Fa oltà di S ienze Matemati he,
Fisi he e Naturali
Corso di Laurea Spe ialisti a in S ienze Fisi he
Anno a ademi o 2010/2011
Tesi di Laurea Spe ialisti a
Chaos Characterization of Pulse-Coupled Neural
Networks in Balanced State
Candidato
Relatore
Esterno
Relatore
Interno
grazieaVoi,so he sidà
Introdu tion 7 ComplexNetworks . . . 7 NeuralNetworks . . . 8 Ergodi theory . . . 10 ThesisStru ture . . . 12 1 Mathemati al Framework 15 1.1 DierentiableDynami alSystemsandErgodi ity . . . 15
1.2 Chaoti Dynami sand LyapunovExponents. . . 18
1.3 Cal ulatingLyapunovExponents . . . 21
1.4 Cal ulatingthe(lo al)Lyapunovve tors. . . 24
1.5 MatrixCal ulation . . . 25
1.6 InformationandEntropy . . . 27
1.7 InformationDimensionoftheAttra tor . . . 28
2 Model 31 2.1 SingleNeuronDynami s . . . 31
2.2 NetworkDynami s . . . 37
2.3 TheBalan edState . . . 38
2.4 ModelAr hite ture. . . 40
2.5 EquationsforaQIFNetwork . . . 42
2.6 Cal ulatingtheJa obian. . . 43
3 Computation and Results 47 3.1 NetworkComputation . . . 47
3.2 OutputData . . . 49
3.3 SpikeTrain . . . 50
3.4 Colle tiveDynami s . . . 53
3.5 Attra torDimensionandEntropy . . . 55
3.6 LyapunovVe torsConvergen e . . . 56
3.7 AnglesandHyperboli ity . . . 61
3.8 Parti ipationRatioandChaosIndex . . . 64
4 Dis ussion 71
4.1 Con lusions . . . 71
4.2 Extensions . . . 72
A Temporal Flu tuationsin Balan ed State 75
Fromthersttheoriesofmatteras onstitutedbyelementaryparti les[
Dem-o ritus( a. 460BC a. 370BC)℄uptothetwentieth entury,thedominating
trendinphysi al,and,moregenerally,ins ienti resear hhasbeen
redu tion-ism: the prin iplethat any omplexsystem anbede omposedin elementary
parts,whi harethenstudiedsingularly,sothatthefullknowledgeoftheglobal
phenomenon is expressed in terms of the elementary intera tions between its
simpleentities.
Su h approa h hasprovided, and still provides, marvellous insights, ru ial
fortherealizationofmostofthes ien eand thete hnology weseearoundus.
Nevertheless,itprovesnotsu ientwhenit omestodealingwiththestudyof
theemergent phenomena.
Emergen e anbedened,innaturalandso ials ien e,asthearisingofnovel
and oherent stru tures,patterns andpropertiesin omplexsystems. In
appar-ent ontradi tion tothethirdprin ipleof thermodynami s, wearesurrounded
byemergent,highlyorderedstru tures,lifeitselfbeingthemostastonishing
ex-ampleofallofthem. Unfortunately,aredu tionistapproa hdoesnotproveto
besu ient,orevenapt,forthestudy ofsu hphenomena. Thenew hallenge
fors ien eandresear histondmodelsandrulesforemergentphenomena,and
applythem toreality. Complexnetworks aretypi almodelsshowingthearise
of omplexbahviourand olle tivephenomenafromanaggregationofrelatively
simple onstituents.
Complex Networks
Althoughnetworks anreveal ari h variety ofbehaviours, andemergent
phe-nomena,theirdenitionisquitesimple:anetworkisasetofitems,usually alled
nodes,with onne tionsbetweenthem,referredtoasedges [1℄. Thetopologi al
studyofnetworksstartedwithEuler,andhis solutionto theKönisberg Bridge
Problem, whi h laiddownthefoundationsof themathemati alGraph Theory.
Sin ethen,extensivestudiesonahighvarietyofnetworksandonthedynami s
taking pla e on them have been ondu ted, and now many s ienti models
relyon network stru tures; just to name afew: food webs, geneti networks,
powergrids, modelsfor the spreadingof omputer virusesorhumandiseases,
et . [2℄. A surprising number of examples on the importan e of networks in
s ien e,te hnology, so ietyandevery-daylife anbefound in apopularbook
Neuralnetwork onstitutea lassof omplexnetworkinspiredbythestru ture
of the nervous system. The nodes are identied with neurons, the edges
rep-resentthelinks betweenthem (axon andsynapses)andthe ex hangedsignals
aretheele tri alpeaks,orspikes,(a tionpotentials)triggeredwhenthe
inter-nalpotentialofthe neuronrea hesa ertain threshold. Theywere introdu ed
mainlytosolvearti ialintelligen eproblem,su hasautomatedlearning,visual
pattern re ognition,adaptivity, fault toleran e,et . orasmodels fornetworks
ofrealnervous ells[4℄.
Inthepresentwork,we onsiderthelatter ategoryofneuralnetworks,
aim-ingat abetterunderstandingand modelling ofreal biologi alstru tures. The
rst exampleof su h approa h datesba kin 1975, when Peskin modelled the
pea emaker ellsof the heart asa fully onne ted network of identi al leaky
integrate-and-reneurons(LIF),andveriedthattheyspontaneously
syn hro-nisetheirsignals,asrealpea emaker ellsdo[5℄. TheLIFmodelisthesimplest
onedimensionalmodelfortheele tri al a tivityof aneuron. Theuseofmore
omplex and rened models would add more biophysi al value to the
inves-tigation, but therequired omputations would in reasedramati ally with the
networksize;moreoverphenomenaemergingfromthe olle tivedynami softhe
systemshouldnotdepend toomu honthepre ise andrealisti re onstru tion
ofitssingle omponents.
Thespe i systemweareinterestedinisnotthehearth,butthepyramidal
neuronsoftheneo ortexintheirstationarystate. Manyexperimentsonliving
ells,bothinvitroandin vitro,showedthat,althoughverydense,anddensely
onne ted(
∼ 10
6
neuronsonamm3
,with∼ 10
4
onne tionsea h),theyexhibit
verylowaverageringrates(
∼ 1
Hz),withaspikepatternveryvariableintime, andveryweakly orrelated,evenwhen sampledfromneurons loseea hother;atthesametime,theresponseintimeseemssurprisinglyfastwhen omparedto
rea tiontimes forisolated,individualneurons. Thesystemisthereforedoubly
e ient: thelowspikingrateredu estheenergy onsumption,but,atthesame
time, itisableto rea teven fasterthananyofitsisolated omponentswould.
Admittedly,brainsstilldetainthere ordofmoste ientand omplexparallel
al ulatorsintheknownuniverse.
Goingba ktothetheneo ortex,themodelusedtoexplaintheobserved
a -tivity is the balan ed statemodel. Namely ea h neuron, for most of thetime,
is kept lose to its ring threshold by a balan e between ex itatory and
in-hibitoryinputsfromotherneurons,sothatevenverysmalldeviationsresultin
aqui kresponsebytheinterestedneuron. Dierentexperimentsandmeasures
onrmed this theory (see se tion 2.3 ). Our aim isnow to onstru ta
neu-ralnetworkwhi hrepli ates,atleast qualitatively,this kindof behaviour,and
studyit.
The reation of a neural network in stationary balan ed state requires just
themixingoftherightingredients,namely:
a) apropermodelforsingleneurondynami s, assimpleaspossible;
b) averylargenumberofneurons,andnumberof onne tionsbetweenthem
still verylargebut smallwhen omparedto thesize ofthesystem, soto
graph. Theimageshows
∼
60neuronswithanaverageof1.5
onne tionsea h. The simulatednetworksrangefrom200
to1000
neuronswith20 − 50
average onne tions. Theorientationofthe onne tionsisnotshown.) bothex itation andinhibitiona tingonneurons.
The rst hoi e for point a) would be the LIF neuron, however they have
been ex luded for an intrinsi aw in their dynami s: when the membrane
potentialofarealneuron rossesa ertainthreshold,apositive-feedba kpro ess
furtherly in reases it, produ ing a peaked signal; LIF neurons, on the other
end, are arti ially reset in the instant they tou h theirthreshold. With this
qualitativedieren e, LIF neural networks have thetenden y to syn hronize.
Innetworksmadeof inhibitoryneuronssubje tedtopositiveexternal urrents
thisphenomenon hasbeenprovedanalyti ally [6℄. Forthat reasonthe hosen
model is thesimples anoni al 1D neuron model: thequadrati integrateand
redes ribedin hapter2.
Conditionb) is obtained by onne tinga large numberof neurons with an
Erd®s-Rény random graph stru ture, as represented in gure 1. If
N
is the totalnumberof neuronsandK
theaverage onne tionsper neuron,our ideal onditionswould beN ≫ K ≫ 1
. In thea tual modelN
rangesfrom200
to1000
, whileK from20
to100
. Finally all onne tionsareassumed to be one-dire tional,asforrealneurons. Forgraphi alreasonstheguredoesnotshowonne tionorientations,andtheparametersareredu edto
N ≈ 60
,K ≈ 1.5
. Asregardsthepoint ), it hasbeenfound thatanetworkof onlyinhibitoryneurons, subje ted to a onstant, ex itatory external urrent, su es to full
it. This simpliesthe model andthe al ulations,sothat we an on entrate
onmoreessentialparameters.
The system dened with those requisites shows deterministi haos. For a
un-expansionis odedbytherstLyapunovve tor.
derstanding, we rely onthe mathemati al framework typi ally used for large,
omplex,dynami alsystems: theErgodi Theory.
Ergodi theory
Sin e hapter 1 ontainsmostoftheformalism,the orre tdenitions andthe
mathemati alrigourne essarytodeneanddes ribetheergodi propertiesand
quantitiesweareinterestedin,inthisse tionwegiveamoreintuitive(andless
formally orre t)des riptionofwhat ergodi ityisabout.
Givenanite-dimensional,deterministi dynami alsystem,itsphasespa eis
dened as then-dimensional manifoldin whi h everysingle pointfully
repre-sentstheexa tstate of thesystemat agiventime. As the systemevolves, it
drawsatraje torythephasespa ewhi h annotinterse titself, ortraje tories
asso iatedto othertime evolutions.
Thebasi ideaof ergodi theoryis that, for su ientextents oftime,
aver-agingovertimeanyquantityasso iatedtothedynami alsystemis ompletely
equivalentto performinganaverageonthewhole phasespa e,aslongassu h
manifoldisweightedwithanergodi measure. Intuitively,su hameasureshould
negle ttheareasofphasespa enever(oralmostnever)tou hedintheevolution
of oursystem,and givemoreimportan e to thosein whi h thesystemlingers
mu hmorein time(theattra tingmanifold).
Withthisapproa hwe anndquantitiesthatareglobalandinvariantforthe
system just followinga single time evolutionstarting at arandom point: the
information gained from su h traje tory would beequivalent to studying the
systemasawhole,ortowhatwewouldndusinganyothertraje torystarting
fromdierentpoints;aslongas,of ourse,thesysteminquestionisergodi . One
ofthemost ru ialquantitiesforthe hara terisationofadynami alsystemis
therst ovariantLyapunovexponent,asitisstri tlyboundwiththedenition
ofdeterministi haos.
The ideaof a dynami al systemboth deterministi and haoti , - i.e. with
a time evolution ompletely and uniquely determined by its variables on one
hand, but omplexand not predi table in the longterm on the other hand
-is only an apparent paradox, easily resolved stating that deterministi haos
d(t
0
) = δ ≪ 1
, evolve in time on paths diverging with an exponential rated(t) ≃ δe
λ
1
t
,
λ
1
> 0
(see gure 2) . The result is that, sin e we an know thestate of asystemonly with nite pre ision, any longterm fore astwouldeventuallybeoutgrownbytheexponentiallyin reasingerrorasso iatedto our
initial un ertainty. This is basi ally what Lorentz originally intended as the
butteryee t.
The rst Lyapunov exponent is, roughly speaking, the
δ → 0
limit of the exponent asso iated tothat divergen e, averagedfortimet → ∞
. If positive, thesystemis learly haoti . Ifnegative,we ansaythat separatetraje torieswouldeventually onverge,althoughforaparti ular lassofdynami alsystems
thetimeneededgrowssofastwiththesystemdimensionalitythatthedynami s
appearpra ti ally haoti . Su h propertyis alledstable haos [7℄.
Going ba k to gure 2, wenoti e that the maximum divergen e denes, in
time,alsoadire tion. We allthedire tionasso iatedtothemaximum
expan-siontherstLyapunovve tor
v
(1)
(t)
. Itislo al,sin eitdepends onthepoint
inphasespa eweare onsidering.
Apartfrompoints,we an onsidern-dimensionalphasespa evolumes. This
introdu estheideaofdierentordersofLyapunovexponents,aswellas
dier-entasso iateddire tions: a
k
dimensionalhyper ubewouldin rease(orshrink) itssizeasV (t) ≃ V (t
0
) exp (t (λ
1
+ . . . + λ
k
))
,andthedeformationofitsshape wouldfollow,intime,thedire tionsoftheasso iatedLyapunovve tors. Figure3representsanexampleofthis: therstdire tionisexpanding,these ond
or-respondstoa0exponent,nallythethirdis ontra ting. Thevolumeexpands
asthesumof thethree exponents,and itsshape hanges: weexpe t that the
dire tion asso iatedto the 0exponentkeepsbeing parallelto the motion,but
nothing anbesaid,ingeneral,abouttheexpandingand ontra tingdire tions,
andtheanglebetweenthem. Thesystemsweare interestedin aredissipative:
thesumofallLyapunovexponentsisnegative,sothatavolumeinphasespa e
shrinksduringthedynami alevolution,progressivelyfalling onasubsetofthe
totalphasespa emanifold alledattra tor.
The attra tor has, in general, a very omplex shape, possibly fra tal. A
usefulpropertyoftheexponentsisthatthey angiveanesteemoftheattra tor
dimension. Ifthe rst exponent is positive, but the totalsum isnegative,for
a ertaininteger
k
we haveλ
1
+ . . . + λ
k
≃ 0
(if thesumisnot exa tlyzero, we anaddasmallnon-integer orre tiontok
). Then, forwhat statedabove, avolume ofdimensionk
will neitherexpand nor ontra tin time, thus giving anupper bound totheattra tordimension.Anotheresteem derived from the exponents is the entropy produ tionrate.
Assuming we know the initial state of the system with nite pre ision, the
exponentialspreadingoftraje toriesfrompointsinitiallyindistinguishableadds
informationregardingtheirinitialstate,thus haoti dynami alsystems anbe
seenasprodu ersofinformation, andtheprodu tionrate anbeestimatedas
thesumofthepositiveLyapunovexponents.
The al ulationsofentropyandattra tordimensionarea tuallyesteemsthat,
in general, annot be taken as exa t equalities. There is, however, a lass of
dynami alsystems, alledaxiom-A,forwiththeentropy al ulationisprovedto
be orre t,whiletheattra tordimensionesteemis onje turedtobe. Estimating
ifa omplex, haoti dynami alsystemisaxiom-Aornotis mathemati allya
another, and shrinksinthe third:
λ
> 0
,λ
= 0
,λ
< 0
. Thedire tions are givenbytheLyapunovve torsasso iatedtotheexponents.thepresentworkisinvestigateifthesystemweareinterestedinsatisesoneof
thekeyrequisitesforaxiom-A:hyperboli ity.
Basi ally,inhyperboli alsystemsexpandingand ontra tingdire tionsdono
mixup(seetheendof hapter1forfurtherdetails. Inthepresentworkwe he k
if this requisite is respe ted by studying the angles between ontra tingand
expandingdire tions,givenbytheLyapunovve torsasso iatedto,respe tively,
positiveornegativeglobalexponents.
Thesis Stru ture
The aim of the present work is to simulate and, using the ergodi formalism,
fully hara terisealarge-s aleneuralnetworkofinhibitoryquadrati
integrate-and-reneurons.
Chapter1introdu esandexplainstheergodi theorymoreformally.
Quanti-tiessu hasattra tordimension,Lyapunovexponents,entropyprodu tionrate
andLyapunovve torsaredened,explainingalsohowthey anbenumeri ally
omputed.
Thefollowing hapterdealswiththe onstru tionoftheneuralnetwork.First
wemodel theone-neurondynami s, taking thegeneral Hodgkin-Huxley
equa-tionsasastartingpoint. Troughdierentlevelsofapproximationwerea hthe
simplest anoni al1Ddes riptionforaneuron: thequadrati integrate-and-re
ortheta neuron. Afterwards,starting from the resultsof in-vivomeasures on
largepopulationsofpyramidalneurons,wedenethebalan edstate,andbuild
a model for the omputation of a large-s alenetwork with similar qualitative
behaviour.
Chapter 3 deals with the omputation and the ndings. It starts with a
step-by-stepdes ription of the omputer simulation we used. Then we assess
them theentropy produ tion rateand the attra tor dimension. The ndings,
sofar, aresubstantiallya onrmation ofwhat has alreadybeenpresentedin
re ent works on the subje t [8, 9℄. We onsequently fo us on the Lyapunov
ve tors: to onrm their invarian eand the robustnessof the algorithm used
for their al ulation, dierent onvergen e tests are performed. The minimal
anglesbetweenve tors, orrespondingtoexpandingand ontra tingdire tions
are usedto estimate thehyperboli ity ofthe dynami alsystem. Theresultis
that,asthesystembe omeslarger,theangledistributionismorepeakedfora
nonzerovalue,showingmoretransversalitybetween ontra tingandexpanding
dire tions.
Finally, for ea h ve tor, we measure the average parti ipation ratio, that
ounts the ee tive number of neurons ontributing to the ve tor dynami s,
andthe haos index (dened asthetime averageofthe squareve tor
ompo-nents),that revealswhi h neuronshaveapredominantroleforasingleve tor.
From the interplay between these two parameters (and some others derived
from them)we an hara terizethe network dynami sbothglobally andfrom
the perspe tive of the single neurons. In parti ular, we nd that the
ontri-bution to the expanding dire tions omes from a group with an average size
thats aleswith
K
,andthesingleneuronsthat takeparttoit hange intime, overinguniformlymostofthenetwork;ontheotherhand,strongly ontra tingdire tionstendtobelo alizedonfewneurons,xedin time.
The nal part summarises the novel results, namely the use of Lyapunov
ve tors to assess the hiperboli ity of the neural network and to hara terise
the role of individual neurons in the olle tive dynami s. A list of possible
Mathemati al Framework
Westatedthatasetofinter onne tedneurons anbemodeledasanetworkand
studied asa
n
-dimensional dynami alsystem. Inthis hapter wewill des ribe thegeneralmathemati alframeworkusedtostudysystemsofknowndynami s,hara terizedby haoti behaviorandhigh-dimensionality.
We start with an introdu tion about the statisti al study of dierentiable
haoti dynami alsystems,throughthedenitionofanatural orphysi al
prob-abilisti measureonphase-spa eandtheappli ation ofBirkho'sErgodi
the-orem.
Afterwardswefo usonthe hara terizationof haos,deningbothintuitively
andinamorerigorouswaytheLyapunovexponents.WeshowhowtheOselede
Ergodi theorem impliesthatsu h exponentsexist and arenite, andhowthe
presen eofaOselede splitting ofphase-spa eemergesfromthat.
Usingthe givendenitions, weillustrate the lassi alalgorithmused to
al- ulatethe ompleteset ofLyapunovexponents,whentheequationsregulating
thedynami sarefully known;then weintrodu eaveryre entmethod forthe
al ulationofave torbase orrespondingtotheexpandingdire tionofthe
dy-nami s,i.e. thelo al Lyapunovve tors. Themeaningofsu h ve torsandtheir
onne tionwith thelo al Lyapunovexponents isbrieyexplained.
Inthelastpartwedes ribetwo onje turesofErgodi theorywhi hinvestigate
the onne tion between the Lyapunov spe trum the attra tordimension and
entropy.
Allquantities des ribedinthis hapterwill beexpli itly al ulatedand
ana-lyzedin thespe i dynami almodeldes ribedinthenext hapter.
1.1 Dierentiable Dynami al Systems
and Ergodi ity
A dierentiable dynami alsystem is a time evolutionon a ompa t,
dieren-tiablemanifold
M ⊆ R
m
(thephasespa e),denedbyadierentiablemapping
orow[10℄
f
t
: M → M ,
Figure1.1: Exponentiallydivergingtraje toriesinphasespa e
withtheproperties:
f
0
=
identity andf
s
f
t
= f
s+t
.
(1.1)
Ifthesystemisdissipative,agivenportionofphasespa eisusually ontra ted
bythetimeevolutiononasmallervolume. Theportionofphasespa ewherethe
motionis on entrated,possiblyafter aninitial transient,is alled attra tor.
We andenetheattra torasaset
A
withthefollowingproperties[11,12℄:1. is invariantinthedynami alevolution:
∀t
,f
t
A = A
;
2. attra ts an open set:
∃
opensetU ⊃ A
su hthat∀x ∈ U
thedistan e betweenA
andf
t
x
redu es tozeroin thelimit
t → ∞
.3. is minimal:there isnosubsetof
A
whi hsatisesproperties1and2.Thelargest
U
satisfyingproperty2is alled basinofattra tion ofA,ifU = M
theattra toris alleduniversal.Adynami alsystemissaidtobe haoti whenitpresentshighsensitivityto
initial onditions,i.e. whenthetraje toriesoftwodistin t phasespa e points,
arbitrarilynear at
t = 0
, divergeexponentiallyduring time evolution,as illus-trated in gure 1.1. In the next se tion we elaborate the on ept in greaterdetailandquantitatively;fornowwe annoti ethat,foragivenstartingpoint,
traje tories al ulated on ma hines with slightly dierent pre isionsor
dier-ent round-omethods would ompletely diverge after a relatively short time.
Unlessweusenumbersofinnitepre ision,there willalwaysbesomeintrinsi
randomnoisethat ompromisesanylong-termfore astofsu hsystems.
Never-thelessglobalstatisti alproperties,su hasthepresen eorthestru tureofthe
attra tor,stay un hanged, regardlessthestarting point and thelevelof noise
(assuming, of ourse, that the latter is reasonablysmall). In haoti systems
theattra toristherefore hara terizedbytheadditionalproperty:
4. is stableunder smallrandom perturbations.
Su h aproperty is essential to guarantee that, in experiments and numeri al
simulations, the motion falls asymptoti ally on the attra tor despite the fa t
The attra tors of haoti dynami al system are alled strange. The name
omes fromthe fa t that theyoftenpresentavery omplex, fra talstru ture,
i.e. withnon-integerHausdor dimension. The denitionof fra talsand their
properties is beyond the s ope of the present work (see for example hap 11
in [12℄ or hapter 3 in [13℄) here we just mention that similar attra tors are
quitehard to model andto study dire tly. This, along with the impossibility
of al ulatingthe true evolutionof apointin phasespa e, suggeststhat the
onlypossibleapproa htota klethoseproblemsisstatisti al andprobabilisti .
ThetoolweuseforouranalysisistheErgodi theory. Itbasi allysaysthat
av-eragesonasingletraje toryintimeequalaveragesoverthewholephase-spa e,
wherethephase-spa eisweightedbyanappropriatemeasure
µ
,withthe funda-mentalrequisitesofbeinginvariantundertimeevolutionandergodi . Abstra tergodi theory dealsalot with thestudy and denition of measures that an
satisfythoserequisites. Asphysi ists,hopefullyinterestedin realdierentiable
dynami alsystem,we anlu kilybypassthisproblembyoperationallydene
auniquenatural orphysi al measurein thewaydes ribedbelow.
Westartbytakinginto onsiderationtheprobabilitydensityfun tions(PDF)
on
M
, dened so that, forx
∈ M
,ρ(x) dx
is the probability of nding the system in the small volume of phase spa edx
around the pointx
. Imagine to have, fort = 0
a given ensemble of phase-spa e points, orresponding to a densityρ
0
(x)
, then evolve ea h point in time withf
, and study the series of pdfsρ
1
(x), ρ
2
(x), . . . ρ
t
(x)
. In thet → ∞
limit, we may expe t that they onvergetoadensitywhi hisinvariantunderthea tionofdynami s:ρ
inv
(x)
.This is not always true, but in the systems of our on ern, i.e. haoti and
nite-dimensional, the appli ation of Perron-Frobenius theorem ensures that
su haPDFexists,isunique,andisapproa hedexponentiallyfastintime[14℄:
ρ
t
(x) = ρ
inv(x) + O(e
−α t
)
Intuitively, if webegin with a homogeneousdistribution of points in phase
spa e,andevolvethemforsometime,allpointswouldfallovertheattra torand
stay onnedonit. Underapra ti alpointof view,unstable xedpointsand
y lesdonotplayanyrole,sin epointsonthemaredrivenawaybytherandom
noiseintrinsi tonumeri al omputation. Tobemathemati allymorerigorous,
we ouldfollow theidea expressed byE kmann and Ruelle [10℄ (said to have
been rst formulated by Kolmogorov), and dene ourdensity asymptoti ally,
asthe
ǫ → 0
limitofdensitiesρ
ǫ,
invhara terizedbyadynami s perturbed by
randomnoiseofmagnitude
ǫ
.Thephysi al measure
µ
anbedenedastheprobabilityofndingthemotion inagivenphase-spa earea:µ(B) =
Z
B
ρ
inv(x) dx .
Fromtheway
ρ
inv(x)
hasbeen onstru ted,thismeasurehasthe ru ial
prop-ertyofbeinginvariant underdynami alevolution:
µ(B) = µ(f
−t
B) .
Afterndinganatural andidatefortheinvariantmeasure,we andenethe
Theinvariantmeasure
µ
issaidtobeergodi ifthephasespa eM
ismetri allyinde omposablewithrespe ttoit;i.e. there annot betwodistin tandinvariantsubsets,A
andB
,bothwithpositiveµ
measure. Inotherterms,ifA
isinvariant(f
t
A = A
),then
µ(A) = 1
or0
.We an imagine that in systems with distin t attra tors, the dynami
evolu-tion would sele t one of them, a ording to the spe i starting point, and
moveeternallyoverit. Asa onsequen e,thestatisti alpropertiesderivedfrom
su htraje torywouldbeneitherglobalnorinvariantofthestarting ondition,
invalidatingthenotionofergodi ity.
Ifsu hasimpleandintuitiverequisiteisalreadyenoughtostatethepresen e
of ergodi ity, in pra ti e the ergodi hypothesis is impossible to demonstrate
for the great majority of systems. Howeverwe anmake the reasonable
as-sumptionthatthesystemweare onsideringhasonlyoneattra tor,andgoon
illustratingBirkho'sErgodi theorem:
Foraintegrablefun tion
φ : M → R
,thelimitlim
T →∞
1
T
Z
T
0
dt φ f
t
x
0
=: hφ(x
0
)i
exists. If
µ
isanergodi measure,thenhφ(x
0
)i
isalmosteverywhere onstant(doesnotdepend ontheinitialpointx
0
)andequalto:hφ(x
0
)i =
Z
M
φ(x) µ(dx) =: [φ(x
0
)] .
Foraproofsee, forexample,theoriginalarti lebyBirkho[15℄.
Thepowerfulresultobtainedisthatanyobservableofoursystem,
orrespond-ing to an integrablefun tion
φ
hasa denite global average valueoverphase spa e. This value an be omputed by integrating the fun tion on a randomtraje toryreasonablylongintime,andtheresultdoesnotdependontheinitial
point
x
0
ofthespe i traje tory. Thisfundamentalprin iplegivesmeaningto basi allyalltheanalysisperformedin thepresentwork.1.2 Chaoti Dynami s and Lyapunov Exponents
As mentioned before, a dynami alsystem is said to be haoti when its time
evolution is highly sensitive to the initial onditions. Namely, if we taketwo
points atdistan e
ǫ
arbitrarilysmall, the divergen eof thetwotraje toriesin timewillbe∼ ǫ e
λt
,
λ > 0
. Forλ < 0
thedieren ewouldqui klyde ay,making thedynami sstable;ifλ = 0
nothing anbesaid. Ifthemotionis onnedonan attra tor,thedistan e annotgrowindenitely. Whathappensisthat thetwotraje tories,aftertheinitialstrongdivergen e,be ome ompletelyindependent
oneanother.
Ifweapplythis on epttoapointinphasespa e
x
,withasmallerror asso i-atedδx
,werea hthe on lusionthateventhougha haoti systemisregulated bypre iseequations,theexponentialgrowoferrorsmakesany al ulatedtimeevolutionself-independent of itslong-term pasthistory, i.e. non-deterministi
inanypra ti alsense[11℄.
The parameter
λ
, alled the Lyapunov exponent, is of primary importan e in studying quantitatively the haoti behavior and estimating its long-termpredi tabilityinsu hsystems. Forapre isedenitionand al ulation,weneed
asatisfyingmathemati aldes riptionoftheentities inquestion.
Let
M
beasmooth, ompa tmanifoldandf
amappingorowoverM
with theproperties(1.1);letµ
beameasureinvariantundertimeevolution;x
isour starting point onM
. We anidentify the small dieren e betweenx
and an arbitrarily losepointasave toru
belongingtothetangentve torspa etoM
inpointx
,T
x
M
. Pairsexpressedas(x, u)
,withx
∈ M
andu
∈ T
x
M
anbe interpretedaselementsofthetangentbundle ofM
,thatwe allT M
.We andeneaniterationofagivenamountoftime
t
asamapoftheform:T
t
: T M → T M
(x, u) → f
t
(x), T
x
t
u .
Thelinearoperator
T
t
x
mapstheve toru
∈ T
x
M
inu
˜
∈ T
f
t
(x)
. Thisoperation an be seen asthe push-forward ofu
byf
: this implies that the operator is simply thelinearized versionoff
t
(x)
, i.e. its total derivative,represented by
theJa obian matrix:
T
t
x
:= Df
t
(x)
.Fromthepropertieson
f
and the hainderivationrulewehavethat:T
x
t+s
= T
s
f
t
(x)
T
x
t
(1.2)in other words, along time iteration is equivalent to a produ t of short time
iterations, onsistingofelementsdenedfollowingthetraje toryofthestarting
point
x
.From now onwewill onsider
f
asa dis retemap, witht ∈ N
. Howeverit ispossibleto generalizeallthefollowingdenitions andresultsfor ows,withlittle hangeinthenotation. If
|| · ||
x
istheve tornormoverthespa eT
x
,we anexpresstherateof hangeoftheve toru
foronestep(t = 1)
as:r(x, u) =
||T
1
x
u
||
f
1
(x)
||u||
x
forthenexttimestep,using(1.2),wehave:
r(f
1
(x), T
1
x
u) =
||T
1
f
t
(x)
T
x
1
u
||
f
2
(x)
||T
1
x
u
||
f
1
(x)
=
||T
2
x
u
||
f
2
(x)
||T
1
x
u
||
f
1
(x)
.
Wenwe omputethegeometri mean of the hange rateforthe
n
elementsof thetimeseriesx
,f
1
(x)
,
f
2
(x)
...
f
(n−1)
(x)
,theresultis:
h
r
f
(n−1)
(x), T
x
(n−1)
u
· r
f
(n−2)
(x), T
x
(n−2)
u
· . . . · r(x, u)
i
1
n
=
=
"
||T
n
x
u
|| ||T
(n−1)
x
u
|| . . . ||T
x
1
u
||
||T
x
(n−1)
u
|| . . . ||T
x
1
u
|| ||u||
#
n
1
=
||T
n
x
u
||
||u||
n
1
.
The logarithm of this quantity in the large
n
limit is alled the Lyapunov Chara teristi Exponent (LCE)of(x, u)
λ(x, u) := lim
n→∞
log
||T
n
x
u
||
||u||
1
n
=
lim
n→∞
1
n
log ||T
n
x
u
|| − lim
n→∞
1
n
log ||u|| = lim
n→∞
1
n
log ||T
n
x
u
||
(1.3) Thisdenition satisestheproperties:∀k ∈ R\{0} λ(x, k u) = λ(x, u) ;
∀u, v ∈ T
x
M
λ(x, u + v) ≤ max {λ(x, u), λ(x, v)} .
(1.4)
TheLCEdepend onlyonthedire tion oftheve tor,andnotonitsmagnitude
(adire t onsequen eofthelinearityoftheoperator
T
t
x
).The existen e of su h limits is granted by Oselede 's multipli ative
er-godi theorem[16℄. It anbeexpressedasfollows:
Givena ompa t,dierentiablemanifold
M
,amappingf : M →
M
andameasureµ
invariantoverM
; letT
1
beamap from
M
to thespa eallthem×m
realmatri es,withthenotationT
1
(x) = T
1
x
, su hthatZ
M
µ(dx) log
+
||T
x
1
|| < ∞ ;
wherelog
+
(k) = max{0, log(k)}
and|| · ||
isamatrixnorm; letT
n
x
bedened astheprodu t
T
n
x
= T
f
1
(n−1)
(x)
...T
1
f (x)
T
x
1
. Thenthere is af
-invariantsubspa eN ⊆ M
su hthatµ(N ) = 1
and∀x ∈ N
(indi atingwithA
∗
theadjointof
A
)thematrixΛ
x
:= lim
n→∞
((T
n
x
)
∗
T
x
n
)
1
2n
exists.It has
s ≤ m
distin t, real eigenvalues, that an be ordered asexp λ
(1)
x
> . . . > exp λ
(s)
x
, orresponding to the eigenspa esU
(r)
x
;r = 1, . . . , s
. Theλ
(r)
x
exponentsassumerealvaluesor anbe−∞
ifthe orrespondingeigenvalueis0
.Ifwedene
L
(r)
x
= U
x
(r)
⊕ U
x
(r+1)
⊕ . . . ⊕U
x
(s)
,andL
(s+1)
x
={0}, wehavethatforu
∈ L
(r)
x
\ ∈ L
(r+1)
x
lim
n→∞
1
n
log ||T
n
x
u
|| = λ
(r)
x
,
i.e. thelogarithms of theeigenvalues of
Λ
x
are theset ofall LCE we anndfromx
. Finally,deningd
(r)
x
:=
dimU
(r)
x
,wehavethat thefun tionsx
→ λ
(r)
x
andx
→ d
(r)
x
aref
-invariant(λ
(r)
x
= λ
(r)
f
t
(x)
, et .) and, ifthe systemisergodi , arealmost everywhere onstant(withthepossibleex eptionofasetof
0
measure).Thematrix
Λ
x
is alled Oselede Matrix. If wewrite the ve tornorm asa s alarprodu t||u|| =
phu, ui
;theexpansionrateafterasingleiterationis:||T
1
x
u
||
||u||
=
s
hT
1
x
u, T
x
1
u
i
hu, ui
=
s
h(T
1
x
)
∗
T
x
1
u
, ui
hu, ui
;
thelastequality omesfromthedenitionofadjointmatrix. Now,assuming
u
isaneigenve torof(T
1
x
)
∗
T
1
x
witheigenvaluek
,theresultis:s
h(T
1
x
)
∗
T
x
1
u
, ui
hu, ui
=
s
hu, ui k
hu, ui
=
√
k .
(1.5)This explanation hasthe only purpose of des ribing theidea behindthe
on-stru tionof theOselede matrix. Forafulldes riptionofthe theoremand its
relationwith theLyapunovexponents,see [11,17, 18℄. An exhaustive
mathe-mati aldemonstration anbefoundin[19℄. TheOselede multipli ativeergodi
theorem isnot simplyan alternativewayto express theLyapunovexponents,
butmorethanthat,itstatestheexisten eofsu halimitasaninvariant
prop-ertyofthedynami alsystem,independentfromtheinitialpoint hosenfortime
evolution.
Thefolding ve torspa es
L
(1)
x
⊇ L
(2)
x
...⊇ L
(s)
x
indu e anaturalsplitting on thetangentspa e,knownasOselede splitting. Theimportan eofsu hsplittinganbeillustratedasfollows: ifwetakearandomve tor
u
belongingtoT
x
M
, itsmeangrowratewillbeexp λ
(1)
x
,i.e. theexponentialofthehighestLyapunov exponent. This omesfromthefa t thatu
∈ L
(1)
x
: thesubspa esL
(2)
x
, L
(3)
x
. . .
havezeromeasurewith respe tto thetotalspa e, sotheprobability ofa
ran-domve tortobelimitedtotheminsubstantiallyzero. However,ifwe onsider
arandomve tor
u
˜
fromwhi hwesystemati allyremovethe omponentin the dire tion(s)ofhighestexpansion,i.e. itsproje tiononU
(1)
x
, thenu
˜
∈ L
(2)
x
and its LCE will bethe se ond highest exponent,λ
(2)
x
, and so on. The numeri al al ulationofthe ompletespe trumofLCE isessentiallybasedonthisme h-anism. We an alsonoti e thatthe orderingof thespe trumindu esdierent
ordersofexpansion(andstability,fornegativeLCEs)indierentsubspa es.
Moreover,the basisof ea h
U
(r)
x
subspa e(i.e. theΛ
x
normalized eigenve -tors), representsthe (average) dire tion asso iated to the expanding (oron-tra ting) average rate
exp λ
(r)
x
. Su h dire tions are alled Lyapunov ve tors and,asdes ribedin se tion1.4, anbe omputedandstudied lo ally, inordertogainusefulinformationonthelo aldynami sandglobaldynami s.
1.3 Cal ulating Lyapunov Exponents
The lassi alalgorithmforthe al ulationoftheLyapunovexponentsdatesba k
to 1980 [20℄. It onsiders only dynami al systemswhose governing equations
arefully known and omputable. If thedynami s ishidden, it isstill possible
to estimate some of the exponents by using an empiri al time series of some
observableof the system. In[21℄, forexample, both ases are onsidered (see
also [22℄). However, for the s ope of the present work, we will sti k to the
hypothesisofthe lassi alalgorithm.
Following the reasoning of the original arti le, westart with the pro edure
forthe rst, highestexponent, and then wegeneralize theresult to thewhole
spe trum. Firstofallweremindthatforastartingpointinourmanifold
x
∈ M
andave torinitstangentspa eu
∈ T
x
M
, thelinearoperatorinvolvedin the push-forward(x, u) → (f
t
(u), T
t
x
u)
is the Ja obianoff
t
x
. This valueis basi ally the ompletederivative off
t
: in se tion 2.6of next
hapterwewillshowthe al ulationinourspe i setting. Fornowweassume
the
T
t
x
matri esasgivenforeveryx
inM
.As said in previousse tion, for the rst exponent is su ient to follow the
evolutionofarandomve tor
u
0
. Tohaveanintuitiverepresentationofthe pro- ess,we animaginetode omposeu
0
inthebasisofthedire tionsofexpansion asso iated to ea h exponent (i.e. the Lyapunovve tors). Being random, ourstarting ve torwill haveanonzero omponentfor ea h dire tion. Due to the
linearityofthepro essea h omponentwillin rease,onaverage,exponentially,
a ordingto therespe tiveLyapunov exponent. The exponentialdieren e of
thegrowingrateswill ausethe omponentasso iatedto thehighestexponent
to dominateoveralltheothersafter ashort amountofiterations,so thatit is
theonlyonesele tedinthelarge
n
limit.Theonlydi ultyisthatanexponentiallygrowingve torwouldsoongoout
oftheboundariesofour omputational apabilities(that'sbasi allythereason
whywe annon al ulate
Λ
x
dire tly). It issolvedasfollows. Givenanintegerk & 1
su h thatT
k
x
lies safely in our numeri limits, westart with a random ve toru
0
∈ T
x
M
and al ulate iterativelytheseries:u
1
= T
x
k
u
0
||u
0
||
,
u
2
= T
f
k
k
(x)
u
1
||u
1
||
,
u
3
= T
f
k
2
k(x)
u
2
||u
2
||
, . . .
u
i
= T
f
k
(i−1)k
(x)
u
i−1
||u
i−1
||
=
T
k
f
(i−1)k
(x) T
k
f
(i−2)k
(x) . . . T
k
f
k
(x)
u
0
||u
i−1
|| ||u
i−2
|| . . . ||u
0
||
;
||T
x
ik
u
0
|| = ||u
i
|| ||u
i−1
|| . . . ||u
0
|| ;
nally,fromu
0
∈ L
(1)
x
and(1.3),wehave:λ
(1)
x
= lim
n→∞
1
n
n
X
i=1
log ||u
i
|| .
(1.6)Both thefa t that
T
k
x
isapplied onlytonormalizedve torsandthat the loga-rithmis omputedateverysinglestephavethepositiveee tof ontainingthesizeofthenumbersinvolved,makingthe omputationpossible.
To al ulatethewholespe trum,wemustthinkintermsofvolumevariations.
Weassumethe
Λ
x
matrixto bem
dimensional and,forsimpli ity, that allits eigenvalueshavemultipli ityone(i.e. wehavem
distin tLCE).LetU ⊂ T
x
M
beaopenset ofvolumeVol(U )
; usingthe denitionof LCEand theOselede theorem,weinferthatitsaveragegrowingratein timeis∝ exp
P
m
i=1
λ
(i)
x
. We start fromU
0
, dened for onvenien e as them
-dimensional hyper ube en losed in a random orthonormal basis ofT
x
M
:{u
(1)
0
,u
(2)
0
, ...u
(m)
0
}
; LetA
be the linear operator orresponding to a single time iteration of the set. Followingthesamereasoningasbefore,wedeneiterativelythesetsU
i
as:U
i
=
A(U
i−1
)
Vol
(U
i−1
)
;
sothat, fromthelinearityof
A
,itfollows:Thequantityontherightisthevolumeofourinitialhyper ubeafter
i
timesteps, namely: Vol(AA . . . AU
0
) =
Vol(T
ik
x
u
(1)
0
, T
x
ik
u
(2)
0
, . . . T
x
ik
u
(m)
0
) ;
thisexpressionleadsto thesumofallLCE:
lim
n→∞
1
n
log
Vol(A
n
U
0
)
Vol(U
0
)
= lim
n→∞
1
n
log (
Vol(A
n
U
0
)) =
m
X
j=1
λ
(j)
x
.
Thequantitiesontheleftassumetheform:
Vol
(U
i
) =
Vol
(A(U
i−1
))
Vol(U
i−1
)
=
(1.7)=
Vol(T
k
f
(i−1)k
(x)
u
(1)
i−1
, T
f
k
(i−1)k
(x)
u
(2)
i−1
, . . . T
f
k
(i−1)k
(x)
u
(m)
i−1
)
Vol(U
i−1
)
.
Byiteratingtheve torsasbefore(
u
(j)
i
= T
f
k
(i−1)
(x)
u
(j)
i−1
/||u
(j)
i−1
||
)and omputing thevolumeforea hstep,we an,inprin iple,ndtherightresult;howeverthismethod is not feasible, due to the fa t that all ve tors would soon onverge
onthe dire tion of maximumexpansion, sothat theangles betweenthem are
beyondthenumeri alresolutionandthevolume annotbe al ulated.
To ir umventthisproblem,forea htimestepwere omputetheve tors
den-ingthevolume
U
i
performingaGram-S hmidtorthogonalizationpro edure:U
i
doesnot hange,anditsvolumesimplybe omestheprodu toftheorthogonalve tor norms. In short, if we assume
h·, ·i
asthe s alar produ t in the spa eT
f
ik
(x)
M
,webuildtheseriesofu
(j)
i
as:u
(1)
i
= T
f
k
(i−1)k
(x)
u
(1)
i−1
||u
(1)
i−1
||
;
forj > 1
:u
(j)
i
= T
k
f
(i−1)k
(x)
u
(j)
i−1
||u
(j)
i−1
||
−
j−1
X
r=1
*
T
k
f
(i−1)k
(x)
u
(j)
i−1
||u
(j)
i−1
||
, u
(r)
i
+
u
(r)
i
||u
(r)
i
||
2
.
(1.8)With this denition Vol
p
(T
k
f
(i−1)k
(x)
u
(1)
i−1
, T
f
k
(i−1)k
(x)
u
(2)
i−1
, . . . T
f
k
(i−1)k
(x)
u
(m)
i−1
)
isequivalentto:
m
Y
j=1
||u
(j)
i−1
||
·
Volp
(u
(1)
i
, u
(2)
i
, . . . u
(m)
i
) =
m
Y
j=1
||u
(j)
i−1
||
·
m
Y
j=1
||u
(j)
i
||
.
Thelast equalityisduetotheorthogonalityofthe
u
(j)
i
ve tors. From(1.7)we obtainVolp
(U
i
) =
Q
m
j=1
||u
(j)
i
||
,whi hleadsto theresult:m
X
j=1
λ
(j)
x
= lim
n→∞
1
n
log
m
Y
j=1
||u
(j)
0
|| ·
m
Y
j=1
||u
(j)
1
|| · . . .
m
Y
j=1
||u
(j)
n
||
.
Rearranging the produ t indexes and de omposing the logarithm, we nally
obtain,forthe
s
th exponent:λ
(j)
x
= lim
n→∞
1
n
n
X
i=1
log ||u
(j)
i
|| .
(1.9)Equation(1.9)isanextendedversionof(1.6),withthesigni antdieren ethat
the
u
(j)
i
ve tors,forj > 1
, are al ulateda ordingto theorthonormalization pro edure (1.8) . As we will see in se tion 1.5 this pro ess anbe simplyandstraightforwardlytranslatedinbasi operationonmatri es.
Before moving to the al ulation of Lyapunov ve tors, some onsiderations
about dimensions should to be done. The result of (1.9) is dimensionless, as
it represents the logarithm of average expansion per step. To obtain a more
generalquantity,independentof thesteplengthsused fortheevolutionof the
dynami alsystem(aslongastheintervalsaresmallenoughtofollowproperly
thedynami alevolution),weneedtores aletheexponentsusing theduration
ofatimestep
∆t
:λ
(j)
i
←
λ
(j)
i
∆t
(1.10)The oe ientsarethusexpressedin se onds
−1
.
Thesystemanalyzedthepresentwork hasthepe uliarityofhavingstepsof
dierent time lengths. Sin e the LCE is dened as an average quantity, we
divide itby theintervallengthaveragedonallperformedsteps:
[∆t]
s
(see for examplepag. 120of[14℄).1.4 Cal ulating the (lo al) Lyapunov ve tors
As said, the Lyapunov ve tors, dened as a base for the eigenspa es of
Λ
x
, representthedire tionoftheglobalaverageexpansionasso iatedtoea hexpo-nent. It is then possible to dene thelo al Lyapunovve torsasthepreferred
expanding dire tions forthe (linearized)dynami s of ea h point in time. The
lo alLyapunovexponentsarethenthepun tualexpansionratesinea hofthose
dire tions.
Theidenti ationat anystep
s
ofthe maximumlo alexponentλ
(1)
s
and its asso iated dire tionv
(1)
s
is straightforward: as said, a random ve tor, freely evolvingin time,wouldrapidlyalignwithit. Sothat,usingtheu
(1)
s
ve torsas dened above,weobtain,fortheve torandtheexponent:v
(1)
s
=
u
(1)
s
||u
(1)
s
||
;
λ
(1)
s
= log
||T
k
f
ik
(x)
u
(1)
s
||
||u
(1)
s
||
= log ||u
(1)
(s+1)
|| ,
(1.11)for values of
s
reasonably distant from0
, so thatu
(1)
has su ient time to
align,andassumingthatthemappingissmoothenough,sotheve torproperly
followsthedominatingdire tionateverypointintime.
Movingto
v
(2)
s
andλ
(2)
s
,theideaintheseries(1.8)isthat,afterweapplythe linearoperator,weproje ttheresultingve toronthespa eorthogonaltov
(1)
s
,t
t
v
1
v
2
Figure 1.2: In a haoti dynami al system any perturbation will onverge onthe
dire tion of the rst Lyapunov ve tor. Onthe other hand, a timeinversion inthe
dynami sresultsinthemarkeddominationoftheleastforwardexpandingdire tion,
i.e. these ondve tor.
thusthegrowthinthatdire tiondire tion(expe tedtobedominating)is
om-pletelysuppressed, whilethenexthighestgrowingdire tion,regulatedby
λ
(2)
s
, be omes visible. As aresult,u
(2)
s
/||u
(2)
s
||
hasthesamedire tionasthe proje -tion ofv
(2)
s
onthespa eperpendi ulartov
(1)
s
. Ingeneralu
(2)
s
/||u
(2)
s
|| 6= v
(2)
s
: noreasonfor estheLyapunovve tors,eitherglobalorlo al,tobeperpendi u-larea hother. La kingtheknowledge ofthe orre tdire tion andofthelo al
exponentdenedasabove(it annotbefounditthe orrespondingdire tionis
notknown),we annot al ulatethetrue
v
(2)
s
.Asolutionto ir umventthisproblemhasbeenproposedonlyin2007[23℄. It
isbasedonthequiteintuitiveandwellknowprin iplethatatime-inversionin
thedynami sofoursysteminvertstheLyapunovspe trum. Avolumenormally
expanding in some dire tions, in a ba kward motion would ontra t in those
samedire tionswithinvertedratios,whilearandomve tortravelingba kwards
in timewouldsoonfollowthedire tion oflessexpansionin forwardtime. See
gure1.2to haveagraphi alideaofthepro ess.
Fortheargumentsabove,weknowthat Span
(v
(1)
s
, v
(2)
s
) =
Span(u
(1)
s
, u
(2)
s
)
. Ifwetakearandom ombinationofu
(1)
s
andu
(2)
s
andevolveitba kwards,after someiteration(let's sayh
)it will bealignedwith the lessforward expanding dire tionofoursubspa e,namelyv
(2)
s−h
. Forve torv
(s)
s−h
wesimplystartwitha linear ombinationofve torsv
(1)
s
, u
(2)
s
,...u
(s)
s
,keepinginmindthath
should bebigenoughtolettheve torsalignproperly.1.5 Matrix Cal ulation
Assaid,thereisasimpleandelegantwaytotranslatethealgorithmsdes ribed
westartdeningarandom,orthonormal,
m×m
matrixwhose olumnsrepresent abasisforT
x
M
:Q
0
= (q
(1)
0
| . . . |q
(m)
0
)
. Duetotheirstru ture,theq
(j)
ve tors,
i.e. the olumnsofthe
Q
matrix,are alledthe Gram-S hmidt basis.T
k
x
istheJa obianmatrixasso iatedtoak
-stepsevolution,withk
hosenso thattheexponentialgrowthsarekeptinthenumeri allimitsofour al ulator.Wenowperformthemultipli ation:
T
f
1
k
−1
x
. . . T
1
f
1
x
· Q
0
= (T
x
k
q
(1)
0
| . . . |T
x
k
q
(m)
0
) = (˜
q
(1)
1
| . . . |˜q
(m)
1
) = ˜
Q
1
;
(1.12) whereT
1
x
...T
1
f
k−1
x
are the Ja obianmatri es fora singletimestep, omputed forthespe i dynami alsystemweareanalyzing.Theve tors
q
˜
(j)
1
arenotorthogonal. Toobtainthe orre titeratedQ
1
matrix, we apply aQR
de omposition onQ
˜
1
. TheQR
de omposition on a generi square matrixQ
˜
onsists in writingit astheprodu tQ = Q · R
˜
,su h thatQ
isorthonormal andR
isuppertriangular. Calling˜
q
(j)
and
q
(j)
the olumnsof,
respe tively,
Q
˜
andQ
,thenewmatri esaredened by:q
(1)
=
q
˜
(1)
||˜q
(1)
||
;
q
(j)
=
˜
q
(j)
−
j−1
P
r=1
D
˜
q
(j)
, q
(r)
E
q
(r)
˜
q
(j)
−
j−1
P
r=1
D
˜
q
(j)
, q
(r)
E
q
(r)
{R}
ij
=
D
q
(i)
, ˜
q
(j)
E
,
fori ≤ j .
(1.13)Itisstraightforwardtoseethatthesevalues anbeexpressedin termsofthe
u
(j)
ve tors(1.8)asq
(1)
1
= u
(1)
1
/||u
(1)
1
||
...q
(m)
1
= u
(m)
1
/||u
(m)
1
||
Moreover,from thedenitionof theQR
algorithmdes endsthat thediagonalof theR
matrix orrespondstotheu
(j)
1
norms:{R
1
}
jj
= ||u
(j)
1
||
.Theorthonormalized
Q
1
isnowreadyforanotheriteration:T
f
1
2k−1
x
. . . T
1
f
k
x
· Q
1
= ˜
Q
2
= Q
2
· R
2
...andsoon. Theu
(j)
i
ve tor modules, given by the diagonal ofR
i
, are stored in ea h iterationforthe al ulationoftheLyapunovspe trum.Thissamesetting,alongwiththeresultsalready omputed, anbealsoused
tondtheLyapunovve tors. Theendingpointofour omputation
T ≫ 1
will beassumedasthestartingpointofourba kwarditeration. Thestartingran-domve tors
v
(1)
T
∈
Span{u
(1)
T
}
,v
(2)
T
∈
Span{u
(1)
T
, u
(2)
T
}
...,v
(m)
T
∈
Span{u
(1)
T
, ...u
(m)
T
}
anbedenedastheprodu tofQ
T
witharandomtriangularmatrixC
T
. Consideringthetimeiteration,weobtain:Q
T
· C
T
= T
f
k
(1−T )k
(x)
· Q
T −1
· C
T −1
= ˜
Q
T
· C
T −1
= Q
T
· R
T
· C
T −1
;
sothat
C
T −1
= (R
T
)
−1
· C
T
.
(1.14)The previously al ulated triangular
R
s
matri es must thus be fully stored, invertedandusedtoiterate ba kwardin timetheC
i
matrix:C
i
= (R
i−1
)
−1
· (R
i−2
)
−1
· . . . · (R
T
)
−1
· C
T
.
Sin e weare interestedonly in the ve tordire tions, we an freelynormalize
the olumnsofthe
C
i
matri es. Ifwedothatateverystep,we anidentifythe normalizationfa torsatpointi
withthelo al Lyapunovexponentsλ
(j)
s
, while theLyapunovlo alve torsaresimplythe(normalized) olumns ofQ
s
· C
s
.1.6 Information and Entropy
Considertwopointsvery loseoneanotherinthephase-spa eofa haoti
sys-tem: for any observerwhose instruments have a pre ision oarser than their
distan e,theywillappear ompletely indistinguishable. Eventually,asthe
sys-temevolvesintime,thetraje toryseparationdueto haoti dynami smakesthe
distan esigni ant,sothatthepointsareper eivedasseparate. Thus,systems
verysensitivetoinitial onditions anbeseenasprodu ersofinformation.
Let
{A
1
, A
2
, . . . , A
α
}
beanite,µ
-measurablepartitionofthephasespa e. We anassumeit orrespondstotheresolutionofourinstruments,sothattwopointsinthesamesetofthepartition
A
annotbeseenasdistin t. We anthen denef
−
k
(A
i
)
astheset of pointsx
su h thatf
k
(x) ∈ A
i
, and allf
−
k
(A )
thepartition{f
−k
(A
1
),
...f
−k
(A
α
)}
. Wenally onsider thepartition given bytheleast ommonrenement:A
(n)
= A ∨ f
−
1
(A ) ∨ f
−
2
(A ) ∨ . . . ∨ f
1−n
(A ) ;
itisdenedsothatageneri setin ithastheform:
A
i
1
∩ f
−1
(A
i
2
) ∩ . . . ∩ f
1−n
(A
i
n
)
fori
j
∈ {1, 2, . . . , α}
It is lear that the latter, dynami s-related, partition has a resolution mu h
nerthanthestartingone,sin eanyelementin itisdis riminated byitspast
history,upto
n − 1
steps.Now,we andenetheinformation ontentofthepartition
A
(n)
withrespe t tomeasureµ
as:H(A
(n)
) =
X
i
1
, ...,i
n
µ(A
i
1
∩ . . . ∩ f
1−n
A
i
n
) log µ(A
i
1
∩ . . . ∩ f
1−n
A
i
n
) ;
where we sum overevery element of
A
(n)
. The rate of information reation,
withrespe ttotheinitialpartition
A
isthen givenbythelimit:h
(µ, A) = lim
n→∞
H(A
(n+1)
) − H(A
(n)
)
= lim
n→∞
1
n
H(A
(n)
)
.
TheShannon-Ma Millantheoremguaranteestheexisten eofthis limit.
The Kolgomorov-Sinai entropy
h
(µ)
is then dened as the further limit ofh
(µ, A)
fornerandnerstartingpartitionsA
.As stated at the beginning of this se tion, the information reation rate is
originatedbytheexpandingmotionofthesysteminthephasespa e, onne ted
withits haoti behavior. In1978Ruelledemonstratedthatitsvalue annotbe
greaterthanthetotalpositiveexpansionrateofthesystem,givenbythesum
overallpositiveLyapunovglobalexponents:
h
(µ) ≤
X
λ
(i)
>0
λ
(i)
.
(1.15)Pesin extended this theorem, proving that (1.15) is an identity if (and only
if) the measure
µ
is aSRB measure. Weshall briey dene and dis uss SRB measures at the end of the next se tion, sin e they result ru ial also in the1.7 Information Dimension of the Attra tor
As mentioned at the beginning of this hapter, (see pag 17) the attra torin
the phase spa e
A ⊆ M
has, in general, afra tal stru ture. We assume the notionofHausdordimensionasgiven,withthenotationdimH
,anddenethe informationdimension ofthemeasureµ
as:dim
H
(µ) = inf{
dimH
(S)|µ(S) = 1} .
Young's theorem (1982)shows that, if
µ
is an ergodi measure, this value is equivalentto: dim H(µ) = lim
r→0
log µ(B
x
(r))
log r
.
Here
B
x
(r)
representsthe ball enteredinx
of diameterr
. The expression is validand onstantforeveryx
∈ A
ex ept,possibly,forasetof0 µ
-measure.Now, if
λ
(1)
, . . . , λ
(m)
are the global Lyapunov exponents asso iated to
µ
, andk = max{ i | λ
(1)
+ . . . + λ
(i)
≥ 0}
,we andenetheLyapunovdimension
as:
dim
Λ
(µ) = k +
λ
(1)
+ . . . + λ
(k)
|λ
(k+1)
|
;
(1.16)these ondtermisasmall,noninteger orre tionforthe ase:
P
k
i=1
λ
(i)
> 0
andP
K+1
i=1
λ
(i)
< 0
. Inoursystem,duetogreatdimensionality,itresultsnegligible. The onne tionbetween thetwo quantities dened hereis given by thefol-lowing onje ture byKaplanandYorke: if
µ
isanergodi ,SRBmeasure,then dimH
(µ) =
dim
Λ
(µ) .
(1.17)It is analyti ally proved that this equality holds in some spe i ases, but
ex eptionsarefound.
SRB measures and hyperboli ity
Intheprevioustwose tionswestatedthata ru ialpropertyofthemeasurewe
use,bothfortheexa t al ulationofentropyprodu tionandareasonableesteem
oftheattra tordimension,isbeingaSRBmeasure(fromSinai,Ruelle,Bowen);
namely ameasure whi h is absolutely ontinuous along unstablemanifolds. A
rigorousdenition anbefoundin[10,11℄.
Itis provedthat fora lassof dynami alsystems,namelythe Axiom-A
sys-tems,existsauniqueSRBmeasure,whi h anbeexpressedphysi ally asthe
ergodi average:
ρ = lim
n→∞
1
n
n−1
X
k=0
δ
f
k
x
.
The problemis then transferred ondemonstrating that the dynami alsystem
isAxiom-A:in that asetheSRBmeasurenaturally orrespondstoanaverage
overlongdynami al traje tories, asthe ergodi prin iplestates; onsequently
PesinidentityholdsandKaplan-Yorke onje tureisonsolidground.
The ru ialpropertyofAxiom-Asystemishyperboli ity. Aset
A
ishyperboli foradieomorphismf
(mappingof ow), if∀x ∈ A
thereexists adire t sumde omposition of thetangentspa e betweenstable (expandingovertime)and
unstabledire tions(expandingoninverted time). If
A
isahyperboli attra ting setandtheperiodi pointsoff
aredenseinA
,the onditionsforanAxiom-A dieomorphism are satised. In ase thewhole manifoldM
is hyperboli ,we havethestronger onditionsofAnisovdieomorphism andstru tural stability.ProvingthatasystemisAxiom-Ais, in general,averyhard task. However,
thepresentworkrepresentsanexampleofhowLyapunovVe tors anbeused
toidentify theexpandingand ontra tingdire tionsofthetangentspa e. The
measure of their transversality represents then a quantitative esteem of the
Model
Inthis hapterabasi modelforastati ,larges alenetworkof orti alneurons
is built. Starting form the Hodgkin-Huxley lassi equations, simpler models
areinferred. Itisthenexplainedwhythe anoni alquadrati integrateandre
model(QIF)isasuitabletodes ribethedynami sofverylarges alenetworks.
Inthenext se tionamoregeneralmodelfor pulse- oupled neuronalnetworks
ispresented. Expli it equationsareobtainedfromthe QIFmodeland usedto
al ulateanalyti allydierentglobalparametersofthenetwork;thenaformula
fortheJa obianatanygivenspiketimeispresented. Su hanalyti alexpressions
will be the basis of the a tual network simulation and of all the subsequent
resultspresentedin next hapters.
2.1 Single Neuron Dynami s
Hodgkin and Huxley model
Thebestknownandmostwidelya eptedequationsusedtodes ribethe
poten-tialofaneural ellsoma asafun tionofexternal urrentandinternal
ondu -tan e parameters,dates ba kto the pionieristi work of Hodgkin and Huxley
[24℄. Theequationsintheirstandardformare: