An Introduction to Rough Paths

Antoine Lejay

Projet OMEGA (INRIA Lorraine),

IECN, Campus scientifique,
BP 239,
54506 Vandœuvre-lès-Nancy CEDEX, France
e-mail: [email protected]
http://www.iecn.u-nancy.fr/~lejay

Key words: controlled differential equations, integration against irregular

paths, p-variation, stochastic processes, iterated integrals, Chen series, geo-
metric multiplicative functional

Summary. This article aims to be an introduction to the theory of rough paths, in

which integrals of differential forms against irregular paths and differential equations
controlled by irregular paths are defined. This theory makes use of an extension of
the notion of iterated integrals of the paths, whose algebraic properties appear to
be fundamental. This theory is well-suited for stochastic processes.

1 Introduction
This article is an introduction to the theory of rough paths, which has been
developed by T. Lyons and his co-authors since the early ’90s. The main results
presented here are borrowed from [32, 36]. This theory concerns differential
equations controlled by irregular paths and integration of differential forms
against irregular trajectories. Here, x is a continuous function from [0, 1] to Rd ,
and the notion of irregularity we use is that of p-variation, as defined by
N. Wiener. This means that for some p
1,

k−1
sup |xti+1 − xti |p < +∞.
k
1, 0t0 ···tk 1 i=0
partition of [0,1]

As we will see, the integer p plays an important role in this theory.
In probability theory, most stochastic processes are not of finite variation,
but are of finite p-variation for some p > 2. We show in Sect. 10 how to apply

J. Azéma et al. (Eds.): LNM 1832, pp. 1–59, 2003.

c Springer-Verlag Berlin Heidelberg 2003
2 Antoine Lejay

this theory to Brownian motion. But the theory of rough paths could be used
for many other types of processes, as presented in Sect. 12.
Firstly, we give a meaning to the integral
 t 
f (xs ) dxs , or equivalently, f (1.1)
0 x([0,t])

for a differential form

d
f (x) = fi (x) dxi . (1.2)
i=1

We are also interested in solving the controlled differential equation

dyt = f (yt ) dxt , (1.3)

where f is the vector field

d
∂
f (y) = fi (y) .
i=1
∂xi

This will be done using Picard’s iteration principle, from the result on integra-
tion of one-forms. Using the terminology of controlled differential equations,
x is called a control.
The theory of rough paths also provided some results on the continuity of
the map x → y, where y is given either by (1.1) or (1.3).
The theory of rough paths may be seen as a combination of two families
of results:
(1) Integration of functions of finite q-variation against functions of finite
p-variation with 1/p + 1/q > 1 as defined by L.C. Young in [52].
(2) Representation of the solutions of (1.3) using iterated integrals of x: this
approach is in fact an algebraic one, much more than an analytical one.
Let us give a short review of these notions.

(1) Young’s integral

Let x and y be two continuous functions respectively 1/p and 1/q-Hölder

t
continuous with θ = 1/p + 1/q > 1. Then, Young’s integral s yr dxr of y
k−1
against x is defined as the limit of Is,t (Π) = i=0 yti (xti+1 − xti ) when the
mesh of the partition Π = { ti s  t0  · · ·  tk  t } of [s, t] goes to zero (see
for example [12, 52]). It is possible to choose a point tj in Π such that
1
|Is,t (Π) − Is,t (Π \ { tj })|  C|t − s|θ
(Card Π)θ
for some constant C that depends only on the Hölder norm of x and y. What-
ever the size of the partition Π is, |Is,t (Π)|  |ys (xt − xs )|+ |t− s|θ ζ(θ), where
 An Introduction to Rough Paths 3

ζ(θ) = n
1 1/nθ . The limit of Is,t (Π) as the mesh of Π goes to 0 may be
considered.
One may be tempted to replace y by f (x), where the regularity of f de-
pends on the irregularity of x. But to apply directly the proof of L.C. Young,
one has to assume that f is α-Hölder continuous with α > p − 1, which is
too restrictive as soon
as p 
2. To bypass this limitation, we construct when
d t
xt ∈ Rd the integral j=1 s fj (xr ) dxjr as



k−1 d
d
∂fj1 i,(j ,j )
lim fj (xti )(xjti+1 − xjti ) + (xti )xti ,t2i+11
mesh(Π)→0
i=0 j=1 j ,j =1
∂xj2
1 2

d 
∂ p−1 fj1 p,(j ,...,j1 )
+ ···+ (xti )xti ,ti+1
p (1.4)
j1 ,...,j
p =1
∂xj
p
· · · ∂xj2

with formally 
i,(j ,...,j1 )
xs,t i = dxjsii · · · dxjs11 . (1.5)
ssi ···s1 t

This expression (1.4) is provided by the Taylor formula on f and the more x
is irregular, i.e., the larger p is, the more regular f needs to be.
What makes the previous definition formal is that the “iterated integrals”
of x have to be defined, and there is no general procedure to construct them,
nor are they unique. The terms xk,(i1 ,...,ik ) for k = 2, . . . , p are limits of
iterated integrals of piecewise smooth approximations of x, but they are sen-
sitive to the way the path x is approximated. Due to this property, the general
principle in the theory of rough paths is:
d  t
The integral j=1 s fj (xr ) dxjr is not driven by x but, if it exists, by
x = (x1,(i1 ) , x2,(i1 ,i2 ) , . . . , xp,(i1 ,...,i
p ) )i1 ,...,i
p =1,...,d corresponding
formally to (1.5).

(2) Formal solutions of differential equations

k,(i ,...,i )
Assume now that x is smooth, and let xs,t 1 k
be its iterated integrals de-
fined by (1.5). Given some indeterminates X , . . . , X d , we consider the formal
1

non-commutative power series:

k,(i ,...,ik )
Φ([s, t], x) = 1 + X i1 · · · X ik xs,t 1 .
k
1 (i1 ,...,ik )∈{ 1,...,d }k

As first proved by K.T. Chen in [6], Φ([s, t], x) fully characterizes the path x,
and for all s  u  t,

Φ([s, u], x)Φ([u, t], x) = Φ([s, t], x). (1.6)

4 Antoine Lejay

This relation between iterated integrals is also used to prove that the limit
in (1.4) exists. If exp is the non-commutative exponential (defined by a power
series), then there exists a formal series Ψ ([s, t], x) such that Φ([s, t], x) =
exp(Ψ ([s, t], x)) and


k,(i ,...,ik )
Ψ ([s, t], x) = F(i1 ,...,id ) (X 1 , . . . , X d )xs,t 1
k
1 (i1 ,...,ik )∈{ 1,...,d }k

where F(i1 ,...,id ) (X 1 , . . . , X d) belongs to the Lie algebra generated by the in-
determinates X 1 , . . . , X d , i.e., the smallest submodule containing X 1 , . . . , X d
and closed under the Lie brackets [Y, Z] = Y Z − ZY .
If f = (f1 , . . . , fd ) and each of the fi is linear, i.e., fi (y) = Ci y where Ci
is a matrix, then the solution y of (1.3) is equal to


yt = exp Ψ([s, t], x) ys ,

where Ψ ([s, t], x) is equal to Ψ ([s, t], x) in which X i was replaced by the ma-
trix Ci . If f is not linear, but is for example a left-invariant vector field on
a Lie group, then a similar relation holds, where X i is replaced by fi , and
the Lie brackets [·, ·] are replaced by the Lie bracket between vector fields.
Here, the exponential is replaced by the map defining a left-invariant vector
field from a vector in the Lie algebra, i.e., the tangent space at 0 (see for
example [13]).
This result suggests that when one knows x, he can compute its iterated
integrals and then formally solve (1.3) by replacing the indeterminates by f .
In fact, when x is irregular, the solution y of (1.3) will be constructed using
Picard’s iteration
t principle, i.e., as the limit of the sequence y n defined by
n+1
yt = y0 + 0 f (yr ) dxr . But it corresponds, if (xδ )δ>0 is a family of piecewise
n

smooth approximations of x and f is smooth, to



y = lim y δ with ytδ = exp Ψ [0, t], xδ y0 .
δ→0

However, in the previous expression, we need all the iterated integrals of x.

Yet, even if x is irregular, there exists a general procedure to compute them
all, assuming we know x defined formally by (1.5). However, different families
of approximations (xδ )δ>0 may give rise to different x. Thus, the solution y
of (1.3) given by the theory of rough paths depends also on x and not only
on x, and the general principle stated above is also respected.

Geometric multiplicative functionals

As we have seen, we need to construct an object x corresponding to the

iterated integrals of an irregular path up to a given order p. Since x may be
reached as the limit of smooth paths together with its iterated integrals, x may
be seen as an extension by continuity of the function x → Φ([s, t], x) giving the
 An Introduction to Rough Paths 5

truncated Chen series, where X i1 · · · X ik is set equal to 0 as soon as k > p.

This means in particular that, at the limit, we keep the algebraic relation (1.6).
This means that xs,t may be seen as a formal non-commutative polynomial (in
the text, we use tensor products, but this is equivalent in the finite-dimensional
setting). Set x0s,t = 1 and xks,t = (i1 ,...,ik ) X i1 · · · X ik xk,(i1 ,...,ik ) . Thus, xs,t =
p
1 + x1s,t + · · · + xs,t . The relation (1.6) becomes

xs,t = xs,u xu,t for all 0  s  u  t  1. (1.7)

This means that, if x lies above x,

x1s,t = x1s,u + x1u,t with x1,i

s,t = xt − xs
i i
(1.8)
x2s,t = x2s,u + x2u,t + x1s,u x1u,t ,
.. . .
. = .. + .. + ···

Thus for k = 1, . . . , p, one can compute xks,t from xis,u and xiu,t when these
quantities are known for i = 1, . . . , k.
The objects x that could be reached as an extension of the truncated
Chen series Φ([s, t], x) and satisfying (1.7) are called geometric multiplicative
functionals.
Our goal is to construct  from x new geometric multiplicative functionals z.
For example, the integral f (xr ) dxr will itself be constructed as a geometric
t
multiplicative functional. Remark that for zs,t = s f (xr ) dxr , (1.8) is no more
than the Chasles relation.
The machinery we use to construct z is the following: We construct first
an approximation y of z. For example, if x is of finite p-variation  t with p < 2,
we define ys,t1
by f (xs )(xt − xs ), which is an approximation of s f (xr ) dxr .
The object y is a non-commutative polynomial, but y does not satisfy re-
lation (1.7) in general. Thus, if Π = { ti 0  t1  · · ·  tk  1 } is a partition
of [0, 1], we set
Π
ys,t = ys,ti yti ,ti+1 · · · ytj−1 ,tj ytj ,t .
where i and j are the smallest and the largest integers such that [ti , tj ] ⊂ [s, t].
At the first level of path, this relation reads

d 
Π,1 1,(i) 1,(i) 1,(i) 1,(i)
ys,t = X i ys,ti + yti ,ti+1 + · · · + ytj−1 ,tj + ytj ,t .
i=1

But for yΠ,2 , this relation implies all the ytΠ,2

j ,tj+1
’s an ytΠ,1
j ,tj+1
’s for tj ∈ Π.
Π
Of course, y also fails to satisfy (1.7), except if s, u and t belong to Π.
But provided one has a relation of the type

|ys,t − ys,u yu,t |  ε(s, u, t)

6 Antoine Lejay

Π Π\{ t }
for a nice ε, then one could consider the difference between ys,t and ys,t j
for an element tj in Π ∩(s, t). If tj is well chosen, and the choice of tj is similar
to the one done for the construction of Young’s integral (see above), one could
show that |ys,tΠ
|  C(s, t), for a function C(s, t) that does not depend on the
partition Π. One could then pass to the limit as the mesh of the partition Π
decreases to 0. Of course, it has to be proved that under reasonable conditions,
the limit, which we denote by z, is unique and is a geometric multiplicative
functional. Moreover, the work has to be done iteratively at each level of
iterated integrals. Thus, if z1 , . . . , zk are already constructed, one gains the
fact that (z1 , . . . , zk ) satisfies (1.7), and zk+1 is constructed using the previous
machinery.

Defining the iterated integrals

Since the previoulsy described procedure is general, anybody interested only

in applying this theory could adopt the following point of view:
 t
p either yt = y0 + 0 f (yr ) dxr
1
x = (x , . . . , x ) Black box t
or zt = z0 + 0 f (xr ) dxr

f, ∂f, . . . , ∂ p f
and focus on x. In the previous section, we have seen how to construct new
geometric multiplicative functionals from x, but we have not said how x is
constructed. We have already said that x may be difficult to construct. The
most natural approach is to choose a piecewise smooth approximation xδ of x
k,(i ,...,ik )
and to define xs,t 1 as the limit of

dxiskk,δ · · · dxis11,δ .
ssk ···s1 t

For example, consider a d-dimensional Brownian motion B. As its trajectories

are α-Hölder continuous for any α < 1/2, they are of finite p-variation for
any p > 2. Hence, applying the theory of rough paths requires knowing the
equivalent of the second-order iterated integrals of the Brownian motion. Let
B δ (ω) be an approximation of B(ω). One knows that the convergence of
δ
Is,t (ω) = ss2 s1 t dBsi,δ
2
(ω) dBsj,δ
1
(ω) depends on the choice of (B δ )δ>0 (see
[19, Sect. VI-7, p. 392]). Besides, if for example, B δ (ω) is a piecewise linear
approximation of B(ω) sampled along deterministic
 partitions, then the limit
δ
of Is,t is the Stratonovich integral Is,t = ss2 s1 t ◦dBsi 2 ◦dBsj1 and is defined
only as a limit in probability. (Yet some recent works prove that for dyadic
partitions, the convergence may be almost sure. See for example [24]) Another
difficulty when we want to use rough paths is that we need to prove that B δ
and I δ converge to B and I in the topology generated by the distance in
p-variation, which is more complicated to use than the uniform norm.
 An Introduction to Rough Paths 7

Trajectories of stochastic processes represent a natural class of irregu-

lar paths for which one may require some integration theory. Thus, the
theory of rough paths provides a pathwise theory of integration, but path-
wise with respect to x, and not to the stochastic process x. So, the dif-
ficulty is to construct the iterated integrals of the trajectories of x. How-
ever, it has to be noted that constructing x may be simpler than construct-
ing stochastic
 integrals driven by x. For example, for the Brownian mo-
tion, ss2 s1 t ◦dBsi 2 ◦dBsj1 = As,t (B i , B j ) + 12 (Bti − Bsi )(Btj − Bsj ), where
As,t (B i , B j ) is the Lévy area of (B i , B j ). This functional As,t (B i , B j ), which
represents the area enclosed between the curve of r ∈ [s, t] → (Bri , Brj ) and
its chord, was constructed by Paul Lévy (see for example [29]) before the
construction of Itô or Stratonovich stochastic integrals.
Using the theory of rough paths, one has then to focus on the construction
of x for given trajectories of x. This has some advantages, among which: (i)
t
It is easier to define an object like s ◦dxr ◦dxr for a stochastic process than
t
s
f (xr )◦dxr or solving dyt = f (yt )◦dxt . Section 12 contains a list of types
of stochastic processes for which the theory of rough paths may be used, and
then may be directly applied to solve differential equations. Moreover, the
separation of x and f may be advantageous since we also gain knowledge of
the algebraic structure of x: see [37, 47] for an original application to Monte
Carlo methods. (ii) A support theorem is immediate once we have one on x
(see [24] for an application). (iii) Different piecewise smooth approximations
of a stochastic process lead to different stochastic integrals. This is well known
for Brownian motion, but generalizes immediately to different processes for
which the theory of rough paths may be applied. Besides, this theory provides
some explanations on the form of the corrective drift (see Sects. 6.2 and 10.2).
(iv) Once x has been defined on a probability space (Ω,  F , P), then all the
differential equations dyt = f (yt ) dxt and the integrals f (xt ) dxt are defined
on the same set Ω0 ⊂ Ω of full measure, whatever the function f is.
In this article, we assume that the path x takes its values in V = Rd ,
and that the differential forms or vector fields f take their values in W =
Rm . However, V and W could in fact be any Banach space, even of infinite
dimension.

Motivations

This article does not give a full treatment of the theory of rough paths. But
its aim is to give the reader sufficient information about this theory to help
him to have a general view of it, and maybe to apply it. The reader who is
interested in this theory can read either [32] or [36] to go further.
The theory of rough paths is suitable for trajectories of stochastic pro-
cesses, since there are many types of stochastic processes for which it is possi-
ble to construct their “iterated integrals”. Yet each application to a particular
type of probabilistic problem may require a specific approach. As randomness
8 Antoine Lejay

plays no role in this theory, probability theory takes only a small place in this
article. The reader is refered to Sect. 12 and to the bibliography for applica-
tions to stochastic analysis.

Outline of the article

For the sake of clarity, we explain in Sect. 2 how to integrate a differential

form along a path of finite p-variation with p ∈ [1, 2), and then how to solve
a differential equation controlled by such a path. In Sect. 3, we deal with
paths of finite p-variation with p ∈ [2, 3). This is the most common case a
probabilist could use. Besides, we think that understanding the situation in
this case together with the proofs of Sect. 2 allows us to fully understand the
general theory.
Sections 4 and 5 are devoted to introducing the basic algebraic results
on iterated integrals. Section 6 gives the general definition of geometric mul-
tiplicative functionals, i.e., the objects x previously introduced, and some
convergence results on them. The notion of almost multiplicative functional,
which is the basic element to define an integral, is presented in Sect. 7. The
general results on integration of one-forms and controlled differential equa-
tions are given in Sects. 8 and 9 without proof.
A practical example is presented in Sects. 10 and 11, where the theory is
applied to the Brownian motion. Section 11 also contains a method to compute
the p-variation of a multiplicative functional.
Finally, Sect. 12 contains a list, which attempts to be as complete as
possible at the date of writing, of bibliographic references on works using
the theory of rough paths. This article ends with some bibliographical and
historical notes.

Note.

At first reading, the reader may go directly from the end of Sect. 3 to Sect. 10
for an application to the Brownian motion.

Acknowledgement

I am glad to have been granted a post-doctoral position by the European Union’s

TMR Stochastic Analysis Network and to have benefited from Prof. T. Lyons’ kind
hospitality at Oxford University during the academic year 2000–2001. I have learned
there, among other things, the basis of the theory of rough paths. I wish also thank
all of those, and especially L. Coutin, M. Émery, M. Ledoux and N. Victoir, who
have read the early versions of the manuscript, pointed out some mistakes, and made
valuable comments.
 An Introduction to Rough Paths 9

2 The case of a not so irregular control


Set ∆+ = (s, t) ∈ [0, 1]2 0  s  t  1 . Through all this article, we use a
function ω from ∆+ to R+ satisfying the following assumption.
Assumption 2.1. The function ω : ∆+ → R+ is such that
(i) The function ω is bounded.
(ii) The function ω is continuous near the diagonal, i.e., on { (s, s) s ∈ [0, 1] },
and ω(s, s) = 0 for all s ∈ [0, 1].
(iii) For all 0  s  u  t  1,
ω(s, u) + ω(u, t)  ω(s, t). (2.1)
If follows immediately that for all θ
1, ω θ is also super-additive:
ω(s, u)θ + ω(u, t)θ  ω(s, t)θ for all s  u  t  1.
Moreover, it is easily seen that for all ε, there exists some η small enough
such that |t − s| < η implies that ω(s, t) < ε for all (s, t) ∈ ∆+ .

2.1 Integration of a differential form along an irregular path

In this section, we show that, provided one controls the value of |xt − xs |p for
t
p ∈ (1, 2), then 0 f (xs ) dxs may be defined with Riemann sums.
Assumption 2.2. There exists a real 1  p < 2 such that
|xt − xs |p  ω(s, t) for all (s, t) ∈ ∆+ (2.2)
for a function ω satisfying Assumption 2.1. For example, this is true if x is
1/p-Hölder continuous, in which case, ω(s, t) = C|t − s| for some constant C.
The differential form f defined by (1.2) is identified with (f1 , . . . , fd ) :
Rd → (Rm )d . The function f is bounded and α-Hölder continuous, with α >
p − 1.
Note that (2.2) together with (2.1) exactly means that x is of finite p-
variation for some p ∈ [1, 2).
Of course, (1.1) will be defined as limit of Riemann sums. In order to do
so, set, for all 0  s  t  1,
ys,t = f (xs )(xt − xs ). (2.3)
For all δ > 0, let Π δ be a family of partitions 0  tδ1  · · ·  tδkδ  1 of [0, 1]
whose meshes decrease to 0 as δ decreases to 0. Assume that for all 0 < δ  < δ,

Πδ ⊂ Πδ .
For all δ > 0, set

δ

−1
Π
zs,t = ys,tδj + ytδ ,t + ytδi ,tδi+1 ,
i=j

where j and  are such that Π δ ∩ (s, t) = tδj , . . . , tδ .
10 Antoine Lejay
δ
Π
Proposition 2.1. Under Assumption 2.2, zs,t admits a limit denoted by zs,t
for all 0  s  t  1. Furthermore, (s, t) ∈ ∆+ → zs,t is continuous, and
zs,u + zu,t = zs,t (Chasles’ relation) for all 0  s  u  t  1.
Finally, there exists some constant K depending only on f , p and ω(0, 1)
such that |zs,t |p  Kω(s, t) for all (s, t) ∈ ∆+ . This implies that z has finite
p-variation.
t
Thus, one may define s f (xr ) dxr to be zs,t . The proof relies on the fol-
lowing Lemmas.

Lemma 2.1. There exists a constant C depending only on f such that for all
0  s  u  t  1,
1+α
|ys,t − ys,u − yu,t |  Cω(s, t)θ , with θ = > 1. (2.4)
p

Proof. Since f is Hölder continuous and x satisfies (2.1), it is easily established

that, for C = supx=y |f (x) − f (y)|/|x − y|α ,

|ys,t − ys,u − yu,t |  |f (xu ) − f (xs )||xt − xu |  C|xu − xs |α |xt − xu |

 Cω(s, u)α/p ω(u, t)1/p  Cω(s, t)(1+α)/p .

Hence (2.4) is proved. 

Lemma 2.2. Let 0  s  t  1, and let s < t1  . . .  tk < t be a partition

of (s, t). Then, if k
2, there exists an integer  in { 1, 2, . . . , k } such that
2
ω(t−1 , t+1 )  ω(s, t),
k
with the convention that t0 = s and tk+1 = t.

Proof. The result is clear if k = 2, since ω(t1 , t2 )  ω(s, t). Assume that k
3.

As ω is super-additive, ki=1 ω(ti−1 , ti+1 )  2ω(s, t). So, at least one of the
ω(ti−1 , ti+1 )’s is smaller than 2ω(s, t)/k. 

Proof (Proof of Proposition  2.1). Fix δ > 0 and 0  s  t  1. We have

Π δ ∩ (s, t) = tδj , . . . , tδ .
δ
If Π δ ∩ (s, t) = ∅, then zs,t
Π
= ytδi ,tδi+1 , where the integer i is such that
[s, t] ⊂ [tδi , tδi+1 ].
If Π δ ∩ (s, t) contains at least one point, then we choose an integer k such
that j  k  , and we construct a new partition

Π = tδj , . . . , tδk−1 , tδk+1 , . . . , tδ

by suppressing the point tδk . We use the convention that tδj−1 = s and tδ+1 = t.
According to Lemma 2.2, the point tδk is chosen so that
 An Introduction to Rough Paths 11

2
ω(tk−1 , tk+1 )  ω(s, t).
|Π δ ∩ (s, t)|

Thus, using the previous notations,

δ
Π
zs,t Π
= zs,t + ytδk−1 ,tδk + ytδk ,tδk+1 − ytδk−1 ,tδk+1 .

With Lemma 2.1,

 θ
 Πδ  2
zs,t − zs,t
Π
 Cω(tk−1 , tk+1 )θ  C ω(s, t)θ
|Π δ ∩ (s, t)|

for a constant C which is equal to the Hölder norm of f .

Suppressing a carefully chosen point in Π, and reiterating the process, one
easily obtains that
 Πδ   
zs,t − ys,t   Kω(s, t)θ + ys,t − ytδ ,tδ , (2.5)
 i(δ) i (δ)


where K = C + 2θ C n
1 1/nθ , and i and i are such that [tδi(δ) , tδi (δ) ] is the
smallest interval containing [s, t]. In particular, tδi(δ) increases to s and tδi (δ)
decreases to t as δ decreases to 0. 
Let 0  s  u  t  1. Set Π δ ∩ (s, u) = tδj , . . . , tδk and Π δ ∩ [u, t) =
δ 
tj  , . . . , tδk . So,

Πδ Πδ Πδ
zs,u + zu,t = zs,t − ytδk ,tδ − ytδk ,u − yu,tδ . (2.6)
j j

As f is bounded, |yr,r |  f ∞ ω(r, r )1/p −−−−−−→ 0. Moreover, tδk −−−→

|r −r|→0 δ→0
δ
u and tδj −−−→ u. Set Ztδ = z0,t
Π
. With (2.6), the inequality |yr,r | 
δ→0
f ∞ ω(r, r )1/p and the continuity of ω near its diagonal, it is easily proved
that (Z δ )δ>0 satisfies the conditions of the Ascoli theorem, i.e., for any κ > 0,
there exists some η > 0 such that sup|t−s|<η |Ztδ − Zsδ |  κ. Thus, there ex-
ists a subsequence of (Z δ )δ>0 which converges uniformly to some continuous
function Z on [0, 1].
Πδ Πδ Πδ
One could set zs,t = Zt − Zs . Again with (2.6), zs,t = z0,t − z0,s +
(ytδk ,tδk+1 − ytδk ,s − ys,tδk+1 ), where tδk and tδk+1 are two adjacent points of Π δ
δ
such that s ∈ (tδk , tδk+1 ]. Hence, zs,t
Π
converges to zs,t for all (s, t) ∈ ∆+ . Hence,
it follows from (2.6) that z satisfies the Chasles relation: zs,t = zs,u + zu,t for
all 0  s  u  t  1. Besides, from the continuity of Z, (s, t) ∈ ∆+ → zs,t is
continuous.
Let Z  be another limit of the sequence (Z δ )δ>0 , and set zs,t = Z t − Zs
for all (s, t) ∈ ∆ . As for z, z also satisfies the Chasles relation, and so is
+

∆z = z − z. However, |∆z|  2Kω(s, t)θ . For any partition Π = { t1 , . . . , tk }

of [s, t],
12 Antoine Lejay

k−1

k−1
|∆zs,t |  |∆zti ,ti+1 |  2K ω(ti , ti+1 )θ
i=1 =1

 2Kω(s, t) sup ω(ti , ti+1 )θ−1 −−−−−−−−→ 0.

i=1,...,k mesh(Π)→0

δ
Π
So, the limit of (zs,t )δ>0 is unique.
With (2.5) and the boundedness of f , |zs,t |  Kω(s, t)θ +|ys,t |. But |ys,t | 
f ∞ |xs,t |  f ∞ ω(s, t)1/p . Thus, |zs,t |p  (Kω(0, 1)α + f p∞ )ω(s, t) and
z is of finite p-variation. The proposition is then proved. 

2.2 The semi-norm of p-variation

In the preceding proof, the regularity of x plays in fact no role. The only
condition required is (2.2). Note that (2.2) implies that for any partition
Π = { t0 , . . . , tk } of [s, t],

k−1

k−1
|xti+1 − xti |p  ω(ti , ti+1 )  ω(s, t).
i=0 i=0

Define the semi-norm of p-variation by

k−1 1/p


Var (x) = sup |xti+1 − xti | p
. (2.7)
p,[s,t] Π={ t0 ,...,tk } i=0
partition of [s,t]

Remark 2.1. When one considers x(t) = t and p > 1, it is immediate that
for
k−1 any partition 0  t0  · · ·  tk  1, the following inequality holds:
i=0 |ti+1 − ti |  supi=0,...,k−1 |ti+1 − ti |
p p−1
. The later quantity converges
to 0 with the mesh of the partition. But Varp,[s,t] (x) = 1. This means that in
the definition of the p-variation, we have really to consider a supremum on all
the partitions, and not only on those whose mesh converges to 0.

An interesting property of the p-variation is that as soon as Varp,[s,t] (x)

is finite, then Varq,[s,t] (x)  Varp,[s,t] (x) for all q
p. In other words, any
function of finite p-variation is of finite q-variation for all q
p.
Inequality (2.2) in Assumption 2.2 means that Varp,[s,t] (x)  ω(s, t) for all
0  s  t  1. On the other hand, we know that Varp,[s,u] (x) + Varp,[u,t] (x) 
Varp,[s,t] (x).
Although Varp,[0,1] is only a semi-norm, Varp,[0,1] (·) + ·∞ is a norm on
the space of continuous function. However, the space of continuous functions
with this norm is not separable.
Set for (s, t) ∈ ∆+ and two continuous functions x and y,

δp,[s,t] (x, y) = Var (x − y),

p,[s,t]
 An Introduction to Rough Paths 13

and δp,[s,t] (x) = δp,[s,t] (x, 0). Set also δp (x, y) = δp,[0,1] (x, y) and δp (x) =
δp,[0,1] (x). Note that δp,[0,1] is a not a distance, excepted when restricted to
functions for which x0 is fixed.
Let x be a function such that δp (x) is finite. If there exists a function
ω : ∆+ → R+ satisfying Assumption 2.1 and such that
|xt − xs |p  ω(s, t)
for all (s, t) ∈ ∆+ , then x is said to be of finite p-variation controlled by ω. It
is clear that the function ω defined by ω(s, t) = δp,[s,t] (x)p satisfies Assump-
tions 2.1 and that for all (s, t) ∈ ∆+ , |xt − xs |p  ω(s, t).
The following lemma is related to sequences of functions of finite p-
variation.
Lemma 2.3. Let (xn )n∈N be a sequence of functions of finite p-variation and
let x be a function of finite p-variation such that δp (xn , x) converges to 0.
Then there exists a subsequence (xnk )k∈N and some function ω satisfying
Assumption 2.1 such that xnk and x are of finite p-variation controlled by ω.
Moreover, for any ε > 0, there exists an integer k for which
∀
k, δp,[s,t] (xn , x)  εω(s, t) for all (s, t) ∈ ∆+ .
Proof. There exists a subsequence (nk )k∈N such that δp (x, xnk )  4−k . Hence,
we set  


+∞
ω(s, t) = 2p−1 δp,[s,t] (x)p + 2k δp,[s,t] (xnk , x)p .
k=0
By our choice of the subsequence, this function ω is well defined for all (s, t) ∈
∆+ . As δp,[s,t] (xn , x)  δp (xn , x) −−−−→ 0, and δp,[s,t] (xn , x) is continuous near
+∞ n→∞
the diagonal, (s, t) → k=0 2k δp,[s,t] (xnk , x)p is continuous near the diagonal.
Similarly, (s, t) → δp,[s,t] (xn )p is continuous near the diagonal. Clearly, ω is
super-additive and satisfies Assumption 2.1.
Since δp,[s,t] (xnk )p  2p−1 δp,[s,t] (xnk , x)p + 2p−1 δp,[s,t] (x)p , x and all of the
nk
x ’s are controlled by ω. Furthermore,
1
δp,[s,t] (xnk , x)p  ω(s, t),
2k
and the lemma is proved. 

The proof of the following Lemma is straightforward.

Lemma 2.4 (A convergence and compactness criterion). Let q be a
real number greater than p. Then,


Var (x − xn )q  2p sup (xr − xnr )q−p Var (x)p + Var (xn )p . (2.8)
q,[0,1] r∈[0,1] p,[0,1] p,[0,1]

Moreover, if (xn )n∈N converges pointwise to x, then

Var (x)  lim inf Var (xn ).
p,[0,1] n∈N p,[0,1]
14 Antoine Lejay

Hence, to prove that xn converges to x in q-variation, we have only to

prove that xn converges uniformly to x and that supn∈N Varp,[0,1] (xn ) is finite
for some p < q.
Thus, if (xn )n∈N is equi-continuous, uniformly bounded, and the sequence
(Varp,[0,1] (xn ))n∈N is also bounded, then there exists a subsequence of (xn )n∈N
which converges uniformly to a function x. With (2.8), xn converges in q-
variation to x for any q > p.

Remark 2.2 (Extension of Helly’s selection principle). If (xn )n∈N is a fam-

ily of continuous functions uniformly bounded and of finite p-variation such
that (Varp,[0,1] (xn ))n∈N is bounded, then there exists a function x of finite
p-variation and a subsequence (xnk )k∈N such that xnk converges pointwise
to x. But x is not necessarily continuous. This may be seen as an extension
of Helly’s selection principle (see [7, Theorem 6.1]).

2.3 Continuity

For any bounded and α-Hölder continuous function f with a Hölder constant
α > p−1, we have constructed a map Kf : x → z, where x is a function on [0, 1]

 t
with finite p-variation
 (with 1  p < 2) and z is the function s f (xr ) dxr ;
0st1 .
We have seen in Proposition 2.1 that Kf (x) is also of finite p-variation.
We are now interested in the continuity of Kf . Let x and x  be two func-
tions of finite p-variation, both satisfying Assumption 2.2 with respect to the
same ω.

Proposition 2.2. Assume that there exists some ε > 0 such that for all 0 
s  t  1,
|(xt − x
t ) − (xs − x
s )|p  εω(s, t),
and that x0 = x 0 . Then there exists a function κ(ε) decreasing to 0 as ε
decreases to 0 and depending only on f and p such that

|Kf (x)s,t − Kf (
x)s,t |  κ(ε)ω(s, t)1/p

for all (s, t) ∈ ∆+ .

Proof. The proof is similar to the one of Proposition 2.1. Using the same
Πδ Πδ
notations, define zs,t and zs,t . As previously, we create a new partition Π by
suppressing a carefully chosen point of Π δ . Hence, to estimate

Π δ 
Π 
 zs,t − zs,t
Πδ Π 
− zs,t − zs,t ,

we have only to estimate, for all u ∈ [s, t],

∆ = |(ys,t − ys,u − yu,t ) − (

ys,t − ys,u − yu,t )|,
 An Introduction to Rough Paths 15

with ys,t = f (xs )(xt − xs ) and ys,t = f ( xt − x

xs )( s ). Thus,


 
|∆|   f (xu ) − f (xs ) (xt − xu ) − f ( xu ) − f ( u )
xt − x
xs ) (
 |f (xu ) − f (xs )| |xt − x
t − (xu − xu )|
+ |f (xu ) − f (xs ) − f ( xs )| |
xu ) + f ( xt − x
u |.

Now, if C denotes the α-Hölder constant of f , we remark that

|f (xu ) − f (xs ) − f ( xs )|  2εα/p Cω(0, 1)1/p

xu ) + f (

and that
|f (xu ) − f (xs ) − f ( xs )|  2Cω(s, t)α/p .
xu ) + f (
Choosing β ∈ (0, 1) such that βα + 1 > p, one gets

|∆|  Cε1/p ω(s, t)(1+α)/p + 2Cεα(1−β)/p ω(0, 1)(1−β)/p ω(s, t)(αβ+1)/p .

Hence, as in the proof of Proposition 2.1,

 Πδ 
zs,t − zs,t
Πδ
− ys,t − ys,t   κ (ε)ω(s, t)θ ,

where κ (ε) decreases to 0 with , and depends only on f , x0 and p. On the

other hand, there exists some function κ decreasing to 0 with ε such that
|ys,t − ys,t |  κ (ε)ω(s, t)1/p . In the limit, if κ = κ ω(0, 1)(θ−1)/p + κ ,

|zs,t − zs,t | = |Kf (x)s,t − Kf (

x)s,t |  κ(ε)ω(s, t)1/p

The Proposition is then proved. 

Denote by G p (Rd ) the space of continuous functions in C([0, 1]; Rd) of finite
p-variation and starting at the same given point x0 . Denote by V p the topology
that the distance δp defines on the space G p (Rd ).

Corollary 2.1. Let f be a bounded and α-Hölder continuous function, and

let p ∈ [1, 2) be such that α > p − 1. Let (xn )n∈N be a sequence of continu-
ous functions in G p (Rd ) converging in V p to a function x in G p (Rd ). Then
Kf (xn ) ∈ G p (Rd ) converges in V p to Kf (x). Thus, x → Kf (x) is continuous
with respect to δp .

Proof. Using Lemma 2.3, there exists a function ω : ∆+ → R+ satisfying

Assumption 2.1 that controls x and xn (or maybe a subsequence of it) and
such that for any ε > 0, there exists some integer nε for which

|xt − xs − (xnt − xns )|p  εω(s, t)

for every n
nε and all (s, t) ∈ ∆+ . From Proposition 2.2, for all (s, t) ∈ ∆+
and any n ∈ N,

|Kf (x)s,t − Kf (xn )s,t |  κ(ε)ω(s, t)1/p ,

16 Antoine Lejay

with κ(ε) −−−→ 0. This function κ depends only on f and p. As ω is bounded

ε→0
on ∆+ ,
p
Var Kf (x) − Kf (xn ) −−−−→ 0.
p,[0,1] n→∞

The previous convergences are proved at least along a subsequence, but using
the way Lemma 2.3 is proved, the limit of Kf (xn ) is in fact unique. The
Corollary is then proved. 


As an application, let Π δ = tδi 0  tδ0  · · ·  tδkδ  1 be a family of
partitions of [0, 1] whose meshes go to 0 with δ. Then it is easily seen that the
piecewise linear approximation xδ of some path x ∈ G q (Rd ) for some q ∈ [1, 2)
given by

t − tδi
xδt = xtδi + (x δ − xtδi ) when t ∈ [tδi , tδi+1 ]
tδi+1 − tδi ti+1

converges uniformly to x.
Let 0  s0  . . .  s  1 be a partition of [0, 1]. Then,

δ
k
−1

−1
 δ q
|xδsi+1 − xδsi |q = x − xδsi 
si+1
i=1 j=0 i s.t. si ∈[tδj ,tδj+1 ]
δ
k
−1
 δ q
+ xs − xδsi  .
i+1
j=0 i s.t. tδj ∈(si ,si+1 )

However,
 δ q
x − xδsi   |xtj+1 − xtj |q
si+1
i s.t. si ∈[tδj ,tδj+1 ]

and if i is such that tδj ∈ (si , si+1 ) for a given j, then

 δ q
xs − xδsi   2q−1 |xtδj+1 − xtδj |q + 2q−1 |xtδj − xtδj−1 |q .
i+1

In the previous inequality, we set if necessary, tδ−1 = 0 and tδkδ +1 = 1. It

is now clear that Varq,[0,1] (xδ )  3 Varq,[0,1] (x), and then, from Lemma 2.4,
Varp,[0,1] (xδ − x) converges to 0 for all p > q. If follows that
 t  t
dxδr
f (xδr ) dr −−−→ f (xr ) dxr .
s dr δ→0 s

This convergence holds in fact both in p-variation and uniformly.

 An Introduction to Rough Paths 17

2.4 Solving differential equations

Let f = (f1 , . . . , fd ) be a function from W = Rm to Wd . We are now interested

in solving the differential equation
d 

t
yt = ys + fi (yr ) dxir (2.9)
i=1 s

where x is a continuous function of finite p-variation, with p ∈ [1, 2).

Theorem 2.1. If f is α-Hölder continuous with α > p − 1 and x is in

G p (Rd ), then there exists a solution y in G p (Rm ) to (2.9). Moreover, if f is
bounded, continuous with a bounded derivative which is α-Hölder continuous
with α > p − 1, then y with a given initial condition y0 , is unique. Besides,
the map x → y = If,y0 (x) (called the Itô map) is continuous from G p (Rd )
to G p (Rm ).

Proof. For two continuous paths y, x of finite p-variation, denote by L the

map defined by
   t 

m   d 

 
ys,t 0 fi (yr )
Ls,t (y, x) = = i
dyr + i
dxr
xs,t s 0 1
i=1 i=1

for any (s, t) ∈ ∆+ . Clearly, x = x. Define also I(y, x) = y. For any integer
n
1, set y n = I(y n−1 , x). Of course, if y n converges to some function y in
G p (Rm ), then y is solution to (2.9).
Step 1: Existence. Assume that two paths x and y of finite p-variation are
controlled respectively by ω and γω on a time interval [S, T ], for some con-
stant γ > 0.
A slight modification of the proof of Proposition 2.1 shows that there exists
some constant K, depending only on f and p, such that

|Is,t (y, x) − f (ys )xs,t |  γ α/p Kω(s, t)θ for all S  s  t  T,

1+α
with θ = p > 1. Hence,


|Is,t (y, x)|  Kγ α/p ω(S, T )θ−1/p + f ∞ ω(s, t)1/p .

If γ = 2p f p∞ , and S and T are close enough so that

ω(S, T )θ−1/p Kγ α/p  f ∞ ,

we have proved that on [S, T ], I(y, x) is of finite p-variation controlled

by γω(s, t).
Thus, one may construct a finite number N of intervals [Ti , Ti+1 ] such that
T0  T1  · · ·  TN and ω(Ti , Ti+1 )θ−1/p  2−α K −1 f 1−α
∞ .
18 Antoine Lejay

From a function y 0 of finite p-variation controlled by γω, one may recur-

sively construct functions y n by setting ys,t n
= I(y n−1 , x)s,t and y0n = y0 ,
where y0 is a given point in R .m

On each interval [Ti , Ti+1 ], y n is of finite p-variation controlled by γω.

From the convexity inequality, |ys,tn p
|  N p−1 γω(s, t) for all (s, t) ∈ ∆+ .
So, (yt ; 0  t  1)n∈N is equi-continuous, bounded, and according to
n

Ascoli’s theorem and Lemma 2.4, there exists some y of finite p-variation
such that a subsequence of (y n )n∈N converges to y in q-variation for some
q > p. But y → I(y, x) is also continuous on G q (Rm ). So, we deduce that y is
solution to (2.9) with the initial condition y0 .
Step 2: Uniqueness. In this step, assume that f is continuous, bounded with
a bounded α-Hölder continuous derivative with α > p − 1.
Let y and y be two paths of finite p-variation controlled by ω and starting
from the same point, that is y0 = y0 . Assume also that x is of finite p-variation
controlled by ω, and that y − y is of finite p-variation controlled by γω for
some γ > 0. It is clear that γ may be chosen smaller than 2p .
We are interested in I(y, x) − I( y , x). With our construction, this dif-
k−1
ference is approximated by i=1 (f (yti ) − f ( yti ))xti ,ti+1 on some partitions
Π = { ti 0  t1  . . .  tk  1 } whose meshes go to 0. We follow the proof
of Proposition 2.1 and we set for all s  u  t,



εs,u,t = |(f (ys ) − f (
ys ))xs,t − f (ys ) − f (ys ) xs,u − f (yu ) − f (yu ) xu,t |

 
 
=  f (ys ) − f (yu ) − f (ys ) − f (yu ) xu,t 
 1


  ∇f ys + τ ( ys − ys ) · (
ys − ys ) dτ
0
 1 

 
− ∇f yu + τ ( yu − yu ) · ( yu − yu ) dτ  ω(s, t)1/p .
0

But, as ∇f is α-Hölder continuous, there exists some constant C such that


εs,u,t  ω(s, t)1/p ∇f ∞ (
ys − ys − yu + yu )
 1 
+ C|yu − yu | |ys − yu + τ (
ys − ys − yu + yu )|α dτ
0

 ω(s, t)1/p
∇f ∞ γ 1/p ω(s, t)1/p + Cγ 1/p ω(0, u)1/p ω(s, u)α/p

+ Cγ (1+α)/p ω(0, u)1/p ω(s, u)α/p
 ω(s, t)(1+α)/p γ 1/p (C1 + C2 γ α/p )
where C1 and C2 depend only on f , ω, α and p. We have also remarked that γ
may be chosen smaller than 2p , so εs,u,t  ω(s, t)(1+α)/p γ 1/p (C1 + 2α C2 ).
Following the proof of Proposition 2.1, for all (s, t) ∈ ∆+ , there exist some
constant K depending only on f , ω, α and p such that
 An Introduction to Rough Paths 19

 
Is,t (y, x) − Is,t ( ys ) xs,t   Kγ 1/p ω(s, t)(1+α)/p . (2.10)
y , x) − f (ys ) − f (

On the other hand,


 
ys ) xs,t   ω(s, t)1/p ∇f ∞ |y0,s − y0,s |
 f (ys ) − f (
 ω(s, t)1/p ω(0, s)1/p γ 1/p ∇f ∞ . (2.11)

From (2.10) and (2.11), one can select a time T small enough depending on
α, p, f and ω such that
γ 1/p
|Is,t (y, x) − Is,t (
y , x)|  ω(s, t)1/p .
2
In other words, I(y, x) − I( y , x) is controlled by 2−p γω on [0, T ].
If both y and y are solutions to (2.9), then I(y, x) − I( y , x) = y − y. So,
iterating the procedure, one deduces that y − y is controlled by 2−np ω on
the time interval [0, T ] for each integer n. This proves that y = y on [0, T ].
Similarly, it is possible to construct iteratively a finite sequence of increasing
times Tk for k = 1, . . . , n with T1 = 0, T2 = T and such that Tn = 1 and
y = y on [Tk , Tk+1 ] as soon as yTk = yTk . For that, these times are constructed
so that ω(Tk , Tk+1 ) is smaller than a given constant c small enough, which
explains why this set is finite. We deduce that the solution of (2.9) is unique
on [0, 1].
Step 3: Continuity. Denote by If,y0 the map which at x gives the solution y
to (2.9) with the given initial condition y0 .
For a given y 0 , one may iteratively construct for each integer n
1 a
path y n by setting y n = I(y n−1 , x). In Step 1, we have seen that (y n )n∈N
admits a convergent subsequence, and in Step 2, under stronger hypotheses
on f , that the limit If,y0 (x) of (y n )n∈N is unique. Furthermore, if y 0 , y 1 ,
y 1 − y 0 and x are of finite p-variation controlled by ω, y n − y n−1 are of finite
p-variation controlled by 2−(n−1)p ω. So, If,y0 (x) − y n is of finite p-variation
controlled by 2−(n−2)p ω.
Now, consider two paths x and x  both of finite p-variation controlled by
ω, and such that x − x  is of finite p-variation controlled by εω for some ε > 0.
Let (y n )n∈N and ( y n )n∈N be two sequences of functions of finite p-variation
controlled by ω with y 0 = y0 and constructed by setting y n = I(y n−1 , x) and
yn = I( ). From Proposition 2.2 it is clear that for each n
0, there
y n−1 , x
exists a function ϕn (ε) converging to 0 with ε such that y n − yn is of finite
p-variation controlled by ϕn (ε)ω. But y − y n and y − yn are both of finite
p-variation controlled by 2−(n−1)p ω.
Thus, for all η > 0, there exists n0 large enough so that both If,y0 (x) − y n
and If,y0 ( x) − yn is controlled by ηω for all n
n0 . Besides, if ε is small
enough and x − x  is controlled by εω, then y n0 − yn0 is controlled by ηω.
This means that for ε small enough, If,y0 (x) − If,y0 ( x) is controlled by 3p ηω,
if ε is also chosen smaller than η. With Lemma 3.1, this means that If,y0 is
continuous from G p (Rn ) to G p (Rm ). 

20 Antoine Lejay

Remark 2.3. The previous proof is slightly different from the original proof
of [32], where f was required to be differentiable with a α-Hölder continuous
derivative to prove the existence of a solution.

3 Integration of one-form along trajectories of finite

p-variation with 2  p < 3
In this section, we consider the case of a path x with finite p-variation, with
2  p < 3.

3.1 Second iterated integrals

Let x be a piecewise smooth function from [0, 1] to Rd . For i, j = 1, . . . , d and

(s, t) ∈ ∆+ , set
  t  t
dxir1 dxjr2 = (xir1 − xis ) dxjr1 = (xir1 − xis )(xj )r1 dr1 .
s<r1 <r2 <t s s

dLet e1i, . . . , ed be the canonical basis of V = R , which implies that

d
xt =
t
i=1 ei xt for all t ∈ [0, 1]. In order to simplify expressions, define s dx ⊗ dx
as an element of V ⊗ V by
 t

d 
dx ⊗ dx = ei ⊗ ej dxir1 dxjr2 .
s i,j=1 s<r1 <r2 <t

Remark that for all 0  s  u  t  1,

 t  u  t
dx ⊗ dx = dx ⊗ dx + dx ⊗ dx + (xu − xs ) ⊗ (xt − xu ). (3.1)
s s u

The space V ⊗ V is equipped with a norm ·V⊗V , also denoted by | · |,

such that x ⊗ yV⊗V  xV yV .

3.2 Estimating the error in the approximated Chasles’ relation for

an irregular control

We assume still that x is piecewise smooth. However, the only information we

want to use is that x is continuous, and the following assumptions on x: there
exists a function ω satisfying Assumption 2.1 and a real number p ∈ [2, 3)
such that for all (s, t) ∈ ∆+ ,

|xt − xs |p  ω(s, t), (3.2a)

 p/2
 t 
 dx ⊗ dx  ω(s, t). (3.2b)

s
 An Introduction to Rough Paths 21

Assume also that f : Rd → Rd is bounded, with a bounded derivative. This

derivative is α-Hölder continuous, with α such that 2 + α > p.
Define the p-variation on [s, t] of a function y from ∆+ to Rd by
k−1 1/p


Var (y) = sup |yti ,ti+1 | p
.
p,[s,t] Π={ t0 ,...,tk } i=0
partition of [s,t]

When ys,t = yt − ys , this definition is the same as the one of (2.7).

Under these conditions (3.2a)-(3.2b),
 t 
Var (x) < +∞ and Var dx ⊗ dx < +∞.
p,[0,1] p/2,[0,1] s
t
However, by (3.1), this does not imply that the map t → 0 dx ⊗ dx has finite
p/2-variation.
t
Our goal is now to define a “good approximation” ys,t of s f (xr ) dxr , so
that this integral will be a limit of Riemann sums:
 t
δ
k
−1
Πδ Πδ
f (xr ) dxr = lim zs,t with zs,t = ytδi ,tδi+1
s δ→0
i=0

for a partition Π δ = tδ0 , tδ1 , . . . , tδkδ of [s, t] whose mesh goes to zero with δ.
In the proof of Proposition 2.1, we have seen that one can consider the limit
Πδ
of zs,t , provided that one has a control of the form

|ys,t − ys,u − yu,t |  Cω(s, t)θ for all 0  s  u  t  1,

for some constant C and some θ > 1. The fact that θ > 1 iscrucial, since the
proof of Proposition 2.1 involves the Zeta function ζ(θ) = n
1 1/nθ .
In Lemma 2.1, we used the fact that f is α-Hölder continuous, that α > p
and that |xt − xs |p  ω(t, s). With only (3.2a) if p
2, this no longer works
even if f has a bounded derivative, i.e., α = 1.
We are then forced to use a better estimate. If x is smooth, then for
i = 1, . . . , d,

d  t
∂f i
f i (xt ) = f i (xs ) + (xr ) dxjr ,
i=1 s ∂xj

and then
 t  t d  t

r1
∂f i
f i (xr1 ) dxir1 = f i (xs ) dxir1 + (xr2 ) dxjr2 dxir1 .
s s j=1 s s ∂xj

For any y in V = Rd , denote by ∇f (y) be the bilinear form defined on V⊗V by

22 Antoine Lejay

∂f j
∇f (y), ei ⊗ ej  = (y).
∂xi
t
A first approximation of the integral f (xr ) dxr will be given by
s
 t
ys,t = f (xs )(xt − xs ) + ∇f (xs ) dx ⊗ dx. (3.3)
s

With (3.2a) and (3.2b), |ys,t |p  N (f )ω(0, 1)ω(s, t), where


N (f ) = inf M
0  f ∞  M, ∇f ∞  M

sup |∇f (x) − ∇f (y)|/|x − y|α  M . (3.4)
x=y

So, y is of finite p-variation.

Lemma 3.1. For all 0  s  u  t,
|ys,t − ys,u − yu,t |  2N (f )ω(s, t)θ ,
with θ = (2 + α)/p > 1.
Proof. Let a and b be two points of Rd . Then,

d  1
∂f i
f i (b) = f i (a) + (a + (b − a)r)(bj − aj ) dr
j=1 0
∂xj

d
∂f i
= f i (a) + (a)(bj − aj ) + Ri (a, b) (3.5)
j=1
∂xj

with
 

d  1  
 ∂f i ∂f i
 
|R (a, b)| = 
i
(a) − a + (b − a)r (b − a ) dr
j j
 0 ∂xj ∂xj  (3.6)
j=1
1+α
 N (f ) b − a
since the derivative of f is α-Hölder continuous (the quantity N (f ) has been
defined in (3.4)). Set R = (R1 , . . . , Rd ).
Using (3.5) and (3.1),
  

 t 

|ys,t − ys,u − yu,t |   ∇f (xu ) − ∇f (xs ) dx ⊗ dx
u
+ |R(xu , xs )(xt − xu )|.
t
As ∇f is α-Hölder continuous, x satisfies (3.2a) and s dx⊗dx satisfies (3.2b),
2+α
|ys,t − ys,u − yu,t |  Cω(s, t)θ with θ = > 1.
p
The Lemma is then proved. 

 An Introduction to Rough Paths 23

3.3 Geometric multiplicative functionals

One may re-use the proof of Proposition 2.1 exactly the same way with ys,t
defined by (3.3), and not by (2.3). As we saw in the proof of Proposition 2.1
or in the proof of Lemma 3.1, the  tsmoothness of x plays no role.
However, if x is not smooth, s dx ⊗ dx has to be defined. In fact, there is
no general procedure to construct this term. However, for some particular x,
such as the trajectories of some stochastic process, this is possible, but may be
rather technical (see Sect. 12 for examples of stochastic processes for which the
second order iterated integral has been defined). In the following, we assume
that the second order  iterated integral exists. But the path we consider is not
2
x but the couple (x, dx ⊗ dx), which no longer lives in Rd , but in Rd+d and
whose components satisfy some algebraic relations.
With this end in view, consider xs,t = (x1s,t , x2s,t ) defined for (s, t) ∈ ∆+ ,
such that there exists a function ω satisfying Assumption 2.1 and a constant
p ∈ [2, 3) for which:
x1s,t ∈ V and x1s,t = xt − xs , (3.7a)
|x1s,t |p  ω(s, t), (3.7b)
x2s,t ∈ V ⊗ V and x2s,t = x2s,u + x2u,t + x1s,u ⊗ x1u,t , (3.7c)
|x2s,t |p/2  ω(s, t) (3.7d)
for all 0  s  u  t  1. Such a x is called a multiplicative functional with
p-variation controlled by ω. Condition (3.7a) means that x1 may be identified
with the path x. In this case, we say that x lies above x. Condition (3.7b)
means that x has finite p-variation, and is (3.2a). Condition (3.7c) is equivalent
to (3.1), while (3.7d) is analogue to (3.2b).
t
When x is piecewise smooth, set x1s,t = xt − xs , and x2s,t = s dx ⊗
dx, and (3.7a)-(3.7d) are clearly satisfied. Denote by S 2 (V) the set of such
multiplicative functionals.
The distance δp,[s,t] is extended to p ∈ [2, 3) by

δp,[s,t] (x, y) = Var (x1 − y1 ) + Var (x2 − y2 ).

p,[s,t] p/2,[s,t]

Denote still by V the topology it generates on the space of multiplicative

p

functionals of finite p-variation.

We restrict ourself to multiplicative functionals which may be approxi-
mated by some elements in S 2 (V), where S 2 (V) is the set of multiplicative
t
functionals x such that x1s,t = xt − ss and x2s,t = s dx ⊗ dx for a piecewise
smooth path x. Define G p (V) as the set a multiplicative functionals satisfying
(3.7a)-(3.7d) for a given control ω, and such that
x may be approximated in V p by elements in S 2 (V). (3.7e)
In this case, x is said to be a geometric multiplicative functional. In fact, (3.7e)
is not really necessary in the case 2  p < 3, but provides us with an intuitive
24 Antoine Lejay

view of integral driven by rough paths. This issue is discussed in Sect. 10.3
for Brownian motion.

Remark 3.1. If x belongs to G p (V), and c = (ci,j )i,j=1,...,d is an antisymmetric

matrix, i.e., ci,j = −cj,i , seen as an element on V ⊗ V, then the function
(s, t) ∈ ∆+ → xs,t + c(t − s) is also an element of G p (V). We give further
explanations in Sects. 6.2 and 10.2.

3.4 Integration of a one-form

In the previous sections, we have given all the elements to construct the in-
tegral of a differential one-form along a path x of finite p-variations with
p ∈ [2, 3), given that one also knows a geometric multiplicative x lying above x.
Once Lemma 3.1 has been proved, then one could use the same machinery
as in the proof of Proposition 2.1, to prove that
δ
k
−1
zs,t = lim ytδi ,tδi+1 (3.8)
δ→0
i=0

exists and is unique for all partition tδ0 , . . . , tδkδ of [s, t] when y is given by
(3.3).
Proposition 3.1 below summarizes this result. However, we will give in the
next section a more complete construction of the integral of a one-form along
the path x. In this new definition, the integral belongs to the set of geometric
multiplicative functionals G p (W). This means that this integral could also be
used as a path along which a another differential one-form is integrated.

Proposition 3.1. Let x be an element in G p (V) lying above a continuous

path x for p ∈ [2, 3). Let f be a continuous, bounded function with α-Hölder
continuous, bounded derivatives for α > p−2.  Then, for all 0  s  t  1 and
any family of partitions Π δ = tδ0 , . . . , tδkδ of [s, t] whose meshes decrease
to 0 as δ → 0, the limit zs,t defined in (3.8) exists and is unique, when
ys,t = f (xs )x1s,t + ∇f (xs )x2s,t . The limit zs,t , which does not depend on the
t
partitions Π δ , is denoted by s f (xr )dxr , and is of finite p-variation. Finally,
for all 0  s  u  t  1, zs,t = zs,u + zu,t .

We also have the equivalent of the continuity result of Proposition 2.2,

assuming that x2 and x2 are close enough in the sense given in Proposition 2.2,
where p is replaced by p/2.

t
3.5 The iterated integrals of s f (xr ) dxr
d
Let us consider the differential form f (x) = i=1 fi (x)dxi , where the f i ’s are
t
functions from V = Rd into W = Rm . The integral z1s,t = s f (xr ) dxr takes
 An Introduction to Rough Paths 25

its value in W, and is of finite p-variation. However, to construct the solution

of a differential equation of the type
 t
yt = ys + f (yr ) dxr ,
s

one may first define the integration of one-forms, and then use Picard’s it-
eration principle. However, integrating with respect to a control of finite p-
variation with p ∈ [2, 3) requires an element in G p (V). So, to use a fixed point
t
theorem, we need to construct s f (xr ) dxr not only as an element of W, but
also as an element of G p (W).
To this end, set
1
ys,t = f (xs )x1s,t + ∇f (xs )x2s,t ∈ W, (3.9a)
2
ys,t = f (xs ) ⊗ f (xs ) · x2s,t ∈ W ⊗ W, (3.9b)
1 2
and ys,t = (ys,t , ys,t ). In the definition of y2 , we used a shorthand, which
means in fact that


d
2
ys,t = f i (xs ) ⊗ f j (xs )x2,i,j
s,t .
i,j=1
t
Denote by z1s,t the element of W given by s f (xr ) dxr .
Let 1 denote an element of a one-dimensional space. We use the following
computation rule: If y belongs to W⊗k for some integer k = 1, 2, then 1 ⊗ y =
y ⊗ 1 = y ∈ W⊗k . If y and z belong to W, then y ⊗ z belongs to W⊗2 .
If y belongs to W and z belongs to W⊗2 , then y ⊗ z = z ⊗ y = 0. Set
T2 (W) = 1 ⊕ W ⊕ W⊗2 . By the definition of the tensor product, if x, y and z
belong to T2 (W), then for all α, β ∈ R, (αx + βy) ⊗ z = αx ⊗ z + βy ⊗ z and
z ⊗ (αx + βy) = αz ⊗ x + βx ⊗ y.
Let Π = { ti t0  · · ·  t } be a partition of [s, t]. Set

s,t = (1 + zt0 ,t1 + yt0 ,t1 ) ⊗ · · · ⊗ (1 + zt−1 ,t + yt−1 ,t ).

zΠ 1 2 1 2

The computation rules previously given mean that we keep only the elements
in T2 (W), and not those in W⊗k for k > 2. From Proposition 3.1, the projec-
tion zΠ,1 of zΠ ∈ T2 (W) on W is equal to z1s,t .
The proof that zΠ has a limit when the mesh of the partition Π decreases
to 0 is similar to the proof of Proposition 2.1. But we have also to estimate
2 2 2
the “error” when ys,t is split into ys,u and yu,t .
Lemma 3.2. For all 0  s  u  t  1, set
2
ε(s, u, t) = ys,t − ys,u
2
− yu,t
2
− ys,u
1
⊗ yu,t
1
.
There exists some constant C depending only on N (f ), ω(0, 1) and α such that
|ε(s, u, t)|  Cω(s, t)θ
with θ = (2 + α)/p > 1.
26 Antoine Lejay

Proof. Recall that xt = x10,t . Using (3.7c) and the relation

1
ys,u ⊗ yu,t
1
= f (xs )x1s,u ⊗ f (xu )x1u,t
= f (xs ) ⊗ f (xs ) · x1s,u ⊗ x1u,t + f (xs ) ⊗ (f (xu ) − f (xs )) · x1s,u ⊗ x1u,t ,

we obtain that


ε(s, u, t) = f (xs ) ⊗ f (xs ) − f (xu ) ⊗ f (xu ) · x2s,t


− f (xs ) ⊗ f (xu ) − f (xs ) · x1s,u ⊗ x1u,t .

But

f (xs ) ⊗ f (xs ) − f (xu ) ⊗ f (xu )



= f (xs ) − f (xu ) ⊗ f (xs ) + f (xu ) ⊗ f (xs ) − f (xu ) .

Using the relation f (xu ) − f (xs ) = ∇f (xs )(xu − xs ) + R(xu , xs ) together

with (3.6), the boundedness of f and ∇f , we obtain that

|f (xs ) − f (xu ) ⊗ f (xs )|  2N (f )2 (|xu − xs | + |xu − xs |1+α )


 2N (f )2 ω(s, t)1/p + ω(s, t)(1+α)/p .

Moreover, |x1s,u ⊗ x1u,t |  |x1s,u | · |x1u,t |  ω(s, t)2/p . The Lemma is now easily
proved by combining all the previous estimates. 

Proposition 3.2. As the mesh of Π decreases to 0, zΠ s,t admits a limit, de-

t
noted by zs,t and by s f (xr ) dxr . This limit is of finite p-variation.

Proof. Assume that Π ∩(s, t) has more than one element. Let tk be an element
of Π ∩ (s, t) such that ω(tk−1 , tk+1 )  2ω(s, t)/|Π ∩ (s, t)| (see Lemma 2.2).
We use the convention that tk−1 = s if Π ∩ (s, tk ) = ∅, and that tk+1 = t is
Π ∩ (tk , t) = ∅. Using the computations’ rules on 1 ⊕ W ⊕ W⊗2 provided in
Sect. 3.5, one has

(1 + z1tk−1 ,tk +yt2k−1 ,tk ) ⊗ (1 + z1tk ,tk+1 + yt2k ,tk+1 )

=1 + z1tk−1 ,tk+1 + yt2k−1 ,tk + yt2k ,tk+1 + z1tk−1 ,tk ⊗ z1tk ,tk+1
=1 + z1tk−1 ,tk+1 + yt2k−1 ,tk+1 − ε(tk−1 , tk , tk+1 )
+ z1tk−1 ,tk ⊗ z1tk ,tk+1 − yt1k−1 ,tk ⊗ yt1k ,tk+1 .

Set

δk = z1tk−1 ,tk ⊗ z1tk ,tk+1 − yt1k−1 ,tk ⊗ yt1k ,tk+1

= (z1tk−1 ,tk − yt1k−1 ,tk ) ⊗ z1tk ,tk+1 − yt1k−1 ,tk ⊗ (yt1k ,tk+1 − z1tk ,tk+1 ).

In Proposition 3.1, as in Proposition 2.1, for all (r, u) ∈ ∆+ ,

 An Introduction to Rough Paths 27

|z1r,u − yr,u
1
|  Kω(r, u)(2+α)/p

for some constant K depending on f , α and p. Besides, |yr,u 1

|  Cω(r, u)1/p +


Cω(r, u) , where C is f ∞ + ∇f ∞ . So, |δk |  C ω(s, t)θ , where the
2/p

constants θ > 1 and C  depend only on f , α and p. Thus, if Π  = Π \ { tk }.

 Π 
zs,t − zΠ   θ 
s,t  |ε(tk−1 , tk , tk+1 )| + |δk |  C ω(s, t) ,

where the constants θ > 1 and C  depend only on f , α and p. The end of
the proof is similar to the one of Proposition 2.1.
For the uniqueness of the limit, remark that if z and  z are two multiplica-
tive functionals of finite p-variation such that z1s,t = 
z1s,t for all (s, t) ∈ ∆+ ,
then
ψ(s, t) = z2s,t − 
z2s,t
is additive, i.e., ψ(s, u) + ψ(u, t) = ψ(s, t) for all 0  s  u  t  1. Let z
δ
and  z be two cluster points of (zΠ )δ>0 for a family (Π δ )δ>0 of partitions of
[0, 1]. By construction, z1 = z1 . Moreover, for all integer n
1,

 2

 n−1 
zs,t − 
z2s,t   z2tn ,tn z2tni ,tni+1   2Kω(s, t)
− sup ω(tni , tni+1 )θ−1 ,
i i+1
i=1 i=1,...,n−1

where tni = s + i(t − s)/n. Since ω is continuous near its diagonal, letting n
increase to infinity proves that z2 = 
z2 , and the limit is unique.


Corollary 3.1. The map
 t 
x ∈ G p (V) −→ f (x0 + x10,r ) dxr ; (s, t) ∈ ∆+ ∈ G p (W)
s

is continuous with respect to δp .

It is because this map is continuous that, in view of (3.7e), the integral
belongs to G p (W).
The proof of Corollary 3.1 is similar to the one of Proposition 2.2 and
Corollary 2.1, although a bit more complicated, since the number of terms to
consider is more important.

4 A faithful representation of paths

We have explained in the previous sections how to construct solutions of
differential equations controlled by path of finite p-variations with p < 3. We
have also constructed defined the integration along such irregular paths. We
have seen that the “iterated integrals” appear naturally for defining our new
objects. Our article is now devoted to provide the construction of these objects
for all real number p.
28 Antoine Lejay

In this section and the next one, we consider iterated integrals of piecewise
smooth paths. We present some results, mainly due to K.-T. Chen (see [6]
and related articles), which allows to perform some manipulations on smooth
paths which could be expressed using algebraic computations. These results
provides us with a very powerful tool. The first main result expresses that a
piecewise smooth path x can be uniquely defined by a power series involving
its iterated integrals.

4.1 The Chen series

Let x : [0, t] → V = Rd be a piecewise smooth path. We shall assume that for

all s ∈ (0, t), there exists no ε > 0 such that x([s − ε, s]) = x([s, s + ε]). Such
a path is called irreducible. Let I = (i1 , . . . , ik ) be a multi-index. Denote by
t
0 dI x the iterated integral
 t 
dI x = dxis11 · · · dxiskk .
0 0<si1 <···<sik <t

Introduce some indeterminates X 1 , . . . , X d . Each of these indeterminates cor-

responds to a direction in the space Rd . Thus, X i may be identified with the
vector ei of the canonical basis of Rd . For a multi-index I = (i1 , . . . , ik ), set
X I = X i1 · · · X ik .
The non-commutative power series with indeterminates X i for i = 1, . . . , d
is given by  t


I
Φ([0, t], x) = X dI x
I multi-index 0

provides a faithful representation of x, and Φ is called a Chen series.

Theorem 4.1 (K.-T. Chen). If Φ([0, t], x) = Φ([0, t], y) for two paths x
and y, then x = y on [0, t] up to a translation.

In facts, manipulation on paths may be considered as manipulation on

power series, as we will see.

4.2 Concatenating two paths

If z denotes the path obtained by the concatenation of two paths x : [0, t] → Rd

and y : [t, s] → Rd such that x(t) = y(t), then

Φ([0, s], z) = Φ([0, t], x)Φ([t, s], y). (4.1)

It means that the power series corresponding to z is equal to the formal

non-commutative product of the two power series corresponding to x and y.
Furthermore, if x̂(s) = x(t − s), then Φ([0, t], x̂) = Φ([0, t], x)−1 .
 An Introduction to Rough Paths 29

4.3 Products of iterated integrals

Another interesting property arises when one considers the product of two
iterated integrals:
  
i i 
Ψs,t = dxiu11 . . . dxiujj dxu11 . . . dxujj .
s<u1 <···<uj <t s<u1 <···<uj <t

If I = (i1 , . . . , ij ) and I  = (i1 , . . . , ij  ) are two multi-indices, denote by I  J

the shuffle product of I and J. Then

 
i i
j 
1
Ψs,t = dxu1 . . . dxuj .
K=(i  s<u1 <···<uj <t
1 ,...,ij  ) multi-index ∈IJ

The shuffle product I  J of I and J is the set of all multi-indexes of length

length(I) + length(J) such that for each K in I  J, the elements of K cor-
respond either to the elements of I or J, and I (resp. J) is recovered if the
elements of I  J belonging to I (resp. J) are kept regardless the elements of
J (resp. I).

5 Lie algebra and enveloping algebra

In this section, we continue to manipulate piecewise smooth paths and their
Chen series. The main result of this section is Proposition 5.2, which asserts
that the Chen series Φ([s, t], x) could also be expressed as
  t 


Φ([s, t], x) = exp ΘI dI x ,
I multi-index s

where the coefficients ΘI belong to a particular subspace of the space con-

taining the X I ’s for all multi-index I.

5.1 Enveloping algebra

We present now some aspects of Lie algebras and enveloping algebras. The
relation with Φ([0, t], x) is developed in Sect. 5.3. On this topic, see for example
[40, Chap. 1].
Let A = { a1 , . . . , an } be some letters. In Sect. 5.3, these letters will be
identified with the indeterminates X i . The letters may be used to construct
some words ai1 · · · aik for some multi-index I = (i1 , . . . , ik ) of length k. The
set of words with letters in A for which an empty word 1 is added is denoted
by A∗ . Let K be a ring containing Q. A non-commutative polynomial P is a
linear combination over K of words on A :
30 Antoine Lejay


P = Pw w, Pw ∈ K with only a finite number of terms.
w∈A∗

The set of non-commutative polynomials is denoted by KA. Similarly, we

may define formal series in the same way, but the number of terms in the
sum is countable. The set of formal power series is denoted by KA.
With A may be constructed a Lie algebra with bracket [a, b] = ab − ba for
all a, b ∈ A. This Lie bracket may be extended to the set all non-commutative
polynomials P . The set KA is closed under [·, ·] and corresponds to the Lie
algebra generated by A.
Given this Lie algebra L, it is known that there a unique associative al-
gebra E(L) and a Lie algebra homomorphism ϕ0 : L → E(L) such that for
all other associative algebra B and any Lie algebra homomorphism ϕ (i.e.,
[ϕ(a), ϕ(b)] = ϕ([a, b]) for all a, b ∈ L) from L to B, there exists a unique
algebra homomorphism f : E(L) → B such that ϕ = f ◦ ϕ0 . The associative
algebra E(L) is called the enveloping algebra. So, any algebra homomorphism
from L into some associative algebra B may be extended to an algebra homo-
morphism from E(L) into B.
Denote by LK (A) the smallest submodule of KA containing A and closed
under the Lie bracket.

Proposition 5.1. The algebra KA is the enveloping algebra of LK (A).

Given a Lie algebra L, its enveloping algebra E(L) may be constructed by

quotienting the tensor algebra

T (L) = L⊗n
n
0

by the ideal I generated by the elements of the form x ⊗ y − y ⊗ x − [x, y] for

x, y ∈ L.

5.2 A characterization of the Lie polynomials

An element of L
K (A) is called a Lie polynomial. A formal series S that may
be written S = n
0 Sn where the Sn ’s are Lie polynomials is called a Lie
series.

Remark 5.1. To understand the difference between a Lie polynomial and an

element of KA, consider the following example: If the elements of a tangent
space of a manifold are seen a first-order differential operators, then the Lie
brackets [x, y] of two of them remains a first-order differential operator and
then belongs to the tangent space. If the letters ai are identified with vectors
of a tangent space, a Lie polynomial belongs (formally) to the tangent space.
Yet an element of the enveloping algebra is a general differential operator,
with terms that could be of any order.
 An Introduction to Rough Paths 31

Let δ be the algebra homomorphism from KA → KA × KA

defined by δ(a) = a ⊗ 1 + 1 ⊗ a for all a ∈ A. The existence of δ as an algebra
homomorphism is given by the property of the enveloping algebra.
Theorem 5.1. An element P in KA is a Lie series if and only δ(S) =
S ⊗ 1 + 1 ⊗ S. Furthermore, if S is a Lie series, then the constant term S1
is equal to 0, where 1 is the empty word of A∗ .
From the condition on δ(P ), the Lie polynomials are also called primitive.
For a formal series S with zero constant term (S1 = 0), the exponential is
defined to be

Sn
exp(S) = .
n!
n
0

For a formal series of the form S = 1 + T where T have a zero constant term,
the logarithm is defined to be

(−1)n−1
log(S) = log(1 + T ) = T n.
n
n
1

Theorem 5.2. The following properties hold:

(i) Let S be a series in KA with a constant term equal to 1. Then log(S)
is a Lie series if and only if δ(S) = S ⊗ S.
(ii) The set of series in KA with a constant term equal to 1 such that
log(S) is a Lie series is a group under multiplication.
An element in the group generated by exp(S) where S is a Lie series with
a constant term S0 equal to 0 is called group-like. Theorems 5.1 and 5.2 assert
that the map log is a bijection from group-like elements to primitive elements.

5.3 The series of iterated integrals is a group-like element

Now, take as alphabet A = (e1 , . . . , ed ), where e1 , . . . , ed is the canonical basis
of V = Rd . The ring K is R. The product of two words a and b is replaced
by the tensor product a ⊗ b of a and b. Each of the ei ’s corresponds to the
indeterminates X i , and this identification is used. That is,

 t
Φ([s, t], x) = ei1 ⊗ · · · ⊗ eik dI x.
I multi-index s
I=(i1 ,...,ik )

Thus, the series Φ([0, t], x), which belongs to KA, may be seen as an
element of R ⊕ V ⊕ V⊗2 ⊕ · · · . More precisely, for all integer k, the element
 t ⊗k
I=(i1 ,...,ik ) ei1 ⊗ · · · ⊗ eik s dI x belongs to V . In Sect. 4.2, we have seen
that the concatenation of x : [0, s] → V and x : [s, t] → V creates a new paths
x : [0, t] → V characterized by the series Φ([0, t], x) given by the product of
Φ([0, s], x) and Φ([s, t], x). With our convention, (4.1) is rewritten Φ([0, t], x) =
Φ([0, s], x) ⊗ Φ([s, t], x).
32 Antoine Lejay

Remark 5.2 (Another notation for the iterated integrals). If e1 , . . . , ed is the
canonical basis, identified
t with e1 , . . . , ed , of the dual V∗ of V, then define a
multi-linear form s dx ⊗ · · · ⊗ dx on V∗ × · · · × V∗ by
 t  
 
dx ⊗ · · · ⊗ dx, (ei1 , . . . , eik ) = dxis11 · · · dxiskk .
s s<s1 <···<sk <t
t
Thus, s dx ⊗ · · · ⊗ dx is an element of the dual of V∗ × · · · × V∗ , which is
t
identified with V⊗k , and s dx ⊗ · · · ⊗ dx is identified with

 t
ei1 ⊗ · · · ⊗ eik dI x.
s
I=(i1 ,...,ik )∈{ 1,...,d }k

The following Proposition may be proved using the properties of the shuffle
product and the Campbell–Hausdorff formula, and links the series constructed
in Sect. 4 with our constructions of objects related to Lie algebras.
Proposition 5.2 ([5]). For any irreducible, piecewise smooth path x : [0, t] →
Rd , Φ([0, t], x) is a group-like element when the elements ei of the basis of Rd
are identified with the indeterminates X i . Moreover,

 t
log Φ([0, t], x) = ΘI dI x, (5.1)
I multi-index s

where ΘI belongs to LK (A) = 0 ⊕ V ⊕ [V, V] ⊕ [V, [V, V]] ⊕ · · · and does not
involve more that length(I) Lie brackets.
Remark 5.3. There are very nice algebraic properties that can be considered
on the series of type Φ([0, t], x). In particular, two structures of bi-algebra may
be considered, one corresponding to concatenation of paths, the other one to
product of the series and then using shuffle products (see for example [40,
Sect. 1]). See also [47] for an example of use of the Poincaré-Birkhoff-Witt
Theorem that gives a basis of Lie algebras.

6 (Geometric) multiplicative functionals

We have already encountered geometric multiplicative functionals in Sect. 3.3.
In this section, we give a definition of geometric multiplicative functionals of
any order.
Roughly speaking, a geometric multiplicative functional

x = (1, x1s,t = xt − xs , x2s,t , . . . , xks,t )(s,t)∈∆+

lying above a path x corresponds to x together with its first “iterated inte-
grals”, and such that the iterated integrals xδ, = dxδ ⊗· · ·⊗dxδ of piecewise
smooth approximations xδ of x converge to x for  = 2, . . . , k.
 An Introduction to Rough Paths 33

If x is a piecewise smooth path, we have seen that one can construct its
Chen series Φ([0, t], x), which fully characterize x. Moreover, given a Chen
series of a path x, one could formally solve a differential equation controlled
by x or consider the integral of a one-form along the path x, by writing these
objects with the Chen series of x (For example, see Sect. 6.3 below).
If x is irregular, then knowing x is not sufficient to define its iterated
integrals (see Remark 3.1 and Sect. 6.2 for example). However, when one
knows a (geometric) multiplicative functional x = (1, x1 , x2 , . . . , xk ) lying
above a path x of finite p-variation with k = p, then it will be proved that
there exists a procedure to extend x in a (geometric) multiplicative functional
(1, x1 , x2 , . . . ), and that this extension has some nice properties, especially
with respect to the topology of generated by the norm of p-variation. Thus,
one can construct an extension of the notion of Chen series for irregular paths,
provided enough information is known on the path, i.e., its first “iterated in-
tegrals”. And, in view of the results of Sects. 2 and 3, one will not be surprised
by the results of Sects. 8, where integrals of one-form along irregular paths are
constructed, and 9, where differential equations controlled by irregular paths
are solved.
For any integer k
1, set

Tk (V) = R ⊕ V ⊕ V⊗2 ⊕ · · · ⊕ V⊗k ,

which is a truncated tensor algebra. Let also Ak (V) ⊂ Tk (V) containing all
the elements of A(V) = 0 ⊕ V ⊕ [V, V] ⊕ [V, [V, V]] ⊕ . . . , where all the terms
involving more than k Lie brackets are set to 0. Similarly, computations on
Tk (V) are done by setting to 0 all tensor products involving more than k
terms.
The norm we choose on V⊗k is such that

x1 ⊗ · · · ⊗ xk V⊗k  |x1 | × · · · × |xk |.

This norm is also denoted by | · |, and there are different possibilities for
constructing such a norm (see for example [41]).

Definition 6.1. A multiplicative functional x = (x0 , x1 , . . . , xk ) of order k

is a function from ∆+ into Tk (V) such that

x : ∆+ → Tk (V) is continuous, (6.1a)

xs,t = xs,u ⊗ xu,t for all 0  s  u  t  1. (6.1b)

Furthermore, x is said to be geometric if

log xs,t ∈ Ak (V), (6.1c)

and x0 = 1.
34 Antoine Lejay

Of course, in the previous definition, xi denotes the projection of x on V⊗i .

The set of multiplicative functionals with values in Tk (V) is denoted
by Tk (V). The subset of Tp (V) of of multiplicative functionals of finite p-
variation is denoted by Mp (V). Let us also denote by G p (V) ⊂ Mp (V) the set
of geometric multiplicative functionals of finite p-variation taking their values
in Tp (V).
Clearly, for a piecewise smooth path, xs,t = Φ([s, t], x) is a geometric
t
multiplicative functional, and xis,t = s dx ⊗ · · · ⊗ dx. Denote by S k (V) the
set of geometric multiplicative functionals in Tk (V) lying above a piecewise
smooth path given by the projection of Φ(x) on Tk (V).
Definition 6.1 extends the one given previously by (3.7a)-(3.7e), and (6.1c)
replaces (3.7e). We will see below in Proposition 6.1 that these conditions are
equivalent, provided that Mp (V) is equipped with the good norm.

6.1 A norm on multiplicative functionals

We use for multiplicative functionals in Tk (V) the norm

1/p
xp,[s,t] = max βΓ (i/p) Var (xi )1/i .
i=1,...,k p/i,[s,t]

and xp = xp,[0,1] . Here β is a positive constant, and Γ is the Gamma

function. Note that xq is finite for all q > p as soon as xp is finite.
The space Tk (V) equipped with the norm ·p is complete, but not sepa-
rable.
Remark 6.1. Note that any rough path x lies above the path x defined by
xt = x0,t . For such a path, x0 = 0, and Varp,[0,1] is a norm on this space of
functions from x : [0, 1] → V with x0 = 0, and not only a semi-norm.
Definition 6.2. We say that x in Tk (V) is of finite p-variation controlled
by ω : ∆+ → R+ (satisfying Assumption 2.1) when xpp,[s,t]  ω(s, t) for
all (s, t) ∈ ∆+ .
Lemma 6.1. A multiplicative functional x in Tk (V) is of finite p-variation
controlled by ω if and only if
ω(s, t)i/p
|xis,t |  for i = 1, . . . , k. (6.2)
βΓ (i/p)
Moreover, Lemma 2.3 also holds for multiplicative functionals: If (xn )n∈N is
a sequence of multiplicative functionals converging to x in ·p , then there
exists an ω : ∆+ → R+ satisfying Assumption 2.1 such that xn (or possibly
only for the element of a subsequence of (xn )n∈N ) and x are controlled by ω,
and for all ε > 0, there exists some n such that for all m
n,
ω(s, t)i/p
|xm,i
s,t − xs,t |  ε
i
. (6.3)
βΓ (i/p)
 An Introduction to Rough Paths 35

Let x be a smooth path, and x be its associated geometric multiplicative

t
functional, i.e., xis,t = s dx ⊗ · · · ⊗ dx. Hence, there exists a constant C such
that |xs,t |  C(t − s) /i!. Thus, the choice of the norm ·p is coherent with
i i

that fact that a smooth function is controlled by ω(s, t) = C|t − s|.

We said earlier that (6.1c) replaces (3.7e). In fact, the two hypotheses may
be seen as equivalent.

Proposition 6.1 ([32, Lemma 2.3.1, p. 259]). The space G p (V) is the closure
of S p (V) (i.e., geometric multiplicative functionals lying above piecewise
smooth trajectories) with respect to ·p .

Practically, geometric multiplicative functionals will be constructed by ap-

proximating irregular trajectories by piecewise linear functions whose iterated
integrals converge. This is why (3.7e) is generally more useful than (6.1c).

Lemma 6.2 (A convergence and compactness result). Let p be a fixed

real number. Let (xn )n∈N be a family of multiplicative functionals in Tk (V)
such that for some q < p and for i = 1, . . . , k,

∀ε > 0, ∃ η > 0 such that |s − t| < η =⇒ |xn,i

s,t | < ε, (6.4a)
n,i
sup Var (x ) < +∞. (6.4b)
n∈N q/i,[0,1]

Then there exist a subsequence (n )∈N and a multiplicative functional x

in Tk (V) such that
xn − xp −−−→ 0
→∞

In other words, (x )n∈N is relatively compact in (Tk (V), ·p ).

n

Moreover, if xn lies above a path xn and (xn0 )n∈N is relatively compact

in R, then any possible x limit of (xn )n∈N lies above a path x, which is also
a possible limit of the sequence (xn )n∈N in the space of continuous functions.

Remark 6.2. In particular, (6.4a) and (6.4b) are true if (xn0 )n∈N is bounded
and there exists a function ω satisfying Assumption 2.1 such that xn q,[s,t] 
ω(s, t) for all (s, t) ∈ ∆+ and any n ∈ N.

Remark 6.3. Thanks to the Ascoli theorem and the relation (6.1b), the con-
dition (6.4a) is also equivalent to saying that the sequences of functions
(t → xn,i
0,t )n∈N are relatively compact in the space of continuous functions
for i = 1, . . . , k. This is a fact we use in the proof of this Lemma.

Proof. If y is a multiplicative functional in Tk (V), (6.1b) implies that for all

(s, t) ∈ ∆+ and i = 1, . . . , k,


i
y0,t i
= y0,s + k
y0,s ⊗ ys,t

. (6.5)
k+=i,
1

Hence,
36 Antoine Lejay


|y0,t
i
− y0,s
i
| |y0,s
k
| · |ys,t

|
k+=i,
1

i
 C y∞ |ys,t

|  C 2 y∞ yq,[s,t] .
=1

where C depends only on q and i.

With (6.4a) and (6.5), it is clear that (t → xn,i0,t )n∈N is bounded and equi-
continuous for i = 1, . . . , k. Thus, Ascoli’s Theorem implies that there exists
2 k
a continuous function t → x0,t ∈ Rd+d +···+d such that, at least along a
subsequence, t → xn0,t converges uniformly to t → x0,t .
Using (6.5) recursively, one can construct from t → x0,t a function (s, t) ∈
∆+ → xs,t that is a multiplicative functional and such that, at least along a
subsequence, xns,t −−−−→ xs,t uniformly in s and t.
n→∞
It is now easily seen that Varp/i (xi )  lim inf n→∞ Varp/i (xn,i ) for i =
1, . . . , k. Thus, x is of finite p-variation. With a variation of (2.8), one obtains
that for all p > q, x − xn p,[s,t] goes to 0, at least along a subsequence.
The second part of the Lemma is clear from Ascoli’s Theorem. 

Corollary 6.1 (A tightness result). Let (Xn )n∈N be a family of random

variables taking their values in Tk (V). Assume that the family of stochastic
processes (Xn0,t ; t ∈ [0, 1])n∈N is tight in the space of continuous functions with
the uniform norm, and that for all ε > 0, there exists some constant C large
enough so that  
sup P Xn q > C < ε (6.6)
n∈N

for some real number q
1. Then (X)n is tight in (Tk (V), ·p ) (hence in
Mp (V) if k = p) for all p > q.

Remark 6.4. Since (Tk (V), ·p ) is not separable, a sequence (Xn )n∈N may be
tight in this space but fails to satisfy (6.6).

Remark 6.5. Owing to (6.1b), the tightness of (t → Xn,i 0,t )n∈N for all i ∈
{ 1, . . . , k } is equivalent to saying that for all ε > 0 and any C > 0, there
exists some η > 0 small enough such that
 
i,n
sup sup P sup |Xs,t | > C  ε.
n∈N i=1,...,k |t−s|<η

Proof. The proof is immediate from Lemma 6.2 and Remark 6.5, since the
subsets K of Tk (V) of the form K0 ∩ K1 for a given C > 0 are relatively
compact in (Tk (V), ·p ), where the sets of functions (t → xi0,t )x∈K0 are equi-
continuous for i = 1, . . . , k and K1 contains the multiplicative functionals such
that supx∈K1 xq < C for some q < p and a given constant C.
 An Introduction to Rough Paths 37

6.2 Back to the case 2  p < 3

To provide a better understanding of what a geometric multiplicative func-

tional could be, consider a multiplicative functional x in T2 (V) with V = Rd .
Set
1
Ss,t (x) = x1s,t ⊗ x1s,t and As,t (x) = x2s,t − Ss,t (x).
2
Thus, A(x) = (Ai,j (x))i,j=1,...,d with Ai,j 2,i,j 2,j,i
s,t (x) = 2 (xs,t − xs,t ). Remark that
1

S(x) and A(x) are respectively the symmetric part and the antisymmetric
part of x2 . Moreover, S(x) depends only on x1 , and if x1 lies above a path x,
then S(x) = S(x). Moreover, if x is of finite p-variation, the map x → S(x)
is continuous for the topology generated by Varp/2,[s,t] (·) + ·∞ .
On the other side, if x is a natural geometric multiplicative functional in
S 2 (V) lying above a smooth path x, then Ai,js,t (x) is the area contained between
the curve (xir , xjr )r∈[s,t] and the chord (xis , xjs )(xit , xjt ). Denote also this area
by Ai,j
s,t (x).

Lemma 6.3. For all 0  s  t  1, the map x ∈ S 2 (Rd ) → As,t (x) is not
continuous with respect to the uniform norm (except if d = 1, in which case
As,t (x) = 0).

Proof. Assume that d = 2 and identify R2 with the complex plane C. Set
2
xnt = n−1 ein t . Then, A0,2π (xn ) = π, but xn converges uniformly to 0. 

Thus, to construct a geometric multiplicative functional x lying above x of

finite p-variation with p ∈ [2, 3), one has to focus only on the construction of
the antisymmetric part A(x) of x. This also provides us a with nice geometric
interpretation. However, note that this choice is not unique.

Lemma 6.4. If x is in G p (V) for p ∈ [2, 3), and ϕ = (ϕi,j )i,j=1,...,d is a func-
tion from [0, 1] to d × d-antisymmetric matrices, and of finite p/2-variation.
Then x  defined by

1s,t = x1s,t and x

x 2,i,j 2,i,j
s,t = xs,t + ϕi,j (t) − ϕi,j (s).

 is also in G p (V).
Then x

Proof. If x is a geometric multiplicative functional, and e1 , . . . , ed is the canon-

ical basis of V = Rd , then rewrite xs,t as

d

d
xs,t = 1 + ei x1,i
s,t + ei ⊗ ej x2,i,j
s,t .
i=1 i,j=1

The quantity log(xs,t ), which belongs to A2 (V) (see (6.1c)), may be explicitly
computed:
38 Antoine Lejay

d
1

d
log(xs,t ) = ei x1,i
s,t + [ei , ej ]Ai,j
s,t (x),
i=1
2 i,j=1

where [ei , ej ] = ei ⊗ ej − ej ⊗ ei is the Lie bracket of ei and ej . It is clear

that x  is a multiplicative functional in Mp (V), and as ϕ is antisymmetric,
i.e., ϕi,j (t) = −ϕj,i (t) for all t ∈ [0, 1],

1

d
xs,t ) = log(xs,t ) +
log( [ei , ej ](ϕi,j (t) − ϕi,j (s)).
2 i,j=1

 belongs to G p (V).
xs,t ) belongs to A2 (V) for all (s, t) ∈ ∆+ . Then, x
Thus, log(


We have then a way to construct as many geometric multiplicative func-

tionals lying above a path x as we
 want. The difference
 between the geometric
 = f (xs ) d
multiplicative functionals y = f (xs ) dxs and y xs is immediate
in view of (3.9a)-(3.9b):
d 

t
∂fi
 1s,t = ys,t
y 1
+ (xr ) dψj,i (r)
i,j=1 s ∂xj
 t
 2s,t = ys,t
and y 2
+ f (xr ) ⊗ f (xr ) · dψ(r).
s

In Sect. 10.2 below, these results will be used to compare the theory of integra-
tion given by this theory and the Stratonovich integral for Brownian motion
(see especially (10.2) and (10.3)).
The fact that ϕ was taken additive (i.e., ϕ(s, u) + ϕ(u, s) = ϕ(s, t) for
all 0  s  u  t  1) in Lemma 6.4 is justified by the following Lemma.
Lemma 6.5 ([32, Lemma 2.2.3, p. 250]). If x and x  are two multiplicative
functionals in Mp (V) which agree for all order smaller than k = p (i.e.,
is,t for i = 0, 1, . . . , k − 1), then ϕ(s, t) = xks,t − x
xis,t = x  ks,t is additive, and
is of finite p/k-variation.

6.3 Intermezzo: solving linear differential equations

It is now time to justify the usefulness of geometric multiplicative functionals,

and the choice of the semi-norm ·p,[s,t] .
Assume that x is a piecewise smooth path, and that x is its Chen series
(in T∞ (V), i.e., consider all its iterated integrals).
Let C be a m × m-matrix and assume to begin with that d = 1. Then it
is well known that the solution of the differential equation dyt = Cyt dxt is
yt = exp(Cxt )y0 , where exp is the exponential of matrices.
Now, consider a family C1 , . . . , Cd of m × m-matrices, and the differential
equation
 An Introduction to Rough Paths 39

d
dyt = Ci yt dxit . (6.7)
i=1
As y appears in the right-hand side of (6.7), one may replace it by its value
given by (6.7). Thus,

d  t d 


yt = ys + Ci ys dxis + Ci Cj ys dxir1 dxjr2 .
i=1 s i,j=1 s<r1 <r2 <t

For a multi-index I = (i1 , . . . , ik ), set CI = Ci1 Ci2 · · · Cik .

Re-iterating the substitution, one obtains the formal power series
  t 


yt = Id + CI dI x ys . (6.8)
I multi-index s

As the Ci ’s appear at the same place as the indeterminates X i in the formal

Chen’s series Φ([s, t], x) of x, (6.8) may also be written, using (5.1),
  t 


yt = exp ΘI (C1 , . . . , Cd ) dI x ys ,
I multi-index s

where ΘI (C1 , . . . , Cd ) is a linear combination of terms in the smallest space of

matrices containing { C1 , . . . , Cd } and closed under the Lie brackets [A, B] =
AB − BA.
Of course, one may wonder if the series

 t
Ξ(x) = Id + CI dI x
I multi-index s

converges. But there are d

 i multi-indexes
 I of length i, and CI   ci =
 t 
(supi=1,...,d C) . As  s dI x  C|t − s| /i! for some constant C which de-
i i

pends on the bounds of the derivatives of x, then

+∞ i i
d c |t − s|i
Ξ(x)  < +∞.
i=0
i!
Thus, using the condition that a rough path x is of finite p-variation con-
trolled by ω, one may construct the solution y of


d
dyt = Ci yt dx1,i
s,t , for t
s, (6.9)
i=1

by setting  


yt = Id + CI xlength
s,t
I,I
ys . (6.10)
I multi-index
Inequality (6.2) implies that the series that appears in this expression is con-
vergent, in the sense of the norm of operators.
40 Antoine Lejay

6.4 Extending multiplicative functionals to any order

We have defined geometric multiplicative functionals as elements of the “trun-

cated” tensor algebra Tk (V), while an expression like (6.10) requires to know
a geometric multiplicative functionals in the tensor algebra T∞ (V). Does one
need to know all the terms of a geometric multiplicative functional?

Theorem 6.1 ([32, Theorem 2.2.1]). Let x be a multiplicative functional

in Tk (V), where k = p for some p
1. Assume that there exists a function ω
satisfying Assumption 2.1 such that for all (s, t) ∈ ∆+ ,

ω(s, t)i/p
|xis,t |  (6.11)
βΓ (i/p)

for i = 1, . . . , k. Then, if β is large enough (however, the choice of β depends

only on p), for all integer  > k, there exists a procedure to construct a
multiplicative functional y ∈ T (V) extending x (i.e., xi = yi for i = 1, . . . , k)
and satisfying (6.11) for i = 1, . . . , . Moreover, the extension y of x given
by this procedure is unique.

Proof (Sketch of the proof ). The idea is to construct y() ∈ T (V) recursively,
by setting y(k) = x ∈ Tk (V)  and, once y
()
has been defined, set z(+1)
in T+1 (V) by z(+1)
= y + (i1 ,...,i+1 ) 0 · ei1 ⊗ e+1 . Thus, y(+1) is defined
()

by
(+1) (+1) (+1) (+1)
ys,t = lim ztδ ,tδ ⊗ ztδ ,tδ ⊗ · · · ⊗ ztδ ,tδ ,
δ→0 0 1 1 2 iδ −1 iδ

where Π =δ
s   ··· 
tδi tδ0  t is a partition of [s, t] whose mesh
tδiδ
goes to 0 as δ decreases to 0. Thanks to the multiplicative property of y() ,
remark that y(+1),i = y(),i for i = 1, . . . , . In fact, this idea was already
used in the proof of Proposition 3.2, and will be used later in the proof of
Theorem 7.1. In T (V), the extension of x is then y() . 

This Theorem means that there exists a function Ψp transforming multi-

plicative functionals in Mp (V ) to multiplicative functionals in T∞ (V). There
are many ways to extend a multiplicative functional. For example, if ϕ is
a smooth function with values in V⊗2 , then (1, 0, ϕ(t) − ϕ(s))(s,t)∈∆+ be-
longs to M2 (V) and extends the multiplicative functional (1, 0). But the
function Ψ1 applied to smooth paths yields the series of iterated integrals,
i.e., Ψ1 (x)s,t = Φ([s, t], x). Moreover, the next Theorem states that Ψp is con-
tinuous on Mp (V). So, given a multiplicative functional x ∈ Mp (V), we call
Ψp (x) its extension.

Theorem 6.2 ([32, Theorem 2.2.2]). Let x and y be two multiplicative func-
tionals in Tk (V) satisfying the hypotheses of Theorem 6.1 for the same ω.
Assume that there exists a constant β large enough (depending only on p)
and a constant 0 < ε < 1 such that
 An Introduction to Rough Paths 41

ω(s, t)i/p
|xis,t − ys,t
i
|ε for i = 1, . . . , k. (6.12)
βΓ (i/p)

Then, their extensions x  and y

 to T∞ (V) given by Theorem 6.1 also sat-
isfy (6.12) for all integer i.

In view of Lemma 3.1, this means that the map giving the extension
of a multiplicative functional in Tp (V) is continuous with respect to the
norm ·p .
Combined with Proposition 6.1, this means that the extension of a geo-
metric multiplicative functional in G p (V) to T (V) is also a geometric multi-
plicative functional in this space for all  > p.
Using Theorem 6.1, if x is a geometric multiplicative functional in Tp (V)
of finite p-variation (i.e., xp is finite: there is no difficulty to find a function ω
such that x is controlled by ω), then one may solve (6.9) by density using (6.10)
and the previous Theorem.
The idea behind Theorem 6.1 is that the more irregular is a trajectory
(“rough”), the more “iterated integrals” have to be considered. But on the
other side, once one knows enough iterated integrals, then the whole set of
iterated integrals could be known. Hence, when one deals with a path x of finite
p-variation, then he needs to know a geometric multiplicative functional x
lying above x and belonging to the truncated tensor algebra Tp (V).

7 Almost multiplicative functionals

When p < 2, the geometric multiplicative functionals we consider are xs,t =

(1, x1s,t ), with x1s,t = xt − xs for a continuous path of finite p-variation. In this
t
case, we have defined integrals of the type s f (xr ) dxr for Hölder continuous
functions f , which are, as we have seen, of finite p-variation. This integral is
a multiplicative functional in Mp (W), since from Chasles’ relation,
 u  t  t
f (xr ) dxr + f (xr ) dxr = f (xr ) dxr . (7.1)
s u s

However, we have seen that f (xr ) dxr may be constructed as limit of Rie-
mann sums, and for that, we have used the approximation
 t
ys,t = f (xs )(xt − xs )  f (xr ) dxr .
s

Of course, (ys,t )(s,t)∈∆+ fails to satisfy (7.1), but the error εs,u,t = ys,t − ys,u −
yu,t was easily controlled. And the estimate on εs,u,t was the key of the proof
of Proposition 2.1. So, (ys,t )(s,t)∈∆+ may be called an almost multiplicative
functional.
42 Antoine Lejay

Definition 7.1. A continuous function x : ∆+ → Tk (V) for some integer k is

called an almost multiplicative functional if it is of finite p-variation controlled
by ω (see Definition 6.2) and, for i = 1, . . . , k,

|(xs,t − xs,u ⊗ xu,t )i |  Kω(s, t)θ for all 0  s  u  t  1. (7.2)

for some θ > 1 and some constant K.

Thus, an almost multiplicative functional fails to satisfy (6.1b), but the

error is in some sense close to a multiplicative functional.

Theorem 7.1 ([32, Theorem 3.3.1]). Let x be an almost multiplicative func-

tional taking its values in Tk (V) with finite p-variation controlled by ω. Then
there exists a unique multiplicative functional z in Tk (V) such that for all
(s, t) ∈ ∆+ , there exists a constant C depending on K, θ (defined by (7.2)),
the degree k and the control ω(s, t) such that

|(xr,u − zr,u )i |  Cω(r, u)θ for s  r  u  t and i = 1, . . . , k. (7.3)

Furthermore, there is at most one multiplicative functional z in Tk (V) satis-

fying (7.3) regardless of the choice of C.

Proof (Sketch of the proof ). The idea was already used in the proofs of Propo-
sition 3.2 and Theorem 6.1: Construct z by setting z0 = z(0) = 1 and recur-
() (−1)
sively for  = 1, . . . , k, ys,t = zs,t + xs,t , where z(−1) = (z0 , z1 , . . . , z−1 ) ∈
T−1 (V). Hence, define z ∈ T (V) by
()

() () () ()

zs,t = lim ytδ ,tδ ⊗ ytδ ,tδ ⊗ · · · ⊗ ytδ ,tδ
,
δ→0 0 1 1 2 iδ −1 iδ


where Π δ = tδi s  tδ0  · · ·  tδiδ  t is a partition of [s, t] whose mesh
goes to 0 as δ decreases to 0. The multiplicative functional z is then z(k) .

8 Integration of a one form

t
We now have all the elements to define an integral like 0 f (xs ) dxs against
a geometric multiplicative functional x ∈ Tp (V) of finite p-variation for an
arbitrary p > 1. We want this integral to belong to Tp (W) if f is a one-
form taking its values in a Banach space W, and in fact that (s, t) ∈ ∆+ →
t
s
f (xr ) dxr belongs to G p (W).

8.1 Lipschitz functions

The notion of Lipschitz functions we use is that of E.M. Stein (see for example
the book [44]).
 An Introduction to Rough Paths 43

Definition 8.1. Let F be a closed subset of the normed space U, and α > 0.
Let W be a separable Banach space, and f a function with values in W.
Assume that k < α  k + 1. Then f belongs to Lip(α, F, W) if there exist
some functions f J where J is a multi-index of length length(J)  k and some
functions RJ : F × F → W such that for all x, y ∈ F,

1
f J (x) − f J (y) = f (J,L) (y)(x1 − y 1 ) · · · (xm − y m )
1 ! · · · m !
L=(1 ,...,m ),
length(J,L)k

+ RJ (x, y),
where (J, L) denotes the concatenation of the multi-indexes J and L. By
definition, f ∅ = f . The functions f J shall satisfy
|f J (x)|  M and |RJ (x, y)|  M |x − y|α−length(J) (8.1)
for all x, y ∈ F and any J with length(J)  k. Denote by f Lip the small-
est M such that (8.1) is true. With this norm, Lip(α, F, W) is a Banach
space.
This definition requires some comments. If F = V = Rd , then the functions
in Lip(α, Rd , W) are from Rd to W with bounded derivatives up to order

α. Moreover, f (i1 ,...,i ) = ∂xi ∂···∂x
f
i
, and f (i1 ,...,i
α) is (α − α)-Hölder
1
continuous.
If F is a strict subset of Rd , then a function f ∈ Lip(α, F, W) may be
extended continuously to a function in Lip(α, Rd , W), but the family of the
f J ’s is not necessarily unique. In this case, by a function f in Lip(α, F, W),
we denote not only f , but the whole family (f J )J multi-index, length(J)α .

8.2 Integration
To start with, let f = (f1 , . . . , fd ) be a smooth function defined on the Banach
t
space V = Rd . The idea to define 0 f (xs ) dxs is to construct an almost multi-
t
plicative function y such that ys,t gives a first approximation of s f (xr ) dxr ,
and then to transform y to a geometric multiplicative functional using Theo-
rem 7.1.
The value of ys,t will be computed as previously using a Taylor expansion
of f when it is assumed that x is smooth. So, in a first approach, assume that x
is piecewise smooth and that x is the geometric multiplicative functional given
by its iterated integrals.
Set

k
 t
yt − ys = ys,t =
1
DI (f )(xs ) dI x
j=1 I multi-index, I=(i1 ,...,ij ) s

∂ j−1 fi1
with DI (f )(xs ) = (xs ).
∂xij · · · ∂xi2
44 Antoine Lejay

j
Once this is done, one may define y forjt = 2, . . . , k using the iterated
j
integrals of y: ys,t = I multi-index, I=(i1 ,...,ij ) s dI y. But this involves expres-
sions such as
  s1   s 
S(J1 , . . . , J ) = d dJ1 x · · · d dJ x ,
ss1 ···s t 0 0

where J1 , . . . , J are themselves multi-indexes. But it is possible to express

such a sum S(J1 , . . . , J ) as the sum of

 t
S(J1 , . . . , J ) = εK dK x,
K multi-index, s
length K=length(J1 )+···+length(J )

where εK ∈ { 0, 1 } (see [32, Predefinition 3.2.1, p. 283]). Denote by J1  · · · J

the set of multi-indexes K that appear really in the previous sum, i.e., for
which εK = 1.

Definition 8.2. Let x be a geometric multiplicative functional of finite p-

variation lying above a path x, and let f be a differential form in Lip(α, V, W)
for some α > p − 1, that is, v ∈ V → f (·)v is linear and for all v ∈ V, f (·)v
t
belongs to Lip(α, V, W) for some α > p − 1. The integral 0 f (xs ) dxs of f
along the path x is defined to be the geometric multiplicative functional of finite
p-variation given by Theorem 7.1 corresponding to the almost multiplicative
functional


length(K),K
i
ys,t = DJ1 (f )(xs ) ⊗ · · · ⊗ DJi (f )(xs )xs,t .
J1 ,...,Ji multi-indexes, K∈J1
···
Ji
length(J1 )+···+length(Ji )k

Here, DJ (f )(x) is defined by to be an element in the dual of V⊗ length(J) (with

values in W): If e1 , . . . , ed is the canonical basis of V and e1 , . . . , ed is its
dual basis, then

(x)eij ⊗ · · · ⊗ ei1 .
(i ,...,i2 )
D(i1 ,...,ij ) (f )(x) = fi1 j

Theorem 8.1 ([32, Theorem 3.2.1, p. 285]). Definition 8.2 is valid, i.e., the
definition of y gives rise to an almost multiplicative functional, which is con-
trolled by Kω if x is controlled by ω, where the constant  K depends only
on α, p, f Lip and sup(s,t)∈∆+ ω(s, t). Moreover, x → f (x0 + x10,s ) dxs is
continuous from G p (V) to G p (W).

9 Solving differential equations

We can now consider solving differential equations of the form
 An Introduction to Rough Paths 45
 t
yt = a + f (ys ) dxs , (9.1)
0

where x belongs to G p (V), and yt belongs to a Banach space W for all t

0. Thus, f is a differential form, such that for all v ∈ V, f (·)v belongs to
Lip(α, W, W) for some α > p − 1.
Equation (9.1) will be solved using Picard’s iteration principle: Construct
t
a sequence (y n )n∈N by ytn+1 = a + 0 f (ysn ) dxs , and prove that y n converges
to some element y. But such a principle requires that y n+1 and y n belong
to the same space. So, consider some multiplicative functional z = (x, y) in
Tp (V ⊕ W), and define

K(z) = h(x0 + x10,s , a + y0,s 1
) dzs ,

where h(x, y) is the differential form h(x, y) = fi (y) dxi . Thus, Picard’s
iteration principle will be applied on elements of Tp (V ⊕ W).

Definition 9.1. The solution of (9.1) is an extension z in G p (V ⊕ W) of

x ∈ G p (V) such that z lies above (x, y) with (x0 , y0 ) = (x0 , a) and z satisfies
z = K(z).

Note that although the projection y on G p (W) of z can be seen as the

solution of (9.1), z also keeps track of the “interactions” between x and y
using the iterated integrals.

Theorem 9.1 ([32, Theorem 4.1.1, p. 298]). If f is a linear form on V with

values in Lip(α, W, W) for some α > p − 1, then there exists a solution
to (9.1) when x belongs to G p (V). Moreover, if f is a linear form on V with
values in Lip(α, W, W) for α > p, then this solution z is unique, and the map
I : x → z, called the Itô map, is continuous from G p (V) to G p (V ⊕ W).

Remark 9.1. To prove the existence of a solution under the assumption that f
belongs to Lip(α, W, W) with α > p − 1, one only has to act as in Step 2 in
the proof of Theorem 2.1: In Step 3 of the proof of Theorem 4.1.1 in [32], it is
proved that the paths y n given by the Picard iteration principle are of finite
p-variations controlled by the same ω. Hence Lemma 6.2 can be used.

10 A practical example: Brownian motion

We show in this section how the theory of rough paths may be used to define
a stochastic integral against Brownian motion. It is well known that almost
surely, a trajectory of Brownian motion is α-Hölder continuous for all α < 1/2.
Thus, a trajectory of Brownian motion is then of finite p-variation for all p > 2.
Given a Brownian trajectory B(ω), the main difficulty is to create a ge-
ometric multiplicative functional B(ω) lying above B, where the Brownian
46 Antoine Lejay

motion is defined on a probability space (Ω, F , P) and lives in V = Rd . In

view of Proposition 6.1, one has only to construct a piecewise smooth approx-
imation B δ (ω) converging to B(ω) as δ → 0, and to study the convergence of
Bδ with B1,δ (ω) = B δ (ω) and
 t
2,i,j,δ
Bs,t (ω) = Bri,δ (ω) dBrj,δ (ω).
s

But we know that: (i) The limit of B2,i,j,δ depends on the choice of the
approximation. (ii) When it converges, B2,i,j,δ does not converge almost surely
but only in probability or in L2 (P) (however, it is proved that for dyadic
partitions, the convergence may be almost sure. See [24] for example).
Point (ii) is contained in the classical result from E. Wong and M. Zakai
in [51] for some piecewise linear approximation of the Brownian motion, while
point (i) is related to the extensions of such a result (see [19, Sect. VI,-7,
p. 392] or [22, Chap. 5.7, p. 274] for example). In fact, the problems with
(i) are similar to the results given in Sect. 6.2: There are different geometric
multiplicative functionals lying above the same path B.

10.1 The “natural” choice

The natural choice for B δ is given by

Btδ (ω) = Bti (ω) + (ti+1 − ti )−1 (t − ti )(Bti+1 (ω) − Bti (ω)) (10.1)

for t ∈ [ti , ti+1 ], where Π δ = { ti 0  t1  · · ·  tk  1 } is a determin-

istic partition [0, 1] whose mesh goes to 0 with δ. It is clear that B δ (ω)
converges uniformly to B(ω). We have seen at the end of Sect. 2.3 that
Varq,[0,1] (B δ (ω))  3 Varq,[0,1] (B(ω)) for all q > 2. According to Lemma 6.2,
B δ (ω) converges in the topology generated by Varp,[0,1] (·) + ·∞ to B(ω) for
all p > 2.
For such an approximation, we know that
 t
uniformly in (s, t) ∈ ∆+ 2,i,j def
B2,i,j,δ
s,t −− − −− −− −− − − − − −−→ B s,t = (Bri − Bsi )◦dBrj
δ→0 s

in probability. Here, the stochastic integral is a Stratonovich integral. We prove

below in Sect. 11.2 that Bδ converges to B in the topology generated by ·p
if the partitions Π δ are dyadic. Thus, for this choice of B2 , the geometric
multiplicative functional B belongs to G p (V) for all p > 2.
Now, let f be a linear function on V taking its values in Lip(α, V, W)
for some α > 1, with W = Rm . A direct consequence of Proposition 3.1 or
t
Theorem 8.1 is that Xs,t = s f (Br ) dBr is well defined and belongs to G p (W)
for all p ∈ (2, 1 + α).
Remark 10.1. A practical feature of the theory of rough paths is that X is
defined on a subset Ω0 ⊂ Ω of full measure whatever the function f is.
 An Introduction to Rough Paths 47

Let X be a path such that X lies above X, with X0 = x0 . Then, a direct

consequence of Theorem 8.1 is that
 t
probability
Xtδ = x0 + f (Bsδ ) dBsδ −−−−−−−→ Xt ,
0 δ→0

the convergence holding with respect to both the uniform norm and the norm
of p-variation. But from a theoremof Wong-Zakai type, it is also known that
·
X δ converges in probability to x0 + 0 f (Bs )◦dBs . Thus, this integral is almost
surely equal to X. There is in fact a deep relation between the Stratonovich
integral and the one given by the theory of rough paths.

10.2 Stratonovich integrals and rough paths theory

We develop in this section the link between stochastic integrals given by the
theory of rough paths and Stratonovich integrals for Brownian motion. It also
explains the influence of the term B2 , where B is a geometric multiplicative
functional lying above the Brownian motion, but different from the one given
by the “natural” construction of Sect. 10.1. Note that such geometric multi-
plicative functionals may arise naturally. For example (see e.g., [26, 28] in the
homogenization theory), there exist some families (X ε )ε>0 of semi-martingales
converging, thanks to a central limit theorem, to a Brownian motion B, but
such that As,t (X ε ) converges to As,t (B) + c(t − s) for some matrix c.
t
By definition, the Stratonovich integral 0 fj (Bs )◦dBsj is the limit in prob-
ability of

k−1
(fj (Bti+1 ) + fj (Bti )) j
δ def
I = (Bti+1 − Btji )
i=1
2


k−1

k−1
(fj (Bti+1 ) − fj (Bti )) j
= fj (Bti )(Btji+1 − Btji ) + (Bti+1 − Btji ),
i=1 i=1
2

where Π δ = { ti 0  t1  · · ·  tk  t } is a deterministic partition of [0, t].

The functions f belongs to Lip(α, W, W) with α > 1. Let p ∈ (2, 1 + α)
d ∂fj 
be fixed. By definition, fj (x) − fj (y) = =1 ∂x (y − x ) + Rj (x, y) and


Rj (x, y)  |y − x| . Thus,
α

k−1
Iδ = fj (Bti )(Btji+1 − Btji )
i=1

1

∂fj
k−1 d
+ (Bti )(Bti+1 − Bti )(Btji+1 − Btji ) + εδ ,
2 i=1 ∂x
=1

with
48 Antoine Lejay

k−1
εδ = Rj (Bti , Bti+1 )(Btji+1 − Btji )
i=1


k−1
C |ti+1 − ti |(α+1)/p  Ct sup |ti+1 − ti |(α+1−p)/p −−−→ 0.
i=1,...,k−1 δ→0
i=1

This constant C is such that |Bt − Bs |  C 1/(1+α) |t − s|1/p for the considered
trajectory of the Brownian motion.
Now, let B be a geometric multiplicative functional lying above B. There
is no necessity to choose the previous one, and we have seen in Sect. 6.2 how
to construct as many areas as we want. Then
1 
(B − Bti )(Btji+1 − Btji ) = B2,,j ,j
ti ,ti+1 − Ati ,ti+1 (B),
2 ti+1
where A(B) is the antisymmetric part of B2 . Moreover, we have seen that
d k−1 j j 1 ∂fj 2,,j
j=1 i=1 fj (Bti )(Bti+1 − Bti ) + 2 ∂x (Bti )Bti ,ti+1 converges almost surely
to Xt − X0 , where X is the path above which f (Bs ) dBs ∈ G p (W) lies. So,
we deduce that

d  t
X0 + fj (Bs )◦dBsj = Xt − Qt (B)
j=1 0

d
k
∂fj δ ,j
with Qt (B) = lim (Bti )Ati ,ti+1 (B).
δ→0
i=1
∂x
j,=1

The limit defining Qt (B) is a limit in probability.

Remark 10.2. Using the antisymmetry of A,j

s,t (B), one has

d k−1  
1

∂fj ∂f
Qt (B) = lim − (Btδi )A,j
ti ,ti+1 (B).
δ→0 2 ∂x ∂xj
i=1 j,=1

∂fj
Thus, if ∂x − ∂f
∂xj = 0, Qt (B) = 0, then X depends only on B and not
2 ∂F
on the choice of B . In particular, this is true if fi = ∂x i
for some function
δ
F . In such a case, this could be shown directly, if (B )δ>0 is a family of
geometric multiplicative function lying above an approximation B δ of B and
converging to G p (V) to B, then the change of variables’ formula reads: F (Btδ )−
t
F (B0δ ) = 0 fi (Bsδ ) dBsδ . Thus, F (Btδ ) − F (B0δ ) converges to F (Bt ) − F (B0 ),
t
while 0 fi (Bsδ ) dBsδ converges to Xt .
Now, if Bnat is the “natural” rough path lying above B (see Sect. 10.1),
then A,j
s,t (B
nat
) is the Lévy area A,js,t (B) of the 2-dimensional Brownian mo-
tion (Br , Brj )r∈[s,t] , i.e., the area enclosed between the curve of (B  , B j ) and
its chord:
 An Introduction to Rough Paths 49
 t  t 
1
A,j
s,t (B) = (Br − Bs )◦dBrj − (Brj − Bsj )◦dBr .
2 s s

The result given at the end of Sect. 10.1 implies that Qt (Bnat ) = 0 almost
surely.
Moreover, one knows from Sect. 6.2 that there exists a function ϕ =
(ϕi,j )i,j=1,...,d from [0, 1] to the space of antisymmetric matrices (i.e., ϕi,j (t) =
−ϕj,i (t) for all t ∈ [0, 1]) and of finite p/2-variation such that As,t (B) =
As,t (B) + ϕ(t) − ϕ(s). We deduce that
d 

t
∂fj
Qt (B) = (Bs ) dϕ,j (s).
0 ∂x
j,=1

To summarize,
 if B δ is a piecewise smooth approximation of B such that
(1, B , dB ⊗ dB δ ) converges to the geometric multiplicative functional B
δ δ

in G p (V), and As,t (B) = As,t (B) + ϕ(t) − ϕ(s), then we obtain directly that
 t  t d 

t
probability ∂fj
f (Bsδ ) dBsδ −−−−−−−→ f (Bs )◦dBs + (Bs ) dϕ,j (s)
0 δ→0 0 0 ∂x
j,=1
 d   
t
1
t ∂fj ∂f
= f (Bs )◦dBs + − (Bs ) dϕ,j (s). (10.2)
0 2 0 ∂x ∂xj
j,=1

When one considers the solution Y δ in W = Rm of the ordinary differential

equations  t
Ytδ = y0 + f (Ysδ ) dBsδ ,
0
then Y δ converges in probability to the solution Y of
 t
Yt = y0 + f (Ys )◦dBs
0

 t 
1 ∂fj k ∂fj 
+ f − f (Ys ) dϕk, (s). (10.3)
2 0 ∂x  ∂xk k
,k=1,...,m
j=1,...,d

Here, the drift term is different from the  one in (10.2), since it comes from
the cross iterated integrals of the type dY ⊗ dB, which may be computed
first for smooth paths, and then by passing to the limit.
Thus, the theory of rough paths provides us with some new light on the
results presented in Sect. VI-7 in [19, p. 392] (see also Historical Note 5 below).
In this book, the results concern the case where ϕ,j (t) = c,j where c =
(c,j ),j=1,...,d is an antisymmetric matrix, whose terms are given by
 δ  δ 
1 1  ,j
δ 
c,j = lim E Bs dBs −
,δ j,δ j,δ ,δ
Bs dBs = lim E A0,δ B ,
δ→0 2δ 0 0 δ→0 2δ
50 Antoine Lejay

where B δ (ω) is given by an interpolation of B(ω) sampled at points 0, δ, 2δ, . . .

The matrix c depends on the way the trajectory is interpolated.
Remark 10.3. In the construction of the Stratonovich integrals, it is important
that the partitions Π δ are deterministic. If not, a drift term may appear, which
may be computed using the expression of Qt (B). This result has also been
widely used to construct approximations of solutions of SDEs using partitions
whose meshed goes to 0 at a speed that depends on the considered trajectory
of the underlying Brownian motion (see for example [16]).

10.3 Itô stochastic integrals

When one constructs integrals, only the fact that x is multiplicative is used.
The fact that x is geometric is not really used, except that it allows us to
construct first the objects for smooth paths, and then to deduce what the
result should be for general geometric multiplicative functionals. It is the way
the integral was defined in Sect. 8. However, when p < 3, as we have seen in
Sect. 3, one may directly set, given a multiplicative functional x,
1
ys,t = f (xs )x1s,t + ∇f (xs )x2s,t and ys,t
2
= f (xs ) ⊗ f (xs )x2s,t ,
1 2
and prove that y =  (y , y ) may be transformed into amultiplicative func-
tional denoted by f (xs ) dxs . Moreover, the map x → f (x0 + x0,s ) dxs is
continuous in Mp (V).
For a N -dimensional Brownian motion B, we know that
 t  t
(Br − Bs )◦dBs =
i i j
(Bri − Bsi )dBsj + δi,j (t − s)
s s
where δi,j is the Kronecker symbol. Thus, one may define a multiplicative
functional Bitô ∈ Mp (V) by (Bitô
s,t )
1,i
= Bti − Bsi , (Bitô
s,t )
2,i,j
= (Bnat
s,t )
2,i,j
−

δi,j (t−s) and define the (pathwise) Itô stochastic integral to be f (Bs ) dBitô s .
 
Remark 10.4. Of course, f (Bs ) dBs and f (Bs ) dBitô s are defined pathwise,
but pathwise means “pathwise with respect to B” and not pathwise with
respect to B”. And the definition of B from B is not pathwise, and requires
that some stochastic integration, here of Itô or Stratonovich type, is already
defined.

11 How to compute p-variation?

It is generally difficult to compute the p-variation of a function. We give in this
section a trick which has been introduced in [17] and allows us to compute the
p-variation of a multiplicative functional x provided one has a nice estimate
on |xji2−n ,(i+1)2−n |p/j for j = 1, . . . , p, for all integers n and i = 0, . . . , 2n −1.
We give an example in the case of the Brownian motion, that allows us
to complete the results of Sect. 10.1. In fact, this approach was successful in
many cases: See [38, 3, 24, 1, 27] for various applications.
 An Introduction to Rough Paths 51

11.1 Using dyadics

For any integer n and any k = 0, . . . , 2n , set tnk = i/2n, that is (tnk )k=0,...,2n is
the dyadic partition of [0, 1] at level n. Let (s, t) belongs to ∆+ , and construct
recursively a sequence (sm )m∈Z, m=0 of elements in (tnk )n∈N, k=0,...,2n by the
following way: Let n0 be the smallest integer such that [tnk 0 , tnk+10
] ⊂ [s, t] for
some integer k. set s−1 = tnk 0 and s1 = tnk+1 0
. Hence, construct sm for m
1
by setting, if sm < t,
n 
nm = inf n
nm−1 ∃k ∈ N, tnk m = tk m−1 , tm k+1  t .
n

Denote by sm the value tnk+1m

, where k is the unique integer for which tnk m =
nm−1
tk and tnk+2
m
> t. If sm = t then sn = t for all n
m.
Construct sm for m < −1 using a similar procedure, where sm decreases
to s instead of increasing to t.
This construction ensures that the sequences (nm )m∈N and (nm )m∈Z∗ , m<0
are increasing, and

[s, t] = · · · ∪ [s−m−1 , s−m ] ∪ · · · ∪ [s−1 , s1 ] ∪ · · · ∪ [sm−1 , sm ] ∪ · · · .

Then, for all x in Mp (V),

xs,t = lim xs−m−1 ,s−m ⊗ · · · ⊗ xs−1 ,s1 ⊗ · · · ⊗ xsm ,sm+1 .

m→∞

Thus, if k = p and j = 1, . . . , k,

j


xjs,t = xrs1m1 ,sm1 +1 ⊗ · · · ⊗ xrsim ,smi +1 .
i
i=1 r1 ,...,ri =1,...,j m1 <···<mi
r1 +···+ri =j m1 ,...,mi ∈Z∗

In the previous expression, we use the convention that mi + 1 = 1 if mi = −1.

So, for j = 1, . . . , k,
   
j

j



|xs,t |  |xsm ,sm+1 | · · ·
r1
|xsm ,sm+1 | .
ri

i=1 r1 ,...,ri =1,...,j m∈Z∗ m∈Z∗

r1 +···+ri =j

Using the Hölder inequality, for r = 1, . . . , k and any β > (p − 1)/p,

 r/p



|xrsm ,sm+1 |  C(r) m |xsm ,sm+1 |
nβp/r r p/r

m∈Z∗ m∈Z∗
 r/p


j

C m |xsm ,sm+1 |
nβp/r r p/r
,
r=1 m∈Z∗
52 Antoine Lejay
 (p−r)/p
−βp/(p−r)
where C(r) = m∈Z ∗ n m and C = supr=1,...,k C(r). Our
choice of β ensures that C is finite. Then, there exists a constant K depending
only on k and C such that
 j/p


j

|xjs,t |  K m |xsm ,sm+1 |
nβp/r r p/r

r=1 m∈Z∗

and then that

 j/p

j


|xjs,t |  K 
 nβp/r |xrtni ,tni+1 |p/r 
 , (11.1)
r=1 n
n(s,t) i=0,...,2n −1
i ∈[s,t], ti+1 ∈[s,t]
tn n

where n(s, t) is the smallest integer n such that there exists some integer k
for which [tnk , tnk+1 ] ⊂ [s, t].
Inequality (11.1) is useful, since it allows to estimate both

sup |xjs,t |, and Var (xj )

(s,t)∈∆+ ,t−s<η p/j,[0,1]

which satisfies
 j/p


j
n
2
−1
Var (xj )  K  nβp/r |xrtni ,tni+1 |p/r  (11.2)
p/j,[0,1]
r=1 n
0 i=0

provided one knows xtni ,tni+1 for all dyadic point tni = i/2n .

11.2 Application to Brownian motion

We have constructed in Sect. 10.1 a piecewise linear approximation B δ (ω)

of a Brownian motion trajectory B(ω). Let Bδ (resp. B) be the geometric
multiplicative functional in G p (V) lying above B δ (resp. B) and constructed
as in Sect. 10.1.
Proposition 11.1. The sequence (Bδ )δ>0 converges in probability to B in
Mp (Rd ) for all 2  p < 3. Moreover, if Bδ is a piecewise linear approx-
imation of B along dyadic partitions (consider only the δ’s of type 1/2n ),
then Bδ converges almost surely to B in Mp (Rd ).
Proof. To simplify, we do not give a complete proof here. We prove only
that Bδ converges in probability to B where δ = 2−n for some integer n, and
the partition Π δ we use is (tnk )k=0,...,2n with tnk = k2−n .
There is no real difficulties in extending the proof when Π δ is not dyadic
partitions (see [27] for example), although it requires a bit more computa-
tions. Furthermore, still using the ideas to compute the p-variation of Bδ by
 An Introduction to Rough Paths 53

estimating Bδs,t , where s, t are dyadic points (s, t) = (k/2m , (k + 1)/2m ) for
all m
1 and any k ∈ { 0, . . . , 2n }, it could be shown that if Π δ is the dyadic
partition, then Bδ converges almost surely to B (see [24]).
Let q be a real number in (2, p). If m  n, then according to the Doob
inequality, there exists a constant C depending only on q such that for a =
1, . . . , d,
 q !   !
E Bta,δ m − Bta,δ
m
  E Btam − Btam q  C(tm
k+1 k k+1 − tk )
m q/2
 C2−mq/2 .
k+1 k

If m > n, then again by the Doob inequality, there exists a constant C de-
pending only on q such that for a = 1, . . . , d,
 q
  q !  tm − tm    !
E Bta,δ
m − Bta,δ
m
   k+1

k 
 E Bta,δ a,δ q
n − Btn
k+1 k ti+1 − ti
n n i i

 C2−mq 2nq/2  C2−mq/2

k , tk+1 ] ⊂ [ti , ti+1 ]. So, we deduce that for a = 1, . . . , d and

if i is such that [tm m n n
−n
any δ = 2 for some integer n,
 q !
E Ba,1,δ
m
t ,t m
  C2−mq/2 .
k k+1

Let Aa,b δ
s,t (B (ω)) be the area enclosed between the curve defined by
(Br (ω), Br (ω))srt and its chord for a, b ∈ { 1, . . . , d }.
a,δ b,δ

Let m be an integer such that m  n, and let k be an integer such that

k ∈ { 0, . . . , 2m − 1 }. Let r belongs to [tni , tni+1 ) with tni
tm
k . Then, it follows
easily from (10.1) that
 a,δ   a   
Br − B a,δ   − Btani  + Btani − Btam 
k
tm  Btn i+1 k

for a = 1, . . . , d, since Btδni = Btni for all i ∈ { 0, . . . , 2n }. The Doob and the
convexity inequalities imply that there exists a constant C depending only on
q such that for a = 1, . . . , d,
   
 a,δ  
E sup  Br − Btm a,δ q
 C2q−1 E |B a tni+1 − B a tni |q/2

k
r∈[tm m
k ,tk+1 ]

+ E |B a tni − B a tm |q/2
k (11.3)
 C2q−1 (2−nq/2 + 2−mq/2 )
 C2q 2−mq/2 .

The last inequality holds since m  n.

Let j and j  be such that tnj = tm n m
k and tj  = tk . With the Doob inequality,

there exists some constant C such that, when one uses (11.3), for a, b ∈
{ 1, . . . , d },
54 Antoine Lejay
"j  −1  #


a,δ 
b,δ q/2
E  Btn − Btm Btn − Btn 
a,δ b,δ
i k i+1 i
i=j
"j  −1  #


2
$ b % $ b %  q/4

CE Btani − Btam
k
B tn − B tn
i+1 i
i=j
 1/2
 a 
CE 
sup Br − Btam q |tm
k+1 − tk |
m q/2
k
r∈[tm m
k ,tk+1 ]
√
 C  C2q/2 2−mq/4 .
(11.4)
But it is easily verified that

δ
Aa,b
tm ,tm B
k k+1
 
j −1 −1
1

a 
b  1 j

b 
a 
= Btni − Btam Btn − Bt
b
n − Btni − Btbm Btni+1 − Btani .
4 i=j k i+1 i 4 i=j k

So, the inequality (11.4) implies that there exists some constant C depending
only on q such that
  !
E Aa,b δ q/2
tm ,tm (B )  C2−mq/2 for all m  n. (11.5)
k k+1

If m
n, the trajectory of B δ between the times tm m

k and tk+1 is a straight
a,b
line, so Atm ,tm (B δ ) = 0. So, (11.5) is true for all integers n and m.
k k+1

Since Aa,b (B δ ) is the antisymmetric part of Ba,b,2,δ , the convexity inequal-

ity implies that for a, b ∈ { 1, . . . , d }
  1  a,1,δ   b,1,δ   q/2
 a,b,2,δ q/2 q/2−1  a,b δ 
Btm m   B m m  · B m m  + 2 A m m (B )
k ,tk+1 2 tk ,tk+1 tk ,tk+1 tk ,tk+1

 C2−mq/2

for some constant C that depends only on q. For any β > 0,

m
2
−1

m β
2−mq/2 = mβ 2m(1−q/2) < +∞
m
1 i=0 m
1

since 1 − q/2 < 0. So, one deduces from (11.2) that


q !
q/2 !
sup E Var B1,δ + E Var B2,δ < +∞.
δ>0 q q/2

For 0  u  v  1, let m(v−u) is the smallest integer m such that [tm

k , tk+1 ] ⊂
m

[u, v] for some integer k. This quantity is deterministic and depends only on
u − v. Owing to (11.1), for β large enough and for all η > 0,
 An Introduction to Rough Paths 55
    

 q  2,δ q/2
sup E sup B1,δ
s,t
 + E sup  B s,t
 C mβ 2−m(1−q/2)
δ>0 |t−s|<η |t−s|<η
m
m(η)

for
 some constant C that depends only on q and β. Consequently, the series
β −m(1−q/2)
m
m(η) m 2 may be arbitrary small if η is chosen small enough.
Corollary 6.1 proves that (Bδ )δ>0 is tight in G p (R2 ) for all p > q > 2. But
we already know from the Wong-Zakai theorem that Bδs,t converges in proba-
bility to Bs,t uniformly in (s, t) ∈ ∆+ . Thus, (Bδ )δ>0 converges in probability
to B in G p (Rd ). 

12 Applications to stochastic analysis

The trajectories of stochastic processes are generally of finite p-variation with
p > 2. The typical case is of course that of the Brownian motion, whose
trajectories are α-Hölder continuous for all α < 1/2, and then of finite p-
variation as soon as p > 2. To apply the theory of rough paths to stochastic
processes, the main difficulty is generally to construct the equivalent of the
iterated integrals of the trajectories of the process.
The theory of rough paths has proved successful in many situations:
— Brownian motion and semi-martingales [32, 43].
— Reversible Markov Processes [1].
— Brownian motion on fractals [17].
— Fractional Brownian motion with Hurst exponent greater than 1/4 (note
that if the Hurst exponent h belongs to (1/4, 1/3], then third order iterated
integrals have to be considered) [9, 10].
— Stochastic processes generated by divergence form operators [1, 27].
— Lévy processes [48, 49, 50].
— Gaussian processes in infinite dimension [24].
— Free Brownian Motion [3].
Further results or extensions of previous results to geometric multiplicative
functionals also follow:
— Flow of diffeomorphisms [32, 34].
— Calculus of variations [35].
— Large and small deviation principle [25, 42],
— Support theorems [25].
— Application to homogenization [28].
Some applications to numerical analysis are also provided:
— Pathwise construction of solutions of SDEs using path-dependant meshes
[4, 15, 16].
— “Cubature formula” for computing weak solutions of SDEs numerically
[37, 47].
56 Antoine Lejay

Bibliographical and historical remarks

1. The article [32] is the synthesis of a series of works from T. Lyons and his co-
authors: [30, 31, 33, 43]. . .
2. The idea of “pathwise” stochastic calculus is an old idea: see for example [12,
14]. . . But the theory of rough paths brings for the first time a theory of pathwise
stochastic calculus valid for a large number of processes.
3. The use of the representation of the solution of some SDE using formally expo-
nentials of iterated integrals have been also widely used: see for example the works
[2, 11, 18, 21, 45] and related papers.
4. Stochastic Taylor expansions applied to numerical computations of solutions of
SDEs have also given rise to an abundant literature: see for example [20] and refer-
ences within.
5. (Related to Sect. 10.2). Shortly after being stated in [51], the theorem from
E. Wong and M. Zakai on the approximations of SDEs by ordinary differential equa-
tions attracted many interest, and was extended in many directions. E. J. McShane
was the first to show in [39] that different approximations of the trajectories may
lead not to construct the Stratonovich integral, but the Stratonovich integral and
a drift. An explicit construction is also given. There is now an important literature
on such a corrective drift: See for example [8, 23, 46] and references within.

References
Link to an article with a doi on the publisher site: <dx.doi.org/
doi>.
Link to an article on ArXiv: <front.math.ucdavis.edu/
arkiv>.
[1] Bass, R.F., Hambly, B., Lyons, T.J. (2002): Extending the Wong-Zakai theorem
to reversible Markov processes. J. Eur. Math. Soc., 4:3, 237–269. <doi:10.1007/
s100970200040>.
[2] Ben Arous, G. (1989): Flots et séries de Taylor stochastiques. Probab. Theory
Related Fields, 81:1, 29–77.
[3] Capitaine, M., Donati-Martin, C. (2001): The Lévy area process for the free
Brownian motion. J. Funct. Anal., 179:1, 153–169. <doi:10.1006/jfan.2000.
3679>.
[4] Castell, F., Gaines, J. (1995): An efficient approximation method for stochastic
differential equations by means of the exponential Lie series. Math. Comput.
Simulation, 38:1-3, 13–19. Probabilités numériques (Paris, 1992).
[5] Chen, K.-T. (1957): Integration of paths, geometric invariants and a generalized
Baker-Hausdorff formula. Ann. of Math. (2), 65, 163–178.
[6] Chen, K.-T. (1958): Integration of paths–a faithful representation, of paths by
noncommutative formal power series. Trans. Amer. Math. Soc., 89:2, 395–407.
[7] Chistyakov, V.V., Galkin, O.E. (1998): On maps of bounded p-variations with
p > 1. Positivity, 2, 19–45.
[8] Cohen, S., Estrade, A. (2000): Non-symmetric approximations for manifold-
valued semimartingales. Ann. Inst. H. Poincaré, Probab. Stat., 36:1, 45–70.
[9] Coutin, L., Qian, Z. (2000): Stochastic differential equations for fractional Brow-
nian motions. C. R. Acad. Sci. Paris Sér. I Math., 331:1, 75–80.
[10] Coutin, L., Qian, Z. (2002): Stochastic analysis, rough path analysis and
fractional brownian motions. Probab. Theory Related Fields, 122:1, 108–140.
<doi:0.1007/s004400100158>.
 An Introduction to Rough Paths 57

[11] Doss, H. (1977): Liens entre équations différentielles stochastiques et ordinaires.

Ann. Inst. H. Poincaré Sect. B (N.S.), 13:2, 99–125.
[12] Dudley, R.M., Norvais̆a, R. (1998): An introduction to p-variation and
Young integrals — with emphasis on sample functions of stochastic processes.
<www.maphysto.dk/cgi-bin/w3-msql/publications/genericpublication.
html%?publ=60>. Lecture given at the Centre for Mathematical Physics and
Stochastics, Department of Mathematical Sciences, University of Aarhus.
[13] Fliess, M., Normand-Cyrot, D. (1982): Algèbres de Lie nilpotentes, formule de
Baker-Campbell-Hausdorff et intégrales itérées de K. T. Chen. In: Séminaire de
Probabilités, XVI, pp 257–267. Springer, Berlin.
[14] Föllmer, H. (1981): Calcul d’Itô sans probabilités. In: Séminaire de Probabilités,
XV (Univ. Strasbourg, Strasbourg, 1979/1980), vol. 850 of Lecture Notes in
Math., pp 143–150. Springer, Berlin.
[15] Gaines, J. G., Lyons, T. J. (1994): Random generation of stochastic area inte-
grals. SIAM J. Appl. Math., 54:4, 1132–1146.
[16] Gaines, J. G., Lyons, T. J. (1997): Variable step size control in the numerical
solution of stochastic differential equations. SIAM J. Appl. Math., 57:5, 1455–
1484.
[17] Hambly, B. M., Lyons, T. J. (1998): Stochastic area for Brownian motion on
the Sierpinski gasket. Ann. Probab., 26:1, 132–148.
[18] Hu, Y. Z. (1990): Calculs formels sur les EDS de Stratonovitch. In: Séminaire de
Probabilités, XXIV, 1988/89, vol. 1426 of Lecture Notes in Math., pp 453–460.
Springer, Berlin.
[19] Ikeda, N., Watanabe, S. (1989): Stochastic Differential Equations and Diffusion
Processes. North-Holland, 2nd ed.
[20] Kloeden, P. E., Platen, E. (1992): Numerical solution of stochastic differential
equations. Springer-Verlag, Berlin.
[21] Kunita, H. (1980): On the representation of solutions of stochastic differential
equations. In: Séminaire de Probabilités, XIV (Paris, 1978/1979), vol. 784 of
Lecture Notes in Math., pp 282–304. Springer, Berlin.
[22] Kunita, H. (1990): Stochastic flows and stochastic differential equations. Cam-
bridge University Press.
[23] Kurtz, T. G., Protter, P. (1991): Wong-Zakai corrections, random evolutions,
and simulation schemes for SDEs. In: Stochastic analysis, pp 331–346. Academic
Press, Boston, MA.
[24] Ledoux, M., Lyons, T., Qian, Z. (2002): Lévy area of Wiener processes in
Banach spaces. Ann. Probab., 30:2, 546–578.
[25] Ledoux, M., Qian, Z., Zhang, T. (2002): Large deviations and support theo-
rem for diffusions via rough paths. Stochastic Process. Appl., 102:2, 265–283.
<doi:10.1016/S0304-4149(02)00176-X>.
[26] Lejay, A. (2002): On the convergence of stochastic integrals driven by processes
converging on account of a homogenization property. Electron. J. Probab., 7:18,
1–18. <www.math.washington.edu/~ejpecp>.
[27] Lejay, A. (2002): Stochastic differential equations driven by processes gener-
ated by divergence form operators. <www.iecn.u-nancy.fr/Preprint/publis/
lejay.Fri_Jun_21_17_03_23.html>. Institut Élie Cartan, Université Nancy I,
preprint no. 23.
[28] Lejay, A., Lyons, T.J. (2002): On the importance of the Lévy area for systems
controlled by converging stochastic processes. Application to homogenization.
In preparation.
58 Antoine Lejay

[29] Lévy, P. (1965): Processus stochastiques et mouvement brownien. Gauthier-

Villars & Cie, Paris, 2nd ed.
[30] Lyons, T. (1994): Differential equations driven by rough signals. I. An extension
of an inequality of L. C. Young. Math. Res. Lett., 1:4, 451–464.
[31] Lyons, T. J. (1995): The interpretation and solution of ordinary differential
equations driven by rough signals. In: Stochastic analysis (Ithaca, NY, 1993),
pp 115–128. Amer. Math. Soc., Providence, RI.
[32] Lyons, T.J. (1998): Differential equations driven by rough signals. Rev. Mat.
Iberoamericana, 14:2, 215–310.
[33] Lyons, T. J., Qian, Z. M. (1996): Calculus for multiplicative functionals, Itô’s
formula and differential equations. In: Itô’s stochastic calculus and probability
theory, pp 233–250. Springer, Tokyo.
[34] Lyons, T. J., Qian, Z. (1997): Flow equations on spaces of rough paths. J.
Funct. Anal., 149:1, 135–159.
[35] Lyons, T. J., Qian, Z. M. (1997): Calculus of variation for multiplicative func-
tionals. In: New trends in stochastic analysis (Charingworth, 1994), pp 348–374.
World Sci. Publishing, River Edge, NJ.
[36] Lyons, T., Qian, Z. (2002): System Control and Rough Paths. Oxford Mathe-
matical Monographs. Oxford University Press.
[37] Lyons, T., Victoir, N. (2002): Cubature on Wiener space. Oxford University
(preprint).
[38] Lyons, T., Zeitouni, O. (1999): Conditional exponential moments for iterated
Wiener integrals. Ann. Probab., 27:4, 1738–1749.
[39] McShane, E. J. (1972): Stochastic differential equations and models of ran-
dom processes. In: Proceedings of the Sixth Berkeley Symposium on Mathemat-
ical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971),
Vol. III: Probability theory, pp 263–294. Univ. California Press.
[40] Reutenauer, C. (1993): Free Lie algebras, vol. 7 of London Mathematical Society
Monographs. New Series. Oxford University Press.
[41] Ryan, R.A. (2002): Introduction to Tensor Products of Banach Spaces.
Springer-Verlag.
[42] Simon, T. (2002): Small deviations in p-variation norm for multidimen-
sional Lévy processes. <www.maths.univ-evry.fr/pages_perso/simon/publi.
html>. Université d’Évry (preprint).
[43] Sipiläinen, E.-M. (1993): A pathwise view of solutions of stochastic differential
equations. PhD thesis, University of Edinburgh.
[44] Stein, E. M. (1970): Singular integrals and differentiability properties of func-
tions. Princeton University Press, Princeton, N.J.
[45] Sussmann, H.J. (1978): On the gap between deterministic and stochastic ordi-
nary differential equations. Ann. Probability, 6:1, 19–41.
[46] Sussmann, H. J. (1991): Limits of the Wong-Zakai type with a modified drift
term. In: Stochastic analysis, pp 475–493. Academic Press, Boston, MA.
[47] Victoir, N. (2002): Asymmetric cubature formulae with few points in high
dimension for symmetric measures. Oxford University (preprint).
[48] Williams, D.R.E. (1998): Solutions of differential equations driven by cádlág
paths of finite p-variations. PhD thesis, Imperial College, London.
[49] Williams, D.R.E. (2000): Diffeomorphic flows driven by Lévy processes.
<arxiv:math.PR/0001016>. Imperial College, London (preprint).
[50] Williams, D.R.E. (2002): Path-wise solutions of SDE’s driven by Lévy pro-
cesses. Rev. Mat. Iberoamericana, 17:2, 295–330.
 An Introduction to Rough Paths 59

[51] Wong, E., Zakai, M. (1965): On the convergence of ordinary integrals to stochas-
tic integrals. Ann. Math. Statist., 36, 1560–1564.
[52] Young, L.C. (1936): An inequality of the Hölder type, connected with Stieltjes
integration. Acta Math., 67, 251–282.