You are on page 1of 17

14/11/16

Estimation and Detection


Lecture 1: Introduction and Minimum
Variance Unbiased Estimators

Dr. ir. Richard C. Hendriks & Dr. Sundeep P. Chepuri – 15/11/2016


1

Introduction – Why Estimation & Detection


Many signal processing applications involve estimation and/or detection tasks, e.g.:

• Hearing aid: Estimation of target speech in a noisy environment.

• localisation of sensors/sources in a wireless sensor network.

• Estimation/detection of biomedical signals.

• Radar: Detection of a target

• Etc.

1
14/11/16

Example: Estimation

• Sensitivity
• Dynamic range

Gain and Noise Gain and


compression reduction compression

• Due to the gain, inaudible noise/disturbances can suddenly become audi-


ble.
• Hearing impaired people have a worse spectral and temporal resolution,
and therefore more difficulties to understand a target under noisy condi-
tions. 3

Example: Estimation
A simplified model:

y 1 = s + n1

s Noise Gain and
reduction compression

n y 2 = s + n2
• ŝ1 = y1
1
PN
• ŝ2 = N i=1 yi
How to determine an estimate ŝ? • How good are these estimators?
• Are there better estimators?
PN yi
i=1 2
ni
What about ŝ3 = PN 1 ?
i=1 2
ni
4

2
14/11/16

y 1 = s + n1
Example: Estimation ŝ
Noise Gain and
reduction compression
let’s assume
y 2 = s + n2
• s is deterministic
• ŝ1 = y1
• Noise ni at every microphone is a zero-mean iid random process.
1
PN
• ŝ2 = N i=1 yi
PN
V ar[ŝ]? i=1
yi
2
ni
• ŝ3 = PN 1
i=1 2
ni
2 2
• V ar[ŝ1 ] = ni If indeed iid ) V ar[ŝ1 ] = n
PN P 2
• V ar[ŝ2 ] = V ar[ N1 i=1 yi ] = 1
N2
2
ni . If indeed iid ) V ar[ŝ2 ] = n
N
" PN yi #
i=1 2 2
• V ar[ŝ3 ] = V ar PN
ni
1 = P 11 . If indeed iid ) V ar[ŝ3 ] = n
i=1 2 2
N
ni ni

y 1 = s + n1
Example: Estimation ŝ
Noise Gain and
reduction compression
y 2 = s + n2
What if the noise is not iid, but only independent?
2 2 2 2
• n1 = n and n2 =2 n

V ar[ŝ]?
2
• V ar[ŝ1 ] = n
PN 3 2
• V ar[ŝ2 ] = V ar[ N1 i=1 yi ] = n
4 . Questions:
" PN yi #
i=1 2
ni 2 2 • How to derive good (optimal) estimators?
• V ar[ŝ3 ] = V ar PN 1 = 3
n
.
i=1 2
ni
• How to quantify their performance?
• What is the best estimator for a certain estimation task?
6

3
14/11/16

Example: Detection
Characterization of abnormal heart rhythms (Atrial fibrillation)
Common treatment: Ablation. Only successful in 50 % of cases. . .
Research project on atrial fibrillation:
• Measure signals directly on the heart using a high dimensional sensor
(188).

• Get understanding of the real problem in Atrial fibrillation.


• Develop less invasive methods.

Example: Detection
5000
Response at 5 di↵erent positions of the sensor
0

-5000
without atrial fibrillation.
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000

0
• How to detect the moment of atrial activity
-5000
in each signal?
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000

-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000

-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000

-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

time samples (f s = 10 kHz)

4
14/11/16

Example: Detection
5000
Response at 5 di↵erent positions of the sensor
0
WITH atrial fibrillation.
-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000
• How to detect the moment of atrial activity
0
in each signal?
-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000

-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000

-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000

-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
time samples (f s = 10 kHz)

Example: Detection
5000 Derivative of the response at 5 di↵erent
0 positions of the sensor WITH atrial fibrillation.
-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000 • How to detect the moment of atrial activity
0 in each signal?
-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000 • Check whether y[n] > ?
0

-5000
• How to chose the threshold ?
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
5000

0
• What is the optimal detector?
-5000

5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
• How to evaluate the performance of the
0 detector?
-5000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

time samples (f s = 10 kHz)


Threshold

10

5
14/11/16

This Course
Theory on estimation and detection:
• Generally, the (optimal) solution to estimation and detection problems
depends on the underlying signal model.
• Data is often random: Random noise and sometimes random target as
well. Models are therefore often based on statistical descriptions.
Key Questions:
• How to determine optimal estimators.
• What is the estimation bound?

• How to determine the optimal detector?


• How to express the detection performance?

11

This Course – Estimation Theory

Typical problem formulation for estimation theory:

Consider the estimation of ✓ from observations y[n] = ✓ + w[n]

While the noise is typically assumed to be stochastic, the target can be


• ✓ can be an unknown deterministic quantity ) Classical estimation (lec-
ture 1 - 5).
• ✓ can be an (unknown) random quantity ) Bayesian estimation (lecture
6 - 8).

12

6
14/11/16

This Course – Detection Theory


Typical binary detection problem: Determine whether a certain signal that is embedded in
noise is present or not.
H0 x[n] = w[n]
H1 x[n] = s[n] + w[n]
Typical detector:
T (x[n]) >

• How to make an optimal decision? (that is, how to chose )

• Detection for deterministic signals (lecture 9).

• Detection for random signals (lecture 10).

• Detection with unknown parameters (lecture 11).


13

Course Information - 1

• Course website: http://ens.ewi.tudelft.nl/Education/courses/et4386/index.php

• Blackboard

• Exam:

– Written

– closed book, with one 2-sided HANDWRITTEN A4 formula sheet.

– Mini-project counts for 20 % of the final grade.

14

7
14/11/16

Course Information - 2
• books:

– Fundamentals of Statistical Signal Processing, Volume I: Estimation Theory;


S.M. Kay, Prentice Hall 1993; ISBN-13: 978-0133457117.

– Fundamentals of Statistical Signal Processing, Volume II: Detection Theory;


S.M. Kay, Prentice 1993; ISBN-13: 978-0135041352.

• Instructors:

– Richard C. Hendriks

– Sundeep Chepuri

• Contact: et4386-EWI@TuDelft.nl
15

Course Information – Mini Projects


• Grade for the project counts for 20 % of the final grade.

• Can be carried out in pairs of students.

• There will be several projects to chose from. Let us know your ranked preference
before November 25th (to the course e-mail address).

• Final report deadline: before January 9th!

• Access to final exam will only be granted if report is handed in!

• Project presentation during last two lectures.

16

8
14/11/16

Estimation Theory – Lecture1

Minimum Variance Unbiased


(MVU) Estimation

17

Introductory Example
Given a process (DC in noise):

x[n] = A + w[n] , n = 0, · · · , N 1, w[n] zero mean noise

How can we estimate A?

1. Â1 = ...? What is the signal model?

2. Â2 = ...? • w[n] is a zero-mean random uncorrelated process.

3. ... • A: Let’s for now assume A is deterministic.

18

9
14/11/16

Introductory Example
Given a process (DC in noise):

x[n] = A + w[n] , n = 0, · · · , N 1, w[n] zero mean noise

How can we estimate A? Consider the case A = 1.


Some candidates: Which estimate is better?

1. Â1 = x[0] • Â1 = x[0] = 0.93


PN 1 PN 1
2. Â2 = 1
N n=0 x[n] • Â2 = 1
N n=0 x[n] = 0.9

Can the performance of an estimator be based on a single realisation?


No! The data is random. As a result, the estimator is a random variable too!
19

Introductory Example
Use statistical descriptors: Expected value and variance of the estimator.

• E(Â1 ) = A
⇣ PN ⌘ PN
1 1 1 1
E(Â2 ) = E N n=0 x[n] = N n=0 E(x[n]) =A
Both estimators are unbiased.

• Look at the variance, define 2


= var(w[n]):
2
var(Â1 ) =
⇣ PN ⌘ PN 2
1 1 1 1
var(Â2 ) = var N n=0 x[n] = N2 n=0 var(x[n]) = N

Â2 has smaller variance. Also, var(Â2 ) ! 0 as N ! 1: the estimator is consistent.

Conclusion: Estimators are random variables.


Question: What are good (or optimal) estimators?
20

10
14/11/16

Unbiased Estimators
Let ✓ˆ = g(x), x = [x[0], x[1], ..., x[N 1]]T .
An unbiased estimator implies that
Z
ˆ =
E[✓] g(x)p(x; ✓)dx = ✓.

Is being unbiased sufficient? No, hence the example:

1. Â1 = x[0]
PN 1
2. Â2 = 1
N n=0 x[n]

21

Minimum Variance Criterion


A natural criterion is the Mean Square Error (MSE):
h i ⇢h i2
mse(✓)ˆ = E (✓ˆ ✓)2 = E (✓ˆ E(✓))ˆ + (E(✓)
ˆ ✓)
h i
= E (✓ˆ E(✓))
ˆ 2 + (E(✓)
ˆ ˆ +(E(✓)
✓)2 = var(✓) ˆ ✓ )2
| {z } | {z }
variance bias

The MSE consist of errors due to

• bias of the estimator

• variance of the estimator.

22

11
14/11/16

Minimum Variance Criterion


h i
ˆ = E (✓ˆ E(✓))
mse(✓) ˆ 2 + (E(✓)
ˆ ˆ +(E(✓)
✓)2 = var(✓) ˆ ✓ )2
| {z } | {z }
variance bias
PN 1
Consider the (biased) estimator: Â3 = a
N n=0 x[n], with constant a.

• E[Â3 ] = aA

a2 2
• var[Â3 ] = N

a2 2
• MSE(Â3 ) = N + (a 1)2 A2 .

A2
Find the optimal estimator requires finding the optimal a: aopt = A2 + 2 /N
.
2 PN 1
The optimal estimator: Â3 = N AA2 + 2 n=0 x[n].

) This estimator depends on A and is therefore unrealisable!!


23

Minimum Variance Unbiased Estimator

The optimal estimator depends on A (and thus unrealisable), because the bias term in the
MSE depends on A:
a2 2
MSE(Â3 ) = + (a 1)2 A2 .
N | {z }
bias

We therefore abandon the MSE and constrain the bias to be zero.

ˆ = ✓, then
Hence, if E(✓)
h i h i
ˆ = E (✓ˆ E(✓))
mse(✓) ˆ 2 + (E(✓)
ˆ ✓)2 = E (✓ˆ E(✓))
ˆ 2 .

This is what we call: the minimum variance unbiased estimator (MVU).

24

12
14/11/16

Minimum Variance Unbiased Estimator

Remark: The MVU does not always exist and is generally difficult to find.

ˆ
var(✓) ˆ
var(✓)
✓ˆ1 ✓ˆ1
✓ˆ2
✓ˆ3 ✓ˆ2
✓ˆ3
✓ ✓

If ✓1 , ✓2 and ✓3 are all unbiased MVU does not exist!


estimators, then ✓3 is the MVU.
25

MVU: Example
8
< N (✓, 1) , ✓ 0
Consider two independent random processes x and y: x ⇠ N (✓, 1) , y⇠
: N (✓, 2) , ✓<0
Consider two possible estimators for ✓: ✓ˆ1 = 1
2 (x + y) , ✓ˆ2 = 2 1
3x+ 3y

Both estimators are unbiased. Look at the variances:


8
1 < 18
, ✓ 0
var(✓ˆ1 ) = (var(x) + var(y)) = 36
4 : 27
, ✓<0
36
8
4 1 < 20
, ✓ 0
var(✓ˆ2 ) = var(x) + var(y) = 36
9 9 : 24
, ✓<0
36

Between these two estimators, there is no MVU estimator.

26

13
14/11/16

Finding the MVU

Even if the MVU exists, there is no standard "recipe" to find it.

1. Determine Cramer-Rao-Bound (next week, Ch. 3)

2. Apply Rao-Blackwell-Lehmann-Scheffe theorem (will not be discussed)

3. Restrict estimators to be both unbiased AND linear (Ch. 6)

27

Towards the CRB (Ch. 3): Estimator accuracy

Goal: Estimate A
x[0] = A + w[0],

with w[n] ⇠ N (0, 2


).

⇥ ⇤
The pdf of x[0] is given by: p(x[0]; A) = p 1 exp 1
2 2
(x[0] A)2
2⇡ 2
Question: What determines now how accurate we can estimate A?

The CRB is a bound on the variance V ar[Â] of an estimator.

28

14
14/11/16

Towards the CRB (Ch. 3): Estimator accuracy


p(x[0] = 3; A; < 2 = 1=3])
Likelihood function: p(x[0]; A) as function of A.
0.5

0.4

0.3 Consider the two cases 2


= 1 and 2
= 1/3.
0.2

0.1 The "sharpness" (curvature) of the likelihood


0
0 1 2 3
A
4 5 6 7 function determines how accurately the unknown
0.5
p(x[0] = 3; A; < 2 = 1])
parameter can be estimated.
0.4

0.3

0.2
How to measure sharpness/curvature?
0.1

0
0 1 2 3 4 5 6 7
A

29

Towards the CRB (Ch. 3): Estimator accuracy


@ 2 ln[p(x[0];A)]
Measure curvature by @A2
p(x[0] = 3; A; < = 1])
0.2

0.15
p(x[0]; A)
0.1

0.05

0
0 1 2 3 4 5 6 7
A

6
ln [p(x[0]; A)]
4 @ ln[p(x[0];A)]
@A
2 @ 2 ln[p(x[0];A)]
@A2
0

-2

-4

-6
0 1 2 3 4 5 6 7
A

30

15
14/11/16

Towards the CRB (Ch. 3): Estimator accuracy


@ 2 ln[p(x[0];A)]
Measure curvature by @A2

h i
In this example: 1
@ 2 ln[p(x[0];A)]
= 2
= var Â
@A2
2
@ ln[p(x[0];A)]
In general, as @A2
can depend on x[0], curvature is measured by

@ 2 ln [p(x[0]; A)]
E
@A2

The Cramer-Rao bound:


h i 1
var  h i
@ 2 ln[p(x;A)]
E @A2
31

Towards the CRB (Ch. 3): Example

Calculate the CRB:

x[n] = A + w[n] , n = 0, · · · , N 1, with w[n] ⇠ N (0, 2


)

PN 1
Remember estimator Â2 = 1
N n=0 x[n].

What is the Cramer-Rao bound


h i 1
var  h i?
@ 2 ln[p(x;A)]
E @A2

32

16
14/11/16

Next week:
Cramer-Rao Lower bound

• Gives a bound on the variance of an unbiased estimator:

ˆ 1 1
var(✓)  2 = "✓ ◆2 #
@ ln p(x; ✓) @ ln p(x; ✓)
E E
@✓2 @✓

• Can be used to determine whether an estimator is the MVU. This is the case if an
estimator reaches the bound.

• Can be used as a benchmark to compare the performance of unbiased estimators.

• Can be used to determine the non-existence of better estimators.


33

17

You might also like