Professional Documents
Culture Documents
After more than 40 years the so-called Hurst effect remains an open problem in stochastic
hydrology.Historically, its existencehas been explainedeither by preasymptoticbehaviorof the
resca!edadjustedrangeR'n, certainclassesof nonstationarity
in time series,infinitememory,or
erroneous estimation of the Hurst exponent. Various statistical tests to determine whether an
observedtime seriesexhibitsthe Hurst effectare presented.The testsare basedon the fact that for
thefamilyof processes
in theBrownian
domain
of'attraction,
R*n/((On))•/2
converges
in distribution
to a nondegeneraterandom variable with known distribution(functional central limit theorem). The
scaleof fluctuation0, definedas the sumof the correlationfunction,plays a key role. Applicationof
the teststo severalgeophysicaltime seriesseemsto indicatethat they do not exhibit the Hurst effect,
althoughthoseserieshave beenusedas examplesof its existence,and furthermore the traditionalpox
diagrammethodto estimatethe Hurst exponentgivesvalueslargerthan 0.5. It turned out that the
coefficient
in therelationofR*nversus
n, whichis directlyproportional
to thescaleof fluctuation,
was
more important than the exponent.The Hurst effect motivated the popularizationof 1/f noisesand
related ideas of fractals and scaling.This work illustrateshow delicate the proceduresto deal with
infinity must be.
that, in general,R*ngrowslike n to a powerof the orderof standarddeviationof the Xi, say, or.In that case,R*• is
0.72 with some variations for the different records, but in all scaled
bycrbecause
it isthelimitofR*•which
isscaled
by
the casesthe exponentwaslargerthan0.5. This resultwasin Dn.
contradictionwith his own theoretical analysiswhich indi- A consequenceof the above definition and of the invari-
catedthat the exponentshouldbe asymptotically0.5. Feller ance principle is that for a sequenceto exhibit the Hurst
[ 1951] was quick to providefirm theoreticalcomputationsfor effect,it is necessarythat at leastone of the conditionsofthe
the caseof independentidenticallydistributed(IID) random functionalcentral limit theorem be violated. Becauseof the
variableswith finite secondmoments.He computedthe limit generality of these hypotheses something quite dramatic
mean and variance of the rescaled adjusted range, and must be happeningfrom the physical point of view anda
suggestedthat Hurst's [1951] empirical findings might be searchfor the cause of the violation is very much in order.
explained by some kind of Markovian dependence.The The discoveryof anomaliessuchas a nonstationarity or a
discrepancybetween the empirical observationshowingthe strong dependence in that large a class of geophysical
increaseof the rescaledadjustedrangelike a power of 0.7 or recordswill have profound physical implications.
so, and the theoretical expectation that for a wide class of As may be expected,from a retrospectiveviewpoint,the
processesthe exponent is asymptotically 0.5 has becometo explanationsproposed by different authors were relatedto
be known as the Hurst effect. The exponent H in the the violationof the hypothesesof the invarianceprinciple.
empirical relation There have been some theories essentially related to the
dependencestructureof the process,to its stationarity,to
a*n"- nH (5) the existence of an infinite second moment, and to the
preasymptoticbehavior of the limiting process.
is called the Hurst exponent.
The strongdependenceexplanation was initiated by Hurst
To appreciate the different argumentsthat have been put
[1951] himself, and by Feller's [195!] suggestionaboutthe
forward as possibleexplanationsof the Hurst effect, it is
Markovian character of the series being responsiblefor the
necessaryto use a more precisedefinition.Suchprecisionis
Hurst effect. However, soon Barnard [1956] showed the
important also for the proper understandingof the estimation
existence of a 0.5 convergence for that case. Nevertheless,
problem of the next section. To that end, we will use the
Matalas and Huzzen [ 1967]presentedreports of estimatesof
definition of Bhattacharya et al. [1983, p. 651]: "A sequence
H between 0.58 and 0.87 for simulation of first-order autore-
of random variables is said to exhibit the Hurst effect with
exponent
H > 0 5 ff (1/nU)R* converges
' as gressiveMarkovian models with different correlation coeffi-
in distribution,
t!
cients. At that time the Markovian explanation seemed
n goesto infinity, to a non-zero random variable." This is in
adequate, and the autoregressive models were becoming
contrast with a very general result, known as the invariance
fashionable.However, Mandelbrot and Van Ness [ 1968]put
principle or the functional central limit theorem [Ibragimov,
things back in proper perspective by recalling that for
1962;BilIingsley, 1968], which implies that under conditions
processesin the domain of attraction of the Brownian
of stationarity andweakdependence R*•/nø'5converges in
distribution to a randomvariableR*•/n0.5with mean motion, the rescaled adjusted range grows like n to the 0.5
power (notice that Barnard's result goes back to 1956,and
E(n-ø'SR*•)= (0•r/2)1/2 (6) that the invarianceprincipledatesback to Ibragimov [1962]).
Mandelbrot and Wallis [!968, 1969] and Mandelbrot and
and variance Van Ness [1968]proposedan explanationof the Hursteffect
by the strong dependenceof the geophysical seriesand
Var (n-ø'SR*•)= 0 (z-2/6- z-/2) (7) introduced the so-called fractional Brownian motion and the
where 0 is a positive constant, the so-calledscale of fluctu- fractional Brownian noise (1/f noise), which are processes
ation, or correlation length scale, a parameter first intro- with infinitememory, as modelsthat may be usedto simulate
duced by Taylor [1921]. It can be shown that 0 is the sum of the Hurst effect. These important theoretical contributions
the correlation function were not accompaniedby either a physical explanationnor
an investigationof the structure of dependenceof the
differentgeophysical series.The lackof physicaljustification
0= Z p(m) (8) of thistheory was pointedout by authorssuchas Scheideger
rn-- -• [1970]and Klem•s [1974]. There has not been appreciation
for thiskind of questionas the followingquote shows:"I am
where p(m) denotesthe correlationcoefficientbetweenXn preparedto arguethata lackof seriousmotivationina model
and Xn+ m. that fitsand workswell is muchpreferableto lack of fit in a
Equations (6) and (7) correspondto Feller's [195!] previ- model that seems well motivated" [Mandelbrot, !982, p.
ous result, for in that case 0 is one. Besides the mean and the 253].
variance, the whole asymptotic distribution of the rescaled On the estimation side, the contribution of Mandelbrot
adjustedrange are ready availablefrom known resultsin the and Wallis[1968]wasalsoimportant.They pointedoutthat
BrownJan motion case [Bhattacharya et al., 1983]. The the way Hurst [195!] estimated the exponent H was not
condition of weak dependenceimplies a correlation function adequate. Hurst [1951] used the equation
decreasingfast enoughto ensureconvergenceof the seriesin
(8). This explains the use of terms such as strong depen- H l = log(R*n)/log
(n/2) i9)
dence, infinite memory (infinite 0), and short memory(finite whichpresupposes
that the relation(R*) versusn in loga-
0). Normally, existence of finite second moments for the rithmicpaperpasses
throughthepointn = 2 R* = I This
sequence of Xi is assumed. Then the ergodic theorem wasmotivated by weakempiricalarguments,
becauseit is
applied
toX•2 implies
thatDn converges
inprobability
tothe easy to see that using the biased estimator for O'n, one
MESA AND POVEDA: HURST EFFECT 3997
obtains
R*n= 1 whenn = 2. Mandelbrot andWallis[1968] els were used in hydrology to model Hurst's exponents
the so-called"pox" diagram.This and other largerthan 0.5. The broken line model [Rodrlguez-Iturbeet
introduced
waysofestimating
H will bediscussedin thenextsection. al., 1972; Mejia et al., 1972] and the ARMA(1, 1) model
Attempts of [O'Connell, 1974]are the standardpractice. Indeed, Hipel
to explainthe Hurst effectas a consequence
aninfinitesecondmomentwereproposedby Moran [ 1964] and McLeod [1978]demonstratedthat ARMA(p, q) models
andBoesand Salas [1973]. However, Mandelbrotand statisticallypreservethe rescaledadjustedrangeor equiva-
Taqqu
[1979]demonstrated
thatanasymptotic with lently the Hurst exponentestimatedas (9). One can interpret
relation
//= 0.5 holdsfor a sequence of IID randomvariableswith the philosophyof thesemodelsas an attemptto increasethe
distributionand characteristicexponentstrictly less correlation length scale 0 so as to obtain preasymptotic
stable
than2 (seeFeller [1971, p. 169] for a discussionof stable estimates of the Hurst coefficient similar to the observed in
dist•butions). empirical records. The conclusionsof the chapter on the
Hurst [1951] recognized the nonstationaritiesin his origi- Hurst effect in the textbook by Bras and Rodriguez-Iturbe
nalgeophysical
series.He evendesigned
anexperiment
with [1984,p. 265]are very illustrativeof the standardpracticein
"probability
cards"whichproduced sudden
changesin the engineeringhydrology. This is most noteworthy if one
meanof the process,and obtainedempiricalestimationsof considersthat the physical problem is unsolved.
theexponentH near 0.71. Klemds[1974]andPotter [1975] In fact, there is no physical explanation yet for the
developed simulationswith nonstationary modelsthat pro- occurrenceof infinitememoryin geophysicalseries.There is
ducedempirical series exhibiting the Hurst effect. Some- no systematic(physicalor empirical)studyof the correlation
thingsimilarwasobtainedby Boesand Salas[1978]with the structureof theseprocesses.Even thoughthe nonstationar-
shiftinglevels model. It is worth noting that all of these ity of some geophysicalprocessesmay be argued on some
nonstationarymodels belong to the Brownian domain of physicalgrounds[Leopold et al., 1964, p. 61], it is only in
attraction. very generalterms and it remains to be explained why the
Bhattacharya et al. [1983] provided clear mathematical trendsproducethe sameHurst exponentin variousgeophys-
demonstration
of the existence of the Hurst effect for weakly ical series. Moreover, the issues of estimation raised by the
processesperturbedby smallmonotonictrends. preasymptoticexplanationsare very relevant and have no
dependent
Asan example, let Y• be a sequenceof weakly dependent definiteanswers.The importance of the estimationproblem
randomvariables, say IID normal variables with zero mean is reinforced if one considers the claims about the robustness
andunit variance, then of the range analysis in current literature on fractals [Man-
delbrot, 1982, p. 382; Feder, 1988, p. 194].
Xn -- Yn+ c(m+ n)• (10)
willexhibit the Hurst effect with exponent H dependent on 3. TESTS FOR THE EXISTENCE OF THE HURST EFFECT
•e valueof the parameter/3as follows:for - 1/2 </3 < 0, H
isequalto 1 + /3, for/3 > 0, the exponentH is 1, and for Notice that the whole puzzle rests on the empirical evi-
/3< -1/2, and for /3 = 0, the Hurst exponentis I/2 [see dence of the exponent H being larger than 0.5. Also, recall
Bhattacharyaet al., 1983]; notice the discontinuityat/3 = 0. thatthequestion isnotrelatedtominorthings: no.7isalmost
In all the cases,m and c are arbitrary parameters(c > 0 and 4 timesn0.sfor n = 1000, andthe factorkeepsincreasing
m -> 0). The significanceof this demonstrationis twofold. with n. With all the history behind but the perspective of
First, the class of processeswith infinite memory is no huge unsolved questions present, the least that can be done
longerthe only theoretically proved class of processes is to look further into the empirical evidence.
exhibitingthe Hurst effect. Second, the estimationproblem
3.1. Estimators of H
becomes very important, for it is possiblethat very small
trendsmay be responsiblefor the appearanceof the Hurst As was pointed out before, Hurst [1951] originally esti-
effect. mated H by means of (9), but this practice was shortly
In additionto the infinite memory and the nonstationarity abandonedand substitutedby the least squaresslope in the
explanations,there have been theories that present the linear relationof log (n) versuslog R* alongwith some
Hursteffectonly as a preasymptoticbehavior.This means other variations. How good are those estimators?Very little
theconvergence to the theoretical0.5 exponentis slow,and theoretical work has been done along these lines. As an
thereforethe empiricalobservations for finitesamplesizes alternative for the complex theoretical issues involved,
maygive Hurst exponentslarger than 0.5 [Lloyd, 1967]. extensivecomputer experimentswere performed by Poveda
SaMsand Boes [1974] consideredthe equation for the [1987] using Bhattacharya et al.'s [1983] nonstationary
expectedvalue of R n for finite n and the case of IID model(equation(10)), given its capacityto producevaluesof
variablesshowingthe preasymptoticbehaviorof the ad- H, at will, by fixing fl. This analysis consideredmost of the
justedrange,andproposed a differentway of estimation of estimators reported in the literature [see also Poveda and
theexponent H. Gomide[1975,1978]considered the caseof Mesa, 1993]; Table 1 presents Poveda's [1987] results for
MarkovJan processes and showedhow the preasymptoticone caseand someof the estimators.The performanceof all
regionis expandedwith valuesof p(1) near one, and pro- estimatorsin all caseswas poor. Slightly better resultswere
poseda new way of estimationof the exponentH. Salas et obtainedwith a new estimator proposed by Poveda [1987],
al. [1979a, hi, using autoregressive movingaverage which consistsof the slope Sn of the regressionof sample
(ARMA)(1,1)models,showed region valuesof R*mversusm takingonlyvaluesof m largerthann,
thatthepreasymptotic
isexpanded because of eitherasymmetricmarginal
distribu- but some degree of arbitrariness remains regarding the
tion,largebutfinitememory(large0) andnonstationarity.choice of n.
It is worth notingthat parallelto the preasymptotic A conclusionof these computerexperimentsis that all the
explanation
of the Hurst effect,variousshort-memory
rood- estimatorsof the Hurst's [1951] exponent H performed
3998 MESA AND POVEDA: HURST EFFECT
8'
o
.................................
;...............
..........
........
.............................................
2 o []
0
' ' "" '-' "i0 ....... ....... 0 20 40 60 80 1O0 120
n
Fig. I•. GEOS dia•m for a time se•s of tree •ngs of a Fig. 1c. GEOS diagram for St. Lawrence river discharges.
•u•ss Hr(Snake•ver). HoHzont• dashedlinescodespondto the
•ymptoticme• •d •2 stand•d deviations.
-
the factorsmultiplyingq,• and q, are affectedby the new
situationof 0* being an estimate. In a processexhibiting the
SinceF(n) is not known, finding an estimate 0* of the
scaleof fluctuation will require some iteration. Using an
Hursteffect,theincrease
of R*•/nø'5withn will eventually
dominate. The hypothesis of the test is also weak depen-
approximate
modelfor the variancefunction(F(T) • O/T,
dence and stationarity.
forlargeT) provides the following expression[Vanmarcke,
Notice that the estimation of 0 is by itself a test of the
1988,p. 337]:
existence of the Hurst effect for stationary processes. In
F*(T)Tn fact, because of the functional central limit theorem, if 0 is
O* = T<n. (14) finite the exponent h is 0.5 and there is no Hurst effect.
- r*(r)]n- T
Otherwise, if 0 is infinite then there is Hurst effect and it is
Anotherpossibleway of estimating 0 is by means of the not necessaryto perform the tests. In fact, for long memory
one-sideunit area spectral density function at zero. Also, if time series the estimation of the scale of fluctuation in (14)
a shortmemory theoretical model is adjusted to the data, 0 does not converge to a finite limit and therefore test 2 is not
couldbe estimated from the model, according to the theo- suitable. For the nonstationary process of Bhattacharya et
reticalexpressionsfor 0 in terms of the parameters of the al. [1983] the estimation of 0 by any of the means presented
model. above also show divergence to infinity, in concordance with
the theoretical result about the existence of the Hurst effect.
3.3.2. Test2 (Outline). Supposethe scaleof fluctuation
0 is not given, estimate it by means of 0* using any of the For those processestest 2 is not applicable either.
methodsdiscussed above, estimate also the size of the In view of the above, the recommendation to deal with an
variance of R*n/(O*n) ø'5 andperforma testin classical observed time series is to proceed first to the estimation of 0
terms. Even though some technical details need to be usingfor instancethe variancefunction approachof (14) (see
workedout, the test is in the same spirit of test 1. For this Figure 3b). Stabilization of the estimator with n indicates
reason,formal substitution of 0* for the scale of fluctuation finite memory and test 2 may be performed. If there is no
intest1 is proposed.The idea is that this will not affectthe stabilizationthere may be three possible causes:the series
powerof the test significantly.Only small modificationsin comesfrom an infinite memory process, or it comes from a
nonstationaryprocess or the length of the record is insuffi-
cient to estimate 0. As an easy cheek, if the value of n (the
length of the record) divided by the estimated 0 is less than,
say, 15, the record is short. In these cases, if possible, the
length of the record should be increased, and the estimation
of 0 repeated. On the other side, there are various ways of
testing and removing nonstationarities.If the problem re-
mains, no conclusion can be inferred from the data alone. In
fact, all extrapolationsof the behavior of either the range or
the estimator of 0 are equally arbitrary from a statistical
point of view, and any decisionshouldbe basedon physical
reasoning.
In a related problem, Burg [1967] observed that the
0.5 problemwith conventionalFourier spectralanalysisof finite
time seriesis that only a finite numberof correlationlagsare
estimableand that the truncation in lag spaceresults in a
0 5• 1•)0 "1• 2(•)0 '2•0 300 smoothingof the true spectralfunctionin frequencyspace.
n
Burg [1967]arguedthat the criterion for extrapolationshould
Fig.lb. GEOSdiagram
fortemperature
incentral
England
series. be to obtainthe spectraldensityestimatethat correspondsto
4000 MESA AND POVEDA:HURSTEFFECT
5-
1--Ea j
1 •4-
0* = -- (15)
2 •3-
1- qbi
i=1
work by Hurst et al. [1965]. Additionally, resultsby Hipel Fig. 2. G EOS diagramfor Lake Saki mud varves.
MESA AND POVEDA: HURST EFFECT 4001
10000• 16
14'
12-
lOOO
10- Im
m c•
6-
100:
lO ' t i i i i
lO •oo ..... i'6'oo.... i 6• i •36ooo 00 3_ • •) • 1'0 1'2' 1'•' 1'6 1'8' 20
TIME(0.1SECONDS) TIME (0.1 SECONDS)
(Thousands)
Fig. 3a. Pox diagramfor a vertical wind velocity time series.
Thedottedline is a least squaresfit with slopeof 0.773,and the solid Fig. 3c. GEOS diagram for a vertical wind velocity time series.
lineisa 0.5slopewith((•r0/2))1/2
intercept.
,,,
REFERENCES
0 1• 260 360 46O 560 600
TIME (0.1SECONDS) Barnard,G. A., Discussionof Hurst, Proc. Inst. Civ. Eng., 5(5),
552-553, 1956.
Fig. 3b. Scale of fluctuation estimation for a vertical wind
Bhattacharya,R. N., V. K. Gupta, and E. C. Waymire,The Hurst
velocity
timeseries.The solidcurveis TF(T), andthedottedcurve effectundertrends, J. Appl. Probab., 20(3), 649-662, 1983.
istheestimator
givenby(14).Unitsofthescaleoffluctuation
arein Billingsley,
P., Convergence
of ProbabilityMeasures,JohnWiley,
0. Is. New York, 1968.
4002 MESA AND POVEDA: HURST EFFECT