Extremum statistics: A framework for data analysis

28 Pages·0.21 MB·

by  Sandra Chapman

2
0
0
2

n
a
J

2
2

]
h
c
e
m
Extremum Statistics: a Framework for Data Analysis.
-
t
a
t
s
.
t S. C. Chapman,G. Rowlands1 and N. W. Watkins2
a
m
1 PhysicsDept. Univ. ofWarwick, CoventryCV4 7AL, UK
-
d 2BritishAntarcticSurvey,High Cross,MadingleyRd., CambridgeCB3 0ET,UK
n
o
c
[

3
v
5
1
Manuscriptsubmittedto
0
6
NonlinearProcesses inGeophysics
0
1
February 1, 2008
0
/
t
a
m
-
d
n
o
c
:
v
i
X
r
a
Abstract
Recent work has suggested that in highly correlated systems, such as sandpiles,
turbulent fluids, ignited trees inforest firesandmagnetization inaferromagnet close
to a critical point, the probability distribution of a global quantity (i.e. total energy
dissipation, magnetization andsoforth)thathasbeennormalizedtothefirsttwomo-
mentsfollowsaspecificnonGaussiancurve. Thiscurvefollowsaformsuggestedby
extremumstatistics, whichisspecifiedbyasingleparametera(a =1corresponds to
theFisher-Tippett TypeI(“Gumbel”)distribution.)
Here, we present a framework for testing for extremal statistics in a global ob-
servable. In any given system, we wish to obtain a in order to distinguish between
the different Fisher Tippett asymptotes, and to compare with the above work. The
normalizationsoftheextremalcurvesareobtainedasafunctionofa. Wefindthatfor
realisticrangesofdata,thevariousextremaldistributionswhennormalizedtothefirst
twomoments aredifficult todistinguish. Inaddition, the convergence tothelimiting
extremal distributions for finite datasets is both slow and varies with the asymptote.
However, whenthethirdmomentisexpressed asafunction ofathisisfound tobea
moresensitivemethod.
1
1 Introduction
The study of systems exhibiting non Gaussian statistics is of considerable current interest (see
e.g. Sornette (2000) and references therein). These statistics are often observed to arise in finite
size many body systems exhibiting correlation over a broad range of scales; leading to emergent
phenomenology suchasselfsimilarityandinsomecasesfractionaldimension(Bohretal.,1998).
The apparent ubiquitous nature of this behavior has led to interest in self organized criticality
(Bak,1997;Jensen,1998)asaparadigm;otherhighlycorrelatedsystemsincludethoseexhibiting
fully developed turbulence. In solar terrestrial physics in particular, problems of interest include
MHD turbulence in the solar wind and in the earth’s magnetotail. Irregular or bursty transport
and energy release in the latter has recently led to complex system approaches such as SOC (see
the review by Chapman and Watkins, Space. Sci. Rev., 2001). These complex systems are often
characterized by a lack of scale, and in particular, by the exponents of the power law probability
distributions (PDF)ofpatches ofactivity inthesystem. Examplesofthese patches ofactivity are
energy dissipated by avalanches in sandpiles, vortices in turbulent fluids, ignited trees in forest
fires and magnetization in a ferromagnet close to the critical point. In the earth’s magnetotail,
patchesofactivityintheauroraasseenbyPOLARUVIhavebeenusedasaproxyfortheenergy
released in bursty magnetotail transport in order to infer its scaling properties (Lui et al., 2000;
Uritsky et al., 2001). The challenge is to distinguish the system from an uncorrelated Gaussian
process, by demonstrating self similarity; and to determine the power law exponents. To do this
directly is nontrivial, requiring measurements of the individual patches or activity events over
manydecades. Hereweconsider whatmaybeamorereadily accessible measure, thestatistics of
aglobalaveragequantity suchasthetotalenergydissipation, magnetization andsoforth.
Animportanthypothesisthatisthesubjectofthispaperisthatthedataarisefromanextremum
process; i.e. thatsomeunknownselection processoperates suchthattheobserved globalquantity
isdominatedbythelargesteventsselectedfromensemblesofindividual‘patches’ofactivity. This
is a real possibility for two reasons. First, measurements of physical systems, and in particular,
observationsofnaturalsystems,inevitablyincorporateinstrumentalthresholdsandthismayaffect
the statistics of a global quantity comprising activity summed over patches. Second, there has
recently been considerable interest in a series of intriguing results from turbulence experiments
(Labbe et al., 1996; Pinton et al., 1999; Bramwell et al., 1998), and numerical models exhibiting
correlations (Bramwellet al. (2000), but see also Ajiand Goldenfeld (2001); Zheng and Trimper
(2001); Bramwell et al. (2001)). These studies reveal statistics of a global quantity (i.e. E) that
2
follows curves that are of the form of one of the limiting extremal distributions: (Gumbel, 1958;
FisherandTippett,1928)
P(E) = K(ey−ey)a
y = b(E s) (1)
−
where K,b and s are obtained by normalizing to the first two moments (M = 1, M = 0,
0 1
M = 1),andthesingleparameteraappearstobeclosetothevalueπ/2.
2
Foraninfinitelylargeensemble, therearetwolimitingdistributions thatweconsider here. The
Fisher-Tippett type I (or ‘Gumbel’) extremal distribution is of the form (1) but with a = 1 and
arisesfromselectingthelargesteventsfromensembleswithdistributionsthatfalloffexponentially
or faster. Since we wish to construct a framework that could encompass all highly correlated
systems we also treat the case where the distribution of ‘patches’ is power law. An example is
thePottsmodel(Cardy,1996)formagnetizationwhereconnectedbondsformclusters,thesizeof
which is powerlaw distributed at thecritical point. In this case the relevant extremal distribution
isFisher-Tippett typeII(or’Frechet’).
Here we provide a framework for comparing data with Fisher Tippett type I and II extremal
curves. This essentially requires obtaining the normalizations of these curves in terms of the
momentsofthedataandultimatelyasfunctions ofthesingleparametera.
Wefindthatthecurvesofform(1)whichareobtainedbynormalizingtothefirsttwomoments
are difficult to distinguish if a is in the range [1,2] or from Frechet curves given a realistic range
ofdata. Furthermorewedemonstratethatslowconvergence withrespecttothesizeofthedataset,
tothe limiting a = 1extremal distribution has the consequence that, for alarge but finite ensem-
ble, the extremal distribution of an uncorrelated Gaussian process is indistinguishable from the
a = π/2 curve. Toovercome these limitations wesuggest two much more sensitive methods for
determining whether or not the curve is of the form (1), and, if so, the corresponding value of a.
Thesemethodsarebasedonthethirdmoment,andthepeakofthedistribution, bothofwhichwe
obtainhereasafunctionofa.
2 Extremumstatistics: generalresults.
To facilitate the work here we first develop some results from extremum statistics (for further
background reading see Sornette (2000); Gumbel (1958); Bouchaud and Potters (2000)). If the
∗
maximum Q drawn from an ensemble of M patches of activity Q with distribution N(Q) is
3
∗ ∗
Q = max Q ,..Q ,thentheprobability distribution (PDF)forQ isgivenby
1 M
{ }
P (Q∗)= MN(Q∗)(1 N (Q∗))M−1 (2)
m >
−
whereM isthenumberofpatchesintheensembleand
∞
∗
N (Q )= N(Q)dQ (3)
>
ZQ∗
Wenow obtain P for large M,Q. Forgeneral PDFN(Q) wecan write(for appropriate choice
m
∗
ofthefunction g(Q )):
(1 N )M = e−Mg(Q∗) (4)
>
−
∗
andforsmallN (Q )wehave
>
N2
g(Q∗) = ln(1 N (Q∗)) N + > (5)
> >
− − ∼ 2
Wenowconsideracharacteristic valueofQ∗,namelyQ˜∗,suchthatbydefinition
Mg(Q˜∗) = q (6)
sothat
N2(Q˜∗)
q = Mg(Q˜∗) MN (Q˜∗)+M > + (7)
>
≈ 2 ···
Wenowexpandg(Q∗)aboutQ˜∗ toobtain
g′′(Q˜∗)
g(Q∗) = g(Q˜∗)+g′(Q˜∗)∆Q∗+ (∆Q∗)2+ (8)
2 ···
andfrom(5)wehave
′ ∗ ∗ ∗
g (Q )= N(Q ) N(Q )N + (9)
>
− − ···
g′′(Q∗)= N′(Q∗) N′(Q∗)N +N2(Q∗)+ (10)
>
− − ···
where g′,g′′ denote differentiation with respect to Q∗, ∆Q∗ = Q∗ Q˜∗, and we have used
−
′ ∗
N = dN /dQ = N. Inverting expansion (7)gives
> > −
MN>(Q˜∗)= q 1 q + M 1 e−Mq (11)
− 2M ··· ≈ −
(cid:20) (cid:21) (cid:16) (cid:17)
∗
Weobtainfrom(5)anditsderivatives withrespecttoQ :
q q 1 q 2
g(Q˜∗) = 1 + +
M − 2M 2 M ···
(cid:18) (cid:19) (cid:18) (cid:19)
q q 3
= +0 (12)
M M
(cid:18) (cid:19)
4
whichtorelevantorderisconsistent with(6),and
q
g′(Q˜∗)= N(Q˜∗) 1+ + (13)
− M ···
(cid:20) (cid:21)
Forq finiteasM thisgivesg′(Q˜∗) = N(Q˜∗)andMN (Q˜∗) =q.
>
→ ∞ −
Wecan now consider the extremal statistics of specific PDFN(Q), and importantly show that
∗
P (Q )canbewrittenintheuniversalform(1).
m
2.1 GaussianandExponential N(Q)
If N(Q) falls off sufficiently fast in Q, i.e. is Gaussian or exponential it is sufficient to consider
∗
lowest order only in (5) giving g(Q ) N (Gumbel, 1958; Bouchaud and Mezard, 1997) and
>
∼
q = MN (Q˜∗). Expanding (3)inQ∗ nearQ˜∗ givestothisorder:
>
∞
MN (Q∗) = M N(Q)dQ MN(Q˜∗)∆Q∗
>
ZQ˜∗ −
= q 1 MN(Q˜∗)∆Q∗+ qe−MN(Qq˜∗)∆Q∗ (14)
" − q ···# ≈
∗
Expanding N(Q)aboutQ yields
N′(Q˜∗)
N(Q∗)= N(Q˜∗) 1+ ∆Q∗+
" N(Q˜∗) ···#
N(Q˜∗)eNN′((QQ˜˜∗∗))∆Q∗ (15)
≈
Astothisorder(1 N )M−1 e−MN> wethenhavefrom(2)
>
− ≈
P (Q∗)= MN(Q∗)(1 N (Q∗))M−1 MN(Q∗)e−MN>
m >
− ≈
(eu−eu)a (16)
∼
with
N′(Q˜∗)N (Q˜∗)
>
a= (17)
− N2(Q˜∗)
and
MN (Q˜∗) N(Q˜∗)
> ∗
u= ln ∆Q (18)
a !− N>(Q˜∗)
SincethroughoutweareconsideringQ˜∗large(M ,qfinite)wehavetheeffectivevalueof
→ ∞
aasthatgivenby(17)inthelimitQ˜∗ . ForN(Q)exponential theabovegivesa= 1. Inthe
→ ∞
5
particularcaseoftheexponentialallthesummationswhichintheabovewehavetruncatedcanbe
resummedexactlyandgivea 1,recovering theresultofBouchaudandMezard(1997).
≡
ForN(Q) Gaussian wecannot obtain a exactly in this way but as we shall see it is instructive
tomakeanestimate. GivenN(Q)= N exp( λQ2)andexpandingequations(14),(15)and(16)
0
−
tonextorderweobtain
P = P¯ eR(u)
m m
ln2(q) 2ln(q) u¯2
R = +u¯ 1+ eu¯ (19)
−4λQ˜∗2 4λQ˜∗2!− 4λQ˜∗2 −
where we have used u = 2λQ˜∗∆Q∗ and u¯ = u + ln(q). To lowest order in ∆Q∗/Q˜∗ (i.e.
−
Q˜∗ )wehaveauniversalPDFwitha = 1,buttonextorder,thatis,neglectingonlytheterm
→ ∞
inu¯2 in(19)wehaveauniversal distribution ofform(1,16)with
2ln(q)
a 1+ = 1 (20)
≡ 4λQ˜∗2! 6
2.2 PowerlawN(Q)
ThePDFofpatchesN(Q)mayhoweverbeapowerlawandinthiscaseitwillfalloffsufficiently
slowlywithQthatweneedtogotonextorderasin(7). IfweconsideranormalizablesourcePDF
N
0
N(Q) = (21)
(1+Q2)k
thenforlargeQ(Q >>1)wehaveN(Q) N /Q2k andthenusing(3)and(7)
0
∼
q q
Q˜∗N(Q˜∗) =(2k 1)N (Q˜∗) = (2k 1) (1 ) (22)
>
− − M − 2M
which with the above general expressions for g(Q˜∗) and its derivatives substituted into (8) gives
∗
anexpression forg(Q )
∗ ∗
q ∆Q ∆Q
g(Q∗) = 1 (2k 1) +k(2k 1)( )2 (23)
M " − − Q˜∗ − Q˜∗ ···#
WealsorequireanexpressionforN(Q∗),againexpandingaboutQ˜∗andobtainingthederivatives
ofN(Q˜∗)fromthoseofg(Q˜∗)andvia(11)gives
∗ ∗
∆Q ∆Q
N(Q∗)= N(Q˜∗) 1 2k +k(2k+1)( )2 (24)
" − Q˜∗ Q˜∗ #
whichcanberearranged as
∗ ∗
N(Q∗)= N(Q˜∗)e −2k∆Q˜Q∗ +k(∆Q˜Q∗ )2 (25)
h i
6
Aftersomealgebra(23)canberearranged togive
Mg(Q∗) = qe −(2k−1)∆Q˜Q∗∗+2k2−1(∆Q˜Q∗∗)2 (26)
h i
Thesetwoexpressions combinetofinallygive
P (Q) P (Q∗) (eu¯−eu¯)a (27)
m m
≡ ∼
with
∗ ∗
∆Q ∆Q
u¯ = ln(a) ln(q) (2k 1) (1 ) (28)
− − − − Q˜∗ − 2Q˜∗
and
2k
a= (29)
2k 1
−
Tolowestorder, neglecting the(∆Q∗/Q˜∗)2 term(28)reducesto(18).
Hence a power law PDF has maximal statistics P (Q) which, when evaluated to next order,
m
can be written in the form of a universal curve (i.e. of form (1,16)) with a correction that is non
negligible atthe asymptotes. This canbe seen (Jenkinson, 1955; Bouchaud and Potters, 2000) to
beconsistentwiththewellknownresultduetoFrechetwhere(followingthenotationofBouchaud
andPotters(2000)) ifwehavePDF
1
N(x) (30)
∼ x 1+µ
| |
then
1
N (31)
> ∼ xµ
Pm(x∗)= (x∗µ)1+µe−(x∗1)µ
whichwecanwriteintheform
Pm(x∗)= µeµ+µ1ln(µ+µ1)(eu−eu)a (32)
µ+1
∗
u= µln(x ) ln (33)
− − µ
(cid:18) (cid:19)
whichisofuniversal form (1,16) inu. Noting thathere µ = 2k 1anda = (µ+1)/µ and that
−
tosecond order
∗ ∗ ∗
∆Q ∆Q ∆Q
(1 )= ln 1+ (34)
Q˜∗ − 2Q˜∗ Q˜∗ !
wesimply identify 1+∆Q∗/Q˜∗ withx˜∗ toobtain (28). Tonext order in∆Q∗/Q˜∗ the analogue
of(28)stillyieldstherighthandsideof(34).
7
2.3 Convergence tothelimitingdistributions
TheaboveresultsshouldbecontrastedwiththederivationofFisherandTippett(FisherandTippett
, 1928). Central to (Fisher and Tippett , 1928) and later derivations is that a single ensemble of
NM patcheshasthesamestatisticsastheN ensembles(ofM patches),ofwhichitiscomprised.
The fixed point of the resulting functional equation (Bhavsar and Barrow, 1985) for arbitrarily
largeN andM isa = 1fortheexponentialandGaussianPDF,andtheFrechetresultforpowerlaw
PDF.Here,weconsider afinitesizedsystemsothatalthough thenumberofrealizable ensembles
of the system can be taken arbitrarily large, the number of patches M per ensemble is always
large but finite. Importantly, the rate of convergence with M depends on the PDF N(Q). For
an exponential or power law PDF we are able to resum the above expansion exactly to obtain
a; and convergence will then just depend on terms O(1/M) and above. This procedure is not
possible for N(Q) Gaussian, instead we consider the characteristic Q∗, that is Q˜∗ which for M
arbitrarilylargeshouldbelargealso. Rearranging(7)tolowestorderforN(Q) = N exp( λQ2)
0
−
yields √λQ˜∗ ln(M)implying significantly slowerconvergence. Thisisfurther discussed in
∼
Sornette(2000)(ppp. 19-21).
The extremal distributions are thus essentially a family of curves that are approximately of
universalform(1,16)andareasymmetricwithahandednessthatjustdependsonthesignofQ;we
haveassumedQpositivewhereasonecouldchooseQnegativeinwhichcaseN(Q) N( Q ).
→ | |
This would correspond to, say, power absorbed, rather than emitted, from a system. The single
parameter a that distinguishes the extremal PDF then just depends on the PDF of the individual
events. For N(Q) exponential we then recover exactly the well known result (Gumbel, 1958;
Bouchaud and Mezard, 1997) a = 1. For a power law PDF a is determined by k via (29). We
havealsodemonstratedthatforaGaussianPDFwithfinitebutlargeM andN thata= 1andwill
6
explorethesignificance ofthisinSection3.1.
3 Normalization tothefirsttwomoments
To compare these curves with data we need P(Q¯) P (Q∗) in normalized form. This has
m
≡
moments
∞
M = ynP¯(y)dy (35)
n
−∞
Z
whichwewillobtainasafunction ofaandtheninsistthatM = 1,M = 0andM = 1.
0 1 2
8
Setting M = 0 (and M = 1, M = 1) in our analysis of extremal distributions does not
1 0 2
requireanyassumptionsabouttheformofthePDFexceptthatthemomentsexist. Itwillallowus
towritetheanalytically obtained extremaldistributions asfunctions ofsingleparameter a.
3.1 Extremaldistributions arisingfromGaussianandexponential N(Q)
ForGaussianandexponential PDFwehave
P¯(y) = K(eu−eu)a (36)
u= b(y s) (37)
−
Thishasmomentswhichconverge foralln. FromAppendixAwehavethatthenth moment:
1 ∞ [ln(a)+bs η]n dn
M = P¯(y)dη − = Ke−aln(a) Γ(a) (38)
n b −∞ bn dan
Z
whereη = ln(a) u.
−
To normalize we insist that M = 1, M = 0 and M = 1. The necessary integrals can be
0 1 2
expressed in terms of derivatives of the Gamma function Γ(a) (Gradshteyn and Ryzhik (1980))
andweobtaininAppendixA:
b2 = Ψ′(a)
b
K = ealn(a) (39)
Γ(a)
(Ψ(a) ln(a))
s = −
− b
where
1 dΓ(a)
Ψ(a) =
Γ(a) da
dΨ
′
Ψ(a) =
da
The ambiguity in the sign of b (and hence s) corresponds to the two solutions for P(Q¯) for
positiveandnegativeQ.
Wecannowplotthecurves,thatis,normalizedtothefirsttwomomentsandtheseareshownin
Figure1. Experimental measurements ofaglobal PDFP(E)normalized toM wouldbeplotted
0
M P versus(E M )/M . Inthemainplotweshownormalizeddistributions oftheform(1,16)
2 1 2
−
for a = 1,π/2 and 2. It is immediately apparent that the curves are difficult to distinguish over
severaldecadesinP¯(y)andthustoobtainagoodestimatefora,thenumericalorrealexperiments
9