Lecture Notes Radar Systems

Week-7
Dr Sajid Ahmed

Information Technology University

Lahore, Pakistan
e-mail: sajid.ahmed@itu.edu.pk
April 12, 2020
 Probability Density Function (1/3)
• The Gaussian/normalized probability density function is defined as
1 − (x−m)
2

p(x) = √ e 2σ 2

2πσ 2

0.8
2
=0.5
2
0.7 =1
2
=2

0.6

0.5
p(x)

0.4

0.3

0.2

0.1

0
-5 -4 -3 -2 -1 0 1 2 3 4 5
x

1
 Probability Density Function (2/3)
• The area under the curve
Z ∞ ∞
1
Z 2
− (x−m)
p(x)dx = √ e 2σ 2 dx
−∞ −∞ 2πσ 2

= 1

• Area under the half region of the curve

Z 0 Z ∞
1 (x−m)2
− 2σ2 1 − (x−m)
2
√ e dx = √ e 2σ2 dx
−∞ 2πσ 2 0 2πσ 2
= 0.5

2
 Probability Density Function (3/3)
0.4

0.2

0
-5 -4 -3 -2 -1 0 1 2 3 4 5
0.4
p(x)

0.2

0
-5 -4 -3 -2 -1 0 1 2 3 4 5
0.4

0.2

0
-5 -4 -3 -2 -1 0 1 2 3 4 5
x

R∞ (x−m)2 R∞ −(x−m)2 R∞ −(x−m)2

−
√ 1 e 2σ 2 dx, √ 1 e 2σ 2 dx, √ 1 e 2σ 2 dx
0.5 2πσ 2 1 2πσ 2 −1 2πσ 2

3
 Some Useful Results (1/3)
• The error-function, erf(·) is defined as
x
1
Z
2
erf(x) = √ e−y dy
π −x

• The right-hand-side can be considered as the Gaussian pdf with σ 2 = 1/2.

Therefore,
Z x
1 2
erf(x = ∞) = √ e−y dy = 1
π −x
−y 2
• Since e is a symmetric we can write
x
1 1
Z
−y 2
erf(x) = √ e dy
2 π 0

4
 Some Useful Results (2/3)
• Moreover,
x ∞
1 1 1
Z Z
−y 2 2
√ e dy + √ e−y dy =
π 0 π x 2
Z ∞
1 1 2 1
erf(x) + √ e−y dy =
2 π x 2
Z ∞
1 2 1 1 1 1
√ e−y dy = − erf(x) = (1 − erf(x)) = erfc(x)
π x 2 2 2 2
• Therefore,
∞
1 1
Z
2
erfc(x) = √ e−y dy
2 π x

5
 Some Useful Results (3/3)
• For normalized (Gaussian) distribution the above integrals can be found using
Q function defined as
∞
1 −y2
Z
Q(x) = √ e 2 dy
x 2π
x−m
 
1
= erfc √
2 2σ 2
x−m
  
1
= 1 − erf √
2 2σ 2

6
 Detection in the Presence of Noise (1/6)
15 Detectable Noise
Targets
Marginal
Relative Power (dB)
Threshold
10
Undetectable
5

-5

-10

-15

-20
0 10 20 30 40 50 60 70 80 90 100
Range Gate

• The radar return is sampled at regular interval of time

• The sampled returns are due to the noise and may contain the signal of interest
• A threshold is used to reject noise
7
 Detection in the Presence of Noise (2/6)
• Detection is a binary hypothesis problem where we decide either

Hypothesis H0 is true and only noise is present (decision D0 ),

OR
Hypothesis H1 is true and a target is present in noise (decision D1 ).
• The associated probabilities are
– PD = p[D1 |H1 ]
– PF A = p[D1 |H0 ]
– Pmiss = p[D0 |H1 ]

• In radar, the decision process is designed to maximise probability-of-detection

(PD ), under the constraint that probability-of-false alarm (PF A ) does not
exceed a preset value (say ǫ) is called Neyman-Pearson criterion.
8
 Detection in the Presence of Noise (3/6)
• The binary hypothesis problem can be expressed as follows

H0 : x = v,
H1 : x = α + v,

where v ∼ N (µ, σ 2 ) is normal distributed random variable with mean µ and

variance σ 2 and the received signal from the target is α.

• The conditional densities under both hypotheses are given by:

1 v
− 2σ
2
p(x|H0 ) = √ e 2

2πσ 2

1 − (x−α)
2

and p(x|H1 ) = √ e 2σ . 2

2πσ 2

9
 Detection in the Presence of Noise (4/6)
0.4

0.3
p(x)

0.2

0.1

0
-5 -4 -3 -2 -1 0 1 2 3 4 5
x

• The probability of detection can be derived as

Z ∞
PD = p(x|H1)dx
T

10
 Detection in the Presence of Noise (5/6)
0.4

0.3
p(x)

0.2

0.1

0
-5 -4 -3 -2 -1 0 1 2 3 4 5
x

• The probability of false-alarm can be derived as

Z ∞
PF A = p(x|H0)dx
T

11
 Detection in the Presence of Noise (6/6)
0.4

0.3
p(x)

0.2

0.1

0
-5 -4 -3 -2 -1 0 1 2 3 4 5
x

• The probability of miss can be derived as

Z T
PM ISS = p(x|H1)dx
−∞

12
 Optimal Threshold (1/4)
• To maximise the probability of detection the Neyman-Pearson criterion is
J = PD + λ(PF A − ǫ)
= −ǫλ + PD + λPF A
Z ∞ Z ∞
= −ǫλ + p(x|H1)dx + λ p(x|H0)dx
T T

• The terms ǫλ is constant, therefore to maximise J, p(x|H1) + λp(x|H0 ) > 0

p(x|H1)
l = > −λ = τ,
p(x|H0)
H1
2αx−α2 >
= e 2σ 2 τ.
<
H0
• The LRT, l, can be used to know whether the data is generated under H0 or H1 .
13
 Optimal Threshold (2/4)
• The log-LRT, using the natural log, ( ln(·) ), is often used as an equivalent test
because the natural logarithm is a monotonic function for positive values

H1
2αx − α2 >
Λ = ln l = 2
ln τ = η. (1)
2σ <
H0
• Using (1), decision can be made on the basis of observation only as

H1
> 2
 
σ α
x η+ .
< α 2

H0

14
 Optimal Threshold (3/4)
 
σ2 α
• Above equation says that to satisfy Neyman Pearson criterion, T = αη + 2
is the optimal value of threshold.

• The PF A can be found as

Z ∞
PF A = p(x|H0)dx = ǫ
T
1  √ 
= erfc T/ 2σ 2 .
2
• Using the above equation
√
T = 2σ 2 erfc−1 (2ǫ).
√ −1
 2
σ α

• The optimum value of η can be found using 2σ 2 erfc (2ǫ) = α η + 2 .

15
 Optimal Threshold (4/4)
• The probability-of-detection in terms of erfc(·) can be derived as
Z ∞
1 − (x−α)
2

PD = √ e 2σ .2
(2)
T 2πσ 2
√ −1
!
T−α 2
2σ erfc (2ǫ) − α
 
1 1
= erfc √ = erfc √
2 2σ 2 2 2σ 2
r ! r !
1 −1 α2 1 −1 SN R
= erfc erfc (2ǫ) − = erfc erfc (2PF A) − .
2 2σ 2 2 2

16
Problem 6.1

17
Problem 6.2

18
Problem 6.3

19
Problem 6.4

20
Problem 6.5

21
Problem 6.6

22
Problem 6.7

23
 Log-Likelihood Ratio Test (1/7)
• Now consider the characterization of the log likelihood ratio test:
α α2
Λ = 2x − 2, (3)
σ 2σ
where it can be seen that Λ is itself a Gaussian random variable.
• To find the PDF of Λ under both hypothesises the mean of Λ can be calculated
as
α α2 α2
µΛ|H0 = E{Λ|H0 } = 2 E{x|H0 } − 2 = − 2 ,
σ 2σ 2σ
α α2 α2
µΛ|H1 = E{Λ|H1 } = 2 E{x|H1 } − 2 = 2 .
σ 2σ 2σ

If x ∼ N (µ, σ 2 ), then
E{x} = µ,
2 2 2 2
E{(x − µ) } = E{x } − (E{x}) = σ .
24
 Log-Likelihood Ratio Test (2/7)
• Using (3), we can write
α2
2 2 α3 α4
E{Λ } = 4 E{x } − 4 E{x} + 4 . (4)
σ σ 4σ
• The conditional expected value of the square of Λ under both hypothesis can be
calculated as
2 3 4 2 4
2 α 2 α α α α
E{Λ |H0 } = 4 E{n } − 4 E{n} + 4 = 2 + 4 ,
σ σ 4σ σ 4σ
2 3 4 2 4
2 α 2 α α α α
E{Λ |H1 } = 4 E{(α + n) } − 4 E{(α + n)} + 4 = 2 + 4 .
σ σ 4σ σ 4σ
• Therefore, using (4), the variance of Λ under both hypothesis can be written as
2 2 2 α2
σΛ|H = E{Λ |H0 } − (E{Λ|H0 }) = 2 ,
0
σ
2
2 2 2 α
σΛ|H = E{Λ |H 1 } − ( E{Λ|H 1 }) = 2.
1
σ
25
 Log-Likelihood Ratio Test (3/7)
• The PDF of Λ can be written as
!
2
1 (Λ − µΛ|H0 )
p(Λ|H0 ) = q exp − 2 ,
2
2πσΛ|H0 2σ Λ|H0
!
1 (Λ − µΛ|H1 )2
p(Λ|H1) = q exp − 2 .
2
2πσΛ|H 2σΛ|H1
1

26
 0.4

0.3 p( |H0 )
p( |H1 )
P( )

0.2

0.1

0
-5 -4 -3 -2 -1 0 1 2 3 4 5

2 2
Figure 1: PDF of log-LRT for both hypothesises, where σΛ|H0
= σΛ|H1
= 1.

27
 Log-Likelihood Ratio Test (4/7)
• The probability of detection (PD ) and the probability-of-false alarm (PF A ) can
be calculated using the statistics we have just derived.

• Both of these probabilities are associated with exceeding the threshold.

– PD is related to exceeding the threshold under hypothesis H1 .
– PF A is related to exceeding the threshold under hypothesis H0 .

!
∞ 2
1 (y − µΛ|H1 )
Z
PD = p(Λ > η|H1 ) = q exp − 2 dy,
y=η 2
2πσΛ|H 2σΛ|H1
1
!
∞ 2
1 (y − µΛ|H0 )
Z
PF A = p(Λ > η|H0 ) = q exp − 2 dy.
y=η 2
2πσΛ|H0 2σ Λ|H0

28
 Log-Likelihood Ratio Test (5/7)
• The PD and PF A in terms of area under the PDF curves can be represented as

0.5

0.45

0.4

0.35

0.3
p (Λ)

p(Λ|H 0) p(Λ|H 1)
0.25

0.2

0.15

0.1
P miss
0.05
PF A
0
−6 −4 −2 0 η 2 4 6
Λ
2
Figure 2: Probability of detection and false alarm, where σΛ|H1
= 1.

29
 √ !
η − µΛ|H1
  σ α η SNR
PD = p(Λ > η|H1 ) = Q =Q η− =Q √ −
σΛ|H1 α 2σ SNR 2
√ !
η − µΛ|H0
   σ α  η SNR
PF A = p(Λ > η|H0 ) = Q =Q η+ =Q √ +
σΛ|H0 α 2σ SNR 2

30
 Log-Likelihood Ratio Test (6/7)
• By increasing SNR the PF A can be decreased and PD can be increased.

0.16

0.14

0.12

0.1
p(Λ|H 0) p(Λ|H 1)
0.08

0.06

0.04

0.02

0
−15 −10 −5 0 η 5 10 15
Λ

2
Figure 3: Probability of detection and false alarm, where σΛ|H1
= 8.

31
 Log-Likelihood Ratio Test (7/7)
• By increasing SNR the PF A can be decreased and PD can be increased.

0.9

0.8

0.7

0.6
SNR = −10dB
0.5 SNR = −5dB
PD

SNR = 0dB
0.4 SNR = 5dB
SNR = 10dB
0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
PFA

32