An Introduction to Signal Detection and

Estimation - Second Edition

Chapter II: Selected Solutions
H. V. Poor
Princeton University
March 16, 2005
Exercise 2:
The likelihood ratio is given by
L(y) =
3
2(y + 1)
, 0 y 1.
a. With uniform costs and equal priors, the critical region for minimum Bayes error is
given by {y [0, 1]|L(y) 1} = {y [0, 1]|3 2(y + 1)} = [0, 1/2]. Thus the Bayes rule
is given by

B
(y) =
_
1 if 0 y 1/2
0 if 1/2 < y 1
.
The corresponding minimum Bayes risk is
r(
B
) =
1
2
_
1/2
0
2
3
(y + 1)dy +
_
1
1/2
dy =
11
24
.
b. With uniform costs, the least-favorable prior will be interior to (0, 1), so we examine
the conditional risks of Bayes rules for an equalizer condition. The critical region for the
Bayes rule

0
is given by

1
=
_
y [0, 1]

L(y)

0
1
0
_
= [0,

],
where

=
_

_
1 if 0
0

3
7
1
2
_
3

0
5
_
if
3
7
<
0
<
3
5
0 if
3
7

0
1
.
1
Thus, the conditional risks are:
R
0
(

0
) =
_

0
2
3
(y + 1)dy =
_

_
1 if 0
0

3
7
2

3
_

2
+ 1
_
if
3
7
<
0
<
3
5
0 if
3
7

0
1
,
and
R
1
(

0
) =
_
1

dy =
_

_
0 if 0
0

3
7
1

if
3
7
<
0
<
3
5
1 if
3
7

0
1
.
By inspection, a minimax threshold

L
is the solution to the equation
2

L
3
_

L
2
+ 1
_
= 1

L
,
which yields

L
= (

375)/2. The minimax risk is the value of the equalized conditional

risk; i.e., V (
L
) = 1

L
.
c. The Neyman-Pearson test is given by

NP
(y) =
_

_
1 if
3
2(y+1)
>

0
if
3
2(y+1)
=
0 if
3
2(y+1)
<
,
where and
0
are chosen to give false-alarm probability . Since L(y) is monotone
decreasing in y, the above test is equivalent to

NP
(y) =
_

_
1 if y <

0
if y =

0 if y >

,
where

=
3
2
1. Since Y is a continuous random variable, we can ignore the random-
ization. Thus, the false-alarm probability is:
P
F
(
NP
) = P
0
(Y <

) =
_

0
2
3
(y + 1)dy =
_

_
0 if

0
2

3
_

2
+ 1
_
if 0 <

< 1
1 if

1
.
The threshold for P
F
(
NP
) = is the solution to
2

3
_

2
+ 1
_
= ,
which is

1 + 3 1. So, an -level Neyman-Pearson test is

NP
(y) =
_
1 if y

1 + 3 1
0 if y >

1 + 3 1
.
The detection probability is
P
D
(
NP
) =
_

0
dy =

1 + 3 1, 0 < < 1.
2
Exercise 4:
Here the likelihood ratio is given by
L(y) =

e
y
y
2
2

2e

(y1)
2
2
, y 0.
a. Thus, Bayes critical regions are of the form

1
=
_
y 0

(y 1)
2

_
,
where

=
_

2e
log
_

0
1
0
_
. There are three cases:

1
=
_

_
if

< 0
_
1

, 1 +

_
if 0

1
_
0, 1 +

_
if

> 1
.
The condition

< 0 is equivalent to

0
<
0
1, where

0
=

2e

1+

2e

; the condition
0

1 is equivalent to

0

0

0
, where

0
=

1+

; and the condition

> 1 is
equivalent to 0
0
<

0
.
The minimum Bayes risk V (
0
) can be calculated for the three regions:
V (
0
) = 1
0
,

0
<
0
1,
V (
0
) =
0
_
1+

e
y
dy+(1
0
)

_
_
1

0
e

y
2
2
dy +
_

1+

y
2
2
dy
_
,

0

0

0
,
and
V (
0
) =
0
_
1+

0
e
y
dy + (1
0
)

_

1+

y
2
2
dy, 0
0
<

0
.
b. The minimax rule can be found by equating conditional risks. Investigation of the
above shows that this equality occurs in the intermediate region

0

0

0
, and thus
corresponds to a threshold

L
(0, 1) solving
e

L
e

L
= 2e(1 + (1 +
_

L
) (1
_

L
)).
The minimax risk is then either of the equal conditional risks; e.g.,
V (
L
) = e
1+

L
e
1

L
.
c. Here, randomization is unnecessary, and the Neyman-Pearson critical regions are
of the form

1
=
_
y 0

(y 1)
2

_
,
3
where

=
_

2e
log(). There are three cases:

1
=
_

_
if

< 0
_
1

, 1 +

_
if 0

1
_
0, 1 +

_
if

> 1
.
The false-alarm probability is thus:
P
F
(
NP
) = 0,

< 0
P
F
(
NP
) =
_
1+

e
y
dy = e
1+

e
1

=
2
e
sinh
_
_

_
, 0

1,
and
P
F
(
NP
) =
_
1+

0
e
y
dy = 1 e
1

> 1.
From this we see that the threshold for level NP testing is

=
_
_
_
_
sinh
1
_
e
2
__
2
if 0 < 1 e
2
[1 + log(1 )]
2
if 1 e
2
< < 1
.
The detection probability is thus
P
D
(
NP
) = 2
_

_
1 +
_

_
1
_

__
= 2
_

_
1 + sinh
1
_
e
2
__

_
1 sinh
1
_
e
2
___
, 0 < 1 e
2
,
and
P
D
(
NP
) = 2
_

_
1 +
_

1
2
_
= 2
_
(2 + log(1 ))
1
2
_
, 1 e
2
< 1.
Exercise 6 a & b:
Here we have p
0
(y) = p
N
(y + s) and p
1
(y) = p
N
(y s), which gives
L(y) =
1 + (y + s)
2
1 + (y s)
2
.
a. With equal priors and uniform costs, the critical region for Bayes testing is
1
=
{L(y) 1} = {1 + (y + s)
2
1 + (y s)
2
} = {2sy 2sy} = [0, ). Thus, the Bayes
test is

B
(y) =
_
1 if y 0
0 if y < 0
4
The minimum Bayes risk is then
r(
B
) =
1
2
_

0
1
[1 + (y + s)
2
]
dy +
1
2
_
0

1
[1 + (y s)
2
]
dy =
1
2

tan
1
(s)

.
b. Because of the symmetry of this problem with uniform costs, we can guess that
1/2 is the least-favorable prior. To conrm this, we can check that this answer from Part
a gives an equalizer rule:
R
0
(
1/2
) =
_

0
1
[1 + (y + s)
2
]
dy =
1
2
_
0

1
[1 + (y s)
2
]
dy = R
1
(
1/2
).
Exercise 7:
a. The densities under the two hypotheses are:
p
0
(y) = p(y) = e
y
, y > 0,
and
p
1
(y) =
_

p(y s)p(y)ds =
_
y
0
e
sy
e
s
ds = ye
y
, y > 0.
Thus, the likelihood ratio is
L(y) =
p
1
(y)
p
0
(y)
= y, y > 0.
b. Randomization is irrelevant here, so the false-alrm probability for threshold is
P
F
(
NP
) = P
0
(Y > ) = e

,
which gives the threshold = log , for level Neyman-Pearson testing. The corre-
sponding detection probability is
P
D
(
NP
) = P
1
(Y > ) =
_

ye
y
dy = ( + 1)e

= (1 log ), 0 < < 1.

c. Here the densities under the two hypotheses become:
p
0
(y) =
n

k=1
p(y
k
) =
n

k=1
e
y
k
, 0 < min{y
1
, y
2
, . . . , y
n
},
and
p
1
(y) =
_

_
n

k=1
p(y
k
s)
_
p(s)ds =
_
min{y
1
,y
2
,...,yn}
0
_
n

k=1
e
sy
k
_
e
s
ds
=
p
0
(y)
n 1
_
e
(n1) min{y
1
,y
2
,...,yn}
1
_
, 0 < min{y
1
, y
2
, . . . , y
n
}.
5
Thus, the likelihood ratio is
L(y) =
1
n 1
_
e
(n1) min{y
1
,y
2
,...,yn}
1
_
, 0 < min{y
1
, y
2
, . . . , y
n
}.
d. The false-alarm probability incurred by comparing L(y) from Part c to a threshold
is
P
F
(
NP
) = P
0
(L(Y ) > ) = P
0
_
min{Y
1
, Y
2
, . . . , Y
n
} >

log((n 1) + 1)
n 1
_
= P
0
(
n

k=1
(Y
k
>

)) =
n

k=1
P
0
(Y
k
>

) =
n

k=1
e

= e
n

,
from which we have

=
1
n
log , or, equivalently,
=
e
(n1)

1
n 1
=

(n1)/n
1
n 1
.
Exercise 15
a. The LMP test is

lo
(y) =
_

_
1 if
p

(y)

|
=0
> p
0
(y)
, if
p

(y)

|
=0
= p
0
(y)
0 if
p

(y)

|
=0
< p
0
(y) .
we have
p

(y)

|
=0
p
0
(y)
= sgn(y) ;
thus

lo
(y) =
_

_
1 if sgn(y) >
, if sgn(y) =
0 if sgn(y) < .
To set the threshold , we consider
P
0
(sgn(Y ) > ) =
_

_
0 if 1
1/2 if 1 < 1
1 if < 1 .
This implies that
=
_
1 if 0 < < 1/2
1 if 1/2 < 1 .
6
The randomization is
=
P
0
(sgn(Y ) > )
P
0
(sgn(Y ) )
=
_
2 if 0 < < 1/2
2 1 if 1/2 < 1 .
The LMP test is thus

lo
(y) =
_
2 if y > 0
0 if y 0
for 0 < < 1/2 ; and it is

lo
(y) =
_
1 if y 0
2 1 if y < 0
for 1/2 < 1.
For xed > 0, the detection probability is
P
D
(

lo
; ) = P

(sgn(Y ) > ) + P

(sgn(Y ) = )
=
_
2
_

0
1
2
e
|y|
dy if 0 < < 1/2
_

0
1
2
e
|y|
dy + (2 1)
_
0

1
2
e
|y|
dy if 1/2 < 1
=
_
(2 e

) if 0 < < 1/2

1 + ( 1)e

if 1/2 < 1
.
b. For xed , the NP critical region is

= {|y| |y | >

}
=
_

_
(, ) if

<
((

+
2
), ) if

if

> ,
from which
P
0
(

) =
_

_
1 if

<
1
2
e
(

+)/2
if

0 if

> .
Clearly, we must know to set

, and thus the NP critical region depends on . This

implies that there is no UMP test.
The generalized likelihood ratio test uses this statistic:
sup
>0
e
|y||y|
= exp{sup
>0
(|y| |y |)}
=
_
1 if y < 0
e
y
if y 0
.
7
Exercise 16:
We have M hypotheses H
0
, H
1
. . . , H
M1
, where Y has distribution P
i
and density p
i
under hypothesis H
i
. A decision rule is a partition of the observation set into regions

0
,
1
, . . . ,
M1
, where chooses hypothesis H
i
when we observe y
i
. Equivalently, a
decision rule can be viewed as a mapping from to the set of decisions {0, 1, . . . , M1},
where (y) is the index of the hypothesis accepted when we observe Y = y.
On assigning costs C
ij
to the acceptance of H
i
when H
j
is true, for 0 i, j (M1),
we can dene conditional risks, R
j
(), j = 0, 1, . . . , M 1, for a decision rule , where
R
j
() is the conditional expected cost given that H
j
is true. We have
R
j
() =
M1

i=0
C
ij
P
j
(
i
).
Assuming priors
j
= P(H
j
occurs), j = 0, 1, . . . , M 1, we can dene an overall average
risk or Bayes risk as
r() =
M1

j=0

j
R
j
().
A Bayes rule will minimize the Bayes risk.
We can write
r() =
M1

j=0
M1

i=0

j
C
ij
P
j
(
i
) =
M1

i=0
_
_
M1

j=0

j
C
ij
P
j
(
i
)
_
_
=
M1

i=0
_
_
M1

j=0

j
C
ij
_

i
p
j
(y)(dy)
_
_
=
M1

i=0
_

i
_
_
M1

j=0

j
C
ij
p
j
(y)
_
_
(dy).
Thus, by inspection, we see that the Bayes rule has decision regions given by

i
=
_
_
_
y

M1

j=0

j
C
ij
p
j
(y) = min
0kM1
M1

j=0

j
C
kj
p
j
(y)
_
_
_
.
Exercise 19:
a. The likelihood ratio is given by
L(y) =

n
k=1
1

2
1
e
(y
k

1
)
2
/2
2
1

n
k=1
1

2
0
e
(y
k

0
)
2
/2
2
0
=
_

1
_
n
e
n
2
_

2
0

2
0

2
1

2
1
_
e
_
1
2
2
0

1
2
2
1
_

n
k=1
y
2
k
e
_

2
1

2
0
_

n
k=1
y
k
,
8
which shows the structure indicated.
b. If
1
=
0
and
2
1
>
2
0
, then the Neyman-Pearson test operates by comparing
the quantity

n
k=1
(y
k
)
2
to a threshold, choosing H
1
if the threshold is exceeded and
H
0
otherwise. Alternatively, if
1
>
0
and
2
1
=
2
0
, then the NP test compares

n
k=1
y
k
to a threshold, again choosing H
1
when the threshold is exceeded. Note that, in the rst
case, the test statistic is quadratic in the observations, and in the second case it is linear.
c. For n = 1,
1
=
0
and
2
1
>
2
0
,, the NP test is of the form

NP
(y) =
_
1 if (y
1
)
2

0 if (y
1
)
2
<

,
where

> 0 is an appropriate threshold. We have

P
F
(
NP
) = P
0
((Y
1
)
2
>

) = 1 P
0
(
_

Y
1

_

)
= 1
_

0
_
+
_

0
_
= 2
_
1
_

0
__
.
Thus, for size we set

=
_

1
_
1

2
__
2
,
and the detection probability is
P
D
(
NP
) = 1 P
1
(
_

Y
1

_

) = 2
_
1
_

1
__
= 2
_
1
_

1
_
1

2
___
, 0 < < 1.
9