8

Linear equations with additive noise I

Let A be the generator of a C0 -semigroup S on E. In the previous lecture we

have seen that the inhomogeneous abstract Cauchy problem

u′ (t) = Au(t) + f (t), u(0) = x,

is solved by the convolution formula

Z t
u(t) = S(t)x + S(t − s)f (s) ds,
0

We now turn to the stochastic analogue of this equation,

(
dU (t) = AU (t) dt + B dWH ,
u(0) = x,

where WH is an H-cylindrical Brownian motion, and B ∈ L (H, E) is a

bounded operator. In concrete examples, WH models space-time white noise
and B ‘injects’ this noise into the state space E. Reasoning by analogy, this
equation should be solved by the stochastic convolution
Z t
U (t) = S(t)x + S(t − s)B dWH .
0

We shall see that this is indeed correct, provided the L (H, E)-valued function
S(·)B is stochastically integrable with respect to WH .

8.1 Stochastic preliminaries

In this section we collect several results which are needed in the proof of the
main result of this lecture, Theorem 8.6. We begin with an integrations by
parts formula.
106 8 Linear equations with additive noise I

Lemma 8.1 (Integration by parts). For all φ ∈ C 1 [0, T ] and h ∈ H,

almost surely the following identity holds:
Z T Z T
φ′ (t)WH (t)h dt = φ(T )WH (T )h − φ ⊗ h dWH .
0 0
Before we prove the lemma we clarify the meaning of the integral on the left
hand side. Recalling that WH h is a Brownian motion, using Corollary 6.10 we
select a version of WH h whose trajectories are continuous almost surely. Then
the integral on the left hand side is well defined almost surely as a Lebesgue
integral.
Proof. We may assume that φ′ (0) = 0; this somewhat simplifies the calcula-
tions below. RT RT
We begin by noting that 0 φ ⊗ h dWH = 0 φ dWH h. For step functions
this is clear from the definitions and the general case follows by approximation.
Rescaling h to unit length, it is therefore enough to prove the almost sure
identity
Z T Z T
φ′ (t)W (t) dt = φ(T )W (T ) − φ dW
0 0
for functions φ ∈ C 1 [0, T ], where W is a scalar Brownian motion. This identity
is a special case of Itô’s formula, but for those readers who are not familiar
with it we shall give a self-contained argument (which is indeed nothing but
the proof of Itô’s formula in the special case considered here). Let
N
X N X
X n
g := cn 1(tn−1 ,tn ] , G := cm (tm − tm−1 )1(tn−1 ,tn ]
n=1 n=1 m=1

with c1 , . . . , cN scalars and 0 = t0 < · · · < tN = T . Then, almost surely,

Z T X N Z tn
g(t)W (t) dt = cn W (t) dt
0 n=1 tn−1

and
Z T
G(T )W (T ) − G dW
0
N
X N X
X n
= cm (tm − tm−1 )W (T ) − cm (tm − tm−1 )(W (tn ) − W (tn−1 ))
m=1 n=1 m=1
N
X N X
X N
= cm (tm − tm−1 )W (T ) − cm (tm − tm−1 )(W (tn ) − W (tn−1 ))
m=1 m=1 n=m
N
X N
X
= cm (tm − tm−1 )W (T ) − cm (tm − tm−1 )(W (T ) − W (tm−1 ))
m=1 m=1
N
X
= cm (tm − tm−1 )W (tm−1 ).
m=1
 8.1 Stochastic preliminaries 107
PNk ′
Now let φ ∈ C 1 [0, T ] be given and put gk := n=1 φ (tk,n−1 )1(tk,n−1 ,tk,n ] ,
assuming that limk→∞ sup16n6Nk (tk,n − tk,n−1 ) = 0. Then limk→∞ gk = φ′
uniformly on (0, T ]. Defining the functions Gk in terms of the gk as above, we
have limk→∞ Gk = φ uniformly on (0, T ]. The above computation gives the
following identity, which almost surely holds for all k:
Z T  Z T 
gk (t)W (t) dt − Gk (T )W (T ) − Gk dW
0 0
Nk
X Z tk,n Nk
X
= φ′ (tk,n−1 ) W (t) dt − φ′ (tk,n−1 )(tk,n − tk,n−1 )W (tk,n−1 ) .
n=1 tk,n−1 n=1

RT
As k → ∞, the left hand side tends to | 0 φ′ (t)W (t) dt − (φ(T )W (T ) −
RT
0
φ dW )| in L2 (Ω) and hence in measure, whereas the right hand side tends
to 0 almost surely by path continuity. This proves the lemma. ⊓
⊔
We continue with a Fubini theorem for interchanging a Bochner integral
and a stochastic integral of an H-valued function. In this context it is natural
to impose an integrability condition which is L1 with respect to the vari-
able of Bochner integration and L2 with respect to the variable of stochastic
integration.

Lemma 8.2 (Stochastic Fubini theorem). Let φ : (0, T ) × (0, T ) → H be

a strongly measurable function satisfying
Z T Z T  12
kφ(s, t)k2H dt ds < ∞.
0 0

(1) t 7→ φ(s, t) belongs to L2 (0, T ; H) for almost all s ∈ (0, T ), and the L2 (Ω)-
RT
valued function s 7→ 0 φ(s, t) dWH (t) belongs to L1 (0, T ; L2 (Ω));
(2) s 7→ φ(s, t) belongs to L1 (0, T ; H) for almost all t ∈ (0, T ), and the H-
RT
valued function t 7→ 0 φ(s, t) ds belongs to L2 (0, T ; H);
(3) in L2 (Ω) we have
Z T Z T  Z T Z T 
φ(s, t) dWH (t) ds = φ(s, t) ds dWH (t).
0 0 0 0

Proof. (1): By assumption we have φ ∈ L1 (0, T ; L2(0, T ; H)), and therefore

(1) is an immediate consequence of the Itô isometry (6.2).
(2): We claim that for a step function φ : (0, T ) × (0, T ) → H we have

kφkL2 (0,T ;L1 (0,T ;H)) 6 kφkL1 (0,T ;L2 (0,T ;H)) .
PM PN
It suffices to prove this for T = 1. If φ = j=1 k=1 1(sj−1 ,sj ) 1(tk−1 ,tk ) ⊗ hjk ,
then
108 8 Linear equations with additive noise I
N
X X
M 2
kφk2L2 (0,T ;L1 (0,T ;H)) = (tk − tk−1 ) (sj − sj−1 )khjk k
k=1 j=1
M X
X M N
X
= (si − si−1 )(sj − sj−1 ) (tk − tk−1 )khik kkhjk k
i=1 j=1 k=1

and similarly
X
M X
N  12 2
kφk2L1 (0,T ;L2 (0,T ;H)) = (sj − sj−1 ) (tk − tk−1 )khjk k2
j=1 k=1
M X
X M
= (si − si−1 )(sj − sj−1 )
i=1 j=1
X
N  21  X
N  21
× (tk − tk−1 )khik k2 (tk − tk−1 )khjk k2 .
k=1 k=1

In view of the Cauchy-Schwarz inequality, this proves the claim. It follows that
the identity mapping on step functions extends to a continuous embedding of
L1 (0, T ; L2(0, T ; H)) into L2 (0, T ; L1(0, T ; H)). This gives (2).
(3): For step functions φ the identity follows by a trivial computation, and
its extension to functions φ ∈ L1 (0, T ; L2 (0, T ; H)) is obtained by approxima-
tion using (1) and (2). ⊓
⊔

8.2 Semigroup preliminaries

Let A be the generator of a C0 -semigroup S on E. Define

E ⊙ := D(A∗ ),

the closure being taken with respect to the norm topology of E ∗ . Note that
E ⊙ is a closed and weak∗ -dense subspace of E ∗ . We let A⊙ be the part of A∗
in E ⊙ , that is,

D(A⊙ ) := {x∗ ∈ D(A∗ ) : A∗ x∗ ∈ E ⊙ },

A⊙ x∗ := A∗ x∗ , x∗ ∈ D(A⊙ ).

Proposition 8.3. Let A be the generator of a C0 -semigroup S on E. The

adjoint semigroup S ∗ restricts to a C0 -semigroup S ⊙ on E ⊙ whose generator
equals A⊙ .
Proof. For t ∈ [0, T ], x ∈ E, and x∗ ∈ D(A∗ ) we have
Z t
∗ ∗ ∗
|hx, S (t)x − x i| 6 |hx, S ∗ (s)A∗ x∗ i| ds 6 tkxk · sup kS(s)k · kA∗ x∗ k.
0 s∈[0,T ]
 8.2 Semigroup preliminaries 109

Taking the supremum over all x ∈ E of norm kxk 6 1 gives

lim sup kS ∗ (t)x∗ − x∗ k 6 lim t · sup kS(s)k · kA∗ x∗ k = 0.

t↓0 t↓0 s∈[0,T ]

Since D(A∗ ) is invariant under S ∗ (by duality we have A∗ S ∗ (t)x∗ = S ∗ (t)A∗ x∗

for x∗ ∈ D(A∗ ) and t > 0) and S ∗ (t) is uniformly bounded on [0, T ], it follows
that S ∗ restricts to a C0 -semigroup S ⊙ on E ⊙ .
Let B denote the generator of S ⊙ . If x⊙ ∈ D(B), then for all x ∈ D(A)
we have
1 1
hx, Bx⊙ i = lim hx, S ⊙ (t)x⊙ − x⊙ i = lim hS(t)x − x, x⊙ i = hAx, x⊙ i.
t↓0 t t↓0 t

Hence x⊙ ∈ D(A∗ ) and A∗ x⊙ = Bx⊙ . Since Bx⊙ ∈ E ⊙ it follows that

x⊙ ∈ D(A⊙ ) and A⊙ x⊙ = A∗ x⊙ = Bx⊙ . Conversely, if x⊙ ∈ D(A⊙ ), then
A⊙ x⊙ ∈ E ⊙ and s 7→ S ⊙ (s)A⊙ x⊙ is strongly continuous and, for all x ∈ E,
D E Z
⊙ ⊙ 1 ⊙ ⊙ ⊙ ⊙ ⊙ 1 t ⊙
x, A x − (S (t)x − x ) 6 kxk A x − S (s)A⊙ x⊙ ds .
t t 0

Hence,
Z t
1 1
A⊙ x⊙ − (S ⊙ (t)x⊙ − x⊙ ) 6 A⊙ x⊙ − S ⊙ (s)A⊙ x⊙ ds .
t t 0

⊙ ⊙ ⊙
Since A x ∈ E , the right hand side tends to 0 as t ↓ 0 by strong continuity.
This proves that x⊙ ∈ D(B) and Bx⊙ = A⊙ x⊙ . ⊓
⊔

This proposition will be used in combination with the next approximation

result.

Lemma 8.4. For k = 0, 1, 2, . . . , linear combinations of the functions φ ⊗ x

with φ ∈ C k [0, T ] and x ∈ E are dense in C k ([0, T ]; E).

Proof. We begin with the case k = 0, which is proved by a standard partition

of unity argument. Let f ∈ C([0, T ]; E) be arbitrary. Let ε > 0. Since f
is uniformly continuous we may choose δ > 0 such that kf (t) − f (s)k < ε
whenever |t − s| < δ. Let I1 , . . . , IN be open intervals of length < δ covering
[0, T ] and let φ1 , . . . , φN be a partition of unity with respect to this cover,
PN
that is, 0 6 φn 6 1, φn is supported in In , and n=1 φn = 1. Choose points
P
tn ∈ [0, T ] ∩ In , let xn := f (tn ), and put fε := N n=1 φn ⊗ xn . Fix t ∈ [0, T ].
If t ∈ In , then |t − tn | < δ and therefore kf (t) − xn k < ε. If t 6∈ In , then
P
φn (t) = 0. Hence, using that f = N n=1 φn f ,

N
X X
kf (t) − fε (t)k 6 φn (t)kf (t) − xn k 6 ε φn (t) 6 ε.
n=1 n: t∈In
110 8 Linear equations with additive noise I

This proves that kf − fε k 6 ε.

The general case is proved with induction on k. Suppose the lemma has
been proved for k = 0, . . . , l and let f ∈ C l+1 ([0, T ]; E) be arbitrary. Then
f ′ ∈ C l ([0, T ]; E) and therefore we can find functions gj ∈ C l ([0, T ]; E) of
PNj
the form gj = n=1 φjn ⊗ xjn with φjn ∈ C l [0, T ] and xjn ∈ E such that
′ l
Rt
limj→∞ gj = f in C ([0, T ]; E). Let ψjn (t) := 0 φjn (s) ds, put x0 := f (0),
and set
Nj
X
fj := 1 ⊗ x0 + ψjn ⊗ xjn .
n=1

Then limj→∞ fj = f in C([0, T ]; E) and limn→∞ fj′ = f ′ in C l ([0, T ]; E), so

limj→∞ fj = f in C l+1 ([0, T ]; E). ⊓
⊔

8.3 Existence and uniqueness: cylindrical Brownian

motion
We consider the stochastic abstract Cauchy problem
(
dU (t) = AU (t) dt + B dWH (t), t ∈ [0, T ],
(SACP)
U (0) = x.

Here A is the generator of a C0 -semigroup {S(t)}t>0 on E, WH is an H-

cylindrical Brownian motion on (Ω, F , P), and B ∈ L (H, E) is a given
bounded operator.
An E-valued process {U (t)}t∈[0,T ] will be called strongly measurable if it
has a version which is strongly B([0, T ]) × F -measurable on [0, T ] × Ω.

Definition 8.5. A weak solution of the problem (SACP) is an E-valued pro-

cess {U x (t)}t∈[0,T ] which has a strongly measurable version with the following
properties:
(i) almost surely, the paths t 7→ U x (t) are integrable;
(ii) for all t ∈ [0, T ] and x∗ ∈ D(A∗ ) we have, almost surely,
Z t
hU x (t), x∗ i = hx, x∗ i + hU x (s), A∗ x∗ i ds + WH (t)B ∗ x∗ .
0

In order not to overburden notations, we do not distinguish notationally

the process {U x (t)}t∈[0,T ] from its version with the properties (i) and (ii).

Theorem 8.6. The following assertions are equivalent:

(1) the problem (SACP) has a weak solution {U x (t)}t∈[0,T ] ;
(2) t 7→ S(t)B is stochastically integrable on (0, T ) with respect to WH .
 8.3 Existence and uniqueness: cylindrical Brownian motion 111

In this situation, for every t ∈ (0, T ) the function s 7→ S(t − s)B is stochasti-
cally integrable on (0, t) with respect to WH and almost surely we have
Z t
U x (t) = S(t)x + S(t − s)B dWH (s). (8.1)
0

Proof. We start by noting that t 7→ U x (t) is a weak solution corresponding

to the initial value x if and only if t 7→ U x (t) − S(t)x is a weak solution cor-
responding to the initial value 0. Without loss of generality we shall therefore
assume that x = 0 and write U (t) := U 0 (t) for convenience.
(1) ⇒ (2): We will show first that for all t ∈ [0, T ] and x∗ ∈ D(A⊙2 ),
almost surely we have
Z t
hU (t), x∗ i = B ∗ S ∗ (t − s)x∗ dWH (s). (8.2)
0
⊙ ⊙
Fix t ∈ [0, T ] and x ∈ D(A ). By Fubini’s theorem, almost surely the
identity Z s
hU (s), x⊙ i = hU (r), A⊙ x⊙ i dr + WH (s)B ∗ x⊙ (8.3)
0
holds for almost all s ∈ (0, t); here we use that both terms on the right hand
side are jointly measurable on (0, t) × Ω. In combination with Lemma 8.1 this
gives, for any C 1 -function φ : [0, t] → R,
Z t
φ′ (s)hU (s), x⊙ i ds
0
Z t Z s  Z t
′ ⊙ ⊙
= φ (s) hU (r), A x i dr ds + φ′ (s)WH (s)B ∗ x⊙ ds
0 0 0
Z t Z t
⊙ ⊙
= φ(t) hU (s), A x i ds − φ(s)hU (s), A⊙ x⊙ i ds
0 0
Z t
∗ ⊙
+ φ(t)WH (t)B x − φ(s)B ∗ x⊙ dWH (s)
0

almost surely. Multiplying both sides of (8.3) with φ(t), putting f := φ ⊗ x⊙

and rewriting, we obtain
Z t Z t
hU (t), f (t)i = hU (s), f ′ (s) + A⊙ f (s)i ds + B ∗ f (s) dWH (s) (8.4)
0 0

almost surely. By Lemma 8.4 applied to the Banach space D(A⊙ ), this identity
extends to arbitrary functions f ∈ C 1 ([0, t]; D(A⊙ )). In particular we may
take f (s) = S ⊙ (t − s)x⊙ , with x⊙ ∈ D(A⊙2 ). For this choice of f , the identity
(8.4) reduces to (8.2).
So far we have proved that (8.2) holds for functionals x∗ ∈ D(A⊙2 ). We
shall prove next that (8.2) holds for functionals x∗ ∈ E ∗ . Then the stochastic
integrability of s 7→ S(t − s)B on (0, t) follows from Theorem 6.17.
112 8 Linear equations with additive noise I

The extension of (8.2) from functionals x∗ ∈ D(A⊙2 ) to functionals x∗ ∈

E ∗ is not entirely straightforward since in general D(A⊙2 ) is only weak∗ -dense
in E ∗ . Let x∗ ∈ E ∗ be arbitrary and fixed, and let weak∗ -limn→∞ x∗n = x∗
with all x∗n ∈ D(A⊙2 ) (for instance, take x∗n = λ3n R(λn , A∗ )3 x∗ with suitable
λn → ∞). By dominated convergence, for all f ∈ L2 (0, t; H) we have

lim [f, B ∗ S ∗ (t − ·)x∗n ]L2 (0,t;H) = [f, B ∗ S ∗ (t − ·)x∗ ]L2 (0,t;H) .

n→∞

It follows that for all N > 1, B ∗ S ∗ (t − ·)x∗ belongs to the weak closure in
L2 (0, t; H) of the tail sequence (B ∗ S ∗ (t − ·)x∗n )∞
n=N . By the Hahn-Banach
theorem, B ∗ S ∗ (t − ·)x∗ belongs to the strong closure in L2 (0, t; H) of the
∗
convex hull of this sequence. It follows that there exist vectors yN , belonging
∗ ∞
to the convex hull of (xn )n=N , such that

1
kB ∗ S ∗ (t − ·)yN
∗
− B ∗ S ∗ (t − ·)x∗ kL2 (0,t;H) < .
N
The isometry (6.2) implies that
Z t Z t
lim B ∗ S ∗ (t − s)yN
∗
dWH (s) = B ∗ S ∗ (t − s)x∗ dWH (s)
N →∞ 0 0

in L2 (Ω). By passing to a subsequence and using that weak∗ -limN →∞ yN∗

= x∗
(this follows from the fact that we used the tail sequence (x∗n )∞
n=N to define
∗
yN ), we obtain
Z t
∗ ∗
hU (t), x i = lim hU (t), yN j
i = lim B ∗ S ∗ (t − s)yN
∗
j
dWH (s)
j→∞ j→∞ 0
Z t
= B ∗ S ∗ (t − s)x∗ dWH (s)
0

almost surely.
(2) ⇒ (1): Suppose now that the function t 7→ S(t)B is stochastically
integrable on (0, T ). This implies the stochastic integrability of s 7→ S(t −
s)B on (0, t) for all t ∈ (0, T ]. We check that the process U defined by the
convolution (8.1) with x = 0 has a strongly measurable version which is a
weak solution of the problem (SACP) with initial value x = 0.
To prove that U has a strongly measurable version we argue as follows.
As in the proof of Step 1 of Theorem 6.17 (3)⇒(1) we may assume that H
is separable. Then by Proposition 5.14 the γ(L2 (0, T ; H), E)-valued function
t 7→ Rt is strongly measurable, where Rt is the integral operator associated
with s 7→ 1(0,t) (s)S(t − s)B. By covariance domination, kRt kγ(L2 (0,T ;H),E) 6
kRT kγ(L2 (0,T ;H),E) . Applying the Itô isometry of Theorem 6.14 we see that
U defines an element of L∞ (0, T ; L2 (Ω; E)). The existence of a strongly mea-
surable version follows from this (cf. Example 1.21).
Fix x∗ ∈ D(A∗ ) and t ∈ [0, T ]. Then almost surely
 8.4 Existence and uniqueness: Brownian motion 113
Z t
hU (t), A∗ x∗ i = B ∗ S ∗ (t − s)A∗ x∗ dWH (s).
0
By the stochastic Fubini theorem applied to φ(s, t) := 1{06s6t6T } B ∗ S ∗ (t −
s)x∗ , the L2 (Ω)-valued function t 7→ hU (t), A∗ x∗ i is integrable on (0, T ) and
Z t Z tZ s
∗ ∗
hU (s), A x i ds = B ∗ S ∗ (s − r)A∗ x∗ dWH (r) ds
0 0 0
Z tZ t
= B ∗ S ∗ (s − r)A∗ x∗ ds dWH (r)
0 r
Z t
= B ∗ S ∗ (t − r)x∗ − B ∗ x∗ dWH (r)
0
= hU (t), x∗ i − WH (t)B ∗ x∗ ,
where all identities are understood in the sense of L2 (Ω). In particular the
identities hold almost surely.
It remains to check that the trajectories of U are integrable almost surely.
Let µt be the distribution of U (t) and let Qt be its covariance operator. We
have
Z t
hQt x∗ , x∗ i = kB ∗ S ∗ (s)x∗ k2H ds 6 hQT x∗ , x∗ i = hRx∗ , x∗ i.
0
Hence by Fubini’s theorem and covariance domination, for arbitrary but fixed
1 6 p < ∞ we obtain
Z T Z TZ Z
p p
E kU (t)k dt = kxk dµt (x) dt 6 T kxkp dµT (x) < ∞.
0 0 E E
This implies that almost all trajectories t 7→ U (t, ω) belong to Lp (0, T ; E). ⊓
⊔
Note that theorem 8.6 contains the following uniqueness assertion: if U x
and Ue x are both weak solutions of (SACP), then U x and U e x are versions of
x e x
each other: both U (t) and U (t) equal the right hand side of (8.1) almost
surely. This justifies us to speak of ‘the’ solution of (SACP).
Comparing the proof of Theorem 8.6 with that of Theorem 7.17 we observe
that the existence proofs are essentially identical, whereas the uniqueness parts
are very different. The reason is that the exceptional sets in the definition of
a weak solution of the stochastic problem (SACP) depend on t and x∗ , which
prevents us from applying Proposition 7.14 almost surely. Because of this it
is no longer clear whether a weak solution is always a strong solution (cf.
Proposition 7.16).

8.4 Existence and uniqueness: Brownian motion

Next we consider the problem (SACP) under the assumption that B ∈
γ(H, E). In this situation the term ‘B dWH ’ may be replaced by ‘dW B ’, where
W B is an E-valued Brownian motion canonically associated with B.
114 8 Linear equations with additive noise I

Definition 8.7. An E-valued process (W (t))t∈[0,T ] is called an E-valued

Brownian motion if it enjoys the following properties:
i) W (0) = 0 almost surely;
ii) W (t − s) and W (t) − W (s) are identically distributed Gaussian random
variables for all 0 6 s 6 t 6 T ;
iii) W (t) − W (s) is independent of {W (r) : 0 6 r 6 s} for all 0 6 s 6 t 6 T .

Proposition 8.8. Let (WH (t))t∈[0,T ] be an H-cylindrical Brownian motion

and let B ∈ γ(H, E). If (hn )∞ ⊥
n=1 is an orthonormal basis of (ker(B)) , then:

(1) the sum

∞
X
W B (t) := WH (t)hn ⊗ Bhn
n=1

converges almost surely and in Lp (Ω; E), 1 6 p < ∞, for all t ∈ [0, T ];
(2) up to a null set, W B (t) is independent of the choice of the basis (hn )∞
n=1 ;
(3) the process (W B (t))t∈[0,T ] defines an E-valued Brownian motion.

The proof involves a straightforward application of Theorem 5.15, noting

that for 0 6 s 6 t 6 T the covariance operator of W B (t) − W B (s) equals
(t − s)BB ∗ .
This proposition shows that for operators B ∈ γ(H, E) the problem
(SACP) may be restated as
(
dU (t) = AU (t) dt + dW B (t), t ∈ [0, T ],
U (0) = x.

In the converse direction, every E-valued Brownian motion is of the form

W B for canonical choices of H and B ∈ γ(H, E) (Exercise 2).

Definition 8.9. Let B ∈ γ(H, E). A strong solution of (SACP) is a strongly

measurable E-valued process (U x (t))t∈[0,T ] with the following properties:
i) the trajectories of U x are integrable almost surely;
Rt
ii) for all t ∈ [0, T ], almost surely we have 0 U x (s) ds ∈ D(A) and
Z t
x
U (t) = x + A U x (s) ds + W B (t).
0

Theorem 8.10. Let B ∈ γ(H, E). The following assertions are equivalent:
(1) the problem (SACP) has a strong solution;
(2) the problem (SACP) has a weak solution.
In this situation, the weak and strong solutions are versions of each other, and
both are given by (8.1).
 8.4 Existence and uniqueness: Brownian motion 115

Proof. We only need to prove that (2) implies (1). We may assume that x = 0.
Let U be a weak solution of (SACP) with initial value x = 0. Fix t ∈ [0, T ].
We claim that the function Ψt : (0, t) → L (H, E),
Z t
Ψt (r)h := S(s − r)Bh ds,
r

is stochastically integrable with respect to WH and

Z t Z t
Ψt (r) dWH (r) = U (s) ds. (8.5)
0 0

To see this, note that for all x∗ ∈ E ∗ the stochastic Fubini theorem gives
Z t Z tZ t
Ψt∗ (r)x∗ dWH (r) = B ∗ S ∗ (s − r)x∗ ds dWH (r)
0 0 r
Z tZ s Z t
= B ∗ S ∗ (s − r)x∗ dWH (r) ds = hU (s), x∗ i ds,
0 0 0

where the last identity follows from the assumption that U is a weak solution
and therefore satisfies (8.1). The claim now follows from Theorem 6.17.
Also, from Ψt (r)h ∈ D(A) and AΨt (r)h = S(t − r)Bh − Bh it follows that
AΨt : (0, t) → L (H, E) is stochastically integrable with respect to WH and
Z t Z t
AΨt (r) dWH (r) = (S(t − r)B − B) dWH (r) = U (t) − W B (t),
0 0

where in the second identity we used that WH (t)B ∗ x∗ = hW B (t), x∗ i.

Combining these facts it follows that Ψt is stochastically integrable as a
function from (0, t) to L (H, D(A)). It follows that the left hand side of (8.5)
defines a D(A)-valued Gaussian random variable. Moreover, as A is bounded
from D(A) to E, almost surely we have
Z t Z t Z t
A U (s) ds = A Ψt (r) dWH (r) = AΨt (r) dWH (r) = U (t) − W B (t).
0 0 0

This shows that U is a strong solution. ⊓

⊔
We may now apply the result of Exercise 5.4 as follows:
Corollary 8.11. Let E have type 2 and assume that B ∈ γ(H, E). Then the
problem (SACP) has a unique strong solution, and this solution is given by
the convolution (8.1).
It can be shown that this solution has a version with continuous trajecto-
ries; this follows from the Da Prato-Kwapień-Zabczyk factorisation principle
which will be discussed later on. It appears to be an open problem whether,
in the more general situation of Theorems 8.6 and 8.10, a solution (if it exists)
always has a continuous version.
116 8 Linear equations with additive noise I

8.5 Non-existence

In this section we present an example of a stochastic evolution equation driven

by a rank one Brownian motion which has no (weak or strong) solution.

Example 8.12. Let E = Lp (T), where T denotes the unit circle in the complex
plane with its normalized Lebesgue measure. We let A = d/dθ denote the
generator of the rotation (semi)group S on Lp (T), S(t)f (θ) = f (θ+t mod 2π).
Consider the stochastic Cauchy problem
(
dU (t) = AU (t) + φ dW, t ∈ [0, 2π],
(8.6)
U (0) = 0,

where W is a standard real Brownian motion and φ ∈ Lp (T) is a fixed element.

This problem has a weak solution if and only if the operator R := R2π :
L2 (T) → Lp (T) of Theorem 8.6 (with T = 2π) is γ-radonifying. Let (hn )∞n=1
be an orthonormal basis for L2 (T). For all N > M > 1, by Fubini’s theorem
and the Khintchine inequality we have
N
X p
Z 2π N
X p
E γn Rhn = E γn Rhn (θ) dθ
Lp (T) 0
n=M n=M
Z 2π  X
N  p2 X
N  12 p
hp |Rhn (θ)|2 dθ = |Rhn |2 .
0 Lp (T)
n=M n=M

Now,
N
X N Z
X 2π 2 N
X 2
|Rhn (θ)|2 = hn (t)φ(θ + t mod 2π) dt = [hn , φθ ]L2 (T)
n=M n=M 0 n=M

where φθ (t) := φ(θ + t mod 2π). Via an application of the Kahane-Khintchine

inequality we deduce that R ∈ γ(L2 (T), Lp (T)) if and only if φ ∈ L2 (T). In
particular, for p ∈ [1, 2) and φ ∈ Lp (T) \ L2 (T) the resulting initial value
problem has no weak solution.

It is not a coincidence that a nonexistence is obtained in the range p ∈ [1, 2)

only. Indeed, for p ∈ [2, ∞) the space Lp (T) has type 2, and therefore Corollary
8.11 guarantees the existence of a strong solution for (8.6).

8.6 Exercises

1. This exercise offers an alternative approach to the integration by parts

formula of Lemma 8.1. The starting point is the fact that if H is a real
 8.6 Exercises 117

Hilbert space, φ : [0, T ] → R is of bounded variation, and ψ : [0, T ] → H

is continuous, then
Z T Z T
ψ(t) dφ(t) = φ(T )ψ(T ) − φ(t) dψ(t),
0 0

where both integrals are interpreted as Riemann-Stieltjes integrals in H .

Let (W (t))t∈[0,T ] be a standard Brownian motion.
a) Show that the function ψ : [0, t] → L2 (Ω), ψ(t) := W (t), is continuous.
b) Deduce Lemma 8.1 from the above integration by parts formula.
2. Let (W (t))t∈[0,T ] be an E-valued Brownian motion. Show that there exists
a unique Gaussian covariance operator Q ∈ L (E ∗ , E) such that

EhW (s), x∗ ihW (t), y ∗ i = min{s, t}hQx∗ , y ∗ i

for all 0 6 s, t 6 T and x∗ , y ∗ ∈ E ∗ .

Hint: Consider Q := QT /T , where QT is the covariance of W (T ).
3. We consider the problem (SACP) with initial value x = 0 and assume
that it admits a weak solution U . Prove that U is a Gaussian process with
covariance
Z min{s,t}
EhU (s), x∗ ihU (t), y ∗ i = [B ∗ S ∗ (s − r)x∗ , B ∗ S ∗ (t − r)y ∗ ] ds
0

for all 0 6 s, t 6 T and x∗ , y ∗ ∈ E ∗ .

4. We consider the problem (SACP) with initial value x and assume that it
admits a weak solution U x .
a) Prove that the solvability of the problem (SACP) is independent of
the time T . More precisely, show that if (SACP) has a weak (resp.
strong) solution on some interval [0, T ], then it has a weak (resp.
strong) solution on every interval [0, T ].
Hint: Use the semigroup property and Theorem 8.6.
By a) and uniqueness, U x extends to a solution on [0, ∞). For f ∈ Cb (E)
and t > 0 we define the function P (t)f : E → R by

P (t)f (x) := Ef (U x (t)), x ∈ E.

b) Explain why for all f ∈ Cb (E) and t > 0 we have the identity
Z
Ef (U x (t)) = f (S(t)x + y) dµt (y),
E

where µt denotes the distribution of the random variable U 0 (t).

c) Deduce that P (t)f ∈ Cb (E).
118 8 Linear equations with additive noise I

d) Prove the identity

µt+s = µt ∗ S(t)µs ,
where ∗ denotes convolution and S(t)µs is the image measure of µs
under the operator S(t).
Hint: Use Fourier transforms and observe that for the covariances Qt
of U 0 (t), t > 0, we have the identity

Qt+s = Qt + S(t)Qs S ∗ (t).

e) Deduce that P = (P (t))t>0 is a semigroup of operators on Cb (E), in

the sense that P (0) = I and P (t)P (s) = P (t + s) for all t, s > 0.
f) Prove that for all x ∈ E and f ∈ Cb (E) we have

lim P (t)f (x) = f (x)

t→0

uniformly on compact subsets K of E.

Hint: By the remark in Exercise 6.4, the process
Z t
V x (t) := S(t)x + S(s)B dWH (s), t ∈ [0, T ],
0

has a continuous version (a proof will be given later in this course).

Now use b) together with the observation that for each fixed t ∈ [0, T ]
the random variables U x (t) and V x (t) are identically distributed.
Remark: By considering (real and imaginary parts of) trigonometric poly-
nomials of the form x 7→ exp(ihx, x∗ i) it is not hard to show that P fails
to be a C0 -semigroup on Cb (E) (and even on the closed subspace U Cb (E)
of all bounded uniformly continuous functions) unless A = 0.
5. In addition to the assumptions of the previous exercise, let us assume that
there exists a Borel probability measure µ∞ on E such that limt→∞ µt =
µ∞ in the sense that
Z Z
lim f (x) dµt (x) = f (x) dµ∞ (x)
t→∞ E E

for all f ∈ Cb (E).

a) Prove the identity
µ∞ = µt ∗ S(t)µ∞ .
b) Prove that µ∞ is an invariant measure in the sense that for all f ∈
Cb (E) and t > 0 we have
Z Z
P (t)f dµ∞ = f dµ∞ .
E E

c) Prove that P extends to a C0 -semigroup of contractions on the space

Lp (E, µ∞ ), 1 6 p < ∞.
 8.6 Exercises 119

Notes. The theory of (linear and non-linear) stochastic evolution equations in

Hilbert spaces dates back to the 1970s and was developed extensively through
the efforts of the Italian and Polish schools around Da Prato and Zabczyk.
A comprehensive overview is given in the monographs [27, 28] by these two
authors. Parts of the theory have been extended to (martingale-)type 2 spaces;
we refer to the review paper by Brzeźniak [15] and the references given there.
The results of Section 8.3 are taken from [84] and generalise known Hilbert
space results and improve the preliminary Banach space results of [16]. The
proof of theorem Theorem 8.6 essentially follows the Hilbert space proof in
[27]. The theory of adjoint semigroups was initiated by Phillips, who proved
Proposition 8.3 and noted as a consequence that E ⊙ = E ∗ if E is reflexive.
The equivalence of weak and strong solutions in the case where B is γ-
radonifying is taken from an unpublished note by Veraar.
The example in Section 8.5 is from [84]. Such examples cannot exist in
Hilbert spaces, due to Corollary 8.11.
The semigroup P of Exercises 4 and 5 is called the Ornstein-Uhlenbeck
semigroup associated with A and B. The literature on this class of semigroups
is extensive, with contributions by many mathematicians. Using Itô’s formula
it can be shown that the infinitesimal generator L of P is given, on a suitable
dense subspace of D(L) consisting of cylindrical functions, by
1
Lf (x) = Tr (BB ∗ D2 f (x)) + hAx, Df (x)i, x ∈ D(A),
2
where D denotes the Fréchet derivative and Tr the trace. The first term in the
right hand side is the ‘diffusion part’ corresponding to B WH and the second
is the ‘drift part’ corresponding to A.
The clever argument in part f) of Exercise 4 is due to Veraar. A self-
contained analytic proof can be found in see [42].
For a systematic account on invariant measures for stochastic evolution
equations we refer to Da Prato and Zabczyk [28].