IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO.

4, APRIL 2006

743

Transactions Papers System Evaluation of Optimal Downlink Beamforming with Congestion Control in Wireless Communication
Rickard Stridh, Mats Bengtsson, and Bj rn Ottersten o

Abstract We investigate the use of congestion control and joint optimal downlink beamforming, power control, and access point allocation, in a multi-cell wireless communication system. The access points of the system employ smart antennas and single antennas are used at the terminals. The possibility to send messages to multiple terminals at the same frequency in the same time slot is exploited. We show how previously proposed algorithms for optimal downlink beamforming easily can be extended to determine also the optimal access point for each mobile terminal. In order to assign resources, optimal beamforming requires a feasible set of mobiles, i.e. that all admitted users can be offered the required Signal-to-Interference-and-Noise Ratio. Therefore, an algorithm for deciding which mobile terminals to admit or reject from a congested system is proposed and evaluated. Using the proposed congestion algorithm, joint optimal downlink beamforming is evaluated and the throughput increase as compared to decentralized beamforming algorithms and other congestion control strategies is assessed from a system point of view. The results show that the proposed strategy can almost double the throughput compared to decentralized beamforming algorithms and give a vefold increase in throughput compared to conventional beamforming without any interference suppression. Index Terms Mobile communication, communication systems, antenna arrays, array signal processing, transmitting antennas, power control, resource management, optimization methods.

capacity requirements as an alternative to, for example, exploiting new frequency bands. Smart antennas allow the reduction of co-channel interference which may be exploited by decreasing the reuse distance for frequency channels. Also, since different mobile terminals in the system will, in general, have different spatial signatures as seen from the access point, each access point may be able to communicate with several mobile terminals, using the same time and frequency resource. Such access schemes are denoted Space Division Multiple Access (SDMA). This paper considers the downlink of such an SDMA system. Downlink communication is treated since data trafc is expected to be asymmetric leading to a capacity bottleneck in the communication from the access points to the terminals. The issue of downlink beamforming has been investigated in e.g. [1] and [2]. A commonly used optimality criterion for downlink beamforming is to minimize the total transmitted power, given that all users have acceptable Quality of Service (QoS), as rst formulated in [3]. The idea of minimizing the total transmitted power is to keep the total interference level in the system as low as possible. Of course, other optimality criteria exist. The optimization problem above, can be solved using the algorithms in [4][6]. These algorithms have been extended to perform joint optimal power control, beamforming and access point allocation in [7]. In [8], the solutions in [5], [7] for joint optimal power control and beamforming with and without access point assignment are assessed from a system perspective. These results are reviewed and extended in this paper. The area of power control in combination with congestion control by stepwise addition or removal of mobile terminals, is also investigated in e.g. [9] for single antenna base stations and [10], [11] for antenna arrays. The resource allocation algorithm herein will require global knowledge of channel statistics and may therefore mainly be interesting for use in system simulations in search for the fundamental performance limits of a wireless system. However, for systems of limited size and in environments with low mobility (e.g indoor systems), it may be possible to implement centralized beamforming based on knowledge

I. I NTRODUCTION

IRELESS communications is currently in a transition from carrying only voice communication to providing both voice and data communication resulting in increasing capacity requirements, especially on the downlink. Smart antennas have been proposed in order to meet the increased

Manuscript received November 15, 2002; revised February 13, 2004, November 17, 2004, and February 3, 2005; accepted February 22, 2005. The associate editor coordinating the review of this paper and approving it for publication was V. Leung. This work is supported by the Swedish Foundation for Strategic Research through the Personal Computing and Communication program. The authors are with the Department of Signals, Sensors & Systems, Royal Institute of Technology, SE100 44 Stockholm, Sweden (e-mail: [email protected]). Digital Object Identier 10.1109/TWC.2006.04008.

1536-1276/06$20.00 c 2006 IEEE

744

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 4, APRIL 2006

of global channel statistics in a practical system. As stated above, several algorithms for solving the optimal downlink beamforming problem have been presented. However, in a congested system, it may happen that no solution exists, since too many terminals want to communicate at the same time. Then, some mobiles have to be refused access to the system to allow for a feasible solution. This introduces the problem of congestion control, i.e., how to admit or remove mobiles in the system. In this work we suggest a heuristic algorithm that will reach a good feasible set of mobiles to optimize over. The algorithm aims at removing or not admitting, the specic mobile that is most critical to the system, in terms of the possibility to achieve the Qualityof-Service requirement, here dened as the target Signal-toInterference-and-Noise Ratio (SINR) for all remaining mobile terminals. This strategy can be extended to a full scheduling algorithm. However, considering all the aspects of scheduling is outside the scope of this article. In addition, a system assessment of joint optimal beamforming, power control, and access point assignment, is performed in the context of cellular networks taking into account trafc and radio resource management issues. Results are given in terms of system throughput as a function of the trafc load in the system. The results are also compared to previously presented beamforming techniques. The paper is organized as follows: First the system and signal assumptions are reviewed, followed by the algorithms used for beamforming, power control, and access point assignment. In order to explain the congestion control algorithm, a couple of general optimization theory results are presented, followed by the actual algorithm. Finally, some simulation results and a summary conclude the paper. II. S YSTEM A SSUMPTIONS A cellular system, with several terminals sharing the same frequency channel is investigated. Each access point communicates with several terminals. We assume a controlled environment where all mobile terminals request a channel resource of a certain quality, which may be different for different users. When connecting, the terminals are allocated to an appropriate access point. It is assumed that the access points are equipped with array antennas with nt elements and that the channel is a low rank (at fading) narrowband random channel modeled with local independent scattering at the terminal. However, the techniques can be easily be extended to wideband systems, see [2], [4]. It is further assumed that the second-order statistics of the vector channels between all transmitters and receivers are known. In a Frequency Division Duplex system, the uplink and downlink channels will typically fade independently which means that the instantaneous downlink channel can in general not be estimated from uplink measurements. However, it may still be reasonable to estimate the second-order statistics of the fast fading of the downlink channel, based on uplink measurements averaged over the fast fading. Several methods have been proposed to compensate for the frequency offset between the uplink and downlink channels, see for example [2], and will not be described further here.

The user requirement on QoS is assumed to be fullled if the Signal-to-Interference-and-Noise Ratio (SINR) at the user terminal i exceeds a specied SINR requirement i . A. Signal Model Assume that there are d mobile terminals requesting downlink communication in the system. There are K access points serving these mobile terminals and it is possible for each access point to serve several mobile terminals. The ith mobile terminal is assigned to access point (i) and the set I(j) = {i; (i) = j} contains all mobile terminals assigned to the jth access point. In the system, all users utilize the same time and frequency resources. The nt 1 signal xj (t), transmitted from access point j has the form xj (t) =
iI(j)

wi si (t),

(1)

where wi is the nt 1 transmit weight vector for terminal i. The signal received at terminal i is modeled as
K

ri (t) =
j=1

(vi,j ) xj (t) + ni (t).

(2)

The nt 1 baseband vector channel from access point j to terminal i, vi,j is assumed to be a zero mean and circular symmetric random vector with E[vi,j ] = 0
T E[vi,j vi,j ]

(3) (4)

=0

and a nt nt channel covariance matrix denoted by

(5)

If the instantaneous channel is known exactly, this knowledge is easily incorporated by setting Ri,j = vi,j vi,j . The user requirement on QoS is assumed to be fullled if the Signal-to-Interference-and-Noise Ratio (SINR) at the user terminal i exceeds the SINR requirement i for every terminal i.
wi Ri,(i) wi d n=1 n=i 2 wn Ri,(n) wn + i

i ,

i = 1 . . . d,

(6)

2 where the scalar i is the noise power received at receiver i.

III. B EAMFORMING A. Suboptimal beamforming methods The performance of the joint optimal beamforming and power control algorithm will be compared to some alternative beamforming methods. These are briey described below. Conventional beamforming is optimal, with respect to Signal-to-Noise Ratio (SNR) in absence of interference and with spatially white noise. The weight vectors, when transmitting to terminal i, are given by wi = arg max
w

w Ri,(i) w . w w

(7)

STRIDH et al.: SYSTEM EVALUATION OF OPTIMAL DOWNLINK BEAMFORMING WITH CONGESTION CONTROL IN WIRELESS COMMUNICATION

745

The maximum can be calculated using a scaled version of the principal eigenvector of Ri,(i) . The generalized eigenvalue-based beamformer is a decentralized beamforming strategy that gives the maximum signal power at the desired terminal under the constraint that the total transmitted power to all other terminals is kept below a certain value [1], [2], [12]. The beamformer is given by wi = arg max
w

vectors are obtained. To fully use joint optimal beamforming and power control in a practical system, a congestion control or scheduling algorithm must be added.

D. Joint Optimal Beamforming, Power Control and Access Point Assignment Above, the assignment of mobile terminals to access points was assumed to be given. This can be done in a decentralized way, where the mobile terminal is assigned to the access point to which it has the best link gain. However, the optimal beamforming algorithm can be extended to determine not only the transmit weights, but also the optimal access point assignment of the mobile terminals [7]. First, view the antenna arrays of all the access points as a single joint access point and determine the optimal beamforming vectors from this combined access point to all the terminals. The covariance matrix of the channel from the joint access point to terminal i is dened as 0 ... 0 Ri,1 0 0 Ri,2 . . . (11) Ri = . . . .. . . . . 0 . 0 ... 0 Ri,K The optimal weight vector from the joint access point to the ith terminal, may then be written wi = [wi,1 , wi,2 , . . . , wi,K ], where wi,j is the weight vector from access point j to terminal i. It can be shown that for the optimal solution, wi,j = 0 for all j = k where k corresponds to the optimal access point assignment for terminal i, i.e. only one access point will be used for each terminal, see [7]. Thus every mobile terminal is assigned to a single access point. Note that any of the algorithms [2], [4][6], [14] for solving the ordinary downlink beamforming problem can be used. In all these algorithms, it is easy to exploit the known zeros in (11) to reduce the computational complexity, see [7]. IV. H ANDLING C ONGESTED S YSTEMS Semidenite optimization [15], [16] has shown to be a powerful tool for solving a large variety of problems. Here follows a short review of optimization over convex semidenite cones, highlighting the results which are used when introducing the heuristic congestion control algorithm presented below.

w Ri,(i) w w (
d n=1 n=i

Rn,(i) + i I)w

(8)

where i I may be interpreted as an approximation of channels unknown to the access point or as a robustication to channel uncertainties [1], [2]. The problem can be solved as a generalized eigenvalue problem [13]. The choice of i will be discussed further in Sect. V. B. Power Control Power control decoupled from beamforming may be performed using the following criteria [9]. The requirement in (6) may be rewritten as SINRi = Pi Gi,(i)
d n=1 n=i

Pn Gi,(n) + i

i ,

i = 1...d

(9)

where Gi,j represents the total link gain, including beamforming, from access point j to user i and Pi represents the transmit power for terminal i. The optimal power vector is obtained by replacing the inequalities with equalities in (9), solving for Pi . The solution provides an SINR-balanced system i.e. all terminals will have reach their target SINR. When there is no feasible solution, one possible strategy is to remove terminals until all remaining terminals reach the target SINR as suggested in [9]. However, this method cannot be applied directly in combination with optimal beamforming, since the Gi,j depend on the beamformers and it is not even clear which beamforming strategy to use when there is no feasible solution. An alternative is to start from an empty system and iteratively add mobiles to ll the system. Two such strategies, similar to what has been proposed in [10] are used as reference in this investigation, see Sections IV-E and V. C. Joint Optimal Beamforming and Power Control Optimal beamforming is dened as the set of beamformers that minimize the total transmitted power in the system, while providing sufcient received Signal-to-Interference-andNoise Ratio (SINR) at each terminal. The problem may be formulated
d

min
n=1

wn
d n=1 n=i

(10) i , i = 1 . . . d.

A. Introduction to Optimization over Semidenite Convex Cones Optimization techniques for convex functions and constraints has been a very active area of research for the last two decades. A number of, so-called, interior point or path following methods have appeared that can solve not only linear optimization problems but almost all convex optimization problems, in polynomial time. One class of convex problems that is of great interest in signal processing and control theory is the class of semidenite problems. A general formulation of a semidenite problem is

s.t.

wi Ri,(i) wi 2 wn Ri,(n) wn + i

2 where i is the noise variance at terminal i and i is the target SINR for user i. Algorithms for solving (10) are described in [2], [4][6], [14]. When too many users are present in the system, the optimization problem cannot be solved and no transmit weight

746

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 4, APRIL 2006

min
Xk k=1 K

Tr[Ck Xk ] Tr[Ak,l Xk ] = bl , l = 1, . . . , L Xk = X k
k=1

but also provides the necessary theoretical framework to derive the congestion control algorithm. Dene the matrix Wi = wi wi . Then the relaxation of (10) is (12)
d

s.t.

min
i=1

Tr[Wi ]

(15)

0, k = 1, . . . , K,

s.t. Tr[Ri,(i) Wi ]
d

where the matrices Ak,l , Ck and scalars bl are constants describing the problem, K is the number of matrix valued optimization variables Xk , L is the number of linear constraints and X 0 denotes a matrix X that is positive semidenite. Note that even though this last set of constraints is highly non-linear, the problem is still convex since the set of positive semidenite matrices form a convex cone. In practice, several program packages are available to efciently nd the solution of (12), for example the SeDuMi toolbox [17]. Introducing the Lagrange multipliers l , the Lagrange dual of (12) is dened as

i
n=1 n=i

2 Tr[Ri,(n) Wn ] i i ,

i = 1...d

Wi = Wi

0,

where Wi 0 denotes that Wi is positive semidenite. With the additional constraints rank[Wi ] = 1 for all i = 1 . . . d, (15) is easily shown to be equivalent to (10). Here, however the problem is relaxed to allow for any rank. Using the fact that optimality is reached at equality in the constraints, it is straightforward to write (15) in the form (12) and solve numerically. In general, such a relaxation technique K L K can only provide a lower bound on the original problem but max min Tr[Ck Xk ] l Tr[Ak,l Xk ] bl . for this specic problem formulation, it can be shown [2], [5] l 0 Xk 0 k=1 l=1 k=1 (13) that (15) is feasible if and only if (10) is and in that case it has at least one optimal solution where rank[Wi ] = 1 for This dual can be shown [15] to be equivalent to all i = 1 . . . d, i.e. Wi = wi wi , where the wi solve (10). L This formulation of the optimal beamforming problem may l bl max be extended to include additional constraints or robustness to l 0 l=1 (14) channel uncertainties [2]. In the case of additional constraints, L though, there are no guarantees to nd the desired rank one s.t. Ck l Ak,l 0, k = 1, . . . , K solutions, but in practice there are very few exceptions. l=1 We say that an optimization problem is feasible when there is at least one point that satises all constraints. We will use the following Farkas-type lemma from [18], which relates the feasibility of the primal problem (12) to properties of the dual problem (14). Lemma. The primal problem (12) is strongly infeasible if and only if the dual problem is unbounded so there exists a dual improving feasible direction (also called feasible ray), i.e. a point 0 and a direction + such that the set of vectors [1 , . . . , L ]T = 0 + t+ are feasible solutions of (14) for all t > 0 and the cost function of (14) goes to innity when t grows. Proof Proofs can be found in in [18] or [19]. B. Solution of the Beamforming Problem The optimization problem (10) can be solved using any of the algorithms in [4], [6], [14]. These algorithms are based on a reformulation into a, so-called, virtual uplink problem. An extension to detect infeasibility is included in [6]. Here, though, an alternative solution strategy is considered using so-called Lagrange relaxation [20] techniques. For general quadratic non-convex problems, such techniques can only provide lower bounds on the cost function. However, because of the specic structure of (10) it can be shown [2], [5] that the Lagrange relaxation is exactly equivalent to the original problem, in this specic case. This reformulation into a convex problem is not only useful as a practical method to solve (10) C. Terminal Admission/Removal If no feasible solution exists to the optimization problem, the constraints have to be relaxed, either by decreasing the target SINR, or by removing mobile terminals. In this work we treat the case of removing one or several mobile terminals and the goal is to remove as few as possible in order to keep a high throughput. In order to nd the mobile terminal, i.e. the constraint, that is most critical to the system feasibility we examine the dual problem. The dual of (15) is given by
d

max
i 0 i=1

2 i i i d

(16) n n Rn,(i)
n=1 n=i

s.t. I i Ri,(i) +

0,

i = 1 . . . d.

This is equivalent to the virtual uplink problem described in [4], [6], if i i is interpreted as the virtual uplink power. According to the Lemma above, if the primal does not have a solution, the dual must have an improving direction + = [+ , . . . , + ]T . From the denition it follows directly that this 1 d improving direction fullls + 0 i
d n=1 n=i

+ n Rn,(i) + Ri,(i) n i

0,

i = 1 . . . d,

(17)

STRIDH et al.: SYSTEM EVALUATION OF OPTIMAL DOWNLINK BEAMFORMING WITH CONGESTION CONTROL IN WIRELESS COMMUNICATION

747

which is equivalent to i 0 i u Ri,(i) ui i u ( i

E. Alternative Congestion Handling Strategies The following alternative congestion/admission strategies will be used in the numerical examples for comparison. Random removals: First try to solve the beamforming problem with all available mobiles. If the scenario is not feasible, remove one randomly chosen mobile at a time until a feasible solution is found. Random additions: Starting from an empty system, add one mobile at a time, randomly chosen from the pool of users. If the addition of one mobile makes the scenario infeasible, reject it and try the next one, until all available mobiles have been tested. This is similar to the First Fit algorithm proposed in [10] when only a single time slot is available. Best additions: Start from an empty system. In each iteration, test all mobiles from the pool of users that have not yet been admitted and add the mobile that makes least harm to the system. More precisely, to test a mobile, it is temporarily admitted to the system and then the algorithm in [6, Table 1] is used to maximize the minimal SINR among all admitted users. The mobile that permits the largest such SINR value is admitted. This is similar to the Best Fit algorithm proposed in [10] applied to a single time slot. The algorithm is stopped when no more mobile can be added. To give a rough comparison of the complexity of the algorithms, assume that a total of davail mobiles are available in the system, whereof dadm are admitted to the system by the congestion control algorithm. Then it is easy to see that the number of downlink beamforming problems to solve is davail dadm for the dual based strategy proposed in IV-C and for the random removal strategy, davail for the random additions strategy and (2davail dadm )(dadm + 1)/2 for the best additions strategy. However, since most of the downlink problems involve a small number of mobiles for the strategies starting from an empty system whereas a large number of mobiles is involved for the strategies starting with all available mobiles, the random additions algorithm is often fastest even in scenarios where most mobiles are admitted. V. R ESULTS The following beamforming techniques are assessed below Conventional beamforming, followed by power control. Generalized eigenvalue based beamforming [1], followed by power control. Joint optimal beamforming and power control. Joint optimal beamforming, power control and access point assignment. In the rst three cases, the user is assigned to the access point from which it has the highest average channel gain, called maximum gain assignment, while in the fourth case the assignment is included in the optimization algorithm. In the two latter cases, the congestion control proposed in Section IV-C IV-D, here named heuristic removal, is included and in all four cases the congestion handling algorithms of Section IV-E are used for comparison. A. Performance of the Congestion Control Strategy The only guaranteed method to determine the optimal subset of mobile terminals is an exhaustive search over all

n Rn,(i) )ui

< i ,

i = 1 . . . d,

(18)

for all ui Cd and with i dened as i = + i . To i summarize, the constraints in (18) have a solution if and only if our original problem (10) does not have any feasible solution. Removing mobile m corresponds to removing constraint i = m and also removing the term m Rm,(i) in the denominator of the remaining constraints. We suggest removing the mobile corresponding to the largest i . This will decrease the denominator of all the other constraints in (18) and thus make it more difcult to nd a feasible point of (18). Also, a large value of i indicates that constraint i of (18) was easy to fulll. D. Implementation Aspects Even though the semidenite formulation of the problem was necessary in the derivation of the removal strategy, the strategy can be implemented both in combination with the iterative scheme of [6], as well as with interior point methods for semidenite programming. When the beamforming problem is solved with the interior point software packet SeDuMi [17] or some other convex optimization package using self-dual embedding techniques, the algorithm provides a certicate of the infeasibility in the form of the dual improving direction (Farkas-dual) [18]. In order to nd the mobile terminal that is most critical for the problem feasibility, a good heuristic is thus to nd the largest element of the dual improving direction. The terminal with the SINR constraint corresponding to this dual variable can be considered one of the most limiting for system feasibility. If two terminals are located closely together such that no feasible solution can be found, both will typically get a large Farkas-dual variable but it is often sufcient to remove one of them to reach a feasible situation. Thus, we suggest to repeatedly remove the most critical terminal, one at a time, until a feasible solution can be found. Below, we show that this heuristic method can give a signicant increase in system capacity compared to random removals and give performance close to that of the more complex best additions algorithm presented below. Since the Farkas-dual solution is not unique in most cases, the exact result will depend on the specic interior point algorithm. The iterative scheme proposed in [6] solves (10) in two stages. The rst stage nds the largest possible values of i . If these are smaller than the target SINR values, the problem is marked as infeasible. However, in such a case, the so-called virtual uplink powers ui (using the notation of [6]) can be interpreted as dual variables which provide a feasible solution to (18) and thus be used exactly the same way as the + values i in the convex formulation to determine a good candidate for removal. In other words, the mobile with the highest virtual uplink power is removed. Actually, (18) can be used directly as a stopping criterion when the largest possible i are smaller than the target SINR values.

748

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 4, APRIL 2006

TABLE I C OMPARISON OF THE PROPOSED HEURISTIC CONGESTION CONTROL TO THE OPTIMAL SOLUTION AND OTHER SUBOPTIMAL APPROACHES . Strategy Average number of users: Full search 9.00 Best t 8.97 Heuristic SeDuMi 8.83 Heuristic [6] 8.86 First t 8.71

0.9

=0 dB
f
0.8

=2 dB f f =4 dB =6 dB f

0.7

0.6

f =8 dB f =10 dB

C.D.F.

0.5

combinations. The full system simulations, described below, include too many mobile terminals to do a full search in reasonable time. Therefore, the congestion control strategies were evaluated and compared to the optimal solution in a small scenario with two access points and 10 mobile terminals, placed randomly. Joint optimal beamforming and power control was used and the SINR threshold was determined such that the average number of users was 9 out of 10 using the optimal solution. The results, averaged over 1000 random scenarios, is summarized in Table I. The results show that the performance of the best t and exhaustive search strategies are almost identical, so the best t algorithm provides a relevant benchmark in the results below. Note also that the heuristic method provides almost identical performance on the average, no matter if the decisions are based on the Farkas-dual of the semidenite formulation or the virtual uplink powers of the iterative solution. B. Simulations Setup A multi-cellular system, corresponding to the assumptions above, is simulated with uniform circular arrays at all access points. Omni-directional antenna elements are assumed, both at the mobile terminals and in the access point arrays. Mobile terminals arrive to the system according to a Poisson process and they are uniformly distributed over the area. The performance of the different algorithms is evaluated as a function of the load, i.e. the average number of users per cell. In order to take signals from surrounding cells into account a wrapped system is used, i.e. all cells are surrounded by all other cells. Thus the whole simulated area corresponds to a torus shape. To simulate the channel vectors, a local scattering model is used [21]. Thus the channel statistics for the channel from access point j to terminal i is parameterized by the nominal azimuth angle i,j and the standard deviation i,j of the angular spread. The total covariance matrix including the propagation properties in the gain matrix, is given by Ri,j = Gi,j R(i,j , i,j ), (19)

0.4

=12 dB
f
0.3

0.2

0.1

10

12

14

16

18

20

SINR [dB]

Fig. 1. Cumulative Distribution Function for the resulting SINR on channel realizations for different fade margins, f . The target SINR in the simulations is 7 dB. TABLE II N UMBER OF MOBILE TERMINALS IN THE SYSTEM FOR DIFFERENT FADE
MARGINS

f [dB] # mobiles / cell

0 3.7

2 3.1

4 2.9

6 2.7

8 2.5

10 2.1

12 2.0

For the conventional beamforming and the generalized eigenvalue method, the weight vectors are created according to (7) and (8). Based on the gain matrix and the beamformer antenna diagrams, power control is applied. The joint beamforming and power control algorithm provides power controlled weights directly. Just as in [5], the parameter i in (8) was chosen more or less arbitrary as i = 0.1 Tr Rn , nT n=1
n=i d

(20)

where R(i,j , i,j ) is the channel covariance matrix. Here the gain matrix Gi,j contains the inuence of distance dependent path loss with path loss exponent att and log-normal shadow fading, while the inuence of beamforming and power control is included in the weight vectors given by the beamforming algorithms and power control algorithms. The covariance matrices between all access points and terminals are assumed known. The noise in the system is assumed very low, as the main issue investigated here is the interference from other transmitters. The parameters used in the system simulations are shown in Table III.

Another possible choice of the parameter i , as suggested in [22], would have been to choose it 2025 dB below the signal strength expressed as Tr[Ri ]. The jointly optimal beamforming and power control problems are solved using the algorithm in [6]. The possibility in this algorithm to constrain the total power transmitted from all base stations, was not used here, i.e. the power constraint was set to a high value. The results are given as probability of blocking, i.e. probability of not being admitted to the system and outage probability i.e. P (SINRi < i ) for all i = 1 . . . d. In the simulations it is assumed that all users request the same target SINR i.e. i = 0 for i = 1 . . . d. The throughput is calculated as Load (Probability of no blocking), and excludes the retransmission that would be possible in a packet switched system. Assuming that only second-order statistics are known to the system and not the actual channel realizations, the algorithms described above only take the average SINR into account

STRIDH et al.: SYSTEM EVALUATION OF OPTIMAL DOWNLINK BEAMFORMING WITH CONGESTION CONTROL IN WIRELESS COMMUNICATION

749

TABLE III S IMULATION PARAMETERS .

0.5 0.45 Opt. BF, MG alloc., random add. Opt. BF, MG alloc., heuristic rem . Opt. BF, MG alloc., best add. Opt. BF, opt. alloc., random add. Opt. BF, opt. alloc., heuristic rem. Opt. BF, opt. alloc., best add.

Blocking Rate

Number of cells Reuse factor Noise level Shadow fading standard deviation Path loss exponent Number of antenna elements Angular spread variance Uniform circular array radius

9 and 16 1 -90 dBm 8 dB att = 3.5 8 9 degrees2 0.4

0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0

0.6

0.5

Conv. BF, best add. Gen. Eig. BF, random add. Gen. Eig. BF, best add. Opt. BF, random rem . Opt. BF, random add. Opt. BF, heuristic rem . Opt. BF, best add.

Blocking Rate

0.4

0.5

1.5 2 Load [mobiles/cell]

2.5

0.3

Fig. 3. Blocking rate for optimal beamforming with different congestion handling and access point assignment schemes. Required SINR is 7 dB and outage probability is 10%.

0.2

0.1

0 0

0.5

1.5 2 Load [mobiles/cell]

2.5

Fig. 2. Blocking rate for different beamforming methods and congestion handling with maximum gain access point assignment. Required SINR is 7 dB and outage probability is 10%.

and the actual SINR may often fall below the prescribed i , resulting in an outage. In order to keep the probability of outage low a fade margin is used. The fade margin is added to the target SINR in (10), i.e. i f is used instead of i . The results for optimal beamforming with the heuristic congestion control algorithm for a setup with an average load of four mobile terminals per cell, are presented in Fig. 1. It is shown that for 0 = 7 dB, the probability of outage is 60% and that a fade margin f = 10 dB, will keep the probability of outage below 10%. This result is theoretically supported in [23], [24], which also shows that designing with a fading margin is close to optimal also in terms of outage probability, given knowledge of only the second order statistics of the fading. Interestingly enough, the bounds derived in [23], [24] only depend on the fact that the fading is Raleigh distributed, not on the particular beamforming strategy. The increase of the fade margin certainly will inuence the system performance and the number of admitted mobile terminals for different fade margins, f is given in Table II. In the simulations below, a fade margin f = 10 dB is used and in addition to the plotted throughput and blocking rates, the outage probability of the system is 10%. In practice, such a fade margin is typically determined in an outer power control loop. C. Simulation Results The performance results in terms of blocking rate as a function of load have been split into two gures for clarity.

Figure 2 compares the performance of the different beamforming methods using maximum gain assignment of the access points whereas Figure 3 shows results of jointly optimal beamforming and power control, comparing maximum gain assignment to the optimal assignment. Note that the same results for optimal beamforming with maximum gain assignment are included in both gures. 1) Congestion Handling: To avoid cluttering the gures, only the most advantageous congestion handling options are shown in Figure 2 for conventional beamforming and generalized eigenvalue based beamforming, but the results follow the same trends as shown for the optimal beamforming method. For all the beamforming methods, the best additions congestion handling provides the lowest blocking rate, followed by the random additions, whereas random removals performs the worst. This is not surprising, since a mobile with bad conditions will stay long on the average in the system when using random removals but will often not even be admitted using the two other schemes. As is shown in Figure 2, the proposed heuristic removal strategy for optimal beamforming performs better than doing random additions but worse than doing best additions. As indicated in Section IV-E, the average computational complexity of the heuristic removal strategy will also typically be higher than for random additions but lower than for best additions. Note also that random removals clearly is a bad choice since it has the same computational complexity as our heuristic removals but clearly worse performance than any of the other methods. 2) Throughput Results: Figure 4 shows the same results, but presented in terms of throughput instead of load. As can be seen, the use of joint optimal beamforming and power control allows 56 times higher load or throughput, respectively, than conventional beamforming combined with optimal power control at a blocking rate of 10%. The corresponding gure over the generalized eigenvalue method is a 50% increase in load and throughput. Note that we used the most favorable

750

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 4, APRIL 2006

0.4 Conv. BF, best add. Gen. Eig. BF, best add. Opt. BF, MG alloc., random add. 0.3 Opt. BF, MG alloc., heuristic rem . Opt. BF, MG alloc., best add. 0.25 Blocking Rate Opt. BF, opt. alloc., random add. Opt. BF, opt. alloc., heuristic rem . 0.2 Opt. BF, opt. alloc., best add.

0.35

0.15

0.1

0.05

0 0

0.5

1 1.5 Throughput [mobiles/cell]

2.5

Fig. 4. Relation between blocking rate and throughput for joint optimal beamforming and power control with and without access point assignment and with different congestion handling schemes. Required SINR is 7 dB and outage probability is 10%. Results for the suboptimal beamforming methods with best additions congestion handling and maximum gain assignment are given as reference.

iterative algorithm of [6]. Such a congestion control strategy can be used directly but also provide input to more complete scheduling and radio resource management algorithms. The algorithms are not primarily intended for practical systems but can be used as benchmark in system level simulations. A system assessment of these optimum beamforming strategies has been performed, attempting to nd an upper limit to the throughput on a single channel in a multi-cell system with smart antennas. The results show that the proposed algorithms substantially increase the system performance compared to using suboptimal beamforming strategies and that the proposed congestion handling strategy provides a good trade-off between two previously proposed techniques, both in terms of performance and computational complexity. Note that the results presented here correspond to a situation where a rough scheduling has already taken place, so that most of the users in the current time and frequency slot can be admitted to the system. The results and conclusions may change when these strategies are incorporated fully into a scheduling strategy where a large pool of users are scheduled on a number of time slots. R EFERENCES

congestion handling strategy for the conventional beamforming and generalized eigenvalue based beamforming schemes in this comparison, the gains will be even larger when comparing to congestion handling strategies with lower complexity. Note also that all these conclusions may be different in other scenarios. Adding optimal access point allocation will increase the possible load and throughput with an additional 1520%, still assuming an allowed blocking rate of 10%, as shown in Fig. 3. Finally, it is worth noting that the congestion handling mechanism has a large impact on the performance, especially in scenarios with a large pool of potential users (high load), where a clever congestion handling algorithm can nd a selection of mobiles that t well together. This can have a larger impact on the system throughput than the actual beamforming strategy or access point assignment, which is clearly seen for example in Figure 4 where several of the curves cross each other for higher blocking rates. Still, these effects can be even more pronounced when the pool of users is even larger. VI. S UMMARY In this paper we have demonstrated the use of optimal beamforming, power control and access point allocation in the downlink of a wireless communication system. We show that any of the previously published algorithms for joint optimal beamforming and power control can easily be extended to determine also the optimal assignment of mobile terminals to access points. When a feasible solution does not exist for the optimization problem, terminals must be removed/blocked from the system. Exploiting a convex formulation of the optimization problem, critical constraints can be identied from the dual of the problem and based on these, a heuristic congestion control algorithm is proposed and evaluated. The resulting strategy may be applied also in combination with the

[1] P. Zetterberg, A comparison of two systems for downlink communication with base station antenna arrays, IEEE Trans. Veh. Technol., vol. 48, no. 5, pp. 1356-1370, Sept. 1999. [2] M. Bengtsson and B. Ottersten, Optimal and suboptimal transmit beamforming, in Handbook of Antennas in Wireless Communications, L. C. Godara, ed. CRC Press, 2001. [3] C. Farsakh and J. A. Nossek, Channel allocation and downlink beamforming in an SDMA mobile radio system, in Proc. Personal, Indoor, and Mobile Radio Communications, pp. 687-691, Sept. 1995. [4] F. Rashid-Farrokhi, K. J. Liu, and L. Tassiulas, Transmit beamforming and power control for cellular wireless systems, IEEE J. Selct. Areas Commun., vol. 16, no. 8, pp. 1437-1450, Oct. 1998. [5] M. Bengtsson and B. Ottersten, Optimal downlink beamforming using semidenite optimization, in Proc. 37th Annual Allerton Conference on Communication, Control, and Computing, pp. 987-996, Sept. 1999. [6] M. Schubert and H. Boche, Solution of the multiuser downlink beamforming problem with individual SINR constraints, IEEE Trans. Veh. Technol., vol. 53, no. 1, pp. 18-28, Jan. 2004. [7] M. Bengtsson, Jointly optimal downlink beamforming and base station assignment, in Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing, May 2001. [8] R. Stridh, M. Bengtsson, and B. Ottersten, System evaluation of optimal downlink beamforming in wireless communication, in Proc. IEEE Vehicular Technology Conference, Fall, pp. 343-347, Oct. 2001. [9] J. Zander, Performance of optimal transmitter power control in cellular radio systems, IEEE Trans. Veh. Technol., vol. 41, no. 1, pp. 57-62, Feb. 1992. [10] F. Shad, T. D. Todd, V. Kezys, and J. Litva, Dynamic slot allocation (DSA) in indoor SDMA/TDMA using a smart antenna basestation, IEEE/ACM Trans. Networking, vol. 9, no. 1, pp. 69-81, Feb. 2001. [11] D. Bartolom and A. I. P rez-Neira, Performance analysis of schedule e ing and admission control for multiuser downlink SDMA, in Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing 2004, vol. 2, pp. 333-336. [12] J. M. Goldberg and J. R. Fonollosa, Downlink beamforming for spatially distributed sources in cellular mobile communications, Signal Processing, vol. 65, no. 2, pp. 181-197, Mar. 1998. [13] G. H. Golub and C. F. van Loan, Matrix Computations, third ed. Baltimore: The John Hopkins University Press, 1996. [14] E. Visotsky and U. Madhow, Optimum beamforming using transmit antenna arrays, in Proc. IEEE Vehicular Technology Conference, Spring, pp. 851-856. [15] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2005 (see also http://www.stanford.edu/ boyd/cvxbook.html) [16] H. Wolkowicz, R. Saigal, and L. Vandenberghe, Handbook of Semidenite Programming. Theory, Algorithms and Applications. Kluwer Academic Publishers, 2000.

STRIDH et al.: SYSTEM EVALUATION OF OPTIMAL DOWNLINK BEAMFORMING WITH CONGESTION CONTROL IN WIRELESS COMMUNICATION

751

[17] J. F. Sturm, Using SeDuMi 1.0x, a MATLAB toolbox for optimization over symmetric cones, Department of Econometrics, Tilburg University, The Netherlands, 1999. [18] Z.-Q. Luo, J. F. Sturm, and S. Zhang, Duality results for conic convex programming, Report 9719/A, Econometric Institute, Erasmus University Rotterdam, The Netherlands, Apr. 1997. [19] E. de Klerk, T. Terlaky, and K. Roos, Handbook of Semidenite Programming. Theory, Algorithms and Applications, ch. 5: Self-Dual Embeddings, H. Wolkowicz, R. Saigal, and L. Vandenberghe, eds.: Kluwer Academic Publishers, 2000. [20] L. Vandenberghe and S. Boyd, Semidenite programming, SIAM Review, vol. 38, no. 1, pp. 49-95, Mar. 1996. [21] T. Trump and B. Ottersten, Estimation of direction of arrival and angular spreading using and array of sensors, Signal Processing, vol. 50, no. 1-2, pp. 57-69, Apr. 1996. [22] D. Giancola, F. Margherita, S. Parolari, A. Piciriello, and U. Spagnolini, Analysis of the spectral efciency of a fully adaptive antenna array system in GSM/DCS networks, in Proc. IEEE Vehicular Technology Conference 1999, pp. 812-815. [23] M. Bengtsson and B. Ottersten, Signal waveform estimation from array data in angular spread environment, in Proc. 30th Asilomar Conference on Signals, Systems & Computers 1996, pp. 355-359, Nov. 1996. [24] S. Kandukuri and S. Boyd, Optimal power control in interferencelimited fading wireless channels with outage-probability specications, IEEE Trans. Wireless Commun., vol. 1, no. 1, pp. 46-55, Jan. 2002. Rickard Stridh received his M.Sc. in Electrical Engineering from Lund University in 1995. In 1995-96 he was a Graduate Student at University of California, San Diego (UCSD). In 1996-98 he was working in the area of airborne sensor arrays at the Swedish Defence Materiel Administration. He received his Eng. Lic. degree in 2001 and his Ph.D. degree in 2003. Dr. Stridhs research interests include Smart Antennas in Future Generation Wireless Systems. Main focii are the impact of Smart Antennas on the System Architecture including the use of Radio Resource Management and Dual Arrays. Dr Stridh is currently managing R & D in the area of command, control, communications and information systems in the Swedish Armed Forces.

Mats Bengtsson (S96-M00) received the M.S. degree in computer science from Link ping University o in 1991 and the Tech. Lic. and Ph.D. degrees in electrical engineering from the Royal Institute of Technology (KTH), Stockholm, Sweden, in 1997 and 2000, respectively. From 1991 to 1995, he was with Ericsson Telecom AB Karlstad. He currently holds a position as Research Associate at the Royal Institute of Technology. His research interests include statistical signal processing and its applications to antenna array processing and communications, radio resource management and propagation channel modeling. Bj rn Ottersten Bj rn Ottersten was born in Stocko o holm, Sweden, 1961. He received the M.S. degree in electrical engineering and applied physics from Link ping University, Link ping, Sweden, in 1986. o o In 1989 he received the Ph.D. degree in electrical engineering from Stanford University, Stanford, CA. Dr. Ottersten has held research positions at the Department of Electrical Engineering, Link ping o University, the Information Systems Laboratory, Stanford University, and the Katholieke Universiteit Leuven, Leuven. During 96/97 Dr. Ottersten was Director of Research at ArrayComm Inc, San Jose, California. He has authored papers that received the Signal Processing Society Paper Award in 1993 and 2001. In 1991 he was appointed Professor of Signal Processing at the Royal Institute of Technology (KTH), Stockholm and he is currently dean of the school of Electrical Engineering at KTH. Dr. Ottersten is also a visiting professor at the University of Luxembourg. Dr. Ottersten has served as Associate Editor for the IEEE Transactions on Signal Processing and a member of the editorial board of EURASIP Journal of Applied Signal Processing. He is currently co-editor in chief of EURASIP Signal Processing Journal and a member of the editorial board of IEEE Signal Processing Magazine. Dr. Ottersten is a Fellow of the IEEE. His research interests include wireless communications, stochastic signal processing, sensor array processing, and time series analysis.