Finite samples, scaling and non-stationarity Khurom Kiyani and Sandra Chapman CFSA Sunday, 16 November 2008 1 The question What effects does the the partitioning of data into smaller segments have on the estimate of the scaling exponents? Sunday, 16 November 2008 2 The question What effects does the the partitioning of data into smaller segments have on the estimate of the scaling exponents? • On the way to answering these questions we will explore: 1. apparent or pseudo-nonstationarity, 2. limit theorems, 3. and the role of extreme values from heavy-tailed statistics. Sunday, 16 November 2008 3 The method Stationary Brownian motion (Gaussian) N=106 #!! ! !#!! !%!! !'!! !+!! !"!!! !"#!! !"%!! ! " # $ ,-./ % & ' * ()"! Partition time series i = 1 · · · L parts Calculate scaling exponents in these L parts Sunday, 16 November 2008 4 Test statistic and its scaling increments yi (t, τ ) = xi (t + τ ) − xi (t) )!! '!! %!! τ 12-3 #!! ! !#!! !%!! !'!! ! " # $ % & -./0 ' ( ) * "! & +,"! " # $ % & -./0 ' ( ) * "! & +,"! ' % +1-2!3 # ! !# !% !' ! Sunday, 16 November 2008 5 Test statistic and its scaling increments yi (t, τ ) = xi (t + τ ) − xi (t) )!! '!! %!! τ 12-3 #!! ! !#!! !%!! !'!! ! " # $ % & -./0 ' ( ) * "! & +,"! " # $ % & -./0 ' ( ) * "! & +,"! ' % +1-2!3 # second moment N ! 1 2 2 Mi (τ ) = (yi )j N j=1 ! !# !% !' ! Sunday, 16 November 2008 6 Test statistic and its scaling increments yi (t, τ ) = xi (t + τ ) − xi (t) )!! '!! %!! τ 12-3 #!! ! !#!! !%!! !'!! ! " # $ % & -./0 ' ( ) * "! & +,"! ' % +1-2!3 # second moment N ! 1 2 2 Mi (τ ) = (yi )j N j=1 moment scaling ! !# 2 Mi (τ ) !% !' ! " # $ % & -./0 ' ( ) * "! & +,"! 2 log Mi (τ ) = = 2 ζi (2) Mi (1)τ 2 log Mi (1) + ζi (2) log τ via ordinary least-squares regression Sunday, 16 November 2008 7 Test statistic and its scaling increments yi (t, τ ) = xi (t + τ ) − xi (t) )!! '!! %!! τ 12-3 #!! ! !#!! !%!! !'!! ! " # $ % We also use the iterative conditioning technique to handle second moment heavy-tailed statistics; see: N ! & -./0 ' ( ) * "! & +,"! 1 2 = (yi )j N j=1 K. Kiyani, S. C. Chapman and B. Hnat, Phys. Rev. Emoment 74, 051122 (2006). scaling 2 Mi (τ ) ' % # ! !# 2 Mi (τ ) !% !' ! " # $ % & -./0 ' ( ) * "! & +,"! 2 log Mi (τ ) = = 2 ζi (2) Mi (1)τ 2 log Mi (1) + ζi (2) log τ via ordinary least-squares regression Sunday, 16 November 2008 8 The model time series Lévy process )!! '!! • • • • %!! 12-3 #!! ! Infinite variance Stationary increments Heavy-tailed Self-similar -- monofractal !#!! !%!! !'!! ! Stationary Levy process with !=1.4 (N=106) " # $ % & -./0 ' ( ) * "! & +,"! "%!!! "#!!! Standard Brownian motion "!!!! +!!! Finite variance All moments finite Stationary increments Self-similar -- monofractal '!!! 0 • • • • %!!! #!!! ! !#!!! !%!!! ! Sunday, 16 November 2008 " # $ ,-./ % & ' * ()"! 9 The model time series # # p-model !"#!$%&%'"#&()*+(",#'&#*-"%'"#* ./&0$$'&#*'#1(2-2#%$*,'%3*4'#2&(4) 5&()'#6*!*,'%3*%'-27*!89:; '(&! • • • • " ! . !" !# Finite variance Nonstationary increments Heavy-tailed Self-similar -- multifractal !$ &#! !% " # *+,- $ % Nonstationary Brownian motion • Finite variance • Nonstationary increments PDF broader with time • Self-similar -- monofractal &"! &! ) '(&! &!! '()*+('(,-./*+0 !&! ! - 1!23456-+*25-05(*50 718!9$:-;8&!$ %! $! #! "! !- Sunday, 16 November 2008 10 Pseudo-nonstationarity &#! - &"! '()*+('(,-./*+0 &!! 1!23456-+*25-05(*50 718!9$:-;8&!$ %! $! #! "! !8 7 5 N=10 6 N=104 M2i (!=1) 5 4 3 2 1 0 Sunday, 16 November 2008 11 Pseudo-nonstationarity &#! - 0.2 &"! %! 0.15 !i(2) '()*+('(,-./*+0 &!! 1!23456-+*25-05(*50 718!9$:-;8&!$ $! 4 <!2> for N=10 0.1 6 <!2> for N=10 <!2> for N=105 #! 0.05 !2 for N=104 "! 5 !2 for N=10 0 !8 7 0.24 0.22 5 N=10 6 0.2 N=104 0.18 4 3 2 !i(2) M2i (!=1) 5 0.16 0.14 0.12 0.1 1 0.08 0 0.06 0 Sunday, 16 November 2008 2 4 time 6 8 10 5 x 10 12 Pseudo-nonstationarity &#! - 0.2 &"! 0.15 %! !i(2) '()*+('(,-./*+0 &!! 1!23456-+*25-05(*50 718!9$:-;8&!$ $! 4 <!2> for N=10 0.1 6 <!2> for N=10 <!2> for N=105 #! 0.05 !2 for N=104 "! 5 !2 for N=10 0 !8 7 0.24 0.22 5 N=10 6 0.2 N=104 0.18 !i(2) M2i (!=1) 5 4 3 0.16 0.14 0.12 2 0.1 1 0.08 0 0.06 0 2 points Sunday, 16 November 2008 2 4 time 6 8 10 5 x 10 13 Quantifying scatter/variation of ζ(2) with N !* % '(&! Standard Brownian + $ 6-78!8"9:59 ( ) # * " & !( ! " # $ ,-./01(2341(5 % &! ) '(&! Power law? Sunday, 16 November 2008 14 Quantifying scatter/variation of ζ(2) with N !* % '(&! Standard Brownian + $ 6-78!8"9:59 ( ) !! !" . # !' * !" & !( ! " # $ ,-./01(2341(5 % 3)45!5'6726 " !# !" &! !" '(&!) !$ !% !" Power law? !& !" . # !" Answer: yes Sunday, 16 November 2008 $ !" % ()*+,-./01-.2 !" & !" 15 Quantifying scatter/variation of ζ(2) with N −1 a.) !* % '(&! −1.5 ( −2 Var(ζ(2)) + 10 ) −3 −3.5 log 6-78!8"9:59 $ −2.5 # Standard Brownian p−model (p=0.6) fBm H=0.8 Non−stationary Brownian Cyclic−stationary Brownian −ve unit gradient −4 −4.5 * −5 " −5.5 3 & !( b.) ! " # $ ,-./01(2341(5 % &! ) '(&! 3.5 4 4.5 log10 N 5 5.5 5 5.5 1 0 log10 Var(ζ(2)) −1 −2 −3 Levy α=1.2 Levy α=1.4 Levy α=1.6 Levy α=1.8 LFSM H=0.44 α=1.4 LFSM H=0.9 α=1.6 −ve unit gradient −4 −5 −6 −7 Sunday, 16 November 2008 3 3.5 4 4.5 log10 N 16 Quantifying error on ζ(2) !* % '(&! ( + 6-78!8"9:59 $ ) !! !" . # * !' !" & !( ! " # $ ,-./01(2341(5 ! V ar(ζ(2)) ! 0.05 ζ(2)|L=1 3)45!5'6726 " !# !" % &! ) '(&! !$ !" !% !" !& !" . # $ !" !" % ()*+,-./01-.2 !" & !" V ar(ζ(2)) ! (0.05ζ(2)|L=1 ) 2 Sunday, 16 November 2008 17 Quantifying error on ζ(2) p-model N~105 Var(ζ(2)) −2 10 fBm, stationary & nonstationary Brownian N~103; −1.5 −2.5 −3 −3.5 log finite variance −1 a.) Standard Brownian p−model (p=0.6) fBm H=0.8 Non−stationary Brownian Cyclic−stationary Brownian −ve unit gradient −4 −4.5 −5 −5.5 3 Lévy α=1.8 & LFSM (H=0.9,α=1.6) N~103; Lévy α=1.6 N~104; Lévy α=1.4, 1.2 & LFSM (H=0.44,α=1.4) N~105; Sunday, 16 November 2008 4 4.5 log10 N 5 5.5 5 5.5 1 0 −1 log10 Var(ζ(2)) infinite variance b.) 3.5 −2 −3 Levy α=1.2 Levy α=1.4 Levy α=1.6 Levy α=1.8 LFSM H=0.44 α=1.4 LFSM H=0.9 α=1.6 −ve unit gradient −4 −5 −6 −7 3 3.5 4 4.5 log10 N 18 Quantifying error on V ar(ζ(2)) b.) a.) 0 −2 Var(ζ(2)) −3 10 −4 log log10 Var(ζ(2)) −1 −5 −6 −2 −3 −4 3 3.5 4 4.5 5 −5 5.5 3 3.5 4 4.5 log10 N log10 N c.) −1.5 d.) 5 5.5 0 −2 −1 log10 Var(ζ(2)) log 10 Var(ζ(2)) −2.5 −3 −3.5 −4 −4.5 −5 −5.5 Sunday, 16 November 2008 −3 (Mean) Standard Brownian linear fit ( y = −1.06*x + 1.39 ) 3 3.5 4 4.5 log10 N −2 5 5.5 −4 linear fit ( y = −1.4*x + 4.35 ) (Mean) Levy α=1.4 3 3.5 4 4.5 log N 5 5.5 6 10 19 Real-world data 0 log10 Var(ζ(2)) −1 −2 −3 −4 Levy α=1.4 ACE 2000 B2 WIND 1996 |V| WIND 1996 Bz Standard Brownian −ve unit gradient −5 2.5 3 3.5 4 4.5 log10 N 5 5.5 N~105 introduces an error of ~12% Sunday, 16 November 2008 20 Limit theorems ‘All epistemological value of the theory of probability is based on this: that large scale random phenomena in their collective action create strict, non-random regularity.’ (Gnedenko and Kolmogorov, Limit Distributions for Sums of Independent Random Variables) Sunday, 16 November 2008 21 Limit theorems Central Limit Theorem (De Moivre, Laplace, Lyapunov) SN N ! 1 yi =√ N i=1 Sunday, 16 November 2008 lim SN → Gaussian N →∞ 22 Limit theorems Central Limit Theorem (De Moivre, Laplace, Lyapunov) SN N ! 1 yi =√ N i=1 lim SN → Gaussian N →∞ Generalized Central Limit Theorem (Lévy) SN = Sunday, 16 November 2008 1 N 1/α N ! i=1 yi lim SN → Lévy N →∞ 23 Limit theorems Central Limit Theorem (De Moivre, Laplace, Lyapunov) SN N ! 1 yi =√ N i=1 lim SN → Gaussian N →∞ Generalized Central Limit Theorem (Lévy) SN = 1 N 1/α N ! yi i=1 and many others! Sunday, 16 November 2008 lim SN → Lévy N →∞ Stable processes 24 Limit theorems and self-similar processes SN N ! 1 yi =√ N i=1 SN = Sunday, 16 November 2008 1 N 1/α N ! i=1 yi Coarse-graining/averaging Scaling study of limit theorems and stable processes have a very profound link to self-similar processes 25 !(2) H(!(2)) 3 2.5 2.5 !(2) !(2) 4 1.4 Statistics of ζ(2) b ii.) 6 1.6 1.2 2 1 0 2 0 2 4 6 time 0.8 H(!(2)) b i.) !(2) 7 x 10 1 0.4 0 1.5 Levy "=1.4 N=1000 Gumbel MLE (!=1.15,#=0.25) 0.6 2 1.5 2 4 6 time 7 x 10 Levy "=1.4 N=10,000 Gumbel MLE (!=1.32,#=0.13) 1 0.5 0.2 1.5 1 1.5 !(2) 7 2 4 6 time 1 1.1 !(2) 1.2 1.3 #! ' .-#! 7 x 10 ! !"# !"$ !"% !"& !*$+ 2.5 !"' !"( 2.5 !(2) 3 3.5 2 , " 0 2 4 6 time 7 x 10 Levy "=1.4 N=10,000 Gumbel MLE (!=1.32,#=0.13) 1 0 4 !"( ) !&( ! #% ! ' /012 #! % ' .-#! 3!14526-*37!"(+-87#!!! 8491:6-;<= *!7!"#(>-"7!"!)+ $ !"' !"( !") , d i.) # !)#* $ /012 lim !"# !"#$ N →∞ !"% !"%$ !*%+ !"&$ !"' d ii.) 1 0.5 ! 0 1.5 "&' "&0123 "&( "&. ( /,"! 1.6 456,#!!!,7 ,8,",9!! 8:;2<=,>?6 )!@!&''A,#@!&$"* " # $ 1.8 1.9 8 x 10 ACE 2000 B2 N " 5000 Normal MLE (!=0.78, #=0.23) 0.5 % 1.7 time 1 # ! !"& 0 0.5 1 !)#* 1.5 2 !(2) √ #! $ .-#! N (ζ(2) − ζ̂(2)) ∼ N (0, σ 2 ) !"& !"&$ !"' via ordinary least-squares regression 1.5 2 " 1 0.5 ! 1.5 "&' "&0123 "&( "&. ( /,"! !(2)) !" Sunday, 16 November 2008 ! d ii.) "&# !"%$ !*%+ 3!14526-*37!"(+-87#!9!!! 84:1;6-<=> *!7!"#$)9-"7!"!%'+ ' !- !"% !"& !*$+ !"% !"# ( !- !"$ !"#$ !"#$ !"!$ ) % !"# !"# !"% #! # ! $ 1.5 !&' !, #! .-#! !"%$ !"$ Theorem: & 2.5 - #' !!"$ 2 #( ,*!*%++ !*$+ ' 1.5 c ii.) !"& ( 1 $ /012 3!14526-*37!"(+-87#!9!!! 84:1;6-<=> *!7!"#$)9-"7!"!%'+ 2 !&% !(2) - ! ) !- !") " 1.5 !*%+ 2 #! ' 1.5 !(2) 1.5 !"# !"!$ ( # !&# 1 !"% !"#$ % 0.5 0 ,*!*$++ ' /012 "&# 0.2 !)#** ! 3!14526-*37!"(+-87#!!! 8491:6-;<= *!7!"#(>-"7!"!)+ d i.) 3 1 Levy "=1.4 N=1000 Gumbel MLE (!=1.15,#=0.25) 0.4 !&( #% H(!(2)) 0 0.6 " ! $ !- 0.9 !(2) !(2) 0 0.8 c i.) #' & 2.5 !"%$ #( !" 1 2 - c ii.) !"$ !!"$ 1.5 !(2) # 2 2 1 % +)!)#** H(!(2)) 6 x 10 Standard Brownian N=10,000 Normal MLE (!=0.99, "=0.04) 4 1.2 4 6 2.5 1.4 2 time b ii.) 6 1.6 ' 0 0 2 H(!(2)) b i.) 1 0.9 2 0.5 0 ( 4 Standard Brownian N=1000 Normal MLE (!=0.974, "=0.127) 0 4 !(2) 7 ) 1 8 6 x 10 3.5 !"( !)#* 4 time H(!(2)) H(!(2)) 2 3 ,*!*%++ 10 0 2 2.5 !(2) !"& 2.5 0.5 2 !*$+ 1 !(2) 3 1.1 12 1.5 - c i.) ,*!*$++ a ii.) 1.5 !(2) a i.) 3.5 1 !*%+ 0 0 1.6 1.7 time 1.8 1.9 8 x 10 26 Statistics of ζ(2) 7 #% !!"$ ! ' /012 & #! ' .-#! 3!14526-*37!"(+-87#!!! 8491:6-;<= *!7!"#(>-"7!"!)+ $ !- 0.9 1 1.1 !(2) 1.2 1.3 4 time 0.8 6 7 x 10 Levy "=1.4 N=1000 Gumbel MLE (!=1.15,#=0.25) 0.6 0.4 ! !"# !"$ !"% !"& !*$+ !"' !"( , 2.5 !)#* " 2 " 1.5 0 2 4 time 6 7 x 10 Levy "=1.4 N=10,000 Gumbel MLE (!=1.32,#=0.13) 1 !&( 2.5 !(2) 3 3.5 4 Theorem: 0 1 1.5 2 "&' "&0123 "&( "&. ( /,"! !&' !, !"#$ !"% !"%$ !*%+ 456,#!!!,7 ,8,",9!! 8:;2<=,>?6 )!@!&''A,#@!&$"* ! " # $ lim N →∞ !"&$ !"' 1 0 1.6 1.7 time 1.8 1.9 8 x 10 1 ACE 2000 B2 N " 5000 Normal MLE (!=0.78, #=0.23) 0.5 % !"& 0.5 0 0.5 1 !)#* √ $ 1.5 2 # !&% 2.5 !(2) !"# 1.5 !&# 2 #! .-#! 3!14526-*37!"(+-87#!9!!! 84:1;6-<=> *!7!"#$)9-"7!"!%'+ ! 1.5 1 $ /012 d ii.) "&# 0.2 1.5 ! ) !- !") # 0.5 1 #! ' H(!(2)) 2 !"# !"!$ % d i.) 3 !(2) !(2) 0 !"#$ ( !" 0 !"% !*%+ ! # 2 2 1 0 #' % +)!)#** H(!(2)) 7 !"$ What about explaining finite N? 4 1.2 6 x 10 Standard Brownian N=10,000 Normal MLE (!=0.99, "=0.04) 2.5 1.4 4 6 b ii.) 6 1.6 2 time 0 2 H(!(2)) b i.) 1.5 !(2) 0 2 0.5 1 ' 4 Standard Brownian N=1000 Normal MLE (!=0.974, "=0.127) 0 0.9 8 6 x 10 1.5 1 ( !(2) 4 time H(!(2)) H(!(2)) 2 1 !"%$ #( ,*!*%++ 10 0 2 ) - c ii.) !"& 2.5 0.5 !"( !*$+ 1 !(2) 3 1.1 12 - c i.) ,*!*$++ a ii.) 1.5 !(2) a i.) 3.5 1.5 2 !(2) N (ζ(2) − ζ̂(2)) ∼ N (0, σ 2 ) via ordinary least-squares regression Sunday, 16 November 2008 27 Statistics of ζ(2) -- finite N N ! 1 2 2 Mi (τ ) = (yi )j N j=1 2 log Mi (τ ) Sunday, 16 November 2008 = 2 log Mi (1) + ζi (2) log τ 28 Statistics of ζ(2) -- finite N N ! 1 2 2 Mi (τ ) = (yi )j N j=1 2 log Mi (τ ) = 2 log Mi (1) + ζi (2) log τ 1 = Tlog Z + √ ! N ! t "−1 t 2 Z = Tlog Tlog Tlog Mlog 2 Mlog Sunday, 16 November 2008 29 Statistics of ζ(2) -- finite N N ! 1 2 2 Mi (τ ) = (yi )j N j=1 2 log Mi (τ ) = 2 log Mi (1) + ζi (2) log τ 1 = Tlog Z + √ ! N ! t "−1 t 2 Z = Tlog Tlog Tlog Mlog 2 Mlog ζi (2) = k ! j=1 Sunday, 16 November 2008 aj " # 2 log Mi (τj ) 30 Statistics of ζ(2) -- finite N Sum 1 Sum 2 N ! 1 2 2 Mi (τ ) = (yi )j N j=1 ζi (2) = k ! j=1 Sunday, 16 November 2008 aj " # 2 log Mi (τj ) 31 Statistics of ζ(2) -- finite N Sum 1 finite variance N ! 1 2 2 Mi (τ ) = (yi )j N j=1 infinite variance ? Sunday, 16 November 2008 32 Statistics of ζ(2) -- finite N Sum 1 finite variance N ! 1 2 2 Mi (τ ) = (yi )j N j=1 infinite variance Gaussian Sunday, 16 November 2008 33 Statistics of ζ(2) -- finite N Sum 1 finite variance N ! 1 2 2 Mi (τ ) = (yi )j N j=1 infinite variance Frechet! Extreme value distribution for the maximum Gaussian P Sunday, 16 November 2008 ! Mi2 " = Λ 1+α/2 2 2 (Mi ) # exp − Λ α/2 2 α (Mi ) $ 34 Statistics of ζ(2) -- finite N ζi (2) = k ! aj j=1 finite variance Gumbel min-stable x −ex P (x) ∼ e e Sunday, 16 November 2008 " # 2 log Mi (τj ) infinite variance Gumbel max-stable −x −e−x P (x) ∼ e e 35 Statistics of ζ(2) -- finite N Sum 2 ζi (2) = k ! j=1 finite variance Gaussian Convergence fast Sunday, 16 November 2008 aj " # 2 log Mi (τj ) infinite variance Gaussian Convergence slow 36 2 4 8 6 time H(!(2)) H(!(2)) 0 2 7 x 10 1.5 H(!(2)) 0.9 1 1.1 !(2) b ii.) !(2) 1.2 1 0 2 0 2 4 6 time 0.8 7 x 10 0.4 2 1.5 1 0 1.5 Levy "=1.4 N=1000 Gumbel MLE (!=1.15,#=0.25) 0.6 1.3 2.5 4 2 1.2 3 2.5 1.4 7 Standard Brownian N=10,000 Normal MLE (!=0.99, "=0.04) 0 2 6 1.6 6 x 10 6 !(2) b i.) 1.5 !(2) H(!(2)) Statistics of ζ(2) 1 4 2 0.5 0 2 time 4 Standard Brownian N=1000 Normal MLE (!=0.974, "=0.127) 1 0 2 4 6 time 7 x 10 Levy "=1.4 N=10,000 Gumbel MLE (!=1.32,#=0.13) 1 0.5 0.2 7 x 10 1.5 1.5 !(2) 7 #% ! ' /012 #! ' .-#! 3!14526-*37!"(+-87#!!! 8491:6-;<= *!7!"#(>-"7!"!)+ $ !- 0.9 1 1.1 !(2) 1.2 1.3 2 4 time 6 7 x 10 0.6 0.4 ! !"# !"$ !"% !"& !*$+ !"' !"( , " 1.5 2 4 6 time 7 x 10 Levy "=1.4 N=10,000 Gumbel MLE (!=1.32,#=0.13) 1 2.5 !(2) 3 3.5 0 4 !&( !"( !"$ #' ! #% !!"$ ! ' /012 & #! ' .-#! 3!14526-*37!"(+-87#!!! 8491:6-;<= *!7!"#(>-"7!"!)+ !"&$ !"' 1 0 1.5 "&' "&0123 "&( "&. ( /,"! 1.6 456,#!!!,7 ,8,",9!! 8:;2<=,>?6 )!@!&''A,#@!&$"* " # !)#* $ 1.8 1.9 8 x 10 ACE 2000 B2 N " 5000 Normal MLE (!=0.78, #=0.23) 0.5 % 1.7 time 1 # ! !"& 0 0.5 1 1.5 2 !(2) !"% !"# !"!$ #! ! $ /012 #! $ .-#! ) 3!14526-*37!"(+-87#!9!!! 84:1;6-<=> *!7!"#$)9-"7!"!%'+ ( ' # !"%$ !*%+ 0.5 !&' !, !"% !"#$ % $ 2.5 !"#$ !"%$ #( ,*!*%++ !*$+ ' 2 - c ii.) !"& ( 1.5 !"# ! !&% !(2) - ) 1 !*%+ 2 $ 1.5 !&# 1.5 #! .-#! 3!14526-*37!"(+-87#!9!!! 84:1;6-<=> *!7!"#$)9-"7!"!%'+ 2 1.5 0 $ /012 d ii.) " 2 0.2 1 ! ) !- !") # 0.5 0 !"!$ #! ' "&# 2.5 1 Levy "=1.4 N=1000 Gumbel MLE (!=1.15,#=0.25) !"# ( H(!(2)) 0 !"% !"#$ % d i.) 3 !(2) !(2) 0 0.8 ,*!*$++ ! !" 1 c i.) #' & 2.5 !"%$ #( # 2 2 2 - c ii.) !"$ !!"$ 1.5 % +)!)#** H(!(2)) 6 x 10 Standard Brownian N=10,000 Normal MLE (!=0.99, "=0.04) 4 1.2 4 6 2.5 1.4 2 time b ii.) 6 1.6 0 0 2 H(!(2)) b i.) 1 ' 2 0.5 0 0.9 4 Standard Brownian N=1000 Normal MLE (!=0.974, "=0.127) 1 ( 1 !(2) !"( ) 1 8 6 0 4 !(2) 4 time H(!(2)) H(!(2)) 2 3.5 ,*!*%++ 10 0 2 3 !"& 2.5 0.5 2.5 !(2) !)#* 1 !(2) 3 12 2 - c i.) 1.1 1.5 !*$+ a ii.) 1.5 ,*!*$++ 3.5 !(2) a i.) 1 !*%+ 0 % Sunday, 16 ! - November 2008 ! !"# !"$ !"% !"& !"' !"( !") !- !"# !"#$ !"% !"%$ !"& !"&$ !"' 37 Conclusions • Finite N behaviour studied here. • Apparent or pseudo nonstationarity is introduced due to small sample size -intuitively makes sense. • Important note not considered in detail here: the errors determined here depend on the statistical estimator used -- other estimators might be better; conversely others might be worse. • Distinguish scaling stationarity and parameter stationarity. If forecasting then might be better to consider latter. But for investigating fundamental universal feature (scaling) the former should suffice as long as one is sensible. • BUT need to distinguish between more compact PDFs as opposed to the more hazardous heavy or fat-tailed PDF -- some common results might not apply; in particular convergence to limiting statistics is slow. Sunday, 16 November 2008 38 Conclusions Relevant papers are now on the ISSI group directory: ‣K. Kiyani, S. C. Chapman and N. Watkins, preprint (2008), submitted to Phys. Rev. E. ‣K. Kiyani, S. C. Chapman and B. Hnat, Phys. Rev. E 74, 051122 (2006). Sunday, 16 November 2008 39