Serii Tempoale Arma Stationare
Serii Tempoale Arma Stationare
Serii Tempoale Arma Stationare
> AirPassengers
Jan Feb.Mar.Apr.May Jun.Jul.Aug.Sep.Oct.Nov.Dec.
1949
112 118 132 129 121 135 148 148 136 119 104 118
1950
115 126 141 135 125 149 170 170 158 133 114 140
1951
145 150 178 163 172 178 199 199 184 162 146 166
1952
171 180 193 181 183 218 230 242 209 191 172 194
1953
196 196 236 235 229 243 264 272 237 211 180 201
1954
204 188 235 227 234 264 302 293 259 229 203 229
1955
242 233 267 269 270 315 364 347 312 274 237 278
1956
284 277 317 313 318 374 413 405 355 306 271 306
1957
315 301 356 348 355 422 465 467 404 347 305 336
1958
340 318 362 348 363 435 491 505 404 359 310 337
1959
360 342 406 396 420 472 548 559 463 407 362 405
1960
417 391 419 461 472 535 622 606 508 461 390 432
>
>plot(AirPassengers)
>
(h) ;
(h) ; 8t 2 Z
(W N 0;
) daca
(t) = 0; 8t 2 Z
2
;
0;
(h) =
h=0
h 6= 0
1 Xt 1
:::
p Xt p
= f t ; t 2 Zg este W N 0;
i; j
1 t 1
+ ::: +
2 R 8i; j;
q t q;
6= 0;
8t; unde
6= 0:
Notatii
B j Xt = Xt j ; j = 0; 1; 2:::::
p
(z) = 1
:::
1z
pz
(z) = 1 + 1 z + ::: + q z q
Cu aceste notatii, ecuatia cu diferente ARM A se scrie
(B) Xt = (B)
(z) = 1;
Xt = (B)
q
X
jM
t j)
=0
j=0
cov (Xt+h ; Xt ) =
8
>
<
>
:
q jhj
j=0
j+jhj ;
jhj
j h j> q
AR (p) : (z) = 1;
(B) Xt =
1 Xt 1
=
Xt =
Xt =
k
X
j
1) t j
t 1
1 Xt 2 )
k+1
1)
+(
Xt
= ::::::
k 1
j=0
2 def
Xt
1)
t j
2(k+1)
1)
=(
j=0
Seria
1
X
j
1) t j
kXt
2
k 1k
! 0 pt k ! 1
j=0
1
X
j
1) t j
j=0
M (Xt ) =
1
X
j
1)
M(
t j)
j=0
00
1
X
cov (Xt+h ; Xt ) = lim M @@
(
n!1
1)
t+h j
j=0
1
X
jhj
= 2 ( 1)
(
2j
1)
j=0
(1
= 0; 8t
10
A@
1
X
1)
t j
j=0
jhj
1)
2)
1
11
AA =
(h) ; 8t
Totusi, scriind
Xt =
1
t+1
1
1
Xt+1 = :::
Xt =
1)
j
t+j
j=1
Observam ca Xt astfel obtinut este corelat cu f s ; s > tg; ceea ce este nenatural (corelarea starii prezente a procesului cu inovatiile viitoare).
Daca j 1 j = 1; ecuatia cu diferente AR(1) nu are solutie stationara.
Denitie
O serie temporala ARM A (p; q) de forma
(B) Xt = (B)
Xt =
1
X
j t j;
j=0
t2Z
si
(z) =
1
X
(z)
pentru jzj
(z)
zj =
j=0
Demonstratie:
Pp. ca (z) 6= 0 pentru jzj
admite dezvoltarea in serie
X
1
=
cj z j = c (z) ; jzj < 1 + ":
(z) j=0
j
"
2
8j = 0; 1; 2; :::
In particular,
1
X
j=0
c (z)
jcj j < 1
(z) = 1 pentru jzj
1
X
j t j
j=0
cu
(z) =
1
X
(z)
pentru jzj
(z)
zj =
j=0
(B) Xt =
Notand
(z) =
(z)
(z) =
1
X
j=0
putem scrie
q
X
j t j
j=0
Inmultind cu
t k
(B)
(B) t ; 8t
z j pentru jzj
1
X
1;
j t j
j=0
k;
0;
k = 0; 1; :::; q
;
k>q
deci
(z) = (z) =
jzj
(z)
Cum polinoamele (z) si (z) nu au radacini comune iar j (z)j < 1 pentru
1; rezulta ca (z) 6= 0 pentru jzj 1:
Observatie:
Daca X = fXt ; t 2 Zg este o solutie stationara a ecuatiei cu diferente
1
X
ARM A (p; q) cu (z) 6= 0 pentru jzj 1; atunci rezulta Xt =
j t j:
j=0
Reciproc, daca Xt =
1
X
j t j;
atunci (B) Xt =
(B)
(B)
= (B) t .
j=0
2) AR (1) ; Xt
1 Xt 1 = t ; este cauzal () polinomul
radacina de modul mai mare ca 1; adica
j
(z) = 1
z 1
z+1
y+1
z=
y 1
y=
z2
= [ 1; 1] () y > 0
y2 (
y+1
y 1
y+1
y 1
1) + y (2
+ 2) +
+ 1)
>0
1
2
1
1
2+ 1
>0
1
2
1
2
2
are
j< 1
3) AR (2) ; Xt
1 Xt 1
2 Xt 2 = t ; este cauzal () polinomul
2
z
z
are
radacinile
inafara
cercului unitate.
1
2
Consideram
1+
1z
<1
1 <1
1< 2<1
1
=0
2
(z) =
Denitie:
O serie temporala ARM A (p; q) denita de ecuatia cu diferente (B) Xt =
(B) t se numeste inversabila daca exista un sir de constante f j ; j = 0; 1; :::g
asa incat
1
X
j jj < 1
j=0
1
X
j Xt j ;
j=0
8t 2 Z
1
X
(z)
pentru jzj
(z)
zj =
j=0
1 ; :::;
p
X
i Xt i
t;
i=1
unde = f t ; t 2 Zg este W N 0;
Xt2
(0) =
p
X
i=1
p
X
t2Z
: Avem M (Xt ) = 0; 8t 2 Z
i Xt i Xt
t Xt
(i) + M ( t Xt )
i=1
Dar
t Xt =
p
X
i Xt i t
i=1
M ( t Xt ) = M
2
t
2
t
(0) =
p
X
(i) +
i=1
1;
p
X
i Xt i Xt k
i=1
p
X
(k) =
(k
t Xt k
i)
i=1
(0) =
(i) +
(k
i=1
p
X
(k) =
i) ; k = 1; :::; p
i=1
(k) =
(0) = 1
Structura de corelatie a seriei temporale este descrisa de relatiile
Yule - Walker:
p
X
(k) =
(k i) ; k = 1; :::; p
i
i=1
j< 1:
2
(0) =
k
1
(k) =
(1) =
Cazul AR (2) cauzal, cu
(0) =
(1) =
(2) =
+
1
1
1
2
1
(0) ; k
1
1
< 1;
(1) +
(0) +
(1) +
2
2
2
2
< 1;
(2) +
(1)
(0)
1<
2
<1:
(0) =
(1
2)
2
2
2
1
3
2
(1) =
(2) =
3
2
2
1
2
3
2
2
1
2
2
1
2+
2
2
+1
2
1 2
2
2
2
1 2
+1
2
2
(1) =
(2) =
2
1 2
+1
(1)
(1) + 2
2
(1) =
(2) =
2
1
+
1
2
2
2
q
X
t+
j t j
j=1
(k) =
8
>
<
>
:
(0) =
(k) =
8 q
>
< X
>
:
j=0
q
X
j j+k ;
j=0
0;
1+
j j+k
j=0
k>q
2
1
, q
X
+ ::: +
2
j
2
q
k = 1; 2; :::; q
j=0
0;
k>q
(0) = 1
j< 1:
(0) =
1+
(1) = 2 1
(k) = 0; k > 1
10
2
1
= 1:
1 Xt 1
+ ::: +
1 t 1
q t q
+ ::: + p Xt p Xt + t Xt + 1 t 1 Xt + ::: + q t q Xt
(1) + ::: + p (p) + M ( t Xt ) + 1 M ( t 1 Xt ) + ::: + q M (
1 Xt 1 t
2
M ( t Xt ) = 0 +
t 1 Xt
M(
p Xt p
1 Xt 1 Xt
t Xt
M(
+ ::: +
1 Xt 1 t 1
t 1 Xt )
t 2 Xt
1 Xt 1 t 2
t 2 Xt )
1M
(Xt
2
t
p Xt p t 1
p Xt p t
+ ::: +
1 t 1 t
q t q t
+0
+ ::: +
t q Xt )
+ ::: +
t t 1
2
1 t 1
+ ::: +
q t q t 1
1)
2 Xt 2 t 2 :::
1 t 2)
+
2)
p Xt p t 2
2
+
1
t t 2
1)
+(
1 t 1 t 2
2
2 ))
etc:
Putem insa utiliza reprezentarea procesului ARM A (p; q) ca proces cauzal,
Xt =
1
X
j t j;
t2Z
j=0
Obtinem
(0) =
(1) + ::: +
(p) +
q
X
j=0
(k
1) + ::: +
(k
p) +
q
X
j=k
k < maxfp; q + 1g
(k) = 1 (k 1) + ::: +
k maxfp; q + 1g
11
(k
p)
j k;
si mediind
2
2 t 2
+ ::: +
q t q t 2
(1) =
(0) =
(1) =
(k) =
1 Xt 1
1 Xt 1 Xt
2
1 Xt 1
1
1
1
t Xt
(0) +
(1) +
(1) +
t Xt 1
1
1 t 1
1 t 1 Xt
1
1 t 1 Xt 1
(1 +
(0) + 1
(k 1) ; k
(0) =
(1) =
(k) =
(1) =
(k) =
1 + 2 1 1 + 12
2
1
1
2
(1 + 1 1 ) ( 1 +
2
1
1
(k
1) ; k
1)
(1 + 1 1 ) ( 1 + 1 )
(1 + 2 1 1 + 12 )
1
(k
12
1)
1) ; k
1 ))
(F
(Q1
Q1
13
B ! [0; 1]
Construim
F
(Q1
::::: (Qn
Qn ))
1 ; :::; X0 )
= P (Xn+1 2 B j Xn ) = Qn (Xn
1 ; B)
P a:s:; 8B 2 B
B1 ::: Bn ) = F
8B0 ; B1 ; :::; Bn 2 B
(Q
::::: (Q
Q)) (B0
B1
Denitii:
Fie X = fXt ; t 2 N g un lant Markov omogen, cu repartitia de trecere Q:
Q se numeste ergodica daca exista o probabilitate pe (R; B) asa incat
pentru 8B 2 B
lim P (Xn 2 B) = (B) ;
n!1
Lantul Markov se numeste regulat daca repartitia sa de trecere este ergodica, iar repartitia limita este stationara pentru Q:
Model parametric pentru un lant Markov regulat, cu multimea
starilor (R; B)
( ; K; P ) ;
14
Rk ; k
:::
Bn )
f (x; y; )
0; 8y 2 R
f (x; y; ) dy = 1
(densitatea de trecere)
(B) =
f (x; ) dx
f (x; )
0; 8x 2 R
Observatie
Densitatea de repartitie a vectorului (X0 ; X1 ; :::; Xn ) este
f (x0 ; x1 ; :::; xn ; ) = f (x0 ; )
n
Y
f (xi
1 ; xi ;
i=1
; ;
2 R;
15
ce va precizata
2 ( 1; 1) ;
2 (0; 1)
1
2
2
(1) =
(0) =
2
2
; k
(1)
=
(0)
Propozitie
Procesul AR (1) cauzal, cu inovatii gaussiene este un lant Markov regulat,
cu
Q (x; ) = N
+ x;
=N
2
2
(y
( + x))
Demonstratie:
Consideram
Q (x; B) = p
1
2
exp
dy
P (Xn+1 2 B j Xn = x) = P ( + x +
=p
1
2
exp
n+1
2 B) =
1
(y
( + x+
2
( + x))
n+1
(!)) dP (!) =
dy = Q (x; B)
Proprietatea Markov :
Notam Fn = B (X0 ; :::; Xn ) si e A 2 Fn arbitrar ales.
Z
Z
P (Xn+1 2 B j Fn ) (!) dP (!) = M
X 1 (B) j Fn (!) dP (!) =
n+1
1
Xn+1
(B)
16
(!) dP (!) = P
1
A \ Xn+1
(B)
a:s:
1
A \ Xn+1
(B) = P d 1 (f(! 0 ; !) j ! 0 2 A; + Xn (! 0 ) + n+1 (!) 2 Bg) =
= (P
P ) (f(! 0 ; !) j ! 0 2 A; + Xn (! 0 ) + n+1 (!) 2 Bg) =
Z
= P (f! j + Xn (! 0 ) + n+1 (!) 2 Bg) dP (! 0 ) =
A
Q (Xn (! 0 ) ; B) dP (! 0 )
Rezulta
P (Xn+1 2 B j Fn ) = Q (Xn ; B) P
a:s:
+ x;
Repartitia stationara:
Notam cu ' (t) functia caracteristica a repartitiei stationare : Determinarea
analitica a lui ' se face pornind de la conditia
'Xn+1 (t) = ' (t) ; 8t
M (exp (it ( + Xn +
n+1 )))
= ' (t) ; 8t
n+1 ))
= ' (t) ; 8t
= ' (t) ; 8t
1 2
t
2
= ' (t) ; 8t
it
1
exp
verica aceasta conditie.
- Notam
2
2)
' (t)
g (t) =
exp
it
1
t2
2(1
2
2)
17
t2
2 (1
it
1
2)
=N
1
2
exp
1
2
(y
M (Xn+1 j Xn = xn ) =
2
D (Xn+1 j Xn = xn ) =
x)
; x2R
; y2R
+ xn
2
Observatie:
Din proprietatile lui R ca spatiu Hilbert (cu produsul scalar < X; Y >=
cov (X; Y )), stim ca cel mai bun predictor liniar al lui Xn+1 ; cunoscand
fXt ; t = n; n 1; :::g este
\
X
n+1 = M (Xn+1 j Xn ) =
+ Xn
18
(F
R2
1)
= Qn ((Xn
1 ; Xn ) ; B)
= x2 ) =
f (x1 ; x2 ; x3 ; )
0; 8x3 2 R
(B) =
f (x1 ; x2 ; )
0; 8x1 ; x2 2 R
R2
19
1;
2;
1 Xt 1
2 Xt 2
< 1;
t;
t2N
< 1; 1 <
< 1;
2 (0; 1)
M (Xt ) = 0; 8t
2
(0) =
3
2
2
1
(1
2
2
2
2
(1) =
(k) =
3
2
2
1
(k
2)
2
1 2
+1
2
1 2
+1
2
2
1) +
(k
2) ; k
Propozitie
Procesul AR (2) cauzal, cu inovatii Gaussiene este un lant Markov de ordinul
2, cu
Q (x1 ; x2 ; ) = N 1 x2 + 2 x1 ; 2
=N
2;
0
0
(0)
(1)
(1)
(0)
Generalizare:
Procesul AR (p) cauzal, cu inovatii Gaussiene este un lant Markov de ordinul
p, cu
Q (x1 ; :::; xp ; ) = N 1 xp + ::: + p x1 ; 2
0 0
1 0
11
0
(0)
(1)
:::
(p 1)
B B 0 C B
C
(1)
(0)
:::
(p 2) C
B
C B
CC
=NB
@p; @ ::: A ; @
AA
:::
:::
:::
:::
0
(p 1)
(p 2) :::
(0)
20
si functie de
b (X1 ; ::::Xn ) = X n =
1X
Xt
n t=1
n h
1X
Xt
n t=1
Xn
b (h) =
Propozitie
Xt+h
Xn ; 0
b (h)
b (0)
M Xn =
D2 X n = M X n
nD2 X n
1
X
h= 1
! 0 daca
(n) ! 0 pentru n ! 1
1
X
h= 1
(h) j< 1
Demonstratie:
Nedeplasarea estimatorului X n este imediata.
nD2 X n =
n
X
1 X
cov (Xi ; Xj ) =
n i;j=1
jhj<n
21
jhj
n
(h)
jhj<n
(h) j
Daca
deci
D2 X n
1 X
j
n
(h) j! 0:
jhj<n
Daca
1
X
h= 1
n!1
(n) j= 0
n!1
jhj<n
jhj
n
(h) =
1
X
(h)
h= 1
Comentariu:
Propozitia arata ca daca (n) ! 0 pentru n ! 1; atunci X n converge in
medie patratica (deci si in probabilitate) la :
1
X
j (h) j< 1 (conditie
Daca este indeplinita conditia mai puternica
h= 1
1
1 X
n
(h)
h= 1
Propozitie
Daca procesul X = fXt ; tg este (slab) stationar si Gaussian, atunci
0
1
X
p
jhj
n Xn
N @0;
1
(h)A
n
jhj<n
e matricea (n
0
Y1
:::
:::
Y1
Y2
:::
Yn
Atunci, 8a 2 Rn ;
Y2
:::
:::
0
2n) dimensionala
1
::: Yn
Yn 0 C
C
::: ::: A
::: 0
bn = 1 T T 0
n
1 0
0
(a T ) (a0 T )
0
n
Matricea de autocovarianta de selectie b n si matricea de autocorelatie de
bn = b n = b (0) sunt semipozitiv denite.
selectie R
a0 b n a =
Desigur, nu pot construiti estimatori pentru (h) cu h n dintr-o traiectorie observata (X1 ; :::; Xn ).
Se adopta ca estimator al functiei de autocovarianta
8
n jhj
>
< 1 X
Xt X n Xt+h X n ; j h j< n
n
b (h) =
>
t=1
:
0;
jhj n
Functia b (h) este semipozitiv denita, deci ea poate functia de autocovarianta a unui proces stationar.
=========================
Dar, functia de autocovarianta a unui proces stationar are urmatoarea proprietate:
Proprietate
Daca (0) > 0 si (h) ! 0 pentru h ! 1; atunci matricea n = k (i j)ki;j=1;:::;n
este nesingulara pentru orice n (lungime a traiectoriei observate)
Demonstratia se face prin absurd (Brockwell & Davis, propozitia 5.1.1., pag
167 - 168)
=========================
Aplicand acest rezultat, obtinem det b n > 0 daca b (0) > 0:
23
p
X
(i) +
(k
i=1
(k) =
p
X
i) ; i = 1; :::; p
i=1
Utilizand notatia
0
= ( 1 ; :::; p ) ;
= k (i j)ki;j=1;:::;p ;
0
j)ki;j=1;:::;p =
(i
j)
(0)
i;j=1;:::;p
(0)
bp
=b
c2 ; b1 ; :::; bp
= b (0)
c2 = b (0)
(b)
24
bp
b= R
bp
c2 = b (0) 1
(b)
bp
R
f t ; tg
+ ::: + b(m)m Xt
2
W N 0; d
(m)
1=2
pentru m > r
b(1)1 = b (1) ;
2
d
2
(b (1)) ;
(1) = b (0) 1
0
1
m
1
X
1
b(m 1) j b (m j)A
b(m)m =
@b (m)
2
\
j=1
(m 1)
0
1
0
1
b(m 1)m 1
b(m)1
@
A = b(m 1) b(m)m @
A
:::
:::
b(m)m 1
b(m 1)1
d
2
2
\
(m) = (m 1) 1
b(m)m
Comentariu
In cazul seriilor AR (p) cu inovatii Gaussiene, estimarea indicelui de dependenta p se face prin alegerea modelului care minimizeaza creiteriul informational
Akaike (AIC).
25
(m)m
EXEMPLU
Seria temporala observata "Yearly Sunspot Data, 1700-1988"
sunspot.year package: datasets R Documentation
Yearly Sunspot Data, 1700-1988
Description: Yearly numbers of sunspots.
Usage: sunspot.year
Format: The univariate time series sunspot.yearcontains 289 observations,
and is of class "ts".
> library(utils)
> sunspot.year
Time Series:
Start = 1700
End = 1988
Frequency = 1
[1] 5.0 11.0 16.0 23.0 36.0 58.0 29.0 20.0 10.0 8.0 3.0 0.0
[13] 0.0 2.0 11.0 27.0 47.0 63.0 60.0 39.0 28.0 26.0 22.0 11.0
[25] 21.0 40.0 78.0 122.0 103.0 73.0 47.0 35.0 11.0 5.0 16.0 34.0
[37] 70.0 81.0 111.0 101.0 73.0 40.0 20.0 16.0 5.0 11.0 22.0 40.0
[49] 60.0 80.9 83.4 47.7 47.8 30.7 12.2 9.6 10.2 32.4 47.6 54.0
[61] 62.9 85.9 61.2 45.1 36.4 20.9 11.4 37.8 69.8 106.1 100.8 81.6
[73] 66.5 34.8 30.6 7.0 19.8 92.5 154.4 125.9 84.8 68.1 38.5 22.8
[85] 10.2 24.1 82.9 132.0 130.9 118.1 89.9 66.6 60.0 46.9 41.0 21.3
[97] 16.0 6.4 4.1 6.8 14.5 34.0 45.0 43.1 47.5 42.2 28.1 10.1
[109] 8.1 2.5 0.0 1.4 5.0 12.2 13.9 35.4 45.8 41.1 30.1 23.9
[121] 15.6 6.6 4.0 1.8 8.5 16.6 36.3 49.6 64.2 67.0 70.9 47.8
[133] 27.5 8.5 13.2 56.9 121.5 138.3 103.2 85.7 64.6 36.7 24.2 10.7
[145] 15.0 40.1 61.5 98.5 124.7 96.3 66.6 64.5 54.1 39.0 20.6 6.7
[157] 4.3 22.7 54.8 93.8 95.8 77.2 59.1 44.0 47.0 30.5 16.3 7.3
[169] 37.6 74.0 139.0 111.2 101.6 66.2 44.7 17.0 11.3 12.4 3.4 6.0
[181] 32.3 54.3 59.7 63.7 63.5 52.2 25.4 13.1 6.8 6.3 7.1 35.6
[193] 73.0 85.1 78.0 64.0 41.8 26.2 26.7 12.1 9.5 2.7 5.0 24.4
[205] 42.0 63.5 53.8 62.0 48.5 43.9 18.6 5.7 3.6 1.4 9.6 47.4
[217] 57.1 103.9 80.6 63.6 37.6 26.1 14.2 5.8 16.7 44.3 63.9 69.0
[229] 77.8 64.9 35.7 21.2 11.1 5.7 8.7 36.1 79.7 114.4 109.6 88.8
[241] 67.8 47.5 30.6 16.3 9.6 33.2 92.6 151.6 136.3 134.7 83.9 69.4
[253] 31.5 13.9 4.4 38.0 141.7 190.2 184.8 159.0 112.3 53.9 37.5 27.9
[265] 10.2 15.1 47.0 93.8 105.9 105.5 104.5 66.6 68.9 38.0 34.5 15.5
[277] 12.6 27.5 92.5 155.4 154.7 140.5 115.9 66.6 45.9 17.9 13.4 29.2
[289] 100.2
> plot(sunspot.year)
26
0:5038Xt
0:1931Xt
3:
+ 0:0630Xt
4:
+ 49:1297 +
Functia ar
(in ipoteza suplimentara de inovatii Gaussiene)
ar {stats}
R Documentation
Fit Autoregressive Models to Time Series
Description: Fit an autoregressive time series model to the data, by default
selecting the complexity by AIC.
Usage
ar(x, aic = TRUE, order.max = NULL, method=c("yule-walker", "burg",
"ols", "mle", "yw"), na.action, series, ...)
> sunspot.ar<-ar(sunspot.year)
> sunspot.ar
Call:
ar(x = sunspot.year)
Coe cients:
123456789
1.1305 -0.3524 -0.1745 0.1403 -0.1358 0.0963 -0.0556 0.0076 0.1941
Order selected 9, sigma^2 estimated as 267.5
28
b (0) =
q
X
j j+k ;
k = 1; :::; q
j=0
2
1+
2
1
2
q
+ ::: +
- Calculam bq ; bq
1 ; :::; 2 ; 1
bj = b (j)
c2
j = q; q
1 j+1
1; :::; 2; 1
:::
q j q
+ b(m)1
t 1
2
W N 0; d
(m)
+ ::: + b(m)m
t m
(m)m
Pentru m = 1; 2; :::
b(m)m
0
1 @
b (m
=
d
2
k
X1
k)
j=0
(k)
k = 0; 1; :::; m
1
k
X1
d
2
(m) = b (0)
b(m)m
b(m)m
j=0
2
j
b(k)k
d
2 A;
(j)
d
2
(j)
b(1)1 = b (1) =
0:67832
d
2
(1) = b (0)
b(1)1
d
2
(0) = 4:0785
b(2)1 = 1 b (1)
d
2
(1)
d
2
(2) = b (0)
b(2)2
2
b(2)2 b(1)1 d
(0) =
2
d
2
(0)
30
b(2)1
1:0268
d
2
(1) = 3:002
1 ; :::;
pg
b (q + 1) = 1 b (q) + 2 b (q 1) + ::: +
b (q + 2) = 1 b (q + 1) + 2 b (q) + ::: +
::::::::::::::::::::::::
b (q + p) = 1 b (q + p 1) + 2 b (q + p
p b (q
p b (q
p+
p + 1)
p + 2)
2) + ::: +
p b (q)
+ ::: + bp Xt
1 t 1
+ ::: +
q t q
bj = c (j)
c2
j = q; q
1+
1 j+1
1; :::; 2; 1
31
c (0)
+ ::: +
2
1
:::
2
q
q j q
X0
Xt =
=
f (x; ) = q
2
exp
2
2
f (x; y; ) = p
+ Xt
2 0
; ;
2 R;
t;
; x2R
; y2R
2
Pentru traiectoria observata (X0 ; X1 ; :::; Xn ) (!) = (x0 ; x1 ; :::; xn ) ; desitatea
de repartitie este
2
f (x0 ; x1 ; :::; xn ; )
(
1
1
=q
exp
2
2
21 2
2 1 2
exp
o
2 (0; 1)
)
2 ( 1; 1) ;
1
21
t2N
x0
(y
x)
n
Y
i=1
1
2
n
log (2 )
2
n
log
2
n
1 X
2
(xi
i=1
8
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
:
n +
n
X
xi
i=1
1
n
n
X
xi
i=1
n
X
n
X
x2i
i=1
(xi
i=1
32
=
1
n
X
xi
i=1
n
X
xi
1 xi
i=1
xi
2
1)
exp
xi
2
1)
xi
2
1)
(xi
xi
2
1)
EVM este
bn = bn ; bn ; c2 n
bn (x0 ; x1 ; :::; xn ) =
n
X
xi
i=1
!2
n
X
xi
i=1
n
X
x2i
n
X
i=1
n
bn (x0 ; x1 ; :::; xn ) =
n
X
xi
xi
n
X
xi
xi
i=1
n
X
1 xi
i=1
i=1
n
X
x2i
i=1
i=1
n
X
xi
bn
n
X
bn xi
xi
1 xi
i=1
!2
n
X
i=1
X
c2 n (x0 ; x1 ; :::; xn ) = 1
xi
n i=1
xi
i=1
!2
2
1
Teorema (Billingsley)
Notam estimatorul de verosimilitate maxima cu
bn (X0 ; :::; Xn ) = bn (X0 ; :::; Xn ) ; bn (X0 ; :::; Xn ) ; c2 n (X0 ; :::; Xn )
p
Vectorul aleator n bn (X0 ; :::; Xn )
1 la un vector aleator repartizat normal,
N 3 : 0; I1 1 ( ) ;
unde I1 ( ) este matricea informationala Fisher asociata lantului Markov cu
densitatea de trecere f (x; y; ) ;
I1 ( ) = kiuv ( )ku;v=1;2;3
i11 ( ) = M
@ log f (X0 ; X1 ; )
@
i12 ( ) = i21 ( ) = M
i13 ( ) = i31 ( ) = M
i22 ( ) = M
@ log f (X0 ; X1 ; )
@
33
i23 ( ) = i32 ( ) = M
I1 ( ) =
@ log f (X0 ; X1 ; )
@ 2
i33 ( ) = M
1
2
+
0
1
4
1
C
A
Exemplicare a calculului:
1
log (2 )
2
log f (x0 ; x1 ; ) =
1
log
2
@ log f (x0 ; x1 ; )
1
= 2 (x1
@
@ log f (x0 ; x1 ; )
x0
= 2 (x1
@
1
4
X0
x0 )
x0 )
X0 )
X0 ) j X 0
X0
2
1
2
34
(X1
2
X0 (X1
=
X0 )
(X1
(X1
M
M
=M
1
i12 ( ) = M
(X1
=M
x0 )
i11 ( ) = M
(x1
X0 ) j X 0
=
1
2
X0 )
=
1;
2;
1 Xt 1
2 Xt 2
< 1;
t;
t2N
< 1; 1 <
< 1;
2 (0; 1)
0
0
2;
(0)
(1)
(1)
(0)
1
2
n
Y
i=2
log (2 )
1
2
log
exp
(xi
n
1 X
2
2
2 xi 2 )
1 xi 1
(xi
1 xi 1
2
2 xi 2 )
i=2
@ log L
@ log L
@ log L
=
=
=0
@ 1
@ 2
@ 2
se scrie sub forma
8
>
>
>
>
>
>
>
>
<
EVM este
>
>
>
>
>
>
>
>
:
n
X
i=2
n
X
x2i
1+
xi
i=2
xi
1 xi 2
i=2
2
n
X
1
n 1
n
X
1 xi
n
X
x2i
=
=
i=2
(xi
n
X
i=2
n
X
xi
1 xi
xi
2 xi
i=2
1 xi 1
2
2 xi 2 )
i=2
Pn
Pn
Pn
Pn
2
(
x
x
)
x
(
x
x
)
(
i
i
1
i
i
2
i
2
i=2
i=2
i=2
i=2 xi
c1 (x0 ; :::; xn ) =
Pn
P
P
2
n
n
2
2
( i=2 xi 1 xi 2 )
i=2 xi 1
i=2 xi 2
Pn
Pn
Pn
Pn
2
xi 2 )
( i=2 xi xi 1 ) ( i=2 xi
i=2 xi 1
c2 (x0 ; :::; xn ) = ( i=2 xiP
Pn
Pn
2
n
2
2
( i=2 xi 1 xi 2 )
i=2 xi 1
i=2 xi 2
c2 (x0 ; :::; xn ) =
n
X
1 i=2
35
xi
c1 xi
c2 xi
1 xi 2 )
1 xi 2 )