Submitted to Bernoulli

Fractional stochastic wave equation driven by

a Gaussian noise rough in space
JIAN SONG1 , XIAOMING SONG2 and FANGJUN XU3,4
1
School of Mathematics, Shandong University, Jinan, Shandong 250100, China.
E-mail: [email protected]
2
Department of Mathematics, Drexel University, Philadelphia, PA 19104, USA.
E-mail: [email protected]
3
Key Laboratory of Advanced Theory and Application in Statistics and Data Science - MOE,
School of Statistics, East China Normal University, Shanghai 200062, China.
E-mail: [email protected]
4
NYU-ECNU Institute of Mathematical Sciences, NYU Shanghai, Shanghai 200062, China.
E-mail: [email protected]

MSC 2010: 60H07; 60H15; 60G15.

In this article, we consider fractional stochastic wave equations on R driven by a multiplicative

Gaussian noise which is white/colored in time and has the covariance of a fractional Brownian
motion with Hurst parameter H ∈ ( 41 , 12 ) in space. We prove the existence and uniqueness of the
mild Skorohod solution, establish lower and upper bounds for the p-th moment of the solution
for all p ≥ 2, and obtain the Hölder continuity in time and space variables for the solution.

Keywords: Fractional Brownian motion, Malliavin calculus, Skorohod integral, stochastic wave
equation, intermittency, Hölder continuity.

1. Introduction
Consider the following fractional stochastic wave equation (SWE) on R
 2
∂ u κ
(t, x) = −(−∆) 2 u(t, x) + uẆ (t, x), t > 0
∂t2 (1.1)
u(0, x) = 1, ∂u (0, x) = 0,
∂t

with κ ∈ (0, 2], where Ẇ (t, x) is a Gaussian noise with covariance

E[Ẇ (t, x)Ẇ (s, y)] = f0 (t − s)f (x − y).

In this article, we assume that the noise is rough in space, i.e., f (x) = (|x|2H )00 with H ∈
( 41 , 21 ), where (|x|2H )00 means the second derivative of |x|2H in the sense of distribution.
Note that for fractional Brownian motion B H with Hurst parameter H ∈ (0, 1), its
1
imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019
2 J. Song, X. Song and F. Xu

derivative (in the sense of distribution) Ḃ H has the covariance E[Ḃ H (t)Ḃ H (s)] = H(2H −
1)f (t − s). We also assume that the temporal covariance function f0 (t) is either the
Dirac delta function δ(t) or a nonnegative and nonnegative-definite function such that
f0 (t) ∼ |t|2H0 −2 with H0 ∈ ( 21 , 1), i.e., c|t|2H0 −2 ≤ f0 (t) ≤ C|t|2H0 −2 for some constants
0 < c < C < ∞.
The Itô-type probabilistic approach for stochastic partial differential equations (SPDEs)
was established in [43], where Walsh introduced martingale measures and defined stochas-
tic integrals with respect to the martingale measures, and then SPDEs driven by space-
time white noise were investigated. Following Walsh’s approach, SWEs on Rd with d ≤ 2
were studied, for instance in [19, 36, 37]. In [18], Dalang extended Walsh’s stochastic
integral and applied it to solve SPDEs whose Green’s function is not a function but a
Schwartz distribution. In particular Dalang’s theory is applicable to SWEs in d-dimension
with d ≥ 3, and we refer to [15, 22, 26] and the references therein for the study of SWEs
in high dimensions.
For SPDEs driven by a multiplicative Gaussian noise which is colored in time (i.e,
the temporal covariance f0 is not the Dirac delta function), the probabilisitic approach
based on martingale properties cannot be applied directly since the noise does not have
martingale structure in time. An alternative approach is to apply Malliavin calculus to
study the chaos expansion of the Skorohod solution, see, for instance, [27, 29, 30, 41] for
stochastic heat equations (SHEs) and [1, 6] for SWEs.
In this article, we shall prove the existence and uniqueness of the mild Skorohod
solution to (1.1) (Theorem 3.2), establish lower and upper bounds for the p-th moment
of the solution for all p ≥ 2 (Proposition 4.1), and obtain the Hölder continuity for
the solution in time and space variables (Proposition 5.1). In the following, we briefly
describe some related recent development on SHEs and SWEs driven by multiplicative
Gaussian noise.
Hu and Nualart [29] investigated SHEs driven by a multiplicative fractional Brownian
sheet that is colored in time and white in space. Hu et al. [30] obtained Feynman-Kac
formulae for solutions of SHEs driven by a fractional Brownian sheet, and used them to
investigate the regularity of the solutions. The result in [30] then was extended to SHEs
driven by a general Gaussian noise in Hu et al. [27] and to SHEs with the Laplacian
operator being replaced by the infinitesimal generator of a symmetric Lévy process in
Song [41]. The noise considered in the papers [27, 29, 30, 41] is not “rough”, as its spatial
covariance corresponds to that of fractional Brownian motion with Hurst parameter
H ≥ 12 . The SHEs and SWEs on R driven by a Gaussian noise that is white in time and
rough in space were investigated in Hu et al. [25] and in Balan et al. [3], respectively.
Recently, Chen conducted a systematic investigation on SHEs with noise that is rough
in space and/or in time in [11, 12].
If the solution of a dynamic system with noise develops very high peaks, it is said that
the system possesses the intermittency property. The concept of intermittency arose in
physics, and in mathematics it is related to the long-term asymptotics of the moments
of the solution. The intermittency property was studied, for instance, in [2, 8, 9, 17, 23,
27, 40] for heat equations, and in [2, 16, 20] for wave equations. In particular, precise
long-term asymptotics for SHEs was obtained in [10, 13, 14, 31, 32], and the second order

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 3

Lyapunov exponent for SWEs was obtained in [7].

For the Hölder continuity of SHEs driven by multiplicative Gaussian noise colored in
time, we refer to [5, 27, 28, 41] and the references therein. For SWEs with noise white in
time, Hölder continuity of the solutions was studied in [19] for the spatial dimension d = 2,
in [22, 26] for d = 3, and in [15] for general dimensions; for SWEs with noise correlated
in time and space, Hölder continuity was established in [1, 6] for general dimensions.
Finally, we would like to make some comments on our results.
(a) Note that we require H > 14 for SWEs in Theorem 3.2, which was also assumed
in [31] for SHEs with rough spatial noise. Nevertheless, the approach used in the
proof of Theorem 3.2 can be also applied to SHEs and relax the condition H > 14
to H0 + H > 43 (see Remark 3.3).
(b) The rate of the bounds for the p-th moments obtained in Proposition 4.1 is consis-
tent with the known results in, for instance, [2, 7]. The lower bound is relatively
more difficult to establish. One of the obstacles is that the Fourier transform of the
Green’s function of the fractional wave equation is not a nonnegative function, and
this issue is resolved by showing that the integral of the Fourier transform of the
Green’s function is positive (see Lemma 6.5).
(c) The Hölder continuity obtained in Proposition 5.1 is consistent with the known
results (e.g., [1, Theorem 5.1], [6, Proposition 8.3], and [15, Theorem 7.6]) which
dealt with SWEs driven by the noise that is not rough in space. The major dif-
ference/difficulty of obtaining the Hölder continuity for SWEs with rough spa-
tial noise is the following. Denoting the spectral measure of the spatial covariance
µ(dξ) = fb(ξ)dξ, the condition
Z
1
sup µ(dξ) < ∞ (1.2)
η∈R R 1 + |ξ − η|2

plays a critical role in obtaining the Hölder continuity of the solution when the
spatial covariance f is a nonnegative, nonnegative definite, and locally integrable
function (see, e.g., [1, 6, 15]). However, when H < 21 , the spatial covariance (|x|2H )00
is a genuine distribution (see, e.g., [33]), and condition (1.2) is indeed violated (see
[3, Lemma A.1]).
This article is organized as follows. In Section 2, some preliminary results on Malli-
avin calculus associated with the noise W are provided. In Section 3, the existence and
uniqueness of the solution to (1.1) is obtained under proper conditions. In Section 4, we
derive the lower and upper bounds for the p-th moment of the solution for p ≥ 2 and
then deduce the weak intermittency. In Section 5, the Hölder continuity of the solution
in time and space is obtained. Finally, some lemmas used in the preceding sections are
gathered in Section 6.

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

4 J. Song, X. Song and F. Xu

2. Preliminaries
In this section, we recall some preliminaries on Malliavin calculus associated with the
Gaussian noise Ẇ . We refer to [38] for more details.
Let H be the completion of the Schwartz space S(R+ × R) under the inner product
Z Z
hϕ, φiH = CH f0 (r − s)ϕ(r,
b ξ)φ(s,
b ξ)µ(dξ)drds, (2.1)
R2+ R

where
Γ(2H + 1) sin(πH)
CH = (2.2)
2π
and µ(dξ) = |ξ|1−2H dξ with H ∈ (0, 21 ). Here, ϕb is the Fourier transform of ϕ in the
space, i.e., for ϕ ∈ S(R+ × R),
Z
ϕ(s,
b ξ) = e−iξx ϕ(s, x)dx.
R

In particular, if ϕ is a measurable function such that ϕ

b is also a measurable function
and Z Z
f0 (r − s)|ϕ(r, b ξ)|µ(dξ)drds < ∞,
b ξ)||ϕ(s,
R2+ R

then ϕ ∈ H. Note that H may contain distributions rather than just measurable functions
if f0 (r − s) ∼ |r − s|2H0 −2 for some H0 ∈ ( 12 , 1) (see [33, 39]).
In a complete probability space (Ω, F, P ), let W = {W (ϕ), ϕ ∈ H} be an isnormal
Gaussian process with the covariance
E[W (ϕ)W (φ)] = hϕ, φiH ,
and we also denote Z Z
W (ϕ) = ϕ(t, x)W (dt, dx).
R+ R

We also call W (ϕ) the Wiener integral of ϕ with respect to W . In light of [39, Theorem
3.1] and [33, Proposition 4.1], the Gaussian family {W (ϕ), ϕ ∈ H} coincides with the
linear expansion of the Gaussian family {W (t, x), (t, x) ∈ R+ × R)} with the covariance
1  2H Z tZ s
E[W (t, x)W (s, y)] = |x| + |y|2H − |x − y|2H f0 (r1 − r2 )dr1 dr2 ,
2 0 0

and in particular W (t, x) = W (I[0,t]×[0,x] ) with the convention I[0,t]×[x,0] = −I[0,t]×[0,x]

for x < 0.
For the smooth and cylindrical random variables of the form F = h(W (ϕ1 ), . . . , W (ϕn ))
with h being smooth and its partial derivatives having at most polynomial growth, the
Malliavin derivative DF of F is the H-valued random variable defined by
n
X ∂h
DF = (W (ϕ1 ), . . . , W (ϕn ))ϕk .
∂xk
k=1

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 5

Noting that D is closable from L2 (Ω) to L2 (Ω; H), we define the Sobolev space D1,2 as
the closure of the space of the smooth and cylindrical random variables under the norm

1
kDk1,2 = E[F 2 ] + E[kDF k2H ] 2 .
The divergence operator δ, which is also known as the Skorohod integral, is the adjoint
of the Malliavin derivative operator D defined by the duality
E[F δ(u)] = E[hDF, uiH ], ∀F ∈ D1,2 , ∀u ∈ Dom δ.
Here Dom δ is the domain of the divergence operator δ, which is the space of the H-
valued random variables u ∈ L2 (Ω; H) such that |E[hDF, uiH ]| ≤ cF kF k2 with some
constant cF depending on F , for all F ∈ D1,2 . Thus, for u ∈ Dom δ, δ(u) ∈ L2 (Ω). In
particular, E[δ(u)] = 0. We also use the following notation
Z Z
δ(u) = u(t, x)W (dt, dx), u ∈ Dom δ.
R+ R

Now we recall the Wiener chaos expansion. Let H0 = R, and for any integer n ≥ 1,
let Hn be the closed linear subspace of L2 (Ω) containing the set of random variables
{Hn (W (ϕ)), ϕ ∈ H, kϕkH = 1}, where Hn is the n-th Hermite polynomial, i.e., Hn (x) =
2
dn −x2
(−1)n ex dx n (e ). Then Hn is called the n-th Wiener chaos of W . Denoting by F the
σ-field generated by {W (ϕ), ϕ ∈ H}, then we have the following Wiener chaos decompo-
sition
L2 (Ω, F, P ) = ⊕∞
n=0 Hn .

For n ≥ 1, denote by H⊗n the n-th tensor product of H, and let H e ⊗n be the symmetriza-
⊗n ⊗n
tion of H . Then the mapping In (h ) = Hn (W (h)) for any h ∈ H can be extended
e ⊗n and the n-th Wiener chaos Hn . Thus, for any random
to a linear isometry between H
2
variable F ∈ L (Ω, F, P ), it has the following unique Wiener chaos expansion in the
sense of L2 (Ω),
∞
X
F = E[F ] + e ⊗n .
In (fn ) with fn ∈ H
n=1
Throughout the paper, the generic constant C varies at different places.

3. Existence and uniqueness of the solution

In this section, we obtain the existence and uniqueness of the mild Skorohod solution to
(1.1) under some conditions in Theorem 3.2, and we show that in Proposition 3.4 these
conditions are also necessary if the noise is white in time.
∂2 κ
d
Let Gt (x) be the fundamental solution of the equation ∂t 2 u + (−∆) u = 0 on R ,
2

then its Fourier transform in space Gb t (ξ) solves the following equation

∂2G
b t (ξ)
+ |ξ|κ G
b t (ξ) = 0,
∂t2

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

6 J. Song, X. Song and F. Xu

and it is given by (see [24, Section 2.2]; [18, Example 6] and [42, Chapter 1 Section 7] for
the case κ = 2)
κ/2
b t (ξ) = sin(t|ξ| ) .
G
|ξ|κ/2

Recall that when κ = 2, the Green’s function Gt (x) is a measurable function for d ≤ 2,
1
 I{|x|<t} , if d = 1,
2


Gt (x) =
 1 1
I{|x|<t} , if d = 2,


 p
2π t2 − |x|2
1
Gt (·) = 4πt σt for d = 3, where σt is the surface measure on the sphere {x ∈ R3 ; |x| = t},
and Gt (·) is a genuine distribution with compact support in Rd if d ≥ 3. Note that when
κ ∈ (1, 2) and d = 1, the Green’s function Gt (x) ∈ L2 (R) for all t ≥ 0 as its Fourier
κ/2
transform sin(t|ξ|
|ξ|κ/2
)
∈ L2 (R).
We consider the following filtration
Ft = σ{W (I[0,s] ϕ), 0 ≤ s ≤ t, ϕ ∈ S(R)} ∨ N ,
where N denotes the collection of null sets.
Definition 3.1 An adapted random field u = {u(t, x), t ≥ 0, x ∈ R} is a mild Skorohod
solution to (1.1) if E[u2 (t, x)] < ∞ for all (t, x) ∈ R+ × R and it satisfies the following
integral equation
Z tZ
u(t, x) = 1 + Gt−s (x − y)u(s, y)W (ds, dy), (3.1)
0 R

where the integral on the right-hand side is a Skorohod integral.

Note that if E[|u(t, x)|2 ] < ∞, the solution has a unique Wiener chaos expansion
∞
X
u(t, x) = In (gn (·, t, x))
n=0

e ⊗n . Now assume that u(t, x) is a mild Skorohod solution to (1.1).

with gn (·, t, x) ∈ H
Let Pn be the set of permutations on {1, 2, . . . , n}. Following the approach used in [29,
Section 4.1], we get
1
gn (s1 , . . . , sn , x1 , . . . , xn , t, x) = Gt−sρ(n) (x−xρ(n) ) · · · Gsρ(2) −sρ(1) (xρ(2) −xρ(1) ), (3.2)
n!
where ρ ∈ Pn is the permutation such that 0 < sρ(1) < sρ(2) < · · · < sρ(n) < t. Thus, to
prove the existence and uniqueness of the solution to (1.1) is equivalent to prove
∞
X
E[|u(t, x)|2 ] = n!kgn (·, t, x)k2H⊗n < ∞. (3.3)
n=0

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 7

Theorem 3.2 Assume that H0 ∈ [ 21 , 1), H ∈ ( 14 , 12 ) and κ ∈ (3 − 4H, 2]. Then there
exists a unique square integrable mild Skorohod solution to (1.1).
Proof It suffices to prove (3.3). We use the notation ξ := (ξ1 , . . . , ξn ) and similarly for
s, r and µ(dξ).
We first consider the case H0 ∈ ( 12 , 1). Since we assume that f0 (s) ∼ |s|2H0 −2 for
H0 ∈ ( 21 , 1), throughout the rest of the article, we will simply assume f0 (s) = |s|2H0 −2
in this case. Note that

n!kgn (·, t, x)k2H⊗n

Z Z n
Y
=n! Fgn (s, ·, t, x)(ξ)Fgn (r, ·, t, x)(ξ) |sj − rj |2H0 −2 dsdrµ(dξ) (3.4)
Rn [0,t]2n j=1

with
n
1 −ix(ξ1 +···+ξn ) Y sin((sρ(j+1) − sρ(j) )|ξρ(1) + · · · + ξρ(j) |κ/2 )
Fgn (s, ·, t, x)(ξ) = e ,
n! j=1
|ξρ(1) + · · · + ξρ(j) |κ/2

where we use the convention sρ(n+1) = t. Thus, by Lemma B.3 in [2] (see also [35]) and
a change of variables, we have

n!kgn (·, t, x)k2H⊗n

Z Z  2H1 2H0
2 0
≤n! |Fgn (s, ·, t, x)(ξ)| µ(dξ) ds
[0,t]n Rn
n 2H0
sin2 ((sρ(j+1) − sρ(j) )|ξρ(1) + · · · + ξρ(j) |κ/2 )  2H1
Z Z
1 Y 0
= µ(dξ) ds
n! [0,t]n Rn j=1 |ξρ(1) + · · · + ξρ(j) |κ
n 2H0
sin2 ((sj+1 − sj )|ξ1 + · · · + ξj |κ/2 )  2H1
Z Z Y
=(n!)2H0 −1
0
µ(dξ) ds
[0,t]n
< Rn j=1 |ξ1 + · · · + ξj |κ
n 2H0
sin2 ((sj+1 − sj )|ηj |κ/2 )  2H1
Z Z Y
=(n!)2H0 −1 1−2H 0
|η j − η j−1 | dη ds ,
[0,t]n
< Rn j=1 |ηj |κ
(3.5)

where [0, t]n< = [0 = s0 < s1 < · · · < sn < sn+1 = t].

Let An be a subset of the index set (α1 , . . . , αn ) ∈ {0, 1, 2}n such that
n n
α
Y X Y
x1 (xj + xj−1 ) = xj j .
j=2 α∈An j=1

Then #(An ) = 2n−1 , and for each α ∈ AnP , we have the following properties: α1 ∈
n
{1, 2}, αn ∈ {0, 1}, α2 , . . . , αn−1 ∈ {0, 1, 2}, j=1 αj = n and αj + αj+1 ∈ {1, 2, 3} for

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

8 J. Song, X. Song and F. Xu

j = 1, 2, . . . , n − 1. Hence, noting that |a + b|1−2H ≤ |a|1−2H + |b|1−2H , we get

n
Y n
Y n
X Y
1−2H 1−2H 1−2H
|ηj − ηj−1 | = |η1 | |ηj − ηj−1 | ≤ |ηj |(1−2H)αj . (3.6)
j=1 j=2 α∈An j=1

1 1
2H0
P P
Using (3.6) and the fact that ( xm ) 2H0 ≤ xm for all xm ≥ 0, the estimation (3.5)
now becomes

n!kgn (·, t, x)k2H⊗n

n 2H0
sin2 ((sj+1 − sj )|ηj |κ/2 )  2H1
Z Z Y
2H0 −1 1−2H 0
≤(n!) |ηj − ηj−1 | dη ds
[0,t]n
< Rn j=1 |ηj |κ
n 2H0
sin2 ((sj+1 − sj )|ηj |κ/2 )  2H1
Z  X Z Y
2H0 −1 αj (1−2H) 0
≤(n!) |ηj | dη ds
[0,t]n n |ηj |κ
< α∈An R j=1
n 2H0
sin2 ((sj+1 − sj )|ηj |κ/2 )  2H1
Z X Z Y
≤(n!)2H0 −1 αj (1−2H) 0
|η j | dη ds
[0,t]n Rn j=1 |ηj |κ
< α∈An
Z X  2  2Hn 0 Yn
1 2 2
=(n!)2H0 −1 (sj+1 − sj ) 2H0 [2− κ − κ αj (1−2H)]
[0,t]n
< α∈An
κ j=1
 2H1 0 !2H0
sin2 (η) 2 αj (1−2H)+ 2 −1
Z
|η| κ κ dη ds . (3.7)
R η2

It follows from Lemma 6.4 and the condition κ > 3 − 4H that for all αj ∈ {0, 1, 2}

sin2 (η) 2 αj (1−2H)+ 2 −1

Z
|η| κ κ dη < ∞.
R η2

Lemma 6.4 is applicable here, since the condition κ > 3−4H implies κ2 αj (1−2H)+ κ2 −3 ∈
(−3, −1) for all αj ∈ {0, 1, 2}.
Therefore, one can find a positive constant C depending only on (κ, H0 , H) such that

n!kgn (·, t, x)k2H⊗n

Z n 2H0
1 2 2
−κ
X Y [2− κ αj (1−2H)]
n 2H0 −1
≤C (n!) (sj+1 − sj ) 2H0
ds .
[0,t]n
< α∈An j=1

1
2 − κ2 − κ2 αj (1 − 2H) , j = 1, . . . , n, and
 
For each fixed α ∈ An , denote βj = 2H 0
Pn
β = j=1 βj = Hn0 1 − κ2 + 2H 1 1 1


κ . Note that H0 ∈ ( 2 , 1), H ∈ ( 4 , 2 ) and κ > 3 − 4H
implies that βj > 0 and β > 0. By Lemma 6.2 we have
n Qn
Γ(1 + βi )tn+β
Z Y
βj
(sj+1 − sj ) ds = i=1 .
[0,t]n
< j=1
Γ(n + 1 + β)

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 9

Therefore, from (6.1) in Lemma 6.3 with a = 1 + H10 1 − κ2 + 2H



κ and b = 1, and the
n−1
fact #(An ) = 2 , it follows that there exists some positive constant C such that,

n!kgn (·, t, x)k2H⊗n

2H0
tn+β

≤ C n (n!)2H0 −1
Γ(n + 1 + β)
2H0 −1 n(2H0 +2[(1− κ )+ 2H
κ ]) C n tn(2H0 +2[(1− κ )+ κ ])
2 2 2H
n
C (n!) t
= ∼ (3.8)
Γ(an + 1)2H0 (n!)2[(1− κ )+ κ ]+1 a2H0 an+H0 nH0 (1−a)
2 2H

when n → ∞. Notice that there exists λ > 1 such that λ−n ≤ a2H0 an+H0 nH0 (1−a) ≤ λn
for all n. Hence, by (6.2) in Lemma 6.3, there exists a positive constant C such that
∞ ∞
C n tn(2H0 +2[(1− κ )+ κ ])
2 2H
X X
E[|u(t, x)|2 ] = n!kgn (·, t, x)k2H⊗n ≤
(n!)2[(1− κ )+ κ ]+1
2 2H
n=0 n=0
 2κH0 +2(κ−2)+4H 
≤ C exp Ct 3κ−4+4H . (3.9)

Next, we consider the case H0 = 12 , i.e, f0 (t) = δ(t), and we have

n!kgn (·, t, x)k2H⊗n

Z Z
2
=n! |Fgn (s, ·, t, x)(ξ)| dsµ(dξ)
Rn [0,t]n
n n
sin2 ((sρ(j+1) − sρ(j) )|ξρ(1) + · · · + ξρ(j) |κ/2 ) Y 1−2H
Z Z
1 Y
= |ξ| dξds
n! [0,t]n Rn j=1 |ξρ(1) + · · · + ξρ(j) |κ j=1
n n
sin2 ((sj+1 − sj )|ξ1 + · · · + ξj |κ/2 ) Y 1−2H
Z Z
Y
= |ξ| dξds
[0,t]n
< Rn j=1 |ξ1 + · · · + ξj |κ j=1
n n
sin2 ((sj+1 − sj )|ηj |κ/2 ) Y
Z Z Y
= κ
|ηj − ηj−1 |1−2H dηds.
n n
[0,t]< R j=1 |η j | j=1

The last term in the above equation equals the right-hand side of (3.5) with H0 = 12 .
Analogue to the arguments in (3.6)-(3.9), we shall get the following estimation for the
second moment
∞
X
E[|u(t, x)|2 ] = n!kgn (·, t, x)k2H⊗n ≤ C exp (Ct) . (3.10)
n=0

We complete the proof.

Remark 3.3 For SHEs on R driven by a multiplicative Gaussian noise that is rough in
space, the existence and uniqueness of the mild Skorohod solution was obtained in [25] for
the noise white in time and in [31] for the noise colored in time. The condition H > 14

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

10 J. Song, X. Song and F. Xu

was assumed in both [25] and [31]. However, the method used in the proof of the above
theorem suggests that the condition can be reduced to H0 + H > 34 , and this is consistent
with the result in [11]. Indeed, for the following SHE on R,
 h
 ∂u
(t, x) = 12 ∆uh (t, x) + uh Ẇ (t, x), t > 0
∂t
uh (0, x) = 1,

|x|2
the Green’s function is the heat kernel Ght (x) = √ 1 e− 2t . Consequently, the Wiener
2πt
chaos expansion of the solution is
∞
X
uh (t, x) = In (gnh (·, t, x)),
n=0

where
1 h
gnh (s1 , . . . , sn , x1 , . . . , xn , t, x) = G (x − xρ(n) ) · · · Ghsρ(2) −sρ(1) (xρ(2) − xρ(1) )
n! t−sρ(n)
and
n  
1 −ix(ξ1 +···+ξn ) Y 1
Fgnh (s, ·, t, x)(ξ) = e exp − (sρ(j+1) − sρ(j) )|ξρ(1) + · · · + ξρ(j) |2 .
n! j=1
2

Now, the second moment of each chaos is

n!kgnh (·, t, x)k2H⊗n

Z Z n
Y
=n! h h
Fgn (s·, t, x)(ξ)Fgn (r, ·, t, x)(ξ) |sj − rj |2H0 −2 dsdrµ(dξ),
Rn [0,t]2n j=1
Z Z  2H1 2H0
|Fgnh (s, ·, t, x)(ξ)|2 µ(dξ)
0
≤n! ds
[0,t]n Rn
Z Z n  2H1 2H0
1 Y h i
exp − [sρ(j+1) − sρ(j) ]|ξρ(1) + · · · + ξρ(j) |2 µ(dξ)
0
= ds
n! [0,t]n Rn j=1
Z Z n
Y  2H1 2H0
2H0 −1 2 0
 
=(n!) exp − [sj+1 − sj ]|ξ1 + · · · + ξj | µ(dξ) ds
[0,t]n
< Rn j=1
Z Z n
Y h i  2H1 2H0
2H0 −1 2 1−2H 0
=(n!) exp − [sj+1 − sj ]|ηj | |ηj − ηj−1 | dη ds .
[0,t]n
< Rn j=1

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 11

Then, using similar argument as in the proof of Theorem 3.2 and Lemma 6.1, we have

n!kgnh (·, t, x)k2H⊗n

n 2H0
 (1−2H)αj  2H1 0
Z  X Z Y
2H0 −1 2

≤(n!) exp − (sj+1 − sj )|ηj | |ηj | dη ds
[0,t]n
< α∈An Rn j=1
 2H0
Z n
1
− 4H
X Y [1+(1−2H)αj ]
≤C n (n!)2H0 −1  (sj+1 − sj ) 0 ds .
[0,t]n
< α∈An j=1

1
For each fixed α ∈ An , denote βj = 4H0 [1 + (1 − 2H)αj ] ∈ (0, 1) noting that H0 ∈ ( 21 , 1),
Pn
H ∈ (0, 12 ) and H0 + H > 3/4, and then β = j=1 βj = n(1−H)2H0 . By Lemma 6.2 we have
n Qn
Γ(1 − βi )tn−β
Z Y
−βj i=1
(sj+1 − sj ) ds = .
[0,t]n
< j=1
Γ(n + 1 − β)

Therefore, since #(An ) = 2n−1 , there exists some positive constant C such that,

n!kgnh (·, t, x)k2H⊗n

2H0
tn−β

1
≤C n (n!)2H0 −1 = C n (n!)2H0 −1 (2H +H−1)n tn(2H0 +H−1) .
Γ(n + 1 − β) Γ( 0
+ 1) 2H 0
2H0

It follows from Lemma 6.3 and a similar argument in dealing with (3.8) that there exists
a positive constant C such that
∞
X  2H0 +H−1 
h 2
E[|u (t, x)| ] = n!kgnh (·, t, x)k2H⊗n ≤ C exp Ct H .
n=0

Proposition 3.4 When H0 = 21 and d = 1, the condition κ > 3 − 4H is also a necessary

condition for the existence of the square integrable solutions to (1.1). When H0 = 12 and
Hj < 12 , j = 1, · · · , d, the equation (1.1) has a solution only if d = 1.
Proof When d = 1, the L2 -norm of the second chaos of the solution is

kg2 (·, t, x)k2H⊗2

sin2 ((s2 − s1 )|η1 |κ/2 ) sin2 ((t − s2 )|η2 |κ/2 )
Z Z
=2 |η1 |1−2H |η2 − η1 |1−2H dηds
[0,t]2< R2 |η1 |κ |η2 |κ
sin2 ((s2 − s1 )|η1 |κ/2 ) sin2 ((t − s2 )|η2 |κ/2 )
Z Z
≥ |η1 |1−2H |η2 + η1 |1−2H dηds
[0,t]2< R2+ |η1 |κ |η2 |κ
sin2 ((s2 − s1 )|η1 |κ/2 ) sin2 ((t − s2 )|η2 |κ/2 )
Z Z
≥ |η1 |2(1−2H) dηds,
[0,t]2< R2+ |η1 |κ |η2 |κ

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

12 J. Song, X. Song and F. Xu

where the last integral is infinity if κ ≤ 3 − 4H due to Lemma 6.4.

For general dimension d, the above estimation becomes

kg2 (·, t, x)k2H⊗2

d
sin2 ((s2 − s1 )|η1 |κ/2 ) sin2 ((t − s2 )|η2 |κ/2 ) Y j 1−2Hj j
Z Z
=2 κ κ
|η1 | |η2 − η1j |1−2Hj dηds
2
[0,t]< R 2d |η1 | |η 2 | j=1
d
sin2 ((s2 − s1 )|η1 |κ/2 ) sin2 ((t − s2 )|η2 |κ/2 ) Y j 1−2Hj j
Z Z
≥ |η1 | |η2 + η1j |1−2Hj dηds
[0,t]2< R2d
+
|η1 |κ |η2 |κ j=1
d
sin2 ((s2 − s1 )|η1 |κ/2 ) sin2 ((t − s2 )|η2 |κ/2 ) Y j 2(1−2Hj )
Z Z
≥ |η1 | dηds.
[0,t]2< R2d
+
|η1 |κ |η2 |κ j=1

Now, by the change of variables

 1
η = r cos(θ1 )
 η12 = r sin(θ ) cos(θ )


1 2
 η13 = r sin(θ ) sin(θ ) cos(θ )



 1 1 2 3
..
 .


 η1d−1 = r sin(θ1 ) · · · sin(θd−2 ) cos(θd−1 )



 d
η1 = r sin(θ1 ) · · · sin(θd−2 ) sin(θd−1 ),

we have
d
d Z ∞
sin2 (|η1 |κ/2 ) Y j 2(1−2Hj )
P
2(1−2Hj )−κ+d−1
Z
κ
|η1 | dη1 =Cd sin2 (rκ/2 )rj=1 dr,
Rd
+
|η1 | j=1 0

d
P
which by Lemma 6.4 is infinite when d > 1 since 2(1 − 2Hj ) − κ + d − 1 ≥ −1 for
j=1
Hj ∈ (0, 12 ).

4. Moments of the solution and weak intermittency

In this section, we first obtain the lower bound and upper bound for the p-th moment of
the solution to (1.1) for p ≥ 2, and then deduce the weak intermittency.
Proposition 4.1 Under the conditions in Theorem 3.2, there exist 0 < C1 , C2 < ∞
such that for all p ≥ 2
2κH0 +2(κ−2)+4H 2κH0 +2(κ−2)+4H
   κ

C1 exp C1 t 3κ−4+4H ≤ ku(t, x)kp ≤ C2 exp C2 p 3κ−4+4H t 3κ−4+4H (4.1)

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 13

and
2κH0 +2(κ−2)+4H
C1 ≤ lim inf t− 3κ−4+4H log ku(t, x)kp
t→∞
2κH0 +2(κ−2)+4H κ
≤ lim sup t− 3κ−4+4H log ku(t, x)kp ≤ C2 p 3κ−4+4H . (4.2)
t→∞

2κH0 +2(κ−2)+4H
In particular, 3κ−4+4H = 1 if H0 = 12 .
Proof We shall prove (4.1) for H0 ∈ ( 12 , 1). The proof for H0 = 1
2 is similar and thus
omitted.
By (3.4), we have

n!kgn (·, t, x)k2H⊗n

Z Z n
Y
=n! Fgn (s, ·, t, x)(ξ)Fgn (r, ·, t, x)(ξ) |sj − rj |2H0 −2 dsdrµ(dξ)
Rn [0,t]2n j=1
n
sin((sj+1 − sj )|ξ1 + · · · + ξj |κ/2 )
Z Z Y
=n!
Rn ([0,t]n 2
< ) j=1
|ξ1 + · · · + ξj |κ/2
n n
Y sin((rj+1 − rj )|ξ1 + · · · + ξj |κ/2 ) Y
κ/2
|sj − rj |2H0 −2 dsdrµ(dξ)
j=1
|ξ 1 + · · · + ξ j | j=1
n
sin((sj+1 − sj )|ηj |κ/2 ) sin((rj+1 − rj )|ηj |κ/2 )
Z Z Y
=n!
Rn ([0,t]n 2
< ) j=1
|ηj |κ/2 |ηj |κ/2
n
! n
Y Y
2H0 −2
|sj − rj | dsdr |ηj − ηj−1 |1−2H dη
j=1 j=1
n
sin((sj+1 − sj )|ηj |κ/2 ) sin((rj+1 − rj )|ηj |κ/2 )
Z Z Y
≥n!
Dn ([0,t]n 2
< ) j=1
|ηj |κ/2 |ηj |κ/2
n
! n
Y Y
2H0 −2
|sj − rj | dsdr |ηj |1−2H dη,
j=1 j=1

noting that in the last step we used the facts that the inner integral with respect to dsdr
is nonnegative and that |ηj − ηj−1 |1−2H ≥ |ηj |1−2H on Dn with Dn = {(η1 , . . . , ηn ) ∈
Rn : η1 ≥ 0, η2 ≤ 0, η3 ≥ 0, η4 ≤ 0, . . . }.

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

14 J. Song, X. Song and F. Xu

Now, we have

n!kgn (·, t, x)k2H⊗n

Z Z Y n n
!
Y
κ/2 κ/2 1−2H−κ
≥ n! sin((sj+1 − sj )|ηj | ) sin((rj+1 − rj )|ηj | ) |ηj | dη
([0,t]n
<)
2 Dn j=1 j=1
n
Y
|sj − rj |2H0 −2 dsdr
j=1
Z n Z
!
Y
= n! sin((sj+1 − sj )|η|κ/2 ) sin((rj+1 − rj )|η|κ/2 )|η|1−2H−κ dη
([0,t]n
<)
2
j=1 R+

Yn
|sj − rj |2H0 −2 dsdr
j=1
 2
 n n
Z Y Z  Y
Y n
≥ n!tn(2H0 −2) κ/2
|ηj |1−2H−κ dη, (4.3)

 sin((sj+1 − sj )|η j | )ds
n
 n
R+ j=1 [0,t]< j=1
 j=1


where the last step holds due to |sj − rj | ≤ t and the fact that the integral with respect
to η is nonnegative by Lemma 6.5.
Let
 2
Z Z n 
Y
κ/2
 Yn
|ηj |1−2H−κ dη.

An (t) =  sin((sj+1 − sj )|η j | )ds  (4.4)
n
 n
R+  [0,t]< j=1
 j=1


Make the change of variables s0j = sj /t and ηj0 = ηj t2/κ , and we have the scaling
1−H
An (t) = t4n(1− κ )
An (1). (4.5)

Now we estimate E[An (τ )] where τ is an exponential random time with parameter 1. By

Fubini’s Theorem and Jensen’s inequality, we obtain
Z ∞
E[An (τ )] = e−t An (t)dt
0
 2
Z Z ∞ Z Yn  Yn
−t  κ/2
|ηj |1−2H−κ dη
 
= e  sin((sj+1 − sj )|ηj | )ds dt
Rn j=1
+ 0  [0,t]n< j=1 
 2
Z Z ∞ Z n  Y
Y n
−t κ/2
|ηj |1−2H−κ dη.
 
≥  e sin((sj+1 − sj )|ηj | )dsdt (4.6)
Rn

0 [0,t] n j=1
+  < j=1 

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 15

Applying the change of variables rj = sj+1 − sj for j = 0, 1, . . . , n with the convention

s0 = 0 and sn+1 = t and using Lemma 6.6, we have
Z ∞ Z Yn
−t
e sin((sj+1 − sj )|ηj |κ/2 )dsdt
0 [0,t]n
< j=1
Z n
Y
= e−(r0 +r1 +···+rn ) sin(rj |ηj |κ/2 )dr0 dr1 . . . drn
Rn+1
+ j=1
n
Y |ηj |κ/2
= . (4.7)
j=1
1 + |ηj |κ

Now, combining (4.6) and (4.7), we get

n
!n
|ηj |1−2H |η|1−2H
ZY Z
E[An (τ )] ≥ dη = dη = cn ,
Rn
+ j=1
(1 + |ηj |κ )2 R+ (1 + |η|κ )2

R |η|1−2H
where c = R+ (1+|η|κ )2
dη ∈ (0, ∞). Together with the scaling property (4.5), we have
1−H
cn ≤ E[An (τ )] = E[τ 4n(1− κ )
]An (1). (4.8)

Therefore, it implies from (4.3), (4.4), (4.5), (4.8) and the fact E[τ x ] = Γ(x + 1) that

n!kgn (·, t, x)k2H⊗n ≥ n!tn(2H0 −2) An (t)

1−H
= n!tn(2H0 −2) t4n(1− κ )
An (1)
4n(1− 1−H 1
≥ C n n!tn(2H0 −2) t κ )
1−H
E[τ 4n(1− κ ) ]
4−4H n!
= C n tn(2H0 +2− κ )
Γ(4n(1 − 1−H
κ ) + 1)
4−4H
C n tn(2H0 +2− κ )
∼ 4(1−H) 1−a
,
(n!) 3− κ aan+ 21 n 2

where the last step follows from (6.1) in Lemma 6.3 with a = 4 1 − 1−H


κ and b = 1.
1 1−a
Noting that there exists λ > 1 such that λ−n ≤ aan+ 2 n 2 ≤ λn , we obtain
4−4H
C n tn(2H0 +2− κ )
n!kgn (·, t, x)k2H⊗n ≥ 4(1−H)
.
(n!) 3− κ

Therefore, applying (6.2) in Lemma 6.3, we have

∞ 4−4H
! 21
X C n tn(2H0 +2− κ )  2κH0 +2(κ−2)+4H

ku(t, x)kp ≥ ku(t, x)k2 ≥ 4(1−H)
≥ C1 exp C1 t 3κ−4+4H ,
n=0 (n!)3− κ

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

16 J. Song, X. Song and F. Xu

for some C1 > 0.

n
For the upper bound, noting that kIn (gn )kp ≤ (p − 1) 2 kIn (gn )k2 (see the last line on
Page 62 in [38]), by Minkowski’s inequality and similar arguments in (3.8)-(3.9) we have
∞ ∞
n
X X
kukp ≤ kIn (gn )kp ≤ (p − 1) 2 kIn (gn )k2
n=0 n=0
∞
(p − 1) 2 C 2 tn(H0 +(1− κ )+ κ )
n n 2 2H
X
≤
(n!)[(1− κ )+ κ ]+ 2
2 2H 1
n=0
2κH0 +2(κ−2)+4H
 κ

≤ C exp Cp 3κ−4+4H t 3κ−4+4H ,

and hence, we get

2κH0 +2(κ−2)+4H
 κ

kukp ≤ C2 exp C2 p 3κ−4+4H t 3κ−4+4H ,

for some C2 > 0.

Remark 4.2 Note that the lower and upper bounds (lower bounds, resp.) for the p-th
moment were studied in [20] (in [2], resp.) based on the probabilistic representation for
the second moment of the solution to wave equations obtained in [21].
Recall that the lower Lyapunov exponent Ll (p) and the upper Lyapunov exponent
Lu (p) of order p ≥ 2 of the solution u(t, x) are defined, respectively by
1
Ll (p) = lim inf inf log E[|u(t, x)|p ]
t→∞ R(t) x∈R
and
1
Lu (p) = lim sup sup log E[|u(t, x)|p ]
t→∞ R(t) x∈R
for some positive function R(t). If Ll (2) > 0 and Lu (p) < ∞ for all p ≥ 0, we say that
u(t, x) possesses the weak intermittency. Heuristically speaking, if a process u(t, x) is
weakly intermittent, it concentrates on a few of very high peaks (see, e.g., [34] and the
2κH0 +2(κ−2)+4H
references therein). Taking R(t) = t 3κ−4+4H , the Proposition below follows directly
from Proposition 4.1.
Proposition 4.3 Under the conditions in Theorem 3.2, the solution u(t, x) to (1.1) is
weakly intermittent.

5. Hölder continuity
In this section, the Hölder continuity in time and space for the solution u(t, x) to the SWE
(1.1) is obtained in Proposition 5.1. The result is consistent with the Hölder continuity
for SWEs with noise that is not rough in space obtained in [1], [6], and [15]. Note that

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 17

exponents of Hölder continuity in both time and space are independent of the temporal
covariance function f0 (t). Similar phenomenon occurs also for SHEs (see [5, Theorem
3.2]).
Proposition 5.1 Assume the same conditions as in Theorem 3.2. Then on any set
[0, T ] × M where M ⊂ R is a compact set, u(t, x) has a modification which is θ1 -Hölder
continuous in time for all θ1 ∈ (0, 1 − κ2 + 2H
κ ) and θ2 -Hölder continuous in space for all
θ2 ∈ (0, H + κ2 − 1).
In particular, when κ = 2, the solution has a version that is θ-Höder continuous both
in time and in space for all θ ∈ (0, H).
Proof First we show the Hölder continuity in space. Noting that kIn (gn )kp ≤ (p −
n
1) 2 kIn (gn )k2 , by Minkowski’s inequality, we have

ku(t, x + z) − u(t, x)kp

∞
n 1
X
≤ (p − 1) 2 (n!) 2 kgn (·, t, x + z) − gn (·, t, x)kH⊗n
n=1
∞  Z Z h i
n
X
= (p − 1) 2 n! F gn (s, ·, t, x + z)(ξ) − gn (s, ·, t, x)(ξ)
n=1 Rn [0,t]2n

h n
iY  12
2H0 −2
F gn (r, ·, t, x + z)(ξ) − gn (r, ·, t, x)(ξ) |sj − rj | dsdrµ(dξ)
j=1
∞ n
sin2 ((sj+1 − sj )|ξ1 + · · · + ξj |κ/2 )
Z hZ
n 1
X Y
≤ (p − 1) 2 (n!)H0 − 2 κ
|ξj |1−2H
n=1
n
[0,t]< n
R j=1 |ξ1 + · · · + ξj |
 2 i 2H1  H0
1 − e−iz(ξ1 +···+ξn )  dξ
0
ds
 

∞ n
sin2 ((sj+1 − sj )|ηj |κ/2 )
Z hZ
n
H0 − 21
X Y
= (p − 1) (n!)
2
κ
|ηj − ηj−1 |1−2H
n=1
n
[0,t]< n
R j=1 |ηj |
i 1 H0
1 − e−izηn  dη 2H0 ds
 
.

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

18 J. Song, X. Song and F. Xu

Now, by changing of variables, we have

n  2H1
sin2 ((sj+1 − sj )|ηj |κ/2 )
Z Z Y
0
1−2H 1 − e−izηn  dη
 
|ηj − ηj−1 | dsdt
[0,t]n
< Rn j=1 |ηj |κ
n  2H1
sin2 ((sj+1 − sj )|ηj |κ/2 )
Z  X Z Y 0
αj (1−2H)
≤ κ
|ηj | (|zηn | ∧ 2)dη dsdt
[0,t]n n |ηj |
< α∈An R j=1
Z n "
X  2  2H0 n−1 Z 2  2H1 0
Y 1
[2− 2
− 2
α (1−2H)] sin (η) 2
α (1−2H)+ 2
−1
≤ (sj+1 − sj ) 2H0 κ κ j
|η| κ j κ dη
[0,t]n
< α∈An
κ j=1 R η2
 2H1 0 #
sin2 ((t − sn )y) 2 αn (1−2H)+ 2 −1
Z
2
|y| κ κ ((|z||y| κ ) ∧ 2)dy ds. (5.1)
R y2

Recall that αn ∈ {0, 1}, and by Lemma 6.7 with λ = κ2 , β = κ2 (αn (1 − 2H) + 1) − 1, γ = 0
for αn = 1 and γ = 1−2H κ for αn = 0, we have
sin2 ((t − sn )y) 2 αn (1−2H)+ 2 −1
Z
2 2(1−2H)

2
|y| κ κ ((|z||y| κ ) ∧ 2)dy ≤ C(1 ∨ (t − sn ) κ )|z|2H+κ−2
R y
≤ C|z|2H+κ−2 .
Thus, there exists a positive constant C depending only on (p, H0 , H, T ) such that
 H0
∞ Z n−1
1 1 2 2
(sj+1 − sj ) 2H0 [2− κ − κ αj (1−2H)] ds .
X X Y
ku(t, x+z)−u(t, x)kp ≤ C n (n!)H0 − 2 
n=1 [0,t]n
< α∈An j=1

When n = 1, the integral on the right-hand side of the above inequality equals t.
1 2 2

When n ≥ 2, for each fixed α ∈ An , denote βj = 2H 0
2 − κ − κ αj (1 − 2H) , j =
Pn
1, . . . , n. Noting that αn ∈ {0, 1} and j=1 αj = n, we have
n−1
(
(κ−2+2H)n
X
H0 κ − κ−2+2H
H0 κ , when αn = 1,
βj = (κ−2+2H)n κ−1
(5.2)
j=1 H0 κ − H0 κ , when αn = 0.
Using Lemma 6.2, we obtain
Z n−1
Y Z tZ n−1
Y
(sj+1 − sj )βj ds = (sj+1 − sj )βj ds
[0,t]n
< j=1 0 0<s1 <···<sn j=1
Qn−1 t
Γ(1 + βj )
Z
j=1
= sβ1 +···+βn−1 +n−1 dsn
Γ(β1 + · · · + βn−1 + n) 0 n
Qn−1
j=1 Γ(1 + βj )
= tβ1 +···+βn−1 +n
(β1 + · · · + βn−1 + n)Γ(β1 + · · · + βn−1 + n)
Qn−1
j=1 Γ(1 + βj )
= tβ1 +···+βn−1 +n .
Γ(β1 + · · · + βn−1 + n + 1)

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 19

κ−2+2H
Then applying (6.1) in Lemma 6.3 with a = 1 + H0 κ ∈ (1, 2) and either b = 1 −
κ−2+2H κ−1
H0 κ ∈ (0, 1) or b = 1 − H 0κ
∈ [0, 1), we have

n−1
C n−1
Z Y
(sj+1 − sj )βj ds ≤ tβ1 +···+βn−1 +n
[0,t]n
< j=1
Γ(β1 + · · · + βn−1 + n + 1)
C n−1 a(n−1)+ κ−2+2H +b
= t H0 κ
Γ(an + b)
C n−1 a(n−1)+ κ−2+2H +b
∼ an+b− 1
b− 1
− a t
H0 κ
a
(n!) a 2 n 2 2

κ−2+2H
+b a(n−1)
C n−1 T H0 κ
t
≤ .
(n!)a

Therefore, there exists a positive constant C depending only on (p, H0 , H, T ) such that
∞ (H0 κ+κ−2+2H)(n−1)
n
!
κ 1
X (p − 1) 2 C n−1 t κ
ku(t, x + z) − u(t, x)kp ≤|z|H+ 2 −1 (p − 1) 2 Ct + 3κ−4+4H
n=2 (n!) 2κ
∞ (H0 κ+κ−2+2H)n
κ
X C nt κ κ
≤ C|z|H+ 2 −1 3κ−4+4H ≤ C|z|H+ 2 −1 . (5.3)
n=0 (n!) 2κ

Then the θ2 -Hölder continuity for θ2 ∈ (0, H + κ2 − 1) follows from the Kolmogorov’s
continuity criterion.
Now we consider the Hölder continuity in time.
∞
n
X
ku(t + h, x) − u(t, x)kp ≤ (p − 1) 2 kgn (·, t + h, x) − gn (·, t, x)kH⊗n
n=1
∞ hp i
n 1
X p
≤ (p − 1) 2 (n!) 2 An (t, h) + Bn (t, h) , (5.4)
n=1

where
An (t, h) = kgn (·, t + h, x)I[0,t]n − gn (·, t, x)k2H⊗n
and
Bn (t, h) = kgn (·, t + h, x)I[0,t+h]n \[0,t]n k2H⊗n .

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

20 J. Song, X. Song and F. Xu

For An (t, h), we have

An (t, h) = kgn (·, t + h, x)I[0,t]n − gn (·, t, x)k2H⊗n

Z Z h i
= F gn (s, ·, t + h, x)(ξ) − gn (s, ·, t, x)(ξ)
Rn [0,t]2n
h n
iY
F gn (r, ·, t + h, x)(ξ) − gn (r, ·, t, x)(ξ) |sj − rj |2H0 −2 dsdrµ(dξ)
j=1
"Z n−1 n
2
sin ((sj+1 − sj )|ξ1 + · · · + ξj |κ/2 ) Y
Z Y
≤(n!)2H0 −2 |ξj |1−2H
[0,t]n
< Rn j=1 |ξ1 + · · · + ξj |κ j=1
# 2H1
 sin((t + h − s )|ξ + · · · + ξ |κ/2 ) sin((t − s )|ξ + · · · + ξ |κ/2 ) 2
  !2H0
0
n 1 n n 1 n
−  dξ ds
 
|ξ1 + · · · + ξn |κ/2 |ξ1 + · · · + ξn |κ/2

 
" Z n−1 
n
Z Y sin2 ((sj+1 − sj )|ηj |κ/2 ) Y
=(n!)2H0 −2 
κ
 |ηj − ηj−1 |1−2H
[0,t]n
< R n
j=1
|η j | j=1
# 2H1 !2H0
| sin((t + h − sn )|ηn |κ/2 ) − sin((t − sn )|ηn |κ/2 )|2 0

dη ds
|ηn |κ
"  
Z X Z n−1
Y sin2 ((sj+1 − sj )|ηj |κ/2 ) Y n
≤(n!)2H0 −2 
κ
 |ηj |αj (1−2H)
n
[0,t]< Rn
j=1
|ηj | j=1
α∈An
# 2H1 !2H0
Cγ2 ((|h|2γ |ηn |γκ ) ∧ (|h|2 |ηn |κ )) 0

dη ds ,
|ηn |κ

2 2H
where the last step follows from Lemma 6.8 with γ ∈ (0, 1 − κ + κ ). Note that when
γ ∈ (0, 1 − κ2 + 2H
κ ),

((|h|2γ |ηn |γκ ) ∧ (|h|2 |ηn |κ ))

Z
|ηn |αn (1−2H) dηn ≤ C(|h|2 + |h|2γ ) for αn ∈ {0, 1}
R |ηn |κ

and following the approach in the analysis of (5.1)-(5.3), we can show that
∞
n 1
X p
(p − 1) 2 (n!) 2 An (t, h) ≤ C(|h| + |h|γ )≤ C|h|γ (5.5)
n=1

for γ ∈ (0, 1 − κ2 + 2H
κ ) with C depending on (p, κ, H0 , H, T, M, γ).
Now we consider the term Bn (t, h). Denote Et,h = [0, t + h]n \[0, t]n , and then
[ n o
Et,h = (s1 , . . . , sn ) : sρ(1) ≤ sρ(2) ≤ · · · ≤ sρ(n) , t < sρ(n) ≤ t + h .
ρ∈Pn

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 21

Therefore, we have

Bn (t, h) = kgn (·, t + h, x)I[0,t+h]n \[0,t]n k2H⊗n

Z Z h i h i
= F gn (s, ·, t + h, x)(ξ) F gn (r, ·, t + h, x)(ξ)
Rn [0,t+h]2n
n
Y
IEt,h (s)IEt,h (r) |sj − rj |2H0 −2 dsdrµ(dξ)
j=1
n
t+h
sin2 ((sρ(j+1) − sρ(j) )|ξρ(1) + · · · + ξρ(j) |κ/2 )
X Z Z Z Y
−2
≤(n!)
t 0<sρ(1) <sρ(2) <···<sρ(n) Rn j=1 |ξρ(1) + · · · + ξρ(j) |κ
ρ∈Pn
n  2H1 !2H0
Y 0
1−2H
|ξj | dξ ds
j=1
n n  2H1 !2H0
t+h
sin2 ((sj+1 − sj )|ξ1 + · · · + ξj |κ/2 ) Y
Z Z Z
Y 0
−2 1−2H
= (n!) n! |ξj | dξ ds
t 0<s1 <s2 <···<sn Rn j=1 |ξ1 + · · · + ξj |κ j=1
Z t+h Z n n
X  2  2H0 Y 1 2 2
= (n!)2H0 −2
(sj+1 − sj ) 2H0 [2− κ − κ αj (1−2H)]
t 0<s1 <s2 <···<sn α∈A κ j=1
n

Z 2  2H1 0 !2H0
sin (η) 2 αj (1−2H)+ 2 −1
|η| κ κ dη ds
R η2

where by convention sj+1 = t + h.

Note that in the case n = 1 we have
 2− 2 − 2 (1−2H)+2H0
sin2 (η) 2 (1−2H)+ 2 −1
Z
2 h κ κ
Bn (t, h) ≤ 2
|η| κ κ dη
κ R η (β + 1)2H0
1
with β = 2H 0
[2 − κ2 − κ2 (1 − 2H)], and in the case n ≥ 2, it is easy to see from (5.2) that
β1 + · · · + βn−1 + n − 1 > 0 and hence by Lemma 6.2, we have
Z t+h Z n
Y
(sj+1 − sj )βj ds
t 0<s1 <s2 <···<sn j=1
Qn−1 t+h
Γ(1 + βj )
Z
j=1
= sβn1 +···+βn−1 +n−1 (t + h − sn )βn dsn
Γ(β1 + · · · + βn−1 + n) t
Qn−1
j=1 Γ(1 + βj ) 1
≤T β1 +···+βn−1 +n−1 hβn +1
Γ(β1 + · · · + βn−1 + n) βn + 1
Qn−1
(β1 + · · · + βn−1 + n) j=1 Γ(1 + βj ) 1
= T β1 +···+βn−1 +n−1
hβn +1 ,
Γ(β1 + · · · + βn−1 + n + 1) βn + 1

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

22 J. Song, X. Song and F. Xu

1
where βj = 2H 0
[2 − κ2 − κ2 αj (1 − 2H)]. Now similar to the calculus below the equation
(5.2) and recalling that αn ∈ {0, 1}, we can show that
n 1p 1 2 2H
X
(p − 1) 2 (n!) 2 Bn (t, h) ≤ C(|h|1− κ +H0 + |h|1− κ + κ +H0 ) (5.6)
n≥0

with C depending on (p, H0 , H, T ).

Finally, combining inequalities (5.4), (5.5) and (5.6), for |h| ≤ 1, we have

ku(t + h, x) − u(t, x)kp ≤ C|h|θ1

for any θ1 ∈ (0, 1− κ2 + 2H

κ ) where C is a constant depending only on (p, κ, H0 , H, T, M, θ1 ),
and the Hölder continuity in space follows from the Kolmogorov’s continuity criterion.
The proof is concluded.

6. Appendix
In this section, we collect the lemmas that were used in the preceding sections. Some of
the proofs are obvious and hence omitted.
Lemma 6.1 For a > 0 and θ > −1,
Z Z
2 θ − 21 (1+θ)
exp(−ax )|x| dx = a exp(−x2 )|x|θ dx.
R R

Lemma 6.2 Suppose αi ∈ (−1, ∞), i = 1, . . . , n and let α = α1 + · · · + αn . Then

n Qn
Γ(αi + 1)tα+n
Z Y
αi i=1
(ri+1 − ri ) dr1 . . . drn = ,
[0<r1 <···<rn <rn+1 =t] i=1
Γ(α + n + 1)
R∞
where Γ(x) = 0
tx−1 e−t dt is the Gamma function.
Lemma 6.3 For any a > 0 and b ∈ [0, 1], we have

Γ(an + b)
lim 1 1 a = 1, (6.1)
n→∞ (n!)a aan+b− 2 nb− 2 − 2

and
∞
 1
 X xn  1

c1 exp c2 x a ≤ ≤ C 1 exp C2 x a , ∀x > 0, (6.2)
n=0
(n!)a
where c1 > 0, c2 > 0, C1 > 0 and C2 > 0 are some constants depending on a.
Proof The proof of (6.1) follows from Stirling’s formula (see also (68) in [2] which is
(6.1) in the case of b = 1). See Lemma A. 1 in [2] for the upper bound in (6.2) and
Lemma 5.2 in [4] for the lower bound in (6.2).

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 23

Lemma 6.4 Z ∞
sin2 (x)x−α dx < ∞
0
if and only if α ∈ (1, 3).
Proof The sufficiency is obvious. The necessity follows from the estimation
Z ∞ Z π4 ∞ Z (n+3/4)π
sin2 (x) 2−α X
sin2 (x)x−α dx ≥ x dx + sin2 (x)x−α dx
0 0 x2 n=0 (n+1/4)π
Z π4 2 ∞
sin (x) 2−α 1 1−α X
≥ x dx + π (n + 3/4)−α .
0 x2 4 n=0

The proof is completed.

Lemma 6.5 For H ∈ (0, 1) and r, s > 0,
Z
1
sin(r|η|) sin(s|η|)|η|−1−2H dη = |r + s|2H − |r − s|2H .


CH
R 4
with CH given by (2.2). In particular, this integral is positive.
Proof Let XH be the Hilbert space associated with fractional Brownian motion {B H (x), x ∈
R} with Hurst parameter H ∈ (0, 1), i.e., it is the linear expansion of indicator functions
under the inner product
1
hI[0,x] , I[0,y] iXH = (|x|2H + |y|2H − |x − y|2H ).
2
Using the convention I[0,x] = −I[x,0] for x < 0, the linear mapping B H : I[0,x] → B H (x)
extends to a linear isometry between XH and the Gaussian space {B H (ϕ), ϕ ∈ XH }
spanned by B H . Furthermore, for ϕ, φ ∈ XH , we have (see [33])
Z
1−2H
hϕ, φiXH = CH ϕ(η)
b φ(η)|η|
b dη.
R

Now, noting that (FI[−r,r] (·))(ξ) = (FI[−r,r] (·))(ξ) = 2 sin(r|ξ|)

|ξ| , we have
Z
CH sin(r|η|) sin(s|η|)|η|−1−2H dη
R
Z
1
= CH (FI[|x|≤r] )(η)(FI[|x|≤s] )(η)|η|1−2H dη
4 R
1
= hI[−r,r] (·), I[−s,s] (·)iXH
4
1
= E[B H (I[−r,r] (·))B H (I[−s,s] (·))]
4
1
= E[(B H (r) − B H (−r))(B H (s) − B H (−s))]
4
1
= |r + s|2H − |r − s|2H .


4

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

24 J. Song, X. Song and F. Xu

We complete the proof.

Lemma 6.6 For any a ≥ 0, we have
Z ∞
a
e−t sin(at)dt =
0 1 + a2

Proof The result follows by integration by parts.

Lemma 6.7 Let a and b be two positive constants, and λ ∈ [1, ∞), β ∈ [0, 1), γ ∈ [0, 1]
such that 1 − λ < β + 2γ < 1. Then we have
Z
1
sin2 (ax)|x|−2+β ((b|x|λ ) ∧ 2)dx ≤ Ca2γ b λ (1−β−2γ) ,
R

where
(Z Z )
1 1
λ (−1+β+2γ) λ (−1+β+2γ)−1
C = Cλ,β,γ = max |y| dy, |y| dy .
|y|≤2 |y|>2

Proof We write
Z
sin2 (ax)|x|−2+β ((b|x|λ ) ∧ 2)dx
ZR Z
= sin2 (ax)|x|−2+β b|x|λ dx + 2 sin2 (ax)|x|−2+β dx.
b|x|λ ≤2 b|x|λ >2

Noting that | sin(x)| ≤ |x|γ for γ ∈ [0, 1], we have

Z Z
sin2 (ax)|x|−2+β b|x|λ dx ≤ |ax|2γ |x|−2+β b|x|λ dx
λ
b|x| ≤2 λ
b|x| ≤2
Z
1 2γ 1 (1−β−2γ) 1
= a bλ |y| λ (−1+β+2γ) dy,
λ |y|≤2

and
Z Z Z
2 −2+β 2γ −2+β 1 1 1
sin (ax)|x| dx ≤ |ax| |x| dx = a2γ b λ (1−β−2γ) |y| λ (−1+β+2γ)−1 dy.
b|x|λ >2 b|x|λ >2 λ |y|>2

Thus, the proof is concluded.

Lemma 6.8 For any t, h ∈ R and γ ∈ (0, 1], there exists a constant Cγ such that

| sin((t + h)x) − sin(tx)| ≤ Cγ |hx|γ .

In particular,
| sin((t + h)x) − sin(tx)| ≤ Cγ (|hx|γ ∧ |hx|).

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 25
1−γ
Proof By the mean value theorem and the fact of y ≤ 2 γ y γ for y ∈ [0, 2] and γ ∈ (0, 1],
we have
| sin((t + h)x) − sin(tx)| ≤ |hx|| cos(sx)| ≤ |hx|,
and
21−γ
| sin((t + h)x) − sin(tx)| ≤ | sin((t + h)x) − sin(tx)|γ = Cγ |hx|γ | cos(sx)|γ ≤ Cγ |hx|γ ,
γ
where s is a number between t and t + h. The desired results can be obtained.

Acknowledgements
F. Xu is partially supported by National Natural Science Foundation of China (Grant
No.11871219, No.11871220) and 111 Project (B14019).

References
[1] Raluca M. Balan. The stochastic wave equation with multiplicative fractional noise:
a Malliavin calculus approach. Potential Anal., 36(1):1–34, 2012.
[2] Raluca M. Balan and Daniel Conus. Intermittency for the wave and heat equations
with fractional noise in time. Ann. Probab., 44(2):1488–1534, 2016.
[3] Raluca M. Balan, Maria Jolis, and Lluı́s Quer-Sardanyons. SPDEs with affine mul-
tiplicative fractional noise in space with index 41 < H < 12 . Electron. J. Probab.,
20:no. 54, 36, 2015.
[4] Raluca M. Balan, Maria Jolis, and Lluı́s Quer-Sardanyons. Intermittency for the
hyperbolic Anderson model with rough noise in space. Stochastic Process. Appl.,
127(7):2316–2338, 2017.
[5] Raluca M. Balan, Lluı́s Quer-Sardanyons and Jian Song. Hölder continuity for the
Parabolic Anderson Model with space-time homogeneous Gaussian noise. Preprint.
[6] Raluca M. Balan and Jian Song. Hyperbolic Anderson model with space-time ho-
mogeneous Gaussian noise. ALEA Lat. Am. J. Probab. Math. Stat., 14(2):799–849,
2017.
[7] Raluca M. Balan and Jian Song. Second order Lyapunov exponents for parabolic
and hyperbolic Anderson models. Bernoulli, in press.
[8] Lorenzo Bertini and Nicoletta Cancrini. The stochastic heat equation: Feynman-Kac
formula and intermittence. J. Statist. Phys., 78(5-6):1377–1401, 1995.
[9] Le Chen and Robert C. Dalang. Moments and growth indices for the nonlinear
stochastic heat equation with rough initial conditions. Ann. Probab., 43(6):3006–
3051, 2015.
[10] Xia Chen. Moment asymptotics for parabolic Anderson equation with fractional
time-space noise: in Skorokhod regime. Ann. Inst. Henri Poincaré Probab. Stat.,
53(2):819–841, 2017.

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

26 J. Song, X. Song and F. Xu

[11] Xia Chen. Parabolic Anderson model with rough or critical Gaussian noise. Preprint.
[12] Xia Chen. Parabolic Anderson model with a fractional Gaussian noise that is rough
in time. Preprint.
[13] Xia Chen, Yaozhong Hu, Jian Song and Xiaoming Song. Temporal asymptotics for
fractional parabolic Anderson model. Electron. J. Probab., 23:Paper No. 14, 39,
2018.
[14] Xia Chen, Yaozhong Hu, Jian Song and Fei Xing. Exponential asymptotics for
time-space Hamiltonians. Ann. Inst. Henri Poincaré Probab. Stat., 51(4):1529–1561,
2015.
[15] Daniel Conus and Robert C. Dalang. The non-linear stochastic wave equation in
high dimensions. Electron. J. Probab., 13:no. 22, 629–670, 2008.
[16] Daniel Conus, Mathew Joseph, Davar Khoshnevisan and Shang-Yuan Shiu. Inter-
mittency and chaos for a nonlinear stochastic wave equation in dimension 1. In
Malliavin calculus and stochastic analysis, volume 34 of Springer Proc. Math. Stat.,
pages 251–279. Springer, New York, 2013.
[17] Daniel Conus and Davar Khoshnevisan. On the existence and position of the farthest
peaks of a family of stochastic heat and wave equations. Probab. Theory Related
Fields, 152(3-4):681–701, 2012.
[18] Robert C. Dalang. Extending the martingale measure stochastic integral with ap-
plications to spatially homogeneous s.p.d.e.’s. Electron. J. Probab., 4:no. 6, 29 pp.
1999.
[19] Robert C. Dalang and N. E. Frangos. The stochastic wave equation in two spatial
dimensions. Ann. Probab., 26(1):187–212, 1998.
[20] Robert C. Dalang and Carl Mueller. Intermittency properties in a hyperbolic An-
derson problem. Ann. Inst. Henri Poincaré Probab. Stat., 45(4):1150–1164, 2009.
[21] Robert C. Dalang, Carl Mueller and Roger Tribe. A Feynman-Kac-type formula for
the deterministic and stochastic wave equations and other P.D.E.’s. Trans. Amer.
Math. Soc., 360(9):4681–4703, 2008.
[22] Robert C. Dalang and Marta Sanz-Solé. Hölder-Sobolev regularity of the solu-
tion to the stochastic wave equation in dimension three. Mem. Amer. Math. Soc.,
199(931):vi+70, 2009.
[23] Mohammud Foondun and Davar Khoshnevisan. On the stochastic heat equation
with spatially-colored random forcing. Trans. Amer. Math. Soc., 365(1):409–458,
2013.
[24] Mohammud Foondun, Davar Khoshnevisan and Eulalia Nualart. A local-time cor-
respondence for stochastic partial differential equations. Trans. Amer. Math. Soc.,
363(5):2481–2515, 2011.
[25] Yaozhong Hu, Jingyu Huang, Khoa Lê, David Nualart, and Samy Tindel. Stochastic
heat equation with rough dependence in space. Ann. Probab., 45(6B):4561–4616,
2017.
[26] Yaozhong Hu, Jingyu Huang, and David Nualart. On Hölder continuity of the
solution of stochastic wave equations in dimension three. Stoch. Partial Differ. Equ.
Anal. Comput., 2(3):353–407, 2014.
[27] Yaozhong Hu, Jingyu Huang, David Nualart and Samy Tindel. Stochastic heat

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019

Fractional SWE driven by a Gaussian noise 27

equations with general multiplicative Gaussian noises: Hölder continuity and inter-
mittency. Electron. J. Probab., 20:no. 55, 50, 2015.
[28] Yaozhong Hu and Khoa Lê. Hölder continuity of Parabolic Anderson Model.
Preprint.
[29] Yaozhong Hu and David Nualart. Stochastic heat equation driven by fractional
noise and local time. Probab. Theory Related Fields, 143(1-2):285–328, 2009.
[30] Yaozhong Hu, David Nualart and Jian Song. Feynman-Kac formula for heat equation
driven by fractional white noise. Ann. Probab., 39(1):291–326, 2011.
[31] Jingyu Huang, Khoa Lê and David Nualart. Large time asymptotics for the parabolic
Anderson model driven by space and time correlated noise. Stoch. Partial Differ.
Equ. Anal. Comput., 5(4):614–651, 2017.
[32] Jingyu Huang, Khoa Lê and David Nualart. Large time asymptotics for the parabolic
Anderson model driven by spatially correlated noise. Ann. Inst. Henri Poincaré
Probab. Stat., 53(3):1305–1340, 2017.
[33] Maria Jolis. The Wiener integral with respect to second order processes with sta-
tionary increments. J. Math. Anal. Appl., 366(2):607–620, 2010.
[34] Davar Khoshnevisan. Analysis of stochastic partial differential equations, volume
119 of CBMS Regional Conference Series in Mathematics. Published for the Con-
ference Board of the Mathematical Sciences, Washington, DC; by the American
Mathematical Society, Providence, RI, 2014.
[35] Jean Mémin, Yulia Mishura and Esko Valkeila. Inequalities for the moments of
Wiener integrals with respect to a fractional Brownian motion. Statist. Probab.
Lett., 51(2):197–206, 2001.
[36] Annie Millet and Marta Sanz-Solé. A stochastic wave equation in two space dimen-
sion: smoothness of the law. Ann. Probab., 27(2):803–844, 1999.
[37] Carl Mueller. Long time existence for the wave equation with a noise term. Ann.
Probab., 25(1):133–151, 1997.
[38] David Nualart. The Malliavin calculus and related topics. Probability and its Ap-
plications (New York). Springer-Verlag, Berlin, second edition, 2006.
[39] Vladas Pipiras and Murad S. Taqqu. Integration questions related to fractional
Brownian motion. Probab. Theory Related Fields, 118(2):251–291, 2000.
[40] Jian Song. Asymptotic behavior of the solution of heat equation driven by fractional
white noise. Statist. Probab. Lett., 82(3):614–620, 2012.
[41] Jian Song. On a class of stochastic partial differential equations. Stochastic Process.
Appl., 127(1):37–79, 2017.
[42] François Trèves. Basic linear partial differential equations. Academic Press [A
subsidiary of Harcourt Brace Jovanovich, Publishers], New York-London, 1975. Pure
and Applied Mathematics, Vol. 62.
[43] John B. Walsh. An introduction to stochastic partial differential equations. In École
d’été de probabilités de Saint-Flour, XIV—1984, volume 1180 of Lecture Notes in
Math., pages 265–439. Springer, Berlin, 1986.

imsart-bj ver. 2014/10/16 file: bj-FSWE.tex date: April 20, 2019