Professional Documents
Culture Documents
SINAPE General
SINAPE General
Don Percival
1
Overview of Talk
• review of wavelets
– wavelet filters
– wavelet coefficients and their interpretation
• wavelet variance decomposition of sample variance
• theoretical wavelet variance for stochastic processes
– stationary processes
– nonstationary processes with stationary differences
• sampling theory for Gaussian processes with an example
• sampling theory for non-Gaussian processes with an example
• use on time series with time-varying statistical properties
• summary
2
Wavelet Filters & Coefficients: I
3
Wavelet Filters & Coefficients: II
where
τj −1
1
X t(τj ) ≡ Xt−l
τj
l=0
4
Haar LA(8)
Figure 1: Haar and LA(8) wavelet filters {h̃j,l } for scales indexed by j = 1, 2, . . . , 7.
5
Wavelet Filters & Coefficients: III
6
Empirical Wavelet Variance
j for levels j = 1, 2, . . . , J0
j,t into W
• collect W
J of scaling coefficients:
• also compute vector V 0
LJ0 −1
V J0,t ≡ g̃J0,l Xt−l , t = 0, 1, . . . , N − 1;
l=0
σ̂X ≡
2
Xt − X = j
2 +
V
W J
2 − X 2
0
N t=0 N j=1
J
2/N = X 2).
(if N = 2J0 , can argue that
V 0
1 2
• N
W j
2
portion of σ̂X due to changes in averages over scale τj ;
i.e., ‘scale by scale’ analysis of variance
• cf. ‘frequency by frequency’ analysis of variance:
N −1
1 2 1
2
σ̂X =
F
2 − X with Fk ≡ √ Xte−i2πtk/N
N N t=0
7
Haar LA(8)
Figure 2: Haar and LA(8) scaling filters {g̃J0 ,l } for scales indexed by J0 = 1, 2, . . . , 7.
8
75
50
25
0
-25
-50
0 48 96 144 192
t
80 O
40 O
O O
O O O
0 O
1 2 4 8 16 32 64 128
scale
6000
3000
0
0.000 0.125 0.250 0.375 0.500
f
Figure 3: Time series of subtidal sea levels (top plot), along with associated empirical wavelet
j
2 /N versus scales τj = 2j−1 for j = 1, . . . , 8 (middle) and periodogram versus
variances
W
frequency (bottom).
9
Theoretical Wavelet Variance
10
Rationale for Wavelet Variance
11
Variance Decomposition
12
SDF Substitute/Complement: I
13
SDF Substitute/Complement: II
• νX
2
(τj ) easier to estimate than SDF
• basic estimator of SDF is periodogram: given X0, . . . , XN −1,
2
1
N −1
(p) −i2πfk t k
ŜX (fk ) ≡ (Xt − X)e , fk ≡
N t=0 N
(p)
– inconsistent because var{ŜX (fk )} ≈ SX
2
(fk )
(i.e., does not decrease to 0 as N → ∞)
– need smoothers etc. to get consistency
– can be badly biased
• basic estimator of νX
2
(τj ) is
N −1
1 2
2
ν̃X (τj ) ≡ W
N t=0 j,t
14
Substitute for Variance: I
15
Substitute for Variance: II
2
for chosen N ; i.e., σ̂X badly biased even for very large N
16
Substitute for Variance: III
17
3
−3
0 500 1000
t
Figure 4: Realization of a fractional Gaussian noise (FGN) process with Hurst coefficient
H = 0.9. The sample mean of approximately 0.53 and the true mean of zero are indicated
by the thin horizontal lines (taken from Figure 300, Percival and Walden, 2000, copyright
Cambridge University Press).
18
Substitute for Variance: IV
19
Generalization to Nonstationary Processes
• if L is properly chosen, νX
2
(τj ) well-defined for processes with
stationary backward differences
• let B be such that BXt ≡ Xt−1 ⇒ B k Xt = Xt−k
• Xt has dth order stationary backward differences if
d
d
Yt ≡ (1 − B)dXt = (−1)k Xt−k
k
k=0
20
Xt (1 − B)Xt (1 − B)2 Xt
0 (a)
0 (b)
0 (c)
0 (d)
21
Wavelet Variance for Processes with
Stationary Backward Differences: I
• proof: {h̃j,l } ⇔ L
2 first differences & then {aj,l } so
22
Wavelet Variance for Processes with
Stationary Backward Differences: II
23
Unbiased Estimator of Wavelet Variance
• recall that
Lj −1
W j,t ≡ h̃j,l Xt−l , t ∈ Z,
l=0
where Mj ≡ N − Lj + 1
24
2
Statistical Properties of ν̂X (τj ) (Gaussian)
25
Estimation of Aj
26
2
Confidence Intervals for νX (τj ): I
27
Three Ways to Set η
28
6 X
t
−6
0.25 (1)
Xt
0.0
−0.25
450 600 750 900
depth (meters)
(1)
Figure 6: Vertical shear measurements and associated backward differences {Xt } (taken
from Figure 328, Percival and Walden, 2000, copyright Cambridge University Press).
29
101 η̂1 η2 η3
100 o o o
o o o
o o o
10−1
o o o
10−2 o o o
o o o
10−3 o o o
o o o
o o o
−4
10
10−1100 101 102 10−1100101 102 10−1100101 102
τj ∆t (meters)
Figure 7: 95% confidence intervals for the D(6) wavelet variance for the vertical ocean
shear series. The intervals are based upon χ2 approximations to the distribution of the
unbiased wavelet variance estimator with EDOFs determined by, from left to right, η̂1 , η2
using a nominal model for SX (·) and η3 (taken from Figure 333, Percival and Walden, 2000,
copyright Cambridge University Press).
30
2
Statistical Properties of ν̂X (τj ) (Non-Gaussian)
and
∗ Mnm(W j ) is σ-algebra for W j,m, . . . , W j,n
∗ χ > (2 + δ)/δ
2
• let Zj,t ≡ W j,t have SDF SZj (·) such that 0 < SZj (0) < ∞
• ν̂X
2 2
(τj ) asymptotically normal with mean νX (τj ) & large sample
variance SZj (0)/Mj
• can estimate SZj (0) using standard SDF estimators
– multitaper SDF estimator with K = 5 tapers
– autoregressive SDF estimator (moderate order p)
• Fig. 8: surface albedo of spring pack ice in Beaufort Sea
31
0.8
0.6
0.4
(a)
0.2
0.0
0 3000 6000 9000
observation index
6
(b)
4
0
8
(c)
6
2
0
101 102 103 104
scale (meters)
Figure 8: (a) surface albedo of pack ice in the Beaufort Sea (sampling interval between mea-
surements is 25 meters, and there are 8428 measurements in all); (b) estimated LA(8) wavelet
variance (thick solid curve), along with upper and lower 90% confidence intervals based upon
Gaussian (thin dotted curves) and non-Gaussian theory (tbin solid curve); (c) ratio of esti-
mated non-Gaussian large sample standard deviations to estimated Gaussian large sample
standard deviations (adapted from Figure 3, A. Serroukh, A. T. Walden and D. B. Percival,
‘Statistical Properties and Uses of the Wavelet Variance Estimator for the Scale Analysis of
Time Series,’ Journal of the American Statistical Association, 95, pp. 184–96).
32
Wavelet Variance Analysis of Time Series
with Time-Varying Statistical Properties
33
5
9
600 1300
year
100
o
o x
10-1 o
o
10-2
1 2 4 8
scale (years)
Figure 9: Nile River yearly minima (top plot), along with estimated Haar wavelet variance
before and after year 715.5 (x’s and o’s, respectively) and 95% confidence intervals (thin and
thick lines, respectively) based upon a chi-square approximation with EDOFs determined
by η3 (taken from Figures 192 and 327, Percival and Walden, 2000, copyright Cambridge
University Press).
34
S 7
7
D
6
D
5
D
4
D
3
D
2
D
1
D
80
60
40
20
0 X
−20
−40
1980 1984 1988 1991
years
Figure 10: LA(8) MODWT multiresolution analysis for Crescent City subtidal variations
measured in centimeters (taken from Figure 186, Percival and Walden, 2000, copyright Cam-
bridge University Press).
35
100.0
10.0
1.0
o
0.1
1980 1984 1988 1991
years
Figure 11: Estimated time-dependent LA(8) wavelet variances for physical scale τ2 ∆t =
1 day for the Crescent City subtidal sea level variations, along with a representative 95%
confidence interval based upon a hypothetical wavelet variance estimate of 1/2 and a chi-
square distribution with ν = 15.25 (taken from Figure 324, Percival and Walden, 2000,
copyright Cambridge University Press).
36
100 1 day 2 days
o
o o
o
o o o o
10 o o o
o o
o
o o
o
o o
o o
o
o o
1
100 4 days 8 odays
o o o o
o o o
o o
o o
o
o
o
10 o o o
o o o
o o o
1
100 16 days 32 days
o o
o o o
o
o o o
o o o
10 o o
o
o
o
o o o
o o
o o
1
J F MAMJ J A S O ND J F MAMJ J A S O ND
month month
Figure 12: Estimated LA(8) wavelet variances for physical scales τj ∆t = 2j−2 days,
j = 2, . . . , 7, grouped by calendar month for the subtidal sea level variations (taken from
Figure 326, Percival and Walden, 2000, copyright Cambridge University Press).
37
Summary
38