t
+ 450
257
Proof. <O is the result of carrying out the above analysis of orthogonal
turbulence. The remainder follows from the argument below. 0
Then with probability one there exists a sequence of times (a,) with a, / 00
such that Y,,,
= 0 for every n.
h ( t ) := ( 2 t l n l n t ) i
258
then
h(tn)-lWrt, + f(r) 7
uniformly over r in [ O , 1 ] .
We show that for each w with t , = t n ( w ) we have h(t,)-2t;1yt, + $.
Let us consider each of the terms that comprise the stochastic process y t ( w ) .
i) . a&h(tn)-2t,lWt, + 0.
ii) .
iii).
Taking
the above notation for the globally reducible deterministic map, wit,*
X: = @;zo, and with X: = z ; ( z , t ) ,z = X: = zt(Xo), z; the repeated
root vector. When Xf = 0, there is a very simple result for the small noise
stochastic turbulence process. (There are numerous examples with X: = 0
in the free case.) Let the corresponding processes be <
for stochastic tur- <:
bulence a t the cusp on the deterministic caustic, z t ( X 0 ) . These are simple
deterministic functions coming from the reduced classical action, ft’,t),)A.(
x; = zE(z, t ) ,z = z t ( X 0 ) .
where z.~(zt(Xo),t ) is the repeated root vector evaluated at zt(X0) the cusp
o n the deterministic caustic, (2 the reduced classical action at zt(Xo) and
.;;(zt(Xo), t ) .
sin(wl(r - t ) )0 aW, -
260
M W ) =
w2
sin(2w2t) cosec2(wlt){sin(cjit).Y(cr,) + w1 cos(w1t)I2
1 2
+ecosec(wlt)Rt(w) - -a,wl sin(2wlt) - c ,
4
where &(w) is a stochastic process well defined for all t.
As t 4 we have cosec2(wlt) + m. Let { t k } denote an increasing
sequence a t which cosec2(wltk) = 00,then limt,t, [t = -m if 4&cSd
d’(4
>0
but limt+tk [t = +m if s ’ n ( 2 w z t k ) < 0. However, we can find an infinite
g” (at)
increasing subsequence {tk,} such that It is continuous on (tk, , tk,+l) and
sgn ( s i n ( 2 ~ Z t k ~ ,
=) - sgn (~in(2w2tk,+~)
so that limt->t, <t successively switches between plus and minus infinity.
Hence, by continuity and the intermediate value theorem, there will exist
an increasing sequence {t3} with t, /” co a t which st,
= 0 almost surely.
We remark that the above argument fails if
is the same for all k E Z+. This will only be the case if % = 2n7r, namely
w2 = nwl, for some n E Z.
Remark 5.2. This means that t h e stochastic turbulence at cusp zt(Xo) will
be intermittent as long as xL(xt(Xo),t ) is t h e minimising critical repeated
root.
Acknowledgement
It is a pleasure for one of us (AT) to acknowledge helpful conversations with
Professor Costas Dafermos (Brown), Professor Mark Freidlin (Maryland)
a n d Professor Oleg Smolyanov (Moscow).
References
1. S. F. Shandarin and Ya. B Zeldovich, The large-scale structure of the uni-
verse: turbulence, intermittency, structures in a self gravitating medium,
Rev. Mod. Phys. 6,185-220 (1989).
2. W.E, K. Khanin, A. Maze1 and Ya Sinai, Invariant measures for Burgers
equation with stochastic forcing, Ann. Math. 151, 877-960 (2000).
3. I.M. Davies, A. Truman and Huaizhong Zhao. Stochastic heat and Burgers
equations and their singularities I - geometrical properties, J. Math. Phys.
43,3293-3328 (2002).
4. I.M. Davies, A. Truman and Huaizhong Zhao. Stochastic heat and Burg-
ers equations and their singularities - geometrical and analytical p r o p
erties (the fish and the butterfly, and why.), UWS MRRS preprint,
http://www.ma.utexas.edu/mp~arc-bin/mpa?yn=Ol-45, 2001.
5 . A. Truman and H.Z. Zhao. Stochastic Burgers’ equations and their semi
classical expansions, Comm. Math Phys. 194,231-248 (1998).
6. H. Kunita. “Stochastic Differential Equations and Stochastic Flows of
Homeomorphisms” in Stochastic Analysis and Applications, edited by M.
A. Pinsky, Advances in Probability and Related Topics (Marcel Dekker,
New York, 1984), Vol. 7,pp. 269 - 291.
7. C. Reynolds. On the polynomial swallowtail and cusp singularities of
stochastic Burgers equation, PhD thesis, University of Wales, Swansea, 2002.
8. V. N. Kolokoltsov, R. L. Schilling, A. E. Tyukov. Estimates for multiple
stochastic integrals and stochastic Hamilton-Jacobi equations, to appear in
Revista Matematica Iberoamericana.
9. V. N . Kolokoltsov, A. E. Tyukov. Small time and semiclassical asymptotics
for stochastic heat equation driven by LBvy noise, Stoch. Stoch. Rep. 75,
1-38 (2003).
10. K.D. Elworthy, A. Truman and H.Z. Zhao. Stochastic elementary formulae
on caustics I: One dimensional linear heat equations, UWS MRRS preprint.
262
ARMEN SHIRIKYAN
Laboratoire de Mathe'matiques
Universite' de Paris-Sud X I , Bhtiment 425
91405 Orsay Cedex, France
E-mail: [email protected]
We establish a version of the strong law of large numbers (SLLN) for mixing-type
Markov chains and apply it to a class of random dynamical systems with additive
noise. The result obtained implies the SLLN for solutions of the 2D Navier-Stokes
system and the complex Ginzburg-Landau equation perturbed by a non-degenerate
random force.
1. Introduction
We study the 2D Navier-Stokes (NS) system perturbed by an external
random force:
li - Au + (u,
0 ) u + Vp = q(t,z), div u = 0, z E D, (1)
u = 0, x E dD. (2)
Here D c R2 is a bounded domain with smooth boundary d D and 7 is a
random process of the form
k=l
I ~ f ( u ( k )-
) (f,p)l I conSte+, k 2 1, (4)
263
264
k-I
We refer the reader to 12,’ for a detailed discussion of the results obtained
in this direction.
The aim of this article is to derive the SLLN (5) from the mixing prop-
erty (4) without using the coupling of solutions and to estimate the rate of
convergence. To this end, we establish a simple version of SLLN for a class
of Markov chains (Section 2) and show that it applies to the problem in
question (Section 3). We note that the result of this paper remains valid for
the 2D NS system perturbed by a random force white in time and smooth
in the space variables.
Notation
Let H be a real Hilbert space with norm 1) . 1). We shall use the following
notation:
B H ( R )is the ball in H of radius R > 0 centred a t zero;
B ( H ) is the Bore1 a-algebra in H ;
P ( H ) is the family of probability measures on ( H ,B ( H ) ) ;
C ( H ) is the space of continuous functions f: H -+ R;
Cb(H)is the space of bounded functions f C ( H )endowed with the norm
llflloo := SUP lf(.>I.
uEH
C ( H ) is the space of Lipschitz-continuous functions f E Cb(H) with norm
Definition 2.1. We shall say that the family ('ilk, P,) is uniformly mixing
if it has a unique stationary measure p E P ( H ) and there is a continuous
function p : R+ + R+ and a sequence { Y k } of positive numbers such that,
for any f E C ( H ) and u E H,we have
Theorem 2.1. Let (Uk, P,) be a uniformly mixing family of Markov chains
in H such that
k=O
We note that Theorem 2.1 remains valid (with trivial modifications) for
Markov processes with continuous time. Moreover, under some additional
assumptions, one can take in (9) functionals f with polynomial growth at
infinity.
We also note that inequality (9) immediately implies the following esti-
mate:
where M ( w ) = D + 2 K ( w )g-'.
2.2. Proof of Theorem 2.1
Let us fix an arbitrary function f E L ( H ) and set
k-1
By the Markoa-property,
-
''' = k i B f 6 , it follows from (14)
and (16) that
1=1 1=2
268
where we used inequalities (13), (15) and the definition of m ( w ) and G,.
Since K = [m3+O], we see that, for 0 < T < 36,
E,KT 5 E,mT(3+O) 5 1 + a - r (c4
3+p) w4)IlfllL
5 1 + 3(& w-11)IlfllL.
The proof of Theorem 2.1 is complete.
3 . Applications
3.1. Dissipative P D E ’ s perturbed by a bounded kick force
Let H be a real Hilbert space with norm 11 . 11 and orthonormal base {ej}.
We consider the random dynamical system (RDS)
uk = s(uk-1) + qk, (17)
where S : H + H is a continuous operator such that S(0) = 0 and { q k }
is a sequence of i.i.d. random variables. As was explained in RDS of 8,9110,
the form (17) naturally arise in the study of dissipative PDE’s perturbed
by the random force (3), and in this case S is the time-one shift along
trajectories of the unperturbed equation. We assume that S satisfies the
following three conditions introduced in 8,10:
(A) For any R > T > 0 there are positive constants a = a ( R , r ) < 1
and C = C(R)and an integer no = no(R,r ) 2 1 such that
IIS(u1) - S(m)115 C ( R ) \ l U l - U Z I I for all ul, uz E B H ( R ) ,
IISn(u)ll 5 max{aIIu.II,r} for u E BH(R),n 2 no
(B) For any compact set K c H and any bounded set B c H there
is R > 0 such that the sets A k ( K , B ) defined recursively by the
formulas do(K,B) = B and d k ( K , B ) = S ( d k - l ( K , B ) ) K are +
contained in the ball B H ( R )for all k 2 0.
(C) For any R > 0 there is an integer N 2 1 such that
l l Q N ( s ( ~ 1-
) S(uz))llI illui - uzll for all ~ 1 , 2 1 2E B H ( R ) ,
qk = bjtjkej, (18)
j=1
269
j=1
H =
{u E L 2 ( D , R 2 :) divu = 0, (u,")IaD = 0},
where v is the unit normal to 6'D (see l5 for further details on the space H).
Let S : H 4 H be the time-one shift along trajectories of the NS system (l),
(2) with 7 z 0. Setting Uk = u ( k ,z), we obtain (17) (see ',lo for details).
Let (uk,pu) be the family of Markov chains associated with the
RDS (17). As is shown in '114, if the non-degeneracy condition (20) is
satisfied for N >> 1, then the family (uk,pu)has a unique stationary mea-
sure p , and (21) holds with p(d) = C , ( l + d ) , where C1 and p are positive
constants not depending on f, u,and k . Moreover, by Theorem 1.3 in ',
we have
References
1. J. Bricmont, A. Kupiainen, and R. Lefevere, Ergodicity of the 2D Navier-
Stokes equations with random forcing, Comm. Math. Phys. 224 (2001), 65-
81.
2. J. Bricmont, A. Kupiainen, and R. Lefevere, Exponential mixing for the
2D stochastic Navier-Stokes dynamics, Comm. Muth. Phys. 230 (2002),
no. 1, 87-132.
271
MARIAN SLODICKA
Department of Mathematical Analysis, Ghent University,
Galglaan 2, B-9000 Ghent, Belgium
E-mail: [email protected]. be, web page: http://cage.rug.ac. be/-ms
The aim of this paper is to present various mathematical models for wells. Special
attention is paid to a non-standard description using nonlocal boundary conditions
(BCs). We also develop numerical algorithms t o handle nonlocal BCs. The choice
of the appropriate model depends, of course, on concrete situation.
1. Introduction
Many ground-water hydrologists are interested in the determination of
water-table elevations resulting from inputs and outputs such as natural
replenishment, artificial recharges and pumping. Some of them are inter-
ested in the general flow pattern in the whole aquifer, other study the details
in a vicinity of a well. Here, wells represent inputs or outputs, which affect
flow in a soil matrix. These sinks/sources are concentrated, i.e., their di-
ameters are relatively small compared with the whole aquifer. This feature
makes the modeling more complicated. Of course, wells are not only used
in the ground-water hydrology, but also by oil extraction or soil venting,
which is used for soil remediation (for cleaning of unsaturated zone from
chlorinated hydrocarbons or other volatile organic compounds). The main
difference among all these applications are (a) the substance (water, oil,
gas) for which wells are used, (b) different geological conditions.
2. Point sources
Let us consider the steady-state case with a single extraction well with an
infinitely small diameter located at the origin. We suppose that our do-
main is infinite in all directions and we consider a homogeneous unconfined
aquifer with the conductivity KO.Then a fundamental solution of
-v . (KOVUO) = ss
272
273
uo(x)= rL 25KO
lnlxl
4.rrKoI 2 I
in 2D
in 3D.
This solution so far has not included any realistic BCs and it generates
drawdownsa everywhere. Further, the seepage face at the well is omitted
(1)
because of a negligible well radius. This is not realistic for a small vicinity
of the wellbore.
Method of images is a simple technique to create some basic BCs.
Adding imaginary wells to the real point sink at strategic locations allows
to generate infinitely long straight equipotentials or no-flow boundaries (cf.
'i2).For the analytical description of a single-phase flow caused by a single
extraction well for a perfectly layered subsurface we refer the reader to '.
Bounded domains. We consider a bounded domain R E Co?l in IRN
( N = 2,3) with boundary r = I'D U r N , where J?D has a positive measure.
We study
-V . ( K V u ) = SS in R
u=O o n r o
K V u . v = 0 on r N .
Problem (2) is linear, but the right-hand side does not belong t o the H-l(R)
(dual space to H1(R)), thus we cannot directly apply the theory of linear
elliptic equations. When the conductivity K is Holder continuous (with the
coefficient a , Q > 0 in 2D, a > in 3D) near the well, then one can use
the method of subtraction of singularities. Then ( 2 ) is rewritten in terms
of an new unknown function ii = u - U O , uo being defined by (1). The
reformulated problem will contain the right-hand side from Lz(R), due to
the Holder continuity of K near the origin.
The case when the conductivity is not Holder continuous is more diffi-
cult. Such a situation can appear, e.g., when a well is located at an interface
of two different layers, or there is a rock at the well tube. In such situa-
tion, we cannot suppose the regularity of K , thus the right-haad side of
the modified problem (after subtraction of singularities) will not belong to
the Lebesgue space. Nevertheless, one can overcome this using the so-called
very weak solution as it has been proposed in '. Here, the solution is defined
aPumping from a phreatic aquifer removes water from the void space leaving there a
certain quantity of water which is held against gravity. As a result, the watertable at
each point is lowered with respect to its initial position by a vertical distance called
drawdown.
274
bModels describing air-, water- or oil-pumping wells differ from each other.
275
R rN (impervious layer)
nR2&W(t>= 2nR
ID q .u -Q ,
where 0 denotes the saturation, K conductivity, p pressure, q the
(5)
mass flow, R the well radius, Q the discharge of the well. D is the
thickness of the aquifer. The Neumann, Dirichlet and Signorini
boundaries (see lo) are denoted by r N , r D , rs,respectively.
(D) Discharge Condition. It is assumed that a constant but un-
known pressure builds up on the well boundary such that the pre-
scribed discharge is obtained. Such a type of BC can be used for
active wells by soil venting. Here, the total discharge of the well is
given and one can assume that the pressure along the well tube is
constant.
We demonstrate this on the following study case:
v . (-KVU) = f in R
U = gD On r D
-KVU, Y = gN on r N
u = unknown constant on r, (6)
G(u)= L, (-KVu) . v = s ER
276
Thus, taking this value of Q for u,,we see that u, solves (6).
277
out without pumping. The pressure at the bottom of the well is unknown.
It varies in time depending on situation in the aquifer. One can measure
the total flux through the well tube, but we cannot expect that it will
remain constant. The total flux through the well clearly depends on the
pressure at the bottom of the well. This can be taken as a space constant
(along the well boundary at the bottom). In fact, this value can change in
time and it is a priori unknown. But the dependence of the total flux on
278
where q is the flux vector and u is for the outer normal vector at rn.
The derivation of a flow equation for water in a confined aquifer can be
found in 1 3 . We assume that the flow is governed by Darcy’s law. Thus,
the flow equation for a saturated flow reads as
8tP = AP +f in R
P=PD on r D
vp.v=o on rN
p = unknown space constant on r, (13)
in R,
The symbol lrnl denotes the measure of the boundary part rn.
Taking into
account (14) and (7), one can easily deduce the well-posedness of a weak
solution p t o the IBVP 1.
Throughout the paper we tacitly assume that the data-functions ap-
pearing in the problem setting are sufficiently smooth.
In that follows C, E and C, denote generic positive constants depending
only on the data, where E is a small one and C, is a large one.
schemes need to start close to the exact solution. This implicitly means that
the time step T is small. One can use the following linearization scheme,
which is robust and converges for any initial datum pi,^ ( k E N,p E V )
The problem (18) is linear and well-posed. This follows from the V -
ellipticity of the left-hand side and from the Lax-Milgramm lemma.
For a given time-index i we perform relaxation iterations for k =
1 , .. . , ki,maa: until the stopping criterion
Lemma 4.1. There exist positive constants COand TO such that for any
r 1. TO and for all k E N the following estimates hold:
Proof. (2) We subtract (20) from (18) and set p = p i & - u i . We get
28 1
Thus
2 lrnl 2 2
- +2 IrnI I I V ( P ~ , ~ - ui)llo,n+ L I I P ~- ,uiIIi,r,
IIpi,k - uiI10,n ~
T
2 (21)
5 L l l p i , k - l - uillo,r,, *
The generalized Friedrichs inequality implies that the following relation is
valid for any w E H1(R) (due to the fact that > 0)
ll'Ulll0,n 5 c ll~'UlIl0,n~ (22)
The combination of the trace inequality and (22) yields for some Co > 0
2 2
COI I P~,~ - utIIo,r, 5 coI I P ~ ,-~ uiIIo,r i 2 IrnI I I V ( P ~-, ~ui)11;,n. (23)
Now, we deduce from (21) and (23)
( L + CO)i h , k
2 2
- uiIlo,r, 5 L lIpi,k-l - ui))o,r,. (24)
This iterative relation gives rise to the following estimate
(ii) The desired result is a consequence of the part (i) and (21).
We point out that the choice of the time step r is free. Next, the
relaxation iterations can start from any starting datum from H1(R) and
they converge in the H1(R)-norm to a function u, E V , which is defined
by (20). Please note that u, = p ( t , ) . We stop the relaxation iterations if
l l p z , k t , m a z - P z , k t , m a z - i /lo,r, I7'. Moreover, we know from (24) that
IIv(Pi,k%,maz c
- ui)llo,n I ~ ~ P i , k z , m a s-
- lUiII o,r, I CTV1 (26)
which are valid for any i = 1,.. . , n.
Now, we derive suitable a priori estimates for ui.
Lemma 4.2. L e t q 2 1. We a s s u m e (19) for all i = 1 , .. . , n. T h e n
i=l i=l
take place f o r all m = 1 , .. . , n
Now, we set 'p = u i r and sum the relation up for i = 1 , .. . , m. W-e get
The lower bound for the left-hand side is (the last term is nonnegative)
1 cm
2 ( ~ ~ ~ m+l ~ ~ , n ui-ll/i,n +
i= 1
IIU~
- c
i=l
m
) .
I I ~ ~ ~ I I : , ~ T
Applying the Cauchy and Young inequalities and (25) to the right-hand
side we easily get the upper bound
283
Applying the Cauchy and Young inequalities and (25) to the right-hand
side of (28) we easily get the upper bound
m m
C&(1 + 3q-1)) + €Cll~uzll:,n 7 5 C&+ ll~~ill02,n
7,
i=l i=l
-
un(0)= uo, z,(t) = uil for t E (ti-l,ti].
Exactly in the same way we also define the step functions ji, and7, as well
as the piecewise linear function p,. Using this notation we rewrite (27) into
Proof. We subtract (16) from (29), set 'p = Ti, - p , integrate the equality
over the time interval (0, t ) and get
The last term on the left-hand side is nonnegative due to the monotonicity
of g. Further, Lemma 4.2(ii) implies
285
i”
The last term of (30) containing the function f can be estimated analo-
gously taking into account the properties of f . Therefore, we can write
t t
llUn(t) - P(t)ll&2 + J’0 llV(% - P)ll:,sl Ic (T + IIun - PIli$)
Proof. The desired result is a consequence of Lemma 4.3, (25) and (26)u
The error estimates from Theorem 4.1 fully correspond to the known
results for semilinear problems starting from po E H1(R). When starting
from more regular initial datum po E H2(R), assuming that po is compatible
with BCs and taking 7 2 2 , one can get the convergence rate 0 (r2).
286
References
1. H.M. Haitjema. Analytic Element Modeling of Groundwater Flow. Academic
Press, Inc, New York, 1995.
2. D.J. Wilson. Modeling of Insitu Techniques for Treatment of Contaminated
Soils: Soil Vapor Extraction, Sparging, and Bzoventing. Technomic Publish-
ing Company, Inc., Lancester-Basel, 1995.
3. R. Nieuwenhuizen, W. Zijl, and Van Veldhuizen. Flow pattern analysis for a
well defined by point sinks. Bansport Porous Med., 21:209-223, 1995.
4. M. SlodiEka. Finite elements in modeling of flow in porous media: How to
describe wells. Acta Mathematica Universitatis Comenianae, LXVII( 1):197-
214, 1998.
5 . M. SlodiEka. Mathematical treatment of point sources in a flow through
porous media governed by Darcy's law. Tramport Porous Med., 28(1):51-
67, 1997.
6. I.S. Papadopulos and H.H. Cooper Jr. Drawdown in a well of large diameter.
Water Resources Research, 3(1):241-244, 1967.
7. S.P. Neumann and P.A. Witherspoon. Analysis of nonsteady flow with a free
surface using the finite element method. Water Resources Research, 7(3):611-
623, 1971.
8. E.A. Sudicky, A.J.A. Unger, and S. Lacombe. A noniterative technique for the
direct implementation of well bore boundary conditions in three dimensional
heterogeneous formations. Water Resources Research, 31 (2):411-415, 1995.
9. S. Schumacher, M. Sloditka, and U. Jaekel. Well modeling and estimation of
hydraulic parameters. Computat. Geosci., 1(3,4):317-331, 1997.
10. C. Baiocchi and A. Capelo. Variational and Quasivariational Inequalities.
John Wiley and Sons, Chichester . New York, 1984.
11. M. SlodiEka and R. Van Keer. A nonlinear elliptic equation with a nonlocal
boundary condition solved by linearization. International Journal of Applied
Mathematics, 6(1):1-22, 2001.
12. M. SlodiEka. Error estimates of a n efficient linearization scheme for a non-
linear elliptic problem with a nonlocal boundary condition. M 2 A N , Math.
Model. Numer. Anal., 35(4):691-711, 2001.
13. J.W. Delleur (Editor-in Chief). The handbook of groundwater engineering.
Springer, Heidelberg, 1998.
14. M. SlodiEka. A robust and efficient linearization scheme for doubly nonlinear
and degenerate parabolic problems arising in flow in porous media. S I A M J .
Sci. Comput., 23(5):1593-1614, 2002.
STOCHASTIC CASCADES APPLIED TO THE
NAVIER-STOKES EQUATIONS
1. Introduction
The study of properties of solutions of the Navier Stokes equations remains
one of the most notable problems in mathematics. While a substantial body
of literature is available on this subject, see e.g. Ternamlo and Galdi4, the
recent work of LeJan and Sznitman' has opened new opportunities for anal-
ysis and a novel application of branching processes. Indeed, in their work,
LeJan and Sznitman obtained a representation of the Fourier transform
of the solution of the Navier-Stokes equations as an expected value of a
multiplicative functional defined on a branching random walk. This repre-
sentation uses exponential random variables with means depending on wave
number in a way naturally related to the equation. The distribution of the
offsprings at each branching is on the other hand determined by a kernel
conveniently introduced to use the quadratic nonlinearity of the equation.
Three basic extensions of this approach are presented in this paper.
First, the notion of majorizing kernels is introduced in order to analyze and
control the regularity of solutions of the Navier-Stokes equations. While
the solutions determined in the work of LeJan and Sznitman have to be
understood in a weak sense, we show that it is possible to use an appropriate
majorizing kernel to maintain or improve the regularity of the solutions.
287
288
where
290
<
Notice that for each fixed with h * h(<)# 0, the convolution h * h(<)
simply normalizes the product h(ql)h(772) to be a probability kernel on
+
the set 71 772 = (. In particular, while a majorizing kernel need not be
integrable, it is required that the convolution h * h(<)be finite for each
< E R3\{O}. It is then possible to show the existence of globally defined
solutions of the (FNS) equations, the regularity of which depends on the
particular majorizing kernel applied as follows.
Introduce the Banach space Fh,T as the completion of the set
in the given norm. In the case h = ho this is the Besov type space intro-
duced by Cannone and Planchon2.
Note that considering h = hz with p < 1, from Proposition 2.1, the
Banach space corresponding to such a majorizing kernel contains initial
data 210 which are infinitely differentiable functions of compact support.
One of the main results obtained using majorizing kernels is the follow-
ing theorem.
ICo(<)l 5 Ch(<)ve-A’v.
In the case that h ( ( )= 1/1[12 G ho((),this result was obtained by LemariQ-
Rieusset’. However, the following proposition shows that there are majoriz-
ing kernels that exhibit a stronger singularity at the origin and decay slower
at infinity than ho(<).
lim G(w) = 00
w-iv
v = u,o~,~{i,
2}j = {el< 1 >, < 2 >, < 11 >, . . .}, (6)
where {1,2}O = {O}. Also let aV = n~o{l, 2) = (1, 2}N.
A stochastic model consistent with (3) is obtained by consideration of
a multitype branching random walk of nonzero Fourier wavenumbers <,
<
thought of as particle types, as follows: A particle of type # 0 initially
at the root O holds for an exponentially distributed length of time So with
holding time parameter A(<) = V I J ~ ~ ; i.e. ESo =&. When this exponen-
tial clock rings, a coin KO is tossed and either with probability the event
[ K O = 01 occurs and the particle is terminated, or with probability f one
has [ K O = 11 and the particle is replaced by two offspring particles of types
71, r/2 selected from the set 71 + <
r/z = according t o the probability kernel
This process is repeated independently for the particle types 71,772 rooted
at the vertices < 1 >, < 2 >, respectively.
Now, recalling (4), for given initial data and forcings xo(<)and p(<,t ) ,
( # 0, t 2 0, define a functional X(O, t ) by the following stochastic recursion:
xo(Ee), if so 2 t
where 71
X(O,t) =
+ 72
{
=
cp(t - S O , < ) ! if SO< t , K O = 0,
m((o)X(<1 >,t - So) @,cs X(< 2 >, t - So) else
are distributed according to K c s ( d q l , d 7 2 ) and
7<2> are the trees defined by re-rooting a t the vertices < 1 >, < 2 > of
r<1>,
292
where
Ivl-1
B~ = 0, B, = C s,,~, e # v E v.
j=O
(9)
where
k=O
and let
) ][R, > t Vv E {u E 7 e ( t ) : 1
A,(@,t ) = [lvl 5 n Vv E ~ ( t n ~ =1 n}],
with l [ n ;19,t]being the indicator of the event A,(O, t ) .
Prop 2.1. Let
uk(t,t) = h(E)Xk(t,t)
= h ( < ) E d l [ k E, tIX(I9,4 )
and denote by & ( t ,t ) the Fourier transform of the kth iterate of the itera-
tion scheme defined in (12). Then W k ( < , t ) = file([, t ) .
is given by
r Nt 1
294
Bv(t)= z + c
Ivl-1
j =O
XVlj(TVlj)+ Xv(t - RV).
Finally, let M ( y e ( 2 ,t ) )= max{ IvI : v E ye} and let 1[k;z, t] the indicator
of the event [ M ( y o ( < , t )5) k].
Let
~ ( zt ),= Ex [n [uo(Bv(t))l
l[k : z, ti] . (15)
It follows that
u ( z , t )= lim u k ( z , t )
k+cc
On the other hand, consideration of the Fourier transform of the KPP
equation leads after a simple integration to the integral equation
where
1
A(<) = 1 + ~ 1 < 1 2 .
Proceeding as done with the Navier Stokes equations, scale (16) by
l/h(<) to obtain
where
295
where < 1 >, < 2 > are re-rooted trees a t vertices of types <<I>,Ec2> re-
+
spectively and the distribution of types is given on v~ q 2 = & by
Note that the only difference with the recursive functional corresponding
t o the Navier-Stokes equations is the node operation which for the K P P
equations is standard multiplication.
Using the strong Markov property it follows that the solution of (16) is
given by
fi(<, t ) t ) ) ].
h(t)EIX(TB(t,
provided the expected value is finite. The analogue of a majorizing kernel
for the KPP equation is given by
1
( h * h)(E) I B(1+ 21<I2)h(<).
where
that
Using the same branching Brownian process used by McKean for the K P P
equation as done in section 3, it is possible to define a recursive multiplica-
tive functional such that the solution of (17) is obtained as an expected
value. Indeed, let
Acknowledgments
The work presented here is joint work with R. Bhattacharya, L. Chen, S.
Dobson, R. Guenther, C. Orum and E. Waymire and it is partially funded
by US NSF Grant 0073958.
References
1. R. N. Bhattacharya, L. Chen, S. Dobson, R. B. Guenther, C. Orum, M. Os-
siander, E. Thomann, and E. C. Waymire, Majorizing Kernels & Stochastic
Cascades With Applications To Incompressible Navier-Stokes Equations. To
appear in Transactions of the AMS.
2. Cannone, M. and F. P1anchon:On the regularity of the bilinear term for
solutions to the incompressible Navier-Stokes equations Revista Matema’tica
Iberoamericana 16 1-16, (2000).
3. Larry Chen, Scott Dobson, Ronald Guenther, Chris Orum, Mina Ossiander,
Enrique Thomann, Edward Waymire. On ItB’s Complex Measure Condition
For a Feynman-Kac Formula. To appear in IMS Lecture-Notes Monographs
Series, Papers in Honor of Rabi Bhattacharya, eds. K. Athreya, M. Majumdar,
M. Puri, E. Waymire.
4. G. Galdi, “An Introduction to the Mathematical Theory of the Navier-Stokes
Equations” Vol 1 and 2. Springer Tracts in Natural Philosophy, Vol 38 and
39. Springer 1994.
5. Gundy, R and M.L. Silverstein: “On a probabilistic interpretation for the
Riesz Transforms” in Functional analysis i n Markov processes, Lecture Notes
in Mathematics, 923, 199-203. Springer 1982.
6. LeJan, Y. and A.S. Sznitman: Stochastic cascades and 3-dimensional Navier-
Stokes equations, Prob. Theory and Rel. Fields 109 343-366, (1997).
7. LemariB-Rieusset, P.G. Une remarque sur l’analyticitB des soutions milds des
Bquations de Navier-Stokes dans R3,C.R. Acad. Sci. Paris, t.330, SBrie 1,
183-186, (2000).
8. Kato, T.: Strong L p solutions of the Navier-Stokes equations in Rm with
applications to weak solutions, Math. Z., 187 471-480, (1984).
9. H. McKean, Applications of Brownian motion to the equation of Kolmogorov,
Petrovskii and Piskunov. Comm. Pure and Applied Math. Vol 28, 323-331,
(1975).
10. R. Temam, “Navier-Stokes equations and nonlinear functional analysis”.
SIAM 1995.
STOCHASTIC BURGERS EQUATION WITH LEVY
SPACE-TIME WHITE NOISE
1. Introduction
This paper is mainly concerned with the Cauchy problem for the following
stochastic Burgers equation
on the given domain [0, m ) x R with L2 initial condition, where Ft,z is the so-
called Lkvy space-time white noise consisting of Gaussian space-time white
noise (i.e. a Brownian sheet on [0, m) x R) and Poisson space-time white
noise (see 52 for the definition). There has recently been increasing interest
in solving stochastic partial differential equations with non-Gaussian white
noise (see, e.g. Bertoin5i6, Giraud18, Winke143, Mueller31, M ~ t n i kand ~~
Shlesinger et a136 and references therein).
In particular, Gaussian white noise driven parabolic SPDEs have been
intensively studied (see e.g. W a l ~ and
h ~ references
~ therein). SPDEs driven
by Poisson white noise are less well known and were first investigated in
298
299
Albeverio et all. Let us also mention that Saint Loubert BiB 35 formulated a
parabolic SPDE driven by a Poisson random measure in a different way from
Albeverio et all and he obtained very.interesting results on the existence of
the unique solution. Moreover, parabolic SPDEs driven by L6vy space-time
white noise are studied in Applebaum and Wu' and besides the existence
of the unique solution, the flow property of the system obtained is also
discussed. While very interesting studies of heat equations driven by Q-
stable LBvy noise have been carried out by Mueller31 and by Mytnik3'.
On the other hand, as is well-known (see e.g. Burgers7), the Burgers
equation
au +--=-
- la(u2) d'u
at 2 ax 8x2
has been used extensively, under the name of Burgers turbulence, to model
a variety of physical phenomena where shock creation is an important in-
gredient. The solution to Burgers equation is then called Burgers turbu-
lent fluid flow. In recent years there appears to be a great interest to
investigate Burgers turbulence in the presence of random forces, that is,
to study stochastic Burgers equations with (Gaussian) white noise as ran-
dom forces and/or with random inital data, see e.g. Bertini et a13, Bertini
and Giacomin4, Bertoin5y6, Da Prato et als, Da Prato and Gatarekg, Da
Prato and Zabczykl', Davies et all', E et all', Giraud17>18,Gyongy and
Nualart2', Holden et a123,24,Kifer27, Le6n et a12', Sinai37>38,Tribe and
Z a b o r o n ~ k iTruman
~~, and Zhao40,41,Winke143. Burgers equation has,also
been used to study efficient stock markets, see Hodges and Carverhill" and
references cited there.
One of the main investigations of Burgers equation is based on the
intriguing connection between the (nonlinear) Burgers equation and the
somehow simpler linear heat equaiton, via the celebrated Hopf-Cole trans-
formation. This technique can be still adapted to stochastic Burgers equa-
tion with additive Gaussian white noise (see e.g. Bertini et a13, Holden
et a123124),but it is no longer available in the case of stochastic Burg-
ers equations driven by more general Gaussian white noise (for instance,
multiplicative Gaussian space-time white noise). Another method is used
successfully, e.g. in Da Prato et a18, Da Prato and Gatarekg, Da Prato and
Zabczyk" and Gyongy and Nualart" (here we just mention a few refer-
ences), to study the mild solutions to stochastic Burgers equations driven
by Gaussian space-time white noise.
In this paper we introduce a stochastic Burgers equation driven by L6vy
space-time white noise which generalizes all stochastic Burgers equations
with white noises considered in the literature mentioned above. We will
300
E [ N ( A ,B , .)] = p ( A ) u ( B ) , A E I ,B E B ( U ) .
Moreover, N is nothing but a Poisson random measure on the Cartesian
product measure space ( E x U , & x B ( U ) ,p @ v ) as formulated e.g. in Ikeda
and Watanabe25. Hence, by a similar argument to that of Theorem 1.8.1of
Ikeda and Watanabe25, we have the following existence result for Poisson
white noise, namely, given any a-finite measure p on the measurable space
(E,&), there exists a Poisson white noise N on ( E ,&, p ) with mean measure
E[N(A,B , .)] = p ( A ) v ( B )A, E E , B E B ( U ) . In fact, N can be constructed
as follows
7 n( W )
N ( A ,B , W ) := CC
,EN j=1
l ( A n E , , ) x ( B n l r n ) ( [ ~ ) ( ~ ) ) l ( w E R : a n ( W ) ~ i ~ ( ~(2)
)
P{w E R : vn(w)= k } =
e-p(En)V(Un) [p(~~)v( u,)]
k! , k E N U {0}u {a};
Thus, given any a-finite measure p on ( E ,&) and any a-finite measure Y
on (U,B( U ) ) ,there is always a Poisson random measure N on the product
measure space ( E x U,& x B ( U ) ,p @ v )which can be constructed in the above
manner. We call such a N canonical Poisson random measure associated
with p and v.
302
for (t,x) E [O, m) x Rd. We call Nt,z Poisson space-time white noise.
In the sequel in this section, we will take the measurable space (I?,&)
in Definition 2.1 to be a product space ( [ O , m) x E , B([0,m)) x &) where
( E , & ) in the latter is a Lusin space. Let p be a o-finite measure
on ( E , & ) (note that from the next section onwards, ( E , & , p )will be
taken to be (EX,B(B), d x ) ) , then there exists a Poisson white noise N on
([O, m ) x E , B([0,m)) x E , d t 8 p ) with mean measure
E{N([O,t] x A, B , .)} = t p ( A ) v ( B )(,t ,A, B ) E [0,03) x & x B(U).
Let { 3 t } t E [ 0 , 0 0 ) be a right continuous increasing family of sub a-algebras
of 3,each containing all P-null sets of 3,such that the Poisson white noise
N has the property that (i) N([O,t ] x A, B , . ) : R + N U (0) U {co} is
Ft/P(N U (0) U {co})-measurable V ( t ,A, B ) E [0,m) x & x B ( U ) and (ii)
{N([O,t + s] x A, .) - N([O,t ]x A , . ) } ~ > o , ( A , B ) E E ~ Bis( uindependent
) of 3 t
for any t 2 0, where P(NU (0) U {m}) is the power set of N U (0) U {co}.
(For instance, we may directly take
3 t := m ( { N ( [ O ,t] x A, B , .) : ( A ,B ) E E x B ( U ) } )V N , t E [0, a)
where N denotes the totality of P-null sets of F.)
In what follows, let us set up related stochastic integrals by following
the procedure of Section 11.3 of Ikeda and Watanabe25 (see also Applebaum
and Wu')). First of all, for those integrands f : [0,00)x E x U x R + R
which are {Ft}-predictable and satisfy
M ( t , A, B , U ) := N([O,t], A, B , w ) - t p ( A ) v ( B ) (3)
for any (t, A, B ) E [0,co)x E x B(U) with p(A)v(B)< co. Obviously,
E[M(t,A, B , ,)I = 0
and
, B , .)I2) = tP(A)V(B).
E ( [ M ( tA, (4)
For any {Ft}-predictable integrand f : [0, co) x E x U x R --+ R which
satisfies
Moreover, stochastic integrals with respect t o M are also well defined for
{&}-predictable integrands f satisfying
( g &)u(t,x,w ) + + g [ u 2 ( tx,1 w ) l =
- a(t,x,u(t,x,w ) ) +
+qt, 2,u(t,5 , w))Ft,,(w) I ( t ,5 , w ) E (0, 03) x R x R (9)
u(O,x,w ) = uo(x,w ) , (x,w ) E R x R
305
where a , b : [0, 00) x R x R ---f R are measurable and the initial condition uo
is .&measurable, and F is an Lkvy space-time white noise, which includes
terms not only controlled by a Gaussian space-time white noise but also by
a Poisson space-time white noise, so that in fact the noise we shall consider
has a formal structure similar to that of a Lkvy process:
for t > 0; and Go(x,z ) = S(z - 2). We will need the following facts:
(i) J
, G t ( z ,z ) d z = 1, Jw[Gt(zlz ) I 2 d t = (27rt)-? , Vt E [0, oo),Vz E R ;
(ii) Gt(z,t)= Gt(z,z) , t E [0, oo), 2,t E R;
(iii) J, G t ( x ,z’)G,(z‘, z)dz’ = Gt+s(x,z ) , s, t E [0, co), z, z E R ;
(iv) Vm,n E N U {0}, there exist some constants K , C > 0 such that
For the property (iv), cf. e.g., Friedman14, or Ladyzhenskaya et a128. Based
on (iv), we have the following particular estimates which will be used later
on
and
1 a u y , ., w )
t ,U ) = (G * U O ( . , w ) ) ( t ,Z)
~ ( Z, + {G * [-z az
+a(*,., U ( . , * , w ) ) -k b(., ., u(.,
.,w))F.,.(U)]>(t, 7 (14)
then u satisfies (9). Furthermore, (14) is formally equivalent to the follow-
ing stochastic integral equation
since from (10) we have the following (heuristic) derivation for the second
term in the right hand side of (14)
308
Pt P
x M ( d s ,dz, d y , w )
aHere and in the sequel, by “const.” we mean a generic positive constant whose value
might vary from line to line.
310
Idt, 2, .)I2 + / U
Ih(t,2, z ; Y)l2Y(dY)I L2(x)+ const.1zI2 (20)
and Lipschitz conditions
Iq(t, 5 , a )- 4 ( t ,2, z2)12 + If ( t ,x,z1) - f (t,5, .2)12
I +
[L3(x) const.(lzl12 Iz21')]1z1 - z2I2+ (21)
(22)
hold f o r all ( t , x ) E [O,T]x R and z , z l , z 2 E R. Then f o r every Fo-
measurable uo : R x R -+ R with EJw(luo(x,.)12)dz < co, there exists a
unique local solution u to Equation (17) with the following property
We need some preparation before the proof to Theorem 3.1. For any
fixed n E N,let the mapping 7rn : L 2 ( R )4 B, := {u E L2(R) : 11ullL2 5 n }
be defined via
II7rn(u)IILz 5 TL.
Moreover, it is clear that the norm
and
312
in particular,
Proof of Theorem 3.1 We will carry out the proof by the following three
steps.
Step 1 Suppose that u : [O,T]x R x f2 -+ R is an L2(R)-valued, {Ft}-
adapted, c&dl&gprocess. For any fixed n E N,set
with
and
313
On the other hand, by our Proposition 2.2 (Fubini’s theorem) and It6’s
isometry property for stochastic integrals with respect to (both continuous
and c&dl&g)martingales, we have
and
Step 2 Now let X > 0 be arbitarily fixed. For any L’(IR)-valued, {&}-
adapted, c&dl&gprocess u : [O,T] x IR x R 4 IR with initial condition
u(0, z, w ) = UO(Z, w ) , we define
315
Then ( B ,(1. Ilx) is a Banach space. Now Vu E B , Ju is well defined and for
any fixed t E [O, TI
Thus
<_ const.
00
( t i + t')e-xtdt fi
= const.[-A
-3
+2~-31
2
I const.(A-$ +A - ~ ) < 00
Therefore we define
{4t7X,W) : ( t , z , w )E [0,7,(W)) x JR x 0)
is a local solution to Equation (17).
Finally, for the uniqueness of the local solution to Equation (17), sup-
pose that there are two local solutions u and v to Equation (17). Then u
and v must satisfy Equation (23) for any fixed n E N. On the other hand,
by the uniqueness of solution to Equation (23), we get
U(t,X,W) =W(t,X,W), V(t,X,W) E [O, 7,(w)) x R x a.
319
--
Remark 3.3. In the case M 0, Equation (17) becomes a Burgers equa-
tion with Gaussian (space-time) white noise. Unique global L2 solutions
are obtained by Gyongy and Nualart2' in the whole space line and by Da
Prato, Debussche and Teman in Da Prato et al' (see also Da Prato and
Zabczyk'') in bounded space intervals. Their methods depended critically
on some uniform estimates which employed Burkholder's inequalities for
continuous martingales. We observe that we are unable to follow this route
herein as the corresponding inequalities for cadlag martingales (see e.g Ja-
cod and Shiryaev26) do not behave so nicely.
Finally let us consider the flow property of the local solution to Equation
(17) starting with an L2 function as the initial condition. We refer to the
references Fujiwara and Kunita15 and Fujiwara" for investigations of LBvy
flows associated with (ordinary) stochastic differential equations of jump
type. For ( T , t , w)E { ( T , t , w)E [0,TI x [0,TI x R : T F t < T,(w) I TIw E
Q}, and 'p E L2(R), we define
Proposition 3.4.
x M ( d s ,d z , d y , W ) , P - U.S.
(33) is obtained by a straightforward derivation using the (usual) Fubini's
theorem (see, e.g. Theorem 7.8 in R ~ d i n ~integration
~), by parts, Theorem
2.6 of W a l ~ h our
~ ~ Proposition
, 2.2, property (iii) of the Green's function
G, and our equality (32). Q.E.D.
321
Acknowledgements
We thank Ian M. Davies for the great help in setting of the manuscript.
References
1. S. Albeverio, J.-L. Wu, T.-S. Zhang, Parabolic SPDEs driven by Poisson white
noise, Stoch. Proc. Appl. 74 (1998), 21-36.
2. D. Applebaum, J.-L. Wu, Stochastic partial differential equations driven by
LBvy space-time white noise, Random Oper. Stochastic Equations 8 (2000),
245-259.
3. L. Bertini, N. Cancrini and G. Jona-Lasinio, The stochastic Burgers equations,
Commun. Math. Phys., 165 (1994), 211-232.
4. L. Bertini, G. Giacomin, Stochastic Burgers equations and KPZ equations
from particle systems, Commun. Math. Phys., 183 (1997), 571-607.
5. J. Bertoin, The inviscid Burgers equation with Brownian initial velocity, Com-
mun. Math. Phys., 193 (1998), 397-406.
6. J. Bertoin, Structure of shocks in Burgers turbulence with stable noise initial
data, Commun. Math. Phys., 203 (1999), 729-741.
7. J.M. Burgers, The Nonlinear Diffusion Equations, Reidel, Dordrecht, 1974.
8. G. Da Prato, A. Debussche, R. Teman, Stochastic Burgers equation, NoDEA,
1 (1994), 389-402.
9. G. Da Prato, D. Gatarek, Stochastic Burgers equation with correlated noise,
Stochastics Stochastics Rep., 52 (1995), 29-41.
10. G. Da Prato, J. Zabczyk, Ergodicity for Infinite Dimensional Systems. LMS
Lect. Notes 229, Cambridge Univ. Press, 1996.
11. I.M. Davies, A. Truman, D. Williams and H.-Z. Zhao, Singularities of stochas-
tic heat and Burgers equations and intermittence of stochastic turbulence,
Swansea preprint, 2001.
12. W. E, K. Khanin, A. Maze1 and Ya Sinai, Invariant measures for Burgers
equations with stoachastic forcing, A n n . Math., 151 (2000), 877-960.
13. S.N. Ethier and T.G. Kurtz, Marlcov Processes: Characterization and Con-
vergence. John Wiley and Sons, New Yrok, 1986.
14. A. Friedman, Partial Differential Equations of Parabolic Type. Prentice-Hall
Inc., Englewood Cliffs, NJ, 1964.
15. T. Fujiwara, H. Kunita, Stochastic differential equations of jump type on
manifolds and LBvy processes in diffeomorphisms group, J . Math. Kyoto Univ.
25 (1985), 71-106.
16. T. Fujiwara, Stochastic differential equations of jump type on manifolds and
Lkvy flows, J. Math. Kyoto Univ. 31 (1991), 99-119.
17. C. Giraud, Genealogy of shocks in Burgers turbulence with white noise initial
velocity, Commun. Math. Phys. 223 (2001), 67-86.
18. C. Giraud, On regular points in Burgers turbulence with stalbe noise initial
data, A n n . Inst. H. Poincare' Probab. Statist. 38 (2002), 229-251.
19. I. Gyongy, Existence and uniqueness results for semilinear stochastic partial
differential equations, Sdoch. Proc. Appl. 73 (1998), 271-299.
20. I. Gyongy, D. Nualart, On the stochastic Burgers equation in the real line,
A n n . Prob., 27 (1999), 782-802.
322
41. A. Truman and H.-Z. Zhm, Stochastic Burgers equation and their semiclas-
sical expansions, Commun. Math. Phys., 194 (1998), 231-348.
42. J.B. Walsh, An introduction to stochastic partial differential equations. In:
Ecole d 'e'te'de Probabilite's d e St. Flour X I V , pp. 266-439, Lect. Notes in Math.
1180,Springer-Verlag, Berlin, 1986.
43. M. Winkel, Burgers turbulence initialized by a regenerative impulse, Stoch.
Proc. Appl., 93 (2001), 241-268.
A COMPARISON THEOREM FOR SOLUTIONS OF
BACKWARD STOCHASTIC DIFFERENTIAL EQUATIONS
WITH TWO REFLECTING BARRIERS AND ITS
APPLICATIONS
T.S.ZHANG
Department of Mathematics, University of Manchester, Oxford Road,
Manchester M13 9PL, England.
In this note, we prove a comparison result for the solutions of backward stochastic
differential equations with two reflecting barrier processes. The result is then
applied t o obtain an existence result for solutionsof a backward SDE with reflecting
barrier processes under weak assumptions on the coefficients.
1. Introduction
The notion of backward stochastic differential equation was introduced by
Pardoux and Peng' in (1990), there they obtained existence and unique-
ness of adapted solutions under suitable conditions on the coefficients and
terminal random variables. Certain backward SDE were also independently
used by Duffie and Epstein in (1992) to study stochastic differential utili-
ties in economics models. This subject has attracted a lot of attention and
has developed rapidly in recent years, which is partly due to the applica-
tions found in the theory of partial differential equations and mathematical
finance, etc. See 2,5,8 and references therein.
We particularly mention the paper by J.Cvitanid and I.Karatzasl, which
motivates our work here. In this paper they studied backward stochastic
differential equations with two reflecting barrier processes and obtained
existence and uniqueness of solutions under various conditions. They also
proved that the solution of a backward SDE with two reflecting barrier
processes is the value function of certain Dynkin game. If the coefficients
depend also on the state variable, the Lipschitz condition is required.
The aim of this paper is to prove a comparison theorem for solutions of
backward stochastic differential equations with two reflecting barrier pro-
324
325
cesses. The theorem is applied to obtain a new existence result for solutions
of a backward SDE with reflecting barrier processes.
L ( t ) 5 U ( t ) , YO 5 t 5 T and L ( T ) 5 5 5 U ( T ) a s (1)
These two processes will serve as two reflecting barriers. Consider the
backward SDE with two reflecting barrier processes.
+ +
d X ( t ) = - f ( t , w , X ( t ) ) d t - d K 1 ( t ) dK2(t) Y'(t)dB(t) (2)
As in we introduce the following
-
Definition 2.1. We say that ( X ,Y,K 1 ,K 2 ) : X : [0,TI x R 4 R , Y :
[0, T ] x R 4 Rd, and K 1 ,K 2 : [0,T ] x R R is a solution of the backward
SDE (2) with reflecting barriers U ( . ) ,L ( . ) and terminal condition if the <
following holds
(i) X , Y,K 1 and K 2 are continuous and F-progressive,
(ii) K1(t),K 2 ( t ) t, 2 0 are increasing with K1(0) = K 2 ( 0 )= 0,
(iii)
X(t)= <+ 1 T
f(S, w,X(s))ds + K 1 ( T )- K 1 ( t )
- ( K 2 ( T )- K 2 ( t ) )- /
0
T
Y'(s)dB(s), 0 It IT, (3)
(iv) LJt) 5 X ( t ) I U ( t ) , 0 5 t 5 T,
(v) J, ( X ( t )- L ( t ) ) d K l ( t )= J?(U(t) - X ( t ) ) d K 2 ( t )= 0,
almost surely.
Theorem 2.2. Assume one of fl and f2, say f2, satisfies a Lipschitz
condition in x uniformly w.r.t. (s,w ) , i.e.,
lf2(s,w7x1) -f2(s,w,x2)1 I c151 - 5 2 1
326
This gives
327
where we have used the fact that 5 &, and (iv), (v) in the definition 2.1.
Letting n + 00,we obtain
1
E[(Xl(S) - XZ(S))+]I M--c"(T - t)" (9)
n!
Ifn(t,w,.) B CnI.-Yyl
- fn(t,w,Y)I
for some constant c, and that for any fixed ( t ,w ) E [0, T ] x R, f n ( t ,w ,).
converges to f ( t ,w , x) uniformly on bounded intervals in R. It is easy to
see that such a sequence f,, n 2 1 exists under our assumptions on f . It
'
was proved in that when f is repaced by fn the equation (2) has a unique
solution. Let us denote it by ( X n ( t ) ,Y,(t), KA(t),K:(t)). By Theorem 2.1,
we have
X , ( t ) 2 X2(t) 2 X 3 ( t ) 2 . . . 2 X n ( t ) 2 . . . 2 X M ( t ) , a.s (12)
where X M stands for the solution to the backward SDE with reflecting bar-
riers corresponding to the coefficient given by the lower bound of f (t,w ,.).
329
This shows that the sequence { X n ( t ) } n > lhas a limit, which we will denote
by k(t).Observe that for fixed ( t , w ) , k n ( t )lies in a compact interval of
R for all n 2 1. Hence,
fn(t,w,X n ( t ) )- f ( t ,w,X ( t ) )
lim
n-+0 Ell T
( f n ( 4w , X n ( t ) )- f ( 4 w, m)2 dtl =0 (14)
Let ( X ,Y,K 1 ,K 2 ) be the unique solution to the backward SDE with re-
flecting barriers with g ( t , w)= f ( t ,w,X ( t ) )replacing f ( t , w , .). Then
X(t)=I + 1 T
+
f ( s , w ,X ( s ) ) d s K 1 ( T )- K 1 ( t )
- ( K 2 ( T )- K 2 ( t ) )- /t
T
Y ' ( s ) d B ( s ) , b'0 5 t 5 T , (15)
L ( t ) 5 X ( t ) 5 U ( t ) , "0 5 t 5 T , (16)
l T ( X ( t )- L ( t ) ) d K l ( t=
)
LT ( U ( t )- X ( t ) ) d K 2 ( t )= 0 , (17)
almost surely.
Next we show that X n ( t ) converges to X ( t ) ,and hence, X ( t ) = X ( t ) . From
the It6 rule,
+2 LT ( X n ( S )- X ( S ) ) ( Y ' ( S ) - YL(s))dB(s-
)
LT lY'(s) - Y;(s)l2ds.
- 2 L T ( X n ( S )- X ( s ) ) d K ' ( s )
= -2 1T
( X n ( S )- L ( s ) ) d K l ( s )+ 2
4'
( X ( s )- L ( s ) ) d K l ( s )
= -2 1T
( X n ( S )- L ( s ) ) d K l ( s )
10
References
1. J.Cvitanic and LKaratzas: Backward Stochastic Differential Equations With
Reflection and Dynkin Games, The Annals of Probability 21:4 (1996) 2024-
2056.
2. Duffie.D and Epstein.L, Stochastic differential utility, Econometrica 60
(1992) 353-394.
3. N.Ikeda and Swatanable, Stochastic Differential Equations and Diffusion
Processes, North-Holland and Kodansha,Amsterdam and Tokyo, 1981.
4. LKaratzas and S.E.Shreve, Brownian Motion and Stochastic Calculus,2nd
ed. Springer, New York, 1991.
5. N.E1 Karoui, C.Kapoudjian, E.Pardoux, S.Peng and M.C.Quenez, Re-
flected solutions of beckward SDEs and related obstacle problems for PEDs.
Preprint,l995.
6. E. Pardoux and S. Peng, Adapted solutions of backward stochastic differen-
tial equations, Systems and Control Letters 14 (1990) 55-61.
7. T.S.Zhang, On the strong solution of one-dimensional stochastic differential
equations with reflecting boundary, Stochastic Processes and Their Applica-
tions 50 (1994) 135-147.
8. T.S.Zhang, On the quasi-everywhere existence of the local time of the solution
of a stochastic differential equation, Potential Analysis 5 (1996) 231-240.
BURGERS EQUATION AND THE WKB-LANGER
ASYMPTOTIC L2 APPROXIMATION OF
EIGENFUNCTIONS AND THEIR DERIVATIVES
A. TRUMAN
Department of Mathematics, University of Wales Swansea, Singleton Park,
Swansea SA2 8PP, UK
Email: A . [email protected]
H.Z. ZHAO
Department of Mathematical Sciences, Loughborough University, Loughborough,
L E l l 3 T U , UK
Email: [email protected]
1. Introduction
Consider the following second order Schrodinger operator on IR1
1 a2
H = --h2-
2 ax2
+V(x)
We are interested in studying approximate eigenfunctions and their deriva-
tives by using our Hamilton Jacobi methods, extending the results in Tru-
man and Zhao 21 to more general potentials than linear harmonic oscillator
ones.
It is well known that the WKB method leads to approximate eigen-
functions. However, the WKB solution is not finite at the turning points.
Langer therefore introduced Bessel functions to study the wave functions
332
333
4 sin( is," b(z)dz- :)&(z)h is the dominant term at these zeros, not
b y )
1 cos(i s," b(z)dz - 2). However, the first term in the WKB-Langer
bz (T)
semiclassical asymptotic expansion approximates both the eigenfunction
and its derivative in L2(R), the natural norm of quantum mechanics.
There will be many applications of the results of this paper, e.g. quan-
tum probability, quantum tunnelling problems (Jona-Lasinio, Martinelli
and Scoppola ', Simon l a ) etc. The Hopf-Cole transformation applied in
this setting should also yield new results for Burgers equation and its in-
viscid limit. This is not so surprising since the Hamilton-Jacobi continuity
equations first arose in this context. We do not include these results here
due to the length of this paper. But we aim to study some of these appli-
cations in our future publications.
d2
h2-Q;(z)
dx2
+ Q 2 ( ~ ) Q k =( z0).
Here Q2(z)= 2(E - V(z)). Let the real function be defined by
First we consider the simplest case where we have only one turning point
in order to obtain some useful formulae. In the next section we will apply
these formulae to more complicated but practical and useful situations. Let
T denote a turning point and for simplicity here we suppose r is a simple
zero of Q 2 ( z ) .In this section we consider the case when V ( z )> E if z < T
and V ( x )< E if x > 7- for a T E R1.Then Q ( x )is simply
G-dy) = V - l ( y ) ,
Gn(y) = G k - l ( ~ )n, = 0 , 1 , 2 , . . .,
and
& E ( Y ) = @E(V-'(y)).
=
1-
--Go(y)
3
+ -31. (2(y-E))+
Ly Gl(y)(2(y- E ) ) Q d y .
G-l(x) = lx &x)dx.
337
and
G(Y) = c.(v-'(Y)).
Similar to (2.4), using the integration by parts formula and induction prin-
, each n = 0 , 1 , 2 , . . .,
ciple, we have for y E [ E ,V ( T - ) ]and
Applying (2.8) and (2.9) to al, a2,. . . , we have the smoothness of aj and
pj in z and Lipschitz continuity in E . The rest of the lemma follows
from van der Corput's fundamental theorem on asymptotic series and its
consequence on asymptotic series with a parameter (Theorem 4.1 and P391,
van der Corput 2z). See also 0lverl3>l4.
We follow Langer to define wave functions using Bessel functions
9110111
near the turning point. Alternatively one can use Airy functions (Olver
13,14,15,Heading 6 ) . Define
K"4
1
= C-ESJ-&) 6 +c+[+J+(,)
6 (2.10)
=I e$i(Jz," b(z)dx)i
x(C-J-g(-fi J,'b(z)dz)
(J,"b(z)dz)i
+ C + J + ( - i J , ' b ( s ) d z ) ) , if x < 7 ,
d2 1
-Qo(x)
dx2
+
jgQ2(x)Qo(x)= O(x)Qo(x). (2.12)
j=1 j=1
For x < T , recall some standard results about Bessel functions (c.f. e.g.
Whittaker and Watson 25)
= (lT 1 2 a
b(z)dz)s(-C-K1(-
7r
3 ;lr b(x)dz)- (C+ - C - ) I ; ( i s,'b(z)ds)).
Notice that I ; (
z E [T-,T - is].
JZr b(z)dz) -
exp{ JzT b(z)dz}when fi is small for
In order to have a L2(R) solution, we have to choose
C- = C+. Recall for z E [ T - , T - $61,when ti is small,
(2.16)
(2.17)
with
j=1 j=1
and
(2.19)
Notice that L1 and L2 in (2.19) for J-; are the same as those in (2.17) for
J; .
+
So for z E [T $5, T + ] , using the same C- and C+ as in the region
xE [T-,T - $71, i.e. C- = C+ = &,we have from (2.10) and (2.11), as
340
h is small.
Qo(x) = (2.20)
1 1
( x ) d x- -.rr)LF(x)
4
+ b(x)dx - -.rr)Lg(x)).
4
It is noted that the term sin( sTx b(x)dx- ;7r)Lg(x) = O ( h ) should not be
neglected.
For simplicity in the following we only consider bound states where we
take C- = Cf = 6.So for r- 5 x < r,
it follows that
I t follows that
d
-K"z)
dx
So for T < rc 5 T + ,
(2.22)
+
with R f i ( z )= 1 O(h) and having an asymptotic expansion. It turns out
from (2.16), (2.21) and (2.23) that for z E [ T - , T - 461
(XI
(2.25)
where
F ( z , h) = M"z)B(x)+h-M
@'(x) A(%)
fL (x)T+b(z)Ryz)- A ( x ) = l+O(h).
@(x) h
It is obvious that P - ( x ) has an asymptotic expansion in powers of ti, say
Dc)
P - ( z , h) N 1 +=-pi)j$j(z), (2.26)
j=1
343
for some smooth and bounded functions $j on [ T - , T - ;6] which are com-
binations of aj and ,Bj and the asymptotic expansions of M h ( x ) and Rh(x).
In particular we have fked values of $j ( T - ) .
Note if we take
rv1 WI
A, = aj(z)h2jand Bm(z)= 1 + ,Bj(z)ti2j
and define
Then
For z E [T+ $S,T+], when ti is small, again from P362 in Whittaker and
Watson 2 5 ,
(2.27)
and
Similarly
(2.28)
Note that R1 and R2 in (2.28) are the same as the ones in (2.27). It
turns out that
344
Clearly we have
and
where
+
for some smooth and bounded function 4j on [r i6,r+].In particular we
have fixed values for $j (r+).
We study the WKB asymptotic expansion outside [r-,r f ] .From Theo-
rem 26.3 in Wasow 24, for x < r - :6, the Schrodinger equation (2.1) which
can be reduced to a system of 2-dimensional singular perturbed differential
equations possesses a solution of the form
(2.33)
345
(2.34)
+j(X) =
i 4 q - ) -
+
q + ( ~ + ) $+
j = 1,2,...,
J
;s,'
- b-~(y)~(b-3(y)$~-~(y))d y , < 7-,
ifx
y , > r+, (2.37)
: b - + ( y ) ~ ( b - q ( y ) ~ ~ - ~ ( y ) ) difx
00
~ * z ( z )N i ) Jz, > T + .
C ( ~ ~ ( z ) ( * h for (2.39)
j=O
*‘(z~) = b-+(z)exp{--
k LT b(y)dy} x ~ - ( z ) , (2.40)
+
where P-(x)= 1 O(fi) for small fi uniformly in z for z < 7-, and
9 = exp{-$i}Q+i + e x p ( $ i } W for z > T + , i.e.
(2.41)
where
N +
P l ( z ) 1 Cj”=1(-l)j$~j(z)h2j and P2(x) Cj”=1(-l)j$2j-l(z)h2j-1
N
+
and Pl(z)= 1 O(h2)and P2(z) = O ( h ) for small h uniformly in z for
+
z > T + . And for z near 7, 9 ’ ( z ) = Qh(z)A(z) * o ( z ) B ( z )as in Proposi-
tion 2.3. From our construction we know that 9 is smooth on R’. Similar
combinations were also used in Furry 4 , Heading and Berry t o exploit
the Stokes’ phenomenon in physics literature. We formulate a proposition.
347
9 + ( x )= b-$(x)exp{ -
:L b(y)dy}P+(x),
m
= QE,O(Z)(l+ X(-h)j$j(Z) + 0(hrn+l)) (3.1)
j=1
= Qk,,(x) + QE,o(z)0(hm+l),
with a uniform 0(hm+')for x E ( - w , T ~ ] . Hence Q;(x),Qk,,(x) and
Q;,,(x) are exponentially small for z < 71. For 71 5 x 5 T:, take
Langer's construction and apply Lemma 2.4,
Qk(4= Q E , O b ) % 4 + QL,,(x)A(x)
IYl
= QE,&)(l+ c
j=1
Pj(Z)h2j +U ( P + l ) ) (3.2)
rw1
+(hQk,o(x))( c
j=1
CYj(z)h2j-l + 0(hrn+l))
= Qk,,(z) + QE,o(x)O(II'"+').
For 71' < x < T;, set
(3.3)
= Qk,,(x) + 0(hrn+l),
where $ j are defined by (2.37) with $j(r:) derived from (3.2) as initial
conditions, i.e. $j(x) = $ j ( ~ ; ) f $ JTy+ b4(~/>A(b-+(y)$j-l(y))dy.On the
other hand we should also have
349
= Gk,,(z) + GE,o(x)o(hm+l).
For x 2 r z , take
' k k ( x )= b-i(z)exp{--
:1: b(y)dy} x p - ( x )
= !i&,o(x)(l+ c
j=1
m
$j(Z)(-h)j + O(hrn+l))
= Gk,,(z) + GE,O(z)O(hm+l)l
with a uniform O(hm+l)in z for x > 7-2'. Here +k(x) and G;,,(x) are
both exponentially small when x > r2$ is large. Here P- in (3.1) and P -
in (3.6) are defined as in Section 2. From Section 2, we know that 9 is
smooth for z E (-00,r;] and ?t is smooth for x E [r:, 00).
Remarks. (i) From Section 2, especially Remark (ii) following Proposi-
tion 2.6, Equation (2.1) also possesses a solution Q+ # L2(R) given by
(2.42), i e . 9 i ( x ) = b-*(x)exp{i szT1
b ( y ) d y } P + ( s ) , for x E ( - m l ~ , - ] .
350
The smooth extension of the solution to the whole space (-00, m) can be
done by using the same method as (3.1)-(3*6) and (3.8). The solution is
linearly independent of the L2 solution 9 given in (3.1)-(3.6) and (3.8).
Furthermore any solution 9 1 of (2.1) is of the f o r m 9 1 = c19- +c29'+ for
constants c1 and c2. But for a L2 solution 91,c2 = 0 is satisfied, whence
91 = c1Q. That means any L2 solution Q1 of (2.1) is linearly dependent
on 9, which is equivalent to the vanishing Wronskian determinant property
for any x ,
d d
-d@x 1 ( x ) 9 ( x ) - 9 1 ( 2 ) d- Q
x ( x ) = 0.
I n particular, we can choose c1 = 1/11911 so that 11Q111 = I. Therefore 9 is
the unique L2 solution up to normalization. Here we state our results for
the WKB-Langer solution 9 . One can give our results for the normalized
wave function if one likes.
(ii) The semi-classical WKB-Langer approximate solution ~ E , ~ ( isx )
given by the first m terms of the series in (3.1)-(3.6) in five different regions
respectively. Note
m
lim ~ E , ~ ( z )
ZTT; j=l
and
lim 9 ~ , ~ =
( x ) b(y)dy}P;(r;, fi)
ZIT;
from (2.26). These two limits are different, but the difference is 0(fim+')
as P - ( z , f i ) - PG(x,fi) = 0(hm+'), so are limxfT; hQk,m(z) and
limxLT; k9b,m(x).A similar remark applies for x = I-:, 72, I-;. However,
(-Q;(x))@;(x)
d - *;(x)-QE(x)
d -ii = 0.
dx dx (3.7)
We will see soon that quantization condition gives the exact eigenvalue E ,
i.e. { E ( N ,~ ) } N = o , I....
, Now we transform (3.7) to an explicit equation of
E. For this we first differentiate (3.3) and (3.4),
and
It is crucial here that the leading term in (3.8) has a different sign from
the leading term in (3.9). Substituting (3.3)-(3.4) and (3.8) and (3.9) into
(3.7) we obtain
= -H"x). (3.10)
352
Here the formula of H can be given explicitly if one wants to. We note
here that H " ( z ) is bounded for all z E [ T ~ , T ; ] uniformly in h. It turns
out that
7r
sin(; b(y)dy - -) = 4hHh(z). (3.11)
2
Recall that the Wronskian determinant (3.7) is vanishing for all z E [T:, T;]
is equivalent to that the Wronskian determinant (3.7) is vanishing at a
particular point ( e g . see Hartman 5). Therefore (3.11) is equivalent to
6 J2(E - V(y))dy = ( N
1
+ -)7rh
2
+ harcsin(4hHfi(M)),
for N = 0 , 1 , . . .. Here M E [T?, T;] is the minima of V(z). The solution
(3.12)
E = E ( N ,h) of the above equation gives the exact N-th eigenvalue of the
Schrodinger operator.
We take the first m terms in the asmptotics expansions of Q i ( z ) and
*;(x), denoted by 9;,,(x) and *;,,(x). We require the Wronskian de-
terminant vanishes a t x = M I i.e.
(y5,7r(4)*t,m(4
d - Q 5 , 7 J ( ~d) ~-ii M0.
Q E , , ( ~ ) / s ==
(3.13)
Similar to (3.7),
we will see that this gives discrete values
{E,(N, as follows. First we go through all the calculations
~ ) } N = o , J...
,
of (3.8)-(3.12) for Q;,(z) and *;,,(x), then we derive
7r
b(y)dy - -) = 4hHk(M). (3.14)
2
Similar t o H ( z ) , H k ( M ) is also bounded uniformly in ti. It is followed
from (3.14) that
1
72 ( E)
+ harcsin(4hHL(M)),
/
71( E)
/ 2 ( E - V(y))dy = ( N + -)7rh
2
for N = 0,1, . . .. The solution of the above equation depends on m, denoted
(3.15)
Here TI(&) and 72(E0) are the only two roots of V(a:)= Eo. We will
analyze the solution in ascending order, setting E = Eo(N,h) for N =
0,1, . . .. The following result is uniform for all N if E ( N ,h) is in a compact
subset of { E : J V ( 2 ) < E J
- da: < +m}. For low lying eigenvalues
2(E-V(z))
( N is fixed), similar estimate was obtained by Simon (1983).
Lemma 3.2. Suppose the same conditions as in Lemma 3.1 and V
i s analytic. Assume the E ( N , h ) and E r n ( N I h )satisfies following travel
T z ( E (N,
< ST1(E(N,h))J 2 ( E ( N , f i ) - q y ) ) d y < +GO and 0 <
ti)) 1
time inequality
72 (Em( N ,h)) 1
d y < +aand Eo(N,ti) is the solution of the
JTl ( E m( N J 3 ) J2( Em( N , h )- V(y))
Bohr-Sommerfeld quantization equation (3.17) and 0 < (V’(x)(< 00 and
IV”(a:)I< 00 f o r 2 between
and
and between
and
Then
That is
(3.21)
355
> 0.
Together with (3.20), we have
sTI
72(E(N,ft)) 1
&(E(N, ti) - v(Y))
(E(N,W)
dY(E(N, ti) - Eo(N,ti))= 0 ( h 2 )
We need a spectral gap result. This can be proved by using the Bohr-
Sommerfeld quantization rule and Lemmas 3 . 2 . We first prove the following
lemma.
Lemma 3.2. I f 0 < JV'(x))< 00 f o r x E [71(E0(N+l,ti)),71(Eo(N,h))]U
[72(EO", ti)), 72(EO(N + 1, and 0 < S7 E';:" d mdx < 00 for
E = Eo(N,Ti)) and E = Eo(N + 1,h ) ) , then for suficiently small ti > 0,
356
But also
357
And
for sufficiently small ti. Here M2 in above is a constant. The lower bound
+
of E o ( N 1, ti) - E o ( N ,ti) in (3.22) follows. 0
358
It follows that there exist constants C1 > 0 and C2 > 0 such that for
sufficiently small h > 0
Clh 5 E ( N + 1,h) - E ( N ,h) 5 Czh,
and
Clh 5 Em(N + 1,h) - E m ( N ,h) 5 Czh.
Therefore there exists a neighbourhood I N of Eo(N,h) of which the length
is O ( h 2 ) ,there exists one and only one E and Em which are E ( N ,h) and
E,(N, h) respectively. But it is easy to see that
d I d -
).(
h(~QE(N,fi),m(x)\E(N,~),m - ~\E(N,fi),m(x)\~(~,fi),m)(x)I~=M
+O(Pfl) = 0. (3.23)
Recall
d d -
-
d x QE ,
I
w,ft),m ( x ) Q E( ~N , R ) (x>
,~ - z , m %,,, ( N , R ) (x)
QE, ( ~ , f i ) (x) ,~ L=M
= 0. (3.24)
But from the construction of Q m we know that there exist constants L1 > 0
and L2 > 0 such that
d d -
h2I (-QE,m
dx (x)*E,m(x) - z Q E , m ( Z ) Q E , m(x))~ z = M
d d -
- ( z Q E , , n ( Z ) * E , , , , m ( x ) - -QEm,m
dx
(z)8~m , m ) ( T )Iz=M I
2 (LI - Lzh)(E- Em(. (3.25)
This can be seen from the fact that
/ n(Ei)
TI( E I )
d2(E1- V ( x ) ) d x- 1 Q(Ez)
71(Ez
~ ( E- Vz ( z ) ) d x 2 CIES- Ezl,
and
1
(z)),
-tC+Ee(N,h)(x)’JQ(ilEE(N,h)
1
G m ( N , t i ) ( 4= c-E~_(N,h)jZ)J-;(~EE,,(N,h)(~))
1
+C+<k,(N,h)(4J;
(ZEE,”(N,h) (x)).
And similarly
for x E [T;, ‘,.I So from the Lipschitz continuity of cxj and with respect
to E l we know for x E [T,,T,’],
36 1
(3.32)
The same estimate is true for x E [T,, ~ $ 1 . For z < 71 and x > T$ we know
that the approximate wave functions Q E , and ~ QE,,~ are exponentially
small. Thus the L2 estimate (3.27) follows immediately. The derivative
estimate (3.28) can be proved by a similar argument. Here, similar to
(3.29), for rt
< z < 72,
h % ( N , h ) , m ( X ) - hQL,(N,h),rn(X) = 0(hrn+')
(3.33)
Proof. We have shown (3.7) holds for any z E (7-1,7-2). From the asymp-
totics *E in (3.1)-(3.6)in the different regions respectively, we have
-
(9E , O ( x ) o( 1:
72
(Q E,O (z)0 (ti"f1
)>2dx.
Note that J?: (QE,O(x)0 (hmfl ) ) dx and JT' (Q E ,0 (x)0 (ti"+')) dx are
exponentially small because of exponentially small integrand. Then using
Lemma 2.4, (3.34) follows easily.
To prove (3.35), calculate the derivative of Qk(z) in different regions
respectively. For x 5 71,
d
+
hQ.lE(X) = hQb,o(x)P-(x) h @ E , o ( z ) z P - ( s ) .
j=1
q!((x)(-h)j, therefore
Here similarly]
363
and
M
j=1
therefore
x )F L Q ~ , ~ ( X=) b+(x)O(hm+')
~Q&(-
d 1 1
+(-(r)
dX bT(x)
,-))o(hm+2).
bT(x)
+
(3.39)
hQ'lE(2)- hQ&,,(x) = ( h 2 Q ~ , o ( x ) ) o ( h m + ~ )
+hQ&,o(z)O (hm+' ) + QE,O(x)O(h m f 2.()3.40)
And for x 2 r;,
~Q&(- +~). +
x )h\IIL,,(x) = h\IIL,o(x))0(hm+') Q E , O ( X ) O ( ~ ~ (3.41)
Then (3.35) follows from (3.37)-(3.41)and similar argument as (3.36). Here
we use Lemma 2.4. 0
Remark. The quantity T = JV(x)lE d-2 ( E -'- V ( x ) )dx in Lemma 3.4 i s the
classical travel time between turning points. If {x : V ( x )= E } consists of
two simple zeros, then the classical travel time T is finite.
The main result of this paper is the following result.
Theorem 3.5. Assume conditions of Lemma 9.1, then the exact N-th
eigenvalue E ( N ,ti) of the Schrodinger operator H is approximated by the
m-th order approximate N-th eigenvalue E m ( N ,h) which satisfies the m-
th order quantization condition in the sense that E ( N , h ) - E m ( N , h ) =
O(hm+2),and the corresponding exact L 2 ( R )wave function Q E ( N , ~ )and its
derivative T L Q & ( ~ , ~are
) approximated by the WKB-Langer semi-classical
approximate wave function Q E , ( N , ~ L ) , ~associated with E m ( N ,h) and its
derivative hQ",,(N,h),m in L2(R), i.e.
l l Q E ( N , h ) - Q E m ( N , h ) , m l I L 2 ( R )= o ( h m f l ) i (3.42)
and
I I%
'( N,h) - hQkm( N , f i ),m 1 ILz(a) = ). (3.43)
TZ(EO)
In particular, set Eo(F) to be the solution of JTl(Eo)
(2(Eo- V ( y ) ) i d y =
7rF for any given F > 0, then the exact eigenvalue E ( N , h ) has semi-
364
classical limit Eo(F) in the sense that lim E ( N ,h) = Eo(F), and has
h-0
N-CC
(N+$)h=F
1
lim
h-0
km IIQE(N, W
- - QE,(N,fi),mIILz(R) = 0, (3.45)
N-CC
(N+$)h=F
+I I Q E (N,FL) ,m - @ E , II
( N , h ),m L2(R)
and applying Lemma 3.4 and Lemma 3.3 we have (3.42). Similarly we have
(3.43) by using
and
(3.48)
Acknowledgement
It’s our great pleasure t o thank D.Elworthy, B.Simon, D.Williams and
W. Zheng for helpful conversations.
References
1. C.M. Bender, K. Olanssen and P.S. Wang (1977), Numerological analysis of
the WKB approximation in large order, Physical Review D, Vol. 16, 1740-
1748.
2. M. V. Berry (1990), Waves near Stokes lines, Proc. Roy. SOC.London Ser.
A , VO~.427, 265-280.
3. J.L. Dunham (1932), The Wentzell-Brillouin-Kramersmethod of solving the
wave equation, Physical Review, Vol. 4, 713-720.
4. W.H. Furry (1947), Two notes on phase-integral methods, Physical Review,
Vol. 71, 360-371.
5 . P. Hartman (1964), Ordinary Differential Equations, John Wiley & Sons,
Inc, New York, London, Sydney.
6. J. Heading (1962), An introduction to phase integral methods, Methuen and
Co Ltd, London.
7. B. Helffer, A. Martinez and D. Robert (1987), Ergodicitk et limite semi-
classique, Commum. Math. Phys., Vol. 109, 313-326.
8. G. Jona-Lasinio, F. Martinelli and E. Scoppola (1981), The semi-classical
limit of quantum mechanics: a qualitative theory via stochastic mechanics,
Physical Reports, Vol. 77, 313-327.
9. R.E. Langer (1931), On asymptotic solutions of ordinary differential equa-
tions, with an application to the Bessel functions of large order, Transactions
of American Mathematical Society, Vol. 33,23-64.
10. R.E. Langer (1937), On the connection formulas and the solutions of the
wave equation, Physical Review, Vol. 51,669-676.
11. R.E. Langer (1949), The asymptotic solutions of ordinary linear differential
equations of second order, with special reference to a turning point, Trans-
actions of American Mathematical Society, Vol. 67, 461-490.
12. V.P. Maslov and M.V. Fedoriuk (1981), Semi-classical approximations in
quantum mechanics, Reidel, Dordrecht.
366