Special Issue on The International Conference on Information and Communication Systems (ICICS 2011)

ITERATIVE ALGORITHMSFOR RADAR SIGNAL PROCESSING

Dib Samira, Grimes Morad, Ghemit Amal and Hamel Sara
Dept. of Electronics Engineering, University of Jijel, Algeria
samiradib@yahoo.fr
Barkat Mourad, IEEE Fellow
Dept. of Computer Engineering, King Saud University, Riyadh, KSA

ABSTRACT
In this paper, we are interested in the application and the evaluation of the
performances of adaptive recursive subspace-based algorithms of linear
complexity for the suppression of interferences in Space Time Adaptive
Processing (STAP), namely PAST, OPAST, API and FAPI. To highlight their
application in the STAP, we present the reduction of the rank by the principal
components (PC) method. The simulation results will be presented and the
performances of the STAP for a reduced rank will be discussed with a comparative
show that these
study made between the used methods.Performance curves
algorithms allow good indeed detection of slow moving targets even with a low
rank covariance matrix and in a Doppler ambiguous environment.

Keywords:PAST, OPAST, API, FAPI, STAP,airborne radar.

INTRODUCTION

A target in a scenario of airborne surveillance is

darkened by the clutter of ground and the jammer in
multiple dimensions. It was shown in [1,2] that the
target is separated from the clutter and jammer in
two-dimensional angle/Doppler.
Space-time processing can provide a rejection of
such clutter and thus, be able to detect the slow
targets. Brennan and Reed [3] first introduced STAP
to the radar community in 1973. With the recent
advancement of high speed, high performance digital
signal processors, STAP is becoming an integral part
of airborne or space-borne radars for MTI functions.
However, it is well-known that optimum STAP
detection implies a large computational cost, since it
utilizes complex matrix operations and often in an
iterative fashion. In fact, the methods of the STAP
with full rank use all the available degrees of freedom
to eliminate the interferences, so requiring a cost of
high calculation. For this reason, some reduced-rank
STAP algorithms have been developed. In [2-10] and
references therein, it was shown that STAP has the
unique property of compensating for the Doppler
spread induced by the platform motion and thus,
making the detection of slows targets possible.
In this paper, we analyze, at first, the STAP with
reduced rank by using the method of the principal
components (PC). Then, we apply four iterative and
adaptive algorithms of subspace tracking to reduce

UbiCC Journal

the rank. Finally a comparison is made to justify the

use of these algorithms in the radar processing.
The paper is organized as follows. In Section 2,
the mathematical model is given. In Section 3, we
give a brief description of STAP with reduced rank
and define the PC method. The iterative algorithms
proposed are treated in Section 4. Results and
discussions are presented in Section 5. Finally, the
conclusion is presented in Section 6 highlighting the
main results obtained in the paper.
2

MATHEMATICAL MODEL OF DATA

We consider a space time network with N

antennas uniformly spaced and M delay elements for
any antenna at a constant pulse-repetition frequency
(PRF). The data are then processed on one range of
interest which corresponds to one slice of the data
cube as shown in Fig. 1.
A space time snapshot at range k in the presence
of a target is given by [1]
X = S + X i

(1)

where, Xi is the vector of interferences (noise,

jamming and clutter), is the target amplitude and S
is the space-time steering vector given by

Page 18

www.ubicc.org

Special Issue on The International Conference on Information and Communication Systems (ICICS 2011)

and thermal noise. More details for the computation

of these matrices are given in [4].
The performance of the processor can be
discussed in terms of the Improvement Factor (IF). IF
is defined as the ratio of the SINR of the output to
that of the input of the Direct Form Processor (DFP)
and is given by [1]

IFopt =

H
H
W S.S W .tr( R)

(5)

R.W .S H S

(a)
Where W is the optimum weights of the interference
plus noise rejection filter.
Note that a notch, which is a reversed peak of
the clutter, appears at the frequency in the direction
of sight of the radar, while the width of this notch
gives a measurement of the detection of slow moving
targets.
3

The fully adaptive techniques of signal processing

cannot be applied for a real-time processing. The
methods with reduced rank realize the adaptation on a
space of reduced dimensions obtained after an
adaptive transformation on the data. They also exploit
the nature of the low rank matrix of interferences.

(b)
Figure 1: (a) Conventional chain of STAP;
(b) Data cube of STAP

S = St S s , Stand Ss are the temporal and space

vectors given respectively by
St = [1;e

j 2Ft

;e j 2 2 Ft ;...;e j 2 ( M 1) Ft ]

(2)

j 2Fs

;e j 2 2 Fs ;...;e j 2 ( N 1) Fs ]

(3)

and
Ss = [1;e

where Ft = Fd / PRF and Fs = d.sin / are,

respectively, the normalized Doppler and spatial
frequency. d is the distance between the antennas and
is the azimuth angle.
The optimum weight of the STAP, which
maximizes the signal to interference noise ratio
SINR, is obtained to be [1]
Wopt = R 1S

(4)

Where R is the covariance matrix of the

interferences. It is supposed to be known and is the
sum of covariance matrices of the clutter, jammers

UbiCC Journal

STAP WITH REDUCED RANK

The idea is the separation of the overall space

into an interference subspace and a noise subspace
[4,5]. A common method to obtain these subspaces is
via Singular Value Decomposition (SVD) of the
interference-plus-noise covariance matrix. Such
method can reduce the sample support requirement
to O(2r) (number of snapshots), where r is the rank
of the covariance matrix to achieve an SINR of 3 dB
below the optimum, but at the expense of a
considerable computational complexity due to the
SVD O((MN)3). This complexity prevents the realtime applications.
The partially adaptive algorithms of the STAP
consists in transforming the data with a matrix
V C MN r where r<<MN. There are several methods
for the covariance matrix rank reduction [3-10],
which may differ in the shape of the processor as well
as in the selection of the columns of the matrix.
The principal component is based on the
eigenvectors conservation of the matrix of covariance
of interferences corresponding to the dominant
eigenvalues [4]. If we assume that the r columns of V
are a subset of the eigenvectors of R, the
improvement factor of the reduced rank can then be
written as

Page 19

www.ubicc.org

Special Issue on The International Conference on Information and Communication Systems (ICICS 2011)

IF RR = S H V (V

RV ) 1V

tr(V

RV )

S H V .V

Table 1: PAST and OPAST Algorithms.

(6)
Initialization :

W (0) = I MXN

ITERATIVE / ADAPTIVE ALGORITHMS

Z (0) = I r

For t = 1 to Number Snapshots do

The iterative/adaptive algorithms of subspace
tracking allow to follow the temporal variations of
the subspace and to update it in every new
observation. One very large number of algorithms
were proposed in the literature [11-17] and
references therein. They showed their efficiency in
several domains of signal processing in particular in
antennas processing and in spectrum analysis. They
offer interesting perspectives in the STAP. They can
be classified according to their computational
complexity. Schemes requiring O(MN2) or
operations will be classified as high
O((MN)2r),
complexity; algorithms with complexity O(MNr2) as
medium complexity, and schemes requiring only
O(MNr) operations are said of low complexity [17].
In this context, we consider the class of the
fastest, most robust and effective algorithms said of
linear or low complexity because they are the most
important ones due to their suitability for real-time
applications. We propose the application and the
evaluation of the performances of the algorithm
PAST and its orthogonal version (OPAST) and the
algorithm API and its Fast version (FAPI). A brief
description of key idea of these algorithms is as
follows.

PAST Section:

y(t ) = W (t 1) H x(t )
h(t ) = Z (t 1) y(t )
h(t )
g(t ) =
+ y(t ) H h(t )
1
Z (t ) = ( Z (t 1) g(t ) y(t ) H Z (t 1))

e(t ) = x(t ) W (t 1) y(t )

W (t ) = W (t 1) + e(t) g(t ) H
OPAST Section:

(t ) = (t )(x (t ) W (t 1) y (t ))

2
1
(t ) =

2
h(t ) 1 + 1 ( t) 2 h (t ) 2

(t )
(t )
2
e(t ) =
W (t 1)h(t ) + (1+ 2 h(t ) (t )

1
1
Z (t ) = ( Z (t 1) g(t ) y(t ) H Z (t 1))

W (t ) =W (t 1) + e (t )g H (t )
2

4.1 PAST and OPAST algorithms

The Projection Approximation Subspace Tracking
(PAST) [12] is based on the optimization of the
following criterion:

J (W (t )) =

t i x(i) W (t ). y(i)

(7)

i =1

where x is the observed data vector and W is the

estimated interference subspace basis and is a
forgetting factor, 0 < 1 and y(i) = W H (i 1).x(t )
this cost function has a global minimum which yields
a non orthonormal basis of the interference subspace
and which may be attained by a RLS adaptive
algorithm given in table 1.
In order to resolve the problem of convergence
and orthogonality of the weighty matrix, the
Orthonormal PAST (OPAST) was presented [13].

UbiCC Journal

End

4.2 API and FAPI algorithms

The Approximated Power Iteration (API) and
Fast Approximated Power Iteration (FAPI)
algorithms derive from the power method [14]. A
less restrictive approximation than for PAST is used.
Indeed it concerns the projection on the estimated
subspace instead of the estimated subspace itself:
W (i).W

(i) W (i 1).W

(i 1)

(8)

The obtained subspace is found to be

orthonormalized. The FAPI algorithm is a fast
implementation of API. API and FAPI codes are
presented in table 2.

Page 20

www.ubicc.org

Special Issue on The International Conference on Information and Communication Systems (ICICS 2011)

Table 2: API and FAPI Algorithms.

(CNR) is set equal to 8dB. This clutter covers the

band [30, 30]. For the iterative algorithms, we use
the improvement factor given by Eq. (5) where
W = (I ww H )S and w is the estimated subspace of
interferences. The forgetting factor of both algorithms
is fixed at = 0.99 .
Fig. 2 shows the eigenspectra for the covariance
matrix. We note a clear distinction between the
interferences subspace and noise subspace. An
extension to Brenanns rule, for a number of an
effective rank for the covariance matrix of a sidelooking radar, has been derived recently in [8] and is
given by r=N+(+J)(M1), where J denotes the
number of jammers. The high-input eigenvalue is
about max=CNR+ 10 log(NM)(dB)). This causes a
slow convergence which is inherent to adaptive
algorithms [17]. These values can actually be read
directly from Fig. 2.

Initialization :
;

Z (0) = I r

PAST Section(see Table 1)

API Section:
1

(t ) =

I + e(t ) g(t ) g(t ) H

Z (t ) =

(I g(t ) y(t ) H Z (t 1)(t ) H )

W (t ) = (W (t 1) + e(t) g(t ) H )(t )

FAPI Section:

2 (t) = x(t)

y(t)

1
2 (t )
(t ) =
2
2
1 + (t ) g(t ) 1 + 2 (t ) g(t )

We recall that the aim of this paper is to

demonstrate the robustness of the proposed iterative
algorithms PAST, OPAST, API and FAPI by
comparing their performances to PC method for low
rank covariance matrices and in an ambiguous
scenario.

(t) = 1 (t) g(t)

y' (t ) = (t ) y(t ) + (t ) g(t )

We start by studying the influence of the

dimension of the STAP on the number of eigenvalues
of the covariance matrix that has given us guidance
on environment statistics in which the target is to be
tested. It can be seen clearly from Fig.2that an
increase in MN leads to an increase in powers and in
the number of degrees of freedom of the system
filter. We also note that the interference-noise space
can be separated into two subspaces: the interference
subspace and the noise subspace.

h' (t ) = Z (t 1) H y' (t )
(t)
(t) =
(Z(t 1)g(t) (h'(t)g(t))g(t)
(t)
1
Z (t ) = (Z (t 1) g(t )h' H (t )+ (t ) g H (t ))

e' (t ) = (t ) x(t ) W (t 1) y' (t )

W (t ) = W (t 1) + e' (t ) g H (t )
End

RESULTS AND DISCUSSIONS

In this Section, we discuss the influence of the PC

method, based on the reduction of the rank of the
covariance matrix, and the iterative algorithms
(PAST, OPAST, API and FAPI) on the detection of a
target with a low power (SNR=0dB) and with a slow
speed. The simulated environment is a linear side
looking network of N=8 antennas spaced out by half
of the emitted wavelength d = / 2 and M=10
impulses in the coherent processing cube. The
dimension of the adaptive process is thus MN = 80.
The elevation angle is fixed to 20. The speed of the
airborne radar is VR=100m/s, and the frequency of
transmission is 0.3GHz. The environment of
interferences consists of five jammers and ground
clutter. The jammers are at azimuth angles of 0,
180, 60, 90, and 72, respectively, and with
respective ratios of jammer to noise (JNRs) of 13dB,
12dB, 11dB, 10dB and 9dB. The clutter to noise ratio

UbiCC Journal

As the number of eigenvalues is a measure of the

degree of freedom of clutter elimination filter, we can
notice that the number of small eigenvalues is high,
while the number of those with a low value is
significantly large, which increases the degree of
freedom.

Page 21

60
MN=20
MN=40
MN=80

50

MN=160
40
Eigenvalues (dB)

W (0) = I MXN

For t = 1 to Number Snapshots do

30

20

10

-10

20

40

60
80
100
Numbers of eigenvalues

120

140

160

Figure 2: Eigenvalues of the space-time covariance

matrix R in the presence of two jammers: with
JNR=35 dB, CNR = 25 dB

www.ubicc.org

Special Issue on The International Conference on Information and Communication Systems (ICICS 2011)

UbiCC Journal

Page 22

40
0.4

Normalised Doppler Frequency, Ft

35
0.3
30

0.2
0.1

25

20

-0.1
15
-0.2
10
-0.3
5

-0.4

-80

-60

-40

-20

0
20
Azimut ()

40

60

80

Normalised Doppler Frequency, Ft

(a)

0.4

40

0.3

35

0.2

30

0.1
25
0
20
-0.1
15
-0.2
10

-0.3

-0.4

-80

-60

-40

-20

0
20
Azimut ()

40

60

80

(b)

45

0.4

Normalised Doppler Frequency, Ft

If we work with all size of the covariance matrix

(MNxMN), there will be difficulties in implementing
real-time especially since it also requires very
cumbersome equipment. So the use of rank reduction
methods reduce the computational complexity and
therefore the cost, and this by taking a dimension
(rxr) with r the rank which must be very inferior to
MN. This choice is motivated by the fact that the
observation space is divided into two areas (noise
and interference) as it was commented on in Fig. 2.
On the other hand, it is important to consider the
performances of the optimal processor according to
the values of the PRF. Using the spectral analysis and
varying the PRF, Fig. 3shows the minimum variance
spectra of the received signal in the presence of two
jammers for different values of PRF. There is the
presence of aliasing of the clutter for high values of
PRF. We can therefore say that the choice of the PRF
is essential to ensure good detection. We can
conclude that high values of PRF lead to Doppler
ambiguities and the clutter spectrum becomes
ambiguous so that blind velocities occur.
This analysis can be addressed by using the
Improved Factor and varying the PRF (Fig. 4 and
Fig. 5). In the case where PRF is equal to (2VR/),
we see the appearance of ambiguous notches with
speeds associated are called blind velocities and the
detection becomes relatively difficult. However, for
the case of PRF equal to (4VR/), we notice that the
ambiguous notches are eliminated and thus the
detection is improved.
Fig. 4 illustrates the effect of the rank on the
detection of slow targets for an environment without
ambiguities. We notice that there is a strong
degradation in the performances for a small rank. The
same detection results are obtained when r = 20 and
r = 40, and thus to save calculation time, we choose
r=20 as a good reduction of the rank. We observe
that if the rank is very small, the slow moving targets
will be removed with the clutter, and thus will not be
detected.
Similar findings were made for an ambiguous
scenario, Fig.5. We notice the appearance of new
weak undulations in the bandwidth due to the
estimate used for the rank reduction. Also, the
appearance of ambiguous notches and the width of
the notch do not change. Thus the rank reduction does
not eliminate the ambiguities of the clutter during the
suppression of the noise. We can then say that we
should not reduce up to the rank to very low values.

0.3

40

0.2

35

0.1

30

25

-0.1

20

-0.2

15

-0.3

10

-0.4

-80

-60

-40

-20

0
20
Azimut ()

40

60

80

(c)
Figure 3: Angle/Doppler spectrum for known R in
the presence of two jammers at -60 and 60 with
JNR=45 dB, N=8, M=10, CNR = 20 dB:
(a) PRF = 8.VR / ;(b) PRF = 4.VR / and
(c) PRF = 2.V R /

www.ubicc.org

10

10

-5

-5

-10

-10

-15

-15

IF(db)

IF(dB)

Special Issue on The International Conference on Information and Communication Systems (ICICS 2011)

-20
-25

-25

-30

-30
Optimal
r=40
r=20
r=8
-0.4

-0.3

-0.2

-0.1

0
Ft

0.1

0.2

0.3

0.4

-40

r=8
-45
-2

0.5

Figure 4: Improvement factor for the PC-DFP with

different values of r and with PRF = 4.VR /
For the iterative algorithms, we use the
improvement factor given by the Eq. (5) where
W = (I ww H )S , and w is the estimated subspace of
interferences. The factor of forgetting of both
algorithms is fixed at = 0.99 .
To compare the different algorithms presented in
this paper, we consider the two cases of scenarios and
two limited values of the rank. In fact, we first
consider
the
case
without
ambiguities, PRF = 4.VR / , and we draw the
variations of the improvement factor, IF, according to
the normalized Doppler frequency (Ft). For medium
value of rank r = 20, we can see from Fig. 6(a) that all
the tested algorithms are globally equivalent and
present good performances: the notch is relatively
thin comparing to the optimal processor which leads
to the detection of slow targets. We observe from
Fig.6(b) that recursive algorithms perform better in
the case of low rank value.
In the presence of ambiguities ( PRF = 2.VR / ),
we see from Fig. 7, the appearance of ambiguous
notches. This is due to the aliasing phenomenon well
known, which becomes more important with the
decline in PRF and the effect of Doppler ambiguities
will be more important and makes target detection
more difficult. Fig. 7(a) shows that the iterative
algorithms present acceptable detection performances
than those of PC method.
We notice from Fig. 7(b) that there is a strong
degradation in the performances for a small rank and
the detection becomes impossible for the PC method.
Contrarily, the use of recursive algorithms
overcomes this problem and improves the detection.
In this case, PAST and API performances are
comparable to that obtained by the optimal
processor. In conclusion, we can note that PAST,
OPAST, API and FAPI algorithms give a much better
performance than the PC method, in terms of
detection when the radar operates in an unambiguous
environment and with low rank of reduction.

UbiCC Journal

-1.5

-1

-0.5

0
Ft

0.5

1.5

Figure 5: Improvement factor for the PC-DFP with

different values of r and with PRF = 2.V R /

Page 23

5
0
-5
-10
-15
IF(dB)

-40

Optimal
r=40
r=20

-35

-20
-25
Optimal
PC
PAST
OPAST
API
FAPI

-30
-35
-40
-45
-0.5

-0.4

-0.3

-0.2

-0.1

0
Ft

0.1

0.2

0.3

0.4

0.5

(a)
5
0
-5
-10
-15
IF(dB)

-35

-45
-0.5

-20

-20
-25
Optimal
PC
PAST
OPAST
API
FAPI

-30
-35
-40
-45
-0.5

-0.4

-0.3

-0.2

-0.1

0
Ft

0.1

0.2

0.3

0.4

0.5

(b)
Figure 6: Improvement factor for the iterative
algorithms and PC-DFP with PRF = 4.VR / :
(a) r = 20; (b) r = 8

www.ubicc.org

Special Issue on The International Conference on Information and Communication Systems (ICICS 2011)

algorithms which can estimate recursively the

weights of the clutter rejection filter. With a very low
covariance matrix rank and with Doppler
ambiguities, the simulation results confirmed the
superior performance of the proposed algorithms
compared to the PC method even with the optimal
filter.

5
0
-5
-10

IF(dB)

-15
-20
-25
Optimal
PC
PAST
OPAST
API
FAPI

-30
-35
-40
-45
-2

-1.5

-1

-0.5

0
Ft

0.5

1.5

This comparative study proved that iterative

algorithms could be applied for the reduction of the
rank for the STAP because they give similar
performances as those given by the methods of rank
reduction, but gave a much better performance for
low rank values. In addition, they present a very low
computational complexity. In fact, it can be viewed
From Tables 1 and 2, that the complexity burden is
O(MNr) instead of O((MN)3) for the PC processor.
Thats why these algorithms can be considered as an
economical approach in comparison with the other
techniques.

(a)
5

REFERENCES

0
-5
-10

IF(dB)

-15
-20
-25
Optimal
PC
PAST
OPAST
API
FAPI

-30
-35
-40
-45
-2

-1.5

-1

-0.5

0
Ft

0.5

1.5

(b)
Figure 7: Improvement factor for the iterative
algorithms and PC-DFP with PRF = 2.VR / :
(a) r = 20; (b) r = 8

CONCLUSION

In this paper, we considered the application and

the evaluation of the performances of four adaptive
recursive subspace-based algorithms of linear
complexity and a comparison with reduced rank
methods and the optimal filter.
We observed that the presence of ambiguities and
the reduction of rank to low values degraded the
performance of STAP in suppressing interferences
and detecting slow targets. We showed that we can
mitigate these problems by using recursive

UbiCC Journal

[1] J. Ward: Space-Time Adaptive Processing for

airborne radar, Technical Report 1015, Lincoln
Laboratory MIT (1994).
[2] R. Klemm: Space Time Adaptive Processing
Principles and applications, The Institution of
Electrical Engineers, London (1998).
[3] L.E. Brennan and I.S. Reed: Theory of radar,
IEEE transactions on Aerospace and Electronics,
Vol. AES-9, n2, pp. 237-252(1973).
[4] H. Nguyen: Robust Steering Vector Mismatch
Techniques for reduced Rank Adaptive Array
Signal Processing, PhD Thesis in Electrical and
Computer Engineering, Virginia, USA (2002).
[5] J.S. Goldstein and I.S. Reed: Theory Theory of
partially adaptive radar, IEEE Transactions on
Aerospace and Electronics Systems In
Proceedings of the IEEE National Radar
Conference, Vol. 33, n4, pp. 1309-1325 (1997).
[6] S.D. Berger, and B.M. Welsh: Selecting a
reduced-rank transformation for STAP, a direct
form perspective, IEEE Transactions on
Aerospace and Electronics Systems, Vol. 35, n2,
pp. 722-729 (1999).
[7] J.R. Guerci et al.,: Optimal and adaptive
reduced-rank STAP, IEEE Transactions on
Aerospace and Electronics Systems, Vol. 36,
n2, pp. 647-663 (2000).
[8] P. G. Richardson: STAP covariance matrix
structure and its impact on clutter plus jamming
suppression solutions, Electronics Letters, Vol.
37, n2, pp. 118119 (2001).
[9] S. Dib et al.,: A Reduced Rank STAP with
Change of PRF, Eusipco2007, Poznan, Poland,
pp. 95-122 (2007).

Page 24

www.ubicc.org

Special Issue on The International Conference on Information and Communication Systems (ICICS 2011)

[10] S. Dibet al.,: Two-dimensional signal adaptive

processing for airborne radar, ICDIPC 2011, Part
II, CCIS 189, Springer, pp.244-257 (2011).
[11] P. Comon and G.H. Golub:Tracking a few
extreme singular values and vectors in signal
processing, Proc. Of the IEEE, Vol. 78, n8, pp.
1327-1343 (1990).
[12] B.
Yang:
Projection
Approximation
SubspaceTracking, IEEE-T-SP,Vol. 44, n1, pp.
95-107 (1996).
[13] K. Abed-Meraim A. Chkeif and Y. Hua: Fast
Orthonormal PAST Algorithm, IEEE Signal
Processing Letters, Vol. 7, pp. 6062 (2000).
[14] R. Badeau et al.,: Fast approximated power
iteration subspace tracking, IEEE Transactions
on Signal Processing, Vol. 53, pp. 29312941
(2005).

UbiCC Journal

[15] A.Valizadeh and M. Karimi: Fast Subspace

tracking algorithm based on the constrained
projection approximation, EURASIP Journal on
Advances in Signal Processing (2009).
[16] M. Kamenetsky and B. Widrow: A variable
leaky LMS adaptive algorithm, in Proceedings
of the IEEE conference of the 38thAsilomar on
Signals, Systems and Computers, Vol. 1, pp.
125128, Pacific Grove, Calif, USA (2004).
[17] L. Yang: Design and analysis of adaptive noise
subspace estimation algorithms, PhD Thesis,
Department of Electrical and Computer
Engineering National University of Singapore
(2009).

Page 25

www.ubicc.org