The stochastic wave equation

by

Robert C. Dalang1

Abstract. These notes give an overview of recent results concerning the non-linear
stochastic wave equation in spatial dimensions d ≥ 1, in the case where the driving
noise is Gaussian, spatially homogeneous and white in time. We mainly address issues
of existence, uniqueness and Hölder-Sobolev regularity. We also present an extension
of Walsh’s theory of stochastic integration with respect to martingale measures that
is useful for spatial dimensions d ≥ 3.

Key words and phrases. Stochastic partial differential equations, sample path
regularity, spatially homogeneous random noise, wave equation.
MSC 2000 Subject Classifications. Primary 60H15; Secondary 60J45, 35R60,
35L05.

1
Partially supported by the Swiss National Foundation for Scientific Research.
1 Introduction
The stochastic wave equation is one of the fundamental stochastic partial differential
equations (s.p.d.e.’s), of hyperbolic type. The behavior of its solutions is signifi-
cantly different from those of solutions to other s.p.d.e.’s, such as the stochastic heat
equation. In this introductory section, we present two real-world examples that can
motivate the study of this equation, even though in neither case is the mathematical
technology sufficiently developed to answer the main questions of interest. It is how-
ever pleasant to have such examples in order to motivate the development of rigorous
mathematics.

Example 1: The motion of a strand of DNA

A DNA molecule can be view as a long elastic string, whose length is essentially
infinitely long compared to its diameter. We can describe the position of the string
by using a parameterization defined on R+ × [0, 1] with values in R3 :
 
u1 (t, x)
~u(t, x) =  u2 (t, x) 

.
u3 (t, x)

Here, ~u(t, x) is the position at time t of the point labelled x on the string, where
x ∈ [0, 1] represents the distance from this point to one extremity of the string if the
string were straightened out. The unit of length is chosen so that the entire string
has length 1.
A DNA molecule typically “floats” in a fluid, so it is constantly in motion, just
as a particle of pollen floating in a fluid moves according to Brownian motion. The
motion of the particle can be described by Newton’s law of motion, which equates the
sum of forces acting on the string with the product of the mass and the acceleration.
Let µ = 1 be the mass of the string per unit length. The acceleration at position x
along the string, at time t, is
∂ 2~u
(t, x),
∂t2
and the forces acting on the string are mainly of three kinds: elastic forces F~1 , which
include torsion forces, friction due to viscosity of the fluid F~2 , and random impulses
F~3 due the the impacts on the string of the fluid’s molecules. Newton’s equation of
motion can therefore be written
∂ 2~u
1· = F~1 − F~2 + F~3 .
∂t2
This is a rather complicated system of three stochastic partial differential equa-
tions, and it is not even clear how to write down the torsion forces or the friction term.

1
Elastic forces are generally related to the second derivative in the spatial variable,
and the molecular forces are reasonably modelled by a stochastic noise term.
The simplest 1-dimensional equation related to this problem, in which one only
considers vertical displacement and forgets about torsion, is the following one, in
which u(t, x) is now scalar valued:

∂2u ∂2u Z 1
(t, x) = (t, x) − k(x, y) u(t, y) dy + Ḟ (t, x), (1.1)
∂t2 ∂x2 0

where the first term on the right hand side represents the elastic forces, the second
term is a (non-local) friction term, and the third term Ḟ (t, y) is a Gaussian noise,
with spatial correlation k(·, ·), that is,

E(Ḟ (t, x) Ḟ (s, y)) = δ0 (t − s) k(x, y),

where δ0 denotes the Dirac delta function. The function k(·, ·) is the same in the
friction term and in the correlation.
Why is the motion of a DNA strand of biological interest? When a DNA strand
moves around and two normally distant parts of the string get close enough together,
it can happen that a biological event occurs: for instance, an enzyme may be released.
Therefore, some biological events are related to the motion of the DNA string. Some
mathematical results for equation (1.1) can be found in [18]. Some of the biological
motivation can be found in [7].

Example 2: The internal structure of the sun

The study of the internal structure of the sun is an active area of research. One
important international project is known as Project SOHO [8]. Its objective was to
use measurements of the motion of the sun’s surface to obtain information about the
internal structure of the sun. Indeed, the sun’s surface moves in a rather complex
manner: at any given time, any point on the surface is typically moving towards or
away from the center. There are also waves going around the surface, as well as shock
waves propagating through the sun itself, which cause the surface to pulsate.
A question of interest to solar geophysicists is to determine the origin of these
shock waves. One school of thought is that they are due to turbulence, but the
location and intensities of the shocks are unknown, so a probabilistic model can be
considered.
A model that was proposed by P. Stark of U.C. Berkeley is that the main source
of shocks is located in a spherical zone inside the sun, which is assumed to be a ball
of radius R. Assuming that the shocks are randomly located on this sphere, the
equation for the pressure variations throughout the sun would be

∂2u
! !
~ · 1 ~ ~ · F~ (t, x) ,
2
(t, x) = c2 (x) ρ0 (x) ∇ ∇u − ∇ (1.2)
∂t ρ0 (x)

2
where x ∈ B(0, R), the ball centered at the origin with radius R, c2 (x) is the speed of
wave propagation at position x, ρ0 (x) is the density at position x and F~ (t, x) models
the shock that originates at time t and position x.
A model for F~ that corresponds to the description of the situation would be 3-
dimensional Gaussian noise concentrated on the sphere ∂B(0, r), where 0 < r < R.
A possible choice of the spatial correlation for the components of F~ would be

δ(t − s) f (x · y),

where x · y denotes the Euclidean inner product. A problem of interest is to estimate

r from the available observations of the sun’s surface. Some mathematical results
relevant to this problem are developed in [3].

2 The stochastic wave equation

Equation (1.2) is a wave equation in vector form. The (simpler) real-valued stochastic
wave equation, that we will be studying in these notes, reads as follows:
  2 
∂ u
 ∂t2 − ∆u (t, x) = σ(t, x, u(t, x)) Ḟ (t, x) + b(t, x, u(t, x),
 (t, x) ∈ [0, t] × Rd
∂u
 u(0, x) = v0 (x),
 (0, x) = ṽ0 (x),
∂t
(2.1)
where Ḟ (t, x) is a Gaussian noise, which we take to be space-time white noise for the
moment, and σ, b : R+ × Rd × R → R are functions that satisfy standard properties,
such as being Lipschitz in the third variable. The term ∆u denotes the Laplacian of
u in the x-variables.

Mild solutions of the stochastic wave equation

It is necessary to specify the notion of solution to (2.1) that we are considering.

We will mainly be interested in the notion of mild solution, which is the following
integral form of (2.1):
Z
u(t, x) = G(t − s, x − y) [σ(s, y, u(s, y)) Ḟ (s, y) + b(s, y, u(s, y))] ds dy
[0,t]×Rd
!
d
+ G(t) ∗ v0 (x) + (G(t) ∗ ṽ0 )(x). (2.2)
dt

In this equation, G(t − s, x − y) is the Green’s function of (2.1), which we discuss

next, and ∗ denotes convolution in the x-variables. For the term involving Ḟ (s, y), a
notion of stochastic integral is needed, that we will discuss later on.

Green’s function of a p.d.e.

3
 We consider first the case of an equation with constant coefficients. Let L be a
partial differential operator with constant coefficients. A basic example is the wave
operator
∂2f
Lf = 2 − ∆f.
∂t
Then there is a (Schwartz) distribution G ∈ S 0 (R+ × Rd ) such that the solution of
the p.d.e.
L u = ϕ, ϕ ∈ S(Rd ),
is
∗ ϕ
u = G (t,x)

where ∗
(t,x) denotes convolution in the (t, x)-variables. We recall that S(Rd ) denotes
the space of smooth test functions with rapid decrease, and S 0 (R+ × Rd ) denotes the
space of tempered distributions [13].
When G is a function, this convolution can be written
Z
u(t, x) = G(t − s, x − y) ϕ(s, y) ds dy.
R+ ×Rd

We note that this is the solution with vanishing initial conditions.

In the case of an operator with non-constant coefficients, such as

∂2f ∂f ∂2f
Lf = 2 + 2c(t, x) + 2 (d = 1),
∂t ∂t ∂x
the Green’s function has the form G(t, x ; s, y) and the solution of

Lu = ϕ

is given by the expression

Z
u(t, x) = G(t, x ; s, y) ϕ(s, y) ds dy.
R+ ×Rd

Example 2.1 The heat equation. The partial differential operator is

∂u
Lu = − ∆u, d ≥ 1,
∂t
and the Green’s function is
|x|2
!
−d/2
G(t, x) = (2πt) exp − .
2t

This function is smooth except for a singularity at (0, 0).

4
Example 2.2 The wave equation. The partial differential operator is
∂2u
Lu = − ∆u.
∂t2
The form of the Green’s function depends on the dimension d. We refer to [16] for
d ∈ {1, 2, 3} and to [6] for d > 3. For d = 1, it is
1
G(t, x) = 1{|x|<t} ,
2
which is a bounded but discontinuous function. For d = 2, it is
1 1
G(t, x) = √ q 1{|x|<t} .
2π t2 − |x|2

This function is unbounded and discontinuous. For d = 3, the “Green’s function” is

1 σt (dx)
G(t, dx) = ,
4π t
where σt is uniform measure on ∂B(0, t), with total mass 4πt2 . In particular, G(t, R3 ) =
t. This Green’s function is in fact not a function, but a measure. Its convolution with
a test function ϕ is given by
1 Zt Z
σs (dy)
(G ∗ ϕ)(t, x) = ds ϕ(t − s, x − y)
4π 0 ∂B(0,s) s
1 Zt Z
= ds s ϕ(t − s, x − sy) σ1 (dy).
4π 0 ∂B(0,1)

Of course, the meaning of an expression such as

Z
G(t − s, x − y) F (ds, dy)
[0,t]×Rd

where G is a measure and F is a Gaussian noise, is now unclear: it is certainly outside

of Walsh’s theory of stochastic integration [9].
In dimensions greater than 3, the Green’s function of the wave equation becomes
even more irregular. For d ≥ 4, set
d−3


 2
if d is odd,
N (d) =
 d−2

2
if d is even.

For d even, set

1
σtd (dx) = q 1{|x| <t} dx,
t2 − |x|2

5
and for d odd, let σtd (dx) be the uniform surface measure on ∂B(0, t) with total mass
td−1 . Then for d odd, G(t, x) can formally be written
!N (d)
σsd
!
1 ∂
G(t, x) = cd ds,
s ∂s s

that is, for d odd,

!N (d)
σ d (dy)
!
Z t 1 ∂ Z
(G ∗ ϕ)(t, x) = cd ds ϕ(t − s, x − y) r
0 r ∂r Rd r r=s

while for d even,

!N (d) Z 
Z t 1 ∂ dy
(G ∗ ϕ)(t, x) = cd ds  ϕ(t − s, x − y) q  .
0 r ∂r B(0,r) r2 − |y|2 r=s

G(t − s, x − y) F (ds, dy) is even less clear in these cases!

R
The meaning of [0,t]

The case of spatial dimension one

Existence and uniqueness of the solution to the stochastic wave equation in spatial
dimension 1 is covered in [17, Exercise 3.7 p.323]. It is a good exercise that we leave
to the reader.
Problem 1. Establish existence and uniqueness of the solution to the non-linear wave
equation on [0, t] × R, driven by space-time white noise :

∂2u ∂2u
− 2 = σ(u(t, x)) Ẇ (t, x)
∂t2 ∂x
with initial conditions
∂u
u(0, ·) =
(0, ·) ≡ 0.
∂x
The solution uses the following standard steps, which also appear in the study of the
semilinear stochastic heat equation (see [17] and [9]):
- define the Picard iteration scheme;
- establish L2 -convergence using Gronwall’s lemma;
- show existence of higher moments of the solution, using Burkholder’s inequality
p/2
E(|Mt |p ) ≤ cp E(hM it ); (2.3)

- establish Hölder continuity of the solution.

It is also a good exercise to do the following calculation.

6
Problem 2. Let G be the Green’s function of the wave equation, as defined in Example
2.2. For d = 1 and d = 2, check that
Z t Z
u(t, x) = ds dy G(t − s, x − y) ϕ(s, y)
0 Rd

satisfies
∂2u
(t, x) − ∆u(t, x) = ϕ(t, x).
∂t2

Space-time white noise in dimension d = 2

Having solved the non-linear stochastic wave equation driven by space-time white
noise in dimension d = 1, it is tempting to attempt the same thing in dimension
d = 2. We are going to show that there is a fundamental obstacle to doing this.
To this end, consider the linear case, that is, σ ≡ 1 and b ≡ 0. The mild solution
given in (2.2) is not an equation in this case, but a formula:
Z
u(t, x) = G(t − s, x − y) W (ds, dy)
[0,t]×R2
Z
1 1
= √ q 1{|y−x| <t−s} W (ds, dy),
[0,t]×R2 2π (t − s)2 − |y − x|2

where W (ds, dy) is space-time white noise.

The first issue is whether this stochastic integral well-defined. For this, we would
need (see [9, Exercise 5.5]) to have
Z t Z
ds dy G2 (t − s, x − y) < +∞.
0 R2

The integral is equal to

Z t Z
dy Z t Z
dz
ds = dr
0 |y−x|<t−s (t − s)2 + |y − x|2 0 |z|<r r 2 − |z|2
Z t Z r
2πρ
= dr dρ 2
0 0 r − ρ2
Z t 0
= π dr ln(r2 − ρ2 )
0 r

= +∞.
In particular, there is no mild solution to the wave equation (2.2) when d = 2.
There have been some attemps at overcoming this problem [11], but as yet, there
is no fully satisfactory approach to studying non-linear forms of the stochastic wave
or heat equations driven by space-time white noise in dimensions d ≥ 2.
A different tack is to consider spatially homogeneous noise, which we introduce
now.

7
3 Spatially homogeneous Gaussian noise
Let Γ be a non-negative and non-negative definite tempered measure on Rd , so that
Γ(dx) ≥ 0, Z
Γ(dx) (ϕ ∗ ϕ̃)(x) ≥ 0, for all ϕ ∈ S(Rd ),
Rd
def
where ϕ̃(x) = ϕ(−x), and there exists r > 0 such that
Z
1
Γ(dx) < ∞.
Rd (1 + |x|2 )r
According to the Bochner-Schwartz theorem [13], there is a nonnegative measure
µ on Rd whose Fourier transform is Γ: we write Γ = Fµ. By definition, this means
that for all ϕ ∈ S(Rd ),
Z Z
Γ(dx) ϕ(x) = µ(dη) Fϕ(η) .
Rd Rd

We recall that the Fourier transform of ϕ ∈ S(Rd ) is

Z
Fϕ(η) = exp(−i η · x) ϕ(x) dx,
Rd

where η · x denotes the Euclidean inner product. The measure µ is called the spectral
measure.
Definition 3.1 A spatially homogeneous Gaussian noise that is white in time is an
L2 (Ω, F, P )−valued mean zero Gaussian process
 
F (ϕ), ϕ ∈ C0∞ (R1+d )
such that
E(F (ϕ) F (ψ)) = J(ϕ, ψ),
where Z Z
def
J(ϕ, ψ) = ds Γ(dx) (ϕ(s, ·) ∗ ψ̃(s, ·))(x).
R+ Rd

In the case where the covariance measure Γ has a density, so that Γ(dx) = f (x) dx,
then it is immediate to check that J(ϕ, ψ) can be written as follows:
Z Z Z
J(ϕ, ψ) = ds dx dy ϕ(s, x) f (x − y) ψ(s, y).
R+ Rd Rd

Using the fact that the Fourier transform of a convolution is the product of the Fourier
transforms, this can also be written
Z Z
J(ϕ, ψ) = ds µ(dη) Fϕ(s)(η) Fψ(s)(η).
R+ Rd

Informally, one often writes

E(Ḟ (t, x)Ḟ (s, y)) = δ0 (t − s) f (x − y).

8
Example 3.2 (a) If Γ(dx) = δ0 (x), where δ0 denotes the Dirac delta function, then
the associated spatially homogeneous Gaussian noise is simply space-time white noise.
(b) Fix 0 < β < d and let
dx
Γβ (dx) = β .
|x|
One can check [15, Chapter 5] that Γβ = Fµβ , with

dη
µβ (dη) = cd,β .
|η|d−β

We point out that if β ↑ d, then the spatially homogeneous Gaussain noise Fβ with
the covariance measure Γβ converges weakly to space-time white noise. Indeed, the
spectral measure µβ converges weakly to a multiple of Lebesgue measure on Rd , which
is the spectral measure of space-time white noise, since F(dη) = δ0 .

Extension of F (ϕ) to a worthy martingale measure

From the spatially homogenenous Gaussian noise, we are going to construct a

worthy martingale measure M = (Mt (A), t ≥ 0, A ∈ Bb (Rd )), where Bb (Rd ) denotes
the family of bounded Borel subsets of Rd . For this, if A ∈ Bb (Rd ), we set
def
Mt (A) = lim F (ϕn ),
n→∞

where the limit is in L2 (Ω, F, P ), ϕn ∈ C0∞ (Rd+1 ) and ϕn ↓ 1[0,t]×A .

One checks [2] that (Mt (A), t ≥ 0, A ∈ Bb (Rd )) is a worthy martingale measure
in the sense of Walsh; its covariation measure Q is given by
Z Z
Q(A × B×]s, t]) = (t − s) dx dy 1A (x) f (x − y) 1B (y),
Rd Rd

and its dominating measure is K ≡ Q. The key relationship between F and M is

that Z
F (ϕ) = ϕ(t, x) M (dt, dx),
R+ ×Rd

where the stochastic integral on the right-hand side is Walsh’s martingale measure
stochastic integral.
The underlying filtration (F t , t ≥ 0) associated with this martingale measure is
given by  
F t = σ Ms (A), s ≤ t, A ∈ Bb (Rd ) ∨ N , t ≥ 0,
where N is the σ-field generated by all P -null sets.

9
4 The wave equation in spatial dimension 2
We shall consider the following form of the stochastic wave equation in spatial di-
mension d = 2:
∂2u
!
− ∆u (t, x) = σ(u(t, x)) Ḟ (t, x), (t, x) ∈ ]0, T ] × R2 , (4.1)
∂t2

with vanishing initial conditions. By a solution to (4.1), we mean a jointly measurable

adapted process (u(t, x)) that satisfies the associated integral equation
Z
u(t, x) = G(t − s, x − y) σ(u(s, y)) M (ds, dy), (4.2)
[0,t]×R2

where M is the worthy martingale measure associated with Ḟ .

The linear equation

A first step is to examine the linear equation, which corresponds to the case where
σ ≡ 1:
∂2u
!
− ∆u (t, x) = Ḟ (t, x), (4.3)
∂t2
with vanishing initial conditions. The mild solution should be
Z
u(t, x) = G(t − s, x − y) M (ds, dy).
[0,t]×R2

We know that the stochastic integral on the right-hand side is not defined for space-
time white noise, so let us determine for which spatially homogeneous Gaussian noises
it is well-defined. This is the case if
Z t Z Z
ds dy dz G(t − s, x − y) f (y − z) G(t − s, x − z) < +∞,
0 R2 R2

or, equivalently, if Z t Z
ds µ(dη) |FG(s)(η)|2 < +∞. (4.4)
0 R2

Calculation of FG

In principle, the Green’s function of a p.d.e. solves the same p.d.e. with δ(0,0) (t, x) =
δ0 (t) δ0 (x) as right-hand side :

∂2G
− ∆G = δ0 (t) δ0 (x). (4.5)
∂t2

10
For fixed t > 0, the right-hand side vanishes. We shall take the Fourier transform in
x on both sides of this equation, but first, we observe that since
Z
FG(t)(ξ) = Ĝ(t)(ξ) = ei ξ·x G(t, x) dx,
R2

it is clear that
∂ 2 G(t) ∂ 2 Ĝ(t)
!
F (ξ) = (ξ),
∂t ∂t2
and, using integration by parts, that
Z
F(∆G(t)) = ei ξ·x ∆G(t, x) dx
R2
Z
= ∆(ei ξ·x ) G(t, x) dx
R2
= −|ξ|2 FG(t) (ξ).

Therefore, we deduce from (4.5) that for t > 0,

∂ 2 Ĝ(t)
2
(ξ) + |ξ|2 Ĝ(t) (ξ) = δ0 (t).
∂t
For fixed ξ, the solution to the associated homogeneous ordinary differential equation
in t is
sin(t|ξ|) cos(t|ξ|)
Ĝ(t)(ξ) = a(ξ) + b(ξ) .
|ξ| |ξ|
The solution that we seek [16, Chapter I, Section 4] is the one such that Ĝ(0)(ξ) = 0
and dĜ(0)
dt
(ξ) = 1, so we conclude that for t ≥ 0 and ξ ∈ R2 ,

sin(t|ξ|)
FG(t) (ξ) = . (4.6)
|ξ|

This formula is in fact valid in all dimensions d ≥ 1.

Condition on the spectral measure

Condition (4.4) for existence of a mild solution on [0, T ] to the linear wave equation
(4.3) becomes
Z T Z
sin2 (s|η|)
ds µ(dη) < +∞.
0 R2 |η|2
Using Fubini’s theorem, one can evaluate the ds-integral explicitly, or simply check
that
c1 Z T
sin2 (s|η|) c2
2
≤ ds 2
≤ ,
1 + |η| 0 |η| 1 + |η|2

11
so condition (4.4) on the spectral measure becomes
Z
1
µ(dη) < +∞. (4.7)
R2 1 + |η|2

Example 4.1 Consider the case where f (u) = |x|−β , 0 < β < d. In this case,
µ(dη) = cd,β |η|β−d dy, so one checks immediately that condition (4.7) holds if and
only if β < 2. Therefore, the spatially homogeneous Gaussian noise is defined for
0 < β < d, but a mild solution of the linear stochastic wave equation (4.3) exists if
and only if 0 < β < 2.

Reformulating (4.7) in terms of the covariance measure

Condition (4.7) on the spectral measure can be reformulated as a condition on the

covariance measure Γ. It is shown in [10] that in dimension d = 2, (4.7) is equivalent
to !
Z
1
Γ(dx) ln < +∞,
|x|≤1 |x|
while in dimensions d ≥ 3, (4.7) is equivalent to
Z
1
Γ(dx) < +∞.
|x| ≤1 |x|d−2

In dimension d = 1, condition (4.7) is satisfied for any non-negative measure µ such

that Γ = Fµ is also a non-negative measure.

The non-linear wave equation in dimension d = 2

We consider equation (4.1). The following theorem is the main result on existence
and uniqueness.

Theorem 4.2 Assume d = 2. Suppose that σ is a Lipschitz continuous function and

that condition (4.7) holds. Then there exists a unique solution (u(t, x), t ≥ 0, x ∈ R2 )
of (4.1) and for all p ≥ 1, this solution satisfies

sup sup E (|u(t, x)|p ) < ∞.

0≤t≤T x∈Rd

Proof. This proof follows a classical Picard iteration scheme. We set u0 (t, x) = 0,
and, by induction, for n ≥ 0,
Z
un+1 (t, x) = G(t − s, x − y) σ(un (s, y)) M (ds, dy).
[0,t]×R2

12
Before establishing convergence of this scheme, we first check that for p ≥ 2,

sup sup sup E (|un (s, x)|p ) < +∞.

n≥0 0≤s≤T x∈R2

We apply Burkholder’s inequality (2.3) and use the explicit form of the quadratic
variation of the stochastic integral [9, Theorem 5.26] to see that
 Z t Z Z
E (|un+1 (t, x)|p ) ≤ cE ds dy dz G(t − s, x − y) σ(un (s, y))
0 R2 R2
p 
2
×f (y − z) G(t − s, x − z) σ(un (s, z)) .

Since G ≥ 0 and f ≥ 0, we apply Hölder’s inequality in the form

Z p Z p/q Z 
p
p
f dµ ≤ 1 dµ |f | dµ , where =p−1 (4.8)
q
and µ is a non-negative measure, to see that E (|un+1 (t, x)|p ) is bounded above by
Z t Z Z  p −1
2
c ds dy dz G(t − s, x − y) f (y − z) G(t − s, x − z)
0 R2 R2
Z t Z Z
× ds dy dz G(t − s, x − y) f (y − z) G(t − s, x − z)
0 R2 R2
 p 
×E |σ(un (s, y)) σ(un (s, z))| 2 .

We apply the Cauchy-Schwartz inequality to the expectation and use the Lipschitz
property of σ to bound this by
Z t Z  p −1
2
2
C ds µ(dη) |FG(t − s)(η)|
0 R2
Z t Z Z
× ds dy dz G(t − s, x − y) f (y − z) G(t − s, x − z)
0 R2 R2
× (E (1 + |un (s, y)|p ))1/2 (E (1 + |un (s, z)|p ))1/2 .

Let t
Z Z
2
Z
1
J(t) = ds µ(dη) |FG(t − s)(η)| ≤ C µ(dη) .
0 R2 R2 1 + |η|2
Then
!
p
Z t
p −1 p
E (|un+1 (t, x)| ) ≤ C (J(t)) 2 ds 1 + sup E (|un (s, y)| )
0 y∈R2
Z
× µ(dη) |FG(t − s)(η)|2
R2
!
Z t
≤ C̃ ds 1 + sup E(|un (s, y)|p ) .
0 y∈R2

13
Therefore, if we set
Mn (t) = sup E (|un (t, x)|p ) ,
x∈R2

then Z t
Mn+1 (t) ≤ C̃ ds (1 + Mn (s)) .
0
Using Gronwall’s lemma, we conclude that

sup sup Mn (t) < +∞.

n∈N 0≤t≤T

We now check L2 -convergence of the Picard iteration scheme. By the same rea-
soning as above, we show that
Z t
sup E (|un+1 (t, x) − un (t, x)|p ) ≤ C ds sup E (|un (s, y) − un−1 (s, y)|p ) .
x∈R2 0 y∈R2

Gronwall’s lemma shows that (un (t, x), n ≥ 1) converges in L2 (Ω, F, P ), uniformly
in x ∈ R2 .
Uniqueness of the solution follows in a standard way: see [9, Proof of Theorem
6.4]. 2

Hölder-continuity (d = 2)

In order to establish Hölder continuity of the solution to the stochastic wave

equation in spatial dimension 2, we first recall the Kolmogorov continuity theorem. It
is a good idea to compare this statement with the equivalent one in [9, Theorem 4.3].

Theorem 4.3 (The Kolmogorov Continuity Theorem). Suppose that there is q > 0,
ρ ∈ ] dq , 1[ and C > 0 such that for all x, y ∈ Rd ,

E (|u(t, x) − u(t, y)|q ) ≤ C |x − y|ρq . (4.9)

Then x 7→ u(t, x) has a ρ̃-Hölder continuous version, for any ρ̃ ∈ ]0, ρ − dq [.

In order to use the statement of this theorem to establish (ρ−ε)-Hölder continuity,

for any ε > 0, it is necessary to obtain estimates on arbitrarily high moments of
increments, that is, to establish (4.9) for arbitrarily large q.

Lq -moments of increments

From the integral equation (4.2), we see that

ZZ
u(t, x) − u(s, y) = (G(t − r, x − z) − G(s − r, y − z)) σ(u(r, z)) M (dr, dz),

14
and so, by Burkholder’s inequality (2.3),
E (|u(t, x) − u(s, y)|p )
Z t Z Z
≤CE dr dz dv (G(t − r, x − z) − G(s − r, y − z)) f (z − v)
0 R2 R2
p/2 !
× (G(t − r, x − v) − G(s − r, y − v)) σ(u(r, z)) σ(u(r, v))
Z Z Z  p −1
2
≤C dr dz dv |G() − G()| f ( ) |G() − G()|
Z Z Z
× dr dz dv |G() − G()| f ( ) |G() − G()|
 
×E |σ(u(r, z))|p/2 |σ(u(r, v))|p/2 ,
where the omitted variables are easily filled in. The Lipschitz property of σ implies a
bound of the type “linear growth”, and so, using also the Cauchy-Schwartz inequality,
we see that the expectation is bounded by
C sup (1 + E(|u(r, z)|p )) .
r≤T, z∈R2

Define
Z t Z Z
J(t, x ; s, y) = dr dz dv |G(t − r, x − z) − G(s − r, y − z)| f (z − v)
0 R2 R2
×|G(t − r, x − v) − G(s − r, y − v)|.
We have shown that
E (|u(t, x) − u(s, y)|p ) ≤ (J(t, x ; s, y))p/2 .
Therefore, we will get Hölder-continuity provided, for some γ > 0 and ρ > 0, we can
establish an estimate of the type
J(t, x ; s, y) ≤ c(|t − s|γ + |x − y|ρ ).
Indeed, this will establish γ21 -Hölder continuity in time, and ρ1
2
-Hölder continuity in
in space, for all γ1 ∈ ]0, γ[ and ρ1 ∈ ]0, ρ[.

Analysis of J(t, x ; s, y)

If there were no absolute values around the increments of G, then we could use the
Fourier transform to rewrite J(t, x ; s, y), in the case x = y and s > t, for instance,
as
Z s Z
J(t, x ; s, x) = dr µ(dη) |FG(t − r)(η) − FG(s − r)(η)|2
0 R2
Z t Z
+ dr µ(dη) |FG(t − r)(η)|2 .
s R2

15
We could then analyse this using the specific form of FG in (4.6). However, the
presence of the absolute values makes this approach inoperable. By a direct analysis
of J(t, x ; s, x), Sanz-Solé and Sarrá [12] have established to following results. If
Z
1
µ(dη) < ∞, for some a ∈ ]0, 1[,
R2 (1 + |η|2 )a
then t 7→ u(t, x) is γ1 -Hölder continuous, for
1
 
γ1 ∈ 0, ∧ (1 − a) ,
2
and x 7→ u(t, x) is γ2 -Hölder continuous, for γ2 ∈ ]0, 1 − a[.
When µ(dη) = |η|−β dη, these intervals become
# " # "
1 2−β 2−β
γ1 ∈ 0, ∧ and γ2 ∈ 0, .
2 2 2

We note that the best possible interval for γ1 is in fact ]0, 2−β
2
[ (see [5, Chapter 5]).

5 A function-valued stochastic integral

Because the Green’s function in spatial dimension 3 is a measure and not a function,
the study of the wave equation in this dimension requires different methods than those
used in dimensions 1 and 2. In particular, we will use a function-valued stochastic
integral, developed in [4].
Our first objective is to define a stochastic integral of the form
Z
G(s, x − y) Z(s, y) M (ds, dy),
[0,t]×Rd

where G(s, ·) is the Green’s function of the wave equation (see Example 2.2) and
Z(s, y) is a random field that plays the role of σ(u(s, y)).
We shall assume for the moment that d ≥ 1 and that the following conditions are
satisfied.

Hypotheses

(H1) For 0 ≤ s ≤ T, Z(s, ·) ∈ L2 (Rd ) a.s., Z(s, ·) is F s −measurable, and s 7→ Z(s, ·)

from R+ → L2 (Rd ) is continuous.

(H2) For all s ≥ 0,

Z T Z
ds sup µ(dη) |FG(s)(ξ − η)|2 < +∞.
0 ξ∈Rd Rd

16
 We note that FG(s)(ξ − η) is given in (4.6), so that (H2) is a condition on the
spectral measure µ, while (H1) is a condition on Z.
Fix ψ ∈ C0∞ (Rd ) such that ψ ≥ 0, supp ψ ⊂ B(0, 1) and
Z
ψ(x) dx = 1.
Rd

For n ≥ 1, set
ψn (x) = nd ψ(n x).
In particular, ψn → δ0 in S 0 (Rd ), and F ψn (ξ) = F ψ(ξ/n), so that |F ψn (ξ)| ≤ 1.
Define
Gn (s, ·) = G(s) ∗ ψn ,
so that Gn is a C0∞ -function. Then
Z
def
vGn ,Z (t, x) = Gn (s, x − y) Z(s, y) M (ds, dy)
[0,t]×Rd

is well-defined as a Walsh-stochastic integral, and

 
E kvGn ,Z (t, ·)k2L2 (Rd ) = IGn , Z , (5.1)

where
Z  
IGn ,Z = dx E (vGn ,Z (t, x))2
Rd
Z Z t Z Z
= dx ds dy dz Gn (s, x − y)Z(s, y) f (y − z)
Rd 0 Rd Rd
× Gn (s, x − z) Z(s, z).

Using the fact that the Fourier transform of a convolution (respectively product) is
the product (resp. convolution) of the Fourier transforms, one easily checks that
Z t Z  Z
2
IGn ,Z = ds dξ E |FZ(s, ·)(ξ)| µ(dη) |FGn (s, ·) (ξ − η)|2 .
0 Rd Rd

We note that:
(a) the following inequality holds:

IGn ,Z ≤ I˜Gn ,Z , (5.2)

where
Z t   Z
def
I˜Gn ,Z = ds E kZ(s, ·)k2L2 (Rd ) sup µ(dη) |FGn (s, ·)(ξ − η)|2 ; (5.3)
0 ξ∈Rd Rd

(b) the equality (5.1) plays the role of an isometry property;

17
 (c) by elementary properties of convolution and Fourier transform,

I˜Gn ,Z ≤ I˜G,Z < +∞,

by (H2) and (H1).

In addition, one checks that the stochastic integral
def
vG,Z (t) = lim vGn ,Z
n→∞

exists, in the sense that

 
E kvG,Z (t) − vGn ,Z (t, ·)k2L2 (Rd ) −→ 0,

and  
E kvG,Z (t)k2L2 (Rd ) = IG,Z ≤ I˜G,Z .
We use the following notation for the stochastic integral that we have just defined:
Z
vG,Z (t) = G(s, · − y) Z(s, y) M (ds, dy).
[0,t]×Rd

For t fixed, vG,Z (t) ∈ L2 (Rd ) is a square-integrable function that is defined almost-
everywhere.
The definition of the stochastic integral requires in particular that hypothesis (H2)
be satisfied. In the case where

Γ(dx) = kβ (x) dx, with kβ (x) = |x|−β , β > 0, (5.4)

this condition becomes

Z T Z
β−d sin2 (s|ξ − η|)
ds sup dη |η| < +∞.
0 ξ∈Rd Rd |ξ − η|2

One checks [4] that this is the case if and only if 0 < β < 2.

6 The wave equation in spatial dimension d ≥ 1

We consider the following stochastic wave equation in spatial dimension d ≥ 1, driven
by spatially homogeneous Gaussian noise Ḟ (t, x) as defined in Section 3:
  2 
∂ u
 ∂t2 − ∆u (t, x) = σ(x, u(t, x)) Ḟ (t, x),

 t ∈ ]0, T ], x ∈ Rd ,

(6.1)
 u(0, x) = v (x), ∂u
(0, x) = ṽ0 (x),

0 ∂t

18
where v0 ∈ L2 (Rd ) and ṽ0 ∈ H −1 (Rd ). By definition, H −1 (Rd ) is the set of square-
integrable functions ṽ0 such that
def
Z
1
kṽ0 k2H −1 (Rd ) = dξ |F ṽ0 (ξ)|2 < +∞.
Rd 1 + |ξ|2
We shall restrict ourselves, though this is not really necessary (see [4]) to the case
where Γ(dx) is as in (5.4), and 0 < β < 2.

The past-light cone property

Consider a bounded domain D ⊂ Rd . A fundamental property of the wave equa-

tion is that u(T, x), x ∈ D, only depends on v0 |K D and ṽ0 |K D , where

K D = { y ∈ Rd : d(y, D) ≤ T }

and d(y, D) denotes the distance from y to the set D, and on the noise Ḟ (s, y) for
y ∈ K D(s) , 0 ≤ s ≤ T , where
n o
K D (s) = y ∈ Rd : d(y, D) ≤ T − s .

Therefore, the solution u(t, x) in D is unchanged if we take the s.p.d.e.

∂2u
!
− ∆u (t, x) = 1K d (t) (x) σ(x, u(t, x)) Ḟ (t, x).
∂t2

We shall make the following linear growth and Lipschitz continuity assumptions
on the function σ.

Assumptions
(a) |σ(x, u)| ≤ c(1 + |u|) 1K D (x) , for all x ∈ Rd and u ∈ R;

(b) |σ(x, u) − σ(x, v)| ≤ c |u − v|, for all x ∈ Rd and u, v ∈ R.

Definition 6.1 An adapted and mean-square continuous L2 (Rd )-valued process (u(t), 0 ≤
t ≤ T ) is a solution of (6.1) in D if for all t ∈ ]0, T ],

d
u(t) 1K D (t) = 1K D (t) G(t) ∗ v0 + G(t) ∗ ṽ0
dt
Z !
+ G(t − s, · − y) σ(y, u(s, y)) M (ds, dy) .
[0,t]×Rd

Theorem 6.2 Let d ≥ 1. Suppose 0 < β < 2 ∧ d and that the assumptions above on
σ are satisfied. Then (6.1) has a unique solution (u(t), 0 ≤ t ≤ T ) in D.

19
Proof. We use a Picard iteration scheme. Set
d
u0 (t, x) = G(t) ∗ v0 + G(t) ∗ ṽ0 .
dt
We first check that u0 (t) ∈ L2 (Rd ). Indeed,
!
d d
G(t) ∗ v0 = F G(t) · Fv0
dt L2 (Rd )
dt L2 (Rd )
Z
cos(t|ξ|) 2
= dξ |ξ| · Fv0 (ξ)
Rd |ξ|
≤ kv0 kL2 (Rd ) ,

and, similarly,
kG(t) ∗ ṽ0 kL2 (Rd ) ≤ kṽ0 kH −1 (Rd ) .
One checks in a similar way that t 7→ u0 (t) from [0, T ] into L2 (Rd ) is continuous.
We now define the Picard iteration scheme. For n ≥ 0, assume that (un (t), 0 ≤
t ≤ T ) has been defined, and satisfies (H1). Set

un+1 (t) = 1K D (t) (u0 (t) + vn+1 (t)) , (6.2)

where Z
vn+1 (t) = G(t − s, · − y) σ(y, un (s, y)) M (ds, dy). (6.3)
[0,t]×Rd

By induction, Zn (s, y) = σ(y, un (s, y)) satisfies (H1). Indeed, this process is
adapted, and since

kσ(·, un (s, ·)) − σ(·, un (t, ·))kL2 (Rd ) ≤ Ckun (s, ·) − un (t, ·)kL2 (Rd ) ,

it follows that s 7→ un (s, ·) is mean-square continuous. One checks that un+1 also
satisfies (H1): this uses assumption (a).
Therefore, the stochastic integral (6.3) is well-defined. Let
 
Mn (r) = sup E kun+1 (t) − un (t)k2L2 (K D (t))
0≤t≤r
 
= sup E kvn+1 (t) − vn (t)k2L2 (K D (t))
0≤t≤r

Z
= sup E G(t − s, · − y)
0≤t≤r [0,t]×Rd
!
2
× (σ(y, un (s, y)) − σ(y, un−1 (s, y))) M (ds, dy)
L2 (K D (t))
Z t  
≤ sup ds E kσ(·, un (s, ·)) − σ(·, un−1 (s, ·))k2L2 (K D (t)) J(t − s),
0≤t≤r 0

20
where Z
sin2 (s|ξ − η|)
J(s) = sup dη |η|β−d .
ξ∈Rd Rd |ξ − η|2
A direct calculation shows that

sup J(s) < +∞,

0≤s≤T

since 0 < β < 2, so

Z t  
Mn (r) ≤ C sup ds E kun (s, ·) − un−1 (s, ·)k2L2 (K D (t)) ,
0≤t≤r 0

that is, Z r
Mn (r) ≤ C Mn−1 (s) ds.
0

Because M0 (T ) < +∞, Gronwall’s lemma implies that

+∞
(Mn (r))1/2 < +∞.
X

n=0

Therefore, (un (t, ·), n ∈ N) converges in L2 (Ω × Rd , dP × dx), uniformly in t ∈ [0, T ],

to a limit u(t, ·). Since un satisfies (H1) and un converges uniformly in t to u(t, ·),
it follows that u(t, ·) is a solution to (6.1): indeed, it suffices to pass to the limit in
(6.2) and (6.3).
Uniqueness of the solution follows by a standard argument. 2

7 Spatial regularity of the stochastic integral (d =

3)
We aim now to analyze spatial regularity of the solution to the 3-dimensional stochas-
tic wave equation (6.1) driven by spatially homogeneous Gaussian noise, with covari-
ance given by a Riesz kernel f (x) = |x|−β , where 0 < β < 2. For this, we shall first
examine the regularity in the x-variable of the function-valued stochastic integral
defined in Section 5 when d = 3.
We recall that studying regularity properties requires information on higher mo-
ments. With these, one can use the Kolmogorov continuity theorem (Theorem 4.3)
or the Sobolev embedding theorem, which we now recall.

Theorem 7.1 (The Sobolev Embedding Theorem). Suppose that g ∈ W p,q (O).
Then x 7→ g(x) is ρ̃-Hölder continuous, for all ρ̃ ∈ ]0, ρ − dq [ .

21
 We recall [14] that the norm in the space W p,q (O) is defined by

kgkqW p,q (O) = kgkqLq (O) + kgkqp,q,O , (7.1)

where
Z
kgkqLq (O) = |g(x)|q dx
O
Z Z
|g(x) − g(y)|q
kgkqp,q,O = dx dy .
O O |x − y|d+ρq
Our first objective is to determine conditions that ensure that
 
E kvG,Z kqLq (O) < +∞.

For ε > 0, we let

n o
Oε = x ∈ R3 : ∃ z ∈ O with |x − z| < ε

denote the ε-enlargement of O, and use the notation

Z
t
vG,Z = G(t − s, · − y) Z(s, y) M (ds, dy).
[0,t]×R3

An estimate in Lp -norm

Theorem 7.2 Suppose 0 < β < 2. Fix T > 0, q ∈ [2, +∞[ and let O ⊂ R3 be a
bounded domain. Suppose that
Z t  
ds E kZ(s)kqLq (Ot−s ) < +∞.
0

Then   Z t  
E t
kvG,Z kqLq (O) ≤C ds E kZ(s)kqLq (Ot−s ) .
0

Proof. We present the main ideas, omitting some technical issues that are handled
in [5, Proposition 3.4]. First, we check inequality with G replaced by Gn :
 
t
E kvG kq
n ,Z Lq (O)
Z Z !
q
= dx E Gn (t − s, x − y) Z(s, y) M (ds, dy)
O [0,t]×R3
Z  Z t Z Z
≤ dx E ds dy dz Gn (t − s, x − y) Z(s, y) f (y − z)
O 0 R3 R3

q/2
×Gn (t − s, x − z) Z(s, z) .

22
Let
Z t Z Z
µn (t, x) = ds dy dz Gn (t − s, x − y) f (y − z) Gn (t − s, x − z).
0 R3 R3

Assume that
sup µn (t, x) < +∞. (7.2)
n, x, t≤T

By Hölder’s inequality, written in the form (4.8), we see, since Gn ≥ 0, that

  Z
q
Z t Z Z
E t
kvG kq
n ,Z Lq (O)
≤ dx (µn (t, x)) 2
−1
E ds dy dz Gn (t − s, x − y)
O 0 R3 R3
!
q/2 q/2
×f (y − z) Gn (t − s, x − z) |Z(s, y)| |Z(s, z)|

= IG , |Z 1Ot−s+1/n |
q/2 .
n

We apply (5.2), then (5.3), to bound this by

Z t  
I˜G , |Z 1Ot−s+1/n |
q/2 = ds E k |Z(s)|q/2 1Ot−s+1/n k2L2 (R3 )
n 0
Z
× sup µ(dη) |FGn (s, ·)(ξ − η)|2 .
ξ∈R3 R3

Since 0 < β < 2, the supremum over ξ is finite, therefore

  Z t  
E t
kvG kq
n ,Z Lq (O)
≤C ds E kZ(s)kqLq (Ot−s+1/n ) .
0

By Fatou’s lemma,
   
t
E kvG,Z kqLq (O) t
≤ lim inf E kvG n
kq q
,Z L (O)
k→∞ k
Z t  
≤ lim inf ds E kZ(s)kqLq (Ot−s+1/nk )
k→∞ 0
Z t  
= ds E kZ(s)kqLq (Ot−s ) .
0

It remains to check that (7.2) holds. Since

sin2 ((t − s)|η|)

|FGn (t − s)(η)|2 ≤ |FG(t − s)(η)|2 = ,
|η|2

it follows that for t ∈ [0, T ] and x ∈ R3 ,

Z t Z
sin2 ((t − s)|η|)
µn (t, x) ≤ ds dη |η|β−3 ≤ C(T ),
0 R3 |η|2

23
since 0 < β < 2. This completes the proof. 2

An estimate in Sobolev norm

We consider here a spatially homogeneous Gaussian noise Ḟ (t, x), with covariance
given by f (x) = |x|−β , where 0 < β < 2. We seek an estimate of the Sobolev norm
t
of the stochastic integral vG,Z . We recall that the Sobolev norm is defined in (7.1).
Theorem 7.3 Suppose 0 < β < 2. Fix T > 0, q ∈ ]3, +∞[, and let O ⊂ R3 be a
bounded domain. Fix γ ∈ ]0, 1[, and suppose that
Z t  
ds E kZ(s)kqW γ,q (Ot−s ) < +∞.
0

Consider !
2−β 3 h
i
ρ ∈ 0, γ ∧ − .
2 q
Then there exists C < +∞ (depending on ρ but not on Z) such that
  Z t  
E t
kvG,Z kqρ,q,O ≤C ds E kZ(s)kqW ρ,q (Ot−s ) .
0

Remark 7.4 In the case of the heat equation, spatial regularity of the stochastic
t
integral process, that is, of x 7→ vG,Z (x), occurs because of regularity of the heat
t
kernel G, even if Z is merely integrable. Here, the spatial regularity of vG,Z is due to
the regularity of Z.
Proof of Theorem 7.3. The key quantity that we need to estimate is
t t
(y)|q
!
Z Z
|vG,Z (x) − vG,Z
E dx dy .
O O |x − y|3+ρq
Let ρ̄ = ρ + 3q , so that 3 + ρ q = ρ̄ q. If we replace G by Gn , then the numerator above
is equal to
Z t Z q
ds (Gn (t − s, x − u) − Gn (t − s, y − u)) Z(s, u) M (ds, du) ,
0 R3

so by Burkholder’s inequality (2.3),

 
t t
E |vG n ,Z
(x) − vG n ,Z
(y)|q
 Z t Z Z
≤CE ds du dv Z(s, u) f (u − v) Z(s, v) (7.3)
0 R3 R3
× (Gn (t − s, x − u) − Gn (t − s, y − u))
q/2 
×(Gn (t − s, x − v) − Gn (t − s, y − v)) .

24
If we had G instead of Gn , and if G were smooth, then we would get a bound involving
an exponent of |x − y|, even if Z were merely integrable.
Here we use a different idea: we shall pass the increments on the Gn over to the
factors Z f Z by changing variables. For instance, if there were only one factor involv-
ing increments of Gn , we could use the following calculation, where Gn is generically
denoted g and Z f Z is denoted ψ:
Z
du (g(x − u) − g(y − u)) ψ(u)
R3
Z Z
= du g(x − u) ψ(u) − du g(y − u) ψ(u)
R3 R3
Z Z !
= dũ g(ũ) ψ(x − ũ) − dũ g(ũ) ψ(y − ũ)
R3 R3
Z
= dũ g(ũ) (ψ(x − ũ) − ψ(y − ũ)).
R3

Using this idea, it turns out that the integral on the right-hand side of (7.3) is equal
to
4
t
X
Ji,n (x, y),
i=1
where
Z t Z Z
t
Ji,n (x, y) = ds du dv Gn (s, u) Gn (s, v) hi (t, s, x, y, u, v),
0 R3 R3

and
h1 (t, s, x, y, u, v) = f (y − x + v − u) (Z(t − s, x − u) − Z(t − s, y − u))
×(Z(t − s, x − v) − Z(t − s, y − v),

h2 (t, s, x, y, u, v) = Df (v − u, x − y)Z(t − s, x − u)
×(Z(t − s, x − v) − Z(t − s, y − v)),

h3 (t, s, x, y, u, v) = Df (v − u, y − x)Z(t − s, y − v)
×(Z(t − s, x − u) − Z(t − s, y − u)),

h4 (t, s, x, y, u, v) = −D2 f (v − u, x − y) Z(t − s, x − u) Z(t − s, x − u),

and we use the notation
Df (u, x) = f (u + x) − f (u),
D2 f (u, x) = f (u − x) − 2 f (u) + f (u + x).
We can now estimate separately each of the four terms
t
Z Z
E(|Ji,n (x, y)|q/2 )
Tni (t, O) = dx dy , i = 1, . . . , 4.
O O |x − y|ρ̄ q

25
The term Tn1 (t, O). Set
Z Z
µn (x, y) = sup du dv Gn (s, u) Gn (s, v) f (y − x + v − u)
s∈[0,T ] R3 R3
Z
= sup µ(dη) ei η·(x−y) |FGn (s)(η)|2 ,
s∈[0,T ] R3

so that
sup µn (x, y) < +∞
n,x,y

since β < 2. Therefore, since Gn (s, u) ≥ 0, by Hölder’s inequality,

  q
t
E |J1,n (x, y)|q/2 ≤ (T µn (x, y)) 2 −1
Z t Z Z
×E ds du dv Gn (s, u) Gn (s, v) f (y − x + v − u)
0 R3 R3

×|Z(t − s, x − u) − Z(t − s, y − u)|q/2

!
q/2
×|Z(t − s, x − v) − Z(t − s, y − v)| .

Apply the Cauchy-Schwarz inequality with respect to the measure

dP dx dy ds du dv Gn (s, u) Gn (s, v) f (y − x + v − u)

to see that  1/2

Tn1 (t, O) ≤ Tn1,1 (t, O) Tn1,2 (t, O) ,
where
Z t Z Z Z Z
Tn1,1 (t, O) = ds dx dy du dv Gn (s, u) Gn (s, v) f (y − x + v − u)
0 O O R3 R3

E (|Z(t − s, x − u) − Z(t − s, y − u)|q )

× ,
|x − y|ρ̄q

and there is an analogous expression for Tn1,2 (t, O). We note that for x ∈ O, when
Gn (s, u) > 0 (resp. for y ∈ O, when Gn (s, v) > 0), x − u ∈ Os(1+1/n) (resp. y − u ∈
Os(1+1/n) ), so
Z t  
Tn1,1 (t, O) ≤ ds E kZ(t − s)kqρ,q,Os(1+1/n) sup µn (x, y).
0 n,x,y

The same bound arises for the term Tn1,2 (t, O), so this gives the desired estimate for
this term.
We shall not discuss the terms Tn2 (t, O) and Tn3 (t, O) here: the interested reader
may consult [5, Chapter 3], but we consider the term Tn4 (t, O).

26
The term Tn4 (t, O). In order to bound Tn4 (t, O), we aim to bring the exponent q/2
inside the ds du dv integral, in such a way that it only affects the Z factors but not f .
Set
Z Z
|D2 f (v − u, x − y)|
µn (x, y) = sup du dv Gn (s, u) Gn (s, v) . (7.4)
s∈[0,T ] R3 R3 |x − y|2ρ̄
We will show below that
sup µn (x, y) ≤ C < +∞, (7.5)
n≥1, x,y∈O

which will turn out to require quite an interesting calculation. Assuming this for the
moment, let p = q/2. Then, by Hölder’s inequality,
 
t
E |J4,n (x, y)|p
|x − y|2pρ̄
Z t Z Z
≤ sup (µn (x, y))p−1 ds du dv Gn (s, u) Gn (s, v)
n,x,y 0 R3 R3
|D2 f (v − u, x − y)|
× E (|Z(t − s, x − u)|p |Z(t − s, x − v)|p ) .
|x − y|2ρ̄
This quantity must be integrated over O × O. We apply the Cauchy-Schwarz inequal-
ity to the measure ds du dv(· · · )dP , and this leads to
Z t  
p
Tn4 (t, O) ≤ sup (µn (x, y)) ds E kZ(s)kqLq (O(t−s)(1+1/n) ) .
n,x,y 0

This is the desired bound for this term.

It remains to check (7.5). The main difficulty is to bound the second order differ-
ence |D2 f (v − u, x − y)|. We explain the main issues below.

Bounding symmetric differences

Let g : R → R. Suppose that we seek a bound on

D2 g(x, h) = g(x − h) − 2g(x) + g(x + h).
Notice that if g is differentiable only once (g ∈ C 1 ), then the best that we can do is
essentially to write
|D2 g(x, h)| ≤ |g(x − h) − g(x)| + |g(x + h) − g(x)|
≤ c h.
On the other hand, if g is twice differentiable (g ∈ C 2 ), then we can do better:
|D2 g(x, h)| ≤ c h2 .

27
 In the case of a Riesz kernel f (x) = |x|−β , x ∈ R3 , we can write

|D2 f (u, x)| = |u − x|−β − 2|u|−β + |u + x|−β

00
≤ C |f (u)| |x|2
= C |u|−β−2 |x|2 .

Taking into account the definition of µn (x, y) in (7.4), this inequality leads to the
bound
|x − y|2
Z Z 
µn (x, y) ≤ sup du dv Gn (s, u) Gn (s, v) |u − v|−(β+2) .
s∈[0,T ] R3 R3 |x − y|2ρ̄

However, the double integral converges to +∞ as n → ∞, since β + 2 > 2.

Since differentiating once does not necessarily give the best bound possible and
differentiating twice gives a better exponent but with an infinite constant, it is natural
to want to differentiate a fractional number of times, namely just under 2 − β times.
If we “differentiate α times” and all goes well, then this should give a bound of the
form µn (x, y) ≤ C|x − y|α , for α ∈ ]0, 2 − β[. We shall make this precise below.

Riesz potentials, their fractional integrals and Laplacians

def
Let α = 2ρ̄. We recall that
2−β 3 3
ρ< − and ρ̄ = ρ + , so α < 2 − β.
2 q q
The Riesz potential of a function ϕ : Rd → R is defined by
1 Z ϕ(y)
(Ia ϕ)(x) = dy, a ∈ ]0, d[,
γ(a) Rd |x − y|d−a

where γ(a) = π d/2 2a Γ(a/2)/Γ( d−a

2
). Riesz potentials have many interesting properties
(see [15]), of which we mention the following.
(1) Ia+b (ϕ) = Ia (Ib ϕ) if a + b ∈ ]0, d[. Further, Ia can be seen as a “fractional
integral of order a”, in the sense that
!
1 Fϕ(ξ)
F(Ia ϕ)(ξ) = Fϕ(ξ) F (ξ) = .
| · |d−a |ξ|a

(2) Our covariance function kβ (x) = |x|−β is a Riesz kernel. These kernels have
the following property:
Z
|x|−d+a+b = dz kd−b (x − z) |z|−d+a (7.6)
Rd
 
= Ib | · |−d+a . (7.7)

28
This equality can be viewed as saying that |z|−d+a is “bth derivative (or Laplacian)”
of |z|−d+a+b , in the sense that
 
(−∆)b/2 |z|−d+a+b = |z|−d+a .

Indeed, taking Fourier transforms, this equality becomes simply

|ξ|b |ξ|−a−b = |ξ|−a .

Recall the notation

Df (u, y) = f (u + y) − f (u).
From (7.6), one can easily deduce (see [5, Lemma 2.6]) that
Z
b
Dkd−a−b (u, cx) = |c| dw kd−a (u − cw) Dkd−b (w, x)
Rd

and !
2 b
Z
2 x
|D kd−a−b (u, x)| ≤ |x| dw kd−a (u − |x|w) D kd−b w, .
Rd kxk
Set b = α = 2ρ̄ and a = 3 − α − β, where α + β ∈ ]0, 2[ . Looking back to (7.4),
these two relations lead to the following estimate:
1 Z Z
µn (x, y) ≤ sup du dv Gn (s, u) Gn (s, v) |x − y|α
s∈ [0,T ] |x − y|α R3 R3
!
Z
2 x
× dw kα+β (v − u − |y − x|w) × D k3−α w,
R3 |x|
Z Z !
≤ sup sup du dv Gn (s, u) Gn (s, v) kα+β (v − u − |y − x|w)
s∈ [0,T ] x,y,w R3 R3
!
Z
2 x
× sup dw D k3−α w, .
x kxk

The double integral above is finite since α+β < 2. Indeed, taking Fourier transforms,
the shift −|y − x|w introduces a factor eiη·|y−x|w , which is of no consequence. The
second integral is finite (and does not depend on x). For this calculation, see [5,
Lemma 2.6]. 2

8 Hölder-continuity in the 3-d wave equation

We consider the stochastic wave equation (6.1) for d = 3, driven by spatially homo-
geneous Gaussian noise with covariance f (x) = |x|−β , where 0 < β < 2.
The main idea for checking Hölder continuity of the solution is to go back to
the Picard iteration scheme that was used to construct the solution, starting with

29
a smooth function u0 (t, x) (whose smoothness depends only on the regularity of the
initial conditions), and then check that regularity is preserved at each iteration step
and passes to the limit. The details are carried out in [5, Chapter 4]. The main result
is the following.

Theorem 8.1 Assume the following three properties:

(a) the initial value v0 is such that v0 ∈ C 2 (R3 ) and ∆v0 is Hölder continuous with
exponent γ1 ;

(b) the initial velocity ṽ0 is Hölder continuous with exponent γ2 ;

(c) the nonlinearities σ, b : R → R are Lipschitz continuous.

Then, for any q ∈ [2, ∞[ and

2−β
α ∈ ]0, γ1 ∧ γ2 ∧ [, (8.1)
2
there is C > 0 such that for all (t, x), (s, y) ∈ [0, T ] × D,

E (|u(t, x) − u(s, y)|q ) ≤ C (|t − s| + |x − y|)αq .

In particular, (t, x) 7→ u(t, x) has a Hölder continuous version with exponent α.

We observe that the presence of γ1 ∧ γ2 in (8.1) can be interpreted by saying that

the (ir)regularity of the initial conditions limits the possible regularity of the solution:
there is no smoothing effect in the wave equation, contrary to the heat equation.
We note that this result is sharp. Indeed, if we consider the linear wave equation,
in which we take σ ≡ 1 and b ≡ 0 in (6.1), with vanishing initial condition v0 ≡ ṽ0 ≡ 0,
then it is possible to show (see [5, Chapter 5] that
 
E |u(t, x) − u(t, y)|2 ≥ c1 |x − y|2−β

and  
E |u(t, x) − u(s, x)|2 ≥ c2 |t − s|2−β .
This implies in particular that t 7→ u(t, x) and x 7→ u(t, x) are not γ-Hölder continu-
ous, for γ > 2−β
2
.

References
[1] R.C. Dalang: Extending the martingale measure stochastic integral with applica-
tions to spatially homogeneous spde’s. Electronic J. of Probability, Vol 4, 1999.

30
 [2] R.C. Dalang, N.E. Frangos: The stochastic wave equation in two spatial dimen-
sions. Annals of Probab. 26, 1, 187-212, 1998.

[3] R.C. Dalang, O. Lévèque: Second-order linear hyperbolic SPDEs driven by

isotropic Gaussian noise on a sphere. Annals of Probab. 32, 1068–1099, 2004.

[4] R.C. Dalang, C. Mueller: Some non-linear SPDE’s that are second order in time.
Electronic J. of Probability, Vol 8, 1, 1-21, 2003.

[5] R.C. Dalang, M. Sanz-Solé: Hölder-Sobolev regularity of the solution to the

stochastic wave equation in dimension 3. Memoirs of the AMS (2007, to appear).

[6] G.B. Folland: Introduction to Partial Differential Equations. Princeton Univ.

Press, 1976.

[7] O. Gonzalez, J.H. Maddocks: Extracting parameters for base-pair level models
of DNA from molecular dynamics simulations. Theoretical Chemistry Accounts,
106, 76-82, 2001.

[8] A. Hellemans: SOHO Probes Sun’s Interior by Tuning In to Its Vibrations. Sci-
ence May 31 1996: 1264-1265.

[9] D. Khoshnevisan: A Primer on Stochastic Partial Differential Equations. In:

this volume, 2007.

[10] A. Karkzewska, J. Zabczyk: Stochastic PDE’s with function-valued solutions.

In: Infinite-dimensional stochastic analysis (Clément Ph., den Hollander F., van
Neerven J. & de Pagter B., eds), pp. 197-216, Proceedings of the Colloquium of
the Royal Netherlands Academy of Arts and Sciences, 1999, Amsterdam.

[11] M. Oberguggenberger & F. Russo: Nonlinear stochastic wave equations. In:

Generalized functions—linear and nonlinear problems (Novi Sad, 1996). Inte-
gral Transform. Spec. Funct. 6, no. 1-4, 71–83, 1998.

[12] M. Sanz-Solé, M. Sarrà: Hölder continuity for the stochastic heat equation with
spatially correlated noise. In: Stochastic analysis, random fields and applications
(R.C. Dalang, M. Dozzi & F. Russo, eds), pp. 259-268, Progress in Probability
52, Birkhäuser, Basel, 2002.

[13] L. Schwartz: Théorie des distributions. Hermann, Paris, 1966.

[14] N. Shimakura: Partial differential operators of elliptic type. Translations of

Mathematical Monographs, 99. American Mathematical Society, 1992.

[15] E. M. Stein: Singular Integrals and Differentiability Properties of Functions.

Princeton University Press, Princeton, 1970.

31
[16] F. Treves: Basic Linear Partial Differential Equations. Academic Press, 1975.

[17] J.B. Walsh: An introduction to stochastic partial differential equations, École

d’été de Probabilités de Saint Flour XIV, Lecture Notes in Mathematics, Vol.
1180, Springer Verlag, 1986.

[18] J. Zabczyk: A mini course on stochastic partial differential equations. In:

Stochastic climate models (P. Imkeller & J.-S. von Storch, eds), pp. 257–284,
Progr. Probab., 49, Birkhuser, 2001.

32