Professional Documents
Culture Documents
22
Power Spectral Density (PSD) of Random sequences
∞
−ωk
X
S(ω) = r (k )e (1)
k =−∞
23
Power Spectral Density (PSD) of Random sequences (cont’d)
This second definition of the PSD is based on the fact that, for a wide sense ergodic process :
1 N−1
∗
X
r (k ) = lim E y (n)y (n − k )
N→∞ N
n=k
24
Non parametric Spectral Estimation
Periodogram and correlogram methods
The periodogram :
2
N−1
1 X −ωn 1 2
ŜP (ω) = y (n)e = |YN (ω)| (3)
N n=0 N
The correlogram :
N−1
−ωk
X
ŜC (ω) = r̂ (k )e (4)
k =−(N−1)
Bias : How far the estimate’s mean value µθ̂ is from the true value θ.
n o
def
bias{θ̂} = E θ̂ − θ = µθ̂ − θ.
Variance : How spread the estimates θ̂ lie around their mean value µθ̂ .
n o n o
2 def 2 2 2
var{θ̂} = σθ̂ = E (θ̂ − µθ̂ ) = E θ̂ − µθ̂
Let θ̂N be an estimator of a quantity θ computed from a set of N samples of available data.
Several statistical properties of the estimator can be defined:
Unbiasedness
n o
The estimator is said to be unbiased if E θˆN = θ, i.e., if bias{θN } = 0.
n o
The estimator is asymptotically unbiased if lim E θˆN = θ, i.e., if lim bias{θˆN } = 0.
N→∞ N→∞
Consistency
h i
The estimator is said to be consistent if lim var θˆN = 0.
N→∞
The estimator is MSE consistent if lim MSE{θˆN } = 0.
N→∞
MSE consistency implies variance consistency and asymptotic unbiasedness.
N−1
1 X ∗
r̂ (k ) = y (n)y (n − k ), 0≤k ≤N−1 (5)
N − k n=k
N−1
1 X ∗
r̂ (k ) = y (n)y (n − k ), 0≤k ≤N−1 (6)
N n=k
|k |
Biased because E{r̂ (k )} = 1− N r (k )
The biased ACS estimator is preferred :
it presents reduce variance for k large (take for example k = N − 1 ...)
is guaranteed to be semi-definite positive (important when introducing the covariance matrix !)
the estimation error in r̂ (k ) is on the order of √1
N
Proof: in tutorial
Consequence: both can be analyzed simultaneously
n o n o N−1
−ωk
X
From (4) : E ŜP (ω) = E ŜC (ω) = E{r̂ (k )} e
k =−(N−1)
k
Then, from (6 ) :E{r̂ (k )} : 1− r (k ), k ≥ 0. Hence
N
N−1
|k |
n o
−ωk
X
E ŜP (ω) = 1− r (k )e
N
k =−(N−1)
which is the DTFT of a product, and hence is the convolution of the true DSP and the Triangular
window in the frequency domain.
Z π
def 1
n o
E ŜP (ω) = S(ω) ∗ WB (ω) = S(ξ)WB (ω − ξ)dξ.
2π −π
The mean ŜP (ω) : convolution of the ”Bartlett kernel” and the real DSP
N−1 2
N − |k |
X −2πfk 1 sin(ω N/2)
WB (ω) = e =
N N sin(ω/2)
k =−(N−1)
n o
E ŜP (ω) should be as close as possible to S(ω), hence, WB (ω) should be as close as possible
to a Dirac δ(ω) : OK, limN→∞ WB (ω) = δ(ω)
positive effect : narrower main lobe → narrower smearing → improved frequency resolution.
negative effect :ore prominent sidelobes → increased power leakage
DTFT of rectangular window (left) and Bartlett window (right) for N = 25 samples.
Note : interested reader, refer to Stoica’s book, section 2.4.2, for the details.
Consequences :
Variance cannot be improved by increasing the sample size N.
Inconsistency - as spectral smearing - has an adverse effect on resolvability properties.
These results will be illustrated in the computer lab.
Motivation
Periodogram: based on the FFT → simple, computationally efficient PSD estimate.
Asymptotically unbiased but inconsistent estimator: variance does not improve with observed
sample size N.
Periodogram variance mainly caused by ACS estimates with large |k |, which have large
variance due to small number of contributing samples.
To improve variance, two main families of approaches:
periodogram smoothing: Blackman-Tuckey, Daniell
periodogram averaging: Bartlett, Welch.