Professional Documents
Culture Documents
Signal Detection
Signal Detection
Prior information
Observations
Hypothesis
Testing
Decision outcome
available. Based on the likelihood ratio (LR)1 hypothesis test; three well-known
hypothesis tests are given as follows:
(1) Maximum a posteriori probability (MAP) hypothesis test.
(2) Baysian hypothesis test.
(3) Maximum likelihood (ML) hypothesis test.
In the following section, the hypothesis tests in the above are illustrated respectively.
For example, suppose that 10 balls are drawn from each box as shown in Fig. 2.2.
Based on the empirical distribution results in Fig. 2.2, conditional distributions of the
number on balls are given by
4
;
10
3
Pr (2|H0 ) =
;
10
3
Pr (3|H0 ) =
;
10
Pr (1|H0 ) =
and
Note that in Chaps. 2 and 3, we use LR to denote the term likelihood ratio, while in the later
chapters of the book, the LR is used to represent lattice reduction.
{3,1,2,1,1,3,1,2,3}
{3,1,3,4,1,2,4,2,3,3}
Observation (Knowledge)
1
D
A new observation
2
;
10
2
Pr (2|H1 ) =
;
10
4
;
Pr (3|H1 ) =
10
2
Pr (4|H1 ) =
.
10
Pr (1|H1 ) =
where Pr (n) denotes the probability that the ball with number n is drawn. Taking
Pr(Hk ) as the a priori probability (APRP) of Hk , the a posteriori probability (APP)
of Hk is shown as follows:
2
Pr (H0 |1) = ;
3
3
Pr (H0 |2) = ;
5
3
Pr (H0 |3) = ;
7
Pr (H0 |4) = 0,
and
1
;
3
2
Pr (H1 |2) = ;
5
4
Pr (H1 |3) = ;
7
Pr (H1 |4) = 1.
Pr (H1 |1) =
Here, Pr (Hk |n) is formed as the conditional probability that the hypothesis Hk is
true under the condition that the number on the drawn ball is n. For example, if the
number of the ball is n = 1, since Pr (H0 |1) = 23 is greater than Pr (H1 |1) = 13 , we
can decide that the ball is drawn from box A, where the hypothesis H0 is accepted.
The corresponding decision rule is named as the MAP hypothesis testing, since we
choose the hypothesis that maximizes the APP.
Generally, in the binary hypothesis testing, H0 and H1 are referred to as the null
hypothesis and the alternative hypothesis, respectively. Under the assumption that
the APRPs Pr(H0 ) and Pr(H1 ) are known and the conditional probability, Pr(Y |Hk ),
is given, where Y denotes the random variable for an observation, the MAP decision
rule for binary hypothesis testing is given by
(2.2)
Pr (H1 )
Pr (Y = y|H0 )
>
;
H0 : Pr
Pr (H0 )
(Y = y|H1 )
(2.3)
Pr (H1 )
Pr (Y = y|H0 )
H1 :
<
.
Pr (Y = y|H1 )
Pr (H0 )
0) =
(0,2 )
f(y
9
1) =
(s,2 )
exp (x)
2
2
N , 2 =
.
2 2
(2.4)
Let the noise n be a Gaussian random variable with mean zero and variance , while
s be a positive constant. Consider the case that a constant signal, s, is transmitted,
while a received signal, y, may be corrupted by the noise, n, as shown in Fig. 2.3.
Then, we can have the following hypothesis pair to decide whether or not s is present
when y is corrupted by n:
H0 : y = n;
(2.5)
H1 : y = s + n.
Then, as shown in Fig. 2.3, we have
f (y|H0 ) = N 0, 2 ;
f (y|H1 ) = N s, 2 ,
(2.6)
s(2y s)
f (Y = y|H0 )
,
= exp
f (Y = y|H1 )
2 2
(2.7)
and
Pr(H0 )
,
Pr(H1 )
10
Table 2.1 MAP decision rule
H0
H1
H0 is true
H1 is true
Correct
Type II (miss)
2 ln
s
;
H0 : y < +
2
s
2
H1 : y > s + ln .
2
s
Error probability
PA
PB
(2.8)
Since the decision rule is a function of y, we can express the decision rule as
follows:
r (y) = 0 : y A0 ;
(2.9)
r (y) = 1 : y A1 ,
where A0 and A1 represent the decision regions of H0 and H1 , respectively. Therefore, for the MAP decision rule in (2.8), the corresponding decision regions are
given by
2 ln
s
;
A0 = y | y <
2
s
(2.10)
2 ln
s
A1 = y | y >
,
2
s
where A0 and A1 are regarded as the acceptance region and the rejection/critical
region, respectively, in the binary hypothesis testing.
Table 2.1 shows four possible cases of decision, where type I and II errors are
usually carried out to analyze the performance. Note that since the null hypothesis,
H0 , normally represents the case that no signal is present, while the other hypothesis,
H1 , represents the case that a signal is present, the probabilities of type I and II errors
are regarded as the false alarm and miss probabilities, respectively. The two types of
decision errors are summarized in Table 2.2. Using the decision rule r (y), PA and
PB are given by
PA = Pr (Y A1 |H0 )
= r (y) f (y|H0 )dy
= E [r (Y )|H0 ]
and
(2.11)
(2.12)
(2.13)
11
PB = Pr (Y A0 |H1 )
= (1 r (y)) f (y|H1 )dy
(2.14)
= E [(1 r (Y ))|H1 ] ,
(2.16)
(2.15)
(2.17)
(2.18)
(2.19)
(2.20)
(2.21)
Let Ac0 denote the complementary set of the decision region, A0 , and assume that
A1 = Ac0 for convenience. Since
Pr (D1 |Hk ) = 1 (D0 |Hk ) ,
(2.22)
A0
g1 (y) g0 (y)dy,
(2.23)
(2.24)
Then, it is possible to minimize the average cost G by properly defining the acceptance regions, while the problem is formulated as
12
min G.
(2.25)
A0 ,A1
G = Constant +
we can show that
min G min
A0
A0
A0
g1 (y) g0 (y)dy,
A0
(2.26)
g1 (y) g0 (y)dy ,
(2.27)
while the optimal regions that minimize the cost are given by
(2.28)
Hence, we can conclude the Baysian decision rule that minimizes the cost as follows:
or
Pr(H0 ) G 01 G 11
f (y|H0 )
H1 :
<
,
f (y|H1 )
Pr(H1 ) G 10 G 00
(2.29)
(2.30)
where (G 10 G 00 ) and (C01 C11 ) are positive. More importantly, for binary
G 01 G 11
,
hypothesis testing, we can find out that the ratio of the cost differences, G
10 G 00
is able to characterize the Baysian decision rule rather than the values of individual
G 01 G 11
= 1, the Baysian hypothesis test becomes the
costs, G ik s. Specifically, as G
10 G 00
MAP hypothesis test.
f 0 (y) = f (y|H0 );
f 1 (y) = f (y|H1 ).
(2.31)
13
Notice that the likelihood function is not a function of y since y is given, but a
function of the hypothesis. With respect to the ML function, the ML decision rule is
to choose the hypothesis as follows:
f 0 (y)
H0 : f (y) > 1;
H0 : f 0 (y) > f 1 (y);
1
H1 :
< 1,
f 1 (y)
f 0 (y)
f 1 (y) ,
(2.32)
f 0 (y)
f 1 (y)
(2.33)
the e-based log-likelihood ratio (LLR), the ML decision rule can be rewritten as
H0 : LLR(y) > 0;
H1 : LLR(y) < 0.
(2.34)
Note that the ML decision rule can be considered as a special case of the MAP
decision rule when the APRPs are the same, i.e., Pr(H0 ) = Pr(H1 ). In this case, the
MAP decision rule is reduced to the ML decision rule.
f 0 (y)
H0 : f (y) > ;
1
f 0 (y)
H1 :
< ,
f 1 (y)
(2.35)
f 0 (y) = N (0 , 2 );
f 1 (y) = N (1 , 2 ),
(2.37)
14
MAP
decision
rule
ML
decision
rule
0 +1
2
(2.38)
0 + 1
<
H0 : y
2
0 + 1
>
H1 : y
2
2 ln
;
1 0
2 ln
.
1 0
(2.39)
Letting
=
(2.39) can be rewritten as
2 ln
0 + 1
,
+
1 0
2
Specifically, if we let
=
H0 : y < ;
H1 : y > .
Pr(H0 )(G 10 G 00 )
,
Pr(H1 )(G 01 G 11 )
(2.40)
(2.41)
(2.42)
then the LR-based decision rule becomes the Baysian decision rule. In summary, The
LR-based decision rule can be regarded as a generalization of the MAP, Baysian,
and ML decision rules, where the relationship of various decision rules for binary
hypothesis testing is shown in Fig. 2.4.
15
subject to
PA (d) ,
(2.43)
where d and represent a decision rule and the maximum false alarm probabilities,
respectively. In order to find the decision rule, d, with the maximum false alarm
probability constraint, the NeymanPearson Lemma is presented as follows.
Lemma 2.1. Let the decision rule of d (s) be
1,
if f 1 (s) > f 0 (s);
1,
with
probability
p;
d (s) = (s) =
if f 1 (s) = f 0 (s);
0, with probability 1 p,
0,
if f 1 (s) < f 0 (s),
(2.44)
where > 0 and p is decided such that PA = . The decision rule in (2.44) is
named as the NeymanPearson (NP) rule and becomes the solution of the problem
in (2.43).
Proof. Under the condition of PA , we can assume that the decision rule becomes
In order to show the optimality of the problem in (2.43), we need to verify that
d.
we have PD (t ) PD (t). Denote by S the observation set, where s S.
for any d,
Using the definition of d , for any s S, it shows that
d (s) d(s)
( f 1 (s) f 0 (s)) 0.
(2.45)
d (s) d(s)
( f 1 (s) f 0 (s)) ds 0
(2.46)
and
sS
d (s) f 1 (s)ds
sS
f 1 (s)ds
d(s)
sS
d (s) f 0 (s)ds
sS
f 0 (s)ds ,
d(s)
(2.47)
PA (d ) PA (d)
0.
PD (d ) PD (d)
(2.48)
16
From (2.43), we can show that the NP decision rule is the same as the LR-based
decision rule with the threshold = 1 except the randomized rule, i.e., f 1 (s) =
f 0 (s).
Example 2.2. Let us consider the following hypothesis pair:
H0 : y = n;
H1 : y = s + n,
(2.49)
where s > 0 and n N (0, 2 ). Using (2.41), the type I error probability is shown as
=
=
d(y) f 0 (y)dy
f 0 (y)dy
y2
exp 2 dy
=
2 2
=Q
,
Q(x) =
x
1
2
et /2 dt.
2
(2.50)
Note that the function Q(x), x 0, is the tail of the normalized Gaussian pdf (i.e.,
N (0, 1)) from x to . Then, it follows that
=Q
or
(2.51)
= Q1 ().
(2.52)
=
f 1 (y)dy
(y s)2
exp
2N02
2 N02
1
dy
17
1
0.9
0.8
0.7
PD
0.6
0.5
0.4
0.3
s=2
s=1
s = 0.5
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
s
=Q
,
= Q Q1 ()
(2.53)
which shows that the detection probability is a function of the type I error probability.
In Fig. 2.5, the receiver operating characteristics (ROCs) of the NP decision rule
are shown for different values of s. Note that the ROCs are regarded as the relationship
between the detection and type I error probabilities.
H0 : y = s + n;
H1 : y = s + n,
(2.54)
18
where y is the received signal, s > 0, and n N (0, 2 ). According to the symmetry
of transmitted signals s and s, the probabilities of type I and II errors become the
same and given by
PError = Pr(Accept H0 | H1 )
= Pr(Accept H1 | H0 ).
(2.55)
(2.56)
1
= 2 (Y s)2 (Y + s)2
2
2s
= 2 Y.
LLR(Y ) = log
(2.57)
H0 : Y > 0;
H1 : Y < 0.
(2.58)
From this, we can show that the ML detection is simply a hard-decision of the
observation Y = y.
(2.59)
19
2
2
0
(2.60)
and
1
Y2
dY
=
exp
s
2
2
.
=Q
PError
(2.61)
Letting the signal-to-noise ratio (SNR) be SNR = s 2 , then we can have PError =
Q( SNR). Note that the error probability decreases with the SNR, since Q is a
decreasing function.
exp y 2 dy
(2.62)
2
exp t 2 dt, for y > 0,
=
y
(2.63)
where the relationship between the Q-function and the complementary error function
is given by
2y ;
y
1
Q(y) = erfc .
2
2
erfc(y) = 2Q
(2.64)
Then, for a given Y = y, we can show that the complementary error function has
the following bounds:
20
1
1
2Y 2
exp Y 2
exp Y 2
< erfc(Y ) <
,
Y
Y
(2.65)
where the lower bound is valid if Y > 1/ 2. Accordingly, the Q-function Q(Y ) is
bounded as follows:
exp Y 2 /2
1 exp Y 2 /2
1 2
< Q(Y ) <
,
(2.66)
Y
2Y
2Y
where the lower bound is valid if x > 1. Thus, the upper bound of error probability
is given by
exp (SNR/2)
PError = Q( SNR) <
.
(2.67)
2 SNR
In order to obtain an upper bound on the probability of an event that happens rarely,
the Chernoff bound is widely considered, which can be used for any background
noise.
Let p(y) denote the step function, where p(y) = 1, if y 0, and p(y) = 0, if
y < 0. Using the step function, the probability for the event that Y y, which can
be also regarded as the tail probability, is given by
Pr(Y y) =
=
f Y ()d
p( y) f Y ()d
= E[ p(Y y)],
(2.68)
where f Y () represents the pdf of Y . From Fig. 2.6, we can show that
p(y) exp(y).
(2.69)
(2.70)
Pr(Y y) E exp ((Y y))
(2.71)
or
for 0. By minimizing the right-hand side with respect to in (2.71), the tightest
upper bound can be obtained which is regarded as the Chernoff bound and given by
Pr(Y y) min exp (y) E exp (Y ) .
0
In (2.72), E exp (Y ) is called the moment generating function (MGF).
(2.72)
21
p(y)
0
Let Y be a Gaussian random variable with mean and variance 2 . The MGF of
Y is shown as
1
(2.73)
E[exp (Y )] = exp + 2 2 ,
2
while the corresponding Chernoff bound is given by
1 2 2
Pr(Y y) min exp (y) exp +
0
2
1 2 2
= min exp ( y) + .
0
2
The solution of the minimization is found as
1 2 2
(2.74)
(2.75)
(2.76)
With respect to the error probability in (2.60), the Chernoff bound is given by
PError
SNR
s2
.
exp 2 = exp
2
2
(2.77)
22
2
y
,
Q(y) exp
2
(2.78)
lT
L
(l1)T
L
lT
L
(l1)T
L
lT
L
y(t)dt,
sm (t)dt,
n(t)dt,
(2.81)
H0 : yl = s0,l + nl ;
H1 : yl = s1,l + nl .
(2.82)
(l1)T
L
where
In addition, since N (t) is a white process, the nl s are independent, while the mean
of nl is zero and the variance can be derived as
23
2 = E[nl2 ]
= E
=
=
lT
L
(l1)T
L
lT
L
lT
L
(l1)T
L
(l1)T
L
lT
L
(l1)T
L
lT
L
(l1)T
L
2
N (t)dt
N0 T
.
2L
(2.83)
(2.84)
L
l=1
L
l=1
log
f 0 (yl )
f 1 (yl )
1
2
2
(yl s0,l ) (yl s1,l )
log exp
N0
1
(yl s1,l )2 (yl s0,l )2
N0
l=1
L
L
1
2
2
=
2yl (s0,l s1,l ) +
(s1,l s0,l ) .
N0
l=1
(2.85)
l=1
1 T
2y (s0 s1 ) (s0T s0 s1T s1 ) .
N0
(2.86)
24
H
:
y
0
1
0
Pr(H2 )
Pr(H1 )
T
2
+
H1 : y (s0 s1 ) < log
Pr(H2 )
For the LR-based decision rule, we can replace
1 T
(s s0 s1T s1 );
2 0
1 T
(s s0 s1T s1 ).
2 0
Pr(H0 )
Pr(H1 )
(2.87)
(2.88)
As the number of samples during T seconds is small, there could be signal information loss due to sampling operation using the approach in (2.81). To avoid any
signal information loss, suppose that L is sufficiently large to approximate as
yT si
1
T
y(t)sm (t)dt.
(2.89)
T
1 T
2
s0 (t) s12 (t) dt;
s02 (t) s12 (t) dt,
H0 : T y(t)(s0 (t) s1 (t))dt > WT ;
0
H : T y(t)(s (t) s (t))dt < W ,
1
0
1
T
0
if we let
1
WT = log +
2
(2.90)
(2.91)
(2.92)
Note that the decision rule is named as the correlator detector, which can be implemented as in Fig. 2.7.
Let us then analyze its performance. For the ML decision rule, we can have = 1
T
in the LR-based decision rule, which leads to WT = 21 0 s02 (t) s12 (t) dt. Letting
T
X = 0 y(t) (s0 (t) s1 (t)) dt WT , we can show that
(2.93)
25
In order to derive the error probabilities, the random variable X has to be characterized. Note that X is a Gaussian random variable, since N (t) is assumed to be a
Gaussian process. On the other hand, if Hm is true, the statistical properties of G
depend on Hm as Y (t) = sm (t)+ N (t). Then, letting Hm be true, to fully characterize
X , the Gaussian mean and variance of X are given by
E[X |Hm ] =
=
(2.94)
and
2
= E[(X E[X |Hm ])2 |Hm ],
m
(2.95)
(g E s (1 ))2
exp
2N0 E s (1 )
f 0 (g) =
2 N0 E s (1 )
(g + E s (1 ))2
exp
2N0 E s (1 )
f 1 (g) =
.
2 N0 E s (1 )
(2.97)
26
By following the same approach in Sect. 2.7.1, the error probability can be derived as
PError = Q
E s (1 )
.
N0
(2.98)
For a fixed signal energy, E s , since the error probability can be minimized when
= 1, the corresponding minimum error probability can be written as
PError = Q
2E s
N0
.
(2.99)
Note that the resulting signals that minimize the error probability are regarded as
antipodal signals, which can be easily shown as s0 (t) = s1 (t). In addition, for an
orthogonal signal set where = 0, we have the corresponding error probability to be
PError = Q
Es
N0
.
(2.100)
Therefore, it can be shown that there is a 3 dB gap (in SNR) between the antipodal
signal set in (2.99) and orthogonal signal set in (2.100).
log2 M
T
(2.101)
bits per seconds. In (2.101), it can be observed that the transmission rate increases
with M, while a large M would be preferable. However, the detection performance
becomes worse as M increases in general.
Let the received signal be
y(t) = sm (t) + n(t)
(2.102)
for 0 t < T under the mth hypothesis. The likelihood with L samples is shown as
27
f m (y) =
L
f m (yl )
l=1
L
l=1 exp
(y s )2
l N0m,l
L
( N0 ) 2
(2.103)
Taking the logarithm on (2.103), we can rewrite the log-likelihood function as follows:
log f m (y) = log
= log
1
L
( N0 ) 2
1
( N0 )
L
2
L
l=1
(yl sm,l )2
log exp
N0
(yl sm,l )2
,
N0
(2.104)
(2.105)
l=1
by canceling the common terms for all the hypotheses. In addition, (2.105) can be
rewritten as
T
1
1 T 2
log f m (y(t)) =
y(t)sm (t)dt
sm (t)dt ,
(2.106)
N0
2 0
0
T
when L goes infinity. Note that E m = 0 sm2 (t)dt in (2.106) represents the signal
T
energy, while 0 y(t)sm (t)dt denotes the correlation between y(t) and sm (t).
Based on the log-likelihood functions, the ML decision rule can be found by
If log f m (y(t)) log f m (y(t)) accept Hm
or
If log
f m (y(t))
0 accept Hm ,
f m (y(t))
(2.107)
(2.108)
28
http://www.springer.com/978-3-319-04983-0