ebook img

Fractional advection-dispersion equations for modeling transport PDF

15 Pages·2009·0.39 MB·English
by  
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 Fractional advection-dispersion equations for modeling transport

JOURNAL OF GEOPHYSICAL RESEARCH, VOL.114,F00A07, doi:10.1029/2008JF001246,2009 Click Here for Full Article Fractional advection-dispersion equations for modeling transport at the Earth surface Rina Schumer,1 Mark M. Meerschaert,2 and Boris Baeumer3 Received6January2009;revised3June2009;accepted25August2009;published18December2009. [1] Characterizing the collective behavior of particle transport on the Earth surface is a key ingredientindescribing landscapeevolution. We seekequations thatcaptureessential features of transport of an ensemble of particles on hillslopes, valleys, river channels, or river networks, such as mass conservation, superdiffusive spreading in flow fields with large velocity variation, or retardation due to particle trapping. Development of stochastic partial differential equations such as the advection-dispersion equation (ADE) begins withassumptionsabouttherandombehaviorofasingleparticle:possiblevelocitiesitmay experience in a flow field and the length of time it may be immobilized. When assumptions underlying the ADE are relaxed, a fractional ADE (fADE) can arise, with a non-integer-order derivative on time or space terms. Fractional ADEs are nonlocal; they describe transport affected by hydraulic conditions at a distance. Space fractional ADEs arise when velocity variations are heavy tailed and describe particle motion that accounts for variation in theflow field over theentire system. Time fractional ADEs arise as a result of power law particle residence time distributions and describe particle motion with memory in time. Here we present a phenomenological discussion of how particle transport behavior may be parsimoniously described by a fADE, consistent with evidence of superdiffusive and subdiffusive behavior in natural and experimental systems. Citation: Schumer,R.,M.M.Meerschaert,andB.Baeumer(2009),Fractionaladvection-dispersionequationsformodelingtransport at theEarthsurface,J.Geophys.Res., 114, F00A07,doi:10.1029/2008JF001246. 1. Introduction thantheaveragemotionduetospatiotemporal variationsin themechanisms inducingtheirmotion. Tracermotions thus [2] An important class of problems in the Earth surface maybeconsideredasconsistingofquasi-randomwalkswith sciences involves describing the collective behavior of restperiods. particles in transport. Familiar examples include transport [3] During motion, tracers may experience characteristi- ofsoluteandcontaminantparticlesinsurfaceandsubsurface callydifferent‘‘hop’’or‘‘jump’’behaviors.Considerreleas- waterflows,thebehaviorofsoilparticlesandassociatedsoil ing at some instant an ensemble of tracers within a system, constituentsundergoingbiomechanicaltransportandmixing say, marked particles within a river. Assuming all tracers by bioturbation, and the transport of sediment particles and undergo and remain in motion, then due to turbulence sediment-bornesubstancesinturbulentshearflows,whether fluctuations in the case of suspended particles, or to effects involving shallow flows over soils, deeper river flows, or of near-bed turbulence excursions together with tracer-bed oceancurrents.Theseandmanyotherexamplessharethree interactionsinthecaseofbedloadparticles,duringasmall essential features. First, is the behavior of a well defined intervaloftimeDt,mostofthetracersmoveshortdistances ensemble of particles. These particles may be considered downstream whereas a few move longer distances. The ‘‘tracers’’ whose total mass is conserved or otherwise displacement(orhop)distanceDxgenerallymaybeconsid- accountedforifradioactive decay,physicaltransformations ered a random variable. If the probability density function or chemical reactions are involved. Second, these tracers (pdf)describinghoplengthhasaformthatdeclinesatleastas typically alternate between states of motion and rest over fast as an exponential distribution, then it possesses finite manytimescales,andindeed,mosttracersofinterestinEarth meanandvariance.Moreover,afterfinitetimet,thecenterof surfacesystemsareatrestmuchofthetime.Third,whenin massofthetracersiscenteredatthepositionx=x +vt,where transport,sometracersmovefaster,andsomemoveslower, 0 x isthestartingpositionandvisthemeantracerspeed.The 0 spread or ‘‘dispersion’’ around average tracer position, de- 1Division of Hydrologic Sciences, Desert Research Institute, Reno, scribedbystandarddeviation,grows ataconstantratewith Nevada,USA. time s = (Dt)1/2 where D is the dispersion coefficient. 2Department of Statistics and Probability, Michigan State University, EastLansing,Michigan,USA. As described below, this behavior, know as Fickian or 3Department of Mathematics and Statistics, University of Otago, Boltzmann scaling, may be characterized as being ‘‘local,’’ Dunedin,NewZealand. inthatduringasmallinterval,tracersinmotionmostlymove from location x to nearby positions. Conversely, during the Copyright2009bytheAmericanGeophysicalUnion. sameinterval,tracersarrivingatpositionxmostlyoriginate 0148-0227/09/2008JF001246$09.00 F00A07 1 of 15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 fromnearbypositions.LettingC(x,t)denotethelocaltracer tion that reflects the presence of unusually ‘‘long’’ rest concentration,theninthislocalcasethebehaviorofCmaybe periods for tracers in storage, then as described below a adequately described by the classic advection-dispersion suitable phenomenological description of the behavior of equation(ADE). the tracer concentration C(x, t) may be obtained from a [4] Consider, in contrast, the possibility that the hop fractionalADE(fADE)withafractionaltimeterm@@gtCg,where length pdf declines with distance in such a way that the 0<g<1.Dispersionnowscalesatarateproportionaltotg/2. varianceisundefined,forexample,apowerlawdistribution. Because s grows at a rate slower than normal Fickian In this case, long displacement distances, albeit relatively dispersion, this behavior is referred to as ‘‘subdiffusion.’’ rare, are nonetheless more numerous than with, say, an Thischaracterizesa‘‘longmemory’’effect,analogoustoan exponentialdistribution,sothathoplengthdensitypossesses influence from far upstream as described above with refer- a so-called ‘‘heavy tail.’’ For example, during downstream ence to the fractional space derivative; the fractional time transport,sedimenttracersdispersedbyturbulenteddiesand derivative characterizes an influence from far back in time. bymomentaryexcursionsintopartsofthemeanflowthatare Like variations in tracer speed, rest times contribute to either faster or slower than the average may occasionally dispersion. As such, these effects may not be separable experience‘‘superdispersive’’eventsduringexcursionsinto withoutancillaryinformation. sites with unusually high flow velocities [Bradley et al., [6] As elaborated below, if both the hop length density 2009]. In these situations where the hop length density is andthewaitingtimedensityareheavytailed,thenthefADE heavytailed,thescalingofdispersionisnon-Fickian,thatis, can include both a time and a space fractional derivative. s=(Dt)1/a,where1<a<2.Becausesthereforegrowsata Dispersion now scales at a rate proportional to tg/a. Note ratefasterthan‘‘normal’’Fickiandispersion,thisbehavioris that a particular value of the ratio g is a nonunique com- a referredtoassuperdiffusive.Moreover,thisbehaviormaybe binationofvaluesofaandg.Thismeansthatiftheratiois characterized as being ‘‘nonlocal’’ in that during a small obtainedempiricallyfrommeasurementsoftracerdispersion, interval, tracers released from position x mostly move to say, in a flume or river, additional information, empirical nearby positions, but also involve an ‘‘unusually large’’ and/or theoretical, is needed to distinguish whether the number of motions to positions far from x. Conversely, source of the anomalous dispersion resides with the hop duringthesameinterval,tracersarrivingatpositionxmostly lengths, waiting times, or both. On this note we point out originatefromnearbypositions,butalsoinvolveasignificant that the development below provides a mostly phenome- numberormotionsoriginatingfarfromx.Thismeansthatthe nological viewpoint of how tracer behavior may be parsi- behaviorataparticularpositiondoesnotnecessarilydepend moniouslydescribedbyafADE,consistentwithevidence onlyonlocal(nearby)conditions,butrathermayalsodepend ofsuperdiffusiveandsubdiffusivebehaviorinnaturaland on conditions upstream or downstream. For example, be- experimental systems [e.g., Benson et al., 2001, 2000a, causethelocalrateofchangeintracerconcentrationconsists 2000b; Haggerty et al., 2002; Schumer et al., 2003]. A of the divergence of the flux of tracers, then if the flux is particularly exciting opportunity resides in clarifying the relatedtothebedstress,say,inthecaseofsedimenttransport, underlying physics that lead to these behaviors, including thepossibilityexiststhatchanges inthelocalconcentration themechanismsthatgiverisetoheavytaildistributionsof dependonstressconditions,andthereforeonsystemconfig- hopdistancesandwaitingtimes,perhapsbuildingonrecent uration,‘‘far’’upstream.Asdescribedbelow,thebehaviorof insights from the hydrologic literature. In section 3, we thetracerconcentrationC(x,t)inthisnonlocalcasemaybe look at details underlying the application of fractional describedbyafractionalderivativeversionoftheadvection- advection-dispersion equations for describing collective dispersionequation. behavior of particle transport at the Earth surface. Much [5] Considernowanensembleof tracers whosemotions, of this manuscript is a review of basic information about afterrelease,involvestatesofmotionandrest.Inthecaseof fADE for nonspecialists that is not available in a single tracerparticlesmovingasbedload,reststatesmightinvolve reference.Newmaterialfillsintheoreticalgaps,compares momentary disentrainment on the bed surface associated solution characteristics andapplicationof various fADEs, with turbulence fluctuations, or burial beneath the surface and links landscape evolution processes to the conceptual for longer periods followed by reentrainment with scour modelsunderlyingfADEs. [Nikora et al., 2002; Parker et al., 2000]. Even longer rest periods might consist of storage of tracers within the bars and floodplain of a river [Malmon et al., 2002, 2003]. 2. Fractional Derivatives Similarly,resttimesofnutrientsinopen-channelflowmight [7] Before presenting fractional ADEs, we provide some consist of temporary excursions into so-called dead zones, useful introductory material about fractional derivatives. or into hyporheic zone storage, with eventual release back Three characteristics of fractional derivatives will be used into the main fluid column [Haggerty et al., 2002; Gooseff to interpret and solve fractional ADEs: (1) the Grunwald et al., 2003]. With bioturbation of soil particles, rest times fractional difference quotient that approximates a fractional consist of the intervals between displacement by biotic derivative;(2)thefractionalderivativeastheconvolutionof activity. The rest time (or waiting time, or residence time) aninteger-orderderivativewithamemoryfunction;and(3) may be considered a random variable. If random waiting theFourierandLaplacetransformsoffractionalderivatives. times are, in general, small, then the speed V may be consideredanaverage‘‘virtualspeed’’thatincludesperiods 2.1. Grunwald Finite Difference of motion and rest, and the local rate of change in tracer [8] Howcanwegeneralizeourunderstandingofderivatives concentration, @C, in the ADE accommodates this. If, in to include non-integer-order cases? The first derivative of a @t contrast the waiting time density is a heavy-tailed distribu- function y = f(x) is approximated by the difference quotient 2 of15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 Figure1. (a) Theusualderivative islocal,but(b)the fractional derivative isnonlocal.(c) Thelog-log plot demonstrates the power law decay in Grunwald weights. Dy/DxwhereDx=handDy=f(x)(cid:2)f(x(cid:2)h)(Figure1a).The The Grunwald definition of the fractional derivative is the secondderivativef00(x)isapproximatedbyD2y/Dx2where noninteger variant of (2) D2y¼D½fðxÞ(cid:2)fðx(cid:2)hÞ(cid:3) da fðxÞ(cid:4) 1 X1 (cid:2)a(cid:3)ð(cid:2)1Þjfðx(cid:2)jhÞ; ð3Þ ¼½fðxÞ(cid:2)fðx(cid:2)hÞ(cid:3)(cid:2)½fðx(cid:2)hÞ(cid:2)fðx(cid:2)2hÞ(cid:3) dxa ha j j¼0 ¼fðxÞ(cid:2)2fðx(cid:2)hÞþfðx(cid:2)2hÞ ð1Þ where the fractional binomial coefficients are theseconddifference.Continuinginthismannershowsthat the nth-order derivative is approximated by the nth finite (cid:2)a(cid:3) Gðaþ1Þ difference quotient ¼ j Gðjþ1ÞGða(cid:2)jÞ dn fðxÞ(cid:4) 1 Xn (cid:2)n(cid:3)ð(cid:2)1Þjfðx(cid:2)jhÞ; ð2Þ and the approximation becomes exact as h = Dx ! 0. To dxn hn j seethattheinteger-orderformula(2)isreallyaspecialcaseof j¼0 thefractionalformula(3),notethatthebinomialcoefficients are0=1/1forj>n,sothatwecouldhavewrittenthesumin where the binomial coefficients (from Pascal’s triangle) (2)withtheupperlimitofinfinity. (cid:2)n(cid:3) n! Gðnþ1Þ ati[v9]efSoellvoewraflroinmtertehsetiGngrupnrwoapledrtdieesfinoiftitohne(f3r)a.cMtioonsatlimdeproivr-- ¼ : j j!ðn(cid:2)jÞ! Gðjþ1ÞGðn(cid:2)jÞ tantisthatafractionalderivativeisanonlocaloperatorthat 3 of15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 depends on function values far away from x (Figure 1b). Thea-orderfractionalderivativeofebxisbaebxandsoforth. The importance given to a far away point x (cid:2) jh is Thepointisthatfractionalderivativesarenaturalanalogues measured by the Grunwald weight w = ((cid:2)1)jG(a + 1)/ oftheirinteger-ordercousins. j (G(j + 1) (cid:5)G(a(cid:2)j)) so that w =1,w =(cid:2)a,w =a(a(cid:2) 0 1 2 1)/2 and w = w (cid:5) (j (cid:2)a)/(j + 1). An appeal to Stirling’s j+1 j approximation [Meerschaert and Scheffler, 2002] shows 3. Classical Random Walk, Central Limit that w (cid:4) j(cid:2)(1+a)/G((cid:2)a) (Figure 1). Hence the Grunwald j Theorem, and the ADE definitionisessentiallyadiscreteconvolutionwithapower law. The finite difference formula (3) is also the basis for [13] Sediment transport has long been modeled as a numerical codes to solve the fractional ADE [Tadjeran et stochastic sequence of hops and rest periods [Einstein, al., 2006]. 1937; Sayre and Hubbell, 1965] leading to emergent prop- erties after long time. It is also common to represent Earth 2.2. Convolution of the Derivative With a Power Law surfacetransportprocessusingdiffusion-typemodelsarising Memory Function fromthecombinationofacontinuityequationwithtransport [10] ThecontinuumlimitoftheGrunwaldfinitedifference (flux)laws(sinceCulling[1960]).Randomwalkanddiffu- formulaisaconvolutionintegral[see,e.g.,Meerschaertand sion models are related. Specifically, diffusion equations Tadjeran, 2004]. Writing (3) in terms of the Grunwald govern the stochastic processes that arise when the scaling weightsandusingw (cid:4)j(cid:2)(1+a)/G((cid:2)a)suggests limitofrandomwalkistaken.Herewedescribethetypeof j particlemotionthatmaybewellmodeledbyaclassicalADE da fðxÞ(cid:4) 1 X1 fðx(cid:2)jhÞðjhÞ(cid:2)a(cid:2)1h sboetchleaatrt.hWeeexcaacntanpapturroeacohfothuerafrdavcetciotinoanl-gdeisnpeerrasliioznatieoqnusatwioilnl dxa Gð(cid:2)aÞ j¼0 fromeitheraLagrangianorEulerianpointofview.Byusinga 1 Z 1 (cid:4) fðx(cid:2)yÞy(cid:2)a(cid:2)1dy ð4Þ Lagrangian(followingtheprogressofanindividualparticle Gð(cid:2)aÞ 0 withtime)approach,theexactnatureofthetransportprocess isclarified. whichcanbewritteninvariousequivalentforms[Samkoet al., 1993; Mainardi, 1997]. If f(t) is defined on t (cid:6) 0 then 3.1. Lagrangian Approach to the ADE integration by parts yields the Caputo form [14] Ourgoalistodescribethemotionofanensembleof particlesasmeasuredbyitsconcentration(ormass)inspace da 1 Z t fðtÞ¼ f0ðt(cid:2)yÞy(cid:2)ady ð5Þ and time C(x, t). The distribution of particles within the dta Gð1(cid:2)aÞ 0 plumewillberepresentedbythepdfgoverningthelocation of a single particle in space. The discrete stochastic model which is very useful for fractional time derivatives. Noting we choose to develop this distribution is the random walk. that the integral (5) is a convolution we can write We write the random particle location X(t) as sum of sta- tionary random jump lengths Y separated by time steps of i dafðtÞ dfðtÞ t(cid:2)a length Dt: ¼ ? ð6Þ dta dt Gð1(cid:2)aÞ t=Dt X XðtÞ¼ Y : whichrelatesthefractionalderivativetoapowerlawmemory n n¼1 function. Then we can use the Central Limit Theorem (CLT) to 2.3. Transforms of Fractional Derivatives estimate the distribution of X(t) at long time. The law of [11] For partial differential equations, it is useful to use large numbers (LLN) says that the average of independent FouriertransformsinspaceandLaplacetransformsintime. and identically distributed (IID) random variables with Sincethetransform ofapower lawisanotherpowerlaw,it mean m converges to the theoretical average: is easy to check using (4) and (6) the Fourier and Laplace transforms pairs: Y þ(cid:5)(cid:5)(cid:5)þY 1 n(cid:2)m!0 ð8Þ n dafðxÞ $ðikÞa^fðkÞ dxa ð7Þ as n becomes large. The CLT refines this statement by dgfðtÞ$sg~fðsÞ(cid:2)sg(cid:2)1fð0Þ approximatingthedeviationbetweenthesetwoterms.Itsays dtg that when appropriately rescaled, the sum of IID random using t(cid:2)g/G (1 (cid:2) g) $ sg(cid:2)1. variableswithmeanmandstandarddeviationsconvergesto astandardnormaldistribution: 2.4. Examples of Fractional Derivatives [12] UsetheLaplacetransformpairt(cid:2)g/G(1(cid:2)g)$sg(cid:2)1 Y1þ(cid:5)(cid:5)(cid:5)þYn(cid:2)nm!Nð0;1Þ: alongwiththeLaplacetransform(7)tocheckthatthea-order sn1=2 fractionalderivativeoftpistp(cid:2)aG(p+1)/G(p+1(cid:2)a)which WeusethistheoremtoapproximatethesumofIIDparticle reducestotheusualformwhenaisaninteger.TheFourier jumps transform(7)canbeusedtoshowthatthea-orderfractional derivativeofsinxissin(x+pa/2),whichreducestotheusual formulaforintegerausingstandardtrigonometricidentities. Y1þ(cid:5)(cid:5)(cid:5)þYn(cid:4)nmþsn1=2Nð0;1Þ: ð9Þ 4 of15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 that governs particle location after a number of jumps. The right-handsideof(10)iscalledBrownianmotionwithdrift. Itcombinesadeterministic advectivedriftwithastochastic diffusion term Z(t) = (2Dt)1/2N(0, 1) that spreads like t1/2 (Fickianscaling).TheADEsolutionC(x,t)givenby(11)is theprobabilitydensityoftheprocessvt+Z(t)thatrepresents the location of a randomly selected particle at time t > 0, assumingthatparticlelocationx=0attheinitialtimet=0. Figure 2a illustrates the simple random walk. Figure 2b showsthesamerandomwalkatalongertimescale.Figure2c shows the Brownian motion that emerges as the long-time scaling limit of this simple random walk. The continuous graphoftheparticlepathisarandom(notgeneratedfroma deterministic pattern) fractal with Hausdorff dimension 3/2 [Mandelbrot, 1982]. The abrupt jumps seen in the random walk disappear in the scaling limit. Figure 3 illustrates the probability density C(x, t) of particle locations for the Brownianmotionlimitprocess. 3.2. Eulerian Approach to the ADE [15] Using an Eulerian approach, we choose a specific location in space and describe particle motion through that location with time. From this perspective, we begin with a conservation of particle mass equation that equates the rate ofmasschangeatalocationwiththedifferencebetweenthe mass of particles entering and leaving: @Cðx;tÞ @Fðx;tÞ n ¼(cid:2) ; ð12Þ @t @x where effective porosity n = 1 for transport at the surface, particleconcentrationCisinmasspervolume,andfluxFin mass per area per time. If particle flux is assumed to be by advection and Fickian dispersion: @C F ¼vnC(cid:2)nD ; ð13Þ @x where v is average particle velocity and D is a dispersion coefficient, then we obtain the advection-dispersion equa- tion (also known as the Fokker Planck equation [Feller, 1971]): @Cðx;tÞ @C @2Cðx;tÞ ¼(cid:2)v þD : ð14Þ @t @x @x Figure2. Comparisonofrandomwalktracesatincreasing time scale and in the scaling limit as a Brownian motion. (a) A simple random walk X(t) simulates particle motion. [16] ThesolutiontotheADEcanbeobtainedusingFourier transforms: (b) Simple random walk X(t) at a longer time scale. (c) Brownian motion as the scaling limit of a random walk X(t). Particle graph is a random fractal with dimension 3/2. @C^ðk;tÞ¼(cid:2)vðikÞC^ðk;tÞþDðikÞ2C^ðk;tÞ: ð15Þ @t Letthenumberofjumpsbethetotaltimedividedbytimeper Solvefor Cand useinitial conditions C(x, 0) =d(0) to find jumpn= t andrecasttheCLTtosuittherandomsum Dt Y þ(cid:5)(cid:5)(cid:5)þY (cid:4)t m þps t1=2Nð0;1Þ: ð10Þ C^ðk;tÞ¼eð(cid:2)vtðikÞþDtðikÞ2Þ ð16Þ 1 n Dt ffiDffiffiffitffiffi theFouriertransform of aGaussian density (11)withmean Letv= m andD= s2 andwefindaGaussiandensity Dt 2Dt m = vt and variance s2 = 2Dt. [17] It is interesting that the deterministic solution to a Cðx;tÞ¼pffiffiffiffiffi1ffiffiffiffiffiffiffiffiffiffiffiffie(cid:2)ðx2(cid:2)(cid:5)2vDtÞt2 ð11Þ partial differential equation should beapdf. This important 2p(cid:5)2Dt observation means that random behavior exhibits a deter- 5 of15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 [20] 2. The characteristics of the flux at position x are bell-shapedbreakthroughcurves,leadingandtrailingedges that decay rapidly, breakthrough curve width that grows like x1/2, and area under breakthrough curve that remains constant. 4. Fractional-in-Space ADEs [21] A variety of field and theoretical studies suggest superdiffusivenonlocalityintransportoftracersattheEarth surface.Forexample,reanalysisoftracerstudiesinsandand gravel bed streams revealed hop length distributions with heavytails[Bradleyetal.,2009].E.Foufoula-Georgiouetal. (A non-local theory for sediment transport on hillslopes, submittedtoJournalofGeophysicalResearch,2009b)argue thattransportonhillslopesisnonlocalandthatsedimentflux mustbecalculatedusingnotjustthelocalgradient,butalso Figure 3. Brownian motion density function C(x, t) that of upslope topography. Stark et al. [2009] use a frac- describing particle spreading away from plume center of tionallyintegrated fluxtermtocapturenonlocaleffectsina massattimet=1,4,and9inthescalinglimit.Notethesquare modelofbedrock channelevolutionthatincludestheeffect rootspreadingrateandfasttaildecay. of hillslope sediment production on channel bed sediment buffering and bedrock erosion. These studies all suggest fractionalfluxtermstorepresentnonlocalityintransport. ministic regularity at long time. The location vt + Z(t) of a [22] K. M. Hill et al. (Particle size dependence of the probability distribution functions of travel distances of randomly selected particle is unknown, but even so, the gravel particles in bed load transport, submitted to Journal relative concentration of particles is predictable. We obtain ofGeophysicalResearch,2009)suggestthesourceofpower the same solution using the Eulerian perspective of Fickian law hop lengths in bed load transport. They present and particle jumps and conservation of solute mass, or the reviewexperimentaldatatojustifyanexponentialsteplength Lagrangian random walk formulation taken to its scaling for particles of a given size. Since mean step length varies limit by applying the CLT.The assumptions underlying the withparticlesize,theoverallsteplengthforallparticlesisa CLT must hold for a Gaussian limit to emerge. This means mixture of those exponential distributions. Using standard thatthereareassumptionsburiedinFickian/massconserving models (e.g., Gamma distribution) for the pdf of grain size ADEthatmustholdforittobeapplicable.Ourfocushereis leads to an overall step length pdf with a heavy power law that, in order for the ADE to emerge, there can be only tail. Even though neither the grain size distribution nor the moderate deviation from the average jump size. Since it is well known that the ADE and its solutions do not always reproduce essential features of sediment or solute transport (e.g., Benson [1998], Gooseff et al. [2003], Wo¨rman et al. [2007], Neuman and Tartakovsky [2009], and E. Foufoula- Georgiouetal.(Normalandanomalousdispersionofgravel tracerparticlesinrivers,submittedtoJournalofGeophysical Research,2009a), as well as referencesin thisspecial issue on stochastic transport and emergent scaling at the Earth surface), we will now consider models that allow less stringentassumptions.Wewillalsoexplorethelongmemory case where a time component is introduced by allowing random waiting times between jumps. In sections 4 and 5, we introduce fractional-in-space ADEs caused by heavy- tailed velocity distributions, and fractional-in-time ADEs causedbyheavy-tailedresidencetimes. 3.3. Characteristics of ADE Solutions [18] If tracer transport is well modeled by a classical ADE, thenlong-term transport characteristics (of apulseof tracer) will resemble the Green’s function (point source) solutions [Arfken and Weber, 1995] to traditional integer- order advection-dispersion equations (Figure 4). These characteristics include the following. Figure4. Comparisonofthesolutionsinspacetointeger- [19] 1. The spatial snapshot characteristics are Gaussian order and fractional ADEs at time t = 1,t = 10, and t =20. (symmetric bell-shaped) concentration profiles, plume For both models, v = 1 and D = 0.1. The heavy leading edges that decay rapidly, snapshot width that spreads like edgesofthea-stablesolutionstothe fractionalADE decay t1/2, and total mass that remains constant over time. as C(x) (cid:4) x(cid:2)a(cid:2)1. 6 of15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 steplengthpdfforgrainsofagivendiameterareheavytailed, suming that particle location x = 0 at the initial time t = 0. the mixture distribution turns out to have a heavy tail. This The diffusion term Z(t) = t1/aS (0, 1, a, b) exhibits non- a result shows that a fractional Exner equation (Foufoula- Fickianscaling,acharacteristicoftenseeninlaboratoryand Georgiou et al., submitted manuscript, 2009a) is applicable fieldstudies[Bensonetal.,2000a,2000b,2001].Sincea<2 tobedloadtransport.Italsohighlightstheneedforadditional thescalingfactort1/agrowsfaster(theplumespreadsfaster) experimentsandanalysistoimproveestimatesofgrainsize than the classical ADE case a = 2, so this model is often distribution, step length for moving particles, as well as called superdiffusive. Aside from its super-Fickian scaling, entrainment rate for particles of varying size. Here we Le´vy motion has other features that distinguish it from its developspacefractionalADEsusedtomodelsuperdiffusive close cousin, Brownian motion. Le´vy motion probability transport using both the Lagrangian (particle tracking) and densities are positively skewed, with a long leading tail Eulerian(conservationofmass)approach. [Samorodnitsky and Taqqu, 1994]. In fact, the probability massadistancerunitsormorefromtheplumecenterofmass 4.1. Lagrangian Approach to Fractional-in-Space fallsofflikex(cid:2)asothatparticlesaremuchmorelikelytorace ADEs aheadofthemean. de[n23s]ityIfpt(hxe),dtehnesnitytheofinjutemgprallensg2th=s RYi1is(tXhe(cid:2)heamv)y2p-t(axi)ldedx (cid:2)1 i 4.2. Eulerian Approach to Fractional-in-Space ADEs diverges. Then the classical CLT we use to determine the density of the sum of jump lengths X(t) does not hold. A [27] TheEulerianapproachaccountsforbulkmassmove- ments into and out of a finite size control volume of edge moregeneralformoftheCLTexists,however,andsaysthat sizeDx=hduringashorttimestepDt,asinFigure7.Ina the properly rescaled sum of stationary IID random varia- fractal porous medium (and perhaps, in a fractal river bles witheveninfinitevarianceconvergesindistribution to network), the velocity of different particles in the control ana-stabledensity,denotedS ,withmeanm,spreads,tail a volume can vary widely, so that a particle can enter the parameter a, and skewness b [Feller, 1971; Gnedenko and control volume from long distances upstream via high- Kolmogorov,1968].Hereweassumethatparticlejumpsare velocity paths. The fractal medium imposes a power law heavy tailed in the direction of flow only. velocitydistribution,leadingtoafractionalFick’sLawF= (cid:2)D@a(cid:2)1C/@xa(cid:2)1fordispersiveflux[Schumeretal.,2001], Y þ(cid:5)(cid:5)(cid:5)þY (cid:2)nm 1 n !S ðm¼0;s¼1;a;b¼1Þ: ð17Þ since dispersion comes from variations in velocity. The sn1=a a Grunwald weights in Figure 1 code the velocity distribu- tion. A fractional dispersive flux in the solute mass con- NotethattheGaussiandistributionisastablewitha=2(in servationequation(12)leadstoafractional-in-spaceADE which case the skewness is irrelevant), making this CLTa true generalization of the classical version. [24] As before, recast the generalized CLT (17) to @Cðx;tÞ¼(cid:2)v@Cðx;tÞþD@aCðx;tÞ: ð19Þ approximate the sum of random jumps: @t @x @xa Use Fourier transforms to solve the space fADE (19): Y þ(cid:5)(cid:5)(cid:5)þY (cid:4)nmþsn1=aS ð0;1;a;bÞ 1 n a m s ð18Þ Y1þ(cid:5)(cid:5)(cid:5)þYn(cid:4)tDtþDt1=at1=aSað0;1;a;bÞ @C^ðk;tÞ¼(cid:2)vðikÞC^ðk;tÞþDðikÞaC^ðk;tÞ @t and let v = m and D = sa to find an a-stable density C^ðk;tÞ¼exp½(cid:2)vtðikÞþDtðikÞa(cid:3) Dt Dt C^ðk;tÞ¼eð(cid:2)vtðikÞþDtðikÞaÞ which is the Fourier transform of an a-stable density with shift m = vt, spread sa = Dt, skewness b = 1, and tail in Fourier space. The a-stable density cannot be written in parameter 1 < a (cid:7) 2. More extreme deviations from the closed form in real space. A variety of methods exist to mean velocity are represented by heavier-tailed jump estimate the inverse transform numerically [Nolan, 1998]. distributions and, in turn, are governed by fractional ADEs [25] Figure 5a illustrates a simple random walk with with smaller order fractional derivatives on the flux term heavy-tailed particle jumps. Figure 5b shows the same [Clarke et al., 2005]. random walk at a longer time scale. Figure 5c shows the [28] AnalternativeEulerianderivationofthespacefADE stable Le´vy motion that emerges as the long-time scaling uses a fractional conservation of mass with a traditional limit of this heavy-tailed random walk. The graph of the Fickian flux [Meerschaert et al.,2006]. Foufoula-Georgiou particle path is a random fractal with dimension 2 (cid:2) 1/a. etal.(submitted manuscript, 2009a) developaprobabilistic Thelargejumpsseenintherandomwalkpersistinthescaling Exner equation for sediment transport, which is applicable limit. Figure 6 illustrates the probability density C(x, t) of whenparticlesteplengthsfollowaheavy-tailed(powerlaw) particle locations for the stable Le´vy motion limit process. pdf. This asymptotic governing equation for the movement Thisparticlemotionprocessexhibitsasuperdiffusivespread- of entrained particles is valid at scales long enough to ing rate, skewness, and power law right tail (early arrivals encompass many particle jumps. Using an active layer downstream). approach, neglecting porosity, and assuming equilibrium [26] The right-hand side of (18) is called a-stable Le´vy (steady,uniform)bedloadtransportofgrainsofuniformsize motionwithdrift.Thislimitprocessvt+Z(t)representsthe overabed,theprobabilistic Exner equationinthepresence location of a randomly selected particle at time t > 0, as- ofaheavy-tailedsteplengthisLa@fa ¼(cid:2)v@faþD@afa,where E @t @x @xa 7 of15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 Figure6. Le´vymotiondensityfunctionC(x,t)describing particlespreadingawayfromplumecenterofmassattime t=1,4,and9inthescalinglimit.Notethesuperdiffusive spreading rate, skewness, and power law tail. 4.3. Characteristics of Space fADE Solutions [29] Characteristics of the Green’s function solutions to the advection-dispersion equation with a fractional-order term in space (Figure 4) are as follows. [30] 1. The spatial snapshot characteristics are a-stable concentration profiles skewed with long right tail, plume leading edges that decay like a power law x(cid:2)a(cid:2)1, snapshot width that spreads like t1/a, and total mass that remains constant over time. [31] 2. The characteristics of the flux at position x are asymmetric breakthrough curves, long leading edge, break- through curve width that grows like x1/a, and area under breakthrough curve that remains constant. 5. Fractional-in-Time ADEs [32] TracerparticlesattheEarthsurfacespendmoretime at rest than in motion [Sadler, 1981; Leopold et al., 1964; Tipper, 1983]. For example, frequency distributions of bed loadpathlengtharepositivelyskewed,indicatingthatmany particlesdonotmovefromthepointoftracerinputoverthe measurementperiod[HassanandChurch,1992;Pryceand Ashmore, 2003]. Small deviations in waiting times will not affect long-term dispersion rates, but heavy-tailed waiting times will. This affects overall bed load rates [Singh et al., Figure5. Comparisonofheavy-tailedrandomwalktraces at increasing time scale and in the scaling limit as a stable Le´vymotion.(a)Aheavy-tailedrandomwalkX(t)simulates anomalous particle motion. (b) Heavy-tailed random walk X(t) at a longer time scale. (c) Stable Le´vy motion as the scaling limit of a random walk X(t). Particle graph is a random fractal with heavy-tailed jumps. Figure 7. When governed by a fractional-in-space ADE, f isthefractionoftracerparticlesintheactivelayer,v,Dare particle hop length during a small time step Dt may be a as in the space fADE (19), L is the thickness of the active much larger than the mean size. In finite difference imple- a layer,andEistheentrainmentrate.Dividingbothsidesby mentations for fractional ADEs, probability of particle hop L /E yields an equation mathematically identical to the length decays as a power law with distance [after Schumer a spacefADE(19),sothatthesamesolutionmethodsapply. et al., 2001]. 8 of15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 2009] and the resulting transport characteristics can be foru>0.Iff(x,u)isthepdfofu1/aZthentakingaweighted accommodated by time-fractional ADE, which will be average with respect to u [Meerschaert and Scheffler, developedfromboththeLagrangianandEulerianview.Evi- 2004]showsthatthelimitparticlelocationxattimethas denceforheavy-tailedwaitingtimesintransportiscaptured pdf inthedepositionalrecord[SchumerandJerolmack,2009].A promising avenue for future research is to examine waiting Z 1 Cðx;tÞ¼ fðx;uÞqðu;tÞdu ð22Þ times between entrainment events, to determine whether a 0 time-fractional Exner equation may be useful to incorporate particles outside the active layer, including deeply buried ascalemixtureofstabledensities.Since particles,thoseincorporatedintothehyporheiczone,andpar- ticles outside the normal flow path of a river (sandbars, ^fðk;uÞ¼eDuðikÞa ð23Þ floodplain)(e.g.,seediscussionbyFoufoula-Georgiouetal. ~qðu;tÞ¼sg(cid:2)1e(cid:2)usg (submittedmanuscript,2009a)). (fortheLaplacetransformformula,seeMeerschaertetal. [2002b]) a simple integration shows that the Fourier- 5.1. Lagrangian Approach to Fractional-in-Time LaplacetransformofC(x,t)is ADEs [33] A continuous time random walk (CTRW) imposes a sg(cid:2)1 rTahnedopmartwicaleitilnogcattiimoneaJtntbimefeorteitshe particle jump Yn occurs. Cðk;sÞ¼sg(cid:2)DðikÞa whichwillleadtothespace-timefractionalADE.Rearrange XðtÞ¼Y þ(cid:5)(cid:5)(cid:5)þY ; 1 n to get sgC(k, s) (cid:2) sg(cid:2)1 = D(ik)aC(k, s) and invert the LaplaceandFouriertransformsusing(7)toget where n = N(t) is the number of jumps by time t. Suppose the particle jumps have mean zero so that the centered (generalized) CLTapplies: @gCðx;tÞ @aCðx;tÞ ¼D @tg @xa Y þ(cid:5)(cid:5)(cid:5)þY (cid:4)n1=aZ; 1 n usingtheCaputoderivativeintime,wherethepointsource initial condition implies C^(k, t = 0) = 1. Adding an where Z is stable (normal if a = 2). The sum J + (cid:5)(cid:5)(cid:5) + J 1 n advectivedriftyieldsthespace-timefractionalADE gives the time of the nth particle jump. If the waiting times haveafinitemeanvthenthelawoflargenumbers(8)shows that the nth jumps happens at time t = Tn (cid:4) nn and @gCðx;tÞ¼(cid:2)v@Cðx;tÞþD@aCðx;tÞ ð24Þ thenX(t)(cid:4)Y +(cid:5)(cid:5)(cid:5)+Y leadstotheclassicalADE.Onthe @tg @x @xa 1 t/n other hand, if the waiting times J are heavy tailed with n whichgovernstheCTRWparticledensityinthelong-time power law index 0 < g < 1 and scale s = 1 then the limit.Itisimportanttonotethattheparameteracodesthe generalizedCLTyields largeparticlejumps(earlyarrivals)andgcontrolsthelong waiting times (residence times). A more detailed analysis J þ(cid:5)(cid:5)(cid:5)þJ (cid:2)nn 1 n !W; ð20Þ using the CLT leads to higher-order temporal derivatives n1=g [Baeumeretal.,2005;BaeumerandMeerschaert,2007]. whereWisdistributedlikeS (m=0,s=1,g,b=1).Since [34] The limit process X(t) = Z(T(t)) is a subordinated g g < 1, we find nn/n1/g ! 0, and so the centering term in Le´vy (or Brownian) motion. The outer process is the random walk limit discussed previously. The inner process (20)canbeneglected.Thenthenthjumpoccursattime T(t) = (t/W)g adjusts for the time a randomly selected particle spends in motion (e.g., see recent work by Ganti t¼J þ(cid:5)(cid:5)(cid:5)þJ (cid:4)n1=gW ð21Þ 1 n etal.[2009]whoproposeasubordinatedBrownian motion model for sediment transport). We call u = T(t) the opera- for large n. Solving for n shows that the particle location tional time, since it counts the number of particle jumps X(t)=Y +(cid:5)(cid:5)(cid:5)+Y (cid:4)n1/aZwheren(cid:4)(t/W)g forlargen. 1 n by time t. Since the operational time process incorpo- Puttingthisalltogether,weseethat rates memory effects, it is non-Markovian, and it is simpler to understand its inverse process t = W(u) that XðtÞ(cid:4)ðt=WÞg=aZ maps operational time back to clock time t. The process W(u) = u1/gW comes from (21) and it is another Le´vy andthisapproximationisexactinthescalinglimit,where motion with index g < 1. Then the operational time the number of jumps tends to infinity [Meerschaert and process u = T(t) is the inverse of this Le´vy process. The Scheffler, 2004]. If g(t) is the pdf of W, a change of scaling W(u) = u1/gW is consistent with the inverse variables[MeerschaertandScheffler,2004]showsthatu= scaling T(t) = tgT already noted. The scaling X(t) = (t/W)ghaspdf Z(T(t)) = Z(tgT) = tg/aZ(T) = tg/aX shows that particles followingthespace-timefractionalADEspreadatratetg/a qðu;tÞ¼ tu(cid:2)1(cid:2)1=ggðtu(cid:2)1=gÞ whichisslowerthantheADEifa=2.Thismodelisoften g called subdiffusive. The random time change T has finite 9 of15 F00A07 SCHUMERETAL.:FRACTIONALADES F00A07 walklimitZ(u)accordingtothePDFq(u,t)oftheinverse Le´vyprocessT(t). [35] Figure 8a illustrates a CTRW with heavy-tailed waitingtimes.Figure8bshowsthesameCTRWatalonger time scale. Figure 8c shows the continuous particle path of thesubordinatedBrownianmotionthatemergesasthelong- time scaling limit of this CTRW. The long resting times (particle retention) seen in the CTRW persist in the scaling limit. 5.2. Eulerian Approach to Fractional-in-Time ADEs [36] The Eulerian approach accounts for bulk mass movements into and out of a control volume of edge size Dx = h during a short time step Dt, as in Figure 7. The time fADE includes memory effects, allowing particle residencetimesforlongperiods.Thismeansthatparticles can enter the control volume at the current time step from locationsupstreamintimeasdepictedinFigure9.Apower law memory function [Haggerty and Gorelick, 1995; Haggerty et al., 2000, 2002] implies that the time flux canbecodedusingadiscreteconvolutionintimeSC(x,t(cid:2) jh)w usingthepowerlawGrunwaldweightsfromFigure1. j This leads to a time-fractional analogue n@gC(x, t)/@tg = (cid:2)@(x,t)/@xtotheconservationofsolutemassequation(12) thatcombineswiththeusualadvectiveanddispersiveflux equation(13)toproducethetimefADE @gCðx;tÞ @Cðx;tÞ @2Cðx;tÞ ¼(cid:2)v þD : ð25Þ @tg @x @x2 Ifpowerlawresidencetime(memoryfunction)iscombined with power law velocity distribution, we recover the same space-time fADE (24) derived from the Lagrangian model. The time-fractional parameter g codes retention, since it impliesthatthememoryfunctiont(cid:2)gisapowerlaw,andso it can be estimated from the late-time tail of the break- through curve [Schumer et al., 2003]. [37] The solution C(x, t) to the time fADE (25) is the probability density of the subordinated process X(t) = Z(T(t)) that models the location of a randomly selected particle.Thesolutionformula(22)mixesthedensityofZ(u) accordingtotheoperationaltimeprocessu=T(t).Theouter processZ(u)isaBrownianmotionconnectedwiththeright- handsideof(25),andtheinnerprocesscompensatesforthe memoryeffectsof thefractional timederivative ontheleft- hand side of (25). 5.3. Characteristics of Time (and Space) fADE Solutions Figure 8. Comparison of heavy-tailed waiting time [38] Characteristics of the Green’s function solutions to CTRW traces at increasing time scale and in the scaling the advection-dispersion equation with a fractional-order limit as a subordinated Brownian motion. (a) CTRW with term in time and possibly space are as follows. heavy-tailed waiting times simulates subdiffusive particle [39] 1. The spatial snapshot characteristics are subordi- motion. (b) CTRWat a longer time scale. (c) Subordinated nateda-stableconcentrationprofilesskewedwithlongright Brownian motion as the CTRW scaling limit. Long resting tail if a < 2, plume leading edges that decay like a power times persist in the limit process. law x(cid:2)a(cid:2)1 if a < 2, snapshot width that spreads like tg/a, and total mass that remains constant over time. [40] 2. The characteristics of the flux at position x are moments of all orders,sointheheavy-tailedcase a<2it breakthrough curve tail that decays as t(cid:2)g(cid:2)1; breakthrough does not affectthetailsofX(t)whichstillfalloff likex(cid:2)a curve width that grows like xg/a; if a < 2, immediate spike [Meerschaert and Scheffler, 2004]. The mixture integral in flux then strong leading edge; and area under break- (22)isaweightedaverageofthePDFf(x,u)oftherandom through curve that is constant with time. 10of 15

Description:
We seek equations that capture essential features of transport of .. function (point source) solutions [Arfken and Weber, 1995] to traditional integer-.
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.