ebook img

Auto-correlation function and frequency spectrum due to a super-position of uncorrelated exponential pulses PDF

0.59 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Auto-correlation function and frequency spectrum due to a super-position of uncorrelated exponential pulses

Auto-correlation function and frequency spectrum due to a super-position of uncorrelated exponential pulses O. E. Garcia∗ and A. Theodorsen Department of Physics and Technology, UiT The Arctic University of Norway, N-9037 Tromsø, Norway (Dated: February2,2017) 7 1 Abstract 0 2 The auto-correlation function and the frequency power spectral density due to a super-position of un- b e F correlated exponential pulses are considered. These are shown to be independent of the degree of pulse 1 overlapandtherebytheintermittencyofthestochasticprocess. Forconstantpulsedurationandaone-sided ] h exponentialpulseshape,thepowerspectraldensityhasaLorentzianshapewhichisflatforlowfrequencies p - and apower law at high frequencies. The algebraic tail is demonstrated to result from the discontinuity in m s the pulse function. For a strongly asymmetric two-sided exponential pulse shape, the frequency spectrum a l p isabroken powerlawwithtwoscaling regions. Inthecaseofasymmetricpulse shape, thepowerspectral . s c density is the square of a Lorentzian function. The steep algebraic tail at high frequencies in these cases i s y is demonstrated to follow from the discontinuity in the derivative of the pulse function. A random distri- h p bution of pulse durations is shown to result in apparently longer correlation times but has no influence on [ theasymptotic powerlawtailofthefrequency spectrum. Theeffectofadditional randomnoise isalsodis- 1 v cussed,leadingtoaflatspectrumforhighfrequencies. Theprobability densityfunctionforthefluctuations 5 0 1 isshowntobeindependentofthedistributionofpulsedurations. Thepredictionsofthismodeldescribethe 0 0 varietyofauto-correlationfunctionsandpowerspectraldensitiesreportedfromexperimentalmeasurements . 2 inthescrape-off layerofmagnetically confinedplasmas. 0 7 1 : v i X r a 1 I. INTRODUCTION Theboundaryregionofmagneticallyconfinedplasmasisgenerallyfoundtobeinaninherently fluctuating state with order unity relative fluctuation levels of the plasma parameters.1–10 The frequencypowerspectraofthesefluctuationsaregenerallyfoundtobecharacterizedbyaflatlow- frequencyregionandasteephigh-frequencytail.11–27 Inmanypreviousworks,powerlawscaling relationshipshavebeeninferredfromexperimentalmeasurementsandtheresultsinterpretedinthe frameworks of scale-free inertial range cascade, self-similar processes and self-organized critical behavior.11–21 By contrast, it will in this contribution be demonstrated that the auto-correlation functions and frequency spectra reported from experimental measurements can be described by a super-positionofuncorrelated exponentialpulses. Large-amplitude fluctuations in the boundary region are attributed to the radial motion of blob-like plasma filaments through the scrape-off layer, most clearly demonstrated by gas puff imagingdiagnostics.22–35 Duringtheirradial propagation,theblob-likestructuresdevelopasteep front and a trailing wake, which can also be observed in numerical simulations of isolated blob structures36–50 and simulations of scrape-off layer turbulence.51–61 Conditional averaging of ex- perimental measurement data have shown that large-amplitude fluctuations are well described by an exponentialwave form.62–83 This universalobservationof front steepening in interchangemo- tionsofplasmafilamentsmotivatesthepresent studyofexponentialpulses. The statistical properties of large-amplitude fluctuations in the scrape-off layer of tokamak plasmas have recently been elucidated by means of exceptionally long data time series under stationaryplasmaconditions.78–84 From single-pointmeasurementsit has been demonstratedthat there is an exponential distribution of both the peak amplitudes and the waiting times between them. Moreover, the average duration time does not depend on the amplitudeand also appears to berobustagainstchangesinplasmaparameters.75–84 Basedontheseproperties,astochasticmodel for the plasma fluctuations has been developed by a super-position of uncorrelated exponential pulses.85–90 The underlying assumptions and predictions of this model have recently been found tocomparefavorably withexperimentalmeasurements.78–84 In this contribution, the auto-correlation function and frequency spectrum are derived and dis- cussed in detail for one- and two-sided exponential pulses. These are shown to be independent of the amplitude distribution of the pulses as well as the degree of pulse overlap and thereby the intermittency of the stochastic process. For constant pulse duration and a one-sided exponential 2 pulse shape, the power spectral density has a Lorentzian shape, which is flat for low frequencies andapowerlawspectrumathighfrequencies. Foratwo-sidedexponentialpulseshape,thepower spectral density is the product of two Lorentzian functions. The power law tails at high frequen- cies are shown to result from discontinuitiesin the pulse function or its derivative. A distribution of pulsedurationtimes is demonstratedto result in apparently longercorrelation timesbut has no influenceon thepowerlawtailofthefrequencyspectrum. This paper is organized as follows. In the following section the stochastic model describing fluctuations due to a super-position of uncorrelated pulses is presented. In Sec. III the mean and variance of the random variable are calculated and the intrinsic intermittency of the process is discussed. General expressions for the auto-correlation function and the power spectral density are derivedin Sec. IV. Thecases ofone- and two-sidedexponentialpulsefunctionswithconstant pulse duration is considered in Sec. V. In Sec. VI the algebraic tail in the frequency spectra are demonstratedto resultfromthediscontinuityin thepulsefunctionoritsderivative. A distribution of pulse durations is in Sec. VII shown to result in apparently longer correlation times but has no effect on the asymptoticpowerlaw tail in the frequency spectrum. The contribution of additional random noiseis discussedin Sec. Cand finallya discussionof theresultsand theconclusionsare giveninSec.VIII.TheprobabilitydensityfunctionsinthecaseofexponentialandLaplaceampli- tudedistributionsarederivedin AppendixA. A discussionoftherelationbetween discontinuities inthepulseshapeoritsderivatviesandpowerlawspectraisgiveninAppendixB.Finally,therole ofadditionalrandom noiseisdiscussedinAppendixC. II. STOCHASTICMODEL Considerastochasticprocess givenby thesuper-positionofa randomsequence ofK pulsesin atimeintervalofdurationT,85–94 K(T) F (t)= (cid:229) A f t−tk , (1) K k t k=1 (cid:18) k (cid:19) whereeachpulselabeledk ischaracterized byanamplitudeA ,arrivaltimet ,anddurationt ,all k k k assumed to be uncorrelated and each of them independent and identically distributed. The pulse arrivaltimest areinthefollowingassumedtobeuniformlydistributedonthetimeintervalunder k consideration,thatis,theirprobabilitydensityfunctionis givenby1/T. Thepulsedurationtimes t k are assumed to be randomly distributed with probability density Pt (t ), and the average pulse 3 durationtimeisdefined by ¥ t d =ht i= dtt Pt (t ). (2) 0 Z Here and in the following, angular brackets denote the average of the argument over all random variables. TheresultspresentedhereareindependentofthedistributionofpulseamplitudesP (A), A it is only assumed that the mean hAi and variance hA2i are finite. The role of the pulse amplitude distributionis discussedfurtherin AppendixA. Thepulseshapef (q )istakento bethesameforalleventsinEq. (1). Thisfunctionisnormal- ized suchthat ¥ dq |f (q )|=1. (3) −¥ Z Thus,forconstantdurationeachpulsecontributesequallytothemeanvalueofF (t). Theintegral K ofthen’thpowerofthepulseshapeisdefined as ¥ I = dq [f (q )]n, (4) n −¥ Z for positive integers n. It is assumed that T is large compared with the range of values of t for which f (t/t ) is appreciably different from zero, thus allowing to neglect end effects in a given realization of the process. Furthermore, the normalized auto-correlation function of the pulse functionisdefined by90 1 ¥ r f (q )= dcf (c )f (c +q ), (5) I2 −¥ Z and theFouriertransformofthisgivesthefrequency spectrum, ¥ 1 ̺f (J )= dqr f (q )exp(−iJq )= |j |2(J ), (6) −¥ I2 Z wherethetransformofthepulsefunctionisdefined by ¥ j (J )= dqf (q )exp(−iJq ). (7) −¥ Z In this contribution, the auto-correlation function and the power spectral density for the process defined by Eq. (1) will be investigated for the case of exponential pulses. The influence of var- ious distributions of the pulse duration times as well as additive random noise will be explored. However,first thetwolowestordermomentsoftheprocess willbederived. III. MEAN,VARIANCEANDINTERMITTENCY Thetwo lowestordermomentsofarealizationofthestochasticprocessdefined by Eq.(1)can becalculated inastraightforward mannerby averagingoverallrandom variables.85–93 4 A. Meanvalue Starting with thecase of exactly K eventsin a timeintervalwith duration T, themean valueis givenby ¥ ¥ T dt hF Ki= dA1PA(A1) dt 1Pt (t 1) 1 −¥ 0 0 T Z Z Z ···Z−¥¥ dAKPA(AK)Z0¥ dt KPt (t K)Z0T dTtK k(cid:229)=K1Akf (cid:18)t−t ktk(cid:19), (8) using that the pulse arrival times are uniformly distributed. Neglecting end effects by taking the integrationlimitsfor thearrivaltimest inEq. (8)to infinity,themeanvalueofthesignalfollows k directly, K hF i=t I hAi . (9) K d 1 T Herea changeofintegrationvariabledefined by q =(t−t )/t has been made,giving k k ¥ ¥ t−t ¥ ¥ dt kPt (t k) dtkf t k = dt kt kPt (t k) dqf (q )=t dI1. (10) Z0 Z−¥ (cid:18) k (cid:19) Z0 Z−¥ Takingintoaccount thatthenumberofpulsesK isalsoarandom variableandaveraging overthis as wellgivesthemean valueforthestationaryprocess, t hF i= d I hAi, (11) t 1 w where t =T/hKi is the average pulse waiting time. For a non-negative pulse function, I = 1, w 1 themean valueoftheprocess isgivenby theaverage pulseamplitudeand theratiooftheaverage pulsedurationand waitingtimes. Themean valuevanishesboth foranti-symmetricpulseshapes, I =0,andforpulseamplitudedistributionswithvanishingmean,hAi=0. Itshouldalsobenoted 1 thatthemean valuesdoesnotdepend on thedistributionofpulsedurationtimes. B. Variance The variance can similarly be calculated by averaging the square of the random varible. From thedefinitionofF (t)itfollowsthat K F 2(t)= (cid:229)K (cid:229)K A A f t−tk f t−tℓ . (12) K k ℓ t t k=1ℓ=1 (cid:18) k (cid:19) (cid:18) ℓ (cid:19) 5 Averagingthisoverpulseamplitudesas wellas durationand arrivaltimesfor fixedt and K gives ¥ ¥ T dt hF 2Ki= −¥ dA1PA(A1) 0 dt 1Pt (t 1) 0 T1 Z Z Z ···Z−¥¥ dAKPA(AK)Z0¥ dt KPt (t K)Z0T dTtK k(cid:229)=K1ℓ(cid:229)=K1AkAℓf (cid:18)t−t ktk(cid:19)f (cid:18)t−t ℓtℓ(cid:19). (13) There are two contributionsto the variance from thedouble sum. When k6=ℓ there are K(K−1) termswiththevalue ¥ ¥ T dt t−t Z−¥ dAkPA(Ak)Z0 dt kPt (t k)Z0 Tk Akf (cid:18) t k k(cid:19) ¥ ¥ T dt t−t ×Z−¥ dAℓPA(Aℓ)Z0 dt ℓPt (t ℓ)Z0 TℓAℓf (cid:18) t ℓ ℓ(cid:19), (14) whilefork=ℓthereare K termsgivenby theintegral ¥ ¥ T dt t−t Z−¥ dAkPA(Ak)Z0 dt kPt (t k)Z0 TkA2kf 2(cid:18) t k k(cid:19). (15) Neglecting end effects due to the finite duration of individual pulses by extending the integration limitsfort andt toinfinity,thevarianceforlargeT in thecase ofexactlyK pulsesisgivenby k ℓ K K(K−1) hF 2i=t I hA2i +t 2I2hAi2 . (16) K d 2 T d 1 T2 By averagingoverall realizationswherethenumberofpulsesK isstatisticallydistributedgives t hF 2i= d I hA2i+hF i2, (17) t 2 w wherehK(K−1)i=hKi2 hasbeenassumed. ThisisanexactrelationforaPoissondistributionof the number of pulses, and approximatewhen there is a large number of pulses in each realization oftheprocess. Itfollowsthatthestandarddeviationorrootmeansquare(rms)valueoftherandom variableis givenby t F 2 = d I hA2i. (18) rms t 2 w Thus, the absolute fluctuation level is large when there is significant overlap of pulse events, that is, for long pulsedurations and short pulse waitingtimes. As for the mean valuegiven above, the variancedoesnot dependon thedistributionofthepulsedurationtimes. 6 C. Intermittency Theratio oftheaverage pulsedurationand waitingtimes, t g = d, (19) t w determinesthedegreeofpulseoverlapinthestochasticprocess. Wheng issmall,pulsesgenerally appear isolated in realizations of the process, resulting in a strongly intermittent signal where most of the time is spent at small values. When g is large, there is significant overlap of pulses and realizations of the process resemble random noise. Indeed, it can be demonstrated that the probabilitydensityfunctionfortherandomvariableF (t)approachesanormaldistributioninthe K limitofinfinitelylargeg , independentlyofthepulseshapeand amplitudedistribution.85,86 Foranon-zero mean valueoftheprocess,therelativefluctuationlevelisgivenby F 2 1I hA2i rms = 2 . (20) hF i2 g I2hAi2 1 This is large for long pulse waiting times and short pulse durations. In Appendix A it is shown that alsotheskewnessand flatness momentsbecomelarge forsmallvaluesof g . In thefollowing, therescaled variablewithzero meanand unitstandarddeviationwillbefrequentlyconsidered, F −hF i F (t)= . (21) F rms e Some realizations of this process are presented in Fig. 1 for one-sided exponential pulses with constant duration and exponentially distributed amplitudes. It is clearly seen that the signal is strongly intermittent for small values of g , while pulse overlap for large values of g makes the signals resemble random noise. Further description of the intermittency effects in this process is givenin AppendixAand are discussedin Refs. 85–90. IV. CORRELATIONSANDSPECTRA In the same way as the variance was calculated above, the auto-correlation function for F (t) K can be calculated by a straight forward average over all random variables and the power spectral densityisthen givenbya transformationtothefrequency domain. 7 5 g =102 g =10 g =1 g =10-1 5 0 ) 0 5 t ( ~ F 5 0 0 00 2200 4400 6600 8800 110000 t/t d FIG. 1. Realizations of the stochastic process for a one-sided exponential pulse shape with constant dura- tion t and exponentially distributed pulse amplitudes. The degree of pulse overlap is determined by the d intermittency parameterg =t /t . d w A. Auto-correlation function Considering first the signal F (t) defined by Eq. (1), the auto-correlation function for a time K lag r isgivenby adoublesum, ¥ ¥ T dt hF K(t)F K(t+r)i= dA1PA(A1) dt 1Pt (t 1) 1··· −¥ 0 0 T Z Z Z Z−¥¥ dAKPA(AK)Z0¥ dt KPt (t K)Z0T dTtK k(cid:229)=K1ℓ(cid:229)=K1Akf (cid:18)t−t ktk(cid:19)Aℓf (cid:18)t−ttℓℓ+r(cid:19). (22) Thereareagaintwocontributionstothedoublesum,comprisingK(K−1)termswhenk6=ℓgiven by hAi2 (cid:229)K ¥ dt kPt (t k) T dTtk f t−t tk ¥ dt ℓPt (t ℓ) T dTtℓf t−ttℓ+r , (23) k,ℓ=1Z0 Z0 (cid:18) k (cid:19)Z0 Z0 (cid:18) ℓ (cid:19) k6=ℓ and K termswhen k=ℓ, hA2i (cid:229)K ¥ dt kPt (t k) T dTtkf t−t tk f t−ttk+r . (24) k=1Z0 Z0 (cid:18) k (cid:19) (cid:18) k (cid:19) Neglecting end effects due to thefinite pulseduration by taking theintegrationlimitsfort andt k ℓ toinfinityforthecaseofexactlyK pulseeventsresultsin K(K−1) K ¥ hF K(t)F K(t+r)i=t d2I12hAi2 T2 +I2hA2iT dtt Pt (t )r f (r/t ). (25) 0 Z 8 Averaging over the number of pulses occurring in the interval of duration T, it follows that the auto-correlationfunctionforthestationaryprocessis givenby90 1 ¥ RF (r)=hF i2+F 2rmst dtt Pt (t )r f (r/t ), (26) d Z0 wherer f (q )isthenormalizedauto-correlationfunctionforthepulsefunctiondefinedbyEq.(5). The expression of the auto-correlation function is simplified by considering the rescaled variable defined by Eq.(21), ¥ 1 RF (r)= t dtt Pt (t )r f (r/t ). (27) d 0 Z It is emphasized that this expresseion for the auto-correlation function does not rely on a Poisson distribution of the number of pulse events. The auto-correlation function is determined by the pulseshapethroughr f (q ) andtheprobabilitydensityfunctionforthepulsedurationtimes. Equation(26) emphasizes theroleof thepulseshape in determiningthe auto-correlationfunc- tion for the process. However, the time lag r in the integrand can be transfered to the distribution function for thepulsedurations timesby thechange of variables q =|r|/t , which givesthealter- nativeformulation 1 ¥ dq r RF (r)= t q 2 q |r|Pt (|r|/q )r f (q ), (28) d 0 Z wheretheprobabilitydensityfeunctionforthepulsedurationtimesis normalizedsuchthat ¥ ¥ dq dt Pt (t )= |r|Pt (|r|/q )=1. (29) q 2 0 0 Z Z In the case of constant pulse duration, the latter is given by the degenerate distribution Pt (t ) = d (t −t ), where d is the delta distribution. The auto-correlation function is then R (r) = d F r f (r/t d), that is, it is simply given by the normalized auto-correlation function of the individ- e ual pulsesintheproecss. B. Powerspectraldensity From the auto-correlation function given above, thepower spectral density follows directly by aFouriertransformationto thefrequency domain,90 ¥ 1 ¥ W F (w )= drRF (r)exp(−iw r)=2p hF i2d (w )+F 2 dtt 2Pt (t )̺f (tw ), (30) rmst Z−¥ dZ0 where w is the angular frequency and ̺f (J ) is the Fourier transform of the normalized auto- correlation function defined by Eq. (6). Here the first term on the right hand side of the second 9 equality is a zero-frequency contribution from the mean value of the random variable. The ex- pression for the power spectral density is simplified by considering the rescaled random variable defined by Eq.(21), 1 ¥ W F (w )= t dtt 2Pt (t )̺f (tw ). (31) dZ0 e It should be noted that this frequency spectrum is independent of the amplitude distribution P A and does not depend on the intermittency parameter g . The latter property is evidently due to the assumption of uncorrelated pulses. Moreover, the above expression is not restricted to a Poisson distributionfor the number of pulses K(T). The only assumptionsmade are that the pulse arrival timeshaveauniformdistributionandthatthetwolowestordermomentsfortheprocessarefinite. Inthespecialcaseofconstantpulseduration,theexpressionforthepowerspectraldensitybecome W F (w ) = t d ̺f (t dw ), that is, the spectrum is simply determined by that of the pulse function f (q ). e By the simple change of variables, J = t |w |, Eq. (31) for the power spectral density can be writtenin thealternativeform 1 ¥ dJ J 2 W F (w )= t |w |w 2 Pt (J /|w |)̺f (J ), (32) dZ0 e wheretheprobabilitydensityfunctionforthepulsedurationtimesis normalizedsuchthat ¥ ¥ dJ dt Pt (t )= Pt (J /|w |)=1. (33) |w | 0 0 Z Z Equation (32) is suggestive of a power law spectrum for sufficiently broad distributions of pulse durations. In general, the power spectral density is determined by both the pusle shape and the distributionfunctionforpulsedurations.90 V. EXPONENTIALPULSES The above expressions for the auto-correlation function and the power spectral density were derived for an arbitrary pulse function under the assumption that the two lowest order moments exist. In this section, the properties of the stochastic process will be investigated for the case of one-and two-sidedexponentialpulseswithconstantduration. 10

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.