4020 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO.

8, AUGUST 2008

Rate Optimization for Multiuser MIMO Systems

With Linear Processing
Shuying Shi, Student Member, IEEE, Martin Schubert, Member, IEEE, and Holger Boche, Senior Member, IEEE

Abstract—This paper focuses on linear transceiver design which offer a good performance at less implementation com-
for rate optimization in multiuser Gaussian multple-input mul- plexity. But finding an optimum linear scheme is still an open
tiple-output (MIMO) systems. We focus on two design criteria: problem.
1) maximizing the weighted sum rate subject to a total power
constraint; 2) maximizing the minimum user rate subject to a In this paper, we consider a multiuser downlink MIMO
total power constraint. For these problems, new power allocation system with users and each user can perform spatial mul-
strategies are derived, which can be formulated as geometric tiplexing with several data streams (layers). The discussion is
programs (GPs) involving mean square errors (MSEs). Based focused on linear equalization, since this problem is still open.
on these solutions, we propose iterative algorithms, where each However, the approach can also be combined with DPC and
iteration contains the optimization of the uplink power, uplink
receive filters, and downlink receive filters. Monotonic conver-
successive interference cancellation (SIC), as discussed later.
gence of the algorithms is proved. Simulations show that the We aim at jointly optimizing the powers, transmit and receive
algorithms outperform existing linear schemes. Additionally, we filters with respect to following design criteria:
extend the results to other variations of the problems, such as, the P1: weighted sum-rate maximization under a total power
problem of sum-rate constrained or user-rate constrained power constraint;
minimization, and the problem of sum-rate maximization subject
P2: max-min balancing the user rate under a total power
to user-rate constraints and a total power constraint.
constraint.
Index Terms—Max-min fairness, multiuser multiinput-multi- We will also discuss variations of these problems, such as rate-
output (MIMO), sum-rate optimization, transceiver design. constrained power minimization, etc.
One commonly used strategy for the above optimization
problems is to diagonalize or block-diagonalize the channel
I. INTRODUCTION [11]–[15]. However, such a zero-forcing receive or transmit
strategy suffers from noise or power enhancement and has a re-
striction on the number of transmit and receive antennas. Thus,
zero-forcing is generally not a capacity-achieving strategy.
T HE capacity region of multiuser multiple-input mul-
tiple-output (MIMO) Gaussian broadcast channels has
been characterized in [1]–[4]. However, the search for effi-
In recent years, duality between uplink and downlink chan-
nels has been proven useful for the development of optimum
cient practical schemes to achieve the limit is still ongoing. transceiver strategies. A duality based on signal-to-interference
Existing schemes are usually under the assumption of “dirty ratios (SIR) was already observed in a power control context
paper coding” (DPC). Examples are schemes for sum-capacity [17]. Later it was used for beamforming optimization [18]. This
optimization [5]–[7] and individual rate optimization [8]–[10]. was extended and formalized in [4] and [19]. Later, a connection
However, dirty paper precoding techniques are difficult to with a Lagrangian duality was observed in [20], and per-antenna
realize in practice, and there are still open issues, like finding power constraints were investigated in [21]. Independently, an
the optimal precoding order [8], [10]. Thus, much research has information-theoretical duality was shown for capacity regions
been spent on linear equalization schemes (see, e.g., [11]–[16]), of MIMO Gaussian broadcast channels [1], [22]–[24].
Given these results, it is perhaps not surprising that duality
also holds in a minimum mean square error (MMSE) context.
Manuscript received July 7, 2007 revised December 2, 2007. This work was For MMSE regions of multiuser MIMO channels, it was shown
supported in part by the STREP project No. IST-026905 (MASCOT) within the in [25] and [26] that MIMO uplink and downlink channels
sixth framework programme of the European Commission. The work of S. Shi
was supported by the Deutsche Forschungsgemeinschaft (DFG) by Grant SCHU share the same MSE region. This was extended in [27]. Based
2107/2-1. The associate editor coordinating the review of this manuscript and on MSE duality, many known uplink strategies for transceiver
approving it for publication was Dr. Geert Leus. optimization can be transferred to the downlink. For example,
S. Shi is with the Heinrich-Hertz-Chair for Mobile Communications, Tech-
nical University of Berlin, 10587 Berlin, Germany (e-mail: Shuying.Shi@hhi.
the problem of sum-MMSE minimization subject to a total
fhg.de). power constraint was solved in [26], [27]. There is also research
M. Schubert is with the Fraunhofer German-Sino Laboratory for Mobile on power efficient designs under individual MMSE constraints
Communications MCI, 10587 Berlin, Germany (e-mail: Martin.Schu-
[email protected]).
[27]–[30]. Uplink designs with respect to MMSE are also
H. Boche is with the Heinrich-Hertz-Chair for Mobile Communications, studied, e.g., in [31]–[33].
Technical University of Berlin, 10587 Berlin, Germany. He is also with the However, all these MMSE-based strategies are not optimal
Fraunhofer German-Sino Laboratory for Mobile Communications MCI, and
the Fraunhofer Institute for Telecommunications, Heinrich-Hertz-Institut HHI,
with respect to the rate optimization problems considered in this
10587 Berlin, Germany (e-mail: [email protected]). paper. Even finding the sum-rate optimal MIMO transceivers is
Digital Object Identifier 10.1109/TSP.2008.924796 still an open problem for the linear case. Partial results appeared
1053-587X/$25.00 © 2008 IEEE
SHI et al.: MULTIUSER MIMO SYSTEMS WITH LINEAR PROCESSING 4021

in [16] and [34], but only for beamforming (MISO) channels,

and in [35], where power allocations were studied.
It was noted in [35] that the sum-rate optimization problem
leads to a problem formulation involving ratios of posynomials,
which is generally an intractable, NP hard problem. The prob-
Fig. 1. Downlink channel.
lems studied here are even more complicated, since we consider
a joint optimization over powers, transmit filters and receive The following notations are used. Boldface letters denote ma-
filters. Thus, the existence of efficient and optimal algorithms trices and vectors. The superscripts and denote the
cannot be expected. The goal should be the development of transpose and the conjugate transpose, respectively. The nota-
transceiver designs with good performance/complexity tradeoff. tion returns a diagonal matrix with the elements of a
The main contributions of the paper are as follows. vector on the diagonal, or return a vector containing the diag-
1) We derive new power allocation algorithms for problems onal elements of a matrix, and returns a block diag-
P1 and P2. Our approach exploits the connection between onal matrix with matrices on its diagonal. is the identity matrix
capacity and MMSE for Gaussian distributed signals. and is the all one vector. and are the -norm and
Thereby we can reformulate the problem as a geometric -norm, respectively. The notation stands for infimum and
program (GP) involving the geometric MSE, for which for supremum.
the global optimum can be found efficiently. The relation
between capacity and MMSE is ensured by updating the II. SYSTEM MODEL
receivers as MMSE receivers.
Note, that this approach differs from [35], where a We consider a standard multiuser downlink MIMO model
high-SINR approximation was used. (Fig. 1) with transmit antennas at the base station and
The performance comparison with our proposed strategy decentralized receivers, each with antennas. The
channel matrix is where
under a heuristic optimization framework, is given in
models the channel between the th user and the base station.
Section VI. The proposed power allocation algorithms
Without loss of generality, we assume that the transmit data
also differ from the MMSE optimization problems con-
symbols are independent and with unity average powers, i.e.,
sidered in [26], [28]–[30], where the optimization is with with .
respect to the sum of MSEs. Each user can multiplex data streams (each data stream
2) Iterative algorithms for jointly optimizing powers, transmit is also referred to as a layer), i.e., is the data
filters, and receive filters with respect to P1 and P2 are pro- vector to be transmitted to the th mobile. We denote the data
posed. The algorithms are based on the aforementioned op- symbol of the th stream (layer) as and the total number of
timal power control strategies and the MSE duality [26]. In the transmit (active) data streams (layers) as .
each iteration, the following optimization steps are carried Zero-mean complex white Gaussian noise is denoted by
out successively: a) uplink power control by GP; b) up- . The data and the noise are
link receive MMSE filtering; c) dual transformation from statistically independent.
the uplink to the downlink; d) downlink receive MMSE fil- For convenience, the transmit and receive filters are sep-
tering. By exploiting the optimality of the power allocation arated in two parts, matrices with unity-norm columns
step, this joint strategy is proved to be convergent. and diagonal matrices. In particular, the transmit filter is
3) We pay special attention to problems P1 and P2. However, and receive filters are ,
it will be seen later that the results can also be extended to , where the diagonal matrix
other variations of the problems, such as, the problem of contains the transmit powers of all users. The matrices
sum-rate constrained or user-rate constrained power mini- and are with
mization, and the problem of sum-rate maximization sub- normalized columns, i.e., and , . The
ject to user-rate constraints and a total power constraint. matrix is a diagonal matrix with
4) If SIC is applied in the uplink or DPC in the downlink, then positive entries. The system model is given by
we observe from simulations that the proposed algorithm
achieves the global sum-rate optimum. This indicates that (1)
the proposed algorithm can be considered as an alternative
to the iterative waterfilling algorithm [6], [7].
It should be noted that the results of this paper are not re-
stricted to the downlink broadcast channel. The assumption of III. PROPOSED ALGORITHMS
a total power constraint is typical for a downlink channel, so The proposed algorithms are based on the MSE duality be-
our discussion will focus on this case. But besides the power tween the uplink and downlink channels [25], [26] (see also
constraint, all results hold equivalently for uplink and downlink Appendix B).
channels. It will be seen later, that a joint treatment of both links To this end, we first introduce an equivalent uplink channel
is useful and even necessary, since our algorithms are based on (Fig. 2), which is obtained by switching the role of the nor-
the duality between uplink and downlink MMSE regions [25], malized transmit and receive filters. The th transmit filter is
[26]. By exploiting the duality, transmit powers and filters can and the receive filter is .
be computed efficiently. The quantities , , and are the same as for the downlink
4022 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 8, AUGUST 2008

the theoretical optimum is suitable as a good starting point in

most cases.
Moreover, we have the well known relation between the
achieved SINR and MMSE

Fig. 2. Uplink channel.

where
model. Additionally, we assume that the uplink noise power is
the same as the downlink noise power. The effective channel
(including transmit and receiver filters) is the transpose of the
effective downlink channel. In general, different power alloca-
tions are needed in the uplink and downlink in order to achieve
the same MSE values. We use a diagonal matrix to denote (3)
. We also write the powers in vector form
as and , and assume that both links
fulfill the same sum-power constraint, i.e., . This holds under the assumption of an optimum MMSE receiver
With this uplink channel model, we are ready to derive iter- (see, e.g., [39])
ative algorithms for P1 and P2, where the variables to be opti-
mized are , , , and .
(4)

It is known that with Gaussian input and noise, and under the
A. P1: Weighted Sum-Rate Maximization assumption of linear processing, linear MMSE estimation is the
optimal receive strategy. Therefore, the rate per layer can be
The weighted sum-rate maximization can be formulated as expressed as a function of MMSE, i.e.

By introducing (layer-wise) weighting coefficients

, the weighted sum rate in terms of MMSE is
(2) given by

where is the total power constraint, and

, , are the layer-rates (rate per channel use) (5)
and , , are weighting factors, which can be used to model
individual user priorities.
In the following, we will first explain the optimization in the Thus, maximizing the weighted sum rate is equivalent to min-
uplink and downlink channel separately. Then a joint transceiver imizing the product of MMSEs, or equivalently the weighted
optimization based on MSE duality is proposed. geometric MMSE. The MMSE (3), however, is coupled by the
1) Optimization in the Uplink Channel: Under the common powers and transmit filters, which makes the joint optimization
assumption of Gaussian signaling and noise, there is a very difficult. In order to minimize the weighted geometric
one-to-one monotone relationship between the maximum MMSE, we perform two steps for the uplink optimization:
achievable layer-rate and the SINR first, for fixed transmit and receive filters and , the
weighted geometric MSE is minimized by finding the optimal
uplink power allocation ; then, the receive filter
is updated as the MMSE filter (4).
For fixed filters and , the MSE of the th layer is given
hboxwhere by (see Appendix B)

(6)

with a diagonal matrix

Note that the relation between rate and SINR requires an infinite (7)
signaling time interval. For practical designs, the SNR gap ap-
proximation [36] can be utilized to predict the achievable rate and a matrix defined by
depending on the modulation and coding scheme. Another al-
ternative approach is the use of heuristic bitloading algorithms.
(8)
Examples can be found in [37], [38]. The authors also show that
SHI et al.: MULTIUSER MIMO SYSTEMS WITH LINEAR PROCESSING 4023

It can be observed that the weighted geometric MSE is TABLE I

a posynomial (see Appendix A). Hence, the power optimization ALGORITHM 1, 2: TRANSCEIVER OPTIMIZATION FOR SUM-RATE MAXIMIZATION
problem can be formulated as a geometric program (GP) (2) AND MAX-MIN USER RATE (13)

(9)

The global optimum of (9) can be found by existing optimiza-

tion tools, e.g., MOSEK [40] or GGPLAB [41]. An overview
on geometric programming is given in Appendix A.
For the optimal allocation and transmit filter
, the MSE of each layer can be minimized independently by
the MMSE receive filters (4). This increases the weighted sum
rate (5).
2) Optimization in the Downlink Channel: Similar results as
for the uplink channel also hold for the downlink channel. That
is, for fixed transmit and receive matrices, the downlink power
optimization can be formulated as a GP as well. For a given
power allocation and transmit filter , the optimal receiver
of each user is the MMSE receiver In addition, when switching from the downlink channel to the
uplink channel, the powers are computed to achieve the same
MSE values as in the downlink channel (see Appendix B)
(10)
3) Joint Transceiver Optimization: When we consider a (12)
joint optimization of the transceivers, one possible strategy
The receivers are updated as the MMSE receivers (4). The pur-
is to switch the optimization between the uplink channel and
pose of the power allocation step (12) is to update the uplink
the downlink channel. The MSE duality [26], [27] facilitates
MMSE receiver (4). Based on this update, a new uplink power
the computation of the downlink transmit filter via a “virtual”
allocation is computed by geometric programming. From simu-
uplink channel. Likewise, uplink transmitters can be found via
lations, we observe that this additional receiver update reduces
a “virtual” downlink channel.
the overall number of iteration steps.
The described optimization steps can be carried out in an
The detailed algorithm (Algorithm 1) is summarized in
alternating manner. There are many possible orders in which
Table I. Superscript denotes the th iteration.
we can optimize transmitters, receivers and powers. In the fol-
lowing, we will focus on a particular algorithm for a downlink
B. P2: Per-User Rate Balancing
channel, where transmitters and powers are computed via a “vir-
tual” uplink channel. The problem of maximizing the minimum user rate can be
Specifically, we start with the optimization in the uplink. written as
Power allocation is performed with GP (9) and the receivers
are updated with (4). Intuitively, before proceeding the opti-
mization in the downlink, these two optimization steps (power
allocation + MMSE receiver) should be repeated until they (13)
converge. Note that the geometric programming step (9) has a
higher computational complexity compared with other steps in where is the index set containing the layer indices associated
the algorithm. We observe a lower overall complexity, if these with the user .
two steps are only performed once in each iteration. Different uplink power allocation strategies influence the
Then, we switch the optimization in the downlink channel. power distribution among layers according to the different de-
As it has been mentioned, we first ensure the same performance sign goals under consideration. Therefore, the key step in each
as in the uplink channel by MSE duality. Assume that the iteration is the specific uplink power strategy. The remaining
achieved uplink MSE is . The down-
steps in Algorithm 1 will be the same for different problems.
link power allocation achieving the given targets is given by
In the following, we specify the power strategy for P2. We
(see Appendix B)
introduce to denote a certain rate target and as-
(11) sume that the minimum achieved user rate is
, or equivalently, , . Due to
Since the power allocation with GP has high complexity, in each , , to
iteration we only perform power allocation with (11) and then ensure all users achieve the rate value , we can con-
update the receivers with (10). strain , , i.e., , .
4024 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 8, AUGUST 2008

Additionally, with fixed transmit and receive filters in the uplink iteration, the powers allocated to data streams are distributed in
the geometric MSE an uneven fashion. Even though the optimization begins with
a fixed number of data streams, it is possible that the power
of one data stream approaches zero. This will happen automat-
ically when the initially given number of data streams is not
is a posynomial as well. Thus, we can formulate the uplink favored by the spatial structure of the channel. Simulations in
power optimization for (13) as a GP Section VI show that the sum rate or user rate contributed by
the chosen active layers is very close to the maximum.

(14)
D. Convergence
By exchanging the uplink power allocation (9) in Algorithm
1 with (14), we obtain a new algorithm for P2 (see Algorithm 2 Proof of the Convergence of Algorithm 1: Exploiting the
MSE duality [26], we can derive the following (in)equalities:
in Table I).

C. Comments on the Number of Active Data Streams

Beamforming, i.e., one data stream per user ( , ),
is the optimal transmit strategy for line-of-sight (LoS) channels.
This is the consequence of the rank one channel matrix of the
LoS channel. In this case, only one data stream can be supported.
For multipath channels, the throughput can be increased by spa-
tially multiplexing several data streams per user. However, the
number of supportable data streams depends on the structure of
the channel, in particular on the rank of the channel matrices.
But even with full-rank matrices, it can be advisable to switch
off one data stream and to focus the available power on the re-
maining streams. This not only depends on the structure of the The notation refers to the uplink layer MSE
channel matrices, but also on the chosen design goal and the achieved at step 5.a in the th iteration in Table I. The equal-
power constraints. ities hold due to the MSE duality (see Appendix B) and the
The proposed algorithms are able to deal with this effect, first three inequalities due to the uplink or downlink MMSE
since the power allocation adjusts iteratively to the given filtering. The optimal power allocation by GP (9) leads to the
channel. Although the chosen MMSE approach does not permit last inequality. Note that the decrease of the weighted geometric
powers to be set to zero exactly, powers can be arbitrarily close MMSE values implies the increase of the weighted sum rates.
Thus the weighted sum rate is monotonically increasing in each
to zero, which means that the associated data streams cannot
iteration.
be supported. In this way, the algorithm adapts to the channel
Proof of the Convergence of Algorithm 2: In the th iteration,
conditions. There are no constraints on the number of users, we have the following (in)equalities:
antennas, or data streams. A special case is the aforementioned
beamforming mode, which is obtained for LoS channels.
Very small powers are no problem from a practical point of
view, since streams below a certain threshold can be removed
after convergence. But in the course of the iteration, powers
are strictly positive. This is the consequence of the property of
the GP. Actually, the MMSE is only defined for existing data
streams. If a stream is switched off, then a common approach is
to set the MMSE of this stream to one. However, this exception
has not been included in this paper. Instead we assume strictly
positive powers, which simplifies the discussion and notation. The equalities hold due to the MSE duality and the first three
If a power becomes very small, then the MMSE of this stream inequalities due to uplink or downlink MMSE filtering. The last
inequality is due to optimal power allocation by GP (14). There-
will approach 1, and its rate zero.
fore, the individual rate per user is monotonically increasing in
Both power allocation step [(9) or (15)] and the receiver op- each iteration.
timization step result in a certain SINR for each data stream. The above proofs show that Algorithm 1 and Algorithm 2
The transmitter is optimized with respect to a dual power allo- converge. However, due to the nonconvexity of the problems,
cation (11) associated with these SINRs. If one data stream has different initializations of , , can lead to different limit
a small SINR, then also its virtual power will be small. In this points. Therefore, choosing a suitable initialization is impor-
way, receivers and transmitters depend on each other. During the tant. How to choose an initialization which ensures the global
SHI et al.: MULTIUSER MIMO SYSTEMS WITH LINEAR PROCESSING 4025

optimum is an interesting problem for future work. The rec- downlink, respectively, we can use the same optimization frame-
ommended initialization of is the singular vectors of the work for computing the transmit filters, powers, and receive fil-
channel matrix or an identity matrix. ters for problems P1–P5. We also assume an arbitrary fixed DPC
precoding order to 1 in the downlink, and an SIC decoding
order 1 to in the uplink.
IV. VARIATIONS OF THE PROBLEMS
Now, consider the uplink channel. The interference caused
In this section, we consider other variations of the problems, by the previous layers is subtracted from the th layer.
for which the uplink power optimization problems can be for- Thus, the MMSE filter for the th layer only needs to suppress
mulated as GPs as well. Thus, the global optimum power alloca- the interference from the remaining layers. The uplink
tion can be found for fixed transmitters and receivers. We only MMSE filter with SIC is
specify the uplink power allocations for the following problems:
P3: sum-rate constrained power minimization;
P4: user-rate constrained power minimization;
P5: sum-rate maximization subject to individual user-rate
targets.
For the problem of sum-rate constrained power minimization (18)
P3, the total power has to be minimized for the given sum-rate Similarly, in the downlink the first layer sees no interference and
target the second layer sees interference from the first layer, and so on.
Therefore, the MMSE filter for the th layer is

(19)
(15) As a consequence of the uplink SIC/downlink DPC, a triangular
structure is imposed on the effective channel. The resulting cou-
It is clear that (15) is a GP problem as well. pling matrix is , where denotes the upper
For problem P4, the total transmit power is minimized to triangular part of a matrix.
achieve individual user-rate constraints, i.e. Exchanging with , the uplink filters (4) with (18), and
the downlink filters (10) with (19), we obtain new algorithms for
P1–P5 with SIC/DPC. Following a similar reasoning as for the
linear case, we can show that the new algorithms converge as
well. For convenience, we denote the algorithm with SIC/DPC
for P1 and P2 as Algorithm 1-DPC and Algorithm 2-DPC, re-
(16) spectively.
where is the rate target for the th user. This is a GP as well.
The GP formulation of the power optimization for P5 is given VI. SIMULATION RESULTS
by
Now we compare our algorithms with other schemes, which
are derived either based on the result [15], or by combining the
result [35] with our alternating optimization framework. A short
description of these schemes are given as follows.
1) TxNullSVD-SumRate/RateBalancing: Based on
nullspace-directed SVD [15], the transmit filters, channel
and receive filters together form an effective diagonal channel.
(17)
It has been shown in [15] that this approach outperforms other
zero-forcing approaches published in [11]–[14]. Here, we take
With power allocations obtained by solving (15), (16), and it as a performance benchmark for the category of orthogonal
(17), new algorithms for P3–P5 are at hand. Detailed algo- processing.
rithms are omitted here. The monotonic convergence of the An additional power allocation over the parallelized channels
algorithms can be shown by a similar reasoning as the proofs can be easily performed, either to maximize the sum rate (called
in Section III-D. TxNullSVD-SumRate) or to balance the rates among all users
(called TxNullSVD-RateBalancing).
V. EXTENSION: OPTIMIZATION WITH SIC/DPC 2) GPSINR-SumRate: This scheme follows the alternating
approach Algorithm 1 and 2, with a different power allocation
In the uplink channel, known interference can be partially proposed in [35]. Specifically, for high SINRs, we can approx-
canceled by SIC. The downlink counterpart is known as DPC. imate . With fixed
DPC can partially compensate interference by appropriately en- transmit and receiver filters, the power allocation for sum-rate
coding the signals. Assuming that ideal SIC and DPC (no error optimization can be formulated as a GP problem with respect to
propagation and precoding loss) are applied to the uplink and , , (see [35] for more information), i.e.
4026 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 8, AUGUST 2008

Fig. 3. Average sum rate versus P = .  =1 ,K =2 ,N =4 Fig. 4. Average balanced rate level versus P = .  =1 ,K =2 ,
N =2 , 8k . Algorithm 1 (with linear processing) outperforms other existing
,
N =4 ,N =2 , 8k . Algorithm 2 (with linear processing) outperforms
linear schemes and Algorithm 1-DPC achieves the global optimum. TxNullSVD-RateBalancing and the performance of Algorithm 2-DPC is very
close to the global optimum.

(20)

where , . We can exchange the power

allocation step 5.a) in Table I with (20) to obtain another itera-
tive algorithm, which we call GPSINR-SumRate. Note that due
to the high-SNR approximation, it is difficult to prove that the
power allocation returned by (20) indeed improves the sum rate
as compared with that in the previous iteration. Thus, it is un-
clear whether this algorithm converges monotonically. Fig. 5. Average sum rate & average balanced rate level versus number of iter-
The performance improvement over above schemes is shown ations. P = = 10 dB , =1,K =2,N =4 ,N =2 , 8k .
in Figs. 3 and 4. We consider a two-user downlink MIMO
channel, with 4 transmit antennas at the base station and 2
receive antennas at each mobile. We assume an independently rates are averaged over 1000 channel realizations. With non-
and identically distributed Rayleigh fading channel model. The linear processing, the performance is improved over the case
performance measures, the sum rate (Fig. 3) and the balanced with linear processing, which is expected.
rate level (Fig. 4), are averaged over 1000 channel realizations. The required numbers of iterations to achieve accuracy
It can be observed in Fig. 3 that our proposed Algorithm 1 or are listed in Table II.
outperforms TxNullSVD-SumRate and GPSINR-SumRate. For Algorithm 1 and Algorithm 1-DPC is defined as
Interestingly, Algorithm 1-DPC with DPC achieves the channel and for Algorithm 2 and Algorithm 2-DPC
capacity. From Fig. 4, we can see that the proposed Algorithm
2 achieves a higher balanced level than TxNullSVD-RateBal- as . The total power is
ancing. The algorithm with DPC (a simple ordering strategy assumed to be 10 and noise power be 1. Two scenarios are
is used, where the users are ordered according to their channel considered. One is 2-user MIMO, with 4 transmit antennas and
norms), has a performance, which is very close to the global each user with 2 receive antennas, denoted as (4,2,2), the other
optimum. The global optimum is denoted by the line with one is 3-user MIMO, with 6 transmit antennas and each user
asterisks. The optimal values of sum rates and balanced user with 2 receive antennas, denoted as (6,3,2). Two data streams
rates are obtained by the algorithms proposed in [6] and [9], are transmitted per user. Generally, with DPC, the required
respectively. numbers of iterations is less than without DPC. With linear pro-
Fig. 5 shows the monotonic convergence behavior of the pro- cessing, transmitting more data streams increases the number
posed Algorithm 1 and Algorithm 2. The total power constraint of iterations. The GP in Algorithm 1 or Algorithm 1-DPC
is 10 and noise power is 1. The sum rates and balanced user has higher order of degree of difficulty (DoD) [42] (see also
SHI et al.: MULTIUSER MIMO SYSTEMS WITH LINEAR PROCESSING 4027

TABLE II the new power allocation strategies, which are carried out by
NUMBER OF ITERATIONS AND DEGREE OF DIFFICULTY FOR DIFFERENT optimizing the product of layer-MSEs together with MMSE
0
CHANNEL SETUPS.  < 1E 3=1E 5, P 0 = 10,  = 1, SETUP 1
(4,2,2):K = 2, N = 4, N 8
= 2, k ; SETUP 2 (6,3,2): K = 3, N = 6,
filtering. Monotonic convergence of the iteration sis proven by
N = 2, k 8 exploiting properties of this new power control strategy. It is
shown that the proposed algorithms outperform other existing
linear schemes.

APPENDIX A
GEOMETRIC PROGRAMMING
Let a vector with real positive variables
, . A real valued function of , with the form

where and , , are real numbers, is called a monomial

function, or a monomial (of the variables ). A sum of
one or more monomials, i.e.

where , is called a posynomial function, or simply, a

posynomial (with terms, in variables ).
A geometric program (GP) is an optimization problem of the
form

Fig. 6. Rate and MSE comparison of different schemes. Scheme 1: sum-rate

(21)
maximization; Scheme 2: max-min user rate; Scheme 3: sum MSE minimiza-
tion; Scheme 4: min-max user MSE. P = = 10 dB,  = 1, K = 2, where , , are posynomials:
N = 4, N 8
= 2, k .
, , , are monomials:
, and are the optimization
Appendix A) than Algorithm 2 or Algorithm 2-DPC. Degree of variables. There is an implicit constraint that the variables are
difficulty is defined as the difference between the total number positive. Problem (21) is referred to as a GP in posynomial
of the monomial terms in the objective and constraints, and one form.
plus the number of variables. Via a logarithmic change of variables: ,
In Fig. 6, we illustrate the achieved sum rates/user rates and , , and a logarithmic transformation of the
sum MSEs/user MSEs for a randomly chosen channel realiza- objective and constraint functions, the GP (21) becomes
tion with difference schemes: Scheme 1: sum-rate maximiza-
tion (Algorithm 1); Scheme 2: max-min user rate (Algorithm 2);
Scheme 3: sum-MSE minimization [26]; Scheme 4: min-max
user MSE [29]. The total power is 10 and noise power is 1. For
each scheme, there are three bars. The first and second bar de-
note the user rates or user MSEs, and the third bar denotes the
sum rate or sum MSE. It can be observed that Scheme 1 achieves
(22)
a higher sum rate, whereas Scheme 3 achieves a smaller sum
MSE than others. Scheme 2 balances the user rates and Scheme where optimization variables and
4 balances the user MSEs. However, both schemes are neither .
optimal in terms of sum rate nor in terms of sum MSE.
A GP in form (21) is not a convex optimization problem, be-
cause posynomials are not convex functions, whereas (22) is
VII. CONCLUSION convex. We refer to (22) as a GP in convex form. Consider a
We study the problem of transceiver design for a multiuser special case where all the posynomials are simply monomials.
MIMO systems with linear transmitters and receivers. We Then this GP in convex form reduces to a linear program (LP).
focus on two different rate optimization problems P1 and P2 Hence, geometric programming can be considered to be a gen-
and show that the results can be extended to other variations of eralization of linear programming.
the problems. These problems are nonconvex, and no efficient As we have known that GP can be transformed into a convex
solutions are available. Here, we propose iterative algorithms problem. Therefore, a local optimum for GP is also a global op-
based on the MSE duality and a heuristic alternating opti- timum, and the duality gap is zero under mild technical con-
mization framework. The crucial steps in the iterations are ditions. Detailed discussion on numerical methods for GP can
4028 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 8, AUGUST 2008

be found in, e.g., [43], [44]. Two major approaches to solve a and
GP are the interior-point method in [45], and the infeasible in-
terior-point algorithm in [46]. User-friendly software packages
for GP are available publicly, such as the MOSEK package [40]
and the GGPLAB toolbox [41].
If , then there exists a strictly
To measure how “difficult” a GP is, the degree of difficulty
positive power allocation
was introduced by Duffin, Peterson and Zener in [42]. The de-
gree of difficulty is defined as the difference between the total (25)
number of the monomial terms in the objective and constraints,
and one plus the number of variables. However, in [46], the au- such that , . This is an immediate consequence of
thors reported that the proposed infeasible interior-point algo- the convergence properties of the Neumann series in which (25)
rithm for solving primal and dual GP is effective regardless of can be decomposed. Conversely, by a similar reasoning as in the
the degree of difficulty. context of SINR [50], we know that if there exists a such
The GP can be solved very efficiently by standard interior- that , , then holds.
point methods with a worst case polynomial-time complexity. In the same way it can be shown that downlink targets , ,
Roughly speaking, standard interior-point methods can solve can be achieved if and only if . If
a GP with 1000 variables and 10 000 constraints in under a
the targets are feasible, then they can be achieved by a strictly
minute, on a small desktop computer [44]. The detailed discus-
positive power allocation
sion on computational complexity via self-concordant functions
and parameter choices can be found in [47]. The upper bound on
the total number of Newton steps in the barrier method [47] does
not depend on the dimension of the variables, or the number of Since , we
equality constraints, or the particular values of the problem data, know that this implies that exists, if and only if
i.e., the objective and constraint functions. This upper bound exists.
shows that the barrier method converges at least linearly. By ap- Both allocations have the same total power which can be ver-
propriately choosing the parameters, the bound on the number ified by
of Newton steps can grow as , instead of [47].
The infeasible algorithm [46], which solves both the primal
and dual GP simultaneously, is reported to produce very com-
petitive numerical efficiency for a wide range of GPs. This algo-
rithm is tested on most challenging 19 GP problems and shown
to be faster than the earlier methods [46].
It has been also known that the GP is sensitive to perturbation
procedures that may for example arise in solution methods. A
careless perturbation may eliminate dual optimal solutions and
i.e., the same feasible layer-MSE values can be achieved in both
make it impossible to solve the original problem [48]. A suc-
links with the same total transmit power .
cessful attempt at developing a stable perturbation method ap-
pears in [49]. Since duality holds layer-wise, it obviously holds user-wise, and
also for the total MSE (sum of all layers).
APPENDIX B
REFERENCES
MSE DUALITY
[1] S. Vishwanath, N. Jindal, and A. Goldsmith, “Duality, achievable rates,
Given , , and a total power limit , the same MSE and sum-rate capacity of Gaussian MIMO broadcast channels,” IEEE
Trans. Inf. Theory, vol. 49, no. 10, pp. 2658–2668, Oct. 2003.
values can be achieved in the downlink channel [2] W. Yu and J. Cioffi, “Sum capacity of Gaussian vector broadcast chan-
(Fig. 1) and uplink channel (Fig. 2). nels,” IEEE Trans. Inf. Theory, vol. 50, no. 9, pp. 1875–1892, Sep.
With matrices (7) and (8), we can rewrite the downlink 2004.
[3] H. Weingarten, Y. Steinberg, and S. Shamai(Shitz), “The capacity re-
MSE and uplink MSE as gion of the Gaussian MIMO broadcast channel,” in Proc. CISS’04 ,
Princeton University, NJ, Mar. 2004.
[4] P. Viswanath and D. Tse, “Sum capacity of the vector Gaussian broad-
(23) cast channel and uplink-downlink duality,” IEEE Trans. Inf. Theory,
vol. 49, no. 8, pp. 1912–1921, Aug. 2003.
[5] W. Yu, W. Rhee, S. Boyd, and J. M. Cioffi, “Iterative water-filling for
and Gaussian vector multiple access channels,” IEEE Trans. Inf. Theory,
vol. 50, no. 1, pp. 145–151, Jan. 2004.
[6] N. Jindal, W. Rhee, S. Vishwanath, S. A. Jafar, and A. Goldsmith,
(24) “Sum power iterative water-filling for multi-antenna Gaussian broad-
cast channels,” IEEE Trans. Inf. Theory, vol. 51, no. 4, pp. 1570–1580,
Apr. 2005.
respectively. [7] T. Michel and G. Wunder, “Sum rate iterative water-filling for Gaussian
MIMO broadcast channels,” in Proc. WPMC, San Diego, CA, 2006.
Collecting all layer-MSEs in a diagonal matrix [8] J. Oh, S. J. Kim, R. Narasimhan, and J. Cioffi, “Transmit power opti-
, we can rewrite (23) and (24) as mization for Gaussian vector broadcast channels,” in Proc. ICC, Seoul,
Korea, May 2005.
[9] J. Lee and N. Jindal, “Symmetric capacity of MIMO downlink chan-
nels,” in Proc. ISIT, 2006.
SHI et al.: MULTIUSER MIMO SYSTEMS WITH LINEAR PROCESSING 4029

[10] C. F. Fung, W. Yu, and T. J. Lim, “Precoding for the multi-antenna [35] M. Chiang, C. W. Tan, D. Palomar, D. O’Neill, and D. Julian, “Power
downlink: Multi-user gap approximation and optimal user ordering,” control by geometric programming,” IEEE Trans. Wireless Commun.,
IEEE Trans. Commun., vol. 55, no. 1, pp. 188–197, Jan. 2007. Jul. 2007.
[11] M. Rim, “Multiuser downlink beamforming with multiple transmit and [36] J. R. Barry, E. A. Lee, and D. G. Messerschmitt, Digital Communica-
receive anteanns,” Electron. Lett., vol. 38, pp. 1725–1726, Dec. 2002. tions, 3rd ed. Norwell, MA: Kluwer, 2004.
[12] R. L. U. Choi and R. D. Murch, “A transmit pre-processing technique [37] T. Haustein and H. Boche, “Optimal power allocation for MSE and
for multiuser MIMO systems: A decomposition approach,” IEEE bit-loading in MIMO systems and the impact of correaltion,” in Proc.
Trans. Wireless Commun., vol. 3, pp. 20–24, Jan. 2003. ICASSP’03, Hong Kong, China, Apr. 2003.
[13] K. K. Wong, R. D. Murch, and K. B. Letaief, “A joint-channel diago- [38] T. Haustein, H. Boche, and G. Lehmann, “Bit-loading for the SIMO
nalization for multiuser MIMO antenna system,” IEEE Trans. Wireless multiple access channel,” in Proc. PIMRC’03, Berlin, Germany, Sep.
Commun., vol. 2, pp. 773–786, Jul. 2003. 2003.
[14] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing [39] M. H. Hayes, Statistical Digital Signal Processing and Modeling.
methods for downlink spatial multiplexing in multiuser MIMO chan- New York: Wiley, 1996.
nels,” IEEE Trans. Signal Process., vol. 52, no. 2, pp. 461–471, Feb. [40] MOSEK: Solutions Through Mathematical Optimization, [Online].
2004. Available: http://www.mosek.com.
[15] Z. Pan, K. Wong, and T. Ng, “Generalized multiuser orthogonal space- [41] GGPLAB: A Simple MATLAB Toolbox for Geometric Programming,
divison multiplexing,” IEEE Trans. Wireless Commun., vol. 3, no. 6, [Online]. Available: http://www.stanford.edu/ boyd/ggplab/
pp. 1969–1973, Nov. 2004. [42] R. J. Duffin, E. L. Peterson, and C. Zener, Geometric Programming:
[16] M. Stojnic, H. Vikalo, and B. Hassibi, “Maximizing the sum-rate of Theory and Applications. New York: Wiley, 1967.
multi-antenna broadcast channels using linear preprocessing,” IEEE [43] M. Chiang, “Geometric programming for communication systems,”
Trans. Wireless Commun., vol. 5, no. 9, pp. 2338–2342, Sep. 2006. Foundations and Trends in Communications and Information Theory,
[17] J. Zander and M. Frodigh, “Comment on performance of optimum vol. 2, no. 1–2, pp. 1–154, Aug. 2005.
transmitter power control in cellular radio systems,” IEEE Trans. Veh. [44] S. Boyd, S. J. Kim, L. Vandenberghe, and A. Hassibi, “A tutorial on
Technol., vol. 43, no. 3, p. 636, Aug. 1994. geometric programming,” Optimiz. Eng., vol. 8, no. 1, pp. 67–127,
[18] F. Rashid-Farrokhi, K. J. Liu, and L. Tassiulas, “Transmit beamforming 2007.
and power control for cellular wireless systems,” IEEE J. Sel. Areas [45] Y. Nesterov and A. Nemirovskii, Interior Point Polynomial Algorithms
Commun., vol. 16, no. 8, pp. 1437–1449, Oct. 1998. in Convex Programming. Philadelphia, PA: SIAM, 1994.
[19] M. Schubert and H. Boche, “Solution of the multi-user downlink beam- [46] K. O. Kortanek, X. Xu, and Y. Ye, “An infeasible interior-point al-
forming problem with individual SINR constraints,” IEEE Trans. Veh. gorithm for solving primal and dual geometric programs,” Math. Pro-
Technol., vol. 53, no. 1, pp. 18–28, Jan. 2004. gramm., vol. 76, pp. 155–181, 1996.
[20] A. Wiesel, Y. C. Eldar, and S. Shamai, “Linear precoding via conic [47] S. Boyd and L. Vandenberghe, Convex Optimization. New York:
optimization for fixed MIMO receivers,” IEEE Trans. Signal Process., Cambridge Univ. Press, 2004.
vol. 54, no. 1, pp. 161–176, Jan. 2006. [48] S. C. Fang, E. L. Peterson, and J. R. Rajasekera, “Controlled dual
[21] W. Yu and T. Lan, “Transmitter optimization for the multi-antenna perturbations for central for posynomial programs,” Europ. J. Operat.
downlink with per-antenna power constraints,” IEEE Trans. Signal Res., vol. 35, pp. 111–117, 1988.
Process., Jun. 2007. [49] J. Zhu, K. O. Kortanek, and S. Huang, “Controlled dual perturbations
[22] G. Caire and S. Shamai (Shitz), “On the achievable throughput of a for central path trajectories in goemetric programming,” Europ. J. Op-
multi-antenna Gaussian broadcast channel,” IEEE Trans. Inf. Theory, erat. Res., vol. 73, pp. 524–531, Mar. 1992.
vol. 49, no. 7, pp. 1691–1706, Jul. 2003. [50] H. Boche and M. Schubert, “A general duality theory for uplink and
[23] W. Yu and J. M. Cioffi, “Sum capacity of Gaussian vector broadcast downlink beamforming,” in Proc. VTC, Fall, Vancouver, Canada, Sep.
channels,” IEEE Trans. Inf. Theory, vol. 50, no. 9, pp. 1875–1892, 2002.
2004.
[24] H. Weingarten, Y. Steinberg, and S. Shamai (Shitz), “The capacity
region of the Gaussian MIMO broadcast channel,” IEEE Trans. Inf.
Theory, 2004.
[25] S. Shi and M. Schubert, “MMSE transmit optimization for multiuser Shuying Shi (S’05) received the M.Sc. degree in
multi-antenna systems,” in Proc. ICASSP, Philadelphia, PA, Mar. 2005. electrical engineering from the Technische Univer-
[26] S. Shi, M. Schubert, and H. Boche, “Downlink MMSE transceiver op- sität Kaiserslautern, Germany, in 2002.
timization for multiuser MIMO systems: Duality and sum-MSE mini- She is currently pursuing the Ph.D. degree in
mization,” IEEE Trans. Signal Process., vol. 55, no. 11, pp. 5436–5446, electrical engineering. In 2003, she joined the Fraun-
Nov. 2007. hofer German-Sino Lab for Mobile Communications
[27] A. Mezghani, M. Joham, R. Hunger, and W. Utschick, “Transceiver (MCI). Since 2006, she has been with Technische
design for mutli-user MIMO systems,” in Proc. Int. ITG/IEEE WSA Universität Berlin, Germany. Her research focuses
2006, Ulm, Germany, Mar. 2006.
on optimal transmit and receive strategies for mul-
[28] M. Schubert, S. Shi, and H. Boche, “Iterative transceiver optimiza-
tion for linear multiuser MIMO channels with per-user MMSE require- tiuser MIMO systems.
ments,” in Proc. 14th Europ. Signal Process. Conf., Italy, Sep. 2006.
[29] S. Shi, M. Schubert, and H. Boche, “Transceiver optimization for multi-
user MIMO systems: Min-max relative user-MSE under total power
constraint,” in Proc. ISITA, Seoul, Korea, Oct. 2006.
[30] S. Shi, M. Schubert, and H. Boche, “Computational efficient trans- Martin Schubert (M’06) received the M.Sc. and
ceiver optimization for multiuser MIMO systems: Power minimization Ph.D. degrees in electrical engineering from the
with user-MMSE requirements,” in Proc. Asilomar CSSC, CA, Oct. Technische Universität Berlin, Germany, in 1998
2006. and 2002, respectively.
[31] E. Jorswieck and H. Boche, “Transmission strategies for the MIMO In 1998, he joined the Heinrich-Hertz Institute for
MAC with MMSE receiver: Average MSE optimization and achievable Telecommunications (HHI), Berlin, as a Research
individual MSE region,” IEEE Trans. Signal Process., Special Issue on Assistant. Since 2003, he has been with the Fraun-
MIMO Wireless Commun., vol. 51, no. 11, pp. 2872–2881, Nov. 2003. hofer German-Sino Lab for Mobile Communications
[32] S. Serbetli and A. Yener, “Transceiver optimization for multiuser (MCI), where he is working as a Senior Researcher
MIMO systems,” IEEE Trans. Signal Process., vol. 52, no. 1, pp.
and Project Leader. He is also a lecturer with the
214–226, Jan. 2004.
[33] Z. Luo, T. N. Davidson, G. B. Giannakis, and K. Wong, “Transceiver Technical University of Berlin. His research interests
optimization for block-based multiple access through ISI channels,” include multiantenna signal processing, resource allocation, and multiuser
IEEE Trans. Signal Process., vol. 52, no. 4, pp. 1037–1052, Apr. 2004. communication theory.
[34] M. Schubert and H. Boche, “Throughtput maximization for uplink and Dr. Schubert was a corecipient of the VDE Johann-Philipp-Reis-Award in
downlink beamforming with independent coding,” in Proc. CISS’03. 2007. He coauthored the 2007 Best Paper Award of the IEEE Signal Processing
Baltimore, MD: Johns Hopkins Univ., 2003. Society.
4030 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 8, AUGUST 2008

Holger Boche (M’04–SM’07) received the M.Sc. German-Sino Lab for Mobile Communications, and since 2005, has been Di-
and Ph.D. degrees in electrical engineering from rector of the Fraunhofer Institute for Telecommunications, Berlin. He was vis-
the Technische Universität Dresden, Germany, in iting professor at the ETH Zurich during winter term 2004 and 2006 and with
1990 and 1994, respectively. In 1998, he received KTH Stockholm during summer term 2005.
the Ph.D. degree in pure mathematics from the Prof. Boche received the Research Award “Technische Kommunikation”
Technische Universität Berlin, Germany. from the Alcatel SEL Foundation in October 2003, the “Innovation Award”
From 1994 to 1997, he did postgraduate studies from the Vodafone Foundation in June 2006, and the Gottfried Wilhelm
in mathematics with the Friedrich-Schiller Univer- Leibniz Prize 2008 from the DFG. He was a corecipient of the 2006 IEEE
sität Jena, Germany. In 1997, he joined the Heinrich- Signal Processing Society Best Paper Award and the recipient of the 2007 IEEE
Hertz-Institut (HHI) für Nachrichtentechnik Berlin. Signal Processing Society Best Paper Award. He is currently an Associate
He is head of the Broadband Mobile Communication Editor of the IEEE TRANSACTIONS ON INFORMATION THEORY. He is a member
Networks Department, HHI. Since 2002, he has been a Full Professor for Mo- of IEEE Signal Processing Society SPCOM and SPTM Technical Committee.
bile Communication Networks, Institute for Communications Systems, Tech-
nische Universität Berlin. Since 2003, he has been Director of the Fraunhofer