Spectrum Estimation

Dr. Hassanpour
Payam Masoumi
Mariam Zabihi
Advanced Digital Signal Processing Seminar
Department of Electronic Engineering
Noushirvani University
1

Course Outlines
Introduction
Fourier Series and Transform
Time/Frequency Resolutions
Autocorrelation & spectrum estimation
Non-parametric Methods
Periodogram
Modified Periodogram
Bartletts Method
Welchs Method
Blackman-Tukey Method
Parametric Methods

Fourier Series and Transform

Im

sin( k0t )

jk0t

cos(k0t )

Re

Fourier basis functions:

real and imaginar parts of a complex sinusoid
vector representation of a complex exponential.
3

Fourier Series:
x(t )

ck e

jk0 t

1
ck
T0

T0 2

T0 2

x(t )e

jk0t

dt

k=,-1,0,1,

x(t )

c(k )
T0 n

T0 Ton Toff

T0 ff

1
T0

Fourier Transform:

x(t ) X ( f )e

j 2ft

df

x(t )

X ( f ) x(t )e j 2ft dt

X(f )

T0 ff
t

Discrete Fourier Transform (DFT)

x(0)

X ( 0)

DFT

x(1)

N 1

X (1)

X ( k ) x ( m )e

2km
N

m 0

x( N 1)

X ( N 1)

N 1

x ( m ) X ( k )e

2km
N

k ....,1,0,1,....

m 0

Autocorrelation & Spectrum estimation

Autocorrelation:

N
1
{
x(n k ) x * (n)} rx (k )

lim
N 2 N 1 n N

Power spectrum :
Px (e

jk

jk
r
(
k
)
e
x

Spectrum estimation is a problem that involves

j t
P
(
e
)from finite number of noisy
estimating x
measurements of x(n).
7

Nonparametric methods
Peroidogram
Modified periodogram
Bartlett method
Welch method
Blackman-Tukey method

The periodogram
Estimated autocorrelation:

1
rx (k )
N

N 1 k

x ( n k ) x * ( n)
n 0

Estimated power spectrum or periodogram:

jk

Pper (e )

jk

rx (k )e

The periodogram cont.

x(n)

x N (n) wR (n) x(n)

1
1

rx (k) x N (n k)x N (n) x N (k) x N ( k)

N n
N

1
1
j
j

j
j 2

Pper (e ) X N (e ) X N (e ) | X N (e ) |
N
N
10

The periodogram of white noise

x(n)
x N (n)

DFT

: white noise with a variance 2 , length N=32

X N (k )

1
| . |2
N

Pper (e j 2k

1
| X N ( k ) |2
N

11

The periodogram of white noise cont.

The estimated autocorrelation sequence

White noise power spectrum

12

Periodogram of sinusoid in noise

x(n) A sin( n0 ) v(n)

1 2
A
2

Px (e ) v

1 2
A ( 0 )
2
13

Periodogram of sinusoid in noise cont.

win(1)
win( 2)

y1 (n)

y 2 ( n)

1
| . |2
N
1
| . |2
N

z1 (n)
z 2 ( n)

P (e j1 ) z ( N 1)
X
1
PX (e j2 ) z 2 ( N 1)

x(n)

win(L)

y L (n)

1
| . |2
N

z L (n)

P (e jL ) z ( N 1)
X
L
14

Periodogram Bias
rx (k )

E{}

1
N

E{Pper (e j )} rx (k ) wB (k )e _ jk

1
E{Pper (e j )}
Px (e j ) WB (e j )
2

N k
rx (k )
rx (k )

N
n 0

N 1 k

N | k |
| k | 0

w B (k)
N
0
o.w

lim E{Pper (e j )} Px (e j )

Thus, the bias is deference between estimated and actual

Power spectrum.
15

Periodogram of sinusoid in noise cont.

x(n) A sin( n0 ) v(n)

1 2
A
4
j
2 1 2

E{Pper (e )} v A [WB (e j ( 0 ) ) WB (e j ( 0 ) )]
4

2
k 0

2
k
N

16

Example: x(n) 1sin(0.4n ) v(n)

N 128

N 512

17

Periodogram Resolution
x(n) A1 sin(n1 1 ) A1 sin( n2 2 ) v(n)
j

Px (e ) v

1 2
1
2
A1 ( 1 ) A2 ( 2 )
2
2

1 2
j
2

E{Pper (e )} v A1 [WB (e j ( 1 ) ) WB (e j ( 1 ) )]
4
1 2
A2 [WB (e j ( 2 ) ) WB (e j ( 2 ) )]
4
Set equal to the width of main lobe of the spectral window
at its half power or 6dB point.

2
j

Res[ Pper (e )] 0.89

N

18

Example:

x(n) 1sin(0.4n 1 ) 1sin(0.45n 2 ) v(n)

N 128

N 512

19

Properties of the periodogram

1
j

Pper (e )
N

Bias:

n 1

jn
x
(
n
)
w
(
n
)
e

n 0

1
E{Pper (e j )}
Px (e j ) WB (e j )
2

Resolution:
Variance:

2
Res[ Pper (e j )] 0.89
N

Var {Pper (e j )} Px2 (e j )

20

Modified Periodogram
1
1
Pper (e j ) | X N (e j ) |2
N
N

x ( n) w

(n)e jn

Would there be any benefit in replacing the rectangular window

with other windows? (for example triangular window)

1
j
j
j 2

Pper (e )
Px (e ) * WR (e ) ,
2N
1
PM (e j )
NU
1
U
N

N 1

sin( N 2) j ( N 1) 2
WR (e )
e
sin( 2)
j

x ( n) w

jn
(
n
)
e
R

1
w
(
n
)

2N
n 0

W (e

) d

21

Example:

x(n) 0.1sin(0.2n 1 ) 1sin(0.3n 2 ) v(n)

N=128
Rectangular Window

N=128
Hamming Window

22

Properties of the M-periodogram

1
PM (e j )
NU

jn
x
(
n
)
w
(
n
)
e

1
U
N

Bias:

N 1

w(n)

Variance:

n 0

E{PM (e j )}

Resolution:

1
j
j 2
Px (e ) W (e )
2NU

window dependent
Var {PM (e j )} Px2 (e j )
23

Bartletts method (periodogram averaging)

1
i
j

Pper (e )
L

L Points

L 1

jn
x
(
n
)
e
; i 1 , 2 , ..., k
i
n 0

L Points

L Points

....
x1 (n)

x2 ( n )

xk (n)

24

Properties of Bartletts method

1
j

PB (e )
N

Bias:

jn
x
(
n

iL
)
e

i 0 n 0

1
E{PB (e j )}
Px (e j ) WB (e j )
2

Resolution:
Variance:

k 1 L 1

2
Res[ PB (e j )] 0.89k
N

1 2 j
j

Var {Pper (e )} Px (e )
k
25

Example:x(n) 1sin(0.25n 1 ) 3sin(0.45n 2 ) v(n)

N 512
k 1

N 512
k 4

N 512
k 16 26

Example:x(n) 1sin(0.25n 1 ) 3sin(0.45n 2 ) v(n)

N 128
k 4

N 512
k 4

N 1024
k 4
27

Welchs method (M-periodogram averaging)

xi (n) x(n iD )

n 0 ,1,....., L 1

Overlap = L-D

....
28

Properties of Welchs method

PW (e j )

1
KLU

k 1 L 1

x(n iD)e

k 1
1
PW (e ) PM( i ) (e j )
L i 0

E{PB (e j )}

Resolution

jn

i 0 n 0

Bias

1 L 1
2
U w(n)
L n 0

1
j
j 2
Px (e ) W (e )
2LU

Window dependent
Variance

9 L 2 j
j

Var {Pper (e )}
Px (e )
16 N

with 50% overlap 29

Example:

x(n) 1sin(0.2 1 ) 3 sin(0.25 2 ) v(n)

N 512
k 4

N 512 , L 128

30
overlap 50% , hamming

Resolution:x(n) 1sin(0.2n 1 ) 3sin(0.25n 2 ) v(n)

N 512 , L 64
overlap 50% , hamming

N 512 , L 128
overlap 50% , hamming

N 512 , L 256
overlap 50% , hamming

31

windowing:x(n) 1sin(0.2n 1 ) 3sin(0.25n 2 ) v(n)

N 512 , L 128
overlap 50% , Rectangular

N 512 , L 128
overlap 50% , Bartlett 32

windowing:x(n) 1sin(0.2n 1 ) 3sin(0.25n 2 ) v(n)

N 512 , L 128
overlap 50% , Hanning

N 512 , L 128
overlap 50% , Hamming

N 512 , L 128

33

overlap 50% , Blackman

Blackman-Tukeys method (Periodogram smoothing)

Note: Bartlett & Welch are design to reduce the variance if the priodogram
by averaging and modified it.
Periodogram is computed by taking the Fourier transform of a consistent
estimate of the auto correlation sequence.
For any finite data record of length N, the variance of r (k ) will be large
x
for values of k that are close to N. for example:

rx ( N 1)

1
x( N 1) x(0)
N

at lag k n 1

In Bartlett & Welch, the variance is decreased by reducing the variance

of autocorrelation estimate by averaging.

34

Blackman-Tukeys method cont.

In the Blackman-Tukey method, the variance is decreased by applying a
window to r ( k ) in order to decrease the contribution of the unreliable
x
estimates to the periodogram.
Specifically, the Blackman-Tukey spectrum estimation is:

PBT (e )
j

jk

r
(
k
)
w
(
k
)
e
x

k M

For example, if w(k) is a rectangular window extending from M to M

with M<N-1 , then r (k ) having the largest variance are set to zero and
x
consequently, the power spectrum estimation will have a smaller variance.

35

Properties of B-Ts method

PBT (e j )

Bias

jk

r
(
k
)
w
(
k
)
e
x

k M

1
j

E{PBT (e )}
Px (e j ) W (e j )
2

Resolution

Window dependent

Variance Var {PBT (e )} P (e )

N
j

2
x

2
w
(k )
M

36

windowing: x(n) 3 sin(0.2 1 ) 1sin(0.25 2 ) v(n)

N 512 , M 128
Rectangular

N 512 , M 128
Hanning

N 512 , M 128
Bartlett

N 512 , M 128
Blackman

37

Performance comparisons
We can summarized the performance of each technique in terms of
two criteria.

(I) Variability (which is a normalized variance)

Var {Px (e j )}
2
E {Px (e j )}
(II) Figure of merit

That is approximately the same for all of the nonparametric methods

38

Summery
variability

Periodogram
Bartlett
Welch
Blackman
Tukey

1
1
k
91
8k
2M
3 N

Resolution

2
0.89
N
2
0.89k
N
2
1.28
L
2
0.64
M

Figure of merit

2
0.89
N
2
0.89
N
2
0.72
N
2
0.43
N

***50% overlap and the Bartlett window***

39