Anomalous Lévy diffusion: From the flight of an albatross to optical lattices. Eric Lutz Abteilung für Quantenphysik, Universität Ulm

Transcription

1 Anomalous Lévy diffusion: From the flight of an albatross to optical lattices Eric Lutz Abteilung für Quantenphysik, Universität Ulm

2 Outline 1 Lévy distributions Broad distributions Central limit theorem Physical examples 2 Lévy diffusion Normal diffusion Anomalous diffusion (fractional derivative) Experimental observations 3 Statistical mechanics Anomalous Lévy diffusion in an optical lattice Fluctuations: how large is large? Ergodicity

3 Broad distributions Physical point of view Many phenomena: well-defined average behavior with fluctuations around average value fluctuations described by narrow distributions ex: Gauss distribution Other phenomena: properties dominated by fluctuations (average not important, may not even exist) fluctuations described by broad distributions ex: Lévy distributions "broad"=power-law tail

4 Broad distributions Mathematical point of view: Central limit theorem S N = X 1 + X X N Sum of N independent random variables X i Case 1: finite second moment X 2 i < S N Gauss distribution N Case 2: divergent second moment X 2 i S N Lévy distribution N Stable probability distributions

5 Lévy distribution Definition and main properties: Lévy 1937 Characteristic function ϕ(k) = dx P(x) e ikx = e c k α 0 < α 2 Asymptotic power law decay P(x) 1 x 1+α, x α < 2 Second moment diverges x 2 = dx x 2 P(x) α < 2 Scale-free distributions

6 Physical examples α = 2 Gauss e x2 /2 2π (Errors in astronomical observations) 1809 α = 1 Breit-Wigner 1 π 1 1+x 2 (Cross-section of resonant scattering) 1936 α = 1 2 Schrödinger e 1/2x 2πx 3 α = 3 2 Holtzmark (First-passage time distribution) 1915 (Random gravitational force on a star) 1919

7 Outline 1 Lévy distributions Broad distributions Central limit theorem Physical examples 2 Lévy diffusion Normal diffusion Anomalous diffusion (fractional derivative) Experimental observations 3 Statistical mechanics Anomalous Lévy diffusion in an optical lattice Fluctuations: how large is large? Ergodicity

8 Normal diffusion Example: small particles in a glass of water Perrin 1909 Gaussian dynamics: Diffusion equation: Solution: P t P(x, t) = = D 2 P x 2 1 4πDt e x 2 /(4Dt) Mean-square displacement: Irregular, random motion Brownian motion Brown 1827 x 2 = 2Dt linear in time

9 Transport equations: Gauss System S in contact with bath B (temperature T) Langevin 1908 mẍ + U (x) + ηẋ = F (t) η = friction F (t) = Gaussian random force (central-limit theorem) F(t) = 0 F(t)F(t ) = 2D δ(t t ) D = ηkt

10 Lévy diffusion Definition: mẍ + U (x) + ηẋ = F(t) F(t) = Lévy random force (central-limit theorem) Properties: algebraic tail of force distribution large fluctuations long jumps (Lévy flights) Enhanced diffusion ( x 2 )

11 Flight of an albatross Flight-time measurement of 30 albatrosses at Bird Island Typical trajectories: Stanley et al., Nature 1996 Distribution of flight-time intervals: P(t i ) 1 (t i + 1) µ µ = 1+α 2 Power-law distribution

12 Experiment 1: Diffusion in micelles Micelles = polymer-like molecules that break and recombine rapidly Ott et al., PRL 1991 α = 1.5 Measurement: intensity of light emitted by fluorescent atoms after photobleaching characteristic function Diffusion equation: Gauss P t = D 2 P x 2 e P t = Dk 2 P P(k, t) = e Dt k 2 Lévy P t = D α P x α e P t = D k α P P(k, t) = e Dt k α

13 Fractional derivative Letter from Leibniz to L Hospital: d 1/2 dx 1/2 x =? d m dx m x n = n! (n m)! x n m = m = 1/2, n = 1 d 1/2 dx 1/2 x = 2 π x 1/2 Riemann-Liouville operator: Γ(n + 1) Γ(n m + 1) x n m d α dx α F(x) d (ik)α F(k) α Ex: Riesz operator: d α d x α F(x) d k α F(k) α Ex: d α d x α F(x) = 1 d 2 κ α dx 2 dy dx α eax = a α e ax d x α cos ax = a α cos ax F (y) x y α 1 non-local

14 Experiment 2: 2D rotating flow Rotating tank (water + glycerol) 1.5 Hz Weeks et al., PRL 1993 Typical trajectory: Regime of chaotic advection chain of vortices Pdf of duration of flight: α = 1.3

15 Experiment 3: Surface diffusion Senft et al., PRL 1995 Palladium atoms on a tungsten lattice T = 133K α = 1.251, s = 7.4 Measurement: displacement of single atom on a 1D lattice (field microscopy) Low diffusion barrier longs jumps (2, 3 or more lattice sites) ϕ(k) = e sk α Realization of 1D discrete random walk

16 Experiment 4: Saccadic eye movements Brockmann and Geisel 2004 Magnitude of visual angle: 0.1 < x < 100 P exp (x > 75 ) = Pdf of saccadic magnitude larger than x: P > (x) = x dx x (1+α) x α (α = µ)

17 Outline 1 Lévy distributions Broad distributions Central limit theorem Physical examples 2 Lévy diffusion Normal diffusion Anomalous diffusion (fractional derivative) Experimental observations 3 Statistical mechanics Anomalous Lévy diffusion in an optical lattice Fluctuations: how large is large? Ergodicity

18 Optical lattice Spatially periodic optical potential: obtained, for instance, by superposition of two counterpropagating laser beams "optical crystal" (band structure, Bragg scattering,...) Advantages: potential is exactly known (no defects) can be tuned in a precise way (spacing, amplitude) Atomic transport?

19 Atomic transport in optical lattice Transport equation for the semiclassical Wigner function: Dalibard and Cohen-Tannoudji 1989 W (p, t) t = [K (p)w ] + p p [D(p) W p ] Drift : K (p) = [Friction force] ᾱ p 1+(p/p c) 2 Diffusion : D(p) = D 0 + (Sisyphus effect) D 1 1+(p/p c) 2 [Momentum fluctuations] Cooling Heating Range of validity: large momentum, p k defines semiclassical limit small saturation, s 1 low laser intensity large kinetic energy, p 2 /2m U 0 allows spatial averaging

20 Power-law tail distribution in momentum Stationary momentum distribution: W s (p) = 1 Z Power-law tail: W s (p) p (ᾱp2 c )/D 0 Exponent can be rewritten as: [ 1 + D 0 p 2 ] (ᾱp 2 c )/(2D 0 ) D 0 + D 1 pc 2 (ᾱp 2 c)/d 0 = U 0 /(22E R ) (U 0 = potential depth, E R = recoil energy) Statistical properties of W s (p) can be changed: from normal statistics for U 0 > 66E R to Lévy statistics for U 0 < 66E R In fact, W s (p) = Tsallis distribution Lutz PRA (R) 2003

21 Lévy statistics and diverging moments Second moment: diverges if U 0 < 66E R. p 2 = dp p 2 W s (p) mean kinetic energy, E K = p 2 /2m, infinite. Fourth moment: diverges if U 0 < 110E R. p 4 = dp p 4 W s (p) mean square kinetic energy, E 2 K = p4 /4m 2, infinite.

22 Experiment Single 24 Mg + ions in a one dimensional optical lattice: Katori et al., PRL 1997 Experimental observation of the divergence of the atomic kinetic energy

23 Fluctuations: how large is large? Statistical properties of rare but extreme events: How tall should one design an embankment so that the river reaches this level only once in 50 years? How large will be the largest earthquake in Los Angeles in the next 20 years? How large is the largest atomic momentum in 1000 observations? Extreme Value Theory Gumbel 1958

24 Fluctuations: how large is large? For independent events: Value x max = λ of the maximum among N realizations that will not be exceeded with probability p satisfies: for exponential P(x) = e x λ dx P(x) ln(1/p) N [ λ ln N ln(1/p) for power-law P(x) = x (1+α) λ [ N ln(1/p) ] ] 1/α Example: P(x) = C (1 + x 2 ) (1+α)/2 N = 1000 p = 99% U 0 /E R α λ ( p 4 diverges) ( p 2 diverges) ( p diverges)

25 Ergodicity Motivation: Two recent experiments "Experimental Investigation of Nonergodic Effects in Subrecoil Laser Cooling" Saubamea et al., PRL 1999 "Statistical Aging and Nonergodicity in the Fluorescence of Single Nanocrystals" Stockmann et al., PRL 2003 Lévy statistics induces ergodicity breaking Question: connection between nonergodicity and diverging moments? Advantage of optical lattice: Ordinary linear Fokker-Planck equation

26 Ergodicity Definition: "Ensemble average and time average of observables are equal in the infinite-time limit." A = A as t Ensemble average: A = dp A(p) W (p, t) in the long time limit: A s = dp A(p) W s (p) Time average : A = 1 t t 0 dτ A(p(τ)) Ergodicity in the mean-square sense: σ 2 A (t) = (A A ) 2 0 when t

27 Ergodicity Ergodicity breaking: Lutz PRL 2004 Ergodicity depends in general on the observable A(p) for A(p) = p n σ 2 A (t) t µ with µ = 1 (2n + 1)D 2D σ 2 A (t) 0 as t only if D < D n = 1/(2n + 1) Existence of a noise threshold D n above which ergodicity is broken (rescaled variables ᾱ = p c = 1, D 1 = 0, D 0 = D = 22E R /U 0 )

28 Ergodicity Moments of stationary momentum distribution: p m = dp p m W s (p) are finite for D < D m = 1/(m + 1) Direct relationship between the loss of ergodicity for A(p) = p n and the divergence of the 2nth moment of the stationary momentum distribution In particular: for n = 1 p p when p 2 diverges (D = 1/3, U 0 = 66E R )

29 Calculation of σ 2 A (t) σa(t) 2 = 1 t t t 2 dt ] dt 2 [[ A(p(t 1 ))A(p(t 2 )) A(p(t 1 )) A(p(t 2 )) Two-time correlation function: A(p(t 1 ))A(p(t 2 )) = dp 1 dp 2 A(p 1 )A(p 2 ) W 2 (p 1, t 1 ; p 2, t 2 ) Two-point joint probability density function: ( W 2 (p 1, t 1 ; p 2, t 2 ) = ψ 0 (p 1 )ψ 0 (p 2 ) ψ 0 (p 1 )ψ 0 (p 2 ) + 0 dk ψ k (p 1 )ψ k (p 2 ) e E k t 1 t 2 ) Finally σ 2 A (t) = 2 t 2 t 0 dτ (t τ) C A(τ) With C A (τ) = 0 dk [ dp A(p) ψ0 (p)ψ k (p) ] 2 e E k τ

30 Power-law tail distribution in time First passage time distribution: Divide momentum space in two regions: low momentum region p < 1, K 1 (p) p high momentum region p > 1, K 2 (p) 1/p First passage time = time at which the momentum of the system first exits a certain momentum interval, given that it was originally in that interval In the high momentum region: g 2 (t) t γ with γ = (3D + 1)/(2D) Power-law distribution in time

31 Power-law tail distribution in time Moments of the first passage time distribution: t n = dt t n g 2 (t) converge for D < D n = 1/(2n 1) Ergodicity is broken for A(p) = p n when the (n + 1)th moment of the first passage time distribution in the high momentum region becomes infinite In particular: for n = 1 p p when t 2 diverges (D = 1/3, U 0 = 66E R )

32 Summary Lévy distributions = power-law tail describe large fluctuations Strange properties (diverging moments) but observed in nature! Direct link between diverging moments and nonergodicity New statistical physics is needed!