Introduction To Spectral Analysis Sm-Slides-1ed

Basic Denitions and The Spectral Estimation Problem
Lecture 1
Lecture notes to accompany Introduction to Spectral Analysis by P. Stoica and R. Moses, Prentice Hall, 1997
Slide L11
Informal Denition of Spectral Estimation
Given: A nite record of a signal. Determine: The distribution of signal power over frequency.
signal t t=1, 2, ... + spectral density
! = (angular) frequency in radians/(sampling interval) f = !=2 = frequency in cycles/(sampling interval)
Slide L12
Applications Temporal Spectral Analysis Vibration monitoring and fault detection Hidden periodicity nding Speech processing and audio devices Medical diagnosis Seismology and ground movement study Control systems design Radar, Sonar
Spatial Spectral Analysis Source location using sensor arrays

Lecture notes to accompany Introduction to Spectral Analysis by P. Stoica and R. Moses, Prentice Hall, 1997 Slide L13
Deterministic Signals
fy(t)g1 ;1 = t=
If:
discrete-time deterministic data sequence
t=;1
jy(t)j2 < 1
X
Then:
Y (! ) =
t=;1
y(t)e;i!t
exists and is called the Discrete-Time Fourier Transform (DTFT)
Slide L14
Energy Spectral Density
Parseval's Equality:
jy(t)j2 = 21 ; S (!)d! t=;1

X Z
where
4 S (!) = jY (!)j2 = Energy Spectral Density

We can write
S (!) =
where
k=;1
X
(k)e;i!k
(k) =
t=;1
y(t)y (t ; k)
Slide L15
Random Signals
Random Signal
random signal probabilistic statements about future variations t current observation time
Here:
t=;1
n
jy(t)j2 = 1
o
But:
E jy(t)j2 < 1
E f g = Expectation over the ensemble of realizations E jy(t)j2

n o
= Average power in y
(t)
PSD = (Average) power spectral density
Slide L16
First Denition of PSD
(! ) =
where r
k=;1
r(k)e;i!k
(k) is the autocovariance sequence (ACS) r(k) = E fy(t)y (t ; k)g r(k) = r (;k) r(0) jr(k)j
Z
Note that
r(k) = 21 ; (!)ei!k d!
n o
(Inverse DTFT)
Interpretation:
so
r(0) = E jy(t)j2 = 21 ; (!)d!

Z
(!)d! =
innitesimal signal power in the band
! d! 2
Slide L17
Second Denition of PSD
1 N y(t)e;i!t (!) = Nlim E N !1 t=1

X > :
8 > <
9 > = >
Note that
1 jY (!)j2 (!) = Nlim E N N !1

YN (!) =
X
where
is the nite DTFT of
fy(t)g.
t=1
y(t)e;i!t
Slide L18
Properties of the PSD
P1:
(!) = (! + 2 ) for all !.

Thus, we can restrict attention to
!2 ;
P2:
] () f 2 ;1=2 1=2]
(!) 0 (t) is real, Then: (! ) = (;! ) Otherwise: (! ) 6= (;! )
P3: If y
Slide L19
Transfer of PSD Through Linear Systems
System Function: where q ;1
H (q) =
k=0
hk q;k
= unit delay operator: q;1y(t) = y(t ; 1)

H (q ) y (t) 2 y (! ) = jH (! )j
e(
e(t) e (! )
Then
!)
y(t) =
k=0
hk e(t ; k) hk e;i!k
H (! ) =
y (!) = jH (!)j2 e(!)
k=0
Slide L110
The Spectral Estimation Problem
The Problem: From a sample
fy(1) : : : y(N )g
(!): f ^(!) ! 2 ; ]g
Find an estimate of
Two Main Approaches :
Nonparametric: Derived from the PSD denitions.
Parametric: Assumes a parameterized functional form of the PSD
Slide L111
Periodogram and Correlogram Methods
Lecture 2
Slide L21
Periodogram Recall 2nd denition of (! ):
1 N y(t)e;i!t (!) = Nlim E N !1 t=1

X > :
8 > <
9 > = >
Given : Drop
fy(t)gN t=1
lim
and E
N !1
f g to get
X
^p(!) = 1 y(t)e;i!t N t=1

Natural estimator
Used by Schuster ( 1900) to determine hidden periodicities (hence the name).
Slide L22
Correlogram
Recall 1st denition of (! ):
(! ) =
P
Truncate the and replace r
k=;1 N ;1 X
r(k)e;i!k
(k) by ^(k): r
r ^(k)e;i!k
^c(!) =
k=;(N ;1)
Slide L23
Covariance Estimators (or Sample Covariances)
Standard unbiased estimate:
1 N y(t)y (t ; k) k 0 ^ r(k) = N ; k t=k+1

X
Standard biased estimate:
1 N y(t)y (t ; k) k 0 r(k) = N ^ t=k+1

X
For both estimators:
r ^(k) = ^ (;k) k < 0 r
Slide L24
Relationship Between ^p (! ) and ^c(! )
If: the biased ACS estimator r Then:
^(k) is used in ^c(!),

N
^p(!) = 1 y(t)e;i!t N t=1

X
N ;1 X
= ^c(!)
k=;(N ;1)
r ^(k)e;i!k
^p(!) = ^c(!)
Consequence: Both p ! and
^( )
^c(!) can be analyzed simultaneously.

Slide L25
Statistical Performance of ^p(! ) and ^c(! ) Summary: Both are asymptotically (for large N ) unbiased:
E ^p(!) ! (!) as N ! 1
n o
Both have large variance, even for large N . Thus,
^p(!) and ^c(!) have poor performance.
Intuitive explanation:
r(k) ; r(k) may be large for large jkj ^

Even if the errors r k r k Nj;1 are small, jk =0 there are so many that when summed in ! , the PSD error is large. p!
f^( ) ; ( )g
^ ( ) ; ( )]
Slide L26
Bias Analysis of the Periodogram
E ^p(!) = E ^c(!) = E f^(k)g e;i!k r k=;(N ;1) N ;1 jkj r(k)e;i!k = 1; N k=;(N ;1) 1 = wB (k)r(k)e;i!k k=;1
n o n o X ! X
N ;1 X
wB (k) =
8 < :
=
Thus,
n o
k 1 ; jNj
jkj N ; 1 jkj N
Bartlett, or triangular, window
E ^p(!) = 21 ; ( )WB (! ; ) d
Z
Ideally:
WB (!) = Dirac impulse (!).

Slide L27
Bartlett Window WB (! )
1 sin(!N=2) 2 WB (!) = N sin(!=2)

" #
0
WB (!)=WB (0), for N = 25
10
20
dB
30
40
50
60
1 0 1 ANGULAR FREQUENCY
Main lobe 3dB width For small N , WB
1=N .
(!) may differ quite a bit from (!).

Slide L28
Smearing and Leakage
Main Lobe Width: smearing or smoothing Details in resolvable.

() smearing
(!) separated in f by less than 1=N are not

^ ()
<1/
Thus:
Periodogram resolution limit =
1=N .
^ ()W () B
Sidelobe Level: leakage

()() leakage
Slide L29
Periodogram Bias Properties
Summary of Periodogram Bias Properties: For small N , severe bias As N , WB ! !, so ! is asymptotically unbiased.
!1 ^( )
( )! ( )
Slide L210
Periodogram Variance As N
!1 E ^p(!1) ; (!1) ^p(!2) ; (!2) 2(!1) !1 = !2 = 0 !1 6= !2

nh ih (
io
Inconsistent estimate Erratic behavior
^ () ^ ()
asymptotic mean = () + 1 st. dev = () too -
Resolvability properties depend on both bias and variance.
Slide L211
Discrete Fourier Transform (DFT)
Finite DTFT: YN
(!) =
t=1
y(t)e;i!t
2 k and W = e;i 2 . N Let ! = N

Then YN
( 2 k) is the Discrete Fourier Transform (DFT): N

X
Y (k) =
t=1
y(t)W tk
k = 0 ::: N ; 1
Direct computation of O N 2 ops
( )
;1 fY (k)gN=0 from fy(t)gN : t=1 k
Slide L212
Radix2 Fast Fourier Transform (FFT)
N = 2m N=2 N Y (k) = y(t)W tk + y(t)W tk t=1 t=N=2+1 N=2 2 = y(t) + y(t + N=2)W Nk ]W tk t=1 Nk +1 for even k with W 2 = ;1 for odd k
Assume:
X X X
Let N For k
~ ~ = N=2 and W = W 2 = e;i2 =N . ~

4 = 0 2 4 : : : N ; 2 = 2p:
Y (2p) =
For k
~ N
= 1 3 5 : : : N ; 1 = 2p + 1:
X
t=1
~ ~ y(t) + y(t + N )]W tp
Y (2p + 1) =
Each is a N
~ N
~ = N=2-point DFT computation.

Slide L213
t=1
~ ~ f y(t) ; y(t + N )]W tgW tp
FFT Computation Count
Let ck Then
= number of ops for N = 2k point FFT. 2k + 2c ck = 2 k;1 k2k ) ck = 2

ck = 1 N log2 N 2
Thus,
Slide L214
Zero Padding Append the given data by zeros prior to computing DFT (or FFT):
fy(1) : : : y(N ) 0 : : : 0g
|
{z
Goals: Apply a radix-2 FFT (so N Finer sampling of
= power of 2)
^(!): 2 k N ;1 N k=0
sampled, N=8
2 k N ;1 ! N k=0
^ ()
continuous curve
Slide L215
Improved Periodogram-Based Methods
Lecture 3
Slide L31
Blackman-Tukey Method
Basic Idea: Weighted correlogram, with small weight applied to covariances r k with large k .
^( )
jj
^BT (!) =
M ;1 X k=;(M ;1)
w(k)^(k)e;i!k r
fw(k)g =
Lag Window
w(k)
1
-M
Slide L32
Blackman-Tukey Method, con't
^BT (!) = 1 ^p( )W (! ; )d 2 ;

Z
W (!) = =
Conclusion: Effect:
DTFT
fw(k)g
Spectral Window
^BT (!) = locally smoothed periodogram
Variance decreases substantially Bias increases slightly By proper choice of M : MSE
= var + bias2 ! 0 as N ! 1
Slide L33
Window Design Considerations Nonnegativeness:
^BT (!) = 1 ^p( ) W (! ; )d 2 ; 0

Z | {z }
If W
(!) 0 (, w(k) is a psd sequence) ^BT (!) 0

(which is desirable)
Then:
Time-Bandwidth Product
M ;1 X
;1) Ne = k=;(M (0) w

Z
w (k )
= equiv time width
1 2 ; W (!)d! = equiv bandwidth e= W (0)
Ne e = 1
Window Design, con't
e = 1=Ne = 0(1=M )
is the BT resolution threshold. As M increases, bias decreases and variance increases.
Choose M as a tradeoff between variance and bias.
Once M is given, Ne (and hence xed.
e) is essentially
Choose window shape to compromise between smearing (main lobe width) and leakage (sidelobe level).
The energy in the main lobe and in the sidelobes cannot be reduced simultaneously, once M is given.
Slide L35
Window Examples Triangular Window, M
= 25
10
20
dB
30
40
50
60
Rectangular Window, M
0 10
= 25
20
dB
30
40
50
60
Slide L36
Bartlett Method Basic Idea:

1 2 ...
^1() ^2()
... average
^L()
^B()
Mathematically:
yj (t) = y((j ; 1)M + t) t = 1 : : : M = the j th subsequence 4 (j = 1 : : : L = N=M ])
^j (!) = 1 yj (t)e;i!t M t=1

X
1 L ^ (!) ^B (!) = L j =1 j
X
Slide L37
Comparison of Bartlett and Blackman-Tukey Estimates
1 L ^B (!) = L j =1
X
8 > < > :
M ;1 X
'
Thus:
k=;(M ;1) M ;1 X k=;(M ;1)
k=;(M ;1) 8 M ;1 < 1 X L X

:
r ^j (k)e;i!k
9 =
9 > = >
L j =1
r ^j (k) e;i!k
r ^(k)e;i!k
^B (!) ' ^BT (!)

Since B ! implicitly uses method has
with a rectangular lag window wR k
()
^( )
fwR(k)g, the Bartlett
High resolution (little smearing) Large leakage and relatively large variance
Welch Method Similar to Bartlett method, but allow overlap of subsequences (gives more subsequences, and thus better averaging) use data window for each periodogram; gives mainlobe-sidelobe tradeoff capability
2 . . . subseq subseq #1 #2 subseq #S
Let S # of subsequences of length M . (Overlapping means S > N=M better averaging.)
])
Additional exibility: The data in each subsequence are weighted by a temporal window Welch is approximately equal to non-rectangular lag window.
^BT (!) with a

Slide L39
Daniell Method
By a previous result, for N
1,
f ^p(!j )g are (nearly) uncorrelated random variables for 2 j N ;1 ! =

N j =0 Idea: Local averaging of (2J + 1) samples in the
j
(2 + 1)
frequency domain should reduce the variance by about J .
k +J 1 ^D(!k ) = ^p (!j ) 2J + 1 j=k;J

X
Slide L310
Daniell Method, con't
As J increases: Bias increases (more smoothing) Variance decreases (more averaging)
= 2J=N . Then, for N 1, ^D(!) ' 1 ^p(!)d! 2 ; Hence: ^D (! ) ' ^BT (! ) with a rectangular spectral
Let
Z
window.
Slide L311
Summary of Periodogram Methods Unwindowed periodogram

reasonable bias unacceptable variance
Modied periodograms
Attempt to reduce the variance at the expense of (slightly) increasing the bias.
BT periodogram
Local smoothing/averaging of spectral window.
^p(!) by a suitably selected ^(k) using a lag
Implemented by truncating and weighting r window in c !
^( )
Bartlett, Welch periodograms
Approximate interpretation: BT ! with a suitable lag window (rectangular for Bartlett; more general for Welch).
^ ()
Implemented by averaging subsample periodograms.
Daniell Periodogram
Approximate interpretation: spectral window.
^BT (!) with a rectangular

Slide L312
Implemented by local averaging of periodogram values.

Parametric Methods for Rational Spectra
Lecture 4
Slide L41
Basic Idea of Parametric Spectral Estimation

Observed Data Assumed functional form of (,) Estimate parameters in (,)
Estimate PSD
^ ^ =(,) ()
possibly revise assumption on ()
Rational Spectra
jkj m k e;i!k (!) = P jkj n k e;i!k (!) is a rational function in e;i! .

P
By Weierstrass theorem, ! can approximate arbitrarily well any continuous PSD, provided m and n are chosen sufciently large. Note, however: choice of m and n is not simple some PSDs are not continuous
( )
AR, MA, and ARMA Models
By Spectral Factorization theorem, a rational factored as
(!) can be
B (!) 2 2 (!) = A(!) A(z) = 1 + a1z;1 + + anz;n B (z) = 1 + b1z;1 + + bmz;m and, e.g., A(! ) = A(z )jz =ei!
Signal Modeling Interpretation:
e(t) e (! ) =
B (q) A(q)
white noise
B (!) 2 2 A(!) ltered white noise
y(t)y (! ) =
ARMA: AR: MA:
A(q)y(t) = B (q)e(t) A(q)y(t) = e(t) y(t) = B (q)e(t)

Slide L43
ARMA Covariance Structure
ARMA signal model:
y(t)+
i=1 n
aiy(t ; i) =
j =0
bj e(t ; j )
(b0 = 1)
Multiply by y
(t ; k) and take E f g to give:

X
r (k ) +
2 m bj hj ;k = j =0 = 0 for k > m B (q) = 1 h q;k (h = 1) where H (q ) = A(q ) 0 k k=0

X X
i=1
air(k ; i) =
j =0
bj E fe(t ; j )y (t ; k)g
Slide L44
AR Signals: Yule-Walker Equations
AR: m
= 0.
2 1 a1 = 0 . . . . an 0
3 7 7 7 5 2 6 6 6 4
Writing covariance equation in matrix form for k : : : n:
=1 r(0) r(;1) : : : r(;n) . r(1) r(0) . . . . . r (;1) . r(n) : : : r(0) 1 = 2 R 0

2 6 6 6 4 7 7 7 5 " # "
32 6 6 6 4
3 7 7 7 5
These are the YuleWalker (YW) Equations.
Slide L45
AR Spectral Estimation: YW Method
Yule-Walker Method: Replace r

2 6 6 6 4
(k) by r(k) and solve for fîg and ^2: ^ a r ^(0) ^(;1) : : : ^(;n) 1 r r ^2 . . r ^(1) r(0) ^ ^.1 = 0 a . . . . . ^(;1) . . r . r ^(n) : : : r(0) ^ ^n a 0
32 7 7 7 5 6 6 6 4 3 7 7 7 5 2 6 6 6 4
3 7 7 7 5
Then the PSD estimate is
^2 ^(!) = ^ 2 jA(!)j
Slide L46
AR Spectral Estimation: LS Method
Least Squares Method:
e(t) = y(t) +
4 = y(t) + y(t) ^ where '(t) = y (t ; 1) : : : y (t ; n)]T .

Find
i=1
aiy(t ; i) = y(t) + 'T (t)
= a1 : : : an]T to minimize
f( ) =
X
This gives
2
y=6 4
y(n + 1) y(n + 2) y(N )

. . .
^ = ;(Y Y );1(Y y) where

3 7 5 2 6
t=n+1
je(t)j2
y(1) y(2) y(N ; n)
. . .
3 7 5
=4
y(n) y(n ; 1) y(n + 1) y(n)

. . .
y(N ; 1) y(N ; 2)
Slide L47
LevinsonDurbin Algorithm
Fast, order-recursive solution to YW equations

2 6 6 6 4 |
0 ;1 1 .0 .
.
..
... ...
{z
;n . ;1
. .
32 7 7 7 5 } 6 6 6 4
1
n
3 7 7 7 5
2 6 6 6 4
^ k = either r(k) or r(k).

Direct Solution: For one given value of n: For k
Rn+1
= 0 . . 0
2 n
3 7 7 7 5
O(n3) ops
= 1 : : : n: O(n4) ops
LevinsonDurbin Algorithm: Exploits the Toeplitz form of Rn+1 to obtain the solutions for k : : : n in O n2 ops!
=1
( )
Slide L48
Levinson-Durbin Alg, con't
Relevant Properties of R:
Rx = y $ Rx = y, where x = xn : : : x1]T ~ ~ ~
Nested structure
2
Rn+2 =
Thus,
2
Rn+1
n+1
r ~n
n+1 r ~n
0
3 5
r ~n =
6 4
. .
3 7 5
Rn+2 4
1 n = 0
3 5
2 4
Rn+1
n+1
r ~n
n+1 r ~n
0
32 54
1 n n = 0 0 n
3 5 2 4
3 5
where
r n = n+1 + ~n n
Slide L49
Levinson-Durbin Alg, con't

2
Rn+2 4
82 < 4 +2 :
1 n n = 0 0 n
3 5 2 4
3 5
Rn+2 4
39 = 5 2
0 ~n = 0n 2 1 n
3 5 2 4 3 5 2
3 5
Combining these gives:
Rn
0 1 n + kn ~n 0 1
3 5 2 4
n + kn n
2
0 n + kn
2 n
2 +1
3 5
0 0
Thus,
2 kn = ; n = n )
n+1 =
0 + kn 2+1 = n + kn n = n(1 ; jknj2) 2 2 n
"
"
~n 1
Computation count: k ops for the step k
2 )
n2 ops
!k+1 2 to determine f k
k gn=1 k
Slide L410
This is O
(n2) times faster than the direct solution.
MA Signals
MA: n
=0 y(t) = B (q)e(t) = e(t) + b1e(t ; 1) +
+ bme(t ; m)
Thus,
r(k) = 0 for jkj > m

and
2 2 = m r(k)e;i!k (!) = jB(!)j k=;m

X
Slide L411
MA Spectrum Estimation Two main ways to Estimate (! ): 1. Estimate
fbkg and 2 and insert them in (!) = jB(!)j2 2
nonlinear estimation problem
^(!) is guaranteed to be 0
2. Insert sample covariances
(! ) =
This is length
f^(k)g in: r
r(k)e;i!k
k=;m
^BT (!) with a rectangular lag window of 2m + 1. ^(!) is not guaranteed to be 0 =0
Both methods are special cases of ARMA methods described below, with AR model order n .
ARMA Signals
ARMA models can represent spectra with both peaks (AR part) and valleys (MA part).
A(q)y(t) = B (q)e(t) B (!) 2 = m=;m k e;i!k k (!) = 2 A(!) jA(!)j2

P
where
= E f B(q)e(t)] B(q)e(t ; k)] g = E f A(q)y(t)] A(q)y(t ; k)] g =

X
j =0 p=0
aj ap r(k + p ; j )
Slide L413
ARMA Spectrum Estimation
Two Methods:
2g in (!) = 2 B(!) 2 1. Estimate fai bj A(!)

nonlinear estimation problem; can use an approximate linear two-stage least squares method
^(!) is guaranteed to be 0
2. Estimate
fai r(k)g in (!) =
m ;i!k k=;m k e
jA(!)j2
linear estimation problem (the Modied Yule-Walker method).
^(!) is not guaranteed to be 0

Two-Stage Least-Squares Method Assumption: The ARMA model is invertible:
A e(t) = B(q) y(t) (q ) = y(t) + 1y(t ; 1) + 2y(t ; 2) + = AR(1) with j kj ! 0 as k ! 1 Step 1: Approximate, for some large K e(t) ' y(t) + 1y(t ; 1) + + K y(t ; K )
1a) Estimate the coefcients modelling techniques.
f kgK=1 by using AR k
1b) Estimate the noise sequence
^(t) = y(t) + ^1y(t ; 1) + e

and its variance
+ ^K y(t ; K )
^2 =
N 1 j^(t)j2 e N ; K t=K +1
X
Slide L415
Two-Stage Least-Squares Method, con't
fe(t)g by ^(t) in the ARMA equation, e A(q)y(t) ' B (q)^(t) e and obtain estimates of fai bj g by applying least squares
Step 2: Replace techniques. Note that the ai and bj coefcients enter linearly in the above equation:
y(t) ; ^(t) ' ;y(t ; 1) : : : ; y(t ; n) e ^(t ; 1) : : : ^(t ; m)] e e = a1 : : : an b1 : : : bm]T
Slide L416
Modied Yule-Walker Method
ARMA Covariance Equation:
r(k) +
2 6 4
In matrix form for k
i=1
air(k ; i) = 0 k > m
3 7 5 "
r(m) r(m + 1)
. . .
: : : r(m ; n + 1) r(m ; n + 2)
... . . .
= m + 1 ::: m + M
a1 an
. . .
#
r(m + M ; 1) : : : r(m ; n + M )
;6 4
Replace
fr(k)g by f^(k)g and solve for faig. r f^ g
. 5 . . r(m + M )
r(m + 1) r(m + 2)
3 7
If M n, fast Levinson-type algorithms exist for obtaining ai . If M > n overdetermined squares solution for ai .
f^ g
Y W system of equations; least
f^ g
Note: For narrowband ARMA signals, the accuracy of ai is often better for M > n
Slide L417
Summary of Parametric Methods for Rational Spectra
Method AR: YW or LS low low-medium medium-high medium-high medium low-medium No No Yes
Computational Burden low Accuracy medium
Guarantee ^(! ) 0 ? Yes
MA: BT
ARMA: MYW
ARMA: 2-Stage LS
Use for Spectra with (narrow) peaks but no valley Broadband spectra possibly with valleys but no peaks Spectra with both peaks and (not too deep) valleys As above
Slide L418
Parametric Methods for Line Spectra Part 1
Lecture 5
Slide L51
Line Spectra Many applications have signals with (near) sinusoidal components. Examples: communications radar, sonar geophysical seismology ARMA model is a poor approximation Better approximation by Discrete/Line Spectrum Models
2 6(2 1) 2
(!)
6
(2 2) 2
(2 2) 3
!
Slide L52
; !1
!2
!3
An Ideal line spectrum

Line Spectral Signal Model Signal Model: Sinusoidal components of frequencies !k and powers 2 , superimposed in white noise of k 2. power
f g
f g
y(t) = x(t) + e(t) t = 1 2 : : : n x(t) = k ei(!k t+ k ) k=1 xk (t)

X | {z }
Assumptions: A1: A2:
k>0
!k 2 ;
(prevents model ambiguities)
f'k g = independent rv's, uniformly distributed on ; ]

(realistic and mathematically convenient)
A3:
e(t) = circular white noise with variance 2 E fe(t)e (s)g = 2 t s E fe(t)e(s)g = 0

(can be achieved by slow sampling)
Slide L53
Covariance Function and PSD
Note that:
E E
i'p e;i'j o = 1, for p = j e i'p e;i'j o = E nei'p o E ne;i'j o e

Z
1 i' d' 2 = 0, for p 6= j =2 ; e

E xp(t)xj (t ; k) = 2 ei!pk p j p
n o
Hence,
r(k) = E fy(t)y (t ; k)g = n=1 2ei!pk + 2 k 0 p p

P
and
(! ) = 2
p=1
2 (! ; !p) + 2 p
Slide L54
Parameter Estimation Estimate either:
f!k
k 'k gn=1 k
(Signal Model)
f!k 2gn=1 2 k k
Once
(PSD Model)
^ Major Estimation Problem: f!k g
f!kg are determined: ^

X
f ^2g can be obtained by a least squares method from k n 2ei!pk + residuals ^ ^ r(k) = p
p=1
OR: Both k and 'k can be derived by a least squares method from
f^ g
f^ g
X
y(t) =
with
k=
k=1 k ei'k .
^ ei!k t + residuals k
Slide L55
Nonlinear Least Squares (NLS) Method
min' g f!
Let:
k k k t=1 |
y(t) ;
k=1 {z F (! ')
k ei(!k t+'k )
2
}
= kei'k = 1 : : : n]T Y = y(1) : : : y(N )]T ei!1 ei!n . . . . B = eiN!1 eiN!n

k
2 6 4
3 7 5
Slide L56
Nonlinear Least Squares (NLS) Method, con't
Then:
F = (Y ; B ) (Y ; B ) = kY ; B k2 = ; (B B);1B Y ] B B] ; (B B);1B Y ] +Y Y ; Y B(B B);1B Y

This gives:
^ = (B B);1B Y !=^ !
and
! = arg max Y B (B B );1B Y ^ !
Slide L57
NLS Properties
Excellent Accuracy:
6 2 (for N var (^k ) = 3 2 ! N k Example: N = 300 SNRk = 2= 2 = 30 dB k

q
1)
Then
var
(^k ) 10;5. !
Difcult Implementation: The NLS cost function F is multimodal; it is difcult to avoid convergence to local minima.
Slide L58
Unwindowed Periodogram as an Approximate NLS Method
For a single (complex) sinusoid, the maximum of the unwindowed periodogram is the NLS frequency estimate: Assume: Then:
n=1
X
B B=N
N t=1
BY=
y(t)e;i!t = Y (!)
(nite DTFT)
1 Y B (B B );1B Y = N jY (!)j2 = ^p(!) = (Unwindowed Periodogram)

So, with no approximation,
! = arg max ^p(!) ^ !
Slide L59
Unwindowed Periodogram as an Approximate NLS Method, con't
Assume: Then:
n>1
f!kgn=1 ' the ^ k

provided that
locations of the peaks of p !
^( )
largest
inf j!k ; !pj > 2 =N

which is the periodogram resolution limit. If better resolution desired then use a High/Super Resolution method.
Slide L510
High-Order Yule-Walker Method Recall:
y(t) = x(t) + e(t) =
k ei(!k t+'k ) +e(t) | {z } k=1 xk (t)

X
Degenerate ARMA equation for y (t):

n
(1 ; ei!k q;1)xk (t) = k ei(!kt+'k) ; ei!k ei !k(t;1)+'k] = 0

o
Let
B (q) = 1 +
4 bk q;k = A(q)A(q) k=1 A(q) = (1 ; ei!1 q;1) (1 ; ei!n q;1) A(q) = arbitrary
X
Then B
(q)x(t) 0 ) B (q)y(t) = B (q)e(t)

Slide L511
High-Order Yule-Walker Method, con't
Estimation Procedure: Estimate
fîgL=1 using an ARMA MYW technique b i

^( )
Roots of B q give spurious roots.
f!kgn=1, along with L ; n ^ k
Slide L512
High-Order and Overdetermined YW Equations
ARMA covariance:
r (k ) +
i=1
bir(k ; i) = 0 k > L
In matrix form for k

2 6 6 6 4 |
= L + 1 ::: L + M
3 7 7 7 5 2 6 6 6 4 3 7 7 7 5 }
: : : r(1) r(L + 1) : : : r(2) b = ; r(L + 2) . . . . . . r(L + M ; 1) : : : r(M ) r(L + M ) 4 4 = =

{z } | {z
r(L) r(L + 1)
This is a high-order (if L > n) and overdetermined (if M > L) system of YW equations.
Slide L513
High-Order and Overdetermined YW Equations, con't Fact: SVD of rank :
( )=n =U V
U = (M n) with U U = In V = (n L) with V V = In
= (n n), diagonal and nonsingular

Thus,
(U V )b = ;
The Minimum-Norm solution is
b = ; y = ;V ;1U Important property: The additional (L ; n) spurious zeros of B (q ) are located strictly inside the unit circle, if the Minimum-Norm solution b is used.
HOYW Equations, Practical Solution
Let
^= ^ ^ ^ ^
but made from
f^(k)g instead of fr(k)g. r

, V from the
Let U , , V be dened similarly to U , SVD of . Compute
^ = ;V ^ ;1U ^ ^ ^ b ^(q) that
Then !k n=1 are found from the n zeroes of B k are closest to the unit circle.
f^ g
When the SNR is low, this approach may give spurious frequency estimates when L > n; this is the price paid for increased accuracy when L > n.
Slide L515
Parametric Methods for Line Spectra Part 2
Lecture 6
Slide L61
The Covariance Matrix Equation Let:
a(!) = 1 e;i! : : : e;i(m;1)! ]T A = a(!1) : : : a(!n)] (m n)

rank
Note: Dene
(A) = n
y(t) y(t ; 1)
(for m
3 7 7 7 5
n)
46 y(t) = 6 ~
6 4
y(t ; m + 1)
. .
= Ax(t) + ~(t) ~ e
where
x(t) = x1(t) : : : xn(t)]T ~ ~(t) = e(t) : : : e(t ; m + 1)]T e

4 R = E fy(t)~ (t)g = APA + 2I ~ y
Then
with
2
P = E fx(t)~ (t)g = ~ x
6 4
2 1
...
0
2 n
3 7 5
Slide L62
Eigendecomposition of R and Its Properties
R = APA + 2I (m > n)
Let:
2 :::
m: eigenvalues of R
fs1 : : : sng: orthonormal eigenvectors associated with f 1 : : : ng fg1 : : : gm;ng: orthonormal eigenvectors associated with f n+1 : : : m g S = s1 : : : sn] (m n) G = g1 : : : gm;n] (m (m ; n))
Thus,
2
R = S G]
6 4
"
...
7 5
S G
Slide L63
Eigendecomposition of R and Its Properties, con't As rank
(APA ) = n: k > 2 k = 1 ::: n k = 2 k = n + 1 ::: m 0 1; 2 . .. = = nonsingular 0 n; 2

2 6 4 3 7 5 2 6 4
Note:
RS = APA S + 2S = S
...
0
n
3 7 5
with C (since rank S , Therefore, since S G
j j 6= 0
4 ;1) = AC S = A(PA S
( ) = rank(A) = n). =0 A G=0

Slide L64
MUSIC Method
a (!1) . . A G= G=0 a (!n) ) fa(!k )gn=1 ? R(G) k

2 6 4 3 7 5
Thus,
f!kgn=1 are the unique solutions of k a (!)GG a(!) = 0.

Let:
1 N y(t)~ (t) ^ ~ y R = N t=m ^^ S G = S G made from the ^ eigenvectors of R

X
Slide L65
Spectral and Root MUSIC Methods Spectral MUSIC Method:
f!kgn=1 = the locations of the n highest peaks of the ^ k

pseudo-spectrum function:
1 ^G a(!) ! 2 ; a (!)G ^
Root MUSIC Method:
f!kgn=1 = the angular positions of the n roots of: ^ k

^^ aT (z;1)GG a(z) = 0
that are closest to the unit circle. Here,
a(z) = 1 z;1 : : : z;(m;1)]T

Note: Both variants of MUSIC may produce spurious frequency estimates.
Pisarenko Method
Pisarenko is a special case of MUSIC with m (the minimum possible value). If:
=n+1
m=n+1
Then:
^ g G = ^1, ) f!kgn=1 can be found from the roots of ^ k aT (z;1)^1 = 0 g

no problem with spurious frequency estimates computationally simple (much) less accurate than MUSIC with m
n+1
Slide L67
Min-Norm Method
Goals: Reduce computational burden, and reduce risk of false frequency estimates. Uses m n (as in MUSIC), but only one vector in G (as in Pisarenko).
R( )
Let
"
1 = the vector in R(G), with rst element equal ^ ^ to one, that has minimum Euclidean norm. g
#
Slide L68
Min-Norm Method, con't
Spectral Min-Norm
f!gn=1 = ^k
the locations of the n highest peaks in the pseudo-spectrum
2 1 = a (!) 1 ^ g
" #
Root Min-Norm
f!gn=1 = ^k
the angular positions of the n roots of the polynomial
aT (z;1) 1 ^ g
"
that are closest to the unit circle.
Slide L69
Let S
^= S g1 ;1 gm
" # " #
Min-Norm Method: Determining g ^
Then:
1 2 R(G) ) S 1 = 0 ^ ^ ^ ^ g g ) S ^= ; g
" #
Min-Norm solution: As:
^ = ;S(S S );1 g
+ S S , (S S );1 exists iff = k k2 6= 1 (This holds, at least, for N 1.) ^^ I=S S=

Multiplying the above equation by gives:
(1 ; k k2) = (S S ) ) (S S );1 = =(1 ; k k2) ) ^ = ;S =(1 ; k k2) g

ESPRIT Method
Let
A1 = Im;1 0]A A2 = 0 Im;1]A
Then A2
Also, let
= A1D, where e;i!1 0 ... D= 0 e;i!n S1 = Im;1 0]S S2 = 0 Im;1]S

2 6 4 |
3 7 5
Recall S
= AC with jC j 6= 0. Then S2 = A2C = A1DC = S1 C ;1DC

{z
So has the same eigenvalues as D . determined as
is uniquely
= (S1S1);1S1S2
ESPRIT Implementation From the eigendecomposition of R, nd S , then S1 and S2 .
The frequency estimates are found by:
where
f^kgn=1 are the eigenvalues of k ^ = (S1S1);1S1S2 ^^ ^^
f!kgn=1 = ; arg(^k ) ^ k
ESPRIT Advantages: computationally simple no extraneous frequency estimates (unlike in MUSIC or MinNorm) accurate frequency estimates
Slide L612
Summary of Frequency Estimation Methods
Method Periodogram Nonlinear LS Yule-Walker Pisarenko MUSIC Min-Norm ESPRIT
Computational Accuracy / Risk for False Burden Resolution Freq Estimates small medium-high medium very high very high very high medium high medium small low none high high medium medium high small medium very high none
Recommendation:
Use Periodogram for medium-resolution applications
Use ESPRIT for high-resolution applications

Slide L613
Filter Bank Methods
Lecture 7
Slide L71
Basic Ideas
Two main PSD estimation approaches: 1. Parametric Approach: Parameterize nite-dimensional model.
(!) by a
2. Nonparametric Approach: Implicitly smooth ! !=; by assuming that ! is nearly constant over the bands
f ( )g
( )
!;
!+ ] N >1
2 is more general than 1, but 2 requires
to ensure that the number of estimated values = = ) is < N . (
=2 2 =1
N > 1 leads to the variability / resolution compromise

associated with all nonparametric methods.
Filter Bank Approach to Spectral Estimation

H 6 (!)
1
!c
-!
Filtered Signal
y (t)
- with varying !
Power Calculation
Bandpass Filter c and xed bandwidth
Power in the band
Division by lter bandwidth
^(!c )
6
y
(! )
;
!c
@Q Q -!
;
Z
6 ~ (! ) @ -!
(b)
Z
!c
^FB (!) ' 1 2

(a)
jH ( )j2 ( )d = ' 21
!+ !;
c ( )d = (') (!)
(a) consistent power calculation
(b) Ideal passband lter with bandwidth
(c)
( ) constant on 2 ! ; 2 ! + 2 ]
Note that assumptions (a) and (b), as well as (b) and (c), are conicting.
Slide L73
Filter Bank Interpretation of the Periodogram
1 N y(t)e;i!t 2 ~ ^p(~) = ! 4 N t=1 2 N 1 ~ = N y(t)ei!(N ;t) t=1

X X
= N
where
(
k=0
hk y(N ; k)
! ;i!k = 1 eiN (~;!) ; 1 H (!) = hk e ! N ei(~;!) ; 1 k=0

X
1 ei!k k = 0 : : : N ; 1 ~ hk = N otherwise 0
1
center frequency of H 3dB bandwidth of H
(!) = ! ~
(!) ' 1=N

Slide L74
Filter Bank Interpretation of the Periodogram, con't
jH (!)j as a function of (~ ; !), for N = 50. !

0 5 10
15 dB
20
25
30
35
40
Conclusion: The periodogram p ! is a lter bank PSD estimator with bandpass lter as given above, and: narrow lter passband, power calculation from only 1 sample of lter output.
^( )
Slide L75
Possible Improvements to the Filter Bank Approach 1. Split the available sample, and bandpass lter each subsample. more data points for the power calculation stage. This approach leads to Bartlett and Welch methods. 2. Use several bandpass lters on the whole sample. Each lter covers a small band centered on ! .
provides several samples for power calculation. This multiwindow approach is similar to the Daniell method.
Both approaches compromise bias for variance, and in fact are quite related to each other: splitting the data sample can be interpreted as a special form of windowing or ltering.
Slide L76
Capon Method
Idea: Data-dependent bandpass lter design.
yF (t) =
k=0
|
hk y(t ; k)
2 6 4 |
. = h0 h1 : : : hm] . y(t ; m) h y(t) ~ E jyF (t)j2 = h Rh R = E fy(t)~ (t)g ~ y

{z } 7 5 {z } n o
y(t)
H (! ) =
where a
(! ) = 1
k=0 e;i! : : : e;im! ]T
hk e;i!k = h a(!)
Slide L77
Capon Method, con't Capon Filter Design Problem:
min(h Rh) h
Solution:
subject to h
a(!) = 1
h0 = R;1a=a R;1a
o
The power at the lter output is:
E jyF (t)j2 = h0Rh0 = 1=a (!)R;1a(!) which should be the power of y (t) in a passband centered on ! .
n
The Bandwidth
1 ' m+1 = (lter 1 length)
Conclusion Estimate PSD as:
+1 ^(!) = m^;1 a (!)R a(!)

with
N 1 ^ y(t)~ (t) ~ y R = N ;m t=m+1

X
Slide L78
Capon Properties
m is the user parameter that controls the compromise

between bias and variance: as m increases, bias decreases and variance increases.
Capon uses one bandpass lter only, but it splits the N -data point sample into (N m) subsequences of length m with maximum overlap.
Slide L79
Relation between Capon and Blackman-Tukey Methods
Consider
^BT (!) with Bartlett window: m m + 1 ; jkj ^BT (!) = r(k)e;i!k ^ k=;m m + 1 1 m m ^(t ; s)e;i!(t;s) = m+1 r t=0 s=0 ^ a (!)Ra(!) R = ^(i ; j )] ^ r = m+1
X X X
Then we have
^ a (!)Ra(!) ^BT (!) = m+1 m+1 ^C (!) = ^ a (!)R;1a(!)
Slide L710
Relation between Capon and AR Methods Let
2 ^k ^k (!) = ^ 2 jAk(!)j be the kth order AR PSD estimate of y (t).

AR Then
^C (!) =
1 1 m 1= ÂR(!) m + 1 k=0 k
X
Consequences: Due to the average over k, C ! generally has less statistical variability than the AR PSD estimator. Due to the low-order AR terms in the average, C ! generally has worse resolution and bias properties than the AR method.
^( )
^( )
Slide L711
Spatial Methods Part 1
Lecture 8
Slide L81
The Spatial Spectral Estimation Problem

Source 1
vSource 2
B B B B N B cccc ccc c c c c c
@ ,,, @, ,,@ , ,R , , ,@ ,
v Source n
@; @;
Sensor 1
@; @; @; @;
Sensor 2 Sensor
Problem: Detect and locate n radiating sources by using an array of m passive sensors. Emitted energy: Acoustic, electromagnetic, mechanical Receiving sensors: Hydrophones, antennas, seismometers Applications: Radar, sonar, communications, seismology, underwater surveillance Basic Approach: Determine energy distribution over space (thus the name spatial spectral analysis)
Simplifying Assumptions
Far-eld sources in the same plane as the array of sensors Non-dispersive wave propagation
Hence: The waves are planar and the only location parameter is direction of arrival (DOA) (or angle of arrival, AOA). The number of sources n is known. (We do not treat the detection problem) The sensors are linear dynamic elements with known transfer characteristics and known locations (That is, the array is calibrated.)
Slide L83
Array Model Single Emitter Case
x(t) =
k=
the signal waveform as measured at a reference point (e.g., at the rst sensor) the delay between the reference point and the kth sensor the impulse response (weighting function) of sensor k noise at the kth sensor (e.g., thermal noise in sensor electronics; background noise, etc.)
hk (t) = ek (t) =
Note:
t 2 R (continuous-time signals). Then the output of sensor k is yk (t) = hk (t) x(t ; k) + ek (t) ( = convolution operator).
Basic Problem: Estimate the time delays known but x t unknown.
()
f kg with hk(t)
This is a time-delay estimation problem in the unknown input case.

Narrowband Assumption
Assume: The emitted signals are narrowband with known carrier frequency !c . Then:
x(t) = (t) cos !ct + '(t)] where (t) '(t) vary slowly enough so that (t ; k) ' (t) '(t ; k ) ' '(t) Time delay is now ' to a phase shift !c k : x(t ; k ) ' (t) cos !ct + '(t) ; !c k ] hk (t) x(t ; k ) ' jHk (!c)j (t)cos !ct + '(t) ; !c k + argfHk (!c)g]
where Hk function
(!) = Ffhk(t)g is the kth sensor's transfer
Hence, the kth sensor output is
yk (t) = jHk (!c)j (t) cos !ct + '(t) ; !c k + arg Hk (!c)] + ek(t)
Slide L85
Complex Signal Representation The noise-free output has the form:
z(t) = (t) cos !ct + (t)] = = (t) ei !ct+ (t)] + e;i !ct+ (t)] 2 Demodulate z (t) (translate to baseband): 2z(t)e;!ct = (t)f ei (t) + e;i 2!ct+ (t)] g
n o | {z } | {z }
Lowpass lter
2z(t)e;i!ct to obtain (t)ei (t)
lowpass
highpass
Hence, by low-pass ltering and sampling the signal
yk (t)=2 = yk(t)e;i!ct ~ = yk(t) cos(!ct) ; iyk(t) sin(!ct) we get the complex representation: (for t 2 Z ) yk (t) = (t) ei'(t) jHk (!c)j ei arg H (! )] e;i! + ek (t)
|
s(t)
{z
} |
Hk (!c)
{z
c k
yk (t) = s(t)Hk (!c) e;i!c k + ek (t) where s(t) is the complex envelope of x(t).
or
Vector Representation for a Narrowband Source
Let
= m =
a( ) =
2 6 4
the emitter DOA the number of sensors

2 6 4
H1(!c) e;i!c
. .
3 7 5
Hm(!c) e;i!c m y1(t) e1(t) . . y(t) = e(t) = . . ym(t) em(t)

3 7 5 2 6 4
3 7 5
Then
y(t) = a( )s(t) + e(t)

NOTE: enters a via both For omnidirectional sensors the on .
()
f kg and fHk(!c)g. fHk(!c)g do not depend
Slide L87
Analog Processing Block Diagram
Analog processing for each receiving array element

SENSOR DEMODULATION
A/D CONVERSION
Internal Noise External Noise XXXXX ? XX ll z l - Cabling and , Filtering : ,, Incoming Transducer Signal x(t ; k )
Re yk (t)] Lowpass Re yk (t)] - !! ~ Filter A/D cos(!ct) 6 y k (t) Oscillator -sin(!ct) ? ~ - Im yk (t)] Lowpass Im yk (t)] - !! Filter A/D
Slide L88
Multiple Emitter Case Given n emitters with received signals: DOAs:
fsk(t)gn=1 k
Linear sensors
) y(t) = a( 1)s1(t) +
+ a( n)sn(t) + e(t)
Let
A = a( 1) : : : a( n)] (m n) s(t) = s1(t) : : : sn(t)]T (n 1)

Then, the array equation is:
y(t) = As(t) + e(t)

Use the planar wave assumption to nd the dependence of k on .
Uniform Linear Arrays
Source h
JJ J J
J J
JJ J ^ J
JJ
JJ+ JJ
JJ
2
] J d sin J ^ JJ J JJ 3 J J
v
4
v ]? v
d -
ppp
vSensor
ULA Geometry Sensor #1 = time delay reference Time Delay for sensor k:
where c = wave propagation speed
d sin k = (k ; 1) c
Slide L810
Spatial Frequency
Let:
sin 4 !s = !c d sin = 2 dc=f = 2 d sin c c = c=fc = signal wavelength a( ) = 1 e;i!s : : : e;i(m;1)!s ]T By direct analogy with the vector a(! ) made from
uniform samples of a sinusoidal time series,
!s = spatial frequency The function !s 7! a( ) is one-to-one for j!sj $ dj sin j 1 d =2

As
=2
=2 is a spatial Shannon sampling theorem.

Slide L811
d = spatial sampling period
Spatial Methods Part 2
Lecture 9
Slide L91
Spatial Filtering
Spatial ltering useful for DOA discrimination (similar to frequency discrimination of time-series ltering) Nonparametric DOA estimation
There is a strong analogy between temporal ltering and spatial ltering.
Slide L92
Analogy between Temporal and Spatial Filtering
Temporal FIR Filter:
yF (t) = hk u(t ; k) = h y(t) k=0 h = ho : : : hm;1] y(t) = u(t) : : : u(t ; m + 1)]T

If u
m;1 X
(t) = ei!t then yF (t) = h a(!)] u(t)

| {z }
lter transfer function
a(!) = 1 e;i! : : : e;i(m;1)! ]T We can select h to enhance or attenuate signals with different frequencies ! .
Slide L93
Analogy between Temporal and Spatial Filtering Spatial Filter:
fyk(t)gm=1 = k
the spatial samples obtained with a sensor array.
Spatial FIR Filter output:
yF (t) =
k=1
hk yk (t) = h y(t)
Narrowband Wavefront: The array's (noise-free) response to a narrowband ( sinusoidal) wavefront with complex envelope s t is:
() y(t) = a( )s(t) a( ) = 1 e;i!c 2 : : : e;i!c m ]T

yF (t) = h a( )] s(t)
| {z }
The corresponding lter output is
lter transfer function
We can select h to enhance or attenuate signals coming from different DOAs.

Analogy between Temporal and Spatial Filtering

u(t) = ei!t
z
1
s q; s qq
h
-
?
1
m;1
q ;1
? -
1 - @ 0 B 1 C h @@ C B B e;i! C @ h a(!)]u(t) R @ C - ? B -P C B C B B B .. Cu(t) C B . C C B C B C B C B C B A @ ;i m; ! hm; | e {z } ? a(!)

1
qq
1)
(Temporal sampling)
(a) Temporal lter

narrowband source with DOA=
Z
Z ~ aa 1reference !!
0 1 B aa 2- B ;i! Be 2 B !! B B B B .. B . B B @ ;i! m | e {z a( ) aa m!
point
( )
qq
( )
1 C C C C C Cs(t) C C C C C A }
? @ ? @ @
h
-
@ R @ -
h a( )]s(t) -
hm;
?
qq
(Spatial sampling)
(b) Spatial lter

Spatial Filtering, con't
Example: The response magnitude h a of a spatial lter (or beamformer) for a 10-element ULA. Here, h a 0 , where 0
j ( )j
= ( )
= 25
30
30
Th et a (d eg )
60
60
90 0 2 4 6 Magnitude 8 10
90
Slide L96
Spatial Filtering Uses
Spatial Filters can be used To pass the signal of interest only, hence ltering out interferences located outside the lter's beam (but possibly having the same temporal characteristics as the signal). To locate an emitter in the eld of view, by sweeping the lter through the DOA range of interest (goniometer).
Slide L97
Nonparametric Spatial Methods
A Filter Bank Approach to DOA estimation. Basic Ideas Design a lter h
( ) such that for each
It passes undistorted the signal with DOA = It attenuates all DOAs
6=
Sweep the lter through the DOA range of interest, and evaluate the powers of the ltered signals:
E jyF (t)j2 = E jh ( )y(t)j2 = h ( )Rh( ) with R = E fy (t)y (t)g.

n o n
The (dominant) peaks of h DOAs of the sources.
( )Rh( ) give the

Slide L98
Beamforming Method
Assume the array output is spatially white:
Then:
R = E fy(t)y (t)g = I E jyF (t)j2 = h h

n o
Hence: In direct analogy with the temporally white assumption for lter bank methods, y t can be considered as impinging on the array from all DOAs.
()
Filter Design:
min (h h) h
Solution:
subject to
h a( ) = 1
h = a( )=a ( )a( ) = a( )=m E jyF (t)j2 = a ( )Ra( )=m2

n o
Slide L99
Implementation of Beamforming
1 N y(t)y (t) ^ R=N t=1

X
The beamforming DOA estimates are:
f^kg =
the locations of the n largest peaks of a Ra .
( )^ ( )
This is the direct spatial analog of the Blackman-Tukey periodogram. Resolution Threshold:
inf j k ; pj > =
wavelength array length array beamwidth
Inconsistency problem: Beamforming DOA estimates are consistent if n inconsistent if n > .
= 1, but
Slide L910
Capon Method Filter design:
min(h Rh) subject to h a( ) = 1 h

Solution:
h = R;1a( )=a ( )R;1a( ) E jyF (t)j2 = 1=a ( )R;1a( )

n o
Implementation:
f^kg =
the locations of the n largest peaks of
^ 1=a ( )R;1a( ):
Performance: Slightly superior to Beamforming. Both Beamforming and Capon are nonparametric approaches. They do not make assumptions on the covariance properties of the data (and hence do not depend on them).
Parametric Methods Assumptions: The array is described by the equation:
y(t) = As(t) + e(t)

The noise is spatially white and has the same power in all sensors:
E fe(t)e (t)g = 2I
The signal covariance matrix
P = E fs(t)s (t)g
is nonsingular.
Then:
R = E fy(t)y (t)g = APA + 2I

Thus: The NLS, YW, MUSIC, MIN-NORM and ESPRIT methods of frequency estimation can be used, almost without modication, for DOA estimation.
Nonlinear Least Squares Method
1 N ky(t) ; As(t)k2 min(t)g N f k g fs t=1

X |
Minimizing f over s gives
f ( s)
{z
^(t) = (A A);1A y(t) t = 1 : : : N s

X
Then
1 N k I ; A(A A);1A ]y(t)k2 f ( ^) = N s t=1 1 N y (t) I ; A(A A);1A ]y(t) = N t=1 ^ = trf I ; A(A A);1A ]Rg ^ Thus, f ^k g = arg max trf A(A A);1A ]Rg
X
f kg
For N , this is precisely the form of the NLS method of frequency estimation.
=1
Nonlinear Least Squares Method
Properties of NLS: Performance: high Computational complexity: high Main drawback: need for multidimensional search.
Slide L914
Yule-Walker Method
e y(t) = y(t) = A s(t) + ~(t) ~ y(t) ~ e(t) A

" # " # "
Assume: Then:
E fe(t)~ (t)g = 0 e
4 E fy(t)~ (t)g = AP A ~ ;= y
(M L)
Also assume:
M > n L > n ( ) m = M + L > 2n)

rank
~ (A) = rank(A) = n
" # " #
(;) = n, and the SVD of ; is ; = U1 U2 ] 0 n n 0 V1 g n ;n 0 V2 g L n M ;n ~ ~ Properties: A V2 = 0 V1 2 R(A)

Then: rank
|{z} |{z}
Slide L915
YW-MUSIC DOA Estimator
f^kg =
where
the n largest peaks of
1=~ ( )V2V2 ~( ) a ^â
~( ), (L 1), is the array transfer vector for y(t) a ~

at DOA
^ V2 is dened similarly to V2, using 1 N y(t)~ (t) ^ ;=N y t=1

X
Properties: Computational complexity: medium Performance: satisfactory if m
2n
Main advantages: weak assumption on e t the subarray A need not be calibrated
f ( )g
Slide L916
MUSIC and Min-Norm Methods
Both MUSIC and Min-Norm methods for frequency estimation apply with only minor modications to the DOA estimation problem. Spectral forms of MUSIC and Min-Norm can be used for arbitrary arrays Root forms can be used only with ULAs MUSIC and Min-Norm break down if the source signals are coherent; that is, if rank
(P ) = rank(E fs(t)s (t)g) < n
Modications that apply in the coherent case exist.
Slide L917
ESPRIT Method Assumption: The array is made from two identical subarrays separated by a known displacement vector. Let
m = # sensors in each subarray A1 = Im 0]A (transfer matrix of subarray 1) A2 = 0 Im]A (transfer matrix of subarray 2)
Then
A2 = A1D, where e;i!c ( 1) 0 ... D= 0 e;i!c ( n) ( ) = the time delay from subarray 1 to
2 6 4
3 7 5
subarray 2 for a signal with DOA = : where d is the subarray separation and is measured from the perpendicular to the subarray displacement vector.
( ) = d sin( )=c
ESPRIT Method, con't ESPRIT Scenario

source
known
subarray 1
displac ement v
ector
subarray 2
Properties: Requires special array geometry Computationally efcient
No risk of spurious DOA estimates

Does not require array calibration Note: For a ULA, the two subarrays are often the rst m and last m array elements, so m m and
;1
;1 A1 = Im;1 0]A
= ;1 A2 = 0 Im;1]A
Slide L919

Introduction To Spectral Analysis Sm-Slides-1ed

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Introduction To Spectral Analysis Sm-Slides-1ed

Hochgeladen von

Copyright:

Verfügbare Formate

Basic Denitions and The Spectral Estimation Problem

Informal Denition of Spectral Estimation

! = (angular) frequency in radians/(sampling interval) f = !=2 = frequency in cycles/(sampling interval)

Spatial Spectral Analysis Source location using sensor arrays

discrete-time deterministic data sequence

exists and is called the Discrete-Time Fourier Transform (DTFT)

Energy Spectral Density

jy(t)j2 = 21 ; S (!)d! t=;1

4 S (!) = jY (!)j2 = Energy Spectral Density

E f g = Expectation over the ensemble of realizations E jy(t)j2

PSD = (Average) power spectral density

First Denition of PSD

r(0) = E jy(t)j2 = 21 ; (!)d!

innitesimal signal power in the band

Second Denition of PSD

1 N y(t)e;i!t (!) = Nlim E N !1 t=1

1 jY (!)j2 (!) = Nlim E N N !1

is the nite DTFT of

Properties of the PSD

(!) = (! + 2 ) for all !.

(!) 0 (t) is real, Then: (! ) = (;! ) Otherwise: (! ) 6= (;! )

Transfer of PSD Through Linear Systems

System Function: where q ;1

= unit delay operator: q;1y(t) = y(t ; 1)

y (!) = jH (!)j2 e(!)

The Spectral Estimation Problem

The Problem: From a sample

Two Main Approaches :

Nonparametric: Derived from the PSD denitions.

Parametric: Assumes a parameterized functional form of the PSD

Periodogram and Correlogram Methods

Periodogram Recall 2nd denition of (! ):

1 N y(t)e;i!t (!) = Nlim E N !1 t=1

^p(!) = 1 y(t)e;i!t N t=1

Used by Schuster ( 1900) to determine hidden periodicities (hence the name).

Recall 1st denition of (! ):

Truncate the and replace r

Covariance Estimators (or Sample Covariances)

Standard unbiased estimate:

1 N y(t)y (t ; k) k 0 ^ r(k) = N ; k t=k+1

Standard biased estimate:

1 N y(t)y (t ; k) k 0 r(k) = N ^ t=k+1

For both estimators:

r ^(k) = ^ (;k) k < 0 r

Relationship Between ^p (! ) and ^c(! )

If: the biased ACS estimator r Then:

^(k) is used in ^c(!),

^p(!) = 1 y(t)e;i!t N t=1

^c(!) can be analyzed simultaneously.

Both have large variance, even for large N . Thus,

^p(!) and ^c(!) have poor performance.

r(k) ; r(k) may be large for large jkj ^

Bias Analysis of the Periodogram

Bartlett, or triangular, window

WB (!) = Dirac impulse (!).

1 sin(!N=2) 2 WB (!) = N sin(!=2)

WB (!)=WB (0), for N = 25

Main lobe 3dB width For small N , WB

(!) may differ quite a bit from (!).

Smearing and Leakage

Main Lobe Width: smearing or smoothing Details in resolvable.

(!) separated in f by less than 1=N are not