Bootstrap Techniques for Signal Processing
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
1
Abdelhak M. Zoubir ¨ Darmstadt Technische Universitat Fachgebiet Signalverarbeitung Institut fur ¨ Nachrichtentechnik Merckstr. 25, 64283 Darmstadt E-Mail:
[email protected] URL: www.spg.tu-darmstadt.de
c A.M. Zoubir
ATiSSP Sommersemester 2007
1
Bootstrap Techniques for Signal Processing
Signal Processing Group
Outline Signal Processing Group
2
➠ Introduction Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Non-Parametric Bootstrap ➠ Examples ➠ Bias Estimation ➠ Variance Estimation ➠ Confidence Interval Estimation for the Mean ➠ The Parametric Bootstrap ➠ An Example: Confidence Interval Estimation for the Mean ➠ An Example of Bootstrap Failure c A.M. Zoubir
ATiSSP Sommersemester 2007
2
Bootstrap Techniques for Signal Processing
Outline (Cont’d) Signal Processing Group
3
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Dependent Data Bootstrap ➠ An Example: Variance Estimation for AR Parameter Estimates ➠ The Principle of Pivoting ➠ An Example: Confidence Interval for the Mean ➠ Variance Stabilisation ➠ Examples ➠ Correlation Coefficient c A.M. Zoubir
ATiSSP Sommersemester 2007
3
Bootstrap Techniques for Signal Processing
Outline (Cont’d) Signal Processing Group
4
➠ Hypothesis Testing with the Bootstrap
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ An Example: Testing the Frequency Response for Zero ➠ Signal Detection ➠ Bootstrap Matched-Filter ➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
4
Bootstrap Techniques for Signal Processing
Introduction Signal Processing Group
5
➠ Most techniques for computing variances of parameter estimators or for setInstitut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
ting confidence intervals assume a large sample size [Bhattacharya & Rao (1976)].
➠ In many signal processing problems large sample methods are inapplicable [Fisher & Hall (1991)].
➠ The bootstrap was introduced [Efron (1979)] to calculate confidence intervals for parameters when few data are available.
➠ The bootstrap has been shown to solve many other problems which would be too complicated for traditional statistical analysis [Hall (1992), Efron & Tibshirani (1993), Shao & Tu (1995)]. c A.M. Zoubir
ATiSSP Sommersemester 2007
5
Bootstrap Techniques for Signal Processing
Introduction (Cont’d) Signal Processing Group
6
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ The bootstrap does with a computer what the experimenter would do in practice, if it were possible.
➠ The observations are randomly re-assigned, and the estimates re-computed. ➠ These assignments and re-computations are done many times and treated as repeated experiments.
➠ In an era of exponentially increasing computational power, computer-intensive methods such as the bootstrap are affordable.
c A.M. Zoubir
ATiSSP Sommersemester 2007
6
Bootstrap Techniques for Signal Processing
Introduction (Cont’d) Signal Processing Group
7
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Applications of the bootstrap to real-life problems have been reported in (see also special session at ICASSP-94 [64])
➠ radar signal processing [Nagaoka & Amai (1990,1991)], ➠ sonar signal processing [Krolik et al. (1991), Reid et al. (1996)], ➠ geophysics [Fisher & Hall (1989,1990,1991), Tauxe et al. (1991)], ➠ biomedical engineering [Haynor & Woods (1989), Banga & Ghorbel (1993)], ➠ control [Lai & Chen (1995)],
c A.M. Zoubir
ATiSSP Sommersemester 2007
7
Bootstrap Techniques for Signal Processing
Introduction (Cont’d) Signal Processing Group
8
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ atmospheric environmental research [Hanna (1989)], ¨ ➠ vibration analysis [Zoubir & Bohme (1991,1995)].
➠ power systems [Herman (1996)], ➠ computer vision [Lange et al. (1998), Kanatani & Ohta (1998)], ➠ image analysis [Archer & Chan (1996)], ➠ nuclear technology [Yacout et al. (1996)], ➠ metrology [Ciarlini (1997), Cox et al. (1997)], ➠ financial engineering [Bhar & Chiarella (1996), Ankenbrand & Tomassini (1996)]. c A.M. Zoubir
ATiSSP Sommersemester 2007
8
Bootstrap Techniques for Signal Processing
Introduction (Cont’d) Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
9
➠ Bootstrap methods are potentially superior to large sample techniques for small sample sizes [Hall (1992)].
➠ A danger exists when applying bootstrap techniques in some circumstances where standard approaches are judged inappropriate and in such circumstances the bootstrap may also fail [Freedman (1981)].
➠ Special care is therefore required when applying the bootstrap in real-life situations [Young (1994)].
c A.M. Zoubir
ATiSSP Sommersemester 2007
9
Bootstrap Techniques for Signal Processing
Problem Statement Signal Processing Group
10
X = {X1 , X2 , . . . , Xn } be a random sample from a completely unspecified distribution F . Let θ denote an unknown characteristic of F , such as its
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Let
mean or variance.
Find the distribution of θˆ, an estimator of sample X . Possible solution:
θ, derived from the
repeat the experiment a sufficient number of times and
approximate the distribution of θˆ by the so obtained empirical distribution. Problem: may be inapplicable for cost reasons or because the experimental conditions are not reproducible.
c A.M. Zoubir
ATiSSP Sommersemester 2007
10
Bootstrap Techniques for Signal Processing
Principle of the Bootstrap Signal Processing Group
11
BOOTSTRAP WORLD
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
REAL WORLD Unknown Probability Model
F
Estimated Probability Model
Observed Data
^
X=(x1 , x2 , ..., xn )
F
θ=s(x ) Statistic of interest ^
Bootstrap Sample
x*=(x*1 , x*2 , ..., x*n )
θ*=s(x *) Bootstrap replication ^
Basic idea: simulate the probability mechanism of the real world by substituting the unknowns with estimates derived from the data. c A.M. Zoubir
ATiSSP Sommersemester 2007
11
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
12
➠ Introduction Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Non-Parametric Bootstrap ➠ Examples ➠ Bias Estimation ➠ Variance Estimation ➠ Confidence Interval Estimation for the Mean ➠ The Parametric Bootstrap ➠ An Example: Confidence Interval Estimation for the Mean ➠ An Example of Bootstrap Failure c A.M. Zoubir
ATiSSP Sommersemester 2007
12
Bootstrap Techniques for Signal Processing
The Non-Parametric Bootstrap Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
13
1. Conduct the experiment to obtain the random sample
{X1 , X2 , . . . , Xn } and calculate the estimate θˆ from X .
X
=
ˆ , which puts equal mass 1/n at each 2. Construct the empirical distribution F observation X1
= x1 , X2 = x2 , . . . , Xn = xn .
3. From the selected
Fˆ , draw a sample X ∗ = {X1∗ , X2∗ , . . . , Xn∗ }, called
the bootstrap (re)sample.
4. Approximate the distribution of θˆ by the distribution of θˆ∗ derived from
c A.M. Zoubir
ATiSSP Sommersemester 2007
13
Bootstrap Techniques for Signal Processing
X ∗.
The Bootstrap Procedure Signal Processing Group
14 Measured Data
-
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Conduct Experiment
-
Compute Statistic
x
- θ(x) ˆ
? Resample
Bootstrap Data
-
x∗ 1 x∗ 2
-
c A.M. Zoubir
ATiSSP Sommersemester 2007
θ1 ˆ∗
Statistic
-
Compute
ˆ ˆ ˆ (θ) F Θ θ2 ˆ∗
Statistic
. . .
-
Compute
ˆ ˆ ˆ (θ) F Θ
. . .
x∗ N
14
-
Compute
Generate EDF,
-
6 - θˆ
-
∗ θˆN
Statistic
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
15
➠ Introduction Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Non-Parametric Bootstrap ➠ Examples ➠ Bias Estimation ➠ Variance Estimation ➠ Confidence Interval Estimation for the Mean ➠ The Parametric Bootstrap ➠ An Example: Confidence Interval Estimation for the Mean ➠ An Example of Bootstrap Failure c A.M. Zoubir
ATiSSP Sommersemester 2007
15
Bootstrap Techniques for Signal Processing
Example: Bias Estimation Signal Processing Group
16
Consider the problem of estimating the variance of an unknown distribution Fµ,σ , Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
based on the random sample
ˆb2 . exist, σ ˆu2 and σ
X = {X1 , . . . , Xn }. Two different estimators
It can be easily shown that
Eσ ˆ u2
=σ
2
and
Eσ ˆ b2
= (1 −
1 2 )σ n
With the bootstrap we estimate the bias b(ˆ σ2)
= Eσ ˆ 2 −σ 2 by E∗ σ ˆ ∗2 −ˆ σ 2 , where σ ˆ 2 is the maximum likelihood estimate of σ 2 , i.e., σ ˆb2 and E∗ is expectation w.r.t. bootstrap sampling. c A.M. Zoubir
ATiSSP Sommersemester 2007
16
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
17
Step 0. Experiment. Collect the data into X Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
ˆb2 . estimates σ ˆu2 and σ
= {X1 , . . . , Xn }. Compute the
Step 1. Resampling. Draw a sample of size n, with replacement, from X . Step 2. Calculation of the bootstrap estimate. Calculate the bootstrap esti-
ˆb2 were computed ˆu2 and σ ˆb∗2 from X ∗ in the same way σ σ ˆu∗2 and σ but with the resample X ∗ .
mates
Step 3. Repetition. Repeat Steps 1 and 2 to obtain a total of
N bootstrap
∗2 ∗2 ∗2 ∗2 ,...,σ ˆu,N . estimates σ ˆu,1 ,...,σ ˆb,N and σ ˆb,1
Step 4. and c A.M. Zoubir
σu∗2 ) Bias Estimation. Estimate b(ˆ σu2 ) by b∗ (ˆ PN ∗2 ∗2 2 ˆb2 . ˆb,i − σ σb ) = 1/N i=1 σ b(ˆ σb ) by b∗ (ˆ
ATiSSP Sommersemester 2007
17
= 1/N
PN
Bootstrap Techniques for Signal Processing
2 ∗2 − σ ˆ σ ˆ b i=1 u,i
Example (Cont’d) Signal Processing Group
18
We considered a sample of size
n = 5 from the standard normal distribution.
N = 999 and 1000 Monte Carlo simulations, we obtained the following σb∗2 ). histograms for b∗ (ˆ σu∗2 ) and b∗ (ˆ 450
200
400
180
160
350
140 300
Frequency of occurence
Frequency of occurence
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
With
250
200
120
100
80
150 60 100
40
50
0 −0.15
20
−0.1
−0.05 0 Bootstrap bias estimate
0.05
0.1
0 −0.9
−0.8
−0.7
−0.6
−0.5 −0.4 −0.3 Bootstrap bias estimate
−0.2
−0.1
0
Bootstrap estimation of b(ˆ σu2 ),
Bootstrap estimation of b(ˆ σb2 ),
sample mean: -4.9400e-04 (=0 !).
sample mean: -0.1573 (=-0.2 !).
c A.M. Zoubir
ATiSSP Sommersemester 2007
18
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
19
➠ Introduction Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Non-Parametric Bootstrap ➠ Examples ➠ Bias Estimation ➠ Variance Estimation ➠ Confidence Interval Estimation for the Mean ➠ The Parametric Bootstrap ➠ An Example: Confidence Interval Estimation for the Mean ➠ An Example of Bootstrap Failure c A.M. Zoubir
ATiSSP Sommersemester 2007
19
Bootstrap Techniques for Signal Processing
Example: Variance Estimation Signal Processing Group
20
Consider the problem of finding the variance σ 2ˆ of an estimator θˆ of θ , based on Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
θ
the random sample X
= {X1 , . . . , Xn } from the unknown distribution Fθ .
➠ If tractable, one may derive an analytic expression for σθ2ˆ. ➠ Alternatively, one may use asymptotic arguments to compute an estimate σ ˆθ2ˆ for σ 2ˆ. θ Problem: In many situations the conditions for the above are not fulfilled. Solution: The bootstrap provides a simple and accurate alternative to approx-
ˆ ∗2 imate σ 2ˆ by σ ˆ . θ
c A.M. Zoubir
θ
ATiSSP Sommersemester 2007
20
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
21
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 0. Experiment. Conduct the experiment and collect the random data into the sample X
= {X1 , . . . , Xn }.
Step 1. Resampling. Draw a random sample of size n, with replacement, from
X. Step 2. Calculation of the bootstrap estimate. Evaluate the bootstrap estimate
θˆ∗ from X ∗ calculated in the same manner as θˆ but with the resample X ∗ replacing X . Step 3. Repetition. Repeat Steps 1 and 2 many times to obtain a total of ∗ bootstrap estimates θˆ1∗ , . . . , θˆB . Typical values for 200 [Efron & Tibshirani (1993)]. c A.M. Zoubir
ATiSSP Sommersemester 2007
21
B
B are between 25 and
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
22
Step 4. Estimation of the variance of θˆ. Estimate the variance σ 2ˆ of θˆ by θ
σˆ
2
BOOT
2 P P B B ˆ∗ 1 1 ∗ ˆ = B−1 b=1 θb − B b=1 θb .
➠ Suppose Fµ,σ is N (10, 25) and we wish to estimate σµˆ based on a random sample X of size n = 50. ➠ Following the above procedure with B = 25, a bootstrap estimate of the 2 = 0.49 as compared to the true σµ2ˆ = 0.5. variance of µ ˆ is found to be σ ˆBOOT c A.M. Zoubir
ATiSSP Sommersemester 2007
22
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
replacements 23
X = {X1 , . . . , Xn }
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Observations
Resamples
X1∗
X2∗
XB∗
Bootstrap statistics
θˆ1∗
θˆ2∗
∗ θˆB
σ ˆθ2∗ ˆ∗
Variance estimation
c A.M. Zoubir
ATiSSP Sommersemester 2007
23
P B 1 ˆ∗ − θ = B−1 b b=1
1 B
PB ˆ∗ 2 b=1 θb
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
24 140
100 Frequency of occurence
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
120
80
60
40
20
0 0.44
∗2(1)
Histogram of σ ˆµˆ and B
c A.M. Zoubir
0.46
∗2(2)
,σ ˆµˆ
0.48
0.5
0.52 0.54 0.56 Bootstrap variance estimate
∗2(1000)
,...,σ ˆµˆ
0.58
0.6
0.62
, based on a random sample of size n
= 25. ATiSSP Sommersemester 2007
24
Bootstrap Techniques for Signal Processing
= 50
Outline Signal Processing Group
25
➠ Introduction Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Non-Parametric Bootstrap ➠ Examples ➠ Bias Estimation ➠ Variance Estimation ➠ Confidence Interval Estimation for the Mean ➠ The Parametric Bootstrap ➠ Confidence Interval Estimation for the Mean ➠ An Example of Bootstrap Failure c A.M. Zoubir
ATiSSP Sommersemester 2007
25
Bootstrap Techniques for Signal Processing
Example: Confidence Interval for the Mean
Signal Processing Group
26
X = {X1 , . . . , Xn } be from some unknown distribution Fµ,σ . We wish to find an estimator and a 100(1 − α)% interval for µ. Let
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Let
µˆ = A confidence interval for
X1 +...+Xn . n
µ is found by determining the distribution of µ ˆ, and
finding µ ˆL , µ ˆU such that
Pr(ˆ µL
≤ µ ≤ µˆ U ) = 1 − α.
The distribution of µ ˆ depends on the distribution of the Xi ’s, which is unknown. If
n is large, the distribution of µ ˆ could be approximated by the normal distribution as per the central limit theorem. What if n is small? c A.M. Zoubir
ATiSSP Sommersemester 2007
26
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
27
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 0. Experiment.
Conduct
the
experiment
and
collect
X1 , . . . , Xn into X . Suppose Fµ,σ is N (10, 25) and X = {−2.41, 4.86, 6.06, 9.11, 10.20, 12.81, 13.17, 14.10, 15.77, 15.79} is of size n = 10. The mean of all values in X is µ ˆ = 9.95. Step 1. Resampling.
Draw a sample of 10 values,
with replace-
X. One might obtain the bootstrap resample X ∗ = {9.11, 9.11, 6.06, 13.17, 10.20, −2.41, 4.86, 12.81, −2.41, 4.86}.
ment, from
Note that some values from the original sample appear more than once while others do not appear at all. Step 2. Calculation of the bootstrap estimate. Calculate the mean of X ∗ . The mean of all 10 values in X ∗ is µ ˆ∗1
c A.M. Zoubir
ATiSSP Sommersemester 2007
27
= 6.54.
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
28
Step 3. Repetition. Repeat Steps 1 and 2 to obtain
N bootstrap estimates
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
ˆ∗N . Let N = 1000. µ ˆ∗1 , . . . , µ µ ˆ. Sort the bootstrap estiˆ∗(N ) . We might get 3.48, ˆ∗(2) ≤ . . . ≤ µ mates to obtain µ ˆ∗(1) ≤ µ 3.39, 4.46, . . . , 8.86, 8.88, 8.89, . . . , 10.07, 10.08, . . . , 14.46, 14.53, 14.66.
Step 4. Approximation of the Distribution of
− α)% bootstrap confidence interˆ∗(q2 ) ), where q1 = ⌊N α/2⌋ and q2 = N − q1 + 1. For val is (ˆ µ∗(q1 ) , µ α = 0.05 and N = 1000, q1 = 25 and q2 = 976, and the 95% confidence interval is found to be (6.27, 13.19) as compared to the theoretical (6.85, 13.05).
Step 5. Confidence Interval. The 100(1
c A.M. Zoubir
ATiSSP Sommersemester 2007
28
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
29 0.3
0.2 Density function
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
0.25
0.15
0.1
0.05
0 4
Histogram of
6
8
10 Bootstrap estimates
12
14
16
ˆ∗1000 , based on the random sample X = {−2.41, 4.86, ˆ∗2 , . . . , µ µ ˆ∗1 , µ
6.06, 9.11, 10.20, 12.81, 13.17, 14.10, 15.77, 15.79}, together with the density function of a Gaussian variable with mean 10 and variance 2.5. c A.M. Zoubir
ATiSSP Sommersemester 2007
29
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
30 1 0.9 0.8
Density function
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 −2
−1.5
−1
−0.5 0 0.5 Boostrap estimates
1
1.5
2
Histogram of 1000 bootstrap estimates of the mean of the t4 -distribution and the kernel probability density function obtained from 1000 Monte Carlo simulations. The 95 % confidence interval is (-0.896,0.902) and (-0.886,0.887) based on the bootstrap and Monte Carlo, respectively. c A.M. Zoubir
ATiSSP Sommersemester 2007
30
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
31
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ The procedure described above can be substantially improved because the interval calculated is, in fact, an interval with coverage less than the nominal value [Hall (1988)].
➠ Later, we shall discuss another way that will lead to a more accurate confidence interval for the mean.
➠ The computational expense to calculate the confidence interval for µ is approximately N times greater than the one needed to compute µ ˆ. ➠ This is acceptable given the ever-increasing capabilities of today’s computers.
c A.M. Zoubir
ATiSSP Sommersemester 2007
31
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
32
➠ Introduction Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Non-Parametric Bootstrap ➠ Examples ➠ Bias Estimation ➠ Variance Estimation ➠ Confidence Interval Estimation for the Mean ➠ The Parametric Bootstrap ➠ An Example: Confidence Interval Estimation for the Mean ➠ An Example of Bootstrap Failure c A.M. Zoubir
ATiSSP Sommersemester 2007
32
Bootstrap Techniques for Signal Processing
The Parametric Bootstrap Signal Processing Group
33
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Bootstrap sampling can be carried out parametrically. ➠ If one has partial knowledge of F , one may use Fˆθˆ instead of Fˆ . ➠ Draw N samples of size n from the parametric estimate of F
Fˆθˆ −→ (x∗1 , x∗2 , . . . , x∗n) and proceed as before.
➠ When used in a parametric way, the bootstrap provides more accurate answers, provided the model is correct.
c A.M. Zoubir
ATiSSP Sommersemester 2007
33
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
34
➠ Introduction Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Non-Parametric Bootstrap ➠ Examples ➠ Bias Estimation ➠ Variance Estimation ➠ Confidence Interval Estimation for the Mean ➠ The Parametric Bootstrap ➠ An Example: Confidence Interval Estimation for the Mean ➠ An Example of Bootstrap Failure c A.M. Zoubir
ATiSSP Sommersemester 2007
34
Bootstrap Techniques for Signal Processing
Example: Confidence Interval for the Mean
Signal Processing Group
35
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 0. Experiment.
Conduct
the
experiment
and
collect
X1 , . . . , Xn into X . Suppose Fµ,σ is N (10, 25) and X = {−2.41, 4.86, 6.06, 9.11, 10.20, 12.81, 13.17, 14.10, 15.77, 15.79} is of size n = 10. The mean of all values in X is µ ˆ = 9.95 and the sample variance is σ ˆ 2 = 33.15. Step 1. Resampling.
Draw
a
sample
of
10
values,
with
Fµˆ,ˆσ . We might obtain X ∗ = {7.45, 0.36, 10.67, 11.60, 3.34, 16.80, 16.79, 9.73, 11.83, 10.95}.
replacement,
from
Step 2. Calculation of the bootstrap estimate. Calculate the mean of X ∗ . The mean of all 10 values in X ∗ is µ ˆ∗1
c A.M. Zoubir
ATiSSP Sommersemester 2007
35
= 9.95.
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
36
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 3. Repetition. Repeat Steps 1 and 2 to obtain
N bootstrap estimates
ˆ∗N . Let N = 1000. µ ˆ∗1 , . . . , µ Step 4. Approximation of the Distribution of µ ˆ. Sort the bootstrap estimates to obtain µ ˆ∗(1)
ˆ∗(N ) . ≤µ ˆ∗(2) ≤ . . . ≤ µ
− α)% bootstrap confidence interˆ∗(q2 ) ), where q1 = ⌊N α/2⌋ and q2 = N − q1 + 1. For val is (ˆ µ∗(q1 ) , µ α = 0.05 and N = 1000, q1 = 25 and q2 = 976, and the 95% confidence interval is found to be (6.01, 13.87) as compared to the theoretical (6.85, 13.05).
Step 5. Confidence Interval. The 100(1
c A.M. Zoubir
ATiSSP Sommersemester 2007
36
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
37 0.3
0.2 Density function
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
0.25
0.15
0.1
0.05
0
Histogram of
4
6
8
10 Bootstrap estimates
12
14
16
ˆ∗1000 , based on the random sample X = {−2.41, 4.86, ˆ∗2 , . . . , µ µ ˆ∗1 , µ
6.06, 9.11, 10.20, 12.81, 13.17, 14.10, 15.77, 15.79}, together with the density function of a Gaussian variable with mean 10 and variance 2.5. c A.M. Zoubir
ATiSSP Sommersemester 2007
37
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
38
➠ Introduction Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Non-Parametric Bootstrap ➠ Examples ➠ Bias Estimation ➠ Variance Estimation ➠ Confidence Interval Estimation for the Mean ➠ The Parametric Bootstrap ➠ An Example: Confidence Interval Estimation for the Mean ➠ An Example of Bootstrap Failure c A.M. Zoubir
ATiSSP Sommersemester 2007
38
Bootstrap Techniques for Signal Processing
An Example of Bootstrap Failure Signal Processing Group
39
X ∼ U(0, θ) and X = {X1 , X2 , . . . , Xn }. We wish to estimate θ by θˆ ˆ . The Maximum Likelihood (ML) estimator of θ is given ˆˆ (θ) and its distribution F Θ by θˆ = X(n) .
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Let
➠ To obtain an estimate of the density function of θˆ we sample with replacement from the data and each time estimate θˆ∗ from X ∗ . ˆ and estimate θˆ∗ from X ∗ (para➠ Alternatively, we could sample from U(0, θ) metric bootstrap).
➠ We ran an example with θ = 1, n = 50 and N = 1000. The ML estimate of θ was found to be θˆ = 0.9843. c A.M. Zoubir
ATiSSP Sommersemester 2007
39
Bootstrap Techniques for Signal Processing
An Example of Bootstrap Failure (Cont’d)
Signal Processing Group
40
The non-parametric bootstrap shows that approximately 62% of the values of θˆ∗
n −→ ∞.
ˆ = 1 − (1 − 1/n)n −→ 1 − e−1 ≈ 0.632 as = θ)
700
120
600
100
Frequency of occurrence
500 Frequency of occurrence
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
equal θˆ. In fact, Pr(θˆ∗
400
300
80
60
40 200
20
100
0 0.65
0.7
0.75
0.8 0.85 theta hat star
0.9
0.95
Non-parametric Bootstrap c A.M. Zoubir
ATiSSP Sommersemester 2007
40
1
0 0.65
0.7
0.75
0.8 0.85 theta hat star
0.9
Parametric Bootstrap
Bootstrap Techniques for Signal Processing
0.95
1
Outline Signal Processing Group
41
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Dependent Data Bootstrap ➠ An Example: Variance Estimation for AR Parameter Estimates ➠ The Principle of Pivoting ➠ An Example: Confidence Interval for the Mean ➠ Variance Stabilisation ➠ Examples ➠ Correlation Coefficient c A.M. Zoubir
ATiSSP Sommersemester 2007
41
Bootstrap Techniques for Signal Processing
The Dependent Data Bootstrap Signal Processing Group
42
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ The assumption of i.i.d. data can break down in practice either because the data is not independent or because it is not identically distributed, or both.
➠ We can still invoke the bootstrap principle if we knew the model that generated the data [Efron & Tibshirani (1993), Bose (1988), Kreiss & Franke (1992), Paparoditis (1996), Zoubir (1993)].
➠ For example, a way to relax the i.i.d. assumption is to assume that the data is identically distributed but not independent such as in autoregressive (AR) models.
c A.M. Zoubir
ATiSSP Sommersemester 2007
42
Bootstrap Techniques for Signal Processing
The Dependent Data Bootstrap (Cont’d)
Signal Processing Group
43
➠ If no plausible model such as AR is available for the probability mechanism Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
generating stationary observations, we could make the assumption of weak dependence.
➠ Strong mixing processesa , for example, satisfy the weak dependence condition.
➠ The moving blocks bootstrap [Kunsch (1989), Liu & Singh (1992), Politis & ¨ Romano (1992,1994)] has been proposed for bootstrapping weakly dependent data. a
Loosely speaking a process is strong mixing if observations far apart (in time) are almost
independent [Rosenblatt (1985)].
c A.M. Zoubir
ATiSSP Sommersemester 2007
43
Bootstrap Techniques for Signal Processing
The Moving Blocks Bootstrap Signal Processing Group
44 Observed Data
Amplitude
5
5
−5
4
1
3
0
10
20
30
40 50 60 Resampled Blocks
70
80
90
100
Amplitude
5
1
2
3
4
5
0
−5
0
10
20
30
40
50 60 Bootstrap Data
70
80
90
100
0
10
20
30
40
50 60 Sample number
70
80
90
100
5 Amplitude
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
2
0
0
−5
The moving block bootstrap randomly selects blocks of the original data (top) and concatenate them together (centre) to form a resample (bottom). In this example, the block size is l c A.M. Zoubir
= 20 and n = 100. ATiSSP Sommersemester 2007
44
Bootstrap Techniques for Signal Processing
Other Block Bootstrap Methods Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
45
➠ The circular blocks bootstrap [Politis & Romano (1992), Shao & Yu (1993)] allows blocks which start at the end of the data and wrap around to the start.
➠ The blocks of blocks bootstrap [Politis et al. (1992)] uses two levels of blocking to estimate confidence bands for spectra and cross-spectra.
➠ The stationary bootstrap [Politis & Romano (1994)] allows blocks to be of random lengths instead of a fixed length.
c A.M. Zoubir
ATiSSP Sommersemester 2007
45
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
46
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Dependent Data Bootstrap ➠ An Example: Variance Estimation for AR Parameter Estimators ➠ The Principle of Pivoting ➠ An Example: Confidence Interval for the Mean ➠ Variance Stabilisation ➠ Examples ➠ Correlation Coefficient c A.M. Zoubir
ATiSSP Sommersemester 2007
46
Bootstrap Techniques for Signal Processing
An Example: Variance Estimation in AR Models 47
We generate n observations xt , t
Signal Processing Group
= 0, . . . , n − 1, from
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Xt + a · Xt−1 = Zt , Zt is white Gaussian noise with EZt = 0, cZZ (u) = σZ2 δ(u), and a such that |a| < 1. where
After de-trending the data, we fit the AR(1) model to the observation
xt . With
Pn−|u|−1
xt xt+|u| for 0 ≤ |u| ≤ n − 1, we calculate the ˆ = −ˆ cxx (1)/ˆ cxx (0), which has Maximum Likelihood Estimate (MLE) of a, a approximate variance‡ σ ˆaˆ2 = (1 − a2 )/n. cˆxx (u) = 1/n
t=0
σ ˆaˆ2 can be found in the nonGaussian case and is a function of a and the variance and kurtosis of Zt [Porat & Friedlander ‡
under some regularity conditions an asymptotic formula for
(1989)]. c A.M. Zoubir
ATiSSP Sommersemester 2007
47
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
48
n observations xt , t = 0, . . . , n − 1, from an auto-regressive process of order one, Xt .
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 0. Experiment.
Conduct the experiment and collect
Step 1. Calculation of the residuals. With the Maximum Likelihood Estimate a ˆ of a, define the residuals zˆt Step 2. Resampling.
= xt + a ˆ · xt−1 for t = 1, 2, . . . , n − 1.
Create a bootstrap sample
x∗0 , x∗1 , . . . , x∗n−1 by
∗ , with replacement, from the residuals zˆ1∗ , zˆ2∗ , . . . , zˆn−1 ax∗t−1 + zˆt∗ , t = zˆ1 , zˆ2 , . . . , zˆn−1 , then letting x∗0 = x0 , and x∗t = −ˆ 1, 2, . . . , n − 1.
sampling
c A.M. Zoubir
ATiSSP Sommersemester 2007
48
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
49
Step 3. Calculation of the bootstrap estimate.
After centring the time se-
ˆ∗ , using the above formulae but based on x∗0 , x∗1 , . . . , x∗n−1 , obtain a x∗0 , x∗1 , . . . , x∗n−1 .
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
ries
Step 4. Repetition. Repeat steps 2–3 a large number of times,
N = 1000,
ˆ∗N . ˆ∗2 , . . . , a say, to obtain a ˆ∗1 , a ˆ∗N , approximate the variance ˆ∗2 , . . . , a Step 5. Variance estimation. From a ˆ∗1 , a of a ˆ by
σˆ aˆ∗2 c A.M. Zoubir
=
ATiSSP Sommersemester 2007
1 N −1
49
PN
∗ (ˆ a i=1 i
−
1 N
PN
∗ 2 a ˆ i=1 i ) .
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
50 6
4 Density function
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
5
3
2
1
0 −0.8
−0.7
ˆ∗1000 for ˆ∗2 , . . . , a Histogram of a ˆ∗1 , a for a was a ˆ
−0.5 −0.4 Bootstrap estimates
−0.3
−0.2
−0.1
a = −0.6, n = 128 and Zt Gaussian. The MLE
= −0.6351 and σ ˆaˆ = 0.0707. The bootstrap estimate was σ ˆaˆ∗ = 0.0712
as compared to σ ˆaˆ c A.M. Zoubir
−0.6
= 0.0694 based on 1000 Monte Carlo simulations.
ATiSSP Sommersemester 2007
50
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
51
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Dependent Data Bootstrap ➠ An Example: Variance Estimation for AR Parameter Estimates ➠ The Principle of Pivoting ➠ An Example: Confidence Interval for the Mean ➠ Variance Stabilisation ➠ Examples ➠ Correlation Coefficient c A.M. Zoubir
ATiSSP Sommersemester 2007
51
Bootstrap Techniques for Signal Processing
The Principle of Pivoting Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
52
➠ A statistic T (X, θ) is called pivotal if it possesses a fixed probability distri´ (1967), Lehmann (1986)]. bution independent of θ [Cramer ➠ Bootstrap confidence intervals or tests have excellent properties even for relatively low fixed resample number [Hall (1992)].
➠ For example, one can show that the coverage error in confidence interval estimation with the bootstrap is Op (n−1 ) as compared to Op (n−1/2 ) when using the normal approximation.
➠ The accuracy claimed holds whenever the statistic is asymptotically pivotal [Hall & Titterington (1989), Hall (1992)].
c A.M. Zoubir
ATiSSP Sommersemester 2007
52
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
53
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Dependent Data Bootstrap ➠ An Example: Variance Estimation for AR Parameter Estimates ➠ The Principle of Pivoting ➠ An Example: Confidence Interval for the Mean ➠ Variance Stabilisation ➠ Examples ➠ Correlation Coefficient c A.M. Zoubir
ATiSSP Sommersemester 2007
53
Bootstrap Techniques for Signal Processing
An Example: Confidence Interval Estimation
Signal Processing Group
54
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
We consider the construction of a confidence interval for the mean. Let
X =
{X1 , . . . , Xn } be a random sample from some unknown FµX ,σX . We wish to find an estimator of µX with a 100(1 − α)% confidence interval. Let
2 µ ˆX and σ ˆX be the sample mean and the sample variance of X , respec-
tively. Alternatively to the previous example, we will base our method for finding a confidence interval for µX on the statistic
µˆ Y =
µ ˆX −µX σ ˆ
,
σ ˆ is the standard deviation of µ ˆX . The statistic has asymptotically for large n a distribution free of unknown parameters. where
c A.M. Zoubir
ATiSSP Sommersemester 2007
54
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
55
Step 0. Experiment. Conduct the experiment and collect the random data into
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
the sample X
= {X1 , X2 , . . . , Xn }.
Step 1. Parameter estimation. Based on deviation σ ˆ , using a nested bootstrap.
X , calculate µ ˆX and its standard
Step 2. Resampling. Draw a random sample, ment, from X .
X ∗ of n values, with replace-
Step 3. Calculation of the pivotal statistic. Calculate the mean of all values in
X ∗ and using a nested bootstrap, calculate σ ˆ ∗ . Then, form
µˆ ∗Y = c A.M. Zoubir
ATiSSP Sommersemester 2007
55
µX µ ˆ∗X −ˆ σ ˆ∗
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
56
Step 4. Repetition. Repeat Steps 2-3 many times to obtain a total of
N boot-
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
strap estimates µ ˆ∗Y,1 , . . . , µ ˆ∗Y,N . Step 5. Ranking. Sort the bootstrap estimates to obtain
... ≤ µ ˆ∗Y,(1000) .
ˆ∗Y,(2) ≤ µ ˆ∗Y,(1) ≤ µ
ˆ∗Y,(q2 ) ) is an interval containing (1 − Step 6. Confidence Interval. If (ˆ µ∗Y,(q1 ) , µ α)N of the means µ ˆ∗Y , where q1 = ⌊N α/2⌋ and q2 = N − q1 + 1, then
(ˆ µX −
∗ σˆ µˆ Y,(q2), µˆ X
−
∗ σˆ µˆ Y,(q1))
is a 100(1 − α)% confidence interval for µX . The interval is called percentile-t confidence interval [Efron (1987), Hall (1988)]. c A.M. Zoubir
ATiSSP Sommersemester 2007
56
Bootstrap Techniques for Signal Processing
Example (Cont’d) Signal Processing Group
57
X as before, we obtained the confidence interval (3.54, 13.94) as compared to (6.01, 13.87). Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
For the same random sample
➠ This interval is larger than the one obtained earlier and enforces the statement that the interval obtained there has coverage less than the nominal 95%.
➠ It also yields better results than an interval derived using the assumption that µ ˆY is N (0, 1)- or the (better) approximation that µ ˆY is tn−1 -distributed. ➠ The interval obtained here accounts for skewness in the underlying population or other errors [Hall (1988,1992), Efron & Tibshirani (1993), Zoubir (1993)]. c A.M. Zoubir
ATiSSP Sommersemester 2007
57
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
58
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Dependent Data Bootstrap ➠ An Example: Variance Estimation for AR Parameter Estimates ➠ The Principle of Pivoting ➠ An Example: Confidence Interval for the Mean ➠ Variance Stabilisation ➠ Examples ➠ Correlation Coefficient c A.M. Zoubir
ATiSSP Sommersemester 2007
58
Bootstrap Techniques for Signal Processing
Variance Stabilisation Signal Processing Group
59
To ensure pivoting, usually the statistic is “studentised”, i.e., we form
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Tˆ =
ˆ θ−θ σ ˆθˆ
The percentile-t method is particularly applicable to location statistics, such as the sample mean, sample median, etc. [Efron & Tibshirani (1993)]. However, for more general statistics, it may not be accurate. Problem:
Studentising results in confidence intervals with erratically varying
lengths and end points. Solution: Pivoting often does not hold unless an appropriate variance stabilising transformation is applied first. How do we “automatically” get a variance stabilising transformation? c A.M. Zoubir
ATiSSP Sommersemester 2007
59
Bootstrap Techniques for Signal Processing
Variance Stabilisation (Cont’d) Signal Processing Group
60
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 1. Estimation of the variance stabilising transformation. (a) Generate B1 bootstrap samples Xi∗ from X and for each calculate the value of the statistic θˆi∗ , i
= 1, . . . , B1 . For example B1 = 100.
B2 bootstrap samples from Xi∗ , i = 1, . . . , B1 , and calculate σ ˆi∗2 , a bootstrap estimate for the variance of θˆi∗ , i = 1, . . . , B1 . For example B2 = 25.
(b) Generate
(c) Estimate the variance function against
ζ(θ) by smoothing the values of σ ˆi∗2
θˆi∗ , using, for example, a fixed-span 50% “running lines”
smoother [Hastie & Tibshirani (1990)].
c A.M. Zoubir
ATiSSP Sommersemester 2007
60
Bootstrap Techniques for Signal Processing
Variance Stabilisation (Cont’d) Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
61
Step 1. Estimation of the variance stabilising transformation.
ˆ from (d) Estimate the variance stabilising transformation h(θ)
h(θ) =
Rθ
{ζ(s)}−1/2ds.
Step 2. Bootstrap quantile estimation. Generate
B3 bootstrap samples and
h(θˆi∗ ) for each sample i. Approximate the distribution of ˆ − h(θ) by that of h(θˆ∗ ) − h(θ) ˆ. h(θ)
compute θˆi∗ and
c A.M. Zoubir
ATiSSP Sommersemester 2007
61
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
62
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Principle of the Bootstrap ➠ The Dependent Data Bootstrap ➠ An Example: Variance Estimation for AR Parameter Estimates ➠ The Principle of Pivoting ➠ An Example: Confidence Interval for the Mean ➠ Variance Stabilisation ➠ Examples ➠ Correlation Coefficient
c A.M. Zoubir
ATiSSP Sommersemester 2007
62
Bootstrap Techniques for Signal Processing
Example: Correlation coefficient Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
63
θ = ̺ be the correlation coefficient of two unknown populations, and let ̺ˆ and σ ˆ 2 be estimates of ̺ and the variance of ̺ˆ, respectively, based on X = {X1 , . . . , Xn } and Y = {Y1 , . . . , Yn }.
Let
X ∗ and Y ∗ be resamples, drawn with replacement from X and Y , respectively, and let ̺ ˆ∗ and σ ˆ ∗2 be bootstrap versions of ̺ˆ and σ ˆ2.
Let
By repeated resampling from
00
- Tˆ (X)
- Tˆ ≷ λ
K H
Receiver Structure c A.M. Zoubir
ATiSSP Sommersemester 2007
91
Bootstrap Techniques for Signal Processing
Receiver
Signal Detection with the Matched Filter
Signal Processing Group
92
-
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
x
- ×
Ps
6
-
1 √ σ s′ s
- ˆ T
s
Tˆ =
c A.M. Zoubir
s√′ x σ s′ s
=
ATiSSP Sommersemester 2007
s′√ Ps x σ s′ s
92
∼N
θ
√
s′ s ′ ′ , 1 , P = ss /s s . s σ
Bootstrap Techniques for Signal Processing
Performance of the Matched Filter Signal Processing Group
93
−5
=1 0
=1 0 FA
P
=1 P
FA
−4
0
−3 FA
P
FA
=1 0
−2
P
FA
P
FA
P 0.99
Detection probability
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
=1 0
=1 0
−1
0.999
−6
0.9999
0.95 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.05
0.01
0
1
2
3
4
5
6
7
8
9
Voltage signal−to−noise ratio
c A.M. Zoubir
ATiSSP Sommersemester 2007
93
Bootstrap Techniques for Signal Processing
Signal Detection with the CFAR Matched Filter
Signal Processing Group
94
s
? -×
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
-P s
x
′
√
s′ s ·? ·
- I −P
s
-k · k2
-×
-
6
6
√
1/(n − 1)
T = c A.M. Zoubir
√ ′ P x σ 2 s′ s s √ ′ s x (I−Ps )x/(n−1)σ 2
ATiSSP Sommersemester 2007
94
∼ tn−1
√
-≷ λ
θ s′ s σ
Bootstrap Techniques for Signal Processing
Performance of the CFAR Matched Filter 95
0.9999
N=∞
N=21 N=11
0.99
Detection probability
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
0.999
N=6
N=3
0.95 0.9 0.8
N=2
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.05
0.01
0
1
2
3
4
5
6
7
8
9
Voltage signal−to−noise ratio
c A.M. Zoubir
ATiSSP Sommersemester 2007
95
Bootstrap Techniques for Signal Processing
Signal Processing Group
Limitations of the Matched Filter Signal Processing Group
96
➠ The matched filter and the CFAR matched filter are designed (and optimal) Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
for Gaussian interference.
➠ Although they show high probability of detection in the non-Gaussian case, they are unable to maintain the preset level of significance for small sample sizes.
➠ The matched filter fails in the case where the interference/noise is non-Gaussian and the data size is small.
➠ The goal is to develop techniques which require little in the way of modelling and assumptions.
c A.M. Zoubir
ATiSSP Sommersemester 2007
96
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
97
➠ Hypothesis Testing with the Bootstrap
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ An Example: Testing the Frequency Response for Zero ➠ Signal Detection ➠ Bootstrap Matched-Filter ➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
97
Bootstrap Techniques for Signal Processing
The Bootstrap in Signal Detection: Motivation
Signal Processing Group
98
➠ The Bootstrap can provide solutions for detecting signals in non-Gaussian Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
interference, e.g., high-resolution radar and radar at low grazing angles.
➠ The bootstrap is particularly attractive when only a limited number of samples is available for signal detection.
➠ The bootstrap is also attractive when detection is associated with estimation, e.g., estimation of the number of signals.
➠ The bootstrap has been successfully applied to signal detection in areas such as wireless communications, landmine detection, radar, sonar and biomedical engineering.
c A.M. Zoubir
ATiSSP Sommersemester 2007
98
Bootstrap Techniques for Signal Processing
The Bootstrap in Signal Detection: Principle 99
Measured Data
-
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Conduct Experiment
-
Compute
ˆ
T -
-
Statistic
x
? Resample
Bootstrap Data
-
x∗ 1 x∗ 2
-
c A.M. Zoubir
ATiSSP Sommersemester 2007
Tˆ1∗
-
Statistic
-
Compute
Tˆ2∗
-
Statistic
. . .
-
Compute
x∗ B
99
-
ˆ ∗ , . . . , Tˆ ∗ Sort T 1 B to get
∗ ∗ Tˆ(1) ≤ . . . ≤ Tˆ(B) Set λ
. . . Compute
Signal Processing Group
∗ , = Tˆ(q)
q = ⌈(1 − α)(B + 1)⌉ ∗ TˆB
-
Statistic
Bootstrap Techniques for Signal Processing
Decision
The Bootstrap Matched Filter Signal Processing Group
100
Step 0. Experiment.
Run the experiment and collect the data
xt , t =
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
0, . . . , n − 1.
ˆθˆ, and Step 1. Estimation. Compute the LSE θˆ of θ , σ
ˆ θ−θ ˆ T = σˆ ˆ θ
θ=0
.
Step 2. Resampling. Compute the residuals
ˆ t, w ˆt = xt − θs
t = 0, . . . , n − 1,
and after centering, resample the residuals, assumed to be i.i.d., to obtain
w ˆt∗ , t = 0, . . . , n − 1. c A.M. Zoubir
ATiSSP Sommersemester 2007
100
Bootstrap Techniques for Signal Processing
The Bootstrap Matched Filter (Cont’d)
Signal Processing Group
101
Step 3. Bootstrap Test Statistic. Compute new measurements
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
ˆ t + wˆ ∗ , x∗t = θs t
t = 0, . . . , n − 1 ,
the LSE θˆ∗ based on the resamples x∗t , t
Tˆ ∗ =
= 0, . . . , n − 1, and
θˆ∗ −θˆ . ∗ σ ˆ ˆ∗ θ
Step 4. Repetition. Repeat Steps 2 and 3 a large number of times to obtain
Tˆ1∗ , . . . , TˆN∗ . ∗ ∗ Step 5. Bootstrap Test. Sort Tˆ1∗ , . . . , TˆN to obtain Tˆ(1)
H if Tˆ c A.M. Zoubir
∗ , where q = ⌊(1 − α)(N + 1)⌋. ≥ Tˆ(q)
ATiSSP Sommersemester 2007
101
∗ ≤ · · · ≤ Tˆ(N ) . Reject
Bootstrap Techniques for Signal Processing
The Bootstrap Matched Filter (Cont’d)
Signal Processing Group
102 Measured Data Estimate
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Conduct Experiment
θ
x
ˆ θ
Compute Statistic
ˆ T
Decision
ˆ ˆ = x − θs w
Resample & Reconstruct Bootstrap Data
x∗ 1
Compute Statistic
ˆ∗ T 1
Compute Statistic
ˆ∗ T 2
ˆ∗ , . . . , Sort T 1
ˆ∗ T N
to get
x∗ 2
ˆ∗ ≤ . . . ≤ T ˆ∗ T (1) (N ) ˆ∗ set T
x∗ B
c A.M. Zoubir
ATiSSP Sommersemester 2007
102
Compute Statistic
ˆ∗ T B
, (q) q = ⌊(1 − α)(N + 1)⌋
Bootstrap Techniques for Signal Processing
Performance of the Bootstrap Matched Filter
Signal Processing Group
103
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ The bootstrap matched filter is consistent, ie, Pd → 1 as n → ∞. ➠ One can show [Ong & Zoubir (2000)] that −1
Pf = α + O p n −1/2 α+O n for the matched filter for |st | < ∞, t ∈ Z.
➠ If st = cos(ωt + φ), then
Pf = α + O p n
−3/2
(α + O (n−1 ) for the matched filter) c A.M. Zoubir
ATiSSP Sommersemester 2007
103
Bootstrap Techniques for Signal Processing
Simulation Results (Gaussian Case)
Signal Processing Group
104
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Let st
= sin(2πt/6), n = 10, N = 999 (25 for variance estimation), α = 5%,
and the number of independent runs be 5,000.
N (0, 1), SNR=7 dB Pˆf [%]
Pˆd [%]
Matched Filter (MF)
5
99
CFAR MF
5
98
Boot. (σ known)
4
99
Boot. (σ unknown)
5
98
Detector
c A.M. Zoubir
ATiSSP Sommersemester 2007
104
Bootstrap Techniques for Signal Processing
Simulation Results (Non-Gaussian Case)
Signal Processing Group
105
Let st
= sin(2πt/6), n = 10, N = 999 (25 for variance estimation), α = 5%,
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
and the number of independent runs be 5,000.
P4
Wt ∼ i=1 ai N (µi , σi2 ) with a = (0.5, 0.1, 0.2, 0.2), µ = (−0.1, 0.2, 0.5, 1), σ = (0.25, 0.4, 0.9, 1.6), and SNR = -6 dB Pˆf [%]
Pˆd [%]
Matched Filter (MF)
8
83
CFAR MF
7
77
Boot. (σ known)
5
77
Boot. (σ unknown)
7
79
Detector
c A.M. Zoubir
ATiSSP Sommersemester 2007
105
Bootstrap Techniques for Signal Processing
Simulation Results (Cont’d) Signal Processing Group
106 (a)
2
fa
P /α
3
0 10
20
30
40
50
60
70
80
90
100
60
70
80
90
100
50 60 Sample size
70
80
90
100
(b)
2
fa
P /α
3
1 0 10
20
30
40
50 (c)
3 2
fa
P /α
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
1
1 0 10
20
30
40
Ratio of actual vs. nominal false alarm rate for the detection of a DC in Gaussian (top), skewed (middle) and heavy tailed (bottom) noise. c A.M. Zoubir
ATiSSP Sommersemester 2007
106
Bootstrap Techniques for Signal Processing
Simulation Results (Cont’d) Signal Processing Group
107 (a)
2
fa
P /α
3
0 10
20
30
40
50
60
70
80
90
100
60
70
80
90
100
50 60 Sample size
70
80
90
100
(b)
2
fa
P /α
3
1 0 10
20
30
40
50 (c)
3 2
fa
P /α
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
1
1 0 10
20
30
40
Ratio of actual vs. nominal false alarm rate for the detection of a sinusoid in Gaussian (top), skewed (middle) and heavy tailed (bottom) noise. c A.M. Zoubir
ATiSSP Sommersemester 2007
107
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
108
➠ Hypothesis Testing with the Bootstrap
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ An Example: Testing the Frequency Response for Zero ➠ Signal Detection ➠ Bootstrap Matched-Filter ➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
108
Bootstrap Techniques for Signal Processing
Tolerance Interval Bootstrap Matched Filter 109
Signal Processing Group
➠ The thresholds calculated in the bootstrap detectors are random variables
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
conditional on the data.
➠ We would like to set β to specify our confidence that the set level is maintained. ∗ must fulfill ➠ The required threshold T(r) ∗ |H) ≤ α) ≥ β. Pr(Pr(Tˆ > T(r)
➠ Distribution free tolerance interval theory gives r as the smallest positive integer which satisfies the relation
FB (α; B − r + 1, r) ≥ β
where FB (x; λ1 , λ2 ) is the beta cdf with parameters λ1 , c A.M. Zoubir
ATiSSP Sommersemester 2007
109
λ2 .
Bootstrap Techniques for Signal Processing
Aside - Tolerance Interval Theory Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
110
➠ Problem : Given an iid sample X1 , . . . , Xn , find two order statistics, X(r) and X(n−s+1) such that [X(r) , X(n−s+1) ] will encompass at least 100(1 − α)% of the underlying distribution, F (x), with probability at least β , R X(n−s+1) dF (x) ≤ 1 ≥ β. Pr 1 − α ≤ X(r) ➠ From the theory of distribution free tolerance intervals we must find r and s such that
1 − FB (1 − α; n − r − s + 1, r + s) ≥ β, where FB (1 − α; λ1 , λ2 ) is the beta cdf with parameters λ1 and λ2 [Wilks
1941].
c A.M. Zoubir
ATiSSP Sommersemester 2007
110
Bootstrap Techniques for Signal Processing
Tolerance Interval Bootstrap Matched Filter
Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
111
➠ Proceed through steps 0 to 4 of previous bootstrap matched filter procedure. ➠ Step 5 is modified for tolerance intervals : Step 5. Bootstrap Test. Let the α be the level of significance, β , be the probability that this level is kept and B be the number of bootstrap resamples. ∗ ∗ T1∗ , . . . , TB∗ into increasing order to obtain T(1) . Re≤ · · · ≤ T(B) ject H if Tˆ ≥ Tˆ ∗ , where r is the smallest positive integer which satisfies
Sort
(r)
FB (α; B − r + 1, r) ≥ β .
c A.M. Zoubir
ATiSSP Sommersemester 2007
111
Bootstrap Techniques for Signal Processing
Simulations Signal Processing Group
112
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ A signal, s = sin(2πt/9), t = 0, . . . , 9, i.e., n = 10, in Gaussian and Laplace noise. B = 1000, SNR = 5dB, α = 0.05 and β = 0.95. ➠ Estimated β using bootstrap detectors with and without tolerance intervals. β (nominal 0.95)
c A.M. Zoubir
ATiSSP Sommersemester 2007
Gaussian
Laplace
BMF
0.5146
0.4003
BCMF
0.4684
0.5549
TBMF
0.9643
0.9398
TBCMF
0.9584
0.9734
112
Bootstrap Techniques for Signal Processing
Example - Real Data Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
113
➠ Detection of sinusoids in non-stationary, non-Gaussian noise, α = 0.05, β = 0.95, window length of 21 samples. BCMF PFA = 0.0562, TBCMF PFA = 0.0463. 5
1.5
x 10
1
0.5
0
−0.5 data, α=0.05 β=0.95 target present BMF detection TBMF detection
−1
−1.5 1000
c A.M. Zoubir
ATiSSP Sommersemester 2007
1200
113
1400
1600
1800
2000 sample
2200
2400
2600
2800
3000
Bootstrap Techniques for Signal Processing
Summary Signal Processing Group
114
➠ Bootstrap detectors, on average, keep the false alarm rate close to the set Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
level when :
➠ The underlying noise distribution is unknown and probably non-Gaussian. ➠ The noise exhibits some non-stationarity. ➠ Using distribution free tolerance intervals we can specify our confidence that the false alarm rate will not exceed the set level.
➠ Simulations verify the accuracy of tolerance interval bootstrap detection. ➠ With real non-Gaussian non-stationary data the tolerance interval based bootstrap detector maintains the set level. c A.M. Zoubir
ATiSSP Sommersemester 2007
114
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
115
➠ Hypothesis Testing with the Bootstrap
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ An Example: Testing the Frequency Response for Zero ➠ Signal Detection ➠ Bootstrap Matched-Filter ➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
115
Bootstrap Techniques for Signal Processing
Applications in Signal Processing Signal Processing Group
116
The relatively simple examples showed that the bootstrap is powerful. This sugInstitut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
gests its use in real-life applications and more complex problems. We show now how the bootstrap can be applied for:
➠ Noise Floor Estimation in Radar [Zoubir & Boashash (1996)] ➠ Confidence Intervals for Flight Parameters in Passive Sonar [Reid et al. (1996), Zoubir & Boashash (1998)].
➠ Detection of Buried Landmines Using GPR [Zoubir et al. (1999,2000,2001,2002)]. ➠ Image Processing [Zoubir & Iskander (2007)]. ➠ Micro-Doppler Radar [Zoubir & Iskander (2007)]. c A.M. Zoubir
ATiSSP Sommersemester 2007
116
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
117
➠ Signal Detection ➠ Bootstrap Matched-Filter
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Image Processing ➠ Micro-Doppler Radar ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
117
Bootstrap Techniques for Signal Processing
Noise Floor Estimation in OTH-Radar
Signal Processing Group
118
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
IONOSPHERE
F LAYER
E LAYER
TRANSMITTER BEAM
BEAM TO RECEIVER FROM TARGET
350 km 150 km
TRANSMITTER EARTH’S SURFACE
800-3000 km RECEIVER
c A.M. Zoubir
ATiSSP Sommersemester 2007
118
Bootstrap Techniques for Signal Processing
Noise Floor Estimation in OTH-Radar (Cont’d)
Signal Processing Group
119
➠ The noise floor is one of the main radar performance statistics. Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ It is computed by taking a trimmed mean of power estimates in the Doppler spectrum, after excluding the low Doppler power associated with any stationary targets or ground clutter.
➠ The radar operator requires an optimal value for the trim and also supplementary information describing the accuracy of the computed noise floor estimate.
➠ The limited number of samples available and the non-Gaussian nature of the noise field make it particularly difficult to estimate confidence measures reliably. This suggests the use of the bootstrap. c A.M. Zoubir
ATiSSP Sommersemester 2007
119
Bootstrap Techniques for Signal Processing
Optimal Selection of the Trim Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
120
➠ The goal is to use the bootstrap to optimally select the amount of trimming, γ , say. ➠ The problem can be stated as one of estimation of θ from a class of estimates ˆ {θ(γ) : γ ∈ A} indexed by a parameter γ , which is selected according to some optimality criterion.
➠ It is reasonable to use the estimate that leads to a minimum variance. Beˆ may depend on unknown parameters, the optimal cause the variance of θ(γ) parameter γo is unknown.
c A.M. Zoubir
ATiSSP Sommersemester 2007
120
Bootstrap Techniques for Signal Processing
Optimal Selection of the Trim (cont’d)
Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
121
ˆ ➠ The objective is to find γ by estimating the unknown variance of θ(γ) and select γ which minimises the variance estimate. ➠ The problem of trimming has been solved theoretically [Jaeckel (1971)] for the mean of symmetric distributions. The trim is, however, restricted to the interval [0, 25]%.
➠ In addition to the fact that for a small sample size asymptotic results are invalid, explicit expressions for the asymptotic variance may not be available outside the interval [0, 25]%.
c A.M. Zoubir
ATiSSP Sommersemester 2007
121
Bootstrap Techniques for Signal Processing
An Example: Selection of the Trim for the Mean
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
122
Signal Processing Group
➠ Let X = {X1 , . . . , Xn } be a random sample drawn from a distribution function F . ➠ Let X(1) , . . . , X(n) denote the order statistics. ➠ For an integer γ less then n/2, the γ -trimmed mean based on the sample X is given by
µ ˆ(γ) =
1 n−2γ
Pn−γ
i=γ+1 X(i)
➠ For an asymmetric distribution we use the γκ-trimmed mean:
µ ˆ(γ, κ) =
c A.M. Zoubir
ATiSSP Sommersemester 2007
122
1 n−γ−κ
Pn−κ
i=γ+1 X(i) ,
γ, κ < n/2 .
Bootstrap Techniques for Signal Processing
Example Signal Processing Group
123
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 1. Initial conditions. Select the initial trim γ
= 0 and κ = 0.
Step 2. Resampling. Using a pseudo-random number generator, draw a large number, say N , of independent samples
∗ ∗ }, . . . , XN∗ = {XN∗ 1 , . . . , XN∗ n } , . . . , X1n X1∗ = {X11 of size n from X . Each sample is taken with replacement. Step 3. Calculation of the bootstrap statistic. For each bootstrap sample
j = 1, . . . , N , calculate the trimmed mean
µ ˆ∗j (γ, κ) c A.M. Zoubir
=
ATiSSP Sommersemester 2007
1 n−γ−κ
123
Pn−κ
∗ X i=γ+1 j(i) ,
Xj∗ ,
j = 1, . . . , N.
Bootstrap Techniques for Signal Processing
Example (cont’d) Signal Processing Group
124
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 4. Variance estimation. Using bootstrap based trimmed mean values, calculate the estimate of the variance
σ ˆ ∗2 =
1 N −1
PN j=1
µ ˆ∗j (γ, κ) −
1 N
2 ∗ . µ ˆ j=1 j (γ, κ)
PN
Step 5. Repetition. Repeat steps 2–4 using different combinations of the trim
γ and κ. Step 6. Optimal trim. Choose the setting of the trim that results in a minimal variance estimate.
c A.M. Zoubir
ATiSSP Sommersemester 2007
124
Bootstrap Techniques for Signal Processing
Noise Floor Estimation in OTH-Radar(Cont’d) 125
Signal Processing Group
➠ In radar, we estimate residuals for the Doppler spectrum by taking the ratio of the periodogram and a spectral estimate of the raw data. The residuals are
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
used for resampling to generate bootstrap spectral estimates as in a previous example.
➠ We proceed as in the example for the trimmed mean, replacing Xi by CˆXX (ωi ), where ωi = 2πi/n, i = 1, . . . , n, are discrete frequencies and n is the number of observations.
➠ The procedure makes the assumption that the spectral estimates are i.i.d. for distinct discrete frequencies.
➠ The optimal noise floor is found by minimising the bootstrap variance estimate of the trimmed spectrum w.r.t. (γ, κ). c A.M. Zoubir
ATiSSP Sommersemester 2007
125
Bootstrap Techniques for Signal Processing
Noise Floor Estimation in OTH-Radar
Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
126
Step 1. Initial conditions. Select the initial trim γ
= κ = 0.
B , of indepenˆ ∗(1) (ω1 ), . . . , Cˆ ∗(1) (ωM )}, . . . , X ∗ = dent samples X1∗ = {C B XX XX ∗(B) ∗(B) {CˆXX (ω1 ), . . . , CˆXX (ωM )}
Step 2. Resampling.
Draw a large number,
say
Step 3. Calculation of the trimmed mean. For each bootstrap sample Xj∗ , j
1, . . . , B , calculate the trimmed mean µ ˆ∗j (γ, κ)
c A.M. Zoubir
=
ATiSSP Sommersemester 2007
1 M −γ−κ
126
∗(j) ˆ i=γ+1 CXX (ωi ),
Pn−κ
j = 1, . . . , B.
Bootstrap Techniques for Signal Processing
=
Optimal Noise Floor in OTHR (cont’d)
Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
127
Step 4. Variance estimation. Using bootstrap based trimmed mean values, calculate the estimate of the variance
σ ˆ ∗2 =
1 B−1
PB j=1
µ ˆ∗j (γ, κ) −
1 B
2 ∗ µ ˆ . j=1 j (γ, κ)
PB
Step 5. Repetition. Repeat steps 2–4 using different combinations of the trim
γ and κ. Step 6. Optimal trim. Choose the setting of the trim that results in a minimal variance estimate.
c A.M. Zoubir
ATiSSP Sommersemester 2007
127
Bootstrap Techniques for Signal Processing
Optimal Noise Floor in OTHR Signal Processing Group
128 X0 , . . . , Xn−1
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Measurements
Variance estimation
∗2 σ ˆµ ˆ ∗ (γ,κ) =
1 = M − Mγ − Mκ
MX −M κ
1
N X
N − 1 j=1
0
128
N 1 X
N j=1
∗
12
µ ˆ j (γ, κ)A
ˆ ∗(j) (ω(i) ) C XX
i=M γ+1
γ,κ
ATiSSP Sommersemester 2007
µ ˆ∗ B (γ, κ)
∗ @µ ˆ j (γ, κ) −
Find arg min
c A.M. Zoubir
∗ XB
ˆ∗ µ ˆ∗ 2 (γ, κ) 1 (γ, κ) µ
Trimmed mean of resamples for (γ, κ)
µ ˆ∗ b (γ, κ)
X2∗
X1∗
Spectral resamples
∗2 σ ˆµ ˆ ∗ (γ,κ)
Bootstrap Techniques for Signal Processing
How to Bootstrap Power Spectra? Signal Processing Group
129
We will consider a kernel estimate of the spectrum CXX (ω) of a stationary signal
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Xt , t = 0, . . . , n − 1, CˆXX (ω; h) =
1 nh
PM
k=−M
➠ K(·) is symmetric and nonnegative,
K
ω−ωk h
IXX (ωk ) ,
➠ IXX (ωk ) is the periodogram at ωk ➠ ωk = 2πk/n, −M ≤ k ≤ M . ➠ h is the bandwidth of the kernel, ➠ M denotes the largest integer ≤ n/2 c A.M. Zoubir
ATiSSP Sommersemester 2007
129
Bootstrap Techniques for Signal Processing
How to Bootstrap Power Spectra? (Cont’d) 130
Signal Processing Group
➠ Asymptotically for large n and with m = ⌊(h n − 1)/2⌋,
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
CˆXX (ω1 ), . . . , CˆXX (ωM ) are independently ∼ CXX (ωk ) χ24 m+2 /(4 m + 2), k
= 1, . . . , M .
¨ ➠ Alternatively, we use (see [Franke & Hardle (1992)]) the approximate regression
IXX (ωk ) , k = 1, . . . , M , εk = CXX (ωk ) ˆXX (ω) or of functions thereof. to find the distribution of C c A.M. Zoubir
ATiSSP Sommersemester 2007
130
Bootstrap Techniques for Signal Processing
How to Bootstrap Power Spectra? (Cont’d)
Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
131
Step 1. Compute Residuals. Choose an hi
> 0 which does not depend on ω
and compute
εˆk =
IXX (ωk ) ˆXX (ωk ;hi ) , C
k = 1, . . . , M.
Step 2. Rescaling. Rescale the empirical residuals to
ε˜k =
εˆk , εˆ
k = 1, . . . , M, εˆ =
1 M
PM
ˆj j=1 ε
.
Step 3. Resampling. Draw independent bootstrap residuals ε˜∗1 , . . . , ε˜∗M from the empirical distribution of ε˜1 , . . . , ε˜M .
c A.M. Zoubir
ATiSSP Sommersemester 2007
131
Bootstrap Techniques for Signal Processing
How to Bootstrap Power Spectra? (Cont’d)
Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
132
Step 4. Bootstrap estimates. With a bandwidth g , find ∗ IXX (ωk ) = CˆXX (ωk ; g) ε˜∗k , k = 1, . . . , M, ∗ P M ω−ω 1 ∗ k CˆXX (ω; h) = n h k=−M K IXX (ωk ). h
For confidence bands estimation, we would repeat Steps 3 − 4 and find c∗U (and
proceed similarly for c∗L ) such that Pr∗
c A.M. Zoubir
√
ATiSSP Sommersemester 2007
ˆXX (ω;g) ˆ ∗ (ω;h)−C C XX nh ˆXX (ω;g) C
132
≤ c∗U = γ .
Bootstrap Techniques for Signal Processing
How to Bootstrap Power Spectra? (Cont’d)
Signal Processing Group
133 Measured Data Conduct Experiment
x
` ´ IXX ωk
Find Periodogram
Smooth Periodogram
` ´ ˆ C XX ωk
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
k = 1, . . . , M `
ǫ ˆ ωk
´
´ ` IXX ωk ´ ` = ˆ ωk C XX
Resample Bootstrap Data
` ´ ǫ ˆ∗ 1 ωk ´ ` ǫ ˆ∗ 2 ωk
´ ` ǫ ˆ∗ B ωk
c A.M. Zoubir
ATiSSP Sommersemester 2007
133
Reconstruct & Estimate
Reconstruct & Estimate
Reconstruct & Estimate
´ ` ˆ ∗(1) ω C k XX ` ´ ˆ ∗(2) ω C k XX
Confidence Sort
` ´ ˆ ∗(1) ω C ,... k XX ` ´ ˆ ∗(B) ω C k XX
´ ` ˆ ∗(N ) ω ...,C k XX
Bootstrap Techniques for Signal Processing
Interval
Noise Floor Estimation in OTH-Radar
Signal Processing Group
134 70
C XX C
NN
TM
50 Power Spectrum [dB]
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
60
40
30
20
10 −0.5
−0.4
−0.3
−0.2
−0.1 0 0.1 Normalized Doppler Frequency
0.2
0.3
0.4
0.5
Doppler spectrum and estimated noise floor for a real over-the-horizon radar return (no target present). With B c A.M. Zoubir
ATiSSP Sommersemester 2007
ˆ γ, κ = 99, θ(ˆ ˆ ) = 2.0738 × 103 . 134
Bootstrap Techniques for Signal Processing
Noise Floor Estimation in OTH-Radar (Cont’d)
Signal Processing Group
135 70
C XX C
65
NN
TM
Target
55 Power Spectrum [dB]
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
60
50
45
40
35
30
25
20 −0.5
−0.4
−0.3
−0.2
−0.1 0 0.1 Normalized Doppler Frequency
0.2
0.3
0.4
0.5
Doppler spectrum and estimated noise floor for a real over-the-horizon radar return (target present). With B c A.M. Zoubir
ˆ γ, κ = 99, θ(ˆ ˆ ) = 8.9513 × 103 .
ATiSSP Sommersemester 2007
135
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
136
➠ Signal Detection ➠ Bootstrap Matched-Filter
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Image Processing ➠ Micro-Doppler Radar ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
136
Bootstrap Techniques for Signal Processing
Confidence Intervals for an Aircraft’s Flight Parameters 137
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
OVERFLYING AIRCRAFT
ACOUSTIC EMISSION
MICROPHONE
Schematic of the passive acoustic scenario.
c A.M. Zoubir
ATiSSP Sommersemester 2007
137
Bootstrap Techniques for Signal Processing
Signal Processing Group
The Passive Acoustic Approach Signal Processing Group
138
➠ Information about the flight parameters is contained in the acoustic signal as Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
heard by the stationary observer.
➠ The frequency of the observed acoustic signal s(t) from the over-flying aircraft undergoes a time varying Doppler shift.
➠ Thus, the phase of s(t) undergoes a time varying rate of change. A simple model for the aircraft acoustic signal, as heard by a stationary observer, is given by j2π
X(t) = z(t) + U (t) = Aejφ(t) + U (t) = Ae where z(t) c A.M. Zoubir
Rt
−∞
f (τ )dτ
= s(t) + j H[s(t)] and H[·] is the Hilbert transform.
ATiSSP Sommersemester 2007
138
Bootstrap Techniques for Signal Processing
+ U (t) ,
Problem Description (Cont’d) Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
139
v
θ(t)
R(t)
h
r(t)
Schematic of the geometric arrangement of the aircraft and observer in terms of the physical parameters of the model. c A.M. Zoubir
ATiSSP Sommersemester 2007
139
Bootstrap Techniques for Signal Processing
Problem Description (Cont’d) Signal Processing Group
140
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Context: Estimation of an aircraft’s flight parameters (height, speed, range and acoustic frequency) from the acoustic signal as heard by a stationary observer.
Problem:
To estimate an aircraft’s flight parameters using passive acoustic
techniques and to determine confidence bounds given only a single acoustic realisation.
Solution: Estimate the aircraft flight parameters from the time varying phase or frequency of the observed acoustic signal. Use bootstrap techniques to estimate the confidence bounds. c A.M. Zoubir
ATiSSP Sommersemester 2007
140
Bootstrap Techniques for Signal Processing
The Passive Acoustic Model Signal Processing Group
141
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
An observer phase model
φ(t) describes the phase of the observed acoustic
signal in terms of the flight parameters and is given by
φ(t) =
f a c2 2π c2 −v2
t−
q
h2 c+v 2 t2 c+2v 2 th c3
+ φ0 , −∞ < t < ∞ ,
where fa is the source acoustic frequency, c is the speed of sound in the medium,
v is the constant velocity of the aircraft, h is the constant altitude of the aircraft and φ0 is an initial phase constant.
c A.M. Zoubir
ATiSSP Sommersemester 2007
141
Bootstrap Techniques for Signal Processing
The Passive Acoustic Model (Cont’d)
Signal Processing Group
142 16000 14000
phase (radians)
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
12000 10000 8000 6000 4000 2000 0
−15
−10
−5
0
5
10
15
time (s)
Typical Time varying acoustic phase profile of an over-flying aircraft as described by the observer phase model. For this example,
v = 75 m/s, h = 300 m, fa = 85 Hz and
t0 = 0 s. The cartwheel symbol marks the time at which the aircraft is directly overhead the observer. c A.M. Zoubir
ATiSSP Sommersemester 2007
142
Bootstrap Techniques for Signal Processing
The Passive Acoustic Model (Cont’d)
Signal Processing Group
143
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
From the phase model, the instantaneous frequency (IF), relative to the stationary observer, can be expressed as
f (t) =
1 dφ(t) 2π dt
fa c2 = c2 −v2 1 − √
v 2 (t+h/c) h2 (c2 −v 2 )+v 2 c2 (t+h/c)2
.
f (t), or φ(t), −∞ < t < ∞ and c, the aircraft parameters collected in the vector θ = (fa , h, v, t0 )′ can be uniquely determined from the For a given
phase or observer IF model.
c A.M. Zoubir
ATiSSP Sommersemester 2007
143
Bootstrap Techniques for Signal Processing
The Passive Acoustic Model (Cont’d)
Signal Processing Group
144 115 110 105
frequency (Hz)
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
100 95 90 85 80 75 70 65
−15
−10
−5
0
5
10
15
time (s)
Typical Time varying acoustic frequency profile of an over-flying aircraft as described by the observer frequency model. For this example, v and t0
= 75 m/s, h = 300 m, fa = 85 Hz
= 0 s. The cartwheel symbol indicates the time at which the aircraft is directly
overhead the observer. c A.M. Zoubir
ATiSSP Sommersemester 2007
144
Bootstrap Techniques for Signal Processing
A Bootstrap Approach Signal Processing Group
145
Step 0. Conduct Experiment. Collect and sample the data to obtain
Xt , t =
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
−n/2, . . . , n/2 − 1. Xt to provide a ˆt which approximates the true phase φt . non-decreasing function φ
Step 1. Phase Unwrapping. Unwrap the phase of the signal
ˆ , an initial estimate of the aircraft parameθ ˆt in a leastters by fitting the non-linear observer phase model φt;θ to φ
Step 2. Initial Estimation. Obtain
squares sense. Step 3. Compute residuals. Compute the residuals
εˆt = φ
ˆ
t;θ
c A.M. Zoubir
ATiSSP Sommersemester 2007
145
− φˆt ,
t = −n/2, · · · , n/2 − 1 .
Bootstrap Techniques for Signal Processing
A Bootstrap Approach (Cont’d) Signal Processing Group
146
Step 4. Variance estimation. Compute σ ˆi , a bootstrap estimate of the standard Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
deviation of θˆi , i
= 1, . . . , 4.
X ∗ = {ˆ ε∗−n/2 , · · · , εˆ∗n/2−1 }, with replacement, from X = {ˆ ε−n/2 , · · · , εˆn/2−1 } and construct
Step 5. Resampling. Draw a random sample
φˆ∗t = φ
ˆ
t;θ
+ εˆ∗t .
Step 6. Bootstrap estimates. Obtain and record the bootstrap estimates of the ∗ ˆ ˆ∗ in a leastaircraft parameters θ by fitting the observer phase model to φ t
squares sense.
c A.M. Zoubir
ATiSSP Sommersemester 2007
146
Bootstrap Techniques for Signal Processing
A Bootstrap Approach (Cont’d) Signal Processing Group
147
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Step 7. Bootstrap variance estimate. Estimate the standard deviation of θˆi∗ using a nested bootstrap step and compute and record
Tˆi∗ =
θˆi∗ −θˆi σ ˆi∗
,
i = 1, . . . , 4 .
Step 8. Repetition. Repeat Steps 5 through 7 a large number of times N . Step 9. Sorting. Order the bootstrap estimates as
∗ ∗ ≤ ··· ≤ ≤ Tˆi,(2) Tˆi,(1)
∗ Tˆi,(N ) and compute the (1 − α)100% confidence interval ∗ ˆi − Tˆ∗ σ (θˆi − Tˆi,(q σ ˆ , θ i i,(q2 ) ˆi ) , 1)
where q1 c A.M. Zoubir
= N − ⌊N α/2⌋ + 1 and q2 = ⌊N α/2⌋.
ATiSSP Sommersemester 2007
147
Bootstrap Techniques for Signal Processing
Simulation Results Signal Processing Group
148
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ An n = 320 point passive acoustic signal zt is generated at three levels of SNR (15 dB, 20 dB and 30 dB).
➠ We set h = 304.8 m, v = 102.89 m/s, fa = 1 Hz, t0 = 0 s, sampling frequency fs = 8 Hz and bootstrap variables B1 = 25, B2 = 100 and B3 = 1000. ➠ The bootstrap confidence bounds are computed and compared with those obtained by Monte Carlo simulation where the aircraft parameters are calculated for 1000 independent realisations of zt .
c A.M. Zoubir
ATiSSP Sommersemester 2007
148
Bootstrap Techniques for Signal Processing
Simulation Results (Cont’d) Signal Processing Group
149
SNR
Actual
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
dB
30
20
15
h [m]
v [m/s]
t0 [s]
fa [Hz]
304.8
102.89
0.000
1.000
BS
MC
BS
MC
BS
MC
BS
MC
Upper Bound
308.40
308.0
102.94
103.04
0.004
0.002
1.000
1.000
Lower Bound
301.4
301.9
102.67
102.74
-0.005
-0.006
1.000
1.000
Interval length
6.6
6.1
0.27
0.30
0.009
0.008
0.000
0.000
Upper Bound
309.5
314.5
103.17
103.34
0.011
0.016
1.000
1.000
Lower Bound
290.8
295.2
102.26
102.44
-0.020
-0.016
0.999
0.999
Interval length
18.7
19.3
0.91
0.90
0.031
0.032
0.001
0.001
Upper Bound
315.7
320.7
103.54
103.65
0.040
0.027
1.001
1.001
Lower Bound
286.0
288.6
102.14
102.12
-0.008
-0.027
0.999
0.999
Interval length
29.7
32.1
1.40
1.53
0.048
0.054
0.002
0.002
The bootstrap derived 95% confidence bounds for each of the parameters are compared with the 95% confidence bounds determined by Monte Carlo simulation.
c A.M. Zoubir
ATiSSP Sommersemester 2007
149
Bootstrap Techniques for Signal Processing
Simulation Results (Cont’d) Signal Processing Group
300
300
200
200
count
count
150 100 0 0.9992
300
100
102.5
103
(b) speed (m/s) (bootstrap)
300
count 290
295
300
305
310
315
(c) height (m) (bootstrap) count
100 0 −0.03
−0.02
−0.01
0
0.01
0.02
(d) time reference (s) (bootstrap)
0.03
1.0008
102.5
103
103.5
(f) speed (m/s) (Monte Carlo)
200 100
300
200
1.0004
100
0 285
320
1
200
300
100
0.9996
(e) source frequency (Hz) (Monte Carlo)
0 102
103.5
200
0 285
0 0.9992
1.0008
count
count
1.0004
200
300
count
1
(a) source frequency (Hz) (Bootstrap)
0 102
count
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
300
0.9996
100
290
295
300
305
310
315
320
(g) height (m) (Monte Carlo)
200 100 0 −0.03
−0.02
−0.01
0
0.01
0.02
0.03
(h) time reference (s) (Monte Carlo)
The histograms of the bootstrap distribution (left column) are compared with the histograms obtained by Monte Carlo simulation (right column) for each of the parameters at 20 dB SNR. The true values are indicated by the dashed line. c A.M. Zoubir
ATiSSP Sommersemester 2007
150
Bootstrap Techniques for Signal Processing
Real Experiment Signal Processing Group
151
seconds of latitude (relative to the ground location)
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
−100
−50
ground location
0
50
100
150
−150
−100
−50
0
50
100
seconds of longitude (relative to the ground location)
The complete flight path of an aircraft acoustic data experiment conducted at Bribie Island showing five fly-overs. The ground location is indicated by the cartwheel symbol. c A.M. Zoubir
ATiSSP Sommersemester 2007
151
Bootstrap Techniques for Signal Processing
The Instantaneous Frequency Signal Processing Group
152
90
frequency (Hz)
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
85
80
75
70
65
60
55 0
5
10
15
20
25
time (s)
Doppler signature of an acoustic signal.
c A.M. Zoubir
ATiSSP Sommersemester 2007
152
Bootstrap Techniques for Signal Processing
Real Experiment (Cont’d) Signal Processing Group
153
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Run 1.
Run 2.
Run 3.
Run 6.
h [m]
v [m/s]
t0 [s]
fa [Hz]
Nominal Value
149.05
36.61
0.00
76.91
Upper Bound
142.03
36.50
0.02
76.99
Lower Bound
138.16
36.07
-0.03
76.92
Interval length
3.87
0.43
0.05
0.07
Nominal Value
152.31
52.94
0.00
77.90
Upper Bound
161.22
53.27
-0.04
78.18
Lower Bound
156.45
52.64
-0.06
78.13
Interval length
4.77
0.62
0.02
0.05
Nominal Value
166.52
47.75
0.00
75.94
Upper Bound
232.30
57.29
0.10
76.48
Lower Bound
193.18
53.47
-0.14
75.87
Interval length
39.13
3.83
0.24
0.60
Nominal Value
233.01
56.56
0.00
77.65
Upper Bound
243.02
57.15
0.74
77.96
Lower Bound
209.72
53.69
0.68
77.80
Interval length
33.30
3.46
0.06
0.16
The bootstrap derived 95% confidence bounds for four real test signals using the unwrapped phase based parameter estimator. c A.M. Zoubir
ATiSSP Sommersemester 2007
153
Bootstrap Techniques for Signal Processing
Interpretation of the Results Signal Processing Group
154
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ The proposed bootstrap and unwrapped phase based estimation techniques can be used to provide a practical flight parameter estimation scheme.
➠ The techniques do not assume any statistical distribution for the parameter estimates and can be applied in the absence of multiple acoustic realisations.
➠ The obtained confidence bounds are in close agreement with those obtained from Monte Carlo simulations.
➠ Similar experiments with Central Finite Difference (CFD) estimates were performed. The confidence bounds presented here are much tighter than those of the CFD based estimates.
c A.M. Zoubir
ATiSSP Sommersemester 2007
154
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
155
➠ Signal Detection ➠ Bootstrap Matched-Filter
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Image Processing ➠ Micro-Doppler Radar ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
155
Bootstrap Techniques for Signal Processing
Detection of Buried Landmines Using GPR
Signal Processing Group
156
➠ In many post-war zones, landmines of little or no metal content have been Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
found. Metal detectors are ineffective in such regions.
➠ Ground Penetrating Radar (GPR) can detect discontinuities in the electric properties of the soil, and therefore has the potential to detect buried plastic landmines.
➠ GPR is sensitive to changes in the soil conditions, e.g. moisture content and soil type, and signal processing may be used to improve detection of targets.
➠ Robust detection requires adaptive techniques to allow for changes in the soil whilst still detecting targets.
c A.M. Zoubir
ATiSSP Sommersemester 2007
156
Bootstrap Techniques for Signal Processing
Application to Real Data Signal Processing Group
157
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ An FR-127-MSCB Impulse GPR (ImGPR) system developed by the CSIRO has been used.
➠ 127 soundings per second, each composed of 512 samples with 12 bit accuracy.
➠ Bistatic bow-tie antennae transmitting wide-band, ultra-short duration pulses. Center frequency of 1.4GHz and 80% bandwidth.
➠ Surrogate targets for PMN, PMN-2 and M14 AP mines with non-metallic casings were used.
c A.M. Zoubir
ATiSSP Sommersemester 2007
157
Bootstrap Techniques for Signal Processing
Impulse GPR System Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
158
The ImGPR runnig over a sand box.
c A.M. Zoubir
ATiSSP Sommersemester 2007
158
Bootstrap Techniques for Signal Processing
and Surrogate Landmines Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
159
Surrogate Landmines.
c A.M. Zoubir
ATiSSP Sommersemester 2007
159
Bootstrap Techniques for Signal Processing
Minelike Surrogate Targets Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
160
Target
Surrogate
Dimensions
Orientation
diam×hght ST-AP(1)
M14
52 × 42 mm
not critical
ST-AP(2)
PMN
118 × 50mm
sensitive
ST-AP(3)
PMN2
115 × 53mm
sensitive
Minelike surrogate targets used.
c A.M. Zoubir
ATiSSP Sommersemester 2007
160
Bootstrap Techniques for Signal Processing
Signal Model and Preliminary Investigations
Signal Processing Group
161
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ It is difficult to define a complete, physically-motivated signal model of the GPR backscatter waveform.
➠ Extensive data analysis has shown that a polynomial amplitude - polynomial phase model is suitable,
gt =
hP
Pa n=0
an tn
i
h P i b exp j Pm=0 bm tm + zt
➠ zt is stationary interference and the amplitude and frequency modulation are described by polynomials of order Pa and Pb respectively.
c A.M. Zoubir
ATiSSP Sommersemester 2007
161
Bootstrap Techniques for Signal Processing
Signal Model Investigations: An Example
Signal Processing Group
162
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ The signal model has been applied to GPR returns from clay soil, containing a small plastic target, denoted ST-AP(1) – a surrogate for the M14 antipersonnel (AP) mine.
➠ ST-AP(1) has a PVC casing and is filled with paraffin wax. ➠ A solid stainless steel cylinder 5 cm in diameter and length is also present, denoted by SS05x05.
➠ Phase and amplitude parameters are estimated using the DPT and the method of least-squares respectively.
c A.M. Zoubir
ATiSSP Sommersemester 2007
162
Bootstrap Techniques for Signal Processing
Results Signal Processing Group
163 4
2
x 10
0
−1
−2
0
2
4
6 Time (ns)
8
10
12
0
2
4
6 Time (ns)
8
10
12
6000 4000 Amplitude
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Amplitude
1
2000 0 −2000 −4000
Top: Fit of the polynomial amplitude - polynomial phase model (bold line) to data (fine line) from non-homogeneous soil. Bottom: Residual of the fit to data.
c A.M. Zoubir
ATiSSP Sommersemester 2007
163
Bootstrap Techniques for Signal Processing
Test on the Parameters Signal Processing Group
164
It is suggested that the mean of this distribution changes in the presence of a target. b0
−0.5
−1
0
100
200
300
400
500
600
700
800
900
1000
0 100 −3 x 10
200
300
400
500
600
700
800
900
1000
0
200
300
400
500 600 Distance (steps)
700
800
900
1000
0.3
b1
0.28 0.26 0.24 0
b2
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
0
−0.5
−1
100
Phase coefficients from an order 2 model. Vertical dotted lines indicate the approximate position of ST-AP(1) (left) and SS05x05 (right) targets. c A.M. Zoubir
ATiSSP Sommersemester 2007
164
Bootstrap Techniques for Signal Processing
Test on the Parameters (Cont’d) Signal Processing Group
165
➠ For the data set shown b2 is seen to decrease in the presence of a target, however, in some cases, in particular when the soil type is loam, b2 increases
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
with the presence of a target. We suggest
H : b2 = b2,0 K : b2 6= b2,0 ➠ The obvious test statistic is T =
ˆb2 −b2,0 σ ˆˆb 2
where σ ˆˆb2 is an estimate of the standard deviation of ˆb2 .
➠ Since the exact distribution of ˆb2 is unknown and has unknown variance, we use the bootstrap to determine thresholds. c A.M. Zoubir
ATiSSP Sommersemester 2007
165
Bootstrap Techniques for Signal Processing
Example Signal Processing Group
166
From a background region of 100 samples, estimates of b2 are seen in a Q − Q plot.
Normal Probability Plot
0.99 0.98 0.95 0.90
0.75 Probability
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
0.997
0.50
0.25
0.10 0.05 0.02 0.01 0.003 −3.8
−3.6
−3.4
−3.2 Data
−3
−2.8
−2.6 −4
x 10
Normal probability plot of third phase coefficient estimate, ˆ b2 , from a background region. c A.M. Zoubir
ATiSSP Sommersemester 2007
166
Bootstrap Techniques for Signal Processing
Bootstrap-Based Detection Signal Processing Group
167
Step 1. If gt for t
= 0, ..., n − 1 is a sampled GPR return signal, fit a polynomial
amplitude - polynomial phase model to the data. From the model, an estimate Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
of the second order phase parameter, ˆ b2 , is made. Step 2. Form the residual signal rt
= gt − gˆt , where gˆt is the polynomial
amplitude-polynomial phase model corresponding to estimated parameters.
Step 3. Whiten rt by removing an AR model of suitable order to obtain the innovations zt . Step 4. Re-sample from zt
n times using the Block Bootstrap [Politis (1998)] to
obtain zt∗ . Step 5. Repeat Step 4. Step 6. Generate c A.M. Zoubir
N times to obtain zt∗1 , ..., zt∗N .
N bootstrap residual signals rt∗i , for i = 1, ..., N by filtering
ATiSSP Sommersemester 2007
167
Bootstrap Techniques for Signal Processing
Bootstrap-Based Detection Signal Processing Group
168
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
zt∗i with the AR process obtained in Step 3. Step 7. Generate N bootstrap signals gt∗i
= gˆt + rt∗i , for i = 1, ..., N .
Step 8. Estimate the third phase coefficient from gt∗i to obtain ˆ b∗i 2 for i Step 9. Calculate the bootstrap statistics T
∗i
=
ˆb∗i −ˆb2 2 σ ˆˆ∗
for i
= 1, ..., N .
= 1, ..., N .
b2
Step 10. Compare the test statistic
T =
ˆb2 −b2,0 σ ˆˆb
to the empirical distribution of
2
T ∗ . Reject H if T is in the tail of the distribution, otherwise retain H .
c A.M. Zoubir
ATiSSP Sommersemester 2007
168
Bootstrap Techniques for Signal Processing
Discussion Signal Processing Group
169
➠ Shown in the following figures are sample results for three surrogate mines Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
as well as various other “targets”. The location of the first target in each scenario is around the 300th trace, while the second target is around the 690’th trace.
➠ Overall, the results are very encouraging. It can be said that the targets have been correctly detected, while false alarms appear to be concentrated in areas near the targets – rather than true false alarms that occur far from the targets in background only areas.
➠ These near-target false alarms may be triggered by disturbances to the soil structure caused by the burying of the target. c A.M. Zoubir
ATiSSP Sommersemester 2007
169
Bootstrap Techniques for Signal Processing
Example Signal Processing Group
170
200 300 400 500 100
200
300
400
500
100
200
300
400 500 Sample
600
700
800
900
−3
x 10 2 Test statistic
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Depth samples
100
1.5 1 0.5 0 0
Top:
600
700
800
900
B -scan with ST-AP(1) and SS05x05 targets buried at 5 cm in clay.
Bottom: Corresponding test statistics (thin line) and detection decision (thick line). c A.M. Zoubir
ATiSSP Sommersemester 2007
170
Bootstrap Techniques for Signal Processing
Example Signal Processing Group
171 Depth samples
100 200 300
500 100
200
300
400
500
600
700
800
100
200
300
400 500 Sample
600
700
800
900
−3
x 10 2 Test statistic
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
400
1.5 1 0.5 0 0
Top:
900
B -scan with ST-AP(2) target (parallel orientation) and an aluminium soft-
drink can buried at 5 cm in clay. Bottom: Corresponding test statistics (thin line) and detection decision (thick line). c A.M. Zoubir
ATiSSP Sommersemester 2007
171
Bootstrap Techniques for Signal Processing
Example Signal Processing Group
172 Depth samples
100 200 300
500 100
200
300
400
500
600
700
800
100
200
300
400 500 Sample
600
700
800
900
−3
x 10 2 Test statistic
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
400
1.5 1 0.5 0 0
Top:
900
B -scan with ST-AP(3) target (perpendicular orientation) and shrapnel buried
at 5 cm in clay. Bottom: Corresponding test statistics (thin line) and detection decision (thick line). c A.M. Zoubir
ATiSSP Sommersemester 2007
172
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
173
➠ Signal Detection ➠ Bootstrap Matched-Filter
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Image Processing ➠ Micro-Doppler Radar ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
173
Bootstrap Techniques for Signal Processing
An Image Processing Example Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
174
➠ We consider an example of fitting functions such as circles or ellipses to the outlines of the pupil and limbus (iris) in eye images [Iskander et al. (2004)].
➠ Performance for extracting eye features can be evaluated with synthetic images.
➠ For real images and particularly for the limbus [Morelande et al. (2002)] these procedures need to be benchmarked against manual operators.
c A.M. Zoubir
ATiSSP Sommersemester 2007
174
Bootstrap Techniques for Signal Processing
An Image Processing Example (Cont’d)
Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
175
Typical example of a slit lamp image of an eye with manually selected points (yellow crosses) around the limbus (i.e., iris outline) area.
c A.M. Zoubir
ATiSSP Sommersemester 2007
175
Bootstrap Techniques for Signal Processing
An Image Processing Example (Cont’d) 176
Signal Processing Group
➠ The number of points (xi , yi ), i = 1, 2, . . . , n, where xi and yi denote the
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
horizontal and vertical point position is small.
➠ A linear least-squares procedure can be applied: 2 2 x 1 + y1 2x1 2y1 1 ε1 p1 . . . . . = .. .. .. .. · p + .. 2 p3 x2n + yn2 2xn 2yn 1 εn where p1
= x0 , p2 = y0 , p3 = x20 + y02 − R2 and εi , i = 1, 2, . . . , n, Y =X ·P +E
➠ A solution is given by ˆ = XT X P c A.M. Zoubir
ATiSSP Sommersemester 2007
176
−1
XT Y .
Bootstrap Techniques for Signal Processing
An Image Processing Example (Cont’d)
Signal Processing Group
177
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ The question is how well a particular operator can fit a circle to the limbus. ➠ One way is to select 8 data point pairs (considered in our example) 1000 times, say.
➠ Although feasible, this task is laborious and the results would most definitely be affected by the subsequently decreasing commitment of the operator.
➠ The alternative is the bootstrap. ➠ The selected data points (xi , yi ), i = 1, 2, . . . , n, form a particular pattern and are not i.i.d, unlike the modelling errors εi , i = 1, . . . , n, collected in the random sample ε = {ε1 , ε2 , . . . , εn }. c A.M. Zoubir
ATiSSP Sommersemester 2007
177
Bootstrap Techniques for Signal Processing
The Bootstrap Procedure Signal Processing Group
178
ˆ and construct an estimate of the Step 1. Estimate x0 , y0 , and R, collected in P
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
ˆ circle using Y
ˆ. =X ·P
ˆ Step 2. Calculate the residuals E
= Y − Yˆ .
∗ ˆ Step 3. Create a set of bootstrap residuals E by resampling with replacement ˆ. from E ∗ ∗ ˆ ˆ ˆ Step 4. Create a new estimate of the circle by Y = X · P + E .
Step 5. Estimate a new set of parameters from the newly created bootstrap sample
∗ ˆ Pb
T
= X X
−1
∗ ˆ X Y . T
∗ ∗ ∗ ˆ ˆ ˆ Step 6. Repeat Steps 3-5 B times to obtain a set of P 1 , P 2 , . . . , P B , from
which empirical distributions can be obtained. c A.M. Zoubir
ATiSSP Sommersemester 2007
178
Bootstrap Techniques for Signal Processing
The Bootstrap Procedure-Results Signal Processing Group
179 120
Frequency of occurrence
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
100
80
60
40
20
0 5.92
5.94
5.96
5.98
6
6.02
6.04
Bootstrap estimates of limbus radius
6.06
An example of the distribution (histogram) of the limbus radius obtained with the bootstrap method is shown
c A.M. Zoubir
ATiSSP Sommersemester 2007
179
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
180
➠ Signal Detection ➠ Bootstrap Matched-Filter
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Image Processing ➠ Micro-Doppler Radar ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
180
Bootstrap Techniques for Signal Processing
Micro-Doppler Analysis Signal Processing Group
181
➠ Doppler often arises in engineering applications where radar, ladar, sonar
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
and ultra-sound measurements are made.
➠ Due to the relative motion of an object with respect to the measurement system.
➠ If the motion is harmonic, for example due to vibration or rotation, the resulting signal can be well modeled by an FM process [Huang et al. 1990].
➠ Estimation of the FM parameters may allow one to determine physical properties such as the angular velocity and displacement of the vibrational/rotational motion which can in turn be used for classification.
➠ Objective: estimate the FM parameters along with a measure of accuracy, such as confidence intervals. c A.M. Zoubir
ATiSSP Sommersemester 2007
181
Bootstrap Techniques for Signal Processing
Micro-Doppler Analysis (Cont’d) Signal Processing Group
182
➠ Assume the following AM-FM signal model:
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
s(t) = a(t) exp{jϕ(t)} where a(t; α)
=
Pq
k α t and α = (α0 , . . . , αq ) . k=0 k
➠ The phase modulation for a micro-Doppler signal is described by a sinusoidal function: ϕ(t) = −D/ωm cos(ωm t + φ). ➠ The instantaneous frequency (IF) of the signal is given by: dϕ(t) ωi (t; β) , = D sin(ωm t + φ), dt where β = (D, ωm , φ) are termed the FM or micro-Doppler parameters. ➠ Given {x(k)}nk=1 of X(t) = s(t) + V (t), the goal is to estimate the microDoppler parameters as well as their confidence intervals.
c A.M. Zoubir
ATiSSP Sommersemester 2007
182
Bootstrap Techniques for Signal Processing
Micro-Doppler Analysis (Cont’d) Signal Processing Group
183
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ The estimation of the phase parameters is performed using a time-frequency Hough transform (TFHT) [Barbarossa and Lemoine1996,Cirillo, Zoubir & Amin 2006].
➠ The TFHT is based on the pseudo Wigner-Ville (PWVD) distribution: (L−1)/2
Pxx [k, ω) =
X
l=−(L−1)/2
h[k]x[k + l]x∗ [k − l]e−j2ω ,
k = −(L − 1)/2, . . . , n − (L − 1)/2, where h[k] is a windowing function of duration L samples.
for
c A.M. Zoubir
ATiSSP Sommersemester 2007
183
Bootstrap Techniques for Signal Processing
Micro-Doppler Analysis (Cont’d) Signal Processing Group
184
➠ The TFHT based on the PWVD is given by: n−(L−1)/2+1
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
H(β) =
X
Pxx [k, ωi (n; β)),
k=−(L−1)/2
where ωi (t) is described above and β
= (D, ωm , φ).
➠ An estimate of β is obtained from the location of the largest peak of H(β). ➠ Once the phase parameters have been estimated, the phase term is demodulated and the amplitude parameters are estimated via least-squares.
➠ Given estimates for α and β , the residuals are obtained. ➠ We whiten the residuals using a fitted AR model. The innovations are resampled, filtered and added to the estimated signal term. c A.M. Zoubir
ATiSSP Sommersemester 2007
184
Bootstrap Techniques for Signal Processing
Micro-Doppler Analysis (Cont’d) Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
185
➠ The results shown here are based on an experimental radar system, operating at carrier frequency fc = 919.82 MHz. ➠ After demodulation, the in-phase and quadrature baseband channels are sampled at fs = 1 kHz. ➠ The radar system is directed toward a spherical object, swinging with a pendulum motion, which results in a typical micro-Doppler signature.
➠ The PWVD of the observations is computed and shown below:
c A.M. Zoubir
ATiSSP Sommersemester 2007
185
Bootstrap Techniques for Signal Processing
Micro-Doppler Analysis (Cont’d) Signal Processing Group
186 PWVD of the radar data
200
200
150
150
100
100 Frequency (Hz)
Frequency (Hz)
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
PWVD of the radar data
50 0 −50
50 0 −50
−100
−100
−150
−150
−200
−200
−250
0.05
0.1
0.15
0.2
0.25 0.3 Time (sec)
0.35
0.4
0.45
−250
(a)
0.05
0.1
0.15
0.2
0.25 0.3 Time (sec)
0.35
0.4
0.45
(b)
The PWVD of the radar data (a). The PWVD of the radar data and the microDoppler signature estimated using the TFHT (b).
c A.M. Zoubir
ATiSSP Sommersemester 2007
186
Bootstrap Techniques for Signal Processing
Micro-Doppler Analysis (Cont’d) Signal Processing Group
187 Real (in−phase) component
Time−domain waveform of residuals 0.01 Re{ r(n) } Im{ r(n) }
Data AM−FM Fit
0.005
Amplitude (V)
0.02 0 −0.02 −0.04
0
−0.005
−0.01
−0.015
0
0.05
0.1
0.15
0.2
0.25 0.3 0.35 Time (sec) Imaginary (quadrature) component
0.4
0.45
0.5
0
0.05
0.1
0.15
0.2
0.25 Time (sec)
0.3
0.35
0.4
0.45
0.5
Estimated spectrum of residuals
0.04
−20 Periodogram AR Fit
Data AM−FM Fit
0.02
−30
−40 Power (dBW)
Amplitude (V)
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
Amplitude (V)
0.04
0
−50
−60
−70
−0.02
−80
−0.04
0
0.05
0.1
0.15
0.2
0.25 0.3 Time (sec)
0.35
0.4
(c)
0.45
0.5
−90 −500
−400
−300
−200
−100
0 f (Hz)
100
200
300
400
(d)
The real and imaginary components of the radar signal are with their estimated counterparts (c). The real and imaginary parts of the residuals and their spectral estimates (d). c A.M. Zoubir
ATiSSP Sommersemester 2007
187
Bootstrap Techniques for Signal Processing
500
Micro-Doppler Analysis (Cont’d) Signal Processing Group
188 Boostrap confidence interval for ω
Boostrap confidence interval for D
m
5
450 Bootstrap dist. Initial estimate 95% CI
4.5
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
4
Bootstrap dist. Initial estimate 95% CI
400 350
3.5
300
3 250 2.5 200 2 150
1.5
100
1
50
0.5 0 71.7
71.8
71.9
72 D/(2π) (Hz)
72.1
72.2
72.3
0 3.037
3.038
3.039
3.04
3.041 3.042 ω /(2π) (Hz)
3.043
3.044
3.045
3.046
m
(a)
(b)
The bootstrap distributions and 95% confidence intervals for the FM parameters
D (a) and ωm (b) using N = 500. c A.M. Zoubir
ATiSSP Sommersemester 2007
188
Bootstrap Techniques for Signal Processing
Outline Signal Processing Group
189
➠ Signal Detection ➠ Bootstrap Matched-Filter
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Tolerance Interval Bootstrap Matched-Filter ➠ Applications in Signal processing ➠ Noise Floor Estimation in OTH-Radar ➠ Confidence Intervals for Flight Parameters in Passive Sonar ➠ Detection of Buried Landmines Using GPR ➠ Image Processing ➠ Micro-Doppler Radar ➠ Summary ➠ Acknowledgements c A.M. Zoubir
ATiSSP Sommersemester 2007
189
Bootstrap Techniques for Signal Processing
Summary Signal Processing Group
190
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ It is often necessary to find the sampling distributions of parameter estimators, so that the respective means and variances can be calculated, and more generally, confidence intervals for the true parameters can be set.
➠ Most techniques for computing variances or confidence intervals assume that the size of the available set of sample values is sufficiently large, so that “asymptotic” results can be applied.
➠ In many signal processing problems this assumption cannot be made because, for example, the process is non-stationary and only small portions of stationary data are considered.
c A.M. Zoubir
ATiSSP Sommersemester 2007
190
Bootstrap Techniques for Signal Processing
Summary (Cont’d) Signal Processing Group
191
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ Bootstrap techniques are an alternative to asymptotic methods. ➠ The bootstrap does with a computer what the experimenter would do in practice, if it were possible: they would repeat the experiment.
➠ With the bootstrap, the observations are randomly re-assigned, and the estimates re-computed. This process is done thousands of times to simulate repeated experiments.
➠ In an era of exponentially increasing computational power, such computerintensive methods are becoming increasingly attractive.
c A.M. Zoubir
ATiSSP Sommersemester 2007
191
Bootstrap Techniques for Signal Processing
Summary (Cont’d) Signal Processing Group
192
➠ The tutorial provides the fundamental concepts and methods needed by the Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
signal processing practitioner to decide when and how to apply the bootstrap successfully.
➠ The tutorial focuses on the independent data bootstrap. The assumption of i.i.d. data can break down in practice either because the data is not independent or because it is not identically distributed, or both.
➠ The bootstrap can still be invoked if we knew the model that generated the data. In other cases we can make the reasonable assumption that the data is identically distributed but not independent such as in autoregressive processes. c A.M. Zoubir
ATiSSP Sommersemester 2007
192
Bootstrap Techniques for Signal Processing
Summary (Cont’d) Signal Processing Group
193
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
➠ For confidence interval estimation or hypothesis testing, it is essential that the statistic used is asymptotically pivotal.
➠ It has been shown that working on a variance stable scale is better than studentising.
➠ When a variance stabilising transformation is not known, bootstrap can be used to estimate it.
➠ Many applications have been presented to demonstrate the power of the bootstrap. Special care is however required when applying the bootstrap in real-life situations.
c A.M. Zoubir
ATiSSP Sommersemester 2007
193
Bootstrap Techniques for Signal Processing
Acknowledgements Signal Processing Group
194
➠ Many of the examples and simulations were implemented by my former and Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
current PhD students. I wish to thank them for their support.
➠ The passive acoustic problem was the subject of Dr. David Reid’s PhD. He collected the data and implemented the algorithms.
➠ The data from the JINDALEE over-the-horizon radar system and the GPR was provided by the Defence Science & Technology Organisation in Salisbury, South Australia.
➠ The Australian Research Council is acknowledged for providing partial financial support for some of the projects.
c A.M. Zoubir
ATiSSP Sommersemester 2007
194
Bootstrap Techniques for Signal Processing
Bibliography Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
195
This presentation is based on
➠ A. M. Zoubir and D. R. Iskander. Bootstrap Techniques for Signal Processing. Cambridge University Press, 2004.
➠ A. M. Zoubir and B. Boashash. The Bootstrap and Its Application in Signal Processing. IEEE Signal Processing Magazine, 15(1):56–76, 1998.
c A.M. Zoubir
ATiSSP Sommersemester 2007
195
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
196
[1] H. Akaike. Comments on “On Model Structure Testing in System Identification”. International J. Control, 27:323–324, 1978. [2] T. W. Anderson. An Introduction to Multivariate Statistical Analysis. John Wiley and Sons, 1984. [3] T. Ankenbrand and M. Tomassini. Predicting Multivariate Financial Time Series Using Neural Networks: the Swiss Bond Case. In Proceedings of the IEEE/IAFE 1996 Conference on Computational Intelligence for Financial Engineering (CIFEr), pages 27–33, New York, USA, 1996. IEEE. [4] G. Archer and K. Chan. Bootstrapping Uncertainty in Image Analysis. In Proceedings of the 12th Symposium in Computational Statistics, pages 193–198, Heidelberg, Germany, 1996. Physica-Verlag. c A.M. Zoubir
ATiSSP Sommersemester 2007
196
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
197
[5] C. Banga and F. Ghorbel. Optimal Bootstrap Sampling for Fast Image Segmentation: Application to Retina Image. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP-93), Minnesota, 1993. [6] R. Bhar and C. Chiarella. Bootstrap Algorithms and Applications. In Proceedings of the IEEE/IAFE 1996 Conference on Computational Intelligence for Financial Engineering (CIFEr), pages 168–182, New York, USA, 1996. IEEE. [7] R. N. Bhattacharya and R. R. Rao. Normal Approximation and Asymptotic Expansions. Wiley, New York, 1976. ¨ [8] J. F. Bohme and D. Maiwald. Multiple Wideband Signal Detection and Tracking from ¨ ¨ editors, Proceedings of the 10th Towed Array Data. In M. Blanke and T. Soderstr om, IFAC Symposium on System Identification (SYSID 94), volume 1, pages 107–112, Copenhagen, July 1994. Danish Automation Soc. [9] A. Bose. Edgeworth Correction by Bootstrap in Autoregressions. The Ann. of Statist., 16:1709–1722, 1988. [10] D. R. Brillinger. Time Series: Data Analysis and Theory. Holden-Day, 1981. [11] P. Ciarlini. Bootstrap Algorithms and Applications. In Advanced Mathematical Tools c A.M. Zoubir
ATiSSP Sommersemester 2007
197
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
198
in Metrology III, pages 12–23, Singapore, 1997. World Scientific.
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
[12] M. G. Cox, P. M. Harris, M. J. T. Milton, and P. T. Woods. A method for evaluating trends in ozone-concentration data and its application to data from the UK Rural Ozone Monitoring network. In Advanced Mathematical Tools in Metrology III, pages 171–177, Singapore, 1997. World Scientific. ´ Mathematical Methods of Statistics. Princeton University Press, 1967. [13] H. Cramer. [14] P. M. Djuri´c and S. M. Kay. Parameter Estimation of Chirp Signals. IEEE Transactions on Acoustics, Speech and Signal Processing, 38(12):2118–2126, December 1990. [15] B. Efron. Bootstrap Methods. Another Look at the Jackknife. The Ann. of Statist., 7:1–26, 1979. [16] B. Efron. Computers and the Theory of Statistics: Thinking the Unthinkable. SIAM Review, 4:460–480, 1979. [17] B. Efron. Better Bootstrap Confidence Interval. J. Amer. Statist. Assoc., 82:171– 185, 1987. [18] B. Efron and R. Tibshirani. An Introduction to the Bootstrap. Chapman and Hall, c A.M. Zoubir
ATiSSP Sommersemester 2007
198
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
199
1993.
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
[19] B. G. Ferguson. A Ground Based Narrow-Band Passive Acoustic Technique for Estimating the Altitude and Speed of a Propeller Driven Aircraft. J. Acoust. Soc. Am., 92(3):1403–1407, 1992. [20] B. G. Ferguson and B. G. Quinn. Application of the Short-Time Fourier Transform and the Wigner-Ville Distribution to the Acoustic Localization of Aircraft. J. Acoust. Soc. Am., 96(2):821–827, 1994. [21] N. I. Fisher and P. Hall. Bootstrap Confidence Regions for Directional Data. J. Amer. Statist. Assoc., 84:996–1002, 1989. [22] N. I. Fisher and P. Hall. New Statistical Methods for Directional Data – I. Bootstrap Comparison of Mean Directions and the Fold Test in Palaeomagnet. Geophys. J. International, 101:305–313, 1990. [23] N. I. Fisher and P. Hall. Bootstrap Algorithms for Small Samples. J. Statist. Plan. Infer., 27:157–169, 1991. [24] N. I. Fisher and P. Hall. General Statistical Test for the Effect of Folding. Geophys. J. International, 105:419–427, 1991. c A.M. Zoubir
ATiSSP Sommersemester 2007
199
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
200
[25] R. A. Fisher. On the “Probable Error” of a Coefficient of Correlation Deduced from a Small Sample. Metron, 1:1–32, 1921.
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
¨ [26] J. Franke and W. Hardle. On Bootstrapping Kernel Estimates. The Ann. of Statist., 20:121–145, 1992. [27] D. A. Freedman. Bootstrapping Regression Models. The Ann. of Statist., 9:1218– 1228, 1981. [28] P. Hall. Theoretical Comparison of Bootstrap Confidence Intervals (with discussion). The Ann. of Statist., 16:927–985, 1988. [29] P. Hall. The Bootstrap and Edgeworth Expansion. Springer-Verlag New York, Inc., 1992. [30] P. Hall and D. M. Titterington. The Effect of Simulation Order on Level Accuracy and Power of Monte Carlo Tests. J. R. Statist. Soc. B, 51:459–467, 1989. [31] S. R. Hanna. Confidence Limits for Air Quality Model Evaluations, as Estimated by Bootstrap and Jackknife Resampling Methods. Atmospheric Environment, 23:1385– 1398, 1989. [32] T. Hastie and R. Tibshirani. Generalized Additive Models. Chapman & Hall, 1990. c A.M. Zoubir
ATiSSP Sommersemester 2007
200
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
201
[33] D. R. Haynor and S. D. Woods. Resampling Estimates of Precision in Emission Tomography. IEEE Transactions on Medical Imaging, 8:337–343, 1989. [34] R. Herman. Using the Bootstrap Method for Evaluating Variations Due to Sampling in Voltage Regulation Calculations. In Proceedings of the Sixth Southern African Universities Power Engineering Conference (SAUPEC-96), pages 235–238, Witwatersrand, South Africa, 1996. Univ. Witwatersrand. [35] A. O. Hero III, J. A. Fessler, and M. Usman. Exploring Estimator Bias-Variance Tradeoffs Using the Uniform CR Bound. IEEE Transactions on Signal Processing, 44(8):2026–2041, 1996. [36] M. J. Hinich. Testing for Gaussianity and Linearity of a Stationary Time Series. J. Time Series Anal., 3:169–176, 1982. [37] L. B. Jaeckel. Some Flexible Estimates of Location. Ann. Math. Statist., 42:1540– 1552, 1971. [38] K. Kanatani and N. Ohta. Optimal Robot Self-Localization and Reliability Evaluation. In Proceedings of the 5th European Conference on Computer Vision (ECCV’98), volume 2, pages 796–808, Berlin, Germany, 1998. Springer-Verlag. c A.M. Zoubir
ATiSSP Sommersemester 2007
201
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
202
[39] S. Kay. Modern Spectral Estimation. Theory and Application. Prentice Hall, 1988.
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
[40] J. P. Kreiss and J. Franke. Bootstrapping Stationary Autoregressive Moving Average Models. J. Time Ser. Anal., 13:297–319, 1992. [41] J.-P. Kreiss and G. Lien. Bootstrapping autoregressions with infinite order. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP-94), volume 6, pages 97–100, Adelaide, South Australia, April 1994. [42] E. Kreyszig. Introductory Functional Analysis with Applications. J. Wiley & Sons, 1989. [43] J. Krolik, G. Niezgoda, and D. Swingler. A bootstrap approach for evaluating source localization performance on real sensor array data. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP91), volume 2, pages 1281–1284, Toronto, Canada, May 1991. [44] H. R. Kunsch. The Jackknife and the Bootstrap for General Stationary Observations. ¨ The Ann. of Statist., 17:1217–1241, 1989. [45] D. Lai and G. Chen. Computing the Distribution of the Lyapunov Exponent from c A.M. Zoubir
ATiSSP Sommersemester 2007
202
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
203
Time Series: The One-dimensional Case Study. International Journal of Bifurcation and Chaos in Applied Sciences and Engineering, 5:1721–1726, 1995. [46] J. M. Lange, H. M. Voigt, S. Burkhardt, R. Gobel, S. Burkhardt, and R. Gobel. The Intelligent Inspection Engine—a Real-Time Real-World Visual Classifier System. In IEEE International Joint Conference on Neural Networks Proceedings, pages 1810– 1815, New York, USA, 1998. IEEE. [47] E. L. Lehmann. Testing Statistical Hypotheses. Wadsworth and Brooks, 1991. [48] R. Y. Liu and K. Singh. Moving Blocks Jackknife and Bootstrap Capture Weak Dependence. In LePage and Billard, editors, Exploring the limits of bootstrap, New York, 1992. John Wiley. [49] S. L. Marple Jr. Digital Spectral Analysis with Applications. Prentice-Hall, 1987. [50] R. G. Miller. The Jackknife - A Review. Biometrika, 61:1–15, 1974. [51] S Nagaoka and O. Amai. A Method for Establishing a Separation in Air Traffic Control Using a Radar Estimation Accuracy of Close Approach Probability. J. Japan Ins. Navigation, 82:53–60, 1990. [52] S Nagaoka and O. Amai. Estimation Accuracy of Close Approach Probability for c A.M. Zoubir
ATiSSP Sommersemester 2007
203
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
204
Establishing a Radar Separation Minimum. J. Navigation, 44:110–121, 1991.
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
[53] H. Ong, D. R. Iskander, and A. M. Zoubir. Detection of a Common Non-Gaussian Signal in Two Sensors Using The Bootstrap. In Proceedings of the IEEE Signal Processing Workshop on Higher-Order Statistics, pages 463–467, Banff, Alberta, Canada, 1997. [54] H. Ong and A. M. Zoubir. Non-Gaussian Signal Detection from Multiple Sensors using the Bootstrap. In Proceedings of the International Conference on Information, Communications & Signal Processing (ICICS’97), volume 1, pages 340–344, Singapore, September 1997. [55] E. Paparoditis. Bootstrapping Autoregressive and Moving Average Parameter Estimates of Infinite Order Vector Autoregressive Processes. J. Multivar. Anal., 57:277– 296, 1996. [56] D. N. Politis and J. P. Romano. A General Resampling Scheme for Triangular Arrays of α-Mixing Random Variables with Application to the Problem of Spectral Density Estimation. The Ann. of Statist., 20:1985–2007, 1992. [57] D. N. Politis and J. P. Romano. Bootstrap Confidence Bands for Spectra and CrossSpectra. IEEE Transactions on Signal Processing, 40:1206–1215, 1992. c A.M. Zoubir
ATiSSP Sommersemester 2007
204
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
205
[58] D. N. Politis and J. P. Romano. The Stationary Bootstrap. Journal Am. Stat. Assoc., 89:1303–1313, 1994.
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
[59] B. Porat and B. Friedlander. Performance Analysis of Parameter Estimation Algorithms Based on High-Order Moments. International J. Adaptive Control and Signal Processing, 3:191–229, 1989. [60] M. B. Priestley. Spectral Analysis and Time Series. Academic Press, 1981. [61] D. C. Reid, A. M. Zoubir, and B. Boashash. The Bootstrap Applied to Passive Acoustic Aircraft Parameter Estimation. In Proceedings of International Conference on Acoustics, Speech and Signal Processing (ICASSP-96), volume VI, pages 3153– 3156, Atlanta, May 1996. [62] D. C. Reid, A. M. Zoubir, and B. Boashash. Aircraft Parameter Estimation based on Passive Acoustic Techniques Using the Polynomial Wigner-Ville Distribution. J. Acoust. Soc. Am., 102(1):207–223, 1997. [63] M. Rosenblatt. Stationary Sequences and Random Fields. Birkhauser, Boston, 1985. [64] Special Session. The Bootstrap and Its Applications. In Proceedings of the IEEE c A.M. Zoubir
ATiSSP Sommersemester 2007
205
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
206
International Conference on Acoustics, Speech and Signal Processing (ICASSP94), volume VI, pages 65–79, Adelaide, Australia, 1994. [65] J. Shao. Bootstrap Model Selection. J. Am. Statist. Assoc., 91:655–665, 1996.
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
[66] J. Shao and D. Tu. The Jackknife and Bootstrap. Springer, 1995. [67] Q. Shao and H. Yu. Bootstrapping the Sample Means for Stationary Mixing Sequences. Stochastic Processes and Their Appl., 48:175–190, 1993. [68] R. H. Shumway. Replicated Time-Series Regression: An Approach to Signal Estimation and Detection. In D. R. Brillinger and P. R. Krishnaiah, editors, Handbook of Statistics, volume 3, pages 383–408. North-Holland, 1983. [69] T. Subba Rao and M. M. Gabr. Testing for Linearity of a Stationary Time Series. J. Time Series Anal., 1:145–158, 1980. [70] A. Swami, J. M. Mendel, and C. L. Nikias. Higher-Order Spectral Analysis Toolbox for use with MATLAB. The MathWorks, Inc., 1995. [71] L. Tauxe, N. Kylstra, and C. Constable. Bootstrap Statistics for Paleomagnetic Data. J. Geophysical Research, 96:11723–11740, 1991. [72] S. Tretter. Estimating the Frequency of a Noisy Sinusoid. IEEE Transactions on ATiSSP
c A.M. Zoubir
Sommersemester 2007
206
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
207
Information Theory, 31:832–835, 1985.
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
[73] J. K. Tugnait. Two-Channel Test for Common Non-Gaussian Signal Detection. IEE Proceedings-F, 140:343–349, 1993. [74] A. M. Yacout, S. Salvatores, and Y. G. Orechwa. Degradation Analysis Estimates of the Time-to-Failure Distribution of Irradiated Fuel Elements. Nuclear-Technology, 113(2):177–189, 1996. [75] G. A. Young. Bootstrap: More Than a Stab in the Dark? Statistical Science, 9:382– 415, 1994. [76] Y. Zhang, D. Hatzinakos, and A. N. Venetsanopoulos. Bootstrapping Techniques in the Estimation of Higher-Order Cumulants from Short Data Records . In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP-93), volume IV, pages 200–203, Minnesota, 1993. [77] A. M. Zoubir. Bootstrap: Theory and Applications. In F. T. Luk, editor, Advanced Signal Processing Algorithms, Architectures and Implementations, volume 2027, pages 216–235, San Diego, July 1993. Proceedings of SPIE. [78] A. M. Zoubir. Multiple Bootstrap Tests and Their Application. In Proceedings of c A.M. Zoubir
ATiSSP Sommersemester 2007
207
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
208
the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP-94), volume VI, pages 69–72, Adelaide, Australia, 1994. [79] A. M. Zoubir and B. Boashash. Advanced Signal Processing for Over-The-Horizon Radar: Application of Bootstrap Methods. Technical report, Signal Processing Research Centre, Queensland University of Technology, March 1996. DSTO Report ref. 96/2–Part I-IV. [80] A. M. Zoubir and B. Boashash. The Bootstrap and Its Application in Signal Processing. IEEE Signal Processing Magazine, 15(1):56–76, 1998. ¨ [81] A. M. Zoubir and J. F. Bohme. Bootstrap Multiple Hypotheses Tests for Optimizing Sensor Positions in Knock Detection. In Proceedings of 13th World Congress on Computation and Applied Mathematics, IMACS’91, pages 955–958, Dublin, 1991. ¨ [82] A. M. Zoubir and J. F. Bohme. Application of Higher Order Spectra to the Analysis and Detection of Knock in Combustion Engines. In B. Boashash, E. J. Powers, and A. M. Zoubir, editors, Higher-Order Statistical Signal Processing, pages 269–290. John Wiley, 1995. ¨ [83] A. M. Zoubir and J. F. Bohme. Multiple Bootstrap Tests: An Application to Sensor Location. IEEE Transactions on Signal Processing, 43:1386–1396, 1995. c A.M. Zoubir
ATiSSP Sommersemester 2007
208
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
209
[84] A. M. Zoubir and D. R. Iskander. A Bispectrum Based Gaussianity Test Using The Bootstrap. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP-96), volume V, pages 3029–3032, Atlanta, 1996. [85] A. M. Zoubir and D. R. Iskander. On Bootstrapping the Confidence Bands for Power Spectra. Technical report, Signal Processing Research Centre, Queensland University of Technology, January 1996. [86] A. M. Zoubir and D. R. Iskander. Bootstrap Model Selection for Polynomial Phase Signals. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP-98), pages 2229–2232, Seattle, USA, May 1998. [87] A. M. Zoubir and D. R. Iskander. Bootstrapping Bispectra: An Application to Testing for Departure from Gaussianity of Stationary Signals. IEEE Transactions on Signal Processing, March 1999. [88] A. M. Zoubir and D. R. Iskander. Bootstrap Based Modelling of a Class of NonStationary Signals. IEEE Transaction on Signal Processing, (under review). [89] I. Koutrouvelis. Regression-type estimation of the parameters of stable laws. Journal of the American Statistical Association, 75(372), 1980. c A.M. Zoubir
ATiSSP Sommersemester 2007
209
Bootstrap Techniques for Signal Processing
Cited References Signal Processing Group
Institut fur ¨ Nachrichtentechnik-Fachgebiet Signalverarbeitung
210
[90] D. Middleton. Statistical-physical models of electromagnetic interference. IEEE Transactions on Electromagnetic Compatibility, EMC-19(3):106–27, August 1977. [91] J. Shao and D. Tu. The Jackknife and Bootstrap. Springer-Verlag, 1996. [92] S.S. Wilks. Determination of sample sizes for setting tolerance limits. Annals of Mathematical Statistics, 12:91–96, 1941.
c A.M. Zoubir
ATiSSP Sommersemester 2007
210
Bootstrap Techniques for Signal Processing