Pathwise stochastic integrals for model free finance

Nicolas Perkowski
CEREMADE & CNRS UMR 7534
Universite Paris-Dauphine
[email protected]

David J. Promel
Humboldt-Universitat zu Berlin
Institut f
ur Mathematik
[email protected]

November 4, 2014

Abstract
We present two different approaches to stochastic integration in frictionless model
free financial mathematics. The first one is in the spirit of Itos integral and based on
a certain topology which is induced by the outer measure corresponding to the minimal
superhedging price. The second one is based on the controlled rough path integral. We
prove that every typical price path has a naturally associated Ito rough path, and
justify the application of the controlled rough path integral in finance by showing that
it is the limit of non-anticipating Riemann sums, a new result in itself. Compared to
the first approach, rough paths have the disadvantage of severely restricting the space of
integrands, but the advantage of being a Banach space theory.
Both approaches are based entirely on financial arguments and do not require any
probabilistic structure.

Contents
1 Introduction

2 Superhedging and typical price paths

2.1 The outer measure and its basic properties . . . . . . . . .
2.2 Arbitrage notions and link to classical mathematical finance
2.3 A topology on path-dependent functionals . . . . . . . . . .
2.4 Relation to Vovks outer measure . . . . . . . . . . . . . . .
3 Model free It
o integration

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

5
5
7
9
10
11

We would like to thank Mathias Beiglb

ock for suggesting us to look at the work of Vovk and at possible
connections to rough paths. We are grateful to Julio Backhoff, Peter Imkeller, and Johannes Ruf for helpful
discussions on the subject matter. The main part of the research was carried out while N.P. was employed by
Humboldt-Universit
at zu Berlin.

Supported by the Fondation Sciences Mathematiques de Paris (FSMP) and by a public grant overseen by
the French National Research Agency (ANR) as part of the Investissements dAvenir program (reference:
ANR-10-LABX-0098).

Supported by a Ph.D. scholarship of the DFG Research Training Group 1845.

4 Rough path integration for typical price paths

4.1 The Lyons-Gubinelli rough path integral . . . . . . . . . . . . . . . . . . . . .
4.2 Typical price paths as rough paths . . . . . . . . . . . . . . . . . . . . . . . .
4.3 The rough path integral as limit of Riemann sums . . . . . . . . . . . . . . .

14
15
18
22

A Pathwise Hoeffding inequality

27

B Davies criterion

28

Introduction

In this paper we use Vovks [Vov12] game-theoretic approach to mathematical finance to

develop two different approaches to stochastic integration in frictionless model free finance.
A priori the integration problem is highly non-trivial in the model free context since we do not
want to assume any probabilistic respectively semimartingale structure. Therefore, we do not
have access to It
o integration and most known techniques completely break down. There are
only two general solutions to the integration problem in a non-probabilistic continuous time
setting that we are aware of. One was proposed by [DS14] who simply restrict themselves to
trading strategies (integrands) of bounded variation. While this already allows to solve many
interesting problems, it is not a very natural assumption to make in a frictionless market
model. Indeed, while in [DS14] a general duality approach is developed for pricing pathdependent derivatives that are Lipschitz continuous in the supremum norm, this approach
does not allow to treat derivatives depending on the volatility.
Another interesting solution was proposed by [DOR13] (using an idea which goes back
to [Lyo95b]). They restrict the set of possible price paths to those admitting a quadratic
variation. This allows them to apply
R Follmers pathwise Ito calculus [Fol81] to define pathwise
stochastic integrals of the form F (S)dS. In [Lyo95b] that approach was used to derive
prices for American and European options under volatility uncertainty. In [DOR13] the given
data is a finite number of European call and put prices and the derivative to be priced is
a weighted variance swap. The restriction to the set of paths with quadratic variation is
justified by referring to Vovk [Vov12], who proved that typical price paths (to be defined
below) admit a quadratic variation.
In our first approach we do not restrict the set of paths and work on the space of ddimensional continuous paths (which represent the possible asset price trajectories). We follow
Vovk who introduces an outer measure on . The crucial point is that this outer measure
is defined as a minimal superhedging price (in a suitable sense), and therefore has a purely
financial interpretation and does not come from an artificially imposed probabilistic structure.
Our first observation is that Vovks outer measure allows us to define a topology on processes
on , and that the natural It
o integral on step functions is in a certain sense continuous
in that topology. This allows us to extend the integral to c`adl`ag adapted integrands, and we
call the resulting integral model free Ito integral. We stress that the entire construction is
based on purely financial arguments.
Let us also stress that it is the continuity of our integral which is the most important
aspect. Without reference to any topology the construction would certainly not be very
useful, since already in the classical probabilistic setting virtually all applications of the It
o
integral (SDEs, stochastic optimization, duality theory, . . . ) are based on the fact that it is
a continuous operator.
2

This also motivates our second approach, which is more in the spirit of [Lyo95b, DOR13,
DS14]. While in the first approach we do have a continuous operator, it is only continuous
with respect to a sequence of pseudometrics and it seems impossible to find a Banach space
structure that is compatible with it. However, using the model free Ito integral we are able to
show that every typical price path has a natural Ito rough path associated to it. Since in
financial applications we can always restrict ourselves to typical price paths, this observation
opens the door for the application of the controlled rough path integral [Lyo98, Gub04] in
model free finance. Controlled rough path integration has the advantage of being an entirely
linear Banach space theory which simultaneously extends
the Riemann-Stieltjes integral of S against functions of bounded variation which was
used by [DS14];
the Young integral [You36]: typical price paths have finite p-variation for every p > 2,
and therefore
for every F of finite q-variation for q < 2 (so that 1/p + 1/q > 1), the
R
integral F dS is defined as limit of non-anticipating Riemann sums;
Follmers [F
ol81] pathwise It
o integral which was used by [Lyo95b, DOR13]. That this
last integral is a special case of the controlled rough path integral is, to the best of
our knowledge, proved rigorously for the first time in this paper, although also [FH14]
contains some preliminary observations in that direction.
In other words, our second approach covers all previously known techniques of integration in
model free financial mathematics, while the first approach is much more general but at the
price of leaving the Banach space world.
There is only one pitfall: the rough path integral is usually defined as a limit of compensated Riemann sums which have no obvious financial interpretation. This sabotages our
entire philosophy of only using financial Rarguments. That is why we show that under some
weak condition every rough path integral F dS is given as limit of non-anticipating Riemann
sums that do not need to be compensated the first time that such a statement is shown for
general rough path integrals. While this will not change anything in concrete applications, it
is of utmost importance from a philosophical point of view. Indeed, the justification for using
the Ito integral in classical financial mathematics is crucially based on the fact that it is the
limit of non-anticipating Riemann sums, even if in every day applications one never makes
reference to that; see for example the discussion in [Lyo95b].
Plan of the paper
Below we present a very incomplete list of solutions to the stochastic integration problem
under model uncertainty and in a discrete time model free context (both a priori much simpler
problems than the continuous time model free case), and we introduce some notations and
conventions that will be used throughout the paper. In Section 2 we briefly recall Vovks
game-theoretic approach to mathematical finance and introduce our outer measure. We also
construct the topology induced by the outer measure. Section 3 is devoted to the construction
of the model free It
o integral. Section 4 recalls some basic results from rough path theory,
and continues by constructing rough paths associated to typical price paths. Here we also
prove that the rough path integral is given as a limit of non-anticipating Riemann sums. We
also compare F
ollmers pathwise It
o integral with the rough path integral and prove that the
latter is an extension of the former. Appendix A recalls Vovks pathwise Hoeffding inequality.
3

In Appendix B we show that a result of Davie which also allows to calculate rough path
integrals as limit of Riemann sums is a special case of our results in Section 4.
Stochastic integration under model uncertainty
The first works which studied the option pricing problem under model uncertainty were
[ALP95] and [Lyo95b], both considering the case of volatility uncertainty. As described above,
[Lyo95b] is using F
ollmers pathwise Ito integral, while in [ALP95] the problem is reduced to
the classical setting by deriving a worst case model for the volatility.
A powerful tool in financial mathematics under model uncertainty is Karandikars pathwise
construction of the It
o integral [Kar95, Bic81] which allows to construct the Ito integral of
a c`adl`ag integrand simultaneously under all semimartingale measures. The crucial point
that makes the construction useful is that the Ito integral is a continuous operator under
every semimartingale measure. While its pathwise definition would allow us to use the same
construction also in a model free setting, it is not even clear what the output should signify
in that case (for example the construction depends on a certain sequence of partitions and
changing the sequence will change the output). Certainly it is not obvious whether the
Karandikar integral is continuous in any topology once we dispose of semimartingale measures.
A more general pathwise construction of the Ito integral was given in [Nut12], but it suffers
from the same drawbacks with respect to applications in model free finance.
A general approach to stochastic analysis under model uncertainty was put forward in
[DM06], and it is based on quasi sure analysis. While this approach is extremely helpful when
working under model uncertainty, it also does not allow us to define stochastic integrals in a
model free context.
In a related but slightly different direction, in [CDGR11] non-semimartingale models are
studied (which do not violate arbitrage assumptions if the set of admissible strategies is
restricted). While the authors work under one fixed probability measure, the fact that their
price process is not a semimartingale prevents them from using Ito integrals, a difficulty which
is overcome by working with the Russo-Vallois integral [RV93].
Of course all these technical problems disappear if we restrict ourselves to discrete time,
and indeed in that case [BHLP13] develop an essentially fully satisfactory duality theory for
the pricing of derivatives under model uncertainty.
Notation and conventions
Throughout the paper we fix T (0, ) and we write := C([0, T ], Rd ) for the space of
d-dimensional continuous paths. The coordinate process on is denoted by St () = (t),
t [0, T ]. For i {1, . . . , d}, we also write Sti () = i (t), where = ( 1 , . . . , d ). The
filtration (Ft )t[0,T ] is defined as Ft := (Ss : s t), and we set F := FT . Stopping times
and the associated -algebras F are defined as usually.
Unless explicitly stated otherwise, inequalities of the type Ft Gt , where F and G
processes on , are supposed to hold for all , and not modulo null sets, as it is usually
assumed in stochastic analysis.
The indicator function of a set A is denoted by 1A .
A partition of [0, T ] is a finite set of time points, = {0 = t0 < t1 < < tm = T }.
Occasionally, we will identify
with the set of intervals {[t0 , t1 ], [t1 , t2 ], . . . , [tm1 , tm ]}, and
P
write expressions like [s,t] .
4

For f : [0, T ] Rn and t1 , t2 [0, T ], denote ft1 ,t2 := f (t2 ) f (t1 ) and define the pvariation of f restricted to [s, t] [0, T ] as
kf kpvar,[s,t] := sup

 m1
X

|ftk ,tk+1 |p

1/p


: s = t0 < < tm = t, m N ,

p > 0,

(1)

k=0

(possibly taking the value +). We set kf kpvar := kf kpvar,[0,T ] . We write T = {(s, t) :
0 s t T } for the simplex and define the p-variation of a function g : T Rn in the
same manner, replacing ftk ,tk+1 in (1) by g(tk , tk+1 ).
For > 0, the space C consists of those functions that are bc times continuously
differentiable, with ( bc)-H
older continuous partial derivatives of order bc. The space
Cb consists of those functions in C that are bounded, together with their partial derivatives,
and we define the norm kkCb by setting
kf kCb :=

bc
X

kDk f k + kDbc f kbc ,

k=0

where kk denotes the -H

older norm for (0, 1), and kk denotes the supremum norm.
P
For x, y Rd , we write xy := R di=1 xi yi for the usual inner product. However, often
we will encounter terms of the form SdS or Ss Ss,t for s, t [0, T ], where we recall that S
denotes
R i jthe coordinate process on . Those expressions are to be understood as the matrix
( S dS )1i,jd , and similarly for Ss Ss,t . The interpretation will be usually clear from the
context, otherwise we will make a remark to clarify things.
We use the notation a . b if there exists a constant c > 0, independent of the variables
under consideration, such that a c b, and we write a ' b if a . b and b . a. If we want to
emphasize the dependence of c on the variable x, then we write a(x) .x b(x).
We make the convention that 0/0 := 0 := 0 and inf := .

2
2.1

Superhedging and typical price paths

The outer measure and its basic properties

In a recent series of papers, Vovk [Vov08, Vov11, Vov12] has introduced a model free, hedging
based approach to mathematical finance that uses arbitrage considerations to examine which
properties are satisfied by typical price paths. This is achieved with the help of an outer
measure given by the cheapest superhedging price.
Recall that T (0, ) and = C([0, T ], Rd ) is the space of continuous paths, with
coordinate process S, natural filtration (Ft )t[0,T ] , and F = FT . A process H : [0, T ] Rd
is called a simple strategy if there exist stopping times 0 = 0 < 1 < . . . , and Fn -measurable
bounded functions Fn : Rd , such that for every we have n () = for all but
finitely many n, and such that
Ht () =

Fn ()1(n (),n+1 ()] (t).

n=0

In that case, the integral

(H S)t () :=

Fn ()(Sn+1 t () Sn t ()) =

n=0

Fn ()Sn t,n+1 t ()

n=0

is well defined for all , t [0, T ]. Here Fn ()Sn t,n+1 t () denotes the usual inner
product on Rd . For > 0, a simple strategy H is called -admissible if (H S)t () for
all , t [0, T ]. The set of -admissible simple strategies is denoted by H .
Definition 2.1. The outer measure of A is defined as the cheapest superhedging price
for 1A , that is
n
o
n
n
P (A) := inf > 0 : (H )nN H s.t. lim inf ( + (H S)T ()) 1A () .
n

A set of paths A is called a null set if it has outer measure zero.

The term outer measure will be justified by Lemma 2.3 below. Our definition of P is
very similar to the one used by Vovk [Vov12], but not quite the same. For a discussion see
Section 2.4 below.
By definition, every It
o stochastic integral is the limit of stochastic integrals against simple
strategies. Therefore, our definition of the cheapest superhedging price is essentially the same
as in the classical setting, with one important difference: we require superhedging for all
, and not just almost surely.
Remark 2.2 ([Vov12], p. 564). An equivalent definition of P would be


n
n
e
P (A) := inf > 0 : (H )nN H s.t. lim inf sup ( + (H S)t ()) 1A () .
n t[0,T ]

Clearly Pe P . To see the opposite inequality, let Pe(A) < . Let (H n )nN H be a sequence
of simple strategies such that lim inf n supt[0,T ] ( + (H n S)t ) 1A , and let > 0. Define
n := inf{t [0, T ] : ++(H n S)t 1}. Then the stopped strategy Gnt () := Htn ()1t<n ()
is in H H+ and
lim inf ( + + (Gn S)T ()) lim inf 1{++supt[0,T ] (H n S)t 1} () 1A ().
n

Therefore P (A) + , and since > 0 was arbitrary P Pe, and thus P = Pe.
Lemma 2.3 ([Vov12], Lemma 4.1). P is in fact an outer measure, i.e. a nonnegative function
defined on the subsets of such that
- P () = 0;
- P (A) P (B) if A B;
S
P
- if (An )nN is a sequence of subsets of , then P ( n An ) n P (An ).
Proof. Monotonicity and P () = 0 are obvious. So let (An ) be a sequence of subsets of .
Let > 0, n N, and let (H n,m )mN be a sequence of (P (An ) + 2n1 )-admissible simple

n,m S) ) 1 . Define for m N the

strategies
such that lim inf m (P (An ) + 2n1 + (H
T
An
P
P
n,m . Then by Fatous lemma
( n P (An ) + )-admissible simple strategy Gm := m
H
n=0

lim inf
m

X

P (An ) + + (G S)T

n=0

k
X

P (An ) + 2n1 + lim inf (H n,m S)T

m

n=0

1Sk

n=0

An

for all k N. Since the left hand side does not depend on k, we can replace 1Sk An by
n=0
1Sn An and the proof is complete.
Maybe the most important property of P is that there exists an arbitrage interpretation
for sets with outer measure zero:
Lemma 2.4. A set A is a null set if and only if there exists a sequence of 1-admissible
simple strategies (H n )n H1 such that
lim inf (1 + (H n S)T ) 1A (),
n

(2)

where we recall that by convention 0 = 0.

Proof. If such a sequence exists, then we can scale it down by an arbitrary factor > 0 to
obtain a sequence of strategies in H that superhedge 1A , and therefore P (A) = 0.
If conversely P (A) = 0, then for every n N there exists a sequence of simple strategies
n1 +lim inf
n,m ) 1 () for all . Define
(H n,m )mN
m (H
T
A
2n1 such that 2
Pm Hn,m
m
m
, so that G H1 . For every k N we obtain
G := n=0 H
lim inf (1 + (Gm S)T )
m

k
X

(2n1 + lim inf (H n,m S)T ) (k + 1)1A .

m

n=0

Since the left hand side does not depend on k, the sequence (Gm ) satisfies (2).
In other words, if a set A has outer measure 0, then we can make infinite profit by investing
in the paths from A, without ever risking to lose more than the initial capital 1.
This motivates the following definition:
Definition 2.5. We say that a property (P) holds for typical price paths if the set A where
(P) is violated is a null set.
The basic idea of Vovk, which we shall adopt in the following, is that we only need to
concentrate on typical price paths. Indeed, non-typical price path can be excluded since
they are in a certain sense too good to be true: they would allow investors to realize infinite
profit while at the same time taking essentially no risk.

2.2

Arbitrage notions and link to classical mathematical finance

Before we continue, let us discuss different notions of arbitrage and argue that our outer
measure is an interesting object to study. We start by observing that P is an outer measure
which simultaneously dominates all local martingale measures on .

Propostion 2.6 ([Vov12], Lemma 6.3). Let P be a probability measure on (, F), such that
the coordinate process S is a P-local martingale, and let A F. Then P(A) P (A).
Proof. Let > 0 and let (H n )nN H be such that lim inf n ( + (H n S)T ) 1A . Then
P(A) EP [lim inf ( + (H n S)T )] lim inf EP [ + (H n S)T ] ,
n

where in the last step we used that + (H n S) is a nonnegative P-local martingale and thus
a P-supermartingale.
This already indicates that P -null sets are quite degenerate, in the sense that they are
null sets under all local martingale measures. However, if that was the only reason for our
definition of typical price paths, then a definition based on model free arbitrage opportunities
would be equally valid. A map X : [0, ) is a model free arbitrage opportunity if X is not
identically 0 and if there exists c > 0 and a sequence (H n ) Hc such that lim inf n (H n
S)T () = X() for all . See [DH07, ABPS14] where (a similar) definition is used in the
discrete time setting.
It might then appear more natural to say that a property holds for typical price paths if
the indicator function of its complement is a model free arbitrage opportunity, rather than
working with Definition 2.5. This arbitrage definition would also imply that any property
which holds for typical price paths is almost surely satisfied under every local martingale
measure. Nonetheless we decidedly claim that our definition is the correct one. First of all
the arbitrage definition would make our life much more difficult, since it seems not very easy to
work with. But of course this is only a convenience and cannot possibly serve as justification
of our approach. Instead, we argue by relating the two notions to classical mathematical
finance.
For that purpose recall the fundamental theorem of asset pricing [DS94]: If P is a probability measure on (, F) under which S is a semimartingale, then there exists an equivalent
measure Q such that S is a Q-local martingale if and only if S admits no free lunch with
vanishing risk (NFLVR). But (NFLVR) is equivalent to the two conditions no arbitrage (NA)
(intuitively: no profit without risk) and no arbitrage opportunities of the first kind (NA1)
(intuitively: no very large profit with a small risk). The (NA) property holds if for every
c > 0 and every sequence (H n ) Hc for which limn (H n S)T () exists for all we have
P(limn (H n S)T < 0) > 0 or P(limn (H n S)T = 0) = 1. The (NA1) property holds if
{1 + (H S)T : H H1 } is bounded in P-probability, i.e. if
lim sup P(1 + (H S)T c) = 0.

c HH1

Strictly speaking this is (NA1) with simple strategies, but as observed by [KP11] (NA1) and
(NA1) with simple strategies are equivalent; see also [IP11]. In the case of continuous S, the
equivalence of (NA1) and (NA1) with simple strategies had previously been shown by [Ank05],
Corollary 8.3.2, although here the result is formulated in a slightly different language.
Now the arbitrage definition of typical price paths corresponds to (NA), while our definition corresponds to (NA1):
Propostion 2.7. Let A F be a null set, and let P be a probability measure on (, F) such
that the coordinate process satisfies (NA1). Then P(A) = 0.

Proof. Let (H n )nN H1 be such that 1 + lim inf n (H n S)T 1A . Then for every c > 0



P(A) = P A lim inf (H n S)T > c sup P({(H S)T > c}).
n

HH1

By assumption, the right hand side converges to 0 as c and thus P(A) = 0.

Remark 2.8. Proposition 2.7 is actually a consequence of Proposition 2.6, because if S
satisfies (NA1) under P, then there exists a dominating measure Q  P, such that S is a
Q-local martingale. See [Ruf13] for the case of continuous S, and [IP11] for the general case.
The crucial point is now that (NA1) is the essential property which every sensible market
model has to satisfy, whereas (NA) is nice to have but not strictly necessary. Indeed, (NA1) is
equivalent to the existence of an unbounded utility function such that the maximum expected
utility is finite [KK07, IP11]. (NA) is what is needed in addition to (NA1) in order to obtain
equivalent local martingale measures. But there are perfectly viable models which violate
(NA), for example the three dimensional Bessel process. By working with the arbitrage
definition of typical price paths, we would in a certain sense ignore these models.

2.3

A topology on path-dependent functionals

It will be very useful to introduce a topology on functionals on . For that purpose let us
identify X, Y : R if X = Y for typical price paths. Clearly this defines an equivalence
relation, and we write L0 for the space of equivalence classes. We then introduce the analog
of convergence in probability in our context: (Xn ) converges in outer measure to X if
lim P (|Xn X| > ) = 0

for all > 0.

We follow [Vov12] in defining an expectation operator. If X : [0, ], then

n
o
E[X] := inf > 0 : (H n )nN H s.t. lim inf ( + (H n S)T ()) X() .
n

In particular, P (A) = E[1A ]. The expectation E is countably subadditive, monotone, and

positively homogeneous. It is an easy exercise to verify that
d(X, Y ) := E[|X Y | 1]
defines a metric on L0 .
Lemma 2.9. The distance d metrizes the convergence in outer measure. More precisely, a
sequence (Xn ) converges to X in outer measure if and only if limn d(Xn , X) = 0. Moreover,
(L0 , d) is a complete metric space.
Proof. The arguments are the same as in the classical setting. Using subadditivity and
monotonicity of the expectation operator, we have
P (|Xn X| ) E[|Xn X| 1] P (|Xn X| > ) +
for all (0, 1], showing that convergence in outer measure is equivalent to convergence with
respect to d.
9

As for completeness, let (Xn ) be a Cauchy sequence with respect to d. Then there exists
a subsequence (Xnk ) such that d(Xnk , Xnk+1 ) 2k for all k, so that
i X
hX
X
d(Xnk , Xnk+1 ) < ,
E
(|Xnk Xnk+1 | 1)
E[|Xnk Xnk+1 | 1] =
k

which means that (Xnk ) converges for typical price paths. Define X := lim inf k Xnk . Then
we have for all n and k
X
d(Xn` , Xn`+1 ) d(Xn , Xnk ) + 2k .
d(Xn , X) d(Xn , Xnk ) + d(Xnk , X) d(Xn , Xnk ) +
`k

Choosing n and k large, we see that d(Xn , X) tends to 0.

2.4

Relation to Vovks outer measure

Our definition of the outer measure P is not exactly the same as Vovks [Vov12]. We find our
definition more intuitive and it also seems to be easier to work with. However, since we rely
on some of the results established by Vovk, let us compare the two notions.
For > 0, Vovk defines the set of processes
X


X
k
k
S :=
H : H Hk , k > 0,
k = .
k=0

For every G =

k0 H

k=0

S , every and every t [0, T ], the integral

(G S)t () :=

(H k S)t () =

k0

(k + (H k S)t ())

k0

is well defined and takes values in [, ]. Vovk then defines for A the cheapest
superhedging price as


Q(A) := inf > 0 : G S s.t. + (G S)T 1A .
This definition corresponds to the usual construction of an outer measure from an outer
content (i.e. an outer measure which is only finitely subadditive and not countably subadditive); see [Fol99], Chapter 1.4, or [Tao11], Chapter 1.7. Here, the outer content is given
by the cheapest superhedging price using only simple strategies. It is easy to see that P is
dominated by Q:
Lemma 2.10. Let A . Then P (A) Q(A).
P
P
Proof. P
Let G = k H k , with H k Hk and k k = , and assume that + (G S)T 1A .
Then ( nk=0 H k )nN defines a sequence of simple strategies in H , such that


lim inf +
n

n
X

  
Hk S
= + (G S)T 1A .
T

k=0

So if Q(A) < , then also P (A) , and therefore P (A) Q(A).

10

Corollary 2.11. For every p > 2, the set Ap := { : kS()kpvar = } has outer
measure zero, that is P (Ap ) = 0.
Proof. Theorem 1 of Vovk [Vov08] states that Q(Ap ) = 0, so P (Ap ) = 0 by Lemma 2.10.
It is a remarkable result of [Vov12] that if = C([0, ), R) (i.e. if the asset price process
is one-dimensional), and if A is invariant under time changes and such that S0 () = 0
for all A, then A F and Q(A) = P(A), where P denotes the Wiener measure. This can
be interpreted as a pathwise Dambis Dubins-Schwarz theorem.

Model free It
o integration

The present section is devoted to the construction of a model free Ito integral. The main
ingredient is a (weak) type of model free Ito isometry, which allows us to estimate the integral
against a step function in terms of the amplitude of the step function and the quadratic
variation of the price path. Based on the topology introduced in Section 2.3, it is then easy
to extend the integral to c`
adl`
ag integrands by a continuity argument.
Since we are in an unusual setting, let us spell out the following standard definitions:
Definition 3.1. A process F : [0, T ] Rd is called adapted if the random variable
7 Ft () is Ft -measurable for all t [0, T ].
The process F is said to be c`
adl`
ag if the sample path t 7 Ft () is c`adl`ag for all .
To prove our weak It
o isometry, we will need an appropriate sequence of stopping times:
Let = ( 1 , . . . , d ) and n N. For each i = 1, . . . , d define inductively


n,i
() := inf t kn,i : | i (t) i (kn,i )| 2n , k N.
k+1
0n,i () := 0,
Since we are working with continuous paths and we are considering entrance times into closed
sets, the maps ( n,i ) are indeed stopping times, despite the fact that (Ft ) is neither complete
nor right-continuous. Denote n,i := {kn,i : k N}. To obtain an increasing sequence of
partitions, we take the union of the ( n,i ), that is we define 0n := 0 and then
n
()
k+1


:= min t >

kn ()

:t

d
[

n,i


() ,

k N,

(3)

i=1

and we write n := {kn : k N} for the corresponding partition.

Lemma 3.2 ([Vov11], Theorem 4.1). For typical price paths , the quadratic variation
along ( n,i ())nN exists. That is,
Vtn,i () :=

2
n,i
i (k+1
t) i (kn,i t) ,

t [0, T ],

n N,

k=0

converges uniformly to a function hS i i() C([0, T ], R) for all i {1, . . . , d}.

For later reference, let us estimate Ntn := max{k N : kn t}, the number of stopping
times kn 6= 0 in n with values in [0, t]:
11

Lemma 3.3. For all , n N, and t [0, T ], we have

2

2n

Ntn ()

d
X

Vtn,i () =: Vtn ().

i=1

Proof. For i {1, ..., d} define Ntn,i := max{k N : kn,i t}. Since i is continuous, we
n,i
n,i
have | i (k+1
) i (kn,i )| = 2n as long as k+1
T . Therefore, we obtain
Ntn ()

d
X

n,i

Ntn,i ()

i=1

()1
d Nt X
X
i=1

k=0

n,i
(k+1
)
22n

2
(kn,i )

2n

d
X

Vtn,i ().

i=1

We will start by constructing the integral against step functions, which are defined similarly as simple strategies, except possibly unbounded: A process F : [0, T ] Rd is called
a step function if there exist stopping times 0 = 0 < 1 < . . . , and Fn -measurable functions
Fn : Rd , such that for every we have n () = for all but finitely many n, and
such that

X
Ft () =
Fn ()1[n (),n+1 ()) (t).
n=0

For notational convenience we are now considering the interval [n (), n+1 ()) which is closed
on the right hand side. This allows us define the integral
(F S)t :=

Fn Sn t,n+1 t =

n=0

Fn Sn t,n+1 t ,

t [0, T ].

n=0

The following lemma will be the main building block in the construction of our integral.
Lemma 3.4 (Model free It
o isometry). Let a > 0 and let F be a step function. Then for all
a, b, c > 0 we have

P {k(F S)k ab c} {kF ()k a} {hSiT c} 2 exp(b2 /(2d)),

where the set {hSiT c} should be read as {hSiT = limn VTn exists and satisfies hSiT c}.
P
Proof. Assume Ft =
n=0 Fn 1[n ,n+1 ) (t) and set a := inf{t > 0 : |Ft | a}. Let n N and
n
define 0 := 0 and then for k N


nk+1 := min t > nk : t n {m : m N} ,
where we recall that n = {kn : k N} is P
the n-th generation dyadic partition generated by
S. For t [0, T ], we have (F S)a t = k Fnk Sa nk t,a nk+1 t , and by the definition of
n () and a we get



sup Fnk Sa nk t,a nk+1 t  a d2n .
t[0,T ]

Hence, the pathwise Hoeffding inequality, Lemma A.1 in Appendix A, yields for every R
the existence of a 1-admissible simple strategy H ,n H1 such that


2 (n )
,n
2n 2
1 + (H S)t exp (F S)a t (Nt
+ 1)2
a d =: E,n
a t
2
12

for all t [0, T ], where

(n )

Nt

( )

:= max{k : nk t} Ntn + Nt

:= Ntn + max{k : k t}.

By Lemma 3.3, we have Ntn 22n Vtn , so that



2 n 2
2 ( )
,n
2n 2
Ea t exp (F S)t VT a d (NT + 1)2
a d .
2
2

If now k(F S)k ab c, kF ()k a and hSiT c, then




Et,n + Et,n
1
2 2
lim inf sup
exp ab c ca d .
n t[0,T ]
2
2
2

The argument inside the exponential is maximized for = b/(a cd), in which case we obtain
1/2 exp(b2 /(2d)). The statement now follows from Remark 2.2.
Of course, we did not actually establish an isometry but only an upper bound for the
integral. But this estimate is the key ingredient which allows us to extend the model free
Ito integral to more general integrands, and it is this analogy to the classical setting that the
term model free It
o isometry alludes to.
Let us extend the topology of Section 2.3 to processes: we identify X, Y : [0, T ] Rm
if for typical price paths we have Xt = Yt for all t [0, T ], and we write L0 ([0, T ], Rm ) for the
resulting space of equivalence classes which we equip with the distance
d (X, Y ) := E[kX Y k 1].
Ideally, we would like the stochastic integral on step functions to be continuous with
respect to d . However, using Proposition 2.6 it is easy to see that P (k((1/n) S)k > ) = 1
for all n N and > 0. This is why we also introduce for c > 0 the pseudometric
dc (X, Y ) := E[(kX Y k 1)1hSiT c ] d (X, Y ),
and then
dcpct (X, Y ) :=

2n d2n (X, Y ) d (X, Y ).

n=1

For step functions F, G, we get from Lemma 3.4

dc ((F S), (G S)) P ({k((F G) S)k ab c} {kF Gk a} {hSiT c})

dc (F, G)
+
+ ab c
a
 b2  d (F, G)

c
2 exp
+
+ ab c
2d
a
p
p
whenever a, b > 0. Setting a = dc (F, G) and b = d| log a|, we deduce that

dc ((F S), (G S)) . (1 + c)dc (F, G)1/2

(4)
for all > 0, and in particular
dcpct ((F S), (G S)) .

2n/2 d2n (F, G)1/2 . d (F, G)1/2 .

n=1

13

adl`
ag process with values in Rd . Then there exists
RTheorem 3.5. Let F be an adapted, c`
n
n
F dS L0 ([0, T ], R) such thatR for every sequence of step functions
R (F ) with limn d (F , F ) =
n
0 we have limn dcpct ((F S), F dS) = 0. The integral
process F dS is continuous for typR
ical price paths, and there exists a representative F dS which is adapted, although it may
Rt
R
R
take the values .We usually write 0 Fs dSs := F dS(t), and we call F dS the model
free Ito integral of RF with respect to S.
The map F 7 F dS is linear, satisfies
Z
Z

dcpct
F dS, GdS . d (F, G)1/2
for all > 0, and the model free It
o isometry extends to this setting:
n Z
o


P k F dSk ab c {kF k a} {hSiT c} 2 exp(b2 /(2d))

for all a, b, c > 0.
Proof. Everything follows in a straightforward way from (4) in combination with Lemma 2.9.
We have to use the fact that F is adapted and c`adl`ag in order to approximate it uniformly
by step functions.
Another simple consequence of the model free Ito isometry is a strengthened version of
Karandikars [Kar95] pathwise It
o integral.
Corollary 3.6. In the setting of Theorem 3.5, let (F m )mN be a sequence of step functions
with kF m () F ()k cm for all and all m N. Then for typical price paths
there exists a constant C() > 0 such that
Z


p
m

(5)
(F S)() F dS() C()cm log m

forRall m N. So, if (cm log m) converges to 0, then for typical price paths (F m S) converges
to F dS.
Proof. For c > 0 the model free It
o isometry gives



Z
p

1
m
P
k(F S) F dSk cm 4d log m c {hSiT c} 2 .
m
Since this is summable in m, the claim follows from Borel Cantelli (which only requires
countable subadditivity).
Remark 3.7. The speed of convergence (5) is better than the one that can be obtained using
the arguments in [Kar95], where the summability of (cm ) is needed.

Rough path integration for typical price paths

Our second approach to model free stochastic integration is based on the rough path integral,
which has the advantage of being a continuous linear operator between Banach spaces. The
disadvantage is that we have to restrict the set of integrands to those locally looking like S,
modulo a smoother remainder. Our two main results in this section are that every typical
price path has a naturally associated Ito rough path, and that the rough path integral can be
constructed as limit of Riemann sums.
Let us start by recalling the basic definitions and results of rough path theory.
14

4.1

The Lyons-Gubinelli rough path integral

Here we follow more or less the lecture notes [FH14], to which we refer for a gentle introduction
to rough paths. More advanced monographs are [LQ02, LCL07, FV10]. The main difference
to [FH14] in the derivation below is that we use p-variation to describe the regularity, and not
Holder continuity, because it is not true that all typical price paths are Holder continuous.
Also, we make an effort to give reasonably sharp results, whereas in [FH14] the focus lies
more on the pedagogical presentation of the material. We stress that in this subsection we
are merely collecting classical results.
Definition 4.1. A control function is a continuous map c : T [0, ) with c(t, t) = 0 for
all t [0, T ] and such that c(s, u) + c(u, t) c(s, t) for all 0 s u t T .
Observe that if f : [0, T ] Rd satisfies |fs,t |p c(s, t) for all (s, t) T , then the pvariation of f is bounded from above by c(0, T ).
Definition 4.2. Let p (2, 3). A p-rough path is a map S = (S, A) : T Rd Rdd such
that Chens relation
S i (s, t) = S i (s, u) + S i (u, t)

and

Ai,j (s, t) = Ai,j (s, u) + Ai,j (u, t) + S i (s, u)S j (u, t)

holds for all 1 i, j d and 0 s u t T and such that there exists a control function
c with
|S(s, t)|p + |A(s, t)|p/2 c(s, t)
(in other words S has finite p-variation and A has finite p/2-variation). In that case we call
A the area of S.
Remark 4.3. Chens relation simply states that S is the increment of a function, that
is S(s, t) = S(0, t) S(0, s) =: S(t) S(s), and that for all i, j there exists a function
j
f i,j : [0, T ] R such that Ai,j (s, t) = f i,j (t) f i,j (s) S i (s)Ss,t
. Indeed, it suffices to set
j
i,j
i,j
i
f (t) = A (0, t) + S (0)S0,t .
Remark 4.4. The (strictly speaking incorrect) name area stems from the fact that if S is
smooth and two-dimensional and if
Z t
Z t Z r2
i
Ai,j (s, t) =
dS i (r1 )dS j (r2 ) =
Ss,r
dS j (r2 ),
2
s

then the antisymmetric part of A(s, t) corresponds to the algebraic area enclosed by the curve
(S(r))r[s,t] . It is a deep insight of Lyons [Lyo98], proving a conjecture of F
ollmer, that the
area is exactly the additional information which is needed to solve differential equations driven
by S in a pathwise continuous manner, and to construct stochastic integrals as continuous
maps. Actually, [Lyo98] solves a much more general problem and proves that if the driving
signal is of finite p-variation for some p > 1, then it has to be equipped with the iterated
integrals up to order bpc 1 to obtain a continuous integral map. The for us relevant case
p (2, 3) was already treated in [Lyo95a].
Example 4.5. If S is a continuous semimartingale and if we set S(s, t) = Ss,t as well as
Z t Z r2
Z t
i,j
i
j
i
A (s, t) =
dS (r1 )dS (r2 ) =
Ss,r
dS j (r2 ),
2
s

15

where the integral can be understood either in the Ito or in the Stratonovich sense, then almost
surely S = (S, A) is a p-rough path for all p (2, 3). This is shown in [CL05], and we will
give a simplified model free proof below (indeed we will show that every typical price path is
a p-rough path for all p (2, 3), from where the statement about continuous semimartingales
easily follows).
From now on we fix p > 2 and we assume that S is a p-rough path. Gubinelli [Gub04]
observed that for every rough path there is a naturally associated Banach space of integrands,
the space of controlled paths. Heuristically, a path F is controlled by S, if it locally looks
like S, modulo a smooth remainder. The precise definition is:
Definition 4.6. Let q > 0 be such that 2/p + 1/q > 1. Let F : [0, T ] Rn and F 0 : [0, T ]
Rnd . We say that the pair (F, F 0 ) is controlled by S if the derivative F 0 has finite q-variation,
and the remainder RF : T Rn , defined by
RF (s, t) = Fs,t Fs0 Ss,t ,
has finite r-variation for 1/r = 1/p + 1/q. In this case, we write (F, F 0 ) CSq = CSq (Rn ), and
define
k(F, F 0 )kC q := kF 0 kqvar + kRF krvar .
S

Equipped with the norm |F0 | +

|F00 |

+ k(F, F 0 )kC q , the space CSq is a Banach space.

S

Naturally, the function F 0 should be interpreted as the derivative of F with respect to S.

The reason for considering couples (F, F 0 ) and not just functions F is that the smoothness
requirement on the remainder RF usually does not determine F 0 uniquely for a given path F .
For example, if F and S both have finite r-variation rather than just finite p-variation, then
for every F 0 of finite q-variation we have (F, F 0 ) CSq .
Note that we do not require F or F 0 to be continuous. We will point out in Remark 4.10
below why this does not pose any problem.
To gain a more quantitative feeling for the condition on q, let us assume for the moment
that we can choose p > 2 arbitrarily close to 2 (which is the case in the example of a continuous
semimartingale rough path). Then 2/p + 1/q > 1 as long as q > 0, so that the derivative F 0
may essentially be as irregular as we want. The remainder RF has to be of finite r-variation
for 1/r = 1/p + 1/q, so in other words it should be of finite r-variation for some r < 2 and
thus slightly more regular than the sample path of a continuous local martingale.
Example 4.7. Let (0, 1] be such that (2 + )/p > 1. Let Cb1+ and define Fs := (Ss )
p/
and Fs0 := 0 (Ss ). Then (F, F 0 ) CS : Clearly F 0 has finite p/-variation. For the remainder,
we have
|RF (s, t)|p/(1+) = |(St ) (Ss ) 0 (Ss )Ss,t |p/(1+) kkC 1+ c(s, t),
b

where c is a control function for S. As the image of the continuous path S is compact, it is
not actually necessary to assume that is bounded. We may always consider a C 1+ function
of compact support, such that agrees with on the image of S.
This example shows that in general RF (s, t) is not a path increment of the form RF (s, t) =
G(t) G(s) for some function G defined on [0, T ], but really a function of two variables.

16

Example 4.8. Let G be a path of finite r-variation for some r with 1/p + 1/r > 1. Setting
(F, F 0 ) = (G, 0), we obtain a controlled path in CSq , where 1/q = 1/r 1/p. In combination
with Theorem 4.9 below, this example shows in particular that the controlled rough path
integral extends the Young integral and the Riemann-Stieltjes integral.
R
The basic idea of rough path integration is that if we already know how
R to define SdS,
and if F looks like S on small scales, then we should be able to define F dS as well. The
precise result is given by the following theorem:
Theorem 4.9 (Theorem 4.9 in [FH14], see also [Gub04], Theorem 1).
R Let q be such that
2/p + 1/q > 1. Let (F, F 0 ) CSq . Then there exists a unique function F dS C([0, T ], Rn )
which satisfies

Z t


Fu dSu Fs Ss,t Fs0 A(s, t) . kSkpvar,[s,t] kRF krvar,[s,t] + kAkp/2var,[s,t] kF 0 kqvar,[s,t]

s

for all (s, t) T . The integral is given as limit of the compensated Riemann sums
Z

Fu dSu = lim
0

m
[s1 ,s2



Fs1 Ss1 ,s2 + Fs01 A(s1 , s2 ) ,

(6)

] m

where ( m ) is any sequence of partitions

of [0, t] with mesh size going to 0.
R
The map (F, F 0 ) 7 (G, G0 ) := ( Fu dSu , F ) is continuous from CSq to CSp and satisfies
k(G, G0 )kC p . kF kpvar + (kF 0 k + kF 0 kqvar )kAkp/2var + kSkpvar kRF krvar .
S

Remark 4.10. To the best of our knowledge, there is no publication in which the controlled
path approach to rough paths is formulated using p-variation regularity. The references on
the subject all work with H
older continuity. But in the p-variation setting, all the proofs work
exactly as in the H
older setting, and it is a simple exercise to translate the proof of Theorem 4.9
in [FH14] (which is based on Youngs maximal inequality which we will encounter below) to
obtain Theorem 4.9.
There is only one small pitfall: We did not require F or F 0 to be continuous. The rough
path integral for discontinuous functions is somewhat tricky, see [Wil01]. But here we do not
run into any problems, because the integrand S = (S, A) is continuous. The construction based
on Youngs maximal inequality works as long as integrand and integrator have no common
discontinuities, see the Theorem on page 264 of [You36].
If now Cb1+ for some > 0, then using a Taylor expansion one can show that there
exist p > 2 and q > 0 with 2/p + 1/q > 0, such that (F, F 0 ) 7 ((F ), 0 (F )F 0 ) is a locally
bounded map from CSp to CSq . Combining this with the fact that the rough path integral is a
bounded map from CSq to CSp , it is not hard to prove the existence of solutions to the rough
differential equation
dXt = (Xt )dSt ,
X0 = x0 ,
(7)
p R
t [0, T ], where X CS , (Xt )dSt denotes the rough path integral, and S is a typical price
path. Similarly, if Cb2+ , then the map (F, F 0 ) 7 ((F ), 0 (F )F 0 ) is a locally Lipschitz
continuous from CSp to CSq , and this yields the uniqueness of the solution to (7) at least
among the functions in the Banach space CSp . See Section 5.3 of [Gub04] for details.
17

A remark is in order about the stringent regularity requirements on . In the classical It

o
theory of SDEs, the function is only required to be Lipschitz continuous. But to solve a
Stratonovich SDE, we need better regularity of . This is natural, because the Stratonovich
SDE can be rewritten as an It
o SDE with a Stratonovich correction term: the equations
dXt = (Xt ) dWt and
1
dXt = (Xt )dWt + 0 (Xt )(Xt )dt
2
are equivalent (where W is a standard Brownian motion, dWt denotes Ito integration, and
dWt denotes Stratonovich integration). To solve the second equation, we need 0 to be
Lipschitz continuous, which is always satisfied if Cb2 . But rough path theory cannot
distinguish between It
o and Stratonovich integrals: If we define the area of W using It
o
(respectively Stratonovich) integration, then the rough path solution of the equation will
coincide with the It
o (respectively Stratonovich) solution. So in the rough path setting,
the function should satisfy at least the same requirements as in the Stratonovich setting.
The regularity requirements on are essentially sharp, see [Dav07], but the boundedness
assumption can be relaxed, see [Lej12]. See also Section 10.5 of [FV10] for a slight relaxation
of the regularity requirements in the Brownian case.
Of course, the most interesting result of rough path theory is that the solution to a rough
differential equation depends continuously on the driving signal. This is a consequence of the
following observation:
Propostion 4.11 (Proposition 9.1 of [FH14]). Let p > 2 and q > 0 with 2/p + 1/q > 0.
= (S,
A)
be two rough paths of finite p-variation, let (F, F 0 ) C q and
Let S = (S, A) and S
S
q
(F , F 0 ) CS , and let (s, t) T . Then for every M > 0 there exists CM > 0 such that
Z
Z



Fs dSs
Fs dSs

0

pvar


CM |F0 F0 | + |F00 F00 | + kF 0 F 0 kqvar

pvar + kA Ak
p/2var ,
+ kRF RF krvar + kS Sk

as long as
pvar , kAk
p/2var } M.
max{|F00 | + k(F, F 0 )kC q , |F00 | + k(F , F 0 )kCq , kSkpvar , kAkp/2var , kSk
S

In other words, the rough path integral depends on integrand and integrator in a locally
Lipschitz continuous way, and therefore it is no surprise that the solutions to differential
equations driven by rough paths depend continuously on the signal.

4.2

Typical price paths as rough paths

Our second approach to stochastic integration in model free financial mathematics is based
on the rough path integral. Here we show that for every typical price path, the pair (S, A) is
a p-rough path for all p (2, 3), where
Z t
Z t
Z t

j
Sri dSrj Ssi Ss,t
.
A(s, t) =
Ss,r dSr :=
Ss,r dSr :=
s

1i,jd

The main ingredient in the proof will be our speed of convergence (5).
18

Theorem 4.12. For (s, t) T , , and i, j {1, . . . , d} define

Z s
Z t
Z t
j
j
i,j
i
j
i
j
i
Sri dSrj () Ssi ()Ss,t
().
Sr dSr ()
Sr dSr () Ss ()Ss,t () :=
As,t () :=
s

Let p > 2. Then for typical price paths, A = (Ai,j )1i,jd has finite p/2-variation, and in
particular S = (S, A) is a p-rough path.
Proof. Define the dyadic stopping times (kn )n,kN by 0n := 0 and
n
k+1
:= inf{t kn : |St Skn | = 2n },

P
n Sk
n . Accorcing to (5), for typical
n ) (t), so that kS
and set Stn := k Skn 1[kn ,k+1
2
price paths there exists C() > 0 such that
Z


p
n

(S S)() SdS() C()2n log n.

Fix such a typical price path , which is also of finite q-variation for all q > 2 (recall from
Corollary 2.11 that this is satisfied by typical price paths). Let us show that for such , the
process A is of finite p/2-variation for all p > 2.
We have for (s, t) T , omitting the argument of the processes under consideration,
Z t



|As,t | 
Sr dSr (S n S)s,t  + |(S n S)s,t Ss Ss,t |
s
p
C()2n log n + |(S n S)s,t Ss Ss,t | . C()2n(1) + |(S n S)s,t Ss Ss,t |
for every n N, > 0. The second term on the right hand side can be estimated, using an
argument based on Youngs maximal inequality (see [LCL07], Theorem 1.16), by
max{2n c(s, t)1/q , (#{k : kn [s, t]})12/q c(s, t)2/q + c(s, t)2/q },

(8)

where c(s, t) is a control function with |Ss,t |q c(s, t) for all (s, t) T . Indeed, if there exists
no k with kn [s, t], then |(S n S)s,t Ss Ss,t | 2n c(s, t)1/q , using that |Ss,t | c(s, t)1/q .
This corresponds to the first term in the maximum in (8).
Otherwise, note that at the price of adding c(s, t)2/q to the right hand side, we may
suppose that s = kn0 for some k0 . Let now kn0 , . . . , kn0 +N 1 be those (kn )k which are in [s, t).
Without loss of generality we may suppose N 2, because otherwise (S n S)s,t = Ss Ss,t .
Abusing notation, we write kn0 +N = t. The idea is now to successively delete points (kn0 +` )
from the partition, in order to pass from (S n S) to Ss Ss,t . By super-additivity of c, there
must exist ` {1, . . . , N 1}, for which
c(kn0 +`1 , kn0 +`+1 )

2
c(s, t).
N 1

Deleting kn0 +` from the partition and subtracting the resulting integral from (S n S)s,t , we
get
|Skn +`1 Skn +`1 ,kn +` + Skn +` Skn +` ,kn +`+1 Skn +`1 Skn +`1 ,kn +`+1 |
0
0
0
0
0
0
0
0
0
 2
2/q
= |Skn +`1 ,kn +` Skn +` ,kn +`+1 | c(kn0 +`1 , kn0 +`+1 )2/q
c(s, t)
.
0
0
0
0
N 1
19

Successively deleting all the points except kn0 = s and kn0 +N = t from the partition gives
N 
2/q
X
2
. N 12/q c(s, t)2/q ,
|(S S)s,t Ss Ss,t |
c(s, t)
k1
n

k=2

and therefore (8). Now it is easy to see that #{k : kn [s, t]} 2nq c(s, t) (compare also the
proof of Lemma 3.3), and thus
|As,t | . C()2n(1) + max{2n c(s, t)1/q , (2nq c(s, t))12/q c(s, t)2/q + c(s, t)2/q }
= C()2n(1) + max{2n c(s, t)1/q , 2n(2q) c(s, t) + c(s, t)2/q }.

(9)

This holds for all n N, > 0, q > 2. Let us suppose for the moment that c(s, t) 1 and let
> 0 to be determined later. Then there exists n N for which 2n1 < c(s, t)1/(1) 2n .
Using this n in (9), we get
n
o
|As,t | .,, c(s, t) + max c(s, t)1/(1) c(s, t)/q , c(s, t)(2q)/(1)+ + c(s, t)2/q


q+(1)
2q+(1)
2/q
q(1)
1
= c(s, t) + max c(s, t)
, c(s, t)
+ c(s, t)
.
We would like all the exponents in the maximum on the right hand side to be larger or equal
to 1. For the first term, this is satisfied as long as < 1. For the third term, we need q/2.
For the second term, we need (q 1 )/(1 ). Since > 0 can be chosen arbitrarily
close to 0, it suffices if > q 1. Now, since q > 2 can be chosen arbitrarily close to 2, we
see that can be chosen arbitrarily close to 1. In particular, we may take = p/2 for any
p > 2, and we obtain
|As,t |p/2 ., c(s, t)(1 + c(s, t) ) c(s, t)(1 + c(0, T ) )
for a suitable > 0.
It remains to treat the case c(s, t) > 1, for which we simply estimate
 Z

Z
p/2
p/2




Sr dSr + kSkp
Sr dSr + kSkp c(s, t).
|As,t |p/2 .p
0

So for every interval [s, t] we can estimate |As,t |p/2 .,p c(s, t), and the proof is complete.
Remark 4.13. To the best of our knowledge, this is one of the first times that a non-geometric
rough path is constructed in a non-probabilistic setting, and certainly we are not aware of any
works where rough paths are constructed using financial arguments.
We also point out that, thanks to Proposition 2.6, we gave a simple, model free, and
pathwise proof for the fact that a local martingale together with its It
o integral defines a
rough path. While this seems intuitively clear, the only other proof that we are aware of is
somewhat involved: it relies on a strong version of the Burkholder-Davis-Gundy inequality, a
time change, and Kolmogorovs continuity criterion; see [CL05] or Chapter 14 of [FV10].
The following auxiliary result will allow us to obtain the rough path integral as a limit of
Riemann sums, rather than compensated Riemann sums which are usually used to define it.

20


Lemma 4.14. Let (cn )nN be a sequence of positive numbers such that (cn log n) converges
n
n
n
n
to 0 for all > 0. For
P n N define 0 := 0 and k+1 := inf{t k : |Snt Sk | = cn },
n
n ) (t). Then for typical price paths, ((S
k N, and set St = k Skn 1[kn ,k+1
S)) converges
R
uniformly to SdS. Moreover, for p > 2 and for typical price paths there exists a control
function c = c(p, ) such that
|(S n S)kn ,`n () Skn ()Skn ,`n ()|p/2
1.
c(kn , `n )
n k<`
R
Proof. The uniform convergence of ((S n S)) to SdS follows from Corollary 3.6. For the
second claim, fix n N and k < ` such that `n T . Then
Z








n
n
|(S S)kn ,`n Skn Skn ,`n | . (S S)
Ss dSs + Akn ,`n 

0
p
n n 2/p
. cn log n + vp/2 (k , ` ) . cn1 + vp/2 (kn , `n )2/p , (10)
sup sup

where > 0 and the last estimate holds by our assumption on the sequence (cn ), and where
p/2
vp/2 (s, t) := kAkp/2var,[s,t] for (s, t) T . Of course, this inequality only holds for typical
price paths and not for all .
On the other side, the same argument as in the proof of Theorem 4.12 (using Youngs
maximal inequality and successively deleting points from the partition) shows that
|(S n S)kn ,`n Skn Skn ,`n | . cn2q vq (kn , `n ),

(11)

where vq (s, t) := kSkqqvar,[s,t] for (s, t) T .

Let us define the control function c := vq + vp/2 . Take > 0 to be determined below. If
cn > c(s, t)1/(1) , then we use (11) and the fact that 2 q < 0, to obtain
2q

|(S n S)kn ,`n Skn Skn ,`n | . (

c(kn , `n )) (1) vq (kn , `n ) c(kn , `n )

2q+(1)
(1)

The exponent is larger or equal to 1 as long as (q 1 )/(1 ). Since q and can be

chosen arbitrarily close to 2 and 0 respectively, we can take = p/2, and get
|(S n S)kn ,`n Skn Skn ,`n |p/2 . c(kn , `n )(1 + c(0, T ) )
for a suitable > 0.
On the other side, if cn c(s, t)1/(1) , then we use (10) to obtain
|(S n S)kn ,`n Skn Skn ,`n | . c(kn , `n ) + c(kn , `n )2/p ,
so that also in this case we may take = p/2, and thus we have in both cases
|(S n S)kn ,`n Skn Skn ,`n |p/2 c(kn , `n ),
where c is a suitable (-dependent) multiple of c.

21

4.3

The rough path integral as limit of Riemann sums

Theorem 4.12 shows that we can apply the controlled rough path integral in model free
financial mathematics, since every typical price path is a rough path. But there remains
a
R
philosophical problem: As we have seen in Theorem 4.9, the rough path integral F dS is
given as limit of the compensated Riemann sums
Z t
X 

Fs dSs = lim
Fr1 Sr1 ,r2 + Fr01 A(r1 , r2 ) ,
m

[r1 ,r2 ] m

where ( m ) is an arbitrary sequence of partitions of [0, t] with mesh size going to 0. The
term Fr1 Sr1 ,r2 has an obvious financial interpretation as profit made by buying Fr1 units of
the traded asset at time r1 and by selling them at time r2 . However, for the compensator
Fr01 A(r1 , r2 ) there seems to be no financial interpretation, and therefore it is not clear whether
the rough path integral can be understood as profit obtained by investing in S.
However, we observed in Section 3 that along suitable stopping times (kn )n,k , we have
Z t
X
n t .
Ss dSs = lim
Skn Skn t,k+1
n

By the philosophy of controlled paths, we expect that also for F which looks like S on small
scales we should obtain
Z t
X
n t ,
Fs dSs = lim
Fkn Skn t,k+1
n

n t) in the Riemann sum. With

without having to introduce the compensator F0 n A(kn t, k+1
k
the results we have at hand, this statement is actually relatively easy to prove. Nonetheless,
it seems not to have been observed before.
For the remainder of this section we fix S C([0, T ], Rd ), and we work under the following
assumption:

Assumption (rie). Let n = {0 = tn0 < tn1 < < tnNn = T }, n N, be a given sequence of
partitions such that sup{|Stnk ,tnk+1 | : k = 0, . . . , Nn 1} converges to 0, and let p (2, 3). Set
Stn :=

NX
n 1

Stnk 1[tnk ,tnk+1 ) (t).

k=0

We assume that the Riemann sums

exists a control function c for which
sup
(s,t)T

(S n

S) converge uniformly to

SdS, and that there

|(S n S)tnk ,tn` Stnk Stnk ,tn` |p/2

|Ss,t |p
+ sup sup
1.
c(s, t)
c(tnk , tn` )
n 0k<`Nn

(12)

Remark 4.15. The coarse-grained regularity condition (12) has recently been independently
rediscovered in [Kel14]; see also [GIP14].
Our proof that the rough path integral is given as limit of Riemann sums is somewhat indirect. We translate everything from Ito type integrals to related Stratonovich type integrals,
for which the convergence follows from the continuity of the rough path integral, Proposition 4.11. Then we translate everything back to our Ito type integrals. To go from Ito to
Stratonovich, we need the quadratic variation:
22

Lemma 4.16. Under Assumption (rie), let 1 i, j d, and define

Z

hS i , S j it := Sti Stj S0i S0j

Sri dSrj

Srj dSri .

Then hS i , S j i is a continuous function and

i

hS , S it = lim hS
n

, S j int

NX
n 1

(Stin

= lim

k+1 t

Stin t )(Stjn

k+1 t

Stjn t ).

(13)

k=0

The sequence (hS i , S j in )n is of uniformly bounded total variation, and in particular hS i , S j i

is of bounded variation. We write hSi = hS, Si = (hS i , S j i)1i,jd , and call hSi the quadratic
variation of S.
Proof. The function hS i , S j i is continuous by definition. The specific form (13) of hS i , S j i
follows from two simple observations:
Sti Stj S0i S0j =

NX
n 1 

Stin

k+1

j
t Stn

k+1

j
i
t Stn t Stn t
k

k=0

for every n N, and

Stin

k+1 t

Stjn

k+1 t

Stin t Stjn t = Stin t Stjn t,tn

k

k+1 t

+ Stjn t Stin t,tn

k

k+1 t

+ Stin t,tn
k

k+1 t

Stjn t,tn
k

k+1 t

R
so that the convergence in (13) is a consequence of the convergence of (S n S) to SdS.
To see that hS i , S j i is of bounded variation, note that

2 
2 
1  i
j
i
j n
i
j n
n
n
Stn t,tn t Stn t,tn t =
(S + S )tk t,tk+1 t (S S )tk t,tk+1 t
k
k+1
k
k+1
4
(read hS i , S j i = 1/4(hS i +S j ihS i S j i)). In other words, the n-th approximation of hS i , S j i
is the difference of two increasing functions, and its total variation is bounded from above by
NX
n 1 

(S i + S j )tnk ,tnk+1

2

NX
m 1 

2 

+ (S i S j )tnk ,tnk+1
. sup
(Stim ,tm )2 + (Stjm ,tm )2 .
m

k=0

k+1

k+1

k=0

Since the right hand side is finite, also the limit hS i , S j i is of bounded variation.
Given the quadratic variation, the existence of the Stratonovich integral is straightforward:
Lemma 4.17. Under Assumption (rie), define Sn |[tnk ,tnk+1 ] as the linear interpolation of Stnk
R
and Stn for k = 0, . . . Nn 1. Then ( Sn dSn ) converges uniformly to
k+1

Sr dSr :=
s

Moreover, setting An (s, t) :=

Rt
s

1
Sr dSr + hSis,t .
2

n dS
n for (s, t) T , we have supn kAn kp/2var < .
Ss,r
r

23

(14)

Proof. Let n N and k {0, . . . , Nn 1}. Then for t [tnk , tnk+1 ] we have
Stn = Stnk +
so that
Z

tn
k+1

tn
k

t tnk
Stn ,tn ,
tnk+1 tnk k k+1

1
Srn dSrn = Stnk Stnk ,tnk+1 + Stnk ,tnk+1 Stnk ,tnk+1 ,
2

(15)

from where the uniform convergence and the representation (14) follow by Lemma 4.16.
To prove that An has uniformly bounded p2 -variation, consider (s, t) T . If there exists
k such that tnk s < t tnk+1 , then we estimate
p/2
Z t
p/2  Z t
|Stn ,tn |2



n
p/2
n
n

|A (s, t)| = 
Ss,r dSr   (r s) n k k+1n 2 dr
|

t
|t
s
s
k
k+1
=

2p/2

|t s|

|Stnk ,tnk+1 |p
|tnk+1 tnk |p

|t s|
kSkppvar,[tn ,tn ] .
k k+1
|tnk+1 tnk |

(16)

Otherwise, let k0 be the smallest k such that tnk (s, t), and let k1 be the largest such k. We
decompose
n
tnn
n S
An (s, t) = An (s, tnk0 ) + An (tnk0 , tnk1 ) + An (tnk1 , t) + Ss,t
k0

n
k0 ,tk1

n
tnn ,t .
n S
+ Ss,t
k1

k1

We get from (15) that

|An (tnk0 , tnk1 )|p/2 . |(S n S)tnk

,tn
k

Stnk Stnk
0

,tn
k

|p/2 + (hSintn

n
k0 ,tk1

)p/2 ,

where hSin denotes the n-th approximation of the quadratic variation. By (12) and Lemma 4.16,
there exists a control function c so that the right hand side is bounded from above by c(tnk0 , tnk1 ).
n
n
n n
Combining this with (16) and a simple estimate for the terms Ss,t
n Stn ,tn and Ss,tn Stn ,t ,
k0
k0 k1
k1
k1
we deduce that kAn kp/2var . c(0, T ) + kSk2
, and the proof is complete.
pvar

We are now ready to prove the main result of this section.

Theorem 4.18. Under Assumption (rie), let q > 0 be such that 2/p + 1/q > 1. Let
q
0
(F,
R F ) CS be a controlled path such that F is continuous. Then the rough path integral
F dS which was defined in Theorem 4.9 is given by
Z

Fs dSs = lim
0

NX
n 1

Ftnk Stnk t,tnk+1 t ,

k=0

where the convergence is uniform in t.

Proof. For n N define F n as the linear interpolation of F between the points in n . Then
n of F n with
(F n , F 0 ) is controlled by Sn : Clearly kF n kqvar kF kqvar . The remainder R
F n
n F 0S
n for (s, t) . We need to show that R
n (s, t) = Fs,t

n
respect to Sn is given by R
T
s s,t
F n
F n
has finite r-variation for 1/r = 1/p + 1/q.

24

If tnk s t tnk+1 , we have

r

n n (s, t)|r =  t s Ftn ,tn F 0 t s Stn ,tn 
|R
s n
F
tnk+1 tnk k k+1
tk+1 tnk k k+1
 t s r




r/p
0 r/q
n ,tn
 n
kR
k
+
kF
k
kSk

n
n
n
F
rvar,[t
]
qvar,[tk ,s]
pvar,[tk ,tk+1 ]
k k+1
tk+1 tnk

|t s|
kRF krvar,[tnk ,tnk+1 ] + kF 0 kqvar,[tnk ,tnk+1 ] + kSkpvar,[tnk ,tnk+1 ] ,
n
tk |

|tnk+1

(17)

where in the last step we used that 1/r = 1/p + 1/q, and thus r/q + r/p = 1.
Otherwise, there exists k {1, . . . , Nn 1} with tnk (s, t). Let k0 and k1 the smallest
and largest such k, respectively. Then
0
n n (tn , t)|r +|Fs,t
n n (tn , tn )|r +|R
n n (s, t)|r .r |R
n n (s, tn )|r +|R
n Stn
|R
k1
k0 k1
k0
k
F
F
F
F
k0

,tn
k

0
r
n Stn ,t | .
|r +|Fs,t
k
k1

n (tn , tn ) = RF (tn , tn ), and therefore we can use (17), the assumption on RF , and
Now R
k0 k1
F n k0 k1
the fact that 1/r = 1/p + 1/q (which is needed to treat the last two terms on the right hand
side), to obtain
n n krvar .r kRF krvar + kF 0 kqvar + kSkpvar .
kR
F
n ) converges uniformly to (F, RF ).
On the other side, since F and RF are continuous, (F n , R
F n
Now for continuous functions, uniform convergence with uniformly bounded p-variation implies convergence in p0 -variation for every p0 > p. See Exercise 2.8 in [FH14] for the case of
Holder continuous functions.
Thus, using Lemma 4.17, we see that if p0 > p and q 0 > q are such that 2/p0 +1/q 0 > 0, then
n
((S , An )n ) converges in (p0 , p0 /2)-variation to (S, A ), where A (s, t) = A(s, t) + 1/2hSis,t .
n )) converges in (q 0 , p0 , r0 )-variation to (F, F 0 , RF ), where 1/r0 = 1/p0 +
Similarly, ((F n , F 0 , R
F n
0
1/q .
R
R
Proposition 4.11 now yields the uniform convergence of F n dSn to F dS, by which we
denote the rough path integral of the controlled path (F, F 0 ) against the rough path (S, A ).
But for every t [0, T ] we have
Z t
X 1
lim
Fsn dSsn = lim
(Ftn + Ftnk+1 )Stnk ,tnk+1
n 0
n
2 k
k:tn
t
k+1
 X

1 X
n
n
n
n
n
n
n
Ftk ,tk+1 Stk ,tk+1 .
= lim
Ftk Stk ,tk+1 +
n
2 n
n
k:tk+1 t

k:tk+1 t

Using that F is controlled by

R t S,0 it is easy to see that the second term on the
P right hand side
converges uniformly to 1/2 0 Fs dhSis , t [0, T ]. Thus, the Riemann sums k Ftnk Stnk ,tnk+1
R
R
converge uniformly to F dS 1/2 F 0 dhSi, and from the representation ofR the rough
Rpath integral Ras 0limit of compensated Riemann sums (6), it is easy to see that F dS =
F dS + 1/2 F dhSi, which completes the proof.
Theorem 4.18 is reminiscent of Follmers pathwise Ito integral [Fol81]. Follmer assumes
that the quadratic variation hSi of S exists along a given sequence of partitions and is continuous, and uses this to prove an It
o formula for S: if F C 2 , then
Z t
Z
1 t 2
D F (Ss )dhSis ,
(18)
F (St ) = F (S0 ) +
F (Ss )dSs +
2 0
0
25

R
where the integral 0 F (Ss )dSs is given as limit of Riemann sums along that same sequence
of partitions.
Friz and Hairer [FH14] observe that if for p (2, 3) the function S is of finite p-variation
and hSi is an arbitrary continuous function of finite p/2-variation, then setting
1 i j
Sym(A)(s, t) := (Ss,t
Ss,t + hSis,t )
2
one obtains a reduced rough path (S, Sym(A)). They continue to show that if F is controlled
R
by S with symmetric derivative F 0 , then it is possible to define the rough path integral F dS.
This is not surprising since then we have Fs0 As,t = Fs0 Sym(A)s,t for the compensator term in
the definition of the rough path integral. They also derive
R an Ito formula for reduced rough
paths, which takes the same form as (18), except that now F (S)dS is a rough path integral
(and therefore defined as limit of compensated Riemann sums).
So both the assumption and the result of [FH14] are slightly different from the ones
in [Fol81], and while it seems intuitively clear, it is still not shown rigorously that Follmers
pathwise It
o integral is a special case of the rough path integral. We will now show that
Follmers result is a special case of Theorem 4.18. For that purpose we only need to prove
that Follmers condition on the convergence of the quadratic variation is a special case of the
assumption in Theorem 4.18, at least as long as we only need the symmetric part of the area.
Definition 4.19. Let f C([0, T ], R) and let n = {0 = tn0 < tn1 < < tnNn = T }, n N
be such that sup{|ftnk ,tnk+1 | : k = 0, . . . , Nn 1} converges to 0. We say that f has quadratic
variation along ( n ) in the sense of F
ollmer if the sequence of discrete measures (n ) on
([0, T ], B[0, T ]), defined by
NX
n 1
|ftnk ,tnk+1 |2 tnk ,
n :=
(19)
k=0

converges weakly to a non-atomic measure . We write [f ]t for the distribution function of

(in general will not be a probability measure). The function f = (f 1 , ..., f d ) C([0, T ], Rd )
has quadratic variation along ( n ) in the sense of F
ollmer if this holds for all f i and f i + f j ,
1 i, j d. In this case, we set
1
[f i , f j ]t := ([f i + f j ]t [f i ]t [f j ]t ),
2

t [0, T ].

Lemma 4.20 (see also [Vov11], Proposition 6.1). Let p (2, 3), and let S = (S 1 , ..., S d )
C([0, T ], Rd ) have finite p-variation. Let n = {0 = tn0 < tn1 < < tnNn = T }, n N, be a
sequence of partitions such that sup{|Stnk ,tnk+1 | : k = 0, . . . , Nn 1} converges to 0. Then the
following conditions are equivalent:
1. The function S has quadratic variation along ( n ) in the sense of F
ollmer.
2. For all 1 i, j d, the discrete quadratic variation
hS

, S j int

:=

NX
n 1

Stin t,tn
k

k+1 t

Stjn t,tn

k=0

converges uniformly in C([0, T ], R) to a limit hS i , S j i.

26

k+1 t

PNn 1 i
Stn 1[tnk ,tnk+1 ) , i {1, . . . , d}, n N, the Riemann sums (S n,i S j ) +
3. For S n,i := k=0
k
R
R
(S n,j S i ) converge uniformly to a limit S i dS j + S j dS i . Moreover, the symmetric
part of the approximate area,
Sym(An )i,j (s, t) =

1
j
i
(S n,i S j )s,t +(S n,j S i )s,t Ssi Ss,t
Ssj Ss,t
, 1 i, j d, (s, t) T ,
2

has uniformly bounded p/2-variation along ( n ), in the sense of (12).

If these conditions hold, then [S i , S j ] = hS i , S j i for all 1 i, j d.
Proof. Assume 1. and note that
S jn n
k+1 t tk t,tk+1 t

Stin t,tn
k

1
((S i + S j )tnk t,tnk+1 t )2 (Stin t,tn t )2 (Stjn t,tn t )2 .
k
k+1
k
k+1
2

Thus, the uniform convergence of hS i , S j in and the fact that hS i , S j i = [S i , S j ] follow once we
show that F
ollmers weak convergence of the measures (19) implies the uniform convergence of
their distribution functions. But since the limiting distribution is continuous by assumption,
this is a standard result.
Next, assume 2. The uniform convergence of the Riemann sums (S n,i S j ) + (S n,j S i ) is
shown as in Lemma 4.16. To see that Sym(An ) has uniformly bounded p/2-variation along
( n ), note that for 0 k ` Nn and 1 i, j d we have
|(S n,i S j )tnk ,tn` + (S n,j S i )tnk ,tn` Ssi Stjn ,tn Ssj Stin ,tn |p/2 = |Stin ,tn Stjn ,tn hS i , S j intn ,tn |p/2
k

kSkpvar,[tnk ,tn` ] + khS i , S j in k1var,[tnk ,tn` ] .

That khS i , S j in k1var is uniformly bounded in n is shown in Lemma 4.16.
That 3. implies 1. is also shown in Lemma 4.16.
Remark 4.21. With Theorem 4.18 we can only derive an It
o formula for F C 2+ , since
we are only able to integrate F (S) if F C 1+ . But this only seems to be due to the fact
that our analysis is not sharp. We expect that typical price paths have an associated rough
path of finite 2-variation, up to logarithmic corrections. For such rough paths, the integral
extends to integrands F C 1 , see Chapter 10.5 of [FV10]. For typical price paths (but not
for the area), it is shown in [Vov12], Section 4.3, that they are of finite 2-variation up to
logarithmic corrections.

Pathwise Hoeffding inequality

In the construction of the pathwise Ito integral for typical price processes we needed the
following result, a pathwise formulation of the Hoeffding inequality which is due to Vovk.
Here we present a slightly adapted version.
Lemma A.1 ([Vov12], Theorem A.1). Let (n )nN be a strictly increasing sequence of stopping
times with 0 = 0, such that for every we have n () = for all but finitely many
n N. Let for n N the function hn : Rd be Fn -measurable, and suppose that there
exists a Fn -measurable bounded function bn : R, such that
sup |hn ()Sn t,n+1 t ()| bn ()
t[0,T ]

27

(20)

for all . Then for every R there exists a simple strategy H H1 such that

 X
Nt

2 X
2
bn
1 + (H S)t exp
hn Sn t,n+1 t
2

n=0

n=0

for all t [0, T ], where Nt := max{n N : n t}.

Proof. Let R. The proof is based on the following deterministic inequality: if (20) is
satisfied, then for all and all t [0, T ] we have that


2 2
exp hn ()Sn t,n+1 t () bn () 1
2
 bn ()
 2
e
ebn ()

hn ()Sn t,n+1 t ()
exp b2n ()
2
2bn ()
=: fn ()Sn t,n+1 t ().
(21)
P
This inequality is shown in (A.1) of [Vov12]. We define Ht := n Fn 1(n ,n+1 ] (t), with Fn
that have to be specified. We choose F0 := f0 , which is bounded and F0 -measurable, and on
[0, 1 ] we obtain


2
1 + (H S)t exp h0 Sn t,n+1 t b20 .
2
Observe also that 1 + (H S)1 = 1 + f0 S0 ,1 is bounded, because by assumption h0 S0 ,1 is
bounded by the bounded random variable b0 .
Assume now that Fk has been defined for k = 0, . . . , m 1, that

 X
Nt

2 X
2
bn
1 + (H S)t exp
hn Sn t,n+1 t
2

n=0

n=0

for all t [0, m ], and that 1 + (H S)m is bounded. We define Fm := (1 + (H S)m )fm ,
which is Fm -measurable and bounded. From (21) we obtain for t [m , m+1 ]
1 + (H S)t = 1 + (H S)m + (1 + (H S)m )fm Sm t,m+1 t


2 2

(1 + (H S)m ) exp hm Sm t,m+1 t bm

2
 X

N

t
2 X
b2n ,
exp
hn Sn t,n+1 t
2
n=0

n=0

where the last step follows from the induction hypothesis. From the first line of the last
equation we also obtain that 1 + (H S)m+1 is bounded because fm Sm ,m+1 is bounded for
the same reason that f0 S0 ,1 is bounded.

Davies criterion

It was already observed by Davie [Dav07] that in certain situations the rough path integral
can be constructed as limit of Riemann sums and not just compensated Riemann sums. Davie
shows that under suitable conditions, the usual Euler scheme (without area compensation)
28

converges to the solution of a given rough differential equation. But from there it is easily
deduced that then also the rough path integral is given as limit of Riemann sums. Here we
show that Davies criterion implies our assumption (rie).
Let p (2, 3) and let S = (S, A) be a 1/p-Holder continuous rough path, that is |Ss,t | .
|t s|1/p and |A(s, t)| . |t s|2/p . Write := 1/p and let (1 , 2). Davie assumes
that there exists C > 0 such that the area process A satisfies
 `1

X


A(jh, (j + 1)h) C(` k) h2 ,


(22)

j=k

whenever 0 < k < ` are integers and h > 0 such that `h T . Under these conditions,
Theorem 7.1 of [Dav07] implies that for F C with > p and for tnk = kT /n, n, k N, the
Riemann sums
n1
X
F (Stnk )Stnk t,tnk+1 t , t [0, T ],
k=0

converge uniformly to the rough path integral. But it can be easily deduced from (22) that
the area process A is given as limit of non-anticipating Riemann sums along (tn )n . Indeed,
letting h = T /n,
Z t
  n1  Z tn t

n1
X

 X

k+1




n
n
n
n
n
n
S
dS

S
S
=
S
dS

S
S
s
s
t
t
t,t
t
s
s
t
t
t
t,t
t

 

k
k
k+1
k
k
k+1
n
0

k=0

tk t

k=0

  bt/hc1

 n1

  X
X
n
n



Akh,(k+1)h  + |A(bt/hc, t)|
=
A(tk t, tk+1 t) 
k=0

. Cbt/hc h

k=0
2

+ h kAk2 . Cth

+ h2 kAk2 .

Since < 2, the right hand side converges to 0 as n goes to (and thus h goes to 0).
Futhermore, (22) implies the uniformly bounded p/2-variation condition (12):

  Z tn
 X

 Z tn
 n
  `
  `1

j+1
(S S)tn ,tn Stn Stn ,tn  


Ss dSs Stnj Stnj ,tnj+1 

 n Ss dSs Stnk Stnk ,tn`  + 
k `
k
k ` 
n
tk

kAk2 |tn`

kAk2 |tn`

tnk |2

tnk |2

j=k

tj

X

 `1


+
Atnk ,tnk+1  kAk2 |tn` tnk |2 + C(` k) h2
+

j=k
C|tn`

tnk |2 .

References
[ABPS14] Beatrice Acciaio, Mathias Beiglbock, Friedrich Penkner, and Walter Schachermayer, A model-free version of the fundamental theorem of asset pricing and the
super-replication theorem, Math. Finance (2014+).
[ALP95]

Marco Avellaneda, Arnon Levy, and Antonio Paras, Pricing and hedging derivative
securities in markets with uncertain volatilities, Appl. Math. Finance 2 (1995),
no. 2, 7388.
29

[Ank05]

Stefan Ankirchner, Information and semimartingales, Ph.D. thesis, HumboldtUniversit

at zu Berlin, 2005.

[BHLP13] Mathias Beiglb

ock, Pierre Henry-Labord`ere, and Friedrich Penkner, Modelindependent bounds for option pricesa mass transport approach, Finance Stoch.
17 (2013), no. 3, 477501. MR 3066985
[Bic81]

Klaus Bichteler, Stochastic integration and lp -theory of semimartingales, Ann.

Probab. (1981), 4989.

[CDGR11] Rosanna Coviello, Cristina Di Girolami, and Francesco Russo, On stochastic calculus related to financial assets without semimartingales, Bull. Sci. Math. 135
(2011), no. 6, 733774.
[CL05]

Laure Coutin and Antoine Lejay, Semi-martingales and rough paths theory,
Electron. J. Probab. 10 (2005), no. 23, 761785 (electronic). MR 2164030
(2006i:60042)

[Dav07]

A. M. Davie, Differential equations driven by rough paths: an approach via discrete

approximation, Appl. Math. Res. Express. AMRX (2007), no. 2, Art. ID abm009,
40. MR 2387018 (2009e:34172)

[DH07]

Mark HA Davis and David G Hobson, The range of traded option prices, Math.
Finance 17 (2007), no. 1, 114.

[DM06]

Laurent Denis and Claude Martini, A theoretical framework for the pricing of
contingent claims in the presence of model uncertainty, Ann. Appl. Probab. 16
(2006), no. 2, 827852. MR 2244434 (2007j:60100)

[DOR13]

Mark Davis, Jan Obl

oj, and Vimal Raval, Arbitrage bounds for prices of weighted
variance swaps, Mathematical Finance (2013), nono.

[DS94]

Freddy Delbaen and Walter Schachermayer, A general version of the fundamental

theorem of asset pricing, Math. Ann. 300 (1994), no. 1, 463520.

[DS14]

Yan Dolinsky and H. Mete Soner, Robust Hedging and Martingale Optimal Transport in Continuous Time, Probab. Theory Related Fields 160 (2014), no. 12,
391427.

[FH14]

Peter Friz and Martin Hairer, A Course on Rough Paths: With an Introduction
to Regularity Structures, Springer, 2014.

[Fol81]

Hans F
ollmer, Calcul dIt
o sans probabilites, Seminar on Probability, XV (Univ.
Strasbourg, Strasbourg, 1979/1980) (French), Lecture Notes in Math., vol. 850,
Springer, Berlin, 1981, pp. 143150. MR 622559 (82j:60098)

[Fol99]

Gerald B. Folland, Real analysis, second ed., Pure and Applied Mathematics (New
York), John Wiley & Sons Inc., New York, 1999, Modern techniques and their
applications, A Wiley-Interscience Publication. MR 1681462 (2000c:00001)

[FV10]

Peter Friz and Nicolas Victoir, Multidimensional stochastic processes as rough

paths. Theory and applications, Cambridge University Press, 2010.
30

[GIP14]

Massimiliano Gubinelli, Peter Imkeller, and Nicolas Perkowski, A Fourier approach to pathwise stochastic integration, arXiv preprint arXiv:1410.4006 (2014).

[Gub04]

Massimiliano Gubinelli, Controlling rough paths, J. Funct. Anal. 216 (2004), no. 1,
86140.

[IP11]

Peter Imkeller and Nicolas Perkowski, The existence of dominating local martingale measures, arXiv preprint arXiv:1111.3885 (2011).

[Kar95]

Rajeeva L. Karandikar, On pathwise stochastic integration, Stochastic Process.

Appl. 57 (1995), no. 1, 1118.

[Kel14]

David TB Kelly, Rough path recursions and diffusion approximations, arXiv

preprint arXiv:1402.3188 (2014).

[KK07]

Ioannis Karatzas and Constantinos Kardaras, The numeraire portfolio in semimartingale financial models, Finance Stoch. 11 (2007), no. 4, 447493.

[KP11]

Constantinos Kardaras and Eckhard Platen, On the semimartingale property of

discounted asset-price processes, Stochastic Process. Appl. 121 (2011), no. 11,
26782691. MR 2832419 (2012j:91127)

[LCL07]

Terry J. Lyons, Michael Caruana, and Thierry Levy, Differential equations driven
by rough paths, Lecture Notes in Mathematics, vol. 1908, Springer, Berlin, 2007.
MR 2314753 (2009c:60156)

[Lej12]

Antoine Lejay, Global solutions to rough differential equations with unbounded

vector fields, Seminaire de Probabilites XLIV, Lecture Notes in Math., vol. 2046,
Springer, Heidelberg, 2012, pp. 215246. MR 2953350

[LQ02]

Terry Lyons and Zhongmin Qian, System control and rough paths, Oxford University Press, 2002.

[Lyo95a]

Terry J. Lyons, The interpretation and solution of ordinary differential equations

driven by rough signals, Stochastic analysis (Ithaca, NY, 1993), Proc. Sympos.
Pure Math., vol. 57, Amer. Math. Soc., Providence, RI, 1995, pp. 115128. MR
1335466 (96d:34076)

[Lyo95b]

, Uncertain volatility and the risk-free synthesis of derivatives, Appl. Math.

Finance 2 (1995), no. 2, 117133.

[Lyo98]

, Differential equations driven by rough signals, Rev. Mat. Iberoam. 14

(1998), no. 2, 215310.

[Nut12]

Marcel Nutz, Pathwise construction of stochastic integrals, Electron. Commun.

Probab 17 (2012), no. 24, 17.

[Ruf13]

Johannes Ruf, Hedging under arbitrage, Math. Finance 23 (2013), no. 2, 297317.

[RV93]

Francesco Russo and Pierre Vallois, Forward, backward and symmetric stochastic
integration, Probab. Theory Related Fields 97 (1993), no. 3, 403421. MR 1245252
(94j:60113)
31

[Tao11]

Terence Tao, An introduction to measure theory, Graduate Studies in Mathematics, vol. 126, American Mathematical Society, Providence, RI, 2011. MR 2827917
(2012h:28003)

[Vov08]

Vladimir Vovk, Continuous-time trading and the emergence of volatility, Electron.

Commun. Probab. 13 (2008), 319324. MR 2415140 (2010c:60125)

[Vov11]

, Ito calculus without probability in idealized financial markets, arXiv

preprint arXiv:1108.0799 (2011).

[Vov12]

, Continuous-time trading and the emergence of probability, Finance Stoch.

16 (2012), no. 4, 561609. MR 2972235

[Wil01]

David R. E. Williams, Path-wise solutions of stochastic differential equations

driven by Levy processes, Rev. Mat. Iberoamericana 17 (2001), no. 2, 295329.
MR 1891200 (2003h:60102)

[You36]

Laurence C. Young, An inequality of the H

older type, connected with Stieltjes
integration, Acta Math. 67 (1936), no. 1, 251282.

32