3.

11 CRLB Examples
Well now apply the CRLB theory to several examples of
practical signal processing problems.
Well revisit these examples in Ch. 7 well derive ML
estimators that will get close to achieving the CRLB
1.

Range Estimation

2.

sonar, radar, robotics, emitter location

Sinusoidal Parameter Estimation (Amp., Frequency, Phase)

3.

sonar, radar, communication receivers (recall DSB Example), etc.

Bearing Estimation

4.

sonar, radar, emitter location

Autoregressive Parameter Estimation

speech processing, econometrics

1/19

Ex. 1 Range Estimation Problem

Transmit Pulse:

s(t)

nonzero over t[0,Ts]

Receive Reflection:

s(t o)

Measure Time Delay: o

C-T Signal Model

x ( t ) = s ( t o ) + w( t )
$
!#!
"

0 t T = Ts + o,max

s ( t ; o )

s(t)

Bandlimited
White Gaussian

BPF
& Amp

x(t)

PSD of w(t)
No/2
B

Ts

s(t o)
f

2/19

Range Estimation D-T Signal Model

ACF of w(t)

PSD of w(t)
No/2
B

2 = BNo

f
1/2B

x[n ] = s ( n o ) + w[n ]

1/B 3/2B

Sample Every = 1/2B sec

w[n] = w(n)

n = 0,1, , N 1

DT White
Gaussian Noise
Var 2 = BNo

s[n;o] has M non-zero samples starting at no

w[n ]

x[n ] = s ( n o ) + w[n ]

w[n ]

0 n no 1
no n n o + M 1
no + M n N 1

3/19

Range Estimation CRLB

Now apply standard CRLB result for signal + WGN:
Plug in and keep
non-zero terms

var(o )

2
N 1

s[n; o ]
o

n = 0

no + M 1

n = no

2
no + M 1

n = no

(
)
s
t

t t = n
o

Exploit Calculus!!!

s ( n o )

2
M 1

(
)
s
t

t
=
n

n =0

Use approximation: o = no
Then do change of variables!!
4/19

Range Estimation CRLB (cont.)

Assume sample spacing is small approx. sum by integral

var(o )

No / 2

1 Ts s (t )

dt

Ts s ( t )

dt

Ts s ( t )

Es
No / 2

dt
t
Es
Ts

var(o )
Es
No / 2

Ts

2
2
(
)

2
(
)
f
S
f
df
0

Es

Es
No / 2

A type of SNR

(2f )

Define a BW measure:

FT Theorem
& Parseval

E s = s 2 (t )dt

S ( f ) df

S ( f )

dt

Parseval

Brms =

2
2
(
)

2
(
)
f
S
f
df

S ( f ) dt

Brms is RMS BW (Hz)

5/19

Range Estimation CRLB (cont.)

Using these ideas we arrive at the CRLB on the delay:
var(o )

1
2
SNR Brms

(sec 2 )

To get the CRLB on the range use transf. of parms result:

R
CRLBR =
o

CRLB
o

with R = co / 2

CRLB
CRLBisisinversely
inverselyproportional
proportionalto:
to:
SNR
SNRMeasure
Measure
RMS
RMSBW
BWMeasure
Measure

var( R )

c2 / 4
2
SNR Brms

(m 2 )

So the CRLB tells us

Choose signal with large Brms
Ensure that SNR is large
Better on Nearby/large targets
Which is better?
Double transmitted energy?
Double RMS bandwidth?
6/19

Ex. 2 Sinusoid Estimation CRLB Problem

Given DT signal samples of a sinusoid in noise.
Estimate its amplitude, frequency, and phase
x[n ] = A cos( o n + ) + w[n ]

n = 0, 1, , N 1

o is DT frequency in
cycles/sample: 0 < o <

DT White Gaussian Noise

Zero Mean & Variance of 2

Multiple parameters so parameter vector: = [ A o

]T

Recall SNR of sinusoid in noise is:

Ps A2 / 2
A2
=
=
SNR =
2
Pn

2 2
7/19

Sinusoid Estimation CRLB Approach

Approach:
Find Fisher Info Matrix
Invert to get CRLB matrix
Look at diagonal elements to get bounds on parm variances

Recall: Result for FIM for general Gaussian case specialized to

signal in AWGN case:
1 s s

[I()]ij = 2
i j
=

N 1

s[n; ] s[n; ]

i
j
n =0
8/19

Sinusoid Estimation Fisher Info Elements

Taking the partial derivatives and using approximations given in
book (valid when o is not near 0 or ) :
= [ A o ]T
[I()]11 =

N 1

cos (o n + ) =

n =0

N 1

2 2

(1 + cos(2o n + 2 ))

n =0

[I()]12

1 N 1
A
= [I()]21 = 2 An cos( o n + ) sin( o n + ) =
2 2
n =0

[I()]13

1 N 1
A
= [I()]31 = 2 A cos( o n + ) sin( o n + ) =
2 2
n =0

[I()]22 =

N 1

n =0

[I()]23 = [I()]32 =
[I()]33 =

N 1

n =0

( n ) sin ( o n + ) =
1

N 1

A n sin

n =0
2

A2
2 2

N 1

2 2

N 1

n sin(2o n + 2 ) 0

n =0

N 1

sin(2o n + 2 ) 0

n =0

A2

N 1

n (1 cos(2o n + 2 )) 2 2 n 2
2

n =0

( o n + )

sin ( o n + )

n =0

A2
2

N 1

n =0

NA2
2 2

9/19

Sinusoid Estimation Fisher Info Matrix

= [ A o ]T

Fisher Info Matrix then is:

N
2
2

I( ) 0

Recall SNR =

A2
2

0
N 1

A2
2 2
A2
2

n2

n =0

N 1

n =0

A2 N 1
n
2
2 n =0

NA2
2 2

and closed form results for these sums

10/19

Sinusoid Estimation CRLBs

(using co-factor & det

approach helped by 0s)

Inverting the FIM by hand gives the CRLB matrix and then
extracting the diagonal elements gives the three bounds:
2
2

var( A )
N

)
var(
o
var()

( volts2 )
12

SNR N ( N 2 1)

(( rad/sample) 2 )

2( 2 N 1)
4

SNR N ( N + 1) SNR N

To convert to Hz2
multiply by (Fs /2)2

( rad 2 )

Amp. Accuracy: Decreases as 1/N, Depends on Noise Variance (not SNR)

Freq. Accuracy: Decreases as 1/N3, Decreases as as 1/SNR
Phase Accuracy: Decreases as 1/N, Decreases as as 1/SNR
11/19

Frequency Estimation CRLBs and Fs

Not in Book

The CRLB for Freq. Est. referred back to the CT is

var( fo )

12 Fs2
( 2 ) 2 SNR N ( N 2 1)

( Hz 2 )

Does that mean we do worse if we sample faster than Nyquist?

NO!!!!! For a fixed duration T of signal: N = TFs
Also keep in mind that Fs has effect on the noise structure:
ACF of w(t)

PSD of w(t)
No/2
B

2 = BNo

f
1/2B

1/B 3/2B

12/19

Ex. 3 Bearing Estimation CRLB Problem

Figure 3.8
from textbook:

Emits or reflects
signal s(t)

s(t ) = At cos(2f o t + )
Simple model

Uniformly spaced linear array with M sensors:

Sensor Spacing of d meters
Bearing angle to target radians

Propagation Time to

nth

Sensor: tn = t0 n

d
cos
c

n = 0, 1, , M 1

sn (t ) = s (t tn )

Signal at nth Sensor:

= A cos 2f o t t0 + n cos +
c

13/19

Bearing Estimation Snapshot of Sensor Signals

Now instead of sampling each sensor at lots of time instants
we just grab one snapshot of all M sensors at a single instant ts

sn (t s ) = A cos 2f o t s t0 + n cos +
c

2f
~
~

cos d n + = A cos s n +
= A cos
c
$!

!
#
!
!
"

!#!!!
"
$!!

Spatial sinusoid w/
s

spatial frequency s

Spatial Frequencies:
s is in rad/meter
s is in rad/sensor

For sinusoidal transmitted signal Bearing

Est. reduces to Frequency Est.
And we already know its FIM & CRLB!!!
14/19

Bearing Estimation Data and Parameters

Each sample in the snapshot is corrupted by a noise sample
and these M samples make the data vector x = [x[0] x[1] x[M-1] ]:

x[n ] = sn (t s ) + w[n ] = A cos s n + + w[n ]

Each w[n] is a noise sample that comes from a different sensor so
Model as uncorrelated Gaussian RVs (same as white temporal noise)
Assume each sensor has same noise variance 2

So the parameters to consider are:

which get transformed to:

Parameter of interest!

= [ A s ]T

A
A

c s

= g() = = arccos

2f o d

15/19

Bearing Estimation CRLB Result

Using the FIM for the sinusoidal parameter problem together
with the transform. of parms result (see book p. 59 for details):
var( )

12
2

M +1 L
2
( 2 ) 2 SNR M
sin ( )
M 1
L = Array physical length in meters
M = Number of array elements
= c/fo Wavelength in meters (per cycle)

( rad 2 )

Define: Lr = L/
Array Length in
wavelengths

Bearing Accuracy:
Decreases as 1/SNR

Depends on actual bearing

Decreases as 1/M

! Best at = /2 (Broadside)

Decreases as 1/Lr2

! Impossible at = 0! (Endfire)

Low-frequency (i.e., long wavelength) signals need

very large physical lengths to achieve good accuracy

16/19

Ex. 4 AR Estimation CRLB Problem

In speech processing (and other areas) we often model the
signal as an AR random process and need to estimate the AR
parameters. An AR process has a PSD given by
u2

Pxx ( f ; ) =
1+

a[m]e j 2fm

m =1

AR Estimation Problem: Given data x[0], x[1], , x[N-1]

estimate the AR parameter vector
= [a[1]

a[ 2 ] & a [ p ]

2T
u

This is a hard CRLB to find exactly but it has been published.

The difficulty comes from the fact that there is no easy direct
relationship between the parameters and the data.
It is not a signal plus noise problem
17/19

AR Estimation CRLB Asymptotic Approach

Approach: The asymptotic result we discussed is perfect here:
An AR process is WSS is required for the Asymp. Result
Gaussian is often a reasonable assumption needed for Asymp. Result
The Asymp. Result is in terms of partial derivatives of the PSD and
that is exactly the form in which the parameters are clearly displayed!

Recall: [I()]ij

1
2

1
2

[ln Pxx ( f ; )] [ln Pxx ( f ; )]

df
i
j

u2

ln Pxx ( f ; ) = ln
1+

a[m]e

2
j 2fm

= ln u2 ln 1 +

a[m]e j 2fm

m =1

m =1

18/19

AR Estimation CRLB Asymptotic Result

After taking these derivatives you get results that can be
simplified using properties of FT and convolution.
Complicated
The final result is:

var(a[k ])
var( u2 )

[R ]
N

u2

1
xx kk

k = 1, 2, , p

2 u4
N

dependence on
AC Matrix!!

Both Decrease
as 1/N

To get a little insight look at 1st order AR case (p = 1):

var(a[1])

Im(z)

a[1]

Re(z)

1
(1 a 2 [1])
N

Improves as pole
gets closer to unit
circle PSDs
with sharp
peaks are easier
to estimate

19/19