Space-Time Adaptive Processing (STAP) in Wireless Communications

Bobby Barnes
EE360
Spring 2003

Abstract
Space-Time Adaptive Processing (STAP) algorithms have been proven to be a very
effective way to mitigate the effects of multipath and interference in wireless
communication systems. However, when implementing the STAP algorithm several
design issues arise. This paper will examine two proposed implementations of STAP
algorithms. These implementations will include covariance matrix estimation (a
statistical method) and Direct Data Domain (D3) processing approach (a non-statistical
method). These two STAP methods will be compared and contrasted and the
implications of STAP on wireless networks will be examined.

Contents
1. Background
2. Interference Covariance Matrix Approach
3. Direct Data Domain Approach
4. Comparison/Contrast
5. Conclusion
6. References

Background: What is Space-Time Adaptive Processing (STAP)?

Space-Time Adaptive Processing (STAP) is a signal processing technique used to

suppress the effects of co-channel interference, ISI, and jammers in wireless
communications systems. From the implementation of STAP algorithms, greater
capacity gains and communication quality can be realized [3].

The fundamental principle in all STAP algorithms involves the usage of multiple receive
antennas on the receiving platform. Spacing the antennas apart by at least half the
wavelength of the desired signal provides space diversity, which helps mitigate the
effects of fading. Furthermore, the incoming signals on each antenna element are
adaptively weighted using a variety of algorithms in order to steer the antenna gain
towards the desired signals while nulling the signals from unwanted noise and
interference. Figure 1 below shows the baseline setup for most STAP implementations.
 Antenna

Receiver

Adaptive
Weight
X X X X

Figure 1. Baseline STAP receiver configuration.

STAP algorithms will typically begin with the configuration above in Figure 1 and add
extensions such as tapped delay lines behind each antenna receiver or different antenna
array topologies.

Two algorithm approaches to STAP will be analyzed in this paper. They are interference
covariance matrix estimation and direct data domain (D3) processing.

Interference Covariance Matrix Estimation Approach

Interference covariance matrix estimation is the well-known classical approach to STAP

implementation. Research using this approach has been around since the 70’s. The
approach begins by utilizing the initial antenna structure in Figure 1 and adding tapped
delay lines after each receiver. Figure 2 illustrates this new configuration.

Antennas

Receivers

Adaptive X X X X
Weights

τ τ τ τ

X X X X Output
Σ
τ τ τ τ

X X X X

Figure 2. N-element antenna array with M taps.

As observed in the above figure, each of the taped delay signals is adaptively weighted as
before according to the designer’s processing algorithm that minimizes interference and
noise while preserving or enhancing the desired signal. The final output signal is the sum
of all the weighted taps.

For a system of N antenna elements and M taps per element, a total of NxM signals are
received and weighted. Thus, the input signal has the form:

XT(k) = [x1(k∆ ), x2(k∆ ), … xNM(k∆ )]

where (T) denotes transpose and k∆ is the multiple of the time the sample is taken. The
received voltages on the taps are the sum of the desired signal in the direction of interest
plus noise. Hence,

X(k) = L(k) + N(k)

where L(k) is the contribution on each tap from the desired signal in the “look direction”
and N(k) is the noise contribution on each tap[6].

Again, because of the NxM signal taps, L(k) & N(k) have the form:

L(k) = [l11(k∆ ) , l12(k∆ ), …l1M(k∆ ) …

l21(k∆ - τ ) , l22(k∆ - τ ), …l2M(k∆ - τ ) …
…
…
lN1(k∆ -(M-1)τ ), lN2(k∆ -(M-1)τ ), …lNM(k∆ -(M-1)τ )]

and

N(k) = [n1(k∆ ), n2(k∆ ), …nNM(k∆ )].

The weight vector containing the weights after each tap has a similar form:

WT = [w11,…w1M, …, wN1,…wNM].

It is shown in [2] and [6] that noise and interference power not in the “look direction” of
the desired signal can be minimized by minimizing the power of the system output. The
system output is given by:

y(k) = Σ n=1 Σ m=1 wnmxn(k+M-m)

and the expected power is

P = E{|y(k)|2} = E{(Σ n=1 Σ m=1 wnmxn(k+M-m))( Σ i=1 Σ j=1 wijxi(k+M-j))

= E{Σ n=1 Σ m=1 Σ i=1 Σ j=1 wnmxn(k+M-m) xi(k+M-j) wij}
However, since the weights and tap voltages can be expressed as vectors, the output
signal simplifies to:

y(k) = WTX(k)

and the expected output power becomes

P = E[y2(k)]
= E[WTX(k)XT(k)W]
= WTRXXW.

The covariance structure (RXX) of the tap voltages is easily recognizable in this solution.

Now adapting the weights to minimize the expected output power would yield the trivial
result of zeros. This would also completely cancel the desired signal. Thus, an additional
constraint must be placed on the weight calculation.

Both [2] and [6] show that a linear constraint must also be placed on the weights using
this STAP algorithm. This linear constraint has the form

cHw = 1

where c is the NMx1 constraint vector (designed to maintain the response of the filter to
the desired target), (H) is the conjugate transpose, and w is the vector of weights
containing the NM filter weights [2].

Minimizing the output power now according to the linear constraint yields the following
solution:

w = (R-1c)/(cHR-1c)

with R representing the covariance structure (RXX) as before.

This methodology for STAP has proven to be very effective at reducing the effects of
noise and interference in wireless communication systems and is especially useful in
Pseudo-randomly coded systems with signals below the noise floor.

Direct Data Domain (D3) Approach

The direct data domain (D3) STAP approach begins with the baseline STAP receiver
setup as shown previously in Figure 1. The are no tapped delay lines following the
weight on each receiver element; therefore, the voltages are then summed and the output
y(t) is given. See Figure 3 below.
 Antenna

Receiver

Adaptive X X X X
Weight

Output
Figure 3. N-element antenna array.

Let s(t) denote the desired signal. The observed signal voltage on each antenna element
is the sum of the desired signal plus noise and interference. Hence,

xk(t) = sk(t) + nk(t) k = 1,2, … N

where xk(t) is the received voltage on an antenna and sk(t) & nk(t) are the desired signal
and noise contributions on the antenna respectively.

If θ is the direction of the desired signal, then the received signal on each antenna
element can be modeled by

sk(t) = s(t)exp{j(2π kd/λ )sinθ }

where d is the distance between antenna elements and λ is the wavelength of the desired
signal [7].

The received signal after weighting then becomes

y(t) = Σ κ=1 wkxk(t)

or in matrix form

y(t) = wTx

where (T) denotes transpose and

wT = [w1,w2, … wM]
xT = [x1, x2, … xM]

M is the number of degrees of freedom set by M = (N-1)/2 [7]. In [7] it is assumed that
the weights, wk, perform a weighted average of the received signal, xk.

If the actual received voltages are written as a vector X and the modeled received desired
signals are written as a vector S, then the difference

[ ][ ]
X1 X2 … XM S1 S2 … SM
X2 X3 … Xm+1 S2 S3 … Sm+1
: -α :
: :
Xm Xm+1 … XN Sm Sm+1 … SN

contains just the noise and interference which are desired to be eliminated [7]. This can
also be written as
_ _
X - αS

Again, since we wish to minimize this term, the goal of D3 is to find an adaptive
weighting vector that sets this to zero. ie,

[X - αS] x [w] = [0]

This equation is already constrained in the desired signal direction since the S matrix is
defined using the direction of arrival, θ , of the desired signal as described earlier.

Lastly, this equation can be rewritten to show its eigenvalue form:

Xw = αSw

Comparison/Contrast between Interference Covariance Matrix

Estimation and D3 Algorithms

Although not initially apparent, on the surface interference covariance matrix estimation
and D3 are very similar, specifically with respect to antenna array receiver structure and
data collection.

As mentioned before, both algorithms utilize the baseline STAP receiver configuration
shown in Figure 1. It may appear, however, that the two structures differ since the D3
approach does not incorporate multiple tapped delay lines on each antenna element.
Nevertheless, this still results in a similarity between the two when observed in
conjunction with the data collection process as described below.

D3 algorithms are often used for radar implementations of STAP. In these radar
implementations, M pulses are transmitted and received by the antenna array. For each
pulse repetition interval, R time samples are collected to cover the desired range interval
[4]. Thus, R becomes the number of range cells. Therefore with M pulses, N antenna
elements, and R time samples per pulse, the total received data for over one processing
interval consists of M x N x R samples. Figure 4a below shows what this looks like.
M-Pulses

M-Taps

es
es

pl
pl

m
am

Sa
eS

e
m

m
Ti

Ti
R-

N-Antenna Elements N-Antenna Elements

Figure 4a. D3 Data Cube Figure 4b. Classical Data Cube

On the other hand, in the interference covariance estimation STAP approach, there are M
taps for each antenna element in a given time sample. With T samples in a given
observation time, the total received data per processing interval under this algorithm is
MxNxT samples, as diagramed above in Figure 4b. Hence, if the number of taps and
time samples (T) equals the number of pules and time samples (R), then the total amount
of data needed to perform the signal processing calculations to form the adaptive weights
are the same in both cases. Ironically, however, the processing of these similar data
cubes causes significant performance differences between these two algorithms.

The classical interference covariance estimation approach to STAP has been proven to
have excellent interference cancellation performance in homogeneous correlated
interference scenarios. On the other hand, according to [8], non-statistical algorithms
such as D3 are known to have sharply inferior performance in similar scenarios.

However, [8] also submits that statistical algorithms, such as the one in this comparison,
fail when the data that is collected comes from non-homogeneous environments. This is
due to the fact that the secondary data does not contain the statistics of the interference.
Since D3 algorithms do not calculate covariance matrices, non-homogeneous scenarios do
not affect their performance.

Furthermore, an additional consideration between these two algorithms is convergence

time. By re-examining the adaptive weighting formulas presented earlier, it can be seen
that a matrix inversion of a NM x NM matrix is required in order to calculate the
covariance and adaptive weights for the interference covariance matrix estimation
approach. Even for relatively small values of N and M, a large matrix inversion is
required, which causes a high computational burden and a low convergence rate. This
makes adaptive processing very difficult to do in real-time systems.

Conclusion
It appears that choosing between these two algorithms for STAP systems is highly
problem dependent. On one hand, statistical methods (covariance estimation) greatly
outperform non-statistical methods (D3) when the interference environment is known to
be homogeneous. However, if the interference scenario ever becomes heterogeneous,
statistical methods will fail. In addition, the computational burden of statistical STAP
algorithms makes real-time adaptive processing difficult or even impossible in highly
transient environments. This lends non-statistical methods such as D3 as the algorithm of
choice with inferior homogeneous performance as a tradeoff. Solutions to this problem
may lie in the development of algorithms that combine previous statistical and non-
statistical algorithms. Recent research such as that in [8] is looking into this possibility.

Implications of STAP in Wireless Networks

Despite which STAP implementation is chosen, the implications of STAP are substantial
and worth looking into for development. We will begin looking at the effects of STAP
by first noting the capacity equation for CDMA below.

N = 1 + (W/R)/(Eb/N0)

N is the number of users, W is the RF bandwidth, R is the bit rate, Eb is the energy per
bit, and N0 is the noise power. Research has shown that STAP algorithms can
theoretically reduce noise power by up to 6dB. Thus, a non-STAP system with a
minimum threshold Eb/N0 = 10dB would benefit with a reduction of the threshold SNR to
4dB with the implementation of a STAP algorithm. This would in turn increase the
system’s capacity of users, N.

For example,

Let R = 9600bps, W = 1.25MHz, and Eb/N0 (threshold) = 10dB. Then, capacity, N = 1 +

1.25x106/9600/10 = 14 users without STAP. With STAP, the threshold SNR would
become 4db and the system capacity, N = 1 + 1.25x106/9600/4 = 33 users. As this
example shows, the capacity of the system is more than doubled with the implementation
of a STAP algorithm. However, the benefits of STAP implementations in wireless
systems are not limited to capacity enhancements. Using a similar exercise, by holding
the number of users fixed, it is easily shown that similar increases in the bit rate, R, can
be obtained instead.
References

1. T.K. Sarkar, “Space-time adaptive processing using circular arrays”, in IEEE

Antennas and Propagation Magazine, Vol: 43 Issue 4, Aug. 2001 pp793-794.

2. P.G. Richardson, “STAP covariance matrix structure and its impact on clutter plus
jamming suppression solutions”, Electronics Letters, Vol: 37 Issue: 2, Jan. 2001
pp118-119.

3. Kehu Yang, “A signal subspace-base subband approach to space-time adaptive

processing for mobile communications”, IEEE Transactions on Signal Processing,
Vol: 49 Issue: 2, Feb. 2001 pp401-413.

4. Tapan Sarkar, “A deterministic least-squares approach to space-time adaptive

processing (STAP)”, IEEE Transactions on Antennas and Propagation, Vol: 49 Issue:
1, Jan 2001 pp91-103.

5. G.A. Fabrizio, “An advanced STAP implementation for surveillance radar systems”,
IEEE Signal Processing Workshop on Statistical Signal Processing, 2001 pp134-137.

6. Otis Lamont Frost, “An algorithm for linearly constrained adaptive array
processing”, Proceedings of the IEEE, Vol: 60, No. 8, August 1972.

7. Sheeyun Park, “Prevention of signal cancellation in an adaptive nulling problem”,

Antennas and Propagation Society International Symposium, 1997. IEEE., 1997
Digest, Vol: 2, 1997, pp1040-1043.

8. Adve, “A two stage hybrid space-time adaptive processing algorithm”, Radar

Conference, 1999. The Record of the 1999 IEEE, 1999, pp279-284.