HAL Id: hal-00476240
https://hal.archives-ouvertes.fr/hal-00476240
Submitted on 25 Apr 2010
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
Asymptotic distribution of global errors in the
numerical computations of dynamical systems
Giorgio Turchetti, Sandro Vaienti, Francesco Zanlungo
To cite this version:
Giorgio Turchetti, Sandro Vaienti, Francesco Zanlungo. Asymptotic distribution of global errors in the
numerical computations of dynamical systems. Physica A: Statistical Mechanics and its Applications,
Elsevier, 2010, 389 (21), pp.4994-5006. �10.1016/j.physa.2010.06.060�. �hal-00476240�
numeri al omputations of dynami al systems G.Tur hetti
∗
,S. Vaienti†
,F. Zanlungo‡
Mar h7, 2010 Abstra tWeproposeananalysisoftheee tsintrodu edbynitea ura yand
roundoarithmeti onnumeri al omputationsofdis retedynami al
sys-tems. Ourmethod,thatusesthe statisti altoolofthede ayof delity,
omputestheerror omparingdire tlythenumeri alorbitwiththeexa t
one (or, more pre isely, with another numeri al orbit omputed with a
mu hhighera ura y). Furthermore,asamodelofthe ee tsofround
oarithmeti onthemap,wealso onsiderarandomperturbationofthe
exa torbitwithanadditivenoise,forwhi hexa tresults anbeobtained
for someprototype maps. Weinvestigatethe de ay laws of delity and
their relationshipwiththe errorprobability distributionfor regular and
haoti maps, both for additive and numeri al noise. Inparti ular, for
regularmapswendanexponentialde ayforadditivenoise,andapower
lawde ayfornumeri alnoise. For haoti mapsnumeri alnoiseis
equiv-alenttoadditivenoise,andourmethodissuitabletoidentifyathreshold
for thereliability of numeri al results,i.e. a numberofiterationsbelow
whi h global errors an be ignored. This threshold grows linearly with
thenumberofbitsusedtorepresentrealnumbers.
1 Introdu tion
The reliability of numeri al omputations in dynami al systems is a relevant
longstandingquestion. For ontinuoussystemstheglobalerrorontheorbitis
theresultof thea umulationof twotypes oferrors: therstoneis thelo al
errordue to the dis retization algorithms, the se ond oneis due to the nite
pre isionrepresentationofreals inany omputationaldevi eand totheround
oarithmeti . Inthis paperwepropose astatisti al analysisof theglobal
er-rorsdue to nite a ura yandround o in dis retedynami al systems. Even
∗
DepartmentofPhysi s,UniversityofBologna
†
Centre de Physique Théorique, CNRS, Universités d'Aix-MarseilleI,II, Universitédu Sud,Toulon-VarandFRUMAM
‡
DepartmentofPhysi s,Universityof BolognaandATRIntelligent Roboti sand Com-muni ationLaboratories,Kyoto,Japan
the omparisonwithrandomlyperturbedmaps,forwhi hanalyti alresultsare
available,allowsto understandthekeyfeatures. Formapswith hyperboli
at-tra torstheexisten eoftrueorbits losetopseudo-orbitsisofgreattheoreti al
importan ebutdoesnotprovideanystatisti alinformationonthedis repan ies
between apseudo-orbit and a true orbitwith the same initial ondition. We
showthatfor haoti mapsthereisasharptransitionbetweentworegions,one
inwhi htheerrorisnegligibleandoneinwhi hitsvarian eis omparablewith
thesize of the attra tor,bothfor the numeri al and the randomlyperturbed
map. Forregularorquasiregularaniso hronousmapsthetransitionissmooth
and thea ura y de ay followsapowerlawfor thenumeri al map and
expo-nentiallawfortherandomlyperturbedmap.
Inapreviouswork[1℄,we omparedthevalueofanobservable omputedalong
theorbitofagivenmapandtheorbitofthesamemapwitharandom
pertur-bationforthe sameinitial ondition. Thedieren eis arandompro essand,
forlargeiterationtimes,itsprobabilitydensityfun tion(pdf) onvergestoward
awidespreaddistribution,whi h depends onthe hoi e ofthe observable. For
haoti maps,therelaxationissharp,in thesense thatthereexistsathreshold
timebelowwhi hthewidthofthepdfisverysmallapproximatinga
δ
fun tion andabovewhi hithasalreadyrea heditsasymptoti value. Theseresultswereobtainedanalyti allyand numeri allyby omputing theinverse Fourier
trans-form of asuitable orrelation integralbetween the twoobservables, alled
-delity. For haoti systemsthedelityde aysatleastexponentiallyfastandthis
allowstore overtheaverageoftheobservableswithrespe ttotheinvariantand
thestationarymeasures. FortheBernoullimapsthede ayissuper-exponential
and there is numeri al eviden e for this type of de ay for most haoti maps
numeri allyinvestigated. Forregularmapssu hasthetranslationsonthetorus
thede ayissmootherandfollowsanexponentiallaw. Numeri aleviden ewas
foundthatthesameexponentialde aypersists forquasiintegrablemaps.
In the subsequent letter [2℄ we applied the previous theoreti al framework to
ompare the exa t orbitof amap with its numeri al omputation (on a
ma- hine). The latter ould be onsidered as a pseudo-random orbit where the
noiseisgivenbytheroundoerrors.
In the present paper wedevelop with more details and formal onsiderations
theresultsannoun edin[2℄andprovidesomenewresults.
Thedieren e between the valuesof an observable omputed along the exa t
andthenumeri alorbitsshowsagaintheexisten eofanabrupttransitiontime
n
∗
forthe orrespondingpdf,whi hhasaninterpretationof pra ti alinterest.
We ould in fa t argue that below thetransition time the numeri al orbitsis
faithfultotheexa torbit. Bytheway,thisapproa hrequirestheknowledgeof
theexa tmapwhi hisnota essible. Thisdi ultyisover ameby omparing
thenumeri almap omputedforagivena ura y(simpleoating point
pre i-sion)withthesamemapevaluatednumeri allywithahighera ura y(double
orhigher oatingpoint pre ision), thatwe onsider asthereferen emap. We
showthattheresultsofthesingle-higherpre ision omparisonareequivalent,at
asinglepre isionandtheexa tmap. Weshowthatthethreshold
n
∗
growslike
− ln ε
,whereε
isthea ura yspe iedbytheleastsigni antbitusedto rep-resentareal number. Asa onsequen ethedelitythresholdn
∗
growslinearly
with the length of the string of bits used to represent real numbers. Beyond
thisthresholdtheerrordistributionspreadsqui klyoveritsa essiblerange.
Thebehaviorjustdes ribedistypi alof haoti systems. Forregularmapsthe
situation is slightlydierent: for iso hronousmaps the delityerror does not
de ay,whereasforaniso hronousmapsthede ayfollowsapowerlaw. Forthese
systemsitisnotpossibleto identifyasharptransitionfromthedelta fun tion
totheasymptoti one;thetransitionisagradualpro esswhoselengthdepends
on
ε
, i.e. getslonger with thenumberof bits used to representreal numbers. Asimilar de ayo urs forthe orrelationswhere itis due to thelo al mixingproperties,apropertytypi alofaniso hronousmaps.
2 Additive noise
Webeginbyre allingthemainresultsobtainedapplyingthedelitytorandom
perturbations ofdynami alsystems. We onsideramap
T
dened onaphase spa eX
whi hisasubsetofR
d
,endowedwithaninvariantphysi almeasure
µ
denedbylim
n→∞
Z
X
Ψ(T
n
(x))dm(x) =
Z
X
Ψ(x)dµ(x)
(1)where
m
denotes the Lebesgue measure andΨ
a ontinuous observable. Let usthen onsiderasequen eofindependentandidenti allydistributedrandomvariables
ξ
i
withvaluesintheprobabilityspa eΞ
andwithprobabilitydensity distributionη(ξ)
su hthatT x + ε ξ
stillmapsX
intoitself. Theiterationofthe mapT
isthereforerepla edbya ompositionofmaps hosenrandomly loseto it(notethatT
itselfisre overedwhenε = 0
):T
n
ε
(x) = (T + εξ
n
) ◦ (T + εξ
n−1
) ◦
. . . ◦ (T + εξ
1
) (x)
andthestationarymeasureµ
ε
ofthepro essisdenedby[3℄lim
n→∞
Z
X,Ξ
dm(x)
Y
i
η(ξ
i
)dξ
i
Ψ(T
ε
n
(x)) =
Z
X
Ψ(x)dµ
ε
(x)
(2)Wewant to study thestatisti al properties of the errorat the
n
-th iteration, dened as∆
n
ε
(x) = f (T
n
x) − f(T
ε
n
x)
, wheref
isasmoothobservable; in the following we takef (x) = x
for one dimensional maps andf (x) = x
i
, i =
1, · · · , d
,x
≡ (x
1
, · · · , x
d
)
, for multidimensional maps. We are in parti ular interestedinstudyingρ
n
ε
,theprobabilitydistribution fun tionof∆
n
ε
(x)
. This probabilitydistribution anbestudieddire tly,throughaMonteCarlosamplingover initial onditions
x
and random perturbationsξ
, or indire tly, using the delity.Fidelityisdened throughthefollowingintegral:
F
ε
n
=
Z
X,Ξ
dm(x)
Y
i
η(ξ
i
)dξ
i
Ψ(T
n
(x))Φ(T
ε
n
(x))
(3)the delity onvergesto
R
X
Ψ(x)dµ(x)
R
X
Φ(x)dµ
ε
(x)
, and the absolute value ofthedieren ebetweentheintegral(3) anditslimitingvalueintermsof theinvariantand stationary measure will be alled thedelity error and denoted
with
δF
n
ε
. Note that by theasymptoti hara terization of theinvariantand stationarymeasures,thedelityerror ouldbeequivalentlydened as:δF
ε
n
= F
ε
n
−
Z
X
Ψ(T
n
(x))dm(x)
Z
X,Ξ
Y
i
η(ξ
i
)dξ
i
dm(x)Φ(T
ε
n
(x))
(4)whi h emphasizesthe role of the Lebesgue measure in the numeri al
ompu-tations. Thisquantityis relevantbe ause theinverse Fouriertransformof the
delity
F
n
ε
(u)
, omputed by hoosingΨ(x) = e
iuf (x)
,
Φ(x) = e
−iuf (x)
is the
probabilityerrordistribution
ρ
n
ε
(x)
. Ifthedelityerror onvergestozero,then thedelity onvergestotheprodu tof theFouriertransformsofthemeasuresR
X
e
iux
dµ(x)
R
X
e
−iux
dµ
ε
(x)
anditsinverseFouriertransformisjustρ
∞
ε
. The initialerror distributionis theDira distributionρ
0
(s) = δ(s)
, whereasit an
beshownthatiftheinvariantandstationarymeasuresareLebesguethe
asymp-toti distribution isthe triangular fun tion
ρ
∞
(s) = (1 − |s|)ϑ(1 − |s|)
. Sin e thede aytimes alesof thedelityerrordepend veryweaklyonthe hoi eoftheobservables
Ψ
andΦ
, the study ofδF
n
ε
usingΨ(x) = Φ(x) = x
is usually the most e ient tool to investigate the onvergen e to the asymptoti errordistribution
ρ
∞
ε
.Fora oupleof systems, translations onthe torus (a regular system) and the
Bernoullimap (a haoti system), analyti al results are available also for the
transient. Choosing the probability distribution
η(ξ) =
1
2
χ
[−1,1]
(ξ)
we found thatthedelityfortranslationsisgivenby[1℄F
ε
n
=
X
k∈Z
Ψ
k
Φ
−k
S
n
(kε)
S(x) =
sin(2πx)
2πx
(5)where
Ψ
k
andΦ
k
are theFourier omponentsoffun tionsΨ
andΦ
, whilefor theBernoullimapT x = qx
mod 1,withintegerq ≥ 2
wehaveF
ε
n
=
X
k∈Z
Ψ
k
Φ
−k
S
n,q
(kε)
S
n,q
(x) =
n−1
Y
j=0
S(q
j
x)
(6)Intherst asethede ayofdelityisexponential,withtimes ale
ε
−2
,whilein
these ond asewehaveaplateauoflength
n
∗
∝ − ln ε
,belowwhi hdelityis almost onstant,followedbyanε
independentsuper-exponentialde ay. Below thethresholdn
∗
theerrorprobabilitydistribution anbeapproximatedbyaδ
fun tion,andtheperturbedsystem onsideredasequivalenttotheunperturbedone. Theasymptoti errordistribution is thesamefor thetwosystems(sin e
thephysi alandstationarymeasure oin idewithLebesgue),andresultstobe
thetriangularfun tion(gures1and2). Thetransitionbetweenthe
δ
andthe triangularfun tions issharpandε
independentfortheBernuollisystem,while0
25
50
75
100
n
0.001
0.01
0.1
1
δ
F
0
10
20
n
0.0001
0.001
0.01
0.1
1
δ
F
Figure1: Comparisonbetweenanalyti alresultsandMonte Carlointegralsfor
thede ayof delity. Left: translations,
ε = 0.1
(analyti al result: bluedashed line;MonteCarlo: ir les). Right:3x
mod 1.ε = 10
−4
in bla kand
ε = 10
−8
in green(analyti alresult: dashed lineand rosses;Monte Carlo: ir les). We
used
N = 10
7
integrationpointsintheMonteCarlomethod. Thetestfun tions
Ψ = Φ
∗
weredened inFourierspa eas
Ψ
k
= k
−1
, trun atedatk = 30
.-1
-0,5
0
0,5
1
0
25
50
75
100
∆
ρ
-1
-0,5
0
0,5
1
0
5
10
∆
ρ
-1
-0,5
0
0,5
1
0
0,5
1
∆
ρ
Figure 2:ρ
n
ε
for3x
mod 1,ε = 10
−8
. Left:
n = 13
; enter:n = 15
, right:itisgradualand
ε
dependentforthetranslations(see[1℄).Our numeri al study of maps for whi h analyti al results are not available
(Hénon,Baker's,Intermittent,Logisti andStandardmaps)[1℄showsthat the
behavior of translations and of Bernoulli maps anbe onsidered as a
proto-typeof,respe tively,regularand haoti maps. Inparti ularforallthestudied
haoti mapsispossibletoidentifyathreshold
n
∗
∝ − ln ε
belowwhi hthe per-turbedsystem anbe onsideredasfaithfultotheunperturbedone;athresholdfollowedbyan
ε
independent,super-exponentialde ay. Tothisthreshold orre-spondsasharptransitionoftheerrorprobabilitydistributionfromaδ
fun tion totheasymptoti distribution(that ingeneralisnotatriangularfun tion butdependsontheinvariantandstationarymeasures).
3 Numeri al noise
Sin e real numbers have to be represented as strings of bits on a omputer,
to ea h dis rete map
T
orresponds a numeri al mapT
∗
. The a tion of the numeri almapdependsonthelengthofbitstringsusedtorepresentrealnum-bersandonthedetails ofroundoalgebra,whi harehardwaredependent[4℄,
nevertheless somegeneral results anbe stated. Inour omputational devi e
anyrealnumber
x ∈ R
isrepresentedbyaoatingpointnumberx
∗p
,that isa
binarystringwith
p
signi antbitsx
∗p
= f 2
e
f = ±0.f
1
f
2
. . . f
p
=
p
X
k=1
f
k
2
−k
e = ±
q
X
k=1
e
k
2
k
where
f
k
, e
k
are0or1. Thesuxp
willbenegle tedwheneverthereisno am-biguity. Theresultofanyarithmeti operationsu hasx
∗p
⊕ y
∗p
orresponding
to
x + y
is aoating pointnumberz
∗p
∈ F
p
and sum⊕
implies a round o. Therelativeerrorr
p
isdenedbyx
∗p
= x(1 + r
p
)
|r
p
| ≤ ǫ = 2
−p
Forinstan ein oatingpointsimplepre isionwehave
ǫ = 2
−25
. We anwrite,
usingthenotationintrodu edforadditivenoise,
T
∗
x = T
ε
x = T x + εξ(x)
whereξ
is now determined in a deterministi way by the initial ondition andε
is a onstant whose magnitude is of the order of the last signi antbit used torepresent
x
. The iterated mapT
n
∗
anbewritten asT
n
ε
using the previously introdu ednotation,butinthis asethesinglesteperrorsξ
i
willben
dierent fun tions oftheinitial onditionx
. Noti ethatT
∗
is dened bytheround o rules of the omputer, and the notationT
n
∗
orrespond to applyn
times the mapasdened bythoseround orules;whileεξ
n
= x
∗
n
− T (x
∗
n−1
) = T
∗
n
(x) − T T
∗
n−1
(x)
(7)
is introdu ed to ompare the round o results with the previous results for
additivenoise. Weintrodu ealsotheglobalerror
∆
n
∗
whi h anbewrittenasTheerror
∆
n
∗
is the umulativeresultofthe lo al errorforn
iterationsof the map. It is interesting to ompare these denitions with the lo al and globalerrorswhen the perturbation to the map
T (x)
is given by theadditive noise. Inthis asethemapT
∗
(x)
isrepla ed byT
ǫ
= T (x) + ǫ ξ
sothat lettingx
n
∗
=
T
n
ǫ
(x) = (T (x) + ǫ ξ
n
) ◦ . . . (T (x) + ǫ ξ
1
)
wehaveǫξ
n
= T
ǫ
(x
n−1
∗
) − T (x
n−1
∗
) = T
ǫ
n
(x) − T T
ǫ
n−1
(x)
Inthis ase we an write an expli it expression for theglobal error, whi h at
rstorder in
ǫ
reads∆
n
ε
= ǫ
n
X
k=1
DT
n−k
(x
1
) ξ
k
+ O(ǫ
2
)
(8)where
DT (x)
isthe ja obianmatrixof themapT
, theξ
k
arerandom ve tors andthe hainrule appliesto omputeDT
n
. Forthe translationsonthetorus
DT = 1
andwe anwrite(8)as∆
n
ε
= ǫ(nξ + w
n
)
(9)where
ξ ≡ hξi
istheaverageofthesto hasti pro essandw
n
aretheu tuations whose absolute value is of order 1. As a onsequen e in this ase we havelim
n→∞
∆
n
ε
/n = ǫξ
. Therandomperturbationswe onsider haveusually zero meansothatξ = 0
;thismaynotbethe aseofroundoerrorswhoseaverage anbenonzero. Inthe aseoftheBernoullimapsDT
isa onstantandif the pro essξ
haszero meanagainξ = 0
unlikethe ase ofround errorsdis ussed below. Wealsooutlinethatwhendealingwithroundoerrors,thesummationin eq. 8 starts with
k = 0
be ause also the initial errorx
0
− x
0
∗
is present. Herex
0
∗
standsforthenitea ura yrepresentationoftheinitial onditionx
0
,
whi h is in generaldierent from
x
0
. As we will see, this may have relevant
onsequen es.
4 Generalizations of Fidelity
Thisobservableisusedto hara terizestatisti allythespreadbetweentheorbit
anditssto hasti perturbation. Aspreviouslyintrodu ed(3)delitywasdened
byanintegraloverallthepossiblesinglesteperrorrealizations
ξ
i
. Nevertheless, from anumeri alpointof view, integralswere performed with aMonte Carlomethod,i.e. hoosing
N
representativerandomve tors(x, ξ
i
)
,apro edurethat ledtoarelativeerroroforderN
−1/2
(seegure1).
Wesupposethatifthedeterministi
ξ
i
(x)
fun tionaldependen eofsinglestep errorson the initial onditionis omplex enough, theve tor(x, ξ
i
(x))
an be onsidered as equivalent to a random sequen e, i.e. the Monte Carlo integraloverinitial onditionsandnoise anbesubstitutedforroundobyanintegral
omputethedelityerrorforasystemperturbedwithnumeri alnoiseas
F
∗
n
=
Z
X
Ψ(T
n
x)Φ(T
∗
n
x)dm(x)
(10)δF
∗
n
= F
∗
n
−
Z
X
Ψ(T
n
x)dm(x)
Z
X
Φ(T
∗
n
x)dm(x)
(11)Thisdenition requirestheknowledgeoftheexa tmap
T
, whi h isin general notavailable. This problem anbesolvedby omparingtheroundomapT
∗
, realizedwithagivenpre ision,i.e. asastringofbitsof agivenlength,with amap realizedat anhigher pre ision,
T
†
, that we allthe referen e map. For exampleT
∗
ouldbeasinglepre ision(8digits)map,andT
†
adoublepre ision (16digits)map. Thenumeri ally omputeddelitywillthusbeF
∗
n
=
Z
X
Ψ(T
†
n
x)Φ(T
∗
n
x)dm(x)
(12)To he ktherelevan eoftheseresultswe an omparethemwiththoseobtained
substituting
T
†
withT
‡
(forexampleamap realized using24 or32signi ant digits). If the results do not depend on the pre ision of the referen e map,we an assumethat they areequivalent to those that ould be obtainedifwe
had a ess to the exa t map. We have applied this pro edure to the results
shown in this paper. Typi ally, the resultsobtainedusing asreferen e map a
doublepre ision
T
†
ora24(32)digitmapT
‡
areequivalentbelowagiventime s ale. This time s ale, that orrespondsto the times ale under whi hT
†
an be onsidered equivalenttoT
‡
, as anbe he kedthroughadire t omparison betweenT
†
andT
‡
,is onsiderablylongerthanthetimes aleatwhi htheerror probability distribution ofT
∗
has rea hed its asymptoti form and thus the resultsof thesingle-double pre ision omparison anbe onsidered equivalenttothosethatwouldbeobtainedbya omparisonbetweenasinglepre isionand
anexa tmap.
Anotherway to avoid the problems related to the ina essibility of the exa t
mapwouldbetorely,forinvertiblemaps,onadierentdenitionofdelityas
˜
F
∗
n
=
Z
X
Ψ(x)Φ(T
∗
n
T
∗
−n
x)dm(x)
(13) whereT
−1
∗
isthenumeri al realizationof theinversemap, whi h is in general dierentfromtheinverseofT
∗
. Notethatinthedeterministi settingandwith invertiblemapsT
∗
preservingthemeasurem
, eq. (13)is equivalentto (3)[5℄. We generalize it by integrating on given dierent realizations respe tively ofT
−n
∗
andT
n
∗
, and by supposing that this integral onvergestoR Ψdm R Φdm
withthesamerateasthedelityerror(11). Forthenumeri alrealizationofaninvertiblemapasthestandardmap,theequivalen ebetweenthetwodenitions
hasbeen he kedwithgoodresults,atleastinthe haoti regime(gure 3).
0
10
20
n
0,01
0,1
1
F
δ
Figure3: De ayofdelityforthestandardmapwith
K = 10
,ε = 10
−8
,Monte
Carlointegralsusing
N = 10
6
;redand rosses: omparisonbetweensingleand
doublepre ision,bla kand ir les: singlepre isionusinginversion.
2GHz Intel Core Duo on a Ma OSX v10.5 (Darwin 9) operating system, a
pro essorIntelPentium 43.00GHzonaLinux Fedora4OS, and apro essor
1500MHZ IP35 ona IRIX 6.5 OS (whennot spe ied, the results shown in
guresareobtainedonthe Darwinar hite ture). Thedetails of theround o
pro ess depend on the ar hite ture, nevertheless our studies show that some
generalrulesabouttheerrordistribution anbestated.
5 Regular maps
Forregularmapsthebehaviorissigni antlydierentfromadditivenoise,
show-ingthat theintegralovertheonlyinitial onditionsis notequivalentforthese
systemstoanintegralalsoonthenoise.Fortranslationsonthetorus,
T x = x+ω
mod1,whi haretheprototypesforintegrablemaps,wehavefoundthatispos-sibletowritetheglobalerror
∆
n
∗
(x, ω) = T
n
x − T
∗
n
x
asin (9)∆
n
∗
(x, ω) = ε(ξ(ω)φ(n) + w
n
(x)),
φ(n) = n + φ
0
+ φ
1
n
−1
+ ...
(14) Hereε
is a onstant that represents the last signi ant bit a ura y,ξ
is a onstant that depends in a non trivial (and ma hine dependent) way onω
, whi haswewillseeisequivalenttoanaverageofthesinglesteperrorsξ
i
,whilew
n
isabounded,periodi fun tionwhi hdependsontheinitial onditionx
and haszeroaveragewithrespe tto theinitial onditions. Noti ethat inthis aseξ
isnotzeroin ontrastwiththe aseofrandomnoise. Wethushavelim
n→∞
∆
n
∗
(x, ω)
0
25
50
75
100
n
-1,5e-06
-1e-06
-5e-07
0
∆
0
250
500
750
1000
n
-1e-05
-5e-06
0
∆
Figure 4:∆
n
∗
(x, ω)
in translations on the Torus for two dierent, randomly hosen,initial onditionsx
(redandbla k), omparedwithanaverageover10
4
values of
x
(green).ω =
√
2/10
. The two gures show the same results on dierenttimes ales.Figure 4shows the global error
∆
n
∗
(x, ω)
for twodierent initial onditionsx
and the sameω
, on two dierent time s ales, showing that the ontributions ofw
n
goes to zero in the limit. Theperiodi nature of thesingle step errorsξ
i
(and thus ofw
n
) is shown by gure5 (periodi ity emerges after somenon periodi iterations). Theasymptoti averageerrorξ(ω)
is learly orrelatedto theroundoerrorδω = ω − ω
∗
ofthetranslationparameterasshownbygure 6. Thisgureshowsatleft thedistribution ofξ(δω)
andat righttheξ
andδω
distributionsobtainedusingaMonteCarlosamplingoftheparameterω
. From thegure at left we an see thatξ(δω)
is givenby adistribution law entred aroundξ = δω
. Thedistributionlawfortheroundo erroronthetranslation parameterη(δω) = 2
p−1
∞
X
k=p
χ
[−2
−
k
,2
−
k
]
(δω)
where
χ
I
(x)
denotesthe hara teristi fun tion oftheintervalI
,isstepwiseas a onsequen e of the exponent in the nite pre isionrepresentations of reals,with a ura y
ǫ = 2
−p
The stepwise shape of
η(δω)
is learly ree ted also in the distributionlaw forξ
, asshown in gure 6. These results suggestthat anuniform integration overω
implies an integration overξ
. This integration overξ
should orrespondmoreorless toarandomorpseudo-randompro ess, sin eis relatedto atrun ation of theless signi ativedigits. The probabilitydistributionfor
ξ
followinganuniformintegrationforω
isnotuniformasshown ingure6,butin thefollowingdis ussionswewilloftenapproximateitwithauniformone.
Thepresen eofanonzeroaverage ontributionofnoise(15)leadstoaquite
dierent situation with respe t to that obtained with random white noise, in
whi hthis ontributionisnotpresent.
Even if these results do not show a qualitative dependen e on the hardware
0
25
50
75
100
i
-2e-08
0
2e-08
4e-08
ξ
ε
Figure5: Singlestep errors
ξ
i
, fortwodierent, randomly hosen,initial on-ditionsx
. TranslationsontheTorus,ω =
√
2/10
.-1e-07
0
-5e-08
0
5e-08
1e-07
4e+07
8e+07
η(δω)
η(εξ)
_
Figure6: Left:
ξ(δω)
distribution(translationontheTorus). Right: omparison betweentheξ
(red)andδω
(bla k)distributions. Bothguresareobtainedwith aMonteCarlosamplingoverω
(10
5
valuesfortherstgure,
10
7
forthese ond
0
2500
5000
7500
10000
n
-4e-06
-2e-06
0
∆
0
2e+05
4e+05
6e+05
n
0
5e-13
1e-12
1.5e-12
2e-12
<w >
2
Figure7: Left:∆
n
∗
(x, ω)
fortwodierent,randomly hosen,x
values(bla kand red), omparedwith anaverageover10
4
valuesof
x
(green).ω =
√
2/5
. Maprealizedasa2D rotation. Right: varian eof
w
n
forω =
√
2/10
, averageover10
4
dierentinitial onditions,maprealizedasa2Drotation. Bla kanddashed
orrespondsto theresultsobtainedontheDarwin ar hite ture,while redand
ontinuousto thoseobtainedontheIRIXar hite ture.
anbefoundsee therightframeof gure7),they andependstronglyon the
algorithm. Theresults just shown, for whi h the period of
w
n
is of order 100 andits magnitudeof order 10,where obtainedby implementing themap asatranslationonthe torus
x
′
= x + ω
mod 1. Fromananalyti al pointof view,
thisisequivalenttoa2Drotationona ir le,butfromanumeri alpointofview
thelatterleadsto asequen e
w
n
whi h hasinitially,foranumberofiterations of order10
5
, properties similar to a randomsequen e (i.e. its varian egrows
approximatelywithalinearlaw),seegure7. Forthetranslationonthetorus
ξ(ω)
isgivenbyh∆
n
∗
(x, ω)i
x
/n
withani ea ura yevenforlowvaluesofn
. Con erning the de ay of delity, for regular maps we distinguish betweeniso hronous maps, for whi h the frequen y is xed, and aniso hronous maps,
su h as the skew map on the ylinder
x
′
= x + ω(y)
,
y
′
= y
, for whi h the
frequen ydependsontheinitial onditions. Intheformer asedelitydoesnot
de ay, whilein the latterit hasapowerlawde ay. Forthe iso hronousmaps
thesystemis basi allyequivalentto adeterministi rotationwitha frequen y
ω + εξ
, and it an be easily shown that in this ase delity does not de ay, but os illates with onstant amplitude. Indeed, ignoring the ontribution oftheterm
w
n
, we anwrite thea tion ofn
iterationsof thenumeri al map asx
n
∗
= x
0
+ n(ω + εξ)
. Substituting intheFourieranalysis ofappendix Bin[1℄ andkeepingξ
xed(i.e.,notperformingtheintegralinthenoise),thisleadsto anos illatingphasetermbutnottoade ayterm(a tuallythepresen eoftheos illatingterm
w
n
an auseade ayofdelityforthehighestFouriermodes, those orrespondingtodistan esoftheorderofthemagnitudeofw
n
).Fortheaniso hronousmap with
ω(y) = y
(to whi h anymap withmonotoniω(y)
anberedu ed),apowerlawde ayforF
n
−F
0
isa tuallyobserved. Indeed themapinthis aseisT (x, y)
denedonthe ylinderT
× [a, b]
orthetorusT
T
x
= x + y
mod1
T
y
= y
andthedelitywithrespe ttoperturbedmapis omputedbyintegratingover
x, y
. IfT
∗
denotes the map with numeri al noise, the map be omesT
∗
(x, y)
whoseiterate anbewrittenx
n
∗
= x
0
+ ǫξ
x
(x
0
) + n(y
0
+ ǫξ
y
(y
0
))
y
n
∗
= y
0
+ ǫξ
y
(y
0
)
InwritingthisequationweusethepreviousresultsfortranslationsontheTorus,
were the role of
ξ
is nowplayedbyξ
y
and that ofω
byy
0
. Wehave already
ingnoredtheboundedterm
w
n
butwehaveespli itelywrotetheroundoerror ofx
0
forsimmetrywithrespe tto
y
0
. Letus onsider
Φ
andP hi
tobefun tions oftheonlyvariablex
,andusetheusualFourieranalysisapproa h. Theintegrals anbe omputed ifwerepla e, followingagainthe previousdis ussion,ξ
x
(x
0
)
and
ξ
y
(y
0
)
with
ξ
x
andξ
y
assuming that they are random variables ranging in[−1, 1]
sothat theintegration overx
0
, y
0
isrepla ed byan integrationoverx
0
, y
0
andξ
x
, ξ
y
. ThedelityF
n
isgivenbyF
∗
n
=
X
k∈Z
Ψ
k
Φ
−k
sin(2πnkε)
2πnkε
sin(2πkε)
2πkε
(16)Thisratherstrongassumptionprovidesaresultwhi hisingoodagreementwith
thenumeri al omputationsseegure8. A mapwhereatanystepboth
x
andy
where ae tedbyarandomerrorT
ǫ x
= x + y + ǫξ
x
T
ǫ y
= y + ǫξ
y
would givea ompletely dierent result. Indeedtaking into a ount that the
iterateoforder
n
isx
n
= x
0
+ ny
0
+ ε
n
X
j=1
ξ
j
x
+
n−1
X
j=1
(n − j)ξ
y
j
y
n
= y
0
+ ǫ
n
X
j=1
ξ
y
j
thefollowingresultforthedelityis found
F
n
∗
=
X
k∈Z
Ψ
k
Φ
−k
sin(2πkε)
2πkε
n n
Y
j=1
sin(2πjkε)
2πjkε
(17)Figure 8showsthat this resultagrees with the delity omputed numeri ally
for a randomly perturbed map and diersdrasti ally from the delity of the
mapwithroundonoise.
6 Chaoti maps
Fidelity de ay an be easily studied for haoti numeri almaps (for example
0
1e+05
2e+05
3e+05
4e+05
n
0
0.025
0.05
0.075
F
δ
0
5e+06
1e+07
n
0,03
0,06
0,09
δ
F
Figure8: Left: De ayofdelityfortheskewmap. Bla k: additivenoise,
om-paredwithitsanalyti alpredi tioneq17(red);greenroundonoise ompared
withitsanalyti alpredi tioneq16(blue). Right: omparisonofroundonoise
andanalyti alpredi tiononalongertimes ale.
N = 10
4
with
K ≫ 1
)and showsalwaysagood qualitativeagreementwiththe results obtainedforadditivenoise,i.e. thepresen eofathresholdbelowwhi hdelityis onstant and the error fun tion is qualitatively a
δ
fun tion (its support is manyorder of magnitude smaller than thesize of the phase spa e, asithap-pened in gure2). Beyondthis threshold,that we an all
n
∗
andthat grows as− ln ε
, i.e. linearly in the number of bits used to represent real numbers (gure9),theerrordistributionspreadsqui klyoveritsa essiblerange,as anbe he ked using also a Monte Carlo sampling of the error distribution. The
resultsfor haoti maps donotshowasigni antdependen e onthe hoi eof
thear hite ture.
Stating that the de ay law of delity is qualitatively similar to the one
ob-servedforadditivenoiseisnotenoughtoprovethatthesequen eofsinglestep
errors
ξ
i
(seeequation 7) anbe onsidered equivalenttoarandomone,aswe laimedinouransatz. Inordertobetterunderstandthenatureofthissequen ewearegoingto usetwoapproa hes. Therstapproa hwouldbetostudy the
distribution of
ξ
i
fordierenti
values. In the aseof additivenoisewe would havea ontinuous uniformdistribution identi al forea hi
, andthus themore dierent is the a tualξ
i
distribution from a ontinuous uniform one, the less valid is ourassumption. The se ond approa h onsists in studying in greaterdetailthede aylaw, omparingitwiththeoneforadditivenoise,toseeifsome
dieren e emerges. In order to do that, an be useful to study the de ay of
delity for quen hed noise. By quen hed noise we mean a perturbation that
usesthe same sequen e
ξ
i
for ea h initial ondition(the sequen e is obtained randomly hoosing thevaluesξ
i
but without integrating onthem). Ifour as-sumptionisvalid,thedelityforroundonoiseshouldbemoresimilartothatobtainedusing additive noisethan to the oneobtained using quen hed noise.
Letus startwiththemap
3x
mod 1. Fromadetailed observationofgure9is possibleto see, in parti ular forsinglepre ision,that somedieren ebetween0
10
20
30
n
0.0001
0.001
0.01
0.1
F
δ
Figure9: De ayofdelityfor
3x
mod 1representedin singlepre ision(bla k, ir les)and doublepre ision(red, diamonds) ompared to areferen emapT
†
using 32 digits; the results are ompared to the de ay of delity for random(additive)noisewith
ε = 2
−25
(blue, rosses)and
ε = 2
−53
(green,squares),the
valueof singleand doublepre isionlast signi antbits forreals onthe torus.
MonteCarlointegralswith
N = 10
7
roundo and additivenoiseis present. Forthis map, wehavefound that the
globalerror anbedes ribedas
∆
n
∗
(x) =
n
X
i=0
3
n−i
εξ
i
(x)
(18)(this equation, aseq. (14), doesnot takein a ount boundary ee ts) where
theinitial onditionroundo
ξ
0
hasastep-wise ontinuumspe trumη(εξ) =
2
p−1
P
∞
k=p
χ
[−2
−
k
,2
−
k
]
(εξ)
distribution on e sampled over the spa e of initial ongurations(duetotheexponentin theoatingpointrepresentation,2
−p
is
thevalueof theleastsigni antbit,that we analso onsiderasequivalentto
the onstant
ε
),whiletheξ
i
withi ≥ 1
haveadis retespe trum,whi hresults to be almost ompletely redu ed to zerofori ≥ 2
(i.e. norelevant errorsare madeafter therst iteration). (gure 10). Due to this ee t,whi his ausedbytheextremely simplealgorithmi nature ofthemap, the sequen eof single
steperrors annotbeusedasarepresentativesequen eforanintegraloverthe
noise.
Itisthanevidentthatonlytheinitial roundohasa ontinuous(evenifnot
uniform)distribution,and anbein someway onsidered asequivalentto the
additivenoise. Inthedis ussionoftheskewmap,wehaveseenthattheinitial
roundodeterminedthede aylawofdelity,andthisistruealsoforthemap
-4e-08
0
-2e-08
0
2e-08
4e-08
25000
50000
75000
1e+05
ξ
η
ε
-2e-07
-1e-07
0
1e-07
2e-07
0
1e+06
2e+06
3e+06
4e+06
ξ
η
ε
-2e-07
-1e-07
0
1e-07
2e-07
0
2e+06
4e+06
6e+06
8e+06
ξ
η
ε
Figure10: Singlesteperrordistributionforthemap
3x
mod1insinglepre ision. Left:ξ
0
; enterξ
1
;rightξ
2
. Monte CarlosamplingwithN = 10
7
theskew map, that the integralontheinitial ondition
x
0
is equivalent toan
uniformintegralontheinitialroundoerror,andletusignorethe ontribution
of therest of thesingle steperror sequen e,
ξ
i
withi ≤ 1
. . It is possible to show using Fourier analysis that the de aylawfor the delity of aqx
mod 1 mapperturbedattheonlyinitialstep,withanuniformintegralovernoiseisδF
ε
n
=
X
k6=0
Ψ
k
Φ
−k
sin(2πεkq
n
)
2πεkq
n
(19)(thisresultistriviallyobtainedfollowingtheFourieranalysispro edurein
ap-pendix Bof [1℄ and performing only therst integral over noise). The de ay
law(19)des ribesbetterthede aylaw fordelity dueto roundo noisethan
to(6),showingthatthemain ontributiontode aydelityisduetotheinitial,
ontinuousspe trum perturbation (gure 11). Toanalyse the ontribution of
ξ
i
withi ≤ 1
letusremovetheinitial round o ( hoosing rationalinitial on-ditionsthat anberepresentedexa tlywiththe pres ribedpre ision). Inthisasedelitypresentsathreshold,butdoesnotde aytozero(afterthethreshold
theerrordistributionexpandsto thewhole phasespa ebut doesnot onverge
to an asymptoti distribution, gure 12). In this ase, i.e. when only initial
onditionpointsthat anberepresentedexa tlyusingthenumeri alpre ision
under examinationand thus theinitial roundo isremoved,the delity
os il-latesafterthethresholdwithoutgoingtozero(gure13). Isnotsurprisingthat
this behaviour is quite similar to that obtained studying the quen hed noise
perturbationofthismap. Indeedit anbeshown,(generalizingaresultin[6℄),
that thedelity for quen hed noise os illates but does notde ayto zero. We
havefound that thisresultisparti ularly di ulttoverifyusing MonteCarlo
numeri alintegration. Asitisshowningure13,athemap3
x
mod1studied with numeri alpre isionε
1
and perturbed withnumeri al noiseε
2
, starts os- illatingin orresponden ewiththedelitythresholdforε
2
(inagreementwith theanalyti alresult)butthendropstozeroin orresponden ewiththedelitythresholdfor
ε
1
(seegure13, omparingwithgure9toknowthesingleand doublepre isionthresholds).A ordingtousthisresultsupportsour on lusion0
10
20
n
0,001
0,01
F
δ
Figure11: De ayof delityfor
3x
mod1. Bla k, ir les: analyti alresult(6), usingε = 2
−25
. Red, squares: analyti alresult(19). Green, diamonds: round
onoise(single pre ision omparedtodoublepre ision),
N = 10
7
-1
-0.5
0
0.5
1
0
5
10
15
20
∆
ρ
Figure12: Errordistribution
ρ
n
obtained omparingasinglepre ision
3x
mod1 mapwithadoublepre isionmap,usinginitial onditionsinsinglepre ision(noinitialround o). Bla k,
n = 30
; red,n = 31
; greenn = 32
; ompared to the triangular fun tion (blue) obtainedusing initial ondition in double pre ision(initialroundo). MonteCarlosamplingsusing
N = 10
7
0
20
40
n
0
0,025
0,05
0,075
δ
F
Figure 13: De ay of delity for
3x
mod 1. Bla k, diamonds: quen hed noise omputedwith 32 digits(ε = 2
−25
). Red, squares: quen hed noise omputed
with double pre ision (
ε = 2
−25
). Green, ir les: de ay of delity obtained
omparingasinglepre isionmapwithadoublepre isionone(roundonoise)
usinginitial onditionsthat anbeexa tlyrepresentedin singlepre ision,i.e.
removingtheinitialroundo. MonteCarlointegrals,
N = 10
6
thatthedelityde aythresholdisagoodmeasureforthemaximumnumberof
iterationsunder whi hthenumeri al omputationsare ompletely reliable.
Forthe other, (slightly)morealgorithmi ally omplexmaps, the sequen eof
error
ξ
i
has a ontinuous spe trum for ea h value ofi
, and onverges qui kly toanasymptoti distribution(gure 14). Forallthesesystemstheasymptotierrordistributionresultsto bealmostindistinguishablefromthat obtained
us-ingadditivenoise(gure 15),suggestingthat forthesemapstheassumptionof
equivalen ebetweenroundoandadditivenoise ouldbevalid(forthesemaps
theinitialround oerrorplaysnospe ialrole).
Inordertobetterverifythisassumption,wehavestudiedthede ayofdelity
formaps perturbed with quen hed noise, i.e. withoutperformingthe integral
overnoise. Nevertheless,forallthetested haoti maps(withtheex eptionof
theBaker'smap)thede aylawsforquen hedandadditivenoisewerenot
dis-tinguishable,at least using the omputationalpre isionallowedbyour Monte
Carlomethod. For theBaker'smap, is possible to see that the delity de ay
ismoreirregularforquen hednoise(duetotheabsen eofthesmoothing
inte-gral). Inthis asethede ayusingroundonoiseismoresimilartotheadditive
-2e-07
-1e-07
0
1e-07
2e-07
100
1000
10000
1e+05
ξ
η
ε
Figure14: Distributionofsinglesteperrors
ξ
i
fortheHénonmap,x
observable,N = 10
7
. Bla k,
i = 0
. Red,i = 1
. Green,i = 20
. Bluei = 50
(greenandblue areverydi ulttodistinguish)-3
-2
-1
0
1
2
3
0
0.1
0.2
0.3
0.4
0.5
∆
ε
8
ρ
ε
8
Figure 15: Asymptoti error distribution for the Hénon map,
x
observable. Bla k, additive noise (ε = 10
−8
), red omparison between single and double
pre ision(thetwofun tionsarealmostidenti al). MonteCarlosamplingsusing
N = 10
7
0
10
20
30
40
n
0,001
0,01
F
δ
Figure 16: De ay of delity for the Baker'smap. Bla k, ir les:the sequen e
errors of order
2
−25
is the same for ea h initial ondition (quen hed noise),
showinganirregularde aylaw;green,squares: singlepre isionroundonoise,
showingthesamesmoothde ayofadditivenoisewith
ε = 2
−25
(red, rosses).
MonteCarlointegralswith
N = 10
7
7 Con lusions
Wehaveusedtheresultsofapreviousworkonadditivenoisetostudytheee ts
ofroundo noiseondis rete systems. Forregularmapsthebehaviordepends
onthe algorithmi realization and on its hara ter: for iso hronousmaps the
delityerrordoesnotde ay, foraniso hronousmapsithasapowerlawde ay,
whereasin presen e of anadditive noisethede ay isalwaysexponential. For
haoti systemsroundoandnoisearealmostequivalent. Wealsoshowedthat
thedelityisane ienttoolforthestudyofthetimes alesforthe onvergen e
totheasymptoti errordistribution,sin e,atleastfor haoti maps,itallowsus
tond athresholdvaluebelowwhi h thenumeri alsystem anbe onsidered
asequivalenttotheexa tone. Thisthresholdlinearlygrowsasthenumberof
Acknowledgments
G. Tur hetti a knowledges nan ial support from the PRIN 2007 grant
Sim-ulazione della produzione di elettroni dalla interazione laser-(pre)plasma nel
ontestodellaignizionevelo e oordinatedbyD. Batani
Referen es
[1℄ P.Marie,G.Tur hetti,S.Vaienti,F.Zanlungo,Chaos 19,043118(2009)
[2℄ G.Tur hetti,S.VaientiandF.Zanlungo,a eptedonEurophysi sLetters
[3℄ Y. Kifer,Ergodi TheoryofRandomTransformations,Birkhäuser,1986
[4℄ D.KnuthTheartof omputerprogrammingVol2,Addison-Wesley,(1969)
[5℄ G.Benenti,G.Casati,G.Velbe,Phys. Rev. E,670,055202-1,(2003)
[6℄ C. Liverani, Ph.Marie, S. Vaienti, Random Classi alFidelity, Journal of
Statisti alPhysi s,128,4,1079,(2007)
[7℄ C.Grebogi,S. Hammel,J.Yorke,J.Complexity, 3,136,(1987)
[8℄ T.Sauer,Physi al ReviewE, 65,(2002),036220
[9℄ G.Tur hettiinCHAOS,SystèmesDynamiques,p.239,HermannÉditeurs,