22 IEEE
1, JANUARY 1969
Detection and Parameter Estimation in an
Amplitude-Comparison Monopulse Radar
EDWARD M. HOFSTETTER, MEMBER, IEEE,
Abstract-This paper considers the problem of detecting and
estimating the unknown angular location of a radar target that is ob-
served simultaneously by a number of antennas. The amplitude of
the signal received by a particular antenna is assumed to depend on
the angular location of the target, but the time of arrival of the signal
is assumed to be the same at all of the antennas.
The generalized likelihood ratio test is used to derive the detec-
tion and parameter-estimation strategy for the radar receiver; Ex-
plicit expressions for the detector and the angle estimates are derived
in a number of important special cases, and the performance of the
detector is evaluated. Accuracy formulas for the angle estimates,
valid for high signal-to-noise ratios, are derived for the general
problem and used to compare the performance of an optimum four-
beam monopulse system with a type of monopulse system commonly
in use.
INTRODUCTION
HIS PAPER is concerned with the problem of
detection and parameter estimation in an amplitude-
T
comparison monopulse radar. The term “amplitude
I comparison” refers to the fact that information about the
angular location of a target in space is derived from a
cluster of antenna beams whose gain patterns are skewed
in angle so that the received signal appears in different
antenna beams with amplitudes that depend on the target
angle relative to the antenna complex. The antennas are so
designed that the arrival times of the received waveform
at the different antennas are nearly identical [I], [2].
The basic goal of this paper is to apply the generalized
likelihood ratio test [3], [4] to derive a strategy for detect-
ing targets whose amplitude, carrier phase, and angular
location are unknown and to obtain estimates of these un-
known parameters. The solution to this problem will be
obtained and then extended to the case where additional
target parameters such as range and velocity are unknown.
The detection characteristic (probability of detection ver-
sus probability of false alarm) of the maximum-likelihood
receiver will be derived in certain special cases, and
expressions for the high signal-to-noise ratio covariance
matrix of the angle estimates will be obtained.
There are actually several different versions of the prob-
lem described above, which depend on the physical
situation at hand. First, the problem can be either one-
or two-dimensional depending on whether the target has
to be located in only one angular coordinate such as
azimuth or elevation or whether it has to be located in
both angular coordinates. In the first case a minimum of
Manuscript received April 3, 1968; revised May 29, 1968.
The authors are with M.I.T. Lincoln Laboratory, Lexington,
Mass. 02173, operated with support from the Advanced Research
Projects Agency.
Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.
TRANSACTIONS ON INFORMATION THEORT, VOL. W-15, NO.
I
AND DARROL F. DELONG, JR., MEMBER, IEEE
two beams is required, in the second case a minimum of
three beams. Second, different mathematical problems
arise depending on whether it is assumed that the target
returns received on the various beams are coherent or
incoherent with respect to one another. Solutions can be
obtained for all of these versions of the problem, but the
degree to which the solution can be made explicit and the
degree to which the performance of the resulting receiver
can be analyzed depend strongly on which version is under
consideration. Fortunately, the analysis can be completed
to a satisfactory degree in most cases of interest.
There does not seem to be a great deal of literature
available on the problem considered in this paper. The
most pertinent reference is a recent paper by McGinn 151,
which discusses the estimation of a single target location
angle and perhaps other unknown target parameters
given two coherent receiver beams. The ‘problems of
detection, two-dimensional angle estimation, and incoher-
ent receiver beams are not considered. Furthermore, the
approach taken by the present paper does not require the
beam-shape assumptions used by McGinn, and by explic-
itly treating carrier phase as an unknown parameter it
obtains a number of new results.
Another pertinent reference is a paper by Urkowitz [6],
which discusses the maximum angular estimation accuracy
achievable by means of a physical antenna aperture. This
is a more general problem than the one treated in the
present paper in that Urkowitz considers antennas that
are both phase and amplitude sensitive, but his accuracy
formulas are not as explicit as those derived here for the
special case of amplitude-sensitive antennas, and he does
not attempt to derive the structure of the optimum re-
ceiver that achieves these accuracies. The problem of
combined detection and angle estimation is not considered
in his paper.
DERIVATION OF THE MAXIMUM-LIKELIHOOD RECEIVER
Incoherent Beams
The case of incoherent beams will be considered first
because the mathematics involved is simpler than in the
coherent case. It is assumed that both angular coordinates
of the target must be estimated and that m beams, m ‘2 3,
are used. The complex envelopes r,(t) of the received
waveforms are of the form
r,(t) = Ae”‘G,(B, cp>s(t)+ n,(t) i = 1, ... , m, (1)
where A 2 0 denotes the unknown received signal ampli-
tude, tii the unknown carrier phase of the signal in the ith
beam, and s(t) the known complex envelope of the trans-
 HOFSTETTER AND DELONG : DETECTION AND PARAMETER ESTIMATION

mitted signal. For convenience this signal is normalized so which can be simplified with the result,
that J Is(t)/’ dt = 1. The ni denote the complex envelopes
of zero-mean white Gaussian noises that are independent 1 = 4N, ln L
from channel to channel. In symbols,’
- A*G:(e,0 I ,
R[n,(t)nt(t’)] = 4N, 6;, 6(t - t’)
(2) (5)
E[n,(t)n,(t’)] = 0.
where
The real function Gi (0, ‘p) denotes the beam pattern of the 7
ith beam measured with respect to the angular coordinates y; = r,(t)s*(t) dt. (6)
0 and cp.These coordinates may be thought of as azimuth /0
and elevation angles, but any two other coordinates capa- The complex voltages ?J<can be formed by passing ri(t)
ble of defining a direction in space could be used. Thus, the through a filter matched to the known waveform s(t).
term AG, (0, ‘p) denotes the amplitude of the signal received The maximization on A is accomplished by differenti-
on the ith beam when the angular coordinates of the target ating the summation appearing in (5) with respect to A
are 0 and cp. The assumption that the G, are real reflects and setting this derivative equal to zero. The resulting
the fact that the arrival times of the received waveform equation is
at the different antennas are assumed to be identical re-
gardless of the angular location of the target. This is never
2 {2G,(0, p) Re [y&j”] - 2AGf(0, p)) = 0
exactly the case in a practical amplitude-comparison r=l
monopulse system, but careful antenna design can yield a
or
system for which it is a reasonable assumption.
The beam-to-beam incoherence of the problem comes
about through the assumption that the carrier phases $(
are unknown and not necessarily the same in each channel. (7)
When the coherent case is treated, a single common un-
known carrier phase will be postulated.
Substitut,ion of (7) in (5) yields the result
Stated in mathematical language, the problem at hand
is to observe the time waveforms y,(t), i = 1, . . . , m in the
observation interval’ 0 5 t < T and then to decide
whether a target is present or not (A 2 0 or A = 0). If the
decision is “target present,” estimates of the unknown
parameters, A, #<, 8, and cpare to be obtained. The gener-
alized likelihood ratio test will be used to perform this The next step is to perform the maximization with
task. This test requires the receiver to calculate the respect to the parameters 4. Inspection of (8) shows that
quantity” it is maximized when
max p[r(O, 0 I t 5 T I 8, 0, cp,41 tii = arg [y,G,(O,~11
L= a.t,s,!s
p[r(t), 0 5 t < T 1 A = 0] ’ (3)
and this maximum is given by
where r(t) denotes the set of values rl(t), . . . , r%(t) and 4
denotes the set of values +$, 3 . . , lClrn,and to compare it [ 2 IYLIlGt(e,‘PII]’
with a present threshold A. If L < h, the decision “target 1 = rnaxLcL-.-. (9)
absent” is made. If L > A, “target present” is announced
a.9 5 (80, ‘PI
and the values of A, #;, 0, and P that achieved the maxi-
mum -in (3) are taken as estimates of the true values of The next step would be to maximize with respect to 6
these paramet’ers. Making use of (1) and (a), (3) can be and cp; however, no way of explicitly performing this op-
written eration has been found when m is greater than 3. To
see the reason for this, note that the Schwarz inequality
applied to (9) yields

1 The asterisk denotes complex conjugation. This complex noise

corresponds to a real white noise with a two-sided power density
of No watts per cycle.
2 T is assumed to be much larger than the duration of s(t). and furthermore, equality is achieved in this bound,
3 Strictlv sDeakina. neither numerator nor denominator of this I = CT=1 jy*j’ if and only if
expression “exists. B&, Grenander has shown that it is possible to
define a suitable likelihood ratio which approaches the quantity
given by (3) in the limit as the noises involved become white. lGd’4cp)l= Q bil i = 1, ..f ,m, 01)

Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.
 24 IEEE TRANSACTIONS ON INFORMATION THEORY, JANUARY 1969 ,_
where a! denotes an arbitrary proportionality constant. If where f&, ‘pO denote the true angular coordinates of the
m = 3, (11) can be satisfied by choosing 8 and 4 such that target. The right-hand side of (16), considered as a func-
tion of $, p0 for fixed 19,cp has its maximum at B0 = 0,
cpO= cp.Thus, it represents a beam pointing in the direction
defined by the angular coordinates 0, 43.The shape of this
and then choosing (Y = ]G,(e, @)]/]ys]. Examples of gain beam depends, of course, on the detailed shapes of the
functions Gi can be constructed for which (12) cannot be original beam patterns Gi. It is now obvious that the quan-
solved for 8, @ for all possible Iv<]. Therefore, it will be tity 1 is the maximum response that results when the
necessary to assume that the functions Gi are of such a synthetic beam is steered to all possible directions in
form that the equations above can be solved for 8, @ for all search of the target. (The denominator of (9) is simply a
non-negative values on their right-hand sides. This as- normalization factor that forces the maximum gains of all
sumption will be in force, wherever it is applicable, the synthetic beams to be equal.)
throughout the remainder of the paper. a
When m > 3, there are more equations than unknowns Coherent Beams
and no solution to (11) will exist in general. In this case, In this’ case, the complex envelopes of the signals re-
ceived on the m beams are given by the expressions
*=I
r,(t) = Aei’Gi(e, v)s(t) + n,(t) i = 1, ... , m,
and no explicit way of performing the maximization re- where the meaning of all symbols is the same as in the
quired by (9) has been found. incoherent case. The only difference between the coherent
The results for the case m = 3 can be summarized by and incoherent cases is that the carrier phase angle 9, al-
saying that the maximum-likelihood strategy leads to a though unknown, is assumed to be the same in all receiver
receiver that computes the quantity beams.
l = IYr12 + IY212+ IY312 The likelihood ratio L can be derived using the same
(13)
techniques as in the incoherent case, with the result
and compares it with a threshold. If I < A, “target absent”
is announced. If 1 > X, “target present” is announced, and 1=4N,lnL
c
the values of 8 and $ that solve (12) are taken as estimates
of the true values of these parameters. = ATa;texpg (2AG,(e, p) Re [yie-‘“I - A’G?(B,(p)].
When only one target, angle is unknown or when fan .I, =
beams are used so that the antenna gains are only a func- 07)
tion of one target angle, then a minimum of two beams The maximization on A proceeds exactly as before and
must be used to estimate the unknown angle. By an leads to
analysis that is identical, almost word for word, to the one
just given, it can be shown that, for m = 2, the receiver
compares the quantity
I’= lYl12 + IY212 (14)
with a threshold and uses the value of 13satisfying
Upon rewriting (18) as follows,
2
(15) Re 2 yiGi(O, ,p)eeip
i=l 11
as the estimate of 8. Once again, it will be assumed that the
Gi are of such a form that (15) has a solution for all non-
1 = max
*.e.r { [’ i=,
negative values of its right-hand side.
Even in those cases where it is not possible to give an it is seen readily that its maximum with respect to J, occurs
explicit expression for 1, it is still possible to give an in- for
teresting and useful physical interpretation of this quan-
tity. The crux of this interpretation lies in the fact that a cm
linear combination of the beam voltages ]yil results in
a synthetic beam pointing in some direction in between the and is given by
directions of the original beams. More precisely, if there is
no system noise,

Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.
HOFSTETTER AND DELONG : DETECTION AND PARAMETER
The expression for 1 given by (21) can be given a syn-
thetic beam interpretation similar to the one given for the
incoherent problem. In fact, the only difference between
the two interpretations is that in the coherent case the
synthetic beams are formed at RF; in the incoherent case
they were formed at video.
The analysis so far has differed very little from that per-
formed for the incoherent case. The next step is to perform
the maximization on e and cpindicated in (al), and it is at
this step that a real difference between, the two cases arises.
In the incoherent case, the Schwarz inequality was used for
this purpose, but in the coherent case the upper bound so
derived cannot be achieved because the Gi are real whereas
the yi are complex.
In one special case (two beams measuring one angle) it is
possible to perform the required maximization directly,
but this process is quite tedious and it seems more fruitful
to attack the problem from another viewpoint.
The starting point for this approach is the expression for
1 given by (18), from which the carrier phase angle 9 has
not yet been eliminated. Application of the Schwarz
inequality to this expression yields the bound
which can be achieved if and only if
Gi(O, 9) = a Re [y,e-‘$1 i = 1, ... ,m,
where CYdenotes an arbitrary proportionality
This bound can be re.written in the form
I 5 + 2i=l IYt12 + rn? + Re [eP+ g yf]
by using the identity Re x Re y = ;Re xy* + $Re xy.
The maximum required by (24) is obviously achieved
when # = 4 where
$ = $ arg
which results in the following
[ 1
2
i=l
yf ,
expression for the upper
bound on I,
To recapitulate, 1 is bounded above as indicated in (26),
and this upper bound is achieved if and only if (23) is
satisfied with # = 4 as given by (25). These conditions can-
not be met in general if m > 3 because there are then more
equations than unknowns. When m = 3, the equations can
be solved by choosing 6, 4 so that
where $ is given by (25), and then setting CY= G,(e, Q)/
Re [y3e-“]. When (27) can be satisfied, 1 is given by the
expression
l = ~[IYJ + 1Y212
+ IYJI + a IYT-i YE + yq.
Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.
ESTIMATION 4”
It will be necessary to assume that the gains Gi are of such
a form that (27) can be solved for 8,$ for all real values of
their right-hand sides.
The expression for I given by (28) has an interesting
interpretation. It is a mixture of an incoherent combina-
tion of the beam voltages as given by the first term of the
expression and a coherent combination given by the second
term. The coherent combination is not the usual one
(1y1 + yz + y312), but rather a coherent combination of
squares of the beam voltages. This has the effect of
weighting the larger beam voltages more heavily than the
smaller ones. It may seem surprising at first that a straight-
forward coherent combination of the beam voltages is not
the right thing to do. The reason for this is that the amp-
litude of the target return is in general not the same in
each of the beams. When this is the case, coherent combi-
nation can yield a significantly lower output signal-to-
noise ratio than incoherent combination.4
The detection function given by (28) and the angle esti-
mates given by (27) are both considerably more difficult
to realize than their counterparts in the incoherent case.
The source of this difficulty is the term jy: + yi + yil.
Several attempts have been made to rewrite the pertinent
expressions in a form not containing such a term, but so far
no success has been achieved except in an important
special case. When attention is restricted to the case of
(23) two beams measuring a single target angle, it is possible to
write the solution in a form that is both elegant and rela-
constant. tively easy to realize.
The same kind of analysis given above leads to the
solution of the two-beam problem as expressed by
(24)
l = ~tlvll” + Iv2l”l+ 4 IYT + Yil (29)
G,(d) = Re [yle-‘“I
-; (39)
G(e) iGj&Fj
d=4arg+d). (YT (31)
(25) This solution is based on the assumption that
GI(N/G,(fa = x
can be solved for 8, given any real value of x. These equa-
tions can be simplified by introducing the following change
of variables5
91 = Yl + jY2
(32)
02 = Yl - jY2.
Substitution of (32) into (29) yields
= $hl + l%l12.
4 The authors are grateful to J. Sawyers of the Hughes Aircraft
Company, Fullerton, Calif., for pointing this fact out to them.
6 The authors are indebted to J. Margolin of Lincoln Laboratory
(28) for suggesting this change of variables.
 26
Similarly, substitution of (32) into (30) and (31) yields6
ew h9 = (a)“”
and
= cot +(arg q1 - arg Q).
The realization of the two-beam receiver is now very
simple. The voltages Q and vz can be obtained from the
matched filter outputs y1 and yz by means of a hybrid junc-
tion. The detection function 1 is then formed with linear
envelope detectors and an adder, and a voltage proportion-
al to 2 cot-l G,(@/G,(& is formed by combining v1 and Q
in a phase detector. This angle estimation scheme is very
similar to some of those described in [l] and [a].
One further point is worthy of mention in connection
with the problem just considered. Many practical two-
beam monopulse systems do not work directly with the
signals yl’ and yz, but rather with the sum and difference
signals yz = y1 $ 2/Sand yA = y1 - yz. The antenna gain
functions associated with these signals are G,(e) =
G,(e) + G,(0) and G,(e) = G,(e) - G,(e), respectively.
Since it is easily verified that
IUW) + ~&(e)12 IYAGA(e)
--- + ydMe)12
G:(e) + G(e) = T(e) $ G2,(8)
it follows that the maximum-likelihood receiver based on
yz and yA performs exactly as well as the receiver based on
y1 and y2 and, in addition, has exactly the same structure.
The advantage of this sum-and-difference beam approach
is that for many monopulse designs the function GA (e)/
G,(e) is approximately linear in 0 for f3 in the range of
interest. This fact greatly simplifies the circuitry needed
to implement the estimation procedure.
Additional Unlcnown Parameters
In this section it will be assumed that the signal wave-
form depends on several other parameters, denoted by the
vector (Y, in addition to those already considered. Thus,
the received signals are of the form
r,(t) = Ae’*‘Gi(e, p)s(t, a) + ni(t) i= 1, .e. ,m
in the incoherent case and of the form
r,(t) = Ae”G,(e, cp)s(t, CY)+ ni(t) i=l;..
6 Either square root may be used.
Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON INFORMATION THEORY, JANUARY 1969
in the coherent case. It is still assumed that the signal is
normalized so that J Is(t, a!)]” dt = 1.
(34) A few moments’ thought will reveal that the only effect
the additional parameters will have on the solutions al-
ready obtained is to replace the old y, by
yi(d = ~hs*(f, 4 dt (38)
and to require that the new 1 be obtained from theold
1 = Z(E) by maximizing the latter on CLFor example, in the
incoherent, two-beam, one-angle problem,
(35)
where B denotes the value of (Ythat achieves the maximum
in (39). The reader should experience no difficulty in ex-
tending all the other results of the previous section to the
case where additional signal parameters are unknown.
The additional maximization on (Ythat is required when
calculating 1 often cannot be performed explicitly. When
this is the case, the usual solution is to use a bank of N-
filters matched to s(t, CQ), * . . , s(t, crN) to simultaneously
calculate the quantities Z(CQ), * . * , Z(CX~).The largest one
of these values is taken as the desired value of I, and the CC<
for which it occurs is used as the estimate of CK.This tech-
nique will give satisfactory results as long as the values of
the CQare sufficiently close together to give a reasonably
accurate picture of the function Z(a). Naturally, if the num-
ber of unknown parameters is large, this method may be-
come too unwieldy to realize in practice.
A concrete example of the procedure just described
’ arises when the additional unknown parameters are range
delay r and Doppler frequency f, i.e.,
s(t, a) = s(t - 7)ei2”lt. (41)
The quantity y%(T, I) can be formed by passing r,(t)
through a filter whose impulse response is s*( - t)ei2”t’ and
then sampling the output at time 7. A finite bank of such
filters spaced by the Doppler resolution of s(t) and span-
ning the range of expected Doppler frequencies often will
yield a close enough approximation to y,(~, f). The oper-
ation of the receiver for, say, the incoherent two-beam,
one-angle problem now can be described. The outputs from
the two beams are processed in two identical filter banks of
the type just considered. The outputs from the corre-
sponding Doppler channels are detected in square-law de-
tectors, added, and then compared with a threshold.
Whenever this threshold is exceeded, “target present” is
(36)
announced. The time of the peak response following the
threshold crossing is, the estimated range delay, and the
frequency of the Dc/ppler channel having this peak re-
,m (37) sponse is the estimated Doppler frequency. The ratio of
the outputs of the square-law detectors is formed and used
to obtain the estimate of target angle according to (40).
HOFSTETTER AND DELONG: DETECTION AND PARAMETER ESTIMATION 27

RECEIVERPERFORMANCE Note that El and ta are statistically independent, identical-

Receiver-Detection Characteristic
The receiver-detection characteristic (detection prob-
ability P, versus false-alarm probability P, or information
equivalent to this) has been calculated for the two-beam,
one-angle and three-beam, two-angle incoherent receivers
and for the two-beam, one-angle coherent receiver. No
analytical way of performing this calculation for any of
the remaining cases has been found.
The detection functions for the incoherent two-beam
and three-beam receivers are given by (14) and (13), re-
spectively. These equations represent the result of sum-
ming the quadratically detected envelopes of the pertinent
matched filter outputs. The probability distributions of
such functions are quite easy to calculate. As the details of
this calculation are readily available in the literature [7],
[S], only the results will be given here. The false-alarm
probability P, is given by
PF = exp (-$) g& (&)*,
n-here X is the threshold setting and m denotes the number
of beams in use. The detection probability P, is given by
P, = Qm( d@%, dWN,J,
where
exp (--+(a? + a2))1,-,(~x) dx
Q&, 0) = Jm x(y
denotes the generalized Q function and E is the received
signal energy,
in the two- and three-beam cases, respectively.
The generalized Q function defined by (44) is closely
related to the incomplete Toronto function which has been
plotted extensively by Marcum [7]. The use of these curves
in conjunction with (42) through (45) makes it possible to
construct curves of P, versus P, for any desired signal-to-
noise ratio E/N,.
The detection function for the two-beam coherent re-
ceiver is given by (29) or, equivalently, (33). Equation (33)
represents the summation of two linearly detected RF
voltages 7, and v2 of the form
TL = A[G,(e) + jG2(e)le” + tl
Q = L4K&(e)- jG,(e)le” + tar
where
i1 = 1 h(t) + jn,(t)ls*(t) dt
E2 = 1 [w(t) - jn,(t)ls*(t) dt.
ly distributed, complex Gaussian voltages and note also
that the means of rll and 11~have the same magnitude,
lE(d = lE(d/ = AKZ’(e)+ G(@l”“. (48)
Thus, the quantity 1~~1+ l+j is the sum of two linearly
detected RF voltages consisting of equal-amplitude RF
signals to which statistically independent, identically dis-
tributed Gaussian noises have been added. The problem of
calculating the probability distribution of such a sum was
attacked by Marcum [7] who, although unable to obtain a
closed-form solution, showed that the difference in per-
formance obtained by adding linearly detected voltages as
compared to quadratically detected voltages is negligible
for practical purposes. The conclusion to be drawn from
this result is that the detection function l~1l2 + Ir)z12will
perform almost as well as the optimum detection function
1011
+ Id. But
h12 + 1%12
= aYI12+ IYz121,
and it follows that the two-beam coherent detector per-
(42)
forms about as well as the two-beam incoherent detector
that has already been analyzed.
Angular Accuracy
(43) The problem of calculating exact expressions for the
means and variances of the angle estimates obtained above
is a formidable one. On the other hand, it is relatively easy
(44) to obtain approximate expressions for these quantities
valid for high signal-to-noise ratios. This has been done by
Swerling [9] for the case of a real scalar signal masked by
additive Gaussian noise. It is a straightforward task to ex-
tend his results to the case where the complex envelopes of
several waveforms are observed simultaneously. W ith this
generalization, Swerling’s results read as follows.
Given the complex envelopes r%(t), i = 1, . . . , m of the
form
r,(t) = s,@; a) + n;(t) i = 1, ... , m, (49)
where s,(t; a) is the complex envelope of a waveform that
is known except for a finite number of unknown parameters
described by the vector (Y and where the noises n;(t) have
the properties described by (2) then, for large signal-to-
noise ratios, the maximum-likelihood estimate B of Q: is
unbiased,
E(G) = (Y (59)
and has the covariance matrix
A = E[(& - a)( B - a)‘]
= 2N,[~Re/~~(~)*dt]-’ (51)
(46)
In (51), the prime denotes the matrix transpose, the as-
terisk denotes conjugate-matrix transpose, and aJ(x)/ax
denotes a column vector whose l&h component is af(x)/axk.
(47) The covariance matrix given by the right-hand side of (51)
is the same as the Cramer-Rao [9], [lo] lower bound for the
covariance matrix of any unbiased estimator of CY;there-

Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.
 28 IEEE TRANSACTIONS ON INFORMATION THEORY, JANUARY i969
fore, it follows that for large signal-to-noise ratios ‘the location into two parts. The first is the term ]]G]], which
maximum-likelihood estimate of a! is optimum in the sense is the same in all the beams and represents a modification
that no other estimate can have a smaller covariance of the signal amplitude A that is brought about by the
matrix.? target location. The second is the term gi (0, c,P),which
This general result now will be specialized to the mono- varies from beam to beam and gives a measure of the ex-
pulse problems considered earlier. For the m-beam co- tent to which target location produces differential changes
herent system, the parameter vector is given by p! = in the signal amplitude received by the various beams.
14 A 4 cpland Viewed in this light, it seems natural that the first term
should affect the angular accuracy only through the
si(t; a) = Ae'$-Gi(t9, p)s(t). (52) signal-to-noise ratio and that all the other factors in-
Making these substitutions in (51) leads to the following fluencing angular accuracy should be expressible in terms
asymptotic covariance matrix, of the normalized gain functions gi. This is exactly what is
expressed by (56).
A = 2N,
The matrix appearing in (56) can be written in the form
llGllz 0 j A@, '3 A@, W
0 A2 jjG11'; 0 0
-------------- ,-_------_- ------ - ,
A(G, Gel 0 j A2 IIGe(lZ AWe, G,) where
-1 A(G, G,) 0 :Aa(Ge, G,) A2 llG,lls 1 D = llgel12
Ilg,l12- (go>g,>“. (61)
(53) The variances of 8 and 8, as given by (60), have an inter-
where G’ = [G1(e, cp), . . * , G,(B, cp)], (x, y) denotes the dot esting geometrical interpretation. Noting the simple
product of the vectors x and y and ]]x]]’ = (x, x). identity,
This matrix can be inverted by making use of Frobenius’
relation for the inverse of a partitioned matrix [ll]. This
relation states that
it is seen that the variance of 8 can be written in the form

CT;= E(8 - e)" = 5 llgo - (gev $$ #$' (63)

The vector appearing in the denominator of (63) is the

where
component of go that is perpendicular to g,. This means
A = A,, - A,,A;:A,,. (55) that when ge and g, are perpendicular, ui is inversely
proportional to ]]ge]]2.
If A is partitioned as shown in (53), it then follows that the
This is a reasonable result because it states that the
covariance matrix for the angle estimates alone is given by
more sensitive g is to changes in 0, the better will be the
A-‘. Making the necessary substitutions now leads to the
accuracy in measuring e. When gs and g, are not perpen-
result,
dicular, however, the change produced in g by a change in
Asa = E[(B - 8, (2 - cp)‘(B - e,l$ - p)] B has a component that could equally well have been pro-
duced by a change in cp, It is reasonable to suppose that
(56) this component will reduce the measurement accuracy of 0
and that the useful sensitivity of g to changes in 6’is given
by the vector ge minus that component of go that lies along
where the normalized gain vector g = g(B, cp) is given by
g,, i.e.,thevectorge
- kh g,/llgBllk,/llg,II. Thiss=ms
g = IIW ’ G , (57) to be the reason why the variance ai is inversely propor-
tional to the squared modulus of the latter vector and not
and the signal-to-noise ratio p is given by
to lIgel12when g B and g, are not perpendicular.
P = U2/‘2Nd lIGIIZ. (58) It will now be shown that the large signal-to-noise ratio
angle-accuracy formulas just derived for the m-beam co-
Some insight into the reason why the particular normal- herent receiver are valid also for an m-beam incoherent
ization defined by (57) arises in this problem can be ob- receiver. This means that at large signal-to-noise ratios
tained by noting that the received waveforms can be writ- coherent processing offers no advantages over incoherent
ten in the form processing.
ri(t> = A d* IlGll gi(4 cpM0 + 4). (59) The starting point for this analysis is (51), where (Yand
si(t, w) are now given by
Thus, this normalization divides the effect of the target
Q: = [J/l, **. , $mn,
A, 0, cpl
7 Two non-negative definite matrices A and B are said to satisfy
A < B if and only if B -A is non-negative definite. si(t, a) = AeiS’Gi(B, cp)s(t) i- 1, 9.. ,m. (64)

Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.
 HOFSTETTER AND DELONG : DETECTION AND PARAMETER ESTIMATION ‘Y

Substitution of (64) into (51) yields the result A similar expression holds for the asymptotic variance of
the elevation estimate and it can be shown that the two
A = 2N,
estimates are asymptotically uncorrelated.
0 rnX3 -1 An interesting special case of (69) can be obtained by
A, IIGl/2Lxm~
--------_- I--_-__-_----------___
assuming that the beam shapes are Gaussian with equal
j ll~ll’ A(G, Go) 4G, G ,) widths.
0 3x.m AZ I/G e l/2 A’(Go, G ,) ’
P,(e) = exp (-e”/22), P&p) = exp (-(p”/2~“). (71)
.i I symmetric A” lI% lla!
(65) This assumption leads to the expression,

where I,,, denotes the (m X m)-dimensional identity N a

+=2Ao”;
matrix and OmXndenotes the (m X n)-dimensional zero 0
matrix.
Application of Frobenius’ relation, (54), to (65) leads . exp (82++ exp (P” + ‘~2 cash (aee,l2) (72)
u u cash (2eeo/u2) ’
easily to the following formula for the covariance matrix
of the estimates for A, 0, and cp, The accuracy formulas also are useful for comparing an
optimum monopulse system with various suboptimal sys-
E([ii - A, 6 - e,$ - p]f[A - A, 8 - e,d - cp])
tems, which may be desirable because of the ease with
[Ii W ” 4G, Gel A @ , 6) --I which they can be implemented or for other reasons. An
example of such a suboptimal system is a “conventional”
= 2N, A* jIGlIZ A*(Ge, G,) . 633) four-beam coherent monopulse system. Such a system
1 symmetric A2 II% lla I starts with four beams of the form

Another application of Frobenius’ relation now leads to 604 4 = G (e- eo,CP- cpo)
the desired covariance matrix for 8 and 4,
Gd3,0 = we - ear01+ 4 (73)
A _ I
8 Q-
ll!#
p F ks,fL)
This covariance matrix is the same as the covariance ma-
ksr g,) --I.
II&II” 1 67) G O , ‘P)=
We, 0 =
G @+
G (e+
eo,CP-
eo,P +
‘~~1
po)
trix that we derived for the coherent receiver, (56). and their associated matched filter outputs yl, .yZ, y3, and
As an application of the preceding accuracy formulas, y4. An estimate of the azimuth angle B is obtained by ap-
consider a four-beam, two-angle monopulse system em- plying two-beam one-angle maximum-likelihood process-
ploying “product beams” arranged on a rectangular grid, ing, (30) and (31), to the derived signals y1 + yz and
i.e., y3 + y4, assuming that the appropriate antenna gain func-
tions are G,(B, 0) + G,(e, 0) and G,(e, 0) + G,(e, 0), re-
G&4 d = me - eo)p2(~ - ~“1 spectively. In other words, the ‘Lconventional” receiver
arrives at an azimuth estimate 8 by solving
03f9
G,(e,0) + Gdfi, 0) Re [(yl + y2Fibl
(74)
Ga(e,0) + G.+(d,0) = Re [(yt3+ yJe-‘j] ’
We, ‘P) = fw + eo)p2(9 + cpo)
where
where P, and P, are known functions and.20, and 2~,0,de-
note the known beam separations in azimuth and eleva- 4 = t arg [(Y, + ~2 + (y3 + yd7. (75)
tion, respectively.
Substitution of (68) into (67) yields, after much tedious In actual practice, the sum-and-difference signals (gl + yJ
but straightforward algebra, the following expression for + (ys + y4) and (yl + y2) - (yB + y4) with the associated
the asymptotic variance of the azimuth estimate, gainsMe, 0) + W e , 011+ M O , 0) + G,(4 011and
Me, 0) + W e , 011- K&(4 0) + G,(B,011are usedto
+
we - 0”) 2 determine 6. This is equivalent to the procedure defined by
u; = - m e + 0,)1 (69)
(74) and (75), as was pointed out earlier.
The elevation estimate C$ is obtained by applying a
- P:(e - e,) ' '
P,(e - 0,) 1 similar procedure to the derived signals y1 + y4 and yZ +
~3 using Gl(O, cp) + G4@, ‘PI and GdO, cp) + G(O, cp>as gain
where the prime denotes differentation and functions.
The reasoning behind the procedure just described is
P = & [p:(e + a + pf(e - e,)l based on the fact that the beam shape G and the beam
0
separations 0,, and po,,usually ,can be chosen so that G, (0, ‘p)
* mcp + CPU)
+ p%cp- % )I. (70) + G, (0, cp)and G, (0, ‘p) + G, (0, ‘p) are approximately inde-

Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.
30 IEEE TRANSACTIONS ON INFORMATION THEORY, JANUARY 1969

pendent of 8. This being the case, it is natural to attempt Comparison of (79) and (69) shows that the ratio of the
to simplify the receiver by assuming the above mentioned estimation variance of the optimum system to that of the
gains are exactly independent of cpor 0 as the case may be. “conventional” system is given by
This assumption leads directly to the estimation procedure
described by (74) and (75). 2.L-=--=-
CT”(e) 1 p’
49 2 P
I[ ~z”(cp
1
- PO>
1 C 1 + Pz(cp + %I 2 1 + wcp + %) -I.
2 Pdcp - cpo)
The large signal-to-noise ratio performance of the sys-
tem decribed by (74) and (75) can be derived by approxi- (81)
mating the right-hand side of (74) to first-order noise terms It is, easy to bound the right-hand side of (81) above and
and the left-hand side to first-order terms in 6 - 8, where 0 below with the result’
denotes the true azimuth angle. The expansion of the
$c < a2(e)/&e) < 1, (82)
right-hand side uses the true gains, G,(e, ‘p) + G,(8, 4”) and
G3(0, cp) + G,(0, cp), rather than the assumed gains, G,(e, 0) which states that the azimuth-estimation variance of the
+ G,(B, 0) and G,(0, 0) + G*(L), 0). The results of this conventional system can be no more than twice the vari-
procedure are that the. aximuth estimate 6 has an asymp- ance of the optimum system. Stated in somewhat different
totic mean given by terms, this means that the asymptotic azimuth-estimation
accuracy of a ‘Lconventional” monopulse system can be
Me)
__- Go, d + G&4 d _ 1
made equal to that of an optimum system by increasing
E(B) = e + we) Gdk d + We. -~ 4
(76) the transmitted signal power by no more than 3 dB.
--___-
r:(e)-’ r,(e) ’

rde) we) REFERENCES

where the prime denotes differentation, 0 and cp are the [l] D. R. Rhodes, Introduction to Monopulse. New York:
McGraw-Hill, 1959.
true azimuth and elevation, and [2] W. Cohen and C. M. Steinmetz, “Amplitude- and phase-
sensing monopulse system parameters,” Microwave J., pt. I,
r,(e) = G&4 0) + Gde, 0) pp. 27-33, October 1959; pt. II, pp. 3338, November 1959.
(77) [3] E. J. Kelly, I. S. Reed, and W. L. Root, “The detection of
radar echoes in noise,” 1. SIAM, vol. 8, pt. I, pp. 309-341,
r,(e) = GO, 0) + Gde, 0). June 1960; pt. II, pp. 481-507, September 1968.
[4] U. Grenander, “Stochastic processes and statistical inference,”
The asymptotic variance of 8 is given by A&iv Matematik, vol. 1, pp. 195-277, April 1950.
[5] J. W. McGinn, Jr., “Thermal noise m amplitude comparison
monopulse systems,” IEEE Trans. Aerospace and Electronic
Systems, vol. AES-2, pp. 550-556, September 1966.
[S] H. Urkowitz, “The accuracy of maximum likelihood angle
estimates in radar and sonar,” IEEE Trans. Military Elec-
tronics, vol. MIL-8, pp. 39-45, January 1964.
[7] J. I. Marcum, “A statistical theory of target detection by
pulsed radar,” IRE Trans. Information Theory, vol. IT-6,
Specialization of these formulas to the case of the prod- pp. 59-267, April 1960.
uct ‘beams defined by (68) yields the result, [S] C. W. Helstrom, Statistical Theory of Signal Detection. New
York: Pergamon, 1960, pp. 171-176.
[9] P. Swerling, “Parameter estimation accuracy formulas,”
23(d) = e IEEE Trans. Information Theory, vol. IT-IO, pp. 302-314,
October 1964.
c(e + 0,) [lo] H. Cramer, Mathematical Methods of Statistics. Princeton,

IT:(e)
=-pzi
me -
II
me +
0,)
0,) (79) [II]
N. J.: Princeton University
E. Bodewig, Matrix
Publishing
Press, 1946, pp. 477-497.
Calculus.
Co., 1956, p. 188.
Amsterdam: North-Holland

[ p,(e+ e,)
where 8 The lower bound is based on the assumption that Pz(q + rpO)/ ,
Pz(q - po) 2 0. This condition will be met, at least for angular
directions of interest, in any well designed monopulse system.
PI = & [P:(e + 0,) + p:(e - e,)i Violation of this condition implies that signals are being added in
phase opposition when the beams Gr + Gs and Ga + Gh .are formed.
This is a situation which surely will be avoided in the design of
. [~z(cp + %I + p&P -. %)12. (80) any practical monopulse system.

Authorized licensed use limited to: UNIVERSIDAD CARLOS III MADRID. Downloaded on June 13,2020 at 17:50:59 UTC from IEEE Xplore. Restrictions apply.