You are on page 1of 2

DETECTION OF RADAR SIGNALS IN NOISE 37$"""

the first and third integrals or Eq. (10.27) are absorbed in the constant k. An a posteriori
:r, if it could be built, is one whose output is given by the above equation. If the receiver
I (the a posteriori probability) is greater than a predetermined threshold, a target is said
nesent. If the a priori probability p(SN) can be considered constant, the computation of
posteriori probability is equivalent to multiplying the received signal y(t) by the signal
inn *,(/) and integrating with respect to time. This is the same process performed by the
:orrelation receiver (Sec. 10.3) and is equivalent to the operation of a matched filter
10.2).
limitation of the method of inverse probability based on the application of Bayes' rule is
Ticulty of specifying the a priori probabilities. In most cases of practical interest, one is
mt of the a priori probabilities. For example, it would be necessary to specify the a priori
bility of finding a target at any particular range at any particular time. This is an almost
sible task. In the absence of better data, it might be assumed that all range intervals are y
probable a priori, and the a priori probability may be considered to be constant.
ver, such an assumption applied blindly to computations involving inverse probability
imetimes lead to erroneous and contradictory conclusions. 29 This difficulty in specifying
priori probability was recognized by Woodward and Davies.27 They suggest, however, ie
a priori factor be omitted from the inverse-probability specification when it is doufet-id
in practice it may be supplied subjectively by the human observer. This merely begs
nestion, for it has not been proved that an operator can supply the necessary a priori
ibility, and in addition, there are many applications where no operator is involved in
ig the detection decision. Nevertheless, it may be stated that whenever the a priori
ibilities are known, the inverse-probability method may be used with confidence. When
priori probabilities are not known, the likelihood-ratio test is usually employed.

ionship of inverse probability and likelihood-ratio receivers. The receiver input y may be
r signal-plus-noise or noise alone. Therefore the probability of the event y may be ex-
ed as
p(y) = p(y | SN)p(SN) + p(y | N)p(S) (10.29)
a posteriori probability of Eq. (10.23) becomes

in terms of the likelihood ratio30

P (y\N)

refore, if a receiver can be built which computes the likelihood ratio and if the a priori
jabilily p(SN) is known, the a posteriori probability can be calculated. Since p(SN\y) is a
lotonic function of Lr(y), the output of the likelihood receiver (or the matched-filter
iver) can be calibrated directly in terms of the a posteriori probability. The chief difference
veen the two representations is that the concept ofinverse probability requires a knowlege
le a priori probabilities whereas the likelihood ratio does not. (The likelihood ratio follows I
inverse probability if the assumption is made that the a priori probabilities are equally
ly.) Both the a posteriori method and the likelihood method may be implemented by
iputing the cross-correlation function between the received signal and the signal s,(t).
8 INTRODUCTION TO RADAR SYSTEMS

Let the event x » SN represent signal-plus-noise, and let the event y be the receiver input,
which may consist of either signal-plus-noise or noise alone. Equation (10.22) may be
rewritten as

p(y) (10.23)
This is Bayes' rule. It expresses the (a posteriori) probability that the signal is present, given
that the receiver input is y. f
For a particular input y, the receiver can assess from Eq. (10.23) the probability that a
particular signal was received. Since y will then be fixed, the denominator p(y) will be constant
and the a posteriori probability is
= kp(SN)p(y\SN) (10.24)

where the constant k is determined by the normalizing condition; that is, the integral of
p{SN | y) over all possible values must be unity. Therefore, if the a priori probability p(SN) is
known, the a posteriori probability may be found directly from Eq. (10.24) once p(y|SN) has
been evaluated. If the received waveform y(() as a function of time consists of the signal
waveform st(t) plus the white gaussian noise waveform n(t) = y(i) — Sj(t)< Woodward and
Davies27 show that

p(y\SN) = pJin{t)] = pn[y(t) - .,,(«)] oc exp - --,- n*(t) dt (10.25)

where pr[n(r)] •» probability-density function for noise waveform n(r) and Na = mean noise
power per unit bandwidth (dimensions of energy). With this substitution, the a posteriori
probability for the signal s,(l) becomes

(10.26)

The integral in this expression is a definite one, with limits defined by the duration of the
observation time (0 -► To).
Equation (10.26) forms the basis of the technique used by Woodward and Davies for the
analysis of radar reception problems when the interference is caused by white gaussian noise.
Except for the a priori weighting factor p[SN), Eq. (10.26) shows that the most probable
w&veform st(t) is the one which has the least-mean-square deviation from the received
waveform y(r).
The computation of the a posteriori probability might be accomplished by computing the
cross-correlation function between the actual waveform and the various possible waveforms
that might be received. Expanding the integral in Eq. (10.26) we get

j4?(l)df (10.27)

Upon reception, the waveform y(t) is known, so that the first integral on the right-hand side of
the equation is constant and can be absorbed in the constant k. The last integral is the energy C
contained within the signal s,(t) and also is a constant. The second integral is not a constant.
The a posteriori probability can be written

p(SN\y) = kp(SN) exp (10.28)

You might also like