You are on page 1of 12

1.

I ntroduction

The Power-inversion Adaptive Array: Concept and Performance


R.T. COMPTON, JR. The Ohio State University ElectroScience Laboratory

Abstract

The use of a power-inversion adaptive array to improve the signalto-interference ratio in a communication system is described. "Power inversion" refers to the ability of an adaptive array to invert the power ratio of two received signals. The power-inversion technique is useful because it does not require detailed knowledge about desired-signal structure or arrival angle. The performance attainable with a power-inversion array is described and typical design curves are presented.

Manuscript received January 11, 1978; revised March 8,1979, and April 25, 1979. The work was supported in part by Naval Air Systems Command under Contract N00019-79-C-0291 and in part by the Joint Services Electronics Program under Contract N00014-78-C-0049. Author's address: Department of Electrical Engineering, The Ohio State University, 2015 Neil Avenue, Columbus, OH 43210. 0018-9251/79/1100-803 $00.75 2979 IEEE

This paper discusses the use of a power-inversion adaptive array to improve the signal-to-interference ratio (SIR) in a communication system. "Power inversion" refers to the ability of an adaptive array to invert the power ratio of two received signals. It does this by nuiling the strong signal in favor of the weak one. The power-inversion technique is useful because it does not require detailed information about desired signal structure or arrival angle. The power-inversion technique has been discussed previously oy Compton [1 ,Scliwegman and Compton [21, and Zahin [3]. The least mean square (LMS) adaptive array (due to Widrow et al. [4] ) requires an estimate of the desired signal for use as the reference signal [5] in the array.' When a desired-signal waveform can be estimated in the receiver, the LMS algorithm is a very useful technique. For certain communication systems it is possible to obtain a suitable reference signal [6, 7] while for others it is not. When a reference signal can be obtained, the LMS adaptive array automatically tracks the desired signal and nulls interference. When no reference signal is available, however, the LMS array cannot be used. Tlie power-inversion adaptive array is basically the same as the Howells-Applebaum array [8, 9]. For power inversion, however, the array is used differently than in its radar application. First, the number of degrees of freedom in the array must equal the number of interfering signals. Second, the steering vector is chosen differently than for radar. (It is assumed that the desired-signal arrival angle is not known in advance.) Finally, the loop gain must be chosen to optimize the output signal-tointerference-plus-noise ratio (SINR) over the dynamic range of the received signals. The purpose of this paper is to examine the performance of a power-inversion array in a form useful for system design. Curves will be presented showing the performance of a two-element power-inversion array as a function of desired-signal power, interference power, thermal-noise power, feedback-loop gain, signal arrival angles, and signal bandwidths. We will show that the power-inversion teclhnique is most useful in two situations: either 1) when the received desired-signal level is nearly constant, or 2) when an SINR of less than 0 dB is required at the array output (as in some spread spectrum systems.) Section II discusses the number of degrees of freedom in an array pattern. Section III discusses the power-inversion concept, and the relation between the LMS algorithm and the Howells-Applebaum array. Section IV provides detailed curves of the SINR performance of the powerinversion array as a function of various parameters, particularly the loop gain.

1 The reference signal is called the "desired response" in Widrow et al. [4 ].

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS

VOL. AES-15, NO. 6 NOVEMBER 1979

803

Ill. Power-inversion Feedback

S (t, 0)

Fig. 1. An N-element linear array.

11. Pattern Degrees of Freedom

Consider an N-element linear array as shown in Fig. 1. Assume for simplicity that each element is omnidirectional and that all mutual impedances are zero. Let a unit amplitude CW signal at frequency w0 propagate into the array from angle 0 relative to broadside. The complex signal produced on the ith element is

Xi= exp [-j(27rQi/X) sin 0 ] exp(fco t)

(1)

where X is the free-space wavelength and Qi is the distance between element 1 and element i. (We arbitrarily define the phase of the signal on element 1 to be zero.) Each xi is multiplied by a complex constant wi and then added to produce the array output S (t, 0)

S(t, 0) = w, {I + (w2/wl) exp[-j(27rQ2/X) sin 0]


+ ... +

The remarks above suggest a method of using an adaptive array to obtain interference protection for a communication system. Assume two signals are to be received, one a weak desired signal and the other a strong interference signal. We wish to suppress the interference signal. Suppose we arrange to receive these signals with a two-element array. Suppose further that the weights wi in this array are adjusted to minimize the array output power (with the constraint that not all wi = 0). The array output power will be minimized when the array has directed its only available null at the strong interference signal. As a result, the interference will be attenuated by the null while the desired signal will not be in a null. The SIR will be improved by the pattern ratio in the two directions. This is the essence of the power-inversion technique in adaptive arrays. How do we adjust the weights so the array output power is minimized? This may be done with the feedback concept due to Howells [8] and Applebaum [9]. This adaptive array concept is closely related to the LMS algorithm of Widrow et al. [4] and for our purposes it will be helpful to compare these two systems. Fig. 2 shows a two-element adaptive array. The signal from each element is split into quadrature components with a quadrature hybrid. Let x1j (t) and xI2 (t) be the in-phase components and xQ 1 (t) and XQ2 (t) be the quadrature components. Each xji(t) or xQi(t) is multiplied by a real weight wji or WQi and then summed to produce the array output s(t). In the LMS algorithm [4] the weights are derived from the control equation

= (wN/wI) exp[-j(27rQN/X) sin 0] } exp(jwo t). (2) dwpi/dt -k(a/awp i

[E2(t)]
i= 1, 2
(3)

S (t, 0) describes the voltage pattern of the array for any given weights wi. We have factored one of the constants, w1, out of the expression, and in this form, the angular dependence of 13j(t, 0) is contained in the bracketed term. Since this term contains N -1 complex constants (w2/wl, w3/w1, ..., wN/wl), there are N - 1 degrees of freedom in the pattern of the array at frequency co0. The Nth constant w1 does not affect the pattern (the relative response versus angle), but merely controls the overall scale constant (magnitude and absolute phase) of the antenna response. Thus an N-element array has N -1 degrees of freedom in its pattern at a given frequency. Therefore, with a two-element array, we may place one null in an arbitrary direction in space (for a given frequency). Once that null is specified, however, there is no further flexibility in the pattern other than its overall amplitude and phase. With a three-element array we may place nulls in two arbitrary directions; or, we may place a null in one direction and a beam maximum (d Is IldO = 0) in another. In either case, the pattern flexibility will be "used up" once we have specified two directions.
804

= 2ke(t) x (t),

where P denotes in-phase or quadrature and where e(t) is the error signal, defined as the difference between a reference signal R(t) and the array output s(t), 2 6(t) = R (t)- s(t) = R (t) - iEl [wI X,I(t)
+ WQ xQ
i
i

(t)]

(4)

The behavior of the weights in this array can most easily be studied by making use of analytic signal notation [10-12]. If the quadrature hybrids are assumed to be broadband quadrature hybrids, then
XQ (t)
=

(3) is shown in Fig. 3.

as shown in Fig. 2. The feedback loop corresponding to

Xs

t)

(5)

where the caret denotes the Hilbert Transform [10]. We define the analytic signal x'(t) associated with the ith antenna element as

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC

SYSTEEMS

VOL. AES-15, NO. 6

NOVEMBER 1979

Xi(t)
Xi(t)
x
=

=XIY() + ix
1-11

i(t) =XI (t) + I(t)

"

(6)

and the signal vector X as the vector with components


lj. l(t), X2(t) I... I .N.(t)l
T

(7)

(T denotes the transpose). Also, we define the complex weight for the ith element as

Wi WI

-W

(8)
Fit,. 2. A two-element adaptive array.
(9)
Fig. 3. The LMS feedback loop.
OTHER CHANNELS

and the weight vector as

W=(Wl,W2,...5,WN)T.
The analytic signal at the array output is then given by2
s (t)

SIGNALS FROM

=wT X=-Xlw

(10)

and the real-valued array output signal is


s(t) = Re {s(t) }-

(1 1)
*

Using this notation, the LMS algorithm in (3) may be written

dwi/dt = k ?(t) e (t) = k xl't) [R(t)

s(t)]

INTEGRATOR TRANSFER FUNCTION - 2k


s

(12)

(j 1 S

(14)
(15)

where the asterisk denotes the complex conjugate, and e (t) and R(t) are the analytic signals associated with the real-valued error signal e(t) and reference signal R(t), respectively3. Substituting (10) for s (t) and writing (12) in vector form, one finds that the weight vector w satisfies the differential equation

where CF is the covariance matrix 4) = E(X*XT)


and S is the vector
S = E [X*R(t)]

dwldt

(16)

+ kX*XTw w

kX*R(t) .

(13)

In general, the solution w to (13) consists of the sum of a mean value and a random jitter component.4 In this paper we are interested only in the mean value of w, whose behavior can be approximated by replacing X*XT and X*R(t) in (13) by their expected values. When this is done, (13) may be solved by well-known techniques. The steady-state weight solution is
2The weights Wi are complex random processes that, in general, are not analytic. However, the spectra of Wi are bandlimited around zero frequency and, in a typical communications application, the xi(t) are narrowband processes at a carrier frequency coo. In this case s(t) is analytic as long as the highest frequency in wi is less than the lowest frequency in x i(t). 3Equation (12) is equivalent to (3) in the sense that kx'(t)e (t) in (12) contains the baseband components of 2ke(t) xpi(t) in (3). The second harmonic terms in (3) are assumed too high in frequency to affect the solution for

The major difficulty in using the LMS algorithm in a practical communication system is the problem of obtaining a suitable reference signal. For certain communication systems [6, 7] it is possible to obtain a reference signal by processing the array output. When a reference signal can be obtained, the LMS adaptive array yields automatic beam tracking of the desired signal and good interference protection for the communication system [13]. On the other hand, for many types of communication systems it appears impossible to obtain a reference signal. In this situation, the power-inversion adaptive array may still be useful. The power-inversion adaptive array is based on a modified LMS algorithm requiring no reference signal. If we assume the reference signal R(t) is zero, (16) becomes
S
=0

(17)

wpi.

possibly R(t) ).

4The random component of w, which results in misadjustment [4 1, occurs because of the ranlom noise components of X (and

so

as may

the LMS algorithm will force all the weights to zero, be seen from (14). To prevent this weight shutdown, in a power-inversion array we make two changes
805

COMPTON: THE POWER-INVERSION ADAPTIVE ARRAY: CONCEPT AND PERFORMANCE

xl p

The feedback loop in (12) and Fig. 3 is the loop suggested by Widrow et al. [4], while the feedback loop in (18) and Fig. 4 is the loop originally proposed by Howells and Applebaum [9]. In the radar application of the Howells-Applebaum array, the steering vector wo is used to obtain a main beam of the quiescent pattern in a specified direction (see discussion below). In the powerinversion application described here, the steering vector will be chosen to obtain coverage over a broad sector where desired signals may be received. The loop gain k will be chosen to obtain a maximum SINR at the array output for the range of signal levels expected. These ideas are discussed in the next section.

*LOWPASS

FUNCTION

FILTER TRANSFER 2k =

IV. Array Performance

Fig. 4. Power-inversion array feedback.

in the feedback. First, we offset the ith weight in the array from zero; second, we add a wi term to the left side of (12). Offsetting the ith weight from zero prevents the weights from going to zero when R(t) = 0. The extra wi term added to (12) controls the weight behavior when no signals are being received, as will be seen below. With these changes, the new weight contro equation is
T

Consider an array of two isotropic elements spaced a distance Q apart with adaptive weight control loops as shown in Fig. 4. Suppose two signals are incident on the array, one desired and one interference. In addition, thermal noise is present on each element signal. Thus the element signals are
x.-j() =d(t+i (+n

)1-1, 2

(22)

(dwi/dt) + Wi = w.0
-

k xl*t)s (t)

[8)

jWQio is the ith offset weight (a where two gain constants r and and complex constant) k are now included. The loop-gain constant k controls the steady-state weights in the array and the constant T in conjunction with k, controls the time constants of the array response. The choice of k will be extensively discussed below. The choice of T is determined by the speed of response requirements on the array, and by allowable weight jitter. In this paper we do not discuss the choice of r. The quadrature-feedback loop corresponding to (18) is shown in Fig. 4. When (10) and (15) are substituted in (18) the weigi vector is found to satisfy
where wio = WIio

where dj(t) is the desired signal component, i j(t) is the interference signal component, and n1(t) is a noise component. nj(t) will be assumed to be bandlimited zero mean Gaussian noise with variance a2 (the same variance for each element), and to be statistically independent between elements, i.e.,

E{n (t) n I*(t)} - u26 i


where

(23)

Thus

6ij is the Kronecker delta. The desired signal is assumed to arrive from angle Od (measured from broadside, see Fig. 1). Moreover, d 1 (t) and d 2 (t) are assumed to differ only by an interelement propagation delay.
(24a)
(24b)

d,(t) =d(t)

d2(t)=d(t- Td)

7(dw/dt) + (I + kF) w=wo


vector

(19)

where d (t) is the desired signal waveform and

where I is the identity matrix and wo is the offset weigl


/Wio\

Td = (Q/c) sin Od
(20)
is the spatial time delay between elements for arrival angle Od. (c is the velocity of propagation.) Similarly, the interference arrives from angle Oi, and i I (t)
=

(24c)

wo

\W20

or, as it is usually called, the steering vector. From (19 it is clear that the steady-state weights are given by

i (t)

(25a) (25b)

i 2 (t) = i (t -Ti)

w = [I+ kflf1 wo
which is not zero except perhaps in the limit as k-*oo.
806

(21)

where i1(t) is the interference waveform and


= (Q/c) sin OV T4

(25c)
NOVEMBER 1979

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-15, NO. 6

The desired and interference signals are assumed to be zero mean and statistically independent of each other and the thermal noise. With these assumptions, the covariance matrix (D in (15) is found to be

The inverse of this is easily found


1

+K+Ktd +Kti

-KtdPd I-Kp
(I + k)-' =(1/D)

(Rd(O) + Ri(O) (F = a2i +

R*(Td) + Ri(Ti)
Rd(O) + Ri(m)
/

(32a)
-K1dpd-KKtipi
I + K +Ktd + Kti/

Rd(Td) + Ri(T1)

(26)

where Rd(r) and Rj(r) are the autocorrelation functions of d (t) and i (t), given by
Rd(T) = E [d (t + r)d *(t)]

where D is the determinant of I + k(F

(27;a)

D=(l

K)2 +K+K&d+Kj

IKdpd + KtiPi2.

(32b)

and

Rj(r) = E [i (t + r) i *(t)J

(27 lb)

Since Rd(O) and Rj(O) are the desired signal and interference power on each element, it is convenient to define

Rd(O) = Sd = desired-signal power per element

Rj(O) = Si = interference power per element


and also
Pd= Rd(Td)lRd(O) = Rd(Td)lSd

The steady-state array weights may now be calculated from (21) for any given steering vector w0. Our goal is to protect a desired signal from a stronger interference signal. To see how this may be done, it will be helpful to proceed in steps. We first consider the array performance with noise alone, then with one signal, and finally with both signals incident on the array. In the process, the power-inversion behavior will be seen, and it will become clear how the steering vector w0 and the feedback gain k (or K) should be chosen.
A. Noise Alone

(28a)
(28b)

Pi = Ri(T1)/Ri(0)
Then
(F

Ri(Ti)/S.

First, suppose no signals are present other than thermal noise. Ifd .(t) = 0 and ii(t) = 0 in (22), then td = 0 and ti = 0 and (21) and (32) yield
w=

[1/(1 +K)]wo

may be
Sd

written
PdSd
d
S

[1/(1 t ka2)] w0

(33)

(D = G P2+ + PdSd

+si

,
I

+piSi

dI
+

By also defining the normalized parameters


K = ko2
=

normalized loop gain

(30a)

td = Sdl/a2 = desired signal-to-noise ratio per element (30b)


ti = 5i/a2
=

interference-to-noise ratio per element

(30c)

the matrix I + k(F in (21) may be written

I+k(F=

1+ K + K#d

+ Kti

KtdPd

+ Ktip7

KtdPd +Ktipi

I +K KKd +Kti/

(31)

Thus when dj(t) and i j(t) are absent, the steady-state weight vector equals the steering vector scaled down by the constant (1 + ku2 )-1 Since this constant is the same for every element of the array, the array pattern without signals (which we call the quiescent pattern) is the same as the pattern that would result from the steering vector w0, except for an overall magnitude reduction. Thus the steering vector should be chosen to obtain the quiescent pattern desired from the array. In radar applications w0 is chosen to obtain a quiescent pattern with a main beam at some desired look angle and with suitable sidelobe performance. In communication applications where the angle of arrival of the desired signal will not be known in advance, the quiescent pattern should cover the sector of space from which desired signals may arrive. Such coverage may be obtained by using elements whose patterns cover this sector. Then by choosing the steering vector to have only one nonzero component (so one element is "on" and the other is "off") the quiescent pattern of the array will just be the element pattern of one element. By this choice, we can assure that
807

COMPTON: THE POWER-INVERSION ADAPTIVE ARRAY: CONCEPT AND PERFORMANCE

a desired signal has access to the receiving system from any direction within the sector.5 In this paper we will assume the desired signal may arrive from any angle within -7r < 0d < r. To provide coverage over this sector, we choose isotropic elements with half-wavelength element spacing.6 Furthermore, to be specific, we will assume that element 1 is the one turned on in the quiescent mode. Thus we choose the steering vector

and the array output interference power is

P = 22 E[s i(t) s (t)]


=

(Si/2D2) {(l + K + K4d + Kk)


[l+K+ K&d(l-2 Re {PdP* })

w=\O(A
=

Kti(l -2 p, 12 )]

I
1 + K + K#d +K
\

PKdPd + Ktip,
[(_

(39)
(40)

(34) Finally, the thermal-noise signal at the array output is


s

(l)

n(t) = [n l (t) n 2 (t)

)I

From (21) and (32) the resulting weights are

and the output-noise power is

p =I E[s

(t) s *(t)I

w = (lID) (

(35)

= (a2/2) ( lw

-KtdPd-Ktipi
where D is given in (32b). With these weights the desired signal at the array output is
s j (t) = [d (t) d (t
-

112 + IW2 12)


(41)

=(a2/2D2) [(I + K+KKd + Kf)2 + 1KVd Kd+K ipI12 ]

Td)I]

=-(1/D){f[I + K +Kd + Kti


(KtdPd
+

Now let us apply these results to two cases: 1) desired signal and noise present in the array and 2) desired signal, itterference, and noise present in the array.
B. Desired Si jnal-PI us-Noise

d(t)
(36)

Ktjp,)

d (t

The array output desired-signal power is


d
-

E[s d(t) S d*(t)]

Sd/2D2){(I + K + K4d + Kti)[l + K + K#d(l-2 2pd 12)

If only desired signal and thermal noise are present in the array, we may obtain the desired-signal output power and noise power from (37) and (41) by letting ti =SIG2 = 0. To be specific, let us first assume the desired signal is CW, with frequency wo. (We consider the case of nonzero bandwidth below). We have
d (t) = VSd exp [(o t + 'hd)I

+
+

Kt,( 1-2 Re {Pip*})]


1KVdpd Kt,pj 12}2
+

(42)
on

(37)

where

4'd

is uniformly distributed
-

[-ir, 7T]. Then

Similarly, the interference signal at the array output is


s

Rd(Td) =E[d(t + Td)

d*(t)]

=Sd

expLfcoTd]

i(t) = [ i (t) i (t - T.)]


=(I/D)[(I
-

and thus
w
,(t)

K+

Kd + IK)
-

Pd = Rd(Td)/Sd = exp(Jc0Td ) _ exp(jqd )


=

(43)

(KVdPd Ktjp,) i (t T,)]


+

where qd is the interelement phase shift at frequency cwo. (38) Substituting |Pd 12 I into (37) and (41), we find the output signal-to-noise ratio (SNR) to be
SNR
=

we should choose the element spacing so the have spurious nulls within the sector of interest (i.e., an interference null at one angle within the sector will not create a grating null at another angle within the sector.) 6Half-wavelength spacing permits only one null in visible

5In addition,

array cannot

Pd/Pl

#d {(l + K)2 /[( +K+Kd)2

K2I

1d

space.

From this equation we may obtain an understanding of the effect of the loop gain and the input SNR on the

(44)

808

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-15, NO. 6

NOVEMBER 1979

array performance. Fig. 5 shows a typical plot of output SNR versus K for several values Of Ud and Fig. 7 shows the output SNR versus td for several values of K. From these curves we can see how the loop gain K should be chosen. Since the signal being received by the array is the desired signal, we do not wish to null it. When the input SNR (td) is small, the weights in (35) are not much different than with noise alone and the array does not null the desired signal. This result may be seen in Fig. 5; for example, for td = -10 dB, K has little effect on the output SNR. With a larger input SNR, increasing K causes the array feedback to null the desired signal and the output SNR drops. For example, in Fig. 5 the curve for td = +30 dB shows this behavior. Fig. 6 illustrates the effect of the desired signal on the pattern. It shows the pattern for several values of td for the case K = 0.1. It is seen that for td < 0 dB, the desired signal is not strong enough to cause a pattern null. At td = +10 dB, however, the array has begun to null the desired signal (which is at 00). The null depth increases as td increases; it is -46.1-dB deep when td = +30 dB. (0 dB in Fig. 6 is defined as the array response when the weight vector equals the steering vector.) Fig. 7 shows the output SNR as a function of the input SNR Ud for several values of the loop gain K. It is seen that for a given K, the output SNR at first increases with Ud (in the range where td is too small to affect the weights) and then decreases with td (when the desired signal is nulled by the array). For strong signals a 10-dB increase in td produces a 20-dB increase in the null depth, so the output SNR drops 10 dB. The problem faced by the designer is to prevent the array from nulling the desired signal. The optimum choice of the loop gain K depends on two factors: 1) the required minimum SNR out of the array (which depends on the receiver and the type of modulation used in the communication system), and 2) the dynamic range of signal level that must be accommodated by the array. For example, if the input signal level varies between 0 dB and 20 dB, and an output SNR of 0 dB is required, Fig. 7 shows that K cannot be larger than about 0.07. If a higher value of K is used, the output SNR will drop below 0 dB before td reaches +20 dB. A lower value of K than 0.07 should also not be used, because lower values of K yield poorer interference protection, as will be seen below. Several other effects may also be seen in Fig. 7. First, for td < -10 dB, the output SNR is independent of K. This behavior results because such a weak desired signal has no effect on the array weights. Increasing K simply lowers the quiescent weights, which lowers both the signal and noise powers out of the array by the same factor but does not affect the output SNR. Second, the curve for K = 10 is essentially the same as the curve for K = oo. For K = 10 the output SNR never exceeds -6.5 dB regardless of the input SNR. Third, the larger the dynamic range of the input signal, the smaller K must be to keep the output SNR above a given minimum over the whole range. If the input SNR

LOOP GAIN

K=2kcT2

Fig. 5. Output SNR versus loop gain K (no interference).


Fig. 6. Array pattern-no interference: K = 0.1,
DESIRED SIGNAL

Od = 00.

dB

Fig. 7. Output SNR versus input SNR (no interference).

:O

30 40 50 SNR C (dB)

varies from 0 dB to +60 dB, for example, we must have K < 0.002 if the output SNR must exceed, say, 0 dB. As will be seen in the next section, such a low value of K offers little interference protection. For this reason, the power-inversion technique is most effective when it is necessary to accommodate only a small dynamic range for the desired signal. Moreover, the ideal application of the power-inversion array is to a communication system where the output SNR can be less than 0 dB. One such application is to spread-spectrum systems, which often operate with the signal below the noise, because of the processing gain of the spread spectrum receiver. With the desired signal below noise, the array does not try to null the desired
809

COMPTON: THE POWER-INVERSION ADAPTIVE ARRAY: CONCEPT AND PERFORMANCE

and

Pi = exp(ico T.) A exp(qoi) .


M

(46)

a:

U,

ej is the interelement phase shift at the interference arrival angle. The output SINR, given by
SINR = Pd (P-I +
may now

Pd)

(47)

Fig. 8. Output SINR K 0.01.


=

versus

input SNR

. 0

=0
=

500,

Fig. 9. Output SINR versus input SNR K = 0.1.

td.

6d

6,

500o

SNR Cd (dB)

Fig. 10. Output SINR K = 1.0.


20

versus

input SNR

d.

Od =00

500,

-tO
30 -30

LIMIT

CURVE:-

20

-20

-10

10 SNR

20

30

40

50

6C

(dB)

signal. In this situation the designer has wide latitude in the choice of K, which can then be chosen for good interference suppression.
C. Desired Signal, Interference, and Noise

Now assume two signals are incident on the array, desired and one interference. (However, the array has no way of knowing which is which!) Assume that both signals are CW. For the desired signal, Pd is given in (43). For the interference we let
one

i (t) =

1-

v/Si

exp

V(Wo t + VPi)]
on

be calculated by substituting Pd and pi into (37), (39), and (41). These equations can be used to compute a variety of curves that are helpful for understanding the behavior of the array. Figs. 8, 9, and 10 are typical examples showing the SINR as a function of Ud for K = 0.01, 0.1, and 1.0, respectively. These figures assume Od = Oi = 50 and half-wavelength element spacing. Each figure shows output SINR for several values of ti, the input interference-to-noise ratio. Several things may be seen from these figures. First, for (i = -100 dB the interference is virtually not present, so the SINR curves for this case are identical to the results in Fig. 7 for the appropriate values of K. Second, as the interference power is increased, we find that the SINR at first drops and then rises. For example in Fig. 8, with U = 10 dB, we have SINR = 9.2 dB for ti = -100 dB, then SINR = -3 dB for ti = +17 dB, and finally SINR -+ 12.2 dB Comparing Fig. 8 with Figs. 9 and 10 shows that as ti as K is increased, there is a less significant drop in the SINR for intermediate values of ti. In fact, for K = 1 there is essentially no drop in SINR as ti is increased. In this respect K = 1 represents the best choice of loop gain. However, K = 1 also results in a narrow range of acceptable desiredsignal levels when no interference is present, as was seen in Fig. 7, and also results in higher desired signal attenuation, as will be seen below. Third, for any given ti there is a fimite range of td over which the output SINR is above any given value. For interference powers close to the desired-signal power, this range is least. For interference power substantially less than or greater than the desired signal power, the acceptable range for Ud is wider. Finally, it is interesting to note that for stronger interference signals, the SINR can be substantially better than it would be without the interference. For example, consider Fig. 9. At td 20 dB if ti = -100 dB we have SINR = -2.6 dB, but if ti = +40 dB we have SINR = +17.7 dB. The reason for this behavior is that, without interference, the adaptive array devotes its single null to the desired signal. However, with strong interference the array is forced to use its null on the interference. The desired signal is then not in a null.7 Fig. 11 shows some patterns that illustrate these remarks. It shows the array patterns for K = 0.1 and for -100 dB, +7 dB, +20 dB, and +40 dB. The shift of i
00,

where ipi is uniformly distributed


=

[-rr, 7r]. Then

array, so there is

R1(Ti) EI[i (t + T7) i *(t)] = Si exp(Qwo T.)


810

7This result is achieved because we are using a two-element only one null. With a three-element array, one fmds no such improvement as tj is increased, because the array nulls both signals at once.

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-15, NO. 6

NOVEMBER 1979

DESIRED SIGNAL

ei

(DEGREES)

tp=0dB,tj=30dB.
Fig. 11. Array patterns.

Fig. 13. Output SINR versus interference angle 0. 6 =0,


6 0

td = +20 dB, K = 0.1,

d =0

i= 50 .

6d =o0

Fig. 12. Desired signal attenuation versus input SNR d= 0j 50 . d 60 =500


=0

td =-20 dB, t

F ig. 14. O utput S I N R versus i nterference angle + 20 dB

00
d

60 70

e js

01o

V T z 40

S s 20 dB100d

z, 20
z
0 0

30(-d (20 dB
ei
(DEGREES)

20

Fig. 15. Output SINR versus interference angle 6. 6 = 50,

X10
0

td =-20 dB, t; = +20 dB.


-20

i'

ed(dB) the null from the desired signal to the interference as ti increases may be seen. It is interesting to note that during this change the array response in the desired signal direction is nearly independent of the interference power. We may also compute the attenuation of the desired signal. If the array weight vector equals the steering vector, the desired-signal output power is8 Pod =Sd/2.
From (37) the desired-signal output power with the array operating Pd will be equal to Pod multiplied by factor q
n2=

-30

-10

10

20

30

40

50

60

61 (DEGREES)

{(I +K + Ktd + Kts)[l + K +Ktd(1-2 lpd 12)

Kti(1-2 Re pip* })] + IKpdP + Kt,p, 12 [(1 + K + Ktd + Kt)2 - IKtdPd + Ktip. 12 ] 2

(48)

which we define as the attenuation of the desired signal. When R7 = 1, the array desired-signal output power is the same as it would be if the weights are given by the steering vector wo.

8Sd is the power behind one element of the array. One half of this power is available in each quadrature channel. The weight vector w = w0 turns on only one of the quadrature channels behind element 1.

Fig. 12 shows 71 in decibels versus the input SNR td for several values of K and ti. In all cases, as td increases, the attenuation 7r increases. It is seen that, for a fixed K, the desired signal attenuation is almost independent of ti. The reason for this was noted in Fig. 11; although the pattern changes drastically with ti, the desired-signal response of the array is almost independent of ti. Finally, we note that the array performance depends on the spatial separation of the desired and interfering signals. All of the curves presented so far have been for Od = 00 and Oi = 500. Figs. 13, 14, and 15 show typical plots of the output SINR as a function of Oi for fixed Od . In Figs. 13 and 14 the desired signal arrives from broadside (Od = 00). (In Fig. 13 td =0 dB and fi =30 dB. In Fig. 14 td = -20 dB and ti = + 20 dB.) In Fig.

15,Od =500, td = -20 dB, and ti = +20 dB. In all cases the SINR drops when the interference signal is near the

COMPTON: THE POWER-INVERSION ADAPTIVE ARRAY: CONCEPT AND PERFORMANCE

811

desired signal, because the desired signal falls in the interference null. In general, when interference is too close to the desired signal, the performance of the array will be unsatisfactory. How close the two signals may be depends on the minimum SINR the receiver can accept.9
D. Power Inversion

Let us now assume the noise power in the signals x (t) is very small compared to the signal powers, i.e., (d > 1 and ti > 1. Also, let us assume the array loop gain k is large enough that kSd > 1 and kSi > 1. (Note from (30) that kSd = KUd and kSi = Kti.) Then we may drop the 1 + K terms in (37) and (39) in comparison with the other terms. The output desired signal-to-interference ratio (SIR) is then
l2 SIR = Sd {(d + ti) IV1
+ .( 1-2 Re {p,p* })j
+

the power-inversion array is not as good as with CW signals. A performance degradation occurs with nonzero bandwidth because the antenna pattern is frequency dependent, so its response varies over the signal bandwidth. In particular, the pattern varies much more rapidly with frequency in the nulls than elsewhere, so it is primarily interference bandwidth that affects the performance. Desired signal bandwidth has only a negligible effect on the results. In this section we briefly illustrate the effects of bandwidth on array performance.1" Let us assume the interference signal i (t) is a bandlimited stochastic signal whose power spectral density is constant over a band of width Awi rad/s centered at frequency wO. Then the autocorrelation function of i (t) is

lPd 1)

Ri(T)

Si [sin(Aw1r/2)/(AwT/2)]

exp(jcor)
'

(51)

Substituting X = Ti and noting that


Aw T-/2 = X
+

(AcolAo)(wo T.)

BiOi

tdPd

+ ipi

}/

s
+

-+i)[4d(l2 Re {PdPi
dPd +

ti(l-2

lpi 12)]
(49)

where Bi is the fractional bandwidth,

hipi

Bi Awilwo
=

we find from (28a) that P= [sin


=

When the signals are CW (Pd- exp(pkd), Pi = exp(jqi)), this simplifies to


SIR = (Sd/Si)(t2/02)(l -Re {d PI })/(l - Re Xd Pi*})
As long as Xi :$ kd (mod 2ir), this is1o

(B,O5)/' (B,O,)] exp(qi)

(52)

Similarly, we may assume the desired signal to have a flat spectral density over a fractional bandwidth Bd, so
(50)

= [sin :

SIR = SlSd

(BdOd)I' (Bd4d)] exp(4id).

(53)

which is the reciprocal of the SIR coming into the array. I.e., an interfering signal 20 dB above the desired signal at the array input comes out 20 dB below the desired signal. This property is the reason we refer to this array
as a

power-inversion adaptive array. For finite gain and nonzero noise, the array approximately inverts the power ratio of two signals, as long as the noise is small and the loop gain is large. For example, Fig. 10 (for K = 1) shows that for td = 20 dB and ti = 40 dB (so the input SIR is -20 dB) the output SINR is +18 dB.

E. Bandwidth Effects

The results presented above were all for CW signals. For signals with nonzero bandwidth, the performance of
9This result assumes we take no advantage of electromagnetic polarization differences between the two signals. If the two signals have different polarizations, we may improve the performance at close separations by including cross-polarized elements in the adapative array. loAs 4 approaches tr, higher and higher values of A are required to make the approximation in (49) hold.

For given bandwidths and arrival angles, these Pd and Pi may be substituted into (37), (39), and (41) and the array output SINR may be calculated from (47). Note that the nonzero bandwidth case differs from the CW case only in that IPd < I or jpij < 1. Fig. 16 shows a typical plot of the output SINR for Od = 00, Oi = 500, K = 0.1, i=40 dB, 0. Bd <0.2, and for several values of interference bandwidth in the range 0 < Bi < 0.2. (The curve for Bi = 0 is the same as the = 40 dB curve in Fig. 9.) It may be seen that as Bi -i increases from zero, the output SINR drops for lower values of input SNR. The reason for this behavior may be understood from the array patterns, shown in Fig. 17. We find that increasing the interference bandwidth causes the magnitude of the pattern to drop. This behavior occurs because the null depth varies with frequency over the interference bandwidth. As the bandwidth increases, more and more interference power appears at the array output;
was

originally suggested by D.M. DiCarlo [ 14 1. See also Baird, Martin, Rassweiler, and Zahm [ 15 ] and Rodgers and Compton [ 16 1 for additional results on bandwidth.

11 The method used here to analyze the effect of bandwidth

812

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-15, NO. 6

NOVEMBER 1979

DESI RED

SIGNAL

20
-10

2
z
(

0.05
-30 -20 -10

20 30 10 INPUT SNR (dB)

40

50

60 dB

Fig. 16. Effect of interference bandwidth on output SINR. d =0 i =50 ,K=0.1,t.=40dB,0.B .0.2.

Fig. 17. Effect of interference bandwidth on array patterns.

od =0
6,

i=50' ,d=OdB,K=0.1,Bd=0,ti=40dB.
on output

to compensate for this, the feedback lowers the values of the array weights. The result is to lower the array response in all directions, including the desired-signal direction. On the other hand, we note that all curves in Fig. 16 coalesce for higher values of input SNR. The reason is that with high input SNR, the array weights are already reduced more by the presence of both desired signal and interference than they are by the bandwidth. I.e., with both strong desired signal and strong interference, since the array cannot null both signals, it turns down the weights and reduces the overall pattern magnitude. At high input SNR, this effect is stronger than the pattern reduction due to bandwidth. We also remark that the output SINR is more sensitive to interference bandwidth the higher the input interference power. Fig. 18 shows a plot of output SINR similar to Fig. 16 except that the input interference-to-noise ratio is now +60 dB instead of +40 dB. It is seen that much smaller bandwidths are required to produce a given SINR degradation when ti = 60 dB than when ti = 40 dB. The curves in Figs. 16-18 have been computed for a desired signal bandwidth Bd = 0. However, it is found that Bd has no noticeable effect on these curves over the range 0 S Bd < 0.2, even if the desired signal and interference arrival angles are interchanged (see discussion below about the effect of arrival angle on bandwidth degradation). The reason that Bd has little effect is that the pattern is much less frequency sensitive in the desired signal direction than in the null, as discussed earlier. Finally, we note that it is the product of the interference bandwidth Bi and the interelement phase shift ki that affects pi (see (52) ). For this reason, for interference at broadside (/j = 0), bandwidth has no effect on array performance, but for interference at endfire (oi = 7r), bandwidth has its greatest effect. For applications where interference bandwidth is significant, the designer may wish to minimize bandwidth effects by positioning the array so its broadside direction is close to the interference arrival angle, if that is possible.

Fig. 18. Effect of interference bandwidth =0 =50 ,K =0.1, =60 dB,0


d

<0.05.

SINR.

3c
20

10

U)
-10

-20

-30 -30

-20

-10

10

30 20 INPUT SNR (dB)

40

50

60

V. Conclusions

This paper has discussed the concept of the powerinversion adaptive array. Such an array is based on the use of the Howells-Applebaum feedback to minimize the array output power. When two signals are received by a two-element array with omnidirectional element patterns and the proper weight feedback, the array will improve the ratio of the weak to the strong signal by nulling the strong signal. This technique is useful so long as the strong signal is interference. The problem for the designer wishing to use this idea is to control the situation so the array does not null the desired signal. In situations where the desired-signal power remains near thermal-noise level, suitable performance can be obtained by appropriate choice of K. The wider the dynamic range of the desired signal above thermal noise, however, the more difficult it becomes to determine a loop-gain constant suitable for all situations. In cases where no single value of loop gain is satisfactory, the designer may wish to consider a system where the power-inversion feedback is switched on or off depending on the presence of interference. Also, schemes involving the continuous adjustment or control of the loop gain K are possible.
813

COMPTON: THE POWER-INVERSION ADAPTIVE ARRAY: CONCEPI AND PERFORMANCE

References

[8] P.W. Howells, "Explorations in fixed and adaptive resolu-

[1] R.T. Compton, Jr., "Adaptive Arrays: On power equalization with proportional control," Ohio State Univ. ElectroScience Lab., Dept. Electrical Eng., Rep. 3234-1, Dec. 1971. [21 C.W. Schwegman and R.T. Compton, Jr., "Power inversion in a two-element adaptive array," Ohio State Univ. ElectroScience Lab., Dept. Electrical Eng., Rep. 3433-3, Dec. 1972, ASTIA document AD 758690. [3] C.L. Zahm, "Application of adaptive arrays to suppress strong jammers in the presence of weak signals," IEEE Trans. Aerosp. Electron. Syst., vol. AES-9, p. 260, Mar. 1973. [4] B. Widrow, P.E. Mantey, L.J. Griffiths, and B.B. Goode, "Adaptive antenna systems," Proc. IEEE, vol. 55, p. 2143, Dec. 1967. [51 R.L. Riegler and R.T. Compton, Jr., "An adaptive array for interference rejection," Proc. IEEE, vol. 61, p. 748, June 1973. [6] R.T. Compton, Jr., R.J. Huff, W.G. Swarner, and A.A. Ksienski, "Adaptive arrays for communication systems: An overview of research at the Ohio State University," IEEE Trans. Antennas Propagat., vol. AP-24, p. 599, Sept. 1976. [7] R.T. Compton, Jr., "An adaptive array in a spread spectrum communication system," Proc. IEEE, vol. 66, p. 289, Mar. 1978.

[9]
[101
[11]
[12]

[13]
[14]

[15]

[16]

tion at GE and SURC," IEEE Trans. Antennas Propagat., vol. AP-24, p. 575, Sept. 1976. S.P. Applebaum, "Adaptive arrays," IEEE Trans. Antennas Propagat., vol. AP-24, p. 585, Sept. 1976. J. Dugundji, "Envelopes and preenvelopes of real waveforms," IRE Trans. Inform. Theory, vol. IT-4, p. 53, Mar. 1958. E. Bedrosian, "The analytic signal representation of modulated waveforms," Proc. IRE, vol. 50, p. 2071, Oct. 1962. T.G. Kincaid, "The complex representation of signals," General Electric Co., Heavy Military Electronics Dept., Syracuse, N.Y. 13201, Tech. Rep. R67EMH5, Oct. 1966. R.T. Compton, Jr., "An experimental four-element adaptive array," IEEE Trans. Antennas Propagat., vol. AP-24, p. 697, Sept. 1976. D.M. DiCarlo, "The effect of interference bandwidth on the performance of quadrature weighted adaptive arrays," private communication. C.A. Baird, G.P. Martin, G.G. Rassweiler, and C.L. Zahm, "Adaptive processing for antenna arrays," Radiation Systems Division, Harris Intertype Corp., Melbourne, Fla. 32901, Final Rep., June 1972. W.E. Rodgers and R.T. Compton, Jr., "Adaptive array bandwidth with tapped delay-line processing," IEEE Trans. Aerosp. Electron. Syst., vol. AES-15, p. 21, Jan. 1979.

R.T. Compton, Jr., was born in St. Louis, Missouri, on July 26, 1935. He received the S.B. degree from M.I.T. in 1958 and the M.Sc. and Ph.D. degrees from The Ohio State University, Columbus, in 1961 and 1964, respectively, all in electrical engineering. He is a Professor of electrical engineering at The Ohio State University. From 1965 to 1967 he was an Assistant Professor of Engineering at Case Institute of Technology, Cleveland, Ohio, and from 1967 to 1968 he was a National Science Foundation Postdoctoral Fellow at the Technische Hochschule, Munich, Germany. Dr. Compton is a member of Sigma Xi and Pi Mu Epsilon.
814
IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-15, NO. 6 NOVEMBER 1979

You might also like