Transcript 1. dia
Scaling functions for finite-size corrections in EVS
Zoltán Rácz
Institute for Theoretical Physics
Eötvös University
E-mail: [email protected]
Homepage: cgl.elte.hu/~racz
Motivation: Do witches exist if there were 2 very large hurricanes
in a century?
Introduction: Extreme value statistics (EVS) for physicists in
10 minutes.
Collaborators:
G. Gyorgyi
N. Moloney
K. Ozogany
I. Janosi
I. Bartos
Problems: Slow convergence to limiting distributions.
Not much is known about the EVS of correlated variables.
Idea:
EVS looks like a finite-size scaling problem of critical
phenomena – try to use the methods learned there.
Results:
Finite size corrections to limiting distributions (i.i.d. variables).
Numerics for the EVS of 1 / f signals ( 0 ).
Improved convergence by using the right scaling variables.
Distribution of yearly maximum temperatures.
Extreme value statistics
Y
is measured:
y1 , y2 ,..., yN
z N maxy1 , y2 ,..., yN
yi h(ti )
Question: What is the distribution
of the largest number?
Aim: Trying to extrapolate to
values where no data exist.
P0 ( y )
Logics:
Assume something about
yi
E.g. independent, identically distributed
y
Use limit argument: ( N )
Family of limit distributions (models) is obtained
Calibrate the family of models
by the measured values of xN
P( z N )
zN
Extreme value statistics: i.i.d. variables
z
Y
is measured:
y1 , y2 ,..., yN
F ( z ) dyP0 ( y)
P0 ( y )
z N maxy1 , y2 ,..., yN
GN (z) probability of z N z
y
z
GN ( z) [F ( z)]
N
Question: Is there a limit distribution for N ?
lim
N
GN (aN x bN )
lim
N
[ F (aN x bN )]N G( x)
Result: Three possible limit distributions depending
on the tail of the parent distribution, P0 ( y ) .
P( z N )
z
zN
z aN x bN
Fisher & Tippet (1928)
Gnedenko (1941)
Extreme value limit distributions: i.i.d. variables
P0 ( y )
Fisher-Tippet-Gumbel (exponential tail)
GFTG ( x) exp( exp( x))
e
ya
y
( y0 y) 1
y0
Fisher-Tippet-Frechet
1
GFTF ( x) exp( x )
0
y
GW ( x) exp( ( x) )
1
x0
x0
W
FTF
5/ 2
( x x ) /
x0
x0
Weibull (finite cutoff)
Characteristic shapes
of probability densities: I ( x) dGI ( x) / dx
FTG
(power law tail)
( x x ) /
1
( x x ) /
Gaussian 1 / f signals
h
Independent, nonidentically distributed Fourier modes
(hk ) ~ e
L
k
k
hk
0 .5
1
1/ f
White
noise
noise
EVS
)
Berman, 1964
L
w2 (h h) 2 L1k hk ~ L 1
2
2
EdwardsWilkinson
0
x
0
hk ~ k
with singular fluctuations
h(x)
2
2
Random
walk
MajumdarComtet, 2004
MullinsHerring
4
Random
acceleration
Single mode,
random phase
0
Slow convergence to the limit distribution (i.i.d., FTG class)
P0 ( y )
0
e
y2
0 .5
1
y
4
The Gaussian results are characteristic for the
whole FTG class
except for
y0
2
P0 ( y) ~ e y
0
Finite-size correction to the limit distribution
de Haan & Resnick, 1996
Gomes & de Haan, 1999
z
P0 ( y )
F ( z ) dyP0 ( y)
e
y2
z
P ( x, N )
y
dG N ( x) d
[ F (a N x bN )] N
dx
dx
substitute
Fix the position and the scale of P( x, N ) by
x 0
x 2 1
aN , bN
P( x, N ) FTG ( x)
expand in N
is determined.
1
1 ( x) ...
ln N
1
1 ( x) FTG ( x) a 0 x a 20 1 3 a 0 x e a 0 x e 2 ( a 0 x )
2
a0 6 /
0.577...
0
Finite-size correction to the limit distribution
For Gaussian P0 ( y )
1 ( x )
1
FTG ( x) a 0 x a 20 1 3 a 0 x e a 0 x e 2 ( a 0 x )
2
Comparison with simulations:
How universal is 1 ( x) ?
Signature of (ln N )2 corrections?
0
Finite-size correction: How universal is 1 ( x) ?
z
P0 ( y )
F ( z ) dyP0 ( y) ~ 1 e f ( z )
e
y2
Determines universality
z
f ( z) ~ z 2
1 ( x)
Gauss class
f ( z) ~ z
1 ( x)
Exponential class
f ( z) ~ z p
p -1
1 ( x)
1 ( x)
p
different (known) function
f ( z) ~ z exp(az)
Exponential class is unstable
f ( z) ~ z z
S
f ( z ) ~ z a ln(1 z )
a 1
Exponential class
a 1
Gauss class
Gauss class eves for 0 s 1
Weibull, Fisher-Tippet-Frechet?!
Maximum relative height distribution ( 2 )
0 .5
0
hm
h
h(x)
hm
x
L
0
Majumdar & Comtet, 2004
1
2
4
maximum height measured
from the average height
P(hm , L) ?
Connection to the PDF of the area under
Brownian excursion over the unit interval
Choice of scaling
x hm /hm 0
~e
hm 0 12wL ~ L
6 x 2
Result: Airy distribution
~ x 5e a / x
2
hm 0 P(hm , L) 0 ( x)
Finite-size scaling 2 :
Schehr & Majumdar (2005)
Solid-on-solid models:
H K 1 hi hi 1
hm 0 P(hm , L) 0 ( x)
0 ( x)
L
1
0 ( x) ...
2L
hm 0 12wL ~ L
x hm /hm 0
p
1
hm c0 hm 0 1
...
2L
0 ( x)
x
Finite-size scaling 2 : Derivation of
hm 0 P(hm , L) 0 ( x)
Cumulant generating function
PL ( y)
dk
2
Scaling with
(ik ) m
exp ik ( y c1 )
cm
m
!
m 2
hm 0 L
y hm 0 x
L ( x) hm 0 PL (hm 0 x)
L ( x)
dq
2
Expanding in
(iq) m ( 0)
(0)
(1) 1
exp iq( x c1 c1 L )
cm
m
!
m2
L1 :
c1(1)
L ( x) 0 ( x) 1 0 ( x)
L
hm 0
Shape relaxes faster than the position
Assumption: hm L carries all
the first order finite size correction.
c1 hm L c1(0) L c1(1) L 1 ...
cm2 cm(0) Lm ...
k q / hm 0
1
0 ( x) …
2L
Finite-size scaling 1 : Scaling with the average
Cumulant generating function
PL ( y)
dk
2
Scaling with
(ik ) m
exp ik ( y c1 )
cm
m 2 m!
c1 hm
y hm x
k q / hm
L ( x) hm PL (hm x)
L ( x)
dq
2
Expanding in
cm( 0)
(iq) m
exp iq( x 1)
( 0)
(1) 1 m
]
m 2 m! [c1 c1 L
L1 :
c1(1)
L ( x) 0 ( x) ( 0) 1 [(x 1)0 ( x) 0 ( x)]
c1 L
Assumption: hm L carries all
the first order finite size correction
(shape relaxes faster than the position).
c1 hm L c1(0) L c1(1) L 1 ...
cm2 cm(0) Lm ...
Finite-size scaling 1 : Scaling with the fluctuations
Assumption: c2 relaxes faster than
any other cm 2 .
Cumulant generating function
PL ( y)
dk
2
Scaling with
L ( x)
dq
2
(ik ) m
exp ik ( y c1 )
cm
m 2 m!
c2
y c1
x
k q /
L ( x) PL (c1 x)
c2 (hm hm )2 L c2(0) L2 c2(1) L2 2 ...
cm3 cm(0) Lm ...
cm( 0)
q 2 (iq) m
exp iq( x 1)
2 m3 m! [c2( 0) c2(1) L 2]m / 2
Expanding in
L 2 :
c2(1)
L ( x) 0 ( x) ( 0) 2 [0( x) x0 ( x) 0 ( x)]
2c2 L
Faster convergence
Finite-size scaling: Comparison of scaling with hm and
hm
scaling
scaling
Much faster convergence
.
Possible reason for the fast convergence for ( 1 )
Width distributions
Antal et al. (2001, 2002)
hm
h
w2 L P( w2 ) x
w 2 ~ hm
h(x)
x
L
0
w2 (h h) 2 L1k hk ~ L 1
2
w2
Cumulants of
( w2 )
m
c
~L
( w2 )
1
c
m ( 1)
( 1)
~L
2,4
bm
1
m ( 1)
~
L
a
...
m
m
m 1
L
n 1 n
x w2 / w2
L
b1
1 1 ...
L
L 1 , L 3
c
( w2 )
2
( 1)
~L
b2
1 2 1 ...
L
L 3 , L 7
Extreme statistics of Mullins-Herring interfaces ( 4 )
and of random-acceleration generated paths
0
0 .5
1
0 ( x)
x hm / hm
x hm / hm
0 ( x)
x (hm hm ) /
x (hm hm ) /
2
4
Extreme statistics for large
.
0
0 .5
1
2
(hk ) ~ e
0 ( x)
4
L
k
k
hk
2
k 2
2
k1
Only the k1 mode remains
x hm / hm
hm hk1
0 ( x)
x exp x 2
2
4
hm 0
(hmax ) ~ e
L hk1
2
| hk1| d | hk1 |
Skewness, kurtosis
Distribution of the daily
maximal temperature
P(Tmax )
Scale for comparability
Tmax 0
s skewness
2
Tmax
Tmax 2 1
Calculate skewness
and kurtosis
Put it on the map
Reference values:
sFTG 1.1... FTG 2.4
curtosis
Yearly maximum temperatures
Distribution in
scaling
Corrections to scaling