Applied Probability Trust (21st November 2002) PETTERI MANNERSALO, * VTT Technical Research Centre of Finland, Helsinki, Finland ILKKA NORROS, ** VTT Technical Research Centre of Finland, Helsinki, Finland RUDOLF H. RIEDI, *** Rice University, Houston Texas, U.S.A. MULTIFRACTAL PRODUCTS OF STOCHASTIC PROCESSES: CONSTRUCTION AND SOME BASIC PROPERTIES Abstract In various fields, such as teletraffic and economics, measured time series have been reported to adhere to multifractal scaling. Classical cascading measures possess mul- tifractal scaling, but their increments form a non-stationary process. To overcome this problem we introduce a construction of random multifractal measures based on iterative multiplication of stationary stochastic processes, a special form of T-martingales. We study L 2 -convergence, non-degeneracy and continuity of the limit process. Establishing a power law for its moments we obtain a formula for the multifractal spectrum and hint at how to prove the full formalism. Keywords: Stochastic processes, random measures, multifractals, teletraffic modeling AMS 2000 Subject Classification: Primary 60G57 Secondary 60G30 * Postal address: VTT Information Technology, P.O. Box 1202, FIN-02044 VTT, Finland * Email address: [email protected]** Postal address: VTT Information Technology, P.O. Box 1202, FIN-02044 VTT, Finland ** Email address: [email protected]*** Postal address: ECE Dept, Rice University MS 380, Houston TX 77251-1892, USA *** Email address: [email protected]1
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Applied Probability Trust(21st November 2002)
PETTERI MANNERSALO,∗
VTT Technical Research Centre of Finland, Helsinki, Finland
ILKKA NORROS,∗∗
VTT Technical Research Centre of Finland, Helsinki, Finland
RUDOLF H. RIEDI,∗∗∗
Rice University, Houston Texas, U.S.A.
MULTIFRACTAL PRODUCTS OF STOCHASTIC PROCESSES:
CONSTRUCTION AND SOME BASIC PROPERTIES
Abstract
In various fields, such as teletraffic and economics, measured time series have been
reported to adhere to multifractal scaling. Classical cascading measures possess mul-
tifractal scaling, but their increments form a non-stationary process. To overcome this
problem we introduce a construction of random multifractal measures based on iterative
multiplication of stationary stochastic processes, a special form of T-martingales. We
studyL2-convergence, non-degeneracy and continuity of the limit process. Establishing
a power law for its moments we obtain a formula for the multifractal spectrum and hint
at how to prove the full formalism.
Keywords:
Stochastic processes, random measures, multifractals, teletraffic modeling
AMS 2000 Subject Classification: Primary 60G57
Secondary 60G30
∗ Postal address: VTT Information Technology, P.O. Box 1202, FIN-02044 VTT, Finland∗ Email address: [email protected]∗∗ Postal address: VTT Information Technology, P.O. Box 1202, FIN-02044 VTT, Finland∗∗ Email address: [email protected]∗∗∗ Postal address: ECE Dept, Rice University MS 380, Houston TX 77251-1892, USA∗∗∗ Email address: [email protected]
1
2 Mannersalo, Norros and Riedi
1. Introduction
This study is strongly motivated by the search of new models for teletraffic. In a sequel
of papers [28, 19, 22, 27, 9] it has been demonstrated that teletraffic has a very rich scaling
structure when the measured traffic traces are looked at the very finest resolutions, usually
resolutions 100 ms and smaller. More precisely, letA(t, s) denote the amount of traffic (bytes)
arriving on interval[t, s). Empirical studies of different network environments suggest that the
scaling law
log EA(t, t + δn)q ≈ c(q) log δn + Cq
holds over a wide range of resolutionsδn with a non-linear functionc(q). In other words, the
large-deviation based multifractal spectrum seems to be non-trivial. As an illustrative example,
we consider a traffic sample measured in an international link of the Finnish University and
Research Network (Funet). An IP traffic trace and the corresponding empirical multiscaling
moment plots are shown in Figure 1. Also in this case, multifractal-type behavior is seen.
Currently, there is no complete physical understanding which and how network elements result
in this phenomenon. However, the most probable candidate is the joint dynamics of TCP
(Transmission Control Protocol) and queues acting in a extremely heterogeneous environment.
0.1 0.2 0.3 0.4 0.5sec
50
100
150
200
250Mb/s Resolution 0.25 ms
0.0001 0.001 0.01 0.1d1
10000
1.×108
1.×1012
1.×1016
1.×1020
A^(t,t+d)q
FIGURE 1: A traffic trace from an international link of FUNET and the corresponding empirical
multiscaling moments withq = 0.5, 1, . . . , 5, starting from the bottom
There are many ways to construct random multifractal measures varying from the simple
binomial measures to measures generated by random branching processes (see e.g., [21, 26,
1, 5, 2, 25, 27]). In teletraffic modeling, we would like to have, in addition to a simple
andcausalconstruction, alsostationarity, i.e., stationarity of the increments of the process
A(0, t). Unfortunately, most of the ‘classical’ multifractal models, in particular tree-based
Multifractal product of stochastic processes 3
cascades, lack both of these properties. It should be noted that a multifractal process with
stationary increments is not a completely new idea. Jaffard [12, 13] showed that certain Lévy
processes are also multifractal. However, since Lévy processes have linear multifractal spectra
and real data traffic exhibits strictly concave spectra, [28, 19, 22, 27] they are unsuitable for
our modeling needs. Processes with stationary increments have also been used by Gupta
and Waymire [10] in the context of scaling and multiplicative structures, yet with different
approach and objectives.
Research on multiplicative cascades has been very active. Especially, Mandelbrot’s martin-
gale [20, 21], a simple tree-based construction with independent random multipliers, has been
considered in a large number of publications; first by Kahane and Peyriére [18], and the story
still continues (see e.g. [11, 5, 30, 23, 2, 3, 24]). Extensions such as relaxing the independence
assumption of the multipliers or randomizing the number of offsprings have also been studied.
To give a short list without intention of being complete we refer to Molchan [23] and Waymire
and Williams [31, 32] regarding dependent multipliers, to Peyriére [26], Arbeiter [1], and Burd
and Waymire [4] regarding random number of offsprings.
We consider a natural and stationary generalization of the multiplicative cascade construc-
tion. This scheme, as well as the genuine cascades, belong to the framework of the positive
T -martingales and multiplicative chaos introduced by Kahane [16, 17, 14]. Let(Λ(i)) be
a sequence of positive independent random functions (i.e. processes) defined on a compact
metric spaceT such thatEΛ(i)(t) = 1 for all t ∈ T and consider finite products
Λn(t) =n∏
i=1
Λ(i)(t).
HereΛn is an indexed martingale, since it is a martingale for eacht ∈ T . By [17], Λn(t) dν(t)
converges weakly to a random measureΛ(t) dν(t) for any positive Radon measureν. Only
partial answers are known regarding the convergence inLp, 1 ≤ p < ∞. Special cases which
have been studied include, for example, Gaussian chaos, i.e., lognormal multipliersΛ(i), by
Kahane [16], Lévy chaos by Fan [6], and random Gibbs measures by Fan and Shieh [8]. Note
that theT -martingale approach works also with random coverings (see e.g. [17, 14, 15, 7]).
In our setting, we restrict ourselves to the Lebesgue case and study convergence and related
questions of the limiting measure ofΛn(t) dt, and in its simplest form our model is based on
the multiplication of independent rescaled stationary stochastic processesΛ(i)(·) dist= Λ(bi·)
which are piecewise constant (heredist= denotes equivalence in distributions). It is instructive
4 Mannersalo, Norros and Riedi
to compare it to a Fourier decomposition where one represents or constructs a process by
superposition of oscillationssin(λit).
In multiplying rather than adding rescaled versions of a ‘mother’ process we obtain a
process with novel properties which are best understood not in an additive analysis, but in
a multiplicative one. Processes emerging from multiplicative construction schemes can easily
be forced to have positive increments and exhibit typically a ‘spiky’ appearance. The so-called
multifractal analysis describes the local structure of a process in terms of scalingexponents,
accounting for the multiplicative structure.
It is tempting to view the multiplicative construction as an additive one — which opens the
possibility to use linear theory — followed by an exponential. Such an approach, however, ob-
scures what happens in the limit. As with the cascades, an infinite product of random processes
will typically (almost surely at almost all times) be zero; equally, its logarithm tends to negative
infinity. A non-degenerate limiting behavior can be observed for the product, though, by taking
a distributional limit rather than pointwise limit. In simpler words, a multiplierΛ(i)(t) should
not be evaluated in points but should be seen as redistributing or re-partitioning the mass. In
the words of teletraffic modeling,Λ(i)(t) can be thought of as a local change in the arrival rate
where one is interested actually in the integrated ‘total load’ process. Consequently, we will
study
An(t) =∫ t
0
n∏i=0
Λ(i) (s) ds,
which converges to a well defined, non-degenerate and continuous process under suitable
conditions.
The paper is structured as follows: We start by studying the construction of multifractal
measures based on iterative multiplication of stochastic processes as inAn above, in particular
convergence and non-degeneracy. Then, we consider a special case where the multipliers are
independent rescaled versions of some mother process, looking at continuity, power laws of
moments as well as long-range dependence (LRD) of the limiting process. Finally, we pro-
vide an application-friendly family based on piecewise constant processes with exponentially
distributed sojourn times.
Multifractal product of stochastic processes 5
2. Multifractal products of stochastic processes
In order to keep the presentation simple, we only consider 1-dimensional processes on the
closed unit intervalT = [0, 1]. Extensions to the real lineR as well as to higher dimensions
are not too difficult.
2.1. Construction
We start fromT -martingales defined by independent multiplication as in [17]. Consider a
family of independentpositiveprocesses{Λ(i)(t)}t∈T with
EΛ(i)(t) = 1 ∀t ∈ T, i = 0, 1, 2 . . . .
Later, when studying particular properties of the process, we will usually assume that theΛ(i)’s
are stationary, but this is not necessary in the general case.
Define the finite product processes
Λn(t) .=n∏
i=0
Λ(i)(t).
and the corresponding cumulative processes
An(t) .=∫ t
0
Λn(s) ds =∫ t
0
n∏i=0
Λ(i) (s) ds, n = 0, 1, . . .
Sometimes, it is easier to consider the corresponding positive measures defined on the Borel
setsB of T :
µn(B) .=∫
B
Λn(s) ds, n = 0, 1, . . . , B ∈ B.
2.2. Convergence
Theorem 1. [17] The random measuresµn converge weakly a.s. to a random measureµ.
Moreover, given a finite or countable family of Borel setsBj on T , we have with probability
one:
∀j µ(Bj) = limn→∞
µn(Bj).
The a.s. convergence ofAn(t) in countably many points at the same time can be extended
to all points inT if we know that the limit processA is almost surely continuous. Conditions
for continuity in a more specific setup are given later in Proposition 4.
6 Mannersalo, Norros and Riedi
Corollary 1. If An → A weakly a.s., andAn andA are continuous almost surely, then with
probability one:∀t ∈ T, A(t) = limn→∞ An(t).
Unfortunately, Theorem 1 does not say anything about theL1-convergence; as is noted in
[17], one of the following two cases is met: eitherAn(t) → A(t) in L1 for each givent
or An(1) converges to0 almost surely. The cases are callednon-degenerateanddegenerate,
respectively. In this paper, we present conditions forL2-convergence which is the easiest to
handle.
Proposition 1. Suppose that the stationary independent processesΛ(i), i = 0, 1, . . ., satisfy