Lecture 12

The Ornstein–Uhlenbeck semigroup

All of us know the importance of the Laplacian operator and of the heat semigroup,
d
X Z
@2f 1 |x y|2 /4t
f (x) = (x), T (t)f (x) = e f (y)dy, t > 0,
i=1
@x2i (4⇡t)d/2 Rd

that serve as prototypes for elliptic di↵erential operators and semigroups of operators,
respectively. Choosing X = L2 (Rd , d ), the Laplacian : D( ) = W 2,2 (Rd , d ) ! X
is the infinitesimal generator of T (t), namely given any f 2 X, there exists the limit
limt!0+ (T (t)f f )/t if and only if f 2 W 2,2 (Rd , d ), and in this case the limit is f . The
W 2,2 norm is equivalent to the graph norm. Moreover, the realization of the Laplacian in
X is the operator associated with the quadratic form
Z
Q(u, v) = ru · rv dx, u, v 2 W 1,2 (Rd , d ).
Rd

This means that D( ) consists precisely

R of the elements u 2 W 1,2 (Rd , d ) such that the
function W (R , d ) ! R, ' 7! Rd ru · r' dx has a linear bounded extension to the
1,2 d

whole X, namely there exists g 2 L2 (Rd , d ) such that

Z Z
ru · r' dx = g ' dx, ' 2 W 1,2 (Rd , d ),
Rd Rd

and in this case g = u. If u 2 W 2,2 (Rd , d ), the above formula follows just integrating
by parts, and it is the basic formula that relates the Laplacian and the gradient. Moreover,
for u 2 W 2,2 (Rd , d ) we have
u = div ru,
where the divergence div is (minus) the adjoint of the gradient r : P W 1,2 (Rd , d ) !
L (R , d ; R ), and for a vector field v 2 W (R , d ; R ) it is given by di=1 @vi /@xi .
2 d d 1,2 d d

In this lecture and in the next ones we introduce the Ornstein–Uhlenbeck operator and
the Ornstein–Uhlenbeck semigroup, that play the role of the Laplacian and of the heat

145
146 Lecture 12

semigroup if the Lebesgue measure is replaced by the standard Gaussian measure d , and
that have natural extensions to our infinite dimensional setting (X, , H). As before, X
is a separable Banach space endowed with a centred nondegenerate Gaussian measure ,
and H is the relevant Cameron–Martin space.

12.1 The Ornstein–Uhlenbeck semigroup

In this section we define the Ornstein–Uhlenbeck semigroup; we start by defining it in the
space of the bounded continuous functions and then we extend it to Lp (X, ), for every
p 2 [1, +1).
The Ornstein–Uhlenbeck semigroup in Cb (X) is defined as follows: T (0) = I, and for
t > 0, T (t)f is defined by the Mehler formula
Z p
T (t)f (x) := f (e t x + 1 e 2t y) (dy). (12.1.1)
X

We list some properties of the family of operators {T (t) : t 0}.

Proposition 12.1.1. {T (t) : t 0} is a contraction semigroup in Cb (X). Moreover, for
every f 2 Cb (X) we have
Z Z
T (t)f d = fd , t > 0. (12.1.2)
X X

Proof. First of all we notice that |T (t)f (x)|  kf k1 for every x 2 X and t 0. The fact
that T (t)f 2 Cb (X) follows by Dominated Convergence Theorem. So, T (t) 2 L(Cb (X))
and kT (t)kL(Cb (X))  1. Taking f ⌘ 1, we have T (t)f ⌘ 1 so that kT (t)kL(Cb (X)) = 1 for
every t 0.
Let us prove that {T (t) : t 0} is a semigroup. For every f 2 Cb (X) and t, s > 0 we
have
Z p
(T (t)(T (s)f ))(x) = (T (s)f )(e t x + 1 e 2t y) (dy)
ZX Z p p
= f (e s (e t x + 1 e 2t y) + 1 e 2s z) (dy) (dz).
X X
p p
Setting now (y, z) = e s p 1 e 2t y + p 1 e
2s
z, and using Proposition 2.2.7(iv), we
1 e 2t 2s 1 e 2t 2s
get
Z Z p
s t 2t 2s
(T (t)(T (s)f ))(x) = f (e x+ 1 e (y, z)) (dy) (dz)
ZX X
p
t s 2t 2s ⇠)(( 1
= f (e x+ 1 e ⌦ ) )(d⇠)
ZX p
t s 2t 2s ⇠)
= f (e x+ 1 e (d⇠)
X
= T (t + s)f (x).
The Ornstein–Uhlenbeck semigroup 147

Let us prove that (12.1.2) holds. For any f 2 Cb (X) we have

Z Z Z p
T (t)f (dx) = f (e t x + 1 e 2t y) (dy) (dx).
X X X
p
Setting (x, y) = pe t x + 1 e 2t y, we apply Proposition 2.2.7(iv) with any ✓ 2 R such
t
that e = cos ✓, 1 e 2t = sin ✓, and we get
Z Z Z
1
T (t)f d = f (⇠)( ⌦ ) (d⇠) = f (⇠) (d⇠).
X X X

We point out that the semigroup {T (t) : t 0} is not strongly continuous in Cb (X),
and not even in the subspace BU C(X) of the bounded and uniformly continuous functions.
In fact, we have the following characterisation.
Lemma 12.1.2. Let f 2 BU C(X). Then

lim kT (t)f f k1 = 0 () lim kf (e t ·) f k1 = 0.

t!0+ t!0+

Proof. For every t > 0 and x 2 X we have

Z p
t
(T (t)f f (e ·))(x) = (f (e t x + 1 e 2t y) f (e t x)) (dy)
X

and the right hand side goes to 0, uniformly in X, as t ! 0+ . Indeed, given " > 0 fix R > 0
such that (X \ B(0, R))  p ", and fix > 0 such that |f (u) f (v)|  " for ku vk  .
Then, for every t such that 1 e 2t R  and for every x 2 X we have
Z p
(f (e t x + 1 e 2t y) f (e t x)) (dy)
X
✓Z Z ◆ p
= + (f (e t x + 1 e 2t y) f (e t x)) (dy)
B(0,R) X\B(0,R)
 " + 2kf k1 ".

The simplest counterexample to strong continuity in BU C(X) is one dimensional: see

Exercise 12.1.
However, for every f 2 Cb (X) the function (t, x) 7! T (t)f (x) is continuous in [0, +1)⇥
X by the Dominated Convergence Theorem. In particular,

lim T (t)f (x) = f (x), 8x 2 X,

t!0+

which is enough for many purposes.

The semigroup {T (t) : t 0} enjoys important smoothing and summability improving
properties. The first smoothing property is in the next proposition.
148 Lecture 12

Proposition 12.1.3. For every f 2 Cb (X) and t > 0, T (t)f is H-di↵erentiable at every
x 2 X, and we have
Z p
e t
[rH T (t)f (x), h]H = p f (e t x + 1 e 2t y)ĥ(y) (dy). (12.1.3)
1 e 2t X
Therefore,
e t
|rH T (t)f (x)|H  p |h|H kf k1 , x 2 X. (12.1.4)
1 e 2t

Proof. Set
e t
c(t) := p .
1 e 2t

For every h 2 H we have

Z p
T (t)f (x + h) = f (e t x + 1 e 2t (c(t)h + y)) (dy)
X
Z n
p |h|2H o
= f (e t x + 1 e 2t z) exp c(t)ĥ(z) c(t)2 (dz),
X 2

by the Cameron–Martin formula. Therefore, denoting by lt,x (h) the right hand side of
(12.1.3),

|T (t)f (x + h) T (t)f (x) lt,x (h)|


|h|H
Z ⇣ n
1 p |h|2 o ⌘
 f (e t x + 1 e 2t y) exp c(t)ĥ(y) c(t)2 H 1 c(t)ĥ(y) (dy)
|h|H X 2
Z n
kf k1 |h|2 o
 exp c(t)⇠ c(t)2 H 1 c(t)⇠ N (0, |h|2H )(d⇠)
|h|H R 2
Z n |h|2 o
= kf k1 exp c(t)|h|H ⌘ c(t)2 H 1 c(t)|h|H ⌘ N (0, 1)(d⌘)
R 2

where the right hand side vanishes as |h|H ! 0, by the Dominated Convergence Theorem.
This proves (12.1.3). In its turn, (12.1.3) yields

|[rH T (t)f (x), h]H |  c(t)kf k1 kĥkL1 (X, )  c(t)kf k1 kĥkL2 (X, ) = c(t)kf k1 |h|H

for every h 2 H, and (12.1.4) follows.

Notice that in the case X = Rd , = d, we have rH = r and formula (12.1.3) reads

as
t Z p
e
Di T (t)f (x) = p f (e t x + 1 e 2t y)y
i d (dy), i = 1, . . . , d. (12.1.5)
1 e 2t
Rd

Let us consider now more regular functions f .

The Ornstein–Uhlenbeck semigroup 149

Proposition 12.1.4. For every f 2 Cb1 (X), T (t)f 2 Cb1 (X) for any t 0, and its
derivative at x is
Z p
0
(T (t)f ) (x)(v) = e t
f 0 (e t x + 1 e 2t y)(v) (dy). (12.1.6)
X

In particular,
@ ⇣ ⇣ @ ⌘⌘
T (t)f (x) = e t T (t) f (x), v 2 X, x 2 X. (12.1.7)
@v @v
For every f 2 Cb2 (X), T (t)f 2 Cb2 (X) for any t 0, and its second order derivative at x
is Z p
(T (t)f )00 (x)(u)(v) = e 2t
f 00 (e t x + 1 e 2t y)(u)(v) (dy), (12.1.8)
X

so that ✓ ◆
@ 2 T (t)f 2t @2f
(x) = e T (t) (x), u, v 2 X, x 2 X. (12.1.9)
@u@v @u@v
p
Proof. Fix t > 0 and set (x, y) = e t x + 1 e 2t y. For every x, v 2 X we have
Z p 1
(T (t)f )(x + v) (T (t)f )(x) e f 0 (e t x + 1 e 2t y)(v) (dy)
t
X kvk
Z
1
 |f ( (x, y) + e t v) f ( (x, y)) f 0 ( (x, y))(e t v)| (dy) .
X kvk

On the other hand, for every y 2 X we have

f ( (x, y)e t v) f ( (x, y)) f 0 ( (x, y))(e t v)

lim = 0,
v!0 kvk

and (see Exercise 12.3)

|f ( (x, y) + e t v) f ( (x, y)) f 0 ( (x, y))(e t v)|

 2e t
sup kf 0 (z)kX ⇤ ,
kvk z2X

and (12.1.6) follows by the Dominated Convergence Theorem. Formula (12.1.7) is an

immediate consequence. The derivative (T (t)f )0 is continuous still by the Dominated
Convergence Theorem. The verification of (12.1.8) and (12.1.9) for f 2 Cb2 (X) follow
iterating this procedure (see Exercise 12.4).

Let us compare Proposition 12.1.3 and Proposition 12.1.4. Proposition 12.1.3 describes
a smoothing property of T (t), while Proposition 12.1.4 says that T (t) preserves the spaces
Cb1 (X) and Cb2 (X). In general, T (t) regularises only along H and it does not map Cb (X)
into C 1 (X). If X = Rd and = d we have H = Rd , this difficulty does not arise,
Proposition 12.1.3 says that T (t) maps Cb (Rd ) into Cb1 (Rd ) and in fact one can check that
T (t) maps Cb (Rd ) into Cbk (Rd ) for every k 2 N, as we shall see in the next lecture.
150 Lecture 12

If f 2 Cb1 (X) we can write rH T (t)f (x) in two di↵erent ways, using (12.1.3) and
(12.1.6): for every h 2 H we have

t Z p
e
[rH T (t)f (x), h]H = p f (e t x + 1 e 2t y)ĥ(y) (dy)
1Z e 2t
X
p
=e t
f 0 (e t x + 1 e 2t y)(h) (dy).
X

We now extend T (t) to Lp (X, ), 1  p < 1.

Proposition 12.1.5. Let t 0. For every f 2 Cb (X) and p 2 [1, +1) we have

kT (t)f kLp (X, )  kf kLp (X, ) . (12.1.10)

Hence {T (t) : t 0} is uniquely extendable to a contraction semigroup {Tp (t) : t 0} in

Lp (X, ). Moreover

(i) {Tp (t) : t 0} is strongly continuous in Lp (X, ), for every p 2 [1, +1);

(ii) T2 (t) is self–adjoint and nonnegative in L2 (X, ) for every t > 0;

(iii) is an invariant measure for {Tp (t) : t 0}.

Proof. For every f 2 Cb (X), t > 0 and x 2 X the Hölder inequality yields
Z p
p
|T (t)f (x)|  |f (e t x + 1 e 2t y)|p d = T (t)(|f |p )(x).
X

Integrating over X and using (12.1.2) we obtain

Z Z Z
|T (t)f |p d  T (t)(|f |p ) d = |f |p d .
X X X

Since Cb (X) is dense in Lp (X, ), T (t) has a unique bounded extension Tp (t) to the whole
Lp (X, ), such that kTp (t)kL(Lp (X, ))  1. In fact, taking f ⌘ 1, Tp (t)f ⌘ 1 so that
kTp (t)kL(Lp (X, )) = 1.
Let us prove that {Tp (t) : t 0} is strongly continuous. We already know that for
f 2 Cb (X) we have limt!0+ T (t)f (x) = f (x) for every x 2 X, and moreover |T (t)f (x)| 
kf k1 for every x. By the Dominated Convergence Theorem, limt!0+ T (t)f = f in
Lp (X, ). Since Cb (X) is dense in Lp (X, ) and kTp (t)kL(Lp (X, )) = 1 for every t, then
limt!0+ Tp (t)f = f for every f 2 Lp (X, ).
Let us prove statement (ii). Let f , g 2 Cb (X), t > 0. Then
Z Z p
hT (t)f, giL2 (X, ) = f (e t x + 1 e 2t y)g(x) (dy) (dx)
X X
The Ornstein–Uhlenbeck semigroup 151
p p
and setting u = e t x + 1 e 2t y, v = 1 e 2t x + e t y, (u, v) =: R(x, y), using
Proposition 2.2.7(iii) and the fact that is centred we get
Z p
hT (t)f, giL2 (X, ) = f (u)g(e t u 1 e 2t v)(( ⌦ ) R 1
)(d(u, v))
X⇥X
Z Z p
= f (u)g(e t u 1 e 2t v) (dv) (du)
ZX ZX p
= f (u)g(e t u + 1 e 2t v) (dv) (du)
X X
= hf, T (t)giL2 (X, ) .

Approximating any f , g 2 L2 (X, ) by elements of Cb (X), we obtain hT2 (t)f, giL2 (X, ) =
hf, T2 (t)giL2 (X, ) .
Still for every f 2 L2 (X, ) and t > 0 we have

hT2 (t)f, f iL2 (X, ) = hT2 (t/2)T2 (t/2)f, f iL2 (X, ) = hT2 (t/2)f, T2 (t/2)f iL2 (X, ) 0.

Statement (iii) is an immediate consequence of (12.1.2).

To simplify some statements we extend the Ornstein–Uhlenbeck semigroup T (t) to

vector valued functions. For v 2 Cb (X; H) and t > 0 we set
Z p
T (t)v(x) = v(e t x + 1 e 2t y) (dy), t > 0, x 2 X.
X

By Remark 9.2.6, for every orthonormal basis {hi : i 2 N} of H we have

1
X
T (t)v(x) = T (t)([v(·), hi ]H )(x)hi .
i=1

Using Proposition 9.2.5(i) we get the estimate

Z p
|T (t)v(x)|H  |v(e t x + 1 e 2t y)|H (dy) = (T (t)(|v|H ))(x), x 2 X. (12.1.11)
X

Notice that the right hand side concerns the scalar valued function |v|H . Raising to the
power p, integrating over X and recalling (12.1.2), we obtain

kT (t)vkLp (X, ;H)  kvkLp (X, ;H) , t 0. (12.1.12)

As in the case of real valued functions, since Cb (X; H) is dense in Lp (X, ; H), estimate
(12.1.12) allows to extend T (t) to a bounded (contraction) operator in Lp (X, ; H), called
Tp (t). We will not develop the theory for vector valued functions, but we shall use this
notion to write some formulae in a concise way, see e.g. (12.1.13).
Lp gradient estimates for Tp (t)f are provided by the next proposition.

Proposition 12.1.6. Let 1  p < 1.

152 Lecture 12

(i) For every f 2 W 1,p (X, ) and t > 0, Tp (t)f 2 W 1,p (X, ) and

rH Tp (t)f = e t Tp (t)(rH f ), (12.1.13)

kTp (t)f kW 1,p (X, )  kf kW 1,p (X, ) . (12.1.14)

(ii) If p > 1, for every f 2 Lp (X, ) and t > 0, Tp (t)f 2 W 1,p (X, ) and
Z Z
p p
|rH Tp (t)f (x)|H d  c(t, p) |f |p d , (12.1.15)
X X

✓ ◆1/p0
R 0 p
with c(t, p) = R |⇠|p N (0, 1)(d⇠) e t/ 1 e 2t .

Proof. (i). Let f 2 Cb1 (X). By Proposition 12.1.4, T (t)f 2 Cb1 (X) and (@h T (t)f )(x) =
e t (T (t)(@h f ))(x) for every h 2 H, namely [(rH T (t)f )(x), h]H = e t T (t)([rH f, h]H )(x).
Therefore, |(rH T (t)f )(x)|H  e t T (t)(|rH f |H )(x), for every x 2 X. Consequently,

|rH T (t)f (x)|pH  e tp

(T (t)(|rH f |H )(x))p  e tp
(T (t)(|rH f |pH )(x))

and integrating we obtain

Z Z Z
|rH T (t)f (x)|pH d  e tp
T (t)(|rH f |H ) d = e p tp
|rH f |pH d ,
X X X

so that

kT (t)f kW 1,p (X, ) = kT (t)f kLp (X, ) + k |rH T (t)f |H kLp (X, )
 kf kLp (X, ) + k |rH f |H kLp (X, ) = kf kW 1,p (X, ) .

Since Cb1 (X) is dense in W 1,p (X, ), (12.1.14) follows.

(ii) Let f 2 Cb (X). By Proposition 12.1.3, T (t)f is H-di↵erentiable at every x, and
we have
|rH T (t)f (x)|H = sup |[rH T (t)f (x), h]H |.
h2H, |h|H =1

Let us estimate |[rH T (t)f (x), h]H | = |lt,x (h)| (where lt,x (h) is as in the proof of Proposi-
tion 12.1.3), for |h|H = 1, using formula (12.1.3). We have
t Z p
e
|lt,x (h)|  p |f (e t x + 1 e 2t y)|ĥ(y) (dy)
1 e 2t
X
t
✓Z p ◆1/p ✓ Z ◆1/p0
e t 2t y)| p p0
p |f (e x + 1 e (dy) |ĥ| d
1 e 2t
X X
t
✓Z ◆1/p0
e p 1/p p0
=p (T (t)(|f | )(x)) |⇠| N (0, 1)(d⇠) .
1 e 2t
R
The Ornstein–Uhlenbeck semigroup 153

By the invariance property (12.1.2) of ,

Z ✓ t
◆p Z ✓Z ◆p/p0
e p0
|rH T (t)f (x)|pH (dx)  p p
(T (t)(|f | ) d |⇠| N (0, 1)(d⇠)
1 e 2t
X X R
✓ t
◆p Z ✓Z ◆p/p0
e p p0
= p |f | d |⇠| N (0, 1)(d⇠) .
1 e 2t
X R

Therefore, T (t)f 2 W 1,p (X, ) and estimate (12.1.15) holds for every f 2 Cb (X). Since
Cb (X) is dense in Lp (X, ), the statement follows.

Note that the proof of (ii) fails for p = 1, since for every h 2 H the function ĥ does not
belong to L1 , and the constant c(t, p) in estimate (12.1.15) blows up as p ! 1. Indeed,
T (t) does not map L1 (X, ) into W 1,1 (X, ) continuously for t > 0, even in the simplest
case X = R, = 1 (see for instance [MPP, Corollary 5.1]).

12.2 Exercises
Exercise 12.1. Let X = R and set f (x) = sin x. Prove that T (t)f does not converge
uniformly to f in R as t ! 0.

Exercise 12.2. Show that the argument used in Proposition 12.1.5 to prove that T (t) is
self–adjoint in L2 (X, ) implies that Tp0 (t) = Tp (t)⇤ for p 2 (1, +1) with 1/p + 1/p0 = 1.

Exercise 12.3. Prove that for every f 2 C 1 (X) and for every x, y 2 X we have
Z 1
f (y) f (x) = f 0 ( y + (1 )x)(y x) d ,
0

so that, if f 2 Cb1 (X),

|f (y) f (x)|  sup kf 0 (z)kX 0 ky xk.

z2X

Exercise 12.4. Prove that for every f 2 Cb2 (X) and t > 0, T (t)f 2 Cb2 (X) and (12.1.8),
(12.1.9) hold.

Bibliography
[B] V. I. Bogachev: Gaussian Measures. American Mathematical Society, 1998.

[MPP] G. Metafune, D. Pallara, E. Priola: Spectrum of Ornstein–Uhlenbeck op-

erators in Lp spaces with respect to invariant measures, J. Funct. Anal. (196) (2002)
40–60.