Probab. Theory Relat.

Fields 118, 251–291 (2000)

Digital Object Identifier (DOI) 10.1007/s004400000080

Vladas Pipiras · Murad S. Taqqu

Integration questions related to fractional

Brownian motion

Received: 9 August 1999 / Revised version: 10 January 2000 /
Published online: 18 September 2000 –
c Springer-Verlag 2000

Abstract. Let {BH (u)}u∈⺢ be a fractional Brownian motion (fBm) with index H ∈ (0, 1)
and Sp(BH ) be the closure in L2 ( ) of the span Sp(BH ) of the increments of fBm BH .
It is well-known that, when BH = B1/2 is the usual Brownian motion (Bm), an element
X ∈ Sp(B1/2 ) can be characterized
by a unique function fX ∈ L2 (⺢), in which case one
writes X in an integral form as X = ⺢ fX (u)dB1/2 (u). From a different, though equivalent,
perspective, the space L2 (⺢) forms a class of integrands for the integral on the real line with
respect to Bm B1/2 . In this work we explore whether a similar characterization of elements of
Sp(BH ) can be obtained when H ∈ (0, 1/2)  or H ∈ (1/2, 1). Since it is natural to define the
integral of an elementary function f = nk=1 fk 1[uk ,uk+1 ) by nk=1 fk (BH (uk+1 ) − BH (uk )),
we want the spaces of integrands to contain elementary functions. These classes of integrands
are inner product spaces. If the space of integrands is not complete, then it characterizes only
a strict subset of Sp(BH ). When 0 < H < 1/2, by using the moving average representation
of fBm BH , we construct a complete space of integrands. When 1/2 < H < 1, howev-
er, an analogous construction leads to a space of integrands which is not complete. When
0 < H < 1/2 or 1/2 < H < 1, we also consider a number of other spaces of integrands.
While smaller and hence incomplete, they form a natural choice and are convenient to work
with. We compare these spaces of integrands to the reproducing kernel Hilbert space of fBm.

1. Introduction

Fractional Brownian motion (fBm) {BH (u)}u∈⺢ with index 0 < H < 1 is a Gauss-
ian, mean-zero and H -self-similar process with BH (0) = 0 and stationary incre-
ments. H -self-similarity means that, for a > 0,
d
{BH (au), u ∈ ⺢} = {a H BH (u), u ∈ ⺢},
d
where = stands for the equality in the finite-dimensional distributions. The fBm
BH with H = 1/2 is the usual Brownian motion (Bm). If EBH 2 (1) = 1, the fBm

V. Pipiras, M.S. Taqqu: Department of Mathematics, Boston University, 111 Cum-

mington St., Boston, MA 02215-2411, USA. e-mail: [email protected];
[email protected]


This research was partially supported by the NSF Grant ANI-9805623 at Boston University.
Mathematics Subject Classification (2000): Primary 60H05, 60G18; secondary 26A33
Key words and phrases: Fractional Brownian motion – Integration in the L2 -sense –
Time and spectral domains – Fractional integrals and derivatives – Inner product spaces –
Completeness
252 V. Pipiras, M.S. Taqqu

BH is called standard. Since we will deal with fractional integration, it will be more
convenient in the sequel to use the parametrization
1
κ=H− .
2
The range 0 < H < 1 now corresponds to −1/2 < κ < 1/2. (The index κ is usu-
ally denoted d in the statistical literature, but in the context of fractional integration
the letter d may be confusing.)
We will denote the fBm BH = {BH (u)}u∈⺢ in terms of the parameter κ as
B κ = {B κ (u)}u∈⺢ . By using stationarity of the increments and κ + 21 – self-simi-
larity of fBm B κ , it is easy to see that the covariance of a standard fBm B κ is given
by
1  2κ+1 
 κ (u, v) = EB κ (u)B κ (v) = |u| + |v|2κ+1 − |u − v|2κ+1 . (1.1)
2
For more information on fBm see, for example, Chapter 7 and references to it in
Samorodnitsky and Taqqu [15].
The fBm model is widely applied in telecommunications (see Leland et al. [13])
and is also of interest in finance, notwithstanding the fact that it is not a semimar-
tingale. Integration with respect to fBm has thus many potential applications. The
difficulty, however, lies in the following. On one hand, the paths of fBm are of
unbounded variation and hence the usual Lebesgue-Stieltjes integration cannot be
applied. One cannot use the usual Itô’s stochastic calculus either because fBm B κ
is not a semimartingale. Hence, some other approaches have to be taken. Recently,
there were numerous attempts to develop some type of stochastic calculus for fBm.
For example, Decreusefond and Üstünel [4] define a stochastic integral with respect
to fBm by using the stochastic calculus of variations (also known as the Malliavin
calculus) and the fact that fBm is a Gaussian process. The Malliavin calculus ideas
along with approximation by stochastic integrals with respect to semimartingales
are also employed in another two works by Carmona, Coutin and Montseny [3],
and Alòs, Mazet and Nualart [1]. The paper by Duncan, Hu and Pasik-Duncan [6]
defines a stochastic integral by using the Wick product but it is still in the spirit of
Decreusefond and Üstünel [4]. A different approach, that of a pathwise integration,
is taken by Dudley and Norvaiša [5], and Zähle [17]. They use specific path proper-
ties of fBm, namely, p-variation in Dudley and Norvaiša [5] and Hölder continuity
in Zähle [17]. Eventually, we note the work by Kleptsyna, LeBreton and Roubaud
[12], where an elementary approach to stochastic calculus for fBm is developed.
The purpose of this paper is to explore some fundamental questions related to the
definition of integrals with respect to fBm. As it will soon become apparent, these
questions do not have trivial answers, even when the integrand f is not random
which is the case considered here.

In applications, one likes to view the integral ⺢ f dB κ as approximated by
n
 n

κ
 
fk B (uk ) = fk B κ (uk+1 ) − B κ (uk ) , (1.2)
k=1 k=1
Integration questions related to fractional Brownian motion 253

where fk and uk < uk+1 are real numbers. This idea can be formalized mathemat-
ically as follows. Write the linear combination (1.2) in an integral form as

κ
I (f ) = f (u)dB κ (u), (1.3)
⺢
where f is the elementary (or step) function given by
n

f (u) = fk 1[uk ,uk+1 ) (u), u ∈ ⺢. (1.4)
k=1

Iκ (f ), in (1.3), is a Gaussian random variable with zero mean and variance which
can be easily evaluated by using (1.1). (We will also use in the sequel simple func-
tions where the intervals in (1.4) are replaced by arbitrary bounded Borel sets.) If
E denotes the set of elementary functions on the real line, then the linear space of
Gaussian random variables {Iκ (f ), f ∈ E} is a subset of the larger linear space
L2
Sp(B κ ) = {X : Iκ (fn ) → X, for some (fn ) ⊂ E}. (1.5)
An element X ∈ Sp(B κ ) is also a Gaussian random variable with zero mean and
variance
Var(X) = lim Var(Iκ (fn )).
n→∞
We can associate with X an equivalence class of sequences of elementary functions
(fn ) such that Iκ (fn ) → X in the L2 ( )-sense. If fX stands for this equivalence
class, then X is usually written in an integral form as

X= fX dB κ (1.6)
⺢
and the right-hand side of (1.6) is called the integral with respect to fBm on the real
line (see, for example, Huang and Cambanis [11], p. 587). Thus the integral (1.6)
is indeed a limit (in the L2 ( )-sense) of the (1.2) type linear combinations.
It is well known, however, that when κ = 0 (in which case B κ corresponds
to the usual Bm) there is a simpler characterization of elements of Sp(B 0 ). If we
assume that Bm B 0 is standard, then independence of its increments implies that,
for f ∈ E, 
Var(I0 (f )) = f 2 (u)du.
⺢
Hence, if (fn ) ⊂ E and if I0 (fn ) converges to X ∈ Sp(B 0 ) in the L2 ( )-sense,
there is a unique function fX ∈ L2 (⺢) such that
 
Var(X) = lim Var(I (fn )) = lim
0
fn (u)du =
2
fX2 (u)du.
n→∞ n→∞ ⺢ ⺢

This is because the space L2 (⺢) is complete. Thus, whereas X ∈ Sp(B 0 ) has been
associated above with a sequence of elementary functions (fn ), it can now be char-
acterized by a single function fX ∈ L2 (⺢). One usually writes X in an integral
form as 
X= fX (u)dB 0 (u).
⺢
254 V. Pipiras, M.S. Taqqu

Conversely, since any function f ∈ L2 (⺢) can be approximated in L2 (⺢) by ele-

mentary functions, there is a random variable Xf ∈ Sp(B 0 ) such that

Xf = f (u)dB 0 (u).
⺢

This implies that the integral on the real line with respect to Bm can be defined for
arbitrary functions f ∈ L2 (⺢). Since, for X, Y ∈ Sp(B 0 ),

E(XY ) = fX (u)fY (u)du,
⺢

we can also say that the Hilbert spaces Sp(B 0 ) and L2 (⺢) are isometric, that is,
there is a linear and onto map between these spaces which preserves inner products.
Observe also that this map is an extension of the map f → I0 (f ), for f ∈ E.
In this work we explore whether a similar characterization of the elements of
the space Sp(B κ ) can be obtained when −1/2 < κ < 0 or 0 < κ < 1/2. In
particular, we are interested in the following question:

Is there a Hilbert space C of functions on the real line

which is isometric to Sp(B κ )?

We impose throughout a natural restriction on this isometry, requiring it to be an

extension of the map f → Iκ (f ), for f ∈ E. We show in this paper that such a
Hilbert space C exists when −1/2 < κ < 0. As in the case κ = 0, this means that:
(i) every element of the space Sp(B κ ) with −1/2 < κ < 0 can be expressed
as an integral of a deterministic function with respect to fBm B κ, and
(ii) the integral on the real line with respect to fBm B κ with −1/2 < κ < 0 is
well-defined for functions from this Hilbert space C.
When 0 < κ < 1/2, however, we do not know the answer to the above question.
This is quite surprising because typically the case 0 < κ < 1/2 is (or, is thought to
be) simpler than the case −1/2 < κ < 0. Instead, when 0 < κ < 1/2, we pursue a
simpler goal which is to characterize elements of some linear subspaces of Sp(B κ )
(the larger, the better). The relevant question in this context is:

What are the inner product spaces of functions on the real line
which are isometric to linear subspaces of Sp(B κ )?

We think here again of those isometries that extend the map f → Iκ (f ), for
f ∈ E. The interest of this question lies in the following. If C is such an inner
product space and SC is the corresponding subspace of Sp(B κ ), then:
(i) every element of the linear subspace SC can be expressed as an integral of
a deterministic function with respect to fBm B κ, and
(ii) the integral on the real line with respect to fBm B κ with 0 < κ < 1/2 is
well-defined for functions from C.
Note that, by (ii), an inner product space C can then be viewed as a class of
possible deterministic integrands. We will show that the spaces of deterministic
Integration questions related to fractional Brownian motion 255

integrands one usually considers, correspond to spaces C that are not complete and
hence cannot be isometric to Sp(B κ ) when 0 < κ < 1/2. If one completed them,
one would fall back to the case discussed in (1.6), namely, the representative fX for
a random variable X ∈ Sp(B κ ) would only be an (equivalent) class of sequences
of functions.
A trivial example of an inner product space C is the set of elementary functions
E itself, where the inner product is defined by E(Iκ (f )Iκ (g)), for f, g ∈ E.
Another example when 0 < κ < 1/2 is the space of functions L1 (⺢) ∩ L2 (⺢) with
the inner product defined by

 
κ(2κ + 1) f (s)g(t)|s − t|2κ−1 dsdt (1.7)
⺢ ⺢

(see Gripenberg and Norros [9], p. 404. The kernel |s −t|2κ−1 appears as |s −t|2H −2
in the H parametrization.) In this work we construct, analyze and compare a number
of such inner product spaces (or classes of integrands). We also unite the various
approaches for both 0 < κ < 1/2 and −1/2 < κ < 0. For instance, depending on
how fBm is represented, we can obtain a class of integrands in the “time domain”
or the “spectral domain” (the “time domain” representation of fBm is given in (3.7)
below and its “spectral domain” representation is given in (3.1) below).
The different spaces C we consider are denoted κ , κ and ||κ . The space

 
κ
|| = f : |f (u)||f (v)||u − v|2κ−1 du dv < ∞ (1.8)
⺢ ⺢

is used when 0 < κ < 1/2. Observe that the space L1 (⺢) ∩ L2 (⺢) with the inner
product (1.7) is a subspace of (1.8) because the functions are required to belong to
L1 (⺢) ∩ L2 (⺢) as well. We will also consider the larger inner product space κ ,
defined, using fractional integration, by (3.15) and (3.17) below when 0 < κ < 1/2
and by (3.30) and (3.31) when −1/2 < κ < 0. We will also analyze a class of
integrands in the “spectral domain”, namely, the inner product space κ defined
by (3.5) and (3.6). For the convenience of the reader, we gather at the end of the
paper the definitions of the various spaces and their inner products.
We show that the “time domain” spaces κ are complete only when −1/2 <
κ < 0 and that the “spectral domain” spaces κ are strict subsets of κ and not
complete for all −1/2 < κ < 1/2, κ = 0. As to ||κ , 0 < κ < 1/2, we show that
it is also a strict subset of κ and is not complete.
The paper is structured as follows. In Section 2 we recall how inner prod-
uct spaces or classes of integrands C are constructed. Then, in Sections 3 and
4, we deal with specific classes C. Most of the results are proved in Section 5.
In Section 6 we compare the classes of integrands obtained in the previous sec-
tions with the so-called reproducing kernel Hilbert space of fBm. This provides
a broader perspective on our results. The summary of the paper can be found in
Section 7.
256 V. Pipiras, M.S. Taqqu

2. General construction of classes of integrands

For the sake of clarity and completeness, we recall, in the proposition below, how
to construct classes of integrands C. This scheme will be used many times in the
sequel to obtain classes C.

Proposition 2.1. Let E be the set of elementary functions (1.4), Iκ (f ) be an inte-

gral of f ∈ E with respect to fBm B κ and −1/2 < κ < 1/2. Suppose that C is a
set of deterministic functions on the real line such that
(1) C is an inner product space with an inner product (f, g)C , for f, g ∈ C,
(2) E ⊂ C and (f, g)C = EIκ (f )Iκ (g), for f, g ∈ E,
(3) the set E is dense in C.
Then
(a) there is an isometry between the space C and a linear subspace of Sp(B κ )
which is an extension of the map f → Iκ (f ), for f ∈ E,
(b) C is isometric to Sp(B κ ) itself if and only if C is complete.

Proof. We first show part (a). Let f ∈ C. By (3), there is a sequence (fn ) ⊂ E
such that fn → f in C. In particular, (fn ) is Cauchy in C and hence, by (2),
(Iκ (fn )) is a Cauchy sequence in L2 ( ). Since the space L2 ( ) is complete,
there is Iκ (f ) ∈ L2 ( ) such that

Iκ (f ) = lim Iκ (fn ),
n

in the L2 ( )-sense. Moreover, since (Iκ (fn )) ⊂ Sp(B κ ) and Sp(B κ ) is a closed
subset of L2 ( ), we obtain that Iκ (f ) ∈ Sp(B κ ). We can thus define the map Iκ
from the space C into the space Sp(B κ ). It is easy to verify that this definition does
not depend on an approximating sequence (fn ). This construction of Iκ and (2)
imply that, for f, g ∈ C,

(f, g)C = EIκ (f )Iκ (g), (2.1)

and, since the map Iκ is linear, we conclude that Iκ is, in fact, an isometry between
the space C and a linear subspace of Sp(B κ ).
We now turn to part (b). If C is isometric to Sp(B κ ) itself, then C is complete
because the space Sp(B κ ) is complete (it is a closed subset of the complete space
L2 ( )). Conversely, if C is complete, then the map Iκ is onto because E is dense
in C and hence C is isometric to Sp(B κ ) itself. 


The isometry map Iκ obtained in the proof above is also denoted


Iκ (f ) = f (u)dB κ (u), (2.2)
⺢

for f ∈ C, and the right-hand side of (2.2) is called the integral on the real line of
f with respect to fBm B κ .
Observe that the isometry map Iκ might depend on the inner product space
C. In other words, if C1 and C2 are two different classes of functions that satisfy
Integration questions related to fractional Brownian motion 257

the conditions of Proposition 2.1, then, a priori, it is not clear whether the corre-
sponding isometry maps, say I1κ and I2κ , are equal on C1 ∩ C2 . (A necessary and
sufficient condition for this fact is (f, g)C1 = (f, g)C2 , ∀f, g ∈ C1 ∩ C2 .) For
this reason, when we consider in the sequel specific classes of integrands, we will
compare their corresponding isometry maps and use the same notation only if they
are equal.
Finally, as an example in the simple case when κ = 0, we can take C = L2 (⺢).
Since L2 (⺢) is complete, the map I0 is an isometry between the spaces C and
Sp(B 0 ). The cases −1/2 < κ < 0 and 0 < κ < 1/2, however, are neither
immediately obvious nor simple as the following sections show.

3. Spectral and time domain integrands

We use different representations of fBm and the scheme described in Proposition

2.1 to construct different classes of integrands. The first class, denoted by κ , is
obtained in the context of the “spectral domain” and the second one, denoted by
κ , arises in the context of the “time domain”. We then show that the former class
is a subset of the latter, that is κ ⊂ κ , and also prove that their corresponding
isometry maps Iκ ’s are equal almost surely on the smaller space κ . The fBm B κ ,
in the sequel, is assumed to be standard.

3.1. Class of integrands in the “spectral domain” for −1/2 < κ < 1/2

Recall that a standard fBm {B κ (t)}t∈⺢ with index −1/2 < κ < 1/2 has the spectral
representation (see Samorodnitsky and Taqqu [15], p. 328)
 itx
d 1 e − 1 −κ
{B κ (t)}t∈⺢ = |x| d B(x) , (3.1)
c2 (κ) ⺢ ix t∈⺢

where
1/2
2π
c2 (κ) = (3.2)
(2κ + 2) sin(π(κ + 1/2))
and B = B 1 + iB 2 is a complex Gaussian measure such that B 1 (A) = B 1 (−A),
B 2 (A) = −B 2 (−A) and E(B 1 (A))2 = E(B 2 (A))2 = |A|/2, for any Borel set
A of finite Lebesgue measure |A| (see Section 7.2.2 in Samorodnitsky and Taqqu
[15]). Observe now that (eitx − 1)/ ix
=  1[0,t) (x), where f denotes the Fourier
 ixu
transform of a function f , i.e. f (x) = ⺢ e f (u)du. Then, for f ∈ E,

d 1
Iκ (f ) = f(x)|x|−κ d B(x) (3.3)
c2 (κ) ⺢
and hence (see (7.2.9) in Samorodnitsky and Taqqu [15]) that, for f, g ∈ E,

1
E(Iκ (f )Iκ (g)) = f(x)
g (x)|x|−2κ dx. (3.4)
c2 (κ)2 ⺢
Based on (3.4), we introduce
258 V. Pipiras, M.S. Taqqu

Definition 3.1.

κ
 = f : f ∈ L (⺢), 2
|f(x)|2 |x|−2κ dx < ∞ , (3.5)
⺢

for −1/2 < κ < 1/2.

The following theorem and (3.4) show that κ satisfies the conditions (1), (2) and
(3) of Proposition 2.1. This theorem is proved in Section 5.1.

Theorem 3.1. For −1/2 < κ < 1/2, the class of functions κ , defined by (3.5),
is a linear space with the inner product

1
(f, g)κ = f(x)
g (x)|x|−2κ dx. (3.6)
c2 (κ)2 ⺢

The set of elementary functions E is dense in κ . Moreover, the space κ is not

complete unless κ = 0.

It follows from Proposition 2.1 that there is an isometry map between

the space
κ and a linear subspace of Sp(B κ ) which is denoted by Iκ (f ) = ⺢ f (u)dB κ (u),
for f ∈ κ. This linear subspace is a strict subset of Sp(B κ ) when κ = 0 because,
by the above theorem, κ is not complete. The integral Iκ so defined satisfies the
relations (3.3) and (3.4) for f, g ∈ κ .
We conclude this section with a proposition which characterizes a large subset
of κ in the case 0 < κ < 1/2 (this result is mentioned in Barton and Poor [2],
p. 945).

Proposition 3.1. For 0 < κ < 1/2, f ∈ L1 (⺢) ∩ L2 (⺢) implies f ∈ κ .

Proof. The proposition follows from the estimate

  
|f(x)|2 |x|−2κ dx =  −2κ
|f (x)| |x| dx +
2
|f(x)|2 |x|−2κ dx
⺢ |x|≤1 |x|>1
 
≤ f L1 (⺢)
2
|x| dx + |f(x)|2 dx
−2κ
|x|≤1 ⺢
1
= f 2L1 (⺢) + 2π f 2L2 (⺢) ,
1/2 − κ


1/2
where as usual f L1 (⺢) = ⺢ |f (u)|du and f L2 (⺢) = ⺢ |f (u)|2 du . 


Remarks.
1. Observe that f ∈ L1 (⺢)∩L2 (⺢) is a condition on f , whereas f ∈ κ involves
a condition on f.
Integration questions related to fractional Brownian motion 259

2. Proposition 3.1 does not hold when −1/2 < κ < 0. As a counterexample, con-
sider the function f (u) = sgn(u)e−|u| /|u|p with κ + 1/2 < p < 1/2. Clearly,
f ∈ L1 (⺢) ∩ L2 (⺢). Its Fourier transform equals (see formula 3.945,(1) in
Gradshteyn and Ryzhik [7], p. 491)
p−1
2(1 − p)(x 2 + 1) 2 sin[(1 − p)Arctan(x)]

and hence it behaves (up to a constant) like |x|p−1 at x = ∞. Since 2p − 2κ −

2 + 1 = 2(p − κ − 1/2) > 0, the function |x|2(p−1) |x|−2κ = |x|2p−2κ−2 is not
integrable around x = ∞. This implies that f ∈ / κ for −1/2 < κ < 0.

3.2. Class of integrands in the “time domain” for 0 < κ < 1/2

Instead of working in the “spectral domain” as in Section 3.1, we now work in the
“time domain”. Recall that a standard fBm {B κ (t)}t∈⺢ with index −1/2 < κ < 1/2
has the moving average representation (see Samorodnitsky and Taqqu [15], p. 320)

d 1  
{B κ (t)}t∈⺢ = (t − s)κ+ − (−s)κ+ dB(s) , (3.7)
c1 (κ) ⺢ t∈⺢

where B is a standard Brownian motion,

 ∞
 2 1 1/2
c1 (κ) = (1 + s)κ − s κ ds + , (3.8)
0 2κ + 1
κ = 0, for a ≤ 0, and a κ = a κ , for a > 0. Observe that, when 0 < κ < 1/2,
a+ +

κ κ
(t − s)+ − (−s)+ = κ 1[0,t) (u)(u − s)κ−1
+ du, (3.9)
⺢

where 1[0,t) is interpreted as (−1[t,0) ), if t < 0. Let I−α φ denote a fractional integral
of order α > 0 of a function φ defined by

1
α
(I− φ)(s) = φ(u)(s − u)α−1
− du, (3.10)
(α) ⺢

1
= φ(u)(u − s)α−1
+ du, s ∈ ⺢ (3.11)
(α) ⺢
(see Chapter 2 in Samko et al. [14] and, in particular, (5.2) on p. 94. While the
right-hand side of (3.10) indicates that we are dealing with the convolution of φ
with uα−1
− , relation (3.11) is more convenient to manipulate.) Then the relation
(3.9) can be expressed as

(t − s)κ+ − (−s)κ+ = (κ + 1)(I−κ 1[0,t) )(s). (3.12)

It follows from (3.7) and (3.12) that, for f ∈ E and 0 < κ < 1/2,
 
κ d (κ + 1)
f (u)dB (u) = (I−κ f )(s)dB(s) (3.13)
⺢ c1 (κ) ⺢
260 V. Pipiras, M.S. Taqqu

and hence that, for f, g ∈ E and 0 < κ < 1/2,


(κ + 1)2
E(Iκ (f )Iκ (g)) = (I−κ f )(s)(I−κ g)(s)ds. (3.14)
c1 (κ)2 ⺢

This leads us to introduce the following class of functions

Definition 3.2.

 κ 2
κ = f : (I− f )(s) ds < ∞
 ⺢
  2 
κ−1
= f : f (u)(u − s)+ du ds < ∞ , (3.15)
⺢ ⺢

for 0 < κ < 1/2.

The inner integral in (3.15) is understood in the Lebesgue sense, which implies in
particular that

|f (u)|(u − s)κ−1
+ du < ∞ a.e. ds. (3.16)
⺢

The following theorem is the “time domain” analogue of Theorem 3.1 and is proved
in Section 5.2.

Theorem 3.2. For 0 < κ < 1/2, the class of functions κ , defined by (3.15), is a
linear space with the inner product

  (κ + 1)2
f, g κ = (I−κ f )(s)(I−κ g)(s)ds. (3.17)
c1 (κ)2 ⺢

The set of elementary functions E is dense in the space κ . The space κ is not
complete.

Theorem 3.2 and the relation (3.14) show that κ satisfies the conditions of
Proposition 2.1. Hence, by Proposition 2.1, κ is isometric to
a strict linear sub-
space of Sp(B κ ). We denote this isometry map by Iκ (f ) = ⺢ f (u)dB κ (u), for
f ∈ κ , as in the “spectral domain”. The use of the same notation is justified in
Section 3.4. Observe also that the relations (3.13) and (3.14) hold for f ∈ κ .
Remarks.
1. In the case 0 < κ < 1/2, by (3.15),
 
κ = f : φf = I−κ f ∈ L2 (⺢) , (3.18)

and hence (3.17) states that

  (κ + 1)2  
f, g κ
= φ f , φg L2 (⺢)
. (3.19)
c1 (κ)2
Integration questions related to fractional Brownian motion 261

2. The fractional integral notation provides a convenient way of interpreting

(3.13). For α > 0 and “sufficiently good” functions ψ and φ, the fractional
integration by parts formula (see (5.17) in Samko et al. [14])
 
(I−α ψ)(s)φ(s)ds = ψ(u)(I+α φ)(u)du (3.20)
⺢ ⺢

holds, where I+α φ,

α > 0, is the fractional integral defined by
 u 
1 φ(s)ds 1
(I+α φ)(u) = = φ(s)(u − s)α−1
+ ds, u ∈ ⺢
(α) −∞ (u − s)1−α (α) ⺢
(3.21)
(see Chapter 2 in Samko et al. [14] and, in particular, (5.3) on p. 94). Then the
relation (3.13) with f ∈ κ can be informally expressed as
 
dB κ (κ + 1) dB
f (u) (u)du = f (u) I+κ (u)du
⺢ du c 1 (κ) ⺢ du
or
dB κ (κ + 1) κ dB
(u) = I+ (u).
du c1 (κ) du
Thus, when 0 < κ < 1/2 the “fractional noise” dB κ /du is a κ–fractional
integral of the “white noise” dB/du. This fact could be formalized by using
generalized processes.
We now establish an analogue to Proposition 3.1.
Proposition 3.2. For 0 < κ < 1/2,

L1 (⺢) ∩ L2 (⺢) ⊂ L2/2κ+1 (⺢) ⊂ κ . (3.22)

Proof. The first inclusion in (3.22) follows from the inequality

|f (u)|2/2κ+1 ≤ |f (u)| + |f (u)|2 ,

which holds for 0 < κ < 1/2. On the other hand, the operator I−κ maps L2/2κ+1 (⺢)
into L2 (⺢) for 0 < κ < 1/2 (Theorem 5.3 in Samko et al. [14], p. 103). This implies
the second inclusion in (3.22). 


Remark. It is easy to verify directly that, for 0 < κ < 1/2, f ∈ L1 (⺢) ∩ L2 (⺢)
implies f ∈ κ . Indeed,
  2
κ−1
|f (u)|(u − s)+ du ds
⺢ ⺢
  2
κ−1
≤2 |f (u)|(u − s)+ 1{u−s>1} (u)du ds
⺢ ⺢
  2
+2 |f (u + s)| uκ−1
+ 1 {u<1} (u)du ds < ∞,
⺢ ⺢

after applying (3.37) below.

262 V. Pipiras, M.S. Taqqu

3.3. Class of integrands in the “time domain” for −1/2 < κ < 0

To construct the class of integrands κ in the “time domain” when 0 < κ < 1/2,
we used the moving average representation (3.7) of fBm and the relation (3.12).
Although (3.12) holds for 0 < κ < 1/2 only, there is an analogous relation in the
case −1/2 < κ < 0. Recall (see Samko et al. [14], p. 111) that Marchaud fractional
derivatives Dα± φ of order α ∈ (0, 1) of a function φ are defined by
Dα± φ = lim Dα±,) φ, (3.23)
)→0
where  ∞
α φ(s) − φ(s ∓ u)
(Dα±,) φ)(s) = du. (3.24)
(1 − α) ) u1+α
We then have
Lemma 3.1. If −1/2 < κ < 0 and t > 0, then, for all s ∈ ⺢,
(t − s)κ+ − (−s)κ+ = (κ + 1)(D−κ
− 1[0,t) )(s). (3.25)
The formula (3.25) also holds for t < 0, if 1[0,t) is interpreted as (−1[−t,0) ).
Proof. Let t > 0. By (3.24), we have
 ∞
 
(κ + 1)(D−κ 1
−,) [0,t) )(s) = −κ 1[0,t) (s) − 1[0,t) (s + u) uκ−1 du
 )
=κ 1[0,t) (v)(v − s)κ−1 κ
+ 1{)<v−s} (v)dv + 1[0,t) (s)) .
⺢
(3.26)
One can verify that, if ) < t, the function (3.26) is equal to (t −s)κ+ −(−s)κ+ ,
s ∈ ⺢,
on the set A) = {s < −)} ∪ {0 ≤ s < t − )} ∪ {s ≥ t}. The relation (3.25) follows
by letting ) → 0. When t < 0, the proof of (3.25) is similar. 

By using the moving average representation (3.7) of fBm and (3.25), we obtain
that, for f ∈ E and −1/2 < κ < 0,
 
d (κ + 1)
κ
f (u)dB (u) = (D−κ
− f )(s)dB(s) (3.27)
⺢ c1 (κ) ⺢
and hence that, for f, g ∈ E and −1/2 < κ < 0,

(κ + 1)2
E(Iκ (f )Iκ (g)) = (D−κ −κ
− f )(s)(D− g)(s)ds. (3.28)
c1 (κ)2 ⺢

Observe now that, if f = I−−κ φ, for some φ ∈ L2 (⺢), then Theorem 6.1 in
Samko et al. [14], p. 125, implies that
D−κ −κ −κ
− f = D− (I− φ) = φ (3.29)
and hence that  
 2
(D−κ
− f )(s) ds = φ 2 (u)du < ∞.
⺢ ⺢
Based on this observation and the relation (3.28), we introduce the next class of
functions.
Integration questions related to fractional Brownian motion 263

Definition 3.3. Let

 
κ = f : ∃ φ ∈ L2 (⺢) such that f = I−−κ φ , (3.30)

for −1/2 < κ < 0.

(Note that the class κ coincides with the space of functions I−−κ (L2 ) from
Samko et al. [14], Chapter 6, p. 122.) Compare this definition of κ with the cor-
responding definition for 0 < κ < 1/2 (see (3.18)). The definition of κ in the
cases 0 < κ < 1/2 and −1/2 < κ < 0 can be viewed as dual to each other.
Theorem 3.2 concerned κ when 0 < κ < 1/2. The following theorem deals
with κ in the case −1/2 < κ < 0 and is proved in Section 5.3.

Theorem 3.3. For −1/2 < κ < 0, the class of functions κ , defined by (3.30), is
a linear space with the inner product

  (κ + 1)2
f, g κ = (D−κ −κ
− f )(s)(D− g)(s)ds. (3.31)
c1 (κ)2 ⺢

The set of elementary functions E is dense in the space κ . The space κ is complete
and hence it is isometric to Sp(B κ ).

Remarks.
1. The definition (3.30) of κ states that to each f ∈ κ there corresponds a
function φf ∈ L2 (⺢) such that f = I−−κ φf . By (3.29), the inner product (3.31)
can be expressed as

  (κ + 1)2  
f, g κ
= φf , φg L2 (⺢) . (3.32)
c1 (κ) 2

While (3.32) also holds when 0 < κ < 1/2 (see (3.19)), φf plays a different
role (see (3.18)).
2. Theorem 3.3 states that, for −1/2 < κ < 0, κ is in fact isometric to the space
Sp(B κ ) itself because κ is complete. This is where the case −1/2 < κ < 0
is different from the case 0 < κ < 1/2 considered in Section 3.2: by Theorem
3.2, κ is not complete for 0 < κ < 1/2. The proofs of Theorems 3.2 and 3.3
show that this difference in completeness is a consequence of the following two
facts:
(a) if −1/2 < κ < 0, then the equation

D−κ
− f =φ

has a solution f = I−−κ φ for every φ ∈ L2 (⺢),

(b) when 0 < κ < 1/2, however, there are functions φ ∈ L2 (⺢) for which the
equation
I−κ f = φ
is not solvable.
264 V. Pipiras, M.S. Taqqu

3. Observe that the relations (3.27) and (3.28) hold for f, g ∈ κ as well. As in
the case 0 < κ < 1/2, we may express the relation (3.27) informally as

dB κ (κ + 1) dB
(u) = D−κ
+ (u).
du c1 (κ) du

Thus, when −1/2 < κ < 0 the “fractional noise” dB κ /du is a (−κ)–fractional
derivative of the “white noise” dB/du.

3.4. “Time domain” versus “spectral domain”

In this section we compare the classes of integrands κ and κ , as well as the

corresponding isometry maps Iκ ’s or, equivalently, the definitions of the integral
κ
⺢ f (u)dB (u) in the “spectral domain” and the “time domain”.

Proposition 3.3. Suppose either −1/2 < κ < 0 or 0 < κ < 1/2. Then:
(1) The inclusion
κ ⊂ κ (3.33)
holds.
(2) For f, g ∈ κ ,
   
f, g κ
= f, g κ . (3.34)

(3) For f ∈ κ , the Fourier transform of I−κ f or D−κ

− f is

C κ (x)f(x)|x|−κ , (3.35)

where
C κ (x) = e−iπκ/2 1{x>0} + eiπκ/2 1{x<0} , x ∈ ⺢. (3.36)

Proof. The proof is in two parts: (1) 0 < κ < 1/2, and (2) −1/2 < κ < 0.
(1) We first assume that 0 < κ < 1/2. In the proof given below we make use
of the following two facts:
(i) If f1 ∈ L1 (⺢) and f2 ∈ L2 (⺢) are two functions, then their convolution
f1 ∗ f2 ∈ L2 (⺢) because
  2 1/2
|f1 (u)| |f2 (s − u)|du ds
⺢ ⺢
  1/2
≤ |f1 (u)| |f2 (s − u)| ds
2 2
du
⺢ ⺢
= f1 L1 (⺢) f2 L2 (⺢) . (3.37)

This last relation follows from the generalized Minkowsky inequality

  p  
  1/p 1/p
 g(u, s)du ds ≤ |g(u, s)|p ds du, (3.38)
 
⺢ ⺢ ⺢ ⺢
Integration questions related to fractional Brownian motion 265

which holds for p ≥ 1 (see (1.33) in Samko et al. [14], p. 9). Moreover, by ap-
proximating f2 with functions f2,n = f2 1{|·|≤n} which are in L1 (⺢) ∩ L2 (⺢) and
converge to f2 in L2 (⺢), one has
f1
∗ f2 (x) = f1 (x)f2 (x).
(ii) If f ∈ L2 (⺢), then the function
 
f (u)(u − s)κ−1
+ du = f (u)(u − s)κ−1
+ 1{u−s≤1} du
⺢ ⺢

+ f (u)(u − s)κ−1 + 1{u−s>1} du

⺢

is well-defined because the first term is well-defined by (3.37) and the second term
by the Cauchy-Schwarz inequality.
Now let f ∈ κ and set, for n ∈ ⺞,

1
h(s) = (I−κ f )(s) = f (u)(u − s)κ−1
+ du,
(κ) ⺢

1
hn (s) = f (u)(u − s)κ−1
+ 1{u−s<n} du.
(κ) ⺢
Since f ∈ L2 (⺢), h and hn are well-defined by part (ii). The functions hn converge
to h a.e. ds. Let us show that they converge to h in L2 (⺢) as well. Since hn is the
convolution (up to a constant) of the L2 (⺢)-function f (u) and the L1 (⺢)-function
uκ−1
− 1{u− <n} , it belongs to L (⺢) by part (i) and
2


 1  
hn (x) = f (x) eixu uκ−1 κ
− 1{u− <n} du = Cn (x)f (x)|x| ,
−κ
(κ) ⺢

where
 n|x|  n|x|
1
Cnκ (x) = e −iv κ−1
v dv1{x>0} + eiv v κ−1 dv1{x<0} .
(κ) 0 0

∞
Since the improper integral 0 e±iv v γ −1 dv equals e±iπγ /2 (γ ), for γ > 0, the
functions Cnκ (x) converge to C κ (x) given by (3.36). Observe also that Cnκ (x)’s are
uniformly bounded in n and x. We then have

hn − hm 2L2 (⺢) = (2π)−1 

hn − 
hm 2L2 (⺢)

= (2π)−1 |f(x)|2 |x|−2κ |Cnκ (x) − Cm
κ
(x)|2 dx.
⺢


Since by assumption ⺢ |f(x)|2 |x|−2κ dx < ∞, the dominated convergence theo-
rem implies that hn −hm 2L2 (⺢) → 0, as n, m → ∞. Therefore, there is h ∈ L2 (⺢)
such that hn − h2L2 (⺢) → 0, as n → ∞. By taking an a.e. convergent subse-
quence, we can deduce that h(s) = h(s) a.e. ds. In particular, h ∈ L2 (⺢) or,
equivalently, f ∈ κ . This establishes the inclusion (3.33) when 0 < κ < 1/2.
266 V. Pipiras, M.S. Taqqu

Observe that hn (x) converges to the function given by (3.35) for all x ∈ ⺢,
x = 0. On the other hand, since hn converges to h in L2 (⺢),  hn converges to  h
in L2 (⺢) as well. Hence, the Fourier transform of h = I−κ f , for f ∈ κ , is given
by (3.35). Since |C κ (x)| = 1, if x = 0, the Parseval’s equality implies that, for
f, g ∈ κ ,
 
1
(I−κ f )(s)(I−κ g)(s)ds = f(x)
g (x)|x|−2κ dx,
⺢ 2π ⺢

or, by (3.17) and (3.6), that

   
f, g κ
= c(κ) f, g κ ,
   
for some constant c(κ). Since both f, g κ and f, g κ are equal to
E(Iκ (f )Iκ (g)), for f, g ∈ E, we get c(κ) = 1. Thus, the relation (3.34) holds
when 0 < κ < 1/2.
(2) We now consider the case −1/2 < κ < 0. Let us show that, if f ∈ κ ,
then f ∈ κ as well, that is, there is φ ∈ L2 (⺢) such that

f = I−−κ φ. (3.39)

Since f(x)|x|−κ is in L2 (⺢), |C −κ (x)| = 1, if x = 0, and C −κ (x) = C −κ (−x),

the function
(x) = (C −κ (x))−1 f(x)|x|−κ
φ (3.40)
(x) = φ
is in L2 (⺢), satisfies φ (−x) and hence is the Fourier transform of some
function φ ∈ L (⺢). Let us show that this function satisfies (3.39). From (3.40) we
2

get
f(x) = C −κ (x)φ
(x)|x|κ = C −κ (x)φ
(x)|x|−(−κ) . (3.41)
This implies that
(x)|2 |x|−2(−κ) = |f(x)|2
|φ
and, since f ∈ L2 (⺢), that φ ∈ −κ . Since 0 < −κ < 1/2, we can apply (3.35) to
evaluate the Fourier transform of I−−κ φ. Since it is identical to (3.41), we conclude
that
f = I−−κ φ,
that is, (3.39) holds and hence f ∈ κ .
To verify (3.35), we must evaluate the Fourier transform of D−κ − f . Observe
that our φ satisfies (3.29) and hence the Fourier transform of D−κ
− f is the Fourier
transform of φ, that is (3.40). Since (C −κ (x))−1 = C κ (x), we obtain (3.35). The
relation (3.34) for −1/2 < κ < 0 follows as in the case 0 < κ < 1/2. 


Corollary

3.1. Let f ∈ κ , for −1/2 < κ < 0 or 0 < κ < 1/2. Then the inte-
grals ⺢ f (u)dB κ (u) defined in the “spectral domain” and the “time domain” are
equal almost surely.
Integration questions related to fractional Brownian motion 267

Proof. The integral ⺢ f (u)dB κ (u) in the “spectral domain” is defined as an


L2 ( )-limit of the integrals ⺢ fn (u)dB κ (u), where (fn ) is a sequence of ele-
mentary functions
such that fn → f in κ . By (3.34), fn → f also in κ . There-
fore, the integral ⺢ f (u)dB κ (u) in the “time domain” is also an L2 ( )-limit of


fn (u)dB κ (u). Since an L2 ( )-limit of a sequence of random variables is almost
surely unique, we obtain the required result. 


Proposition 3.4. The inclusion κ ⊂ κ is strict when −1/2 < κ < 0 or 0 <
κ < 1/2.

Proof. Consider the case 0 < κ < 1/2 first. It is enough to find a function f
such that f ∈ κ but f ∈ / L2 (⺢) (so that its L2 –Fourier transform is not defined
and hence f ∈ / κ ). By Theorem 5.3 in Samko et al. [14], p. 103, the operator
I−κ maps L2/2κ+1 (⺢) into L2 (⺢). Therefore, for f we can take any function from
L2/2κ+1 (⺢) which is not in L2 (⺢). Suppose now that −1/2 < κ < 0. If κ = κ ,
then the inner product spaces κ and κ are identical because, by (3.34), the cor-
responding inner products are equal. However, this is not possible, since κ is not
a complete inner product space (Theorem 3.1) whereas κ is a complete inner
product space (Theorem 3.3). Therefore, the inclusion κ ⊂ κ is strict in the
case −1/2 < κ < 0 as well. 


4. An alternative class of integrands in the “time domain”

If  κ is the covariance function of a standard fBm B κ with −1/2 < κ < 1/2 given
by (1.1), then, for f, g ∈ E,
 
E(Iκ (f )Iκ (g)) = f (u)g(v)d 2  κ (u, v), (4.1)
⺢ ⺢

where the double integral is defined to be linear and to satisfy

  

d 2  κ (u, v) = .2  κ 
[a,b] [c,d] [a,b]×[c,d]
 
=  κ (d, b) −  κ (d, a) −  κ (c, b) −  κ (c, a) ,

numbers a < b and c < d. This may suggest that one can define the
for any real
integral ⺢ f (u)dB κ (u) for functions f from the space
 
κ
|| = f : |f (u)||f (v)|d 2 | κ |(u, v) < ∞ ,
⺢ ⺢

where | κ |
is the total variation measure of  κ . Observe, however, that when
−1/2 < κ < 0 the function  κ is not of bounded variation (around the diagonal
u = v) and hence the measure | κ | is not defined. But in the case 0 < κ < 1/2,
we have
d 2  κ (u, v) = κ(2κ + 1)|u − v|2κ−1 du dv, (4.2)
and hence, we have
268 V. Pipiras, M.S. Taqqu

Definition 4.1.
 
κ
|| = f : |f (u)||f (v)||u − v|2κ−1 du dv < ∞ ,
⺢ ⺢

for 0 < κ < 1/2.

Although the class ||κ arises in the “time domain” as naturally as the class
κ , interestingly, it turns out that ||κ is a strict subset of κ .

Proposition 4.1. Let 0 < κ < 1/2. Then the inclusion

||κ ⊂ κ

holds and it is strict.

Proof. The inclusion ||κ ⊂ κ follows from the relation

  2
|f (u)|(u − s)κ−1
+ du ds
⺢ ⺢
   
= |f (u)||f (v)| (u − s)κ−1
+ (v − s) κ−1
+ ds du dv
⺢ ⺢ ⺢
 
= B (κ, 1 − 2κ) |f (u)||f (v)||u − v|2κ−1 du dv, (4.3)
⺢ ⺢

which can be obtained by making the change of variables

1
s = min(u, v) − |v − u| −1
z

1
above (B(p, q) = 0 (1 − v)p−1 v q−1 dv, p, q > 0, is the beta function). To
κ ⊂ κ is strict, we have to find f ∈ κ for which


that the inclusion ||
show
⺢ ⺢ |f (u)||f (v)||u − v| dudv = ∞. By Proposition 3.3, it is enough to pro-
2κ−1

vide an example of a function f which is in κ but not in ||κ . Take f (u) =

sign(u)|u|−p sin(u) with 1/2 < p < 1, which is in L2 (⺢). To calculate its Fourier
transform, consider a sequence of functions fn (u) = sign(u)|u|−p sin(u)1{|u|<n} ,
which converges to f in L2 (⺢). The Fourier transform of fn is
 n
fn (x) = 2 cos(xu)|u|−p sin(u)du
 n0  n
= u−p sin[(x + 1)u]du − u−p sin[(x − 1)u]du
0 0
 n|x+1|
= sign(x + 1)|x + 1|p−1 v −p sin(v)dv
0
 n|x−1|
p−1
−sign(x − 1)|x − 1| v −p sin(v)dv.
0
Integration questions related to fractional Brownian motion 269

It converges to the limit

  ∞
p−1 p−1
sign(x + 1)|x + 1| − sign(x − 1)|x − 1| v −p sin(v)dv,
0


which is also f. Since 1/2 < p < 1, we have ⺢ |f(x)|2 |x|−2κ dx < ∞ and hence
f ∈ κ . Let us show that, for 1/2 < p < κ + 1/2,
  2
κ−1
I= |f (u)|(u − s)+ du ds = ∞,
⺢ ⺢

that is, by (4.3), f ∈

/ ||κ .
We have
  2
I = |u|−p | sin u|(u − s)κ−1
+ du ds
⺢ ⺢
  2
≥ u −p
| sin u|(u − s)κ−1
+ du ds
⺢ {u>π/4}
 
∞  πk+3π/4
2
1
≥ u −p
(u − s)κ−1
+ du ds
2 ⺢ k=0 πk+π/4
 
∞  πk+3π/4
2
1
≥ u−p (u − s)κ−1
+ du ds
4 ⺢ k=0 πk+π/4
 
∞  πk+π/4 −p
2
1 π π
+ u+ (u + − s)κ−1
+ du ds
4 2 2
⺢ k=1 πk−π/4
 
∞  πk+3π/4
2
1
≥ u−p (u − s)κ−1
+ du ds
4 ⺢ k=0 πk+π/4
  ∞  πk+π/4
2
2−2p κ−1
+ u−p (u − s)+ du ds
4 ⺢ k=1 πk−π/4
 ∞  πk+3π/4
2−2p
≥ u−p (u − s)κ−1
+ du
8 ⺢ k=0 πk+π/4

∞  πk+π/4
2

+ u−p (u − s)κ−1
+ du ds
k=1 πk−π/4
  2
= cp u−p (u − s)κ−1
+ du ds
⺢ {u>π/4}
 ∞  2
≥ cp v −p
(v − 1)κ−1
+ dv s 2κ−2p ds. (4.4)
0 {v≥π/4s}

(We have used the inequalities | sin u| > 1/2, for u ∈ (π k + π/4, π k + 3π/4),
a 2 + b2 ≥ (a + b)2 /2 and (u + π2 )−p ≥ (2u)−p , for u > 3π/4.) Since the lower
270 V. Pipiras, M.S. Taqqu

bound (4.4) diverges around s = ∞ for 1/2 < p < κ + 1/2, we obtain that
I = ∞. 

Observe that by (4.3) we also have
   2 
||κ = f : |f (u)|(u − s)κ−1
+ du ds < ∞ (4.5)
⺢ ⺢

and comparing (4.5) with (3.15), we observe that the difference between ||κ and
κ is that the absolute value of f appears in (4.5). Since ||κ ⊂ κ , one can
 define

an inner product on ||κ in the same way as for κ , that is, f, g ||κ = f, g κ ,
for f, g ∈ ||κ , and then ||κ becomes an inner product space. Since the set E is
dense in κ (Theorem 3.2), it is also dense in ||κ . Then, by Proposition 2.1, there
is an isometry map between the space ||κ and a linear subspace of Sp(B κ ). If we
denote this isometry map by |I|κ , then clearly |I|κ (f ) = Iκ (f ), for f ∈ ||κ ,
κ κ
wherethe isometry
κ
  map I corresponds to the space  . The inner product on
|| , f, g ||κ = f, g κ , can be expressed as in (3.17), or alternatively as
 
 
f, g ||κ = κ(2κ + 1) f (u)g(v)|u − v|2κ−1 du dv. (4.6)
⺢ ⺢


Indeed, by writing as in (4.3), B (κ, 1 − 2κ) |u − v|2κ−1 = ⺢ (u − s)κ−1 + (v −
s)κ−1
+ ds, and applying the Fubini’s theorem, for f, g ∈ || κ , we get
 
B(κ, 1 − 2κ) f (u)g(v)|u − v|2κ−1 du dv
⺢ ⺢
    
= f (u)(u − s)κ−1
+ du g(u)(u − s) κ−1
+ du ds,
⺢ ⺢ ⺢
 
which is f, g κ up to a multiplicative constant which depends on κ only. Thus
the relation (4.6) holds up to a multiplicative constant which also depends on κ
only. This constant, however, equals 1. This can be verified by direct computations
(it is complicated!) or, more simply, by noting that both the left-hand side and
the right-hand side of (4.6), the latter by the relations (4.1) and (4.2), are equal to
E(Iκ (f )Iκ (g)), for f, g ∈ E.
Since ||κ is dense in κ (Lemma 5.5) and κ is not complete (Theorem 3.2),
the inner product space ||κ is not complete either. This is equivalent to saying
that the normed space ||κ , with the norm
  1/2
f ||κ = κ(2κ + 1) f (u)f (v)|u − v|2κ−1 du dv
⺢ ⺢

induced by the inner product, is not complete. However, if we introduce a new norm
on ||κ ,
  1/2
f  = |f (u)||f (v)||u − v|2κ−1 du dv ,
⺢ ⺢
we show in the theorem below that (||κ ,  · ) becomes a complete normed space.
The norm f  is also easier to work with than the norm f ||κ .
Integration questions related to fractional Brownian motion 271

Theorem 4.1. Let 0 < κ < 1/2. Then the function · defines a norm on ||κ and
the space (||κ , ·) is a complete normed space.

Proof. Using (4.3), it is easy to verify that · is a norm. To prove completeness, let
(fn ) be a Cauchy sequence in (||κ , ·), that is, fn − fm  → 0, as n, m → ∞.
We then have to find f ∈ ||κ such that f − fn  → 0, as n → ∞. This is done
in two steps:
(i) there is a subsequence (nl ) and a function f such that fnl →f a.e., as
l→∞, and
(ii) f is indeed the required function.
Let us first prove the step (i). Since, for any k > 0,
 
g ≥ (2k)
2 2κ−1
|g(u)||g(v)|du dv = (2k)2κ−1 g2L1 (−k,k) ,
|u|≤k |v|≤k

the convergence fn − fm  → 0 implies that fn − fm L1 (−k,k) → 0, ∀ k > 0.

Since the spaces L1 (−k, k) are complete for all k > 0, it is easy to see that there
is a function f , defined on the real line, such that fn − f L1 (−k,k) → 0, ∀ k > 0.
To choose an almost everywhere convergent subsequence, we use the well-known
diagonal argument. For k = 1, since fn − f L1 (−1,1) → 0, we can take a subse-
quence (nl (1)) such that fnl (1) → f a.e. on (−1, 1). By mathematical induction, we
can thus construct subsequences (nl (k+1)) ⊂ (nl (k)), k ∈ ⺞, such that fnl (k) → f
a.e. on (−k, k). Now, by taking the indices (nl ) = (nl (l)) in the diagonal, we obtain
that fnl → f a.e. on the whole real line.
The step (ii) follows easily from the step (i) because, by Fatou’s lemma, f  =
liml fnl  ≤ liml fnl  < ∞ and f − fn  ≤ liml fnl − fn  ≤ liml fnl −
n→∞
fn  −→ 0, since (fn ) is Cauchy in the norm  · . 


Proposition 4.2. For 0 < κ < 1/2,

L1 (⺢) ∩ L2 (⺢) ⊂ L2/2κ+1 (⺢) ⊂ ||κ ⊂ κ . (4.7)

Proof. After using (3.22), it remains to show that L2/2κ+1 (⺢) ⊂ ||κ . As noted
in the proof of Proposition 3.2, the operator I−κ is bounded from L2/2κ+1 (⺢) into
L2 (⺢) for 0 < κ < 1/2, implying that I−κ f L2 (⺢) ≤ c1 (κ)f L2/2κ+1 (⺢) and
therefore
f  = c2 (κ)I−κ |f |L2 (⺢) ≤ c3 (κ)f L2/2κ+1 (⺢) (4.8)
for some positive constants ci (κ), i = 1, 2, 3. 


5. Proof of Theorems 3.1, 3.2 and 3.3

5.1. The proof of Theorem 3.1

We let λκ be the measure on the real line defined by λκ (dx) = |x|−2κ dx, for
−1/2 < κ < 1/2. In the proof of Theorem 3.1 we will use the following result.
272 V. Pipiras, M.S. Taqqu

Lemma 5.1. Suppose that fis the Fourier transform of a function f ∈ L2 (⺢). If,
for some −1/2 < κ < 1/2, f ∈ L2 (λκ ), then there is a sequence of elementary
functions ln such that

n→∞
f− ln 2L2 (λκ ) = |f(x) − 
ln (x)|2 |x|−2κ dx −→ 0. (5.1)
⺢

Proof. Since f (u) = (1/2)(f (u) + f (−u)) + (1/2)(f (u) − f (−u)), for u ∈ ⺢,
we may prove the lemma in two cases: (1) f is an even function and, (2) f is an
odd function.
Case 1: If f is an even function, then f is real-valued and f(x) = f(−x). To
prove (5.1), we may show that, for arbitrary small ) > 0, there is an elementary
function l such that f −  lL2 (λκ ) < ). We will provide this approximation in
several steps. As a first step, we approximate fby simple functions. For any ) > 0,

there is a simple function  g (x) = kl=1 gl 1Gl (x), gl ∈ ⺢, Gl ∈ B(⺢), such that
f− g L2 (λκ ) < ). Since f(x) = f(−x), we can take the sets Gl to be symmetric
around the origin x = 0. As a second step, observe that, for a symmetric around the
origin set G and any ) > 0, there is a function  h(x) = m n=1 hn 1[−Hn ,Hn ] (x), hn ∈
⺢, Hn > 0, such that 1G −  hL2 (λκ ) < ). It is therefore enough to show, for ex-
ample, that the function 1[−1,1] (x) can be approximated in L2 (λκ ) by the Fourier
transform of an elementary function. In other words, that, for any ) > 0, there is
an elementary function l such that 1[−1,1] −  lL2 (λκ ) < ).
To construct l, observe first that
 
 −2κ
|1[−1,1] (x) − l(x)| |x| dx =
2
|x1[−1,1] (x) − x l(x)|2 |x|−2κ−2 dx
⺢ ⺢

and the measure λκ+1 (dx) = |x|−2κ−2 dx is finite around x = ∞. The idea then is
to truncate the range of x1[−1,1] (x), perform a periodic extension and observe that
its truncated Fourier series is of the form x
l(x), where  l is the (continuous) Fourier
transform of an elementary function. We thus construct l as follows.
First, choose k > 1 such that

|x|−2κ−2 1{|x|>k} dx < ) 2 /2.
⺢

Now, let U be the function which equals x1 [−1,1] (x) on [−k, k] and is periodically
extended to x ∈ ⺢. It has the Fourier series ∞n=−∞ un e
iπnx/k which converges to

U everywhere on [−k, k] except at the points x = ±1 where U is discontinuous.

Moreover, the partial sum Um (x) = m n=−m nu e iπnx/k can be expressed as

 k  πy 
1
Um (x) = U (x − y)Dm dy,
k −k k

where
1 1
Dm (y) = sin(m + )y/(2 sin y), y ∈ ⺢,
2 2
Integration questions related to fractional Brownian motion 273

is the well-known Dirichlet kernel. Observe that Um (0) = 0. Let us show that the
following two properties of the partial sums Um hold:
(i) supm supx |Um (x)| ≤ const,
(ii) supm |Um (x)| ≤ const|x|, for small enough x.
These properties will be used in the sequel to apply the dominated convergence
theorem.
We first prove the property (i). Since Um is periodic with period 2k, we may
assume that x ∈ (−k, k). Then
  πy 
1 k
Um (x) = (x − y)1{|x−y|≤1} (y)Dm dy
k −k k
  πy 
1 k
+ (x − y − 2k)1{|x−y−2k|≤1} (y)Dm dy
k −k k
  πy 
1 k
+ (x − y + 2k)1{|x−y+2k|≤1} (y)Dm dy.
k −k k
Let us show that the second term is uniformly bounded in m and x ∈ (−k, k) by a
constant. (The proof for the other two terms is similar.) The second term equals
  π
(x − 2k) π k
1{|x− kv −2k|≤1} (v)Dm (v)dv − 2 v1 kv (v)Dm (v)dv.
π −π π π −π {|x− π −2k|≤1}
Since the kernel Dm is an even function, it is enough to show that the integrals

ξ
ξ
0 Dm (v)dv and 0 vDm (v)dv are uniformly bounded in m and ξ ∈ [0, π ]. This
fact for the first integral is proved in Zygmund [18], (8·2) Lemma, p. 57. It holds for
the second integral as well because the function vDm (v) can be bounded uniformly
in m and v ∈ [0, π] by a constant.
Let us now prove the property (ii). Observe first that, for any 1/2 < |y| <
k, there is a constant c1 such that supm |Dm (πy/k)| ≤ 1/2| sin(πy/2k)| ≤ c1 .

k
Moreover, by the same (8·2) Lemma in Zygmund [18], the integral (1/k) −k
1{|y−x|≤1}∩{|y|≤1} (y)Dm (πy/k)dy can be bounded uniformly in m and x by a con-
stant c2 . Observe also that, for y ∈ (−k, k) and x − y ∈ (−2k + 1, 2k − 1),

U (x − y) − U (−y) = (x − y)1{|y−x|≤1} (y) + y1{|y|≤1}

= x1{|y−x|≤1}∩{|y|≤1} (y) + y1{|y−x|>1}∩{|y|≤1} (y)
+(x − y)1{|y−x|≤1}∩{|y|>1} (y).

Then, for small enough x,

  k  πy  
1 
|Um (x)| = |Um (x) − Um (0)| =  (U (x − y) − U (−y)) Dm dy 

k −k k
  πy  
|x|  k
≤ 1 (y)D dy 
k  −k
{|y−x|≤1}∩{|y|≤1} m
k
   πy 
1 k  
+ 1{|y−x|>1}∩{|y|≤1} (y) Dm  dy
k −k k
274 V. Pipiras, M.S. Taqqu

 k   πy 
1  
+ 1{|y−x|≤1}∩{|y|>1} (y) Dm  dy
k −k k

2c1 k
≤ c2 |x| + 1{|y−1|≤|x|} (y)dy ≤ const|x|,
k −k
which is the property (ii).
By (i) and (ii), the dominated convergence theorem implies that

m→∞
|x1[−1,1] (x) − Um (x)|2 |x|−2κ−2 dx −→ 0.
{|x|≤k}

In particular, there is an integer M such that


|x1[−1,1] (x) − UM (x)|2 |x|−2κ−2 dx < ) 2 /2.
{|x|≤k}

Since U (0) = 0 and U (−x) = −U (x), we have that

 
1 k 2i k
un = U (x)eiπnx/k dx = U (x) sin(π nx/k)dx
k −k k 0
= −ian , an ∈ ⺢, n ≥ 1,

u0 = 0 and un = ian , for n ≤ −1. Hence, UM (x) = M n=1 (−ian )(e
iπnx/k −

e −iπ nx/k ). Since



1[−πn/k,πn/k) (x) = eixu 1[−πn/k,πn/k) (u)du = (−ix −1 )(eiπnx/k − e−iπnx/k ),
⺢

x −1 UM (x) is the Fourier transform of the elementary function l= M n=1 an
1[−πn/k,πn/k) . We thus obtain the required approximation because

1[−1,1] − 
l2L2 (λκ ) = |x1[−1,1] (x) − UM (x)|2 |x|−2κ−2 dx
⺢

≤ |x1[−1,1] (x) − UM (x)|2 |x|−2κ−2 dx
{|x|≤k}

+ |x|−2κ−2 dx < ) 2 .
{|x|>k}

Case 2: If f is an odd function, then f = if and f(−x) = −f(x). By

the same arguments as in the previous case, it is enough to show that, for example,
the function i(1[0,1] (x) − 1[−1,0] (x)) can be approximated by the Fourier trans-
form of an elementary function. Equivalently, for arbitrary small ) > 0, we have
to find an elementary function l such that (1[0,1] − 1[−1,0] ) − i
lL2 (λκ ) < ). The
proof is similar to the previous case and we only outline it. Fix k as in Case 1
and let V be the function which equals x(1[0,1] (x) − 1[−1,0] (x)) = |x|1[−1,1] (x)
on [−k, k]and is periodically extended to x ∈ ⺢. Its truncated Fourier series
Vm (x) = m n=−m vn e
iπnx/k converges to V everywhere on [−k, k] except at the

points x = ±1. It is not enough here to focus on Vm (x) for small x because Vm (0) is
Integration questions related to fractional Brownian motion 275

not zero. Therefore, instead of dealing with Vm (x), we will consider Vm (x)−Vm (0).
The function Vm (x) − Vm (0) also converges to V (x) a.e. dx and one can show that
supm supx |Vm (x) − Vm (0)| ≤ const and supm |Vm (x) − Vm (0)| ≤ const|x|, for

small enough x. Moreover, Vm (x) − Vm (0) = m n=1 bn (e
iπnx/k + e−iπnx/k − 2),

for some bn ∈ ⺢, and hence 

m(Vm (x) − Vm (0))/x = i lm , where lm is the elementary
function given by lm = n=1 bn (1[0,πn/k) − 1[−πn/k,0) ). The conclusion follows
as in Case 1. 


Remarks.
1. In the case 0 < κ < 1/2, a second way to see that the function 1[−1,1] can
be approximated by the Fourier transform of an elementary function is as fol-
lows. Observe first that 1[−1,1] (x)
= f0 (x), where, by taking the inverse Fou-
rier transform, f0 (u) = (1/2π) ⺢ e−iux 1[−1,1] (x)dx = (sin u/π u). Since
f0 ∈/ L1 (⺢), the function f0 is not easy to deal with. Consider instead the
function fa (u) = e−au sin u/π u, for a > 0, which is in L1 (⺢) ∩ L2 (⺢). Since
functions in L1 (⺢) ∩ L2 (⺢) can be approximated by simple and hence ele-
mentary functions (in the two norms simultaneously), by using the estimate
obtained in the proof of Proposition 3.1, fa can be approximated in L2 (λκ ) by
an elementary function. It is then enough to show that 1[−1,1] − fa L2 (λκ ) → 0,
as a → 0. By formula 3.947,(3) in Gradshteyn and Ryzhik [7],
 ∞
2 du 1 2a
fa (x) = e−au cos xu sin u = Arctan 2 + 1{x 2 ≤1−a 2 } ,
π 0 u π x + a2 − 1

and hence fa (x) → 1[−1,1] (x) a.e. dx, as a → 0. Moreover, since |Arctan(y)| ≤
min( π2 , |y|),
 
 
Arctan 2a  ≤ π 1{|x|≤1} (x) + 2a
1{|x|>1} (x)
 x +a −1
2 2  2 x + a2 − 1
2

π 2
≤ 1{|x|≤1} (x) + 2 1{|x|>1} (x),
2 x −1

if 0 < a < 1, and the bound is in L2 (λκ ). The required convergence then
follows from the dominated convergence theorem.
2. In the case 0 < κ < 1/2, a third way of proving the lemma is to use Proposition
3.3 and Theorem 3.2.

κ
Proof of Theorem 3.1. To show
 that  is an inner product space, we check the
least obvious condition. If f, f κ = 0, for f ∈ κ , then f(x) = 0 a.e. dx and
hence f (u) = 0 a.e. du, since f 2L2 (⺢) = (2π )−1 f2L2 (⺢) = 0.
The set of elementary functions E is dense in κ , if f ∈ κ can be approxi-
mated in L2 (λκ ) by the Fourier transform of an elementary function. This follows
from Lemma 5.1.
276 V. Pipiras, M.S. Taqqu

Finally, we show that κ is not complete if κ = 0. Suppose first that 0 < κ <
1/2. The functions

fn (x) = |x|−p 1{1<|x|<n} (x), p > 0,

are in L2 (⺢), fn (x) = fn (−x), and hence they are the Fourier transforms of func-
tion fn ∈ L2 (⺢). It is clear that fn ∈ κ and fn − fm → 0 in κ , as n, m → ∞, if
−2p−2κ +1 < 0 or 1/2−κ < p. Let 1/2−κ < p < 1/2 and suppose that there is
f ∈ κ such that fn → f in κ . It is then necessary that f(x) = |x|−p 1{1<|x|} (x),
x ∈ ⺢, but this f is not in L2 (⺢) which is a contradiction. When −1/2 < κ < 0,
an example of a Cauchy sequence in κ which does not converge is (fn ) ⊂ L2 (⺢)
with
fn (x) = |x|−p 1{1/n<|x|<1} (x), 1/2 < p < 1/2 − κ.



5.2. The proof of Theorem 3.2

We assume throughout this section that 0 < κ < 1/2. The fractional integrals I±α ,
α > 0, and the inner product spaces κ and ||κ are defined in Sections 3.2 and 4.
The fractional Marchaud derivatives Dα± , 0 < α < 1, were introduced in Section
3.3. In the proof of Theorem 3.2 we will use the following results.

Lemma 5.2. Let a < b be real numbers. Then the function

fa,b (u) = (Dκ− 1[a,b) )(u) (5.2)

 
= ((1 − κ)) −1
(b − u)−κ
+ − (a − u)−κ
+ (5.3)

satisfies the equation

(I−κ fa,b )(s) = 1[a,b) (s), ∀s ∈ ⺢. (5.4)

Proof. The functions (5.2) and (5.3) are equal by Lemma 3.1 (where κ is replaced
by −κ). To show that the function (5.3) satisfies the equation (5.4), we have to
verify that

 −κ 
J1 (s) : = (b − u)−κ
+ − (a − u)+ (u − s)+ du
κ−1
⺢
= (κ)(1 − κ)1[a,b) (s) =: J2 (s).

If s ≥ b, then J1 (s) = J2 (s) = 0. If s < b, we use the following identity, valid for
t > 0:
 t  1
(t − v)−κ
+ + v κ−1
dv = t −κ+κ−1+1
(1 − s)−κ s κ−1 ds = B(κ, 1 − κ)
0 0
(κ)(1 − κ)
= = (κ)(1 − κ).
(κ + 1 − κ)
Integration questions related to fractional Brownian motion 277

If a ≤ s < b, then
 b  b−s
J1 (s) = (b−u)−κ κ−1
+ (u−s)+ du = (b−s−v)−κ κ−1
+ v+ dv = (κ)(1−κ),
s 0

which is also J2 (s). In the case when s < a, we show in a similar way that
 b  a
J1 (s) = (b − u)−κ κ−1
+ (u − s)+ du − (a − u)−κ κ−1
+ (u − s)+ du = 0 = J2 (s).
s s




Lemma 5.3. Let a < b be real numbers. Then the function

ψa,b (s) = (Dκ+ 1(a,b] )(s) (5.5)

 −κ 
= ((1 − κ)) −1
(s − a)−κ
+ − (s − b)+ (5.6)

satisfies the equation

(I+κ ψa,b )(u) = 1(a,b] (u), ∀u ∈ ⺢. (5.7)

Proof. The equality of the functions (5.5) and (5.6) follows from Lemma 3.1 by us-
ing the following property of fractional derivatives Dα± : QDα± ψ = Dα∓ Qψ, where
(Qψ)(x) = ψ(−x), x ∈ ⺢ (see (5.61) in Samko et al. [14], p. 111, after correcting
the typo in the reference). Indeed, by applying Lemma 3.1 (where κ is replaced by
−κ, s by (−s) and 1[0,t) by 1[−b,−a) ), we get
 −κ 
((1 − κ))−1 (s − a)−κ
+ − (s − b)+ = (Dκ− 1[−b,−a) )(−s)

= (Dκ− Q1(a,b] )(−s) = (Dκ+ 1(a,b] )(s).

The relation (5.7) follows from (5.4) because for fractional integrals I±α we similarly
have QI±α f = I∓α Qf (see (5.9) in Samko et al. [14], p. 95). 


The following lemma shows that the fractional integration by parts formula
holds for functions in κ , when associated with ψa,b .

Lemma 5.4. If f ∈ κ , then, for any a < b,

 
ψa,b (s)(I−κ f )(s)ds = (I+κ ψa,b )(u)f (u)du (5.8)
⺢ ⺢

or, equivalently,
 
ψa,b (s)(I−κ f )(s)ds = 1(a,b] (u)f (u)du. (5.9)
⺢ ⺢
278 V. Pipiras, M.S. Taqqu

Proof. The equivalence of the relations (5.8) and (5.9) follows from Lemma 5.3.
To establish (5.8), it is enough to show by Fubini’s theorem that, for f ∈ κ ,
 
|f (u)||ψa,b (s)|(u − s)κ−1
+ dsdu < ∞. (5.10)
⺢ ⺢

Since the function ψa,b is zero on (−∞, a), positive on (a, b) and negative on
(b, ∞), we have that
|ψa,b | = ψa,b − 2ψa,b 1(b,∞) .
Hence, by Lemma 5.3,

((κ))−1 |ψa,b (s)|(u − s)κ−1
+ ds
⺢
= (I+κ |ψa,b |)(u) = (I+κ ψa,b )(u) − 2(I+κ ψa,b 1(b,∞) )(u)
= (I+κ ψa,b )(u) − 2(I+κ ψa,b 1(b,∞) )(u)1(b,∞) (u)
= (I+κ ψa,b )(u) − 2(I+κ ψa,b )(u)1(b,∞) (u)
+2(I+κ ψa,b 1(a,b) )(u)1(b,∞) (u)
= 1(a,b] (u) − 21(a,b] (u)1(b,∞) (u)
 b
+2((κ)(1 − κ)) −1
(s − a)−κ κ−1
+ (u − s)+ ds 1(b,∞) (u)
a
 b
≤ 1(a,b] (u) + 2((κ)(1 − κ))−1 (s − a)−κ κ−1
+ ds (u − b)+ .
a

We would complete the verification of (5.10), if we can show that


|f (u)|1(a,b] (u)du < ∞ (5.11)
⺢

and 
|f (u)|(u − b)κ−1
+ du < ∞. (5.12)
⺢
By (3.16), f ∈ κ implies that f is integrable on any finite interval and hence
(5.11) holds. Relation (5.12) may
hold as well, but if it does not, by (3.16), we can
take (bn ) such that bn → b and ⺢ |f (u)|(u − bn )κ−1+ du < ∞. Then the relation
(5.9) holds with bn substituted for b. By letting n tend to infinity, we obtain the
R.H.S. of (5.9) with b because f is integrable on any finite interval. We also obtain
the L.H.S. of (5.9) with b because I−κ f ∈ L2 (⺢) and ψa,b − ψa,bn L2 (⺢) → 0, as
n → ∞. 


Lemma 5.5. The set ||κ is dense in the inner product space κ .

Proof. If f ∈ κ , then, by (3.18), the function g = I−κ f belongs to L2 (⺢). Hence,

there is a sequence (gn ) of elementary functions such that g − gn L2 (⺢) → 0. By
Lemma 5.2, the elementary functions gn can be expressed as

gn = I−κ fn ,
Integration questions related to fractional Brownian motion 279

where fn is a linear combination of functions fa,b , for some a < b, and the above
equality holds almost everywhere. The functions fa,b are given in (5.3). Since, by
(3.19),
(κ + 1)
f − fn κ = g − gn L2 (⺢) → 0,
c1 (κ)
||κ is dense in κ , if fn ∈ ||κ . Since fn is a linear combination of functions
fa,b , it is in ||κ , if fa,b ∈ ||κ for every a < b, or, by (4.5), if
  2
κ−1
|fa,b (u)|(u − s)+ du ds < ∞. (5.13)
⺢ ⺢

To establish (5.13), we follow the idea of the proof of Lemma 5.4. Since |fa,b | =
fa,b − 2fa,b 1(−∞,a) (see (5.3)), we obtain by Lemma 5.2 that

((κ)) −1
|fa,b (u)|(u − s)κ−1
+ du
⺢
= I−κ (|fa,b |)(s) = (I−κ fa,b )(s) − 2(I−κ fa,b 1(−∞,a) )(s)1(−∞,a) (s)
= (I−κ fa,b )(s) − 2(I−κ fa,b )(s)1(−∞,a) (s) + 2(I−κ fa,b 1(a,b) )(s)1(−∞,a) (s)
= 1[a,b) (s) + 21[a,b) (s)1(−∞,a) (s)
 b
+ 2((κ)(1 − κ))−1 (b − u)−κ κ−1
+ (u − s)+ du 1(−∞,a] (s)
a
 b
≤ 1[a,b) (s) + 2((κ)(1 − κ))−1 (b − u)−κ κ−1
+ du (a − s)+ 1(−∞,a−1] (s)
a
 b
+ 2((κ)(1 − κ))−1 (b − u)−κ κ−1
+ (u − a)+ du 1(a−1,a] (s). (5.14)
a

Since the upper bound (5.14) is in L2 (⺢), the function fa,b satisfies (5.13) and
hence ||κ is dense in κ . 


The following two lemmas will be used to show that the inner product space
κ is not complete.
Lemma 5.6. The inner product space κ is complete if and only if, for every
φ ∈ L2 (⺢), there is a function fφ ∈ κ such that

φ = I−κ fφ . (5.15)

Proof. Suppose that κ is complete and let φ ∈ L2 (⺢). There is a sequence (φn )
of elementary functions such that φn → φ in L2 (⺢). By Lemma 5.2, the elemen-
tary functions (φn ) can be expressed as φn = I−κ fφn , for some fφn ∈ κ . The
sequence (fφn ) is then Cauchy in κ by (3.17). Since κ is complete by assump-
tion, there is f ∈ κ such that fφn → f in κ . In other words, the convergence
φn = I−κ fφn → I−κ f holds in L2 (⺢). Since φn → φ in L2 (⺢) as well, we obtain
(5.15) with f = fφ .
Conversely, suppose that (5.15) holds and let (fn ) be a Cauchy sequence in κ .
Then the sequence (φn ), given by φn = I−κ fn , is Cauchy in L2 (⺢). Since L2 (⺢) is
280 V. Pipiras, M.S. Taqqu

complete, there is φ ∈ L2 (⺢) such that φn → φ in L2 (⺢). By assumption, there is

fφ ∈ κ such that (5.15) holds. Since φn → φ in L2 (⺢) implies fn → fφ in κ ,
κ is complete. 

Lemma 5.7. Let 0 < κ < 1/2. There is φ ∈ L1 (⺢) ∩ L2 (⺢) such that the function

Uφ (a) = u−κ
+ φ(u + a)du, a ∈ ⺢,
⺢
is not differentiable on a set of positive Lebesgue measure.
Proof. Let b > 1 and 0 < p < κ. Consider the complex-valued function
∞ 

∗ −pn ibn u
φ (u) = c0 b e 1[0,1] (u),
n=1

∞ −κ −1  −1
where c0 = 0 eiv v+ dv =  (1 − κ) eiπ(1−κ)/2 . Since φ ∗ is bounded
with compact support, it belongs to L1 (⺢) ∩ L2 (⺢). Let us show that Uφ ∗ (a) is not
differentiable on [0, 1/2] (that is, its real and imaginary parts are not differentiable).
Since, for 0 < a < 1, we have
 1−a ∞ 
 n u ibn a
−κ −pn ib
Uφ ∗ (a) = c0 u+ b e e du
0 n=1
∞
 
 1−a n na
= c0 b −pn
eib u u−κ
+ du eib
n=1 0

∞  ∞ 
n na
= c0 b −pn
eib u u−κ
+ du eib
n=1 0

∞  ∞ 
n n
−c0 b−pn eib u u−κ
+ du eib a
n=1 1−a

=: y1 (a) + y2 (a),
it is enough to show that the function y1 is not differentiable on [0, 1] and that the
function y2 is differentiable on [0, 1/2]. We deal with y1 first. By making a change
of variables bn u = v, we obtain that
∞
 n
y1 (a) = b−(p−κ+1)n eib a .
n=1

Since b−(p−κ+1) b = bκ−p

> 1, the function y1 is a particular case of the well-
known Weierstrass function whose real and imaginary parts are nowhere differen-
tiable functions (see, for example, Hardy [10] or Zygmund [18] for more details on
the nowhere differentiable Weierstrass function). We now turn to the function y2 ,
which is well-defined because, by Lemma 5.8 below,
∞ 
   ∞

 −pn ∞ ibn u −κ 
ibn a 
b e u+ du e −κ
b−(p+1)n < ∞.
  ≤ 4(1 − a)
n=1 1−a n=1
Integration questions related to fractional Brownian motion 281

It is differentiable on [0, 1/2], since, by using Lemma 5.8,

∞   ∞ 
d ibn u −κ

ibn a 
 b −pn
e u du e
 da + 
n=1 1−a

∞    ∞   
 −pn ibn −κ 
 n −pn ∞ ibn u −κ 
ibn a 
≤ e (1 − a)  + 
b ib b e u+ du e 
n=1 n=1 1−a
∞
 ∞

≤ (1 − a)−κ b−pn + 4 bn b−pn b−n (1 − a)−κ
n=1 n=1
∞

≤C b−pn < ∞,
n=1

where the constant C does not depend on a ∈ [0, 1/2]. Finally, for a function φ of
the lemma, we can take φ = φ ∗ or φ = φ ∗ . 


Lemma 5.8. Let a < b be two real numbers. Suppose that f is a positive non-in-
creasing function on the interval [a, b]. Then, for any x ∈ ⺢, x = 0,
 b 
  4f (a)
 e ixu
f (u)du ≤ . (5.16)
  x
a

Proof. Relation (5.16) follows from the inequalities

 b   
   y 
 sin(xu)f (u)du ≤ f (a) sup 
 sin(xu)du

a a<y<b a

and    
 b   y 
 cos(xu)f (u)du ≤ f (a) sup 
 cos(xu)du ,

a a<y<b a

valid for any x ∈ ⺢. 



Proof of Theorem
 3.2. To show that κ is an inner product space, we verify
that f, f κ = 0 implies f (u) = 0 a.e. du.
If f, f κ
= 0, then, by (3.17),
(I−κ f )(s) = 0 a.e. ds, and hence, by (5.9), ⺢ f (u)1(a,b] (u)du

= 0, for any real
numbers a < b. By an approximation argument, we get that ⺢ f (u)1{f (u)>0}∩{|u|≤n}
(u)du = 0, for any n ∈ ⺞. Consequently, f (u) ≤ 0 a.e. du and, by symmetry,
f (u) = 0 a.e. du.
Let us prove that the set of elementary functions E is dense in κ . By Lemma
5.5, it is enough to show that E is dense in ||κ . If f ∈ ||κ , then, by (4.6),
 
f 2κ = κ(2κ + 1) f (u)f (v)|u − v|2κ−1 du dv.
⺢ ⺢

Now, choose a sequence of simple functions fn which converges  to f a.e.

 and such
that |fn | ≤ |f |. Then, by the dominated convergence theorem, f −fn κ → 0, as
n → ∞. On the other hand, for fixed n, there is a sequence of elementary functions
282 V. Pipiras, M.S. Taqqu

gn,m such that gn,m → fn a.e. and supm |gn,m | ≤cn 1[−kn ,kn ) . The dominated con-
vergence theorem implies again that fn − gn,m κ → 0, as m → ∞. It follows
that the set E is dense in ||κ .
We will show that the inner product space κ is not complete by contradiction.
Suppose that κ is complete. Then, by Lemma 5.6, to every φ ∈ L2 (⺢) there
corresponds a function fφ ∈ κ such that

φ = I−κ fφ .

It follows from (5.9) that

 
φ(s)ψ0,a (s)ds = fφ (u)1(0,a] (u)du, (5.17)
⺢ ⺢

for any φ ∈ L2 (⺢) and a ∈ ⺢. Since the function on the right-hand side of (5.17)
is differentiable a.e. da, the function

 −κ 
φ(s) s+ − (s − a)−κ
+ ds,
⺢

which is the left-hand side of (5.17) up to a multiplicative constant, is differentiable

a.e. da as well. If, in addition, φ ∈ L1 (⺢), then

  1/2
1
−κ −2κ
φ(s)s+ ds ≤ φL2 (⺢) s+ ds + φL1 (⺢) < ∞
⺢ 0

and the function

 
Uφ (a) = φ(u)(u − a)−κ
+ du = u−κ
+ φ(u + a)du
⺢ ⺢

is also differentiable a.e. da. In Lemma 5.7 above we provide an example of a

function φ ∈ L1 (⺢) ∩ L2 (⺢) for which Uφ is not differentiable on a set of posi-
tive Lebesgue measure. We thus obtain a contradiction which shows that the inner
product space κ is not complete. 


5.3. The proof of Theorem 3.3

In the proof of Theorem 3.3 we will use the following result which is a consequence
of Lemma 5.1.

Lemma 5.9. Let −1/2 < κ < 1/2 and suppose that  g is the Fourier transform
of a function g ∈ L2 (⺢). Then there is a sequence of elementary functions ln such
that 
 2 n→∞
 ln (x)|x|−κ  dx −→ 0.
g (x) − 
⺢
Integration questions related to fractional Brownian motion 283

Proof.

Suppose first that −1/2 < κ < 0. If  g) (x) = 
g (x)1{|x|>)} (x), then
⺢ |
g (x) − 
g ) (x)| 2 dx → 0, as ) → 0. Since

 

g) (x) = g (x)|x|κ 1{|x|>)} (x) |x|−κ = f) (x)|x|−κ ,

where f) is the Fourier transform of a function f) ∈ L2 (⺢), the result follows from
Lemma 5.1. The proof in the case 0 ≤ κ < 1/2 is similar. 


Proof of Theorem 3.3. Recall that −1/2 < κ < 0 in Theorem 3.3. It is obvious that
the space κ , defined by (3.30), is an inner product space with the inner product
(3.31). Let us show that E is dense in κ . If f ∈ κ , then

f = I−−κ φf , (5.18)

for some function φf ∈ L2 (⺢). When f is an elementary function, Lemma 5.2

implies that φf equals to a linear combination of φa,b , a < b, where the functions

 −1  
φa,b (s) = (1 − κ) (b − s)κ+ − (a − s)κ+ (5.19)

satisfy the equation

I−−κ φa,b = 1[a,b) . (5.20)

Therefore, by (3.32), the set E of elementary functions is dense in κ , if any

φ ∈ L2 (⺢) can be approximated in L2 (⺢) by a linear combination of functions
φa,b from (5.19). To prove this fact, let us work with Fourier transforms. We have
 ∈ L2 (⺢) can be approximated in L2 (⺢) by a linear combination
to show that any φ

of the functions φa,b . Since, by Lemma 3.1,

φa,b = D−κ
− 1[a,b)

and, clearly, 1[a,b) ∈ κ , Proposition 3.3, (3), implies that

a,b (x) = C κ (x)

φ 1[a,b] (x)|x|−κ .

Then, applying Lemma 5.9 with  (x), the required approximation

g (x) = C κ (x)φ
 by a linear combination of φ
of φ a,b follows.
To show that κ is complete, let (fn ) be a Cauchy sequence in κ and use the
facts that, by (3.32), the corresponding sequence (φfn ) is Cauchy in L2 (⺢) and that
L2 (⺢) is complete. Finally, the space κ is isometric to Sp(B κ ) by Proposition
2.1. 

284 V. Pipiras, M.S. Taqqu

6. Connections to the RKHS of fractional Brownian motion

In this section we look at our results from the standpoint of the theory of reproduc-
ing kernel Hilbert spaces (RKHS). The RKHS of fBm also characterizes the Hilbert
space Sp(B κ ): it is isometric to the space Sp(B κ ) itself, not only for −1/2 < κ < 0
but also for 0 < κ < 1/2. The RKHS can be viewed as representing “the space
of integrals”, since the “simple” integral B κ (u2 ) − B κ (u1 ) belonging to Sp(B κ ) is
represented by the element of the RKHS,  κ (·, u2 ) −  κ (·, u1 ), a function which
depends on κ ( κ is the covariance function of fBm). From this perspective, the
spaces which associate to B κ (u2 ) − B κ (u1 ) the indicator function 1[u1 ,u2 ) can be
regarded as “spaces of integrands”. It is therefore interesting to see how the RKHS
characterization compares with the inner product spaces introduced earlier.
We start by recalling the definition of a RKHS. Suppose that X = {X(t)}t∈⺢
is a second order mean zero stochastic process with the covariance function (s, t),
s,
tk ∈ ⺢, and let Sp(X) be the closure in L ( ) of all linear combinations
2

i=1 ai X(ti ), ai , ti ∈ ⺢. Then with the Hilbert space Sp(X) one can always
associate an isometric Hilbert space ⺘() of deterministic functions. The space
⺘() is called the Reproducing Kernel Hilbert Space (RKHS) and is character-
ized by the following two properties: (1) (·, t) ∈ ⺘(), for all t ∈ ⺢, and (2)
(g, (·, t))⺘() = g(t), for all t ∈ ⺢ and g ∈ ⺘(). It consists of all functions
k
of the form f (s) = i=1 ai (s, ti ), s ∈ ⺢, and their limits under the norm
k
f  =2
i,j =1 ai aj (ti , tj ). (See Grenander [8], p. 93, for a quick reference
to RKHS’s, or Weinert [16], for a thorough introduction.) The isometry map J
between the Hilbert spaces Sp(X) and ⺘() satisfies

k
 k

J : ai (·, ti ) −→ ai X(ti )
i=1 i=1

and (J(g), J(h))L2 ( ) = (g, h)⺘() . If the covariance function  can be expressed
as

(s, t) = fs (u)ft (u)dν(u), (6.1)
U

where (U, U, ν) is a measure space and {ft : t ∈ ⺢} ⊂ L2 (ν), then the RKHS
⺘() is characterized by

⺘() = g : g(t) = g ∗ (u)ft (u)dν(u), for some g ∗ ∈ span{ft , t ∈ ⺢} ,
U
(6.2)

(g, h)⺘() = g ∗ (u)h∗ (u)dν(u), (6.3)
U

where span{ft , t ∈ ⺢} is the closure in L2 (ν) of all linear combinations of ft , t ∈ ⺢

(see Grenander [8]).
Integration questions related to fractional Brownian motion 285

6.1. The case −1/2 < κ < 0

We characterize, in the following proposition, the RKHS of fBm with index −1/2 <
κ < 0.
Proposition 6.1. Let −1/2 < κ < 0 and B κ be a standard fBm. Then the RKHS
⺘( κ ) of B κ is

(κ + 1)2
κ
⺘( ) = g : g(t) = g ∗ (u)(D−κ
− 1[0,t) )(u)du,
c1 (κ)2 ⺢

for some g ∗ ∈ L2 (⺢) (6.4)

with the inner product


(κ + 1)2
(g, h)⺘( κ ) = g ∗ (u)h∗ (u)du. (6.5)
c1 (κ)2 ⺢

Proof. By taking f and g in (3.28) to be indicator functions, one gets


(κ + 1)2
κ
 (s, t) = (D−κ −κ
− 1[0,s) )(u)(D− 1[0,t) )(u)du.
c1 (κ)2 ⺢

Thus, the relation (6.1) holds with U = ⺢, the Lebesgue measure dν(u) (up to a
constant) and
ft (u) = (D−κ
− 1[0,t) )(u).
Since, by Proposition 3.3, (3),

ft (x) = C κ (x)

1[0,t) (x)|x|−κ

and |C κ (x)| = 1, Lemma 5.9 implies that {ft , t ∈ ⺢} spans L2 (⺢). Hence, by (6.2)
and (6.3), the RKHS ⺘( κ ) of fBm B κ with −1/2 < κ < 0 is characterized by
(6.4) and (6.5). 


Let us denote the isometry map between the Hilbert spaces ⺘( κ ) and Sp(B κ )
by Jκ . Then to every X ∈ Sp(B κ ) there corresponds a unique gX ∈ ⺘( κ ) such
that X = Jκ (gX ), that is, every element of Sp(B κ ) can be characterized by a
deterministic function from ⺘( κ ).
Example. The random variable B κ (u2 ) − B κ (u1 ) in Sp(B κ ) is represented by the
function g(·) =  κ (·, u2 ) −  κ (·, u1 ) in ⺘( κ ). The corresponding function g ∗ in
(6.4) is g ∗ = D−κ
− 1[u1 ,u2 ) .
To compare the spaces of functions ⺘( κ ) and κ when −1/2 < κ < 0, ob-
serve that there is a natural map from the space of elementary functions E ⊂ κ
into ⺘( κ ) defined by
n
 n

κ
 
γ : fk 1[uk ,uk+1 ) −→ fk  κ (·, uk+1 ) −  κ (·, uk ) . (6.6)
k=1 k=1
286 V. Pipiras, M.S. Taqqu

To see the connection between our integral Iκ and the RKHS integral Jκ , note
that, for all f, g ∈ E,
Iκ (f ) = Jκ (γ κ (f )), (6.7)
where 
(κ + 1)2
(γ κ (f ))(t) = (D−κ −κ
− f )(u)(D− 1[0,t) )(u)du, (6.8)
c1 (κ)2 ⺢
so that
(f, g)κ = (Iκ (f ), Iκ (g))L2 ( ) = (Jκ (γ κ (f )), Jκ (γ κ (g))L2 ( )
= (γ κ (f ), γ κ (g))⺘( κ ) . (6.9)
To see that the definition (6.6) of γ κ can be extended to an isometry between κ
and ⺘( κ ), observe that, for every g ∗ ∈ L2 (⺢), there is a function f such that
g ∗ = D−κ− f. Indeed, by Theorem 6.1 in Samko et al. [14], p. 125, we can take
f = I−−κ g ∗ . This means that the map γ κ in (6.6) can be extended by the relation
(6.8) involving elementary functions to the map from the space κ onto the space
⺘( κ ). Now E is dense in the space κ (Theorem 3.3) and γ κ (E) is dense in the
space ⺘( κ ) (by construction). This means that the relations (6.7) and (6.9) extend
to f, g ∈ κ . Hence, when −1/2 < κ < 0, the Hilbert spaces κ , ⺘( κ ) and
Sp(B κ ) are isometric. This is illustrated in the following diagram:
γκ
κ ✲ ⺘( κ )
❅
I κ ❅ ✠ Jκ
❘
❅
Sp(B κ )
where the maps Iκ , Jκ and γ κ are all isometries between the corresponding
spaces.
Remarks.
1. The isometry between ⺘( κ ) and κ also follows from the facts that ⺘( κ ) is
isometric to Sp(B κ ) and κ is isometric to Sp(B κ ) (Theorem 3.3). The difficult
step in the proof of either Theorem 3.3 or Proposition 6.1 (which characterizes
⺘( κ )) is Lemma 5.9 (or Lemma 5.1 on which Lemma 5.9 is based).
2. The argument proceeding these remarks provides an explicit identification of
the map γ κ .
3. Since the space of functions κ with −1/2 < κ < 0 is a proper subspace of
κ , it is isometric to a subspace of ⺘( κ ). We can identify this subspace as a
subset of functions g ∈ ⺘( κ ) such that the function g∗ (x)|x|κ is in L2 (⺢).
(The functions g and g ∗ are related by (6.4).) Indeed, if the latter condition
holds, there is a function f ∗ ∈ L2 (⺢) such that
g∗ (x)|x|κ = C κ (x)f∗ (x).
In particular, the function f∗ (x)|x|κ is in L2 (⺢) and, by Proposition 3.3, (3),
g ∗ = D−κ ∗
− f ,
Integration questions related to fractional Brownian motion 287

since both sides of the equation have the same Fourier transforms. This shows
that g = γ κ (f ∗ ) with f ∗ ∈ κ . On the other hand, one can similarly show
that, if f ∗ ∈ κ , the function g ∗ which corresponds to g = γ κ (f ∗ ) satisfies
g∗ (x)|x|κ ∈ L2 (⺢).

6.2. The case 0 < κ < 1/2

Proceeding as in the case −1/2 < κ < 0 considered above, one can show that the
following proposition holds.

Proposition 6.2. Let 0 < κ < 1/2 and B κ be a standard fBm. Then the RKHS
⺘( κ ) of B κ is
 
κ (κ + 1)2
⺘( ) = g : g(t) = g ∗ (u)(I−κ 1[0,t) )(u)du,
c1 (κ)2 ⺢

for some g ∗ ∈ L2 (⺢) (6.10)

with the inner product (6.5).

The result (6.10) coincides with the characterization of Barton and Poor [2],
who studied the case 0 < κ < 1/2 and expressed ⺘( κ ) as
 t s
1
⺘( κ ) = g : g(t) = (s − u)κ−1 g(u)duds, for some g ∈ L2 (⺢) .
(κ) 0 −∞
(6.11)
Indeed, by using (3.21) and the fractional integration by parts formula (3.20), we
can write the function g in (6.11) as
 
g(t) = (I+κ g)(s)1[0,t) (s)ds = g(u)(I−κ 1[0,t) )(u)du
⺢ ⺢

and, hence, recover the representation in (6.10).

As in the case −1/2 < κ < 0, let us denote the isometry map between the Hil-
bert spaces ⺘( κ ) and Sp(B κ ) by Jκ and let also γ κ be the map defined by (6.6).
Then the relations (6.7) and (6.9) hold also for f, g ∈ E ⊂ κ and 0 < κ < 1/2,
while (6.8) becomes

(κ + 1)2
(γ κ (f ))(t) = (I−κ f )(u)(I−κ 1[0,t) )(u)du. (6.12)
c1 (κ)2 ⺢

Using the relation (6.12), we can extend the map γ κ to the map from the space κ
into the space ⺘( κ ). Since the sets of functions E and γ (E) are dense in the spaces
κ and ⺘( κ ), respectively, the extended map γ κ satisfies (6.7) and (6.9) for any
f, g ∈ κ and 0 < κ < 1/2. However, since the inner product space κ is not
complete (Theorem 3.2), κ is isometric only to a subspace of the Hilbert space
⺘( κ ). The case 0 < κ < 1/2 is therefore different from the case −1/2 < κ < 0
288 V. Pipiras, M.S. Taqqu

considered earlier because when −1/2 < κ < 0, there is an isometry between
⺘( κ ) and κ .
A slightly different perspective on the relationship between the spaces ⺘( κ )
and κ when 0 < κ < 1/2 is as follows. Let (γ κ )−1 be the isometry map defined
from a linear subspace of ⺘( κ ) into the space of elementary functions E as the
inverse map of (6.6). We can then ask whether we can extend the map (γ κ )−1 to
the isometry between ⺘( κ ) and some space of deterministic functions. If such an
isometry exists, then the class of functions (γ κ )−1 (⺘( κ )) is isometric to the space
Sp(B κ ) itself (since ⺘( κ ) is isometric to Sp(B κ )) and is also

a class of integrands
for which we can extend the natural definition Iκ (f ) = ⺢ f (u)dB κ (u), f ∈ E,
of the integral with respect to fBm. To obtain this isometry, one would naturally try
to do the following: in view of (6.10) and (6.12), express the function g ∗ in (6.10)
as
g ∗ = I−κ f (6.13)
for some function f , and then define (γ κ )−1 (g) = f . However, as we have shown
in Section 5.2 there are functions g ∗ ∈ L2 (⺢) which do not admit the representa-
tion (6.13). This is why κ = {f : g ∗ = I−κ f ∈ L2 (⺢)} is isometric only to a
subspace of ⺘( κ ). Since by the definition of the RKHS, any function in ⺘( κ )
can be approximated by functions in γ κ (E), ⺘( κ ) is isometric to the space C of
all Cauchy sequences in E. But is C a set of functions? We do not know the answer.
Therefore we do not know whether the map γ κ : E → ⺘( κ ) can be extended to
an isometry between some space of functions C and the space ⺘( κ ).
Remark. When 0 < κ < 1/2, we introduced in Sections 3.1 and 4 two other
classes of integrands, namely, the spaces κ and ||κ which are strict subsets of
the class κ . Their images under the map γ κ are
   
γ κ κ = g ∈ ⺘( κ ) : g∗ (x)|x|κ ∈ L2 (⺢) ,
  
γ κ ||κ = g ∈ ⺘( κ ) : g ∗ = I−κ f, for some function f ∈ ||κ ,
where g ∗ is as in (6.10). These relations follow easily from the relation (6.12)
applied to f ∈ κ and the fact that κ and ||κ are subsets of κ .

7. Conclusion

In this work we have introduced a number of different inner product spaces, namely

κ = f : f ∈ L2 (⺢), |f(x)|2 |x|−2κ dx < ∞ , for − 1/2 < κ < 1/2,
⺢

 κ 2
κ = f : (I− f )(s) ds < ∞ , for 0 < κ < 1/2,
⺢
 
 = f : ∃ φf ∈ L2 (⺢) such that f = I−−κ φf , for − 1/2 < κ < 0,
κ

 
κ
|| = f : |f (u)||f (v)||u − v|2κ−1 dudv < ∞ , for 0 < κ < 1/2,
⺢ ⺢
Integration questions related to fractional Brownian motion 289

with the inner products given by


1
(f, g)κ = f(x) g (x)|x|−2κ dx,
c2 (κ)2 ⺢

  (κ + 1)2
f, g κ
= (I−κ f )(s)(I−κ g)(s) ds,
c1 (κ)2 ⺢

  (κ + 1)2
f, g κ
= φf (s)φg (s)ds,
c1 (κ)2
 ⺢
 
f, g ||κ
= κ(2κ + 1) f (u)g(v)|u − v|2κ−1 dudv,
⺢ ⺢

respectively. (The constants c1 (κ) and c2 (κ) are given in (3.8) and (3.2).)1 The
definition of κ uses the integral fractional operator I−α , α > 0, defined in (3.10).
We have shown that

κ ⊂ κ , −1/2 < κ < 1/2, κ = 0, and ||κ ⊂ κ , 0 < κ < 1/2,

where all the inclusions are strict. In fact, for 0 < κ < 1/2, one has (see (4.7)),

L1 (⺢) ∩ L2 (⺢) ⊂ L2/2κ+1 (⺢) ⊂ ||κ ⊂ κ .

Each of the inner product spaces κ , κ and ||κ is isometric to a linear subspace
of the Hilbert space Sp(B κ ), defined by (1.5), with the usual L2 ( ) inner product.

1
For convenience to the reader, we also display the spaces and the inner products in the
commonly used parametrization H = κ + 1/2:

H = f : f ∈ L2 (⺢), |f(x)|2 |x|−2H +1 dx < ∞ , for 0 < H < 1,
⺢
 ! "2
H− 1
H = f : (I− 2 f )(s) ds < ∞ , for 1/2 < H < 1,
⺢
1 −H
H = f : ∃ φf ∈ L2 (⺢) such that f = I−2 φf , for 0 < H < 1/2,
 
||H = f : |f (u)||f (v)||u − v|2H −2 dudv < ∞ , for 1/2 < H < 1,
⺢ ⺢

and, respectively,

1
(f, g)H = f(x)
g (x)|x|−2H +1 dx,
C2 (H )2 ⺢
  
(H + 1/2)2 H− 1 H− 1
f, g = (I− 2 f )(s)(I− 2 g)(s) ds,
H C1 (H ) 2
⺢
  
(H + 1/2)2
f, g = φf (s)φg (s)ds,
H C1 (H )2
   ⺢
f, g = H (2H − 1) f (u)g(v)|u − v|2H −2 dudv,
||H ⺢ ⺢

where C1 (H ) = c1 (H − 1/2) and C2 (H ) = c2 (H − 1/2).

290 V. Pipiras, M.S. Taqqu

The isometry map Iκ is an extension of the natural definition of the integral with
respect to fBm
 n
 
f (u)dB κ (u) = fk B κ (uk+1 ) − B κ (uk ) , (7.1)
⺢ k=1

where f is an elementary (step) function f (u) = nk=1 fk 1[uk ,uk+1 ) (u), u ∈ ⺢,
and, therefore, it is still denoted by

κ
I (f ) = f (u)dB κ (u) (7.2)
⺢

and called the integral on the real line of a function f with respect to fBm B κ .
The inner product spaces κ , κ and ||κ can thus also be viewed as classes of
integrands. The completeness of an inner product space (a class of integrands) is a
desirable property because the space is then (and only then) isometric to the space
Sp(B κ ) itself and hence every element of the space Sp(B κ ) can, in this case, be
expressed as an integral of a function with respect to fBm B κ . We have shown that
the inner product space κ , when either −1/2 < κ < 0 or 0 < κ < 1/2, and
the spaces ||κ , κ , when 0 < κ < 1/2, are not complete, whereas the space
κ , when −1/2 < κ < 0, is complete. We do not know whether there is an inner
product space of functions isometric to the space Sp(B κ ) itself when 0 < κ < 1/2,
where the isometry extends the natural definition (7.1) of the integral with respect
to fBm for elementary functions.
We compared the classes of integrands κ , κ and ||κ to the reproducing ker-
nel Hilbert space ⺘( κ ) of fBm, which can be regarded as representing the space of
integrals. We obtained an explicit characterization of the isometries between these
classes of integrands and either ⺘( κ ) or subspaces of ⺘( κ ).
The space κ is referred to as the class of integrands in the “spectral domain”
because it is constructed by using the spectral representation (3.1) of fBm B κ and
involves a condition on its Fourier transform f. On the other hand, the space κ is
called the class of integrands in the “time domain” because
it is constructed by using
the time domain representation (3.7) of fBm. The integral ⺢ f (u)dB κ

(u), defined
in the “spectral domain”, is equal almost surely to the integral ⺢ f (u)dB κ (u),
defined in the “time domain”. The space ||κ is a practical, alternative class of
integrands in the “time domain” when 0 < κ < 1/2.

References

1. Alòs, E., Mazet, O., Nualart, D.: Stochastic calculus with respect to fractional Brownian
motion with Hurst parameter less than 21 . Stochastic Processes and their Applications,
86(1), 121–139 (2000)
2. Barton, R.J., Poor, H.V.: Signal detection in fractional Gaussian noise. IEEE Transac-
tions on Information Theory, 34(5), 943–955 (1988)
3. Carmona, P., Coutin, L., Montseny, G.: Stochastic integration with respect to fractional
Brownian motion. Preprint (1999)
4. Decreusefond, L., Üstünel, A.S.: Stochastic analysis of the fractional Brownian motion.
Potential Analysis, 10, 177–214 (1999)
Integration questions related to fractional Brownian motion 291

5. Dudley, R.M., Norvaiša, R.: An introduction to p-variation and Young integrals. Con-
centrated advanced course. Maphysto, Centre for Mathematical Physics and Stochastics,
University of Aarhus, Denmark (1999)
6. Duncan, T. E., Hu, Y., Pasik-Duncan, B.: Stochastic calculus for fractional Brownian
motion I: theory. SIAM Journal on Control and Optimization, 38(2), 582–612 (2000)
7. Gradshteyn, I.S., Ryzhik, I.M.: Table of Integrals, Series, and Products. Academic Press,
New York (1980)
8. Grenander, U.: Abstract Inference. John Wiley & Sons, New York (1981)
9. Gripenberg, G., Norros, I.: On the prediction of fractional Brownian motion. Journal of
Applied Probability, 33, 400–410 (1996)
10. Hardy, G.H.: Weierstrass’s non-differentiable function. Transactions of the American
Mathematical Society, 17, 322–323 (1916)
11. Huang, S.T., Cambanis, S.: Stochastic and multiple Wiener integrals for Gaussian pro-
cesses. The Annals of Proability, 6, 585–614 (1978)
12. Kleptsyna, M.L., LeBreton, A., Roubaud, M.C.: Rudiments of stochastic fractional
calculus and statistical applications. Preprint (1999)
13. Leland, W.E., Taqqu, M.S., Willinger, W. Wilson, D.V.: On the self-similar nature of
Ethernet traffic (Extended version). IEEE/ACM Transactions on Networking, 2, 1–15
(1994)
14. Samko, S.G., Kilbas, A.A., Marichev, O.I.: Fractional Integrals and Derivatives. Gordon
and Breach Science Publishers (1993)
15. Samorodnitsky, G., Taqqu, M.S.: Stable Non-Gaussian Processes: Stochastic Models
with Infinite Variance. Chapman and Hall, New York, London (1994)
16. Weinert, H.L.: Reproducing Kernel Hilbert Spaces: Applications in Statistical Signal
Processing. Hutchinson Ross, Stroudsburg, PA (1982)
17. Zähle, M.: Integration with respect to fractal functions and stochastic calculus. I. Prob-
ability Theory and Related Fields, 111, 333-374 (1998)
18. Zygmund, A.: Trigonometric Series. Cambridge University Press, Cambridge (1979).
Volumes I, II