UNIVERSITY OF NOVI SAD

FACULTY OF SCIENCE
DEPARTMENT OF
MATHEMATICS AND INFORMATICS

Dora Seleši

Generalized Stochastic Processes in

Infinite Dimensional Spaces
with Applications to
Singular Stochastic Partial
Differential Equations

- doctoral dissertation -

Novi Sad, 2007

Introduction

White noise theory, as a discipline of infinite dimensional analysis, had a

fast development due to its broad spectrum of applications in the modeling
of stochastic dynamical phenomena arising in physics, economy, biology etc.
White noise is an excellent model for prompt, extremely large fluctuations,
but since it does not exists in classical sense, it requires the notion of a
generalized random process. In [HKPS], [HØUZ], [Ku] generalized random
processes are considered as functions X(t, ω) generalized by the argument ω
(ω is an element of a probability space Ω) and continuous by the time vari-
able t, while in [SM], [Ul], [Wa] generalized random processes are considered
as generalized functions in the t argument. Various conditions of continu-
ity lead to differences in the structure of these classes of processes, which
demand different methods in order to solve differential equations involving
these types of processes. In [AHR1], [AHR2], [OR1], [OR2], [RO], stochastic
differential equations are solved by regularization methods in the framework
of Colombeau theory. On the other hand, in [HØUZ], [LØUZ], [Øks] equa-
tions are solved using the Wick product and the Hermite–transformation or
the S–transformation. The authors in [MFA] make use of operator semi-
groups, and their solutions are generalized random processes with values in
a separable Hilbert space.
The investigation of the structure of various types of generalized random
processes is a prerequisite to solving various classes of stochastic differential
equations with singularities. Many physical, economical, biological or even
social phenomena can be represented by a mathematical model in form of a
stochastic differential equation involving singular coefficients, singular data
and singular initial values. In order to solve these equations, one needs the
concept of a generalized random process which can deal with these types of
singularities, i.e. one needs processes which are generalized by both argu-
ments (ω and t). Processes of this type were introduced in [Se], [PS1] and
will be subject also of this dissertation. A further step is to consider pro-
cesses taking values in an infinite dimensional space; for this purpose we will
consider a separable Hilbert space as the state space of the process.
2

The dissertation is organized in three chapters. Chapter 1 is expository

and it represents a short overview of generalized functions theory. Spaces of
deterministic and stochastic generalized functions are introduced, which will
be used in the sequel. All theorems are stated without proof since they are
parts of well-developed mathematical theories.
Chapter 2 and 3 contain the original parts of the dissertation. The subject
of Chapter 2 are generalized random processes and their structure represen-
tation. We will use the acronym GRP for ”generalized random process” and
the already established acronyms SDE, SPDE, CONS for ”stochastic dif-
ferential equation”, ”stochastic partial differential equation” and ”complete
orthonormal system”, respectively. Generalized random processes by vari-
ous types of continuity are considered and classified as GRPs of type (O),
(I) and (II). In particular, structure theorems for Hilbert space valued gen-
eralized random processes are obtained: Expansion theorems for GRPs (I)
considered as elements of the spaces L(A, S(H)−1 ) are derived, and structure
representation theorems for GRPs (II) on K{Mp } spaces on a set with arbi-
trary large probability are given. Especially, Gaussian GRPs (II) are proven
to be representable as a sum of derivatives of classical Gaussian processes
with appropriate growth rate at infinity. For GRPs (I) a Colombeau type
extension is introduced, and Wick products are defined. Also, in Chapter 2
applications to some classes of stochastic ordinary differential equations are
presented. In order to emphasize the differences in the concept of generalized
random processes defined by various conditions of continuity, the stochastic
differential equation y 0 (ω, t) = f (ω, t) is considered, where y is a generalized
random process having a point value at t = 0 in sense of Lojasiewicz. Other
SDEs considered in Chapter 2 are P (R)X = g and P (R)X = X¨W̃ + g,
where g is a GRP (I) and W̃ is white noise.
Applications to the eponymous singular stochastic partial differential
equations are the subject of Chapter 3. The main part of it is devoted
to elliptic equations; elliptic SPDEs in economics are typical of steady–state
problems such as perpetual options in multi–factor models, in physics they
describe diffusion processes in anisotropic media, while in medicine they are
a good model for brain functions.
We treat the linear elliptic stochastic Dirichlet problem in the framework
of GRPs (I) combined with Sobolev space and Colombeau algebra methods.
The equation Lu = h + ∇f with given stochastic boundary condition is
uniquely solved, where the boundary condition, as well as the data h and f
are GRPs regarded as linear continuous mappings from the Sobolev space
into the Kondratiev space, or as Colombeau extended GRPs. The operator L
is assumed to be strictly elliptic in divergence form Lu = ∇·(A·∇u+bu)+c·
∇u + du. Its coefficients: the elements of the matrix A and of the vectors b, c
 3

and d are (in the most general setting) assumed to be Colombeau generalized
random processes. Stability and regularity properties of the solutions are also
obtained. Another SDE considered in Chapter 3 is the stochastic Helmholtz
equation, which is solved by means of the Fourier transformation.
I would like to take this opportunity to express my huge thank to my
parents, Erika and György for their love, encouragement, kindness and pa-
tience. Special thanks go to my friends and colleagues for providing constant
support during my work on this dissertation. A complete list of their names
would fill pages. I wish to thank Professor Zagorka Lozanov–Crvenković and
Assistant Professor Danijela Rajter–Ćirić for many fruitful discussions about
Probability and Stochastics. To Professor Marko Nedeljkov I owe a debt of
gratitude for a careful reading of this text and providing a sequence of im-
provements. I would like to mention here his excellent lectures on PDEs;
some of the examples in this dissertation were inspired by things I learned at
the time I was conducting exercises for his lectures. A great thank goes to
Professor Michael Oberguggenberger for his hospitality and support during
my stay at the Insitute of Engineering Mathematics, Geometry and Com-
puter Science in Innsbruck, Austria. In particular, I wish to thank him for
his interesting Seminar on Stochastic Analysis and for our discussions during
my stay in Innsbruck. Last but not least, I want to express my deep ap-
preciation for the constant support, help and encouragement in many areas
given to me by my mentor, Professor Stevan Pilipović. He represents an
inexhaustible source of knowledge, who keeps the doors of his office always
open for his students. I wish to thank him for his comprehensive guidance
through the mathematical universe, showing me some of its most interesting
parts like Functional Analysis, Measure Theory, Generalized Functions (just
to mention some of them), which finally converged to this dissertation.
Contents

Introduction 1

1 Generalized Function Spaces 7

1.1 Nuclear Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 The Space of Tempered Distributions . . . . . . . . . . . . . . 10
1.2.1 Hermite functions . . . . . . . . . . . . . . . . . . . . . 10
1.2.2 Schwartz spaces . . . . . . . . . . . . . . . . . . . . . . 11
1.2.3 Hilbert space valued tempered distributions . . . . . . 12
1.3 Zemanian Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.1 The spaces Ak , A0k , A, A0 . . . . . . . . . . . . . . . . 13
1.3.2 The spaces expp A, expp A0 , ExpA, ExpA0 . . . . . . . 14
1.3.3 Hilbert space valued Zemanian functions . . . . . . . . 15
1.4 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 K{Mp } Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5.1 The weight functions Mp . . . . . . . . . . . . . . . . . 16
1.5.2 The functions fsx− , fsx+ . . . . . . . . . . . . . . . . . . 17
1.5.3 Point value of a distribution . . . . . . . . . . . . . . . 18
1.5.4 H–valued K{Mp } spaces . . . . . . . . . . . . . . . . . 18
1.6 Colombeau Generalized Functions . . . . . . . . . . . . . . . . 19
1.7 Spaces of Generalized Random Variables . . . . . . . . . . . . 20
1.7.1 White noise space . . . . . . . . . . . . . . . . . . . . . 20
1.7.2 The Wiener – Itô chaos expansion . . . . . . . . . . . . 21
1.7.3 The Kondratiev spaces (S)m,N ρ and (S)m,N
−ρ . . . . . . . 21
m,N m,N
1.7.4 The Spaces exp(S)ρ and exp(S)−ρ . . . . . . . . . . 22
1.7.5 Hilbert space valued generalized random variables . . . 23

2 Generalized Stochastic Processes 27

2.1 GRPs of Type (O) and (I) . . . . . . . . . . . . . . . . . . . . 30
2.2 Hilbert Space Valued GRPs (I) . . . . . . . . . . . . . . . . . 33
2.2.1 GRPs (I) on nuclear spaces . . . . . . . . . . . . . . . 40
2.2.2 Differentiation of GRPs (I) . . . . . . . . . . . . . . . . 41
CONTENTS 5

2.2.3 Application to a class of linear SDEs . . . . . . . . . . 42

2.3 The Wick Product of GRPs . . . . . . . . . . . . . . . . . . . 44
2.3.1 The Wick product of H–valued GRPs (I) . . . . . . . . 47
2.3.2 A Class of nonlinear SDEs . . . . . . . . . . . . . . . . 49
2.4 Colombeau Algebras for GRPs (I) and (O) . . . . . . . . . . . 53
2.4.1 Colombeau extension of L(A, (S)−1 ) . . . . . . . . . . 53
2.4.2 Colombeau extension of C ∞ (I; (S)−1 ) . . . . . . . . . . 57
2.5 GRPs of Type (II) . . . . . . . . . . . . . . . . . . . . . . . . 59
2.5.1 Generalized random processes on Lr (G), r >1 . . . . 62
2.5.2 Generalized random processes on K{Mp } . . . . . . . . 64
2.5.3 Examples of GRPs . . . . . . . . . . . . . . . . . . . . 73
2.6 Hilbert Space Valued GRPs (II) . . . . . . . . . . . . . . . . . 75

3 Applications to Singular SPDEs 81

3.1 A Linear Elliptic Dirichlet Problem . . . . . . . . . . . . . . . 81
3.1.1 Solvability of the stochastic Dirichlet problem . . . . . 87
3.1.2 Stability properties of the Dirichlet problem . . . . . . 90
3.1.3 Colombeau–type solutions of the Dirichlet problem . . 95
3.1.4 Colombeau–algebras G(W 2,2 ) and G(W 2,2 ; (S)−1 ) . . . . 95
3.2 A Dirichlet Problem with Stochastic Coefficients . . . . . . . . 106
3.2.1 Wick products . . . . . . . . . . . . . . . . . . . . . . . 107
3.2.2 Interpretation of the operator L . . . . . . . . . . . . . 108
3.2.3 The stochastic weak maximum principle . . . . . . . . 113
3.2.4 Solvability of the stochastic Dirichlet problem . . . . . 115
3.2.5 The Wick–adjoint of L . . . . . . . . . . . . . . . . . . 119
3.2.6 Stability and regularity properties . . . . . . . . . . . . 121
3.2.7 Colombeau–solutions of the Dirichlet problem . . . . . 125
3.3 Applications of the Fourier Transformation to GRPs (I) . . . . 129
3.3.1 The Fourier transformation of GRPs (I) . . . . . . . . 130
3.3.2 The one dimensional stochastic Helmholtz equation . . 132
3.3.3 The Helmholtz equation with stochastic wave number . 135

Epilogue 137

References 138

Index 144

Curriculum Vitae 146

Key Words Documentation 150

6 CONTENTS
 There is no branch of mathematics, however
abstract, which may not some day be applied
to phenomena of the real world.
(Nikolai Lobatchevsky)

Chapter 1

Generalized Function Spaces

In this introductory chapter we give a brief overview of some classes of

deterministic (e.g. Schwartz, Zemanian, Colombeau) and stochastic (Hida,
Kondratiev etc.) generalized function spaces. Definitions of some basic con-
cepts, their most important properties and relations are given, which are
necessary to understand the methods used in the sequent chapters of the
dissertation. Most of the material presented here is familiar and therefore
given without proof but with references for further reading.
Some basic notation we will use throughout the dissertation is the follow-
ing: Let V be a topological vector space, V 0 its dual space, and L(V, U ) be the
space of all linear continuous mappings from V into a topological vector space
U . Let (Ω, F, P ) be a probability space
R and Z p = Lp (Ω), p ≥ 1, be the space
of random variables X such that Ω |X|p dP < ∞. By Lr (Rn ), r ≥ 1, we
denote the space of r–integrable functions with respect to the Lebesgue mea-
sure m, by C k (Rn ) denote the space of k–times continuously differentiable
functions, and by C0 (Rn ) the space of continuous functions with compact
support.

1.1 Nuclear Spaces

For convenience of the reader not so familiar with notions in functional
analysis, we provide a basic overview of some concepts such as nuclear spaces
8 Generalized Function Spaces

or Gel’fand triples, which are fundamental in infinite dimensional analysis

and white noise theory. It is outside the scope of this chapter to give a
comprehensive discussion of all concepts and details related to nuclear spaces.
The reader is referred to [GV], [HKPS] and [Tr] for proofs and further details.

Projective and inductive topologies

Let E be a vector space, {Eα }α∈Λ a family of locally convex topological
vector
T spaces and {hα }α∈Λ a family of linear mappings such that hα : E → Eα
−1
and α∈Λ hα (0) = 0. The projective topology is the strongest locally con-
vex topology on E with respect to which all mappings hα are continuous.
Let {gα }α∈Λ be a family of linear mappings such that gα : Eα → E and
S
α∈Λ gα (Eα ) = E. The inductive topology is the finest locally convex topol-
ogy on E with respect to which all mappings gα are continuous.

Countably Hilbert spaces

Let V and W be separable Hilbert spaces. An operator A : V → W is
of Hilbert-Schmidt type if there exist an orthonormal basis P {en }2 in V and
{fn } in PW , and if there exist λn > 0, n ∈ N, such that n∈N λn < ∞ and
A(x) = n∈N λn (en |x)fn . An operator A : V → W is nuclear , if there exist
an orthonormal basisP {en } in V and {fn } inPW , and if there exist λn > 0,
n ∈ N, such that n∈N λn < ∞ and A(x) = n∈N λn (en |x)fn .
Let E0 be a separable Hilbert space endowed with the inner product (·|·)0
and let (·|·)p be a countable family of inner products. The space E0 endowed
with this family of inner products is calledPa countably Hilbert space if it is
−p |x|p
complete with respect to the metric |x| = ∞ p=0 2 1+|x|p
.
Without loss of generality we may assume the inner products are ordered
in an ascending way, i.e. |·|p ≤ |·|p+1 , p ∈ N0 . Denote by Ep the completition
of E0 with respect to |·|p . Clearly, Ep is T
a separable Hilbert space, . . . Ep+1 ⊆
Ep . . . ⊆ E0 and E = projlimp→∞ Ep = p∈N0 Ep .

Topologies on the dual

On the dual of a countably Hilbert space E 0 the strong topology is defined
by the basis of neighborhoods of zero B(A; ε) = {L ∈ E 0 : supx∈A |hL, xi| <
ε}, where A varies through all bounded sets and ε > 0. The strongest
topology on E 0 such that for each fixed x ∈ E, all mappings E 0 → C given
by L 7→ hL, xi, L ∈ E 0 are continuous, is called the weak topology. The dual
of a countably Hilbert space is complete with respect to the strong topology.
Denote now by E−p the dual of Ep . Thus, we obtain an increasing chain
of Hilbert spaces E0 ⊆ E−1 ⊆ . . . E−p ⊆ E−(p+1) . . . , and for all p ∈ N0 the
1.1 Nuclear Spaces 9

S
inclusion Ep ⊆ E0 ⊆ E−p is continuous. Let E 0 = p∈N0 E−p and endow it
with the inductive topology. Since E is dense in each Ep , the dual of E is
(in algebraic sense) ³E 0 and the´inductive topology on E 0 is equivalent to the
T∞ 0
strong topology on p∈N0 E p .

Nuclear spaces
A countably Hilbert space E is called nuclear , if for every p ∈ N there
exists q ≥ p such that the inclusion mapping Eq → Ep is nulcear. Recall
that the
P inclusion Eq → Ep is nuclear if there exists a CONS {ek } in Eq such
that k∈N |ek |2p < ∞. All finite dimensional spaces are nuclear. Moreover,
nuclear spaces have similar properties as finite dimensional spaces, e.g. if E
is nuclear, then A ⊂ E is compact if and only if it is bounded and closed.
Infinite dimensional Banach spaces are not nuclear. The notion of a nuclear
space is not reduced only to countably Hilbert spaces; one can define also
nuclear topological vector spaces (for this we refer to [Tr]). For example,
the dual of a nuclear countably Hilbert space is a nuclear topological vector
space. Some examples of nuclear spaces we will use in this dissertation are the
Schwartz spaces S, S0 , D, D0 , but also C0 , C0∞ . Also, the Zemanian spaces
A, A0 which are yet to be introduced, are nuclear under some additional
conditions.

Gel’fand triples
Let V be a nuclear space and V 0 its dual. If there exists a Hilbert space H
such that V is dense in H, then the triple V ⊆ H ⊆ V 0 is called a Gel’fand
triple. In fact, here we identified H with its dual H 0 . In case of nuclear
countably Hilbert spaces we have the Gel’fand triple E ⊆ E0 ⊆ E 0 .

Tensor products
In order to define a topological structure on the tensor product of two
topological vector spaces X⊗Y , there are two possibilities at our disposal (see
[Tr]). One possibility is to construct a seminorm topology relying directly on
the seminorm topologies of X and Y (this will be the so called π-topology).
The other possibility is to use an embedding of X ⊗ Y in some space related
to X and Y , over which a topology already exists (this construction will
lead to the so called ε-topology). In general, the π-topology is finer than the
ε-topology.
The π-topology on X ⊗ Y is the strongest locally convex topology such
that the canonical bilinear mapping X × Y → X ⊗ Y , (x, y) 7→ x ⊗ y
10 Generalized Function Spaces

is continuous. In terms of seminorms, the π-topology can be described as

follows: Let p, q be seminorms
Pn on X, Y , respectively and let θ ∈ X ⊗ Y .
Define (p ⊗ q)(θ) = inf j=1 p(xj )q(yP j ), where the infimum is taken over all
representations of θ of the form θ = nj=1 xj ⊗ yj for some xj ∈ X, yj ∈ Y ,
j = 1, 2, . . . n and n ∈ N. Especially, for x ∈ X, y ∈ Y , (p ⊗ q)(x ⊗ y) =
p(x)q(y). The completition of X ⊗ Y with respect to the π-topology is in
[Tr] denoted by X ⊗ ˆ πY .
In order to define the ε-topology, we observe that there exists a canon-
ical bilinear mapping X ⊗ Y → B(X 0 , Y 0 ), defined by (x, y) 7→ [(x0 , y 0 ) 7→
hx0 , xihy 0 , yi]. Here B(X 0 , Y 0 ) denotes the space of continuous bilinear forms
on X 0 × Y 0 . Thus, X ⊗ Y ∼ = B(X 0 , Y 0 ). Denote by Bε (X 0 , Y 0 ) the space of
separately continuous bilinear forms on X ⊗ Y equipped with the topology
of uniform convergence on the products A ⊗ B, where A is an equicontinuous
subset of X 0 and B is an equicontinuous subset of Y 0 . Since B(X 0 , Y 0 ) is a
subspace of Bε (X 0 , Y 0 ), it inherits this topology. Due to the above result,
X ⊗ Y also can be provided with this subspace topology. We call this sub-
space topology on X ⊗ Y the ε-topology. The completition of X ⊗ Y with
respect to the ε-topology is in [Tr] denoted by X ⊗ ˆ εY .
It is known that if X is a nuclear topological vector space, and Y is a
Fréchet space, then X ⊗ ˆ εY = X ⊗ ˆ π Y . Since we will deal only with tensor
products of spaces, where at least one of them is nuclear, we will suppress
the index and simply denote by X ⊗Y the completition of the tensor product
space with respect to either of the topologies.

1.2 The Space of Tempered Distributions

We use notation α = (α1 , . . . , αd ) ∈ Nd0 for multiindeces, Dα = ∂1α1 · · · ∂dαd
for the differential operator, and xα = xα1 1 · · · xαd d for x = (x1 , . . . , xd ) ∈ Rd .
The length of a multiindex α is defined as |α| = α1 + α2 + · · · + αd .

1.2.1 Hermite functions

The Hermite polynomial of order n, n ∈ N0 , is defined by hn (x) =
x2 d n x2
(−1)n e 2 dx n (e
− 2
), x ∈ R. It is well known that the family { √1n! hn :
n ∈ N0 } constitutes an orthonormal basis of the space L2 (R, dµ), where
x2
dµ = √12π e− 2 dx is the Gaussian measure. The Hermite function of order
n + 1, n ∈ N0 , is defined as
1 x2 √
ξn+1 (x) = √ √ e− 2 hn ( 2x), x ∈ R.
4
π n!
1.2 The Space of Tempered Distributions 11

Let α = (α1 , α2 , . . . , αd ) ∈ Nd be a multi-index. Define ξα = ξα1 ⊗ ξα2 ⊗

· · · ⊗ ξαd . The set of multi-indeces α can be ordered in an ascending sequence
as it is described in [HØUZ]. Denote by α(j) the jth multi-index in this
ordering. Hence, the family of vectors ξα can also be enumerated into a
countable family ηj = ξα(j) , j ∈ N. The family of functions {ηj : j ∈ N} is an
orthonormal basis of the space L2 (Rd ).

1.2.2 Schwartz spaces

The Schwartz space of rapidly decreasing functions is defined as

S(Rd ) = {f ∈ C∞ (Rd ) : ∀α, β ∈ Nd0 , kf kα,β < ∞},

and the topology on S(Rd ) is given by the family of seminorms kf kα,β =

supx∈Rd |xα Dβ f (x)|, α, β ∈ Nd0 . It is well known that S(Rd ) is a nuclear space.
The Schwartz space of tempered distributions is the dual space S0 (Rd ).
An equivalent construction of the topology on S(Rd ) can be obtained
using the self-adjoint operator A = −4 + |x|2 + 1, where 4 is the Laplace
operator. The operator A is densely defined on L2 (Rd ); more precisely the
domain of A contains S(Rd ). The Hermite functions {ηn }, n ∈ N, which form
an orthonormal basis of L2 (Rd ) are the eigenvectors of A, i.e. Aηn = λn ηn ,
where {λn = 2(n1 + · · · + nd ) − d + 1 : (n1 , . . . , nd ) ∈ Nd } is the spectar of
the operator A. Since zero does not belong to the spectrum, there exists the
inverse operator A−1 , and moreover kA−1 k = λ1 1 .
For p ∈ N define the norm | · |p on L2 (Rd ) by |f |p = |Ap f |L2 (Rd ) . Let
Sp (Rd ) be the closure of {f ∈ C∞ (Rd ) : |f |p < ∞} in L2 (Rd ). Then Sp (Rd )
is a Hilbert space with scalar product (f |g)p = (Ap f |Ap g)0 . For each p ∈ N,
S(Rd ) is dense in Sp (Rd ) and the inclusion Sp+1 (Rd ) ⊆ Sp (Rd ) is of Hilbert-
Schmidt type. The family of seminorms {| · |α,β : α, β ∈ Nd0 } and the family
of norms {| · |p : p ∈ N} are equivalent on S(Rd ). Thus, the projective limit
of spaces Sp (Rd ) is isomorphic to S(Rd ), i.e.
\
S(Rd ) = Sp (Rd ) = {f ∈ C∞ (Rd ) : ∀p ∈ N, |f |p < ∞}.
p∈N

For p ∈ N define |f |−p = |A−p f |L2 (Rd ) and let S−p (Rd ) be the com-
pletition of S(Rd ) with respect to | · |−p . The | · |−p norm topology is
equivalent to the strong topology on the dual space Sp0 (Rd ) defined by
| · |−p = supψ∈S(Rd ),|ψ|p ≤1 |h·, ψi|. Thus, S−p (Rd ) is isomorphic to Sp0 (Rd ).
S
In set–theoretical sense we have S0 (Rd ) = p∈N S−p (Rd ), and moreover
the locally convex inductive limit of the spaces S−p (Rd ) is isomorphic to
12 Generalized Function Spaces

S0 (Rd ) equipped with the strong topology. Thus, S(Rd ) ⊆ L2 (Rd ) ⊆ S0 (Rd )
form a Gel’fand triple. Moreover, we have continuous inclusions

S(Rd ) ⊆ Sp (Rd ) ⊆ L2 (Rd ) ⊆ Sp0 (Rd ) ⊆ S0 (Rd ).

For each p ∈ N the operator A−p is of P Hilbert-Schmidt type on L2 (Rd ),

∞
with Hilbert-Schmidt norm kA−p k2HS = −2p −p
n=1 λn . The vectors {λn ηn }
form an orthonormal basis of Sp (Rd ).
One can easily characterize the Schwartz spaces by Hermite functions
(recall, λk are the eigenvalues of the operator A, and ηk is the Hermite
2
CONS of LP (Rd )): A functionPf belongs to S(Rd ) if and only if it is of the
p 2
form f = k∈N ak ηk , where k∈N λk |ak | < ∞ for each p ∈ N and the
coefficients are given by ak = hf, ηPk i ∈ C. A function Pf belongs to S0 (Rd ) if
−p
and only if it is of the form f = k∈N ak ηk , where k∈N λk |ak |2 < ∞ for
some p ∈ N and the coefficients are given by ak = hf, ηk i ∈ C.
Note that the space of tempered distributions can be constructed also on
a bounded open set I ⊂ Rd . In this case, S0 (I) coincides with E0 (I), where
E0 (I) denotes the space of distributions having compact support.

1.2.3 Hilbert space valued tempered distributions

Let H be a separable Hilbert space with orthonormal basis {ei : i ∈ N}.
By S(Rd ; H) and S0 (Rd ; H) we will denote the H–valued Schwartz test func-
tion space, and the H–valued Schwartz generalized function space, respec-
tively. Since S(Rd ) is nuclear, we have (see [Tr, p.533]) that

S(Rd ; H) ∼
= S(Rd ) ⊗ H, S0 (Rd ; H) ∼
= (S(Rd ) ⊗ H)0 ∼
= S0 (Rd ) ⊗ H,

where the symbol ⊗ is understood as the tensor product on the test space,
and the π–completition (or in this case the equivalent ²–completition) of
the tensor product of dual spaces in the second case. Also, since H is a
Hilbert space, we may identify it with its dual and thus consider S0 (Rd ; H)
as B(S(Rd ), H) i.e. as the space of bilinear continuous mappings S(Rd )×H →
R. We will denote by h·, ·i this dual bilinear pairing.
Since the family {ei ηk : i, k ∈ N} is an orthogonal basis of L2 (Rd )⊗H, we
get a Hermite basis characterization also in the H–valued P case:
P A function
d
f belongs
P toP S(R ; H) if and only if it is of the form f = i∈N k∈N ai,k ηk ei ,
where i∈N k∈N λpk |ai,k |2 < ∞ for each p ∈ N and the coefficients are given
by ai,k = hf, ηk ei i ∈
PC. PA function f belongsP to S0P
(Rd ; H) if and only if it
is of the form f = i∈N k∈N ai,k ηk ei , where i∈N k∈N λ−p 2
k |ai,k | < ∞ for
some p ∈ N and the coefficients are given by ai,k = hf, ηk ei i ∈ C.
1.3 Zemanian Spaces 13

1.3 Zemanian Spaces

Let I be an open interval in R, and let R be a formally self-adjoint linear
differential operator of the form

R = θ0 Dn1 θ1 · · · Dnν θν = θ̄ν (−D)nν · · · (−D)n2 θ̄1 (−D)n1 θ̄0 (1.1)

where D = d/dx, θk are smooth complex functions without zero-points in I,

and nk are integers k = 1, 2, . . . , ν. Suppose that there exist a sequence of
real numbers hλn in∈N , limn→∞ |λn | = ∞, and a sequence of smooth functions
hψn in∈N in L2 (I) which are the eigenvalues and eigenfunctions, respectively,
of the operator R, i.e. Rψn = λn ψn , n ∈ N. We can enumerate them in an
ascending order: |λ1 | ≤ |λ2 | ≤ |λ3 | ≤ · · · → ∞. This re-ordering is made just
for technical reasons, but it is neither unique nor unavoidable. In Example
2.3.1 we will use a Zemanian space without this ordering. Suppose that
{ψn : n ∈ N} forms a CONS in L2 (I) with respect to the usual inner product
denoted byP(·|·). Each function f ∈ L2 (I) can be represented as an infinite
sum f = ∞ 2
n=1 (f |ψn )ψn converging in L (I). Define inductively: R = I,
0
k+1 k
R = R(R ), k ∈ N. Note that λn = 0 for some n ∈ N implies λk = 0 for
every k < n. From now on, if λn = 0, we replace it with λ fn = 1, else we put
f
λn = λn , n ∈ N.

1.3.1 The spaces Ak , A0k , A, A0

Define:
∞
X ∞
X 2k
Ak = {f = a n ψn : fn
|an |2 λ < ∞}, k ∈ Z.
n=1 n=1

If k ∈ N0 , then Ak ⊆ L2 (I). For each k ∈ N0 , Ak is a Hilbert space

P∞ fn 2k , where f =
when
P∞ provided with
P∞ the inner product (f |g)k = n=1 an b̄n λ
n=1 an ψn , g = n=1 bn ψn ∈ Ak . Denote by k · kk the norm induced by
this inner product. The dual space A0k , equipped with the usual dual norm,
is isomorphic with A−k . Thus, we have a sequence of linear continuous
canonical inclusions

· · · ⊆ Ak+1 ⊆ Ak ⊆ · · · A0 = L2 (I) ⊆ A−1 ⊆ A−2 ⊆ · · ·

The set
m
X
2
S = {f ∈ L (I) : f = an ψn , an ∈ C, m ∈ N}, (1.2)
n=1
14 Generalized Function Spaces

i.e. the linear span of the set {ψn : n ∈ N}, is dense in each Ak , k ∈ Z.
Define:
( ∞ ∞
)
\ X X 2k
2 2f
A= Ak = f ∈ L (I) : f = an ψn , ∀k, |an | λn < ∞ ,
k∈N0 n=1 n=1
( ∞ ∞
)
[ X X −2k
A0 = A−k = f= bn ψn : ∃k, fn
|bn |2 λ <∞ .
k∈N0 n=1 n=1

The Zemanian space of test functions A is equipped with the projective

topology, and its dual A0 , the Zemanian space of generalized functions, is
equipped with the inductive topology which is equivalent P∞ to the strong dual
0
topology. The action
P of a generalized function f = n=1 an ψn ∈PA∞ onto a
test function ϕ = ∞n=1 bn ψn is given by the dual pairing hf, ϕi = n=1 an bn .
−k
fn ψn : n ∈
The orthonormal basis of Ak , k ∈ N0 , is the family of functions {λ
P fn −2p < ∞
N}. Note that A is nuclear if for some p ≥ 0 the condition n∈N λ
holds.
d 2
2
Example 1.3.1 In particular, for the choice R = − dx 2 + x + 1, defined on

a maximal domain in L (R), A is the space of tempered distributions S0 (R).

2 0

1.3.2 The spaces expp A, expp A0 , ExpA, ExpA0

Let p ∈ N. Denote expp x = exp(exp(· · · (exp x)) · · · ) and define (see [Pi])
| {z }
p
expp A as the projective limit of the family
∞
X ∞
X
expp,k A = {ϕ = an ψn : fn )2k < ∞},
|an |2 (expp λ k ∈ N0 ,
n=1 n=1

P f 2k
equipped with the norm kϕkp,k = ∞ 2
n=1 |an | (expp λn ) , k ∈ N0 . Thus,
\ [
expp A = expp,k A, expp A0 = expp,−k A.
k∈N0 k∈N0

Clearly, S is dense in each expp,k A. The canonical inclusions expp,k+1 A ⊆

expp,k A are compact. Moreover, expp A is nuclear if for some c ∈ N0 the
P f −2c converges. Define the pair of test and generalized
series ∞ n=1 (expp λn )
function spaces ExpA and ExpA0 as

ExpA = projlim expp A, ExpA0 = indlim expp A0 .

p→∞ p→∞
1.4 Sobolev Spaces 15

The canonical inclusions expp+1 A ⊆ expp A are continuous and compact.

The set S is dense in expp A for each p ∈ N. Hence, ExpA is dense in each
expp A as well as in A.

1.3.3 Hilbert space valued Zemanian functions

Similarly as we considered H–valued Schwartz functions for a separable
Hilbert space H with orthonormal basis {ei : i ∈ N}, we generalize this
concept also to Zemanian functions. We will assume that A is nuclear, i.e.
P f−2p < ∞. This is necessary
that there exists some p ≥ 0, such that ∞ n=1 λn
in order to have an isomorphism of A0 (I; H) with the tensor product space
A0 ⊗H (refer to [Tr, Prop.50.7]). In general case, A0 ⊗H would be isomorphic
to a subspace of A0 (I; H).
Denote by Ak (I; H) the space of functions f : I → H of the form
P∞ P∞ P∞ P∞ 2f
2k
f = i=1 n=1 ai,n ψn ei such that kf kk;H = i=1 n=1 |ai,n | λn < ∞.
Let A(I; H) = projlimk→∞ Ak (I; H). Clearly, A0k (I; H) is isomorphic to
A−k (I; H) and we may define A0 (I; H) = indlimk→∞ A−k (I; H).
A similar construction can be carried out also for the spaces ExpA and
ExpA0 . We denote their Hilbert space valued versions by ExpA(I; H) and
ExpA0 (I; H).

1.4 Sobolev Spaces

Let I be an open subset of Rd . The αth weak derivative of f , denoted by
α
D f is given by the action
Z Z
D f (x)ϕ(x)dx = − f (x)Dα ϕ(x)dx,
α
I I

for all ϕ ∈ C0∞ (Rd ).

Denote by W k,p (I) the space of weakly differentiable functions f such
that Dα f ∈ Lp (I) for all |α| ≤ k. We endow W k,p (I) with the norm
X
kf kW k,p = kDα f kLp (I) .
|α|≤k

Clearly, W k,2 (I) is a Hilbert space.

Another important space we will consider is W0k,p (I) defined as the closure
of C0∞ (I) in W k,p (I). The dual space of W0k,p (I) will be denoted by W −k,p (I).
An isomorphism between W0k,p (I) and W −k,p (I) can be established via the
16 Generalized Function Spaces

Laplace operator. By its Hilbert structure, we also may identify W0k,2 (I)
with W −k,2 (I). Thus, we obtain a Gel’fand triple

W0k,2 (I) ⊆ L2 (I) ⊆ W −k,2 (I).

For further notions and properties of Sobolev spaces we refer to the mono-
graph [Ad].

1.5 K{Mp} Spaces

Now we give a brief overview of some basic notions from the theory of
K{Mp } spaces, which are constructed similarly as the tempered distributions,
but are more general. For further details refer to [GS] or [PS].

1.5.1 The weight functions Mp

Let (Mp )p∈N0 be a sequence of continuous functions on R such that the
following conditions are satisfied:

1 ≤ Mp (x) ≤ Mp0 (x), x ∈ R, p < p0 . (1.3)

(P) For every p ∈ N0 there is p0 ∈ N0 such that

lim Mp (x)Mp−1
0 (x) = 0.
|x|→∞

(N) For every p ∈ N0 there is p0 ∈ N0 such that Mp Mp−1

0 ∈ L1 (R).

K{Mp } is defined as a space of smooth functions ϕ ∈ C ∞ (R) endowed

with the family of norms

kϕkp = sup{Mp (x)|ϕ(i) (x)| : x ∈ R, i ≤ p} , p ∈ N0 .

We refer the reader to [GS] for the properties of K{Mp } and its strong dual
K0 {Mp }. In this paper we will consider a subclass of such spaces. Namely, as
in [GV, p.82], we will assume that {Mp , p ∈ N0 } are smooth functions such
that
(I) for every k, p ∈ N0 there exist p0 ∈ N0 and C > 0 such that

|Mp(k) (x)| ≤ CMp0 (x), x ∈ R.

1.5 K{Mp } Spaces 17

With the quoted conditions on Mp , p ∈ N0 , the sequence of norms k · kp ,

p ∈ N0 , is equivalent to the sequence of norms
Z
kϕkp,2 = sup{( |Mp (x)ϕ(i) (x)|2 dx)1/2 : i ≤ p}, p ∈ N0 .
R
p
For example, if we choose Mp (x) = (1 + |x|2 ) 2 , we obtain the space of rapidly
decreasing functions S(R) as K{Mp }.
Further on we will also assume that the weight functions Mp satisfy con-
dition
(T) for every p ∈ N0 there exist p̃ ∈ N0 such that

Mp (x − u)Mp (u) ≤ Mp̃ (x), 0 ≤ u ≤ |x|, x ∈ R.

p
Note that the functions Mp (x) = (1 + |x|2 ) 2 , and the functions defined
r
to be Mp (x) = ep|x| , for |x| > x0 > 0 and smooth around zero, r ∈ [1, ∞)
satisfy condition (T).

1.5.2 The functions fsx− , fsx+

We now give a brief overview of the techniques we will use in the sequel.
Let d ∈ N0 and Kd be the closure of K{Mp } with respect to k · kd,2 , i.e.
Kd = {φ ∈ C ∞ (R) : Md |φ(i) | ∈ L2 (R), i = 1, 2, . . . , d}. Denote:

K−
d = {φ ∈ Kd : supp φ ⊂ (−∞, b) for some b ∈ R},
K0d = Kd ∩ C0 ,
K+
d = {φ ∈ Kd : supp φ ⊂ (b, ∞) for some b ∈ R}.

Let for t ∈ R,

 H(t)ts−1

 , s>0
Γ(s)
fs (t) = (1.4)


 (n)
fs+n (t), s ≤ 0, s + n > 0, n ∈ N,

where H denotes the Heaviside function. Note that for s < 0 we have
(s)
fs = δ, where δ denotes the Dirac delta distribution . For s, r ∈ R we
have fs ∗ fr = fs+r .
For fixed x ∈ R denote
fsx− (t) = fs (x − t), t ∈ R,
fsx+ (t) = fs (t − x), t ∈ R.
18 Generalized Function Spaces

Clearly, fsx− , fsx+ ∈ C s−2 (R) and supp fsx− ⊂ (−∞, x], supp fsx+ ⊂ [x, ∞).
Note that for φ ∈ K{Mp } following holds:
µZ ¶(d+2) µZ x ¶(d+2)
x−
fd+2 (t)φ(t)dt = fd+2 (x − t)φ(t)dt = φ(x),
R −∞
µZ ¶(d+2) µZ ∞ ¶(d+2)
x+
fd+2 (t)φ(t)dt = fd+2 (t − x)φ(t)dt = φ(x).
R x
Now, let g ∈ K0 {Mp } such that supp g ⊂ (a, ∞) for some a ∈ R. There
exist C > 0 and d ∈ N such that
|hg, φi| ≤ Ckφkd , φ ∈ K{Mp }.
Then, g can be extended onto every φ ∈ Kd . Moreover, g can be extended
onto every ϕ ∈ C ∞ (R) such that supp ϕ ⊂ (−∞, b), for some b ∈ R. Thus, g
x− x−
can be extended to every fd+2 , x ∈ R. Note that the mapping x 7→ hg, fd+2 i
x−
is continuous. Thus, for G(x) = hg, fd+2 i, x ∈ R, we have

G(d+2) = g, (1.5)
where the derivative is understood in distributional sense. A similar ar-
x+
gument holds for G(x) = hg, fd+2 i, x ∈ R, where g ∈ K0 {Mp } such that
supp g ⊂ (−∞, a) for some a ∈ R.

1.5.3 Point value of a distribution

Recall the definition from [Lo] that a distribution g ∈ K0 {Mp } has at
t = t0 the point value in sense of Lojasiewicz, denoted as g(t0 ) = C, where
C ∈ R, if
Z
limhg(t0 + εx), ϕ(x)i = C ϕ(x)dx, ϕ ∈ K{Mp }.
ε→0 R
0
Let h ∈ K {Mp }, supp h ⊂ (a, ∞), a ∈ R. We say that H is a primitive
of h in a neighborhood of t0 , t0 ∈ (a, ∞), if H 0 = f in a neighborhood of t0 .

1.5.4 H–valued K{Mp } spaces

Conditions (P) and (N) imply (see[PS]) that K{Mp } is a nuclear space.
Thus, if H is a Hilbert space with orthonormal basis {ei : i ∈ N}, we
can consider the H–valued K{Mp } spaces, denoted by K{Mp }(H), as the
tensor product K{Mp }(H) ∼ = K{Mp } ⊗ H. Thus,
P∞ a function φ belongs to
K{Mp }(H) if and
P only if it is of the form φ = i=1 φi ei , φi ∈ K{Mp }, i ∈ N,
and kφk2p,2;H = ∞i=1 kφ 2
i p,2 < ∞ for all p ∈ N0 .
k
1.6 Colombeau Generalized Functions 19

1.6 Colombeau Generalized Functions

Main references for the so called simplified and full Colombeau algebras
are [Co1], [Co2], [Mo], [GKOS]. However, we want to extend this concept
and construct other Colombeau–type algebras of generalized function spaces.
For this approach we refer to [NP] and [DHPV].
Let V be a topological vector space on C with an an increasing sequence
of seminorms {pk : k ∈ N} defining the topology on V . The vector space
EM (V ), called the space of moderate nets, is constituted of functions R :
(0, 1) → V , ² 7→ R(²) = R² , such that for every k ∈ N there exists a > 0
with the property pk (R² ) = O(²−a ). The space of nets R² ∈ EM (V ) with the
property pk (R² ) = O(²a ) for all k ∈ N and all a > 0, is called the space of
negligible nets and will be denoted by N(V ). The quotient space
G(V ) = EM (V )/ N(V )
is called the polynomial Colombeau extension of V . Its elements (equivalence
classes) are denoted by [F² ], [F² ], ... etc. If V is an algebra whose products
are continuous for all seminorms, i.e. pk (ab) ≤ Cpk (a)pk (b), for all a, b ∈ V ,
k ∈ N, then N(V ) is an ideal of the algebra EM (V ). In this case G(V ) is an
algebra.
The following two cases were originally constructed by Colombeau:
If V = C, then G(V ) is called the Colombeau algebra of generalized con-
stants.
If V = C ∞ (I), I is an open subset of Rd , and pk (f ) = sup{|Dα f (x)| :
x ∈ Uk , |α| ≤ k}, where Uk , k ∈ N is an increasing sequence of compact
sets exhausting I, then G(V ) is called the simplified Colombeau algebra of
generalized functions and is usually denoted by G(I). Here one can easily deal
with differentiation and multiplication, which are defined by Dα [F² ] = [Dα F² ]
and [f² ][g² ] = [f² g² ]. The space of Schwartz generalized functions S0 (I) can
be embedded into G(I) by regularization: If f ∈ S0 (I) and has compact
support, then for a fixed mollifier function ρ, we have f ∗ ρ² − f ∈ N(I) and
[f ∗ ρ² ] ∈ G(I).
In the following chapters we will apply this type of construction also to
Sobolev spaces, and spaces of stochastic processes.
One can also consider H–valued Colombeau functions; for the case of a
Banach space H and the simplified Colombeau algebra see [GKOS, p.70]. The
H–valued simplified Colombeau algebra is denoted by G(I; H). Note that H
can be identified with a subspace of G(I; H) via the constant embedding
H 3 f 7→ [(x, ²) 7→ f ] ∈ G(I; H). If HLis finite dimensional with a basis
{hj j = 1, 2, . . . , m}, then G(I; H) ∼ =
m
j=1 G(I; R)hj . This idea will be
used in Chapter 2 to construct Colombeau algebras of generalized random
20 Generalized Function Spaces

processes (we will put some space of generalized random variables in place
of H).

1.7 Spaces of Generalized Random Variables

The spaces of generalized random variables are stochastic analogues of
deterministic generalized functions: They have no point value for ω ∈ Ω,
only an average value with respect to a test random variable. For details
refer to the sources that have been drawn upon: [HKPS], [HØUZ] or [Ku].

1.7.1 White noise space

Consider the Gel’fand triple S(Rd ) ⊆ L2 (Rd ) ⊆ S0 (Rd ), the Borel σ-
algebra B generated by the weak topology on S0 (Rd ) and the characteristic
function C(φ) = exp{− 12 |φ|2L2 (Rd ) }. According to the Bochner-Minlos the-
orem, there exists a unique measure µ on (S0 (Rd ), B) such that for each
φ ∈ S(Rd ) the relation
Z
eihω,φi dµ(ω) = C(φ)
S0 (Rd )

holds. Here hω, φi is the dual pairing of S0 (Rd ) and S(Rd ). The triplet
(S0 (Rd ), B, µ) is called the white noise space and µ is called the white noise
measure or the Gaussian measure on S0 (Rd ).
From now on we assume that the basic probability space (Ω, F, P ) is
the space (S0 (Rd ), B, µ). Put (L)2 = L2 (S0 (Rd ), B, µ). It is a Hilbert space
equipped with the inner product (F |G)(L)2 = Eµ (F G).
QmIn a multi-dimensional
d d
case, for a given m ∈ N, m > 1, define
d 0
Qm 0Sm d=
i=1 iS (R ), where S i (R ) is a copy of S(R ), and let S m = i=1 Si (R ).
0
Equip the space Sm with the product Borel σ-algebra and with the product
measure µm = µ×· · ·×µ. The triple (S0m , B, µm ) is called the m-dimensional
d-parameter white noise L space. Put (L)2,m = L2 (S0m , µm ), and for N ∈ N,
N
N > 1, let (L)2,m,N = k=1 (L)2,m be the direct sum of N identical copies of
the m-dimensional d-parameter white noise space. Here the finite dimension
N is the state space dimension. Later, instead of the N –dimensional state
space, we will also consider the infinite–dimensional case, when the state
space is a separable Hilbert space H.
1.7 Spaces of Generalized Random Variables 21

1.7.2 The Wiener – Itô chaos expansion

Denote by I = (NN0 )c the set of sequences of integers which have only
finitely many nonzero components. For a given Q∞α = (α1 , α2 , . . .) ∈ I define
the Fourier-Hermite polynomial as Hα (ω) = i=1 hαi (hω, ηi i), ω ∈ S0 (Rd ).
In the multi-dimensional
Lm valued case (m > 1) the orthonormal basis of the
2 d
space K = k=1 L (R ) is constituted of vectors of length m of the form
e(k) = (0, . . . , ηj , 0, . . .), where ηj takes the ith place in the sequence, and i, j
are integers such that k = i + (j − 1)m, i ∈ {1, 2, . . . , m}, j ∈ N. In this
(m) Q
case we define Hα (ω) = ∞ (k) 0
k=1 hαk (hω, e i), ω ∈ Sm . The Fourier-Hermite
(m)
polynomials Hα form an orthogonal basis of (L)2,m . It is also known that
(m) √
kHα k(L)2,m = α1 !α2 ! · · ·, α = (α1 , α2 , . . .) ∈ I.
The Wiener-Itô expansion theorem (see [HØUZ]) states that each element
F ∈ (L)2,m,N has a unique representation of the form
X
F (ω) = cα Hα(m) (ω), ω ∈ S0m , cα ∈ RN , α ∈ I,
α∈I
P
such that kF k2(L)2,m,N = α∈I α!c2α , where c2α = (cα |cα ) is the standard inner
product in RN .

Example 1.7.1 Let εj = (0, 0, . . . , 1, 0, . . .) be a sequence of zeros with the

number 1 as the jth component. The one-dimensional d-parameter Brownian
motion B(t, ω) = hω, κ[0,t] i has expansion
∞ µZ
X t ¶
B(t, ω) = ηj (u)du Hεj (ω), t ∈ Rd , ω ∈ S0 (Rd ).
j=1 0

For every t ∈ Rd fixed, B(t, ·) is an element of (L)2 . Brownian motion is a

Gaussian process whose almost all trajectories are continuous, but nowhere
differentiable functions.

1.7.3 The Kondratiev spaces (S)m,N

ρ and (S)m,N
−ρ

We will use notation αβ = α1β1 α2β2 · · · for given multi-indeces α, β ∈ I,

and ∞
Y
γ
(2N) = (2j)γj ,
j=1
P
where (γ1 , γ2 , . . .) ∈ I. Then, α∈I (2N)−pα < ∞ if and only if p > 1,
Pγ = −p(2N) α
and α∈I e < ∞ if and only if p > 0.
22 Generalized Function Spaces

We will use the definition of the Kondratiev spaces given in [HØUZ]

where the authors provide an equivalent construction of the original one
introduced by Kondratiev. The space of the Kondratiev stochastic test func-
tions (space of Kondratiev test random variables) (S)m,Nρ consists of elements
P (m) 2,m,N N
f = α∈I cα Hα ∈ (L) , cα ∈ R , α ∈ I, such that
X
kf k2ρ,p = c2α (α!)1+ρ (2N)pα < ∞ for all p ∈ N0 .
α∈I

The space of the Kondratiev stochastic generalized functions (space of Kon-

dratiev generalized random variables) (S)m,N−ρ consists of formal expansions
P (m) N
of the form F = α∈I bα Hα , bα ∈ R , α ∈ I, such that
X
kF k2−ρ,−p = b2α (α!)1−ρ (2N)−pα < ∞ for some p ∈ N0 .
α∈I
P
The action of F onto a test function f is given by hF, f i = α∈I (bα |cα )α!
where (·|·) is the standard inner product in RN .
The generalized expectation of F is defined as E(F ) = hF, 1i = b0 .
Note that the space (S)m,Nρ canPalso be constructed as the projective
m,N
limit of the family (Sp )ρ = {f = α∈I cα Hα : kf kρ,p < ∞}, p ∈ N0 . Also,
space (S)m,N
−ρ can also be constructed as the inductive limit of the family
m,N P
(S−p )−ρ = {f = α∈I cα Hα : kf k−ρ,−p < ∞}, p ∈ N0 .
In particular, for ρ = 0 the Kondratiev spaces are called the Hida spaces
of test and generalized stochastic functions.
Example 1.7.2 One-dimensional d-parameter singular white noise is de-
fined via the formal expansion
∞
X
W (t, ω) = ηk (t)Hεk (ω)
k=1

where t ∈ Rd , and εk , ηk are as in the previous example. For every t ∈ Rd

fixed, W (t, ·) belongs to the Hida space of generalized stochastic functions.
Singular white noise is the distributional derivative of Brownian motion.

1.7.4 The Spaces exp(S)m,N

ρ and exp(S)m,N
−ρ

In [Se], [PS1] following space of generalized random variables was intro-

duced: The space of stochastic test functions exp(S)m,N
ρ consists of elements
P (m)
f = α∈I cα Hα ∈ (L)2,m,N , cα ∈ RN , α ∈ I, such that
X α
kf k2ρ,p,exp = c2α (α!)1+ρ ep(2N) < ∞, for all p ∈ N0 .
α∈I
1.7 Spaces of Generalized Random Variables 23

The space of stochastic generalized functions exp(S)m,N

−ρ consists of formal
P (m) N
expansions of the form F = α∈I bα Hα , bα ∈ R , α ∈ I, such that
X α
kF k2−ρ,−p,exp = b2α (α!)1−ρ e−p(2N) < ∞, for some p ∈ N0 .
α∈I

Note that for each ρ ∈ [0, 1], exp(S)m,N

ρ is nuclear and exp(S)m,N
−ρ ⊆
m,N m,N m,N
exp(S)−1 (the canonical inclusion exp(S)1 ⊆ exp(S)ρ is compact).
Moreover, following relationship to the Kondratiev spaces holds:

exp(S)m,N
ρ ⊆ (S)m,N
ρ ⊆ (L)2,m,N ⊆ (S)m,N m,N
−ρ ⊆ exp(S)−ρ ,

i.e. the generalized random variable space exp(S)m,N

−ρ is wider than the cor-
responding Kondratiev one. The canonical inclusion exp(S)m,N ρ ⊆ (S)m,N
ρ
m,N
is compact. From the construction it follows that exp(S)ρ is dense in
(L)2,m,N , i.e. exp(S)m,N
ρ ⊆ (L)2,m,N ⊆ exp(S)m,N
−ρ is a Gel’fand triple.

1.7.5 Hilbert space valued generalized

random variables
Let H be a separable Hilbert space with orthonormal basis {ei : i ∈ N}.
As already suggested in Section 1.7.1, we will treat H as the state space, i.e.
we replace the N –dimensional case with an infinitedimensional state space
(see [MFA] and the references therein). While in [MFA] the case m = 1 is
considered, we keep our white noise space dimension to be m ≥ 1.
Recall that the basic probability space is (S0m , B, µm ). Denote by
L2,m (Ω; H) the space of functions on Ω with values in H, which are square in-
(m)
tegrable with respect to µm . The family of functions {Hα ei : i ∈ N, α ∈ I}
is an orthogonal basis of the Hilbert space L2,m (Ω; H). Each element of
L2,m (Ω; H) can be represented in either of the following forms:
∞
X ∞
X
2,m
f (ω) = ai (ω)ei , ai = hf, ei iH ∈ (L) , kai k2(L)2,m < ∞,
i=1 i=1

∞ X
X ∞ X
X
f (ω) = ai,α Hα(m) (ω)ei , ai,α = hf, Hα(m) ei i ∈ R, α!|ai,α |2 < ∞,
i=1 α∈I i=1 α∈I

X X
f (ω) = aα Hα(m) (ω), aα = hf, Hα(m) i(L)2,m ∈ H, α!kaα k2H < ∞.
α∈I α∈I
24 Generalized Function Spaces

Now, one can build up spaces of H–valued generalized random variables

(H–valued Kondratiev spaces and others) over L2,m (Ω; H) following the same
pattern as in the RN –valued case. Let ρ ∈ [0, 1].
Define S(H)m as the space of functions f ∈ L2,m (Ω; H),
P∞ P ρ (m)
f (ω) = i=1 α∈I ai,α Hα (ω)ei , ai,α ∈ R, such that for all p ∈ N0 ,
∞ X
X ∞
XX
kf k2ρ,p;H = α! 1+ρ 2
|ai,α | (2N) pα
= α!1+ρ |ai,α |2 (2N)pα < ∞.
i=1 α∈I α∈I i=1

Define S(H)m as the space of formal expansions

P∞ P−ρ (m)
F (ω) = i=1 α∈I bi,α Hα (ω)ei , bi,α ∈ R, such that for some q ∈ N0 ,
∞ X
X ∞
XX
kF k2−ρ,−q;H = α!1−ρ |bi,α |2 (2N)−qα = α!1−ρ |bi,α |2 (2N)−qα < ∞.
i=1 α∈I α∈I i=1

Note, we can also write

∞
XX X ∞
X
f (ω) = ai,α Hα(m) (ω)ei = aα Hα(m) (ω) = ai (ω)ei ,
α∈I i=1 α∈I i=1
P∞ P (m)
where aα = i=1 ai,α ei ∈ H and ai = α∈I ai,α Hα (ω) ∈ (S)m
ρ . Also,

X ∞
X
kf k2ρ,p;H = α!1+ρ kaα k2H (2N)pα = kai k2ρ,p .
α∈I i=1

The same holds also for

∞
XX X ∞
X
F (ω) = bi,α Hα(m) (ω)ei = bα Hα(m) (ω) = bi (ω)ei ,
α∈I i=1 α∈I i=1
P∞ P (m)
where bα = i=1 bi,α ei ∈ H and bi = α∈I bi,α Hα (ω) ∈ (S)m
−ρ . Also,

X ∞
X
kF k2−ρ,−q;H = α! 1−ρ
kbα k2H (2N)−qα = kbi k2−ρ,−q .
α∈I i=1

The action of F onto f is given by

X
hF, f i = α!hbα , aα iH .
α∈I

Similarly as in the finite-dimensional case, we have

(m) (m) (m) (m) (m)
S(H)1 ⊆ S(H)(m)
ρ ⊆ S(H)0 ⊆ L2,m (Ω; H) ⊆ S(H)−0 ⊆ S(H)−ρ ⊆ S(H)−1 .
1.7 Spaces of Generalized Random Variables 25

The same construction can be carried out for the exponential growth rate
spaces. Let exp S(H)m ρ be the space of functions f ∈ L
2,m
(Ω; H), f (ω) =
P∞ P (m)
i=1 α∈I ai,α Hα (ω)ei , ai,α ∈ R, such that for all p ∈ N0 ,
∞ X
X ∞
XX
α α
kf k2ρ,p,exp;H = α!1+ρ |ai,α |2 ep(2N) = α!1+ρ |ai,α |2 ep(2N) < ∞.
i=1 α∈I α∈I i=1

The corresponding space of stochastic generalized functions exp S(H)m

−ρ
P∞ P (m)
consists of formal expansions of the form F (ω) = i=1 α∈I bi,α Hα (ω)ei ,
bi,α ∈ R, such that for some q ∈ N0 ,
∞ X
X ∞
XX
2 −q(2N)α α
kF k2−ρ,−q,exp;H = α! 1−ρ
|bi,α | e = α!1−ρ |bi,α |2 e−q(2N) < ∞.
i=1 α∈I α∈I i=1

Both S(H)m m
ρ and exp S(H)ρ are countably Hilbert spaces and

exp S(H)m m
ρ ⊆ S(H)ρ ⊆ L
2,m
(Ω; H) ⊆ S(H)m m
−ρ ⊆ exp S(H)−ρ .

In general, S(H)m m
ρ and exp S(H)ρ are not nuclear spaces. They would
be nuclear e.g. if H were finite–dimensional (see [Tr, Prop.50.1,Cor.]).
Note that since (S)m m
ρ and exp(S)ρ are nuclear spaces, by [Tr, Prop.50.7]
we have again (as for the Schwartz spaces in the deterministic case) an iso-
morphism with tensor product spaces:
S(H)m ∼ = (S)m ⊗ H,
−ρ −ρ exp S(H)m ∼ = exp(S)m ⊗ H. −ρ −ρ

As in [MFA] one can define H–valued Brownian motion and singular

white noise:
Example 1.7.3 Let n : N × N → N, (i, j) 7→ n(i, j) be the usual bijection
given in the following table:
1 2 3 4 5 6 7 ··· j ···
1 1 3 6 10 15 21 28 ···
..
2 2 5 9 14 20 27 .
..
3 4 8 13 19 26 .
..
4 7 12 18 25 .
..
5 11 17 24 .
..
6 16 23 .
..
7 22 .
.. .. ..
. . .
i ··· ··· ··· ··· ··· ··· ··· n(i, j) · · ·
.. ..
. .
26 Generalized Function Spaces

Let εj = (0, 0, . . . , 1, 0, . . .) be a sequence of zeros with the number 1 as the

jth component and ξj the Hermite functions. Then,
∞ Z
X t
βi (t, ω) = ξj (s)dsHεn(i,j) (ω), i∈N
j=1 0

is a sequence of independent (one-dimensional d-parameter R–valued) Brow-

nian motions. Rewrite this as
∞
X ½ Rt
ξ (s)ds, k = n(i, j)
0 j
βi (t, ω) = θik (t)Hεk (ω), θik (t) = .
0, k 6= n(i, j)
k=1

The formal sum

∞
X ∞
X Z t
B(t, ω) = βi (t, ω)ei = θk (t)Hεk (ω), θk (t) = δn(i,j),k ξj (s)ds ei
i=1 k=1 0
(1.6)
is an H–valued Brownian motion. Note that while R–valued Brownian mo-
tion was an element of (L)2 for fixed t ≥ 0, now B(t, ω) does not (!) belong
to L2 (Ω; H). However, the sum in (1.6) does converge in S(H)−0 for each
t ≥ 0 fixed.

Example 1.7.4 H–valued (one-dimensional, d-parameter) singular white

noise is defined by the formal sum
∞
X
W(t, ω) = κk (t)Hεk (ω), κk (t) = δn(i,j),k ξj (t) ei .
k=1

It is also an element of S(H)−0 .

 The infinite! No other question has ever moved
so profoundly the spirit of man.
(David Hilbert)

Chapter 2

Generalized Stochastic
Processes

Generalized random (stochastic) processes can be defined in various ways

depending on whether the author regards them as a family of random vari-
ables or as a family of trajectories, but also depending on the type of con-
tinuity implied onto this family. It is well–known that a classical stochastic
process X(t, ω), t ∈ T ⊆ Rd , ω ∈ Ω, can be regarded either as a family of
random variables X(t, ·), t ∈ T , as a family of trajectories X(·, ω), ω ∈ Ω, or
as a family of functions X : T × Ω → Rn (Rn is the state space) such that
for each fixed t ∈ T , X(t, ·) is an Rn –valued random variable and for each
fixed ω ∈ Ω, X(·, ω) is an Rn –valued deterministic function (called a trajec-
tory). For classical stochastic processes these three concepts are equivalent,
but if one replaces the space of trajectories with some space of determin-
istic generalized functions, or if one replaces the space of random variables
with some space of generalized random variables, then different types of gen-
eralized stochastic processes are obtained. Each of these approaches is an
acorn from which a great theory has grown. The classification of generalized
stochastic processes by various conditions of continuity leads to structural
theorems such as integral representations and series expansions, which will
be subject of the following sections. Let us give now a historical overview of
various definitions of the concept of generalized random processes (GRPs).
28 Generalized Stochastic Processes

Processes generalized with respect to the t argument

One possible definition of a GRP is the one used by J.B. Walsh (see [Wa])
as a measurable mapping X : Ω → D0 (T ). For each φ ∈ D(T ), the mapping
Ω → R, ω 7→ hX(ω), φi is a random variable. This definition is motivated by
the fact that trajectories of Brownian motion are nowhere differentiable and
can be considered as elements of D0 (T ).
K. Itô defined in [It] a GRP as a linear and continuous mapping from
L (R) to the space L2 (Ω) of random variables with finite second moments,
2

while Inaba considered in [In] a GRP as a continuous mapping from a certain

space of test functions to the space L2 (Ω). Also, Gel’fand and Vilenkin [GV]
have considered GRPs in this sense. Further on, we will refer to GRPs defined
in this sense as to GRPs of type (I).
On the other hand, O. Hanš, M. Ullrich, L. Swartz and others (see [Ha],
[LP], [LP1], [LP2], [SM], [Ul]) defined a GRP as a mapping ξ : Ω × V → C
such that for every ϕ ∈ V, ξ(·, ϕ) is a complex random variable, and for every
ω ∈ Ω, ξ(ω, ·) is an element in V 0 , where V denotes a topological vector space
and V 0 its dual space. Further on, we will refer to GRPs defined in this sense
as to GRPs of type (II).
Note that if a GRP is of type (I), then we do not have continuity of
sample paths for each fixed ω ∈ Ω, only continuity in distributional sense. If
we assume for a GRP (I) the continuity for a.e. fixed ω ∈ Ω, then it is also of
type (II). However, in [Wa] Walsh proved that, if the underlying test space
V is nuclear, then for a GRP (I) there exists a version which is also a GRP
(II). This result need not hold true if V is not nuclear, e.g. if it is a Hilbert
space, as it will be shown in Example 2.5.1.
Vice versa, if a GRP is of type (II), this does not ensure its continuity as
a mapping from the test space K{Mp } into L2 (Ω) or even L0 (Ω) equipped
with convergence in probability. Nevertheless, we show in Theorem 2.5.2
under which conditions is a GRP of type (II) also a GRP (I) up to a set
of arbitrarily small measure. This result can not be improved, i.e. one can
not obtain a version up to a set of measure zero, as it is shown in Example
2.5.5. In Example 2.5.1 we show that the the result need not hold true if we
consider GRPs (II) on a Hilbert space.
It remains as an enticing open question to investigate the general relation-
ship of GRPs (I) and (II) and to find which conditions one should propose on
not nuclear spaces in order to obtain an equivalence between the two types
of continuity.
Another possibility to generalize the concept of a stochastic process is to
take advantage of Colombeau generalized function spaces. This approach was
recently used by M. Oberguggenberger, F. Russo, S. Albeverio, M. Nedeljkov,
 29

D. Rajter and others (see [OR1], [OR2], [RO], [AHR1], [AHR2], [NR].) In
this framework a GRP is defined as a mapping X : Ω → G(T ), (G(T ) is the
simplified Colombeau algebra of generalized functions) having the property
that there exists a net of functions R² : Ω × T → C, ² ∈ (0, 1), such that
R² (ω, ·) represents X(ω) for a.e. ω ∈ Ω, and for every ² ∈ (0, 1), (ω, t) →
R² (ω, t) is a measurable (classical) stochastic process. Recall that D0 (T ) can
be embedded into G(T ), thus this is a generalization of the Walsh definition
of a GRP.
Also, in [LPP], Colombeau GRPs of second order were considered as
elements of the vector–valued Colombeau algebra G(T ; L2 (Ω)), where L2 (Ω)
is the Hilbert space of random variables with finite second moments.

Processes generalized with respect to the ω argument

T. Hida, Y. Kondratiev, B. Øksendal, H.-H. Kuo, and many others (refer
to [HKPS], [HØUZ], [Ku]) have developed a very general concept of GRPs via
chaos expansions. In [HØUZ] GRPs are defined as measurable mappings T →
(S)−1 , where (S)−1 denotes the Kondratiev space, but one can consider also
some other spaces of generalized random variables instead of it. Thus, they
are pointwisely defined with respect to the parameter t ∈ T and generalized
with respect to ω ∈ Ω. Further on, we will refer to GRPs defined in this
sense as to GRPs of type (O).
Examples of GRPs (O) are Brownian motion and singular white noise
defined via the chaos expansions in Example 1.7.1 and Example 1.7.2, re-
spectively.

Processes generalized with respect to both arguments

GRPs of this type were introduced in [Se] and [PS1], where we in fact
generalized and unified the concept of a GRP in Inaba’s sence and the concept
of a GRP (O), and considered GRPs as linear continuous mappings from
the Zemanian test space into (S)−1 . There we gave structural properties of
these GRPs by series expansions in spaces of Zemanian generalized functions
and a simultaneous chaos expansion. Since these processes, as elements of
L(A, (S)−1 ), are very close to the concept of a GRP (I), we will also call
them GRPs of type (I).
As in the deterministic theory of distributions, where we identify locally
integrable functions and regular distributions, we can identify elements f ∈
L1loc (Rd , (S)−1 ) with the corresponding linear mappings f˜ ∈ L(S(Rd ), (S)−1 ).
In other words, the class of locally Pettis-integrable GRPs (O) can be em-
bedded into the class of GRPs (I).
30 Generalized Stochastic Processes

We will follow a similar idea in this dissertation and replace L2 (Ω) with
the more general (S)−1 and consider the Colombeau extension of a GRP
(I) i.e. we will consider GRPs as elements of G(Rd ; (S)−1 ). Similarly as in
the deterministic theory of distributions, where S0 (Rd ) can be embedded into
G(Rd ), we will have the vector–valued analogy of it: The space of GRPs (I)
L(S(Rd ), (S)−1 ) can be embedded into G(Rd ; (S)−1 ).

2.1 Generalized Random Processes

of Type (O) and (I)
We give now an overview of the results obtained in [Se] and [PS1].
Elements of the spaces L(A, (S)−1 ) and L(A, exp(S)−1 ), but also of
L(ExpA, (S)−1 ) and L(ExpA, exp(S)−1 ) are GRPs (I). As already men-
tioned, these processes are generalized both by the time-parameter t and
by the random parameter ω.
Let t 7→ U (t, ·) ∈ (S)−1 be a GRP (O) which is locally integrable in the
Pettis sense, U ∈ L1loc (Rd , (S)−1 ). Then there is an associated GRP (I), i.e.
a linear, continuous mapping Ũ ∈ L(S(Rd ), (S)−1 ) such that
Z
[Ũ , ϕ](ω) = U (t, ω)ϕ(t)dt, ϕ ∈ S(Rd ), ω ∈ S0 (Rd ), (2.1)
Rd

where [·, ·] denotes the action of Ũ ∈ L(S(Rd ), (S)−1 ) onto a ϕ ∈ S(Rd ); so

that [Ũ , ϕ] ∈ (S)−1 .
Further we will also denote by [·, ·] the action of an element from
L(A, (S)−1 ) or L(A, exp(S)−1 ) onto an element from A, and with h·, ·i the
classical dual pairing in A0 and A.

Series expansion of GRPs (O)

Since GRPs (O) with values in (L)2,m,N are defined pointwisely with re-
spect to the parameter t ∈ Rd , their expansions follow directly from the
Wiener–Itô theorem: Let Z : Rd → (S)m,N−ρ be a GRP (O). Then it has an
expansion X
Z(t) = cα (t)Hα(m) , t ∈ Rd , (2.2)
α∈I

where cα : R → R , α ∈ I are measurable functions. If cα ∈ L1 (Rd , RN ),

d N

α ∈ I, and if there exists p ∈ N0 such that

Z X
kZ(t)k−ρ,−p dt = α!1−ρ |cα (t)|2L1 (Rd ) (2N)−pα < ∞,
Rd α∈I
2.1 GRPs of Type (O) and (I) 31

then Z ∈ L1 (Rd , (S)m,N

−ρ ), that is Z is a Pettis-integrable process.
The expansion theorems of GRPs (I) will give an extension of the ex-
pansion of the Pettis-integrable GRPs (O) in the sense that the coefficients
cα (t), α ∈ I, in (2.2) will be generalized functions from A0 .

Series expansion of GRPs (I)

Consider for example GRPs (I) as elements of the space A∗ =
L(A, (S)−1 ). Elements of A∗k = L(Ak , (S)−1 ) are called GRPs (I) of R-order
k. We have a chain of continuous canonical inclusions
[
(L2 (I))∗ = A∗0 ⊆ A∗1 ⊆ · · · ⊆ A∗k ⊆ A∗ = A∗k .
k∈N0

For technical reasons we assume that the set of multi-indeces I is ordered

in a lexicographic order and denote by αj , j ∈ N, the jth element in this
ordering.

Definition 2.1.1 P Let fj ∈ A0 , j = 1, 2, . . . , m and let θαj ∈ (S)−1 , j =

1, 2, . . . , m. Then m j=1 fj ⊗ θαj is a GRP (I), more precisely an element of
∗
A defined by
" m # m
X X
fj ⊗ θαj , ϕ = hfj , ϕiθαj , ϕ ∈ A. (2.3)
j=1 j=1

Theorem 2.1.1 Let k ∈ N0 . The following conditions are equivalent:

(i) Φ ∈ A∗k .
(ii) Φ can be represented in the form
∞
X
Φ= fj ⊗ Hαj , fj ∈ A−k , j∈N (2.4)
j=1

and there exists k0 ∈ N0 such that for each bounded set B ⊆ Ak

∞
X j
sup |hfj , ϕi|2 (2N)−k0 α < ∞. (2.5)
ϕ∈B
j=1

(iii) Φ can be represented in the form (2.4) and there exists k1 ∈ N0 such
that ∞
X j
kfj k2−k (2N)−k1 α < ∞. (2.6)
j=1
32 Generalized Stochastic Processes

Since A∗ is constructed as the inductive limit of the family A∗k , k ∈ N0 ,

we obtain the following expansion theorem for a GRP (I).
Theorem 2.1.2 Φ ∈ A∗ if and only if there exist k, k0 ∈ N0 such that series
expansion (2.4) and condition (2.5) hold.

Consistency of the expansions of GRPs (O) and GRPs (I)

Let U be a GRP (O) given by the expansion
∞
X
U (t, ω) = ai (t)Hαi (ω), t ∈ R, ω ∈ S0 (R),
i=1

such that ai (t) ∈ L1loc (R), i ∈ N. Then due to (2.1), there is a GRP (I),
denoted by Ũ associated with U , such that
Z X
∞ ∞
X
[Ũ , ϕ](ω) = ai (t)Hαi (ω)ϕ(t)dt = hãi , ϕiHαi (ω), ω ∈ S0 (R),
R i=1 i=1

where ãi ∈ S0 (R) is the generalized function associated with the function
ai (t) ∈ L1loc (R), i ∈ N. Thus, Ũ has expansion
∞
X
Ũ = ãi ⊗ Hαi .
i=1

This means that the expansion theorems for GRPs (I) are consistent with
the expansion theorems for GRPs (O).

exp
Expansion theorems for A∗ = L(A, exp(S)−1 )
Analogously to the previous theorems concerning GRPs (I) with values in
(S)−1 , one can consider GRPs (I) taking values in other spaces of generalized
stochastic functions. The expansion theorems can be carried over, mutatis
mutandis, to these GRPs.
Consider for example the space exp(S)−1 , which will provide a larger class
of GRP (I). Let exp A∗ = L(A, exp(S)−1 ) and exp A∗k = L(Ak , exp(S)−1 ) be
GRPs (I) and GRPs (I) of R-order k, respectively. Further, let all the other
terms be defined analogously as for A∗ ; i.e. we replace (S)−1 with exp(S)−1
in Definition 2.1.1 and else where necessary.

Theorem 2.1.3 Let k ∈ N0 . The following conditions are equivalent:

exp
(i) Φ ∈ A∗k .
2.2 Hilbert Space Valued GRPs (I) 33

(ii) Φ can be represented in the form

∞
X
Φ= fj ⊗ Hαj , fj ∈ A−k , j ∈ N, (2.7)
j=1

and there exists k0 ∈ N0 , such that for each bounded set B ⊆ Ak

∞
X αj
sup |hfj , ϕi|2 e−k0 (2N) < ∞. (2.8)
ϕ∈B
j=1

(iii) Φ can be represented in the form (2.7) and there exists k1 ≥ 0, such
that ∞
X αj
kfj k2−k e−k1 (2N) < ∞.
j=1

For examples of GRPs (I) and definition of the Wick product in spaces
A∗ refer to [PS1].

2.2 Hilbert Space Valued Generalized

Random Processes of Type (I)
Now we expand the results of [PS1] and consider Hilbert space valued
GRPs. Throughout this section, and also further on, H will denote a sepa-
rable Hilbert space with orthonormal basis {ei : i ∈ N}.
We replace the Kondratiev space (S)−1 with the H–valued Kondratiev
space S(H)−1 and define H–valued GRPs (I) as linear continuous mappings
from the Zemanian test space A into S(H)−1 i.e. as elements of
A(H)∗ = L(A, S(H)−1 ).
Elements of A(H)∗k = L(Ak , S(H)−1 ) are called H–valued GRPs (I) of R-
order k. Thus, L ∈ A(H)∗k if and only if there exists k0 ∈ N such that
L ∈ L(Ak , S(H)−1,−k0 ). Note that L(Ak , S(H)−1,−k0 ) is a Banach space with
the usual dual norm
kLk∗−k;H = sup {k[L, g]k−1,−k0 ;H : g ∈ Ak , kgkk ≤ 1} .
Clearly, A(H)0k ⊆ A(H)∗k , and kf k∗−k;H = kf k−k;H if f ∈ A(H)−k . We have
a chain of continuous canonical inclusions
[
L2 (I; H) = A(H)∗0 ⊆ A(H)∗1 ⊆ · · · ⊆ A(H)∗k ⊆ A(H)∗ = A(H)∗k .
k∈N0
34 Generalized Stochastic Processes

Definition 2.2.1 P Let fj ∈ A0 , j = 1, 2, . . . , m and let θαj ∈ S(H)−1 , j =

1, 2, . . . , m. Then m ∗
j=1 fj ⊗ θαj is an element of A(H) defined by
" m # m
X X
fj ⊗ θαj , ϕ = hfj , ϕiθαj , ϕ ∈ A. (2.9)
j=1 j=1

Recall that each θαj ∈ S(H)−1 can be represented as

∞ X
X ∞
θαj (ω) = θjik Hαk (ω)ei , θjik ∈ R. (2.10)
i=1 k=1

Thus, (2.9) can be written in an equivalent form

m
X ∞ X
X ∞ ∞ X
X ∞
hfj , ϕi θjik Hαk (ω)ei = hFik , ϕiHαk (ω)ei
j=1 i=1 k=1 i=1 k=1
"∞ ∞ #
XX
= Fik ⊗ Hαk (ω)ei , ϕ ,
i=1 k=1
Pm
where Fik = j=1 fj θjik ∈ A0 .

Lemma 2.2.1 Let hfn in∈N be a sequence in A0 and hθαj ij∈N be a sequence
in S(H)−1 . If there exists k0 ∈ N0 such that for any bounded set B ∈ A,
∞
X
sup |hfj , ϕi| · kθαj k−1,−k0 ;H < ∞, (2.11)
ϕ∈B
j=1
P∞
then j=1 fj ⊗ θαj defined by
∞
X m
X
fj ⊗ θαj = lim fj ⊗ θαj
m→∞
j=1 j=1

is an element of A(H)∗ .
P Pm
Proof. Denote Υ = ∞ j=1 fj ⊗ θαj and Υm = j=1 fj ⊗ θαj , m ∈ N. Clearly,
∗
Υm ∈ A . If (2.11) holds, then Υm ∈ L(A, S(H)−1,−k0 ).
The sequence of partial sums Υm , m ∈ N, is a Cauchy sequence in A(H)∗
because, for given ε > 0 and m > n,
m
X
k[Υm , ϕ] − [Υn , ϕ]k−1,−k0 ;H = |hfj , ϕi| · kθαj k−1,−k0 ;H < ε,
j=n+1
2.2 Hilbert Space Valued GRPs (I) 35

if n, m are large enough.

Since

kΥm k∗ = sup{k[Υ, ϕ]k−1,−k0 ;H : ϕ ∈ A, kϕk ≤ 1}

m
X
≤ sup{ |hfj , ϕi| · kθαj k−1,−k0 ;H : ϕ ∈ A, kϕk ≤ 1},
j=1

it follows that
à m
!
X
sup kΥm k∗ ≤ sup sup{ |hfj , ϕi| · kθαj k−1,−k0 ;H : ϕ ∈ A, kϕk ≤ 1}
m∈N m∈N
j=1
∞
X
≤ sup{ |hfj , ϕi| · kθαj k−1,−k0 ;H : ϕ ∈ A, kϕk ≤ 1} < ∞,
j=1

by condition (2.11). According to the Banach-Steinhaus theorem, Υ =

limm→∞ Υm also belongs to L(A, S(H)−1,−k0 ). ¤

Theorem 2.2.1 Let k ∈ N0 . The following conditions are equivalent:

(i) Φ ∈ A(H)∗k .
(ii) Φ can be represented in the form
∞ X
X ∞
Φ= fij ⊗ Hαj ei , fij ∈ A−k , i, j ∈ N (2.12)
i=1 j=1

and there exists k0 ∈ N0 such that for each bounded set B ⊆ Ak

∞ X
X ∞
j
sup |hfij , ϕi|2 (2N)−k0 α < ∞. (2.13)
ϕ∈B
i=1 j=1

Proof. Let Φ ∈ A∗k (H) = L(Ak , S(H)−1 ). There exists k0 ∈ N0 , such

that Φ ∈ L(Ak , S(H)−1,−k0 ). The mapping fij : Ak → R given by ϕ 7→
([Φ, ϕ]|Hαj ei )−1,−k0 ;H is linear and continuous for each Hαj ei , i.e. fij ∈ A0k =
A−k for each i, j ∈ N. Thus,

hfij , ϕi = ([Φ, ϕ]|Hαj ei )−1,−k0 ;H , ϕ ∈ Ak , j ∈ N.

Also, [Φ, ϕ] ∈ S(H)−1,−k0 has the expansion

∞ X
X ∞
[Φ, ϕ] = ([Φ, ϕ]|Hαj ei )−1,−k0 ;H Hαj ei . (2.14)
i=1 j=1
36 Generalized Stochastic Processes

The series on the right-hand side of (2.14) converges if and only if

X∞ X ∞ ¯ ¯2 ∞
X
¯ ¯ −k0 αj j
¯([Φ, ϕ]|H e )
α i −1,−k0 ;H ¯
j (2N) = |hfj , ϕi|2 (2N)−k0 α < ∞,
i=1 j=1 j=1

which yields (2.13). Now, by Definition 2.2.1, (2.14) is equal to

∞ X∞
"∞ #
X X
[Φ, ϕ] = hfij , ϕiHαj ei = fij ⊗ Hαj ei , ϕ ,
i=1 j=1 j=1

and this implies (2.12). P P

Conversely, let Φ = ∞ i=1
∞
j=1 fij ⊗ Hαj ei , where fij ∈ A−k , i, j ∈ N,
and let (2.13) hold for any bounded set B ⊆ Ak .
Since [Φ, ϕ] ∈ S(H)−1,−k0 , it has the expansion
∞ X
X ∞
[Φ, ϕ] = ([Φ, ϕ]|Hαj ei )−1,−k0 ;H Hαj ei
i=1 j=1
X∞ X ∞ ∞ X
X ∞
= ([ flk ⊗ Hαk el , ϕ]|Hαj ei )−1,−k0 ;H Hαj ei
i=1 j=1 l=1 k=1
X∞ X ∞ ∞
XX ∞
= ( hflk , ϕiHαk el |Hαj ei )−1,−k0 ;H Hαj ei
i=1 j=1 l=1 k=1
X∞ X ∞
j
= hfij , ϕi(2N)−k0 α Hαj ei , ϕ ∈ Ak ,
i=1 j=1

where the orthogonality of the basis HαjP

ei wasPused in the last step.
The sequence of partial sums Φm = m i=1
m
j=1 fij ⊗ Hαj ei , m ∈ N, is a
Cauchy sequence in L(Ak , S(H)−1,−k0 ) because, for given ε > 0,
m
X m
X j
k[Φm , ϕ] − [Φn , ϕ]k2−1,−k0 ;H = |hfij , ϕi|2 (2N)−k0 α < ε,
i=n+1 j=n+1

if we choose n, m large enough. This yields that hΦm im∈N is a Cauchy se-
quence in A(H)∗k . Also, (2.13) implies that
X∞ X
∞
j
sup kΦm k∗2
−k;H ≤ sup{ |hfij , ϕi|2 (2N)−k0 α : ϕ ∈ Ak , kϕkk ≤ 1} < ∞.
m∈N
i=1 j=1

Thus, due to the Banach-Steinhaus theorem, Φ0 = limm→∞ Φm ∈ A(H)∗k . So

it has to be of the form
∞ X
X ∞
Φ0 = f˜ij ⊗ Hαj ei .
i=1 j=1
2.2 Hilbert Space Valued GRPs (I) 37

It remains to show that Φ0 = Φ. Since

hf˜ij , ϕi − hfij , ϕi = ([Φ0 , ϕ]|Hαj ei )−1,−k0 ;H − ([Φ, ϕ]|Hαj ei )−1,−k0 ;H
= ([ lim Φm , ϕ]|Hαj ei )−1,−k0 ;H − ([Φ, ϕ]|Hαj ei )−1,−k0 ;H
m→∞
= lim ([Φm − Φ, ϕ]|Hαj ei )−1,−k0 ;H
m→∞
X ∞
X
= lim ([ fij ⊗ Hαj ei , ϕ]|Hαj ei )−1,−k0 ;H
m→∞
i=m+1 j=m+1
X X ∞
j
= lim |hfij , ϕi|2 (2N)−k0 α = 0
m→∞
i=m+1 j=m+1

for any ϕ ∈ Ak , it implies that f˜ij = fij , i, j ∈ N. ¤

Corollary 2.2.1 If Φ can be represented in the form (2.12) and there exists
k1 ∈ N0 such that
X∞ X∞
j
kfij k2−k (2N)−k1 α < ∞, (2.15)
i=1 j=1

then Φ ∈ A(H)∗k .
Proof. According to the Cauchy-Schwartz inequality, it is obvious that
(2.15) implies (2.13). ¤

Remark: Note that in the finite dimensional valued case we had an

equivalence between (2.5) and (2.6) in Theorem 2.1.1. But now (2.13) does
not necessarily imply (2.15). This implication would be true only if S(H)−1
were a nuclear space (see [GV, Theorem 1, p. 67]), which it is not.
Since A(H)∗ is constructed as the inductive limit of the family A(H)∗k ,
k ∈ N0 , we obtain the following expansion theorem for a H–valued GRP (I).
Theorem 2.2.2 Φ ∈ A(H)∗ if and only if there exist k, k0 ∈ N0 such that
series expansion (2.12) and condition (2.13) hold.
Let U be a H–valued GRP (O) given by the expansion
∞ X
X ∞
U (t, ω) = aij (t)Hαj (ω)ei , t ∈ R, ω ∈ S0 (R),
i=1 j=1

such that aij (t) ∈ L1loc (R), i, j ∈ N. Then there is a H–valued GRP (I),
denoted by Ũ associated with U , such that
Z X∞ X∞ ∞ X
X ∞
[Ũ , ϕ](ω) = aij (t)Hαj (ω)ei ϕ(t)dt = hãij , ϕiHαj (ω)ei , ω ∈ S0 (R),
R i=1 j=1 i=1 j=1
38 Generalized Stochastic Processes

where ãij ∈ S0 (R) is the generalized function associated with the function
aij (t) ∈ L1loc (R), i, j ∈ N. Thus, Ũ has expansion
∞ X
X ∞
Ũ = ãij ⊗ Hαj ei .
i=1 j=1

The expansion theorems for exp A(H)∗ = L(A, exp S(H)−1 ) can also be
stated as in the case of a one–dimensional state space:

Theorem 2.2.3 Let k ∈ N0 . The following conditions are equivalent:

exp
(i) Φ ∈ A(H)∗k .

(ii) Φ can be represented in the form

∞ X
X ∞
Φ= fij ⊗ Hαj ei , fij ∈ A−k , i, j ∈ N, (2.16)
i=1 j=1

and there exists k0 ∈ N0 , such that for each bounded set B ⊆ Ak

∞ X
X ∞
αj
sup |hfij , ϕi|2 e−k0 (2N) < ∞. (2.17)
ϕ∈B
i=1 j=1

Corollary 2.2.2 If Φ can be represented in the form (2.16) and there exists
k1 ≥ 0, such that
X∞ X ∞
αj
kfij k2−k e−k1 (2N) < ∞,
i=1 j=1
exp
then Φ ∈ A(H)∗k .

Theorem 2.2.4 Φ ∈ exp A(H)∗ if and only if there exist k, k0 ∈ N0 such

that series expansion (2.16) and condition (2.17) hold.
d 2
2
Example 2.2.1 Let R = − dx 2 + x + 1 and I = R. Then A = S(R),
0 0
A = S (R) and ψk (t) = ξk (t), k ∈ N, where ξk are the Hermite functions.

(i) In Example 1.7.4, H–valued one-dimensional d-parameter singular

white noise as a GRP (O) was defined by the formal sum
∞
X
W(t, ω) = κk (t)Hεk (ω), κk (t) = δn(i,j),k ξj (t) ei .
k=1
2.2 Hilbert Space Valued GRPs (I) 39

With the Hermite function ξj weR associate a generalized function

ξ˜j ∈ S0 (R) defined by hξ˜j , ϕi = R ξj (t)ϕ(t)dt, ϕ ∈ S(R). Define
κ̃k = δn(i,j),k ξ˜j (t) ei . Then, white noise W̃ as an H–valued GRP
(I) has the expansion
∞
X
W̃ = κ̃k ⊗ Hεk ∈ L(S(R), S(H)−1 ).
k=1

Condition (2.15) is also satisfied, because

∞ X
X ∞ ∞
X
δn(i,j),k kξj k2L2 (R) (2N)−pεj = (2k)−p < ∞,
i=1 j=1 k=1

if we choose p > 1.
Note also that κ̃k is an element of S0 (R; H).
5
(ii) Let l > 12 and t1 , t2 , t3 , . . . ∈ R such that t1 ≤ t2 ≤ t3 ≤ · · · → ∞.
It is known that the Dirac delta distributions δtj , j ∈ N, belong to
A−l = S−l (R). Let

∆k = δn(i,j),k δtj ei , k∈N

(to avoid confusion: the first delta is the Kronecker symbol, the second
one is the Dirac distribution). With
∞
X
∆k ⊗ Hαk (2.18)
k=1

is given a GRP (I) which is not a GRP (O). We will check condition
P 1
(2.15). Since δtj (x) = ∞n=1 ξn (tj )ξn (x), j ∈ N, and ξn = O(n
− 12
),
n ∈ N, we have
∞ X
X ∞ ∞
X
j k
δn(i,j),k kδtj k2−l (2N)−pα = kδtk k2−l (2N)−pα
i=1 j=1 k=1
X∞ X∞
k
= |ξn (tk )|2 (2n)−2l (2N)−pα
k=1 n=1
X∞ ∞
X 1
−pαk
≤C (2N) n− 6 (2n)−2l < ∞,
k=1 n=1

for some constant C > 0 and p > 1. Hence, the process given by (2.18)
meets the definition of a GRP (I).
40 Generalized Stochastic Processes

2.2.1 GRPs (I) on nuclear spaces

Recall that since (S)−1 is a nuclear space, we have S(H)−1 ∼ = (S)−1 ⊗ H.
Assume now, that A is also a nuclear space (this is not a strict restriction
since in most cases it is one); then we have by Proposition 50.7. in [Tr]
L(A, S(H)−1 ) ∼ = A0 ⊗ S(H)−1 . Combining this with the previous remark, we
can now consider GRPs (I) as elements of A0 ⊗ (S)−1 ⊗ H, or if we regroup
the spaces, also as elements of A0 ⊗ H ⊗ (S)−1 , which is again by nuclearity
of A isomorphic to A0 (I; H) ⊗ (S)−1 . In other words, it is equivalent whether
we consider the state space H as the codomain of the generalized random
variables, or as the codomain of the deterministic generalized functions rep-
resenting the trajectories of the process.
Similarly as we did for GRPs (I) in Theorem 2.2.1, we have a representa-
tion for elements of P A0 (I; H). A function g belongs to AP −k (I; H) if and only
∞ ∞ 2
if it is of the form i=1 gi ⊗ ei , gi ∈ A−k and sup Pϕ∈B i=1 |hgi , ϕi| < ∞
∞
holds for eachP∞bounded set BP⊆∞ Ak . The sum i=1 gi ⊗ ei is defined by
the action h i=1 gi ⊗ ei , ϕi = i=1 hgi , ϕiei , ϕ ∈ A, provided the latter sum
converges in H. P P∞
Thus, if Φ is a GRP (I) given by the expansion Φ = ∞ i=1 j=1 fij ⊗Hαj ei ,
fij ∈ A−k , we can rewrite its action in the following manner:
"∞ ∞ # ∞ X ∞ ∞ X ∞
XX X X
fij ⊗ Hαj ei , ϕ = hfij , ϕiHαj ei = h fij ⊗ ei , ϕiHαj .
i=1 j=1 i=1 j=1 j=1 i=1
P P∞
Also, from supϕ∈B ∞ i=1
2
j=1 |hfij , ϕi| (2N)
−pαj
< ∞, B ⊆ Ak , we get
P∞ P
2
supϕ∈B j=1 khgj , ϕikH (2N) −pαj
< ∞, where gj = ∞ i=1 fij ⊗ ei ∈ A−k (I; H)
P ∞ −k
and kgj k2−k;H = i=1 |fij |2 λ̃i .
In view of these facts we can now reformulate our representation theorem
for GRPs (I):
Theorem 2.2.5 Let k ∈ N0 . The following conditions are equivalent:
(i) Φ ∈ A(H)∗k .
(ii) Φ can be represented in the form
∞
X
Φ= f j ⊗ H αj , fj ∈ A−k (I; H), j∈N (2.19)
j=1

and there exists k0 ∈ N0 such that for each bounded set B ⊆ Ak

∞
X j
sup khfj , ϕik2H (2N)−k0 α < ∞. (2.20)
ϕ∈B
j=1
2.2 Hilbert Space Valued GRPs (I) 41

Corollary 2.2.3 If Φ can be represented in the form (2.19) and there exists
k1 ∈ N0 such that
X∞
j
kfj k2−k;H (2N)−k1 α < ∞, (2.21)
j=1

then Φ ∈ A(H)∗k .

Example 2.2.2 Denote by δy ∈ S0 (R; H) be the H–valued Dirac delta dis-

tribution at y ∈ R. Let tj ∈ R, j ∈ N, such that t1 ≤ t2 ≤ t3 ≤ · · · → ∞.
Then,
X∞
δtj ⊗ Hαj
j=1

defines a H–valued GRP (I). Indeed, since P∞ {ξnP

ei : n, i ∈ N} is an orthogo-
nal basis of S (R; H), we can write δy = i=1 ∞
0 2
n=1 din (y)ξn ei , kδy k−k;H =
P∞ P∞ 2 −k 2
i=1 n=1 |din (y)| (2n) < ∞, and moreover kδy k−k;H does not depend on
P∞ 2 −pαj
P∞ −pαj
y. Thus, j=1 kδt j
k−k;H (2N) = Const j=1 (2N) < ∞ for p > 1,
proving that (2.21) holds.

2.2.2 Differentiation of GRPs (I)

Definition 2.2.2 Let F ∈£ A(H)∗¤. The£ distributional
¤ derivative of F , de-
∂ ∂ ∂
noted by ∂t F is defined by ∂t F, ϕ = − F, ∂t ϕ , for all ϕ ∈ A.
P∞ P∞
Lemma 2.2.2 Let F = i=1 j=1 fij ⊗ Hαj ei ∈ A(H)∗ . Then,
∞
XX ∂ ∞
∂
F = fij ⊗ Hαj ei , (2.22)
∂t i=1 j=1
∂t

∂
where f
∂t ij
is the distributional derivative of fij in A0 .

Proof. The assertion follows from the fact that

"∞ ∞ # ∞ X ∞
XX ∂ X ∂
fij ⊗ Hαj ei , ϕ = h fij , ϕiHαj ei
i=1 j=1
∂t i=1 j=1
∂t
X∞ X ∞ · ¸
∂ ∂
=− hfij , ϕiHαj ei = − F, ϕ
i=1 j=1
∂t ∂t

for all ϕ ∈ A. Obviously, condition (2.13) is satisfied. ¤

42 Generalized Stochastic Processes

It is a well-known fact in the (deterministic) generalized functions theory

that if a function f : R → R is differentiable on R \ {x0 } and has a jump in
x0 ∈ R, then Df = f 0 (x) + Cδx0 , where Df is the distributional derivative
in S0 (R), f 0 (x) is the classical derivative on R \ {x0 }, δx0 ∈ S0 (R) is the Dirac
delta distribution, and C = f (x+ −
0 ) − f (x0 ). In this light, the GRP defined
in Example 2.2.2 is a stochastic analogue of the Dirac delta distribution.
P
Example 2.2.3 Let F = ∞ j=1 fj Hαj , fj : R → R be a GRP (O), i.e. for
P∞ 2 −pαj
each t ∈ R fixed j=1 |fj (t)| (2N) < ∞ for some p > 0. Denote by
0
δy ∈ S (R) be the Dirac delta distribution at y ∈ R. Assume that for each
j ∈ N, the function fj is differentiable on R \ {tj }, has one jump in tj ∈ R,
P
and t1 ≤ t2 ≤ t3 ≤ · · · → ∞. Assume that ∞ 0 2
j=1 |fj (t)| (2N)
−pαj
< ∞ for
+ −
each fixed t ∈ R \ {t1 , t2 , . . .}. Let cj = f (tj ) − f (tj ), j ∈ N, be the jump
heights. Assume there exists C > 0 such that |cj | ≤ C, j ∈ N (i.e. the jump
heights are bounded). Then,
X ∂ ∞ X¡ ∞
∂ ¢
F = fj ⊗ Hαj = fj0 + cj δtj ⊗ Hαj .
∂t j=1
∂t j=1
P P∞
Since ∞ 2
j=1 cj kδtj k−k (2j)
−p
P∞ ≤ 0C
2
j=1 kδtj k−k (2j)
−p
< ∞,
P∞the process
above is well-defined. Note j=1 fj (t)Hα is a GRP (O) and j=1 δtj ⊗ Hαj
j

is the GRP (I) defined in Example 2.2.2.

2.2.3 Application to a class of linear SDEs

Let R be of the form (1.1) and P (t) = pn tn + pn−1 tn−1 + · · · + p1 t + p0 ,
t ∈ I, be a polynomial with real coefficients. We give two examples of
stochastic differential equations using GRPs (I), and a differential operator
P (R) defined as P (R) = pn Rn + pn−1 Rn−1 + · · · + p1 R + p0 I. Note that if ψk
is an eigenfunction of R, then
P (R)ψk = (pk Rn + pn−1 Rn−1 + · · · + p1 R + p0 I)ψk
= pn (λek )n ψk + pn−1 (λek )n−1 ψk + · · · + p1 λek ψk + p0 ψk = P (λek )ψk .
Consider an SDE of the form
P (R)u = g (2.23)
where g ∈ A(H)∗r is a GRP (I).
Let u and g be given by series expansions
∞ X
X ∞ ∞ X
X ∞
u(t, ω) = uij (t) ⊗ Hαj (ω)ei , g(t, ω) = gij (t) ⊗ Hαj (ω)ei ,
i=1 j=1 i=1 j=1
2.2 Hilbert Space Valued GRPs (I) 43

t ∈ I, ω ∈ S0 (R),
P∞respectively, where 0
P∞ uijk, gij ∈ A , i, j ∈ N.
k
Let uij = k=1 aij ψk , gij = k=1 bij ψk , i, j ∈ N. Then we have
∞ X
X ∞
P (R)u = P (R)uij ⊗ Hαj ei
i=1 j=1
∞ X
∞
à ∞
! ∞ X
∞ X
∞
X X X
= P (R) akij ψk ⊗ Hαj ei = akij P (λek )ψk ⊗ Hαj ei .
i=1 j=1 k=1 i=1 j=1 k=1
(2.24)
In order to solve (2.23) we may use the method of undetermined coefficients.
From (2.23) and (2.24) we have
X∞
akij P (λek )ψk = gij , i, j ∈ N.
k=1
Finally, we obtain the system
akij P (λek ) = bkij , i, j, k ∈ N.
b k
First case: If P (λek ) 6= 0 for all k ∈ N, then akij = P (λijf ) , i, j, k ∈ N, and the
k
solution of the equation is given by
∞ X ∞
̰ !
X X bkij
u= ψ ⊗ Hαj ei .
P (λek ) k
i=1 j=1 k=1

In this case the solution exists and it is unique. Note, there exists a constant
C > 0 such that P (λek ) ≥ C, for all k ∈ N. Thus,
°∞
∞ °X
°2 °∞ °2
X∞ X
bk ° 1 X∞ X∞ °X °
° ij ° −pαj ° ° j
° ψk ° (2N) ≤ 2 ° k
bij ψk ° (2N)−pα < ∞,
° e
P (λk ) ° C ° °
i=1 j=1 k=1 −r i=1 j=1 k=1 −r

for some p > 1, because g ∈ A(H)∗r .

Thus, u ∈ A(H)∗r .
Second case: Let P (λek ) = 0 for k = k1 , k2 , . . . km . Then a solution exists
if and only if bkij1 = bkij2 = · · · = bkijm = 0, i, j ∈ N. The solution in this case is
not unique, and the coefficients of the solution u are given by
X ∞ Xm
bkij
uij = ψk + csij ψks , i, j ∈ N,
k=1 P (λek ) s=1
f )6=0
P (λk

where csij , s = 1, 2, . . . m, i, j ∈ N, are arbitrary real numbers.

Remark. In case P (λek ) 6= 0, k ∈ N, the solution is unique in sense that
in Section 1.3.1 we fixed λ̃k = 1 for all k ∈ N such that λk = 0. But we
could have chosen also some other convention, thus in this sense the solution
actually depends on the eigenvalues of the operator R.
44 Generalized Stochastic Processes

2.3 The Wick Product of Generalized

Random Processes
It is a well-known problem that in general one can not define a pointwise
multiplication of generalized functions; thus, it is not clear how to deal with
nonlinearities. In framework of white noise analysis this difficulty is solved
by introducing the Wick product. We will consider the Wick product for
GRPs of type (I) and (O). Also, we carry over the Wick type product also to
the spaces of deterministic generalized functions. In Example 2.3.1 we will
show that in a special case (when the orthonoramal basis are trigonometric
functions) the Wick product coincides with the ordinary product. For GRPs
(II) the Wick product does not make sense, since GRPs (II) are generalized
in the t parameter and not in the ω parameter.

Wick products in spaces of generalized random variables

First we recall the definition and some basic properties of the Wick prod-
uct in the Kondratiev spaces (see [HØUZ]).

Definition
P 2.3.1 Let F, G ∈ P (S)−1 be given by their chaos expansion
F (ω) = α∈I fα Hα (ω), G(ω) = β∈I gβ Hβ (ω), fα , gβ ∈ R. The Wick prod-
uct of F and G is the unique element in (S)−1 defined by:
à !
X X
F ♦G(ω) = fα gβ Hγ (ω).
γ∈I α+β=γ

The Wick product is a commutative, associative operation, distributive

with respect to addition. By the same formula we defined in [PS1] the Wick
product for F, G ∈ exp(S)−1 . It is known that the spaces (S)1 , (S)−1 , exp(S)1
and exp(S)−1 are closed under Wick multiplication.
In the Hilbert space valued case, the Wick product is defined analogously
(see [MFA]); with a slight abuse of notation, we denote with the same symbol
the Wick products in (S)−1 and S(H)−1 .

Definition
P∞2.3.2
P Let F, G ∈ S(H)−1 bePgiven by their chaos expansion
∞ P
F (ω) = i=1 α∈I fi,α Hα (ω)ei , G(ω) = i=1 β∈I gi,β Hβ (ω)ei , fi,α , gi,β ∈
R. The Wick product of F and G is the unique element in S(H)−1 defined
2.3 The Wick Product of GRPs 45

by:
∞ X
à !
X X
F ♦G(ω) = fi,α gi,β Hγ (ω)ei
i=1 γ∈I α+β=γ
X∞
= (Fi ♦Gi (ω))ei ,
i=1
P P
where Fi (ω) = α∈I fi,α Hα (ω), Gi (ω) = β∈I gi,β Hβ (ω) ∈ (S)−1 .
This definition is legal, since S(H)1 and S(H)−1 are closed under Wick
multiplication. In the same manner we can define F ♦G for F, G ∈
exp S(H)−1 and it is an easy exercise to show (combining the methods in
[MFA] and [PS1]) that F ♦G ∈ exp S(H)−1 . Also, if F, G ∈ exp S(H)1 then
F ♦G ∈ exp S(H)1 .
Note that one can also Wick–multiply a R–valued and a H–valued gen-
eralized P
random variable: If F P ∈ (S)P −1 and G ∈ S(H)−1 have the forms
F (ω) = α∈I fα Hα (ω), G(ω) = ∞ g H (ω), then F ♦G ∈ S(H)−1
P ∞ P ³P
i=1 ´ i,β β
β∈I

has expansion i=1 γ∈I α+β=γ fα gi,β Hγ (ω)ei .

The Wick product of GRPs (O) is defined pointwisely (as in [HØUZ]) by
the formula à !
X X
F ♦G(ω) = fα (t)gβ (t) Hγ (ω),
γ∈I α+β=γ
P P
for F (t, ω) = α∈I fα (t)Hα (ω), G(t, ω) = β∈I gβ (t)Hβ (ω).

The Wick product in Zemanian spaces

Now we recall the definition of a deterministic multiplication of Wick
type in A0 , which was introduced in [Se] and [PS1]. From now on, when we
use Wick products, we will always assume that A is nuclear, i.e. there exists
some p ≥ 0, such that
X∞
M := fn −2p < ∞.
λ
n=1

Definition
P∞ 2.3.3 LetPf, g ∈ A0 be generalized functions given by expansions
f = k=1 ak ψk , g = ∞ k=1 bk ψk . Define f ¦ g to be the generalized function
0
from A given by  
∞
X X 
f ¦g =  ai bj  ψn . (2.25)
n=1 i,j∈N
i+j=n+1
46 Generalized Stochastic Processes

P P∞
Proposition 2.3.1 If f = ∞ i=1 ai ψi ∈ A−k and g = i=1 bi ψi ∈ A−l , then
f ¦ g ∈ A−(k+l+p) . Moreover,

|hf ¦ g, ϕi|2 ≤ M kf k2−k kgk2−l kϕkk+l+p , (2.26)

for each test function ϕ ∈ A.

Similarly, one can also define the multiplication of test-functions, under

an additional assumption:

Lemma 2.3.1 Let there exist a constant C > 0, such that

λg ee
i+j ≤ C λi λj , i, j ∈ N.
P P∞
If f = ∞i=1 ai ψi ∈ Ak and g = i=1 bi ψi ∈ Ak , then f ¦ g ∈ A(k−p) .

For example, we have ψi ¦ ψj = ψi+j−1 for arbitrary i, j ∈ N.

Example 2.3.1 Let I = (−π, π), R = −iD. Then, the orthonormal basis of
einx
A is given by the family of trigonometric functions ψn (x) = √ 2π
, and λn = n,
n = 0, ±1, ±2, ±3, . . . etc. In this case, we leave the enumeration running
through the set of integer numbers Z. Instead doing a reordering in the space
A, we make a slight modification in³ the definition ´ of the Wick product and
P∞ P
put instead of (2.25) f ¦ g = n=1 i,j∈Z ai bj ψn . Then,
i+j=n

1
ψi ¦ ψj = ψi+j , ψi · ψj = √ ψi+j i, j ∈ Z,
2π
i.e. the Wick product and the ordinary product coincide up to a constant.

The Wick product of GRPs (I)

The notion of the Wick product to GRPs (I) was also extended in [PS1].
We summarize the basic results here.
∗ ∗
Definition
P∞ 2.3.4 Let Φ ∈ Ak , Ψ ∈ Al be two GRPs P∞(I) given by expansions
Φ = i=1 fi ⊗ Hαi , fi ∈ A−k , i ∈ N, and Ψ = j=1 gj ⊗ Hαj , gj ∈ A−l ,
j ∈ N. The Wick product of Φ and Ψ is defined to be
 
∞
X X 
Φ¨Ψ =  fi ¦ gj  ⊗ Hαn . (2.27)
n=1 i,j∈N
αi +αj =αn+1
2.3 The Wick Product of GRPs 47

Theorem 2.3.1 The Wick product Φ¨Ψ from the previous definition is a
GRP (I), precisely Φ¨Ψ ∈ A∗k+l+p .

The Wick product for GRPs (I) taking values in exp(S)−1 can be defined
in an analogue way.

Theorem 2.3.2 The Wick product Φ¨Ψ defined by formula (2.27) of Φ ∈

exp ∗
Ak and Ψ ∈ exp A∗l is a GRP (I), precisely Φ¨Ψ ∈ exp A∗k+l+p .

The Wick product ¦ in A0 can be embedded into the Wick product ¨ in

A∗ , since each deterministic function can be (trivially) regarded as a stochas-
tic process. Also, the Wick product ¨ acting on A∗ is an extension of the
Wick product ♦ acting on (S)−1 .

2.3.1 The Wick product of H–valued GRPs (I)

Now, the main difference is that we are not able to define the Wick
product ¨ for the whole class of H–valued GRPs (I). This is due to the fact
that S(H)−1 is not a nuclear space. Therefore, the Wick product will be
defined for the class of H–valued GRPs (I) satisfying condition (2.15).
∗ ∗
Definition 2.3.5 Let Φ ∈PA(H) P k , Ψ ∈ A(H)l be two H–valued GRPs (I)
givenP
by expansions Φ= ∞ ∞
j=1 fi,j ⊗ Hαj ei , fi,j ∈ A−k , i, j ∈ N, and
∞ P∞
i=1
Ψ = i=1 j=1 gi,j ⊗ Hαj , gi,j ∈ A−l , i, j ∈ N, and let r1 ≥ 0, r2 ≥ 0 be such
P P∞ P P∞
that ∞ i=1
2
j=1 kfi,j k−k (2N)
−r1 αj
< ∞ and ∞ i=1
2
j=1 kgi,j k−l (2N)
−r2 αj
<
∞. The Wick product of Φ and Ψ is defined to be
 
∞
XX X∞

Φ¨Ψ =  fi,s ¦ gi,r  ⊗ Hαn ei . (2.28)
i=1 n=1 s,r∈N
αs +αr =αn+1

We may write (2.28) also as

∞
X
Φ¨Ψ = (Fi ¨Gi )ei ,
i=1
P P∞
where Fi = ∞ ∗
j=1 fi,j ⊗ Hαj ∈ Ak , Gi =
∗
j=1 gi,j ⊗ Hαj ∈ Al and Fi ¨Gi is
defined as in Definition 2.3.4.

Theorem 2.3.3 The Wick product Φ¨Ψ from the previous definition is an
H–valued GRP (I), precisely Φ¨Ψ ∈ A(H)∗k+l+p .
48 Generalized Stochastic Processes

Proof. Due toPLemma 2.3.1 it follows that fi,s ¦gi,r ∈ A−k−l−p for all i, s, r ∈
N,
P∞and Pthus, αs +αr =αn+1 fi,s ¦ gi,r ∈ A−k−l−p , n ∈ N. Since [Φ¨Ψ, ϕ] =
∞ P
i=1 n=1 αs +αr =αn+1 hfi,s ¦ gi,r , ϕiHαn ei , it remains to prove
¯ ¯2
∞ X
X ∞ ¯ X ¯
¯ ¯ n
sup ¯ hfi,s ¦ gi,r , ϕi¯ (2N)−r3 α < ∞,
ϕ∈B n=1
¯ ¯
i=1 αs +αr =α n+1

for any bounded set B ⊆ A and some r3 ≥ 0. Put r3 = r1 + r2 + q, where

q > 1. Then, according to (2.26)
∞ ¯
¯ ¯2
∞ X
X X ¯
¯ ¯ n
sup ¯ hfi,s ¦ gi,r , ϕi¯ (2N)−r3 α
ϕ∈B ¯ s r n+1 ¯
i=1 n=1 α +α =α

¯ ¯2
∞ X
X ∞ ¯ X ¯
¯ √ ¯ n
≤ sup ¯ kfi,s k−k kgi,r k−l kϕkk+l+p M ¯ (2N)−r3 α ≤
ϕ∈B n=1
¯ ¯
i=1 αs +αr =αn+1

¯ ¯2
∞
X ∞ ¯
X X ¯
¯ r αs r αr ¯
−qαn − 12 − 22
M (2N) ¯ kfi,s k−k (2N) kgi,r k−l (2N) sup kϕkk+l+p ¯
n=1
¯ ϕ∈B ¯
i=1 αs +αr =αn+1

∞ ∞
Ã∞ ∞
!
X X X X
−qαn 2 −r1 αs 2 −r2 αr 2
≤M (2N) kfi,s k−k (2N) kgi,r k−l (2N) sup kϕkk+l+p
ϕ∈B
n=1 i=1 s=1 r=1

∞
X ∞ X
X ∞ ∞ X
X ∞
n s r
=M (2N)−qα kfi,s k2−k (2N)−r1 α kgi,r k2−l (2N)−r2 α sup kϕk2k+l+p
n=1 ϕ∈B
i=1 s=1 i=1 r=1
s +αr s r
< ∞, where we used the property (2N)α = (2N)α (2N)α . ¤

Note that the Wick product ♦ acting on S(H)−1 can be embedded into
¨ acting on A(H)∗ .
Also, following theorem holds, similarly as in finite dimensional case.

Theorem 2.3.4 The Wick product Φ¨Ψ defined by formula (2.28) of Φ ∈

exp
A(H)∗k and Ψ ∈ exp A(H)∗l is a H–valued GRP (I), precisely Φ¨Ψ ∈
exp
A(H)∗k+l+p , provided there exist r1 ≥ 0, r2 ≥ 0 such that
P∞ P∞ j P P∞ j

i=1
2
j=1 kfi,j k−k e
−r1 (2N)α
< ∞ and ∞ i=1
2 −r2 (2N)α
j=1 kgi,j k−l e < ∞.
2.3 The Wick Product of GRPs 49

2.3.2 A Class of nonlinear SDEs

Now we will consider a class of a nonlinear SDEs with Wick products
involving H–valued GRPs (I). Here we illustrate a simple method of solving
by means of series expansions.
Let P (R) be the differential operator defined as in (2.23). Assume that
P (λk ) − 1 6= 0 and P (λek ) 6= 0 for all k ∈ N. Consider a nonlinear SDE of the
e
form
P (R)X = X¨W̃ + g, (2.29)
where W̃ is singular white noise in A(H)∗0 given by the expansion
∞ X
X ∞
W̃ (t, ω) = wi,n (t) ⊗ Hεn (ω)ei , t ∈ I, ω ∈ Ω,
i=1 n=1

where ½
ψj (t), n = m(i, j)
wi,n =
0, else
and m(i, j) is defined as in Example 1.7.3.
Let g ∈ A(H)∗r be of the form
∞ X
X ∞
g(t, ω) = gi,n (t) ⊗ Hεn (ω)ei , t ∈ I, ω ∈ Ω,
i=1 n=1

such that gi,n ∈ A−r , i, n ∈ N; and assume there exist q ≥ 0, such that
∞ X
X ∞ ∞ X
X ∞
kgi,n k2−r (2N)−qεn = kgi,n k2−r (2n)−q < ∞. (2.30)
i=1 n=1 i=1 n=1

Expanding each gi,n in A−r we get

∞
X
gi,n (t) = gi,n,k ψk (t), gi,n,k ∈ R, t ∈ I, i, n ∈ N,
k=1

which yields
∞
X −2r
|gi,n,k |2 λek < ∞. (2.31)
k=1

We will look for the solution X of the equation in the form

∞ X
X ∞
X(t, ω) = ai,n (t) ⊗ Hεn (ω)ei , t ∈ I, ω ∈ Ω,
i=1 n=1
50 Generalized Stochastic Processes

where ai,n ∈ A0 , i, n ∈ N, are the coefficients to be determined. Let ai,n (t),

t ∈ I, be given by expansion
∞
X
ai,n (t) = ai,n,k ψk (t), ai,n,k ∈ R, t ∈ I, n ∈ N.
k=1

Then,

∞ X
∞
̰ !
X X
P (R)X(t, ω) = ai,n,k P (λek )ψk (t) ⊗ Hεn (ω)ei , t ∈ I, ω ∈ Ω,
i=1 n=1 k=1
(2.32)
and due to the definition of Wick product:

X(t, ω)¨W̃ (t, ω) + g(t, ω) =

∞ X∞
à !
X X
ai,s (t) ¦ wi,r (t) + gn (t) ⊗ Hεn (ω)ei , t ∈ I, ω ∈ Ω.
i=1 n=1 r+s=n+1
(2.33)

Since
∞
X ∞
X
ai,s (t) ¦ wi,r (t) = ai,s,k ψk (t) ¦ wi,r (t) = ai,s,k ψk (t) ¦ ψj (t)
k=1 k=1
∞
X
= ai,s,k ψk+j−1 (t),
k=1

for j ∈ N such that r = m(i, j), relation (2.33) becomes

X(t, ω)¨W̃ (t, ω) + g(t, ω) =

∞ X∞
à ∞
!
X X X
ai,s,k ψk+j−1 (t) + gn (t) ⊗ Hεn (ω)ei , t ∈ I, ω ∈ Ω.
i=1 n=1 r+s=n+1 k=1
(2.34)

From (2.32) and (2.34) we obtain

X ∞
X X∞ ∞
X
ai,s,k ψk+j−1 (t)+ gi,n,k ψk (t) = ai,n,k P (λek )ψk (t), i, n ∈ N.
s,r∈N k=1 k=1 k=1
s+r=n+1,r=m(i,j)

(2.35)
2.3 The Wick Product of GRPs 51

From the system of equations (2.35) one can recursively determine the
coefficients ai,n,k , i, n, k ∈ N.
For i = 1, n = 1 we have only one possibility how to get r + s = 2
(s = 1, r = 1) and for r = 1 the corresponding j is j = 1. Thus, (2.35) gives

∞
X ∞
X ∞
X
a1,1,k ψk (t) + g1,1,k ψk (t) = a1,1,k P (λek )ψk (t), t ∈ I,
k=1 k=1 k=1

which implies
g1,1,k
a1,1,k = , k = 1, 2, . . . (2.36)
P (λek ) − 1

For i = 1, n = 2 we have two possibilities how to get r+s = 3 (s = 1, r = 2

and s = 2, r = 1). For r = 1 we have j = 1, and for r = 2 we also get j = 1.
Thus, from (2.35) we obtain

∞
X ∞
X ∞
X
a1,1,k ψk (t) + g1,2,k ψk (t) = a1,2,k P (λek )ψk (t), t ∈ I,
k=1 k=1 k=1

and since a1,1,k are known from the previous step, now we get

a1,1,k + g1,2,k
a1,2,k = , k = 1, 2, . . .
P (λek )

For i = 1, n = 3 and consequently r + s = 4 we get following triples:

s = 1, r = 3, j = 2; s = 3, r = 1, j = 1 and s = 2, r = 2, j = 1. Thus,

∞
X ∞
X ∞
X ∞
X ∞
X
a1,1,k ψk+1 + a1,3,k ψk + a1,2,k ψk + g1,3,k ψk = a1,3,k P (λek )ψk .
k=1 k=1 k=1 k=1 k=1

After reordering the indeces in the first sum we get

a1,2,1 + g1,3,1
a1,3,1 = ,
P (λe1 ) − 1
a1,1,k−1 + a1,2,k + g1,3,k
a1,3,k = , k = 2, 3, . . .
P (λek ) − 1

We follow this schedule for n = 4, 5, . . .. Then we fix i = 2 and obtain

52 Generalized Stochastic Processes

the coefficients for n = 1, 2, 3, . . . given by following recursion formulae:

g2,1,k
a2,1,k = , k = 1, 2, . . .
P (λek ) − 1
a2,1,k + g2,2,k
a2,2,k = , k = 1, 2, . . .
P (λek )
a2,2,1 + g2,3,1
a2,3,1 = ,
P (λe1 ) − 1
a2,1,k−1 + a2,2,k + g2,3,k
a2,3,k = , k = 2, 3, . . .
P (λek ) − 1
Then we fix i = 3, and so on....
Since for each r ∈ N its corresponding j ∈ N from r = m(i, j) always
satisfies j < r, we get a general formula
gi,n,k + L
ai,n,k = , i ∈ N, n ∈ N; k = n, n + 1, . . . (2.37)
P (λek ) − 1
where L is a linear combination of ai,n1 ,k1 for n1 < n, k1 < k.
Since P (λek ) − 1 6= 0, k ∈ N, there exists a constant K > 0 such that
|P (λek ) − 1| ≥ K, k ∈ N. Relation (2.31) yields that there exists C(i, n) > 0,
such that r
|gi,n,k | ≤ C(i, n)λek , k ∈ N.
Similarly, according to (2.30) there exists D > 0, such that
∞
X −2r
kgi,n k2−r = |gi,n,k |2 λek ≤ D(2n)q , i, n ∈ N.
k=1

Hence, for each i, n ∈ N we have C(i, n) ≤ D(2n)q and consequently

r
|gi,n,k | ≤ D(2n)q λek , i, n, k ∈ N. (2.38)

We prove now the estimate

r 1
|ai,n,k | ≤ D(2n)q λek Qn ( ), i, n ∈ N, k ≥ n, (2.39)
K
where Qn is a polynomial of order n. The proof can be done by induction.
r
For i = n = 1 we get from (2.36) and (2.38) that |a1,1,k | ≤ K1 D2q λek , k ∈ N.
Assume now (2.39) holds. Then, from (2.37) and (2.38) we get
1 1 ³ q g
r
´
|ai+1,n+1,k | ≤ (|L| + |gi+1,n+1,k |) ≤ |L| + D(2(n + 1)) λk+1 .
K K
2.4 Colombeau Algebras for GRPs (I) and (O) 53

Since L is a linear combination of n coefficients ai,n1 ,k1 , using the induction

r r
hypothesis we get |L| ≤ D(2n)q λek Qn ( K1 ). Thus, since λek ≤ λ er , and
k+1
q q
(2n) ≤ (2(n + 1)) , we obtain
µ ¶
q g 1 1 1
r r
|ai+1,n+1,k | ≤ D(2(n+1)) λk+1 1 + Qn ( ) = D(2(n+1))q λg k+1 Rn+1 ( ),
K K K

where Rn+1 is some polynomial of order n + 1. This proves (2.39).

Let p be given as in (2.3). Then, due to (2.39)
∞
X ∞
−2r−2p 1 X e −2p
|ai,n,k | λek
2
≤ D2 (2n)2q Q2n ( ) λk < ∞,
k=1
K k=1

and thus,
kai,n k2−(r+p) < ∞, i, n ∈ N,
which yields that all coefficients ai,n , i, n ∈ N, belong to A−(r+p) .
Also, it holds that,
∞ X
X ∞ ∞ X
X ∞
εn 1 −2sn
kai,n k2−(r+p) e−s(2N) 2
≤D M (2n)2q Q2n ( )e .. (2.40)
i=1 n=1 i=1 n=1
K

The series on the right-hand side of (2.40) can be made convergent if we

choose s large enough. Thus, there exists a solution X of equation (2.29) in
the space exp A(H)r+p∗
. Since every GRP (I) is uniquely determined by the
coefficients in its expansion, it follows that the solution is unique. (Again,
we assume the fixed convention λ̃k = 1 for λk = 0).

2.4 Colombeau Algebras for Generalized

Random Processes of Type (I) and (O)
In this section we construct Colombeau type extensions for the spaces of
generalized random processes; we consider the case of GRPs (I) and (O).

2.4.1 Colombeau extension of L(A, (S)−1 )

P e −2s <
Throughout this section we assume that A is nuclear (M = ∞ i=1 λi
∞ for some s > 0) and that there exists a constant C > 0, such that
λg ee
i+j ≤ C λi λj , i, j ∈ N. These assumptions are necessary for the Wick–
multiplication of Zemanian test functions (see Lemma 2.3.1).
54 Generalized Stochastic Processes

• Let EM (A; (S)−1 ) be the

Pvector space of functions R : (0, 1) → A ⊗
(S)−1 , ² 7→ R² (t, ω) = α∈I rα,² (t) ⊗ Hα (ω), rα,² ∈ A, t ∈ I, ω ∈ Ω,
such that for every k ∈ N0 there exist a sequence {Cα }α∈I of positive
numbers, ²0 ∈ (0, 1), p ∈ N0 , and there exists a sequence {aα }α∈I
bounded from above (i.e. there exists a ∈ R such that aα ≤ a, α ∈ I),
with following properties:

krα,² kk ≤ Cα ²−aα , for all α ∈ I, ² < ²0 , (2.41)

X
Cα2 (2N)−pα < ∞. (2.42)
α∈I

Clearly, from (2.41) and (2.42) we have

X X X
krα,² k2k (2N)−pα ≤ Cα2 ²−2aα (2N)−pα ≤ ²−2a Cα2 (2N)−pα = K²−2a ,
α∈I α∈I α∈I
P
for ² < ²0 ,, where K = α∈I Cα2 (2N)−pα .
• Let N(A; (S)−1 ) denote the vector space of functions R² ∈
EM (A; (S)−1 ) with the property that for every k ∈ N0 there exist a
sequence {Cα }α∈I of positive numbers, ²0 ∈ (0, 1), p ∈ N0 , and for all
sequences {aα }α∈I bounded from below (i.e. there exists a ∈ R such
that aα ≥ a, α ∈ I), following hold:

krα,² kk ≤ Cα ²aα , for all α ∈ I, ² < ²0 , (2.43)

X
Cα2 (2N)−pα < ∞. (2.44)
α∈I

Clearly, from (2.43) and (2.44) we have

X X X
krα,² k2k (2N)−pα ≤ Cα2 ²2aα (2N)−pα ≤ ²2a Cα2 (2N)−pα = K²2a ,
α∈I α∈I α∈I
P
for ² < ²0 , where K = α∈I Cα2 (2N)−pα .
• Define the multiplication in EM (A;P (S)−1 ) and N(A; (S)−1 ) in the
following
P way: For F² (x, ω) = α∈I fα,² (x) ⊗ Hα (ω), G² (x, ω) =
α∈I gα,² (x) ⊗ Hα (ω) let
à !
X X
F² ♦G² (x, ω) = fα,² (x) ¦ gα,² (x) ⊗ Hγ (ω), (2.45)
γ∈I α+β=γ
2.4 Colombeau Algebras for GRPs (I) and (O) 55

i.e. we use the Wick product for multiplication in (S)−1 and in A as

well. We recall following estimate
√ from [Se] (see also Lemma 2.3.1): If
k
f, g ∈ Ak , then kf ¦ gkk−s ≤ C M kf kk kgkk .

Lemma 2.4.1 (i) EM (A; (S)−1 ) is an algebra under the multiplica-

tion rule given by (2.45).
(ii) N(A; (S)−1 ) is an ideal of EM (A; (S)−1 ).

Proof. (i) Let F² , G² ∈ EM (A; (S)−1 ) and prove that P F² ♦G² ∈

EM (A; (S)−1 ). Let k ∈ N0 be arbitrary.P Since F² = α∈I fα,² (x) ⊗
Hα (ω) ∈ EM (A; (S)−1 ) and G² = β∈I gβ,² (x)⊗Hβ (ω) ∈ EM (A; (S)−1 ),
for k̃ = k + s there exist a, b ∈ R, p, q ∈ N0 , ²1 , ²2 ∈ (0, 1),
and there exist sequences {Cα }α∈I , {Dα }α∈I , {aα }α∈I , {bα }α∈I , such
−aα
that aα ≤ a, bα ≤ b for all α ∈ I, kfP α,² kk+s ≤ Cα ² for
−bα 2 −pα
²P< ²1 , kgα,² kk+s ≤ Dα ² for ² < ²2 , and α∈I Cα (2N) < ∞,
2 −qα
α∈I D α (2N) < ∞. We will prove that for ² 0 = min{² 1 2 there
, ² }
exist r > 0, a sequenceP {Mγ }γ∈I and a sequence {mγ }γ∈I bounded from
above such that k α+β=γ fα,² (x) ¦ gα,² (x)kk ≤ Mγ ²−mγ for ² < ²0 and
P 2 −rα
γ∈I Mγ (2N) < ∞.
√ P
For a fixed multiindex γ ∈ I put Mγ = C k M α+β=γ Cα Dβ and
mγ = max{aα + bβ : α, β ∈ I, α + β = γ} (note, for γ fixed, there are
only finite many α and β which give the sum γ). Now using the fact
that (A, ¦) is an algebra, we get
X √ X
k fα,² (x) ¦ gα,² (x)kk ≤ C k M kfα,² (x)kk+s kgα,² (x)kk+s ≤
α+β=γ α+β=γ
√ X √ X
Ck M Cα Dβ ²−(aα +bβ ) ≤ ²−mγ C k M Cα Dβ = ²−mγ Mγ .
α+β=γ α+β=γ

Clearly, mγ ≤ a + b, γ ∈ I, thus the sequence {mγ }γ∈I is bounded from

above. Let r = p+q +2. Then, using the nuclearity of (S)−1 , we obtain
à !2
X X X
Mγ2 (2N)−rα ≤ C 2k M Cα Dβ (2N)−(p+q+2)γ
γ∈I γ∈I α+β=γ
X X X
≤ C 2k M (2N)−2γ Cα2 (2N)−pα Dβ2 (2N)−qβ < ∞.
γ∈I α∈I β∈I

(ii) Let us check first that N(A; (S)−1 ) is a subalgebra of EM (A; (S)−1 ).
Let F² , G² ∈ N(A; (S)−1 ) be of the form as in (i). Let k ∈ N0
56 Generalized Stochastic Processes

and {mγ }γ∈I be an arbitrary sequence bounded from below. Put

aγ = bγ = m2γ , γ ∈ I. Now for the sequences {aγ }γ∈I , {bγ }γ∈I also
bounded from below, and for k̃ = k + s, since F² , G² ∈ N(A; (S)−1 ),
there must exist p, q ∈ N0 , ²1 , ²2 ∈ (0, 1), and {Cα }α∈I , {Dα }α∈I
such that kfα,²Pkk+s ≤ Cα ²mα /2 for ² <P²1 , kgα,² kk+s ≤ Dα ²mα /2 for
2 −pα 2 −qα
² < ²2 , and α∈I Cα (2N) < ∞, α∈I Dα (2N)√ < ∞. Let
k
P
²0 = min{²1 , ²2 }, r = p + q + 2 and Mγ =P C M α+β=γ Cα Dβ ,
γ ∈ I. Now we may proceed as in (i) to get α+β=γ kfα,² ¦ gα,² kk ≤
√ P P
C k M α+β=γ kfα,² kk+s kgα,² kk+s ≤ Mγ ²mγ and 2
γ∈I Mγ (2N)
−rα
<
∞.
In order to prove that N(A; (S)−1 ) is an ideal of EM (A; (S)−1 ), we
must check that for all G² ∈ N(A; (S)−1 ) and all F² ∈ EM (A; (S)−1 ),
we have G² ♦F² = F² ♦G² ∈ N(A; (S)−1 ). Let k ∈ N0 and {nγ }γ∈I be
an arbitrary sequence bounded from below (i.e. nγ ≥ n, γ ∈ I). Since
F² ∈ EM (A; (S)−1 ), there exist p ∈ N0 , ²1 ∈ (0, 1), {Dβ }β∈I , b ∈ R and
a sequence
P 2 {bβ }β∈I such that bβ ≤ b, β ∈ I, kfβ,² kk+s ≤ Dβ ²−bβ , ² < ²1 ,
−pβ
and β Dβ (2N) < ∞.
For a fixed multiindex α ∈ I, let aα = b + nα . Clearly, the se-
quence {aα }α∈I is bounded from below by a = b + n, thus since
G² ∈ N(A; (S)−1 ), there exist q ∈ N0 , ²P
2 ∈ (0, 1), {Cα }α∈I , such
−aα 2 −qα
that kgα,² kk+s ≤ Cα ² , ² √< ²2 , and α Cα (2N) < ∞. Let
k
P
²0 = min{²1 , ²2 } and Nγ = C M α+β=γ Cα Dβ , γ ∈ I.
P P
Now for G² ♦F² = γ∈I α+β=γ gα,² ¦ fα,² ⊗ Hγ we have that
X X
kgα,² kk+s kfβ,² kk+s ≤ Cα Dβ ²aα −bβ
α+β=γ α+β=γ

X X 1
≤ Cα Dβ ²b+nα −bβ ≤ Cα Dβ ²bβ +nα −bβ ≤ C −k M − 2 Nγ ²nγ , ² < ²0 .
α+β=γ α+β=γ
P
It is clear that γ∈I Nγ2 (2N)−rγ < ∞ for r = p + q + 2. Thus, G² ♦F² ∈
N(A; (S)−1 ). ¤

• The quotient space

G(A; (S)−1 ) = EM (A; (S)−1 ) / N(A; (S)−1 )

is the (S)−1 –valued Colombeau extension of A, and its elements are

Colombeau generalized random processes. We denote the elements of
G(A; (S)−1 ) (equivalence classes) by [R² ].
2.4 Colombeau Algebras for GRPs (I) and (O) 57

Note that (S)−1 can be embedded into G(A; (S)−1 ) by

X X
(S)−1 3 F (ω) = fα Hα (ω) Ã fα (t) ⊗ Hα (ω) ∈ G(A; (S)−1 ),
α∈I α∈I

where fα (t) = fα is the constant mapping.

Remark. In some special cases, one can take the ordinary product in-
stead of the Wick product in the multiplication rule (2.45). If A is the
Schwartz space S(R) of rapidly decreasing functions (see Example 1.3.1),
then (S(R), ·) is also an algebra and one can construct a Colombeau algebra
G(S(R); (S)−1 ) in the same way but with following multiplication rule:
à !
X X
F² ♦G² (x, ω) = fα,² (x) · gα,² (x) ⊗ Hγ (ω), (2.46)
γ∈I α+β=γ

The same holds true if A is generated by the operator R = −iD (see Example
2.3.1). In this case the Wick product and the ordinary product coincide up
to a constant.

2.4.2 Colombeau extension of C ∞ (I; (S)−1 )

In the previous case we constructed a Colombeau extension for the
space of GRPs (I). Now, we consider the space of infinitely differentiable
∞
GRPs (O) i.e. CP (I; (S)−1 ). Recall that F ∈ C ∞ (I; (S)−1 ) if it is of the
∞
form F (t, ω) =
P α∈I fα (t)Hα (ω), fα ∈ C (I), and for each t ∈ I fixed,
2 −pα
α∈I |fα (t)| (2N) < ∞ for some p ≥ 0.
The (usual) topology on C ∞ (I) is the topology of uniform convergence
of the function and its partial derivatives of each order on compact subsets
of I. Endowed with this topology C ∞ (I) is a nuclear space.
Note that since I is an open subset of Rd and (S)−1 is complete, we have
(see [Tr, Theorem 44.1]) C ∞ (I; (S)−1 ) ∼ = C ∞ (I) ⊗ (S)−1 and C0∞ (I) ⊗ (S)−1
is dense in C ∞ (I; (S)−1 ).
For example, singular white noise and Brownian motion defined in Ex-
ample 1.7.2 and Example 1.7.1 are infinitely differentiable GRPs (O).
The algebra now to be constructed is the (S)−1 –valued version of the
simplified Colombeau algebra for deterministic generalized functions. Recall
that the simplified Colombeau algebra was also constructed on C ∞ (I).

vector space of functions R : (0, 1) → C ∞ (I) ⊗

• Let EM (I; (S)−1 ) be theP
(S)−1 , ² 7→ R² (t, ω) = α∈I rα,² (t)Hα (ω), rα,² ∈ C ∞ (I), t ∈ I, ω ∈ Ω,
such that for each compact subset Uk ⊂⊂ I, k ∈ N, there exist a
58 Generalized Stochastic Processes

sequence {Cα }α∈I of positive numbers, ²0 ∈ (0, 1), p ∈ N0 , and there

exists a sequence {aα }α∈I bounded from above with the property that
for all |β| ≤ k,

sup |Dβ rα,² (t)| ≤ Cα ²−aα , for all α ∈ I, ² < ²0 , (2.47)

t∈Uk
X
Cα2 (2N)−pα < ∞. (2.48)
α∈I

• Let N(I; (S)−1 ) denote the vector space of functions R² ∈ EM (I; (S)−1 )
with the property that for each compact subset Uk ⊂⊂ I, k ∈ N, there
exist a sequence {Cα }α∈I of positive numbers, ²0 ∈ (0, 1), p ∈ N0 ,
and for all sequences {aα }α∈I bounded from below, and for all |β| ≤ k
following hold:

sup |Dβ rα,² (t)| ≤ Cα ²aα , for all α ∈ I, ² < ²0 , (2.49)

t∈Uk
X
Cα2 (2N)−pα < ∞. (2.50)
α∈I

• Define the multiplication in EMP (I; (S)−1 ) and N(I; (S)−1 ) in the
following
P way: For F² (x, ω) = α∈I fα,² (x) ⊗ Hα (ω), G² (x, ω) =
α∈I gα,² (x) ⊗ Hα (ω) let
à !
X X
F² ♦G² (x, ω) = fα,² (x) · gα,² (x) ⊗ Hγ (ω), (2.51)
γ∈I α+β=γ

i.e. we use the Wick product for multiplication in (S)−1 and the ordi-
nary product in C ∞ (I). Clearly, (C ∞ (I), ·) is an algebra.

• Then N(I; (S)−1 ) is an ideal of EM (I; (S)−1 ) and the quotient space

G(I; (S)−1 ) = EM (I; (S)−1 ) / N(I; (S)−1 ),

the (S)−1 –valued Colombeau extension of C ∞ (I) is also an algebra. Its

elements are Colombeau generalized random processes.

Note that (S)−1 can be embedded into G(I; (S)−1 ) by

X X
(S)−1 3 F (ω) = fα Hα (ω) Ã fα (t) ⊗ Hα (ω) ∈ G(I; (S)−1 ),
α∈I α∈I

where fα (t) = fα is the constant mapping.

2.5 GRPs of Type (II) 59

The space G(I; (S)−1 ) is constructed so that it involves singular GRPs

(O) i.e. GRPs (I). Like in deterministic case, where S0 (I) can be embedded
into G(I), here L(S(I), (S)−1 ) = S0 (I) ⊗ (S)−1 can also be embedded into
G(I; (S)−1 ) via convolution. For a fixed mollifier function ρ² , ² ∈ (0, 1) we
have that if f ∈ S(I) ⊗ (S)−1 , then f² (t, ω) = f (·, ω) ∗ ρ² (t) ∈ S(I) ⊗ (S)−1
and clearly, f − f² ∈ N(I; (S)−1 ). Thus, [f² ] ∈ G(I; (S)−1 ). The embedding
ιρ : S(I) ⊗ (S)−1 → G(I; (S)−1 ), f 7→ [f² ], can be extended to an embedding
ιρ : S0 (I) ⊗ (S)−1 → G(I; (S)−1 ), defined by

S0 (I) ⊗ (S)−1 3 F 7→ [F ∗ ρ² ] ∈ G(I; (S)−1 ).

Also, G(I) can be embedded into G(I; (S)−1 ) by

X
G(I) 3 f = [f² (t)] Ã [f² (t)]Hα (ω) ∈ G(I; (S)−1 ).
α=(0,0,0,...)

In Chapter 3 we will construct also some other algebras as polynomial

Colombeau extensions of GRP (I) spaces; especially we will make use of
Sobolev spaces instead of the Zemanian test function spaces. We also note
that this Colombeau–type construction cannot be carried out for H–valued
GRPs (I): Since S(H)−1 is not nuclear, Lemma 2.4.1 need not hold true.

2.5 Generalized Random Processes

of Type (II)
The results presented in this section were mainly subject of [PS2]. We
begin with an example showing that the classes of GRPs of type (I) and (II)
are not contained in each other. As main topic, we examine the structure
representation of GRPs (II) and their correlation operators on the space of
test functions K{Mp }. In order to analyze assumptions for general repre-
sentation theorems, we give in Theorem 2.5.1 a representation theorem for
a GRP (II) on Lr (G), r > 1, where G is an open interval in Rn . Analyzing
the continuity condition |ξ(ω, ϕ)| ≤ C(ω)kϕkp(ω),2 , φ ∈ K{Mp }, ω ∈ Ω, for a
GRP (II), we obtain a structural characterization for it and for its correla-
tion operator on K{Mp }. More precisely, we show that for each ε > 0 there
exists a set M with measure P (M ) ≥ 1 − ε, and there exist d ∈ N0 such
that |ξ(ω, ϕ)| ≤ C(ω)kϕkd,2 , φ ∈ K{Mp }, ω ∈ M . Thus, if C ∈ Lp (Ω), this
assures the continuity of ξ as a mapping K{Mp } → Lp (Ω), where Lp (Ω) is
equipped with convergence in mean of order p for p ≥ 1, and convergence in
probability for p = 0.
60 Generalized Stochastic Processes

Especially, in Theorem 2.5.3 we consider Gaussian GRPs, and using an-

other technique obtain generalized Gaussian GRPs (II) on K{Mp } as a sum
of derivatives of classical Gaussian processes, which have appropriate growth
rate at |t| → ∞.
Several examples show that the assumptions which we made are substan-
tial for the characterization of a GRP (II).
In order to motivate our considerations and demonstrate the preferences
of our structural theorems, we will present the simplest equation y 0 = f in
the frame of GRPs (II). The notation we will use further on is given below.
Let f : Ω × K{Mp } → R be such that for all ϕ ∈ K{Mp }, f (·, ϕ) is
a Gaussian random variable, and for every ω ∈ Ω, f (ω, ·) is an element in
K0 {Mp }. Let ω 7→ X(ω), ω ∈ Ω, be a Gaussian random variable. Consider
the stochastic differential equation

d
y(ω, t) = f (ω, t), y(ω, 0) = X(ω), (2.52)
dt
where the value at zero, y(ω, t) at t = 0, is understood in the sense of
Lojasiewicz.
To solve (2.52) we use the decomposition of K{Mp }: Every φ ∈ K{Mp }
is of the form
Z
φ = Cφ φ0 + ψ, where φ0 , ψ ∈ K{Mp }, Cφ = φ(t)dt,
R

Z
d
φ0 = 1, and ψ = θ, for some θ ∈ K{Mp }.
R dt
Now, fixing ω ∈ Ω, and using the continuity of f (ω, ·) : K{Mp } → R, we
solve this equation in the usual way:

y(ω, φ) = Cφ X(ω) − hf (ω, t), θ(t)i, φ ∈ K{Mp }. (2.53)

But now we obtain that

limhf (ω, εt), θ(t)i = 0, for θ ∈ K{Mp } (2.54)

ε→0

is a necessary condition for the existence of y.

Another procedure is to use the representation stated in Theorem 2.5.3,
which can be used in more general cases: Under certain conditions on f ,
which are weaker than (2.54), we can still find a solution y on a somewhat
smaller set M ⊂ Ω. We will show this explicitly in Corollary 2.5.1.
2.5 GRPs of Type (II) 61

GRPs (II) vs. GRPs (I)

As already defined at the beginning of Chapter 1, we denote by Z 2 the
space of random variables with finite second moment. Recall now the defini-
tion of GRPs of type (II), which we study in this section:
Definition 2.5.1 A GRP (II) is a mapping ξ : Ω × V → C such that for ev-
ery ϕ ∈ V, ξ(·, ϕ) is a complex random variable and for every ω ∈ Ω, ξ(ω, ·)
is an element in V 0 .
Example 2.5.1 Let V be a separable Hilbert space and (φn )n∈N be a complete
orthogonal system in V converging to zero in V – e.g. if (en )n∈N is a CONS
in V , then we can take φn = enn , n ∈ N. We will choose a sequence of random
variables (Xn )n∈N in Z 2 satisfying certain conditions - depending on whether
we want to construct a GRP (I) or (II) – and use the following construction:
Define a GRP on Ω × {φi : i ∈ N} by putting ξ(·, φn ) = Xn , n ∈ N, and
extending this by linearity and continuity (type of continuity depending on
the conditions we implied on the sequence Xn ) onto V .
(a) We construct a GRP of type (I) but not of type (II).
Let (Xn )n∈N be a sequence of random variables in Z 2 such that:
• E(Xn Xm ) = 0 for n 6= m,
• kXn kZ 2 ≤ kφn kV , n ∈ N,
• (Xn )n∈N does not converge to zero almost surely.
Note that from the second condition we have that (Xn )n∈N converges to
zero in Z 2 . Such a sequence exists, e.g. one can take a sequence of
independent random variables with distribution law
µ ¶
−1 0 1
Xn : 1 , n ∈ N.
2n
1 − n1 2n1

The extension by continuity onto V is well defined (we take convergence

in Z 2 ), since
∞
X ∞
X ∞
X
kξ(ω, αi φi )k2Z 2 =k αi Xi (ω)k2Z 2 = αi2 kXi k2Z 2
i=1 i=1 i=1
∞
X ∞
X
≤ αi2 kφi k2V = k αi φi k2V .
i=1 i=1
For any sequence (ϕn )n∈N converging to zero in V we have
kξ(ω, ϕn )kZ 2 ≤ kϕn kV , thus ξ is a GRP (I). But, |ξ(ω, φn )| = |Xn (ω)|
/ V 0.
does not converge to zero for a.e. ω ∈ Ω, thus ξ(ω, ·) ∈
62 Generalized Stochastic Processes

(b) We construct a GRP of type (II) but not of type (I).

Let (Xn )n∈N be a sequence of random variables in Z 2 such that:

• (Xn )n∈N does not converge to zero in Z 2 .

• P (Ω0 ) = 1, where
Ω0 = {ω ∈ Ω : (∃k0 ∈ N)(∀k ∈ N)(k ≥ k0 ⇒ Xk (ω) = 0)}.

For example, the sequence of random variables with distribution law

µ ¶
0 n
Xn : , n ∈ N.
1 − n12 n12

meets both conditions given above.

The extension by continuity onto
P∞V is well defined (we take convergence
in a.e. sense), since for v = i=1 αi φi ∈ V and ω ∈ Ω0 we have
∞
X
2
|ξ(ω, v)| = | αi Xi (ω)|2
i=1

∞
X ∞
X k0
X
|Xi (ω)|2 |Xi (ω)|2
≤ αi2 kφi k2V · = kvk2V .
i=1 i=1
kφi k2V i=1
kφi k2V

For any sequence (ϕn )n∈N converging to zero in V we have |ξ(ω, ϕn )| ≤

P |Xi (ω)|2
kϕn kV ∞ i=1 kφi k2V , which converges to zero for each ω ∈ Ω0 . Thus,
ξ(ω, ·) ∈ V 0 . On the other hand, kξ(ω, φn )kZ 2 = kXn (ω)kZ 2 , which
does not converge to zero. Thus, ξ is not a GRP (I).

Further examples will be given in Section 2.5.3.

As in [GV], we will assume that the expectation of ξ, denoted by

E(ξ(ω, ϕ)) = m(ϕ), ϕ ∈ V, exists and belongs to V 0 . Due to this fact,
throughout the paper, we will assume that E(ξ(·, ϕ)) = 0 for every ϕ ∈ V.
If E(ξ(ω, ϕ)ξ(ω, ψ)) exists for all ϕ and ψ, and if it is continuous in each
argument ϕ and ψ, then the correlation operator of the GRP ξ is defined by
Cξ (ϕ, ψ) = E(ξ(·, ϕ)ξ(·, ψ)), ϕ, ψ ∈ V.

2.5.1 Generalized random processes on Lr (G), r > 1

Denote by κ(E) the characteristic function of E ⊂ Rn . Let x0 ∈ Rn and
let (En )n∈N0 be a sequence of Borel sets, which shrinks nicely to x0 . Recall
from [Ru, Chapter 8.] that a sequence (En )n∈N0 of Borel sets in Rn shrinks
2.5 GRPs of Type (II) 63

nicely to x0 ∈ Rn , if there exists α > 0 such that each En lies in an open ball
B(x0 , rn ) and m(En ) ≥ αm(B(x0 , rn )), n ∈ N, rn → 0.
Put
κ(En )(x)
δn (x − x0 ) = , n ∈ N, x ∈ Rn . (2.55)
m(En )
The proof of the following results can be found in [LP2] and [LPP].
Q
Theorem 2.5.1 Let G = ni=1 (αi , βi ) ⊂ Rn , −∞ ≤ αi < βi ≤ ∞, i =
1, 2, ..., n, and let ξ be a GRP on Ω × Lr (G), r > 1.

a) There exists a function f : Ω × G → C such that

(i) for every x ∈ G, f (·, x) is measurable and for every ω ∈ Ω,

f (ω, ·) ∈ Lp (G), p = r/(r − 1).
(ii) Z
ξ(ω, ϕ) = f (ω, t)ϕ(t)dt, ω ∈ Ω, ϕ ∈ Lr (G).
G

b) If ξ is a Gaussian GRP on Ω×Lr (G) such that for every ω ∈ Ω, x0 ∈ G

and every delta sequence δn (x − x0 ), n ∈ N, of the form (2.55)

lim ξ(ω, δn (x − x0 )) exists,

n→∞

then there exists a Gaussian random process f : Ω × G → C such that

(i) and (ii) in a) hold.

c) Let G = Rn . If there exists A ∈ F such that P (A) = 0 and

|ξ(ω, ϕ)| ≤ C(ω)kϕkr , ω ∈ Ω\A,

then the correlation operator Cξ (·, ·) has a representation

Z Z
Cξ (ϕ, ψ) = ϕ(t)E(f (ω, t)f (ω, s))ψ(s)dtds, ϕ, ψ ∈ Lr (G).
Rn Rn

Proposition 2.5.1 Let G be an open subset of Rn and let ξ be a GRP on

Ω × Lr (G). Then for every ω ∈ Ω there exists a set A(ω) ∈ G such that
m(A(ω)) = 0 and for every x0 ∈ G \ A(ω) there exists the limit

lim ξ(ω, δn (· − x0 )),

n→∞

where δn (x − x0 ) is defined as in (2.55).

64 Generalized Stochastic Processes

2.5.2 Generalized random processes on K{Mp }

Here we will use the techniques given in Section 1.5 for K{Mp } spaces.

Theorem 2.5.2 a) Let ξ be a GRP on K{Mp }. Then for every ε > 0

there exist d ∈ N0 , M ∈ F satisfying P (M ) ≥ 1 − ε, and functions
fα : Ω × R → C, α = 0, 1, ..., d, such that fα (·, t) is measurable for
every t ∈ R, fα (ω, ·) is in L2 (R) for every ω ∈ M , α = 0, 1, ..., d and
d Z
X
ξ(ω, ϕ) = fα (ω, t)Md (t)ϕ(α) (t)dt, ω ∈ M, ϕ ∈ K{Mp },
α=0 R
(2.56)
d
X
kfα (ω, ·)kL2 ≤ d, ω ∈ M. (2.57)
α=0

In particular, if there exist C(ω) > 0, ω ∈ Ω, and d ∈ N such that

|ξ(ω, ϕ)| ≤ C(ω)kϕkd,2 , ω ∈ Ω, ϕ ∈ K{Mp }, (2.58)

then representation (2.56) is valid on the whole Ω.

b) Moreover, if ξ is also a continuous mapping from K{Mp } to Z 2 , then

for almost every t, s ∈ R there exist E(fα (·, t)fβ (·, s)), α ≤ d, β ≤ d and
the correlation operator Cξ (ϕ, ψ), ϕ, ψ ∈ K{Mp } has the representation

Cξ (ϕ, ψ)

d Z Z
d X
X
= E(fα (·, t)fβ (·, s))Md (t)Md (s)ϕ(α) (t)ψ (β) (s)dtds.
α=0 β=0 R R

c) If ξ is a GRP on K{Mp } such that (2.58) holds and ω 7→ C(ω) is in Z 2 ,

then ξ : K{Mp } → Z 2 is continuous and (2.56) holds for every ω ∈ Ω.
Condition C(·) ∈ Z 2 is sufficient but not necessary for the continuity
of ξ : K{Mp } → Z 2 .

Proof. a) Note that in [LP] and [LP1] more restricted problems were con-
sidered (see also [Ha] and [Ul]).
Since for every ω ∈ Ω, ξ(ω, ·) is in K0 {Mp }, it follows that for every ω ∈ Ω
there exist C(ω) > 0 and p(ω) ∈ N such that

|ξ(ω, ϕ)| ≤ C(ω)kϕkp(ω),2 , ϕ ∈ K{Mp }.

2.5 GRPs of Type (II) 65

We can assume that p(ω) ≥ C(ω). For every ϕ ∈ K{Mp } and N ∈ N, put
\
AN (ϕ) = {ω ∈ Ω : |ξ(ω, ϕ)| < N kϕkN,2 }, AN = AN (ϕ).
ϕ∈K{Mp }

Since K{M
T p } is separable, it contains a countable dense subset D and
AN = ϕ∈D AN (ϕ) ∈ F. Thus, from
∞
[
Ω= AN and AN ⊂ AN +1 , N ∈ N,
N =1

it follows that for given ε > 0 there exists an integer d such that P (Ad ) ≥
1 − ε. Denote M = Ad . It follows

|ξ(ω, ϕ)| ≤ dkϕkd,2 , ω ∈ M, ϕ ∈ K{Mp }.

We extend ξ on the whole Ω by

½
ξ(ω, ϕ), ω ∈ M
ξ1 (ω, ϕ) = , ϕ ∈ K{Mp }. (2.59)
0, ω∈6 M

Further, put R = {ϕ ∈ K{Mp } : kϕkd,2 ≤ 1} and

S(ω) = sup |ξ1 (ω, ϕ)| = sup

T
|ξ1 (ω, ϕ)|, ω ∈ Ω.
ϕ∈R ϕ∈D R

It follows that S is measurable on Ω, S(ω) ≤ d, ω ∈ Ω. Thus,

|ξ1 (ω, ϕ)| ≤ S(ω)kϕkd,2 , ϕ ∈ K{Mp }, ω ∈ Ω. (2.60)

d
Inequality (2.60) holds also for the space HM (R) ⊂ H d (R), where H d (R)
is the Sobolev space, and HM = {ϕ ∈ H (R) : Md ϕ(α) ∈ L2 (R), α =
d d

0, 1, . . . , d}, equipped with the topology induced by the norm kϕkd,L2 =

Pd (α)
α=0 kMd ϕ kL2 .
We need the following consequence of (2.60):
d
if (ϕν )ν∈N is a sequence in K{Mp } and ϕν → 0 in HM , (2.61)

then ξ1 (ω, ϕν ) → 0, ν → ∞.
Qd
Let Γd = i=0 L2 (R) and endow it with the scalar product ((ϕα ), (ψα )) =
Pd R
α=0 R ϕα ψα dt, (ϕα ), (ψα ) ∈ Γd . Clearly, Γd is a Hilbert space. Define a
mapping θ : K{Mp } → Γd by θ(ϕ) = (Md ϕ, Md ϕ0 , . . . , Md ϕ(d) ), ϕ ∈ K{Mp },
which is injective, and denote ∆ = θ(K{Mp }). Note that
d
∆ = θ(HM ). (2.62)
66 Generalized Stochastic Processes

Define a mapping Ω × Γd → C, for every ω ∈ Ω, by



 ξ1 (ω, θ−1 (ψ)), ψ∈∆
¯ L2
F (ω, ψ) = lim ν→∞ ξ1 (ω, θ −1
(ψ ν )), ψ ∈ ∆, ψ ν ∈ ∆, ψ ν −→ ψ,

 ¯ ⊥.
0, ψ∈∆

The existence of the limit follows from (2.61) and (2.62). Thus,
¯ ψ⊥ ∈ ∆
F (ω, ψ̃) = F (ω, ψ), ψ̃ ∈ Γd , ψ̃ = ψ + ψ ⊥ , ψ ∈ ∆, ¯ ⊥.

Clearly F (·, ψ̃) is measurable, for any ψ̃ ∈ Γd . Let ϕ ∈ K{Mp }, ω ∈ Ω. We

have
|F (ω, θ(ϕ))| ≤ S(ω)kϕkd,2 = S(ω)kθ(ϕ)kΓd .
So, for every ω ∈ Ω, F (ω, ·) is a continuous linear functional on Γd and
it is of the form
Xd
F (ω, ·) = Fα (ω, ·), ω ∈ Ω.
α=0

Here Fα (ω, ·), ∈ Ω, α ≤ d, are continuous linear functionals on subspaces

Γd,α ⊂ Γd , α ≤ d, where

Γd,α = {ψ = (ψβ ) ∈ Γd : ψβ ∈ L2 (R), ψβ ≡ 0, β 6= α}.

It is endowed with the natural norm such that it is isometric to L2 (R), for
every α ∈ {0, 1, ..., d}. Let ψ ∈ Γd . Denote by [ψ]α the corresponding element
in Γd,α . The βth coordinates of [ψ]α are equal to zero for β 6= α and the αth
coordinate is equal to ψα . Since F is a GRP, it follows that Fα = F |Γd,α is a
GRP on Ω × Γd,α i.e. on Ω × L2 (R), for every α, α ≤ d.
By Theorem 2.5.1, for every α = 0, 1, . . . , d, there exists a function fα :
Ω × R → C such that fα (·, t) is measurable for every t ∈ R, fα (ω, ·) ∈ L2 (R),
ω ∈ Ω, and
Z
Fα (ω, ϕ) = fα (ω, t)ϕ(t)dt, ϕ ∈ L2 (R), ω ∈ Ω.
R

Thus, if ω ∈ Ω and ψ = θ(ϕ) for ϕ ∈ K{Mp }, then

d
X d Z
X
F (ω, ψ) = Fα (ω, [ψ]α ) = fα (ω, t)Md (t)ψ (α) dt (2.63)
α=0 α=0 Rn

and
d
X
kF (ω, ·)k0Γd = kfα (ω, ·)kL2 (R) ≤ S(ω) ≤ d, ω ∈ Ω,
α=0
2.5 GRPs of Type (II) 67

where k · k0Γd is the dual norm. Now, the assertion follows by (2.59).
The proof of the last assertion in a) follows by repeating the previous
proof starting from relation (2.60). It follows that ξ is of the form (2.56) for
every ω ∈ Ω.
b) Obviously, Cξ (ϕ, ψ) = E(ξ(·, ϕ)ξ(·, ψ)), ϕ, ψ ∈ K{Mp } is bilinear. The
continuity follows from

Cξ (ϕ, ψ) = |E(ξ(·, ϕ)ξ(·, ψ))| ≤ kξ(·, ϕ)kZ 2 kξ(·, ψ)kZ 2 ≤

≤ kϕkd,2 kψkd,2 sup{kξ(·, ϕ)kZ 2 , ϕ ∈ K{Mp }, kϕkd,2 ≤ 1}

· sup{kξ(·, ψ)kZ 2 , ψ ∈ K{Mp }, kψkd,2 ≤ 1}.

Fubini’s theorem implies

Cξ (ϕ, ψ) = E(ξ(·, ϕ)ξ(·, ψ))

Xd Z Xd Z
(α)
= E(( fα (·, t)Md (t)ϕ (t)dt)( fα (·, s)Md (s)ψ (α) (s)ds))
α=0 R α=0 R

d X
X d Z Z
= E( fα (·, t)Md (t)ϕ(α) (t)fβ (·, s)Md (s)ψ (β) (s)dtds)
α=0 β=0 R R

d Z Z
d X
X
= ( E(fα (·, t)fβ (·, s))Md (t)Md (s)ϕ(α) (t)ψ (β) (s)dtds.
α=0 β=0 R R

This proves the last assertion in b).

c) If ξ is a GRP on K{Mp } which satisfies (2.58) and C(·) ∈ Z 2 , then ξ is
a continuous mapping K{Mp } → Z 2 . Namely, for any sequence (ϕn )n∈N in
K{Mp } such that ϕn → 0, n → ∞, it follows

kξ(·, ϕn )kZ 2 = E(|ξ(·, ϕn )|2 ) ≤ E(C 2 (·))kϕn k2d,2 → 0.

Example 2.5.3 in the next subsection shows that ξ : K{Mp } → Z 2 may

be a continuous mapping from K{Mp } to Z 2 although C(·) ∈
/ Z 2. ¤

Restriction (2.58) is not a draconian one; it holds for a large class of

processes but not for the whole class of GRPs (II) as it will be seen in
Example 2.5.5. The following theorem for Gaussian GRPs (II) represents
the main result of [PS2]. The point value of a distribution is defined in sense
of Lojasiewicz (see Section 1.5).
68 Generalized Stochastic Processes

Theorem 2.5.3 a) Let ξ be a Gaussian GRP on K{Mp } and ε ∈ (0, 1).

Then there exist M ∈ F such that P (M ) ≥ 1 − ε, si , d˜i ∈ N, i = 1, 2, 3,
and there exist Gaussian processes Ξ1 (ω, t), Ξ2 (ω, t) and Ξ3 (ω, t),
(ω, t) ∈ Ω × R, with the following properties:
for every ω ∈ Ω, t 7→ Ξi (ω, t) is continuous, i = 1, 2, 3;

supp Ξ1 (ω, ·) ⊂ [−∞, −a], supp Ξ2 (ω, ·) ⊂ [−2a, 2a], supp Ξ3 (ω, ·) ⊂ [a, ∞];
|Ξi (ω, t)|
sup < ∞, i = 1, 2, 3;
t∈R Md˜i (t)
ξ(ω, ·) = Ξ1 (ω, ·)(s1 ) + Ξ2 (ω, ·)(s2 ) + Ξ3 (ω, ·)(s3 ) , ω ∈ M. (2.64)
b) In particular, if there exist D(ω) > 0, ω ∈ Ω and d ∈ N0 such that
|ξ(ω, ϕ)| ≤ D(ω)kϕkd,2 , ω ∈ Ω, ϕ ∈ K{Mp }, (2.65)
then representation (2.64) is valid on the whole Ω.
c) Let ξ be as in part a) and ε ∈ (0, 1). Let C(ω), ω ∈ Ω, be a Gaussian
random variable. The GRP ξ has the point value C(ω) at t = t0 ,
denoted by ξ(ω, t0 ) = C(ω), ω ∈ Ω, if and only if there exist M ∈ F
such that P (M ) ≥ 1 − ε, s3 , d3 ∈ N, a ∈ R and a Gaussian process
Ξ3 (ω, t), (ω, t) ∈ Ω × R with following properties:
supp Ξ3 (ω, ·) ⊂ [t0 − 2a, ∞),
ξ(ω, ·) = Ξ3 (ω, ·)(s3 ) , ω ∈ M, in a neighborhood of t0 ,
|Ξ3 (ω, t)|
sup < ∞,
t∈R Md3 (t)
and
Ξ3 (ω, t) C(ω)
lim s
= , ω ∈ M. (2.66)
t→t0 (t − t0 ) 3 s3 !
In particular, if there exist D ∈ Z 2 , D(ω) > 0, ω ∈ Ω and d ∈ N0 such
that
|ξ(ω, ϕ)| ≤ D(ω)kϕ ∗ fd3 + 1 kd,2 , ω ∈ Ω, ϕ ∈ K{Mp }, (2.67)
2

then (2.66) is equivalent to

° °
° Ξ3 (ω, t) C(ω) °
lim °
° − ° = 0. (2.68)
t→t0 (t − t0 )s3 s3 ! °Z 2
2.5 GRPs of Type (II) 69

Proof. a) Let ψ1 , ψ2 and ψ3 be smooth functions on Rd such that supp ψ1 ⊂

(−∞, −a], supp ψ2 ⊂ [−2a, 2a] and supp ψ3 ⊂ [a, ∞), a > 0, ψ1 = 1 on
(−∞, −2a), ψ3 = 1 on (2a, ∞) and ψ1 + ψ2 + ψ3 = 1. With this partition of
unity we obtain Gaussian stochastic processes

ξ1 (ω, φ) = ξ(ω, φψ1 ), φ ∈ K{Mp },

ξ2 (ω, φ) = ξ(ω, φψ2 ), φ ∈ K{Mp },

ξ3 (ω, φ) = ξ(ω, φψ3 ), φ ∈ K{Mp }.
With the same proof as of Theorem 2.5.2 a),we have the existence of di ∈ N0 ,
a set Ai,d such that P (Ai,d ) ≥ 1 − ε/3, M i = Ai,d , i = 1, 2, 3, such that

|ξi (ω, φ)| ≤ di kφkdi ,2 , ω ∈ M i , φ ∈ K{Mp }, i = 1, 2, 3.

We extend ξi out of Mi to be equal to zero, denote this extension with ξ˜i ,

i = 1, 2, 3, and (as in the proof of part a)) we have the existence of measurable
functions Si (ω), ω ∈ Ω such that Si (ω) ≤ di , ω ∈ Ω, i = 1, 2, 3 and

|ξ˜i (ω, φ)| ≤ Si (ω)kφkdi ,2 , ω ∈ Ω, φ ∈ K{Mp }, i = 1, 2, 3. (2.69)

Denote by Kd the completion of K{Mp } with respect to the norm k · kd,2 .

Since the limit of a sequence of Gaussian processes is Gaussian, we extend
(2.69) to elements of Kdi and obtain

|ξ˜i (ω, φ)| ≤ Si (ω)kφkdi ,2 , ω ∈ Ω, φ ∈ Kdi , i = 1, 2, 3.

In the sequel we consider ξ˜3 . Let fs be given as in (1.4) and let κ ∈ C ∞ (R),
supp κ ⊂ [−a, ∞), κ ≥ 0, κ(t) = 1, t ≥ 0.
We have that for s3 = d3 + 2, and every x ∈ R,

t 7→ κ(t)fsx−
3
(t) = κ(t)fs3 (x − t) ∈ K−
d3

and moreover, x 7→ κfsx−

3
, x ∈ R, is continuous, because

κfsx−
3
→ κfsx30 − in K−
d3 , as x → x0 .

So, for every ω ∈ Ω, x ∈ R, we have

ds3
ξ˜3 (ω, ·) ∗ δ(x) = ξ˜3 (ω, ·) ∗ fs(s3 3 ) (x) = s3 ξ˜3 (ω, κfsx− ).
dx 3

For every x ∈ R,
ω 7→ ξ˜3 (ω, κfsx−
3
)
70 Generalized Stochastic Processes

is a Gaussian random variable and for every ω ∈ Ω we have that

x 7→ ξ˜3 (ω, κfsx−
3
), x ∈ R,

is continuous. Denote it by Ξ3 , i.e. Ξ3 (ω, x) = ξ˜3 (ω, κfsx−

3
). By (1.5) we have

Ξ3 (ω, ·)(s3 ) = ξ˜3 (ω, ·), ω ∈ Ω.

From
|ξ˜3 (ω, κfsx−
3
)| ≤ S3 (ω)kκfsx−
3
kd3 ,2 , ω∈Ω
and

Z x
kκfsx−
3
k2d3 ,2 = sup |Md3 (t)[κ(t)fs3 (x − t)](i) |2 dt
i≤d3 −a
Z ¯ · ¸ ¯2
¯
x+a
κ(x − u)u d3 +1 (i) ¯
¯ ¯
= sup ¯Md3 (x − u) ¯ du
i≤d3 0 ¯ Γ(d3 + 2) ¯
Z x+a ¯¯ Xi µ ¶· d3 +1
¸(n) ¯¯2
¯ 1 i u ¯
≤ C1 · Md2˜3 (x) sup ¯ ¯ du
i≤d3 0 ¯ Md3 (u) n Γ(d3 + 2) ¯
n=0

we have

v
u
u Z ¯¯ · d3 +1
¸(i) ¯¯2
|Ξ3 (ω, x)| ¯ 1 u ¯
sup ≤ C2 tsup ¯ ¯ du < ∞,
x∈R Md˜3 (x) i≤d3 R ¯ Md3 (u) Γ(d3 + 2) ¯

for some constants C1 , C2 > 0.

In a similar way we construct continuous Gaussian processes Ξ1 and Ξ2
such that Ξ1 (ω, x) = ξ˜1 (ω, fsx+
1
), Ξ2 (ω, x) = ξ˜2 (ω, fsx+
2
), ω ∈ Ω, x ∈ R, and for
every ω ∈ Ω the functions x 7→ Ξi (ω, x), x ∈ R, i = 1, 2, 3, are of Md˜1 , Md˜2
and Md˜3 growth rate respectively.
This completes the proof of the theorem, since
3
\
(s1 ) (s2 ) (s3 )
ξ(ω, ·) = Ξ1 (ω, ·) + Ξ2 (ω, ·) + Ξ3 (ω, ·) , ω∈M = M i.
i=1

b) The proof is similar as in Theorem 2.5.2.

c) Without loss of generality, we prove the assertion only for t0 = 0. Using
a similar partition of unity as in part a) but choosing supp ψ1 ⊂ (−∞, −3a],
supp ψ2 ⊂ [−4a, −a] and supp ψ3 ⊂ [−2a, ∞), we obtain a representation
ξ(ω, ·) = Ξ1 (ω, ·)(s1 ) + Ξ2 (ω, ·)(s2 ) + Ξ3 (ω, ·)(s3 ) , ω ∈ M.
2.5 GRPs of Type (II) 71

Since we are interested in values in the neighborhood of t0 = 0, we restrict

our attention to Ξ3 (ω, ·), where supp Ξ3 (ω, ·) ⊂ [−2a, ∞), and Ξ3 (ω, x) =
ξ˜3 (ω, fsx−
3
κ) as in part a). Thus,
3
X
limhξ(ω, εx), φ(x)i = lim hξ˜i (ω, εx), φ(x)i = limhξ˜3 (ω, εx), φ(x)i.
ε→0 ε→0 ε→0
i=1

We prove now that

Z
limhξ˜3 (ω, εx), φ(x)i = C(ω) φ(x)dx, ω ∈ M, φ ∈ K{Mp } (2.70)
ε→0 R

is equivalent to
Ξ3 (ω, x) C(ω)
lim s
= , ω ∈ M. (2.71)
x→0 x 3 s3 !
Assume that (2.70) holds. Then, for φ(x) = fsx−
3
, x ∈ R, we obtain

Ξ3 (ω, x) 1 t
lim s
= lim hξ˜3 (ω, t), (1 − )s3 −1 i
x
x→0 3 x→0 xΓ(s3 ) x
Z 1
˜ 1 s3 −1 1 C(ω)
= lim hξ3 (ω, tx), (1 − t) i = C(ω) (1 − t)s3 −1 dt = .
x→0 Γ(s3 ) 0 Γ(s3 ) s3 !
Conversely, assume (2.71) holds. Then, for φ ∈ K{Mp } we have
1 x
limhξ˜3 (ω, εx), φ(x)i = limhξ˜3 (ω, x), φ( )i
ε→0 ε→0 ε ε
1 x 1 x
= limhΞ3 (ω, x)(s3 ) , φ( )i = lim(−1)s3 hΞ3 (ω, x), s3 +1 φ(s3 ) ( )i
ε→0 ε ε ε→0 ε ε

Ξ3 (ω, εx)xs3 (s3 ) s3 C(ω)x

s3
= lim(−1)s3 h , φ (x)i = (−1) h , φ(s3 ) (x)i
ε→0 εs3 xs3 s3 !
µ s3 ¶(s3 ) Z
x
= C(ω)h , φ(x)i = C(ω) φ(x)dx.
s3 ! R

Thus, ξ has a point value in sense of Lojasiewicz at t0 = 0 if and only if (2.66)

holds. If condition (2.67), and thus also (2.65), is satisfied and D ∈ Z 2 , then
ξ is a continuous mapping K{Mp } → Z 2 . Also, from (2.67) we have

Ξ3 (ω, x) |ξ˜3 (ω, κfsx− )| D(ω)

| s
| = s
3
≤ kκfsx−+d + 1 kd,2
x 3 |x| 3 |x|s3 3 3 2
72 Generalized Stochastic Processes

v ¯
u
u Z x ¯ Ã 3
!(i) ¯¯2
D(ω) u ¯ 2d + ¯
¯Md (t) κ(t) (x − t)
3 2
= d3 +2 tsup ¯ dt
|x| ¯
i≤d3 −2a ¯
5
Γ(2d3 + 2 ) ¯
¯
sZ
x
D(ω)
≤ K1 d3 +2 (x − t)2d3 +3 dt ≤ K2 · D(ω),
|x| −2a

for some constants K1 , K2 > 0 and x → 0. Thus, the vector-valued version

(2.68) follows from the Lebesgue dominated convergence. ¤

Now, concerning equation (2.52), we will assume that the GRP on the
right hand side of the equation has a primitive with a point value at t = 0
in sense of Lojasiewicz.

Corollary 2.5.1 Let f be a Gaussian GRP f : K{Mp } → Z 2 such that its

primitive has a point value at t = 0 in sense of Lojasiewicz. Let X(ω), ω ∈ Ω
be a Gaussian random variable. Then for every ε > 0 there exists M ∈ F
satisfying P (M ) ≥ 1 − ε, and there exists a Gaussian GRP y : K{Mp } → Z 2
such that for every ω ∈ M

d
y(ω, ·) = f (ω, ·) and y(ω, 0) = X(ω).
dt
Proof. Theorem 2.5.3 c) ensures the existence of F (ω, t), ω ∈ Ω, t ∈ [−2, ∞)
and k ∈ N such that

f (ω, ·) = F (ω, ·)(k) , ω ∈ M,

where F (ω, ·) is continuous for every ω ∈ Ω, and F (·, t) is Gaussian for every
t ∈ [−2, ∞). From the assumption on the initial point t = 0 we get that
F (ω, t)
must be bounded in a neighborhood of zero. Define
tk
µ Z t ¶(k)
X(ω)tk
y(ω, t) = + F (ω, s)ds , ω ∈ M, t ∈ [−2, ∞). (2.72)
k! 0

Let us check now if y is a solution of (2.52) on M . Indeed, from (2.72) we

³R ´(k+1)
t
have dtd y(ω, t) = 0
F (ω, s)ds = F (ω, t)(k) = f (ω, t). Thus, y is a
primitive for f , and therefore it has a point value at t = 0. Also,
µ Z t ¶ Z
1 X(ω)tk X(ω) 1 t
lim k + F (ω, s)ds = + lim k F (ω, s)ds
t→0 t k! 0 k! t→0 t 0
2.5 GRPs of Type (II) 73

X(ω) 1 X(ω) t F (ω, t) X(ω)

= + lim k−1 F (ω, t) = + lim k
= .
k! t→0 kt k! t→0 k t k!
This means y(ω, 0) = X(ω). ¤

Note that the solution given in (2.72) coincides with the solution given
in (2.53) up to the set M , provided (2.54) holds and thus (2.53) exists.

2.5.3 Examples of GRPs

In the first four examples
R which are to follow we let Ω = R, F be the
dx
Borel field, P (A) = A π(1+x 2 ) for A ∈ F, and X be the identity mapping

from R to R. As usual, δ denotes the Dirac delta distribution.

Example 2.5.2 Let, for x ∈ R, ϕ ∈ K{Mp },

ξ(x, ϕ) = X(x)hδ, ϕi.

It is a GRP of type (II) but not of type (I).

Example 2.5.3 Let, for x ∈ R, ϕ ∈ K{Mp },

ξ(x, ϕ) = X(x)ϕ(X(x)) = X(x)hδ(y − X(x)), ϕ(y)i. (2.73)

It is a GRP (II) and moreover, it is a continuous mapping from K{Mp } to

Z 2 , i.e. it is also a GRP (I). For every x ∈ R we have

|ξ(x, ϕ)| = |X(x)ϕ(x)| ≤ |X(x)|kϕkp,2 , (2.74)

/ Z 2.
although |X| ∈

Example 2.5.4 Let

½ 1
X(x)ϕ( X(x) ), x 6= 0
ξ(x, ϕ) = , x ∈ R, ϕ ∈ K{Mp }.
0, x=0

This is a GRP (II) which does not map K{Mp } to Z 2 , since if the support of
a test function φ equals [0, a] for some a > 0, then for x1 being in this interval,
one has 1/X(x) = 1/x ∈ [0, a) and X(x) = x ∈ [1/a, ∞), thus ξ(·, φ) is not
in Z 2 .

Example 2.5.5 We construct a GRP (II) which does not satisfy (2.58),
and thus can not be represented as a sum of continuous processes on a set of
probability measure one, only on a set of arbitrary large probability measure.
74 Generalized Stochastic Processes

Let (xn )n∈N be a sequence of real numbers strictly increasing to infinity.

For x ∈ R, ϕ ∈ K{Mp }, put

 1
ϕ(0), x 6= xn , n ∈ N,
ξ(x, ϕ) = 1 + X 2 (x)
 ϕ(n) (0), x = xn , n ∈ N.

For every ε > 0, this is a continuous mapping from K{Mp } to Z 2 on

x ∈ R\(−ε, ε). There do not exist C(x) > 0 and p ∈ N such that (2.58) holds,
because it would imply that for every x ∈ R the corresponding distribution is
of finite order, while f = ξ(xn , ·) = (−1)n δ (n) , n ∈ N.

In the following examples we consider the space of tempered distributions

S0 (R) and Z 2 = L2 (S0 (R), B, µ) as defined in Section 1.7.1.

Example 2.5.6 Two examples of GRPs (I) which are also GRPs (II) are
singular pointwise one-dimensional one-parameter white noise, and Brown-
ian motion. Recall from Example 1.7.1 and Example 1.7.2 , they are given
respectively, by
∞
X
W (ω, x) = ξk (x)Hεk (ω), ω ∈ S0 (R), x ∈ R
k=1

and
∞ Z
X x
B(ω, x) = ξk (t)dtHεk (ω), ω ∈ S0 (R), x ∈ R
k=1 0

where {ξn , n ∈ N0 } is the Hermite orthonormal basis of L2 (R).

Actually, they act on S(R) as follows:
∞ Z
X
S(R) 3 ϕ 7→ hW (ω, x), ϕ(x)i = ξk (x)ϕ(x)dxHεk (ω).
k=1 R

∞ Z Z
X x
S(R) 3 ϕ 7→ hB(ω, x), ϕ(x)i = ξk (t)dtϕ(x)dxHεk (ω).
k=1 R 0

d
Since for fixed ω ∈ Ω we have dx B(ω, x) = W (ω, x) in S0 (R) and W (ω, x)
is continuous with respect to x ∈ R (see [HØUZ]), this is indeed compatible
with the assertion in Theorem 3.5.3, i.e. continuity ensures representation
on the whole Ω.
2.6 Hilbert Space Valued GRPs (II) 75

Another process described in [HØUZ] and [HKPS] is the coordinate pro-

cess of white noise defined as
∞ Z
X
Wφ (ω, x) = hω(y), φ(y − x)i = ξk (y)ϕ(y − x)dyHεk (ω),
k=1 R

where x ∈ R, ω ∈ S0 (R), φ ∈ S(R). Note that the same process is obtained if

we apply our partition of unity from Theorem 2.5.3.
In a similar manner we can define a process
∞ Z Z
X y
Bφ (ω, x) = ( ξk (t)dt)ϕ(y−x)dyHεk (ω), x ∈ R, ω ∈ S0 (R), φ ∈ S(R).
k=1 R 0

Since both Wφ (ω, x) and Bφ (ω, x) belong to Z 2 , we have in classical sense

d
B (ω, x) = Wφ (ω, x). In notation of Theorem 2.5.3 (see the proof after
dx φ
applying the partition of unity) this means

d
W (ω, φ̃x ) = B(ω, φ̃x ), where φ̃x (y) = f1x− (y) = H(x − y).
dx
This means that in distributional sense i.e. in S0 (R) we obtain again the
d
well-known result dx B(ω, x) = W (ω, x).

2.6 Hilbert Space Valued Generalized

Random Processes of Type (II)
Recall that H is a separable Hilbert space over C with orthonormal
basis {en : n ∈ N}. While for GRPs (I) on nuclear spaces we had
L(A(H); (S)−1 ) ∼ = L(A; S(H)−1 ), i.e. it was equivalent whether H was the
codomain of the x–variable function space or the ω–variable function space,
for GRPs (II) we have a different situation. Now we state the definition of a
Hilbert space valued GRP (II) and the corresponding structure theorem for
it. We will restrict our attention to GRPs on K{Mp } spaces.

Definition 2.6.1 A H–valued GRP (II) is a mapping ξ : Ω × K{Mp }(H) →

C such that:

(i) for every ϕ ∈ K{Mp }(H), ξ(·, ϕ) is a complex random variable,

(ii) for every ω ∈ Ω, ξ(ω, ·) is an element in K0 {Mp }(H).

76 Generalized Stochastic Processes

For r > 1 denote Lr (Rn ; H) = Lr (Rn ) ⊗ H and recall that its dual is
L (Rn ; H), p = r/(r
p p n r n
R − 1). The dual pairing of f ∈ L (R ; H), ϕ ∈ L (R ; H)
can be written as Rn hf (t), ϕ(t)iH dt. It can easily be checked that following
H–valued version of Theorem 2.5.1 holds.
Q
Theorem 2.6.1 Let G = ni=1 (αi , βi ) ⊂ Rn , −∞ ≤ αi < βi ≤ ∞, i =
1, 2, ..., n, and let ξ be a GRP on Ω×Lr (G; H), r > 1. There exists a function
f : Ω × G → H such that
(i) for every x ∈ G, f (·, x) is measurable and for every ω ∈ Ω, f (ω, ·) ∈
Lp (G; H), p = r/(r − 1).
(ii) Z
ξ(ω, ϕ) = hf (ω, t), ϕ(t)iH dt, ω ∈ Ω, ϕ ∈ Lr (G; H).
G

Following H–valued analogue of Theorem 2.5.2 holds:

Theorem 2.6.2 a) Let ξ be a H–valued GRP (II). Then for every ε > 0
there exist d ∈ N0 , M ∈ F satisfying P (M ) ≥ 1 − ε, and functions
fα : Ω × R → H, α = 0, 1, ..., d, such that fα (·, t) is measurable for
every t ∈ R, fα (ω, ·) is in L2 (R; H) for every ω ∈ M , α = 0, 1, ..., d
and
X d Z
ξ(ω, ϕ) = hfα (ω, t), Md (t)ϕ(α) (t)iH dt, ω ∈ M, ϕ ∈ K{Mp }(H),
α=0 R
(2.75)
d
X
kfα (ω, ·)kL2 (R;H) ≤ d, ω ∈ M. (2.76)
α=0

In particular, if there exist C(ω) > 0, ω ∈ Ω, and d ∈ N such that

|ξ(ω, ϕ)| ≤ C(ω)kϕkd,2;H , ω ∈ Ω, ϕ ∈ K{Mp }(H), (2.77)
then representation (2.75) is valid on the whole Ω.
b) Moreover, if ξ is also a continuous mapping from K{Mp }(H) to Z 2 ,
then for almost every t, s ∈ R there exist E(hfα (·, t), fβ (·, s)iH ), α ≤ d,
β ≤ d and the correlation operator Cξ (ϕ, ψ), ϕ, ψ ∈ K{Mp } has the
representation
Cξ (ϕ, ψ)
X d X d Z Z
= E(hfα (·, t), fβ (·, s)iH )Md (t)Md (s)hϕ(α) (t), ψ (β) (s)iH dtds.
α=0 β=0 R R
2.6 Hilbert Space Valued GRPs (II) 77

c) If ξ is a GRP on K{Mp }(H) such that (2.77) holds and ω 7→ C(ω) is

in Z 2 , then ξ : K{Mp }(H) → Z 2 is continuous and (2.75) holds for
every ω ∈ Ω. Condition C(·) ∈ Z 2 is sufficient but not necessary for
the continuity of ξ : K{Mp }(H) → Z 2 .

Proof. a) Since for every ω ∈ Ω, ξ(ω, ·) is in K0 {Mp }(H), it follows that for
every ω ∈ Ω there exist C(ω) > 0 and p(ω) ∈ N such that

|ξ(ω, ϕ)| ≤ C(ω)kϕkp(ω),2;H , ϕ ∈ K{Mp }(H).

We can assume that p(ω) ≥ C(ω). For every ϕ ∈ K{Mp }(H) and N ∈ N,
put
\
AN (ϕ) = {ω ∈ Ω : |ξ(ω, ϕ)| < N kϕkN,2;H }, AN = AN (ϕ).
ϕ∈K{Mp }(H)

Since K{M
T p }(H) is separable, it contains a countable dense subset D and
AN = ϕ∈D AN (ϕ) ∈ F. Thus, from
∞
[
Ω= AN and AN ⊂ AN +1 , N ∈ N,
N =1

it follows that for given ε > 0 there exists an integer d such that P (Ad ) ≥
1 − ε. Denote M = Ad . It follows

|ξ(ω, ϕ)| ≤ dkϕkd,2;H , ω ∈ M, ϕ ∈ K{Mp }(H).

We extend ξ on the whole Ω by

½
ξ(ω, ϕ), ω ∈ M
ξ1 (ω, ϕ) = , ϕ ∈ K{Mp }(H). (2.78)
0, ω∈6 M

Further, put R = {ϕ ∈ K{Mp }(H) : kϕkd,2;H ≤ 1} and

S(ω) = sup |ξ1 (ω, ϕ)| = sup

T
|ξ1 (ω, ϕ)|, ω ∈ Ω.
ϕ∈R ϕ∈D R

It follows that S is measurable on Ω, S(ω) ≤ d, ω ∈ Ω. Thus,

|ξ1 (ω, ϕ)| ≤ S(ω)kϕkd,2;H , ϕ ∈ K{Mp }(H), ω ∈ Ω. (2.79)

d
Inequality (2.79) holds also for the space HM (R; H) ⊂ H d (R; H), where
d ∼ d
H (R; H) = H (R) ⊗ H is the H–valued Sobolev space, and HM d
= {ϕ ∈
78 Generalized Stochastic Processes

H d (R; H) : Md ϕ(α) ∈ L2 (R; H), α P

= 0, 1, . . . , d}, equipped with the topology
induced by the norm kϕkd,L2 ;H = dα=0 kMd ϕ(α) kL2 (R;H) .
We need the following consequence of (2.79):
d
if (ϕν )ν∈N is a sequence in K{Mp }(H) and ϕν → 0 in HM , (2.80)

then ξ1 (ω, ϕν ) → 0, ν → ∞.
Qd
Let Γd = L2 (R; H) and endow it with the scalar product
Pdi=0 R
((ϕα ), (ψα )) = α=0 R hϕα , ψα iH dt, (ϕα ), (ψα ) ∈ Γd . Clearly, Γd is a
Hilbert space. Define a mapping θ : K{Mp }(H) → Γd by θ(ϕ) =
(Md ϕ, Md ϕ0 , . . . , Md ϕ(d) ), ϕ ∈ K{Mp }(H), which is injective, and denote
∆ = θ(K{Mp }(H)). Note that
d
∆ = θ(HM ). (2.81)

Define a mapping Ω × Γd → C, for every ω ∈ Ω, by



 ξ1 (ω, θ−1 (ψ)), ψ∈∆
L2 (R;H)
F (ω, ψ) = limν→∞ ξ1 (ω, θ−1 (ψν )), ¯
ψ ∈ ∆, ψν ∈ ∆, ψν −→ ψ,

 0, ψ∈∆ ¯ ⊥.

The existence of the limit follows from (2.80) and (2.81). Thus,
¯ ψ⊥ ∈ ∆
F (ω, ψ̃) = F (ω, ψ), ψ̃ ∈ Γd , ψ̃ = ψ + ψ ⊥ , ψ ∈ ∆, ¯ ⊥.

Clearly F (·, ψ̃) is measurable, for any ψ̃ ∈ Γd . Let ϕ ∈ K{Mp }(H), ω ∈ Ω.

We have
|F (ω, θ(ϕ))| ≤ S(ω)kϕkd,2;H = S(ω)kθ(ϕ)kΓd .
So, for every ω ∈ Ω, F (ω, ·) is a continuous linear functional on Γd and
it is of the form
Xd
F (ω, ·) = Fα (ω, ·), ω ∈ Ω.
α=0

Here Fα (ω, ·), ∈ Ω, α ≤ d, are continuous linear functionals on subspaces

Γd,α ⊂ Γd , α ≤ d, where

Γd,α = {ψ = (ψβ ) ∈ Γd : ψβ ∈ L2 (R; H), ψβ ≡ 0, β 6= α}.

It is endowed with the natural norm such that it is isometric to L2 (R; H), for
every α ∈ {0, 1, ..., d}. Let ψ ∈ Γd . Denote by [ψ]α the corresponding element
in Γd,α . The βth coordinates of [ψ]α are equal to zero for β 6= α and the αth
2.6 Hilbert Space Valued GRPs (II) 79

coordinate is equal to ψα . Since F is a GRP (II), it follows that Fα = F |Γd,α

is a GRP (II) on Ω × Γd,α i.e. on Ω × L2 (R; H), for every α, α ≤ d.
By Theorem 2.6.1, for every α = 0, 1, . . . , d, there exists a function fα :
Ω × R → H such that fα (·, t) is measurable for every t ∈ R, fα (ω, ·) ∈
L2 (R; H), ω ∈ Ω, and
Z
Fα (ω, ϕ) = hfα (ω, t), ϕ(t)iH dt, ϕ ∈ L2 (R; H), ω ∈ Ω.
R

Thus, if ω ∈ Ω and ψ = θ(ϕ) for ϕ ∈ K{Mp }(H), then

d
X d Z
X
F (ω, ψ) = Fα (ω, [ψ]α ) = hfα (ω, t), Md (t)ψ (α) iH dt (2.82)
α=0 α=0 Rn

and
d
X
kF (ω, ·)k0Γd = kfα (ω, ·)kL2 (R;H) ≤ S(ω) ≤ d, ω ∈ Ω,
α=0

where k · k0Γd
is the dual norm. Now, the assertion follows by (2.78).
The proof of the last assertion in a) follows by repeating the previous
proof starting from relation (2.79). It follows that ξ is of the form (2.56) for
every ω ∈ Ω.
b) Obviously, Cξ (ϕ, ψ) = E(hξ(·, ϕ), ξ(·, ψ)iH ), ϕ, ψ ∈ K{Mp }(H) is bi-
linear. The continuity follows from

Cξ (ϕ, ψ) = |E(hξ(·, ϕ), ξ(·, ψ)iH )| ≤ kξ(·, ϕ)kZ 2 kξ(·, ψ)kZ 2

≤ kϕkd,2;H kψkd,2;H sup{kξ(·, ϕ)kZ 2 , ϕ ∈ K{Mp }(H), kϕkd,2;H ≤ 1}

· sup{kξ(·, ψ)kZ 2 , ψ ∈ K{Mp }(H), kψkd,2;H ≤ 1}.
Fubini’s theorem implies

Cξ (ϕ, ψ) = E(hξ(·, ϕ), ξ(·, ψ)iH )

d Z
X Xd Z
(α)
= E(( hfα (·, t), Md (t)ϕ (t)iH dt)( hfα (·, s), Md (s)ψ (α) (s)iH ds))
α=0 R α=0 R

d X
X d Z Z
= E( hfα (·, t), Md (t)ϕ(α) (t)iH hfβ (·, s), Md (s)ψ (β) (s)iH dtds)
α=0 β=0 R R

d Z Z
d X
X
= ( E(hfα (·, t), fβ (·, s)iH )Md (t)Md (s)hϕ(α) (t), ψ (β) (s)iH dtds.
α=0 β=0 R R
80 Generalized Stochastic Processes

This proves the last assertion in b).

c) If ξ is a GRP on K{Mp }(H) which satisfies (2.58) and C(·) ∈ Z 2 , then ξ is
a continuous mapping K{Mp }(H) → Z 2 . Namely, for any sequence (ϕn )n∈N
in K{Mp }(H) such that ϕn → 0, n → ∞, it follows

kξ(·, ϕn )kZ 2 = E(|ξ(·, ϕn )|2 ) ≤ E(C 2 (·))kϕn k2d,2;H → 0. ¤

 You see things; and you say, ’Why?’
But I dream things that never were;
and I say, ’Why not?’
(George Bernard Shaw)

Chapter 3

Applications to Singular
Stochastic Partial Differential
Equations

In this chapter we present some applications of generalized random pro-

cesses to obtain solutions of some classes of SPDEs where the data and the
boundary conditions are modeled by generalized random processes. We begin
with a class of linear elliptic equations and note that for technical reasons,
the space dimension is now denoted by n instead of d. Throughout the whole
chapter, I is assumed to be an open, bounded subset of Rn .

3.1 A Linear Elliptic Dirichlet Problem with

Deterministic Coefficients
and Stochastic Data
Our aim is to solve the stochastic Dirichlet problem
n
X
L u(x, ω) = h(x, ω) + Di f i (x, ω), x ∈ I, ω ∈ Ω,
i=1 (3.1)
u(x, ω) ¹∂I = g(x, ω)
82 Applications to Singular SPDEs

where I ⊂ Rn is an open set, (Ω, F, P ) is a probability space, h, g and

f i , i = 1, 2 . . . n are GRPs (I) and L is a linear elliptic operator of the form
n
X n
X
L u(x, ·) = Di ( aij (x, ·)Dj u(x, ·) + bi (x, ·)u(x, ·))
i=1 j=1
n
(3.2)
X
i
+ c (x, ·)Di u(x, ·) + d(x, ·)u(x, ·).
i=1

The simplest example is when the coefficients are constants, e.g. for
aij = δij (the Kronecker symbol), bi = ci = d = 0, i, j = 1, 2 . . . , n, we obtain
the Laplace operator L = ∆. Note that the operator L in (3.2) is given in
divergence form, which will make it suitable to work with in Sobolev spaces
in terms of weak derivatives, since its coefficients are singular. If its principal
coefficients aij , bi , i, j = 1, 2, . . . , n are assumed to be differentiable, then L
can also be written in form
n
X n
X
Lu= âij Dij u + ˆ
b̂i Di u + du, (3.3)
i,j=1 i=1

P P
where dˆ = d + ni=1 Di bi and b̂i = bi + ci + nk=1 Dk aki , i = 1, 2, . . . , n.
A physical interpretation of equation (3.1) in nondivergent form (3.3)
can be given as in [Ev]: u can be interpreted as the density of some quantity
(e.g. a chemical
P concentration) at equilibrium within a region I. The prin-
cipal term ni,j=1 âij Di Dj u represents the diffusion of u within I, and the
coefficients âij describe the anisotropic, heterogeneous
Pn nature of the medium
i ij
(e.g. wood or liquid crystal). The vector F = j=1 a Dj u, i = 1, 2, . . . , n
may
Pn be jinterpreted as the flux density. The ellipticity condition will imply
j=1 F Dj u ≥ 0, meaning the flow direction is from regions of lower to
higher concentration (to get the opposite direction, one hasP to put a minus
sign in front of the coefficients âij ). The first order term ni=1 b̂i Di u rep-
resents transport within I, while the zeroth order term du describes local
creation or depletion (owing, for example, to reactions).
In the framework we consider, the coefficients of L will be stochastic
processes; thus in physical interpretation equation (3.1) can be understood
as a diffusion process in a stochastic anisotropic medium, with transport
and creation also dependent on some random factors, and with a stochastic
boundary value. Example of a stochastic anisotropic medium is a medium
consisting of two randomly mixed immiscible fluids.
In neurology, elliptic PDEs showed as a good model for brain func-
tion measurements. The elliptic PDE defined in the brain at a few points
3.1 A Linear Elliptic Dirichlet Problem 83

(the
Pn location Pn ofijthe sensors) involves the divergence form operator L =
ij
i=1 D i ( j=1 a D j u). The fact that the matrix [a ] is discontinuous along
the boundaries separating different layers of the brain, means that the PDE
must deal with singular data. Now, if there are also some random perturba-
tions happening in the brain, the PDE is no longer a PDE but a SPDE.
If the coefficients of L are deterministic functions, then our assumptions
on L will follow the approach in [GT]: We assume that L is strictly elliptic in
I and has bounded measurable coefficients; that is, there exist λ > 0, Λ > 0,
and ν ≥ 0 such that
n
X
aij (x)ξi ξj ≥ λ|ξ|2 , x ∈ I, ξ ∈ Rn , (3.4)
i,j=1

n
X
|aij (x)|2 ≤ Λ2 , x ∈ I, (3.5)
i,j=1
n
1 X¡ i 2 i 2
¢ 1
2
|b (x)| + |c (x)| + |d(x)| ≤ ν 2 , x ∈ I. (3.6)
λ i=1 λ
In fact, λ and Λ are the minimal and maximal, respectively, eigenvalues of
the matrix [aij ]. Without loss of generality we may and will assume that
λ ≤ 1 (else we divide equation (3.1) by λ to obtain this situation). We
seek for a weak solution of (3.1) in the Hilbert space of generalized random
processes L(W01,2 (I), (S)−1 ), where W01,2 (I) denotes the Sobolev space and
(S)−1 denotes the Kondratiev space. As we will see in Theorem 3.1.6, this
generalized solution exists and is unique. In Proposition 3.1.1 we investigate
some stability and regularity properties of this solution. In [LØUZ] regular-
ization was used and an error-function tending to zero was introduced to give
meaning to a distributional valued solution, as well as to give meaning to the
product of a non-smooth deterministic function (the coefficients of L) and
the distribution u(x, ·). The solution was expressed in terms of the Green
function and the Itô integral. We will have a similar stability result, but
using the Hilbert space structure. The goal in this section is to extend the
approach for even more singular coefficients and data, namely Colombeau
generalized functions. We develop the necessary Sobolev–Colombeau spaces
for generalized stochastic processes and prove existence and uniqueness of
the solution for the Dirichlet problem also in this setting.
However, if the coefficients of the operator L are also generalized ran-
dom processes, then a further problem arises: How to interpret the product
between two generalized random processes? In [Va] the product was inter-
preted as the Wick product, and the solution was found as an element of
84 Applications to Singular SPDEs

the Sobolev-Kondratiev space. A similar approach can be found in [Be],

where the tensor product of the Sobolev space and the space of stochastic
trigonometric functions was used, and the product was taken pointwisely. In
[Bu] the author uses the direct product of the classical Sobolev space and a
generalized Sobolev space as a probability space, which allows the product
to be interpreted as ordinary product of two functions.
In [HØUZ] stochastic partial differential equations are solved by a general
receipt: The H-transform and Wick products are used to convert the SPDE
into a deterministic PDE, which can be solved by known PDE methods, then
one must check the conditions to apply the inverse H-transform, which then
defines the solution of the starting SPDE. This method has the disadvantage
that each SPDE must be solved separately, and it can treat only random
processes which are generalized by the random parameter ω, but continuous
(at least) in the time-space variable x. The Hilbert space methods used
in [Va], [Be], [Bu], have the advantage to solve a wider class of equations,
also dealing with random processes that are more singular at the x-variable.
Existence and uniqueness of the solution are obtained, but one does not
obtain an explicit form of the solution. Due to our definition of generalized
random processes as linear continuous mappings from the Sobolev space into
the Kondratiev space, we are also able to reduce the SPDE into a PDE and
make use of the deterministic theory, even without using the H-transform,
which makes the approach simpler than in [HØUZ].

Preliminary review of the deterministic Dirichlet problem

In [GT] the deterministic case, i.e. a Dirichlet problem of the form
n
X
L u(x) = h(x) + Di f i (x), x ∈ I,
i=1 (3.7)
u(x) ¹∂I = g(x)

is considered. A function u ∈ W 1,2 (I) is called a weak solution of (3.7) if

Z X
n Xn Xn
ij i
( (a Dj u + b u)Di v − ( ci Di u + du)v)dx
I i=1 j=1 i=1
Z Xn
(3.8)
= ( f i Di v − hv)dx,
I i=1

for all v ∈ W01,2 (I), and u − g ∈ W01,2 (I). The concept of the weak solution
is based on the fact that one can identify the operator L with its unique
3.1 A Linear Elliptic Dirichlet Problem 85

extension (denoted by the same symbol) L : W 1,2 (I) → W −1,2 (I) which
generates a bilinear form (u, v) 7→ B(u, v) = −hLu, vi, (u, v) ∈ W 1,2 (I) ×
W01,2 (I) where hLu, vi is defined by the left hand side of (3.8). Thus, u ∈
W 1,2 (I) is a weak solution if B(u, v) = hh+∇·f, vi = hh, vi−hf, ∇vi, for every
v ∈ W01,2 (I). Here h·, ·i denotes the standard scalar product in L2 (I), which
can be extended to the canonical dual pairing on W −1,2 (I) × W01,2 (I). Thus,
existence of a weak solution is equivalent to surjectivity, while uniqueness is
equivalent to injectivity of the mapping L. The standard norm on L2 (I) is
denoted by k · k2 .
We state the following theorem from [GT], which will be needed in the
sequel:
Theorem 3.1.1 Let the operator L given by (3.2) satisfy conditions (3.4),
(3.5) and (3.6). Moreover, assume that
Z n
X
(d(x)v(x) − bi (x)Di v(x)) dx ≤ 0, v ≥ 0, v ∈ W01,2 (I). (3.9)
I i=1

Then for g ∈ W 1,2 (I) and h, f i ∈ L2 (I), i = 1, 2, . . . , n the Dirichlet problem

(3.7) has a unique weak solution u. Moreover, there exists a constant C > 0
(depending only on L and I) such that

kukW 1,2 ≤ C (khk2 + kgkW 1,2 ) , (3.10)

where h = (h, f 1 , f 2 , . . . f n ).
These results are based on Hilbert space methods for PDEs. The Lax-
Milgram theorem and the Fredholm alternative are used to prove existence
and uniqueness of the weak solution. For details, refer to [GT, Chapter 8].
Here we state the three main classical theorems, since they are the keystones
that cap the arch of our proofs.

Theorem 3.1.2 (Lax–Milgram) Let B : H × H → Rn be a bilinear form

on a Hilbert space H, satisfying following properties: There exist K, C > 0
such that
• |B(x, y)| ≤ Kkxk kyk, for all x, y ∈ H, (boundedness)

• |B(x, x)| ≥ Ckxk2 , for all x ∈ H. (coercivity)

Then for every bounded linear functional F ∈ H 0 , there exists a unique ele-
ment f ∈ H such that

B(x, f ) = F (x), for all x ∈ H.

86 Applications to Singular SPDEs

Theorem 3.1.3 (Fredholm alternative) Let V be a normed vector space

and T : V → V be a compact linear mapping. Then,

(i) either the homogeneous equation x − T x = 0 has a nontrivial solution

x∈V,

(ii) or for each y ∈ V the equation x−T x = y has a unique solution x ∈ V .

In case (ii), the operator (I − T )−1 whose existence is asserted there is also
bounded.

Theorem 3.1.4 (Fredholm alternative – spectral behavior) Let V be

a Hilbert space and T : V → V be a compact linear mapping. There exists a
countable set Λ ⊂ R having no limit points except possibly λ = 0, such that
if λ 6= 0, λ ∈
/ Λ the equations

λ x − T x = y, λ x − T ∗x = y (3.11)

have uniquely determined solutions x ∈ V for every y ∈ V , and the inverse

mappings (λI − T )−1 , (λI − T ∗ )−1 are bounded. If λ ∈ Λ, the null spaces of
the mappings λI −T , λI −T ∗ have positive finite dimension and the equations
(3.11) are solvable if and only if y is orthogonal to the null space of λI − T ∗
in the first case and λI − T in the second case.

Preliminary review of generalized random processes

Let the basic probability space (Ω, F, P ) be (S0 (Rn ), B, µ) and consider
the Kondratiev space of generalized random variables (S)−1 . Recall that the
generalized expectation of F ∈ (S)−1 is defined by E(F ) = hF, 1i.
We consider generalized random processes as linear continuous mappings
from the Sobolev space W01,2 (I) into the space of Kondratiev generalized
random variables (S)−1 . In Chapter 2 we considered GRPs (I) as linear
continuous mappings from the Zemanian space A into (S)−1 , but since the
Sobolev spaces are constructed in a similar manner, only using the triple
W01,2 (I) ⊆ L2 (I) ⊆ W −1,2 (I) and the Laplace operator, we can state the
results of Theorem 2.1.1 also in this context. Denote further on the space of
GRPs by
WS∗ = L(W01,2 (I), (S)−1 ).
Note that WS∗ ∼ = L((S)1 , W −1,2 (I)) ∼
= B((S)1 × W0 (I), R) ∼
1,2
= L((S)1 ⊗
1,2 ∼
W0 (I), R) = W −1,2
(I) ⊗ (S)−1 . Thus, we may regard a GRP u(x, ω)
as a bilinear continuous mapping u(ϕ, θ) by ϕ(x) ∈ W01,2 (I), x ∈ I and
θ(ω) ∈ (S)1 , ω ∈ Ω, as well as an element of the tensor product space
3.1 A Linear Elliptic Dirichlet Problem 87

W −1,2 (I) ⊗ (S)−1 . The latter isomorphisms hold since (S)1 is a nuclear space
and W −1,2 (I) is a Frèchet space. Also note that the notation ⊗ stands for the
π–completition of the tensor product space, which is in this case equivalent
to the ²–completition by the nuclearity of (S)1 .
By W0 S we denote W01,2 (I) ⊗ (S)1 , and by WS we denote W 1,2 (I) ⊗ (S)1 .
We state now Theorem 2.1.1 in this new context:
Theorem 3.1.5 Following conditions are equivalent:
(i) u ∈ WS∗ .
(ii) u can be represented in the form
X
u(x, ω) = fα (x) ⊗ Hα (ω), x ∈ I, ω ∈ Ω, fα ∈ W −1,2 (I), α ∈ I,
α∈I
(3.12)
and there exists p ∈ N0 such that for each bounded set B ⊂ W01,2 (I)
X
sup |hfα , ϕi|2 (2N)−pα < ∞. (3.13)
ϕ∈B
α∈I

(iii) u can be represented in the form (3.12) and there exists p ∈ N0 such
that X
kfα k2W −1,2 (2N)−pα < ∞. (3.14)
α∈I
P
Note that α∈I fα (x)Hα (ω), fα ∈ W ±1,2 is also a generalized random
process as defined in [Va]. The tensor product space W −1,2 (I) ⊗ (S)−1 was
used also in [Be]; we find it more convenient to deal with the structure given in
Theorem 3.1.5 and also later in the next section to deal with Wick products,
which are all special cases (Sobolev versions) of the structures introduced in
[PS1] and [Se].

3.1.1 Solvability of the stochastic Dirichlet problem

Now we return to our boundary problem (3.1). The idea is to prove
that there exists a mapping Ξ ∈ L((S)1 , W −1,2 (I)) such that Ξ(θ) is a weak
solution of (3.7) for each θ ∈ (S)1 . This will induce a GRP u(x, ω), x ∈ I,
ω ∈ Ω given by hu(x, ω), θ(ω)i, which will be called a generalized solution of
(3.1).
Theorem 3.1.6 Let the operator L given by (3.2) satisfy conditions (3.4),
(3.5), (3.6) and (3.9). Then for g ∈ WS, h, f i ∈ L((S)1 , L2 (I)), i =
1, 2, . . . , n the Dirichlet problem (3.1) has a unique generalized solution
u ∈ WS∗ .
88 Applications to Singular SPDEs

Proof. Let θ ∈ (S)1 and p > 0. Put gθ (x) = hg(x, ω), θ(ω)i, hθ (x) =
hh(x, ω), θ(ω)i, fθi (x) = hf i (x, ω), θ(ω)i, i = 1, 2, . . . , n. Clearly, gθ ∈
W 1,2 (I), while hθ , fθi ∈ L2 (I), i = 1, 2, . . . , n. According to Theorem 3.1.1
there exists a unique weak solution uθ (x), x ∈ I in W 1,2 (I) of the boundary
problem
n
X
L uθ (x) = hθ (x) + Di fθi (x), x ∈ I,
i=1 (3.15)
uθ (x) ¹∂I = gθ (x).

Due to (3.10) and continuity of h and g, there exist constants C > 0, M > 0
(note: C may depend on the coefficients of L and on I, but not on θ) such
that

kuθ kW 1,2 ≤ C (khθ k2 + kgθ kW 1,2 )

≤ C (khk · kθk1,p + kgk · kθk1,p )
≤ M kθk1,p ,
L
where h = (h, f 1 , f 2 , . . . f n ) ∈ L((S)1 , n+1 2
i=1 L (I)) and hθ =
1 2 n
(hθ , fθ , fθ , . . . fθ ), while k · k stands for the operator norm in
L(W01,2 (I), (S)−1 ). Since uθ ∈ W 1,2 (I), we may identify it as an element
of W −1,2 (I) and by the Riesz isometry we get

kuθ kW −1,2 ≤ M kθk1,p . (3.16)

Define a mapping Ξ : (S)1 → W −1,2 (I) by θ 7→ uθ . Clearly, Ξ is linear,

and by (3.16) it is bounded. Thus, Ξ ∈ WS∗ . ¤

Note, since uθ (x) = hu(x, ω), θ(ω)i is constructed as a weak solution of

the deterministic Dirichlet problem, i.e. as a linear continuous mapping, we
also have continuity in the second variable: There exists N > 0 such that
khu(x, ·), ϕ(x)ik−1,−p ≤ N kϕkW 1,2 , for ϕ ∈ W01,2 . Thus, there also exists a
constant K > 0 such that

|hu(x, ω), ϕ(x)θ(ω)i| ≤ KkϕkW 1,2 kθk1,p , ϕ ∈ W01,2 , θ ∈ (S)1 ,

i.e. our solution is a bilinear mapping on W0 S.

Remark. We can obtain the same result using the chaos expansion from
Theorem 3.1.5. Thus, we provide an alternate proof of Theorem 3.1.6.
3.1 A Linear Elliptic Dirichlet Problem 89

Assume g, h, f i are given by following expansions, respectively:

X
g(x, ω) = gα (x) ⊗ Hα (ω), x ∈ I, ω ∈ Ω, gα ∈ W 1,2 (I), α ∈ I,
α∈I
X
h(x, ω) = hα (x) ⊗ Hα (ω), x ∈ I, ω ∈ Ω, hα ∈ L2 (I), α ∈ I,
α∈I
X
i
f (x, ω) = fαi (x) ⊗ Hα (ω), x ∈ I, ω ∈ Ω, fαi ∈ L2 (I), α ∈ I, i = 1, 2, . . . , n.
α∈I

We seek for the solution u in form of

X
u(x, ω) = uα (x) ⊗ Hα (ω), x ∈ I, ω ∈ Ω, uα ∈ W 1,2 (I), α ∈ I,
α∈I

where the coefficients uα , α ∈ I are to be determined. By linearity of L and

consequently
à n
!
X X X
i
L u(x, ω) = L uα (x) ⊗ Hα (ω) = hα (x) + Di fα (x) ⊗ Hα (ω),
α∈I α∈I i=1

we obtain following system of equations:

n
X
L uα (x) = hα (x) + Di fαi (x), x ∈ I, α ∈ I,
i=1 (3.17)
uα (x) ¹∂I = gα (x).

Due to Theorem 3.1.1, for every α ∈ I there exists a unique uα ∈ W 1,2 (I),
which solves (3.17), and there exists C > 0 (uniformly in α) such that
¡ ¢
kuα k2W 1,2 ≤ C khα k22 + kgα k2W 1,2

where hα = (hα , fα1 , fα2 , . . . fαn ). Now, according to the assumptions made on
h, f i , i = 1, 2 . . . , n, there exists p > 0 such that
à !
X X X
kuα k2W 1,2 (2N)−pα ≤ C khα k22 (2N)−pα + kgα k2W 1,2 (2N)−pα < ∞.
α∈I α∈I α∈I

Thus, u ∈ WS. Again, one can consider it also as an element of the dual
space i.e. u ∈ WS∗ .
90 Applications to Singular SPDEs

Corollary 3.1.1 Let u ∈ WS∗ be the generalized solution of (3.1). Then its
generalized expectation E(u) coincides with the weak solution of the deter-
ministic Dirichlet problem
n
X
L v(x) = h̃(x) + Di f˜i (x), x ∈ I,
i=1 (3.18)
v(x) ¹∂I = g̃(x),

where h̃ = E(h), f˜i = E(f i ), i = 1, 2, . . . , n and g̃ = E(g).

Proof. The assertion follows from the proof of Theorem 3.1.6 if we choose
θ = 1. ¤

3.1.2 Stability properties of the Dirichlet problem

First we prove that our generalized solution of the stochastic Dirichlet
problem is continuously dependent on the data. Let L be a strictly elliptic
operator of the form (3.2) with coefficients aij , bi , ci , d, satisfying conditions
(3.4), (3.5) and (3.6). Let L̃ be another strictly elliptic operator of the form
˜ satisfying all given conditions. Let h, h̃, f i
(3.2) with coefficients ãij , b̃i , c̃i , d,
and f˜i , i = 1, 2, . . . , n be generalized random processes from L((S)1 , L2 (I)).
Let g and g̃ be generalized random processes from WS. For θ ∈ (S)1 fixed,
denote by uθ the solution of the Dirichlet problem
n
X
L uθ (x) = hθ (x) + Di fθi (x), x ∈ I,
i=1 (3.19)
uθ (x) ¹∂I = gθ (x)

as it was obtained in Theorem 3.1.6. Respectively, let ũθ be the generalized

solution of the Dirichlet problem
n
X
L̃ ũθ (x) = h̃θ (x) + Di f˜θi (x), x ∈ I,
i=1 (3.20)
ũθ (x) ¹∂I = g̃θ (x).

Due to the construction of the generalized solutions as elements of Sobolev

spaces, we can prove stability (in the x variable) in weak sense of these
solutions. Every Dirichlet problem can be transformed into a problem of
the same form, but with zero boundary conditions (see [GT]). Thus, for
technical simplicity first we assume that g = g̃ = 0. We will prove that for
3.1 A Linear Elliptic Dirichlet Problem 91

every ϕ ∈ W01,2 (I) the expression |huθ − ũθ , ϕi| is bounded by the operator
norms of kL − L̃k and khθ − h̃θ k, where h = (h, f 1 , f 2 , . . . f n ). Here h·, ·i
stands for the usual scalar product in L2 (I).
By subtracting (3.20) from (3.19) we obtain
n
X
Luθ − L̃ũθ + L̃uθ − L̃uθ = hθ − h̃θ + Di (fθi − f˜θi ),
i=1

that is
n
X
L̃(uθ − ũθ ) = hθ − h̃θ + Di (fθi − f˜θi ) − (L − L̃)uθ .
i=1

Let ϕ ∈ W01,2be arbitrary. Denote by ψ ∈ W01,2 the unique solution of the

deterministic Dirichlet problem L̃∗ ψ(x) = ϕ(x), ψ(x) ¹∂I = 0. This is well
defined due to Theorem 3.1.1 since the adjoint operator L̃∗ inherits from L̃
all properties (3.4), (3.5), (3.6) and (3.9). Thus,

|huθ − ũθ , ϕi| = |huθ − ũθ , L̃∗ ψi| = |hL̃(uθ − ũθ ), ψi|
n
X
= |hhθ − h̃θ + Di (fθi − f˜θi ) − (L − L̃)uθ , ψi|.
i=1

Now, by the Cauchy-Schwartz inequality and by continuity of L and L̃

on W 1,2 (I), we get

n
X
|huθ − ũθ , ϕi| ≤ khθ − h̃θ k2 kψk2 + kfθi − f˜θi k2 kDi ψk2 + kL − L̃k2 kuθ k2 kψk2
i=1

≤ khθ − h̃θ k2 kψkW 1,2 + kL − L̃kW −1,2 kuθ kW 1,2 kψkW 1,2 .
(3.21)

From Theorem 3.1.1 we also know there exists C > 0 such that

kψkW 1,2 ≤ CkϕkW 1,2 .

Thus,

|hu(x, ω) − ũ(x, ω), ϕ(x)θ(ω)i| ≤

³ ´
C kh − h̃kL2 (I)⊗(S)−1,−p + kL − L̃kW −1,2 kukW 1,2 ⊗(S)−1,−p kϕkW 1,2 kθk1,p

where h = (h, f 1 , f 2 , . . . f n ).
92 Applications to Singular SPDEs

Now, consider the general Dirichlet problems (3.19), (3.20) with nonzero
boundary conditions. We transform them into zero boundary conditions in
the following way: uθ is a weak solution
Pn of (3.19) if and only if uP
0,θ = uθ −gθ is
a weak solution ofPLu0,θ = h0,θ − i=1 Di f0,θ , where h0,θ = hθ − ni=1 ci Di gθ −
i
i
dgθ , f0,θ = fθi − ni=1 aij Dj gθ − bi gθ . We do the same for (3.20) and apply
the estimate obtained for |hu0 − ũ0 , ϕθi|.
We summarize the stability result in the following proposition.
Proposition 3.1.1 Let L be a strictly elliptic operator of the form (3.2) with
coefficients aij , bi , ci , d, satisfying conditions (3.4), (3.5) and (3.6). Let L̃
be another strictly elliptic operator of the form (3.2) with coefficients ãij , b̃i ,
˜ satisfying all given conditions. Let h, h̃, f i and f˜i , i = 1, 2, . . . , n be
c̃i , d,
generalized random processes from L((S)1 , L2 (I)). Let g and g̃ be generalized
random processes from WS. Let u, ũ ∈ WS∗ be the generalized solutions of
the Dirichlet problems
n
X
L u(x, ω) = h(x, ω) + Di f i (x, ω), x ∈ I, ω ∈ Ω,
i=1
u(x, ω) ¹∂I = g(x, ω)
and
n
X
L̃ ũ(x, ω) = h̃(x, ω) + Di f˜i (x, ω), x ∈ I, ω ∈ Ω,
i=1
ũ(x, ω) ¹∂I = g̃(x, ω)
respectively. There exist C > 0, p ∈ N0 such that for every θ ∈ (S)1,p :
³
|hu(x, ω) − ũ(x, ω), ϕ(x)θ(ω)i| ≤ C kh − h̃kL2 (I)⊗(S)−1,−p

+kL − L̃kW −1,2 ku − gkW 1,2 ⊗(S)−1,−p + kL − L̃kW −1,2 kg̃kW 1,2 ⊗(S)−1,−p
´
+kg − g̃kW 1,2 ⊗(S)−1,−p kLkW −1,2 kϕkW 1,2 kθk1,p
where h = (h, f 1 , f 2 , . . . f n ).
Remark. In particular, to address a similar stability question as in
[LØUZ], let us consider the net of operators L² , ² ∈ (0, 1], given by
n
X n
X
L² u(x, ·) = Di ( aij i
² (x)Dj u(x, ·) + b² (x)u(x, ·))
i=1 j=1
n
X
+ ci² (x)Di u(x, ·) + d² (x)u(x, ·)
i=1
3.1 A Linear Elliptic Dirichlet Problem 93

and a net of data h² (x, ω), f²i (x, ω), i = 1, 2, . . . , n, where aij ij
² (x) = a ∗ ρ̌² (x),
i i i i
b² (x) = b ∗ ρ̌² (x), c² (x) = c ∗ ρ̌² (x), d² (x) = d ∗ ρ̌² (x),h² (x, ω) = h(·, ω) ∗ ρ̌² (x),
f²i (x, ω) = f i (·, ω)∗ ρ̌² (x), i, jR = 1, 2, . . . , n. Here ρ is a mollifier, i.e. a positive
smooth function such that I ρ(x)dx = 1, ρ̌(x) = ρ(−x), ρ² (x) = ²−n ρ(x/²),
and ∗ denotes convolution with respect to x. Denote by u² the solution of
n
X
L² u² = h² + Di f²i , u² ¹∂I = 0.
i=1

Now, from (3.21) we get that kuθ,² − uθ kW −1,2 is bounded by the sum of the
operator norm kL − L² kW −1,2 and of khθ,² − hθ kL2 . It can be shown (see
[NP]) that L² inherits the strict ellipticity and boundedness properties from
L. Also, it is a well-known property of regularization in Sobolev spaces that
kp ∗ ρ² (x) − p(x)kW 1,2 → 0, ² → 0, for p ∈ W 1,2 (I).
Thus,
kuθ,² − uθ kW −1,2 → 0, ² → 0, (3.22)
which establishes the stability result.
This motivates us to consider the stochastic Dirichlet problem in the
Colombeau setting, which will enable us to solve problem (3.1) also with more
singular coefficients and data, i.e. with Colombeau generalized deterministic
functions. In order to do this, we first have to introduce appropriate spaces
of Colombeau algebras of generalized random processes, which will be done
in the next section.
Proposition 3.1.1 and the Remark after it also relate our approach with
those in [HØUZ] and [LØUZ]. Processes involved therein as data in SPDEs
are first smoothed out in order to get an appropriate form for applying the
S-transform and its inverse. For example, singular white noise given by
∞
X
W (x, ω) = ηj (x)Hεj (ω), x ∈ Rn , ω ∈ Ω,
j=1

where ηj , j ∈ N is the Hermite function basis of Rn and εj = (0, 0 . . . , 1, 0, . . .)

is a multiindex, is an element of (S)−1 for fixed x. In [HØUZ] the ”noise”
part in SPDEs is modeled with smoothed white noise given by
∞
X
W (φ, x, ω) = hω, φx i = ω ∗ φ̌(x) = (ηj , φx )L2 (I) Hεj (ω),
j=1

where φ ∈ S(Rn ), φx = φ(· − x), x ∈ Rn and ω ∈ S0 (Rn ). Formally, one can

think of W (x, ω) as of W (δ, x, ω), where δ denotes the Dirac delta distribu-
tion. Smoothed white noise is an element of (L)2 for fixed φ and x. SPDEs
94 Applications to Singular SPDEs

involving smoothed white noise are then solved by standard methods, and
then letting ϕ → δ in S0 (Rn ) one obtains a solution of the SPDE involving
singular white noise. Due to Theorem 3.1.6 we can directly solve the Dirich-
let problem (3.1) where h, f or g is singular white noise, since singular white
noise belongs to L((S)1 , L2 (I)).
For example, consider the Dirichlet problems Lu = W (ρ² , x, ω) and Lu =
W (x, ω), u ¹∂I = 0, with generalized solutions u² , u, respectively. Now, using
the fact that W (ρ² , x, ω) = W (·, ω) ∗ ρ² (x), we get from the Remark after
Proposition 3.1.1

kuθ,² − uθ kW −1,2 ≤ CkWθ,² − Wθ kL2 → 0, ² → 0.

This stability result shows that the solution of the Dirichlet problem
where the right hand side is smoothed white noise, in fact is an adequate
approximation of the solution where the data is singular white noise.

Regularity properties of the Dirichlet problem

One can also show regularity properties (in the x variable) of the solution
obtained in Theorem 3.1.6. So far, we have found generalized solutions as
elements of WS, i.e. solutions which are in the x-variable elements of W 1,2 (I).
Now we can state (for the proof see [GT]) that this generalized solution is in
the x-variable twice weakly differentiable, i.e. u ∈ W 2,2 (I) ⊗ (S)−1 , provided
the domain I and the data in the equation are sufficiently smooth.

Proposition 3.1.2 Let u ∈ W 1,2 (I) ⊗ (S)−1 be the generalized solution of

Lu = h, u ¹∂I = g.

Assume the coefficients aij , bi , i, j = 1, 2, . . . n, are uniformly Lipschitz con-

tinuous in I, ci , d ∈ L∞ (I), i = 1, 2, . . . n, and h ∈ L2 (I) ⊗ (S)−1 . Then, for
arbitrary I 0 such that I 0 ⊂ I, it follows that u ∈ W 2,2 (I 0 ) ⊗ (S)−1 and there
exists C(n, λ, K, d0 ) > 0 such that for each θ ∈ (S)1

kuθ kW 2,2 (I 0 ) ≤ C(kuθ kW 1,2 (I) + khθ kL2 (I) )

where K = max{kaij , bi kC 0,1 (I) , kci , dkL∞ (I) } and d0 = dist(∂I, I 0 ). Addi-
tionally, u satisfies the equation
n
X n X
X n n
X
ij ij i i
Lu= a Dij u + ( (Dj a + b + c )Di u + ( Di bi + d)u = h
i,j=1 i=1 j=1 i=1

for a.e. x ∈ I.
3.1 A Linear Elliptic Dirichlet Problem 95

The next theorem (in deterministic case) was proved in [MP], but we may
state it also in our setting. The polynomial growth rate estimated in (3.23)
will later be crucial for the Colombeau solutions.

Proposition 3.1.3 Assume additionally to the assumptions of Proposition

3.1.2 that I is of C 2 –class and g ∈ W 2,2 (I) ⊗ (S)−1 . There exist s ∈ R and
C > 0 such that for each θ ∈ (S)1
µ ¶s
Λ
kuθ kW 2,2 ≤C (sup |uθ (x)| |I| + khθ kL2 + kgθ kW 2,2 ). (3.23)
λ x∈∂I

3.1.3 Colombeau–type solutions of the stochastic

Dirichlet problem
Now we solve problem (3.1) for more singular coefficients and data – we
assume they are Colombeau generalized functions in the x-variable. The
deterministic Dirichlet problem for linear elliptic operators in Colombeau
setting has been studied in [MP] and [NP]; now we construct stochastic
Colombeau–type algebras which allow us to solve the stochastic Dirichlet
problem.
Colombeau algebras of Colombeau generalized functions are constructed
in such a way that S0 (Rn ) can be embedded into them – thus, they contain
singularities like the Dirac delta distribution. Moreover, one can deal with
multiplication of generalized functions and other nonlinearities.
This approach is motivated by the stability result we obtained in Propo-
sition 3.1.1. If the right hand side of (3.1) and if the coefficients of L involve
singularities in the x-variable, we consider a family of problems L² u² = h² ,
u² ¹∂I = g² , ² ∈ (0, 1), where h² , g² are smooth enough (in our case twice
weakly differentiable) approximations of h and g. Such nets of approxima-
tions are considered as elements of an appropriate quotient algebra where
certain equivalence relations are introduced, in order to have weak equality
between different approximations. Solving this family of problems, we obtain
a family of solutions which represents a Colombeau–algebra solution of the
original problem.
Throughout this section we assume that n ≤ 3 and that ∂I is of C 2 class
(in this case W 2,2 (I) is an algebra).

3.1.4 Colombeau–algebras G(W 2,2 ) and G(W 2,2 ; (S)−1 )

First we define the Colombeau algebra for deterministic functions.
96 Applications to Singular SPDEs

• Let EM (W 2,2 ) be the vector space of functions r : (0, 1) → W 2,2 (I),

² 7→ r² (x), such that there exists a ∈ N0 with the property that
kr² (x)kW 2,2 = O(²−a ). Elements of EM (W 2,2 ) are called moderate func-
tions.

• Let N(W 2,2 ) denote the vector space of functions r² ∈ EM (W 2,2 ) with
the property that for every a ∈ N0 , kr² (x)kW 2,2 = O(²a ) holds. Ele-
ments of N(W 2,2 ) are called negligible functions.

• Since N(W 2,2 ) is an ideal of the algebra EM (W 2,2 ), the quotient space

G(W 2,2 ) = EM (W 2,2 ) / N(W 2,2 )

is also an algebra. It is called the Colombeau extension of W 2,2 (I), and

its elements will be called Colombeau generalized functions. We denote
the elements of G(W 2,2 ) (equivalence classes) by [r² ].
In a similar manner one can define also G(W 1,2 ) and G(W 0,2 ), but they
are not algebras, only vector spaces (however, G(W 1,2 ) is algebra for n = 1).
For f = [f² ] ∈ G(W 1,2 ) we define ∂xi f = [∂xi f² ], i = 1, 2, . . . , n, and note that
∂xki f ∈ G(W 2−k,2 ), k = 1, 2.
The space G(W 2,2 ) is constructed so that it involves singular data, for
example distributions of the form f1 + Df2 + D2 f3 , fi ∈ L2 (I), i = 1, 2, 3,
are obviously embedded into G(W 2,2 ). But S0 (I) can also be embedded into
G(W 2,2 ) via convolution. Note that S(I) ⊆ W 2,2 (I) ⊆ L2 (I), and for a fixed
mollifier function ρ² , ² ∈ (0, 1) we have that if f ∈ S(I), then f² = f ∗ ρ² ∈
S(I) ⊆ W 2,2 (I) and clearly, f − f² ∈ N(W 2,2 ). Thus, [f² ] ∈ G(W 2,2 ). The
embedding ιρ : S(I) → G(W 2,2 ), f 7→ [f² ], can be extended to an embedding
ιρ : S0 (I) → G(W 2,2 ), defined by

S0 (I) 3 F 7→ [F ∗ ρ² ] ∈ G(W 2,2 ).

We continue with the construction of the appropriate Colombeau–type

algebra for generalized random processes. The construction is similar as in
the previous case, but now we consider (S)−1 –valued functions. The idea is to
use the chaos expansion in (S)−1 : All coefficients in the chaos expansion will
be deterministic Colombeau generalized functions i.e. elements of G(W 2,2 ),
but certain conditions must hold so that the result remains in (S)−1 .

• Let EM (W 2,2 ; (S)−1 ) be the vector space

P of functions R : (0, 1) →
2,2
W (I) ⊗ (S)−1 , ² 7→ R² (x, ω) = α∈I rα,² (x) ⊗ Hα (ω), rα,² ∈
2,2
W (I), x ∈ I, ω ∈ Ω, such that there exist a sequence {Cα }α∈I of
positive numbers, ²0 ∈ (0, 1), p ∈ N0 , and there exists a sequence
3.1 A Linear Elliptic Dirichlet Problem 97

{aα }α∈I bounded from above (i.e. there exists a ∈ R such that aα ≤ a,
α ∈ I), with following properties:

krα,² kW 2,2 ≤ Cα ²−aα , for all α ∈ I, ² < ²0 , (3.24)

X
Cα2 (2N)−pα < ∞. (3.25)
α∈I

Clearly, from (3.24) and (3.25) we have

X X X
krα,² k2W 2,2 (2N)−pα ≤ Cα2 ²−2aα (2N)−pα ≤ ²−2a Cα2 (2N)−pα
α∈I α∈I α∈I
−2a
= K² , ² < ²0 ,
P
where K = α∈I Cα2 (2N)−pα .
• Let N(W 2,2 ; (S)−1 ) denote the vector space of functions R² ∈
EM (W 2,2 ; (S)−1 ) with the property that there exist a sequence {Cα }α∈I
of positive numbers, ²0 ∈ (0, 1), p ∈ N0 , and for all sequences {aα }α∈I
bounded from below (i.e. there exists a ∈ R such that aα ≥ a, α ∈ I),
following hold:
krα,² kW 2,2 ≤ Cα ²aα , for all α ∈ I, ² < ²0 , (3.26)
X
Cα2 (2N)−pα < ∞. (3.27)
α∈I

Clearly, from (3.26) and (3.27) we have

X X X
krα,² k2W 2,2 (2N)−pα ≤ Cα2 ²2aα (2N)−pα ≤ ²2a Cα2 (2N)−pα
α∈I α∈I α∈I
= K²2a , ² < ²0 ,
P
where K = α∈I Cα2 (2N)−pα .
2,2
• Define the multiplication in EM (WP ; (S)−1 ) and N(W 2,2 ; (S)−1 ) in
the following way: For F² (x, ω) = α∈I fα,² (x) ⊗ Hα (ω), G² (x, ω) =
P
α∈I gα,² (x) ⊗ Hα (ω) let
à !
X X
F² ♦G² (x, ω) = fα,² (x) · gα,² (x) ⊗ Hγ (ω), (3.28)
γ∈I α+β=γ

i.e. we use the Wick product for multiplication in (S)−1 and the ordi-
nary product in W 2,2 (I).
98 Applications to Singular SPDEs

Lemma 3.1.1 (i) EM (W 2,2 ; (S)−1 ) is an algebra under the multipli-

cation rule given by (3.28).
(ii) N(W 2,2 ; (S)−1 ) is an ideal of EM (W 2,2 ; (S)−1 ).

Proof. (i) Let F² , G² ∈ EMP (W 2,2 ; (S)−1 ) and prove that F² ♦G² ∈
2,2 2,2
EM (W ; (S)P −1 ). Since F² = α∈I fα,² (x) ⊗ Hα (ω) ∈ EM (W ; (S)−1 )
2,2
and G² = g
β∈I β,² (x) ⊗ H β (ω) ∈ E M (W ; (S)−1 ), there exist
a, b ∈ R, p, q ∈ N0 , ²1 , ²2 ∈ (0, 1), and there exist sequences {Cα }α∈I ,
{Dα }α∈I , {aα }α∈I , {bα }α∈I , such that aα ≤ a, bα ≤ b for all α ∈ I,
kf
Pα,² kW 2,2 ≤ Cα ²−aα forP² < ²1 , kgα,² kW 2,2 ≤ Dα ²−bα for ² < ²2 , and
2 −pα
α∈I Cα (2N) < ∞, α∈I Dα2 (2N)−qα < ∞. We will prove that for
²0 = min{²1 , ²2 } there exist r > 0, a sequence P {Mγ }γ∈I and a sequence
{mγ }γ∈I bounded from above such that k α+β=γ fα,² (x)gα,² (x)kW 2,2 ≤
P
Mγ ²−mγ for ² < ²0 and γ∈I Mγ2 (2N)−rα < ∞.
P
For a fixed multiindex γ ∈ I put Mγ = α+β=γ Cα Dβ and mγ =
max{aα + bβ : α, β ∈ I, α + β = γ} (note, for γ fixed, there are only
finite many α and β which give the sum γ). Now using the fact that
W 2,2 (I) is an algebra for n ≤ 3, we get
X X
k fα,² (x)gα,² (x)kW 2,2 ≤ kfα,² (x)kW 2,2 kgα,² (x)kW 2,2 ≤
α+β=γ α+β=γ
X X
Cα Dβ ²−(aα +bβ ) ≤ ²−mγ Cα Dβ = ²−mγ Mγ .
α+β=γ α+β=γ

Clearly, mγ ≤ a + b, γ ∈ I, thus the sequence {mγ }γ∈I is bounded from

above. Let r = p+q +2. Then, using the nuclearity of (S)−1 , we obtain
à !2
X X X
Mγ2 (2N)−rα ≤ Cα Dβ (2N)−(p+q+2)γ
γ∈I γ∈I α+β=γ
X X X
≤ (2N)−2γ Cα2 (2N)−pα Dβ2 (2N)−qβ < ∞.
γ∈I α∈I β∈I

(ii) Let us check first that N(W 2,2 ; (S)−1 ) is a subalgebra of

EM (W 2,2 ; (S)−1 ). Let F² , G² ∈ N(W 2,2 ; (S)−1 ) be of the form as in (i).
Let {mγ }γ∈I be an arbitrary sequence bounded from below. Put aγ =
bγ = m2γ , γ ∈ I. Now for the sequences {aγ }γ∈I , {bγ }γ∈I also bounded
from below, since F² , G² ∈ N(W 2,2 ; (S)−1 ), there must exist p, q ∈ N0 ,
²1 , ²2 ∈ (0, 1), and {Cα }α∈I , {Dα }α∈I such thatPkfα,² kW 2,2 ≤ Cα ²mα /2
for ² < ²1 , kgα,² kW 2,2 ≤ Dα ²mα /2 for ² < ²2 , and α∈I Cα2 (2N)−pα < ∞,
3.1 A Linear Elliptic Dirichlet Problem 99

P 2 −qα
DP
α∈Iα (2N) < ∞. Let ²0 = min{²1 , ²2 }, r = p + q + 2 and
Mγ = α+β=γ α Dβ , γ ∈ I. Now we may
C proceed as in (i) to get
P mγ
P
α+β=γ kfα,² kW 2,2 kgα,² kW 2,2 ≤ Mγ ² and γ∈I Mγ2 (2N)−rα < ∞.
In order to prove that N(W 2,2 ; (S)−1 ) is an ideal of EM (W 2,2 ; (S)−1 ),
we must check that for all G² ∈ N(W 2,2 ; (S)−1 ) and all F² ∈
EM (W 2,2 ; (S)−1 ), we have G² ♦F² = F² ♦G² ∈ N(W 2,2 ; (S)−1 ). Let
{nγ }γ∈I be an arbitrary sequence bounded from below (i.e. nγ ≥ n,
γ ∈ I). Since F² ∈ EM (W 2,2 ; (S)−1 ), there exist p ∈ N0 , ²1 ∈ (0, 1),
{Dβ }β∈I , b ∈ R and a sequenceP{bβ }β∈I such that bβ ≤ b, β ∈ I,
kfβ,² kW 2,2 ≤ Dβ ²−bβ , ² < ²1 , and β Dβ2 (2N)−pβ < ∞.
For a fixed multiindex α ∈ I, let aα = b + nα . Clearly, the se-
quence {aα }α∈I is bounded from below by a = b + n, thus since
G² ∈ N(W 2,2 ; (S)−1 ), there exist q ∈ N0 , P
²2 ∈ (0, 1), {Cα }α∈I , such
−aα 2 −qα
that kgα,² kW 2,2 ≤ Cα ² ,P² < ²2 , and α Cα (2N) < ∞. Let
²0 = min{²1 , ²2 } and Nγ = α+β=γ Cα Dβ , γ ∈ I.
P P
Now for G² ♦F² = γ∈I α+β=γ gα,² fα,² ⊗ Hγ we have that
X X
kgα,² kW 2,2 kfβ,² kW 2,2 ≤ Cα Dβ ²aα −bβ
α+β=γ α+β=γ

X X
≤ Cα Dβ ²b+nα −bβ ≤ Cα Dβ ²bβ +nα −bβ ≤ Nγ ²nγ , ² < ²0 .
α+β=γ α+β=γ
P
It is clear that γ∈I Nγ2 (2N)−rγ < ∞ for r = p + q + 2. Thus, G² ♦F² ∈
N(W 2,2 ; (S)−1 ). ¤

• Since N(W 2,2 ; (S)−1 ) is an ideal of the algebra EM (W 2,2 ; (S)−1 ), the
quotient space

G(W 2,2 ; (S)−1 ) = EM (W 2,2 ; (S)−1 ) / N(W 2,2 ; (S)−1 )

is also an algebra. It is called the (S)−1 –valued Colombeau extension

of W 2,2 (I), and its elements will be called Colombeau generalized ran-
dom processes. We denote the elements of G(W 2,2 ; (S)−1 ) (equivalence
classes) by [R² ].

Note that (S)−1 can be embedded into G(W 2,2 ; (S)−1 ) by

X X
(S)−1 3 F (ω) = fα Hα (ω) Ã fα (x) ⊗ Hα (ω) ∈ G(W 2,2 ; (S)−1 ),
α∈I α∈I
100 Applications to Singular SPDEs

where fα (x) = fα is the constant mapping.

Also, G(W 2,2 ) can be embedded into G(W 2,2 ; (S)−1 ) by
X
G(W 2,2 ) 3 f = [f² (x)] Ã [f² (x)]Hα (ω) ∈ G(W 2,2 ; (S)−1 ).
α=(0,0,0,...)

The following proposition ensures that it is sufficient to consider only the

cases when the polynomial ²-growth rate is uniform in α ∈ I.
Proposition 3.1.4 (i) R² (x, ω), x ∈ I, ω ∈ Ω, belongs to EM (W 2,2 ; (S)−1 )
if and only if there exist ²0 ∈ (0, 1), C > 0, a ∈ R and q ≥ 0 such that
X Z
kDβ R² (x, ·)k2−1,−q dx ≤ C²−a , ² < ²0 . (3.29)
0≤|β|≤2 I

(ii) R² (x, ω), x ∈ I, ω ∈ Ω, belongs to EM (W 2,2 ; (S)−1 ) if and only if there

exist ²0 ∈ (0, 1), C > 0, a ∈ R and q ≥ 0 such that

kR² (x, ω)kW 2,2 (I)⊗(S)−1,−q ≤ C²−a , ² < ²0 . (3.30)

(iii) R² (x, ω), x ∈ I, ω ∈ Ω, belongs to N(W 2,2 ; (S)−1 ) if and only if there
exist ²0 ∈ (0, 1), C > 0 and q ≥ 0 such that for every a ∈ R,
X Z
kDβ R² (x, ·)k2−1,−q dx ≤ C²a , ² < ²0 . (3.31)
0≤|β|≤2 I

(iv) R² (x, ω), x ∈ I, ω ∈ Ω, belongs to N(W 2,2 ; (S)−1 ) if and only if there
exist ²0 ∈ (0, 1), C > 0 and q ≥ 0 such that for every a ∈ R,

kR² (x, ω)kW 2,2 (I)⊗(S)−1,−q ≤ C²a , ² < ²0 . (3.32)

Proof. (i) Let R² ∈ EM (W 2,2 ; (S)−1 ). Then, there exist b ∈ R, p ≥

0, {bα }α∈I , such thatPbα ≤ b, α ∈ I, krα,² kW 2,2 ≤ Cα ²−bα ,
P {Cα }2α∈I , −pα
α∈I Cα (2N) < ∞. Let C = α∈I Cα2 (2N)−pα and a = 2b. Then,
X Z X Z X
β 2
kD R² (x, ·)k−1,−p dx = |Dβ rα,² (x)|2 (2N)−pα dx
0≤|β|≤2 I 0≤|β|≤2 I α∈I

XZ X X
= |Dβ rα,² (x)|2 dx(2N)−pα = krα,² k2W 2,2 (2N)−pα
α∈I I 0≤|β|≤2 α∈I
X X
≤ Cα2 ²−2bα (2N)−pα ≤ ²−2b Cα2 (2N)−pα = C²−a .
α∈I α∈I
3.1 A Linear Elliptic Dirichlet Problem 101

Conversely,Plet (3.29) hold and let us show that R² ∈ EM (W 2,2 ; (S)−1 ).

Denote K := α∈I (2N)−qα < ∞ for an appropriate q > 1. Let R² (x, ω) =
P
α∈I rα,² (x) ⊗ Hα (ω) as in construction (3.25). Then,

à !4 à !2
X X
krα,² kW 2,2 (2N)−qα ≤ K krα,² k2W 2,2 (2N)−qα
α∈I α∈I

 2
Z X X
≤ K2  |Dβ rα,² (x)|2 (2N)−qα dx
I α∈I 0≤|β|≤2

 2
Z X X
≤ K 2 |I|  |Dβ rα,² (x)|2 (2N)−qα  dx
I α∈I 0≤|β|≤2

Z Ã !2
X X
2 β 2 −qα
≤ 2K |I| |D rα,² (x)| (2N) dx
I 0≤|β|≤2 α∈I
Z X
= K̃ kDβ R² (x, ·)k2−1,−q dx ≤ K̃C²−a .
I 0≤|β|≤2

P p
4 a
Thus, α∈I krα,² kW 2,2 (2N)
−qα
≤ K̃C²− 4 , which implies that for each α ∈ I
a p
krα,² kW 2,2 ≤ Cα ²− 4 , where Cα = (2N)qα 4
K 2 |I|C,

and X
Cα2 (2N)−pα < ∞, for p > 1 + 2q.
α∈I

(ii) Since

kR² (x, ω)k2W 2,2 (I)⊗(S)−1,−q = hR² (x, ω), R² (x, ω)i

X X
= hrα,² , rα,² iW 2,2 hHα , Hα i−1,−q = krα,² k2W 2,2 (2N)−qα ≤ C²−a ,
α∈I α∈I

it follows similarly as in (i) that krα,² k2W 2,2 ≤ C(2N)qα ²−a for each α ∈ I.
(iii) and (iv) follow similarly as (i) and (ii). ¤
102 Applications to Singular SPDEs

Elliptic differential operators on G(W 2,2 ; (S)−1 )

We consider a net of linear differential operators
n
X n
X
L² u² (x, ·) = Di ( aij i
² (x)Dj u² (x, ·) + b² (x)u² (x, ·))
i=1 j=1
n
X
+ ci² (x)Di u² (x, ·) + d² (x)u² (x, ·), ² ∈ (0, 1),
i=1
(3.33)

where the nets of coefficients aij i i

² , b² , c² , d² belong to EM (W
2,2
). Define that
two nets of operators are related: L² ∼ L̃² if and only if (L² u² − L̃² u² ) ∈
N(W 0,2 ; (S)−1 ) for all u² ∈ EM (W 2,2 ; (S)−1 ). Clearly, ∼ is an equivalence
relation, and following holds (refer to [MP]):

Proposition 3.1.5 If u² ∈ N(W 2,2 ; (S)−1 ), then L² u² ∈ N(W 0,2 ; (S)−1 ).

Denote by L the family of all nets of differential operators of the form

(3.33) and let L0 = L/ ∼. For L ∈ L0 we define L : G(W 2,2 ; (S)−1 ) →
G(W 0,2 ; (S)−1 ) by
n
X n
X
L [u² (x, ·)] = Di ( [aij i
² (x)][Dj u² (x, ·)] + [b² (x)][u² (x, ·))]
i=1 j=1
n
(3.34)
X
+ [ci² (x)][Di u² (x, ·)] + [d² (x)][u² (x, ·)].
i=1

The operator L = [L² ] given by (3.34) is strictly elliptic, if there exist

representatives aij i i
² , b² , c² , d² ∈ EM (W
2,2
) of its coefficients, such that
n
X
aij 2 a 2
² (x)ξi ξj ≥ λ² |ξ| ≥ K² |ξ| , x ∈ I, ξ ∈ Rn , (3.35)
i,j=1

where K is a constant independent of ².

Boundary conditions in G(W 2,2 ; (S)−1 )

Let u, g ∈ G(W 2,2 ; (S)−1 ). It is known that W 2,2 (I) ⊆ C(I1 ), where I1 is
open and I ⊆ I1 . We define u ¹∂I = g ¹∂I if there exist representatives u² and
g² of u and g, respectively, such that

u² ¹∂I = g² ¹∂I + n² ¹∂I , ² ∈ (0, 1)

3.1 A Linear Elliptic Dirichlet Problem 103

where n² is a net of continuous GRPs, defined in a neighborhood of ∂I

with the property that there exists p > 0 such that for all a > 0,
supx∈∂I kn² (x, ·)k−1,−p = o(²a ).
It is easy to check that this definition does not depend on the choice of
representatives: Let ũ² and g̃² be another choice of representatives for u and
g, respectively. Then ũ² ¹∂I = g̃² ¹∂I + ñ² ¹∂I , where ñ² ¹∂I = (h² − h̃² ) ¹∂I
+ n² ¹∂I +(u² − ũ² ) ¹∂I . This implies supx∈∂I kn² (x, ·)k−1,−p = o(²a ), for all
a > 0.

The Dirichlet problem in G(W 2,2 ; (S)−1 )

Consider now the stochastic Dirichlet problem

L u(x, ω) = h(x, ω) in G(W 2,2 ; (S)−1 )

(3.36)
u(x, ω) ¹∂I = g(x, ω) ¹∂I ,

where L is defined in (3.34). In order to solve (3.36) in the Colombeau

setting, one has to solve a family of problems:

L² u² (x, ω) = h² (x, ω) in W 2,2 (I) ⊗ (S)−1

(3.37)
u² (x, ω) ¹∂I = g² (x, ω) ¹∂I , ² ∈ (0, 1),

then to check whether the net of solutions u² belongs to EM (W 2,2 ; (S)−1 ),

and finally to check whether equation (3.37) holds with other representatives
of L, h, f , g and u. Uniqueness of a solution means that if u = [u² ] and
v = [v² ] satisfy (3.36), then [u² ] = [v² ].
P P
Theorem 3.1.7 Let h = [h² ] = α∈I [hα,² ] ⊗ Hα and g = [g² ] = α∈I [gα,² ] ⊗
Hα belong to G(W 2,2 ; (S)−1 ). Let the operator L = [L² ] be given by (3.34) with
coefficients [aij i i
² ], [b² ], [c² ], [d² ] in G(W
2,2
). Assume that following conditions
hold:
1. L is strictly elliptic, i.e. (3.35) holds,

2. there exist M > 0 and b > 0 such that for all ² ∈ (0, 1) and α ∈ I:

kaij i i
² kW 2,2 , kb² kW 2,2 , kc² kW 2,2 , kd² kW 2,2 ≤ Λ² ;

khα,² kW 2,2 , kgα,² kW 2,2 ≤ Λ² ;

Λ² ≤ M ²b .
104 Applications to Singular SPDEs

3. for every ² ∈ (0, 1) and v ≥ 0, v ∈ W01,2 (I),

Z n
X
(d² (x)v(x) − bi² (x)Di v(x)) dx ≤ 0.
I i=1

Then the stochastic Dirichlet problem (3.36) has a unique solution in

G(W 2,2 ; (S)−1 ).

Proof. For ² ∈ (0, 1) fixed, there exists a unique generalized solution

u² ∈ W 2,2 (I) ⊗ (S)−1 of problem (3.37). This follows from Theorem 3.1.6,
and
P moreover we know that u²2,2is given by the chaos expansion u² (x, ω) =
α∈I uα,² (x)⊗Hα (ω), uα,² ∈ W (I), x ∈ I, ω ∈ Ω, where uα,² is the solution
of the deterministic Dirichlet problem

L² uα,² (x) = hα,² (x) in W 2,2 (I)

(3.38)
uα,² (x) ¹∂I = gα,² (x) ¹∂I .

Using the estimate derived in (3.23) we obtain that there exist C > 0,
s > 0 (from the proof in [MP] one can see that C and s may depend on n, λ,
|I|, ∂I, and the coefficients of L, but they are uniform in ² and α) such that:
µ ¶s ³ ´
Λ²
kuα,² kW 2,2 ≤ C sup |gα,² (x)| · |I| + kgα,² kW 2,2 + khα,² kW 0,2 . (3.39)
λ² x∈∂I

Now, since gα,² and hα,² , α ∈ I, are all bounded by Λ² , which has polyno-
mial growth with respect to ², we obtain

kuα,² kW 2,2 ≤ C̃²a , α ∈ I,

for an appropriate a ∈ R, C̃ > 0. Thus, there exists p > 1 such that

X X
kuα,² (x)k2W 2,2 (2N)−pα ≤ C̃ 2 ²2a (2N)−pα < ∞.
α∈I α∈I

This proves that [u² (x, ω)] ∈ G(W 2,2 ; (S)−1 ).

The definition of operators in L0 implies that with some other represen-
tatives of h, g and the coefficients of L in (3.37), another representative of
u = [u² ] also satisfies (3.36).
Now, we prove uniqueness of the solution: Let u1,² and u2,² satisfy

L1,² u² (x, ω) = h1,² (x, ω) in W 2,2 (I) ⊗ (S)−1

(3.40)
u² (x, ω) ¹∂I = g1,² (x, ω) ¹∂I , ² ∈ (0, 1),
3.1 A Linear Elliptic Dirichlet Problem 105

and
L2,² u² (x, ω) = h2,² (x, ω) in W 2,2 (I) ⊗ (S)−1
(3.41)
u² (x, ω) ¹∂I = g2,² (x, ω) ¹∂I , ² ∈ (0, 1),

respectively, where the operators Lk,² , k = 1, 2, are of the form (3.33), if we

replace the coefficients with aij i i
k,² , bk,² , ck,² , dk,² , k = 1, 2, respectively, and
(aij ij i i i i
1,² − a2,² ), (b1,² − b2,² ), (c1,² − c2,² ), (d1,² − d2,² ) ∈ N(W
2,2
), and (h1,² − h2,² )
(g1,² − g2,² ) ∈ N(W 2,2 ; (S)−1 ) holds.
We will prove (u1,² − u2,² ) ∈ N(W 2,2 ; (S)−1 ), which establishes the asser-
tion [u1,² ] = [u2,² ].
For technical simplicity, for k = 1, 2, let us denote by Ak,² the ma-
trix [aij 1 n
k,² ]i,j=n×n , and by bk,² , ck,² the n-dimensional vectors (bk,² , . . . , bk,² ),
1 n
(ck,² , . . . , ck,² ), respectively. Then (3.33) can be written as

Lk,² u² = ∇ · (Ak,² · ∇u² + bk,² · u² ) + ck,² · ∇u² + dk,² u² ,

where · denotes the standard scalar product in Rn .

If we insert u1,² into (3.40) and u2,² into (3.41), then subtract (3.40) from
(3.41), we obtain

∇·((A2,² − A1,² ) · ∇u1,² + (b2,² − b1,² ) · u1,² ) + (c2,² − c1,² ) · ∇u1,²

+ (d2,² − d1,² )u1,² + ∇ · (A1,² · ∇(u2,² − u1,² ) + b1,² · (u2,² − u1,² ))
+ c1,² · ∇(u2,² − u1,² ) + d1,² (u2,² − u1,² )
= h2,² − h1,²
(u2,² − u1,² ) ¹∂I = (g2,² − g1,² ) ¹∂I , ² ∈ (0, 1).

Thus,

∇ · (A1,² · ∇(u2,² − u1,² ) + b1,² · (u2,² − u1,² )) + c1,² · ∇(u2,² − u1,² ) + d1,² (u2,² − u1,² )
= h2,² − h1,² − (∇ · ((A2,² − A1,² ) · ∇u1,²
+ (b2,² − b1,² ) · u1,² ) + (c2,² − c1,² ) · ∇u1,² + (d2,² − d1,² )u1,² )
= H²
(u2,² − u1,² ) ¹∂I = (g2,² − g1,² ) ¹∂I , ² ∈ (0, 1),

where H² ∈ N(W 2,2 ; (S)−1 ). Using the estimate as in (3.39) we finally obtain
that there exist p > 1 and C > 0 such that for all a ∈ R
X X
kuα,2,² (x) − uα,1,² (x)k2W 2,2 (2N)−pα ≤ C 2 ²2a (2N)−pα < ∞.
α∈I α∈I
106 Applications to Singular SPDEs

Thus, (u2,² − u1,² ) ∈ N(W 2,2 ; (S)−1 ). ¤

Remark. Considering only the term for α = (0, 0, 0, . . .) in the expan-

sion we get that the generalized expectation E([u² (x, ·)]) coincides with the
Colombeau solution [v² (x)] of the deterministic Dirichlet problem
E([L² ]) [v² ] = E([hα,² ]), [v² ] ¹∂I = E([gα,² ]) ¹∂I .

3.2 A Linear Elliptic Dirichlet Problem with

Stochastic Coefficients and
Stochastic Data
This section is devoted to the Dirichlet problem (3.1), when the coeffi-
cients of the operator L are also generalized random processes. We will give
an interpretation of the operator itself, of the equation and its solution in
terms of Wick products. In praxis, the Wick product showed not just as
typographical phantasy, but also as a good model in physical interpretation.
Historically, the Wick product first arised in quantum physics, later it was in-
troduced in the framework of white noise analysis. It is closely related to the
notion of renormalization and to the S-transform, which converts SPDEs
into PDEs and converts the Wick product into ordinary products. From
probabilistic viewpoint, the solutions of an SPDE may show different prop-
erties depending on the type of product (Wick or ordinary) that was used
for modeling. But in order to investigate these probabilistic properties, one
has to consider each SPDE separately (for examples see [HØUZ]).
The plan of exposition is following: At the very beginning we develop all
necessary tools for the Hilbert space methods to be used. The solution of
the stochastic Dirichlet problem, similarly as in the case of deterministic co-
efficients, exists and is unique. In Theorem 3.2.4 we investigate the stability
and regularity properties of this solution. The aim of Theorem 3.2.8 is to
extend the approach for even more singular stochastic coefficients, namely
Colombeau generalized stochastic processes. We prove existence and unique-
ness of the solution for the Dirichlet problem in this setting, too.
In order to prove existence and uniqueness of the solution of (3.1) provided
the coefficients of the operator L are also generalized random processes, we
have to go deeper into the Hilbert space structure of L(W 1,2 , (S)−1,−p ). The
main reference for this remain [GT] and [Ev]. Throughout this part of the
paper h·, ·i will denote (unless otherwise stated) the dual
P pairing of L2 (I) ⊗
(S)−1,−p and L2 (I) ⊗ (S)1,p . P That is, for F (x, ω) = α∈I Fα (x) ⊗ Hα (ω),
Fα ∈ W −1,2 (I), such that 2
α∈I kFα kW −1,2 (2N)
−pα
< ∞ and f (x, ω) =
3.2 A Dirichlet Problem with Stochastic Coefficients 107

P P
α∈I fα (x)Hα (ω), fα ∈ W01,2 (I) such that α∈I kfα k2W 1,2 (2N)pα < ∞, the
dual action is: X
hF, f i = α!hFα , fα iL2 (I) .
α∈I
qP
Clearly, the norm corresponding to h·, ·i is kF k = α∈I α!kFα k2L2 (I) .

3.2.1 Wick products

The product between two generalized random processes in (3.2) will be
interpreted as the Wick product (recall Definition 2.3.1).
We will now define the Wick product of two generalized random processes
(recall Theorem 3.1.5) in an analogous way. First we introduce the class of
essentially bounded (in the x variable) generalized random processes, for
which this product will be well defined.

Definition 3.2.1 Let F ∈PWS∗ be a generalized random process given by

chaos expansion F (x, ω) = α∈I fα (x)⊗Hα (ω). We will call F an essentially
bounded generalized random process, if fα ∈ L∞ (I) for all α ∈ I and if there
exists p > 0 such that
X
kfα k2L∞ (I) (2N)−pα < ∞.
α∈I

Note that we may consider essentially bounded GRPs as elements of the

tensor product space L∞ (I) ⊗ (S)−1 . This follows from the fact that L1 (I) is
complete and (S)1 is nuclear, thus L∞ (I)⊗(S)−1 ∼ = L(L1 (I); (S)−1 ). Clearly,
for fixed p > 0 we also have L∞ (I) ⊗ (S)−1,−p ∼= L(L1 (I); (S)−1,−p ).

Lemma 3.2.1 F is an essentially bounded GRP if and only if there exists

C ∈ (S)−1 such that

esp F (x, ω) = C(ω) for a.e. ω ∈ Ω,

x∈I
P P
where espx∈I F (x, ω) = α∈I essupx∈I |fα (x)|Hα (ω) = α∈I kfα kL∞ (I) Hα (ω),
provided that essupx∈I |fα (x)| exists for all α ∈ I.
P
Proof. Put C(ω) = α∈I kfα kL∞ (I) Hα (ω). ¤

Definition
P 3.2.2 Let F be an essentially bounded GRP given by F (x, ω) =
f
α∈I α (x) ⊗ Hα (ω), fα ∈ L∞ (I). Let G ∈ WS∗ be given by its chaos
108 Applications to Singular SPDEs

P
expansion G(ω) = β∈I gβ (x) ⊗ Hβ (ω), gβ ∈ W −1,2 (I). The Wick product of
F and G is the unique element in WS∗ defined by:
à !
X X
F ♦G(x, ω) = fα (x)gβ (x) ⊗ Hγ (ω).
γ∈I α+β=γ

The next lemma shows that the Wick product is well defined, and that
for fixed F the mapping G 7→ F ♦G is continuous.

Lemma 3.2.2 If F is essentially bounded, and if G ∈ WS∗ , then F ♦G ∈

WS∗ . Moreover, there exist C, p, q, r > 0 such that:

kF ♦GkW −1,2 ⊗(S)−1,−p ≤ CkF kL∞ ⊗(S)−1,−q kGkW −1,2 ⊗(S)−1,−r .

Proof.
Since F is essentially bounded and G is an element of WS∗ , there exist
q > 0 and r > 0 such that
X X
kfα k2L∞ (2N)−qα < ∞ and kgβ k2W −1,2 (2N)−rα < ∞.
α β

By definition of the
P Wick −kγ
product and using the Cauchy-Schwartz inequality,
we have for C = γ (2N) and p = q + r + k, for arbitrary k > 1
° °2
X°
° X °
°
° fα (x)g β (x) ° (2N)−pγ
° °
γ α+β=γ W −1,2

à !à !
X X X
≤ (2N)−kγ (2N)−(q+r)γ kfα k2L∞ (I) kgβ k2W −1,2
γ α+β=γ α+β=γ

à !à !
X X X
≤ (2N)−kγ kfα k2L∞ (I) (2N)−qα kgβ k2W −1,2 (2N)−rβ .
γ α β ¤

3.2.2 Interpretation of the operator L

According to Definition 3.2.2, we will assume that the coefficients
a , bi , ci , d, for i, j = 1, 2 . . . , n, of the operator L are essentially bounded
ij

GRPs, and will thus interpret the action of L onto a GRP u ∈ WS∗ as:
3.2 A Dirichlet Problem with Stochastic Coefficients 109

n
X n
X
L♦u(x, ω) = Di ( aij (x, ω)♦Dj u(x, ω) + bi (x, ω)♦u(x, ω))
i=1 j=1
n
(3.42)
X
+ ci (x, ω)♦Di u(x, ω) + d(x, ω)♦u(x, ω).
i=1

Now the operator L acts as a differential operator (in the x variable)

and as a Wick–multiplication operator (in the ω variable) as well. Thus,
it is natural to consider the ”package” L♦ as a whole unit. But now we
have to take into account that for Wick multiplication in general hf ♦g, hi 6=
hf, g♦hi, while in the deterministic case we used hf g, hi = hf, ghi for ordinary
multiplication of functions. Nevertheless, we will develop the necessary tools
for Wick calculations.
First we note that Wick multiplication satisfies the chain rule:

Lemma 3.2.3 For an arbitrary GRP f such that Di f , i = 1, 2, . . . , n, are

essentially bounded and arbitrary g ∈ WS∗ we have

Di (f ♦g) = (Di f )♦g + f ♦(Di g), i = 1, 2, . . . , n.

P P
Proof. Let f (x, ω) = α fα (x) ⊗ Hα (ω), g(x, ω) = β gβ (x) ⊗ Hβ (ω).
Then, for arbitrary i = 1, 2, . . . , n (recall, Di denotes the weak derivative
with respect to xi ), we have
X X
Di f (x, ω) = Di fα (x)⊗Hα (ω), and Di g(x, ω) = Di gβ (x)⊗Hβ (ω).
α β

Due³ to definition ´ of Wick multiplication, f ♦g(x, ω) =

P P
γ α+β=γ fα (x)gβ (x) ⊗ Hγ (ω), and thus

à !
X X
Di f ♦g(x, ω) = Di fα (x)gβ (x) ⊗ Hγ (ω)
γ α+β=γ
à !
X X
= Di (fα (x)gβ (x)) ⊗ Hγ (ω)
γ α+β=γ
à !
X X
= Di fα (x)gβ (x) + fα (x)Di gβ (x) ⊗ Hγ (ω)
γ α+β=γ

= Di f (x, ω)♦g(x, ω) + f (x, ω)♦Di g(x, ω). ¤

110 Applications to Singular SPDEs

Fix now an essentially bounded GRP a ∈ L∞ (I) ⊗ (S)−1,−p . Due to

Lemma 3.2.2 we have that the operator A : WS∗ → WS∗ defined by g 7→ a♦g
is a continuous linear operator from the Hilbert space W −1,2 ⊗ (S)−1,−p into
the Hilbert space W −1,2 ⊗ (S)−1,−2(p+1) . (Note, we fixed k = 2 from the proof
of Lemma 3.2.2). For a fixed c ∈ W01,2 ⊗ (S)1,2(p+1) consider the operator
Fc : W −1,2 ⊗ (S)−1,−2(p+1) → R, given by Fc (b) = ha♦b, ci. According to
Lemma 3.2.2 and the Cauchy-Schwartz inequality we have

|Fc (b)| = |ha♦b, ci| ≤

kakL∞ (I)⊗(S)−1,−p kbkW −1,2 ⊗(S)−1,−2(p+1) kckW 1,2 ⊗(S)1,2(p+1)

i.e. Fc is a linear continuous operator on the Hilbert space W −1,2 ⊗
(S)−1,−2(p+1) . Due to the Riesz representation theorem there exists a unique
fc ∈ W01,2 ⊗ (S)1,2(p+1) such that Fc (b) = hb, fc i.
This defines a mapping c 7→ fc , which we will denote by a~ . In fact,
a~ : c 7→ fc is the adjoint operator of A.

Definition 3.2.3 The unique linear continuous mapping a~ : W01,2 ⊗

(S)1,2(p+1) → W01,2 ⊗ (S)1,2(p+1) such that for each b ∈ W −1,2 ⊗ (S)−1,−p
and c ∈ W01,2 ⊗ (S)1,2(p+1)

ha♦b, ci = hb, a~ (c)i (3.43)

holds, is called the Wick–adjoint multiplication operator of the generalized

random process a ∈ L∞ (I) ⊗ (S)−1,−p .

In further notation we will suppress the indexes, and simply denote h·, ·i
and k · k. From the context it is clear whether the index is L∞ (I) ⊗ (S)−1,−p
or L2 (I) ⊗ (S)±1,±p or L2 (I) ⊗ (S)±1,±2(p+1) .
Due to the previous definition we are able to develop something that
might be called a weak Wick calculus: The Wick-adjoint operator inherits
its properties from the classical Wick multiplication – most important, the
chain rule holds in weak sense.

Lemma 3.2.4 For an arbitrary GRP f such that Di f , i = 1, 2, . . . , n, are

essentially bounded and arbitrary g ∈ WS∗ we have

hDi (f ~ g), vi = h(Di f ~ )g + f ~ (Di g), vi,

for all v ∈ W01,2 ⊗ (S)1,p and i = 1, 2, . . . , n.

3.2 A Dirichlet Problem with Stochastic Coefficients 111

Proof. From the definition of the Wick–adjoint multiplication we have

hDi (f ~ g), vi = −hf ~ g, Di vi = −hg, f ♦Di vi
= −hg, Di (f ♦v) − (Di f )♦vi = hDi g, f ♦vi + hg, (Di f )♦vi
= hf ~ (Di g), vi + h(Di f ~ )gi.
where we used the chain rule for Wick multiplication (Lema 3.2.3). ¤

Later we will see that all properties of the operator L (ellipticity and
other assumptions on its coefficients) can be carried over to the formal Wick–
adjoint of L.

Assumptions on L. The concept of generalized weak solutions.

The assumptions on L must also be modified to be compatible with the
Wick calculus. Thus, we imply following assumptions on L: There exists
λ > 0 such that for all u, v1 , v2 , . . . vn ∈ W0 S and all v ≥ 0 ∈ W0 S following
conditions hold:
n X
X n n
X
ij
ha ♦vj , vi i ≥ λ kvi k2 (ellipticity), (3.44)
i=1 j=1 i=1

aij , bi , ci , d, (i, j, = 1, 2, . . . , n) are essentially bounded GRPs , (3.45)

n
X
hd♦u, vi − hbi ♦u, Di vi ≤ 0. (3.46)
i=1
Conditions (3.45) and (3.46) will be adequate to prove existence and
uniqueness of the generalized weak solution of (3.47). Later on we will imply
a stronger condition to (3.45) and a weaker condition to (3.46) in order to
prove stability and regularity properties.
We turn now back to our stochastic Dirichlet problem
n
X
L♦u(x, ω) = h(x, ω) + Di f i (x, ω), x ∈ I, ω ∈ Ω,
i=1 (3.47)
u(x, ω) ¹∂I = g(x, ω)
where the action L♦u is defined as in (3.42). Applying partial integration
we obtain
X n Xn n
X
ij i
hL♦u, vi = hDi ( a ♦Dj u + b ♦u), vi + hci ♦Di u, vi + hd♦u, vi
i=1 j=1 i=1
112 Applications to Singular SPDEs

n X
X n n
X n
X
ij i
=− ha ♦Dj u, Di vi − hb ♦u, Di vi + hci ♦Di u, vi + hd♦u, vi.
i=1 j=1 i=1 i=1

for u ∈ WS∗ , v ∈ W0 S.
As in [GT], we associate a bilinear form B : WS∗ × W0 S → R associated
with L, defined by
B(u, v) = −hL♦u, vi
n X
X n n
X n
X
ij i
= ha ♦Dj u, Di vi + hb ♦u, Di vi − hci ♦Di u, vi − hd♦u, vi.
i=1 j=1 i=1 i=1
∗
We
Pncall u i∈ WS a generalized
Pn weak solution of (3.47) if hL♦u, vi =
i
hh + i=1 Di f , vi = hh, vi − i=1 hf , Di vi for all v ∈ W0 S, i.e. if
n
X
B(u, v) = −hh, vi + hf i , Di vi
i=1

for all v ∈ W0 S.

Lemma 3.2.5 The bilinear form B(·, ·) is continuous.

Proof. Since the Wick product is distributive with respect to addition, B

is indeed bilinear. Continuity (boundedness) follows from (3.45) and the
Cauchy–Schwartz inequality:
n X
X n n
X
ij i
|B(u, v)| ≤ |ha ♦Dj u + b ♦u, Di vi| + |hci ♦Di u + d♦u, v|i
i=1 j=1 i=1

n X
X n Xn Xn
≤ kaij kkDj ukkDi vk+ kbi kkukkDi vk+ kci kkDi ukkvk+kdkkukkvk
i=1 j=1 i=1 i=1
à n !à n
!
X X
≤C kDi uk kDi vk = CkukWS∗ kvkW0 S ,
i=0 i=0

where C = max1≤i,j≤n {ka k, kb k, kci k, kdk}. Thus, B is indeed a bilinear

ij i

continuous mapping B : WS∗ × W0 S → R. ¤

Now, similarly as in the deterministic case, we can identify the operator

L♦ with its unique extension L♦ : WS → WS∗ defined via the bilinear form
B. Existence of a weak generalized solution is equivalent to surjectivity, while
uniqueness is equivalent to injectivity of the mapping L♦.
3.2 A Dirichlet Problem with Stochastic Coefficients 113

3.2.3 The stochastic weak maximum principle

First recall that vP∈ W0 S is called positive, denoted by v ≥ 0, if it has
expansion v(x, ω) = α∈I vα (x) ⊗ Hα (ω), vα ∈ W01,2 , and vα (x) ≥ 0 for all
x ∈ I, α ∈ I. An element u ∈ WS∗ is positive in weak sense, denoted by
u ≥ 0 if hu,Pvi ≥ 0 for all v ∈ W0 S, v ≥ 0. Note that if u ∈ WS has expansion
u(x, ω) = α∈I uα (x)⊗Hα (ω), uα ∈ W 1,2 , and uα (x) ≥ 0 for all x ∈ I, α ∈ I,
then u is also positive in weak senseP (this follows
R from the fact that for each
v ∈ W0 S, v ≥ 0, we have hu, vi = α∈I α! I uα (x)vα (x)dx ≥ 0).
Now we introduce some terminology convention. In a weak sense, u ∈
WS∗ is said to satisfy L♦u ≥ 0 in I, if B(u, v) ≤ 0 for all v ≥ 0. Respectively,
L♦u ≤ 0, if B(u, v) ≥ 0 forPall v ≥ 0.
P For +u ∈ WS, u(x, ω) P = α∈I uα (x)⊗Hα (ω), uα ∈ W 1,2 , define u+ (x, ω) =
α∈I uα (x) ⊗ Hα (ω) = α∈I max{uα (x), 0} ⊗ Hα (ω). Let u(x, ω) ≤ 0 on ∂I
+ +
if u (x, ω) ∈ W0 S i.e. if uα (x) ∈P W01,2 for all α ∈ I.
∗
Also, for u ∈ WS , u(x, ω) = α∈I uα (x)⊗Hα (ω), uα ∈ W −1,2 , we define:
X
Sup u(x, ω) = sup uα (x)Hα (ω)
x∈I x∈I
α∈I
X
= inf{k ∈ R|∀x ∈ I, uα (x) ≤ k}Hα (ω)
α∈I
X
= kα Hα (ω) = K(ω) ∈ (S)−1 .
α∈I
P
Clearly, kα ≤ kuα kW 1,2 , which implies α∈I kα2 (2N)−pα < ∞ for some p > 0
and thus Supx∈I u(x, ω) ∈ (S)−1 . Let Inf x∈I u(x, ω) = − Supx∈I (−u(x, ω)).
In a similar manner on can define Supx∈∂I u(x, ω) and Inf x∈∂I u(x, ω).
Note that Inf and Sup are only notations, they do not mean a classical
infimum or supremum, since for fixed x ∈ I, u(x, ·) is an element of (S)−1
which has no partial ordering. S
We use the same procedure to define Spt u(x, ω) = α∈I supp uα (x), and
note that it is not a support in classical sense since it is not necessarily a
closed set.
Theorem 3.2.1 (i) Let u ∈ WS∗ satisfy L♦u ≥ 0 in I. Then
Sup u(x, ω) ≤ Sup u+ (x, ω)
x∈I x∈∂I

(i.e. Supx∈∂I u+ (x, ω) − Supx∈I u(x, ω) is positive in weak sense).

(ii) Let u ∈ WS∗ satisfy L♦u ≤ 0 in I. Then
Inf u(x, ω) ≥ Inf u− (x, ω).
x∈I x∈∂I
114 Applications to Singular SPDEs

Proof. (i) Let B(u, v) ≤ 0. Then, using (3.46) and boundedness of the
coefficients ci we obtain
Xn X n n
X
ij
ha ♦Dj u, Di vi ≤ hci ♦Di u, vi
i=1 j=1 i=1
n
(3.48)
X
≤C kDi uk kvk
i=1

for some constant C > 0.

If ci = 0, i = 1, 2, . . . , n, then put
v(x, ω) = max{u(x, ω) − Sup u+ (x, ω), 0}
x∈∂I
X
= max{uα (x) − sup u+α (x), 0} ⊗ Hα (ω)
x∈∂I
α∈I

and note that v ≥ 0 (since vα (x) = max{uα (x) − supx∈∂I u+α (x), 0} ≥ 0 for
each α ∈ I), and Di v = Di u, i = 1, 2, . . . , n. Now from (3.48) and the
ellipticity condition (3.44) we retain that
n
X n X
X n
2
λ kDi vk ≤ haij ♦Dj v, Di vi ≤ 0,
i=1 i=1 j=1

i.e. kDi vk = 0, for all i = 1, 2, . . . , n. Thus, v(x, ω) is constant in the x

variable, i.e. v(x, ω) = V (ω) ∈ (S)−1 . Now using the Poincaré inequality
kvk ≤ KkDvk we obtain kvk = 0 i.e. vα (x) = 0 for a.e. x ∈ I and all α ∈ I.
Thus, for every α ∈ I we have supx∈I uα (x) − supx∈∂I u+ α (x) ≤ 0. From this
follows that we have also in weak sense Supx∈I u(x, ω)−Supx∈∂I u+ (x, ω) ≤ 0.
. , n such that ci 6= 0, we follow a similar idea. As-
If there exists i = 1, 2, . .P
+
sume, there exists K(ω) = α∈I kα Hα (ω) ∈ (S)−1 Psuch that supx∈∂I uα (x) ≤
kα < supx∈I uα (x), α ∈ I, and put v(x, ω) = α∈I vα (x) ⊗ Hα (ω), where
vα (x) = max{uα (x) − kα , 0}, α ∈ I.
Now for each α ∈ I we have Di vα = Di uα for uα > kα (i.e. for vα 6= 0) and
Di vα = 0 for uα ≤ kα (i.e. for vα = 0). Now from (3.48) and the ellipticity
condition we retain that
Xn X n X n Xn
2 ij
λ kDi vk ≤ ha ♦Dj v, Di vi ≤ C kDi vk kvk,
i=1 i=1 j=1 i=1
Pn
and consequently i=1 kDi vk ≤ 2C λ
kvk. Now
n
à ! 12 à ! 21
X X 2C X
α!kDi vα k2L2 (I) ≤ α!kvα k2L2 (I)
i=1
λ
α∈I α∈I
3.2 A Dirichlet Problem with Stochastic Coefficients 115

implies
n
X
kDi vα kL2 (I) ≤ C̃kvα kL2 (I) , α ∈ I, (3.49)
i=1

for an appropriate C̃ > 0. But now, using Sobolev type inequalities one can
prove that (3.49) implies that supp Dvα is a set of strictly positive measure.
This is a contradiction with Duα = 0. Thus, supx∈I uα ≤ supx∈∂I u+ α for all
α ∈ I. This proves the assertion. ¤

Remark. Instead of condition (3.46) we could have used (similarly as it

is done in [GT]) the following equivalent condition:
n
X
hd♦u, vi + hci ♦u, Di vi ≤ 0. (3.50)
i=1

Theorem 3.2.1 remains valid also if we replace (3.46) by a weaker condition

n
X
hd, vi − hbi , Di vi ≤ 0 (3.51)
i=1

and at the same time replace (3.45) by a stronger condition: Let there exist
p, Λ, ν > 0 such that for all x ∈ I,
n
X
kaij (x, ·)k2−1,−p ≤ Λ2 and
i,j=1

n
1 X¡ i 2 i 2
¢ 1
2
kb (x, ·)k −1,−p + kc (x, ·)k −1,−p + kd(x, ·)k−1,−p ≤ ν 2 . (3.52)
λ i=1 λ

The uniqueness of the generalized weak solution of the homogeneous

Dirichlet problem now follows directly from the maximum principle:

Corollary 3.2.1 Let u ∈ WS∗ satisfy L♦u(x, ω) = 0 in I × Ω. Then u = 0.

3.2.4 Solvability of the stochastic Dirichlet problem

First we note that it suffices to solve the Dirichlet problem (3.47) for zero
boundary values. Namely, for ũ(x, ω) = u(x, ω)−g(x, ω) we have by linearity
116 Applications to Singular SPDEs

of the L♦ operator:

L♦ũ = L♦u − L♦g

n
à n n n
!
X X X X
=h+ Di f i − Di ( aij ♦Dj g + bi ♦g) + ci ♦Di g + d♦g
i=1 i=1 j=1 i=1
n n
à n
!
X X X
=h− ci ♦Di g − d♦g + Di fi − aij ♦Dj g − bi ♦g
i=1 i=1 j=1
Xn
= h̃ + Di f˜i ,
i=1

P P
where h̃ = h − ni=1 ci ♦Di g − d♦g and f˜i = f i − nj=1 aij ♦Dj g − bi ♦g,
i = 1, 2, . . . , n. Clearly, ũ ¹∂I = 0. Thus, any stochastic Dirichlet problem of
the form (3.47) can be reduced to the zero boundary condition. Moreover, if
h, f i ∈ L2 (I) ⊗ (S)−1 and g ∈ W 1,2 (I) ⊗ (S)−1 , then h̃, f˜i ∈ L2 (I) ⊗ (S)−1
and ũ ∈ W01,2 (I) ⊗ (S)−1 .
The following two lemmas state that the bilinear form B associated with
the operator L can be made coercive by adding a sufficiently large multiple
of the P identity operator to it. (Recall, Du denotes the total differential
Du = ni=1 Di u.)

Lemma 3.2.6 Let L satisfy conditions (3.44) and (3.45). There exist con-
stants K1 , K2 > 0 such that

B(u, u) ≥ K1 kDuk2 − K2 kuk2 . (3.53)

Pn
Proof. Clearly, kDuk2 ≤ 2 i=1 kDi uk2 . Now, using the assumptions on L
we get
n
X n
X
2
λ kDi uk ≤ haij ♦Dj u, Di ui
i=1 i,j=1
n
X n
X
i
= B(u, u) − hb ♦u, Di ui + hci ♦Di u, ui + hd♦u, ui
i=1 i=1
Xn
≤ B(u, u) + hci ♦Di u, ui
i=1
n
X
≤ B(u, u) + kci khDi u, ui ≤ B(u, u) + ChDu, ui,
i=1
(3.54)
3.2 A Dirichlet Problem with Stochastic Coefficients 117

where C = max{kci k, i = 1, 2, . . . , n}. Now by the Schwartz inequality,

1
hDu, ui ≤ εkDuk2 + kuk2 .
4ε
Choose ε such that εC ≤ λ4 . Then from (3.54) we get
λ C
B(u, u) ≥ kDuk2 − kuk2 . ¤
4 4ε
Lemma 3.2.7 There exists σ > 0 such that the operator

Lσ ♦u = L♦u − σu

has a coercive bilinear form Bσ associated with it.

Proof. Using the result of the previous lemma we have

Bσ (u, u) = −hLσ ♦u, ui = −hL♦u, ui + σhu, ui = B(u, u) + σkuk2

¡ ¢
≥ K1 kDuk2 − K2 kuk2 + σkuk2 ≥ K3 kDuk2 + kuk2 = K3 kuk2WS∗
for σ > K2 . Thus, Bσ is coercive on WS∗ . ¤

Lemma 3.2.8 The embedding I : WS → WS∗ defined by

³ ´
u 7→ v 7→ hI(u), vi

is compact.

Proof. From [GT] we know that the embedding W 1,2 (I) → L2 (I) is com-
pact. From [HØUZ] we have that the embeddings (S)1,p → (L)2 are of
Hilbert-Schmidt type for all p > 0. Thus, the embedding of their pro-
jective limit (S)1 → (L)2 is compact. It can be extended to a com-
pact embedding (S)1 → (S)−1 . From these we get that the embedding
W 1,2 (I) ⊗ (S)1 → L2 (I) ⊗ (S)−1 is also compact. ¤

Now we are ready to state the main theorem about the existence and
uniqueness of generalized weak solutions of the stochastic Dirichlet problem.

Theorem 3.2.2 Let the operator L satisfy conditions (3.44), (3.45) and
(3.46). Then for h, f i ∈ L2 (I) ⊗ (S)−1 , i = 1, 2, . . . , n and for g ∈
W 1,2 (I) ⊗ (S)−1 the stochastic Dirichlet problem (3.47) has a unique gen-
eralized weak solution u ∈ WS∗ .
118 Applications to Singular SPDEs

Proof. Choose σ > 0 such that the bilinear form Bσ is coercive (this can
be done due to Lemma 3.2.7). From Lemma 3.2.5 we know that Bσ is also
continuous. According to the Lax–Milgram theorem, for Lσ there exists the
inverse operator
(Lσ ♦)−1 : WS∗ → WS
and it is a continuous, injective mapping.
The mapping F : W0 S → R, v 7→ hF, vi = hf i , Di vi − hh, vi is linear and
continuous, i.e. F ∈ WS∗ .
The equation L♦u = F , u ∈ W0 S, F ∈ WS∗ is equivalent to Lσ ♦u+σIu =
F , i.e. to
u − (−σ(Lσ ♦)−1 I)u = (Lσ ♦)−1 F.
Due to Lemma 3.2.8 the operator T = −σ(Lσ ♦)−1 I is compact. Thus, by
the Fredholm alternative,

(i) either the equation u − T u = 0 has a nontrivial solution u 6= 0,

(ii) or the equation u − T u = (Lσ ♦)−1 F has a unique solution u.

Case (i) is impossible, since the homogeneous equation has a unique trivial
solution according to Theorem 3.2.1. Thus, case (ii) must hold true. ¤

Similarly as in Corollary 3.1.1, we know that the generalized expectation

of the solution u can be obtained as the weak solution of the deterministic
Dirichlet problem, of the form (3.47) where all the stochastic processes (co-
efficients of L, data and boundary value) are replaced by their generalized
expectations.

Corollary 3.2.2 Let u ∈ WS∗ be the generalized solution of (3.47). Then

its generalized expectation E(u) coincides with the weak solution of the de-
terministic Dirichlet problem
n
X
L̃ v(x) = h̃(x) + Di f˜i (x), x ∈ I,
i=1 (3.55)
v(x) ¹∂I = g̃(x),

where h̃ = E(h), g̃ = E(g), f˜i = E(f i ), i = 1, 2, . . . , n and

n
X Xn n
X
ij i
L̃ v = Di ( E(a )Dj v + E(b )v) + E(ci )Di v + E(d)v. (3.56)
i=1 j=1 i=1
3.2 A Dirichlet Problem with Stochastic Coefficients 119

Proof. The assertion follows from the construction of a generalized weak

solution if we choose a test function v ∈ W0 S of the form v(x, ω) = w(x)θ(ω)
and then put θ = 1. ¤

Remark. Since the stochastic Dirichlet problem has a unique solution,

it follows from the Fredholm alternative theorem, that the operator L−1 :
WS∗ → WS is a bounded linear operator on WS∗ . Consequently, we have
the following apriori estimate: Let u be the generalized weak solution of
(3.47). Then there exists p > 0 and a constant C > 0 depending only on L
and I such that
¡ ¢
kukW 1,2 ⊗(S)−1,−p ≤ C khkL2 ⊗(S)−1,−p + kgkW 1,2 ⊗(S)−1,−p (3.57)

where h = (h, f 1 , f 2 , . . . , f n ).

3.2.5 The Wick–adjoint of L

Define the formal Wick–adjoint of L, denoted by L~ , as
n
X Xn n
X
~ ji~ i~
L u= Di ( a Dj u − c u) − bi~ Di u + d~ u, (3.58)
i=1 j=1 i=1

for u ∈ W0 S, where aji~ , bi~ , ci~ and d~ are the Wick–adjoint multiplication
operators (recall Definition 3.2.3) of the coefficients aji , bi , ci and d.

Lemma 3.2.9 For arbitrary u ∈ WS∗ and v ∈ W0 S

hu, L~ vi = hL♦u, vi,

i.e. L~ : W0 S → W0 S is the Hilbert space adjoint of L♦ : WS∗ → WS∗ .

Proof.
n
X Xn n
X
~ ji~ i~
hu, L vi = hu, Di ( a Dj v − c v)i − hu, bi~ Di vi + hu, d~ vi
i=1 j=1 i=1
n
X Xn X n
=− hDi u, aji~ Dj v − ci~ vi − hu, bi~ Di vi + hu, d~ vi
i=1 j=1 i=1
Xn n
X n
X n
X
ji i
=− h a ♦Di u, Dj vi + hc ♦Di u, vi − hbi ♦u, Di vi + hd♦u, vi
i=1 j=1 i=1 i=1

= hL♦u, vi. ¤
120 Applications to Singular SPDEs

It is a routine calculation to show that L~ echoes properties (3.44), (3.45)

and (3.46) of the operator L in following form:
n X
X n n
X
haji~ vi , vj i ≥ λ kvi k2 (3.59)
i=1 j=1 i=1

aij~ , bi~ , ci~ , d~ , (i, j, = 1, 2, . . . , n) are bounded operators on WS∗ ,

(3.60)
n
X
hd~ u, vi + hci~ u, Di vi ≤ 0, v ≥ 0. (3.61)
i=1

Let us check (3.59). From the ellipticity property of L we get

n X
X n n X
X n n
X
ji~ ji
ha vi , vj i = hvi , a ♦vj i ≥ λ kvi k2 .
i=1 j=1 i=1 j=1 i=1

Concerning (3.60), let us prove e.g. that d~ is bounded. For u, v ∈ WS

we have |hd~ u, vi| = |hu, d♦vi| ≤ kdk kuk kvk. P Thus, kd~ k ≤ kdk. And
finally, (3.61) can also be reduced to hu, d♦vi + ni=1 hu, ci ♦Di vi ≤ 0. But
as already stated in (3.50), this is also a sufficient condition for the weak
maximum principle.
Consequently, Theorem 3.2.2 can be applied to get a unique solution of
the Dirichlet problem L~ u = h.
The following theorem describes the spectral behavior of L and is a con-
sequence of the previous considerations and the Fredholm alternative.

Theorem 3.2.3 Let the operator L satisfy conditions (3.44) and (3.45).
There exists a countable, discrete set Σ ⊂ R with following properties:
P
(i) If σ P∈/ Σ, the Dirichlet problems Lσ ♦u = h + ni=1 Di f i and L~ σu =
h + ni=1 Di f i , u ¹∂I = g, are uniquely solvable in WS for arbitrary
h, f i ∈ L2 (I) ⊗ (S)−1 and g ∈ WS. Moreover, there exists C > 0
(depending on L, I and σ) such that kuk ≤ C(khk + kgk).

(ii) If σ ∈ Σ, then the subspaces of solutions of the homogeneous problems

Lσ ♦u = 0, L~ σ u = 0, P
u ¹∂I = 0, are of positive, finite dimension and the
problem Lσ ♦u P = h + ni=1 Di f i , u ¹∂I = g is solvable in WS if and only
if hLσ ♦g − h − ni=1 Di f i , vi = 0 for all v satisfying L~
σ v = 0, v ¹∂I = 0.

Moreover, if condition (3.46) holds, then Σ ⊂ (−∞, 0).

3.2 A Dirichlet Problem with Stochastic Coefficients 121

3.2.6 Stability and regularity properties

Concerning stability properties of the generalized weak solution of (3.47)
we can prove the same results as we did for the deterministic-coefficients
case in the previous section: The generalized weak solution is continuously
dependent on the data and the coefficients of L.

Theorem 3.2.4 Let L be an operator of the form (3.42) with coefficients

aij , bi , ci , d, satisfying conditions (3.44), (3.45) and (3.46). Let L̃ be an-
˜ satisfying all
other operator of the form (3.42) with coefficients ãij , b̃i , c̃i , d,
i ˜i
given conditions. Let h, h̃, f and f , i = 1, 2, . . . , n be generalized random
processes from L2 (I) ⊗ (S)−1 . Let g and g̃ be generalized random processes
from WS. Let u, ũ ∈ WS∗ be the generalized weak solutions of the Dirichlet
problems
Xn
L♦u = h + Di f i , u ¹∂I = g, (3.62)
i=1

n
X
L̃♦ũ = h̃ + Di f˜i , ũ ¹∂I = g̃, (3.63)
i=1

respectively. There exist C > 0, p ∈ N0 such that for every v ∈ W01,2 (I) ⊗
(S)1,p following estimate holds:
³
|hu−ũ, vi| ≤ C kh−h̃kL2 (I)⊗(S)−1,−p +kL−L̃kW −1,2 ⊗(S)−1,−p ku−gkW 1,2 ⊗(S)−1,−p

+kL − L̃kW −1,2 ⊗(S)−1,−p kg̃kW 1,2 ⊗(S)−1,−p

´
+kg − g̃kW 1,2 ⊗(S)−1,−p kLkW −1,2 ⊗(S)−1,−p kvkW 1,2 ⊗(S)1,p

where h = (h, f 1 , f 2 , . . . f n ).

Proof. To keep technicalities to a minimum, we consider the Dirichlet prob-

lems with zero boundary conditions. Let v ∈ W0 S be arbitrary. According
to Theorem 3.2.3 there exists a unique solution w ∈ W0 S of the equation
L̃~ w = v, w ¹∂I = 0, and there exists K > 0 such that kwk ≤ Kkvk. From
(3.62) and (3.63) we obtain
n
X
L̃♦(u − ũ) = h − h̃ + (f i − f˜i ) − (L − L̃)♦u.
i=1
122 Applications to Singular SPDEs

Thus, using the Cauchy–Schwartz inequality, continuity of h, h̃, f , f˜ and

continuity of L and L̃ (Lemma 3.2.2) we get

|hu − ũ, vi| = hu − ũ, L̃~ wi|

= |hL̃♦(u − ũ), wi|
Xn
≤ |hh − h̃, wi| + |hf i − f˜i , wi| + |h(L − L̃)♦u, wi|
i=1
à n
!
X
≤ kh − h̃k + kf i − f˜i k + kL − L̃k kuk kwk
i=1
à n
!
X
≤K kh − h̃k + kf i − f˜i k + kL − L̃k kuk kvk
i=1

˜
Note that kL − L̃k = max1≤i,j≤n {kaij − ãij k, kbi − b̃i k, kci − c̃i k, kd − dk}.
Since v is arbitrary, we finally obtain
³ ´
ku − ũk ≤ K kh − h̃k + kL − L̃kkuk . ¤

In particular, let us consider again a net of operators L² , ² ∈ (0, 1], given

by
n
X Xn n
X
L² ♦u = Di ( aij
² ♦D j u + b i
² ♦u) + ci² ♦Di u + d² ♦u
i=1 j=1 i=1

and a net of data h² , f²i , i = 1, 2, . . . , n, where aij ij

² (x, ω) = a (·, ω) ∗ ρ̌² (x),
i i i i
b² (x, ω) = b (·, ω) ∗ ρ̌² (x), c² (x, ω) = c (·, ω) ∗ ρ̌² (x), d² (x, ω) = d(·, ω) ∗
ρ̌² (x),h² (x, ω) = h(·, ω) ∗ ρ̌² (x), f²i (x, ω) = f i (·, ω) ∗ ρ̌² (x), i, j = 1, 2, . . . , n.
Denote by u² the solution of
n
X
L² ♦u² = h² + Di f²i , u² ¹∂I = 0.
i=1

Now, from Theorem 3.2.4 we get that ku² − ukWS∗ is bounded by the sum of
the operator norm kL − L² kWS∗ and of kh² − hkL2 (I)⊗(S)−1 . Thus,

ku² − ukWS∗ → 0, ² → 0.

This leads us to consider problem (3.47) in the Colombeau setting, just

as we did in the previous section with deterministic coefficients. In Theorem
3.2.8 we will prove existence and uniqueness of generalized weak solutions
of (3.47) when the coefficients of the operator L are Colombeau generalized
3.2 A Dirichlet Problem with Stochastic Coefficients 123

processes i.e. elements of the algebra G(W 2,2 ; (S)−1 ). But in order to do this,
first we have to establish some necessary polynomial growth rate estimates
for the generalized weak solution u of (3.47).
In order to get these polynomial growth rates we need stricter assump-
tions on L: Assume that beside the ellipticity condition (3.44), conditions
(3.51) and (3.52) hold. By careful investigation of the proofs in [GT], we
see that one may carry over the regularity properties also in our setting to
the Hilbert space WS∗ . One only needs to consider (S)−1 –valued Sobolev
functions u(x, ω) ∈ W 2,2 (I) ⊗ (S)−1 , but this is no problem since the chain
rule and all other necessary tools hold (one can define differential quotients
and carry out the calculations as in [GT]). We state now our analogies of
[GT, Theorem 8.8.] and [GT, Theorem 8.12.] without proof.

Definition 3.2.4 Let F ∈P WS∗ be a generalized random process given by

chaos expansion F (x, ω) = α∈I fα (x) ⊗ Hα (ω). We will call F a uniformly
Lipschitz continuous generalized random process, if fα ∈ C 0,1 (I) for all α ∈ I
and if there exists p > 0 such that
X
kfα k2C 0,1 (I) (2N)−pα < ∞.
α∈I

We may consider uniformly Lipschitz continuous GRPs as elements of the

tensor product space C 0,1 (I) ⊗ (S)−1 .

Theorem 3.2.5 Let the operator L satisfy conditions (3.44),(3.51) and

(3.52). Let u ∈ W 1,2 (I) ⊗ (S)−1 be the generalized weak solution of

L♦u = h, u ¹∂I = g.

Assume the coefficients aij , bi , i, j = 1, 2, . . . n, are uniformly Lipschitz

continuous GRPs, ci , d, i = 1, 2, . . . n are essentially bounded GRPs, and
h ∈ L2 (I) ⊗ (S)−1 . Then, for arbitrary I 0 such that I 0 ⊂ I, it follows that
u ∈ W 2,2 (I 0 ) ⊗ (S)−1 and there exist p > 0 and C(n, λ, K, d0 ) > 0 such that

kukW 2,2 (I 0 )⊗(S)−1,−p ≤ C(kukW 1,2 (I)⊗(S)−1,−p + khkL2 (I)⊗(S)−1,−p )

where K = max1≤i,j≤n {kaij , bi kC 0,1 (I)⊗(S)−1,−p , kci , dkL∞ (I)⊗(S)−1,−p } and d0 =

dist(∂I, I 0 ). Additionally, u satisfies the equation
n
X n X
X n n
X
ij ij i i
L♦u = a ♦Dij u + ( (Dj a + b + c )♦Di u + ( Di bi + d)♦u = h
i,j=1 i=1 j=1 i=1

for a.e. x ∈ I and ω ∈ Ω.

124 Applications to Singular SPDEs

Theorem 3.2.6 Assume in addition to the hypothesis of Theorem 3.2.5 that

∂I is of C 2 –class and that there exists g ∈ W 2,2 (I) ⊗ (S)−1 such that u − g ∈
W01,2 (I) ⊗ (S)−1 . Then u ∈ W 2,2 (I) ⊗ (S)−1 and there exist p > 0 and
C(n, ∂I) > 0 such that
K³
kukW 2,2 (I)⊗(S)−1,−p ≤ C 2 kukL2 (I)⊗(S)−1,−p + khkL2 (I)⊗(S)−1,−p
λ ´ (3.64)
+ kgkW 2,2 (I)⊗(S)−1,−p .

The deterministic case of the following theorem was proved in [MP] as an

improvement of [GT, Theorem 8.16.]. By careful investigation of the proof
in [MP], we see that one may carry it over also to WS∗ . Since lack of space,
we will give only a sketch of the proof.

Theorem 3.2.7 Suppose the operator L satisfies (3.44),(3.51) and (3.52).

Assume that f i ∈ Lq (I) ⊗ (S)−1 , i = 1, 2, . . . , n, and h ∈ Lq/2
P(I) ⊗ (S)−1 for
some q > n. If u ∈ W 1,2 (I) ⊗ (S)−1 satisfies L♦u − (h + ni=1 f i ) ≥ 0 in
weak sense, then there exist p > 0, s > 0 and C(n, q, ν, |I|) > 0 such that

sup ku(x, ·)k−1,−p ≤ sup ku+ (x, ·)k−1,−p + Υ

x∈I x∈∂I
³ n
X ´2 (3.65)
−s i
+ Cλ Υ+ kb + ci k2L2 (I)⊗(S)−1,−p +1 ,
i=1
P
where Υ = λ−1 ( ni=1 kf i kLq (I)⊗(S)−1,−p + kgkLq/2 (I)⊗(S)−1,−p ).
P
Proof. From B(u, v) ≤ hh + ni=1 f i , vi, v ≥ 0, v ∈ W0 S, using the chain
rule and (3.51) we get
n
X n
X n
X
haij ♦Dj u, Di i ≤ h(bi + ci )♦Di u, vi + hf i , Di vi − hh, vi,
i,j=1 i=1 i=1

for v ∈ W0 S, u ∈ WS∗ such that u~ v ≥ 0. Let M = supx∈I ku+ (x, ·)k−1,−p .

Define
ku+ (x, ·)k−1,−p
ϕ(x) = p ∈ W01,2 (I)
+
2M + Υ − ku (x, ·)k−1,−p
and consider the test function v(x, ω) = ϕ(x)θ(ω), for arbitrary θ ∈ (S)1,p .
Now we can apply the ellipticity condition etc. and proceed by the same
pattern as in [MP, Theorem 6.] to prove the assertion. ¤

Now it is easy to get an estimate as in Proposition 3.1.3 for the Dirichlet

problem L♦u = h, u ¹∂I = g. By the Sobolev lemma we have W 2,2 (I) ⊂ C β (I)
3.2 A Dirichlet Problem with Stochastic Coefficients 125

for n ≤ 3 and β < 12 . Thus, kukL2 (I)⊗(S)−1,−p ≤ supx∈I ku(x, ·)k−1,−p |I|. Com-
bining the results of Theorem 3.2.6 and Theorem 3.2.7 we get the existence
of s > 0 and C > 0 such that
³ Λ ´s ³
kukW (I)⊗(S)−1,−p ≤ C
2,2 sup ku(x, ·)k−1,−p |I| + kgkW 2,2 (I)⊗(S)−1,−p
λ x∈I
´
+ khkL2 (I)⊗(S)−1,−p + khkLq (I)⊗(S)−1,−p .
(3.66)

3.2.7 Colombeau–solutions of the Dirichlet problem

We assume that I is of C 2 –class and that n ≤ 3. Clearly, then Theorem
3.2.6 holds for q = 4, but as we already assumed in Theorem 3.2.2 that
h ∈ L2 (I) ⊗ (S)−1 , we always have condition h ∈ Lq/2 satisfied. Thus, the
last term in 3.66 vanishes.
We consider again a net of linear differential operators
n
X n
X
L² ♦u² (x, ω) = Di ( aij i
² ♦Dj u² (x, ω) + b² ♦u² (x, ω))
i=1 j=1
n
X
+ ci² ♦Di u² (x, ω) + d² ♦u² (x, ω), ² ∈ (0, 1),
i=1
(3.67)

where the nets of coefficients aij i i

² , b² , c² , d² belong to EM (W
2,2
; (S)−1 ). Define
that two nets of operators are related: L² ∼ L̃² if and only if (L² − L̃² )♦u² ∈
N(W 0,2 ; (S)−1 ) for all u² ∈ EM (W 2,2 ; (S)−1 ). Clearly, ∼ is an equivalence
relation, and following holds:

Proposition 3.2.1 If u² ∈ N(W 2,2 ; (S)−1 ), then L² ♦u² ∈ N(W 0,2 ; (S)−1 ).

Denote by L the family of all nets of differential operators of the form

(3.67) and let L0 = L/ ∼. For L ∈ L0 we define L : G(W 2,2 ; (S)−1 ) →
G(W 0,2 ; (S)−1 ) by
n
X n
X
L♦[u² (x, ·)] = Di ( [aij i
² (x, ω)]♦[Dj u² (x, ω)] + [b² (x, ω)]♦[u² (x, ω))]
i=1 j=1
n
X
+ [ci² (x, ω)]♦[Di u² (x, ω)] + [d² (x, ω)]♦[u² (x, ω)].
i=1
(3.68)
126 Applications to Singular SPDEs

The operator L = [L² ] given by (3.68) is strictly elliptic, if there exist

representatives aij i i
² , b² , c² , d² ∈ EM (W
2,2
; (S)−1 ) of its coefficients, such that
n
X n
X n
X
haij
² ♦vj (x, ω), vi (x, ω)i ≥ λ² 2
kvi k ≥ K² a
kvi k2 , (3.69)
i,j=1 i=1 i=1

for all vi ∈ W0 S, i = 1, 2, . . . , n, and K is a constant independent of ².

The boundary condition is defined in the same way as in Section 3.1.4.
Consider now the stochastic Dirichlet problem

L♦u(x, ω) = h(x, ω) in G(W 2,2 ; (S)−1 )

(3.70)
u(x, ω) ¹∂I = g(x, ω),

where L is defined in (3.68). In order to solve (3.70) in the Colombeau

setting, one has to solve a family of problems:

L² ♦u² (x, ω) = h² (x, ω) in W 2,2 (I) ⊗ (S)−1

(3.71)
u² (x, ω) ¹∂I = g² (x, ω), ² ∈ (0, 1),

then to check whether the net of solutions u² belongs to EM (W 2,2 ; (S)−1 ),

and finally to check whether equation (3.71) holds with other representatives
of L, h, f , g and u. Uniqueness of a solution means that if u = [u² ] and
v = [v² ] satisfy (3.70), then [u² ] = [v² ].

Theorem 3.2.8 Let h and g belong to G(W 2,2 ; (S)−1 ). Let the operator L =
[L² ] be given by (3.68) with coefficients [aij i i
² ], [b² ], [c² ], [d² ] in G(W
2,2
; (S)−1 ).
Assume that following conditions hold:
1. L is strictly elliptic, i.e. (3.69) holds,

2. there exist M > 0 and b > 0 such that for all ² ∈ (0, 1):

kaij i i
² k, kb² k, kc² k, kd² k, kh² k, kg² k ≤ Λ² ≤ M ²
b

where k · k denotes the norm in W 2,2 ⊗ (S)−1,−p for some p > 0 fixed.

3. for every ² ∈ (0, 1) and v ≥ 0, v ∈ W0 S,

n
X
hd² , vi − hbi² , Di vi ≤ 0.
i=1

Then the stochastic Dirichlet problem (3.36) has a unique solution in

G(W 2,2 ; (S)−1 ).
3.2 A Dirichlet Problem with Stochastic Coefficients 127

Proof. For ² ∈ (0, 1) fixed, there exists a unique generalized weak solution
u² ∈ W 2,2 (I) ⊗ (S)−1 of problem (3.71). This follows from Theorem 3.2.2
and Theorem 3.2.6.
Using the estimate derived in (3.66) we obtain that there exist C > 0,
s > 0 (independent of ²) such that:
µ ¶s ³
Λ²
ku² kW 2,2 ⊗(S)−1,−p ≤ C sup kg² (x, ·)k−1,−p |I| + kg² kW 2,2 ⊗(S)−1,−p
λ² x∈∂I
´
+ kh² kL2 ⊗(S)−1,−p .
(3.72)

Now, since g² and h² , are all bounded by Λ² , which has polynomial growth
with respect to ², we obtain

ku² kW 2,2 ⊗(S)−1,−p ≤ C̃²a ,

for an appropriate a ∈ R, C̃ > 0.

Thus, by Proposition 3.1.4 (ii), [u² (x, ω)] ∈ G(W 2,2 ; (S)−1 ).
The definition of operators L0 implies that with some other representa-
tives of h, g and the coefficients of L in (3.71), another representative of
u = [u² ] also satisfies (3.70).
Now, we prove uniqueness of the solution: Let u1,² and u2,² satisfy

L1,² ♦u² (x, ω) = h1,² (x, ω) in W 2,2 (I) ⊗ (S)−1

(3.73)
u² (x, ω) ¹∂I = g1,² (x, ω), ² ∈ (0, 1),

and

L2,² ♦u² (x, ω) = h2,² (x, ω) in W 2,2 (I) ⊗ (S)−1

(3.74)
u² (x, ω) ¹∂I = g2,² (x, ω), ² ∈ (0, 1),

respectively, where the operators Lk,² , k = 1, 2, are of the form (3.67), if we

replace the coefficients with aij i i
k,² , bk,² , ck,² , dk,² , k = 1, 2, respectively, and
(aij ij i i i i
1,² − a2,² ), (b1,² − b2,² ), (c1,² − c2,² ), (d1,² − d2,² ), (h1,² − h2,² ) and (g1,² − g2,² ) ∈
2,2
N(W ; (S)−1 ) holds.
We will prove (u1,² − u2,² ) ∈ N(W 2,2 ; (S)−1 ), which establishes the asser-
tion [u1,² ] = [u2,² ].
If we insert u1,² into (3.73) and u2,² into (3.74), then subtract (3.73) from
(3.74), we obtain
128 Applications to Singular SPDEs

n ³X
X n ´
(aij
1,² ♦Dj (u2,² − u1,² ) + bi1,² ♦(u2,² − u1,² )
i=1 j=1
Xn
+ ci1,² ♦Di (u2,² − u1,² ) + d1,² ♦(u2,² − u1,² )
i=1
³X
n n
¡X ¢
= h2,² − h1,² − (aij ij i i
2,² − a1,² )♦Dj u1,² + (b2,² − b1,² )♦u1,²
i=1 j=1
n
X ´
+ (ci2,² − ci1,² )♦Di u1,² + (d2,² − d1,² )♦u1,²
i=1
= H²
(u2,² − u1,² ) ¹∂I = (g2,² − g1,² ), ² ∈ (0, 1),
where H² ∈ N(W 2,2 ; (S)−1 ). Using the estimate (3.66) we finally obtain
that there exist p > 0 and C > 0 such that for all a ∈ R
ku2,² − u1,² kW 2,2 ⊗(S)−1,−p ≤ C²a .
Thus, by Proposition 3.1.4 (iv), (u2,² − u1,² ) ∈ N(W 2,2 ; (S)−1 ). ¤

The Hilbert space valued stochastic Dirichlet problem

Let H be a separable Hilbert space with orthonormal base {ei : i ∈ N}.
Consider now a Dirichlet problem of the form (3.47) where the coefficients
a, b, c, d and f , g, h are Hilbert space valued GRPs i.e. for fixed x ∈ I
they take value in S(H)−1 . The operator L is interpreted as in (3.42), the
Wick product now taken in S(H)−1 . With largely cosmetic changes, one can
carry out all calculations also in this setting; just interpret h·, ·i as the dual
pairing and k · k asPthe norm in L2 (I) ⊗ S(H)−1,−p for fixed p > 0. For
∞ P
example, u(x, ω) = i=1 α∈I ui,α (x) ⊗ Hα (ω)ei ∈ W −1,2 (I) ⊗ S(H)−1 and
P∞ P 1,2
v(x,
P∞ P ω) = R α∈I vi,α (x) ⊗ Hα (ω)ei ∈ W0 (I) ⊗ S(H)−1 act as hu, vi =
i=1
i=1 α∈I α! I ui,α (x)vi,α (x)dx.
For example we state the H–valued version of Theorem 3.2.2
Theorem 3.2.9 Let the operator L satisfy conditions (3.44), (3.45) and
(3.46). Then for h, f i ∈ L2 (I) ⊗ S(H)−1 , i = 1, 2, . . . , n and for g ∈
W 1,2 (I) ⊗ S(H)−1 the stochastic Dirichlet problem (3.47) has a unique gen-
eralized weak solution in W 1,2 (I) ⊗ S(H)−1 .
Analogue H–valued versions of all the theorems concerning regularity, sta-
bility properties and Colombeau solutions hold.
3.3 Applications of the Fourier Transformation to GRPs (I) 129

3.3 Applications of the Fourier

Transformation to Generalized
Random Processes of type (I)
In this section we define the Fourier transformation for GRPs (I) and
present its application to solve some SPDEs involving singular generalized
stochastic processes. In particular, we solve the stochastic version of the
Helmholtz equation.

The Fourier transformation of tempered distributions

The Fourier transformation S(R) 3 f → fˆ ∈ S(R) is defined by
Z
ˆ 1
f (y) = F(f )(y) = √ f (x)e−ixy dx, y ∈ R, (3.75)
2π R
and Rthe inverse Fourier transformation is given by the formula F−1 (f )(x) =
√1 f (y)eixy dy. The Fourier transformation of a tempered distribution
2π R
F ∈ S0 (R) is defined by the action
hF̂ , ϕi = hF, ϕ̂i, ϕ ∈ S(R).
Clearly, F̂ ∈ S0 (R). The inverse Fourier transformation is defined analo-
gously by hF−1 (F ), ϕi = hF, F−1 (ϕ)i. It is easy to check that the Fourier
transformation of the Hermite functions is given by
ξˆn (y) = (−i)n−1 ξn (y), n ∈ N. (3.76)
Thus, by linearity of the Fourier transformation
P∞ we obtain that if F ∈ S0 (R)
has the formal expansion F (x) = k=1 ak ξk (x), then
∞
X
F̂ (y) = (−i)k−1 ak ξk (y). (3.77)
k=1
P∞ 2 −p
Moreover, if
P k=1 |ak | (2k) < ∞ for some p ≥ 0, then also
∞ k−1 2 2 −p
k=1 |(−i) | |ak | (2k) < ∞. Thus, the Fourier transformation maps
S−p (R) into itself and we have kF k−p = kF̂ k−p . The same considerations
show that the Fourier transformation maps exp S0 (R) into itself.
In S(Rn ) and S0 (Rn ) the R Fourier transformation is defined in a similar
manner by fˆ(y) = (2π)1n/2 Rn f (x)e−ixy dx, where xy is interpreted as the
standard inner product in Rn . For example, in S0 (Rn ) the Dirac delta distri-
bution δ has Fourier transformation δ̂ = (2π)−n/2 .
For convenience of the reader we list some crucial formulae related to the
Fourier transformation. For further properties refer e.g. to [PS].
130 Applications to Singular SPDEs

d
∂ 1
• f (y) = yj fˆ(y), j = 1, 2, . . . n,
∂xj i

d 1 d∂
• x j f (y) = f (y), j = 1, 2, . . . n,
i ∂yj

∗ g(y) = (2π)n/2 fˆ(y)ĝ(y),

• f[ (exchange formula)

• fcg(y) = (2π)−n/2 fˆ(y) ∗ ĝ(y),

\
• f (x − h)(y) = e−ihy fˆ(y), h ∈ Rn , (translation formula)
y
• f\
(λx)(y) = |λ|−n fˆ( ), λ ∈ C. (dilatation formula)
λ
In the Hilbert space valued case, the Fourier transformation on S0 (Rn ; H)
is defined (see [Tr]) by

F ⊗ Id : S0 (Rn ) ⊗ H → S0 (Rn ) ⊗ H

where F is the Fourier transformation on S0 (Rn ) and Id is the identity map-

ping of H.

3.3.1 The Fourier transformation of GRPs (I)

The considerations given above allow us to define the Fourier transforma-
tion for GRPs (I) defined on the Schwartz space of tempered distributions.
Thus, we restrict our attention to GRPs considered as elements of the spaces

L(S(Rn ), (S)−1 ), L(exp S(Rn ), (S)−1 )

and L(S(Rn ), S(H)−1 ), L(exp S(Rn ), S(H)−1 ), respectively in the H–valued

case.
P
Definition 3.3.1 Let Φ be a GRP (I) given by expansion Φ = ∞ j=1 fj ⊗Hαj ,
n
P∞ 2 −pαj
fj ∈ S−k (R ), j = 1, 2, . . . such that j=1 kfj k−k (2N) < ∞ for some
p ≥ 0. The Fourier transformation of Φ, denoted by F(Φ) is defined by the
expansion
∞
X
F(Φ) = fbj ⊗ Hαj , (3.78)
j=1

where fbj is the Fourier transformation of fj , j = 1, 2, . . . in S0 (Rn ).

3.3 Applications of the Fourier Transformation to GRPs (I) 131

Since for each j ∈ N, we have kfbj k−k = kfj kk and consequently

P∞ b 2 −pαj
j=1 kfj k−k (2N) < ∞, the Fourier transformation is well defined i.e.
F(Φ) is also a GRP (I). Moreover, we obtain that F maps L(Sk (Rn ), (S)−1,−p )
into itself.
Note that in [HØUZ] there is a Fourier transformation defined for gener-
alized random processes, but it acts in the ω ∈ Ω variable, while our Fourier
transformation acts in the space variable x.
In the Hilbert space valued case, the definition is similar:

Definition
P∞ 3.3.2
P∞ Let Φ be a H–valued GRP (I) given by the expansion
n
Φ = f ij ⊗ H j
α i e , fij ∈ S −k (R ), i, j = 1, 2, . . . such that
P∞ P∞i=1 j=1 2 −pαj
i=1 j=1 kfij k−k (2N) < ∞ for some p ≥ 0. The Fourier transfor-
mation of Φ, denoted by F(Φ) is the unique H–valued GRP (I) defined by
the expansion
X ∞ X
∞
F(Φ) = fcij ⊗ Hαj ei . (3.79)
i=1 j=1

A simple application
Let A ∈ (S)−1 , δ ∈ S0 (R) be the Dirac delta distirbution, and consider
the SDE
d2
u(x, ω) − u(x, ω) = δ 0 (x) ⊗ A(ω), x ∈ R, ω ∈ Ω. (3.80)
dx2

Proposition 3.3.1 Equation (3.80) has a solution of the form

1
u(x, ω) = sgn(x)e−|x| ⊗ A(ω).
2
P∞
Proof. Let the chaos expansion of A be A(ω) P∞ = j=1 aj Hαj (ω), ω ∈ Ω. We
seek for the solution in form of u(x, ω) = j=1 uj (x) ⊗ Hαj (ω). Now (3.80)
P ³ d2 ´ P
obtains the form ∞ j=1 dx 2 u j (x) − u j (x) ⊗Hαj (ω) = δ 0 (x)⊗ ∞ j=1 aj Hαj (ω),
from which we get the system of ODEs

d2
2
uj (x) − uj (x) = aj δ 0 (x), j ∈ N.
dx
Applying the Fourier transformation we get
1
−y 2 ûj (y) − ûj (y) = aj iy √ ,
2π
132 Applications to Singular SPDEs

i.e.
−iyaj
ûj (y) = √ .
2π(1 + y 2 )
Now we apply the inverse Fourier transformation to get
1
uj (x) = aj sgn(x)e−|x| , j ∈ N,
2
where sgn ∈ S0 (R) is understood in distributional sense. ¤

We can check that this solution is bona fide, also using the results from
Example 2.2.3. The function
½ aj −x
e , x>0
uj (x) = 2
aj x
−2e , x<0

has a jump height aj at point x = 0. Thus, its distributional derivative is

½ aj −x
d −2e , x>0
uj (x) = a + aj δ(x)
dx − 2j ex , x < 0

and ½ aj −x
d2 e , x>0
uj (x) = 2
aj x + aj δ 0 (x) = uj (x) + aj δ 0 (x).
dx2 −2e , x<0

3.3.2 The one dimensional stochastic

Helmholtz equation
Let t1 ≤ t2 ≤ t3 · · · → ∞ and δtj denote the Dirac delta distribution in
tj i.e. δtj (x) = δ(x − tj ), j ∈ N. The formal sum
∞
X
∆(x, ω) = δtj (x) ⊗ Hαj (ω) (3.81)
j=1

defines a GRP (I). This process was introduced in [Se], where it was shown
5
that it is an element of L(Sk (R), (S)−1 ) for k > 12 .
Consider now the one dimensional stochastic Helmholtz equation

d2
Φ(x, ω) + k 2 Φ(x, ω) = ∆(x, ω), x ∈ R, ω ∈ Ω, (3.82)
dx2
where k is a constant.
The (deterministic) Helmholtz equation is closely related to the wave
2 d2
equation: Considering the wave equation dtd 2 u(x, t) − c2 dx 2 u(x, t) = 0 and
3.3 Applications of the Fourier Transformation to GRPs (I) 133

applying a separation of variables u(x, t) = e−jvt X(x) one gets that X(x)
d2 2 v
satisfies ( dx 2 + k )X(x) = 0, k = c . Thus, the solution of the Helmholtz

equation represents the spatial part of solution of the wave equation. If

there is a source f (x)e−jvt (e.g. some source producing acoustic waves), then
d2 2
it appears as the right hand side of the Helmholtz equation ( dx 2 + k )X(x) =

f (x). Thus, one can interpret f as the wave source.

Because of its relationship to the wave equation, the Helmholtz equa-
tion arises in the study of electromagnetic radiation, seismology, acoustics,
thermal and mechanical wave propagation etc.
The constant k, known as the ”wave number”, is the quotient of the
angular frequency v and the wave velocity c. E.g. for electromagnetical waves
in homogeneous conducting media, k is a function of magnetic permeability
and electric conductivity. The right hand side of the Helmholtz equation
describes the wave source; thus, the Dirac delta distribution δ on the right
hand side is a model for waves propagating from a point source. If we now
consider the right hand side of the Helmholtz equation to be the stochastic
process ∆, then (3.82) is a model for waves propagating from point sources,
which are randomly appearing (their presence or lack is due to some random
impulses).
It is also known that any elliptic equation with constant coefficients can be
reduced to the Helmholtz equation. Thus, the stochastic Helmholtz equation
is just a special case of the elliptic problems considered in Section 3.1, but
here we present a new solving technique involving the Fourier transformation.
This will provide an explicit form of the solution (the Hilbert space methods
in Section 3.1 guarantee existence of a solution but no explicit form). For
technical simplicity we consider the one dimensional stochastic Helmholtz
equation, but it is easy to carry over the results to ∆Φ(x, ω) + k 2 Φ(x, ω) =
∆(x, ω), x ∈ Rn , ω ∈ Ω, where ∆ is the n–dimensional Laplace operator.
Proposition 3.3.2 The stochastic Helmholtz equation (3.82) has a solution
in L(exp S(R), (S)−1 ).
P
Proof. We seek for the solution in form of Φ(x, ω) = ∞ j=1 fj (x) ⊗ Hαj (ω),
where the coefficients fj , j ∈ N, are to be determined. Thus, (3.82) is
equivalent to
X ∞ µ 2 ¶ X∞
d 2
fj (x) + k fj (x) ⊗ Hαj (ω) = δtj (x) ⊗ Hαj (ω),
j=1
dx2 j=1

from which we obtain the system of equations

d2
2
fj (x) + k 2 fj (x) = δtj (x), j ∈ N.
dx
134 Applications to Singular SPDEs

Applying the Fourier transformation we obtain −y 2 fˆj (y) + k 2 fˆj (y) =

\
δ(x \
− tj )(y). Since δ(x − tj )(y) = e−iytj δ̂(y) = e−iytj √12π , we obtain that
1 1
fbj (y) = √ e−iytj , j ∈ N.
2π k − y 2
2

we note that fbj ¶∈ exp S0 (R), since it has expansion fbj (y) =
First µ
P∞ − 1 t2 √
e 2 j 4π n
n=1 k2 −y 2
i hn (tj ) ξn (y), where hn are the Hermite polynomials, ξn
−t2j
√ P
are the Hermite functions, and (ke 2 −y2π)2 ∞ 2 2n −2k
n=1 |hn (tj )| (e ) < ∞ for k large
enough. Thus, we can apply the inverse Fourier transformation in exp S0 (R)
to obtain
1
fj (x) = sin(k(x − tj )) sgn(x − tj ),
2k
where sgn is understood again in distributional sense as sgn(x) = 2H(x) − 1,
and H is the Heaviside function.
Thus, the solution of the Helmholtz equation (3.82) is given by
∞
1 X
Φ(x, ω) = sin(k(x − tj )) sgn(x − tj ) ⊗ Hαj (ω), x ∈ R, ω ∈ Ω. (3.83)
2k j=1

It remains to prove that (3.83) is a well-defined GRP (I) i.e. the series
converges in L(exp S(R), (S)−1 ).
Denote by an,j = P hsin(k(x − tj )), ξn i the nth coefficient in the expan-
sion sin(k(x − tj )) = ∞ n=1 an,j ξn (x). By elementary calculus we get a1,j =
1 k2 √ 1 − k2
2 √ k2
−√ 4 πe
− 2 2π sin(ktj ), a2,j = √ 4 πe k 2π cos(ktj ), a3,j = 12 e− 2 (2k 2 −
√ k2
1) 4 π sin(ktj ), ..., an,j = Pn−1 (k)e− 2 f (ktj ), where Pn−1 is some polynomial
of order n − 1 and f is either the sine or the cosine function. By boundedness
of f we get
∞
X ∞
X
¡ ¢
2n −2l −k2
k sin(k(x − tj ))k2−l,exp = |an,j | 2
e =e P2n−2 (k)e−4nl < ∞
n=1 n=1

for l > 0 large enough. Moreover, k sin(k(x − tj ))k2−l,exp does not depend on
j ∈ N. Denote A = k sin(k(x − tj ))k2−l,exp . Thus,
∞
X ∞
X
j j
k sin(k(x − tj )) sgn(x − tj )k2−l,exp (2N)−pα =A (2N)−pα < ∞,
j=1 j=1

for p > 1. ¤
3.3 Applications of the Fourier Transformation to GRPs (I) 135

By considering only the zeroth terms in the expansion of the GRPs in-
volved in the Helmholtz equation we obtain following consequence of Propo-
sition 3.3.2.
Corollary 3.3.1 Let Φ be the solution of (3.82). Its generalized expectation
E(Φ) is the solution of the deterministic Helmholtz equation
d2
u(x) + k 2 u(x) = E(∆(x, ·)), x ∈ R,
dx2
where E(∆) is the generalized expectation of ∆.
Note that a similar consideration can be carried out for the Hilbert space
valued case, if we consider the H–valued GRP (I) defined in Example 2.2.1
(ii) as the right hand side of equation (3.82).

3.3.3 The Helmholtz equation with

stochastic wave number
Now we consider the Helmholtz equation, and assume the wave number
k is also random. In physical interpretation this describes waves propagating
with a random speed from randomly appearing P point sources.
Assume K ∈ (S)−1 has expansion K(ω) = ∞ j=1 kj Hαj (ω), and kj ≥ 0
for all j ∈ N. Consider the Helmholtz SDE
d2
Φ(x, ω) + K(ω)♦Φ(x, ω) = ∆(x, ω), x ∈ R, ω ∈ Ω, (3.84)
dx2
where ∆ is the GRP defined in (3.81).
Proposition 3.3.3 The stochastic Helmholtz equation (3.84) has a solution
Φ in L(exp S(R), (S)−1 ). Moreover, its generalized expectation E(Φ) is the
solution of the deterministic Helmholtz equation
d2
u(x) + E(K(·))u(x) = E(∆(x, ·)), x ∈ R,
dx2
where E(K) and E(∆) are the generalized expectations of K and ∆ respec-
tively.
P
Proof. We seek for the solution in form of Φ(x, ω) = ∞ j=1 fj (x) ⊗ Hαj (ω),
where the coefficients fj , j ∈ N, are to be determined. Thus, (3.84) is
equivalent to the system
d2 X
f n (x) + ki fj (x) = δtn (x), n ∈ N. (3.85)
dx2 i,j∈N
i+j=n+1
136 Applications to Singular SPDEs

For n = 1 we obtain (i = j = 1) the equation f100 + k1 f1 = δt1 which has

the solution
1 p
f1 (x) = √ sin( k1 (x − t1 )) sgn(x − t1 ), x ∈ R.
k1

For n = 2 we get (i = 1, j = 2 and i = 2, j = 1) the equation f200 +

k1 f2 + k2 f1 = δt2 .³ This we solve using´ the Fourier transformation to obtain
fˆ2 (y) = √12π k1 −y
1
2 e−iyt2 − k1k−y
2
2e
−iyt1
and now we apply the inverse Fourier
transformation to obtain
1 p
f2 (x) = √ sin( k1 (x − t2 )) sgn(x − t2 )
2 k1
k2 ³ p p ´
− p 3 sin( k1 (x − t1 )) − k1 (x − t1 ) cos( k1 (x − t1 )) sgn(x − t1 ).
4 k1

For n = 3 we get (i = 1, j = 3 and i = 3, j = 1 and i = j = 2)

the equation f300 + k1 f3 + k3 f1 + k2 f2 = δt3 . Since f1 and f2 are already
known, we ˆ
³ apply the Fourier and its inverse ´ transformation to get f3 (y) =
√1 1
2π k1 −y 2
e−iyt3 − k1k−y
3
2e
−iyt1
− k2k−y
2
2e
−iyt2
and

1 p
f3 (x) = √ sin( k1 (x − t3 )) sgn(x − t3 )
2 k1
k3 ³ p p p ´
− p 3 sin( k1 (x − t1 )) + k1 (t1 − x) cos( k1 (x − t1 )) sgn(x − t1 )
4 k1
k2 sgn(x − t2 ) ³p p p p ´
− √ k1 sin( k2 (x − t2 )) − k2 sin( k1 (x − t2 )) .
2(k1 − k2 ) k1 k2

We proceed by the same procedure to calculate the coefficients

f4 (x), f5 (x), . . . etc. By boundedness of the sine and cosine function, we
obtain that |fn (x)| is bounded by a polynomial of order not greater than n.
Thus, convergence in L(exp S(R), (S)−1 ) follows by customary arguments,
analogously as in Proposition 3.3.2. ¤
Epilogue

White noise theory offers one of the most compelling instances of infinite
dimensional analysis, leading to a clear understanding of many empirical
phenomena. As a contribution to this noble theory, in Chapter 2 of the dis-
sertation, fundamental theorems were obtained characterizing the structure
of generalized stochastic processes. To begin the harvest of consequences, a
few examples have found place in Chapter 3 to illustrate the applications to
singular SPDEs. Further applications to SPDEs serving more concrete de-
mands, and the modeling of probabilistical properties of their solutions (e.g.
such as the martingale or the Markov property) are deferred but certainly
not denied. They remain as enticing possibilities for future investigations.
Bibliography

[Ad] Adams R.A., Sobolev Spaces, Academic Press, 1975.

[AHR1] Albeverio S., Haba Z., Russo F., Trivial solutions for a non–linear
two–space dimensional wave equation perturbed by space–time white
noise, Stochastics Stochastics Rep. 56, No.1–2, pp. 127–160, 1996.

[AHR2] Albeverio S., Haba Z., Russo F., A two–space dimensional heat
equation perturbed by (Gaussian) white noise, Probab. Theory Re-
lat. Fields 121, pp. 319–366, 2001.

[Ar] Arnold L., Stochastic Differential Equations - Theory and Applica-

tions, John Wiley & Sons, 1974.

[Be] Besold P., Solutions to stochastic partial differential equations

as elements of tensor product spaces, PhD Thesis, Universität
Göttingen, 2000.

[Bu] Bulla I., Dirichlet Problem with Stochastic Coefficients, Stochastics

and Dynamics, Vol. 5, No. 4, pp. 555–568, 2005.

[Co1] Colombeau J.F., New Generalized Functions and Multiplication of

Distributions, North Holland, 1983.

[Co2] Colombeau J.F., Multiplication of Distributions, LNM 1532,

Springer, 1992.

[DHPV] Delcroix A., Hasler M., Pilipović S., Valmorin V., Generalized func-
tion algebras as sequence space algebras, arXiv:math.FA/0206039
v1, 2002.

[Ev] Evans L.C., Partial Differential Equations, Grad. Stud. Math., Vol.
19, Amer. Math. Soc., Providence, 1998.

[FW] Floret K., Wloka J., Einführung in die Theorie der lokalkonvexen
Räume, Springer Verlag, 1968.
BIBLIOGRAPHY 139

[GS] Gel’fand I.M., Shilov G.E., Generalized functions, Vol. I, II, Aca-
demic Press, New York and London, 1964.

[GV] Gel’fand I.M., Vilenkin N.Ya., Generalized Functions - Vol.4, Ap-

plications of Harmonic Analysis, Academic Press, 1964.

[GT] Gilbarg D., Trudinger N.S., Elliptic Partial Differential Equations

of Second Order, Springer Verlag, 1998.

[GKOS] Grosser M., Kunzinger M., Oberguggenberger M., Steinbauer R.,

Geometric theory of generalized functions with applications to gen-
eral relativity Mathematics and its Applications, Kluwer Academic
Publishers, 2001.

[GKS] Grothaus M., Kondratiev Y.G., Streit L., Regular Generalized

Functions in Gaussian Analysis, Infinite Dimensional Analysis,
Quantum Probability and Related Topics, Vol.2, No.1, pp 1–25,
1999.

[Ha] Hanš O., Measurability of Extension of Continuous Random Trans-

forms, Amer. Math. Stat., 30, pp 1152 – 1157, 1959.

[Hi] Hida T., Brownian Motion, Springer Verlag, 1980.

[HKPS] Hida T., Kuo H.H., Pothoff J., Streit L., White Noise - An Infinite
Dimensional Calculus, Kluwer Academic Publishers, 1993.

[Hg] Higham N. J., Handbook of Writing for the Mathematical Sciences,

SIAM, 1993.

[HØUZ] Holden H., Øksendal B., Ubøe J., Zhang T., Stochastic Partial
Differential Equations - A Modeling, White Noise Functional Ap-
proach, Birkhäuser, 1996.

[In] Inaba H., Tapley B.T., Generalized Processes: A Theory and the
White Gaussian Process, SIAM J. Control, 13, pp 719–735, 1975.

[It] Itô K., Stationary Random Distributions, Mem. Coll. Sci. Kyoto
Univ., Ser. A, 28, pp 209–223, 1954.

[Ku] Kuo H.H., White Noise Distribution Theory, CRC Press, 1996.

[LØUZ] Lindstrøm T., Øksendal B., Ubøe J., Zhang T., Stability Properties
of Stochastic Partial Differential Equations, Stochastic Analysis and
Applications 13, pp 177–204, 1995.
140 BIBLIOGRAPHY

[Lo] Lojasiewicz S., Sur la valuer et le limite d’une distributions dans

un point, Studia Math., 16, pp 1–36, 1957.

[LPP] Lozanov-Crvenković Z, Perišić D., Pilipović S., Contributions to the

Theory of Generalized Random Processes, Faculty of Science, Novi
Sad, 2002.

[LP] Lozanov-Crvenković Z., Pilipović S., On Some Classes of General-

ized Random Linear Functionals, Journal of Mathematical Analysis
and Applications, Vol.129, No.2, pp 433–442, 1998.

[LP1] Lozanov-Crvenković Z., Pilipović S., Gaussian generalized random

processes on K{Mp }, J. Math. Anal. Appl., 181, pp 755–761, 1994.

[LP2] Lozanov-Crvenković Z., Pilipović S., Representation Theorems for

Gaussian Generalized Random Processes on Ω × C[0, 1] and Ω ×
L1 (G), Math. Japon., 37, pp 897 – 900, 1992.

[MFA] Melnikova I.V., Filinkov A.I., Alshansky M.A., Abstract Stochastic

Equations II. Solutions in Spaces of Abstract Stochastic Distribu-
tions, Journal of Mathematical Sciences, Vol. 116, No. 5, pp 3620–
3656, 2003.

[Me] Meyer P.A., Quantum Probability for Probabilists, Springer Verlag,

1995.

[MP] Mitrović D., Pilipović S., Approximations of linear Dirichlet prob-

lems with singularities, J. Math. Anal. Appl. 313, pp 98–119, 2006.

[NP] Nedeljkov M., Pilipović S., Generalized Function Algebras and

PDEs with Singularities – a survey, Preprint.

[NPR] Nedeljkov M., Pilipović S., Rajter-Ćirić D., Heat equation with sin-
gular potential and singular data, Proc. Roy. Soc. Edinburgh Sect.
A 135, No.4, pp 863–886, 2005.

[NR] Nedeljkov M., Rajter D., Nonlinear stochastic wave equation with
Colombeau generalized stochastic processes, Math. Models Methods
Appl. Sci. 12, No. 5, pp 665–688, 2002.

[Nu] Nualart D., Rozovski B., Weighted Stochastic Sobolev Spaces and
Bilinear SPDEs Driven by Space-Time White Noise, Journal of
Functional Analysis 149, pp 200–225, 1997.
BIBLIOGRAPHY 141

[Mo] Oberguggenberger M., Multiplication of Distributions and Appli-

cations to Partial Differential Equations, Pitman Research Notes
Math. Vol. 259, Longman, Harlow 1992.
[OR1] Oberguggenberger, M., Russo, F., Nonlinear SPDEs: Colombeau
solutions and pathwise limits, Stochastic analysis and related topics,
(VI Geilo, 1996),pp 319–332, Progr. Probab., 42, 1998.
[OR2] Oberguggenberger, M., Russo, F., Singular limiting behavior in
nonlinear stochastic wave equations, Stochastic analysis and math-
ematical physics,pp 87– 99, Progr. Probab., 50, 2001.
[Øk] Øksendal B., Stochastic Differential Equations - An Introduction
with Applications, Springer Verlag, 2000.
[Øks] Øksendal B., Stochastic Partial Differential Equations - A Math-
ematical Connection between Macrocosmos and Microcosmos,
M.Gyllenberg et al.(Eds.): Analysis, Algebra and Computers in
Mathematical Research, pp 365–385, 1994.
[PS] Pilipović S., Stanković B., Prostori distribucija, SANU Ogranak u
Novom Sadu, 2000.
[Pi] Pilipović S., Generalization of Zemanian Spaces of Generalized
Functions Which Have Orthonormal Series Expansions, SIAM J.
Math. Anal., Vol.17, No2, pp 477–484, 1986.
[PS1] Pilipović S., Seleši D., Expansion Theorems for Generalized Random
Processes, Wick Products and Applications to Stochastic Differen-
tial Equations, Infinite Dimensional Analysis, Quantum Probability
and Related Topics, to appear, 2007.
[PS2] Pilipović S., Seleši D., Structure theorems for generalized random
processes, Acta Mathematica Hungarica, to appear, 2007.
[RO] Russo, F., Oberguggenberger, M., White noise driven stochastic
partial differential equations: triviality and non-triviality, Nonlin-
ear theory of generalized functions (Vienna, 1997), pp 315 – 333,
Chapman & Hall/CRC Res. Notes Math., 401, 1999.
[Se] Seleši D., Ekspanzija uopštenih stohastičkih procesa sa primenama
u jednačinama, magistarska teza, Univerzitet u Novom Sadu, 2004.
[Se1] Seleši D., Hilbert Space Valued Generalized Random Processes -
Part I and II, Novi Sad Journal of Mathematics, submitted, 2007.
142 BIBLIOGRAPHY

[Se2] Seleši D., Generalized Solutions of the Stochastic Dirichlet Problem,

Preprint, 2006.

[Ru] Rudin W., Real and Complex Analysis, Tata McGraw-Hill, New
Delhi, 1977.

[SM] Swartz C.H., Myers D.E., Random functionals on K{Mp } Spaces,

Studia Math., 39, pp 233–240, 1971.

[Tr] Treves F., Topological Vector Spaces, Distributions and Kernels,

Academic Press, 1967.

[Ul] Ullrich M., Representation Theorems for Random Schwartz Dis-

tributions, Transactions of the Second Prague Conference on In-
formation Theory, Statistical Decision Function, Random Process,
Prague, pp 661–666, 1959.

[Ze] Zemanian A.H., Generalized Integral Transforms, Nauka, Moskva,

1974.

[Zi] Ziemer W.P., Weakly Differentiable Functions - Sobolev Spaces and

Functions of Bounded Variation, Springer Verlag, 1989.

[Va] Våge G., Hilbert Space Methods Applied to Elliptic Stochastic Par-
tial Differential Equations, Stochastic analysis and related top-
ics, V(Silivri, 1994), pp 281–294, Progr.Probab., 38, Birkhäuser,
Boston, 1996.

[Wa] Walsh J. B., An introduction to stochastic partial differential equa-

tions, Springer LNM, pp 265–437, 1980.
BIBLIOGRAPHY 143
Index

B G
bilinear form, 10, 85, 112, 116 Gaussian
bounded, 85, 112 measure, 10, 20
coercive, 85, 117, 118 process, 21, 60, 63, 68–70
Brownian motion, 21, 22, 25, 26, 28, random variable, 60, 68, 70, 72
29, 57, 74 Gel’fand triple, 9, 12, 16, 20, 23
generalized
C expectation, 22, 86, 90, 106, 118,
chaos expansion, 21, 29, 44, 88, 96, 135
107, 108, 123, 131 function, 7, 10–20, 27–29, 31, 32,
Colombeau algebra, 7, 19–20, 29, 30, 38–40, 42, 44, 45, 57, 83, 95,
53–59, 83, 93, 95–106, 125– 96
128 random process, 27–136
random variable, 20–27, 29, 40,
D 44, 45, 86
Dirichlet problem, 81–128 solution, 87, 90, 92, 94, 104, 112,
distribution, 10–12, 16, 29, 30, 41, 74, 118
83, 96
Dirac delta, 17, 39, 41, 42, 73, 93, H
95, 129, 132, 133 Helmholtz equation, 132–136
point value, 18, 67
tempered, 11, 12, 14, 16, 74, 129
O
E operator
expectation, 62 adjoint, 11, 91, 110
convolution, 59, 93, 96
correlation, 59, 62–64
F differential, 13, 42, 49, 57, 102,
Fourier transformation, 129–136 109, 125
function elliptic, 82, 83, 90, 92, 95, 102
Fourier-Hermite, 21 Hilbert-Schmidt, 12
Heaviside, 17, 134 Laplace, 11, 16, 82, 86, 133
Hermite, 10–12, 26, 38, 74, 93, Wick–adjoint, 110, 111, 119
129 Wick-multiplication, 109
INDEX 145

P
product
tensor, 9–10, 12, 15, 18, 25, 84,
86, 87, 107, 123
Wick, 33, 55, 57, 58, 83, 87, 97,
135

S
space
countably Hilbert, 8, 9, 25
Hida, 7, 22
Hilbert, 8, 9, 11–13, 15, 18, 20,
23, 28, 29, 33, 44, 61, 75, 83–
86, 106, 110, 119, 128, 130,
131, 133
Kondratiev, 7, 21–24, 29, 33, 44,
83, 84, 86
nuclear, 8–12, 14, 15, 18, 23, 25,
28, 37, 40, 45, 47, 53, 55, 57,
59, 87, 98, 107
Schwartz, 7, 9, 11–12, 19, 25, 57,
130
Sobolev, 15–16, 19, 59, 65, 77,
82–84, 86, 87, 90, 93, 115,
123, 124
white noise, 20, 23
Zemanian, 7, 9, 13–15, 29, 33, 45,
53, 59, 86

W
white noise
coordinate process, 75
singular, 22, 25, 26, 29, 38, 39,
49, 57, 74, 93, 94
smoothed, 93, 94
Wick product, 44–49, 106–128
146

Curriculum Vitae
Personal Data
Name and surname: Dora Seleši (in Hungarian: Szöllössy Dóra)
Born on April 22nd 1978 in Zrenjanin, Serbia.

Academic Degrees Awarded

ƒ Master of Sciences (Mathematics), February 2004, Faculty of Science, Novi Sad, GPA:
10,00.
ƒ Bachelor of Sciences (Mathematics), June 2000, Faculty of Science, Novi Sad, GPA: 9,92.

Education
2004 – Present Doctoral studies
2000 – 2004 Master studies (Mathematics, Major: analysis and probability)
1996 – 2000 Undergraduate studies (Mathematics, Major: graduate mathematician)

Study Abroad Experience

February – July 2005, Department of Engineering Mathematics, Geometry and Computer
Science (Institut für Technische Mathematik, Geometrie und Bauinformatik) in Innsbruck,
Austria.

Academic Honors, Awards and Scholarships

ƒ 2005 Ernst Mach scholarship of the ÖAD for study stay in Austria
ƒ 2000 Scholarship of the Royal Norway Embassy
ƒ 2000 University award Mileva Marić-Einstein for best student paper
ƒ 1995 – 2001 Scholar of the Republic Foundation for Development of Scientific and
Artistic Youth

Interests
Language Skills
Hungarian, Serbian, English, German (Grosses Deutsches Sprachdiplom, Goethe Institut, 2005.)

Membership
MENSA Yugoslavia, 1998 – 2005.

Hobbys
Swimming, foreign languages, science fiction.

Work Experience
2004 – Present Assistant, Department of mathematics and informatics, Faculty of
Science, Novi Sad.
Assistant, Chair of Pharmacology, Faculty of Medicine, Novi Sad.
2000 – 2004 Junior Assistant, Department of mathematics and informatics, Faculty of
Science, Novi Sad.
 147

Research
2006 – Present Engaged at the project Methods of functional analysis, ODEs and PDEs
with singularities, No. 144016 financed by the Ministry of science and
environment of Republic Serbia
2002 – 2006 Engaged at the project Functional analysis, ODEs and PDEs with
singularities, No. 1835 financed by the Ministry of science, technology and
development of Republic Serbia

Teaching
2000 – Present Conducting exercises for following courses:
ƒ Probability (for students of Mathematics and Informatics)
ƒ Statistics (for students of Mathematics and Informatics)
ƒ Statistical Modeling (for students of Mathematics)
ƒ Functional Analysis (for students of Informatics)
ƒ Applied Analysis (for students of Informatics)
ƒ Metric and normed spaces (for students of Mathematics)
ƒ Partial Differential Equations (for students of Mathematics)
ƒ Financial Mathematics II (for students of Mathematics)
ƒ Introduction to Calculus (for students of Mathematics)
ƒ Analysis I (for students of Mathematics)
ƒ Mathematics I (for students of Physics)
ƒ Mathematics II (for students of Physics)
ƒ Mathematics (for students of Pharmacology, also one group in English
language)

Scientific Impact
Attended several national and international conferences, seminars, summer and winter schools,
apart from those listed below.

Presentations at Conferences
ƒ D. Seleši, Stochastic Models of Financial Markets, XV Conference on Applied
Mathematics, Zlatibor, 2002.
ƒ D. Seleši, Expansion Theorems for Generalized Random Processes, Wick Products and
Applications to Stochastic Differential Equations, Generalized Functions 2004: Topics
in PDE, Harmonic Analysis and Mathematical Physics, Novi Sad, 2004.

Publication List
ƒ D. Seleši, Expansion of Generalized Stochastic Processes with Applications to
Equations, Master Thesis, University of Novi Sad, 2004.
ƒ S. Pilipović, D. Seleši, Structure Theorems for Generalized Random Processes, Acta
Mathematica Hungarica, 2007, to appear.
ƒ S. Pilipović, D. Seleši, Expansion Theorems for Generalized Random Processes, Wick
Products and Applications to Stochastic Differential Equations, Infinite Dimensional
Analysis, Quantum Probability and Related Topics, 2007, to appear in Vol.10.
ƒ D. Seleši, Hilbert Space Valued Generalized Random Processes, Novi Sad Journal of
Mathematics, submitted,
148
 149

Biografija
Rođena sam 22. aprila 1978. godine u Zrenjaninu, gde sam završila
osnovnu školu i gimnaziju sa prosekom 5,00 i Vukovom diplomom,
kao i nižu muzičku školu. Od 1995. do 2001. bila sam stipendista
Republičke fondacije za razvoj naučnog i umetničkog podmlatka.
Studije matematike sam započela 1996. na Prirodno-matematičkom
fakultetu u Novom Sadu na smeru Diplomirani matematičar.
Diplomirala sam juna 2000. sa prosečnom ocenom 9,92 i nagradom
Mileva Marić-Einstein za najbolji seminarski rad.
Od oktobra 2000. godine sam zaposlena kao asistent-pripravnik, a od
juna 2004. kao asistent na Departmanu za matematiku i informatiku, PMF-a u Novom Sadu.
Tokom proteklih godina držala sam vežbe studentima matematike, informatike i fizike iz
predmeta: Verovatnoća, Statistika, Statističko modeliranje, Funkcionalna analiza, Metrički i
normirani prostori, Parcijalne diferencijalne jednačine, Finansijska matematika II, Uvod u
analizu, Analiza I, Matematika I, Matematika II. Od oktobra 2004. godine zaposlena sam kao
asistent i na katedri za farmaciju Medicinskog fakulteta, gde izvodim vežbe iz predmeta
Matematika, kao i vežbe iz predmeta Mathematics za studijsku grupu na engleskom jeziku.
Postdiplomske studije sam upisala novembra 2000. godine iz oblasti Analiza i verovatnoća. U
toku studija do 2002. položila sam sve ispite sa prosečnom ocenom 10,00. Magistarsku tezu pod
naslovom Ekspanzija uopštenih stohastičkih procesa sa primenama u jednačinama, odbranila
sam 16. februara 2004. na Prirodno-matematičkom fakultetu u Novom Sadu i time stekla zvanje
magistra matematičkih nauka. Od februara do jula 2005. godine boravila sam u Austriji kao
stipendista ÖAD na Institutu za tehničku matematiku, geometriju i informatiku (Institut für
Technische Mathematik, Geometrie und Bauinformatik), univerziteta u Innsbrucku.
U novembru 2005. položila sam sertifikacioni ispit Goethe Instituta Grosses Deutsches
Sprachdiplom za poznavanje nemačkog jezika na nivou maternjeg jezika sa najvišom ocenom.
Tečno govorim i engleski kao strani jezik, a maternji jezik mi je mađarski.
Oblasti mog naučnog interesovanja su stohastički procesi, teorija uopštenih funkcija,
diferencijalne jednačine, finansijska matematika i statistika. Imala sam izlaganja na
konferenciji PRIM XV (Zlatibor, 2002.) i na međunarodnoj konferenciji GF04 (Novi Sad,
2004.). Pored toga učestvovala sam još na više konferencija, seminara i letnjih/zimskih škola.
Autor ili koautor sam tri naučna rada, kao i nekoliko skripti za studente. Od 2002. angažovana
sam na projektu Metode funkcionalne analize, ODJ i PDJ sa singularitetima, koji finansira
Ministarstvo nauke i zaštite životne sredine Republike Srbije.
Oblasti interesovanja su mi pored nauke i naučna fantastika, učenje stranih jezika i plivanje. Od
1998. do 2005. godine bila sam član MENSA Jugoslavije.

Novi Sad, 2007. Dora Seleši

150 Key Words Documentation

UNIVERSITY OF NOVI SAD

FACULTY OF SCIENCE
KEY WORDS DOCUMENTATION

Accession number:
ANO
Identification number:
INO
Document type: Monograph type
DT
Type of record: Printed text
TR
Contents code: PhD dissertation
CC
Author: Dora Seleši, MSc
AU
Mentor: Academic Stevan Pilipović, PhD
MN

Title: Generalized Stochastic Processes in Infinite Dimensional Spaces with

Applications to Singular Stochastic Partial Differential Equations
TI

Language of text: English

LT
Language of abstract: English/Serbian
LA
Country of publication: Serbia
CP
Locality of publication: Vojvodina
LP
Publication year: 2007.
PY
Publisher: Author’s reprint
PU
Key Words Documentation 151

Publication place: Novi Sad, Faculty of Science, Dositeja Obradovića 4

PP
Physical description: 3/155/58/0/0/0/0
(chapters/pages/literature/tables/pictures/graphics/appendices)
PD
Scientific field: Mathematics
SF
Scientific discipline: Analysis and probability
SD
Subject / Key words: Generalized random processes, white noise,
Brownian motion, chaos expansion, Wick product, distributions, Colombeau
algebras, regularization, stochastic differential equations, Sobolev spaces,
Kondratiev spaces, Zemanian spaces, Hilbert spaces, linear elliptic differen-
tial operator, Dirichlet problem, Fourier transformation.
SKW

UC:
Holding data: Library of the Department of Mathematics and Informatics,
Novi Sad
HD
Note:
N
Abstract: Subject of the dissertation are various classes of generalized
stochastic processes and their applications to solving singular stochastic
partial differential equations. Basically, the dissertation can be divided into
two parts. The first part (Chapter 2) is devoted to structural characteri-
zations of generalized random processes in terms of chaos expansions and
integral representations. The second part of the dissertation (Chapter 3)
involves applications of the obtained results to solving a stochastic Dirichlet
problem, where multiplication is modeled by the Wick product, and the
coefficients of the elliptic differential operator are Colombeau generalized
random processes.
AB

Accepted by Scientific Board on: November 3, 2006

ASB
Defended:
DE
Thesis defend board:
152 Key Words Documentation

President: Marko Nedeljkov, PhD, Full Professor, Faculty of Science,

University of Novi Sad
Member: Academic Stevan Pilipović, PhD, Full professor, Faculty of
Science, University of Novi Sad
Member: Michael Oberguggenberger, PhD, Associate Professor, Faculty of
Civil Engineering and Architecture, University of Innsbruck (Austria)
Member: Zagorka Lozanov-Crvenković, PhD, Full Professor, Faculty of
Science, University of Novi Sad
Member: Danijela Rajter-Ćirić, PhD, Assistant Professor, Faculty of
Science, University of Novi Sad
DB
Key Words Documentation 153

UNIVERZITET U NOVOM SADU

PRIRODNO-MATEMATIČKI FAKULTET
KLJUČNA DOKUMENTACIJSKA INFORMACIJA

Redni broj:
RBR
Identifikacioni broj:
IBR
Tip dokumentacije: Monografska dokumentacija
TD
Tip zapisa: Tekstualni štampani materijal
TZ
Vrsta rada: Doktorska disertacija
VR
Autor: mr Dora Seleši
AU
Mentor: Akademik dr Stevan Pilipović
MN

Naslov rada: Uopšteni stohastički procesi u beskonačnodimenzionalnim

prostorima sa primenama na singularne stohastičke parcijalne diferencijalne
jednačine
NR

Jezik publikacije: engleski

JP
Jezik izvoda: engleski/srpski
JI
Zemlja publikovanja: Srbija
ZP
Uže geografsko područje: Vojvodina
UGP
Godina: 2007.
GO
Izdavač: Autorski reprint
IZ
154 Key Words Documentation

Mesto i adresa: Novi Sad, Prirodno-matematički fakultet, Trg Dositeja

Obradovića 4
MA
Fizički opis rada: 3/155/58/0/0/0/0
(broj poglavlja/strana/lit. citata/tabela/slika/grafika/priloga)
FO
Naučna oblast: Matematika
NO
Naučna disciplina: Analiza i verovatnoća
ND
Predmetna odrednica/Ključne reči: Uopšteni stohastički procesi, beli
šum, Braunovo kretanje, haos ekspanzija, Vikov proizvod, distribucije,
Kolomboove algebre, regularizacija, stohastičke diferencijalne jednačine,
prostori Soboljeva, prostori Kondratieva, prostori Zemaniana, Hilbertovi
prostori, linearni eliptički diferencijalni operator, Dirihleov problem, Furi-
jeova transformacija.
PO

UDK:
Čuva se: u biblioteci Departmana za matematiku i informatiku, Novi Sad
ČU
Važna napomena:
VN
Izvod: Doktorska disertacija je posvećena raznim klasama uopštenih
stohastičkih procesa i njihovim primenama na rešavanje singularnih sto-
hastičkih parcijalnih diferencijalnih jednačina. U osnovi, disertacija se
može podeliti na dva dela. Prvi deo disertacije (Glava 2) je posvećen
strukturnoj karakterizaciji uopštenih stohastičkih procesa u vidu haos
ekspanzije i integralne reprezentacije. Drugi deo disertacije (Glava 3) čini
primena dobijenih rezultata na rešavanje stohastičkog Dirihleovog problema
u kojem se množenje modelira Vikovim proizvodom, a koeficijenti eliptičnog
diferencijalnog operatora su Kolomboovi uopšteni stohastički procesi.
IZ

Datum prihvatanja teme od strane NN Veća: 03. 11. 2006.

DP
Datum odbrane:
DO
Key Words Documentation 155

Članovi komisije:
Predsednik: dr Marko Nedeljkov, redovni profesor, Prirodno-matematički
fakultet, Univerzitet u Novom Sadu
Član: Akademik dr Stevan Pilipović, redovni profesor, Prirodno-matematički
fakultet, Univerzitet u Novom Sadu
Član: dr Michael Oberguggenberger, vanredni profesor, Fakultet za
grad̄evinu i arhitekturu, Univerzitet u Insbruku (Austrija)
Član: dr Zagorka Lozanov-Crvenković, redovni profesor, Prirodno-
matematički fakultet, Univerzitet u Novom Sadu
Član: dr Danijela Rajter-Ćirić, docent, Prirodno-matematički fakultet,
Univerzitet u Novom Sadu
KO