Lévy Stable Probability Laws, Lévy Flights, Space-Fractional Diffusion and Kinetic Equations (Two Lectures)

Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Lvy Stable Probability Laws, Lvy Flights, Space-Fractional

Diffusion and Kinetic Equations (Two lectures)


(Paul Pierre Lvy, 1886-1971)
(Stable probability laws, alpha-stable probability laws/distributions)
Math: = alpha-stable Levy random processes, Levy motion.
Mandelbrot: Lvy flights (LFs) = non-Brownian random walks with long jumps (physics, nature,
finance)
We need a probabilistic background which goes beyond the scope of traditional courses for physicists on
probability theory

I.

Stable probability laws (stable distributions) and generalized central


limit theorem

# 1 Definition. Stability = invariance under summation. X1, X2, X3, , X i.i.d. (independent
identically distributed) r.v. (random variables) possess a stable distribution if cn > 0 and dn such
that
n

X 1 + X 2 + . . . + X n = X i = cn X + d n
i =1

(1)

equal in probability

cn scaling factor
Theorem. cn = n

1/a

, 0 < 2, index of stability (math), or Lvy index (phys).

Alternative definition. X1, X2, and X i.i.d. r.v. possess stable distribution if for A > 0, B > 0 C
> 0, D such that
d

AX1 + BX 2 = CX + D

(2)

Then (0,2] such that


C = A + B .

(3)

Noteworthy examples:
1. = 2 (Gaussian distribution)
Probability density function (PDF) pG ( x; , ) =

( x )2
1
exp
, dx pG ( x; , ) = 1

2
2 2

mean, variance.
2

2. = 1 (Cauchy Lorenz probability law)

PDF pC ( x; , ) =

( x )2 + 2

dx pC ( x; , ) = 1 .

These two distributions belong to the class of stable distributions, but they are very different.

Gauss: all the moments are finite.


1
Cauchy: Asymptotics. pC
x2
Cauchy: no integer moments, only fractional moments of the order q < 1 are finite:

dx x pC ( x; , ) dx x q x 2 < if
q

q <1 .

# 2 Characteristic function (CF) of stable distribution


CF

p(k ) =

dxp ( x)eikx = eikx

eik ( X 1+ X 2 +...+ Xn = eikX 1

p( x) =

dk

ikx
2 p(k )e

eikX 2 ... eikXn = eikXcn eikd n

(4)

Functional equation for CF p(k) : [ p (k )] = p (cn k )eid n k .


n

Solution (Lvy; Khintchine; Gnedenko, Kolmogorov):

p , (k ; , ) = exp i k | k | (1 i (k , ) ) ,

(5)

where

tan( / 2) if 1, 0 < < 2


( 2 / ) ln | k | if = 1

(k , ) =

4-parametric: 0 < 2 (Levy index), -1 1 (skewness parameter), > 0 (scale parameter),


real (shift parameter).

# 3 Properties of probability densities.


(1) and are important ! and are less important ( and can be eliminated by proper shift and
scale transformation:

p , ( x; , ) =

;0,1 .
p ,

Notation: p , ( x;0,1) p , ( x)

(2) Symmetry property:

p , ( x) = p , ( x) .
(3) Symmetric stable laws: = 0 p ,0 ( x) = p ,0 ( x) (Next lecture: play important role in random
walk theory).

CF for symmetric stable law:

p ,0 (k ) p (k ) = exp | k |

1
p (k ) 1 | k | + | k |2 ... k 0 .
2
The first non-analytical term is responsible for asymptotics of the PDF at x :

p ( x)

C ( )

| x |1+

, x .

[Here: digression on the derivation of C() with the use of improper integral].

C ( ) =


sin

2
1

(1 + )

Examples. In terms of elementary functions:


Example 1. Gauss <To derive from CF>
p2 ( x) =

1 x2 / 4
e
.
4

Example 2. Cauchy <To derive from the CF>

p1 ( x) =

(1 + x 2 )

(4) Extremal stable laws = 1. Important subclass 0 < < 1. Next lecture: play important role in
random walk theory as distributions of waiting times.

In terms of elementary function: only for = 1/2 , = 1.


Example 3. Lvy-Smirnov.

1
e1/ 2 x , x 0

3
p1/ 2,1 ( x) = 2 x

0 , x<0

B. Gnedenko, A. Kolmogorov, Limit Distributions for Sums of Independent Random Variables


(1949), two citations:
The prophecy: All these distribution laws, called stable, deserve the most serious attention. It is
probable that the scope of applied problems in which they play an essential role will become in due
course rather wide.
The guidance for those who write books on statistical physics: normal convergence to
abnormal (that is, different from Gaussian A. Ch.) stable laws , undoubtedly, has to be
considered in every large textbook, which intends to equip well enough the scientist in the field of
statistical physics.

#4 Generalized central limit theorem GCLT

Generalized Central Limit Theorem: stable distributions are the limit ones for
distributions of sums of random variables (has domains of attraction) appear,
when evolution of the system or the result of experiment is determined by the sum
of a large number of random quantities
X1, X2, X3, .Xn i.i.d.r.v. PDF f(x): what is the distribution of their sum at n ?
- Gauss, = 2: attracts f(x) with finite variance
x2 =

f ( x)dx <

- Lvy, 0 < < 2: attract f(x) with infinite variance and the same asymptotic
behavior
x

x 2 f ( x) dx = ,

f ( x) x 1 , 0 < < 2

More precisely: normal convergence to symmetric stable laws

X1 , ... X n i. i. d . r. v. f ( x ) ~

1
x

1+

, x

0 < < 2
S n =

X 1 +

(x )

...

X n
: n

(x ) ~

1
x

1+

(x )
, x

Also:
gravity field of stars ( = 3/2)
electric fields of dipoles (= 1)
velocity field of point vortices in fully developed turbulence (=3/2)

II. Continuous time random walk and space fractional diffusion equation
# 1 Basics of CTRW
Return to random walk (RM Lecture #2) : P(x,t) PDF to find a particle at position x at time t.

-- random jumps , PDF ( ) , ( k ) = F { ( )} =

dx ( x )e

ikx

-- random waiting times, PDF ( ) , (u ) = L{ ( )} = d ( )e u


0

-- Montroll-Weiss equation (RM lecture #2)


P(k , u ) =

1 (u )
1
u
1 ( k ) (u )

(2.1)

-- No anomaly in waiting time PDF (RM lectures 3-4)

< (u ) 1 u + O (u ) , u 0 , > 1 ,

d
= d ( ) = d ( )e u
du
0
0

(2.2)

d (u )

.
=

du u = 0
u =0

# 2 Two cases for jump PDF ().


Case 1. Jump PDF belongs to domain of attraction of the Gaussian law.

( ) < ( k ) 1

2
2

k 2 + O (k ) , k 0, > 2 .

d2

d 2 ( k )
ikx

=
=
dx ( x )e
.

dk 2
dk 2

k =0
k =0

MW equation gives

P(k , u ) = .... =

1
u + Dk 2

, D=

2
2

(2.3)

uP( k , u ) 1 = Dk 2 P(k , u ) .
Now, recalling that

2
d

L (t ) = u (u ) (t = 0) , F

( x ) = k 2 ( k ) ,
2
dt

dx

we can immediately make an inverse Fourier-Laplace transform of (2.3) to get normal diffusion
equation

2
P( x, t ) = D
P( x, t ) , P( x, t = 0) = ( x ) , < x < , D =
2
t
x 2

(2.4)

Solution:
1

P( x, t ) =

4 Dt

x2
4 Dt

(2.5)

mean square displacement

dx x

P( x, t ) = 2 Dt t1 :

normal diffusion

Case 2
Jump PDF belongs to domain of attraction of the stable law with 0 < < 2.
That means

d 2 ( ) = , ( )

| x |1+

, 0 < < 2 , | x |

(k ) 1 | k | +... , k 0

(2.6)

(Remind digression on the improper integral).


Plug (2.6) and (2.2) into the MW equation (2.1):

P(k , u ) =

1
=

u + |k |
u + D | k |

, D=

or

uP( k , u ) 1 = D | k | P( k , u )

(2.7)

Rewrite (2.7) as

P( x, t ) = D
P( x, t ) , P( x, t = 0) = ( x ) , 0 < < 2
t
| x |

# 3 Riesz fractional derivative

( x) (k ) =

dx e ( x) , ( x) =
ikx

df
dx

d x

d2 f
dx

/2

( )

d2 f

ikf (k )

d f

( ik )2

dk ikx
e (k )
2

f (k ) = k 2 f (k )

f (k )

( )

dk
d2 f
2
ikx
=
k f (k )e
= 2
2
2

dx
d x

Example.

( x) =

1
1 + x2

( +1) cos +1 arctan x


d = dk exp(ikx) k (k ) =

( +1)/2

dx

2
2
1+ x

Conclusion: space-fractional diffusion equations emerge as a long time-space


limit of a random walk with long random jumps whose PDF () belongs to
domain of attraction of stable law:
( )

1+

| |

0 < < 2 ,

2 =

(Section II CTRW and Space-Fractional Diffusion Equation continued)


#4

Space-fractional diffusion equation for Lvy stable jump PDF

Jumps in discrete time:


X ( n) =

ik j

= exp k

, ( ) ( k ) = e

j =1

0 < 2

PX ( x, n) = ?
n

PX (k , n) =

ik j

ik j n

e j =1

eikX (n) =

= e

=e

n k

Go to the continuous time by t = nt

PX (k ,t ) = etD|k |

, D=

Differentiate:

P( k , t ) = D k P(k , t )
t

d x

k (k )

P ( x, t ) = D
P ( x, t )

t
x
#5

Solution of space-fractional diffusion equation, q-th moments and superdiffusion

Go back to Fourier space:


P(k , t ) = exp( D k

t)

Asymptotics by using the same trick as shown in Section I:

P ( x, t )

Dt
| x |1+

Reminder: x 2 = (consider the integrand), but

< , q < .

Time behavior of the q-th moment:


1

q
dx x P( x, t ) = 2 dx x q
0

= 2 dx x q
0

dk ikx D|k | t
= k k = k ( Dt )1/ =
2 e

dk ik x /( Dt )1/ |k |
= x x =
e

( Dt )1/ 2

= 2( Dt ) q / dx xq
0

x
( Dt )1/

dk ik x |k |
2 e

Effective second moment:

M 2 (t ; , q ) = x

q 2/ q

( Dt )

2 /

>1 , 0 < < 2 .

Superdiffusion phenomenon
Examples of LFs:
- Optics photons,
- Levy glass
- Strange diffusion along the DNA
- Scaling law of human travel
Imprisonment of Resonant Radiation in Gases:
Photon Trajectories as Lvy Flights
Compton (1922): transfer of excitation
as a type of Brownian motion
diffusion-type equation for the density of
excited atoms
Essential assumption:

n / t = D2n
p ( r ) exp ( r / )

D = 2 / 3

Mean free path

BUT ! Kenty (1932), Holstein (1947): frequency dependent absorption

p (r )
r

( ln r )

1/ 2

or

p(r )

1
r

3/ 2

Lvy flights of photons in hot atomic vapours


(Mercadier et al. Nature Physics 2009)
Measuring the step
size distribution of
photons
Under Gaussian
assumptions of
emission and
absorption spectra
(Doppler broadening)

1
P(r ) 2
r ln ( r / l0 )

Figure: Double cell configuration to measure the step


size PDF

Figure: power law exponent of the step size


distribution vs number n of scattering events
(multiply scattering)

A Lvy Flight for Light


(P. Barthelemy et al, Nature 2008)

Ideal experimental
system to study Lvy
flights in a controlled way
Precisely chosen
distribution of glass
microspheres of
different diameters d

New optical material in


which light performs a
Lvy flight Lvy glass

Spatial distribution
of the transmitted intensity

P(d)
P(d) ~ d-(2 + )

Paradoxical Diffusion in Chemical Space


Protein diffusion to next neighbor sites on folding DNA (deoxyribonucleic acid)
acid)
Motion of binding proteins or enzymes
along DNA: detach to a volume reattach
before reaching the target
(Berg von Hippel model)
Intersegmental jumps permitted
at chain contact points due to polymer looping

Contoured length |x| stored


( x ) | x |1
in a loop between contact points
= 1/ 2
Optimal Target Search on a FastFast-Folding Polymer Chain (Lomholt
(Lomholt et al.2005)

n ( x, t ) = DB
+ DL
koff n ( x, t ) + kon nbulk j ( t ) ( x )

t
x
|x|

Scaling laws of human travel,


travel, 2
( Br ock ma nn, Huf na g el , Geisel , Na t ur e 2006)
Money as a proxy for human travel:
travel: trajectories of banknotes originated from 4
places reminiscent of L
Lvy flights in the US currency tracking project
www.wheresgeorge.com



r = x2 x1

Geographical displacement r between two report


location of a bank note

W (t , r ) = D

| r |

W (t , r ) , 0.6, = 2 /

III. Lvy Flights, Lvy motion / Alpha stable Lvy processes (symmetric case)
The same phenomenon from the random processes theory point of view.
# 1 Definitions, terminology, examples.
(1) L(t): non-stationary Markov process with stationary independent increments distributed with
the alpha-stable law

p (k , ) = exp ik ( L (t + ) L (t ) = exp D | k |

(2) Stationarity of increments:


d

L (t1 + ) L (t2 + ) = L (t1 ) L (t2 )

(3) Physically: increments = jumps


(4) White Lvy noise = Sequence of increments
(5) Physically : description of processes with large outliers, far from equilibrium.

Examples of the Lvy noises include :


economic stock prices and current exchange rates (1963)
radio frequency electromagnetic noise
underwater acoustic noise
noise in telephone networks
biomedical signals
stochastic climate dynamics
turbulence in the edge plasmas of thermonuclear devices (Uragan
2M,ADITYA,2003; Heliotron J,2005...)
Levy Turbulence
Turbulence in boundary plasma
of thermonuclear devices
Levy statistics of plasma fluctuations:
UraganUragan-3M
ADITYA
L-2M
LHD,
LHD, TJTJ-II
Heliotron J
bursty
bursty character of fluctuations in other
devices (DIIIDIII-D, TCABR,
TCABR, )
Levy turbulence is a widely spread phenomenon Models are strongly needed !

Motion of a test particle in the


Guiding Center Approximation :

 
E(r , t)

: Levy random field

 

   
dr E0 ( r ) B0 E ( r , t ) B0
=
+
dt
B02
B02

#2

First passage and arrival properties of LFs

Two random quantities: first passage time and first passage overshooting (leapover) l: the motion
starts at x0 > 0 the boundary is at x = 0

p(t; x0 ) =

d
d
= dxf ( x, t )
dt
dt
0

Probability to leave the interval (0, ) within the time from t till t+dt.
(1) Universality of the First passage time PDF / Sparre Andersen Universality .

p(t ) t 3/ 2

f ( x, t ) = W ( x x0 , t ) W ( x + x0 , t )

(2) Gaussian case

p(t; x0 ) =

,0< 2

x0
4 Dt 3

2
e x0 /(4 Dt ) t 3/ 2

x2
W ( x, t ) =
exp
4 Dt
4 Dt

t = dt t p(t; x0 ) =
0

(3) LFs:

f ( x, t ) = fim ( x, t ) = W ( x x0 , t ) W ( x + x0 , t ) ?
W ( x, t ) =

dk

exp ( ikx ) exp D k t


2

d
d
pim (t ; x0 ) =
= dxfim ( x, t )
dt
dt

pim (t ) t 11/ t 3/ 2

Method of images breaks down for Lvy flights due to their non-local nature, which manifests
itself in the non-zero leapover

(4) Probability of first arrival time (Problem #3 from the assignment sheet).

f
f
=D
p fa (t ) ( x )

t
|x|

f ( x, t = 0) = ( x x0 ) , x0 > 0 ,

f ( x = 0, t ) = 0

p fa (t ) t 2+1/ t 3/ 2
First arrival first passage for LFs !
(5) First passage overshooting (leapover)

PDF of overshooting:
1
f (l ) 1+ / 2
l
Reminder. For LFs p( x )

1
| x |1+

The PDF of the first overshooting decays slower than the PDF of LFs which produce this
overshooting !

IV. Langevin Equation with Lvy noise and space-fractional Fokker-Planck


Equation
# 1 Langevin equations : 2nd Newtons law
m

dv
dU
=
mv + (t ) ,
dt
dx

dx
=v
dt

(t) : white Lvy noise

overdamped approximation: neglect inertia.


# 2 Fokker-Planck equation

f
dU
f
f +D
=
t x dx
| x |
= 2 : normal Fokker-Planck equation
# 3 Reminder. Boltzmann distribution
d dU
d2 f
+
=0
f
D

dx dx
dx 2

U ( x)
f ( x ) = C exp
, dxf ( x ) = 1
D

# 4 Stationary solution for LFs in Harmonic Potential

x2
U ( x) =
,
2

d
d x

k (k )

<To derive the equation for the CF>


k

f (k ) = exp

sin / 2 ( )

f ( x ) C
,
C
=
,
1+

x2

dx x

2f

( x) =

Properties of stationary PDF: unimodality and slowly decaying asymptotics


7

# 5 Stationary Solution of LFs in Quartic Potential

dx
= x 3 + Y (t )
dt

d 3
d f
=0
x f +

dx
d x

( )

d3 f
dk

= sgn(k ) k

f ( x) =

f (k )

(1 x 2 + x 4 )

xmin = 0

f ( x ) x 4

, =1

xmax = 1/ 2

x2 =

dx x 2 f ( x ) = 1

Confined LFs: finite variance, out of the domain of attraction of stable laws

You might also like