Vous êtes sur la page 1sur 9

3878

IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 63, NO. 9, SEPTEMBER 2015

Array Beamforming Synthesis for Point-to-Point


MIMO Communication
Farnaz Karimdady Sharifabad, Student Member, IEEE, Michael A. Jensen, Fellow, IEEE,
and Adam L. Anderson, Senior Member, IEEE

AbstractWhile transmit and receive beamforming for the


multiple-input multiple-output (MIMO) point-to-point channel is
a well-studied topic, existing algorithms cannot specify effective
beamformers for certain situations. To address this deficiency,
we present an iterative algorithm that can establish beamformers
based on different channel information available to the transmitter
and receiver, different receiver architectures, and different constraints on the power radiated by the transmit array. Simulation
results based on modeled and measured channels demonstrate that
the algorithm effectively accommodates a wide range of conditions
and that the performance achieved with the approach matches
that obtained using existing algorithms that are applicable only
to specific conditions.
Index TermsAntenna arrays, antenna radiation pattern
synthesis, multiple-input multiple-output (MIMO) systems.

I. I NTRODUCTION

RIOR work has established that antenna array beamforming in multipath propagation environments enables
multiple spatial degrees of freedom [1], [2] that can be exploited
by multiple-input multiple-output (MIMO) signaling in which
multiple data streams, each weighted by a unique beamformer,
are simultaneously communicated [3], [4]. However, this simultaneous data transmission leads to sophisticated considerations
regarding the radio system that impact the array synthesis
problem. For example, while linear beamforming maximizes
throughput in a point-to-point MIMO system if both transmitter
and receiver have accurate channel state information (CSI) [3],
nonlinear successive decoding must be added to the receiver
if the transmitter CSI is inaccurate or outdated or the transmitter only has channel distribution information (CDI) in the
form of a spatial covariance matrix [5], [6]. Furthermore, while
most algorithms assume a sum power constraint (SPC) that
limits the total transmitted power, physical power amplifiers
(PA) favor a per-antenna power constraint (PAPC) that limits the average power transmitted from each antenna [7]. The

Manuscript received March 07, 2015; revised May 11, 2015; accepted June
11, 2015. Date of publication June 18, 2015; date of current version September
01, 2015. This work was supported by the U.S. Army Research Office under
Grant W911NF-12-1-0469.
F. K. Sharifabad was with the Electrical and Computer Engineering
Department, Brigham Young University, Provo, UT 84602 USA. She is
now with Qualcomm Technologies, San Diego, CA 92121 USA (e-mail:
fkarimdady@gmail.com).
M. A. Jensen is with the Electrical and Computer Engineering Department,
Brigham Young University, Provo, UT 84602 USA (e-mail: jensen@byu.edu).
A. L. Anderson is with the Tennessee Technological University, Cookeville,
TN 38505 USA (e-mail: aanderson@tntech.edu).
Digital Object Identifier 10.1109/TAP.2015.2447035

difficulty with these practical considerations is that existing


algorithms for transmit beamformer synthesis focus on maximization of the mutual information and, therefore, inherently
assume specific radio capabilities. Since the optimal transmit beamforming weights depend on the receiver capabilities
and the power constraint, existing algorithms cannot construct
optimal beamformers for many practical radio architectures
that differ from the assumptions inherent in this mutual information maximization. Table I summarizes the availability of
beamforming synthesis algorithms for different communication
scenarios.
These observations suggest that in developing beamforming
strategies for the point-to-point MIMO link, the optimization cost function should incorporate practical system design
decisions. Recent work on multiuser MIMO communication
demonstrates how to write the achievable rate to incorporate
linear transmit processing [10] and/or linear receive processing [11]. More recent work provides an iterative solution to
determine the optimal beamformers for cooperative multiuser
MIMO links based on linear or nonlinear transmit and receive
processing [12]. While these prior developments teach some
foundational principles in MIMO beamforming, they are posed
for multiuser MIMO channels and do not accommodate all of
the practical radio topologies listed in Table I.
This work develops a general framework for transmit and
receive beamformer synthesis by adapting the iterative algorithm detailed in [12] to maximize the achievable communication rate for strictly point-to-point MIMO channels with
different types of channel information (CSI or CDI) available at
the transmitter, linear or nonlinear receiver capabilities, and different power constraints. Simulations based on a simple channel
model and on experimental MIMO channel data reveal that
the approach is highly effective and is able to generate transmit beamformers whose performance with reduced-complexity
radios and practical power constraints rivals that achieved with
more sophisticated radio capabilities.

II. P OINT- TO -P OINT MIMO S IGNALING


A. System Model
Consider a MIMO communication link based on Nt transmit antennas and Nr receive antennas. The Nr 1 received
signal vector is given by y = Hx + , where H is an Nr
Nt channel matrix and is an Nr 1 noise vector whose
entries are i.i.d. zero-mean complex Gaussian random variables

0018-926X 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

SHARIFABAD et al.: ARRAY BEAMFORMING SYNTHESIS

3879

TABLE I
E XISTING MIMO T RANSMIT B EAMFORMING A LGORITHMS

C. Existing Beamforming Methods

with variance 2 . We assume that the system communicates


K min(Nt , Nr ) data streams
by the K 1 data
 represented


vector x0 with covariance E x0 x0 = I, where E {} is the


expectation, {} is the conjugate transpose, and I is the identity matrix. The Nt 1 transmit vector x is then formed from
x = Bx0 , where B is the Nt K matrix whose columns are
the transmit beamforming weights for each stream.
The receiver applies an Nr 1 unit-length beamforming
0k =
vector wk when detecting the kth element of x0 , or x
wk yk where yk represents a possibly modified version of the
received vector y. For example, if the receiver uses only linear
beamforming, then yk = y. However, if the receiver also uses
nonlinear successive interference cancellation (SIC), then yk
represents y modified such that the contributions of the previously detected symbols are removed. For notational simplicity,
we let wk represent the kth column of the Nr K matrix W
and bk represent the kth column of B.
Under these assumptions and with B specified, for an optimal receiver the achievable system rate is [13]




1
Copt = log I + 2 HBB H 

(1)

where | | is the matrix determinant. If the transmitter has no


knowledge of the channel information, then for a total transmit
power P we have K = Nt and BB = (P/Nt )I, so that (1)
becomes the capacity for an uninformed transmitter.

B. Power Constraints
Constructing the transmit beamforming matrix B that maximizes the communication rate requires that we constrain
the transmit power. Most work assumes a SPC that

 limits
x
=
the total
power
transmitted
from
all
antennas,
or
E
x


 

= tr BB = P , where tr() is the trace.


E tr xx
However, in most radios, the power transmitted from each
antenna is limited by the PA. If the SPC is used, each PA
must be able to transmit the total power P despite the fact
that, on average, it must only accommodate its proportional
share (P/Nt ). Therefore, we also consider a PAPC where
the average power radiated fromthe ith
 antenna
 is specified


=
i x0 x b
as Pi . The constraint becomes E xi xi = E b
0 i
ib
i represent the ith element of x and
= Pi , where xi and b
b
i
ith row of B, respectively.

Our objective is to develop a framework for generating transmit beamforming vectors under the assumptions in Table I
where currently no algorithms exist. In preparation for this
development, it is instructive to first review existing algorithms
for two of the scenarios listed.
1) CDI-Based Beamforming With SPC: If the transmitter
possesses CDI in the form of the transmit spatial covariance
Rt = E H H /Nr , the beamformer should be constructed
as B = Ut [5], [14] where is a real, nonnegative,

 diagonal power allocation matrix constrained such that tr =
P (SPC) and Ut is the unitary matrix of eigenvectors of Rt .
Recent work has proposed an iterative solution for the power
allocation assuming that the full channel spatial covariance is
separable into a Kronecker product [6], a technique we call CDI
Kronecker. Alternatively, if we assume that the channel matrix
entries are wide-sense stationary random variables and take the
expectation of (1) [11], [15], by Jensens inequality this average
rate becomes




Nr

(2)
C opt = E {Copt } log I + 2 B Rt B .

The accuracy of using Jensens inequality to achieve this upper


bound has been studied previously [16][18], with the finding
that the accuracy depends on the level of element correlation.
However, the tightness of (2) as a bound is not as important
as whether or not the value of B that maximizes the bound also
approximately maximizes the expectation of (1). Using a waterfilling approach to maximize (2) yields a second approach for
specifying [19], a technique we call CDI waterfilling. Our
numerical results that follow help to establish the validity of
this approach.
2) CSI-Based Beamforming With PAPC: The Drop-Rank
algorithm is an iterative solution for finding the transmit beamformers that maximize the rate, which inherently assumes the
availability of nonlinear detection at the receiver [7]. The
method assumes that the transmitter possesses CSI, and there
does not appear to be a straightforward way to extend the
concept to linear receivers or to CDI-based beamforming.

III. I TERATIVE B EAMFORMING


When choosing transmit beamformers to maximize the
achievable rate in (1) or its upper bound in (2), in most cases,
the receiver must combine linear minimum mean squared error
(MMSE) beamforming with nonlinear SIC to realize the rate
[13], [20], [21]. Because the prior work for CDI-based beamforming under an SPC or CSI-based beamforming under a
PAPC focuses on capacity maximization, the transmit beamformers are derived under the assumption of MMSE-SIC at
the receiver and therefore are inappropriate for use when the
receiver uses only linear beamforming.
This observation motivates development of an alternate
framework for constructing beamforming matrices based on
the available channel information, receiver capabilities, and
power constraint. To develop our iterative beamforming (IBF)

3880

IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 63, NO. 9, SEPTEMBER 2015

approach, we incorporate the receive beamforming into the


expression for the achievable rate to obtain [10], [12]
C=

log (1 + SINRk )

(3)

where the summation is over all values of k such that the fixed
index j is within the set L k = Lk k.
We will see that for the power constraints considered, we can
express the gradient of the Lagrange multiplier term as

k=1

|wk Hbk |2
nk
=
SINRk =

dk
2 + iLk |wk Hbi |2

j f = 2f bj
(4)

where (4) represents the signal-to-interference-plus-noise ratio


(SINR) for the kth data stream.
Because the second term in the denominator of (4) represents the interference to one data stream caused by the other
data streams, the set Lk of integers used in the sum depends
on the receiver architecture. Specifically, for linear receive
beamforming, the kth stream experiences interference from all
other streams, or Lk = {1 i K, i = k}. In contrast, if the
receiver uses SIC, then we assume that the kth stream experiences interference only from streams that have not yet been
detected. For simplicity in notation, we assume that the receiver
detects streams in numerical order of their indices so that Lk =
{k < i K}. Section III-D indicates how this choice of ordering impacts initialization of the optimization. Throughout the
development, we use the notation CSIT and CSIR (CDIT and
CDIR) to indicate that CSI (CDI) is available at the transmitter
and receiver, respectively.

To incorporate design decisions into the rate in (3), we can


write the SINR given in (4) using
nk = bk Ak bk

b i Ak b i
dk = 2 +

(5)

k
jL

We now construct the matrix f for the SPC and PAPC.


1) SPC: For IBF with the SPC, we have
K


bk bk P
f =

Ak = H wk wk H.

(7)

To optimize the rate, we form the cost function = C f ,


where f generically represents a Lagrange multiplier term that
will, in the following, be developed for both the SPC and the
PAPC. We then take the gradient j = j C j f with
respect to bj and set the result equal to zero, where {}
represents a conjugate. The gradient of the rate C is
K

n
C
nk [nk + dk ]
k
=

bj
dk
dk [nk + dk ]

(8)

k=1

where
nk = j nk =
dk = j dk =

(9)

2Ak bj , j Lk
0, j  Lk

(10)

and we have used that wk is unit length (wk wk = 1).


Substitution of (9) and (10) into (8) yields

nk
2
Ak b j
j C = Aj bj 2
(11)
dj
dk (nk + dk )
k
jL

(14)

where is a Lagrange variable, leading to j f = 2bj or


f = I.

(15)

The Lagrange variable can be determined by recognizing


that since we will set j = 0, we have bj j = 0 and


j bj j = 0. Since P =
j bj bj , use of (13) leads to

Fj =

K
1
b Fj bj
P j=1 j


nk
1
Ak
Aj
dj
dk (nk + dk )

(16)
(17)

k
jL

which is equivalent to
=

K

j=1

2
nj
.
dj (nj + dj ) P

(18)

2) PAPC: For IBF with the PAPC, we have


f =

Nt

i=1



Pi
ib
i b
i

(19)

where we now have Nt different Lagrange variables i . This


leads to j f = 2diag()bj , where = [1 , 2 , . . . , Nt ]T
and diag() produce a diagonal matrix from the vector argument. Therefore, we have
f = diag().

2Ak bk , j = k
0, j = k

(13)

k=1

(6)

iLk

j C =

where f is a diagonal matrix. Use of (11) and (12) leads to


nk
2
Ak b j .
j = Aj bj 2 f +
dj
dk (nk + dk )

A. CSIT and CSIR

(12)

(20)

The Lagrange variables are constructed as discussed above for


the SPC after first prescaling with indicator matrices Ji that
are defined as all-zero matrices with a 1 on the ith diagonal
element, leading to
K
1
i =
b Ji F j b j .
Pi j=1 j

(21)

Now that we have expressions for f for both power constraints, we can set (13) to zero with the goal of solving

SHARIFABAD et al.: ARRAY BEAMFORMING SYNTHESIS

3881

for the beamforming vector bj . However, we recognize that


because the beamformers appear explicitly within nk and dk ,
directly solving for bj is difficult. Instead, we can rearrange the
resulting equation into the form
1 1
G Aj b j
dj j

nk
Ak .
Gj = f +
dk (nk + dk )
bj =

(22)
(23)

k
jL

Before discussing the solution of this equation, we note that


the above procedure can be applied to determine the receive
beamformers W.
 Specifically,
 incorporating the Lagrange multiplier f = wj wj 1 , taking the gradient of the cost
function with respect to wj , and setting the result equal to zero
leads to
 1 A
j wj
wj = G
j


n
j
j =
k
2 I +
G
A
nj + dj

(24)

(25)

kLj

k = Hbk b H .
A
k

wj = 2 I +

(26)

kLj

1
Hbk bk H

Hbj .

the rate in (3) over the randomly varying channel H and apply
Jensens inequality to obtain [15]
C = E {C}

log (1 + E {SINRk }) .

(28)

k=1

j on the rightHowever, we recognize that expansion of A


hand side of (24) leaves a right-most product of bj H wj that
is simply a constant, and since wj should have unit length,
this constant is removed through normalization. Therefore, for
CSIR, (24) reduces to the MMSE beamformer (with or without
SIC) [11], [13], [20], [22]

Fig. 1. Flow diagram showing implementation of IBF for CSIT and CSIR.

(27)

Examination of these equations reveals that the following:


1) bj appears explicitly on both sides of (22), as well as
implicitly on the right of (22) through Gj , and
2) (22) depends on the vectors wj and (27) depends on the
vectors bj .
These observations suggest that we first initialize B and then
iteratively compute W using (27) and B using (22) until
both matrices have converged [11]. Since the Lagrange multiplier formulation does not guarantee enforcement of the power
constraint, in each iteration, the transmit beamformers are normalized to explicitly enforce the constraint. If we use the
superscript {}(n) to indicate a quantity computed at the nth
iteration, then Fig. 1 shows a simple flow diagram for performing the iteration, where the symbol TC indicates the threshold
for terminating the iteration that is chosen as TC = 104 for all
simulations in this paper.

B. CDIT and CSIR


This beamformer synthesis technique can be extended to the
case where the transmitter has CDI. We take the expectation of

As discussed in Appendix A, we further make the approximation E {SINRk } E {nk }/E {dk } = n
k /dk . We can now
perform the procedure of Section III-A using n
k and dk in
place of nk and dk , respectively. This leads to the same iterative
equation (22) but with the replacement


Aj E H wj wj H .
(29)
When implementing this algorithm, however, we must carefully consider approximation of the expectation in (29). If we
assume a single value of W is valid over the time window
during which the transmit beamformer B based on CDI is
assumed valid, Aj can be computed based on knowledge of
the beamforming
matrix W and the full covariance matrix


R = E hh where h represents H stacked columnwise, as
illustrated in Appendix B. In practice, the receiver computes a
new value of W each time it estimates a new channel matrix H,
and, therefore, it is arguably more accurate to estimate Ak using
a sample mean with the true values of H and W over the time
window. Naturally, this latter approach is more computationally
costly, since we must compute this sample mean at each step in
the iterative computation. Our analysis with the experimental
data discussed in Section IV shows that the accuracy improvement associated with this more costly approach provides very
little performance improvement.
C. CDIT and CDIR
The IBF framework can also specify receive beamformers based on CDI using the iterative computation for wj
j
represented
in(24)(26) and making the substitution A


E Hbj bj H . This is a problem that has been considered


from an information theoretic standpoint, but since practical
techniques for coherent symbol demodulation without CSI at

3882

IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 63, NO. 9, SEPTEMBER 2015

the receiver remain an unsolved problem, we simply point out


the development for completeness and ignore this particular
scenario in the remainder of the paper.
D. Beamformer Initialization and SIC Ordering
As indicated in Section III-A, the iteration begins with initialization of B. Since transmitting multiple streams is generally
best accomplished using orthogonal beamformers, one option
is to form a matrix with independent, zero-mean, unit-variance
complex Gaussian random variables and use its unitary singular
vectors (SVs) as the initial value of B. Since, our goal is simply
to obtain a variety of different unitary matrices, other random
matrix realizations are likely effective. Regardless, since the
cost function, in general, is not convex and convergence to the
beamformers that achieve the global optimum rate is sensitive
to the choice of initialization, we run the iteration with a number
of random initializations and choose the result that achieves the
highest rate. Alternatively, if using IBF with CSIT or CDIT, we
can initialize B as the right SVs of the channel matrix H or the
eigenvectors of the transmit covariance matrix Rt , respectively.
This initialization discussion also relates to the specification
of the order in which symbols are detected when implementing SIC at the receiver. Specifically, because we have assumed
detection in the order of the stream index, achieving the optimal
rate means that the iterative algorithm orders the beamforming
vectors in the optimal way. The algorithms ability to achieve
this ordering depends on the initialization. Since, optimal ordering typically means that the stream with the highest SINR
should be detected first, we have found that the proper ordering
of the initializing unitary matrix corresponds to the case where
the singular values (or eigenvalues) are placed in decreasing
order.
IV. R ESULTS
A. Simulation Approach
The performance of the IBF algorithm is demonstrated here
through simulations. In all cases, the MIMO channel matrix H
is normalized so that
H 2F = Nr Nt

use an adjacent subset of antennas in the transmit and receive


array to generate the Nr Nt channel matrix.
Use of this experimental data does not facilitate systematic
evaluation of the effect of the channel directivity, or equivalently multipath richness, on algorithm performance. We,
therefore, also use a two-ring channel model to analyze the
impact of multipath richness on the performance of different
transmit beamforming techniques. In this model, the transmitter and receiver each lie within distinct circles of radius 40.
A fixed number of scatterers lie on each circle, with the angle
of each scatterer location being specified as a uniform random variable on [0, 2). Scatterers on the transmit and receive
circles are paired so that each propagation path has a single
departure and arrival angle as well as a complex gain specified
as a zero-mean, unit-variance circularly symmetric complex
Gaussian random variable. The linear transmit array remains
stationary at the center of its circle while the linear receive
array moves along a straight line within its circle. Normalized
channel matrices are easily computed in this two-dimensional
propagation model using established techniques [23].
Prior work has demonstrated that optimally allocating power
to different beamforming vectors offers little advantage over
equally allocating power across the beamformers as the SNR
increases [26]. Therefore, unless otherwise specified the simulations use a transmit power of P = 1 and noise variance
2 = 1 for an SNR of 0 dB, with other default parameters
being Nt = Nr = K = 4. When the PAPC is applied, we use
Pi = P/Nt , 1 i Nt . When average results are provided,
for each value of the swept parameter (number of antennas,
number of paths, etc.), we compute the rate for each channel sample and then compute the average of these rates. The
channel samples obtained from the experiments and model are
naturally correlated, with the objective that the results show
what would happen as a radio moves through the environment. When using CDIT, the required covariance is estimated
using a sample mean of these correlated channel observations
to approximate R over a specified window size (in terms of
receiver motion) of 1, and this covariance is used to construct the transmit beamforming matrix B that is fed back to
the transmitter at a regular feedback interval of 2.5.

(30)

where F represents the Frobenius norm. With this normalization, the ratio P/2 represents the single-input single-output
signal-to-noise ratio (SNR) [23].
Most of the analysis is performed with MIMO channels measured at a carrier frequency of 2.45 GHz for linear transmit
and receive arrays of monopoles with half-wavelength element
spacing. The transmitter is held stationary while the receiver
moves at a constant velocity of approximately 30 cm/s, and
channels are sampled at an interval of 2.5 ms corresponding
to a distance moved by the receiver of 0.0062, where is
the wavelength at the carrier frequency. All measurements are
taken within an open area between several buildings on the
Brigham Young University campus with the transmitter and
receiver positioned on either side of a dense stand of trees.
Details concerning the measurement system used to collect the
channel matrices can be found in [24] and [25]. The simulations

B. CSIT and CSIR


To begin, for each of the experimentally obtained channel
matrices, we compute the capacity using the waterfilling solution [3] and the rate using the beamformers obtained with the
IBF algorithm with the SPC (based on an MMSE receiver)
assuming CSIT and CSIR. While the results are not shown for
the sake of brevity, the rate obtained using IBF is indistinguishable from the waterfilling capacity over a variety of system
parameters. This demonstrates that for this simple case, the IBF
algorithm provides the expected result.
While this simple demonstration is encouraging, it is more
interesting to explore IBF performance for the PAPC. Fig. 2
shows the rate averaged over the experimental channels as a
function of the number of antenna elements in the transmit and
receive arrays for the IBF algorithm with a PAPC and the previously published Drop-Rank algorithm [7] with both MMSE and

SHARIFABAD et al.: ARRAY BEAMFORMING SYNTHESIS

Fig. 2. Rate averaged over all experimentally obtained channel matrices for
CSIT and CSIR with the PAPC as a function of the number of antennas in the
transmit and receive arrays (Nr = Nt ).

3883

Fig. 3. Fractional difference (n) between the converged rate and the rate at
each iteration index n in the IBF algorithm for CSIT and CSIR with an MMSESIC receiver averaged over all experimentally obtained channel matrices for
different algorithm initializations.

MMSE-SIC receivers. For the MMSE-SIC receiver, the performance of the two algorithms is almost identical. However,
when an MMSE receiver is used, the rate achieved with the
Drop-Rank beamformer that is designed for a nonlinear receiver
degrades significantly, while that obtained using the IBF beamformer remains almost unchanged. The key point is that the IBF
can accommodate reduced receiver capabilities and, at least for
this case, do so with little performance degradation.
Because IBF requires an iterative computation, it is important to consider the convergence rate of the iteration as well as
what is required to achieve the global optimum solution to the
nonconvex optimization. To explore the convergence rate, for
each iteration, we compute the fractional difference


Cfinal C (n) 
(n)
(31)
=
Cfinal
where C (n) and Cfinal denote the rate at the nth iteration and at
final convergence, respectively. Fig. 3 plots this convergence
metric as a function of the iteration index n for both power
constraints and for two different array sizes. The algorithm is
initialized with the channel right SVs and 20 randomly generated vectors in the top and bottom plots, respectively. As can
be seen, initialization with the channel SVs is highly effective,
even with the PAPC. Furthermore, regardless of the initialization, on average the iteration converges within 20 iterations.
Fig. 4 plots the cumulative distribution function (CDF) of the
number of iterations required for convergence for the same scenarios considered in Fig. 3, where convergence is defined as
the value of n for which (n) 104 . These results confirm
that when initializing with the channel SVs, fewer than 20 iterations are needed to essentially guarantee convergence, while
significantly more iterations are required when the computation is randomly initialized. Fig. 4 also shows the result for the
Drop-Rank algorithm for Nr = Nt = 8 (similar results occur
for Nr = Nt = 4), revealing that for most cases, IBF converges
more rapidly than Drop-Rank.
To explore achievement of the global optimum solution, for
each experimental channel observation, we seed the iteration
with NI random matrices and choose the outcome of the iteration that achieves the highest rate. Fig. 5 plots the achieved

Fig. 4. CDF of the number of iterations required for the IBF algorithm to converge for CSIT and CSIR with an MMSE-SIC receiver using all experimentally
obtained channel matrices for different algorithm initializations, array sizes,
and power constraints.

rate averaged over all experimental channel observations as


a function of NI . The triangular symbols show the average
rate achieved when the channel SVs are used for the initialization. As can be seen, when using the SPC or the PAPC
with an MMSE-SIC receiver, we achieve the maximum in just
a few random initializations, and more importantly the channel SV initialization achieves the same result without requiring multiple iterative optimizations. However, when using the
PAPC with an MMSE receiver, more random realizations are
required. Furthermore, while the channel SV initialization with
an MMSE receiver is not as effective for the large array, the initialization still achieves a result that is within 1.5% of the rate
achieved with the more costly random initializations.
C. CDIT and CSIR
We next compare the performance achieved with IBF to that
obtained using the two previously reported methods for beamforming based on CDIT [6], [19]. Fig. 6 plots the rate averaged
over the ring model realizations as a function of the number

3884

IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 63, NO. 9, SEPTEMBER 2015

Fig. 5. Rate averaged over all experimentally obtained channel matrices for
CSIT and CSIR with the different power constraints and array sizes as a
function of the number of random initializations and, therefore, iterative optimizations. The average rate obtained when initializing with the channel right
SVs is also shown.

Fig. 6. Average rate achieved using different transmit beamformers for CDIT
and CSIR (MMSE-SIC) as a function of the number of paths in the two-ring
channel model.

of multipaths for the three transmit beamforming algorithms


assuming MMSE-SIC at the receiver. As can be seen, the performance of IBF with the SPC matches that achieved with
the previously developed CDI-based beamforming algorithms
(that also use the SPC). Since both CDI waterfilling and IBF
maximize the upper bound on the rate while CDI Kronecker
maximizes the rate, their nearly identical performance suggests
that maximizing the upper bound yields acceptable results.
Furthermore, the considerable difference between the rate for
an uninformed transmitter and that for CDI-based beamforming
illustrates the improvement offered by transmit beamforming, although the benefit is reduced as the channel directivity
decreases (more paths).
When the channel is highly directive (few paths), very specific transmit beamforming is required to fully exploit the
channel. Since the PAPC limits the range of beamformers that
can be achieved, IBF with the PAPC performs about the same as
or slightly better than the uninformed transmitter. As the number of paths increases, however, the results show that it becomes

Fig. 7. Average rate achieved using different transmit beamformers for CDIT
and CSIR (MMSE-SIC and MMSE) with the SPC as a function of SNR using
experimentally obtained channel matrices.

Fig. 8. Average rate achieved using different transmit beamformers for CDIT
and CSIR with the PAPC as a function of SNR using experimentally obtained
channel matrices.

easier to form beams that can take advantage of the propagation environment while still satisfying the PAPC, although these
results still fall short of those achieved with the more flexible
SPC.
Fig. 7 shows the rate achieved using different techniques
with the SPC as a function of SNR over the experimentally
obtained channels using MMSE-SIC and MMSE receivers. The
simulation demonstrates that the algorithm works effectively at
high SNR. Furthermore, all results show that the effectiveness
of transmit beamforming relative to uninformed transmission
decreases with increasing SNR, as suggested in Section IV-A.
Once again, all three algorithms achieve nearly identical performance for an MMSE-SIC receiver. However, when using the
MMSE receiver, the performance of CDI waterfilling and CDI
Kronecker degrades significantly, while IBF is able to maintain competitive performance. Fig. 8 plots similar results for the
IBF algorithm with the PAPC using MMSE-SIC and MMSE
receivers. As there are no existing algorithms for this case, only
the uninformed transmitter performance is shown for comparison purposes. The results are quite similar to those in Fig. 7 for

SHARIFABAD et al.: ARRAY BEAMFORMING SYNTHESIS

3885

the SPC, both in trend and in achieved performance. If we compare the results in Figs. 7 and 8 at 0 dB SNR to those in Fig. 6
for a large number of paths, we see that the values are very similar. This comparison suggests that the experimental channels
generally have a large number of multipath components.

where vec() stacks the matrix argument columnwise into a vector, reshape(, m, n) reshapes the vector into a m n matrix,
and

V. C ONCLUSION

where {}T and represent a transpose and a Kronecker product, respectively. The elements of t and r correspond to
elements of the full covariance matrix R.

This paper proposes an iterative algorithm that specifies


transmit and receive beamformers based on different types
of channel information, different capabilities of the receiver,
and different power constraints, thereby offering beamforming
solutions for specific situations where solutions have not yet
been available. Simulations in measured and modeled MIMO
channels show that the performance achieved with the technique matches that obtained with existing algorithms when
used in conjunction with optimal receiver architectures but
that it further allows construction of near-optimal transmit
beamforming for simple MMSE receivers. The simulations further demonstrate the impact of different system parameters on
communication performance.

A PPENDIX A
A PPROXIMATING THE E XPECTED SINR
The transmit beamformer for CDIT relies on the approximation E {SINRk } E {nk }/E {dk }. While rigorously justifying this assumption is difficult, it can be shown that

E

nk
dk

E {nk } cov(nk , dk ) var(dk )E {nk }

+
(32)
2
3
E {dk }
E {dk }
E {dk }

where cov(x, y) is the covariance of x and y and var(x) is


the variance of x. Examining (4), if the transmit beamformers
are nearly orthogonal, then it is reasonable that cov(nk , dk ) is
small. Furthermore, over a block that might be used for covariance estimation, it is likely that the variance of dk is also small.
These observations support the approximation. Furthermore,
even if the approximation is poor, as long as finding the beamformers that maximize the rate under this approximation also
maximize the actual expected rate, then use of the approximation is justified. Since, our results demonstrate that the
performance of the algorithm with this approximation matches
that of other algorithms (for an MMSE-SIC receiver), use of
this approximation has merit.

A PPENDIX B
IN T ERMS OF R
E XPRESSING A AND A
in terms of elements
It is straightforward to express A and A
of the full covariance matrix R. Specifically,


(33)
Ak = reshape t vec(wk wk ), Nt , Nt


k = reshape r vec(bk b ), Nr , Nr
A
(34)
k



t = E HT H
r = E {H H}

(35)
(36)

R EFERENCES
[1] M. D. Migliore, On the role of the number of degrees of freedom of the
field in MIMO channels, IEEE Trans. Antennas Propag., vol. 54, no. 2,
pp. 620628, Feb. 2006.
[2] M. A. Jensen and J. W. Wallace, Capacity of the continuous-space electromagnetic channel, IEEE Trans. Antennas Propag., vol. 56, no. 2,
pp. 524531, Feb. 2008.
[3] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, Capacity limits
of MIMO channels, IEEE J. Sel. Areas Commun., vol. 21, no. 5, pp. 684
702, Jun. 2003.
[4] M. A. Jensen and J. W. Wallace, A review of antennas and propagation
for MIMO wireless communications, IEEE Trans. Antennas Propag.,
vol. 52, no. 11, pp. 28102824, Nov. 2004.
[5] S. A. Jafar and A. Goldsmith, Transmitter optimization and optimality of beamforming for multiple antenna systems, IEEE Trans. Wireless
Commun., vol. 3, no. 4, pp. 11651175, Jul. 2004.
[6] A. Soysal and S. Ulukus, Optimum power allocation for single-user
MIMO and multi-user MIMO-MAC with partial CSI, IEEE J. Sel. Areas
Commun., vol. 25, no. 7, pp. 14021412, Sep. 2007.
[7] M. Vu, MISO capacity with per-antenna power constraint, IEEE Trans.
Commun., vol. 59, no. 50, pp. 12681274, May 2011.
[8] P. W. Wolniansky, G. J. Foschini, G. D. Golden, and R. A. Valenzuela,
V-BLAST: An architecture for realizing very high data rates over the
rich-scattering wireless channel, in Proc. URSI Int. Symp. Signals Syst.
Electron. (ISSSE98), Pisa, Italy, Sep. 29/Oct. 2, 1998, pp. 295300.
[9] G. D. Golden, C. J. Foschini, R. A. Valenzuela, and P. W. Wolniansky,
Detection algorithm and initial laboratory results using V-BLAST spacetime communication architecture, Electron. Lett., vol. 35, pp. 1416, Jan.
1999.
[10] M. Stojnic, H. Vikalo, and B. Hassibi, Rate maximization in multiantenna broadcast channels with linear preprocessing, IEEE Trans.
Wireless Commun., vol. 5, no. 9, pp. 23382342, Sep. 2006.
[11] Q. H. Spencer, J. W. Wallace, C. B. Peel, T. Svantesson,
A. L. Swindlehurst, and A. Gummalla, Performance of multi-user
spatial multiplexing with measured channel data, in MIMO System
Technology and Wireless Communications. Boca Raton, FL, USA: CRC
Press, 2006.
[12] A. L. Anderson and M. A. Jensen, Sum-rate maximization in distributedantenna heterogeneous MIMO downlinks: Application to measured channels, IEEE J. Sel. Topics Signal Process., vol. 6, no. 3, pp. 270280, Jun.
2012.
[13] P. Viswanath and D. N. C. Tse, Sum capacity of the vector Gaussian
broadcast channel and uplinkdownlink duality, IEEE Trans. Inf.
Theory, vol. 49, no. 8, pp. 26182625, Aug. 2003.
[14] E. Visotsky and U. Madhow, Space-time transmit precoding with imperfect feedback, IEEE Trans. Inf. Theory, vol. 47, no. 6, pp. 26322639,
Sep. 2001.
[15] A. L. Anderson, J. R. Zeidler, and M. A. Jensen, Reduced-feedback linear precoding with stable performance for the time-varying MIMO broadcast channel, IEEE J. Sel. Areas Commun., vol. 26, no. 8, pp. 14831493,
Oct. 2008.
[16] A. Gorokhov, Capacity of multiple-antenna Rayleigh channel with
a limited transmit diversity, in Proc. IEEE Intl. Symp. Inf. Theory,
Sorrento, Italy, Jun. 2530, 2000, p. 411.
[17] S. Loyka and A. Kouki, New compound upper bound on MIMO channel
capacity, IEEE Comm. Lett., vol. 6, no. 3, pp. 9698, Mar. 2002.
[18] S. Loyka and A. Kouk, On the use of Jensens inequality for MIMO
channel capacity estimation, in Proc. Can. Conf. Elect. Comput. Eng.,
Toronto, ON, Canada, May 1316, 2001, pp. 475480.

3886

IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 63, NO. 9, SEPTEMBER 2015

[19] M. Vu and A. Paulraj, Characterizing the capacity for MIMO wireless


channels with non-zero mean and transmit covariance, in Proc. 43rd
Allerton Conf. Commun. Control Comput., Monticello, IL, USA, Sep.
2830, 2005, pp. 268277.
[20] R. Zhang, J. M. Cioffi, and Y.-C. Liang, Throughput comparison of
wireless downlink transmission schemes with multiple antennas, in
Proc. IEEE Intl. Conf. Commun., Seoul, Korea, May 1620, 2005,
pp. 27002704.
[21] M. K. Varanasi and T. Guess, Optimum decision feedback multiuser
equalization and successive decoding achieves the total capacity of the
Gaussian multiple-access channel, in Proc. Conf. Record 31st Asilomar
Conf. Signals Syst. Comput., Pacific Grove, CA, USA, Nov. 25, 1997,
pp. 14051409.
[22] A. L. Anderson, J. R. Zeidler, and M. A. Jensen, Stable transmission
in the time-varying MIMO broadcast channel, EURASIP J. Adv. Signal
Process., vol. 2008, pp. 114, Oct. 2008, doi:10.1155/2008/617020.
[23] M. A. Jensen and J. W. Wallace, MIMO wireless channel modeling
and experimental characterization, in Space-Time Processing for MIMO
Communications, J. D. Gibson, Ed. Hoboken, NJ, USA: Wiley, 2005,
ch. 1, pp. 139.
[24] J. W. Wallace, M. A. Jensen, A. L. Swindlehurst, and B. D. Jeffs,
Experimental characterization of the MIMO wireless channel: Data
acquisition and analysis, IEEE Trans. Wireless Commun., vol. 2, no. 2,
pp. 335343, Mar. 2003.
[25] J. W. Wallace and M. A. Jensen, Time-varying MIMO channels:
Measurement, analysis, and modeling, IEEE Trans. Antennas Propag.,
vol. 54, no. 11, pp. 32653273, Nov. 2006.
[26] M. A. Khalighi, J. Brossier, G. Jourdain, and K. Raoof, Water filling
capacity of Rayleigh MIMO channels, in Proc. IEEE 12th Int. Symp.
Pers. Indoor Mobile Radio Commun., vol. 1. San Diego, CA, USA, Sep.
30/Oct. 3, 2001, pp. 155158.

Farnaz Karimdady Sharifabad (S12) received the


B.S. degree in electrical engineering from Amirkabir
University of Technology (Tehran Polytechnic),
Tehran, Iran, in 2007, the M.S. degree in wireless communications from Lund University, Lund,
Sweden, in 2010, and the Ph.D. degree in electrical and computer engineering from Brigham Young
University, Provo, UT, USA, in 2013.
She is currently a Systems Engineer with
Qualcomm Technologies Incorporated, San Diego,
CA, USA. Her research interests include covariance modeling for multipath propagation channels and multiantenna signal
processing.

Michael A. Jensen (S93M95SM01F08)


received the B.S. and M.S. degrees from Brigham
Young University (BYU), Provo, UT, USA, in 1990
and 1991, respectively, and the Ph.D. degree from
the University of California, Los Angeles, CA, USA,
in 1994, all in electrical engineering.
Since 1994, he has been with the Electrical and
Computer Engineering Department, BYU, where he
is currently a University Professor. His research
interests include antennas and propagation for communications, microwave circuit design, multiantenna
signal processing, and physical layer security.
Dr. Jensen is currently a President-Elect of the IEEE Antennas and
Propagation Society. He was previously the Editor-in-Chief of the IEEE
T RANSACTIONS ON A NTENNAS AND P ROPAGATION as well as an Associate
Editor for the same journal and for the IEEE Antennas and Wireless
Propagation Letters. He has been a Member and Chair of the Joint Meetings
Committee for the IEEE Antennas and Propagation Society, a member of the
society AdCom, a member of the society Publications Committee, and Cochair
or Technical Program Chair for six society-sponsored symposia. In 2002, he
was the recipient of Harold A. Wheeler Applications Prize Paper Award in
the IEEE T RANSACTIONS ON A NTENNAS AND P ROPAGATION. He was elevated to the grade of the IEEE Fellow in 2008 in recognition of his research on
multiantenna communication.

Adam L. Anderson (S00M10SM15) received


the B.S. and M.S. degrees from Brigham Young
University, Provo, UT, USA, in 2002 and 2004,
respectively, and the Ph.D. degree from the University
of California at San Diego, La Jolla, CA, USA, in
2008, all in electrical engineering.
He was a Research Assistant Professor with the
University of South Florida, Tampa, FL, USA, and
is currently an Assistant Professor with Tennessee
Technological University, Cookeville, TN, USA.
Dr. Anderson was the winner of the 2014 DARPA
Spectrum Challenge, recipient of the 2014 Leighton E. Sissom Award for
Creativity and Innovation, and a 2015 ORAU HERE Faculty Fellow at Oak
Ridge National Laboratory.

Vous aimerez peut-être aussi