Applied Mathematical Sciences, Vol. 3, 2009, no.

38, 1885 - 1901

Further Remarks on Mixed

Fractional Brownian Motion
Christoph Thäle1

University of Fribourg
Department of Mathematics
Chemin du Musée 23
CH-1700 Fribourg, Switzerland
[Link]@[Link]

Abstract

We study linear combinations of independent fractional Brownian

motions and generalize several recent results from [10] and [17]. As
a first new result we calculate explicitly the Hausdorff dimension of
the sample paths of such processes. Moreover we compare different
notions of fractional differentiability and calculate as a second new result
explicitly the Cesáro fractional derivative of the critical oder. Moreover
we consider the more interesting absolute Cesáro fractional derivative
and calculate its value explicitly.

Mathematics Subject Classification: 60G17; 60J65; G0G15

Keywords: Fractional Brownian Motion; Fractional Calculus; Hausdorff

Dimension; Self-similar Process

1 Introduction
The classical fractional Brownian BtH motion with Hurst parameter H ∈ [0, 1]
is a mean zero Gaussian process with covariance function

1
cov(BsH , BtH ) = (s2H + t2H − |s − t|2H ), s, t ≥ 0.
2
This generalization of the Wiener process (take H = 12 to obtain this process)
was introduced by Mandelbrot and Van Ness [9]. The study of this process is
1
This work was supported by the Schweizerischer Nationalfonds grant SNF PP002-
114715/1.
1886 Christoph Thäle

originally motivated by problems in finance, telecommunication and engineer-

ing applications. The fractional Brownian motion BtH can also be written as
1/2
the following stochastic integral with respect to the Wiener process Bt = Bt ,
see [3]:  t
BtH = KH (t, u)dBt
0
and KH obeys the differential equation
    12  
∂KH 2HΓ 32 − H 1 u 1
2
−H 3
= 1  −H (t − u)H− 2 .
∂t Γ 2 + H Γ(2 − 2H) 2 t

We like to point out that there exists also an explicit expression of the kernel
KH (t, u) in terms of hypergeometric functions, see [11]. A characteristic prop-
erty of this process is its self-similarity. This means that for any c > 0, the
time-space rescaled process c−H Bct H
has the same distribution as the original
H
process Bt .
Recently several authors started to study processes of the following type:

Bt + aBtH and a1 BtH1 + a2 BtH2 ,

see [10] and [17], where a, a1 , a2 ∈ R and H, H1 , H2 ∈ [0, 1]. The first process
was introduced in [2] to present a stochastic model of the discounted stock
price in some arbitrage-free and complete financial markets. Because of the
deficiencies of this model, the study of the second one was motivated. We will
generalize these approaches further and study linear combinations of N ∈ N
independent fractional Brownian motion, i.e. the process
N
Zt := ak BtHk ,
k=1

where a1 , . . . , aN are real coefficients. Another motivation for our approach are
recent developments in modeling internet traffic using self-similar processes, see
[4] and the references therein. Stochastic differential equations whose driven-
ing is given by a sum of independent fractional Brownian motion (i.e. more or
less the process Zt ) were already considered in [16].
We will start in the next section with some elementary properties, which gen-
eralize the results obtained in [10] and [17]. In Section 3 we consider the
correlation of the increments of Zt and show that the process Zt is long range
dependent. This is a property which makes it interesting for in particular fi-
nancial applications. Certain sample path properties of Zt are considered in
Section 4. As an auxiliary tool we calculate the maximal Hölder exponent,
which allows us to estimate the Hausdorff dimension of the sample paths of
Mixed fractional Brownian motion 1887

Zt from above. The lower estimate, which yields an equality at the end, is
obtained by the well known potential theoretic method based on Frostmann’s
famous lemma (cf. [5]). The last results are also new for the processes of
[10] and [17]. It is the aim of Section 6 to study fractional differentiability
of the sample paths of Zt . This problem was partially treated in the above
mentioned special cases in [10] and [17] and our result will generalize these ap-
proaches further. Moreover we will compare in Section 5 the different notions
of fractional derivatives, which can be found in the literature, see [14]. We
are also interested in local fractional derivatives introduced in [7] and further
developed in [8]. These notions will be compared with the approach of [15],
which is based on measure-geometric methods. We will further show that the
local fractional derivative of the critical order does not exists. However, the
fractional Cesáro derivative of the critical order vanishes, which are both new
result even in the case treated in [10] and [17]. This motivated us to consider
absolute fractional Cesáro derivatives, a concept introduced in [12]. As a re-
sult we will show that the fractional Cesáro derivative of the critical order is
positive and finite. We also calculate its exact value by applying linearity of
fractional differential operators and results from [12] and [13].

2 Elementary Properties
Let N ∈ N, a1 , . . . , aN ∈ R and H1 , . . . , HN ∈ [0, 1] and let
N
Zt := ak BtHk , t ∈ [0, ∞),
k=1

where the BtHk ’s are independent fractional Brownian motion with Hurst pa-
rameter Hk defined on some probability space [Ω, A, P].
It is clear that Zt is a Gaussian process, since linear combinations of Gaussian
processes are again a Gaussian process. Moreover, Zt is centered, i.e.
N N
EZt = E ak BtHk = ak EBtHk = 0.
k=1 k=1

It also follows immediately that

N 2 N N
EZt2 =E ak BtHk = a2k (BtHk )2 = a2k t2Hk .
k=1 k=1 k=1

The covariance function of Zt has the following form:

N N
1
a2k (t2Hk + s2Hk − |t − s|2Hk )
H
cov(Zt , Zs ) = EZt Zs = E aj ak Bt j BsHk =
2
j,k=1 k=1
1888 Christoph Thäle

for all t, s ∈ [0, ∞). We show now that the process Zt has stationary incre-
ments. Since they are Gaussian, it suffices to show that they are uncorrelated.
We obtain for any t1 , t2 , t3 , t4 ≥ 0

E(Zt2 − Zt2 )(Zt4 − Zt3 ) = EZt2 Zt4 + EZt1 Zt3 − EZt1 Zt4 − EZt2 Zt3
N N
1 2Hk 1 2Hk
2 + t4 − |t2 − t4 |
ak (tH 1 + t3 − |t1 − t3 |
Hk
= k
)+ ak (tH k Hk
)
2 2
k=1 k=1
N N
1 2Hk 1 2Hk
− 1 + t4 − |t1 − t4 |
ak (tH k Hk
)− 2 + t3 − |t2 − t3 |
ak (tH k Hk
)
2 k=1
2 k=1
= 0,

which gives the desired result.

We introduce now the family of scaling operators S(c1 ,... ,cN ;H1 ,... ,Hn ) , c1 , . . . , cN ≥
0, which act as follows:

N
 N
 N
fk (t) → S(c1 ,... ,cN ;H1 ,... ,HN ) fk (t) = c−H
k
k
fk (ck t).
k=1 k=1 k=1

Especially for the process Zt we obtain

N N
(S(c1 ,... ,cN ;H1 ,... ,Hn ) Z)t = ak c−H
k
k
BcHkkt = ak BtHk ,
k=1 k=1

using the self-similarity of the fractional Brownian motions BtHk mentioned in

the introduction. This shows that Zt is invariant under the family of transfor-
mations S(c1 ,... ,cN ;H1 ,... ,Hn ) , which is some kind of generalized self-similarity of
the process Zt .
We now turn to the Markov property of Zt . Obviously, if H1 = . . . = HN = 12 ,
Zt is a Markov process (recall that the Wiener process is one). Otherwise,
if 0 < H1 , . . . , HN < 1 and Hk = 12 for all k = 1, . . . , N we show that the
covariance functions do not satisfy the equality

cov(Zs , Zu )cov(Zt , Zt ) = cov(Zs , Zt )cov(Zt , Zu ), s, t, u ≥ 0, (1)

Mixed fractional Brownian motion 1889

which is necessary for the Markov property of Zt , see [6, Prop. 13.7]. We take
s = 12 , t = 1, u = 32 and obtain first
   2Hk 
N 2Hk
1 1 3
cov(Z 1 , Z 3 ) = a2k + −1 ,
2 2 2 2 2
k=1
N
cov(Z1, Z1 ) = a2k ,
k=1
N
1
cov(Z 1 , Z1 ) = a2k ,
2 2
k=1
N
  2Hk  2Hk 
1 3 1
cov(Z1, Z 3 ) = a2k 1+ − .
2 2 k=1
2 2

Then (1) is equivalent to

   2Hk    2Hk  2Hk 
N 2Hk N
1 3 1 3 1
a2k + −1 = a2k 1+ −
k=1
2 2 2 k=1
2 2

and further to
3 + 32Hk − 3 · 22Hk = 0
for all k = 1, . . . , N simultaneously. But the last set of equations has only the
solution Hk = 12 for all k = 1, . . . , N, which was treated above.
We summarize our findings in the following

Theorem 2.1 The process Zt has the following properties:

(a) Zt is a Gaussian process with EZt = 0 and covariance function

N
1
a2k (t2Hk + s2Hk − |t − s|2Hk )
2 k=1

for all t, s ∈ [0, ∞).

(b) Zt has stationary increments.

(c) Zt is S(c1 ,... ,cN ;H1 ,... ,HN ) -invariant.

(d) Zt is not a Markov process, unless H1 = . . . = HN = 12 .

1890 Christoph Thäle

3 Long Range Dependence

For each h > 0, the correlation coefficient of the increments Zt+h − Zt , t > 0,
is defined by

cov(Zt+h − Zt , Zs+h − Zs )
ρ(Zt+h − Zt , Zs+h − Zs ) =  , s, t > 0.
V(Zt+h − Zt )V(Zs+h − Zs )

The results from Theorem 2.1 show that

N 2
k=1 ak U(Hk )
ρ(Zt+h − Zt , Zs+h − Zs ) =  2 2Hk
,
2 N a
k=1 k h

where

U(Hk ) := (t + s − h)2Hk − 2(t − s)2Hk + (t − s − h)2Hk , ∀k = 1, . . . , N.

This implies that Zt has correlated increments. A detailed discussion of the

case N = 2 can be found in [10] or [17] for the more special case N = 2 and
H1 = 12 .
A random process Xt with stationary increments (note that the definition
given in [10], [17] cannot be satisfied by their own processes) is called long
range dependent, if the series
∞
r(n)
n=1

diverges, where
r(n) := cov(Z1 , Zn+1 − Zn ).
For our process Zt we have

Proposition 3.1 Zt is long range dependent if and only if there exists some
k ∈ {1, . . . , N } with Hk > 12 .

Proof The structure of the covariance function (see Theorem 2.1) and Talor’s
formula imply that
N N
r(n) = a2k Hk (2Hk − 1)n 2H1 −2
+ n2Hk −2 fk (n)
k=1 k=1

and
lim fk (n) = 0 ∀k = 1, . . . , N.
n→∞

From this it is easy to see that r(n) = +∞ if and only if there exists
1
k ∈ {1, . . . , N } with Hk > 2 . 2
Mixed fractional Brownian motion 1891

4 Sample Path Properties

4.1 Hölder continuity
The well known Kolmogorov-Chentsov Theorem [6, Thm. 3.23] states that if

E|Zs − Zt |a ≤ const|s − t|1+b , s, t ≥ 0

for some a, b > 0, the process Zt has a modification, which is a.s. α-Hölder
continuous for any α ∈ (0, b/a).
Denote by m the minimal Hurst exponent of components of Zt , i.e. m :=
min1≤k≤N Hk . It is easy to see that for any p ≥ 2 there exists some constant
c, such that
 p/2
E|Zt − Zs |p ≤ c E|Zs − Zt |2
(recall that Zt is a Gaussian process). Using Theorem 2.1 and the definition
of m we can further estimate that
 p/2 2mp
c E|Zs − Zt |2 ≤ c|s − t| 2 = c|s − t|mp = c|s − t|1+mp−1 .

Let now b := mp − 1 and a := p. Then Zt has a modification, whose paths are

α-Hölder continuous for any
mp − 1 1
α= = m − → m.
p p

We summarize the result in the following

Proposition 4.1 The process Zt has a modification, which is α-Hölder con-

tinuous for any
α < min Hk a.s.
1≤k≤N

From now on we will work with this modification.

4.2 Hausdorff Dimension of the Graph

The s-dimensional Hausdorff measure of a set A ⊂ Rd , s ≥ 0 a real number,
is defined by
∞ 

∞
Hs (A) := lim inf |Ak |s : A ⊂ Ak , |Ak | < δ ,
δ→0
k=1 k=1

where |Ak | is the diameter of the set Ak and the infimum is taken over all
coverings (Ak )k∈ of A. If s ∈ N than Hs equals up to a known constant the
1892 Christoph Thäle

s-dimensional Lebesgue measure Ls .

The Hausdorff dimension of the set A ⊂ Rd is defined by

dimH A := sup{s ≥ 0 : Hs (A) = +∞} = inf{s ≥ 0 : Hs (A) = 0}.

It is the purpose of this section to calculate the Hausdorff dimension of the

graph of the process Zt . We will split this task into two parts, upper bound
and the more involved lower bound.

4.2.1 Upper Bound

We start with the following easy Lemma, see [5, Cor. 11.2]: Let f : [a, b] → R2
α-Hölder continuous for some α ∈ (0, 1]. Then we have

dimH graph f ≤ 2 − α.

For a random process Xt the statement remains true, if Xt is α-Hölder contin-

uous with probability 1.
From Proposition 4.1 we know that our process Zt is α-Hölder continuous for
any α < min Hk . This implies that for any fixed T > 0 we have

dimH graph Zt |[0,T ] ≤ 2 − min Hk a.s.

1≤k≤n

And this implies

dimH graphZt ≤ 2 − min Hk a.s.
1≤k≤n

4.2.2 Lower Bound and the Result

A well known method for calculating the Hausdorff dimension of a set A is
the potential theoretic method [5, Thm. 4.13], which will be explained now for
completeness. Let μ be a finite measure on A (or Rd ) and define its s-Energy
(s ≥ 0) by  
s 1
I (μ) := dμ(x)dμ(y).
|x − y|s
Then we have:

If I s (μ) < ∞ for some finite measure on A then dimH A ≥ s.

We use now the potential theoretic method to obtain the lower bound for
the estimate of the Hausdorff dimension. We therefore fix some T > 0 and
consider the occupation measure μω (·) defined by
 T
μω (B) := 1B (t, Zt (ω))dt, B ∈ B2 ,
0
Mixed fractional Brownian motion 1893

where B2 is the Borel σ-algebra on R2 . This is a random measure concentrated

on the graph of Zt , when the time parameter t is restricted to the finite interval
[0, T ]. We show now that
 
1
E dμω (x)dμω (y) < ∞
A A |x − y|β

for A := graph Zt |[0,T ] for all β < min Hk . Let therefore x = (u, Zu ), y =
(v, Zv ) and calculate
 
1 1
E dμω (x)dμω (y)
2 A A |x − y|β
 T T
1  −β/2
= E (u − v)2 − (Zu − Zv )2 dudv
2 0 0
 T  T −v
 2 −β/2
= E u + (Zu+v − Zv )2 dudv
0 0
 T T
 2 −β/2
≤ E u + (Zu+v − Zv )2 dudv
0 0
∞  rn c  T  −β/2
= E u2 + (Zu+v − Tv )2 dudv,
n=−k r n+1 c 0

where k is the smallest natural number, for which T ≤ r −k c and c is constant.

We substitute now u = r n s, v = r n t and use the property that Zt has stationary
increments to obtain
∞   T r −n
c  −β/2
= r n
r n
E r 2n s2 + (Zrn (s+t) − Zrn t )2 dtds
n=−k rc 0
∞   T r −n
c  −β/2
= r n
r n
E r 2n s2 + (Zrn s − Z0 )2 dtds
n=−k rc 0
∞  c T r −n 
1 ∞  2n 2 −β/2
≤ T r n
r s + Nam r 2nm x2
n=−k rc 0 T r −n −∞

1 x2
·√ e− 2s2m dxdtds, (2)
2πsm

where m := min Hk and m := max Hk . Since

 T  
1 1 x2 1 T
√ e− 2s2m dx = Φ <∞
T 0 2πsm 2T 2sm
1894 Christoph Thäle

(here Φ(·) is the distribution function of the standard normal distribution) we

get
∞  c ∞
 2n 2 −β/2
(2) ≤ const · r n
r s + Nam r 2nm x2 dx ds. (3)
−∞
n=−k rc
  
=:I

The inner integral I can be evaluated explicitly:

πs(r2n s2 )−β/2 Γ(−1/2 + β/2)
I= ∼ const · s1−β r n(1−β−m) .
Γ(β/2)(N am r 2nm−2n )1/2
Thus
∞  c 
1−β
(3) ≤ const · r n
s ds r n(1−m−β)
n=−k  rc
 
=const
∞
≤ const · r n(2−m−β) < ∞,
n=−k

since 2 − m − β < 0 by the assumption on β.

Multiplying by a factor 2 we have shown that
 
1
E dμω (x)dμω (y) < ∞.
A A |x − y|
β

The potential theoretic method implies now

dimH graph Zt |[0,T ] ≥ 2 − min Hk

1≤k≤N

for all T > 0 and this implies the desired result

dimH graph Zt ≥ 2 − min Hk .

1≤k≤N

Combining the above lower estimate with the upper estimate from Section
4.2.1 we obtain

Theorem 4.2 The Hausdorff dimension of the graph of Zt equals 2− min Hk

1≤k≤N
with probability 1.

Note that this is a non-trivial statement and it does not follow from the
σ-additivity of the Hausdorff dimension, which says that for sets A1 , A2 , . . .
we have
∞
dimH Ak = sup dimH Ak .
k
k=1
Mixed fractional Brownian motion 1895

5 Elements of Fractional Calculus

Let 0 < α < 1 and fix some bounded interval (a, b). The left- and right-sided
Riemann-Liouville integral of fractional order α of a function f ∈ L1 (a, b) are
defined by
 x
1
α
Ia+ f (x) = (x − y)α−1 f (y)dy,
Γ(α) a
 b
1
α
Ib− f (x) = (y − x)α−1 f (y)f y.
Γ(α) x

The semi-group property of fractional integration reads as follows

α β α+β α β α+β
Ia+ Ia+ f = Ia+ f and Ib− Ib− f = Ib− f,

in any point for f ∈ C(a, b) or f ∈ L1 (1, b) and α + β ≥ 1 or in almost all

points for f ∈ L1 (a, b).
α α
The inverse operators of Ia+ and Ib− can for 0 < α < 1 be defined as
 x
1 d d 1−α
Da+
α
f (x) = (x − y)−αf (y)dy = I f (x),
Γ(1 − α) dx a dx a+
 b  
(−1)α d −α d
Db− f (x) =
α
− (y − x) f (y)dy = (−1) −
α
I 1−α f (x).
Γ(1 − α) dx x dx b−

They are called left- and right-side Riemann-Liouville fractional derivative.

With these definitions we have
α
Ia+ Da+
α α
f = f and Ib− Db−
α
f =f

for suitable functions f and any 0 < α < 1. For general α > 0 (i.e. α ≥ 1) the
above definitions can be extended in the following way: Write α = [α] + {α}
and define
 [α]
d {α}
Da+ f (x) :=
α
(Da+ f )(x), (4)
dx
 [α]
d {α}
Db− f (x) :=
α
− (Db− f )(x), (5)
dx

provided the expressions exist.

For the case, when the interval (a, b) is replaced by the whole real axis, one can
take the limits a → −∞, b → ∞ of the above expressions. These limits can be
understood in the sense of absolute convergence as well as in the more general
sense of conditional convergence (cf. [14]). In both cases, the operators Dαl
1896 Christoph Thäle

and Dαr are called left- and right-sided Wely-Marchaud derivatives of fractional
order α. For 0 < α < 1 and f ∈ C 1 (R) they may be written as
 ∞
α α f (x) − f (x − y)
Dl f (x) = dy, (6)
Γ(1 − α) 0 y 1+α

provided that f and f  decrease at least as |x|α−1−ε, ε > 0, as x → −∞,

see [14, p. 109] (analogously for the right-sided case Dαr f (x)). But (6) does
make sense for more general functions f . We therefore take this expression as
a definition, provided that it exists. For general α > 0 the definition can be
extend as in (4) and (5), respectively.
In [7] and [8], the concept of local fractional derivative was introduced and
considered. The (Kolwankar-Gangal) local α-fractional derivative of a function
f at x is defined as

Dα f (x) := lim Dy+

α
(f (y) − f (x)),
y→x

provided the limit exists. Using the definition of the Riemann-Liouville deriva-
tive, it is easy to see that Dα f (x) can also be written as
 y
1 d
D f (x) =
α
lim (f (t) − f (x))(y − t)−α dt.
Γ(1 − α) y→x dx x
Denoting the inner integral by I(y), the above expression may be interpreted
as the derivative of I at x, i.e. I  (x), provided it exists. An easy calculation
shows now that
f (x + Δ) − f (x)
Proposition 5.1 Dα f (x) = Γ(1 + α) lim .
Δ→0 Δα
(We like to mention at this point that there is an error in the calculations
of [8, Prop. 2].) In [8] this last proposition is used as a definition of the local
fractional derivative. For 0 < α < 1 put
f (x − Δ) − f (x)
Dlα f (x) := −Γ(1 + α) lim ,
Δ→0 Δα
f (x + Δ) − f (x)
Drα f (x) := Γ(1 + α) lim .
Δ→0 Δα
If Drα f (x) = Dlα f (x) = D α f (x), then f is called locally α-differentiable at x.
The real number

α(x) := sup{α ≥ 0 : D α f (x) exists}

is called in [7] the critical oder of differentiability of f at x. For more details

and the connection of this kind of fractional derivative and the Weyl-Marchaud
Mixed fractional Brownian motion 1897

derivative we refer to [8].

A much more general concept of local fractional differentiability was however
already introduced in [15]. In this paper the author defined the so-called
lower- and upper fractional Cesáro derivative of order α, 0 < α < 1 by taking
the Cesáro average in the definitions of Dlα and Drα on the logarithmic scale
(omitting the factor Γ(1 + α)):

α 1 1 f (x + y) − f (x) 1
d f (x) := lim inf dy,
T →∞ T e−T yα y

α
1 1 f (x + y) − f (x) 1
d f (x) := lim sup dy.
T →∞ T e−T yα y

A function f is said to have a fractional Cesáro derivative of order α at x if

dα f (x) = dα f (x) = dα f (x). By [15, Prop. 2] there exists for any x in the
domain of f a unique γ(x), such that
 
1
1 |f (x + y) − f (x)| 1 0 : α < γ(x)
lim sup dy =
T →∞ T e−T y α y ∞ : α > γ(x).

The value γ(x) is called the fractional degree of differentiability of f at x. This

definition is justified by the following fact [15, Thm. 2]:

Proposition 5.2 Let γ(x) be the fractional degree of differentiability of f

at x. If α > γ(x) then the Weyl-Marchaud derivative Dαr f (x) does not exist
and for α < γ(x) it exists iff
 ∞
|f (x + y) − f (x)|
dy < ∞.
1 y 1+α

The additional averaging ensures that the expression exists also in such
situations, where too many rapid oscillations occur and the limit without av-
eraging does not make sense. From the definitions and some standard analysis,
the following comparison result is easily seen:

Theorem 5.3 Let f be a function and x a value in its domain. If f is

locally α-differentiable at x, then Γ(1 + α)dαf (x) = D α f (x) and the fractional
degree of differentiability of f at x equals the critical order α of f at x. More-
over, the Wely-Marchaud derivative of f exists iff
 ∞
|f (x + y) − f (x)|
dy < ∞.
1 y 1+α
1898 Christoph Thäle

Note that the other direction is not necessarily true. In particular there
are examples of functions, for which dα f (x) exists, but not D α f (x). We will
see later that the Wiener process, the fractional Brownian motion and also our
process Zt provide examples for this fact. This point was already stressed in
the introduction of [12] for the case of the Wiener process as a motivation for
the definition of the fractional Cesáro derivative dα .
We apply our findings now to functions, which are Hölder continuous.

Corollary 5.4 Let f be α-Hölder continuous, i.e. there exists a finite con-
stant c > 0 with |f (x + y) − f (x)| ≤ c|x − y|α for all x, y in the domain of f .
Then α(x) = γ(x) = α for all x in the domain of f .

6 Fractional Differentiation of the sample paths

of the process Zt
We want to study the fractional differentiability of the process Zt . It is easy
to see that both, the Kolwankar-Gangal fractional derivative as well as the
fractional Cesáro derivative are linear, which means that

D β (af + bg)(x) = aDβ f (x) + bDβ g(x)

dβ (af + bg)(x) = adβ f (x) + bdβ g(x)

for suitable functions f, g and real numbers a, b, β. Furthermore Corollary 5.4

holds in the random case:

Proposition 6.1 Let Xt , t ∈ [0, ∞), be random process which is α-Hölder

continuous with probability 1, 0 < α < 1. Then the critical order of differen-
tiability and the fractional degree of differentiability agree and are equal to the
order of Hölder continuity, i.e. α(t) = γ(t) = α almost surely for almost all
t ∈ [0, ∞).

From Section 4 we know that Zt is a.s. α-Hölder continuous of any oder

α < min1≤k≤N Hk . Recall further that the random process Zt is a linear combi-
nation of fractional Brownian motion. This together with the last proposition
and the linearity of the fractional differential operators leads to

Theorem 6.2 For the random process Zt we have

γ(t) = α(t) = min Hk a.s.

1≤k≤n
Mixed fractional Brownian motion 1899

In [10] and [17] a similar result for critical order was also obtained, but
these authors used other techniques. Unfortunately, they did not consider the
case of the critical order. In particular the question if D α Zt exist remained
open. In fact the law of iterated logarithm

BtH
lim sup = 1 a.s.
t→0 tH (2 ln | ln t|)1/2

from [1] shows that

H
Bt+Δ − BtH
D H
BtH = Γ(1 + α) lim = ∞ a.s. for a.a. t ∈ [0, ∞)
Δ→0 ΔH
and this immediately implies by using linearity of D α that

0 : α < min Hk
D α Zt =
+∞ : α ≥ Hk

with probability 1 and for almost all t. Therefore we will use here the more
general concept of fractional Cesáro differentiability and determine dα Zt for
α = min Hk . We apply the results from [12] to obtain first

dHk BtHk = 0 a.s.

for a.a. t ∈ [0, ∞). By linearity this shows that

N
α
d Zt = ak dα BtHk = 0
k=1

with probability 1 for almost all t ∈ [0, ∞). It means that the averaging
procedure in the definition of dα is was useful in order to ensure the existence
of fractional derivatives of fractal functions. This fact was already explained
in some detail in [12] and [15]. We summarize now our results:

Theorem 6.3 We have with probability 1 for almost all t ∈ [0, ∞)


α 0 : α ≤ min1≤k≤N Hk
d Zt =
+∞ : α > min1≤k≤N Hk

for a.a. t ∈ [0, ∞). Moreover, by Theorem 5.3 and the law of iterated logarithm
for the fractional Brownian motion, the Weyl-Marchaud derivative Dαr Zt does
not exist with probability 1 and for almost all t ∈ [0, ∞) for any α > 0.
1900 Christoph Thäle

Regarding the last result, it seems therefore to be more promising to con-

sider the absolute fractional Cesáro derivative of a function f order α at x,
which is defined as

1 1 |f (x + y) − f (x)| 1
|d |f (x) := lim
α
dy.
T →∞ T e−T yα y

Again using [12] we obtain for the component processes

|dHk |BtHk = E|B1Hk | a.s. for a.a. t ∈ [0, ∞).

The linearity implies now together with Theorem 6.2

Theorem 6.4 We have with probability 1 for almost all t ∈ [0, ∞)

⎧
⎪0
⎨ : α < min1≤k≤N Hk
|d |Zt =
α
k=1 ak E|B1 |
N Hk
: α = min1≤k≤N Hk
⎪
⎩
+∞ : α > min1≤k≤N Hk .

Note that in [12] (Thm. 1 and its Corollary) certain ergodicity conditions
are assumed. It is easy to verify that they are fulfilled for random processes
with stationary and independent increments and it was shown in [13, pp. 92-
93] that these ergodicity properties are also fulfilled for fractional Brownian
motion. It is easily seen that the same would hold true for our process Zt , but
this does not enable us to apply the results from [12] directly to the random
process Zt , since Zt does not fit into the theory of self-affine random functions.

References
[1] Arcones, M.A.: On the law of iterated logarithm for Gaussian processes,
J. Theor. Prob. 8, 877-904 (1995).

[2] Cheridito, P.: Mixed fractional Brownian motion, Benoulli 7, 913-934

(2001).

[3] Decreusefond, L.; Üstünel, A.S.: Stochastic analysis of the fractional

Brownian motion, Potential Anal. 10, 177-214, (1999).

[4] Drakakis, K.; Radulovic, D.: A discretized version of the self-similar

model for internet traffic, Appl. Math. Sci. 2 (2008), 2743-2756.

[5] Falconer, K.J.: Fractal Geometry - Mathematical Foundations and Ap-

plications, 2nd Edition, Wiley, (2003).
Mixed fractional Brownian motion 1901

[6] Kallenberg, O.: Foundations of Modern Probability, 2nd Edition,

Springer, (2001).

[7] Kolwankar, K.M.; Gangal, A.D.: Fractional differentiability of nowhere

differentiable functions and dimension, Chaos 6 (1996), 505-513.

[8] Li, X.; Essex, C.; Davison, M.: A local fractional derivative, Proceed-
ings of the 1st symposium on fractional derivatives and their applications
(2003).

[9] Mandelbrot, B.; Van Ness, J.W.: Fractional Brownian motion, fractional
noises and applications, SIAM Review 10 (1968), 422-437.

[10] Miao, Y.; Ren, W.; Ren, Z.: On the fractional mixed fractional Brownian
motion, Appl. Math. Sci. 35 (2008) 1729-1938.

[11] Mishura, Y.: Stochastic Calculus for Fractional Brownian Motion and
Related Processes, Lecture Notes in Mathematics 1929, Springer (2008).

[12] Patzschke, N.; Zähle, M.: Fractional differentiation in the self-affine case
I - random functions, Stoch. Proc. Appl. 43 (1992), 165-175.

[13] Patzschke, N.; Zähle, M.: Fractional differentiation in the self-affine case
IV - random measures, Stochastics and Stochastic Reports 49 (1994),
87-98.

[14] Samko, S.G.; Kilbas, A.A.; Marichev, P.I.: Fractional Integrals and
Derivatives - Theory and Applications, Gordon and Breach (1993).

[15] Zähle, M.: Fractional differentiation in the self-affine case V - The local
degree of differentiability, Math. Nachr. 185 (1997) 279-306.

[16] Zähle, M.: Stochastic differential equations with fractal noise, Math.
Nachr. 278 (2005) 1097-1106.

[17] Zili, M.: On the mixed fractional Brownian motion, J. Appl. Math.
Stoch. Anal. vol. 2006 (2006) 1-9.

Received: November, 2008