The Annals of Probability

2006, Vol. 34, No. 4, 1451–1496

DOI: 10.1214/009117905000000800
c Institute of Mathematical Statistics, 2006
arXiv:math/0402307v2 [math.PR] 25 Sep 2006

LOWER ESTIMATES OF TRANSITION DENSITIES AND BOUNDS

ON EXPONENTIAL ERGODICITY FOR STOCHASTIC PDE’S

By B. Goldys and B. Maslowski1

University of New South Wales and Academy of Sciences of Czech Republic
A formula for the transition density of a Markov process defined
by an infinite-dimensional stochastic equation is given in terms of the
Ornstein–Uhlenbeck bridge and a useful lower estimate on the density
is provided. As a consequence, uniform exponential ergodicity and V -
ergodicity are proved for a large class of equations. We also provide
computable bounds on the convergence rates and the spectral gap for
the Markov semigroups defined by the equations. The bounds turn
out to be uniform with respect to a large family of nonlinear drift
coefficients. Examples of finite-dimensional stochastic equations and
semilinear parabolic equations are given.

1. Introduction. The aim of this paper is to study the ergodic properties

of solutions to a semilinear stochastic equation
p
dX x = (AX x + F (X x )) dt + Q dW,
(1.1)
X0x = x ∈ E,
considered in a separable Banach space E, where W is a cylindrical Wiener
process on a Hilbert space H such that E ⊂ H. Under the assumptions listed
below (see Section 2), this equation has a unique Markov solution (Xtx ) with
a unique invariant measure µ∗ .
Ergodic properties of solutions to infinite-dimensional stochastic differ-
ential equations have been extensively studied in recent years. The key
problems in this field are the existence and uniqueness of invariant mea-
sure and the rate of convergence of the time t distribution of the process
to the invariant measure. In the case of dim E < ∞ these questions have

Received February 2004; revised June 2005.

1
Supported in part by the ARC Discovery Grant DP0346406, the UNSW Faculty Re-
search Grant PS05345 and the GAČR Grants 201/01/1197 and 201/04/0750.
AMS 2000 subject classifications. 35R60, 37A30, 47A35, 60H15, 60J99.
Key words and phrases. Ornstein–Uhlenbeck bridge, stochastic semilinear system, den-
sity estimates, V -ergodicity, uniform exponential ergodicity, spectral gap.

This is an electronic reprint of the original article published by the

Institute of Mathematical Statistics in The Annals of Probability,
2006, Vol. 34, No. 4, 1451–1496. This reprint differs from the original in
pagination and typographic detail.
1
2 B. GOLDYS AND B. MASLOWSKI

been studied for a long time and the ergodic theory of finite-dimensional
diffusion processes is relatively well developed, see, for example, a classical
monograph [21]. In this paper we study the ergodic properties of a class of
ordinary and partial stochastic differential equations that includes stochastic
reaction–diffusion equations in bounded domains. First results on the exis-
tence and uniqueness of invariant measures for stochastic reaction-diffusion
equations were obtained in [12, 26, 42], see also [5], the monographs [6, 10]
and references therein. The rate of convergence to the invariant measure in
infinite dimensions became a subject of interest much later and still is not
well understood. Jacquot and Royer [24] proved exponential ergodicity for
a semilinear parabolic equation with bounded nonlinear drift, Shardlow [39]
applied the theory of Meyn and Tweedie to obtain V -uniform ergodicity for
some semilinear equations in Hilbert spaces. Hairer in [14] proved, under dif-
ferent sets of conditions, uniform exponential ergodicity for equations with
drifts growing faster than linearly. Exponential convergence to equilibrium in
a norm intermediate between the total variation metric and the Wasserstein
metric has been obtained in [31] for the stochastic Navier–Stokes equation.
A closely related problem of asymptotic behavior of the Markov semigroup
Pt φ(x) = Eφ(Xtx ) attracted much attention due to its importance in Mathe-
matical Physics. In particular, exponential convergence of the semigroup in
the spaces Lp (E, µ∗ ), p ∈ [1, ∞), and related questions of the existence of
the spectral gap and logarithmic Sobolev inequality have been studied by
numerous authors, see [1, 2, 3, 7, 8, 11, 18, 22, 43].
The aim of the present paper is to prove V -uniform (exponential) ergodic-
ity with V (x) = |x|E + 1 and, if the drift grows faster than linearly, uniform
exponential ergodicity, for equation (1.1). Our method allows us to find ex-
act bounds on convergence (i.e., to give explicit estimates for the rate of
exponential convergence in the total variation norm or V -variation norm).
In this respect, our results seem to be new even for finite-dimensional SDE’s
(which is also due to our method to estimate the transition density that, to
the best of our knowledge, has not been used in finite dimensions so far). If
the Markov semigroup (Pt ) is symmetric, we obtain explicit lower estimates
for the spectral gap in L2 (E, µ∗ ). Stronger results are obtained in the case
of a drift growing faster than linearly: for a symmetric Markov semigroup,
we show uniform estimates on the spectral gap in the spaces Lp (E, µ∗ ) for
all p ∈ [1, ∞) and in the nonsymmetric case, our estimates remain valid for
p > 1, in particular, in L2 (E, µ∗ ).
Unlike in the aforementioned papers, in the present paper a lower bound
measure and a suitable small set for a skeleton process are found explicitly
in terms of the lower estimates of transition densities and the constants in
an ultimate boundedness condition (or, in particular, a suitable Lyapunov
function). This enables us to apply earlier results on computable bounds for
 ERGODICITY FOR STOCHASTIC PDE’S 3

Markov chains, which are expressed in terms of lower bound measures, corre-
sponding small sets and constants from the Lyapunov–Foster geometric drift
condition [33]. The bounds turn out to be uniform with respect to a large
family of drift coefficients, which is important for proving continuous depen-
dence of invariant measures on parameter (cf. Section 8). We also believe
that this uniformity is an important tool for studying the Hamilton–Jacobi–
Bellman equation for the ergodic control problem. On the other hand, the
method employed here has its limitations. Our method strongly relies on
the Girsanov theorem and therefore, we need √ an assumption that F maps
the whole state space into the range of Q. Therefore, any extension to
other types of equations (like stochastic Burgers or Navier–Stokes equations)
would be difficult. Note, however, that in the two recent authors’ papers
[17] and [16] V -uniform ergodicity and spectral gap type results have been
proved for stochastic Burgers, 2D Navier–Stokes and more general reaction–
diffusion equations. Nonetheless, in these papers a different method is used
that allows us neither to find explicit bounds on the convergence constants
nor to show the uniformity of convergence with respect to coefficients.
An important tool for our proofs is a formula for the transition densities
that is derived in this paper. We use this formula to establish suitable lower
estimates on the densities which we believe are of independent interest. They
are obtained by means of the Girsanov theorem and the so called Ornstein–
Uhlenbeck bridge (or pinned Ornstein–Uhlenbeck process). Let us explain
the main idea of this approach.
Let (Ztx ) be an Ornstein–Uhlenbeck process on a separable Hilbert space
H. By this, we mean that (Ztx ) is a solution to a linear stochastic evolution
equation
p
dZtx = AZtx dt + Q dWt ,
(1.2)
Z0x = x ∈ H.
The Ornstein–Uhlenbeck bridge (Zbtx,y ) associated to the Ornstein–Uhlenbeck
process (Ztx ) is informally defined by the formula
P(Ztx ∈ B|Z1x = y) = P(Zbtx,y ∈ B), t < 1,
where x, y ∈ H and B ⊂ H is a Borel set. The importance of various types
of bridge processes for the study of transition densities of finite dimensional
diffusions is well recognised, see, for example, [23]. In infinite-dimensional
framework this concept was developed in [41] in order to study the regularity
of transition semigroups of diffusions on Hilbert spaces. In [28] and [29] an
Ornstein–Uhlenbeck bridge is introduced in order to obtain lower estimates
on the transition kernel of some semilinear stochastic evolution equations.
The basic idea is as follows. Using the equivalence of measures corresponding
4 B. GOLDYS AND B. MASLOWSKI

to Xtx and Ztx and the Girsanov formula, we can write the transition density
of the process Xtx in the form
d(T, x, y) = E(Φ(Z·x )|ZTx = y),
where Φ is a measurable functional defined on trajectories of the Ornstein–
Uhlenbeck process. This form of the density is not suitable for the uniform
estimates that are needed. Therefore, the conditional expectation is trans-
formed into a usual expectation with respect to the measure of the OU
bridge (Zbtx,y ) considered for t ∈ [0, T ]:
d(T, x, y) = EΦ(Zb·x,y ),
which enables us to find the uniform lower estimates. Let us note here a
technical difficulty caused by the fact that we can define the OU bridge for
y in a certain Borel subspace of measure one only, but this turns out to be
sufficient for our needs.
Precise formulations and hypothesis are given in the following Section 2.
In Sections 3 and 4 the properties of the OU bridge, which are needed in
the sequel, are established (some auxiliary results are deferred to the Ap-
pendix). The formula for transition densities is found and the lower estimates
are given in Section 5. These results are applied in Section 6 to establish
our main results, V -uniform ergodicity and uniform exponential ergodicity,
respectively, and find the computable bounds on respective constants. In
Section 7 the corollaries on the Lp (E, µ∗ ) exponential convergence and the
spectral gap are stated. Section 8 is devoted to some extensions and ap-
plications (continuous dependence of invariant measures on a parameter).
Examples (finite-dimensional nonlinear stochastic oscillator and stochastic
parabolic equations) are presented in Section 9.

2. Assumptions and notation. Let H = (H, |·|) be a real separable Hilbert

space and let E = (E, | · |E ) be a separable Banach space densely embedded
into H. In this paper we will study a stochastic semilinear equation
p
dXt = (AXt + F (Xt )) dt + Q dWt ,
(2.1)
X0 = x ∈ E,
where (Wt ) is a standard cylindrical Wiener process on H defined on a
stochastic basis (Ω, F, (Ft ), P) satisfying the usual conditions, A denotes a
linear operator on H generating a strongly continuous semigroup (St ) on
H and F is a nonlinear mapping E → E. The first assumption assures the
existence of an H-valued and strong Feller solution to the linear version of
(2.1), when F = 0; in this case we consider the linear equation
p
dZtx = AZtx dt + Q dWt ,
(2.2)
Z0x = x.
 ERGODICITY FOR STOCHASTIC PDE’S 5

The solution to equation (2.2) is given by formula

Z t p
(2.3) Ztx = St x + St−s Q dWs , t ≥ 0.
0

Hypothesis 2.1. The operator Q ≥ 0 is bounded and symmetric. For

each t > 0, a bounded operator,
Z t
Qt = Ss QSs∗ ds,
0

is of trace class. Moreover,

1/2
(2.4) im(St ) ⊂ im(Qt ), t > 0.

If Hypothesis 2.1 holds, then

(2.5) im(Qt ) = H, t > 0.
It is well known (cf. [9]) that (2.4) is equivalent to the strong Feller property
of the process (Ztx ). Moreover, (2.4) yields
1/2 1/2
(2.6) im(Qt ) = im(Q1 ), t > 0.
The next hypothesis assures that the Ornstein–Uhlenbeck process Z x defined
by equation (2.3) takes values in the Banach space E and is continuous in
E.

Hypothesis 2.2. (a) The part à of A in the space E,

à = A|dom(Ã), dom(Ã) = {y ∈ dom(A) ∩ E : Ay ∈ E},
generates a C0 -semigroup on E, which is again denoted by (St ).
(b) The process Z 0 is P-a.s. E-valued and E-continuous.

We further assume the following:

Hypothesis 2.3.
Z 1
−1/2
(2.7) kQt St Q1/2 kHS dt < ∞,
0

where kT kHS stands for the Hilbert–Schmidt norm of the operator T .

Assumption (2.7) is not standard. It is needed to obtain a formula for

the transition density (cf. Theorem 5.2). We will show that it is satisfied in
many important cases (cf. Lemma 3.3, Remark 3.4 and Section 9).
6 B. GOLDYS AND B. MASLOWSKI

In this paper we consider mild pathwise continuous solutions of (2.1).

A process X defined on a filtered probability space (Ω, F, (Ft ), P) is a solu-
tion to equation (2.1) on an interval [0, T ] if P(X· ∈ C(0, T : E)) = 1 and
Z t Z t p
(2.8) Xt = St x + St−r F (Xr ) dr + St−r Q dWr , t ∈ [0, T ], P-a.s.
0 0
Now we will formulate assumptions involving the nonlinear term F in equa-
tion (2.1).

Hypothesis 2.4. (a) The mapping F : E → E is Lipschitz continuous

on bounded sets of E. For eaxh x ∈ E, there exists a unique mild solution
X to equation (2.1). Moreover, X is a Markov process in E.
(b) im(F ) ⊂ im(Q1/2 ) and there exists a continuous function G : E → H
such that Q1/2 G = F and for some constants K, m > 0,
(2.9) |G(x)| ≤ K(1 + |x|m
E ), x ∈ E.

Remark 2.5. The assumption of local Lipschitz continuity of the map-

ping F is not necessary for our main results. It may be replaced by the
existence and uniqueness conditions for equation (2.1) and approximating
equations. Similarly, the mapping G need not be continuous. Only measura-
bility and the polynomial bound (2.9), are needed but this would make some
proofs technically more complicated.

Hypotheses 2.1–2.4 are standing assumptions of the paper and the results
will be enunciated without recalling them again. Obviously, the hypotheses
are used selectively (e.g., Hypothesis 2.4 is not needed for results on the
Ornstein–Uhlenbeck bridge).
Denote by B, P and bB, the Borel σ-algebra of E, the space of probability
measures on E and the space of bounded Borel functions on E, respectively.
Furthermore,
Pt ϕ(x) := Ex ϕ(Xt ), φ ∈ bB, x ∈ E, t ≥ 0,
and
P (t, x, Γ) := Pt 1Γ (x), x ∈ E, Γ ∈ B, t ≥ 0.
Let (Pt∗ ) denote the adjoint Markov semigroup, that is,
Z
(2.10) Pt∗ ν(Γ) := P (t, x, Γ)ν(dx), t ≥ 0, ν ∈ P, Γ ∈ B.
Γ
An invariant measure µ∗
∈ P is defined as a stationary point of the semi-
group (Pt∗ ), that is, Pt∗ µ∗ = µ∗ for each t ≥ 0. Obviously, Pt∗ ν is interpreted
as the probability distribution of Xt if X0 has the initial distribution is ν.
 ERGODICITY FOR STOCHASTIC PDE’S 7

In our main theorems on V -uniform ergodicity, exponential ergodicity

and spectral gap the solution to equation (2.1) is supposed to be ultimately
bounded. In order to illustrate which systems are covered, it may be useful
to formulate a growth condition on the nonlinear term F which will be
selectively used in some statements below (though it is not needed in our
general theorems). By h·, ·iE,E ∗ , we denote the duality between E and E ∗
and by ∂| · |E , the subdifferential of the norm | · |E . Suppose that there exist
k1 , k2 , k3 > 0, and s > 0 such that, for x ∈ dom(Ã) and x∗ ∈ ∂|x|E , we have
(2.11) hÃx + F (x + y), x∗ iE,E ∗ ≤ −k1 |x|E + k2 |y|sE + k3 , y ∈ E.
For example, if the mapping F : E → E is Lipschitz continuous on bounded
sets in E and Hypotheses 2.1 and 2.2 are satisfied, then the above condi-
tion implies existence of a unique mild solution to the equation (2.1) [i.e.,
Hypothesis 2.4(a)]. If, moreover, the moments of the Ornstein–Uhlenbeck
process Z 0 are bounded on [0, ∞) [condition (6.1) below], then there exists
an invariant measure for the corresponding Markov process.

3. Some properties of the Ornstein–Uhlenbeck process. We will use the

notation µxt for the probability distribution of Ztx and µt if x = 0. Obviously,
µxt is a Gaussian measure N (St x, Qt ). For simplicity of notation, we set
Zs := Zs0 , s ≥ 0. It is easy to check that, for s ≤ t
(3.1) EhZs , hihZt , ki = hSt−s Qs h, ki.

−1/2 1/2
Lemma 3.1. The operator Vt = Q1 S1−t Qt is bounded on H and
(3.2) kVt k < 1, t ∈ (0, 1].
Moreover,
(3.3) lim Vt∗ x = lim Vt x = x, x ∈ H.
t→1 t→1

Proof. Estimate (3.2) was proved in [35]. It follows from (3.2) and a
simple identity
∗
Q1 = Q1−t + S1−t Qt S1−t ,
that
1/2 1/2
(3.4) Q1−t = Q1 (I − Vt Vt∗ )Q1 .
To prove (3.3), we will show first that
(3.5) lim hVt x, yi = hx, yi, x, y ∈ H.
t→0
8 B. GOLDYS AND B. MASLOWSKI

−1/2
Indeed, for y ∈ im(Q1 ), we have
1/2 −1/2
lim hVt x, yi = limhS1−t Qt x, Q1 yi = hx, yi.
t→1 t→1
−1/2
For arbitrary y ∈ H, we may find a sequence (yn ) ⊂ im(Q1 ), such that
yn → y in H and then (3.2) yields
hVt x, y − yn i → 0,
uniformly in t ≤ 1, and (3.5) follows. Next, (3.4) yields
1/2 1/2
hQ1−t x, xi = h(I − Vt Vt∗ )Q1 x, Q1 xi, x ∈ H.
1/2
It follows that, for each y ∈ im(Q1 ), we have
lim (|y|2 − |Vt∗ y|2 ) = 0,
t→1

and since kVt∗ k < 1 for all t, we find that

lim |Vt∗ y| = |y|, y ∈ H.
t→1

Now, invoking (3.5), we obtain the first part of (3.3). It is enough to prove
the second part of (3.3) for x such that |x| = 1. In this case (3.5) implies
hVt x, xi → 1, and thereby, invoking (3.2),
1 = lim inf hVt x, xi ≤ lim inf |Vt x| ≤ lim sup |Vt x| ≤ 1.
t→1 t→1 t→1

Therefore, Vt x → 1 as t → 1. Now, taking into account (3.5), we obtain the

second part of (3.3). 

1/2 −1/2
Clearly, Vt∗ = Qt S1−t
∗ Q
1 and the operator
1/2
(3.6) Kt := Qt Vt∗
is of Hilbert–Schmidt type on H. Then the operator
H ∋ x → Kx(t) := Kt x ∈ L2 (0, 1; H)
is bounded.
Let µ denote the probability distribution of the process {Zt , t ∈ [0, 1]}
concentrated on L2 (0, 1; H) and let L : L2 (0, 1; H) → C(0, 1; H) be defined
by the formula
Z t
Lu(t) = St−s Q1/2 u(s) ds.
0
The space im(L) endowed with the norm
kφk = inf{|u| : u ∈ L2 (0, 1; H), Lu = φ}
 ERGODICITY FOR STOCHASTIC PDE’S 9

may be identified with reproducing kernel Hilbert space of the measure µ,

see [9]. For any t ∈ [0, 1), we define an unbounded H-valued operator
−1/2 1/2
Bt x = Q1/2 S1−t
∗
Q1 x, x ∈ im(Q1 ),
and an unbounded operator
Bx(t) = Bt x, t ∈ [0, 1),
taking values in C([0, 1), H).
The following lemma is crucial for the rest of the paper. Let us recall
that the operator V : H → E, where E is a Banach space, is said to be γ-
radonifying if it transforms any cylindrical Gaussian measure on H into a
Radon Gaussian measure on E.

Lemma 3.2. (a) For every t ∈ [0, 1), the operator Bt with the domain
1/2
dom(B) = Q1 (H) extends to a Hilbert–Schmidt operator Bt : H → H and
Z 1
(3.7) kBt kHS dt < ∞.
0

1/2
(b) The operator B with the domain dom(B) = Q1 (H) extends to a
bounded operator B : H → L2 (0, 1; H) and
(3.8) |Bx|L2 (0,1;H) = |x|H , x ∈ H.
(c) We have K = LB and the operator K : H → C(0, 1; E) is γ-radonifying.
−1/2 1/2
Proof. (a) Note first that kQ1 Q1−t k ≤ 1 and by (2.4), the operator
−1/2
Q1−t S1−t is bounded. Therefore, the operator
−1/2 1/2 −1/2 −1/2
(Q1 Q1−t Q1−t S1−t Q1/2 )∗ = Q1/2 S1−t
∗ Q
1
is bounded. Moreover, taking (2.7) and (2.6) into account, we obtain, for a
certain C > 0,
Z 1 Z 1
−1/2
kBt∗ kHS dt ≤ C kQt St Q1/2 kHS dt < ∞
0 0
and (3.7) follows.
For any h ∈ H, we have
Z 1
1/2
|Q1 h|2 = |Q1/2 S1−t
∗
h|2 dt,
0
−1/2
and therefore, for h = Q1 x, we obtain
Z 1
(3.9) |x|2 = |Bx(t)|2 dt.
0
10 B. GOLDYS AND B. MASLOWSKI

1/2
Using the density of Q1 (H) in H, we can extend (3.9) to the whole of H
and (3.8) follows.
1/2
(c) For x ∈ Q1 (H), we have
∗ −1/2
Kt x = Qt S1−t Q1 x
Z t
∗ ∗ −1/2
= St−s QSt−s S1−t Q1 x ds
0
(3.10) Z t
∗ −1/2
= St−s QS1−s Q1 x ds
0
Z t
= St−s Q1/2 Bx(s) ds = L(Bx)(t),
0

for all t ∈ [0, 1]. By (b), this identity extends to all x ∈ H and we find
that K = LB on H. By Hypothesis 2.2, we have µ(C(0, 1; E)) = 1, hence,
L : L2 (0, 1; H) → C(0, 1; E) is γ-radonifying and therefore, K = LB : H →
C(0, 1; E) is γ-radonifying as well. 

We have left open the question of effective verification of Hypothesis 2.3.

This is addressed in the following lemma.

Lemma 3.3. Assume that either:

(i) dim(H) < ∞ or
1+α
(ii) there exist α ∈ (0, 1) and β < 2 such that
Z 1
(3.11) t−α kSt Q1/2 k2HS dt < ∞
0

and
−1/2 c
(3.12) kQt St k ≤ .
tβ
Then Hypothesis 2.3 is satisfied.

Proof. The proof of (i) extends a classical controllability result from

[38] and may be found in [25].
Assume that (ii) holds. By Hypothesis 2.1,
−1/2 −1/2 1/2 −1/2
Qt St Q1/2 = (Qt Qt/2 )(Qt/2 St/2 )St/2 Q1/2 ,
−1/2 1/2
where kQt Qt/2 k ≤ 1 and thereby,
−1/2 −1/2
kQt St Q1/2 kHS ≤ kQt/2 St/2 kkSt/2 Q1/2 kHS .
 ERGODICITY FOR STOCHASTIC PDE’S 11

Therefore, for a certain c1 > 0,

Z 1
−1/2
kQt St Q1/2 kHS dt
0
Z 1 1/2 Z 1 1/2
−1/2
≤ tα kQt/2 St/2 k2 dt t−α kSt/2 Q1/2 k2HS dt
0 0
Z 1 1/2 Z 1 1/2
1 −α
≤ c1 dt t kSt Q1/2 k2HS dt ,
0 t2β−α 0
and (2.7) follows. 

Remark 3.4. Conditions (3.11) and (3.12) are well known and often
used in the theory of SPDE’s. Condition (3.11) is a standard assumption
that implies the existence of an H-continuous version of the OU process
(Ztx ), while (3.12) is closely related to the existence and integrability of the
gradient of the OU transition semigroup (cf. [9] for details). Hypothesis 2.3
will be checked in more specific cases in Section 9.

4. Ornstein–Uhlenbeck bridge. In Lemma A.2 [applied with H1 = L2 (0, 1;

−1/2
H), T = K and C = Q1 ] an extension of the operator Kt Q1 to a measur-
−1/2
able set M ⊂ H, µ1 (M) = 1 is defined. We use the notation Kt Q1 for
−1/2
this extension in the present section. Note first that Q1 Z1 is a cylindrical
Gaussian random variable on H and therefore, by Lemma 3.2(b), the process
−1/2
(Kt Q1 Z1 ) is well defined and has E-continuous modification. Therefore,
we can define an E-valued process
−1/2
Zbt = Zt − Kt Q1 Z1 , t ∈ [0, 1), and Zb1 = 0,
which has an E-continuous modification for t < 1.

Proposition 4.1. (a) The H-valued Gaussian process (Zbt ) is indepen-

dent of Z1 .
b t of Z
(b) The covariance operator Q bt is given by
1/2
b t = Q (I − V ∗ Vt )Q , 1/2
(4.1) Q t t t t ∈ [0, 1).
(c) The process (Zbt ) is continuous in E for t ∈ [0, 1].
1/2
Proof. (a) For h, k ∈ im(Q1 ), (3.1) yields
−1/2
EhZbt , hihZ1 , ki = EhZt , hihZ1 , ki − EhKt Q1 Z1 , hihZ1 , ki
−1/2
= hS1−t Qt h, ki − hQ1 Q1 Kt∗ h, ki
= hS1−t Qt h, ki − hS1−t Qt h, ki = 0
12 B. GOLDYS AND B. MASLOWSKI

and therefore, the process (Zbt ) and Z1 are independent.

(b) It follows from (a) that
b t + Kt K ∗ .
Qt = Q t

Hence, the definition of Kt and Vt yields

1/2
b t = Qt − Qt S ∗ Q−1 S1−t Qt = Q (I − V ∗ Vt )Q . 1/2
Q 1−t 1 t t t

(c) Using (3.2), we find easily that

(4.2) b t ) = 0.
lim tr(Q
t→0

To prove that

(4.3) b t ) = 0,
lim tr(Q
t→1

we note first that

b t ) = tr((I − V ∗ Vt )(Qt − Q1 )) + tr((I − V ∗ Vt )Q1 ).
tr(Q t t

Next, it is easy to see that

0 ≤ lim tr((I − Vt∗ Vt )(Q1 − Qt )) ≤ lim tr(Q1 − Qt ) = 0.

t→1 t→1

We have also
∞
X 1/2
tr((I − Vt∗ Vt )Q1 ) = tr(Q1 ) − |Vt Q1 ek |2 ,
k=1

and (4.3) follows from (3.3), (3.2) and the dominated convergence. Since the
process (Zbt ) has E-continuous version by Lemma 3.2(b), it follows that

0 = lim Zbt = Zb1 .

t→1

This fact completes the proof of continuity. 

Proposition 4.2. There exists a Borel subspace M ⊂ H such that

µ1 (M) = 1 and for all x ∈ H and y ∈ M, the H-valued Gaussian process
−1/2
(4.4) Zbtx,y = Ztx − Kt Q1 (Z1x − y)

is well defined for all t ∈ [0, 1). Moreover,

−1/2 −1/2
(4.5) Zbtx,y = St x − Kt Q1 S1 x + Kt Q1 y + Zbt , P-a.s.
 ERGODICITY FOR STOCHASTIC PDE’S 13

Proof. By Lemma A.3, we can choose a measurable linear space M

−1/2
such that Kt Q1 is linear on M with µ1 (M) = 1 and the mapping (t, y) →
−1/2 1/2
Kt Q1 y is measurable. By (2.4) we have S1 x ∈ im(Q1 ) and therefore,
−1/2
Kt Q1 (Z1x − y) is well defined for any y ∈ M. Clearly, Zbtx,y may be rewrit-
ten in the form (4.5). 

The process (Zbtx,y ) defined in Proposition 4.2 will be called an Ornstein–

Uhlenbeck bridge on H (connecting points x ∈ H and y ∈ M). We will
bx,y the law of the process {Z
denote by µ b x,y : t ∈ [0, 1]}.
t

Theorem 4.3. There exists a Borel subspace M ⊂ E with µ1 (M) = 1,

such that the process (Zbtx,y ) has E-continuous version for each x ∈ E and
y ∈ M. Moreover, there exists a measurable mapping U : M → R+ and a
random variable k, such that
(4.6) kZb x,y kC(0,1;E) ≤ k(1 + |x|E + U (y)), x ∈ E, y ∈ M,
and
(4.7) EkZb x,y knC(0,1 : E) ≤ L(n)(1 + |x|nE + (U (y))n )
for each n ∈ N, x ∈ E and y ∈ M, where L(n) is a constant depending on n
only.

Proof. It was already shown in Proposition 4.1 that the process (Zbt )
has trajectories in C(0, 1; E) and we have
(4.8) k1 = sup |Zbt |E < ∞, P-a.s.
t≤1

Since (Zbt ) is a Gaussian process, we obtain, for any m > 0,

(4.9) k2 (m) = E sup |Zbt |m
E < ∞.
t≤1
−1/2
The same argument shows that the process t → Kt Q1 y has trajecto-
ries in C(0, 1; E) for every y ∈ M, where M is given by Proposition 4.2
(possibly, after excluding a zero µ1 -measure set). By the strong Feller prop-
1/2
erty we have S1 x ∈ im(Q1 ) ⊂ M, so it follows from Proposition 4.2 that
bx,y (C(0, 1; E)) = 1 for x ∈ E and y ∈ M. Furthermore, using the notation
µ
−1/2 −1/2
from Lemma 3.2, we have Kt Q1 S1 x = K(Q1 S1 x)(t), t ∈ [0, 1]. By (2.4),
−1/2
the operator Q1 S1 is bounded and it is easy to see that (3.3) together with
−1/2
(3.6) yields continuity of the mapping Kt Q1 S1 : E → C(0, 1; E). Hence,
−1/2
setting U (y) = kKQ1 ykC(0,1;E) and taking into account (4.8) and (4.9),
we obtain both (4.6) and (4.7) for all n > 0. 
14 B. GOLDYS AND B. MASLOWSKI

The following theorem justifies the intuitive notion of the OU bridge

(Zbtx,y ) given in the Introduction.

Theorem 4.4. Let Φ : C(0, 1; E) → R be a Borel mapping such that, for

x ∈ E,
E|Φ(Z x )| < ∞.
Then
E(Φ(Z x )|Z1x = y) = EΦ(Zb x,y ), µ1 -a.e.

Proof. By Hypothesis 2.2 and Theorem 4.3, the processes (Ztx ) and
−1/2
(Zbtx,y ) are concentrated on C(0, 1; E) and (Kt Q1 (S1 x − y)) ∈ E. More-
−1/2
over, the processes (Zbtx,y ) and (Kt Q1 Z1x ) are independent by Proposi-
tion 4.1. Therefore, using well-known properties of conditional expectations,
we obtain
−1/2
E(Φ(Z x )|Z1x = y) = E(Φ(Zb x,y + Kt Q1 (Z1x − y))|Z1x = y)
= EΦ(Zb x,y ), µ1 -a.e.
for any x ∈ E. 

Let
Z 1
Yu = S1−s Q1/2 dWs , u ≤ 1,
u
and
−1/2
Hu = Q1−u S1−u Q1/2 , u < 1.

Lemma 4.5. For all u ∈ [0, 1], we have

(4.10) Yu = Q1−s Q−1 b
1 Z1 − S1−s Zs , P-a.s.,
where Q1−s Q1−1 is bounded for all s ∈ [0, 1].

Proof. We have
Z t 
−1/2 1/2 −1/2
Kt Q1 Z1 = St−s Q Hs∗ ds Q1 Z1
0
and
Z t 
−1/2 −1/2
S1−t Kt Q1 Z1 = S1−s Q1/2 Hs∗ ds Q1 Z1
0

= (Q1 − Q1−t )Q−1

1 Z1
= Z1 − Q1−t Q−1
1 Z1 ,
 ERGODICITY FOR STOCHASTIC PDE’S 15

and thereby,
−1/2
Z1 − S1−t Kt Q1 Z1 = Q1−t Q−1
1 Z1 .

Therefore, by definition of Zbt , we obtain

Ys = Z1 − S1−s Zs
−1/2 −1/2
= Z1 − S1−s (Zs − Ks Q1 Z1 ) − S1−s Ks Q1 Z1
= Z1 − S1−s Zbs − (Z1 − Q1−s Q−1
1 Z1 )

= Q1−s Q−1 b
1 Z1 − S1−s Zs . 

Since the operator-valued function t → Qt is continuous in the weak

operator topology and all the operators Qt are compact for t > 0, there
exists a measurable choice of eigenvectors {ek (t) : k ≥ 1} and eigenvalues
{λk (t) : k ≥ 1}. For each n ≥ 1 we define a process
n
X 1
αnu = p hYu , ek (1 − u)iHu∗ ek (1 − u).
k=1
λk (1 − u)

Lemma 4.6. There exists a measurable stochastic process (αu ) defined

on [0, 1) such that, for each a < 1,
Z a
(4.11) lim E |αnu − αu |2 du = 0
n→∞ 0
and for each h ∈ H and a < 1, the series
∞
X 1
(4.12) hαu , hi = p hYu , ek (1 − u)ihek (1 − u), Hu hi
k=1
λk (1 − u)
converges in L2 (0, a) in mean square. Moreover, if 0 ≤ u ≤ v < 1, then, for
all h, k ∈ H,
−1/2 1/2
(4.13) Ehαu , hihαv , ki = hHu h, Q1−u Q1−v Hv ki,
−1/2 1/2
where the operator Q1−u Q1−v is bounded. Finally,
Z 1
(4.14) E |αu | du < ∞.
0

−1/2
In what follows we will use the notation hαu , hi = hQ1−u Yu , Hu hi.

Proof of Lemma 4.6. For u ≤ v ≤ 1,

(4.15) EhYu , hihYv , ki = hQ1−v h, ki, h, k ∈ H.
16 B. GOLDYS AND B. MASLOWSKI

Therefore,
n
X 1
Ehαnu − αm 2
u , hi = EhYu , ek (1 − u)i2 hek (1 − u), Hu hi2
j=m+1
λ k (1 − u)
(4.16)
n
X
= hek (1 − u), Hu hi2 −→ 0,
n,m→∞
j=m+1

hence, the process

∞
X 1
hαu , hi = p hYu , ek (1 − u)ihek (1 − u), Hu hi
k=1
λk (1 − u)
(4.17)
−1/2
= hQ1−u Yu , Hu hi
is well defined and measurable for each h ∈ H and u < 1. Let Pn be an or-
−1/2
thogonal projection on lin{ek (1−v) : k ≤ n} and Hun = Pn Hu . Then Q1−u Hun
−1/2
is bounded on H and we may define αnu = (Q1−u Hun )∗ Yu . By (4.15),
−1/2 −1/2
Ehαnu , hihαnv ki = hQ1−v Q1−u Hun h, Q1−v Hvn ki
−1/2 1/2
= hHun h, Q1−u Q1−v Hvn ki.
−1/2 1/2
By (2.6), the operator Q1−u Q1−v is bounded and, therefore,
−1/2 −1/2
EhQ1−u Yu , Hu hihQ1−v Yv , Hv ki = lim Ehαnu , hihαnv , ki
n→∞
−1/2 1/2
= hHu h, Q1−u Q1−v Hv ki.
It follows from (4.13) that
Ehαu , hi2 = |Hu h|2 ,
hence,
E|αu |2 = kHu k2HS < ∞, u < 1,
and by Hypothesis 2.3,
Z 1
E |αu | du < ∞.
0

Then (4.16) and the dominated convergence yield

Z a
lim E|αnu − αm 2
u | du = 0.
n,m→∞ 0

As a consequence, we find that (4.11) holds for any a ∈ (0, 1). 

 ERGODICITY FOR STOCHASTIC PDE’S 17

Lemma 4.7. The cylindrical process

Z t
ζt = Wt − αu du, t ≤ 1,
0

is a standard cylindrical Wiener process on H independent of Z1 .

Proof. We need to show that, for any h ∈ H, the process

Z t
−1/2
hζt , hi = hWt , hi − hQ1−u Yu , Hu hi
0

is a real-valued Wiener process. Let h, k ∈ H. We will show first that, for

r < t < 1,

(4.18) Ehζt − ζr , hihζr , ki = 0.

We have

Ehζt − ζr , hihζr , ki
Z r
−1/2
= −EhWt − Wr , hi hQ1−u Yu , Hu ki du
0
Z t
−1/2
− EhWr , ki hQ1−u Yu , Hu hi du
r
Z t Z r 
−1/2 −1/2
+E hQ1−u Yu , Hu hi du hQ1−u Yu , Hv ki dv
r 0
= −I1 − I2 + I3 .

We will consider I1 first. Taking into account that the series (4.17) is mean-
square convergent, for each u ∈ (0, 1), we obtain
−1/2
EhWt − Wr , hihQ1−u Yu , Hu ki
∞
X hen (1 − u), Hu ki
= p E(hYu , en (1 − u)ihWt − Wr , hi).
n=1 λn (1 − u)

Next, for u ≤ r,

E(hYu , en (1 − u)ihWt − Wr , hi)

Z 1 Z t
=E hQ1/2 S1−s
∗
en (1 − u), dWs i hh, dWs i
u r
Z t
= hS1−s Q1/2 h, en (1 − u)i ds,
r
18 B. GOLDYS AND B. MASLOWSKI

and therefore,
−1/2
EhWt − Wr , hihQ1−u Yu , Hu ki
∞
X Z
hen (1 − u), Hu ki t
= p hS1−s Q1/2 h, en (1 − u)i ds
n=1 λn (1 − u) r
Z ∞
!
t X hen (1 − u), Hu ki
= p hS1−s Q1/2 h, en (1 − u)i ds
r n=1 λn (1 − u)
Z t
−1/2
= hQ1−u S1−s Q1/2 h, Hu ki ds
r
Z t
−1/2 1/2
= hQ1−u Q1−s Hs h, Hu ki ds
r

and
Z rZ t
−1/2 1/2
(4.19) I1 = hQ1−u Q1−s Hs h, Hu ki ds du,
0 r

−1/2 1/2
where the operator Q1−u Q1−s is bounded. By similar arguments, we find
that I2 = 0 and (4.13) yields
Z rZ t
−1/2 1/2
I3 = hQ1−v Q1−u Hu h, Hv ki du dv
0 r

and in view of (4.19), I1 = I3 , and since I2 = 0, (4.18) follows. We will show,

that for h ∈ H and t < 1,

(4.20) Ehζt , hi2 = t|h|2 .

We have
Z t
−1/2
Ehζt , hi2 = t|h|2 − 2EhWt , hi hQ1−u Yu , Hu hi
0
Z t Z t 
−1/2 −1/2
+E hQ1−u Yu , Hu hi du hQ1−v Hv Yv , ki dv
0 0
2
= t|h| − 2J1 + J3 .

Proceeding in the same way as in the computation of I1 and I3 , we obtain

Z t
−1/2
J1 = EhWt , hihQ1−u Yu , Hu hi du
0
(4.21) Z tZ t
−1/2 1/2
= hQ1−u Q1−s Hs h, Hu hi ds du.
0 u
 ERGODICITY FOR STOCHASTIC PDE’S 19

Invoking again (4.13), we obtain

Z tZ t
−1/2 −1/2
J3 = EhQ1−u Yu , Hu hihQ1−v Yv , Hv hi du dv
0 0
Z tZ v
−1/2 1/2
= hHu h, Q1−u Q1−v Hv hi du dv
0 0
Z tZ t
−1/2 1/2
+ hQ1−v Q1−u Hu h, Hv hi du dv.
0 v

Since t < 1 and the functions under the integrals are continuous, we can
change the order of integration in the first integral and obtain
Z tZ t
−1/2 1/2
J3 = hHu h, Q1−u Q1−v Hv hi dv du
0 u
Z tZ t
−1/2 1/2
+ hQ1−v Q1−u Hu h, Hv hi du dv.
0 v

Hence, J3 = 2J1 and (4.20) follows. Combining (4.18) and (4.20), we find
that, for s, t < 1,

Ehζs , hihζt , ki = min(s, t)hh, ki, h, k ∈ H.

Since

sup E|hζt , hi| = |h|2 , h ∈ H,

t<1

there exists a cylindrical random variable ζ1 such that

lim hζt , hi = hζ1 , hi, h ∈ H,

t→1

for all t ≤ 1. Therefore, (ζt ) is a cylindrical Brownian motion for t ∈ [0, 1]. It
remains to show that, for any t < 1,

(4.22) Ehζt , hihZ1 , ki = 0, h, k ∈ H.

By definition of ζt , it is enough to show that

Z t
(4.23) EhWt , hihZ1 , ki = E hαu , hihZ1 , ki du.
0

Now, we have
Z t
(4.24) EhWt , hihZ1 , ki = hS1−u Q1/2 h, ki du.
0
20 B. GOLDYS AND B. MASLOWSKI

Invoking (4.10) and using the fact that Z1 and (Zbt ) are independent, we
obtain
Z t
E hαu , hihZ1 , ki du
0
Z t
−1/2
= EhQ1−u Yu , Hu hihZ1 , ki du,
0
Z t
−1/2
(4.25) EhQ1−u Q1−u Q−1
1 Z1 , Hu hihZ1 , ki du
0
Z t
1/2
= hQ1−u Hu h, ki du
0
Z t
= hS1−u Q1/2 h, ki du.
0

Comparing (4.24) and (4.25), we obtain (4.23) and the lemma follows. 

Remark 4.8. Lemma 4.7 allows to define the Ornstein–Uhlenbeck

bridge (Zbtx,y ) as a unique solution of a certain linear stochastic evolution
equation. As it is not needed in this paper, it is omitted, see [15] for details.

Proposition 4.9. Let

−1/2 −1/2
B1 (s) = (Q1−s S1−s Q1/2 )∗ Q1−s S1−s ,
−1/2 −1/2
B2 (s) = (Q1 S1−s Q1/2 )∗ Q1 S1 ,
−1/2 −1/2 1/2
B3 (s)y = (Q1 S1−s Q1/2 )∗ Q1 y, y ∈ im(Q1 ), s ∈ (0, 1).

Then
Z 1
(4.26) E |B1 (s)Zbs | ds < ∞,
0
Z 1
−1/2
(4.27) |B2 (s)x|2 ds = |Q1 S1 x|2 , x ∈ H.
0

Moreover, there exists a Borel subspace M ⊂ H with µ1 (M) = 1 such that

B3 extends to a linear mapping B3 : M → L1 (0, 1; H), that is,
Z 1
(4.28) |B3 (s)y| ds < ∞, y ∈ M,
0

−1/2 −1/2
and B3 (s)Z1 has the covariance (Q1 S1−s Q1/2 )∗ (Q1 S1−s Q1/2 ) for each
s ∈ [0, 1).
 ERGODICITY FOR STOCHASTIC PDE’S 21

Proof. Condition (4.26) is a reformulation of (4.14) in Lemma 4.6,

where it was also shown that the function t → B1 (t)Zbt is integrable. Invoking
the definition of the operator B and Lemma 3.2, we obtain
Z 1 Z 1
−1/2 −1/2
(4.29) |B2 (s)x|2 ds = |BQ1 S1 x(s)|2 ds = |Q1 S1 x|2 ,
0 0

and (4.27) follows.

1/2 −1/2
For y ∈ im(Q1 ), we have B3 (t)y = BQ1 y and for any a < 1,
Z a
kBt k2HS dt < ∞,
0

and taking (3.7) into account, we may apply Lemma A.3, which yields the
desired result. 

5. Transition density of semilinear stochastic evolution equation. In this

section a formula for transition densities defined by equation (2.1) will be
derived and some useful lower estimates on transition densities will be es-
tablished.

Proposition 5.1. Assume that equation (2.1) has an invariant measure

µ∗ ∈ P. Then
(5.1) kPt∗ ν − µ∗ kvar → 0, t → ∞, ν ∈ P,
where k · kvar denotes the total variation of measures. Furthermore, for each
x ∈ E and T > 0, the measures P (T, x, ·) and µxT are equivalent and for µT
a.e. y,
dP (T, x, ·)
(y)
dµxT
(5.2)  Z Z  
T 1 T

= E exp hG(Ztx ), dWt i − |G(Ztx )|2 dt ZTx = y .
0 2 0

Proof. By Hypothesis 2.1, the Ornstein–Uhlenbeck process is strongly

Feller, therefore, its distributions (µxt ) are equivalent for x ∈ H, t > 0. If (5.2)
holds for each T > 0 and x ∈ E, we have the equivalence P (T, x, ·) ∼ µxT ,
hence, (P (T, x, ·))T >0,x∈E are equivalent, as well and the convergence (5.1)
follows from well-known results (cf. [37, 40]).
It remains to prove (5.2), which follows from the Girsanov theorem (see,
e.g., [9], Theorem 10.4). In order to apply this result, we must verify (taking,
for simplicity, T = 1)
(5.3) E exp ρ(Z x ) = 1,
22 B. GOLDYS AND B. MASLOWSKI

where
Z 1 Z 1
(5.4) ρ(Z x ) := hG(Zsx ), dWs i − 12 |G(Zsx )|2 ds.
0 0

For n ≥ 1, set

F
(x),  if |x|E ≤ n,
(5.5) Fn (x) = nx
F , if |x|E > n,
|x|E
and let Gn be defined by Fn (x) := Q1/2 Gn (x). Obviously the approximating
equations
p
dXn (t) = (AXn (t) + Fn (Xn (t))) + Q dWt ,
(5.6)
Xn (0) = x,
have uniquely defined solutions P-a.s. in C(0, 1; E) and denoting by P̃X , P̃Xn
and P̃Z x the distributions in C(0, 1; E) of X, Xn and Z x , respectively, we
have
 
(5.7) lim P sup |Xn (t) − X(t)|E > 0 = 0.
n→∞ t∈[0,1]

Hence,
(5.8) kP̃Xn (·) − P̃X kvar → 0, m → ∞,

thus, (P̃Xn ) is a Cauchy sequence in the metric of total variation. Therefore,

the sequence of densities
dP̃Xn
(5.9) = exp ρn (Z x ),
dP̃Z x
where
Z 1 Z 1
x
(5.10) ρn (Z ) := hGn (Zsx ), dWs i − 12 |Gn (Zsx )|2 ds,
0 0

is a Cauchy, hence, convergent, sequence in L1 (Ω). As Gn is bounded for

each n, obviously E exp ρn (Z x ) = 1, so it remains to identify the L1 (Ω)-limit
of exp ρn with exp ρ. Clearly, Gn → G pointwise and |Gn (x)| ≤ K(1 + |x|m E ),
therefore,
Z 1 Z 1
2
 
E hGn (Zsx ), dWs i − hG(Zsx ), dWs i
0 0
Z 1
=E |Gn (Zsx ) − G(Zsx )|2 ds → 0
0
 ERGODICITY FOR STOCHASTIC PDE’S 23

by the dominated convergence theorem and Hypothesis 2.2. Similarly, we

have
Z 1 Z 1
|Gn (Zsx )|2 ds → |G(Zsx )|2 ds, P-a.s.,
0 0
so we obtain (possibly, for a subsequence) exp ρn (Z x ) → exp ρ(Z x ) P-a.s.,
which completes the proof of (5.2). 

We will now state one of our main results, which provides a formula for
dP ∗ ν
the density dµt t for a given time t > 0 (we may take t = 1). It follows from
the Fubini theorem that the density has the form
Z
dP1∗ ν dP (1, x, ·)
(y) = (y)ν(dx)
dµ1 E dµ1
(5.11) Z
dP (1, x, ·) dµx
= x (y) 1 (y)ν(dx), µ1 -a.e.,
E dµ1 dµ1
provided the product of densities inside the integral on the r.h.s. is (x, y)-
measurable. As mentioned in the preceding proof, the Gaussian measures
µx1 and µ1 are equivalent with the density given by the Cameron–Martin
formula
dµx
g(x, y) := 1 (y)
dµ1
(5.12)  
−1/2 −1/2 1 −1/2 2
= exp hQ1 S1 x, Q1 yi − |Q1 S1 x| , x ∈ E,
2
for µ1 -almost all y ∈ E.

Theorem 5.2. For each ν ∈ P, we have

Z
dP1∗ ν
(5.13) (y) = h(x, y)g(x, y)ν(dx), µ1 -a.e.,
dµ1 E
where g is defined by the Cameron–Martin formula (5.12), and for x ∈ E
and µ1 -almost all y ∈ E,

h(x, y) := E exp ρ(Zb x,y )
(5.14) Z 
1
b x,y b
− hG(Zs ), B1 (s)Zs + B2 (s)x − B3 (s)yi ds ,
0
where B1 , B2 and B3 are defined in Proposition 4.9. In particular, for each
x ∈ E, we have
dP (1, x, ·)
(5.15) (y) = h(x, y)g(x, y), µ1 -a.e.
dµ1
24 B. GOLDYS AND B. MASLOWSKI

Proof. Since both g and h are (x, y)-measurable, taking into account
(5.11) and (5.12), we only have to prove that
dP (1, x, ·)
(5.16) (y) = h(x, y), x ∈ E, µ1 -a.e.
dµx1
Assume at first that the mapping G is bounded and let tki := ki for k ∈
N, i = 0, 1, . . . , k, that is, ∆k = {tk0 , tk1 , . . . , tkk } are equidistant divisions of
the interval [0, 1], tk0 = 0, tkk = 1, tki+1 − tki = 1/k (for brevity, the dependence
of tki on k will be suppressed in the notation). Set, for k ≥ 1,
k−1
X Z 1
(5.17) ρk (Z x ) := hG(Ztxi ), Wti+1 − Wti i − 1
2 |G(Zsx )|2 ds.
i=0 0

The mapping G is assumed to be bounded, thus,

Z Z 1 k−1 2

1
X 
(5.18) |G(Zsx )|2 ds +  G(Zti )1[ti ,ti+1 ] (s) ds ≤ 2 sup |G|2 < ∞,
x
0 0  
i=0

so the random variables exp ρk (Z x ) are uniformly integrable on Ω. Clearly,

exp ρk (Z x ) → exp ρ(Z x ) P-a.s. (possibly, for a subsequence), and therefore,
(5.19) exp ρ(Z x ) = lim exp ρk (Z x ) in L1 (Ω),
k→∞

which in view of (5.2) yields

dP (1, x, ·)
(5.20) (y) = E(exp ρ(Z x )|Z1x = y) = lim E(exp ρk (Z x )|Z1x = y)
dµx1 k→∞

for µx1 -almost all y ∈ H. On the other hand, in terms of the cylindrical
Wiener process ζt defined in Lemma 4.7, we have
k−1
X
ρk (Z x ) = hG(Ztxi ), ζti+1 − ζti i
i=0
(5.21) Z Z
ti +1 1
−1/2
+ hG(Ztxi ), Hs∗ Q1−s Ys i ds − 12 |G(Zsx )|2 ds.
ti 0

Invoking (4.10), we obtain

−1/2 −1/2 1/2 −1/2
− Hs∗ Q1−s Ys = H ∗ Q1−s S1−s Zbs − Hs∗ Q1−s Q1 Z1
(5.22)
= B1 (s)Zbs − B3 (s)Z1
for s ∈ (0, 1) P-a.s. [note that both terms on the r.h.s. of (5.22) are well
defined P-a.s. in L1 (0, 1, H) by Proposition 4.9]. Therefore, for y ∈ M, where
M is the intersection of the two full measure sets (denoted in both cases by
M) from Theorem 4.3 and Proposition 4.9, respectively, we obtain
 ERGODICITY FOR STOCHASTIC PDE’S 25

E(exp ρk (Z x )|Z1x = y)
= E(exp ρk (Z x )|Z1 = y − S1 x)
k−1
X
= E exp hG(Ztxi ), ζti+1 − ζti i
i=0
Z ti+1
− hG(Ztxi ), B1 (s)Zbs − B3 (s)Z1 i ds
ti
Z ! !
1 
(5.23) − 1
2 |G(Zsx )|2 ds Z1 = y − S1 x

0

k−1
X
= E exp hG(Zbtx,y
i
), ζti+1 − ζti i
i=0
Z ti+1
− hG(Zbtx,y
i
), B1 (s)Zbs − B3 (s)(y − S1 x)i ds
ti
Z !
1
− 21 |G(Zbsx,y )|2 ds
0

=: EΦk (x, y),

since the processes (Zbt ) and (ξt ) are independent of Z1 . By (5.18), the dom-
inated convergence theorem yields
Z 1 Z 1
Φk (x, y) → exp hG(Zbsx,y ), dζs i − 1
2 |G(Zbsx,y )|2 ds
0 0
Z 1 
(5.24) − hG(Zb x,y ), B1 (s)Zbs + B2 (s)x − B3 (s)yi ds
s
0
=: Φ(x, y), P-a.s.
(possibly, for a subsequence), since B2 (·)x, B3 (·)y ∈ L1 (0, 1, H) for x ∈ H
and y ∈ M. Proposition 4.9 and Gaussianity of the process Zb imply
 Z 1 
(5.25) E exp M b
|B1 (s)Zs | ds < ∞
0
for each M < ∞, and hence, the sequence (Φk (x, y)) is equiintegrable on Ω.
It follows that
(5.26) lim Φk (x, y) = Φ(x, y) in L1 (Ω)
k→∞

and in virtue of (5.23) and (5.20), we find that, for any x ∈ E,

dP (1, x, ·)
(5.27) (y) = EΦ(x, y) = h(x, y), µ1 -a.e.
dµx1
26 B. GOLDYS AND B. MASLOWSKI

Now we drop the assumption of boundedness of G. Proceeding as above, we

obtain, for each N > 0 and x ∈ E,
E(1{kZ x kC(0,1;E) ≤N } exp ρ(Z x )|Z1x = y)
(5.28)
= E1{kZbx,y k Φ(x, y), µ1 -a.e.,
C(0,1;E) ≤N }

because on the set {kZ x kC(0,1;E) ≤ N } we have G(Z x ) = GN (Z x ) [cf. defini-

tion in (5.4)] and GN is bounded. By (5.3) and the dominated convergence
theorem, we get
(5.29) lim 1{kZ x kC(0,1;E) ≤N } exp ρ(Z x ) = exp ρ(Z x ) in L1 (Ω),
N →∞
and (possibly, for a subsequence) it follows that
lim E(1{kZ x kC(0,1;E) ≤N } exp(ρ(Z x ))|Z1x = y)
N →∞
(5.30)
dP (1, x, ·)
= E(exp(ρ(Z x ))|Z1x = y) = (y), µ1 -a.s.
dµx1
On the other hand, by the monotone convergence theorem,
(5.31) lim E1{kZbx,y k Φ(x, y) = EΦ(x, y) = h(x, y)
N →∞ C(0,1;E) ≤N }

for x ∈ E, y ∈ M, so (5.28), (5.30) and (5.31) yield (5.16) and the proof is
completed. 

By means of the formula (5.15), we may find a useful lower estimate on

transition densities.

Theorem 5.3. For x ∈ E,

dP (1, x, ·)
(5.32) (y) ≥ c1 exp(−c2 |x|pE − Λ(y)), µ1 -a.e.,
dµ1
where Λ : M1 → R+ is a measurable mapping, M1 ∈ B(E), µ1 (M1 ) = 1, p =
max(2, 2m) and the constants c1 , c2 > 0 depend only on A, Q and K, m from
Hypothesis 2.4(b).

Proof. From (5.15) in virtue of the Jensen inequality, we obtain, for

x ∈ E,
dP (1, x, ·)
(y)
dµ1
  Z 1
(5.33) ≥ exp E ρ(Zb x,y ) − hG(Zbsx,y ), B1 (s)Zbs + B2 (s)x − B3 (s)yi ds
0

1 −1/2
+ hx, S1∗ Q−1
1 yi − |Q1 S1 x|2 ,
2
 ERGODICITY FOR STOCHASTIC PDE’S 27

for y from a set of µ1 -full measure in E. Note that in the proof of Theorem
5.2, we found a set M, µ1 (M) = 1, such that B3 (·)y ∈ L1 (0, 1 : H) and h(x, y)
−1/2
is well defined for y ∈ M1 . Similarly, S1∗ Q1 is a Hilbert–Schmidt operator
−1/2 −1/2
by (2.4), hence, S1∗ Q1 Q1 y ∈ H is well defined for y ∈ M2 , µ1 (M2 ) = 1
and the density g(x, y) is given by the formula (5.12) for y ∈ M2 . We may
take M1 = M ∩ M2 . It follows from Hypothesis 2.4(b) and (4.7) that the
stochastic integral in ρ(Zb x,y ) is a martingale and, hence, for any x ∈ E,

dP (1, x, ·)
(y)
dµ1
 Z 1
1
≥ exp − E|G(Zbsx,y )|2 ds
2 0
Z 1
−E |G(Zbsx,y )|(|B1 (s)Zbsx,y | + |B2 (s)x| + |B3 (s)y|) ds
0

−1/2 −1/2 1 −1/2
− |x||S1∗ Q1 Q1 y| − |Q S1 x|2
2 1
  Z 1 
2
≥ exp −K 1+ E|Zb x,y |2m ds
s E
0
Z 1
−E K(1 + |Zbsx,y |m b
E )(|B1 (s)Zs | + |B2 (s)x| + |B3 (s)y|) ds
0

1 2 −1/2
− c̃|x|E |S1∗ Q−1
1 y| − c̃ kQ1 S1 k2 · |x|2E ,
2
x ∈ E,

for µ1 -almost all y ∈ M1 , where c̃ is the constant from continuous em-

bedding E ֒→ H. Set U1 , U2 : M1 → R+ , U1 (y) := kB3 (·)ykL1 (0,1:H) , U2 (y) =
|S1∗ Q−1
1 y|; by (4.7) of Theorem 4.3, we further get

dP (1, x, ·)
(y) ≥ exp −K 2 [1 + L(2m)(1 + |x|2m
E + (U (y))
2m
)]
dµ1
Z 1
− KE |B1 (s)Zbs | ds
0
Z 1
−K |B2 (s)x| ds − KU1 (y)
0
Z 1
− KL(m)(1 + |x|m m
E + U (y) ) |B2 (s)x| ds
0
(5.34) − KL(m)(1 + kxkm + U m (y))U1 (y)
28 B. GOLDYS AND B. MASLOWSKI
Z 1
− KL(m)EkZb x,y km
C(0,1;E) |B1 (s)Zbs | ds
0

1 −1/2
− c̃|x|E U2 (y) − c̃2 kQ1 S1 k2 · |x|2E .
2
By Lemma 3.2, we have
Z 1
−1/2
|B2 (s)x| ds = kBQ1 S1 xkL1 (0,1:H)
0
−1/2
(5.35) ≤ kBQ1 S1 xkL2 (0,1 : H)
−1/2 −1/2
= |Q1 S1 x| ≤ c̃kQ1 S1 k · |x|E

and it follows from Proposition 4.9 that

Z 1 q
(5.36) E |B1 (s)Zbs | ds < ∞,
0

for each q < ∞. Therefore, for each η > 0 small enough, there exist constants
c1 (η) > 0 and c3 (η) > 0 and a function Λ = Λη : M1 → R+ such that, x ∈ E
and y ∈ M1 ,

dP (1, x, ·)
(y) ≥ exp −c1 (η) − (K 2 L(2m) + η)|x|2m
E
dµ1
−1/2
− KL(m)kQ1 S1 k · |x|m+1
E
  
1 2 −1/2
− c3 (η)|x|m+η
E − c̃ kQ1 S1 k2 + η |x|2E − Λ(y) ,
2
and the estimate (5.32) follows. 

Remark 5.4. Under more stringent conditions, we may obtain a lower

estimate on the transition density which is more “explicit” in y and has
a more symmetric form. In addition to the conditions of Theorem 5.2, as-
sume that there exists a Banach space Ẽ of µ1 -full measure, continuously
embedded into H such that

(5.37) S1 (Ẽ) ⊂ im(Q1 ),

Z 1
(5.38) |B3 (s)y| ds ≤ a1 |y|Ẽ , y ∈ Ẽ,
0

and
Z t

 
(5.39) U (y) = sup  ∗
St−s QS1−s Q−1 
1 y ds ≤ a2 |y|Ẽ , y ∈ Ẽ,
t∈[0,1] 0 E
 ERGODICITY FOR STOCHASTIC PDE’S 29

for some a1 , a2 > 0. Then for some constants b1 , b2 , b3 > 0 (dependent only
on A, Q, K and m) and p = max(2, 2m), we have
dP (1, x, ·)
(5.40) (y) ≥ b1 exp{−b2 |x|pE − b3 |y|pẼ }, x ∈ E, y ∈ Ẽ a.e.
dµ1
To see (5.40), we check that, under present conditions (5.37)–(5.39), (5.34) im-
plies, for all for x ∈ E and for µ1 -a.e. y ∈ Ẽ,
dP (1, x, ·)
(y) ≥ exp(−C(1 + |x|2m 2m
E + |y|Ẽ − |x|E + |y|Ẽ
dµ1
+ |x|m+1
E + |y|m
Ẽ
|x|E + |x|m
E |y|Ẽ

+ |x|m+η
E + |y|m+η
Ẽ
+ |x|2E + |y|2Ẽ )),

for arbitrary small η > 0 and a universal C = C(η) < ∞, and (5.40) follows.
It may be of interest to mention some particular cases when the conditions
(5.37)–(5.39) are satisfied. A trivial example is a finite-dimensional one,
H = E = Ẽ = Rd , in which case we obtain
dP (1, x, ·)
(5.41) (y) ≥ b1 exp(−b2 |x|pRd − b3 |y|pRd ), x, y ∈ Rd .
dµ1
Note that in this case the only assumptions in Theorem 5.2, Corollary 5.3
and the present remark are the well posedness and growth conditions in
Hypothesis 2.4 and the strong Feller property of the Ornstein–Uhlenbeck
process (2.7).
Suppose that A = A∗ is strictly negative and define Hλ = dom((−A)λ ), λ ≥
0, with the norm |y|λ = |(−A)λ y|, y ∈ dom((−A)λ ). Let Q = I; then A−1
must be compact and it is easy to check that im(S1 ) ⊂ im(Q1 ) = dom(A),
therefore, (5.37) holds with any Ẽ, Ẽ ֒→ H. Furthermore, we have

|Q1/2 S1−s Q−1

1 y| = kS1−s (−A)
1−λ
kk(−A)λ−1 Q−1
1 (−A)
−λ
k · |y|λ
(5.42)
const
≤ kykλ
(1 − s)1−λ

for y ∈ Hλ since A−1 Q−1 1 ∈ L(H), thus, (5.38) holds for Ẽ = Hλ with any
λ > 0. If, in addition, kSt kL(H,E) ≤ const · t−σ , t ∈ [0, 1], for σ > 0 such that
σ < λ, then (5.39) holds as well since

(5.43) |St−s S1−s Q−1

1 y|E ≤ const(t − s)
−σ
(1 − s)λ−1 , 0 < s < t ≤ 1.

6. Exponential convergence to invariant measure. The following uniform

ultimate moment boundedness result will be useful in the sequel.
30 B. GOLDYS AND B. MASLOWSKI

Proposition 6.1. Assume that the growth condition (2.11) holds true
and
(6.1) k(p) := sup E|Zt |pE < ∞, p > 0.
t≥0

Then
k2 k(s) + k3
(6.2) Ex |Xt |E ≤ e−k1 t |x|E + + k(1), t ≥ 0.
k1
Suppose that the following stronger version of (2.11) holds: For each x ∈
dom(Ã), there exists x∗ ∈ ∂|x|E such that, for some k1 , k2 , k3 > 0, s > 0, ε >
0, we have
(6.3) hÃx + F (x + y), x∗ iE,E ∗ ≤ −k1 |x|1+ε s
E + k2 |y|E + k3 , y ∈ E.
Then
(6.4) c,
sup sup Ex |Xt |E ≤ M
x∈E t≥1

where
 1+ε  1/ε 
c = k(1) + max 2(k2 k(s) + k3 ) 1
(6.5) M , +2 .
k1 k1 ε

Proof. Inequality (6.4) has been proven in [14], Proposition 2.1 (see
also a similar result in [20]). The proof of (6.2) follows the lines of similar
proofs based on Yosida approximation techniques (see, e.g., [9]) and we
sketch it only. The process Y x (t) := Xtx − Zt satisfies the equation
Z t
x
(6.6) Y (t) = St x + St−s F (Y x (s) + Zs ) ds, t ≥ 0,
0
and the sequence of approximating processes Yλ (t) is defined by
Z t
(6.7) Yλx (t) = R(λ)St x + R(λ)St−s F (Y x (s) + Zs ) ds, t ≥ 0,
0

where R(λ) := λ(λI − Ã)−1 ∈ L(E) is well defined for λ large enough. It is
well known that
dYλ
(6.8) Yλx → Y x , − ÃYλ − F (Yλ + Z) = σλx → 0, λ → ∞,
dt
in C(0, T ; E) (cf. page 201 of [9]). Since by (2.11)
d− x
(6.9) |Y (t)|E ≤ −k1 |Yλ (t)|E + k2 |Ztx |sE + k3 + |σλ (t)|E ,
dt λ
we obtain
Z t
|Y x (t)|E ≤ e−k1 t |x|E + e−k1 (t−τ ) (k2 |Zτx |sE + k3 ) dτ,
0
 ERGODICITY FOR STOCHASTIC PDE’S 31

and thereby,
Z t
(6.10) E|Y x (t)|E ≤ e−k1 t |x|E + ek1 (t−τ ) (k2 k(s) + k3 ) dτ
0
and (6.2) follows. 

Our next aim is to establish uniform geometric ergodicity and V -uniform

ergodicity results for V (x) = |x|E + 1 using the lower density estimates and
uniform moment boundedness shown above. We will also find explicit bounds
on the convergence rates, hence, the constants below will play some role. We
assume that
(6.11) Ex |Xt |E ≤ k0 e−k1 t |x|E + cb, t ≥ 0,
for some k0 , k1 > 0 and cb ∈ R. Note that, by Proposition 6.1, if (6.1) and the
growth condition (2.11) are both satisfied, then (6.11) holds with k1 given
in (2.11), k0 = 1 and
k2 k(s) + k3
(6.12) cb = + k(1).
k1
Now, take R > 4cb, r > 4(cb + 12 ), and define
 
1 R cb
t0 = − log − ,
k1 2rk0 rk0
(6.13)  
1 1 1
T = max t0 + 1, − log , b = cb +
k1 4k0 2
and
Z
p
(6.14) δ = 21 c1 e−c2 R e−Λ(y) µ1 (dy),
Br

where Br := {y ∈ E, |y|E < r}, and c1 , c2 , p and Λ are defined in the same
way as in Theorem 5.3. In the following proposition, existence of a universal
small set satisfying a uniform geometric drift condition is shown.

Proposition 6.2. Assume (6.11). Then the following holds:

(a) We have
(6.15) inf P (T, x, Γ) ≥ δµ̄(Γ), Γ ∈ B(E),
x∈Br

where
Z −1 Z
(6.16) µ̄(Γ) := e−Λ(y) µ1 (dy) e−Λ(y) µ1 (dy), Γ ∈ B(E),
Br Br ∩Γ
is a probability measure. In particular, Br is a small set of the Markov chain
(X̃n ) := (XnT ), with the lower bound measure δµ̄.
32 B. GOLDYS AND B. MASLOWSKI

(b) We have
(6.17) Ex (|XT |E + 1) ≤ 21 (|x|E + 1) + b1Br (x), x ∈ E,

that is, the chain (X̃n ) satisfies the one-step Lyapunov–Foster condition of
geometric drift toward Br , with the constants 12 and b and the Lyapunov
function V (x) = |x|E + 1.

Proof. (a) By (6.11), we have, for t ≥ t0 , x ∈ E, |x|E ≤ r,

E|Xt |E 1 1
(6.18) P (t, x, BR ) ≥ 1 − ≥ 1 − (k0 re−k1 t + cb ) ≥
R R 2
and therefore, by Theorem 5.3, for each t ≥ t0 , we get
Z Z
P (t + 1, x, Γ) = P (1, y, Γ)P (t, x, dy) ≥ P (1, y, Γ)P (t, x, dy)
E BR
Z Z
dP (1, y, ·)
≥ (z)µ1 (dz)P (t, x, dy)
BR Γ dµ1
(6.19) Z Z
≥ c1 exp{−c2 |y|pE − Λ(z)}µ1 (dz)P (t, x, dy)
BR Γ
Z
p
≥ c1 e−c2 R e−Λ(z) µ1 (dz)P (t, x, BR ), x ∈ E, Γ ∈ B(E).
Γ
Hence,
Z
−c2 Rp
inf P (t + 1, x, Γ) ≥ c1 e e−Λ(z) µ1 (dz) inf P (t, x, BR )
x∈Br Γ x∈Br
Z
p
(6.20) ≥ 12 c1 e−c2 R e−Λ(z) µ1 (dz)
Γ∩Br

= δµ̄(Γ), Γ ∈ B(E)
and (6.15) follows.
To prove part (b), we use again (6.11) to obtain
Ex (|Xt |E + 1) ≤ k0 |x|E e−k1 t + cb + 1 ≤ 14 |x|E + cb + 1
(6.21) ≤ 12 (|x|E + 1) − 14 |x|E − 12 + cb + 1
≤ 12 (|x|E + 1) + (cb + 12 )1Br (x)
for x ∈ E, t ≥ − k11 log 4k10 , which completes the proof. 

In the next theorem our main result on uniform geometric V -ergodicity for
V (x) = |x|E + 1 is stated. It is based on the paper by Meyn and Tweedie [33],
where exact bounds for geometric ergodicity of irreducible Markov chains are
 ERGODICITY FOR STOCHASTIC PDE’S 33

found, and Proposition 6.2 above. Following [33], we introduce the constants
b b
v, Mc , γc , λ, b and ξ̄ as follows:
1/2 + γc
v = r + 1, γc = δ−2 (4b + δv), b=
λ < 1,
1 + γc
(6.22)
b 4 − δ2 2
b = v + γc , ξ̄ = 4b
δ5
and
1
Mc = (1 − λ b + bb 2 + ξ̄( bb(1 − λ
b +b b 2 )) > 1.
b)+b
b 2
(1 − λ)
We will show that the Markov chain (X̃n ) has the geometric rate of conver-
gence to the invariant measure with any constant
 
1
(6.23) ρ∈ 1− ,1 .
Mc
Let bV B denote the Banach space of measurable functions ϕ : E → R such
that
|ϕ(x)|
kφkV = sup < ∞.
x∈E V (x)

Theorem 6.3. Assume (6.11). Then there exists an invariant measure

µ∗ ∈ P and for V (x) = |x|E + 1, we have
 Z 
 
(6.24) sup Pt ϕ(x) − ϕ dµ∗  ≤ M V (x)e−ωt , t ≥ 0, x ∈ E,
kφkV ≤1 E

where
1
ω=− log ρ > 0 and
T
(6.25)
ρ
M = (1 + γc ) −1 (cb + k0 + 1)e− log ρ
ρ + Mc − 1
and
(6.26) kPt∗ ν − µ∗ kvar ≤ M (Lν + 1)e−ωt , t ≥ 0, ν ∈ P,
R
where Lν = E |x|E ν(dx). The constants ω and M may be chosen the same
for all nonlinear terms F satisfying Hypothesis 2.4(b) with the same con-
stants K and m and (6.11) with the same constants k0 , k1 and cb [or, in
particular, satisfying the growth condition (2.11) with the same k1 , k2 , k3
and s].
34 B. GOLDYS AND B. MASLOWSKI

Proof. The existence of an invariant probability measure in presence

of the lower bound measure [cf. (6.15)] and condition (6.11) are well known
(see, e.g., [28]). It follows from Proposition 6.2 that we may apply Theorem
2.3 of [33] to the Markov chain (X̃n ) = (XnT ) (note that the measure µ̄ is
concentrated on Br , so the conditions imposed in [33] are satisfied), which
yields
 Z 
 
sup PnT ϕ(x) − ϕ dµ  ≤ Cρn V (x) = Ce−nT ω V (x),
 ∗
kφkV ≤1
(6.27)
x ∈ E, n ∈ N,
where C = (1 + γc ) ρ+Mρ−1 −1 . Using the semigroup property of (Pt ) and
c
(6.11), we obtain
 Z    Z  
 
∗
 ∗

 
sup PnT +s ϕ(x) − ϕ dµ  ≤ sup Ps PnT ϕ − ϕ dµ (x)
kφkV ≤1 kφkV ≤1

≤ |Ps (Ce−nT ω V (x))| ≤ Ce−nT ω Ex (|Xs |E + 1)

(6.28) ≤ Ce−nT ω (k0 |x|E + cb + 1)
≤ C(cb + k0 + 1)e− log ρ (|x|E + 1)e−ω(nT +s) ,
n ∈ N, x ∈ E, s ∈ [0, T ],
which yields (6.24). Inequality (6.26) is an obvious consequence of (6.24)
since
Z
kPt∗ ν − µ∗ kvar ≤ kP (t, x, ·) − µ∗ kvar ν(dx)
E
Z  Z 
 
∗
(6.29) ≤ 
sup Pt ϕ(x) − ϕ dµ ν(dx)
kφkV ≤1
Z
≤ M (|x|E + 1)e−ωt ν(dx) = M (Lν + 1)e−ωt

for each ν ∈ P, t ≥ 0. The universality of M and ω follows from the fact

that all constants defined in (6.12)–(6.14) and (6.22) (including c1 , c2 , p and
the mapping Λ, cf. Theorem 5.3) are independent of F . 

If the growth of the nonlinear term F is faster than linear, the Markov
process defined by the equation (2.1) may be uniformly ergodic, that is, the
constant Lν in (6.26) may be replaced by another constant independent of
the initial measure ν ∈ P. This has been established earlier in [14] and [20];
however, the lower bound measures are not found there constructively. In
the theorem below explicit bounds are found and, in particular, uniformity
of convergence with respect to coefficients is proven.
 ERGODICITY FOR STOCHASTIC PDE’S 35

Theorem 6.4. Assume (6.1) and let the stronger growth condition (6.3)
hold true. Then there exists an invariant measure µ∗ ∈ P and for any ν ∈ P,
(6.30) kPt∗ ν − µ∗ kvar ≤ (1 − δ)−1 e−b
ωt
kν − µ∗ kvar , t ≥ 0,
where ωb = − 12 log(1 − δ) > 0, δ is defined by (6.14) with R = 2M c and r =
∞ and M c is given by (6.5). In particular, the constants on the r.h.s. of
(6.30) are uniform with respect to all nonlinear terms F satisfying the growth
conditions (2.9) and (6.3) with the same constants K, m, k1 , k2 , k3 , s and ε.

Proof. By Proposition 6.1, we have that

Ex |X1 |E 1
(6.31) inf P (1, x, BR ) ≥ 1 − sup ≥
x∈E x∈E R 2
and similarly, as in (6.19), we get
Z
inf P (2, x, Γ) ≥ inf P (1, y, Γ)P (1, x, dy)
x∈E x∈E E
(6.32) Z
p
≥ 21 c1 e−c2 R e−Λ(z) µ1 (dz) = δµ̄(Γ),
Γ
where µ̄ is defined by (6.16) with r = ∞. For each ν ∈ P, it follows that
P2∗ ν ≥ δµ̄ and a simple computation (cf., e.g., [14], Theorem 2.4) yields
kP2∗ µkvar ≤ (1 − δ)kµkvar , µ = ν1 − ν2 , ν1 , ν2 ∈ P.
By the semigroup property of (Pt∗ ), ∗ µk
we have kP2n n
var ≤ (1 − σ) kµkvar and
for s ∈ [0, 2], it follows that
∗
kP2n+s µkvar ≤ kPs∗ P2n
∗ ∗
µkvar ≤ kP2n µkvar

≤ e−2nb
ω
kµkvar ≤ (1 − δ)−1 e−(2n+s)b
ω
kµkvar . 

7. Uniform spectral gap property. In this section we consider exponen-

tial ergodicity in spaces Lp (E, µ∗ ) for p ∈ [1, ∞). Note first that, by The-
orem 5.2, the transition kernels P (T, x, ·) are equivalent for T > 0, x ∈ E,
they are also equivalent to the invariant measure µ∗ (if it exists) and we
have, for each t > 0,
Z
Pt φ(x) = pt (x, y)φ(y)µ∗ (dy), φ ∈ Cb (E),
E
where the function (x, y) → pt (x, y) is measurable. Let us recall that (Pt )
extends to a contraction semigroup on Lp (E, µ∗ ) for all p ∈ [1, ∞] and is
a C0 -semigroup if p < ∞. Let M(E) ⊂ (Cb (E))∗ denote the space of finite
Borel measures on E with the variation norm. For ν ∈ M(E), we have
Z Z
(7.1) hPt∗ ν, φi = hν, Pt φi = pt (x, y)φ(y)µ∗ (dy)ν(dx).
E E
36 B. GOLDYS AND B. MASLOWSKI

Lemma 7.1. Assume that the equation (2.1) has an invariant measure
µ∗ ∈ P. Then the space Lp (E, µ∗ ) is invariant for (Pt∗ ) for each p ∈ [1, ∞).
Moreover, kPt∗ kp→p = 1 and
Z
(7.2) Pt∗ ψ(y) = pt (x, y)ψ(x)µ∗ (dx), ψ ∈ L1 (E, µ∗ ).
E

Finally, (Pt∗ ) is a C0 -semigroup on Lp (E, µ∗ ) for p ∈ [1, ∞).

Proof. For ψ ∈ L1 (E, µ∗ ), ψ ≥ 0, we define

Z
Gt ψ(y) = pt (x, y)ψ(x)µ∗ (dx),
E

and ν = ψµ∗ . For φ ≥ 0, the Fubini theorem yields

Z Z
hPt∗ ν, φi = pt (x, y)ψ(x)φ(y)µ∗ (dx)µ∗ (dy) = hGt ψ, φi < ∞.
E E

Putting φ = 1, we obtain
kGt ψk1 = hPt∗ (ψµ∗ ), 1i = hψµ∗ , 1i = kψk1 ,
and therefore,
kGt ψk1 ≤ kψk1 , ψ ∈ L1 (E, µ∗ ), ψ ≥ 0.
Clearly, Gt ψ = Pt∗ (ψµ∗ ). All those arguments extend immediately to an ar-
bitrary ψ ∈ L1 (E, µ∗ ) and therefore, Gt is a contraction on L1 (E, µ∗ ). Other
parts follow easily by a standard density argument. 

Let Lp be the generator of (Pt ) acting in Lp (E, µ∗ ). We say that Lp has

the spectral gap in Lp (E, µ∗ ) if there exists δ > 0 such that
σ(Lp ) ∩ {λ : Re λ > −δ} = {0}.
The largest δ with this property will be denoted by gap(Lp ).

Theorem 7.2. Assume (6.1) and let the stronger growth condition (6.3)
be satisfied. Then, for each p ∈ (1, ∞), we have
ω
b
(7.3) gap(Lp ) ≥
p
and
(7.4) kPt φ − hµ∗ , φikp ≤ Cp e−(b
ω /p)t
kφkp ,
b > 0 is defined in Theorem 6.4. If, moreover, the semigroup (Pt ) is
where ω
symmetric in L2 (E, µ∗ ), then (7.3) and (7.4) hold for p = 1.
 ERGODICITY FOR STOCHASTIC PDE’S 37

Proof. By Theorem 6.4, there exist C > 0 such that

(7.5) kPt∗ ν − µ∗ kvar ≤ Ce−b

ωt
,
for any probability measure ν on E or, equivalently,

(7.6) kPt∗ ν − ν(E)µ∗ kvar ≤ Ckνkvar e−b

ωt
,
for any signed measure ν. If ν = ψµ∗ , then (7.6) and Lemma 7.1 imply

kPt∗ ψ − hψ, 1ik1 ≤ Ckψk1 e−b

ωt
,
hence,

(7.7) kPt∗ − Πk1 ≤ Ce−b

ωt
,
where Πψ = hψ, 1i1 and
(7.8) kPt∗ − Πkq→q ≤ 2, q ∈ [1, ∞].
Take q > 2. Then by (7.7), (7.8) and the Riesz–Thorin theorem, we find that

(7.9) kPt∗ − Πkp→p ≤ C θ e−b

ω θt 1−θ
2 ,
where
1 θ 1−θ
= + .
p 1 q
Therefore, taking q → ∞ in (7.9), we obtain

(7.10) kPt∗ − Πkp→p ≤ C2 e−(b

ω /pt)
.
Therefore, (7.4) holds, and since (Pt ) is a C0 -semigroup in L2 (E, µ∗ ), Theo-
rem 3.6.2 in [34] (7.10) implies the spectral gap property with gap(Lp ) ≥ b ω
p
for p ∈ (1, ∞). If (Pt ) is symmetric, then the conclusion of the theorem for
p = 1 follows immediately from (7.5). 

Remark 7.3. (1) In Theorem 7.1 and 7.2 the invariant measure µ∗ ,
hence, the space Lp (E, µ∗ ), depends on the coefficients of equation (2.1). It is
interesting to note that the lower bound on gap(Lp ) and Cp are universal for
all systems satisfying Hypothesis 2.4(b) and (6.3) with the same constants.
(2) By Theorem 7.2, the spectral gap exists for all p ∈ (1, ∞). The fact
that this property holds in L1 (E, µ∗ ) is perhaps surprising. Note that it does
not need to hold in general if F = 0. It is known (cf. [13]) that, for a one-
dimensional Ornstein–Uhlenbeck operator LOU 1 considered in L1 (E, µ∗ ), we
have
σ(LOU
1 ) = {λ : Re λ ≤ 0}.
38 B. GOLDYS AND B. MASLOWSKI

If p = 2 and (Pt ) is symmetric, the stronger growth condition (6.3) is not

needed. We may get an estimate on spectral gap in L2 (E, µ∗ ) under the
standard ultimate boundedness assumption, which is stated in Corollary 7.4
below. Note that the assumption of symmetricity of (Pt ) may not be removed
(cf. Example 9.1 below).

Corollary 7.4. Let the conditions of Theorem 6.3 be satisfied and

assume that (Pt ) is symmetric on L2 = L2 (E, µ∗ ). Then
(7.11) kPt ϕkL2 ≤ e−ωt kϕkL2
R
holds for all t ≥ 0 and ϕ ∈ L2 , ϕ dµ∗ = 0, where ω is defined in (6.25).

Proof. Taking arbitrary t > 0, we get, by Theorem 6.3,

 Z 
 
(7.12) sup Ptn ϕ(x) − ϕ dµ∗  ≤ M V (x)ρb n , x ∈ E,
kφkV ≤1

where ρb = e−ωt , that is, the skeleton (Yn ) := (Xtn ) is V -uniformly ergodic
with the rate ρb. By [36], Theorem 2.1, it follows that
Z
n 2
(7.13) kPtn ϕkL2 ≤ ρb kϕkL2 , n ∈ N, ϕ ∈ L , ϕ dµ∗ = 0

and taking n = 1, we obtain (7.11). 

8. Some extensions.

8.1. Equations nonhomogeneous in time. Some results in the present pa-

per may be easily generalized to the case when the nonlinear term F =
F (t, x) in the equation (2.1) also depends on time, that is, the equation has
the form
p
dXt = (AXt + F (t, Xt )) dt + Q dWt , t ≥ s ≥ 0,
Xs = x,
and defines a nonhomogeneous Markov process. For instance, let Ps,t and Ps,t ∗

denote the corresponding two-parameter Markov semigroup and adjoint

Markov semigroup, respectively, 0 ≤ s ≤ t, and set P (s, x, t, Γ) := Es,x1Γ (Xt ) =
Ps,t 1Γ (x), 0 ≤ s ≤ t, x ∈ E, Γ ∈ B.

Theorem 8.1. Let Hypotheses 2.1, 2.2, 2.3 and condition (6.1) be satis-
fied and let F : R+ × E → E be a jointly measurable mapping such that F (t, ·)
is Lipschitz continuous on bounded sets in E and satisfies Hypothesis 2.4(b)
and the growth condition (6.3) with constants independent of t ∈ R+ . Then
∗
kPs,t ∗
ν1 − Ps,t ν2 kvar ≤ (1 − δ)−1 e−b
ω (t−s)
kν1 − ν2 kvar , 0 ≤ s ≤ t, ν1 , ν2 ∈ P,
where ωb = − 12 log(1 − δ) and δ, ω
b depend only on the constants in Hypothe-
sis 2.4(b) and (6.3).
 ERGODICITY FOR STOCHASTIC PDE’S 39

The proof is just a a slight modification of the above results; similarly to

Theorems 5.2 and 5.3 and Proposition 6.1, we obtain
Z
inf P (s, x, s + 2, Γ) ≥ inf P (s + 1, y, s + 2, Γ)P (s, x, s + 1, dy)
s∈R+ ,x∈E s∈R+ ,x∈E E
Z
p
≥ 12 c1 e−c2 R e−Λ(z) µ1 (dz) = δµ̄(Γ),
Γ

where R = 2M c and M c is defined in Proposition 6.1. Our statement now

follows just as in the proof of Theorem 6.4.

8.2. Continuous dependence of invariant measures on parameter. Uni-

formity of convergences proven in Theorems 6.3 and 6.4 with respect to
nonlinear drifts in a fairly large class may be useful in some cases, for in-
stance, in ergodic control theory. Another application (given below) is a
continuous dependence of invariant measures on a parameter. Consider the
parameter-dependent equation
p
dXtα = (AXtα + Fα (Xtα )) dt + Q dWt ,
(8.1)
X0α = x ∈ E,
where α ∈ A ⊂ Rd . Denote by P α (t, x, ·) and µα the transition probability
kernel and the invariant measure, respectively, associated with the equa-
tion (8.1).

Theorem 8.2. Let Hypotheses 2.1–2.4, condition (6.1) and the growth
condition (2.11) hold for equation (8.1) with the constants independent of
α ∈ A, and assume
(8.2) lim Gα (x) = Gα0 (x), x ∈ E,
α→α0

where Fα = Q1/2 Gα . Then

(8.3) lim kµα − µα0 kvar = 0.
α→α0

Proof. First we prove

(8.4) lim kP α (t, x, ·) − P α0 (t, x, ·)kvar = 0,
α→α0

for each t > 0 and x ∈ E. By (5.2), it suffices to show that exp ρα → exp ρα0
in L1 (Ω), where
Z t Z t
ρα (Z x ) = hGα (Zsx ), dWs i − 21 |Gα (Zsx )|2 ds.
0 s
40 B. GOLDYS AND B. MASLOWSKI

By the dominated convergence theorem, we have

Z t
lim E |Gα (Zsx ) − Gα0 (Zsx )|2 ds = 0,
α→α0 0
hence, exp ρα (Z x ) → exp ρα0 (Z x ) P-a.s. In order to prove uniform integra-
bility of (exp ρα (Z x )), α ∈ A, it is enough to show
(8.5) sup Eρα (Z x ) exp ρα (Z x ) < ∞.
α∈A
Rs
cs := Ws −
Setting W Gα (Zτx ) dτ , s ∈ [0, t], we obtain, in virtue of the Gir-
0
sanov theorem
Z t Z t 
E hGα (Zsx ), dWs i − 12 |Gα (Zsx )|2 ds exp ρα (Z x )
0 0
Z t Z t 
(8.6) =E hGα (Zsx ), dW
cs i + 1
2 |Gα (Zsx )|2 ds exp ρα (Z x )
0 0
Z t Z t
= E 12 |Gα (Xsα )|2 ds ≤ K 2 +K E 2
|Xsα |2m
E ds ≤ N,
0 0
where N < ∞ is a constant independent of α ∈ A, which may be easily seen
similarly as in (6.2) (cf. also Proposition 2.1 of [14]). Thus, (exp ρα (Z x )) are
uniformly integrable, which concludes the proof of (8.4). Now we have
kµα − µα0 kvar ≤ kP α (t, x, ·) − µα kvar
(8.7) + kP α0 (t, x, ·) − P α0 (t, x, ·)kvar
+ kP α0 (t, x, ·) − µα0 kvar
and by Theorem 6.3,
lim sup kP α (t, x, ·) − µα kvar = 0,
t→∞ α∈A

which together with (8.4) yields (8.3). 

9. Examples.

Example 9.1 (Finite-dimensional equation). In the finite-dimensional

case E = H = Rd , the condition (2.7) is satisfied, even if the covariance
matrix Q is degenerate, which may be shown by generalizing a well-known
Seidman’s result [38] (cf. [25], Theorem 5.25). Obviously, Z ∈ C([0, T ], Rd )
for each T > 0, so the only assumptions that are needed in Theorem 6.3 (V -
uniform ergodicity) and, if Pt = Pt∗ , in Corollary 7.4 (spectral gap), are the
strong Feller property for the linear equation (2.4) (which is true if and only
if the matrix Q1 is positive and is implied by positivity of the matrix Q),
Hypothesis 2.4 and the the ultimate boundedness of solutions to (6.11). In
 ERGODICITY FOR STOCHASTIC PDE’S 41

order to apply Theorem 6.4 (uniform ergodicity) and Theorem 7.2 [spectral
gap in Lp (E, µ∗ )], we have to assume the stronger growth condition (6.3).
As a specific example, we consider a nonlinear stochastic oscillator equa-
tion
(9.1) ÿ = f (y, ẏ) + σ ẇt , y(0) = x1 , ẏ(0) = x2 ,
in Rd . We assume that f : Rd × Rd → Rd is a locally Lipschitz function,
x1 , x2 ∈ Rd , σ ∈ L(Rd ) is a regular matrix, and (wt ) is a standard Wiener
process in Rd . Equation (9.1) may be rewritten in the form (2.1) with Xt =
(y(t), ẏ(t)) ∈ R2d = E = H,
   
0 I 0
A= , F (x) = , x ∈ R2d ,
0 0 f (x)
 
1/2 0 0
Q = .
0 (σσ ∗ )1/2
According to the Kalman rank condition (see, e.g., [27]), the matrix Qt
is invertible for each t > 0, so the equation with F = 0 is strongly Feller.
Suppose that f has at most polynomial growth, that is,
|f (x)|Rd ≤ K(1 + |x|m
R2d ), x ∈ R2d ,
for some K, m < ∞. Then by Theorem 6.3, the solution is V -uniformly er-
godic (Theorem 6.3), provided the ultimate boundedness condition (6.11)
holds true. For example, we may take d = 1, f (x) = −α2 x2 − α1 x1 for
x = (x1 , x2 ) ∈ R2d , with some α1 , α2 > 0 (a damped linear oscillator). Then
(6.11) holds with constants which may be easily expressed in terms of α1 , α2
and σ and V -uniform ergodicity holds true. Similar results for a more general
version of equation (9.1) can be found in [30].
Note that the semigroup (Pt ) is not symmetric in this case and Corol-
lary 7.4 (on the spectral gap) is not applicable. Indeed, it follows from the
results in [7, 8] that the spectral gap is zero in the present case. This exam-
ple also shows that the assumption of symmetry of Pt in Corollary 7.4 may
not be removed.

Example 9.2 (Stochastic reaction–diffusion equation with the cylindrical

noise). Consider the system
∂u
= Lu + f (u) + η,
∂t
(9.2) u(0, ξ) = x(ξ),
∂u ∂u
(t, 0) = (t, 1) = 0, (t, ξ) ∈ R+ × (0, 1),
∂ξ ∂ξ
42 B. GOLDYS AND B. MASLOWSKI

where L is a uniformly elliptic operator

 
∂ ∂ ∂
(9.3) [Lϕ](ξ) = a(ξ) ϕ (ξ) + b(ξ) ϕ(ξ) + c(ξ)ϕ(ξ), ξ ∈ (0, 1),
∂ξ ∂ξ ∂ξ
with a, b, c ∈ C 1 ([0, 1]), a(ξ) ≥ a0 > 0, ξ ∈ (0, 1), f : R → R is a locally Lips-
chitz mapping, and η = η(t, ξ) is a nondegenerate noise. Let us note that the
C 1 regularity of the coefficients is made for simplicity only and may be easily
relaxed. The system (9.2) is rewritten in the form (2.1), with the coefficients
defined in an obvious way on the spaces H = L2 (0, 1), E = C([0, 1]),
 
∂ϕ ∂ϕ
A = L, Dom(A) = ϕ ∈ H 2 (0, 1), (0) = (1) = 0 , Q ∈ L(H),
∂ξ ∂ξ
where Q is supposed to be boundedly invertible on H, and F : E → E, F (x(ξ)) =
f (x(ξ)), ξ ∈ (0, 1), x ∈ E. It is well known (see, e.g., [9], A5.2) that A gener-
ates a strongly continuous semigroup on H and Hypothesis 2.2 is satisfied.
With no loss of generality (replacing, if necessary, A and F by A − ωI
and F + ωI, respectively, with ω sufficiently large), we may assume that
hÃx, x∗ iE,E ∗ ≤ 0 for each x ∈ Dom(Ã), x ∈ ∂kxk (recall that à denotes the
part of A on E), then (6.1) is satisfied (see, e.g., [14], Example 3.1). It is
well known that the corresponding Ornstein–Uhlenbeck process is strongly
Feller and for a certain c > 0,
−1/2 c
kQt St k ≤ √ , t ∈ (0, 1),
t
(cf. [9]), hence, (3.12) holds. Moreover, standard estimates on the Green
function of the problem [4] yield kSt kHS ≤ const.t−1/4 , which implies (3.11)
with 0 < α < 12 . Therefore, Hypotheses 2.1 and 2.3 hold and it remains only
to specify the growth conditions on f . We assume that
(9.4) |f (ξ)| ≤ k(1 + |ξ|m ), ξ ∈ R,
and
(9.5) f (ξ + η) sign ξ ≤ −k1 |ξ| + k2 |η|s + k3 , ξ, η ∈ R,
for some constants k, m, k1 , k2 , k3 and s. Now it is easy to check that Hy-
pothesis 2.4 is satisfied and we may apply Theorem 6.3 to get V -uniform
ergodicity with the rate which is specified there. Also, in the case the Markov
semigroup Pt is symmetric in L2 (E, µ∗ ) (e.g., if Q = I), we may apply Corol-
lary 7.4 to obtain a lower bound for spectral gap. If the condition (9.5) is
strengthened to
(9.6) f (ξ + η) sign ξ ≤ −k1 |ξ|1+ε + k2 |η|s + k3 , ξ, η ∈ R,
where ε > 0, then (6.3) holds as well and we may apply Theorem 6.4 on
uniform exponential ergodicity and Theorem 7.2 on the spectral gap in
Lp (E, µ∗ ), p ∈ [1, ∞). For example, if f is a true polynomial, we have ob-
tained the following result:
 ERGODICITY FOR STOCHASTIC PDE’S 43

Corollary 9.3. In Example 9.2, assume that Ψ is a set of polynomials

of the form
2n
X
f (ξ) = −a2n+1 ξ 2n+1 + ai ξ i ,
i=0

where ai , i = 1, . . . , 2n, are in a given bounded set in R2n , a2n+1 ≥ ā, for
a given ā > 0 and n ≥ 0. Then the V -uniform ergodicity (and if n > 0,
uniform exponential ergodicity) holds with constants in (6.24), (6.26) and
(6.30) uniform with respect to f ∈ Ψ. Also, if n > 0, then there is a positive
lower bound on the spectral gap for (Pt ) in Lp (E, µ∗ ), p ∈ (1, ∞), uniform in
f ∈ Ψ. Finally, if b = 0 and Q = I, then this bound holds also for p = 1.

Example 9.4 (The case of Lipschitz drift). Consider (9.2) with the same
differential operator L and initial and boundary conditions in the case when
the noise may degenerate, for simplicity, suppose that c ≤ c0 < 0. For σ ≥
0, let Hσ denote the domain dom((−A)σ ) equipped with the graph norm
|y|σ := |(−A)σ y|. As well known, for σ ∈ (0, 21 ), the norm | · |σ is equivalent
with the norm of Sobolev–Slobodetskii space H 2σ (0, 1),
Z 1Z 1 |y(ξ) − y(η)|2
|y|2H 2σ := |y|2L2 (0,1) + dξ dη, y ∈ Hσ .
0 0 (ξ − η)1+2σ
Assume that f : R → R is Lipschitz continuous. It is easy to check that F :
Hσ → Hσ is continuous and satisfies the growth condition (2.11) if, for some
k̄1 , k̄2 > 0, we have
(f (ξ) − f (η)) sign(ξ − η) ≤ k̄1 |ξ − η| + k̄2 , ξ, η ∈ R.
Assume that Q = (−A)−2∆ for some ∆ ≥ 0. Then setting E = Hσ , we obtain
|G(x)| = |Q−1/2 F (x)| ≤ kQ−1/2 kL(Hσ ,H) |F (x)σ ≤ K(1
c + |x|σ ), x ∈ Hσ ,

for a suitable K c < ∞, provided ∆ ≤ σ. Moreover, the mapping G : Hσ → H

is continuous since F is continuous in E. In view of Remark 2.5, the re-
sults of the paper can be applied to this case. Also, condition (3.11) is
satisfied with α < 12 , as shown in Example 9.2, and Hypothesis 2.2(b) holds
P 2σ−2∆−1
true for E = Hσ , provided αi < ∞, where (αi ) are the eigenval-
ues of the operator (−A) (cf. [9]), which is true (taking into account that
αi ∼ i2 ) if ∆ > σ − 14 . The remaining condition (2.7) is always satisfied be-
−1/2 −1/2 √
cause Qt St Q1/2 = Q1/2 St Qt = 2(−A)1/2 (I − etA )−1/2 etA and by the
previous Example 9.2, we have that kHt kHS ≤ const.t−3/4 . Summarizing,
assume that
1 1
σ− 4 <∆≤σ< 2
44 B. GOLDYS AND B. MASLOWSKI

holds, which may be achieved by a suitable choice of σ ∈ (0, 12 ) for ∆ ∈ [0, 12 ).

Then V -uniform ergodicity follows from Theorem 6.3. Moreover, if b = 0,
then the existence of the spectral gap in L2 (E, µ∗ ) with E = Hσ follows
from Corollary 7.4. In both cases the estimates on the rate of convergence
are specified in Theorem 6.3 and Corollary 7.4, respectively.

APPENDIX
For the reader’s convenience, we collect here some basic facts about mea-
surable linear mapping that are used in the paper. Most of them are well
known.
Let H be a real separable Hilbert space and let µ = N (0, C) be a centered
Gaussian measure on H with the covariance operator C such that im(C) =
H. The space HC = im(C 1/2 ) endowed with the norm |x|C = |C −1/2 x| can
be identified as the reproducing kernel Hilbert space of the measure µ. In the
sequel we will denote by {en : n ≥ 1} the eigenbasis of C and by {cn : n ≥ 1}
the corresponding set of eigenvalues:
Cen = cn en , n ≥ 1.
For any h ∈ H, we define
n
X 1
φn (x) = √ hh, ek ihx, ek i, x ∈ H.
k=1
ck

Lemma A.1. The sequence (φn ) converges in L2 (H, µ) to a limit φ and

Z
|φ(x)|2 µ(dx) = |h|2 .
H

Moreover, there exists a measurable linear space Mh ⊂ H, such that µ(Mh ) =

1, φ is linear on Mh and
(A.1) φ(x) = lim φn (x), x ∈ Mh .
n→∞

We will use the notation φ(x) = hh, C −1/2 xi.

Let H1 be another real separable Hilbert space and let T : H → H1 be

a bounded operator. The Hilbert–Schmidt norm of T will be denoted by
kT kHS . Let
n
X 1
T̃n x = √ hx, ek iT ek , x ∈ H.
k=1
ck
 ERGODICITY FOR STOCHASTIC PDE’S 45

Lemma A.2. Let T : H → H1 be a Hilbert–Schmidt operator. Then the

sequence (T̃n ) converges in L2 (H, µ; H1 ) to a limit T̃ and
Z
|T̃ (x)|2H1 µ(dx) = kT k2HS .
H
Moreover, there exists a measurable linear space MT ⊂ H, such that µ(MT ) =
1, T̃ is linear on MT and
(A.2) T̃ (x) = lim T̃n x, x ∈ MT .
n→∞

We will use the notation T C −1/2 x = T̃ (x).

Lemma A.3. Let K(t, s) : H → H be an operator-valued, strongly mea-

surable function, such that, for each a ∈ (0, 1),
Z 1Z 1 Z aZ a
(A.3) kK(t, s)kHS ds dt + kK(t, s)k2HS ds dt < ∞.
0 0 0 0
Then the mapping (t, s, y) → K(t, s)C −1/2 y
is measurable, and there exists a
measurable linear space M ⊂ H of full measure, such that, for each y ∈ M,
Z 1
|K(t, s)C −1/2 y| ds < ∞, t-a.e.
0

Proof. Let
Kx = K(t, s)x, x ∈ H.
By assumption, the operator K : H → L2 ((0, a)×(0, a); H)
is Hilbert–Schmidt
for any a < 1 and thereby, by Lemma A.2, there exists the space Ma of full
measure such that
X∞
−1/2 1
KC y= √ hy, ek iKek ,
k=1
ck
where the convergence holds in mean-square and for each y ∈ Ma , in L2 ((0, a)×
(0, a); H). Therefore, KC −1/2 is a measurable
T function of (y, s, t) for s, t ≤ a.
Let an → a be increasing and let M = ∞ n=1 Man . Then KC
−1/2 y is well

defined, for all s, t < 1 is clearly measurable in (y, s, t). Moreover, for each
y ∈ M,
Z an Z an 2
In2 (y) = |K(t, s)C −1/2 y| ds dt
Z 0
an Z 0
an
[3pt] ≤ |K(t, s)C −1/2 y|2 ds dt < ∞.
0 0
Since the sequence In (y) is nondecreasing to a limit I∞ (y) for each y ∈ Ω
and
Z Z 1Z 1
I∞ (y)µ(dy) ≤ kK(t, s)kHS ds dt < ∞,
H 0 0
the lemma follows. 
46 B. GOLDYS AND B. MASLOWSKI

REFERENCES
[1] Aida, S. (1998). Uniform positivity improving property, Sobolev inequalities, and
spectral gaps. J. Funct. Anal. 158 152–185. MR1641566
[2] Aida, S. and Shigekawa, I. (1994). Logarithmic Sobolev inequalities and spectral
gaps: Perturbation theory. J. Funct. Anal. 126 448–475. MR1305076
[3] Albeverio, S., Kondratiev, Y. G. and Röckner, M. (1997). Ergodicity of
L2 -semigroups and extremality of Gibbs states. J. Funct. Anal. 144 394–423.
MR1432591
[4] Arima, R. (1964). On general boundary value problem for parabolic equations. J.
Math. Kyoto Univ. 4 207–243. MR0197997
[5] Cerrai, S. (1999). Ergodicity for stochastic reaction–diffusion systems with polyno-
mial coefficients. Stochastics Stochastics Rep. 67 17–51. MR1717811
[6] Cerrai, S. (2001). Second Order PDE ’s in Finite and Infinite Dimensions. A Prob-
abilistic Approach. Springer, Berlin. MR1840644
[7] Chojnowska-Michalik, A. (2001). Transition Semigroups for Stochastic Semilinear
Equations on Hilbert Spaces. Dissertationes Math. 396. MR1841090
[8] Chojnowska-Michalik, A. and Goldys, B. (2002). Symmetric Ornstein–
Uhlenbeck semigroups and their generators. Probab. Theory Related Fields 124
459–486. MR1942319
[9] Da Prato, G. and Zabczyk, J. (1992). Stochastic Equations in Inifnite Dimensions.
Cambridge Univ. Press. MR1207136
[10] Da Prato, G. and Zabczyk, J. (1996). Ergodicity for Infinite Dimensional Systems.
Cambridge Univ. Press. MR1417491
[11] Da Prato, G., Debussche, A. and Goldys, B. (2002). Some properties of invari-
ant measures of non symmetric dissipative stochastic systems. Probab. Theory
Related Fields 123 355–380. MR1918538
[12] Da Prato, G. and Zabczyk, J. (1992). Nonexplosion, boundedness, and ergod-
icity for stochastic semilinear equations. J. Differential Equations 98 181–195.
MR1168978
[13] Davies, E. B. and Simon, B. (1986). L1 -properties of intrinsic Schrödinger operators.
J. Funct. Anal. 65 126–146. MR0819177
[14] Goldys, B. and Maslowski, B. (2001). Uniform exponential ergodicity of stochastic
dissipative systems. Czechoslovak Math. J. 51 745–762. MR1864040
[15] Goldys, B. and Maslowski, B. (2005). On the Ornstein–Uhlenbeck bridge.
Preprint.
[16] Goldys, B. and Maslowski, B. (2004). Exponential ergodicity for stochastic
reaction-diffusion equations. In Stochastic Partial Differential Equations and
Applications VII. Lecture Notes Pure Appl. Math. 245 115–131. Chapman
Hall/CRC Press, Boca Raton, FL. MR2227225
[17] Goldys, B. and Maslowski, B. (2006). Exponential ergodicity for stochastic Burg-
ers and 2D Navier–Stokes equations. J. Funct. Anal. 226 230–255. MR2158741
[18] Gong, F., Röckner, M. and Wu, L. (2001). Poincaré inequality for weighted first
order Sobolev spaces on loop spaces. J. Funct. Anal. 185 527–563. MR1856276
[19] Hairer, M. (2002). Exponential mixing properties of stochastic PDEs through
asymptotic coupling. Probab. Theory Related Fields 124 345–380. MR1939651
[20] Hairer, M. (2002). Exponential mixing for a PDE driven by a degenerate noise.
Nonlinearity 15 271–279. MR1888852
[21] Hasminskii, R. Z. (1980). Stochastic Stability of Differential Equations. Sijthoff &
Noordhoff, Alphen aan den Rijn. MR0600653
 ERGODICITY FOR STOCHASTIC PDE’S 47

[22] Hino, M. (2000). Exponential decay of positivity preserving semigroups on Lp . Osaka

J. Math. 37 603–624. MR1789439
[23] Lyons, T. J. and Zheng, W. A. (1990). On conditional diffusion processes. Proc.
Roy. Soc. Edinburgh 115A 243–255. MR1069520
[24] Jacquot, S. and Royer, G. (1995). Ergodicite d’une classe d’equations aux de-
rivees partielles stochastiques. C. R. Acad. Sci. Paris Ser. I Math. 320 231–236.
MR1320362
[25] Masiero, F. (2005). Semilinear Kolmogorov equations and applications to stochastic
optimal control. Appl. Math. Optim. 51 201–250. MR2117233
[26] Maslowski, B. (1989). Strong Feller property for semilinear stochastic evolution
equations and applications. In Stochastic Systems and Optimization. Lecture
Notes in Control Inform. Sci. 136 210–224. Springer, Berlin. MR1180781
[27] Maslowski, B. and Seidler, J. (2000). Probabilistic approach to the strong Feller
property. Probab. Theory Related Fields 118 187–210. MR1790081
[28] Maslowski, B. and Simão, I. (1997). Asymptotic properties of stochastic semi-
linear equations by the method of lower measures. Colloq. Math. 72 147–171.
MR1425551
[29] Maslowski, B. and Simão, I. (2001). Long time behaviour of non-autonomous
SPDE’s. Stochastic Process. Appl. 95 285–309. MR1854029
[30] Mattingly, J. C., Stuart, A. M. and Higham, D. J. (2002). Ergodicity for
SDEs and approximations: Locally Lipschitz vector fields and degenerate noise.
Stochastic Process. Appl. 101 185–232. MR1931266
[31] Mattingly, J. C. (2002). Exponential convergence for the stochastically forced
Navier–Stokes equations and other partially dissipative dynamics. Comm. Math.
Phys. 230 421–462. MR1937652
[32] Meyn, S. P. and Tweedie, R. L. (1993). Markov Chains and Stochastic Stability.
Springer, London. MR1287609
[33] Meyn, S. P. and Tweedie, R. L. (1994). Computable bounds for geometric conver-
gence rates of Markov chains. Ann. Appl. Probab. 4 981–1011. MR1304770
[34] van Neerven, J. M. A. M. (1996). The Asymptotic Behaviour of Semigroups of
Linear Operators. Birkhäuser, Basel. MR1409370
[35] van Neerven, J. M. A. M. (1998). Nonsymmetric Ornstein–Uhlenbeck semigroups
in Banach spaces. J. Funct. Anal. 155 495–535. MR1624573
[36] Roberts, G. and Rosenthal, J. (1997). Geometric ergodicity and hybrid Markov
chains. Electron. Comm. Probab. 2 13–25. MR1448322
[37] Seidler, J. (1997). Ergodic behaviour of stochastic parabolic equations. Czechoslo-
vak Math. J. 47 277–316. MR1452421
[38] Seidman, T. I. (1988). How violent are fast controls. Math. Control Signals Systems
1 89–95. MR0923278
[39] Shardlow, T. (1999). Geometric ergodicity for stochastic PDEs. Stochastic Anal.
Appl. 17 857–869. MR1714903
[40] Stettner, L. (1994). Remarks on ergodic conditions for Markov process on Polish
spaces. Bull. Polish Acad. Sci. Math. 42 103–114. MR1810695
[41] Simão, I. (1996). Pinned Ornstein–Uhlenbeck processes on an infinite-dimensional
space. In Stochastic Analysis and Applications (Powys, 1995 ) 401–407. World
Sci. Publishing, River Edge, NJ. MR1453146
[42] Sowers, R. (1992). Large deviations for the invariant measure of a reaction–diffusion
equation with non-Gaussian perturbations. Probab. Theory Related Fields 92
393–421. MR1165518
48 B. GOLDYS AND B. MASLOWSKI

[43] Wu, L. (2004). Essential spectral radius for Markov semigroups. I. Discrete time
case. Probab. Theory Related Fields 128 255–321. MR2031227

School of Mathematics Mathematical Institute

University of New South Wales Academy of Sciences of Czech Republic
Sydney 2052 Žitná 25
Australia 11567 Prague 1
E-mail: [email protected] Czech Republic
E-mail: [email protected]