Date post: | 08-Dec-2023 |
Category: |
Documents |
Upload: | independent |
View: | 0 times |
Download: | 0 times |
1
The Performance of a Null-Steering
Beamformer in Correlated Rayleigh Fading
Cornelius van Rensburg , Benjamin Friedlander
C van Rensburg is with Samsung Telecommunications America. E-mail: [email protected]
B. Friedlander is with The University of California Santa Cruz. E-mail: [email protected]. The work of
B. Friedlander was funded by the National Science Foundation under Information Technology Research Grant
CCR-0112508, and by the Office of Naval Research, under contract N00014-01-1-0075.
October 19, 2003 DRAFT
2
Abstract
We consider the performance of a null-steering adaptive beamformer in a distributed Rayleigh fading
environment. To determine the beamformer coefficients, the signal and interference array response vectors
must be estimated. An estimator which makes use of the knowledge of the spatial distribution of the signal
is proposed, and compared with other estimators which do not use this information. As the angular spread
of the signal increases, the value of this information is reduced, until for the case of isotropic scattering
the estimate which uses the spatial distribution is equivalent to the estimate which does not. Performance
results are presented that compare the performance of the different estimators. We derive closed form
expressions for the SINRO and probability of error for the case where the training data are orthogonal.
Keywords
EDICS:1-ACOM,3-MCHA.
Array Processing, Communication, Fading, Estimation and Detection.
October 19, 2003 DRAFT
3
I. Introduction
The need to separate co-channel signals by an antenna array arises in many situations,
and in cellular communications in particular. Separating signals from different users by
their spatial signatures can increase significantly system capacity beyond what can be
achieved by temporal processing alone.
Extensive work has been done on techniques for separating co-channel signals as de-
scribed in [1] and [2]. The overwhelming majority of this work assumes that the signals
are point sources (i.e., having zero angular spread), and that the propagation from trans-
mitter to receiver occurs along one path or a few distinct paths (multi path). A more
practical model used in the cellular communications environment represents the propaga-
tion as a continuum of different paths, where the signal arriving at the receiver antenna can
be characterized by a continuous spatial distribution of energy (sometimes called “angular
spread”) around the nominal direction of arrival.
Relatively little work has been done on co-channel signal separation for the case of
distributed signal sources. Previously [3] developed a performance bound for specifically
the isotropic scattering case. More recently, [4] proposed a structured estimate of the array
response, which performed better in a distributed fading environment to a system that used
an unstructured estimate. [5] looked at the effect that temporal and spatial correlation
has on the DOA estimation, and [6] investigated the performance of the general class of
semi-blind subspace-based estimators, like those that we are using here.
In this paper we study the effect of angular spread on the ability of an array to separate
co-channel signals. In particular we use a framework similar to that presented in [7] and
[8], where an isotropic scattering environment was assumed (essentially an angular spread
of 360 degrees ) which is characteristics of an indoor or pico cell environment. Here we
generalize the work of [7] for the case where the angular spread can take on any value
between zero (point source) and 360 degrees (isotropic scattering). Small to moderate
angular spread is characteristic of outdoor urban and rural communication where the
mobile is typically far from the base station and the fading happens in a relatively small
region around the mobile station.
The angular power distribution of the signal source induces the spatial covariance matrix
October 19, 2003 DRAFT
4
of the received array response. In the case of a point source this covariance matrix is of
rank one, while for isotropic scattering it is the scaled identity matrix. In general the rank
of the covariance matrix increases with the angular spread.
In order to perform co-channel signal separation it is necessary to estimate the array
response vectors of the different users. As in [7] we assume that a training sequence is
available to do the estimation. In most TDMA cellular communication systems such as
GSM and IS-136, there are synchronization bursts at the start of every timeslot, which can
be used for this purpose. However, in addition to the training data it is possible to make
use of the statistical information embodied in the spatial covariance matrix associated
with each user. We consider two situations: (i) The case where the covariance matrix is
known a-priori, and (ii) The case where the covariance matrix is estimated from the data
itself. The first is useful as a reference point to show the best that can be done by using
this additional statistical information. The second is the more practical situation in which
the covariance matrix is not known a-priori. As is expected, using the covariance matrix
in addition to the training data can improve signal estimation performance.
We will consider four estimates of the array response, namely the perfect estimate A,
the estimate with unknown covariance R (UR), A, the estimate with known R (KR), AR,
and the estimate using an estimated covariance, AR, (ER).
II. Problem Statement
A. Notation
We use lower case boldface letters to denote vectors and upper case boldface letters to
denote matrices. In addition we use the following conventions throughout the paper,
X⊗Y Kronecker product
Av The vec() operator, which is the concatenation
of the columns of A
()′ The conjugate transpose
()T Transpose
()∗ Conjugate
<(A) The real part of A
October 19, 2003 DRAFT
5
=(A) The imaginary part of A
E[] Expected value
I Identity matrix
S† Moore-Penrose pseudo inverse
∼ Nc A complex Gaussian distribution
B. The Signal Model
Consider an M-element antenna array with an array response ak, corresponding to a
signal sk(tn) at time tn. The total of K transmitted signals are corrupted by an AWGN
term u(tn), so we can write the received signal (the M-vector) as
x(tn) =K∑
k=1
aksk(tn) + u(tn) (1)
where u(tn) ∼ Nc(0, ηI). The k’th signal has a power of pk = |sk(tn)|2. Defining the total
array response as
A = [a1 . . . aK ], (2)
and the total signal at time tn as
s(tn) = [s1(tn) . . . sk(tn)]T , (3)
we can write
x(tn) = As(tn) + u(tn). (4)
If we now collect Nd data points and define the matrices
X = [x(t1) . . .x(tNd)] (5)
S = [s(t1) . . . s(tNd)] (6)
U = [u(t1) . . .u(tNd)], (7)
then we can finally write the model as
X = AS + U. (8)
Using the estimate for A, namely A, we will reconstruct the signal at the receiver using
a null-steering beamformer that would attempt to null-out the interferers completely by
October 19, 2003 DRAFT
6
implementing this equation:
s(tn) = (A′A)−1A′x(tn), (9)
which is the least squares (LS) minimizer for the cost function ||x(tn)− As(tn)||2F .
C. Channel Model
In a faded multi path environment the array response vector is
ak =∫ π
−πψk(θ)w(θ)dθ, (10)
where w(θ) is the steering vector for the particular array and ψk(θ) is the complex fading
from the signal sk. Note that ψk is independent of ψl for l 6= k, and ψk(θ1) is independent
of ψk(θ2) for θ1 6= θ2, such that
E[ψk(θ1)ψ∗l (θ2)] = Ψ(θ1 − θk)δ(θ1 − θ2)δk−l (11)
E[ψk(θ1)ψl(θ2)] = 0. (12)
The covariance of the array response can be expressed as
Rk = E[aka′k] =
∫ π
−πΨk(θ)w(θ)w′(θ)dθ, (13)
where
Ψk(θ) = E[|ψk(θ)|2], (14)
is the spatial energy distribution of the k’th signal. Therefore, the columns of A will be
random with
ak ∼ Nc(0,Rk), ∀ k = 1, . . . , K (15)
but the fading in the different columns are independent, thus
E[aka′i] = 0 ∀ k 6= i. (16)
The problem now is how to estimate the channel matrix A, using the covariance.
III. Channel Estimation
All the estimators considered here assume that there is a known training sequence St,
of length N available to the receiver to estimate the channel.
October 19, 2003 DRAFT
7
A. The Estimate if R is Known (AR)
In this section, we develop an array response estimate AR in which we use the a priori
knowledge of the covariances of the columns of A,
ak = R12k yk, (17)
yk ∼ Nc(0, ILk), (18)
E[yky′i] = 0, ∀ k 6= i (19)
where Lk is the rank of Rk. Note that when Rk is low rank, some of the eigenvalues will
be zero. Equation (2) can be vectorized and expressed as
Av =
R121 0 0
0. . .
...
0 0 R12K
y1
...
yK
= R12y. (20)
We define the matrix square root as the M × Lk matrix
R12k = ΓkΛ
12k (21)
which means that Rk = R12k R′ 12
k , Γk is the matrix of eigenvectors of Rk and Λ12k is the
diagonal matrix of the square roots of the eigenvalues. This property will be used in the
performance analysis section. Vectorizing (8)
Xv = (STt ⊗ IM)Av + Uv = QR
12y + Uv, (22)
where Q = STt ⊗ IM, is a reformatted version of the training symbols. The LS estimator
can be written as
AvR = R
12 (QR
12 )†Xv = R
12y + R
12 (QR
12 )†Uv (23)
or alternatively
ak = R12k [(QR
12 )†Xv]k = R
12k yk + R
12k [(QR
12 )†Uv]k, (24)
where []k is the k’th M ×Lk block sub matrix. This estimator is the LS minimizer for the
cost function
||Xv −QR12y||2F .
October 19, 2003 DRAFT
8
TABLE I
The Estimators AR, AR, A.
Case Estimator
Known R R12k [(QR
12 )†Xv]k
Estimated R R12k [(QR
12 )†Xv]k
Unknown R [(Q)†Xv]k
B. Estimate using no a-priori Knowledge (A)
This case is just a special case of the estimator presented in Section III.A, where the
Rk = I, thus
Av = (Q)†Xv = y + (Q)†Uv. (25)
C. The Estimate where R is Estimated (AR)
In this section, the covariance is assumed to be unknown and first needs to be estimated.
To do this we first make Na estimates of the channel using the method presented in Section
III.B, which assumes no-prior knowledge of the covariance. The estimated covariance is
then the sample covariance of these estimates :
Rk =1
Na
Na∑
n=1
ak(n)a′k(n), (26)
Once we have the estimate Rk, and thus also R12k , we define the ”Estimated R” estimator,
by substituting the actual R in (23) for the estimated R:
AvR
= R12 (QR
12 )†Xv = R
12y + R
12 (QR
12 )†Uv. (27)
D. Comparison of AR, AR and A estimators
The different estimators are summarized in Table I. In the case of point sources, the
Known R estimator (A) is defined by Γk up to an unknown complex scaling factor. The
training symbols are used only to determine this factor. In the case of isotropic scattering
the three estimators are identical because R12k = I.
October 19, 2003 DRAFT
9
IV. The SINRO of the Different Receivers
Next, we evaluate the output signal to interference plus noise ratio (SINRO), of receivers
using the different estimators. Using the partition, A = [D,v], where D represents the
first K − 1 columns of A and v the last column, we can define the following matrices :
T = I−D(D′D)−1D′ (28)
D = (D′D)−1D′. (29)
Here T is the projection matrix on the null space of D, and D is the left pseudo inverse
of D. It can be shown that
(A′A)−1A′ =1
v′Tv
D[v′TvI− vv′T]
v′T
(30)
and therefore the K’th detected signal can be expressed as
sK(tn) =v′Tx(tn)
v′Tv=
v′T[As(tn) + u(tn)]
v′Tv
=v′Tv
v′TvsK(tn) +
K−1∑
k=1
v′Tak
v′Tvsk(tn) +
v′Tu(tn)
v′Tv
= αsK(tn) +K−1∑
k=1
βksk(tn) + µ′u(tn). (31)
Here any hatted symbol represents an estimated value, α represents how well we matched
the received signal and βk is the component of the interferer that was not canceled. We
can now define the signal to noise ratio at the output of the beamformer (SNRO) as
SNRO =pKE[||α||2]E[||µ||2]η , (32)
where pK is the K’th signal power. The signal to interference ratio at the output (SIRO)
is
SIRO =pKE[||α||2]
∑K−1k=1 pkE[||βk||2]
, (33)
and the signal to interference and noise ratio at the output SINRO is
SINRO = [SIRO−1 + SNRO−1]−1. (34)
October 19, 2003 DRAFT
10
A. The Perfect Channel Estimate
When A is known perfectly there is no residual interference power and β = β = 0,
α = α = 1 in (31). We can therefore write the SNRO in (32) as
SNRO = ρKv′Tv (35)
where ρK = pK/η is the input SNR of the K’th signal.
B. The Estimate when R is known (AR)
An expression for the SINRO of this estimator is derived using a perturbation analysis.
We denote <[Apq] as Apq, =[Apq] as Apq. Since βk = 0, a first order Taylor series expansion
of βk is:
βk 'M∑
p=1
K∑
q=1
∂βk
∂Apq
δApq +M∑
p=1
K∑
q=1
∂βk
∂Apq
δApq. (36)
Using lengthy but straightforward algebraic manipulations we obtain the following expres-
sions:
∂βk
∂Apq
= −δqkv′Tep
v′Tv(37)
∂βk
∂Apq
= −jδqkv′Tep
v′Tv(38)
where ep represents an M vector of zeroes, except for a one at the p-th position. If we now
define:
h(1)k =
∂βk
∂Av= −ek ⊗ (Tv)∗
v′Tv, (39)
h(2)k =
∂βk
∂Av= −jek ⊗ (Tv)∗
v′Tv, (40)
hk =
h(1)k
h(2)k
= −
ek ⊗ (Tv)∗v′Tv
jek ⊗ (Tv)∗v′Tv
, (41)
δa =
δAv
δAv
, (42)
then we can rewrite (36) as
βk ' hTk δa, (43)
October 19, 2003 DRAFT
11
and therefore
E[||βk||2] ' hTk var(δa)h∗k. (44)
To calculate the var(δa), we use (23) to get
E[δaδa′] =η
2R
12 (R′ 12Q′QR
12 )−1R′ 12 , (45)
and this can be simplified if the training signals are orthogonal, to
E[δaδa′] =η
2N
1p1
Γ1Γ′1 0 0
0. . .
...
0 0 1pK
ΓKΓ′K
. (46)
Note that since the fading is correlated, Γk is an M × Lk matrix, where Lk is the rank of
Rk. This gets substituted into (44):
E[||βk||2] ' η
2(v′Tv)2
[eT
k ⊗ (v′T) jeTk ⊗ (v′T)
]
×<[E[δaδa′]] −=[E[δaδa′]]
=[E[δaδa′]] <[E[δaδa′]]
ek ⊗ (Tv)
−jek ⊗ (Tv)
' η
pkN(v′Tv)2v′T(ΓkΓ
′k)Tv. (47)
A similar perturbation analysis done on α, which we will not repeat here, will show that
a good approximation when the SNR is high is α ≈ 1. Next, insert the result above into
(33) to get
SIROKR ' NρK(v′Tv)2
v′T(∑K−1
k=1 ΓkΓ′k)Tv
, (48)
SINROKR ' NρK(v′Tv)2
v′T(∑K−1
k=1 ΓkΓ′k + NIM)Tv. (49)
Note that for the case of a point interferer, SIROKR = ∞, irrespective of the position of
the interferer. This is because E[||βk||2] = 0 when T projects the eigenvectors of D onto
their own null space, (TΓkΓ′k = 0). In this case the SINROKR = SNRO.
C. The case when R is Estimated (AR)
The SINRO for this estimator can be found by rewriting (47) as
E[||βk||2] ' η
Npk(v′Tv)2v′TE[ΓkΓ
′k]Tv (50)
October 19, 2003 DRAFT
12
but note that
E[ΓkΓ′k] =
Lk∑
l=1
E[glg′l], (51)
which was shown in [9] to be
E[glg′l] =
λl
Na
Lk∑
i=1,i6=l
λi
(λl − λi)2gig
′i + glg
′l. (52)
Therefore
E[ΓkΓ′k] =
Lk∑
l=1
λl
Na
Lk∑
i=1,i 6=l
λi
(λl − λi)2gig
′i + glg
′l
= ΓkΛkΓ′k, (53)
where Λk is a diagonal matrix with elements
[Λ]ii = 1 +λi
Na
∑
j 6=i
λj
(λi − λj)2. (54)
Substituting (53) into (50) leads to
E[||βk||2] ' η
Npk(v′Tv)2v′T[ΓkΛkΓ
′k]Tv. (55)
Finally the SIRO and SINRO can be written as
SIROER ' NρK(v′Tv)2
v′T(∑K−1
k=1 ΓkΛkΓ′k)Tv, (56)
SINROER ' NρK(v′Tv)2
v′T(∑K−1
k=1 ΓkΛkΓ′k + NIM)Tv. (57)
It is clear from (57) that as Na increases, Λ → I and SINROER → SINROKR, as expected.
D. The Estimate with no a-priori Knowledge (A)
This estimator is a special case of the Estimator where R is Known with ΓkΓ′k = I in
(49). Therefore,
SINROUR = ρKv′TvN
N + K − 1, (58)
which agrees with the results in [7]. The results of this section is summarized in Table II.
October 19, 2003 DRAFT
13
TABLE II
The SINRO for the four different estimators.
Case SINRO
1. Perfect A ρKv′Tv
2. Known R NρK(v′Tv)2
v′T(∑K−1
k=1ΓkΓ′k+NIM )Tv
3. Estimated R NρK(v′Tv)2
v′T(∑K−1
k=1ΓkΛkΓ′k+NIM )Tv
4. Unknown R ρKv′Tv NK−1+N
V. The Average SINRO
The expressions in Table II represent nonlinear combinations of multiple random vari-
ables. Evaluating the average SINRO analytically is difficult in general. Here we consider
the special cases of a single interferer, and multiple (but well separated) interferers. The
expressions for the SINRO for the cases where either A is known perfectly, or not known
at all, are identical (up to a constant). These two cases represent an upper and lower
bound and are of special interest.
A. The Average SINRO with a single interferer
We study how the SINRO changes with the position of the interferer and the angular
spread. In the case of a single interferer, D is a vector, and the average SINRO can be
written as
E[SINRO] = E[v′Tv] = E[|v|2]− E[v′DD′vD′D
]
= E[|v|2]− trace(E[DD′
D′D]E[vv′])
= trace(ΛK)− trace(RdΓ′1ΓKΛKΓ′KΓ1)
(59)
where
Rd = E[Λ
1/21 y1y
′1Λ
1/21
y′1Λ1y1
]
October 19, 2003 DRAFT
14
=
E[rd1] 0 0
0. . . 0
0 0 E[rdL]
(60)
and where Γ′KΓ1 represents the matrix of cosines of the principle angles between the
subspaces spanned by v and D. We are interested in the mean values of the random
variables rd1 . . . rdL. The elements on the off diagonal are independent and have zero
mean. In Appendix A we show that rdi has a probability density function (pdf) of
frdi(y) =
L∑
l=1,l 6=i
αlλiλ2l
[y(λl − λi) + λi]2, (61)
αi =1
αi
L∏
l=1,l 6=i
αi
αi − αl
, (62)
where 0 < y < 1. The mean is then
E[rdi] =L∑
l=1,l 6=i
λiαlλ2l
(λl − λi)2
[λi
λl
− 1 + ln
(λl
λi
)]. (63)
The average SINRO is therefore a function of the squared cosines of the principle angles,
weighted by the diagonal values of ΛK and Rd. For example, if the subspaces have rank
one, there is only one principle angle, and then
E[SINRO] = 1− cos2(θ) = sin2(θ), (64)
which is a measure of the distance between the two subspaces. If each subspace is of
dimension 2, with
|g′v1gd1|2 = cos2(θ11)
|g′v1gd2|2 = cos2(θ12)
|g′v2gd1|2 = cos2(θ21) (65)
|g′v2gd2|2 = cos2(θ22)
then
E[SINRO] = trace(ΛK)− rd1λ1K cos2(θ11)− rd1λ2K cos2(θ12)
− rd2λ1K cos2(θ21)− rd2λ2K cos2(θ22). (66)
October 19, 2003 DRAFT
15
The average SINRO in (66) can be viewed as the total received power for the signal of
interest minus the proportion of the interfering power that coexist in the subspace of the
signal of interest. In Fig. 1 we plot the normalized SINRO (E[v′Tvv′v ]) for a 10 element
uniformly spaced linear array (ULA) as the position and spread is varied. The array
beamwidth here is 10 degrees.
B. The Average SINRO for Multiple Interferers
For the case where the multiple interferers are well separated, D′D is a diagonal matrix,
and
E[SINRO] = trace(ΛK)−K−1∑
k=1
trace(RKΓ′KΓkRdkΓ′kΓK).
(67)
Here, the interpretation of the average SINRO in (67) generalizes (59) to the total received
power for the signal of interest minus the proportion of all the interfering powers that
coexist in the subspace of the signal of interest.
C. The BER Performance in a fading environment
In general it is difficult to get an expression for the error probability (Pe), since the
SINRO consists of products of multiple random variables. However, if the interferer is not
too close to the signal of interest, we can approximate the distribution of the SINRO to a
chi-square distribution, and then use the Pe expressions available in the literature. Here
we use the formulas derived by [10](in Appendix 7A), where we will use the derivation for
the Binary PSK example that uses a pilot signal estimate. These formulas are repeated
here,
γt = γtraffic + γpilot (68)
r =γtraffic
γpilot
(69)
µperfect = =
√γtraffic
γtraffic + 1(70)
µEstimate =
√rv
(r + 1)
√(1γt
+ rr+1
) (1γt
+ vr+1
) (71)
October 19, 2003 DRAFT
16
Pe(µ) =1
2
1− µ
L−1∑
k=0
2k
k
(1− µ2
4
)k . (72)
Here, v represents the number of training symbols, r represents the ratio of the pilot power
versus the power in the traffic channel, γtraffic and γpilot are the average SNR’s for the
traffic and pilot channels respectively. The reader might be surprised to see references
to a pilot and traffic channel when we are describing a TDMA system. The reason is
that we can use these expressions (which were probably derived for a CDMA system)
with slight modification for our problem. So, rather interpret γtraffic as the average SNR
during the data burst, and γpilot as the average SNR during channel estimation. Note that
the Pe is a function of the parameter µ, so the case of the perfect estimate would use the
µperfect and the case of the KR Estimator would use µEstimate in calculating Pe. However,
these formulas were derived for the case where the fading is independent and identically
distributed (iid). Therefore, we have to redefine the variables r, v, γpilot, γperfect since the
fading is correlated here. In this paper we will simply approximate the eigenvalues as either
ones or zeros, and then transform the problem into one that approximates one where the
variables are iid. Firstly, the Pe for the perfect estimate depends only on the average SNR
for the traffic channel which we define as
γtraffic = ρKE[v′Tv], (73)
and L which is chosen to minimize the mean square error between the approximated
eigenvalues and the true eigenvalues
MSE = ||[λ1, ..λM ]− [0, ..0, 1, ..1]/L||2F . (74)
The optimum L to choose is the one that would minimize the Pe, but since the Pe in (72)
is not too accurate we use the method above. For the case of the KR estimate, we choose
L similar to the perfect estimate. Note that here we project each component that needs
to be estimated (yl) on an M dimensional orthogonal subspace before estimation, which
means that there are effectively NM independent training symbols per estimate, but at
the same time, the average SNR in the pilot channel reduces by M , thus
γtraffic = E[SINROKR] (75)
October 19, 2003 DRAFT
17
γpilot = E[SINROKR]/M (76)
v = NM. (77)
To be accurate, these formulas have to be re-derived for the case of correlated fading. Note
that the Interference component in the SINRO is small in most cases that we describe here
since we are nulling it out, so we can simply approximate
γtraffic = ρKE[v′Tv] (78)
γpilot = ρKE[v′Tv]/M, (79)
for which we have closed form expressions. When we study the structure of µEstimate, we
see that the relationship between SINRO and Pe is not straightforward and that is why
we see a much larger performance gain by the new estimators as compared to what the
SINRO alone would suggest. The reason that we can use these expressions directly for
the KR estimator is because it projects the estimators on orthogonal subspaces, and the
variables can then be approximated as a uncorrelated and iid. However, this cannot be
done for the UR estimator and thus we do not have an expression for the BER in this
case.
VI. Numerical Simulations
In this section, we simulate equal powered BPSK signals, with K = 2, that is one
signal of interest and one interferer. There are M = 6 antenna elements in a ULA with a
beamwidth of approximately 20 degrees. The data burst consists of Nd = 150 symbols. In
Fig.2 we present SINRO results from Monte Carlo simulations and compare them with the
analytic values in Table II. Here we simulated a signal of interest which was at broadside
and an interferer which was moved from -50 degrees to -5 degrees, both were spread over
4 degrees. The channel was estimated using N = 2 training symbols while the input
SNR was 10dB. In Fig. 3, we show BER simulation results where both signals were again
spread over 4 degrees and L = 1 in both cases. Here the channel was estimated with
N = 4 training symbols, while the covariance was estimated over Na = 20 timeslots. he
eigenvalues are
[0, 0, 0, 0, 0.0058, 0.9942].
October 19, 2003 DRAFT
18
Note how close the Known R and Estimated R estimators are to the case where the
channel is known perfectly. In Fig. 4, the signals are both spread over 30 degrees, and
here L = 2 in both cases. The eigenvalues in this case is
[0, 0, 0.0025, 0.0591, 0.3480, 0.5903].
There is still a significant performance improvement here compared to the unknown R
estimator, but less than when the signal has a small angular spread. Note how the analytic
Pe differs from the Monte Carlo results in the high SNR cases. This is because the effect
of the approximation of the eigenvalues as either ones or zeros becomes significant. When
the signals are spread over a wide angle, e.g. over 100 degrees, the 3 estimators will be
identical, this should be clear from Table I.
VII. Conclusions
In this paper we investigated the ability of an array to separate co-channel signals in
a fading environment as a function of angular spread. We compared the performance of
several estimators which use the spatial covariance matrix of the array response (either
known a-priori or estimated from the data) to that of an estimator which uses only training
data. As expected, using the covariance information can improve the system performance.
The performance improvement increases as the number of array elements increases, the
number of interferes increases, and the number of training symbols decreases.
Appendices
I. Weighted Beta Function
We are interested in the distribution and the mean value of the random variable defined
as
Y =X1
X1 + X2 + . . . + XN
(80)
where
Xn ∼ αn
2χ2
2 (81)
with a pdf of
fXn(x) =1
αn
e−x
αn , (82)
October 19, 2003 DRAFT
19
for all unique αn’s. To find fY (y) we proceed in 2 steps. First we assume N = 2, and
generalize it later. Doing a transformation of random variables we define
Y =X1
X1 + X2
(83)
Yt = X1 + X2, (84)
and then x1 = yyt, and x2 = yt − yyt, with 0 < y < 1, and 0 < yt < ∞. The Jacobian is
J =
∣∣∣∣∣∣∣
yt y
−yt 1− y
∣∣∣∣∣∣∣= yt. (85)
The joint pdf is
fX1X2(x1, x2) =1
α1α2
e−(
x1α1
+x2α2
)(86)
fY Yt(y, yt) =1
α1α2
yte−(
yytα1
+yt−yyt
α2)
(87)
fY (y) =1
α1α2
∫ ∞
0yte
−yt(y
α1+ 1−y
α2)dyt
=α1α2
[y(α2 − α1) + α1]2. (88)
The expected value of Y is then
E[Y ] =∫ 1
0
α1α2y1
[y1(α2 − α1) + α1]2dy
=α1α2
(α2 − α1)2
[α1
α2
− 1 + ln(
α2
α1
)]. (89)
Now, let us consider the case of N > 2. Here we will keep the same definitions as above,
except that we will introduce
Xsum =N∑
n=2
Xn. (90)
The pdf of Xsum is the convolution of the respective pdf’s, namely
fXsum(x) = fX2(x) ∗ fX3(x) ∗ . . . ∗ fXN(x)
=N∑
n=2
αne− x
αn , (91)
where
αi =1
αi
N∏
n=2
αi
αi − αn
, ∀ i 6= n. (92)
October 19, 2003 DRAFT
20
The derivation from (86) to (89) is now repeated by replacing fX2(x) with fXsum(x) as
defined in (91) instead. This leads to the pdf and the Expected value
fY (y) =N∑
n=2
αnα1α2n
[y(αn − α1) + α1]2(93)
E[Y ] =N∑
n=2
α1αnα2n
(αn − α1)2
[α1
αn
− 1 + ln(
αn
α1
)]. (94)
References
[1] A. Paulraj and B. Papadias. Space-time processing for wireless communications. IEEE Signal Processing
Magazine, November 1997.
[2] L. Godara. Applications of antenna arrays to mobile communications, part 1: Performance improvement,
feasibility, and system considerations. Proceedings of the IEEE, July 1997.
[3] J. Winters, J. Salz, and R. Gitlin. The impact of antenna diversity on the capacity of wireless communication
systems. IEEE Transactions on Communications, February 1994.
[4] G. Klang and B. Ottersten. A structured approach to channel estimation and interference rejection in multi-
channel systems. In Proceedings of the RVK, June 1999.
[5] R. Raich, J. Goldberg, and H. Messer. Localization of a distributed source which is partially coherent
- modeling and cramer rao bounds. International Conference on Acoustics, Speech and Signal Processing
(ICASSP), Phoenix, March 1999.
[6] V. Buchoux, O. Cappe’, E. Moulines, and A. Gorokhov. On the performance of semiblind subspace-based
channel estimation. IEEE Transactions on Signal Processing, June 2000.
[7] A. J. Weiss and B. Friedlander. Fading effects on antenna arrays in cellular communications. IEEE Transac-
tions on Signal Processing, May 1997.
[8] C. van Rensburg and B. Friedlander. Performance of antenna arrays in an urban multi-path environment.
Conference Record of GLOBECOM2000, San Francisco, CA, USA, November 2000.
[9] D. R. Brillinger. Time Series Data Analysis and Theory. San Francisco, CA: Holden-Day, 1981.
[10] J. G. Proakis. Digital Communications. McGraw-Hill, 1983.
October 19, 2003 DRAFT
21
−60 −40 −20 0 20 40 60−0.2
0
0.2
0.4
0.6
0.8
1
Position of Interferer (degrees)
Nor
mal
ized
SIN
RO
= vH
Tv/
vHv
Normalized SINRO as the position of the interferer is moved.
Fig. 1. The normalized SINRO as the position of the interferer is moved from -50 to +50 degrees off
broadside for an array with a 10 degree beamwidth. The signal of interest is at 0 degrees. The
narrowest curve represents a 2 degree spread of the signal and interferer, each wider curve represents
an extra 4 degree spread until the widest curve represents a 40 degree spread.
−50 −45 −40 −35 −30 −25 −20 −15 −10 −58
9
10
11
12
13
14
15
16
17
18
Position of Interferer (degrees)
SIN
RO
(dB
)
The SINROs for a mobile at broadside with an interferer at different positions
Perfect A: AnalysisPerfect A: MCKnown R: AnalysisKnown R: MCUnknown R: AnalysisUnknown R: MC
Fig. 2. SINRO Curves for 3 estimators using a 6 element ULA with a 20 degree beamwidth. The signal
of interest is at broadside and the interferer position is varied from -50 degrees to -5 degrees, both
signals are spread over 4 degrees. In the plots the lines represent the analytic curves generated from
Table II. The symbols represent the results of Monte Carlo experiments. The input SNR is 10dB. In
this plot the Analytic curves for the Perfect A and the Known R estimates are indistinguishable
October 19, 2003 DRAFT
22
0 1 2 3 4 5 6 7 8 9 1010
−2
10−1
100
Input SNR
BE
R
BER curves when the signals are spread by 3 degrees
Perfect AUnknown RKnown REstimated RPe Perfect APe Known R
Fig. 3. BER Curves for the 4 estimators. There is one interferer, both signals are spread over 4 degrees,
and the interferer is 1 beamwidth away, so that there is no overlap between the 2 signals. In the plots
the lines represent the analytic curves generated from (72) (except for the curve for the Unknown R
case). The symbols represent the results of Monte Carlo experiments.
0 1 2 3 4 5 6 7 8 9 1010
−2
10−1
100
Input SNR
BE
R
BER curves when the signals are spread by 31 degrees
Perfect AUnknown RKnown REstimated RPe Perfect APe Known R
Fig. 4. BER Curves for the 4 estimators. There is one interferer, both signals are spread over 30 degrees,
and the interferer is 1 beamwidth away, so that there is an overlap of 10 degrees between the 2 signals.
In the plots the lines represent the analytic curves generated from (72) (except for the curve for the
Unknown R case). The symbols represent the results of Monte Carlo experiments
October 19, 2003 DRAFT