1.

Introduction

The performance of a radar is often limited by echoes

On the Optimization of from external clutter that are large compared with internal
noise. In many practical situations, targets have a radial
MTI Clutter Re jection velocity component with respect to the clutter scatterer. In
this case, Doppler discrimination can be used to enhance
the signal-to-clutter ratio (SCR). Usually, this type of
J.K. HSIAO
Naval Research Laboratory
Doppler discriminator uses a train of identical pulses
Washington, D.C. 20375 uniformly spaced. This type of processor has been discussed
extensively in the literature [1] -[4]. When the frequency
of the clutter is narrow, two or three pulses are enough to
achieve the required improvement for detection. When the
clutter spectrum broadens, it is generally believed that by
Abstract processing a large number of pulses the same amount of
improvement can be achieved. However, a quantitative
In this paper is formulated the problem of optimization of the analysis of the limitations to this approach is not available.
improvement factor of a nonrecursive MTI by minimization of a When the pulses in a train are equally spaced, the
quadratic form. The minimum normalized clutter output (a Doppler discrimination system described above exhibits
reciprocal of the average improvement factor) is the minimum blind speeds; a target which moves at an integral number of
eigenvalue of this quadratic form, and the corresponding eigenvector half-wavelengths between pulses can not be distinguished
is the optimal weight to be used in this filter. By use of classic
from a fixed scatterer. These blind speeds can be avoided
by varying the interpulse spacing (staggered PRF). This has
matrix theory, some properties of this MTI improvement factor are
been discussed by Emerson [51, Brennan [6] , and Jacomini
shown, namely, that it is bounded and is a monotonic function of [7]. One additional problem associated with the
the clutter spectrum variance. Also discussed is the limit of an MTI staggered-PRF system is the selection of good interpulse
system having a large number of cancellers. Finally, the problem of a spacing. No direct method for optimizing the spacing code
staggered-PR F MTI filter is examined, for which it is shown that its has been found.
improvement factor is bounded by two equivalent constant-PRF This paper is intended to discuss some of these
MTI systems. One of these systems has a PRF equal to the lowest problems. For a uniformnly spaced pulse-train system, the
PRF of the staggered-PRF system, while the other has a PRF equal limitations on SCR improvement will be presented in terms
to the highest PR F of the staggered system. of clutter spectral spread and the number of pulses in the
train. As for the staggered-PRF system, the performance
bounds of such a system will be examined in terms of
various interpulse spacing.

II. AVERAGE SCR

Consider a coherent pulse radar which transmits a train

of N pulses in one direction at times tI, t2, , tn. The
pulse-to-pulse spacing may be varied in the train, but the
minimum interpulse spacing is long enough that returns are
not received simultaneously from different ranges within a
region of extended clutter. Let the signal reflected from a
target be

Sn= SO exp [(2rftn + q] (1)

where f is the Doppler frequency of the target. It is well
know that, for best detection, the received signal train
should be weighted with the following optimal weights [6]:
* (2)
xi = kb.bS.
Ifll

Manuscript received October 22, 1973.

B= [b..] (3)
This work was supported by the U.S. Naval Ship Systems where bi, are the elements of matrix B and B is the inverse
Command. of the clutter covariance matrix. Receiver thermal noise can
622 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-10, NO. 5 SEPTEMBER 1974
 The MTI improvement factor (I.F.) or ratio of output to
input SCR is then
I.F. = (Ex.x.
II
*)I(z22x.x.*aii)
i II ii
(9)
where the elements of the covariance matrix are
normalized, so that ai = 1. This same result has been
derived by Brennan [6].
The quantities aii, elements of the covariance matrix A,
are the Fourier transforms of the clutter power density
function. One of the convenient and widely assumed
models of this power density function is

be included as a real constant added to the diagonal

G (f) =
[I/(-vf2-i-T ac)] exp (- f2 /2orc2) (10)
element of this covariance matrix. The xi are the set of where f is the Doppler frequency and a, is the standard
complex weights to be used in the filter (see Fig. 1). Since deviation. Without losing generality, we have assumed that
these weights are a function of the target Doppler the mean Doppler frequency is zero. Accordingly, the
frequency, a filter bank of many channels is required. Each elements of the covariance matrix A are real and have the
of these channels is designed for a specific Doppler form
frequency.
An alternative approach to this problem is to design a [-2Ir2ac2 (ti-t)2 ].
single filter channel such that the SCR is optimized to the
a,1 = exp (11)
average Doppler frequency band. This approach is less For a constant-PRF MTI filter, this becomes
efficient, but it is more economical. The circuit described in
the first case is usually referred to as a Doppler processor, a,1 = exp (-21r2a2k2r2)
while the latter is called MTI (moving target indication). In (12)
this paper we shall concentrate our attention on the MTI where
system.
The output signal power of this MIT system is k = i-j.
One may notice here that the only difference between a
Ps= 1/2 Ex.S 122 (4) staggered-PRF MTI and a constant-PRF MTI is the
By use of (1), this can be written as difference between these matrix elements. One may also
notice that this matrix is symmetrical; that is,
p = (So2 /2)1 Yx x. cos [27rf(t.-t.)] . (5)
a..
11
= da..gt
11
This function is periodic with a period of
The diagonal element has the following form:
fp =1Ilt-j (6)
where the lp are integers for all i and j'. This fp is
a.. = 1+ n0 (13)
commonly referred to as the blind Doppler frequency. where 71o is the normalized receiver noise level. This
When the target Doppler frequency is unknown, one may diagonal element has the maximum value; hence, one may
assume that it has a uniform distribution. In this case, the normalize this matrix such that
expected signal output is then
0 <aii = 1, i,] = 1, 2, --,N. (14)
Ps=(So2/2)2jx 17. (7)
In practice, the receiver noise component q7o is usually
The average output power, plus noise if noise terms are very small. Furthermore, the effect of this noise is to
included in the covariance matrix A, is reduce the value of the off-diagonal element, which is
equivalent to increasing a, or the correlation time t-t1.
P C = x.x. *a..
I I ii (8) For convenience of discussion, we shall ignore this effect.
Since the covariance matrix is real and symmetrical, and,
where the aii are the elements of the clutter covariance as will be shown later, the optimal filter weights xi are
matrix A. components of the eigenvector of this matrix, these xi are
real.
T. and Tj must be rational numbers. For a detailed discussion, The clutter output Pc is actually a quadratic form.
see [8]. Representing the filter weights xl, x2, , xX by an
HSIAO: MTI CLUTTER REJECTION 623
N-dimensional vector x, this quadratic form can be 3) By virtue of property 2), the quadratic form Q(x) can
compactly written as be represented as

Q(x) - (x, Ax) (15) Q(x) = 1: dI.2 X.I (18)

i
where the parentheses represent the inner product of the where
vector x and the vector Ax.
The improvement factor thus becomes x= E2d.x1
I (19)
I. F- (x, x)/(x, Ax). and X1 are eigenvalues and xi are the associated
eigenvectors. It is evident that the minimum Q(x) occurs
that the reciprocal of this improvement factor is actually when x takes the eigenvector which is associated with the
the clutter output, or minimum eigenvalue Xm in, since

PC -(x,Ax) (16) 1 > 2> > xmin

when x is normalized. due to property 1).
In an MTI filter design, one of the optimization criteria Then.
is to maximize the improvement factor or, equivalently, to
minimize the clutter output factor (x, Ax). The problem Q(x) =Xm in
then becomes to find a vector x such that the quadratic
form (x, Ax) is a minimum. Thus, the quadratic form has a minimum value which is
Before we can find such a vector, we will show that this equal to the smallest eigenvalue of matrix A, and the
quadratic form is positive definite. The elements of the eigenvector associated with this eigenvalue is the required
matrix A can be represented by the inner products of a set minimization vector. This same result has been shown by
of N-dimensional vectors, such that Emerson [5] and Brennan [6] by use of Lagrange's
multiplier.
ai- (al, aJ). (17) If one expands the determinant A -II, one finds the
following valuable relations between the eigenvalues and
This can be done because the diagonal elements in matrix A the elements in matrix A:
are always larger than the off-diagonal elements, and the N N
matrix is symmetrical. According to the theorem of WX.= ia.. (20)
Gram-Schmidt [91, the quadratic form of this kind of i-I i= l
matrix is always positive definite. This can also be seen N
intuitively. Since Q(x) represents the clutter output power, i=n IX. - det A. (21)
it should always be positive, unless all filter weights are
zero. We pointed out earlier that the diagonal elements of matrix
A are unity; hence,
Ill. Minimum Figenvalue and its Eigenvector
N
, X.=N. (22)
In the previous section we discussed the fact that the i=
optimal design of the MTI filter requires finding an Since the quadratic form Q(x) is positive definite, we have
N-dimensional vector x such that the quadratic form Q(x) is the following properties [91:
minimum, under the constraint that this vector lies on a
sphere, det A >0 (23)
N
(x,x)= 1. det A < Hl a... (24)
i Il
This amounts to solving a classic eigenvalue problem. It is Hence,
well known that these eigenvectors and eigenvalues possess
N
the following properties [9 ], [10]: 0 HX. 1. (25)
1) The eigenvalues of a real, symmetrical matrix are real. i= II
Furthermore, if A is positive definite, these eigenvalues are From (22) and (25), one can deduce that
positive.
2) The eigenvectors associated with distinct eigenvalues 0<X < 1. (26)
min
of a real symmetrical matrix A are real and orthogonal.
Moreover, these vectors can be normalized so that they Here, m,in is the smallest eigenvalue, which then, in turn,
form an orthonormal set. represents the optimal clutter output of an MTI filter. The
624 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS SEPTEMBER 1974
equalities in the above equations represent two degenerate F(f) 12 = I2Xx.x cos 27rf(i -j) (31)
cases. When either the clutter spectrum standard deviation ii

-c*0o0 or the correlation time ti - becomes infinite, the is a periodic function when f is an integer. Furthermore, it
off-diagonal elements in matrix element A approach zero can be shown that this power transfer function is
[see (13)]. In this case, symmetrical within each period [8]. Therefore, it is evident
that for a working MTI system, a must be limited to a very
X1 = X = - = X = small value (much smaller than 0.5). Otherwise, the clutter
spectrum spreads so far into the filter passband that no
and any normalized vector can satisfy the eigen equations. effective filtering can be achieved. Thus, the value of a
On the other hand, if ac or the correlation time ti-t plays an important role in an MTI filter design.
approaches zero, then every element of matrix A With this assumption, we shall show several interesting
approaches unity. In this case, the determinant A is zero. properties of the minimum clutter output Xmin of a
According to (21), the smallest eigenvalue is zero. Under constant-PRF MTI system.
this condition, the eigen equations are identical and have
the following form: Monotonic Property
x + x +2 -+ X =0. Assume that, for a given ao, the corresponding optimal
filter weights xi' are inserted into (30) and a new function
There are an infinite number of solutions. One solution is defined:
is to set these x, according to the binominal distribution.
This is well known. One sees that the binominal weighting
MTI minimizes the clutter output only when the standard
Xm (o a) = [2I(iof2aa)1 f exp [2f2 /(2a2)I
-

deviation of the clutter spectrum is zero. *Xi (aO. (uoa) cos 2fri-j). (32)
We thus conclude that the clutter output of an MTI i J
system always lies in a range from 0 to 1. As the clutter
spectrum standard deviation ac or the correlation time ti - The derivative of the filter power transfer function
t1 increases, the clutter output increases.
(6/61) F(f) 12 =27rrE(i -j)x (ao)x '(ao)
sin2ifti-0)
IV. Constant-PRF MTI Filter sin 27rf(i -j) (33)
If the PRF is constant, elements of matrix A, as pointed is zero when f= 0, 1/2, 1, 3/2, , no matter how the xi are
out earlier, can be represented as chosen. Since the clutter power density function is a
bell-shaped function and heavily concentrated at the
a. =exp [-2 r2 ac2 (-)2r2] (27) neighborhood of f = 0, it is evident that the optimal
weights xi must be so chosen that a minimum occurs at f =
One introduces a new variable 0. Furthermore, the x, are constrained such that the average
frequency response of the filter is constant; that is, cIx2
a=as c T = 1. Therefore, the envelope of the power transfer function
must increase monotonically from f = 0 to f = 0.5. The
and sets area under the curve of the function (1 /a) exp [-
f /(2a2)] is a constant. However, as a decreases, more area
t = exp (- 2ir2a2 (28) will be concentrated at the neighborhood f = 0. Therefore,
it is not difficult to see that
Then,
X(o, a) < a(ao, ao ) when a < ao. (34)
aaiJ ~(i_ )2-
= (29)
This may be interpreted as follows. The improvement
With this normalized a, the Xm in (or clutter output) can factor for an MTI filter designed for a certain a0 will not be
be represented as degraded if the actual spectrum standard deviation a is less
than ao.
Xmin =
[2I(v;a)1
f exp [2/(2a2)]
- Furthermore, it is evident that
* S2x I.xI. cos 2irf(i -i) df
ii (30) Xmin(a, a) S Xmin(ao, a) < X(afo, a0)
where the xi are the optimal filter weights for a givean a. if a < ao (35)
Notice that in this equation the independent variablcef is
normalized, so that the filter power transfer function due to the nature of the optimization.
HSIAO: MTI CLUTTER REJECTION 625
 0

C
7M10 N=2

Li- 30
CD 20 l 0

iCD 50
z:- 140
t.LJ

cc
N=2Q
N=6
N=7N~

60

70 ........

0. 00 0.05 0. 10 0. 15 0. 20 0. 25
CLUTTER SPECTRUM/PRF
Fig. 2. Improvement factor versus normalized clutter-spectrum
standard deviation for a number of cancelling pulses from 2 to 20.

N
Hence, one may also conclude that Xmin(U) is a
monotonic function and its value increases from 0 to 1 as a
-a,,I< a.i, j*i. (38)
increases from 0 to infinity. From (36), we see that as N, the number of MTI canceling
pulses, increases, the off-diagonal element aiN approaches
Improving Property zero. Thus, the quantities X- aii approach a limit as N
increases. Since aff = 1, the ai, are positive, and Xm. is
Consider matrixes AN as a sequence of matrixes; that is, always less than unity, one may write
Xi >1 Yaii , i *j.
A2=[ :] (36a)
From (37), one may conclude that although the
improvement factor (/Xm in) improves as N increases, it
I t ,4 reaches a limit as N becomes sufficiently large.
Closely related to this result is the clutter spectrum
1 to (36b) deviation a. One may see that as a increases, all high-order
A3 =

terms become essentially zero. We may thus write

Lt4
A -1 It O O
I t t4 .. tN2
(39)
AN =
. I t - t(N- )
(36c)
O O O -1
Lt I _
Under this condition, the limiting circle in (38) remains
According to the Sturmian separation theory [ 1], about the same no matter what N is. Thus, we can establish
another limit. This limit says that as the a increases to a
tN+ 1(AN+ 1)6 XN(AN) (37) point, the clutter output of an MTI system will be
independent of the number of stages of the MTI system.
where XN(AN) represents the minimum eigenvalue of
matrix AN. This means that the minimum clutter output V. Numerical Results of a Constant-PRF MTI System
decreases as the number of cancelling pulses in the MTI
system increases. The optimal improvement factors are computed versus
the normalized clutter-spectrum standard-deviation factor
Limiting Property a. This is shown in Fig. 2. The number of MTI canceling
pulses N is used as the parameter, and a family of curves is
It is well known that each eigenvalue of a non-negative obtained. For each N at each a point, the minimum
matrix lies in the interior or on the boundary of at least one eigenvalue is computed. The reciprocal of this value, which
of the circles [12] - [151 is the improvement factor, is plotted on a decibel scale. At
626 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS SEPTEMBER 1974
 0

M 10
cc
D 20 + LOWER BOUND

CC)uL 30

LU 40-
LU-
CD 50

60

70 - II1I1I1 p

0. 00 0.05 0. 10 0. 15 0. 20 0. 25
CLUTTER SPECTRUM/PRF
Fig. 3. Staggered-PRF MTI filter with a 3-pulse canceller and 10
percent interpulse variation.

each point for a given a, a different eigenvector or filter In a previous section it was shown that the optimal
weight is used. Thus, each of these curves represents the improvement factor of a constant-PRF MTI filter is a
maximum improvement factor one can expect. It clearly monotonic function of a (a = uT). Increasing the
demonstrates the properties discussed in Section IV. The interpulse correlation time, in general, increases the clutter
curves are monotonic. For a given a, the improvement output Xm in. Since the covariance matrix of a
factor approaches a limit as the number of MTI cancelling staggered-PRF MTI system has properties similar to those
pulses increases. Furthermore, as a increases, all these of a constant-PRF MTI system, we would anticipate that
curves converge closely to a single curve, and the this similar monotonic property would be maintained.
improvement factor approaches 0 dB. Therefore, we conjecture that the minimum clutter output
These curves have significant value in the MTI design. of a staggered-PRF MTI system is bounded by the
For example, if the normalized-clutter spectrum standard performance of two equivalent constant PRF MTI systems.
deviation has a frequency of about 1/10 of the PRF for a One of them has a PRF equal to the lowest PRF of the
10-pulse MTI, one would expect an improvement factor of staggered PRF system, while the other has a PRF equal to
34 dB. A further increase in the number of MTI pulses does the highest PRF of the staggered system. That is,
not appreciably improve this situation. The improvement X (T) = Xj(T ) < X (T2). (41)
factors shown on these curves represent an upper bound.
Better improvement can not be expected. In the next To investigate this effect, many numerical examples
section we will show that the use of a staggered-PRF system were computed. In each of these examples, a lower
does not improve this situation. interpulse time T1 and an upper interpulse time T2 are set.
The interpulse times of the staggered-PRF system are then
VI. Staggered-PRF MTI System generated randomly (with a uniform distribution).
However, they are kept within these bounds. The optimal
To avoid blind velocities, staggered-PRF MTI sytems are improvement factors are then computed as a function of a.
generally used. In these systems, the interpulse durations These improvement factors were then compared with
vary from pulse-to-pulse. The choices of this variation are constant-PRF MTI systems having interpulse times T1 and
numerous. It thus makes the analysis of such a problem T2 . Many samples have been computed with the number of
extremely difficult. However, in a practical radar system cancelling pulses ranging from 2 to 20 pulses and with a
this variation is somewhat bounded by the need for variation of interpulse time from 10 to 40 percent. So far
efficient usage of radar time. One may therefore assume we have not found a single case which violates this
that the interpulse durations are varying between a lower conjecture. Fig. 3 shows a typical example. The lower
and upper bound: bound of these curves, which is represented by crosses, is
the improvement factor of a constant-PRF MTI filter
T1 <Ts<T2 having an interpulse time T1, while the upper bound,
represented by squares, is the corresponding improvement
T2 =T1 (1 + a) (40) factor with an interpulse time T2. In this particular
example, T2 = 1.1 T,. Curves (almost a solid line in Fig. 3)
where represents
a the fractional variation of the interpulse which lie within these two bounds are those of
duration. staggered-PRF MTI systems.
HSIAO: MTI CLUTTER REJECTION 627
 1 0
I9
z i.
081o 3

0n 0. 6 11 1111

0:1PU
~~~~LTE DEV1111 I RT I ON 1111

Fig. 4. Clutter output statistic distribution of a 4-pulse

staggered-PRF MTI filter, a -0.05; interpulse variation: curve 1-
10 percent, curve 2 -20 percent, curve 3 -30 percent, curve 4 -
40 percent; U.B. denotes upper bound.

In Fig. 4 we show the statistical property of the sharply. This means that the majority of the mass of the
improvement factor of a staggered-PRF MTI system. The probability density function lies within a narrrow region of
interpulse duration is again generated randomly according the clutter output. For example, curve 1 shows the samples
to a constraint shown in (40), which is repeated as follows: of a 10 percent variation of interpulse duration. It shows
that a randomly chosen interpulse sample has a probability
T T.<
T1 AS T2 T2 of 0.99 that its clutter output lies in the middle of the
upper and lower bound, while the probability of a sample
T= T1(1 +a) having a clutter output about 10 percent higher than the
lower bound is no more than 0.1.
where T1 and T2 are, respectively, the lower and upper
bound of the interpulse duration, while T. is the interpulse Vil. Conclusion
duration of the staggered-PRF MTI filter. The clutter power
density function is assumed to be Gaussian. In this In this paper, we have formulated the problem of
example, its normalized values of standard deviation for the optimization of the improvement factor of a nonrecursive
lower and upper bounds are, respectively, MTI filter by minimization of a quadratic form. It was
shown that the minimum normalized clutter output (a
aLB acTl 0.05 reciprocal of the improvement factor) is the minimum
eigenvalue of this quadratic form, and that the
OUB = aJT2 0-05(1 a). + corresponding eigenvector is the optimal weight to be used
in this filter. It was also shown that the improvement factor
There are four curves plotted in this figure. The curves of a nonrecursive MTI filter has the following properties:
correspond to maximum interpulse variations of 10, 20, 30,
and 40 percent. In each of these cases the optimal clutter 1) The normalized clutter output is always greater than
output of a constant-PRF MTI having an interpulse zero and less than unity.
duration of the lower bound is computed. Subsequently, a 2) The improvement factor is a monotonic function of
set of interpulse durations of the staggered-PRF MTI is the normalized standard deviation of the clutter spectrum.
generated randomly within the specified bounds, and its As the clutter spectrum increases or the PRF decreases, the
minimum clutter output is computed and then normalized improvement factor is reduced.
with respect to the clutter output of the lower bound. The 3) An increase in the number of cancellers in an MTI
clutter output of the upper bound (constant-PRF with system imploves its performance; however, as this number
interpulse time T2) is also computed and normalized. This increases to a point (say 8 to 10 cancellers), this
is shown by dotted lines in Fig. 4. It is also assumed that at improvement reaches a limit. Furthermore, the normalized
least one of the interpulse durations of the staggered-PRF clutter spectrum standard deviation also limits the
system is equal to the lower bound T1. performance of an MTI system. At a value of
A large number of samples are computed. The approximately 0.25, the improvement factor that an MTI
cumulative probability of these samples is then plotted can achieve is independent of the number of cancelling
against the normalized clutter output. These curves rise pulses used.
628 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS SEPTEMBER 1974
 4) The improvement factor achieved by a staggered-PRF [41 D.F. Delong and E.M. Hofstetter, "On the design of optimum
MTI system lies within two bounds. The lower bound is the radar waveforms for clutter rejection," IEEE Trans.
improvement factor which can be achieved by an equivalent Information Theory, vol. IT-1 3, pp. 454463, July 1967.
[51 R.C. Emerson, "Some pulse Doppler and MTI techniques,"
constant-PRF MTI having a PRF equal to the lowest PRF Rand Rept. R-274, March 1954.
of the staggered MTI system, while the upper bound is [61 L.E. Brennan and I.S. Reed, "Optimum processing of
equal to the equivalent constant-PRF MTI system, which unequally spaced radar pulse trains for clutter rejection,"
has a PRF equal to the highest PRF in the staggered system. IEEE Trans. Aerospace and Electronic Systems, vol. AES-4,
In deriving the above conclusions, it is assumed that the pp. 474-477, May 1968.
clutter power density function is Gaussian. We suspect, [71 O.J. Jacomini, "Weighting factor and transmission time
optimization in video MTI systems," IEEE Trans. Aerospace
however, that other even and bell-shaped clutter power and Electronic Systems, vol. AES-8, pp. 517-527, July 1972.
density functions may have these same properties. [8] J.K. Hsiao and F.F. Kretschmer, Jr., "Design of a
staggered-PRF moving target indication filter," Radio and
Electronic Eng., vol. 43, November 1973.
[91 R. Bellman, Introduction to Matrix Analysis, 2nd ed. New
York: McGraw-Hill, 1970, ch. 4.
1101 C.R. Wylie, Jr., Advanced Engineering Mathematics.
References [lll W.S. Burnside and A.W. Panton, Theory of Equations, vol. 2.
New York: Logmans, Green, 1928, ch. 7.
[1 W.D. Rummler, "Clutter suppression by complex weighting of [121 H. Rohbach, "Bemenkungen zu einem Determinantensatz von
coherent pulse trains," IEEE Trans. Aerospace and Electronic Minkowski," Jahr. Deut. Math. Ver. vol. 40, p.49, 1931.
Systems, vol. AES-2, pp. 689-699, November 1966. [13] E.W. Barankin, "Bounds for the characteristic roots of a
[21 R.L. Mitchell and A.W. Rihaczek, "Clutter suppression matrix," BulL Am. Math. Soc., vol. 51, pp. 767-770, 1945.
properties of weighted pulse trains," IEEE Trans. Aerospace [141 E. Bodewig, Matrix Calculus, 2nd ed. Amsterdam: 1959.
and Electronic Systems, vol. AES-4, November 1968. [151 A. Brauer, "On the characteristic roots of non-negative
[3] L.J. Spafford, "Optimal radar signal processing in clutter, matrices," in Recent Advances in Matrix Theory, H.
"IEEE Trans. Information Theory, vol. IT-14, pp. 734-743, Schneider, Ed. Madison, Wisc.: University of Wisconsin Press,
September 1968. 1964.

James K. Hsiao received the B.E.E. degree from Hunan University, China, the M.S. degree
in electrical engineering in 1957 from Montana State College, and the Ph.D. degree in
1962 from Iowa State University.
He was with the Chinese Nationalist Air Force for 10 years, and he taught electrical
engineering at Taipli Institute, China. From 1962 to 1967 he was with Bell Telephone
Laboratories in the Digital System Department, working on research and application of
super-conductor devices. Since 1967 he has been with the Radar Division of the Naval
Research Laboratory, Washington, D.C., specializing in array antennas. His current
interests are in the areas of phased arrays and radar signal processing.
HSIAO: MTI CLUTTER REJECTION 629