Convergence of Markov Processes

January 13, 2016

Martin Hairer
Mathematics Department, University of Warwick
Email: [email protected]

Abstract
The aim of this minicourse is to provide a number of tools that allow one to de-
termine at which speed (if at all) the law of a diffusion process, or indeed a rather
general Markov process, approaches its stationary distribution. Of particular in-
terest will be cases where this speed is subexponential. After an introduction
to the general ergodic theory of Markov processes, the first part of the course
is devoted to Lyapunov function techniques. The second part is then devoted to
an elementary introduction to Malliavin calculus and to a proof of Hörmander’s
famous ”sums of squares” regularity theorem.

1 General (ergodic) theory of Markov processes

In this note, we are interested in the long-time behaviour of Markov processes,
both in discrete and continuous time. Recall that a discrete-time Markov process
x on a state space X is described by a transition kernel P, which we define as
a measurable map from X into the space of probability measures on X . In all
that follows, X will always be assumed to be a Polish space, that is a complete,
separable metric space. When viewed as a measurable space, we will always endow
it with its natural Borel σ-algebra, that is the smallest σ-algebra containing all open
sets. This ensures that X endowed with any probability measure is a Lebesgue
space and that basic intuitive results of probability and measure theory (Fubini’s
theorem, regular conditional probabilities, etc) are readily available.
We will sometimes consider P as a linear operator on the space of signed mea-
sures on X and / or the space of bounded measurable functions by
Z Z
(Pµ)(A) = P(x, A) µ(dx) , (Pϕ)(x) = ϕ(y) P(x, dy) .
X X

Hence we are using P both to denote the action on functions and its dual action on
measures. Note that P extends trivially to measurable functions ϕ : X → [0, +∞].
The main concept arising in the study in the long-time behaviour of a Markov chain
is that of an invariant measure:
G ENERAL ( ERGODIC ) THEORY OF M ARKOV PROCESSES 2

Definition 1.1 A positive measure µ on X is invariant for the Markov process x if

Pµ = µ.
In the case of discrete state space, another key notion is that of transience, re-
currence and positive recurrence of a Markov chain. The next subsection explores
these notions and how they relate to the concept of an invariant measure.

1.1 Transience and recurrence

Take a Markov process on a countable state space, say Z and denote its transition
probabilities by Pij = P(j, {i}). In this simple case, the qualitative long-time
behaviour of the chain can be described in the following way. For a given state i,
we denote by τi the first return time of the process to {i}, i.e.

τi = inf{k ≥ 1 : xk = i} .

We also denote by Pi the law of the process started at i. With these notations, we
say that the state i is:

• Transient if Pi (τi = ∞) > 0.

• Recurrent if Pi (τi < ∞) = 1.

• Positive recurrent if it is recurrent and Ei τi < ∞.

We assume that P is irreducible in the sense that for every pair (i, j) there exists
n > 0 such that Pijn > 0. Given a state i, we also denote by τi the first return time
to i, that is τi = inf{n > 0 : xn = i}. We have

Proposition 1.2 If P is irreducible, then all of the states are of the same type.

Proof. The proof is based on the following fact. If P is irreducible and j 6= k

are any two states, then Pj (τk < τj ) > 0. The argument goes by contradiction:
assume that the probability vanishes. Then, by the strong Markov property, the
process starting from j would never visit the state k. This is in contradiction with
the irreducibility of P .
We first show that we cannot have one recurrent and one transient state. As-
sume by contradiction that i is transient and that k 6= i is recurrent. Start the
process in the state k so that, by the recurrence of k and the strong Markov prop-
erty, it almost surely visits k infinitely often. The above remark on the other hand
ensure that between each of these visits, the process has a strictly positive proba-
bility of visiting i. Since these ‘trials’ are independent, the process almost surely
visits i infinitely often. From the strong Markov property applied at the first hitting
time of i, we conclude that i must be recurrent as well.
It remains to show that if k is positive recurrent, then i must also be so. Denote
by A the event that the process visits i between two successive visits to k, and
G ENERAL ( ERGODIC ) THEORY OF M ARKOV PROCESSES 3

denote by p the probability of A. If p = 1, then Ei τi ≤ 2Ek τk and we are done, so

we can assume that p ∈ (0, 1), so that
Ek τk Ek τk
Ek (τk | A) ≤ , Ek (τk | Ā) ≤ .
p 1−p
We now break the excursion from i into a trajectory from i to k, a number of
excursions from k to k, and a final piece of trajectory from k back to i. Since the
first and the last piece of trajectory are part of an excursion from k that does visit
i, this yields the upper bound
Ek τk X Ek τk Ek τk
Ei τi ≤ 2 +p (1 − p)n n ≤3 ,
p 1−p p
n≥0

thus concluding the proof.

It turns out that if a Markov chain is positive recurrent, then it has a finite in-
variant measure. If it is only recurrent, then it still has a σ-finite invariant measure.
If on the other hand it is transient, the notion of an invariant measure is useless
since the process visits states only finitely often. To show this, it suffices to make
the following construction. Fix a distinguished state 4 and denote by x̃ the process
that starts at 4 and stops as soon as it reaches 4 again, but is conditioned to spend
at least one time step outside of 4. The transition probabilities P̃ij for x̃ are then
given by 
 Pij if j 6= 4,
P̃ij = 1 if i = j = 4.
0 otherwise.


We now set π4 = 1 and, for j 6= 4,

∞
X ∞
X
πj = P(x̃k = j) = E 1{j} (x̃k ) .
k=1 k=1

Note first that since there is a non-zero probability that the process reaches 4
between any two successive return times to j, this quantity is finite for every j.
Furthermore, it follows from the definition of x̃ that we have the identity
 P∞ P k
πj = k=0 i6=4 P̃ji Pi4 if j 6= 4,
1 if j = 4.
Therefore, for ` 6= 4 we have
XXX XX X
(P π)` = P`4 + P`j P̃jik Pi4 = P`4 + P̃`j P̃jik Pi4 = π` .
j6=4 k≥0 i6=4 j k≥0 i6=4

On the other hand, we have

∞ X
XX
(P π)4 = P44 + P4j P̃jik Pi4 = 1 ,
j6=4 k=0 i6=4

since this is precisely the probability that the process x eventually returns to 4.
G ENERAL ( ERGODIC ) THEORY OF M ARKOV PROCESSES 4

1.2 Foster-Lyapunov criteria

We have the following Foster-Lyapunov drift criteria on countable state space. As
a shorthand notation, we define the ‘discrete generator’ of our Markov process by
Lf = Pf − f .

Proposition 1.3 Consider a Markov process with transition probabilities P on a

countable state space. Then:
• It is transient if and only if there exists a function V : X → R+ and a non-
empty set A ⊂ X such that LV (x) ≤ 0 for all x 6∈ A and there exists x 6∈ A
such that V (x) < infy∈A V (y).

• It is recurrent if and only if there exists a function V : X → R+ such that

{x : V (x) ≤ N } is finite for every N > 0 and such that LV (x) ≤ 0 for all
but finitely many values of x.

• It is positive recurrent if and only if there exists a function V : X → R+ such

that LV (x) ≤ −1 for all but finitely many values of x.

Proof. In all three cases, we will show that the existence of a function V with the
specified properties is sufficient to obtain the corresponding transience / recurrence
properties. We then show how to construct such a function in an abstract way.
Let us first show the criterion for transience. Multiplying V by a positive con-
stant, we can assume without loss of generality that infy∈A V (y) = 1. Consider
now x 6∈ A such that V (x) < 1. Since LV (z) ≤ 0 for z 6∈ A, we have
Z Z
V (x) ≥ V (y)P(x, dy) ≥ P(x, A) + V (y)P(x, dy)
Ac Z
Z Z (1.1)
≥ P(x, A) + P(y, A)P(x, dy) + V (z)P(y, dz)P(x, dy) ≥ . . .
Ac Ac Ac

Taking limits, we see that V (x) ≥ Px (τA < ∞). Since V (x) < 1, this immediately
implies that the process is transient. Conversely, if the process is transient, the
function V (x) = Px (τA < ∞) satisfies the required conditions.
We now turn to the condition for recurrence. For N ≥ 0, set VN (x) = V (x)/N
and set DN = {x : VN (x) ≥ 1}. It follows from the assumption that the sets
DN have finite complements. Denote furthermore A = {x : LV (x) > 0}. A
calculation virtually identical to (1.1) then shows that VN (x) ≥ Px (τDN < τA ) ≥
Px (τA = ∞). In particular, one has Px (τA = ∞) = 0 for every x, so that the
process is recurrent. Conversely, assume that the process is transient and consider
an arbitrary finite set A, as T
well as a sequence of decreasing sets DN with finite
complements and such that N >0 DN = φ. We then set WN (x) = Px (τDN < τA )
with the convention that WN (x) = 0 for x ∈ A and WN (x) = 1 for x ∈ DN . It is
straightforward do check that one does have LWN ≤ 0 for x 6∈ A. Furthermore, it
follows from the recurrence of the process that limN →∞ WN P (x) = 0 for every x.
We can therefore find a sequence Nk → ∞ such that V (x) = k≥0 WNk (x) < ∞
G ENERAL ( ERGODIC ) THEORY OF M ARKOV PROCESSES 5

for every x. Since V grows at infinity by construction, it does indeed satisfy the
required conditions.
We finally consider the criterion for positive recurrence. We define the set
A = {x : LV (x) > −1} which is finite by assumption. For x 6∈ A we now have
Z Z
V (x) ≥ 1 + V (y)P(x, dy) ≥ 1 + V (y)P(x, dy)
A c
Z Z
c
≥ 1 + P(x, A ) + V (z)P(y, dz)P(x, dy) ≥ . . .
Ac Ac

Taking limits again, we obtain that V (x) ≥ Ex τA , so that the first hitting time of
A has finite expectation. Since it follows from our assumption that LV is bounded
from above (in particular LV (x) < ∞ on A), this shows that the return time to any
fixed state has finite expectation. Conversely, if the process is positive recurrent,
we set V (x) = Ex τA for an arbitrary finite set A and we check as before that it
does have the required properties.

Actually the criterion for positive recurrence can be slightly strengthened:

Proposition 1.4 if there exist functions V, F : X → R+ such that LV is bounded

and LV (x) ≤ −1 − F (x) for R all but finitely many values of x, then the unique
invariant measure π satisfies F (x) π(dx) < ∞.

Proof. It suffices to show that π has the desired property. Denoting VN (x) =
V (x) ∧ N , we see that GN = LVN is bounded, negative outside
R a finite set, and
satisfies limN →∞ GN (x) = G(x) = LV (x). Furthermore, GN (x) π(dx) = 0 by
the invariance of π. The claim now follows from Fatou’s lemma.

Recurrence / transience of the random walk. The random walk on Zd has

generator
1 X
Lf (x) = (f (y) − f (x)) .
2d y∼x

Our aim is to use the above criteria to show that it is recurrent if and only if d ≤ 2.
It is clear that it is irreducible and aperiodic. It cannot be positive recurrent
since the density of any invariant probability measure would have to satisfy Lπ =
0, so that there cannot be a point x with p(x) maximal. The question of interest is
whether the random walk is transient or recurrent. We use the fact that if f varies
slowly, then Lf (x) ≈ ∆f (x).
Note now that if we set f (x) = kxk2α , we have

∆f (x) = 2α(d + 2α − 2)kxk2α−2 .

This simple calculation already shows that 2 is the critical dimension for the tran-
sience / recurrence transition since for d < 2 one can find α > 0 so that f satisfies
the conditions for the recurrence criterion, whereas for d > 2, one can find α < 0
G ENERAL ( ERGODIC ) THEORY OF M ARKOV PROCESSES 6

such that f satisfies the conditions for the transience criterion. What happens at the
critical dimension? We consider

f (x) = (log(x21 + x22 ))α ,

so that
4α(α − 1)
∆f (x) = (log(x21 + x22 ))α−2 .
x21 + x22
Choosing α ∈ (0, 1) allows to construct a function such that the criterion for recur-
rence is satisfied.

1.3 Ergodic invariant measures

An important class of invariant measures are the ergodic invariant measures. In-
formally speaking, these are precisely those measures for which a law of large
numbers holds. In order to describe them more precisely, we introduce the notion
of a P-invariant set:

Definition 1.5 Given a Markov kernel P, a measurable set A ⊂ P is P-invariant

if P(x, A) = 1 for every x ∈ A.
With this definition at hand, we have

Definition 1.6 An invariant probability measure µ for P is ergodic if every P-

invariant set has µ-measure either 0 or 1.
The importance of invariant measures can be seen in the following structural
theorem, which is a consequence of Birkhoff’s ergodic theorem:

Theorem 1.7 Given a Markov kernel P, denote by I the set of all invariant prob-
ability measures for P and by E ⊂ I the set of all those that are ergodic. Then, I
is convex and E is precisely the set of its extremal points. Furthermore, for every
invariant measure µ ∈ I, there exists a probability measure Qµ on E such that
Z
µ(A) = ν(A) Qµ (dν) .
E

In other words, every invariant measure is a convex combination of ergodic invari-

ant measures. Finally, any two distinct elements of E are mutually singular.

Remark 1.8 As a consequence, if a Markov process admits more than one invari-
ant measure, it does admit at least two ergodic (and therefore mutually singular)
ones. This leads to the intuition that, in order to guarantee the uniqueness of its
invariant measure, it suffices to show that a Markov process explores its state space
‘sufficiently thoroughly’.
G ENERAL ( ERGODIC ) THEORY OF M ARKOV PROCESSES 7

1.4 Existence and uniqueness criteria

In this section, we present two main existence and uniqueness criteria for the in-
variant measure of a Markov process. First, existence essentially follows from a
continuity and compactness result, which can be formulated in terms of the Feller
property:

Definition 1.9 A Markov operator P is Feller if Pϕ is continuous for every con-

tinuous bounded function ϕ : X → R. In other words, it is Feller if and only if the
map x 7→ P(x, ·) is continuous in the topology of weak convergence.

Theorem 1.10 (Krylov-Bogolioubov) Let P be a Feller Markov operator over a

Polish space X . Assume that there exists x0 ∈ X such that the sequence P n (x0 , ·)
is tight. Then, there exists at least one invariant probability measure for P.

Proof. Let µN be the sequence of probability measures defined by

N −1
1 X n
µN (A) = P (x0 , A) . (1.2)
N
n=0

Since our assumption immediately implies that {µN }N ≥1 is tight, there exists at
least one accumulation point µ∗ and a sequence Nk with Nk → ∞ such that
µNk → µ∗ weakly. Take now an arbitrary test function ϕ ∈ Cb (X ) and denote by
k · k the supremum norm of ϕ. One has

k(Pµ∗ )(ϕ) − µ∗ (ϕ)k = kµ∗ (Pϕ) − µ∗ (ϕ)k = lim kµNk (Pϕ) − µNk (ϕ)k
k→∞
1
= lim kµNk (Pϕ) − µNk (ϕ)k = lim kP N ϕ − ϕk
k→∞ k→∞ Nk
2kϕk
≤ lim =0.
k→∞ Nk

Here, the second equality relies on the fact that Pϕ is continuous since P was
assumed to be Feller. Since ϕ was arbitrary, this shows that Pµ∗ = µ∗ as requested.

Example 1.11 Take X = [0, 1] and consider the transition probabilities defined
by 
δx/2 if x > 0
P(x, ·) =
δ1 if x = 0.
It is clear that this Markov operator cannot have any invariant probability mea-
sure. Indeed, assume that µ is invariant. Clearly, one must have µ({0}) = 0 since
P(x, {0}) = 0 for every x. Since, for x 6= 0, one has P(x, {(1/2, 1]}) = 0,
one must also have µ((1/2, 1]) = 0. Proceeding by induction, we have that
µ((1/2n , 1]) = 0 for every n and therefore µ((0, 1]) = 0. Therefore, µ(X ) = 0
which is a contradiction.
S OME SIMPLE UNIQUENESS CRITERIA 8

Endowing X with the usual topology, it is clear that the ‘Feller’ assumption of
the Krylov-Bogolioubov criteria is not satisfied around 0. The tightness criterion
however is satisfied since X is a compact space. On the other hand, we could add
the set {0} to the topology of X , therefore really interpreting it as X = {0}t(0, 1].
Since {0} already belongs to the Borel σ-algebra of X , this change of topology
does not affect the Borel sets. Furthermore, the space X is still a Polish space
and it is easy to check that the Markov operator P now has the Feller property!
However, the space X is no longer compact and a sequence {xn } accumulating at
0 is no longer a precompact set, so that it is now the tightness assumption that is
no longer satisfied.

2 Some simple uniqueness criteria

The following definition captures what we mean by the fact that a given point of
the state space can be ‘visited’ by the dynamic:

Definition 2.1 Let P be a Markov operator over a Polish space X and let x ∈ X .
We say that x is accessible for P if, for every y ∈ X and every open neighborhood
U of x, there exists k > 0 such that P k (y, U ) > 0.
It is straightforward to show that if a given point is accessible, then it must belong
to the topological support of every invariant measure of the semigroup:

Lemma 2.2 Let P be a Markov operator over a Polish space X and let x ∈ X be
accessible. Then, x ∈ supp µ for every invariant probability measure µ.

Proof. Let µ be invariant for the Markov operator P and define the resolvent kernel
R by X
R(y, A) = 2−n P n (y, A) .
n>0

Clearly, the accessibility assumption implies that R(y, A) > 0 for every y ∈ X
and every neighborhood U ⊂ X of x. Then, the invariance of µ implies that
Z
µ(U ) = R(y, U ) µ(dy) > 0 ,
X

as required.

It is important to realise that this definition depends on the topology of X and

not just on the Borel σ-algebra. Considering again Example 1.11, we see that
the point 0 is reachable when [0, 1] is endowed with its usual topology, whereas
it is not reachable if we interpret the state space as {0} t (0, 1]. Therefore, as
in the previous section, this definition can be useful only in conjunction with an
appropriate regularity property of the Markov semigroup. The following example
shows that the Feller property is too weak to serve our purpose.
S OME SIMPLE UNIQUENESS CRITERIA 9

Example 2.3 (Ising model) The Ising model is one of the most popular toy models
of statistical mechanics. It is one of the simplest models describing the evolution
of a ferromagnet. The physical space is modelled by a lattice Zd and the mag-
netisation at each lattice site is modelled by a ‘spin’, an element of {±1}. The
2
state space of the system is therefore given by X = {±1}Z , which we endow with
the product topology. This topology can be metrized for example by the distance
function
X |xk − yk |
d(x, y) = ,
2
2|k|
k∈Z

and the space X endowed with this distance function is easily seen to be separable.
The (Glauber) dynamic for the Ising model depends on a parameter β and can
be described in the following way. At each lattice site, we consider independent
clocks that ring at Poisson distributed times. Whenever the Pclock at a given site
(say the site k) rings, we consider the quantity δEk (x) = j∼k xj xk , where the
sum runs over all sites j that are nearest neighbors of k. We then flip the spin at
site k with probability min{1, exp(−β δEk (x))}.
Let us first show that every point is accessible for this dynamic. Fix an arbitrary
configuration x ∈ X and a neighbourhood U containing x. By the definition of the
product topology, U contains an ‘elementary’ neighbourhood UN (x) of the type
UN (x) = {y ∈ X | yk = xk ∀ |k| ≤ N }. Given now an arbitrary initial condition
y ∈ X , we can find a sequence of m spin flips at distinct locations k1 , . . . , km , all
of them located inside the ball {|k| ≤ N }, that allows to go from y into UN (x).
Fix now t > 0. There is a very small but nevertheless strictly positive probability
that within that time interval, the Poisson clocks located at k1 , . . . , km ring exactly
once and exactly in that order, whereas all the other clocks located in the ball
{|k| ≤ N + 2} do not ring. Furthermore, there is a strictly positive probability
that all the corresponding spin flips do actually happen. As a consequence, the
Ising model is topologically irreducible in the sense that for any state x ∈ X , any
open set U ⊂ X and any t > 0, one has Pt (x, U ) > 0.
It is also relatively straightforward to show that the dynamic has the Feller
property, but this is outside the scope of these notes. However, despite the fact that
the dynamic is both Feller and topologically irreducible, one has the following:

Theorem 2.4 For d ≥ 2 there exists βc > 0 such that the Ising model has at least
two distinct invariant measures for β > βc .
The proof of this theorem is not simple and we will not give it here. It was a
√ tour de force by Onsager to be able to compute the critical value βc =
celebrated
ln(1 + 2)/2 explicitly in [Ons44] for the case d = 2. We refer to the monograph
[Geo88] for a more detailed discussion of this and related models.
This example shows that if we wish to base a uniqueness argument on the
accessibility of a point or on the topological irreduciblity of a system, we need
to combine this with a stronger regularity property than the Feller property. One
S OME SIMPLE UNIQUENESS CRITERIA 10

possible regularity property that yields the required properties is the strong Feller
property:

Definition 2.5 A Markov operator P over a Polish space X has the strong Feller
property if, for every function ϕ ∈ Bb (X ), one has Pϕ ∈ Cb (X ).
With this definition, one has:

Proposition 2.6 If a Markov operator P over a Polish space X has the strong
Feller property, then the topological supports of any two mutually singular invari-
ant measures are disjoint.

Proof. Let µ and ν be two mutually singular invariant measures for P, so that there
exists a set A ⊂ X such that µ(A) = 1 and ν(A) = 0. The invariance of µ and ν
then implies that P(x, A) = 1 for µ-almost every x and P(x, A) = 0 for ν-almost
every x.
Set ϕ = P1A , where 1A is the characteristic function of A. It follows from
the previous remarks that ϕ(x) = 1 µ-almost everywhere and ϕ(x) = 0 ν-almost
everywhere. Since ϕ is continuous by the strong Feller property, the claim now
follows from the fact that if a continuous function is constant µ-almost everywhere,
it must be constant on the topological support of µ.

Corollary 2.7 Let P be a strong Feller Markov operator over a Polish space X . If
there exists an accessible point x ∈ X for P, then it can have at most one invariant
measure.

Proof. Combine Proposition 2.6 with Lemma 2.2 and the fact that if I contains
more than one element, then by Theorem 1.7 there must be at least two distinct
ergodic invariant measures for P.

2.1 Continuous time Markov processes

A continuous time Markov process is no longer described by a single Markov tran-
sition kernel P, but by a family of transition kernels Pt satisfying the semigroup
property Ps+t = Ps Pt and such that P0 is the identity: P0 (x, ·) = δx . Without
further restrictions, a continuous-time Markov process could have very pathologi-
cal properties. We will therefore always assume that t 7→ Pt (x, A) is measurable
for every x and every measurable set A and that, for every initial condition x ∈ X ,
the process admits a version that is càdlàg (right-continuous with left limits) as a
function of time.
In other words, we will assume that for every x ∈ X , there exists a probability
measure Px on D(R+ , X ) such that its marginals on X n at any finite number of
times t1 < . . . < tn are given by the probability measure

Pt1 (x, dx1 )Pt2 −t1 (x1 , dx2 ) · · · Ptn −tn −1 (xn−1 , dxn ) .
H ARRIS ’ S THEOREM 11

In the case of continuous time, we say that a positive measure µ is invariant if

Pt µ = µ for every t ≥ 0. Note that in theory, it is always possible to restrict
oneself to the case of discrete time in the study of the existence and uniqueness of
an invariant measure:

Proposition 2.8 Let Pt be a Markov semigroup over X and let P = PT for some
fixed T > 0. Then, if µ is invariant for P, the measure µ̂ defined by
Z T
1
µ̂(A) = Pt µ(A) dt
T 0

is invariant for the semigroup Pt .

Remark 2.9 The converse is not true at this level of generality. This can be seen
for example by taking Pt (x, ·) = δx+t with X = S 1 .
In the case of continuous-time Markov processes, it is however often conve-
nient to formulate Lyapunov-Foster type conditions in terms of the generator L
of the process. Formally, one has L = ∂t Pt |t=0 , but it turns out that the natural
domain of the generator with this definition may be too restrictive for our usage.
We therefore take a rather pragmatic view of the definition of the generator L of a
Markov process, in the sense that writing

LF = G ,

is considered
R t to be merely a shorthand notation for the statement that the process
F (xt , t) − 0 G(xs , s) ds is a martingale for every initial condition x0 . Similarly,

LF ≤ G ,
Rt
is a shorthand notation for the statement that F (xt , t) − 0 G(xs , s) ds is a super-
martingale for every x0 .

Remark 2.10 It is possible to have LF ≤ G even in situations where there does

not exist any function H such that LF = H. Think of the case F (x) = −|x|
when the process xt is a Brownian motion. There, one has LF ≤ 0, but F does
not belong to the domain of the generator, even in the weakened sense described
above.

3 Harris’s theorem
The purpose of this section is to show that under a geometric drift condition and
provided that P admits sufficiently large ‘small sets’, its transition probabilities
converge towards a unique invariant measure at an exponential rate. This theorem
dates back to [Har56] and extends the ideas of Doeblin to the unbounded state space
setting. This is usually established by finding a Lyapunov function with “small”
H ARRIS ’ S THEOREM 12

level sets [Has80, MT93]. If the Lyapunov function is strong enough, one then has
a spectral gap in a weighted supremum norm [MT92, MT93].
Traditional proofs of this result rely on the decomposition of the Markov chain
into excursions away from the small set and a careful analysis of the exponential tail
of the length of these excursions [Num84, Cha89, MT92, MT93]. There have been
other variations which have made use of Poisson equations or worked at getting
explicit constants [KM05, DMR04, DMLM03, Bax05]. The proof given in the
present notes is a slight simplification of the proof given in [HM10]. It is very
direct, and relies instead on introducing a family of equivalent weighted norms
indexed by a parameter β and to make an appropriate choice of this parameter
that allows to combine in a very elementary way the two ingredients (existence
of a Lyapunov function and irreducibility) that are crucial in obtaining a spectral
gap. The original motivation was to provide a proof which could be extended to
more general settings in which no “minorisation condition” holds. This has been
applied successfully to the study of the two-dimensional stochastic Navier-Stokes
equations in [HM08].
We will assume throughout this section that P satisfies the following geometric
drift condition:

Assumption 3.1 There exists a function V : X → [0, ∞) and constants K ≥ 0

and γ ∈ (0, 1) such that

(PV )(x) ≤ γV (x) + K , (3.1)

for all x ∈ X .

Remark 3.2 One could allow V to also take the value +∞. However, since we do
not assume any particular structure on X , this case can immediately be reduced to
the present case by simply replacing X by {x : V (x) < ∞}.

Exercise 3.3 Show that in the case of continuous time, a sufficient condition for
Assumption 3.1 to hold is that there exists a measurable function V : X → [0, ∞)
and positive constants c, K such that LV ≤ K − cV .
Assumption 3.1 ensures that the dynamic enters the “centre” of the state space
regularly with tight control on the length of the excursions from the centre. We now
assume that a sufficiently large level set of V is sufficiently “nice” in the sense that
we have a uniform “minorisation” condition reminiscent of Doeblin’s condition,
but localised to the sublevel sets of V .

Assumption 3.4 For every R > 0, there exists a constant α > 0 so that

kP(x, ·) − P(y, ·)kTV ≤ 2(1 − α) , (3.2)

for all x, y such that V (x) + V (y) ≤ R.

H ARRIS ’ S THEOREM 13

Remark 3.5 An alternative way of formulating (3.2) is to say that the bound

|Pϕ(x) − Pϕ(y)| ≤ 2(1 − α) ,

holds uniformly over all functions ϕ with absolute value bounded by 1.

In order to state the version of Harris’ theorem under consideration, we intro-
duce the following weighted supremum norm:

|ϕ(x)|
kϕk = sup . (3.3)
x 1 + V (x)

With this notation at hand, one has:

Theorem 3.6 If Assumptions 3.1 and 3.4 hold, then P admits a unique invariant
measure µ? . Furthermore, there exist constants C > 0 and % ∈ (0, 1) such that the
bound
kP n ϕ − µ? (ϕ)k ≤ C%n kϕ − µ? (ϕ)k
holds for every measurable function ϕ : X → R such that kϕk < ∞.
While this result is well-known, the proofs found in the literature are often quite
involved and rely on careful estimates of the return times to small sets, combined
with a clever application of Kendall’s lemma. See for example [MT93, Section 15].
The aim of this note is to provide a very short and elementary proof of Theo-
rem 3.6 based on a simple trick. Instead of working directly with (3.3), we define a
whole family of weighted supremum norms depending on a scale parameter β > 0
that are all equivalent to the original norm (3.3):

|ϕ(x)|
kϕkβ = sup .
x 1 + βV (x)

The advantage of this scale of norms is that it allows us to prove that:

Theorem 3.7 If Assumptions 3.1 and 3.4 hold, then there exist constants β > 0
and %̄ ∈ (0, 1) such that the bound

kP n ϕ − µ? (ϕ)kβ ≤ %̄n kϕ − µ? (ϕ)kβ ,

holds for every measurable function ϕ : X → R such that kϕkβ < ∞.

The interest of this result lies in the fact that it is possible to tune β in such a
way that P is a strict contraction for the norm k · kβ . In general, this does not imply
that P is a contraction for k · k, even though the equivalence of these two norms
does of course imply that there exists n > 0 such that P n is such a contraction.
H ARRIS ’ S THEOREM 14

3.1 Formulation as a Lipschitz norm

We now introduce an alternative definition of the norm k · kβ . We begin by defining
a metric dβ between points in X by
(
0 x=y,
dβ (x, y) =
2 + βV (x) + βV (y) x 6= y .

Though sightly odd looking, the reader can readily verify that since V ≥ 0 by
assumption, dβ indeed satisfies the axioms of a metric. This metric in turn induces
a Lipschitz seminorm on measurable functions by
|ϕ(x) − ϕ(y)|
|||ϕ|||β = sup .
x6=y dβ (x, y)

It turns out that this norm is almost identical to the one from the previous section.
More precisely, one has:

Lemma 3.8 One has the identity |||ϕ|||β = infc∈R kϕ + ckβ .

Proof. It is obvious that |||ϕ|||β ≤ kϕkβ and therefore |||ϕ|||β ≤ infc∈R kϕ + ckβ , so
it remains to show the reverse inequality.
Given any ϕ with |||ϕ|||β ≤ 1, we set c = infx (1 + βV (x) − ϕ(x)). Observe that
for any x and y, ϕ(x) ≤ |ϕ(y)| + |ϕ(x) − ϕ(y)| ≤ |ϕ(y)| + 2 + βV (x) + βV (y).
Hence 1 + βV (x) − ϕ(x) ≥ −1 − βV (y) − |ϕ(y)|. Since there exists at least one
point with V (y) < ∞ we see that c is bounded from below and hence |c| < ∞.
Observe now that

ϕ(x) + c ≤ ϕ(x) + 1 + βV (x) − ϕ(x) = 1 + βV (x) ,

and

ϕ(x) + c = inf ϕ(x) + 1 + βV (y) − ϕ(y)

y
≥ inf 1 + βV (y) − |||ϕ|||β · dβ (x, y) ≥ −(1 + βV (x)) ,
y

so that |ϕ(x) + c| ≤ 1 + βV (x) as required.

3.2 Proof of main theorem

Theorem 3.9 If Assumptions 3.1 and 3.4 hold there exists an ᾱ ∈ (0, 1) and β > 0
such that
|||Pϕ|||β ≤ ᾱ|||ϕ|||β .

Proof. Fix a test function ϕ with |||ϕ|||β ≤ 1. By Lemma 3.8, we can assume
without loss of generality that one also has kϕkβ ≤ 1. The claim then follows if
we can exhibit ᾱ < 1 so that

|Pϕ(x) − Pϕ(y)| ≤ ᾱdβ (x, y) .

H ARRIS ’ S THEOREM 15

If x = y, the claim is true. Henceforth we assume x 6= y. We begin by

assuming that x and y are such that

V (x) + V (y) ≥ R . (3.4)

Choosing any γ̄ ∈ (γ, 1), we then have from (3.1) and (3.3) the bound

|Pϕ(x) − Pϕ(y)| ≤ 2 + βPV (x) + βPV (y)

≤ 2 + βγV (x) + βγV (y) + 2βK
≤ 2 + βγ̄V (x) + βγ̄V (y) + β(2K − (γ̄ − γ)R) .

It follows that if we ensure that R is sufficiently large so that (γ̄ − γ)R > 2K, then
there exists some α1 < 1 (depending on β) such that we do indeed have

|Pϕ(x) − Pϕ(y)| ≤ α1 dβ (x, y) .

We emphasise that up to now β could be any positive number; only the precise
value of α1 depends on it (and gets “worse” for small values of β). The second
part of the proof will determine a choice of β > 0.
Now consider the case of x and y such that V (x) + V (y) ≤ R. To treat this
case, we split the function ϕ as

ϕ(x) = ϕ1 (x) + ϕ2 (x) ,

where
|ϕ1 (x)| ≤ 1 , |ϕ2 (x)| ≤ βV (x) , ∀x ∈ X .
With this decomposition, we then obtain the bound

|Pϕ(x) − Pϕ(y)| ≤ |Pϕ1 (x) − Pϕ1 (y)| + |Pϕ2 (x)| + |Pϕ2 (y)|
≤ 2(1 − α) + γβV (x) + γβV (y) + 2βK
≤ 2 − 2α + β(γR + 2K)

Hence fixing for example β = α/(γR + 2K), we obtain the bound

|Pϕ(x) − Pϕ(y)| ≤ 2 − α ≤ (1 − α/2)dβ (x, y) ,

simply because dβ (x, y) ≥ 2. Setting ᾱ = max{1 − α/2, α1 } concludes the proof.

Remark 3.10 Actually, setting γ0 = γ + 2K/R < 1, then for any α0 ∈ (0, α) one
can choose β = α0 /K and ᾱ = (1 − α + α0 ) ∨ (2 + Rβγ0 )/(2 + Rβ).
S UBGEOMETRIC RATES OF CONVERGENCE 16

4 Subgeometric rates of convergence

Let x be a strong Markov process on a metric space X with generator L and as-
sociated semigroup Pt . We assume that x has a càdlàg modification and that Pt
is Feller for every t > 0, so that in particular the hitting times of closed sets are
stopping times. The aim of this note is to obtain a short self-contained result on the
convergence rate of such a process to its invariant measure (when one exists).
The main result of this section is the following:

Theorem 4.1 Assume that there exists a continuous function V : X → [1, ∞) with
precompact sublevel sets such that
LV ≤ K − ϕ(V ) , (4.1)
for some constant K and for some strictly concave function ϕ : R+ → R+ with
ϕ(0) = 0 and increasing to infinity. Assume furthermore that sublevel sets of V are
‘small’ in the sense that for every C > 0 there exists α > 0 and T > 0 such that
kPT (x, ·) − PT (y, ·)kTV ≤ 2(1 − α) ,
for every (x, y) such that V (x) + V (y) ≤ C. Then the following hold:

R a unique invariant measure µ for the Markov process x and µ is

1. There exists
such that ϕ(V (x)) µ(dx) ≤ K.
2. Let Hϕ be the function defined by
Z u
ds
Hϕ (u) = .
1 ϕ(s)
Then, there exists a constant C such that for every x, y ∈ X, one has the
bounds
Z ∞
(ϕ ◦ Hϕ−1 )(t)kPt (x, ·) − Pt (y, ·)kTV dt ≤ C(V (x) + V (y))
0
V (x) + V (y)
kPt (x, ·) − Pt (y, ·)kTV ≤ C .
Hϕ−1 (t)

3. There exists a constant C such that the bound

CV (x) C
kPt (x, ·) − µkTV ≤ −1 + ,
Hϕ (t) (ϕ ◦ Hϕ−1 )(t)
holds for every initial condition x ∈ X.

Remark 4.2 Since V is bounded from below by assumption, (4.1) follows imme-
diately if one can check that the process
Z t
t 7→ V (xt ) − Kt + (ϕ ◦ V )(xs ) ds
0
is a local supermartingale.
S UBGEOMETRIC RATES OF CONVERGENCE 17

The remainder of this section is devoted to the proof of Theorem 4.1. In a

nutshell, the proof relies on the fact that if xt and yt are two processes that are both
Markov with transition semigroup Pt , then one has the coupling inequality

kPt (x0 , ·) − Pt (y0 , ·)kTV ≤ 2P(xt 6= yt ) .

Of course, this bound will never converge to 0 if the two processes are independent,
so the aim of the game is to introduce correlations in a suitable way. This will be
done in Section 4.2, after some preliminary calculations that provide the main tools
used in this section.
To the best of my knowledge, Theorem 4.1 was stated in the form given in these
notes for the first time in the recent work [BCG08, DFG09]. However, it relies
on a wealth of previous work, for example [DFMS04] for the same condition in
the discrete-time case, early results by Nummelin, Tuominen and Tweedie [NT83,
TT94], etc. See also [Ver99, RT99] for examples of early results on subgeometric
convergence. The proof given in these notes is a simplification of the arguments
given in [BCG08, DFG09] and is more self-contained as it avoids reducing oneself
to the discrete-time case.

4.1 Preliminary calculations

4.1.1 Some renewal business
We start with the following result:

Lemma 4.3 Let H : R+ → R+ be an increasing strictly log-concave function

such that H(0) = 1. Let Fn be an increasing sequence of σ-algebras over some
probability space and let {Xn }n≥0 be a sequence of positive Fn -measurable ran-
dom variables such that there exists CX > 0 with E(H(Xn ) | Fn−1 ) < CX almost
surely for every n. Finally, let N be a geometrically distributed random variable
such that {N P = n} is independent of Fn for every n ≥ 1. Then, the random
variable X = N n=1 Xn satisfies EH(X) < ∞.

Proof. The strict log-concavity of H implies that for every ε > 0 there exists
K > 0 such that
ε
H(u + v) ≤ H(u)H(v) , (4.2)
CX
Pu ≥ K and v ≥ K.
for every u, v such that
Denoting X (k) = kn=1 Xn , we then have
X Y Y
H(X (k) ) = H(X (k) ) 1{Xn ≤K} 1{Xm >K}
A⊂{1,...,k} n∈A m6∈A
X  X  Y
≤ H |A|K + Xm 1{Xm >K}
A⊂{1,...,k} m6∈A m6∈A
X c|
Y H(Xm )
≤ H(|A|K)ε|A ,
CX
A⊂{1,...,k} m6∈A
S UBGEOMETRIC RATES OF CONVERGENCE 18

where we used (4.2) for the last inequality. It follows that

X X
EH(X (k) ) ≤ H(|A|K)εk−|A| ≤ H(kK) εk−|A| .
A⊂{1,...,k} A⊂{1,...,k}

Note now that it follows from Stirling’s formula for the binomial coefficients that
there exists a constant C such that

√ Xk
X
k−|A| kk
ε ≤C k εn .
nn (k − n)k−n
A⊂{1,...,k} n=0

Setting n = kx for some x ∈ [0, 1], we deduce that

 εx k
EH(X (k) ) ≤ Ck 3/2 H(kK) sup x 1−x
(4.3)
x∈[0,1] x (1 − x)

Since limx→0 xx = 1 and the function x 7→ xx is continuous for x > 0, it is

possible for every δ > 0 to find ε > 0 (and therefore K > 0) such that the
supremum appearing in this expression is bounded by 1 + δ.
On the other hand, since N is geometrically distributed, there exists λ > 1 such
that
P(N = k) = (λ − 1)λ−k , k≥1.
Combining this with (4.3), we deduce that
X X
EH(X) = (λ − 1) λ−k EH(X (k) ) ≤ C λ−k H(kK)(1 + δ)k ,
k≥1 k≥1

where we made use of the fact that {N = k} is independent of Fk to get the first
identity. We can first make K large enough so that (1 + δ) < λ1/3 . Then, we note
that the strict subexponential growth of H implies that there exists a constant C
such that H(kK) ≤ Cλk/3 and the claim follows.

4.1.2 Concave functions of semimartingales

We first make the following remark:

Proposition 4.4 Let y be a real-valued càdlàg semimartingale and let Φ : R+ ×

R → R be a function that is C 1 in its firstRargument, and C 2 and
R tconcave in its sec-
t 0
ond argument. Then, the process Φ(yt )− 0 Φ (ys− , s) dy(s)− 0 ∂t Φ(ys− , s) dy(s)
is decreasing.

Proof. Since y is a semimartingale, we can decompose it as yt = At + Mt with M

a martingale and A a process of finite variation. It follows from Itô’s formula that
Z t Z t
Φ(yt ) = Φ(y0 ) + Φ0 (ys− , s) dy(s) + ∂t Φ(ys− , s) ds
0 0
S UBGEOMETRIC RATES OF CONVERGENCE 19

Z t X
+ Φ00 (ys , s) dhM ict + (Φ(ys+ , s) − Φ(ys− , s) − Φ0 (ys− , s)∆ys ) ,
0 s∈[0,t]

where we denote by hM ict the quadratic variation of the continuous part of M and
we denote by ∆ys the size of the jump of y at time s. The claim then follows from
the facts that hM ict is an increasing process and that Φ00 is negative.

As a consequence, we have that:

Corollary 4.5 Let xt be a continuous-time Markov process with generator L and

let F and G be such that LF ≤ G. Then, if Φ is as in Proposition 4.4 with the
further property that Φ0 ≥ 0, we have LΦ(F ) ≤ ∂t Φ + Φ0 (F )G.

Proof. Setting yt = F (xt , t), it follows from our assumptions that one can write
dyt = G(xt , t) dt + dNt + dMt , where M is a càdlàg martingale and N is a non-
increasing process. It follows from Proposition 4.4 that

dΦ(yt ) ≤ Φ0 (yt− , t)(G(xt , t) dt + dNt + dMt ) + ∂t Φ(yt− , t) dt ,

so that the claim now follows from the fact that dNt is a negative measure and Φ0
is positive.

In order to obtain bounds on the dynamic of x, starting from (4.1), it seems

natural to consider the solution Φ to the ordinary differential equation

∂t Φ = −ϕ ◦ Φ , Φ(u, 0) = u .

This can be solved explicitly, yielding

Φ(u, t) = Hϕ−1 (Hϕ (u) − t) ,

where Hϕ is as in the statement of Theorem 4.1. At first glance, one would

‘naı̈vely’ expect from (4.1) to have a bound of the type EV (xt ) ≤ Φ(V (x0 ), t).
Unfortunately, it is easy to be convinced that the application of Jensen’s inequality
required to obtain such a bound does precisely go in the ‘wrong direction’.
However, it turns out that one can obtain a very similar bound, namely (in the
case K = 0) it is possible to obtain an inequality of the type EΦ−1 (V (xt ), t) ≤
V (x0 )! To see this, note first that one has the identity

Φ−1 (x, t) = Hϕ−1 (Hϕ (u) + t) . (4.4)

Combining this with the fact that Hϕ0 = 1/ϕ, it immediately follows that one has
the identity
∂t Φ−1 (x, t) ϕ(Φ−1 (x, t))
∂x Φ−1 (x, t) = = .
ϕ(x) ϕ(x)
S UBGEOMETRIC RATES OF CONVERGENCE 20

Since ϕ is concave and increasing by assumption, this shows that Φ−1 is increasing
and concave in its first argument for every fixed value of t ≥ 0. It then follows from
Corollary 4.5 that, provided that V satisfies (4.1), one has the bound

LΦ−1 (V (x), t) ≤ ∂t Φ−1 (V (x), t) + ∂x Φ−1 (V (x), t)LV

≤ ∂t Φ−1 (V (x), t) + ∂x Φ−1 (V (x), t)(K − ϕ(V (x)))
= K∂x Φ−1 (V (x), t) . (4.5)

If K = 0, the claimed bound then follows at once.

4.2 The coupling argument

We now turn to the proof of Theorem 4.1. The first point, namely the existence of
an invariant measure follows immediately from the Krylov-Bogoliubov criterion,
Theorem 1.10. Indeed, start with any initial condition x0 such that V (x0 ) < ∞
and define µT as in (1.2) (with the sum replaced by an integral). One then has
Z
1
(EV (xt ) − V (x0 )) ≤ K − ϕ(V (x)) µT (dx) .
T X

The tightness of the measures µT then follows immediately from the compactness
of the sublevel sets of V . The required integrability also follows at one from Fa-
tou’s lemma.
Take two independent copies of the process x and define

W (x, y) = 2V (x) + 2V (y) .

We then have

L2 W (x, y) ≤ 4K − 2ϕ(V (x)) − 2ϕ(V (y)) ≤ 4K − 2ϕ(V (x) + V (y)) .

Since we assumed that ϕ is strictly concave, it follows that for every C > 0 there
exists VC > 0 such that ϕ(2x) ≤ 2ϕ(x) − C, provided that x > VC . It follows that
there exists some VC such that

L2 W (x, y) ≤ −ϕ(W (x, y)) ,

provided that V (x) + V (y) > VC , and L2 W (x, y) ≤ 4K otherwise. We denote

from now on by K the set {(x, y) : V (x) + V (y) ≤ VC } and by τ the first return
time to this set.
Performing exactly the same calculation as in (4.5), but this time with the func-
tion Φ−1 (W (x, y), t), it then follows immediately that

E(x,y) Hϕ−1 (τ ) ≤ W (x, y) . (4.6)

We can now construct a coupling between two copies of the process in the follow-
ing way:
S UBGEOMETRIC RATES OF CONVERGENCE 21

1. If the two copies are equal at some time, they remain so for future times.

2. If the two copies are different and outside of K, then they evolve indepen-
dently.

3. If the two copies are different and inside of K, then they try to couple over
the next unit time interval.

We denote by τC the coupling time for the two processes. By assumption, the
maximal coupling has probability at least δ to succeed at every occurrence of step 3.
As a consequence, we can construct the coupling in such a way that this probability
is exactly equal to δ.
Note now that one has the coupling inequality

kPt (x, ·) − Pt (y, ·)kTV ≤ P(τC ≥ t) , (4.7)

which is the main tool to obtain bounds on the convergence speed. We denote by
τn the n-th return time to K and by zt the pair of processes (xt , yt ). We also denote
by Fn the σ-algebra generated by τn . Note that it then follows from (4.1) that

E(W (zτn +1 ) | Fn ) ≤ 2VC + 4K ,

so that there exists a constant C such that

E(F (τn+1 − τn ) | Fn ) ≤ C ,

almost surely for every n ≥ 1.

It then follows immediately from Lemma 4.3 and from (4.6) that

EF (τC ) ≤ EF (τ1 ) E(F (τC − τ1 ) | F1 ) ≤ CW (x, y) .

Combining this with (4.7), it follows that

Z ∞
F 0 (t) kPt (x, ·) − Pt (y, ·)kTV dt < C(V (x) + V (y)) ,
0

for some constant C. In particular, since the map t 7→ kPt (x, ·) − Pt (y, ·)kTV is
non-increasing, we have the pointwise bound

C(V (x) + V (y))

kPt (x, ·) − Pt (y, ·)kTV ≤ . (4.8)
F (t)
4.3 Convergence to the invariant measure
In this section, we bound the speed of convergence towards the invariant measure,
starting from an arbitrary initial condition x. Note that this is not necessarily the
same speed as the one exhibited in the previous section. Indeed, if we knew that V
S UBGEOMETRIC RATES OF CONVERGENCE 22

is integrable with respect to µ, then we would immediately obtain from (4.8) the
bound
Z
kPt (x, ·) − µkTV ≤ kPt (x, ·) − Pt (y, ·)kTV µ(dy)
Z
C  
≤ −1 V (x) + V (y) µ(dy) .
Hϕ (t)
Unfortunately, all that we can deduce from (4.1) is that ϕ(V ) is integrable with
respect to µ with Z
ϕ(V (x)) µ(dx) ≤ K .

This information can be used in the following way, by decomposing µ into the part
where V ≤ R and its complement:
Z
C  
kPt (x, ·) − µkTV ≤ −1 V (x) + V (y) µ(dy) + µ(V > R) .
Hϕ (t) V ≤R

This bound holds for every value of R, so we can try to optimise over it. Since
we assumed that ϕ is concave, the function x 7→ ϕ(x)/x is decreasing so that on
the set {V ≤ R} one has V (y) ≤ ϕ(V (y))R/ϕ(R). Furthermore, Chebychev’s
inequality yields
K
µ(V > R) = µ(ϕ(V ) > ϕ(R)) ≤ .
ϕ(R)
Combining this with the previous bound, we obtain
C  KR  K
kPt (x, ·) − µkTV ≤ −1 V (x) + + .
Hϕ (t) ϕ(R) ϕ(R)

Setting R = Hϕ−1 (t), we finally obtain for some constant C the bound

CV (x) C
kPt (x, ·) − µkTV ≤ + .
Hϕ (t) (ϕ ◦ Hϕ−1 )(t)
−1

4.4 Convergence in stronger norms

Another standard set of results concerns the convergence towards the invariant
measure in weighted total variation norms. For any two positive measures µ and
ν on X that have densities Dµ and Dν with respect to some common reference
measure dx and any function G : X → [1, ∞), we set
Z
kµ − νkG = |Dµ (x) − Dν (x)| G(x) dx . (4.9)

The usual total variation distance corresponds to the choice G = 1. Note also that
(4.9) is independent of the choice of reference measure.
With this notation, one has the following result:
L OWER BOUNDS ON CONVERGENCE RATES 23

Lemma 4.6 Let Ψ : R+ → R+ be an increasing function such that u 7→ Ψ(u)/u

is increasing. Then, the bound
R
X (Ψ ◦ G)(x)(µ + ν)(dx)

−1
kµ − νkG ≤ kµ − νkTV Ψ ,
kµ − νkTV
holds for any pair of positive measures µ and ν such that the right hand side is
finite.

Proof. From the definition of k · kG , we have

Z
kµ − νkG ≤ Rkµ − νkTV + G(x)(µ + ν)(dx) .
G(x)>R

Since Ψ(u)/u is increasing by assumption, the bound u ≤ Ψ(u)R/Ψ(R) holds for

u ≥ R. We deduce that the bound
Z
R
kµ − νkG ≤ Rkµ − νkTV + (Ψ ◦ G)(x)(µ + ν)(dx) ,
X Ψ(R)
holds for every R > 0. Optimising over R concludes the proof.

5 Lower bounds on convergence rates

The main idea for obtaining lower bounds on the convergence rate towards the
invariant
R measure is the following. Suppose that on the one hand, we know that
F dµ = ∞ for some positive function F , so that we have some lower bound
on the ‘heaviness’ of the tails of the invariant measure µ. Suppose on the other
hand that we can find a function G  F such that we have an upper bound on
the growth of the expected value of G under the transition probabilities starting
from some fixed starting point. Then, this can be turned into a lower bound on the
convergence rate towards µ by arguing that if the convergence was too fast, then
the non-integrability of F under µ would imply some lower bound on the expected
value of G. This lower bound must be compatible with the existing upper bound.

Theorem 5.1 Let xt be a Markov process on a Polish space X with invariant mea-
sure µ? and let G : X → [1, ∞) be such that:
• There exists a function f : [1, ∞) → [0, 1] such that the function Id · f : y 7→
yf (y) is increasing to infinity and such that µ? (G ≥ y) ≥ f (y) for every
y ≥ 1.
• There exists a function g : X × R+ → [1, ∞) increasing in its second argu-
ment and such that E(G(xt ) | X0 = x0 ) ≤ g(x0 , t).
Then, the bound
1
kµt − µ? kTV ≥ f ((Id · f )−1 (2g(x0 , t))) , (5.1)
2
holds for every t > 0, where µt is the law of xt with initial condition x0 ∈ X .
M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 24

Proof. It follows from the definition of the total variation distance and from Cheby-
shev’s inequality that, for every t ≥ 0 and every y ≥ 1, one has the lower bound
g(x0 , t)
kµt − µ? kTV ≥ µ? (G(x) ≥ y) − µt (G(x) ≥ y) ≥ f (y) − .
y
Choosing y to be the unique solution to the equation yf (y) = 2g(x0 , t), the result
follows.

The problem is that unless µ? is known exactly, one does not in general have
sufficiently good information on the tail behaviour of µ? to be able to apply The-
orem 5.1 as it stands. However, it follows immediately from the proof that the
bound (5.1) still holds for a subsequence of times tn converging to ∞, provided
that the bound µ? (G ≥ yn ) ≥ f (yn ) holds for a sequence yn converging to infin-
ity. This observation allows to obtain the following corollary that is more useful in
situations where the law of the invariant measure is not known explicitly:

Corollary 5.2 Let xt be a Markov process on a Polish

R space X with invariant
measure µ? and let W : X → [1, ∞) be such that W (x) µ? (dx) = ∞. Assume
that there exist F : [1, ∞) → R and h : [1, ∞) → R such that:
R∞
• h is decreasing and 1 h(s) ds < ∞.
• F · h is increasing and lims→∞ F (s)h(s) = ∞.
• There exists a function g : X × R+ → R+ increasing in its second argument
and such that E((F ◦ W )(xt ) | x0 ) ≤ g(x0 , t).
Then, for every x0 ∈ X , there exists a sequence of times tn increasing to infinity
such that the bound

kµtn − µ? kTV ≥ h((F · h)−1 (g(x0 , tn )))

holds, where µt is the law of xt with initial condition x0 ∈ X .

R
Proof. Since W (x) µ? (dx) = ∞, there exists a sequence wn increasing to infin-
ity such that µ? (W (x) ≥ wn ) ≥ 2h(wn ), for otherwise we would have the bound
Z Z ∞ Z ∞
W (x) µ? (dx) = 1 + µ? (W (x) ≥ w) dw ≤ 1 + 2 h(w) dw < ∞ ,
1 1

thus leading to a contradiction. Applying Theorem 5.1 with G = F ◦ W and

f = 2h ◦ F −1 concludes the proof.

6 Malliavin calculus and Hörmander’s theorem

One of the main ingredients in the convergence results of the previous sections was
the existence of sufficiently large ‘small sets’ in which the transition probabilities
of our Markov process can be uniformly bounded from below. One situation where
M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 25

this is relatively easy to show is that of a process with transition probabilities that
have a continuous density with respect to some reference measure, i.e.

Pt (x, dy) = pt (x, y) µ(dy) ,

where pt is continuous in both arguments. In that case, every point x ∈ X admits a

neighbourhood Ux such that Ux is ‘small’. If furthermore there exists x? such that
x? ∈ supp Pt (x, ·) for every x ∈ X, then it is a straightforward exercise to check
that every compact subset of X is small.
The aim of this section is to provide a reasonably self-contained proof of Hör-
mander’s criterion which ensures that the transition probabilities of a diffusion with
smooth coefficients have a smooth density with respect to Lebesgue measure. Our
main object of study in this section is a stochastic differential equation of the form
m
X
dx = V0 (x) dt + Vi (x) ◦ dWi , (6.1)
i=1

where the Vi ’s are smooth vector fields on Rn . We assume that these vector fields
assume the coercivity assumptions necessary so that the solution to (6.1) is C ∞ with
respect to its initial condition. An import tool for our analysis will be the lineari-
sation of (6.1) with respect to its initial condition. This is given by the stochastic
process J0,t defined as the solution to the SDE
m
X
dJ0,t = DV0 (x) J0,t dt + DVi (x) J0,t ◦ dWi . (6.2)
i=1

(k)
Higher order derivatives J0,t with respect to the initial condition can be defined
similarly. Throughout this section, we will make the following standing assump-
tion:

Assumption 6.1 The vector fields Vi are C ∞ and all of their derivatives grow at
most polynomially at infinity. Furthermore, they are such that
(k) p −1 p
E sup |xt |p < ∞ , E sup |J0,t | <∞, E sup |J0,t | <∞,
t≤T t≤T t≤T

for every initial condition x0 ∈ Rn , every terminal time T > 0, and every p > 0.
−1
Note here that the inverse J0,t of the Jacobian can be found by solving the SDE
m
X
−1 −1 −1
dJ0,t = −J0,t DV0 (x) dt − J0,t DVi (x) ◦ dWi . (6.3)
i=1

The aim of this section is to show that under a certain non-degeneracy assump-
tion on the vector fields Vi , the law of the solution to (6.1) has a smooth density
with respect to Wiener measure. To describe this non-degeneracy condition, recall
M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 26

that the Lie bracket [U, V ] between two vector fields U and V is the vector field
defined by
[U, V ](x) = DV (x) U (x) − DU (x) V (x) .
This notation is consistent with the usual notation for the commutator between
two linear operators since, if we denote by AU the first-order differential operator
acting on smooth functions f by AU f (x) = hV (x), ∇f (x)i, then we have the
identity A[U,V ] = [AU , AV ].
With this notation at hand, we give the following definition:

Definition 6.2 Given an SDE (6.1), define a collection of vector fields Vk by

V0 = {Vi : i > 0} , Vk+1 = Vk ∪ {[U, Vj ] : U ∈ Vk & j ≥ 0} .
We also define the vector spaces Vk (x) = span{V (x) S : V ∈ Vk }. We
n
say that
(6.1) satisfies the parabolic Hörmander condition if k≥1 Vk (x) = R for every
x ∈ Rn .
With these notations, Hörmander’s theorem can be formulated as

Theorem 6.3 Consider (6.1) and assume that Assumption 6.1 holds. If the corre-
sponding vector fields satisfy the parabolic Hörmander condition, then its solutions
admit a smooth density with respect to Lebesgue measure.
A complete rigorous proof of Theorem 6.3 is far beyond the scope of these
notes. However, we hope to be able to give a convincing argument showing why
this result is true and what are the main steps involved in its probabilistic proof. The
interested reader can find the technical details required to make the proof rigorous
in [Mal78, KS84, KS85, KS87, Nor86, Nua95]. Hörmander’s original, completely
different, proof using fractional integrations can be found in [Hör67]. A yet com-
pletely different functional-analytic proof using the theory of pseudo-differential
operators was developed by Kohn in [Koh78] and can also be found in [Hör85].

6.1 Simplified Malliavin calculus

The main tool in the proof is the Malliavin calculus with its integration by part
formula in Wiener space, which was developed precisely in order to provide a
probabilistic proof of Theorem 6.3. It essentially relies on the fact that the image
of a Gaussian measure under a smooth submersion that is sufficiently integrable
possesses a smooth density with respect to Lebesgue measure. This can be shown
in the following way. First, one observes the following fact, which follows from
standard Sobolev embedding results:

Lemma 6.4 Let µ be a probability measure on Rn such that the bound

Z 
D(k) G(x) µ(dx) ≤ Ck kGk∞ ,
 

Rn
holds for every smooth bounded function G and every k ≥ 1. Then µ has a smooth
density with respect to Lebesgue measure.
M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 27

Consider now a sequence of N independent Gaussian random variables δwk

with variances δtk for k ∈ {1, . . . , N }, as well as a smooth map X : RN → Rn .
We also denote by w the collection {δwk }k≥1 and we define the n × n matrix-
valued map X
Mij (w) = ∂k Xi (w)∂k Xj (w) δtk ,
k
where we use ∂k as a shorthand for the partial derivative with respect to the variable
δwk . With this notation, X being a surjection is equivalent to M(w) being invertible
for every w. We then have the following result:

Theorem 6.5 Let X be as above, assume that M(w) is invertible for every w and
that, for every p > 1 and every m ≥ 0, we have
E|∂k1 · · · ∂km X(w)|p < ∞ , EkM(w)−1 kp < ∞ . (6.4)
Then the law of X(w) has a smooth density with respect to Lebesgue measure.
Before we turn to the proof of Theorem 6.5, note that if Fk and G are square
integrable functions with square integrable derivatives, then we have the integration
by parts formula
X X X
E ∂k G(w)Fk (w) δtk = EG(w) Fk (w) δwk − EG(w) ∂k Fk (w) δtk
k
Zk k
def
= EG(w) F dw . (6.5)
R
Here we defined the Skorokhod integral F dw by the expression on the first line.
This Skorokhod integral satisfies the following isometry:

Proposition 6.6 Let Fk be square integrable functions with square integrable deriva-
tives, then
Z 2 X X
E F dw = EFk2 (w) δtk + E∂k F` (w)∂` Fk (w) δtk δt`
k k,`
X X
≤ EFk2 (w) δtk + E(∂k F` (w))2 δtk δt` ,
k k,`

holds.

Proof. It follows from the definition that one has the identity
Z 2 X
E F dw = E(Fk F` δwk δw` + ∂k Fk ∂` F` δtk δt` − 2Fk ∂` F` δwk δt` ) .
k,`

Applying the identity EG δw` = E∂` G δt` the first term in the above formula (with
G = Fk F` δwk ), we thus obtain
X
... = E(Fk F` δk,` δt` + ∂k Fk ∂` F` δtk δt` + (F` ∂` Fk − Fk ∂` F` ) δwk δt` ) .
k,`
M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 28

Applying the same identity to the last term then finally leads to
X
... = E(Fk F` δk,` δt` + ∂k F` ∂` Fk δtk δt` ) ,
k,`

which is the desired result.

Proof of Theorem 6.5. We want to show that Lemma 6.4 can be applied. For η ∈
Rn , we then have from the definition of M the identity
X
(Dj G)(X(w)) = ∂k (G(X(w)))∂k Xm (w) δtk M−1
mj (w) . (6.6)
k,m

Combining this identity with (6.5), it follows that

XZ
EDj G(X) = EG(X(w)) ∂· Xm (w) M−1
mj (w) dw .
m

Combining this with Proposition 6.6 and (6.4) immediately shows that the re-
quested result holds for k = 1. Higher values of k can be treated similarly by
repeatedly applying (6.6).

6.2 Back to diffusions

The results in the previous section strongly suggest that one can define a “Malli-
avin derivative” operator D, acting on random variables and returning a stochastic
process, that has all the usual properties of a derivative and such that
Z T
Dt f (s) dW (s) = f (t) , t ∈ [0, T ] ,
0

for smooth (deterministic) functions f . If on the other hand f is a stochastic pro-

cess, then the following chain rule holds:
Z T Z T
Dt f (s) dW (s) = f (t) + Dt Dt f (s) dW (s) , t ∈ [0, T ] ,
0 0

Using this identity, it then follows from differentiating (6.1) on both sides that one
has for r ≤ t the identity
Z t m Z
X t
Djr X(t) = DV0 (Xs ) Djr Xs ds + DVi (Xs ) Djr Xs ◦ dWi (s) + Vj (Xr ) .
r i=1 r

We see that this equation is identical (save for the initial condition at time t = r
given by Vj (Xr )) to the equation giving the derivative of X with respect to its initial
condition! Denoting by Js,t the Jacobian for the stochastic flow between times s
and t, we therefore have for s < t the identity

Djs Xt = Js,t Vj (Xs ) . (6.7)

M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 29

(Since X is adapted, we have Djs Xt = 0 for s ≥ t.) Note that the Jacobian has the
composition property J0,t = Js,t J0,s , so that Js,t can be rewritten in terms of the
−1
‘usual’ Jacobian as Js,t = J0,t J0,s .
Rt
We now denote by A0,t the operator A0,t v = 0 Js,t V (Xs )v(s) ds, where v is
a square integrable, not necessarily adapted, Rm -valued stochastic process and V
is the n × m matrix-valued function obtained by concatenating the vector fields Vj
for j = 1, . . . , m. This allows us to define the Malliavin covariance matrix M0,t of
Xt in the same way as in the previous section by
Z t
M0,t = A0,t A∗0,t = Js,t V (Xs )V ∗ (Xs )Js,t
∗
ds .
0

Note that M0,t is a random positive definite n × n matrix. We then have:

Proposition 6.7 Consider a diffusion of the type (6.1) satisfying Assumption 6.1.
If M0,t is almost surely invertible and that EkM−1 p
0,t k < ∞ for every p > 0, then
the transition probabilities of (6.1) have a C ∞ density with respect to Lebesgue
measure for every t > 0.

Proof. This is essentially a version of Theorem 6.5. The technical details required
to make it rigorous are well beyond the scope of these notes and can be found for
example in [Nua95].

The remainder of this section is devoted to a proof of Hörmander’s theorem,

which gives a simple (but essentially sharp!) criterion for the invertibility of the
Malliavin matrix of a diffusion process. Actually, it turns out that for technical
reasons, it is advantageous to rewrite the Malliavin matrix as
Z t
∗ −1 −1 ∗
M0,t = J0,t C0,t J0,t , C0,t = J0,s V (Xs )V ∗ (Xs )(J0,s ) ds ,
0

where C0,t is the reduced Malliavin matrix of our diffusion process. Then since
we assumed that J0,t has inverse moments of all orders, the invertibility of M0,t is
equivalent to that of C0,t . Note first that since C0,t is a positive definite symmetric
matrix, the norm of its inverse is given by
 −1
kC−1
0,t k = inf hη, C0,t ηi .
|η|=1

A very useful observation is then the following:

Lemma 6.8 Let M be a symmetric positive semidefinite n × n matrix-valued ran-

dom variable such that EkM kp < ∞ for every p ≥ 1 and such that, for every
p ≥ 1 there exists Cp such that
sup P(hη, M ηi < ε) ≤ Cp εp , (6.8)
|η|=1

holds for every ε ≤ 1. Then, EkM −1 kp < ∞ for every p ≥ 1.

M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 30

Proof. The non-trivial part of the result is that the supremum over η is taken outside
of the probability in (6.8). For ε > 0, let {ηk }k≤N be a sequence of vectors with
|ηk | = 1 such that for every η with |η| = 1, there exists k such that |ηk −η| ≤ ε2 . It
is clear that one can find such a set with N ≤ Cε2−2n for some C > 0 independent
of ε. We then have the bound
hη, M ηi = hηk , M ηk i + hη − ηk , M ηi + hη − ηk , M ηk i
≥ hηk , M ηk i − 2kM kε2 ,
so that
     1
P inf hη, M ηi ≤ ε ≤ P inf hηk , M ηk i ≤ 4ε + P kM k ≥
|η|=1 k≤N ε
   1
≤ Cε2−2n sup P hη, M ηi ≤ 4ε + P kM k ≥ .
|η|=1 ε

It now suffices to use (6.8) for p large enough to bound the first term and Cheby-
chev’s inequality combined with the moment bound on kM k to bound the second
term.

As a consequence of this, Thoerem 6.3 is a corollary of:

Theorem 6.9 Consider (6.1) and assume that Assumption 6.1 holds. If the corre-
sponding vector fields satisfy the parabolic Hörmander condition then, for every
initial condition x ∈ Rn , we have the bound
sup P(hη, C0,1 ηi < ε) ≤ Cp εp ,
|η|=1

for suitable constants Cp and all p ≥ 1.

Remark 6.10 The choice t = 1 as the final time is of course completely arbitrary.
Here and in the sequel, we will always consider functions on the time interval
[0, 1].
Before we turn to the proof of this result, we introduce a useful notation. Given
a family A = {Aε }ε∈(0,1] of events depending on some parameter ε > 0, we say
that A is ‘almost true’ if, for every p > 0 there exists a constant Cp such that
P(Aε ) ≥ 1 − Cp εp for all ε ∈ (0, 1]. Similarly for ‘almost false’. Given two
such families of events A and B, we say that ‘A almost implies B’ and we write
A ⇒ε B if A \ B is almost false. It is straightforward to check that these notions
behave as expected (almost implication is transitive, finite unions of almost false
events are almost false, etc). Note also that these notions are unchanged under any
reparametrisation of the form ε 7→ εα for α > 0. Given two families X and Y of
real-valued random variables, we will similarly write X ≤ε Y as a shorthand for
the fact that {Xε ≤ Yε } is ‘almost true’.
Before we proceed, we state the following useful result, where k · k∞ denotes
the L∞ norm and k · kα denotes the best possible α-Hölder constant.
M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 31

Lemma 6.11 Let f : [0, 1] → R be continuously differentiable and let α ∈ (0, 1].
Then, the bound
− 1 1 o
n
k∂t f k∞ = kf k1 ≤ 4kf k∞ max 1, kf k∞1+α k∂t f kα1+α

holds, where kf kα denotes the best α-Hölder constant for f .

Proof. Denote by x0 a point such that |∂t f (x0 )| = k∂t f k∞ . It follows from the
definition of the α-Hölder constant k∂t f kC α that |∂t f (x)| ≥ 12 k∂t f k∞ for every
x such that |x − x0 | ≤ (k∂t f k∞ /2k∂t f kC α )1/α . The claim then follows from the
fact that if f is continuously differentiable and |∂t f (x)| ≥ A over an interval I,
then there exists a point x1 in the interval such that |f (x1 )| ≥ A|I|/2.

With these notations at hand, we have the following statement, which is es-
sentially a quantitative version of the Doob-Meyer decomposition theorem. Orig-
inally, it appeared in [Nor86], although some form of it was already present in
earlier works. The proof given here is a further simplification of the arguments in
[Nor86].

Lemma 6.12 (Norris) Let W be an m-dimensional Wiener process and let A and
B be R and Rm -valued adapted processes such that, for α = 13 , one has E(kAkα +
kBkα )p < ∞ for every p. Let Z be the process defined by
Z t Z t
Zt = Z0 + As ds + Bs dW (s) . (6.9)
0 0

Then, there exists a universal constant r ∈ (0, 1) such that one has

{kZk∞ < ε} ⇒ε {kAk∞ < εr } & {kBk∞ < εr } .

Proof. Recall the exponential martingale inequality [RY99, p. 153], stating that if
M is any continuous martingale with quadratic variation process hM i(t), then
 
P sup M (t) ≥ x & hM i(T ) ≤ y ≤ exp(−x2 /2y) ,
t≤T

for every positive T , x, y. With our notations, this immediately implies that for
any q < 21 , one has the almost implication

{khM ik∞ < ε} ⇒ε {kM k∞ < εq } . (6.10)

With this bound in mind, we apply Itô’s formula to Z 2 , so that

Z t Z t Z t
2 2
Zt = Z0 + 2 Zs As ds + 2 Zs Bs dW (s) + Bs2 ds . (6.11)
0 0 0
M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 32

Since kAk∞ ≤ε ε−1/2 (or any other negative exponent for that matter) by assump-
tion and similarly for B, it follows from this and (6.10) that
nZ 1  1
o nZ 1  1
o
{kZk∞ < ε} ⇒ε As Zs ds ≤ ε &  Bs Zs dW (s) ≤ ε 2 .
  2
 

0 0

Inserting these bounds back into (6.11) and applying Jensen’s inequality then yields
nZ 1 o nZ 1 o
2 1/2
{kZk∞ < ε} ⇒ε Bs ds ≤ ε ⇒ |Bs | ds ≤ ε1/4 .
0 0

We now use the fact that kBkα ≤ε ε−q for every q > 0 and we apply Lemma 6.11
with ∂t f (t) = |Bt | (we actually do it component by component), so that

{kZk∞ < ε} ⇒ε {kBk∞ ≤ ε1/17 } ,

say. In order to get the bound on A, note that we can again apply the exponential
martingale inequality to obtain that this ‘almost implies’ the martingale part in (6.9)
is ‘almost bounded’ in the supremum norm by ε1/18 , so that
n Z · o
1/18
{kZk∞ < ε} ⇒ε A ds ≤ ε .

s
0 ∞

Finally applying again Lemma 6.11 with ∂t f (t) = At , we obtain that

{kZk∞ < ε} ⇒ε {kAk∞ ≤ ε1/80 } ,

and the claim follows with r = 1/80.

Remark 6.13 By making α arbitrarily close to 1/2, keeping track of the different
norms appearing in the above argument, and then bootstrapping the argument, it is
possible to show that

{kZk∞ < ε} ⇒ε {kAk∞ ≤ εp } & {kBk∞ ≤ εq } ,

for p arbitrarily close to 1/5 and q arbitrarily close to 3/10. This is a slight im-
provement over the exponent 1/8 that was originally obtained in [Nor86]. (Note
however that the two results are not necessarily comparable since [Nor86] used L2
norms instead of L∞ norms.)
We now have all the necessary tools to prove Theorem 6.9:

Proof of Theorem 6.9. We fix some initial condition x ∈ Rn and some unit vector
η ∈ Rn . With the notation introduced earlier, our aim is then to show that

{hη, C0,1 ηi < ε} ⇒ε φ.

At this point, we introduce for an arbitrary vector field F on Rn the process ZF

defined by
−1
ZF (t) = hη, J0,t F (xt )i ,
M ALLIAVIN CALCULUS AND H ÖRMANDER ’ S THEOREM 33

so that
m Z
X 1 m Z
X 1 2
hη, C0,1 ηi = |ZVk (t)|2 dt ≥ |ZVk (t)| dt . (6.12)
k=1 0 k=1 0

The processes ZF have the nice property that they solve the stochastic differential
equation
m
X
dZF (t) = Z[F,V0 ] (t) dt + Z[F,Vk ] (t) ◦ dWk (t) .
i=1

This can be rewritten in Itô form as

m m
 X 1  X
dZF (t) = Z[F,V0 ] (t) + Z[[F,Vk ],Vk ] (t) dt + Z[F,Vk ] (t) dWk (t) . (6.13)
2
k=1 i=1

Since we assumed that all derivatives of the Vj grow at most polynomially, we

deduce from the Hölder regularity of Brownian motion that ZF does indeed satisfy
the assumptions on its Hölder norm required for the application of Norris’ lemma.
We deduce from it that one has the implication

{kZF k∞ < ε} ⇒ε {kZ[F,Vk ] k∞ < εr } & {kZG k∞ < εr } ,

for k = 1, . . . , m and for G = [F, V0 ] + 12 m

P
k=1 [[F, Vk ], Vk ]. Iterating this bound
a second time, we obtain that
2
{kZF k∞ < ε} ⇒ε {kZ[[F,Vk ],V` ] k∞ < εr } ,

so that we finally obtain the implication

2
{kZF k∞ < ε} ⇒ε {kZ[F,Vk ] k∞ < εr } , (6.14)

for k = 0, . . . , m.
At this stage, we are basically done. Indeed, combining (6.12) with Lemma 6.11
as above, we see that

{hη, C0,1 ηi < ε} ⇒ε {kZVk k∞ < ε1/5 } .

Applying (6.14) iteratively, we see that for every k > 0 there exists some qk > 0
such that \
{hη, C0,1 ηi < ε} ⇒ε {kZV k∞ < εqk } .
V ∈Vk

Since ZV (0) = hη, V (x0 )i and since there exists some k > 0 such that Vk (x0 ) =
Rn , the right hand side of this expression is empty for some sufficiently large value
of k, which is precisely the desired result.
E XAMPLES 34

7 Examples
In this section, we go through a number of examples where it is possible, using the
results presented in these notes, to give very sharp upper and lower bounds on the
rate of convergence to equilibrium.

7.1 Gradient dynamic

This is the simplest case of a reversible dynamic on Rn , namely the solution to the
SDE √
dx = −∇V (x) dt + 2dW (t) , (7.1)
where W is a d-dimensional Wiener process. In this case, the diffusion is elliptic,
so it is straightforward that the assumptions of Hörmander’s theorem are satisfied,
so that it has smooth transition probabilities, provided that V is smooth. We assume
that V (x) is behaves like |x|2k at ∞, in the sense that the bounds
c|x|2k ≤ V (x) ≤ C|x|2k , hx, ∇V (x)i ≥ c|x|2k , |D2 V (x)| ≤ C|x|2k−2 ,
hold for some positive constants c and C outside of some sufficiently large compact
set. We also assume that all derivatives of V grow at most polynomially fast.
It then follows from Itô’s formula that the generator of (7.1) is given by
L = ∆ − ∇V ∇ .
In particular, we have the identity
L exp(αV ) = α(∆V + (α − 1)|∇V |2 ) exp(αV ) .
It follows that if k ≥ 12 , then there exists α > 0 and C > 0 such that
L exp(αV ) ≤ −C exp(αV ) ,
outside of some compact set. In particular, the assumptions of Harris’s theorem
hold and we do obtain exponential converge towards the invariant measure.
What about k ∈ (0, 12 )? In this case, the conditions of Harris’s theorem do not
hold, but if we set W (x) = exp(αV (x)) for some α < 1, then we have
W
LW ≤ −C|x|4k−2 W ≤ −C 1 , (7.2)
(log W ) k −2
so that we can apply Theorem 4.1 with ϕ(x) = Cx/(log x)γ for γ = k1 − 2. In this
case, we have
Z u Z log u
γ dx (log u)γ+1
Hϕ (u) ∝ (log x) = y γ dy = ,
1 x 0 γ+1
so that Hϕ−1 (u) ∝ exp(cu1/(γ+1) ), for some c > 0. This immediately yields an
upper bound on the convergence towards the invariant measure of the type
k
kPt (x, ·) − µ? kTV ≤ C exp(αV (x) − ct 1−k ) . (7.3)
E XAMPLES 35

Let us now turn to the lower bounds. Consider the function

G(x) = exp(κV (x)) ,

for a constant κ to be determined later. It is then a straightforward calculation to

check, similarly to before, that there exists a constant C such that
G(x)
LG(x) ≤ C 1 .
(log G(x)) k −2

Note that the difference between this and (7.2) is that the right hand side is positive
rather than negative. It follows from Jensen’s inequality that the quantity g(t) =
EG(xt ) satisfies the differential inequality

dg(t) Cg(t)
≤ 1 ,
dt (log g(t)) k −2

so that there exist some constants c and C such that

k/(1−k)
EG(xt ) ≤ CG(x0 )ect .

On the other hand, since the invariant measure for our system is known to be given
by
µ? (dx) ∝ exp(−V (x)) dx ,
one can check that the bound

µ? (G > R) = µ? (κV (x) > log R) ≥ CR−γ ,

is valid for some C > 0 and some γ > 0 for sufficiently large values of R. The
value of γ can be made arbitrarily small by increasing the value of κ. This enables
k/(1−k)
us to apply Theorem 5.1 with f (R) = CR−γ and g(x, t) = G(x)ect , so that
we obtain a lower bound of the form
γκ k
kPt (x, ·) − µ? kTV ≥ C exp( V (x) − ct 1−k ) ,
1−γ
which essentially matches the upper bound previously obtained in (7.3).

7.2 Renewal processes

Let ν be a probability distribution on R+ \ {0} with finite expectation. Then the
renewal process with holding time distribution ν is a point process on R such the
times between successive events are independent and distributed with law ν. This
renewal process can be described by the Markov process xt on R+ that is equal
to the time left until the next event. The dynamic of xt is given by ẋt = −1 as
long as xt > 0. When it reaches zero, then it jumps instantaneously to a positive
E XAMPLES 36

value distributed according to ν. It is possible to convince oneself that the invariant

measure µ of x is given by
µ? (dx) = c x ν(dx) , (7.4)
for a suitable normalisation constant c > 0.
The generator of this process is then given by Lf (x) = −∂x f (x), but a differ-
entiable function f belongs to the domain of L only if
Z ∞
f (0) = f (x) ν(dx) . (7.5)
0
Suppose now that ν has a density p with respect to Lebesgue measure such that
c− c+
ζ
≤ p(x) ≤ ζ ,
x x
for some ζ > 2. How fast does such a process converge to its invariant measure? A
natural choice of a Lyapunov function is to take V such that V (x) = xα for x > 1
and to adjust it on [0, 1] in such a way that the compatibility condition (7.5) holds.
Note that this is possible if and only if one chooses α < ζ − 1.
In that case, one has
1
LV (x) ∼ −xα−1 = −V 1− α ,
for sufficiently large values of x, so that the conditions of Theorem 4.1 hold with
ϕ(x) = xγ and γ = 1 − α1 . In this case, Hϕ (u) ∼ u1/α , so that we have an upper
bound on the convergence towards the invariant measure given by
C|x|α
kPt (x, ·) − µ? kTV ≤ . (7.6)
tα−1
Since α can be taken arbitrarily close to ζ − 1, we have
log kPt (x, ·) − µ? kTV
lim sup ≤2−ζ . (7.7)
t→∞ log t
Lower bounds can be obtained in a similar way by using the fact that LV is
bounded from above, so that Exαt ≤ xα0 + Ct for some constant C. Furthermore,
since we know by (7.4) that the invariant measure µ? has a density behaving like
x1−ζ for large values of x, we have the bound
ζ−2
µ? (V > R) = µ? (x > R1/α ) ∼ R− α .
ζ−2
This enables us again to apply Theorem 5.1 with f (R) = CR− α and g(x, t) =
|x|α + Ct, so that we obtain a lower bound of the form
ζ−2
kPt (x, ·) − µ? kTV ≥ C(|x|α + t) ζ−2−α .
Taking again α close to ζ − 1 and combining this with (7.7), we see that one does
actually have the identity
log kPt (x, ·) − µ? kTV
lim =2−ζ .
t→∞ log t
E XAMPLES 37

7.3 Conclusive remarks

We have seen in the previous two examples that Theorems 4.1 and 5.1 are surpris-
ingly sharp in the sense that we are able to match upper and lower bounds almost
exactly. This is not a special feature of these two examples, but a widespread prop-
erty of Markov processes that exhibit subgeometric convergence. In a nutshell, the
reason is that Lyapunov function techniques essentially reduce a potentially very
complicated Markov process to a one-dimensional situation, since all the informa-
tion on the process is simply encoded by the value V (xt ). It is not surprising then
that these techniques tend to be sharp for one-dimensional systems, but one might
wonder whether they still perform well for higher dimensional systems.
The reason why this is still the case is that the convergence of Markov process
that exhibit subgeometric convergence to stationarity is dominated by some rela-
tively rare but very long excursions into “bad regions” from which the process takes
very long times to escape. In both of the previous examples, these were the regions
where |x| is large, but it may potentially be much more complicated regions.
Very often, there will be one “worst case scenario” degree of freedom that
varies on a very slow timescale when the process gets trapped in one of these
bad regions. The aim of the game is then to gain a good understanding of the
mechanisms that allows one to construct a Lyapunov function that captures the
dynamic of this slow degree of freedom. A good example of situations where this
philosophy can be implemented in a more complicated setting is given in [HM09,
Hai09].

References
[Bax05] P. H. BAXENDALE. Renewal theory and computable convergence rates for
geometrically ergodic Markov chains. Ann. Appl. Probab. 15, no. 1B, (2005),
700–738.
[BCG08] D. BAKRY, P. C ATTIAUX, and A. G UILLIN. Rate of convergence for ergodic
continuous Markov processes: Lyapunov versus Poincaré. J. Funct. Anal.
254, no. 3, (2008), 727–759.
[Cha89] K. S. C HAN. A note on the geometric ergodicity of a Markov chain. Adv. in
Appl. Probab. 21, no. 3, (1989), 702–704.
[DFG09] R. D OUC, G. F ORT, and A. G UILLIN. Subgeometric rates of convergence
of f -ergodic strong Markov processes. Stochastic Process. Appl. 119, no. 3,
(2009), 897–923.
[DFMS04] R. D OUC, G. F ORT, E. M OULINES, and P. S OULIER. Practical drift condi-
tions for subgeometric rates of convergence. Ann. Appl. Probab. 14, no. 3,
(2004), 1353–1377.
[DMLM03] P. D EL M ORAL, M. L EDOUX, and L. M ICLO. On contraction properties of
Markov kernels. Probability Theory and Related Fields 126, no. 3, (2003),
395–420.
[DMR04] R. D OUC, E. M OULINES, and J. S. ROSENTHAL. Quantitative bounds on
E XAMPLES 38

convergence of time-inhomogeneous Markov chains. The Annals of Applied

Probability 14, no. 4, (2004), 1643–1665.
[Geo88] H.-O. G EORGII. Gibbs measures and phase transitions, vol. 9 of de Gruyter
Studies in Mathematics. Walter de Gruyter & Co., Berlin, 1988.
[Hai09] M. H AIRER. How hot can a heat bath get? Comm. Math. Phys. 292, no. 1,
(2009), 131–177.
[Har56] T. E. H ARRIS. The existence of stationary measures for certain Markov pro-
cesses. In Proceedings of the Third Berkeley Symposium on Mathematical
Statistics and Probability, 1954–1955, vol. II, 113–124. University of Cali-
fornia Press, Berkeley and Los Angeles, 1956.
[Has80] R. Z. H AS0 MINSKII. Stochastic stability of differential equations, vol. 7 of
Monographs and Textbooks on Mechanics of Solids and Fluids: Mechanics
and Analysis. Sijthoff & Noordhoff, Alphen aan den Rijn, 1980. Translated
from the Russian by D. Louvish.
[HM08] M. H AIRER and J. C. M ATTINGLY. Spectral gaps in Wasserstein distances
and the 2D stochastic Navier-Stokes equations. Ann. Probab. 36, no. 6,
(2008), 2050–2091.
[HM09] M. H AIRER and J. C. M ATTINGLY. Slow energy dissipation in anharmonic
oscillator chains. Comm. Pure Appl. Math. 62, no. 8, (2009), 999–1032.
[HM10] M. H AIRER and J. C. M ATTINGLY. Yet another look at Harris’ ergodic the-
orem for Markov chains. In Seminar on Stochastic Analysis, Random Fields
and Applications VI, Progr. Probab. Birkhäuser, Basel, 2010. To appear.
[Hör67] L. H ÖRMANDER. Hypoelliptic second order differential equations. Acta
Math. 119, (1967), 147–171.
[Hör85] L. H ÖRMANDER. The Analysis of Linear Partial Differential Operators I–IV.
Springer, New York, 1985.
[KM05] I. KONTOYIANNIS and S. P. M EYN. Large deviations asymptotics and the
spectral theory of multiplicatively regular markov processes. Electronic Jour-
nal of Probability 10, (2005), no. 3, 61123 (electronic).
[Koh78] J. J. KOHN. Lectures on degenerate elliptic problems. In Pseudodifferen-
tial operator with applications (Bressanone, 1977), 89–151. Liguori, Naples,
1978.
[KS84] S. K USUOKA and D. S TROOCK. Applications of the Malliavin calculus. I.
In Stochastic analysis (Katata/Kyoto, 1982), vol. 32 of North-Holland Math.
Library, 271–306. North-Holland, Amsterdam, 1984.
[KS85] S. K USUOKA and D. S TROOCK. Applications of the Malliavin calculus. II.
J. Fac. Sci. Univ. Tokyo Sect. IA Math. 32, no. 1, (1985), 1–76.
[KS87] S. K USUOKA and D. S TROOCK. Applications of the Malliavin calculus. III.
J. Fac. Sci. Univ. Tokyo Sect. IA Math. 34, no. 2, (1987), 391–442.
[Mal78] P. M ALLIAVIN. Stochastic calculus of variations and hypoelliptic operators.
Proc. Intern. Symp. SDE 195–263.
[MT92] S. P. M EYN and R. L. T WEEDIE. Stability of Markovian processes. I. Crite-
ria for discrete-time chains. Adv. in Appl. Probab. 24, no. 3, (1992), 542–574.
E XAMPLES 39

[MT93] S. P. M EYN and R. L. T WEEDIE. Markov chains and stochastic stability.

Communications and Control Engineering Series. Springer-Verlag London
Ltd., London, 1993.
[Nor86] J. N ORRIS. Simplified Malliavin calculus. In Séminaire de Probabilités,
XX, 1984/85, vol. 1204 of Lecture Notes in Math., 101–130. Springer, Berlin,
1986.
[NT83] E. N UMMELIN and P. T UOMINEN. The rate of convergence in Orey’s theo-
rem for Harris recurrent Markov chains with applications to renewal theory.
Stochastic Process. Appl. 15, no. 3, (1983), 295–311.
[Nua95] D. N UALART. The Malliavin calculus and related topics. Probability and its
Applications (New York). Springer-Verlag, New York, 1995.
[Num84] E. N UMMELIN. General irreducible Markov chains and nonnegative oper-
ators, vol. 83 of Cambridge Tracts in Mathematics. Cambridge University
Press, Cambridge, 1984.
[Ons44] L. O NSAGER. Crystal statistics. I. A two-dimensional model with an order-
disorder transition. Phys. Rev. (2) 65, (1944), 117–149.
[RT99] G. O. ROBERTS and R. L. T WEEDIE. Bounds on regeneration times and
convergence rates for Markov chains. Stochastic Process. Appl. 80, no. 2,
(1999), 211–229.
[RY99] D. R EVUZ and M. YOR. Continuous martingales and Brownian motion,
vol. 293 of Grundlehren der Mathematischen Wissenschaften [Fundamen-
tal Principles of Mathematical Sciences]. Springer-Verlag, Berlin, third ed.,
1999.
[TT94] P. T UOMINEN and R. L. T WEEDIE. Subgeometric rates of convergence of
f -ergodic Markov chains. Adv. in Appl. Probab. 26, no. 3, (1994), 775–798.
[Ver99] A. Y. V ERETENNIKOV. On polynomial mixing and the rate of conver-
gence for stochastic differential and difference equations. Teor. Veroyatnost.
i Primenen. 44, no. 2, (1999), 312–327.