STRUCTURED COVARIANCE ESTIMATION FOR

SPACE-TIME ADAPTIVE PROCESSING

Timothy A . Barton and Steven T. Smith

Lincoln Laboratory
Massachusetts Institute of Technology
Lexington, Massachusetts 02 173
email: barton@ll . m i t .edu and s t s m i t h @ l l . m i t .edu

ABSTRACT 2. COVARIANCE MATRIX ESTIMATORS

Adaptive algorithms require a good estimate of the inter- 2.1. Sample Covariance M a t r i x Estimator
ference covariance matrix. In situations with limited sam-
ple support such an estimate is not available unless there is The sample covariance matrix is usually used as an esti-
structure to be exploited. In applications such as space-time mate of the true interference covariance matrix in STAP
adaptive processing (STAP) the underlying covariance ma- algorithms. This approach is reasonable because the sam-
trix is structured (e.g., block Toeplitz), and it is possible ple covariance matrix is easy to compute and is the uncon-
to exploit this structure to arrive at improved covariance strained ML estimate under a Gaussian data model. How-
estimates. Several structured covariance estimators have ever, this method yields a poor estimate with insufficient
been proposed for this purpose. The efficacy of several of sample support. Diagonal loading is a common technique
these are analyzed in this paper in the context of a vari- used to improve performance in this scenario; however, this
ety of STAP algorithms. The SINR losses resulting from method is not effective against lower power yet significant
the different methods are compared. An example illustrat- interference sources. Furthermore, neither method incorpo-
ing the superior performance resulting from a new max- rates constraints that arise from consideration of the phys-
imum likelihood algorithm (based upon the expectation- ical phenomena that generate clutter data.
maximization algorithm) is demonstrated using simulation
and experimental data. 2.2. M L S t r u c t u r e d Covariance Estimator
In most ML structured covariance estimation problems no
1. INTRODUCTION closed-form solution is known; this is true for covariance
matrices with generic block Toeplitz or Toeplitz block Toep-
The presence of ground clutter can seriously degrade the litz structure. It is this difficulty that leads one to an iter-
detection performance of airborne surveiIlance radars be- ative numeric technique for ML estimation.
cause the clutter energy may exist in all angular directions Structured covariance estimation is considered in its fulI
and doppler bins. Because the precise structure and loca- generality by Burg et al. [2], who propose an iterative algo-
tion of clutter interference is not known a priori, adaptive
rithm for ML structured covariance matrix estimation based
nulling methods, especially space-time adaptive processing upon the idea of driving the derivative of the log-likelihood
(STAP), are used to mitigate its effects. function to zero. Alternatively, the iterative expectation-
Such adaptive processing methods require knowledge maximization (EM) algorithm of Dempster et al. [3] has
of the interference covariance matrix. If this covariance
been used to generate ML structured covariance matrix esti-
matrix is not known, it must be estimated in some way. mates [6,5,9,4]. The attractive properties of the EM-based
Typically, the sample covariance matrix is used as the es- ML algorithm include: (1) the sequence of covariance esti-
timate of the covariance. However, in situations with in- mates are non-decreasing in likelihood; (2) the constraints
sufficient sample support, this estimate yields poor detec- imposed on the covariance estimates are easily incorporated
tion performance [8]. This problem motivates the consid- into the estimation procedure; (3) the procedure reduces
eration of structured covariance estimation procedures in to a sequence of simple matrix operations. Unfortunately,
which improved adaptive processing performance may be because of its iterative nature, the EM algorithm for struc-
achieved [6, 5 , 9, 41. This is the situation for STAP, where tured covariance estimation is computationally intensive.
the clutter covariance matrix is modeled as having block An EM algorithm for generating block Toeplitz and
Toeplitz or Toeplitz block Toeplitz structure [12, lo], de- Toeplitz block Toeplitz covariance matrix estimates derived
pending on prior knowledge of the array geometry. The by Fuhrmmn and Barton [5] is summarized by the following
subject of this paper is to analyze the efficacy of a variety
sequence;
of structured covariance estimation techniques for STAP,
including a new maximum likelihood (ML) algorithm. xk+l = B(xk -I-xkAHRi'(s- Rk)Rk'AEk), (1)
This work was sponsored by DARPA under Air Force contract Rk+i = Axk+iAH, (2)
F19628-95-C-0002.Opinions, interpretations, conclusions, and
recommendations are those of the authors and are not necessarily where Rk is the structured covariance estimate, S is the
endorsed by the United States Air Force. sample covariance matrix, and B(.) denotes the block di-

0-8186-7919-0/97$10.00 0 1997 IEEE 3493

agonal part. The sequence is initiated by providing a pos- For the Hermitian Toeplitz example, this projection results
itive definite guess for C O(say CO= I), then taking RO = in the formulae
ACoAH. For the case of block Toeplitz covariance matrices
of order MN-by-MN ( M 2 blocks each of order N), diagonal of II(A) =
A11 + A22 n+ . .. + A n n > (7)

A = ( P M ,O M x ( P - M ) I W P ) @IN, (3)
where W p is the P-by-P ( P > M) normalized discrete and so on. That is, take the mean down diagonals. The
Fourier transform matrix. The use of this EM algorithm im- projection for block Toeplitz and Toeplitz block Toeplitz
poses the additional constraint that the estimated MN-by- matrices is similar: simply take the mean down diagonals
MN block Toeplitz covariance matrix has nonnegative def- and block diagonals.
inite block circulant extensions of size P. The problem with this method is that there is no guar-
For the case of Toeplitz block Toeplitz covariance ma- antee that n(A) will be positive definite; in fact, it is very
trices of order MN-by-MN ( M 2 blocks each Toeplitz of easy to generate examples where II(A) fails to be a co-
order N ) variance matrix. Furthermore, our purpose for estimat-
ing structured covariance matrices is to construct adaptive
A = (PA!, O M X ( P - M P P ) @ ( [ I N , ONX(Q--N)IWQ), (4) weight vectors that null interference. There is no guarantee
where P > M and Q > N . that n ( A ) will result in adaptive weights that sufficiently
Given the EM iteration sequence of Eqs. (1)and (2), the null the interference, especially in sample limited scenarios.
number of computations incurred when estimating an ML
block Toeplitz structured covariance matrix is significantly 2.4. Weighted Projected Covariance E s t i m a t o r s
increased over that incurred in generating the unstructured By a simple modification in the case of block Toeplitz and
sample covariance matrix. However, the ML structured co- Toeplitz block Toeplitz matrices, the projected covariance
variance matrix estimates result in improved performance estimator of the previous section almost surely yields posi-
of STAP algorithms. tive definite Hermitian matrices. Instead of taking sample
For the case of block Toeplitz and Toeplitz block Toep- means down (block) diagonals as in Eqs. (7) and (8), take
litz covariance structure, these issues lead one to consider the weighted means
fast non-ML structured covariance estimation procedures,
two of which are discussed in the following sections.

2.3. Projected Covariance Estimators

Projection offers one simple method of estimating struc- .............
tured covariance matrices. This idea comes from the fact
that for any vector space with an inner product, there exists
a unique orthogonal projection onto a given subspace. Here
the covariance matrices are viewed as the vector space of then set the covariance estimate to be R = toeplitz(P0, PI,
Hermitian matrices, and the subspace is viewed as the sub- . . . ,Pn-l). This weighted projection almost surely yields
space of structured Hermitian matrices. Looking ahead, we a positive definite block Toeplitz or Toeplitz block Toeplitz
forecast difficulties with this approach because its assump- covariance estimate. To see this, note that if A = (Aij)
tions are invalid: the covariance matrices do not have a +
is a sample covariance matrix, then P k (IC = -n 1, .. . ,
vector space structure. We shall nevertheless describe and
analyze this method because it is an obvious procedure and
-
n 1) represents a (matrix) autocorrelation sequence 171,
which has a positive power spectrum. This property is very
because its deficiencies will be improved somewhat in sec- attractive when there are fewer samples than degrees of
tion 2.4. freedom.
Given two Hermitian matrices A and B, there is a nat- In addition to producing a positive definite estimate,
ural inner product between A and B defined by: this method also has the advantages of speed and ease of
computation. However, this estimator is not based upon a
(A,B) = t r AB. (5) statistical model, and for the application of adaptive nulling,
The square length of A is written llAl12 = (A, A ) (i.e., the there is no reason to believe that it will yield improved per-
Frobenius norm). Let S be an n-dimensional subspace of formance. In fact, as will be seen in the next section, this
the Hermitian matrices, and E; (i = 1, . . . , n) be an or- method may not yield better performance than the sample
thonormal basis of S,i.e., (Ei, E j ) = 6ij (Kronecker delta). covariance matrix estimate.
For example, if S is the vector space of Hermitian Toeplitz
matrices, then we may take E1 = toeplitz(l,O,. . ., O ) / f i ,
- I

3. RESULTS A N D DISCUSSION
E2 = toeplitz(O,l,O,. . . , O ) / d mand , so on (using
Moler’s Matlab notation). The unique orthogonal projec- We will compare different covariancematrix estimation meth-
tion of an arbitrary Hermitian matrix A = (Aij) onto S ods by examining the the SINR loss achieved by the fully
is defined to be the vector ziaiEi such that the square optimum and several suboptimum STAP algorithms [12].
The SINR loss measures the loss in target power resulting
length I[ziaiEi - All2 is minimized over all ai. This re-
sults in the projection from adaptive nulling. In symbols,
SINR loss = IWHV12
~ I ( A )= C . (I E ; , A ) E ~ . (6) (wHRw)(vHv)’
3494
 0

h
-1 0 -10
m m
s -0
v ..
v) -20
v)

9 -30 "
a
I
-
v)
-40
ML ,

FULLY OPTIMUM
m i J -50 ' I I
PRI-STAGGERED
I

8 ELEMENTS ELEMENT SPACE FFT ELEMENT SPACE 8 ELEMENTS

8 PULSES PRE-DOPPLER POST-DOPPLER 8 PULSES

BEAMFORMER BEAMFORMER
DPC / PRI-STAGGERED
8 ELEMENTS 8 ELEMENTS
8 PULSES BEAM SPACE BEAM SPACE 8 PULSES
POST-DOPPLER 4 SUBAPERTURES
6 PRI STAGGERS

-"" -I"
~

-1l2 -114 0 1I4 112 -112 -114 0 1I4 112

NORMALIZED DOPPLER NORMALIZED DOPPLER

Figure 1: SINR Loss versus doppler (sample support = 1 x degrees of freedom) for four classes of STAP algorithms. The
ideal performance is shown by the thin black curve. Note that the ML's performance (thick black curve) exceeds all other
methods both in the width of the clutter notch and the SINR loss away from mainlobe clutter (at 1/4 doppler), except for
the PRI staggered post-doppler algorithm. In this case, we have 32 blocks each of order 8, which has less structure than
the other algorithms-fully optimum has 82 blocks each of order 8, displaced phase center (DPC) pre-doppler has 82 blocks
each of order 3, and DPC PRI-staggered has 62 blocks each of order 4.

where v is the steering vector, w is the adaptive weight which adaptivity occurs-element or beamspace, and pre-
vector, and R is the covariance of the interference-plus- or post-doppler processing [12]. In any domain, the opti-
noise. Of course all adaptive algorithms will suffer SINR mum (w.r.t. SINR) weight vector is given by the well-known
loss near interference; however good nulling algorithms will equation
not have significant losses away from the interference. w = R-lv, (13)
The scenario considered is a airborne radar with N = 8 whose degrees of freedom is M N . Because this product
elements and M = 8 pulses per CPI, whose velocity is such can be quite large for moderately sized systems, reduced
that 1 half-interelement spacing is traversed per PRI, and dimension suboptimum STAP algorithms are attractive. In
a uniform clutter environment with 40 dB clutter-to-noise this case, the suboptimum weight vector is given by the
ratio. A crab angle of 30' is used, resulting in mainlobe equation
clutter at a normalized doppler frequency of 1/4. There is w = T(T~RT)-'T~V, (14)
no back lobe clutter. The ideal clutter covariance matrix for whose degrees of freedom is the column dimension of T,
this scenario is Toeplitz block Toeplitz; however, we shall where T is a transformation matrix into one of the STAP
only assume a block Toeplitz structure, i.e., no pulse-to- domains shown in Figure 1. If T is block Toeplitz, then
pulse mismatches. block Toeplitz structure is preserved in the reduced di-
STAP algorithms may be classified by the domain in mension space, and structured covariance matrix estimators

3495
may be considered. 60
A set of complex Gaussian data with the ideal clutter
covariance matrix for this scenario [lo] is generated. The DOPPLER = 3/16
number of independent samples available for covariance es-
timation is chosen to be exactly the adaptive degrees of
freedom-a low sample support scenario. This equals 64 for
the fully optimum STAP algorithm, and 24 for the reduced
dimension STAP algorithms. The SINR loss of these STAP
algorithms using the covariance estimation techniques de-
scribed in section 2 are shown in Figure l. For comparison,
the loss using the known clutter covariance matrix is plot-
ted using a thin black curve. Because of the small sample
support, the sample covariance matrix yields a poor esti-
mate of the clutter covariance matrix, resulting in severe
I
I
-YL I I1
I I
I - 1 1

SINR loss (8-25 dB) away from mainlobe clutter. The pro- 30 35 40 30 35 40
jected covariance estimator (thin gray curve) yields abysmal RANGE (nmi) RANGE (nmi)
performance for all algorithms because its estimate is not
necessarily positive definite. The weighted projected co- Figure 2: Residual clutter-plus-noise to signal ratio versus
variance estimator (thick gray curve) yields similar perfor- range using the Mountaintop database [ll] with a DPC pre-
mance to the sample covariance matrix, except that the doppler algorithm (data collected 9 March 1994 at White
notch at mainlobe clutter is significantly wider, resulting in Sands, New Mexico). An adaptive weight vector steered
decreased doppler coverage for all algorithms. In contrast, near and away from mainlobe clutter (DOF= 3 . 16, unit
for all but the PRI-staggered post-doppler algorithm, the gain on target) generated using 61 samples of data between
ML covariance estimator (thick black curve) [for P = 4M 25 and 30 nmi is applied to clutter between 30 and 40 nmi.
in Eq. (3)] suffers only minor losses (2-4dB) away from The algorithms’ performance agrees with the simulation re-
mainlobe clutter, and, for all algorithms, has a compara- sults shown in Figure 1.
ble notch width at mainlobe clutter as the ideal (known
covariance) STAP algorithm.
We observe that the ML’s performance improves as the REFERENCES
amount of structure increases, a trend exhibited in Fig- T. A. BARTON. Covariance Estimation for Multidimen-
ure 1. PRI-staggered post-doppler has 32 blocks each of sional Data. D.Sc. diss., Washington University, 1993.
order 8 and yields 4-10 dB of loss away from mainlobe clut- J. P. BURG,D. G . LUENBERGER, and D. L. WENGER. “Es-
ter, whereas the other algorithms have more structure and timation of structured covariance matrices,” Proc. IEEE
less loss. 70 (9) :963-974, 1982.
Structured covariance matrix estimation algorithms ap- A. P. DEMPSTER, N. M. LAIRD,and D. B. RUBIN.L L M ~ -
plied t o Mountaintop data [ll]are shown in Figure 2. This imum likelihood from incomplete data via the EM algc-
rithm,” J. Royal Statisticnl Society B-36:1-38, 1977.
plot shows residual clutter-plus-noise to signal ratio versus D. R. FUHRMANN. “Application of Toeplitz covariance es-
range for a DPC pre-doppler STAP algorithm (14 elements, timation to adaptive beamforming and detection,” IEEE
16 pulses, 3 subapertures) for an adaptive weight vector Ifans. Signal Processing 39 (IO), 1991.
(unit gain on target) steered near mainlobe clutter. D. R. FUHRMANN and T. A. BARTON. “Estimation of block
Toeplitz covariance matrices,” in Proc. 24th Asilomar Conf.
Signals, Systems, Computers 2 :779-783, 1990.
4. CONCLUSIONS M. I. MILLERand D. L. SNYDER. “The role of likelihood and
entropy in incomplete-data problems: Applications to esti-
mating point-process intensities and Toeplitz constrained
In this paper several structured covariance matrix estima- covariances,” Proc. IEEE 75 (7) :892-907, 1987.
tors were considered in conjunction with the performance A. V. OPPENHEIMand R. W. SCHAFER. Digital Signal Pro-
of STAP algorithms that use them. It is shown using simu- cessing. Englewood Cliffs, NJ: Prentice-Hall, Inc., 1975.
lation and experimental data that with limited sample sup- I. S. REED, J. D. MALLETT,and L. E. BRENNAN.
port for a block Toeplitz covariance matrix, an ML estima- “Rapid convergence rate in adaptive arrays,” IEEE Ifans.
tor based upon the E M algorithm can yield superior adap- Aerospace and Electronic Systems AES-10 ( 6 ) :853-863,
1974.
tive nulling performance to other estimators. Based upon
the results in this paper, we make the following recommen- F. C. ROBEY.A Covariance Modeling Approach to Adap-
tive Beamforming and Detection. D.Sc. diss., Washington
dations to those who wish to estimate block Toeplitz co- University, 1990.
variance matrices for adaptive nulling: (1) make sure that S. T. SMITH. “Space-Time clutter covariance matrix com-
the assumption of block Toeplitz structure is valid (e.g., putation and interferencesubspace tracking,” in Proc. 29th
array errors and multipath reflections ruin spatial and tem- Asilomar Conf. Signals, Systems, Computers, 1995.
poral stationarity, respectively); (2) in situations (such as G . W. TITI. L‘Anoverview of the ARPA/Navy Mountain-
data analysis) where greater computational intensity can top Program,” in Proc. IEEE Adaptive Antenna Systems
be afforded, use an ML estimator; (3) if the computational Symposium, November 1994. See also yvv. 11.mit .edu/
intensity of ML cannot be afforded, use a diagonally loaded 11-darpa-sto/mtn-top.
sample covariance matrix; (4) if the weighted projected es- J. WARD.“Space-Time adaptive processing for airborne
radar,” MIT Lincoln Laboratory Technical Report 1015,
timator is used, verify that it yields improved performance DTIC No. ESC-TR-94-109, 13 December 1994.
in an idealized setting.