The Stochastic Heat Equation Driven by a Gaussian Noise: germ
Markov Property
Raluca Balan
Corresponding author. Department of Mathematics and Statistics,
University of Ottawa, 585, King Edward Avenue, Ottawa, Ontario K1N
6N5 Canada
[email protected]
http://aix1.uottawa.ca/$\sim$rbalan
and
Doyoon Kim
Department of Mathematics,
University of Southern California, Los Angeles, 3620 Vermont Avenue,
KAP108, Los Angeles, CA 90089, USA
[email protected]
http://college.usc.edu/faculty/faculty1017406.html
Abstract.
Let u={u(t,x);t∈[0,T],x∈Rd} be
the process solution of the stochastic heat equation ut=Δu+F˙,u(0,⋅)=0 driven by a Gaussian noise F˙, which
is white in time and has spatial covariance induced by the kernel
f. In this paper we prove that the process u is locally germ
Markov, if f is the Bessel kernel of order α=2k,k∈N+, or f is the Riesz kernel of order α=4k,k∈N+.
Key words and phrases:
Gaussian noise; stochastic integral; stochastic heat
equation; Reproducing Kernel Hilbert Space; germ Markov property
2000 Mathematics Subject Classification:
Primary 60H15; Secondary 60G60
The first author was supported by a grant from the Natural Sciences and
Engineering Research Council of Canada. The second author was
supported by a post-doctoral fellowship at University of Ottawa.
1. Introduction
This article is based on the theory of stochastic partial
differential equations (s.p.d.e.’s) initiated by John Walsh in 1986 [24].
This theory relies on the construction of a stochastic integral with
respect to the so-called “worthy martingale measures” and focuses
mainly on equations driven by a space-time white noise.
Recently, there has been a considerable amount of interest in
s.p.d.e.’s driven by a noise term which is white in time, but is
“colored” in space, in the sense that it has a spatial covariance
structure induced by a kernel f, which is the Fourier transform of
a tempered distribution μ. This line of research was initiated
in [15], [4] for the stochastic wave
equation in the case d=2, and then generalized in [3],
[6] to a larger class of s.p.d.e.’s in arbitrary
spatial dimensions. The new theory requires an extension of Walsh’s
stochastic integral, to include the case when the integrand is a
(Schwartz) distribution; this type of extension is needed for
instance for the stochastic wave equation with d≥3. Under a
certain integrability condition imposed on the measure μ, one
can prove that any second-order s.p.d.e. with constant coefficients
has a process solution, i.e. a solution which can be identified with
a multiparameter stochastic process u. As far as we know, the
literature to date does not contain any study of the germ Markov
property of this process solution. Such a study is of great
importance because it allows us to conclude that the behavior of the
process in any time-space region A is independent of its behavior
outside the region, given the values of the process in a thin area
around the boundary ∂A of the region.
There is an immense amount of literature dedicated to the study of
different types of Markov properties for multiparameter processes
(see [2] and the references
therein). In particular, the germ Markov property of multiparameter
Gaussian processes is an area which received a lot of attention in
the 1970’s, which allows for the use of various tools imported from
Hilbert space analysis, like the method of the Reproducing Kernel
Hilbert Space (RKHS). An excellent all-time reference is the
monograph [20]; earlier references include
[18], [14], [17].
In the 1990’s, the problem of the germ Markov property for
multi-dimensional Gaussian processes appears again in the
literature, this time for processes which arise as solutions of
s.p.d.e.’s driven by a white noise: the case of elliptic equations
was thoroughly treated in [7], while the
systematic study of this problem for the quasi-linear parabolic
equations (in the case d=1) is found in [16].
The case of hyperbolic equations turns out to be the most difficult
one: the only reference here seems to be [5], in
which the authors investigate very carefully the structure of the
germ σ-fields induced by the process solution of the
stochastic wave equation (in the case d=2) driven by a Lévy
noise without Gaussian component. (This type of noise induces a very
particular form for the process solution, which is used in a
fundamental way for deriving its germ Markov property.) Finally, the
more recent work [19] contains a detailed analysis
of the relationship between the germ and sharp σ-fields
(which are used for defining the sharp and germ Markov property,
respectively), in the case of the process solution associated to the
Bessel equation driven by a white noise (in the case d=2).
The goal of the present paper is to investigate the structure of the
RKHS associated to the process solution u of the stochastic heat
equation driven by a spatial covariance kernel f, and to identify
some examples of functions f for which the process u possesses
the germ Markov property. As far as we know, this is the first
attempt to tackle a problem of this type in the literature. Our two
main results (Theorem 4.9 and Theorem
4.15) state that the process solution u has
the germ Markov property, if the function f is either the Bessel
kernel of order α=2k,k∈N or the Riesz kernel of
order α=4k,k∈N. In order to prove these results,
we use some results from the Lp-theory of parabolic equations
with mixed norms, due to [12], [13]. The germ
Markov property of the process solution u in the case of the
Bessel or Riesz kernels with α>0 arbitrary (or in the case of
other kernel functions) remains an open problem.
The paper is organized as follows. In Section 2, we introduce the
framework for the study of s.p.d.e.’s, including the construction of
the extended stochastic integral due to Dalang (as in
[3]). In Section 3, we include all the ingredients which
are necessary to formulate the result about the existence of the
process solution u, and we examine the structure of the Gaussian
space and the RKHS associated to this process. In Section 4, we give
the definition of the germ Markov property and we prove our two main
results, by applying a fundamental result of [14]. The
two appendices contain some technical proofs.
2. The Framework
2.1. Basic Notation
We denote by D(U) the space of all infinitely differentiable
functions on Rn whose support is compact and contained
in the open set U, and by D′(U) the space of distributions, i.e. continuous linear functionals on D(U). We
denote by S(Rn) the Schwartz space of all rapidly
decreasing functions on Rn, and by
S′(Rn) the space of tempered distributions,
i.e. continuous linear functionals on S(Rn). The
space S′(Rn) can be viewed as a subspace of
D′(Rn). An important subspace of
S′(Rn) is the space OC′(Rn) of
all distributions with rapid decrease (see Chapter VII,
section 5, [22]).
For an arbitrary function ϕ on Rn, the
translation by x is denoted with ϕx, i.e.
ϕx(y)=ϕ(x+y), and the reflection by [math] is denoted with
ϕ~, i.e. ϕ~(x)=ϕ(−x). These notions have
obvious extensions to distributions. The Fourier transform of ϕ∈L1(Rn) is defined by Fϕ(ξ)=(2π)−n/2∫Rnexp(−iξ⋅x)ϕ(x)dx.
The map F:S(Rn)→S(Rn)
can be extended to L2(Rn) and
S′(Rn). For every subset U⊂Rn, we
denote
[TABLE]
whenever the integral is defined.
2.2. The Gaussian Noise
Let T>0 and f be a locally integrable function on
Rd. As in [3], for every φ,ψ∈D((0,T)×Rd) we define
[TABLE]
The Gaussian
noise mentioned in the Introduction will be a zero-mean Gaussian
process with covariance Jf. The existence of this process is based on the
following Bochner-Schwartz type result (see Theorem 2, p. 157,
[11]):
Lemma 2.1**.**
The bi-functional Jf is
nonnegative-definite if and only if there exists a tempered measure
μ on Rd such that
[TABLE]
In this case, for every φ,ψ∈D((0,T)×Rd)
[TABLE]
where Fφ(t,⋅)
denotes the Fourier transform of φ(t,⋅).
The basic example of kernel functions is the white noise kernel:
f(x)=δ(x), μ(dξ)=dξ.
More interesting examples of kernel functions f arise when
μ(dξ)=g(∣ξ∣2)dξ, for a certain function g on
(0,∞). In this case, f becomes the inverse Fourier
transform of the tempered distribution ξ↦g(∣ξ∣2)
and we can define the operator g(−Δ):S(Rd)→S′(Rd) by
F[g(−Δ)ϕ](ξ)=g(∣ξ∣2)Fϕ(ξ), or
equivalently
[TABLE]
(see p. 149-152, [10], or p. 117-132, [23]).
Here are some typical examples:
Example 2.2**.**
If μ(dξ)=∣ξ∣−αdξ, then f is
the Riesz kernel of order α:
[TABLE]
where γα is an appropriate constant.
In this case, for every ϕ∈S(Rd) we have
(−Δ)−α/2ϕ=ϕ∗Rα.
Example 2.3**.**
If μ(dξ)=(1+∣ξ∣2)−α/2dξ,
then f is the Bessel kernel of order α:
[TABLE]
where
cα is an appropriate constant. In this case, for every
ϕ∈S(Rd) we have
(1−Δ)−α/2ϕ=ϕ∗Bα.
Example 2.4**.**
If μ(dξ)=e−4π2α∣ξ∣2, then
f is the heat kernel
[TABLE]
In this case, for every
ϕ∈S(Rd) we have
eαΔϕ=ϕ∗Gα.
In what follows, we let F={F(φ);φ∈D((0,T)×Rd)} be a zero-mean Gaussian process with
covariance Jf, i.e. ∀φ,ψ∈D((0,T)×Rd)
[TABLE]
The Gaussian space HF of the process F is defined as the
closed linear subspace of L2(Ω), generated by the
variables {F(φ),φ∈D((0,T)×Rd)}.
2.3. The Stochastic Integral
In this subsection, we summarize the construction of the generalized
stochastic integral M(φ)=∫0T∫Rdφ(t,x)M(dt,dx) with respect to the martingale measure M
induced by the noise F, due to Dalang (see [3]). For
our purposes, it is enough to consider only the case of
deterministic integrands. Here is the construction procedure:
Step 0. For each φ∈D((0,T)×Rd), we set
M(φ)=F(φ)∈HF.
Step 1. Let Bb(Rd) be the class of all bounded Borel subsets
of Rd and E(d) be the class
of all linear combinations of functions 1[0,t]×A, t∈[0,T],A∈Bb(Rd). We endow E(d) with
the inner product ⟨⋅,⋅⟩0 given by formula
(2.1) and we denote by ∥⋅∥0 the
corresponding norm.
For each t∈(0,T),A∈Bb(Rd), there
exists a sequence {φn}n⊂D((0,T)×Rd) such that φn→1[0,t]×A and
suppφn⊆K for all n (see
[4], p. 190).
By the bounded convergence theorem, ∥φn−1[0,t]×A∥+→0, where ∥⋅∥+ is the norm defined in Step
2 below. Hence ∥φn−1[0,t]×A∥0→0 and
E(M(φm)−M(φn))2=∥φm−φn∥0→0 as m,n→∞, i.e. the sequence
{M(φn)}n is Cauchy in L2(Ω). A standard
argument shows that its limit does not depend on
{φn}n. We set Mt(A)=M(1[0,t]×A)=L2(Ω)limnM(φn)∈HF.
(Note that the
process M={Mt(A);t∈[0,T],A∈B(Rd)}
is a worthy martingale measure, in the sense of [24].) We
extend M by linearity to E(d). A limiting argument and
relation (2.2) shows that, for every φ,ψ∈E(d)
[TABLE]
Step 2. Let P+(d)={φ:[0,T]×Rd→R measurable;∥φ∥+<∞}, where ∥φ∥+2:=∫0T∫Rd∫Rd∣φ(t,x)∣f(x−y)∣φ(t,y)∣dydxdt. We endow
P+(d) with the inner product ⟨⋅,⋅⟩0 given by formula (2.1) and we
denote by ∥⋅∥0 the corresponding norm; note that ∥⋅∥0≤∥⋅∥+. An argument similar to that
used in the proof of Proposition 2.3, [24], shows that
E(d) is dense in P+(d) with respect to ∥⋅∥+, and hence for each φ∈P+(d), there
exists a sequence {φn}n⊂E(d) such
that ∥φn−φ∥+→0. As in Step 1, we set M(φ)=L2(Ω)limnM(φn)∈HF. Note that relation (2.3) holds for every
φ,ψ∈P+(d).
Step 3. Let P0(d) be the completion of E(d)
with respect to <⋅,⋅>0. The space P0(d) is
the largest space of integrands φ for which we can define
the stochastic integral M(φ). According to
[3],p. 9, the space P0(d) has the following
alternative definition.
Let P(d)={φ:[0,T]→S′(Rd); Fφ(t,⋅) \mboxfunction∀t∈[0,T], (t,ξ)↦Fφ(t,ξ)\linebreak\mboxmeasurable,∥φ∥0<∞}, where ∥φ∥02=∫0T∫Rd∣Fφ(t,ξ)∣2μ(dξ)dt. Let
E0(d)={φ∈P+(d);φ(t,⋅)∈S(Rd), ∀t∈[0,T]} and
P0(d) be the closure of E0(d) in
P(d) with respect to ∥⋅∥0.
Note that
[TABLE]
For each φ∈P0(d), there exists a
sequence {φn}n⊂E0(d) such that
∥φn−φ∥0→0. As in Step 1,
we set M(φ)=L2(Ω)limnM(φn)∈HF. Note that relation (2.3) holds for every
φ,ψ∈P0(d), and hence
[TABLE]
In summary, the previous construction is based on the diagram
[TABLE]
and the following
3 approximation techniques:
indicator functions can be approximated by functions in
D((0,T)×Rd);
functions in P+(d) can be approximated by
indicator functions;
distributions in P0(d) can be
approximated by “smooth” functions in P+(d).
In particular, we conclude that D((0,T)×Rd) is dense in P0(d) with respect to ∥⋅∥0.
2.4. Alternative Characterization of the Space
P0(d)
As in [9], for every φ,ψ∈D(Rd), define
[TABLE]
and let P0,x(d) be the completion of
D(Rd) with respect to ⟨⋅,⋅⟩0,x.
Note that
[TABLE]
Remark 2.5*.*
Note that for every φ∈P0(d), φ(t,⋅)∈P0,x(d) for a.e. t∈[0,T]. Since
E(d) is dense in P(d), one can prove that the map t↦φ(t,⋅) is strongly measurable from [0,T] to
P0,x(d) (in the sense of Definition on p.649,
[8]). Using (2.5), we conclude that
P0(d)⊂L2((0,T),P0,x(d)) and ∥φ∥0=∥φ∥L2((0,T),P0,x(d)),
∀φ∈P0(d).
3. The Process Solution
3.1. The Equation and its Solution
We consider the stochastic heat equation with vanishing initial
conditions, written formally as:
[TABLE]
The solution of this equation is defined formally as
follows. Suppose for the moment that F˙ is a random variable
with values in S′(Rd+1). For every fixed ω∈Ω,
let {u(φ);φ∈D((0,T)×Rd)}
be the distribution solution of (3.1). It is known that
[TABLE]
where G is the fundamental solution of the heat equation:
[TABLE]
(Since G∈OC′(Rd+1), we have φ∗G~∈S(Rd+1) for every φ∈D(Rd+1) and the convolution G∗F˙ is well
defined; see Theorem XI, Chapter VII, [22].)
Going back to our framework, we have the following lemma.
Lemma 3.1**.**
If
[TABLE]
then: (a) (G_{tx})^{\verb 2~2}\in{\mathcal{P}}_{+}^{(d)} for every (t,x)∈[0,T]×Rd; (b) φ∗G~∈P0(d) for every φ∈D((0,T)×Rd).
Proof.
(a) Set g_{t,x}=(G_{tx})^{\verb 2~2}. Note that gtx is measurable and ∥gtx∥+=∥gtx∥0, since gtx≥0. We have Fgt,x(s,ξ)=cd1t>sexp(−iξx−(t−s)∣ξ∣2), where cd is an appropriate constant. Hence
[TABLE]
where N(T) is a constant depending on T.
(b) We apply Remark 4, [3] to the function ψ=φ∗G~. For this we need to check that: (i) t↦Fψ(t,ξ) is continuous ∀ξ∈Rd; (ii) there
exists a nonnegative function k(t,ξ) which is square-integrable
with respect to dt×μ(dξ) such that ∣Fψ(t,ξ)∣≤k(t,ξ), ∀t∈[0,T],ξ∈Rd.
(i) is clearly satisfied, and (ii) will follow from (3.3)
once we prove that
[TABLE]
where N is a constant. Since (1+∣ξ∣2)∣Fψ(t,x)∣=∣Fϕ(t,ξ)∣≤∥ϕ(t,⋅)∥L1(Rd) where
ϕ=(1−Δ)ψ, relation (3.4) follows if we
prove that ∥ϕ(t,⋅)∥L1(Rd)≤N. Note
that
[TABLE]
since ψ is the unique solution of: −ψt−Δψ=φ in (0,T)×Rd, ψ(T,x)=0.
Thus ϕ(s,y)=∫sT∫RdG(t−s,x−y)(1−Δ)φ(t,x)dxdt. From this we calculate ∥ϕ(s,⋅)∥L1(Rd), which turns out to be bounded
(note that (1−Δ)φ has a compact support in (0,T)×Rd).
∎
A consequence of Lemma 3.1.(b) is the fact that
M(φ∗G~) is well-defined for every
φ∈D((0,T)×Rd). By analogy with
(3.2) (and using a slight abuse of terminology), we
introduce the following definition:
Definition 3.2**.**
The process {u(φ);φ∈D((0,T)×Rd)} defined by
[TABLE]
is called the distribution-valued solution of the stochastic
heat equation (3.1), with vanishing initial conditions.
Theorem 3.3** ([3], [6]).**
Let {u(φ);φ∈D((0,T)×Rd)} be the distribution-valued
solution of the stochastic heat equation (3.1). In order
that there exists a jointly measurable process X={X(t,x);t∈[0,T],x∈Rd} such that
[TABLE]
it is necessary and sufficient that (3.3)
holds. In this case, X is a modification of the process
u={u(t,x);t∈[0,T],x∈Rd} defined by
[TABLE]
Definition 3.4**.**
The process u={u(t,x);t∈[0,T],x∈Rd}
defined by (3.5) is called the process
solution of the stochastic heat equation (3.1), with
vanishing initial conditions.
Note that the process u is a zero-mean Gaussian process.
In the present paper, we are interested in examining the germ Markov
property of the process u.
3.2. The Gaussian space
The Gaussian space Hu of the process u is defined as the
closed linear subspace of L2(Ω), generated by the
variables {u(t,x),t∈[0,T],x∈Rd}.
The next result shows that this space coincides with the space
HF.
Lemma 3.5**.**
We have Hu=HF.
Proof.
a) First, we prove that Hu⊆HF. Recall
that u(t,x)=M((G_{tx})^{\verb 2~2}) and (G_{tx})^{\verb 2~2}\in{\mathcal{P}}_{+}^{(d)}\subset{\mathcal{P}}_{0}^{(d)}, by Lemma
3.1.(a). Since D((0,T)×Rd) is
dense in P0(d) with respect to ∥⋅∥0, there
exists a sequence {φn}n≥1⊆D((0,T)×Rd) such that
\|\varphi_{n}-(G_{tx})^{\verb 2~2}\|_{0}\to 0. Hence
E(M(φn)−u(t,x))2→0. But M(φn)∈HF for all n and therefore u(t,x)∈HF.
b) Let H∗u be the closed linear subspace of L2(Ω),
generated by the variables {u(η),η∈D((0,T)×Rd)}. To prove that HF⊆H∗u, let φ∈D((0,T)×Rd) be arbitrary and
η=−φt−Δφ∈D((0,T)×Rd). Then φ=η∗G~ and F(φ)=M(φ)=M(η∗G~)=u(η)∈H∗u.
c) Finally, we prove that H∗u⊆Hu. Let η∈D((0,T)×Rd) be arbitrary. Note that
u(η)=M(φ) where φ=η∗G~. Let K be a
compact set such that suppη⊂[0,T]×K.
For each n≥1, let {Qm(n),m∈Z} be a partition
of (0,T)×Rd such that each Qm(n)=Rm(n)×Sm(n), where Rm(n) is an interval in [0,T] and
Sm(n) is a cube in Rd. Suppose that maxm∈Z∣Qm(n)∣→0 as n→∞. For each Qm(n), we
choose (tm(n),xm(n))∈Qm(n) such that tm(n)≥t for all t∈Rm(n). We consider the Riemann sum:
[TABLE]
where In={m∈Z;tm(n)>s,xm(n)∈K}.
Clearly φn(s,y)→φ(s,y) for every
(s,y). We claim that
[TABLE]
From here, it follows that
E(M(φn)−M(φ))2→0. This concludes that proof,
since \varphi_{n}=\sum_{m\in I_{n}}a_{m}^{(n)}(G_{t_{m}^{(n)},x_{m}^{(n)}})^{\verb 2~2} and hence
M(φn)=∑m∈Inam(n)u(tm(n),xm(n))∈Hu, where am(n)=∣Qm(n)∣ η(tm(n),xm(n)).
The proof of (3.6) is given in Appendix A.
∎
3.3. The RKHS
Let Hu={hY;hY(t,x)=E(Yu(t,x)),Y∈Hu} be
the RKHS of the process u, endowed with the inner product ⟨hY,hZ⟩Hu:=E(YZ), Y,Z∈Hu. Note
that any function h∈Hu is continuous on [0,T]×Rd and satisfies h(0,⋅)=0. Moreover, Hu is
the closure of {R((t,x),⋅ );(t,x)∈[0,T]×Rd}
with respect to ∥⋅∥Hu, where
R((t,x),(s,y))=E(u(t,x)u(s,y)).
By Lemma 3.5 and (2.4), it follows that
[TABLE]
Moreover, if h(t,x)=E(M(φ)u(t,x)),
g(t,x)=E(M(η)u(t,x)) with φ,η∈P0(d),
then ⟨h,g⟩Hu=⟨φ,η⟩0.
Let g_{t,x}=(G_{tx})^{\verb 2~2}. Using the fact that
u(t,x)=M(gtx) and (2.5), we get:
[TABLE]
Lemma 3.6**.**
Let h(t,x)=E(M(φ)u(t,x)),φ∈P0(d). For η∈D((0,T)×Rd),
let ϕ be a solution to the equation −ϕt−Δϕ=η
in (0,T)×Rd, ϕ(T,⋅)=0. Then
[TABLE]
In particular, for any
ϕ∈D((0,T)×Rd), we have
[TABLE]
Proof.
Using (3.7) and applying Fubini’s theorem (since
φ∈P0(d), gtx∈P+(d), and η
has compact support), we get
[TABLE]
Note that for every r∈(0,T) and y∈Rd
[TABLE]
Hence
[TABLE]
From (3.8) and (3.9), we
conclude that
[TABLE]
(recall that
ϕ∈P0(d) by Lemma 3.1). This finishes the
proof.
∎
4. The Germ Markov Property
4.1. The Definition
Let S⊆[0,T]×Rd be an arbitrary set.
Let FSu be the σ-field generated by the variables
{u(t,x);(t,x)∈S}, KSu be the closed linear subspace
of L2(Ω) generated by the variables {u(t,x);(t,x)∈S}, and KSu be the closed linear subspace of
Hu generated by the functions {R((t,x),⋅ );(t,x)∈S}. Let
[TABLE]
Definition 4.1**.**
The process u={u(t,x);(t,x)∈[0,T]×Rd} is
locally germ Markov if for every relatively compact open set
A⊂[0,T]×Rd, GAu
and GAcu are conditionally independent
given G∂Au, where ∂A=A∩Ac.
Based on the fact that GSu=σ(HSu),
one can prove that u is locally
germ Markov if and only if for every open set A,
Hu=HAu⊕(HAcu⊖H∂Au) (see Lemma 1.3, [14]), or
equivalently, using the isometry between Hu and Hu
[TABLE]
(Here ⊕,⊖ denote the usual operations on Hilbert
subspaces: if H is a Hilbert space, S is a closed subspace and
S⊥ is its orthogonal complement, then we write H=S⊕S⊥ and S⊥=H⊖S.)
On the other hand, h(t,x)=⟨h,R((t,x),⋅ )⟩Hu for every h∈Hu. Hence
[TABLE]
Based on (4.1) and
(4.2), we have the following fundamental result.
Theorem 4.2** (Theorem 5.1, [14]).**
The Gaussian process u is locally germ
Markov if and only if the following two conditions hold:
(i) If h,g∈Hu are such that (supp h)∩(supp g)=∅ and supp h is compact,
then <h,g>Hu=0.
(ii) If ζ=h+g∈Hu, where h,g are such that (supp h)∩(supp g)=∅ and supp h is
compact, then h,g∈Hu.
In the next two subsections, we will suppose that f is the Bessel
kernel of order α>0 (Example 2.3), respectively the
Riesz kernel of order 0<α<d (Example 2.2). In both
cases we must have α>d−2, in order to have condition
(3.3) satisfied. Our goal is to prove that conditions (i)
and (ii) of Theorem 4.2 hold. For this, we will
assume that α=2k,k∈N+ in the case of the the Bessel
kernel, respectively α=4k,k∈N+ in the case of the
Riesz kernel (and d=4k+1, since d−2<α<d).
4.2. The Case of the Bessel Kernel
In this subsection we will assume that f is the Bessel kernel,
i.e. f=Bα with α>max{0,d−2}.
In this case,
P0,x(d)=H2−α/2(Rd), where
[TABLE]
denotes the
fractional Sobolev space of index γ∈R (see e.g.
[10], p.191-192).
By Remark 2.5, P0(d)⊂L2((0,T),H2−α/2(Rd)).
On the other hand, we have the denseness of E0(d) in
L2((0,T),H2−α/2(Rd)) (for a proof, one may see the
proof of Theorem 3.10 in [12]). Thus P0(d)=L2((0,T),H2−α/2(Rd)) and
[TABLE]
For each t∈[0,T], let φ1(t,⋅):=(1−Δ)−α/2φ(t,⋅) and note that the
map φ↦φ1 is an isometry between L2((0,T),H2−α/2(Rd)) and L2((0,T),H2α/2(Rd)).
Since
Fφ1(t,ξ)=(1+∣ξ∣2)−α/2Fφ(t,ξ), it is not difficult to see that: ∀φ∈L2((0,T),H2−α/2(Rd))
[TABLE]
To investigate the relationship between h and φ1,
we need some general results from the Lp-theory of parabolic
equations.
Recall that H20(Rd)=L2(Rd), and H2γ(Rd)⊂H2γ′(Rd) if γ>γ′. For every
γ,β∈R, (1−Δ)γ/2 is a unitary
isomorphism between H2β(Rd) and
H2γ−β(Rd). In particular,
(1−Δ)γ/2 is a unitary isomorphism between
H2γ(Rd) and L2(Rd). For every v∈H2γ(Rd), ϕ∈D(Rd), we denote
[TABLE]
Note that, if v∈H2γ(Rd) and γ≥0, then
(v,ϕ)=⟨v,ϕ⟩L2(Rd) for all ϕ∈D(Rd).
Definition 4.3**.**
If t↦v(t,⋅) is a function from [0,T] to
H2γ(Rd), with γ∈R, we say that v is a
solution of
[TABLE]
if for any t∈(0,T) and for any
ψ∈D(Rd), we have
[TABLE]
We write v∈H2,0γ(T) if vxx∈L2((0,T),H2γ−2(Rd)), v(0,⋅)=0, and there exists g∈L2((0,T),H2γ−2(Rd)) satisfying (4.4). By
∥v∥H2,0γ(T) we mean
[TABLE]
Remark 4.4*.*
It is known that (4.4) implies that: ∀t∈(0,T), ∀ϕ∈D((0,T)×Rd)
[TABLE]
(see
e.g. Proposition 10, [9] for a stochastic
version of this result). In particular, (4.4) implies
that: ∀ϕ∈D((0,T)×Rd)
[TABLE]
(by taking t=T and using the fact that ϕ(T,⋅)=0).
Theorem 4.5** ([12], [13]).**
Given g∈L2((0,T),H2γ(Rd)),
γ∈R, there exists a unique solution
v∈H2,0γ+2(T) to the equation
[TABLE]
Moreover, there exists a constant N (independent of v) such that
[TABLE]
We now return to our framework.
Theorem 4.6**.**
Let h(t,x)=E(M(φ)u(t,x)), where u is defined in
(3.5) and φ∈P0(d). Set
φ1=(1−Δ)−α/2φ. Then h is the unique
solution in
H2,0α/2+2(T) to the equation
[TABLE]
Proof.
Recall that φ1∈L2((0,T),H2α/2(Rd)).
Thus by Theorem 4.5 there exists a unique solution v∈H2,0α/2+2(T) to the equation (4.6).
We are now proving that h=v. This will follow once we prove that
[TABLE]
Let η∈D((0,T)×Rd) be arbitrary and ϕ be
the unique solution of
[TABLE]
By Lemma 3.6 and (4.3)
[TABLE]
On the other hand, by (4.5) we have
[TABLE]
This concludes the proof.
∎
Corollary 4.7**.**
If f is the Bessel kernel of order α, then
[TABLE]
and the norms in the two spaces are equivalent.
Proof.
From the argument at the beginning of subsection 3.3, for every h∈Hu, we have h(t,x)=EM(φ)u(t,x), where u is
defined in (3.5) and φ∈P0(d).
Then by Theorem 4.6, h∈H2,0α/2+2(T).
For v∈H2,0α/2+2(T), there is a φ1∈L2((0,T),H2α/2(Rd)) satisfying (4.4) with
φ1 in place of g. Then
[TABLE]
Set h(t,x)=EM(φ)u(t,x)∈Hu, then by Theorem
4.6 it follows that v=h, so v∈Hu.
Now notice that
[TABLE]
[TABLE]
where ≃ indicates the equivalence of the norms, which follows
from the definition of the norm of H2,0α/2+2(T) and
the estimate in Theorem 4.5. This finishes the proof.
∎
Remark 4.8*.*
Under the conditions of Theorem 4.6, we can
also say that h is the unique solution in W21,2((0,T)×Rd) of:
[TABLE]
Here W21,2((0,T)×Rd) is the space of all measurable functions v:(0,T)×Rd→R, such that the weak derivatives vt,
vxi, vxixj exist and are in L2((0,T)×Rd).
We are now ready to prove the main result of this subsection.
Theorem 4.9**.**
Suppose that f is the Bessel kernel of
order α=2k,k∈N+ such that α>d−2. Then the
process solution u of the stochastic heat equation (3.1)
with vanishing initial conditions is locally germ Markov.
Proof.
We need to verify conditions (i) and (ii) of Theorem
4.2.
We first verify condition (i). Let h,g∈Hu are such that
(supp h)∩(supp g)=∅ and supp h is compact. We have to prove that ⟨h,g⟩Hu=0. We know that there exist φ,η∈P0(d) such that h(t,x)=E(M(φ)u(t,x)) and
g(t,x)=E(M(η)u(t,x)). Then by Theorem
4.6 and Corollary 4.7 (also
recall the definition of the norm of H2,0γ(T)),
[TABLE]
where φ1=(1−Δ)kφ and η1=(1−Δ)kη. Note that (see also Remark 4.8)
[TABLE]
Similar inclusions hold for g and η1. Thus it is clear that
[TABLE]
from which we arrive at ⟨h,g⟩Hu=0. This
finishes the proof of the condition (i).
To prove the condition (ii), let ζ=h+g∈Hu, where
h and g are such that (supp h)∩(supp g)=∅ and supp h is compact. We have to prove that
h∈Hu.
Let χ be an infinitely differentiable function such that
χ=1 on supp h and χ=0 on an open set containing
supp g. By Corollary 4.7, it follows
that ζ∈H2,0k+2(T), and hence h=χζ∈H2,0k+2(T)=Hu . The theorem is proved.
∎
4.3. The Case of the Riesz Kernel
In this subsection, we will assume that f is the Riesz kernel,
i.e. f=Rα with max{d−2,0}<α=4k<d, k∈N+.
According to [21], for any 0<β<d/2 we can define the
Riesz potential
[TABLE]
Let q=qd,β>2 be such
that 1/q=1/2−β/d.
The space Iβ(L2(Rd)) of all Riesz potentials has the
following properties (see [21]):
- (1)
Iβ(L2(Rd))⊂Lq(Rd) and there exists a
constant N>0 such that
[TABLE]
2. (2)
For every f∈Iβ(L2(Rd)), we define the Riesz
derivative Dβf as limε→0Dεβf in L2(Rd), where
[TABLE]
Δylf=(I−τy)lf and
(τyf)(x)=f(x−y). Then
[TABLE]
3. (3)
Iβ(L2(Rd))={f∈Lq(Rd);Dβf \mboxexistsandisin L2(Rd)}.
4. (4)
Dβ is the left inverse of Iβ in
L2(Rd), i.e. Dβ(Iβφ)=φ for all
φ∈L2(Rd).
5. (5)
D(Rd) is dense in Iβ(L2(Rd)) with respect to
the norm ∥f∥Lq(Rd)+\linebreak∥Dβf∥L2(Rd).
In what follows, we will use these properties with
β=α/2=2k, where d−2<4k<d. We let q=qd,4k>2 be such
that 1/q=1/2−2k/d.
Proposition 4.10**.**
Let 2<d/(2k) and 1/q=1/2−2k/d. If f∈I2k(L2(Rd)), where k∈N+, then
[TABLE]
where Ff is an element in S′(Rd).
Proof.
First of all, ∣ξ∣2kFf is well-defined as an element of
S′ since ∣ξ∣2k=(ξ12+⋯+ξd2)k is
infinitely differentiable.
Due to the fact that D(Rd) is dense in I2k(L2(Rd)),
there exists a sequence {fn}⊂D(Rd) such that
[TABLE]
Note that F(D2kfn)=∣ξ∣2kFfn∈S(Rd). Also note that
[TABLE]
[TABLE]
where the convergence is possible because fn→f in
Lq(Rd). This along with the fact that D2kfn→D2kf in L2(Rd) implies that
[TABLE]
∎
Corollary 4.11**.**
Let 2<d/(2k) and 1/q=1/2−2k/d. If f∈I2k(L2(Rd)), where k∈N+, then
[TABLE]
Proof.
Note that
[TABLE]
On the other hand,
[TABLE]
[TABLE]
∎
Lemma 4.12**.**
Let 2<d/2k. There exists a linear operator J:P0(d)→L2((0,T)×Rd) such that
J is one-to-one and onto, and satisfies:
[TABLE]
for any φ∈D((0,T)×Rd).
Proof.
For all φ∈D((0,T)×Rd), set
[TABLE]
where cd,k is an appropriate constant. Denote φ0:=J(φ). We see that φ0(t,x) is measurable in (t,x)∈(0,T)×Rd. Since φ(t,⋅)∈S(Rd) for
each fixed t∈(0,T), we have
[TABLE]
in the sense that
[TABLE]
where ϕ∈S(Rd). Also notice that, for a.e. t∈(0,T),
[TABLE]
because φ∈P0(d) (recall that α=4k), i.e.,
[TABLE]
Thus, for a.e. t∈(0,T), as a function of x∈Rd,
φ0(t,x)∈L2(Rd) and
[TABLE]
Thus
[TABLE]
[TABLE]
From this and the fact that D((0,T)×Rd) is dense in
P0(d), we extend J to all elements in P0(d).
It is easy to see that J is linear as well as one-to-one. To prove
the fact that J is onto, take φ0∈L2((0,T)×Rd). Also take a sequence {φ0n}⊂D((0,T)×Rd) such that φ0n→φ0 in L2((0,T)×Rd). Let φn=(−Δ)kφ0n. Then
[TABLE]
and
[TABLE]
Thus {φn} is a Cauchy sequence in P0(d), so there
is a φ∈P0(d) such that φn→φ in
P0(d). To prove J(φ)=φ0, we only need to
prove that
[TABLE]
which follows from
[TABLE]
∎
From the above result it follows that:
Corollary 4.13**.**
For φ, η∈P0(d), let φ0:=J(φ)
and η0:=J(η). Then
[TABLE]
Now we set
[TABLE]
where φ0∈L2((0,T)×Rd). Notice that φ1
is a measurable function of (t,x)∈(0,T)×Rd. Since
φ0(t,x)∈L2(Rd) for a.e. t∈(0,T),
[TABLE]
for a.e. t∈(0,T), where 1/q=1/2−2k/d. This implies that
[TABLE]
[TABLE]
From the Lp-theory of parabolic equations with mixed norms, there
exists a unique function w∈W2,q1,2((0,T)×Rd)
satisfying
[TABLE]
Lemma 4.14**.**
Let h(t,x)=EM(φ)u(t,x), where
u is defined in (3.5). Let φ0=J(φ), φ1 be a function defined as above, and w be
the solution to (4.7). Then
[TABLE]
Proof.
For η∈D((0,T)×Rd), find a function ϕ, a
unique solution to
[TABLE]
Then by Lemma 3.6
[TABLE]
and
[TABLE]
Find a sequence {φ0n}⊂D((0,T)×Rd) such that φ0n→φ0 in L2((0,T)×Rd). Let ψn=I2k(φ0n) and wn be the
solution to (∂/∂t−Δ)wn=ψn and
wn(0,⋅)=0. Then
[TABLE]
[TABLE]
On the other hand,
[TABLE]
[TABLE]
where φn=(−Δ)2kφ0n. This finishes the
proof.
∎
Now we prove the main result of this subsection.
Theorem 4.15**.**
Suppose that f is the Riesz kernel of order α=4k,k∈N+ and d=4k+1. Then the process solution u of the
stochastic heat equation (3.1) with vanishing initial
conditions is locally germ Markov.
Proof.
Again we need to verify conditions (i) and (ii) in Theorem
4.2.
We first verify condition (i).
Let h(t,x)=EM(φ)u(t,x) and g(t,x)=EM(η)u(t,x), where φ,η∈P0(d), such that
supph∩suppg=∅. We need to
show that ⟨φ,η⟩0=0. Let φ0=J(φ) and η0=J(η), where J is the operator
defined in Lemma 4.12. Also let φ1=I2k(φ0(t,⋅)) and η1=I2k(η0(t,⋅)). For a.e. t∈(0,T), we
have
[TABLE]
where the second equality follows from Corollary
4.11. Let ϕ(t,x)∈D((0,T)×Rd)
such that ϕ(t,x)=0 on suppφ1. Then for
a.e. t∈(0,T),
[TABLE]
Thus
[TABLE]
This shows that suppφ0⊂suppφ1. Similarly, we have suppη0⊂suppη1. Hence we have ⟨φ0,η0⟩L2((0,T)×Rd)=0. This and Corollary
4.13 prove that ⟨φ,η⟩0=0.
We now verify condition (ii).
Assume that ζ=h+g, where ζ(t,x)=