BER-based vs.

Game-theoretic Power
Allocation Strategies for Multiuser
MISO Systems
D. Bartolomé and A. I. Pérez-Neira

Publication: in Proc. 12th European Signal Processing Conference (EUSIPCO 2004)

Vol.: -
No.: -
Date: September 6-10, 2004

This publication has been included here just to facilitate downloads to those people asking for personal use copies.
This material may be published at copyrighted journals or conference proceedings, so personal use of the download
is required. In particular, publications from IEEE have to be downloaded according to the following IEEE note:
c
°2007 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material
for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or
lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.
 BER-BASED VS. GAME-THEORETIC POWER ALLOCATION STRATEGIES FOR
MULTIUSER MISO SYSTEMS
Diego Bartolomé1 and Ana I. Pérez-Neira2
1 [email protected] 2 [email protected]
Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) Universitat Politècnica de Catalunya (UPC)
c/ Gran Capità 2-4, Planta 2 c/ Jordi Girona 1-3, Mòdul D5
08034 Barcelona (Spain) 08034 Barcelona (Spain)

ABSTRACT are being served by the multi-antenna AP. This scheme is especially
Motivated by the extensive use of game-theoretic strategies for up- well-suited for Space Division Multiple Access (SDMA) systems,
link power control in CDMA, we compare in this paper a strategy since the resources granted for the users do not overlap. Note that
based on the widespread utility function used in the literature with a similar idea holds for Time Division Multiple Access (TDMA) or
other traditional schemes based on the BER. Here, we focus on the Frequency Division Multiple Access (FDMA). ZF creates parallel
downlink of a communication system. Basically, that utility func- and orthogonal equivalent channels for the users without inner-cell
tion is a ratio between the frame success rate and the used power. interference with a simple and closed-form solution [7].
It is shown in this paper that the strategy maximizing the utility im- We deal in this paper mainly with the allocation of the limited
plies a higher error rate than for other classical schemes, which was total instantaneous power among the users, where fairness consid-
not shown in the literature to the best of our knowledge. Finally, erations come to the scene [8]. Then, the figure of merit should
we briefly discuss the usefulness of pricing mechanisms in a game- not be an aggregate magnitude, thus a trade-off among the perfor-
theoretic formulation of the power control. mance of the active users should be taken into account [9]. We con-
centrate on techniques based on the BER, and either minimize the
1. INTRODUCTION sum of BER for the users, or all of them are given the same BER.
The former is the Minimum Sum BER (MSB) strategy whereas the
Game-theoretic power control has been widely studied in the lit- latter Minimizes the Maximum BER (MMB). These schemes are
erature not only in the context of Code Division Multiple Access compared to the well-known Uniform Power Allocation (UPA), and
(CDMA) since late nineties, see e.g. [1], [2], and [3], but also for to a utility-based framework, particularly the Maximization of the
digital subscriber lines [4]. Concentrating on CDMA, the authors Sum of Utilities (MSU). The MSU reflects a situation where the AP
model the uplink power control problem as a game. Provided that wishes to maximize the global perceived satisfaction. To the best of
the users are selfish and rational, game theory provides an elegant our knowledge, a similar study has not been conducted in the litera-
mathematical tool to obtain a distributed solution to the problem. ture. The final remark is that all the considered problems have been
However, complete information is needed at the terminals, which solved using convex optimization [10].
means that they shall know the channel from the other users in the In Section 2 we give an overview of the problem, just before
cell. Therefore, the term distributed refers to the computation of the the power allocation criteria exposed in Section 3. Simulation re-
solution. Some advantages of the game-theoretic formulation for sults are shown in Section 4, and then we discuss the usefulness of
the uplink are that it is easy scalable and that it provides fairness pricing schemes and conclude.
among the users since they are granted their maximum satisfaction.
This degree of satisfaction is expressed mathematically in terms 2. PROBLEM STATEMENT
of a convenient utility function, which is a key issue. If data is trans-
mitted, the utility should be increasing with respect to the Signal to In the following, boldface capital (lowercase) letters refer to matri-
Interference Ratio (SIR) if the transmit power is fixed, or it should ces (vectors). The conjugate transpose of a is aH and the element
be decreasing with power if the SIR is kept constant, among other at row ith and column jth of A is denoted by [A]i, j . The square
properties [5]. Therefore, it is sensible to use a ratio between the matrix with diagonal elements given by a1 , a2 , . . . , an is expressed
Frame Success Rate (FSR), that is, the probability that the frame is as diag(a1 , a2 , . . . , an ), and a+ = max(0, a). The cardinality of the
correct, and the transmitted power, as the authors suggest [2]. set K is given by |K |, exp(x) is the exponential function of x, and
Related to this issue, we address the power control in the down- tr(A) denoted the trace of the matrix A.
link of a communication system, where the Access Point (AP) or We focus on the downlink, where a Q-antenna AP communi-
Base Station (BS) is equipped with multiple antennas, whereas cates simultaneously with K single-antenna terminals, which are
the terminals have a single one, as it happens nowadays for most gathered in the set K = {1, . . . , K} and we assume that K ≤ Q.
communication standards. Essentially, we wish to compare the At any time instant, the received signal vector for this model is
widespread utility-based strategy used in e.g. [1], [2], or [3] with
other schemes based on the Bit Error Rate (BER). We show that al-
y = HBs + w ∈ CK×1 , (1)
though the utility-based optimization maximizes the utility within
the cell, while minimizing the power, the FSR is penalized, or
equivalently, the BER is higher than for other methods. This result where the kth position of vector y (s) is the received (transmitted)
reflects the difficulty in choosing convenient utility functions. signal for user k. H is the K ×Q complex flat-fading channel ma-
Within our context, the first issue is the transmit beamforming. trix, the ith row of which contains the 1 × Q vector of the channel
As in [6], we assume a Zero Forcing (ZF) beamforming criterion, gains for the ith user, i.e. hTi , and we assume that the components
which eliminates the inner-cell interference among the users that of the channel matrix are independent and identically distributed
Gaussian random variables with zero mean and unit variance. In a
This work is partially supported by the CTTC, by the European Union TDMA/TDD system at pedestrian speed, it might not be far from
under IST-2003-507581, by the Spanish government under FIT-070000- reality the assumption that the channel matrix is known at the AP,
2003-257, HF2001-0055, and TIC2002-04594-C02 (jointly financed by whereas the receivers are only aware of their own channel
response.

FEDER), and by the Catalan government thanks to grant 2003FI 00190. The noise vector is complex Gaussian, i.e. w ∼ C N 0, σ 2 IQ ,
and the transmit beamvectors for the K users are gathered in the which leads to a lower BER for the users having a better channel.
matrix B = [b1 b2 . . . bK ] ∈ CQ×K . Since this paper assumes that the AP has perfect channel knowl-
With a ZF beamforming criterion, the K channels become par- edge, more efficient power allocation criteria could be applied, as
allel and orthogonal, thus the users receive their transmitted symbol we can see next. We point out that the presented techniques are
corrupted only by additive white Gaussian noise, without inner-cell considered to deliver a best-effort service, since the AP optimizes a
interference. In this problem, it is meaningful to separate the effect function of the BER regardless of the individual Quality of Service
of the channel and the power allocation. Therefore, we normalize (QoS) achieved by the users. For further details, see e.g. [9].
the beamvector, so that the effect of the equivalent channel is cap-
tured by αk . The modified ZF criterion becomes Hb  k = αk 1k , ∀k, 3.1 Minimum Maximum BER (MMB)
where the vector 1k has zeros at all positions but the kth. The nor- A possible optimization criterion consists of minimizing the maxi-
malized beamvector for the kth user is obtained as mum BER among the users. We will see in this subsection that it
 −1 finally reduces to assigning the same BER to all users, regardless of
 k = αk HH HHH
b 1k , (2) their channel quality. The cost function is expressed as

 min max BERk (6)

 βk2 k
where the αk = 1/ (HHH )−1 are real and positive by con-
struction. The K normalized
k,k
beamvectors are gathered in the ma-
s.t. ∑ βk2 ≤ PT , (7)
 k∈K
   
trix B = b1 b2 . . . bK , thus HB = Dα = diag (α1 , α2 , . . . , αK ).
where we have implicitly assumed that the βk2 are non-negative,
The beamforming matrix contains also the power factors βk , i.e. since they are power allocation factors. According to convex op-
B = BD  β , where Dβ = diag (β1 , β2 , . . . , βK ). The signal model in timization theory [10], the previous problem is convex because the
(1) finally reduces to BER approximation is an exponential and the constraints are lin-
ear. In order to properly solve this problem, one should express it
y = Dα Dβ s + w ⇒ yk = αk βk sk + wk , (3) according to the convex formulation. Recalling (5),

in which the equivalent gain αk depends on the channels from the min t (8)
βk2
rest of the users. With this model, the Signal to Noise Ratio (SNR)
for the kth user is given by s.t. ∑ βk2 − PT ≤ 0, (9)
k∈K
αk2 βk2 αk2 βk2
γk = , (4) c1 exp(−c2 ) − t ≤ 0, ∀k ∈ K , (10)
σ2 σ2
where we have assumed that the symbols have unitary mean en- −βk2 ≤ 0, ∀k ∈ K , (11)
ergy, particularly, normalized Quadrature Amplitude Modulation
(QAM) symbols are considered. For the sake of simplicity, which to which we can apply the Karush-Kuhn-Tucker (KKT) conditions
is an important feature for cross-layer designs, we use the easy- [10]. Using these,we can find the following solution1
differentiable BER expression given in [11] for QAM signals, i.e.
PT
αk2 βk2 =
, (12)
BER(γ ) ≈ c1 exp(−c2 γ ), (5) tr (HHH )−1

where c1 and c2 depend on the concrete signal mapping. Since this which implies that all the users attain the same SNR, and thus the
paper assumes that there is no channel coding at the transmitter, same BER. However, in this case the AP is using some resources to
the channel is time-invariant, and the noise is Gaussian, the Frame increase the performance of the worse users.
Error Rate (FER) can be expressed as a function of the BER and
the frame length L in bits as FER = 1 − (1 − BER)L , thus the Frame 3.2 Minimum Sum BER (MSB)
Success Rate (FSR) can be obtained as FSR = 1 − FER. Another possibility is to minimize the total BER, regardless of the
users with poorer channels. In one sentence, we would like to min-
3. POWER ALLOCATION STRATEGIES imize the sum BER of all the users in the cell subject to the power
In this section, we propose several alternatives for allocating the to- constraint, which means
tal available instantaneous power PT among the users. We assume
that they are homogeneous, i.e. their data traffic requirements are min
βk2 k∈K
∑ BERk (13)
the same, and we focus on optimizing a function of the BER while
imposing a constraint on the instantaneous power, in contrast to the s.t. ∑ βk2 − PT ≤ 0, (14)
approach taken e.g. in [6]. Besides, we evaluate the utility-based k∈K
cost function proposed e.g. in [2] among other papers. However, −βk2 ≤ 0, ∀k ∈ K . (15)
it is already pointed out in [12] that further understanding of the
utility functions is needed. In this sense, there is a number of cost We can apply the KKT conditions [10] because the problem is
functions that could be used. Since in this paper a centralized sys- convex, and one can see that the solution is similar to a waterfilling:
tem is considered, we assume that the AP would like to maximize
the sum of utilities for all the users, see Section 3.3, which reflects   +
the situation where the AP obtains the highest global satisfaction. σ2 c1 c2 αk2
βk2 = log − logλ , ∀k ∈ K , (16)
Without any channel knowledge, the best option would be the c2 αk2 σ2
well-known Uniform Power Allocation (UPA), in which the whole
power is divided equally among the active users in the cell, so that where logλ is obtained in order to fulfill (14) with equality. A re-
we do not care about their actual channel gain nor how we can
mark about implementation is that since the αk2 change (increase)
improve the performance. The power allocated to the kth user is
αk2 PT

βk2 = PT
K , thus the SNR for the kth user is given by γk = σ2 K
, 1 Note that ∑k∈K 1/αk2 = tr (HHH )−1 .
when the number of users is reduced, they shall be recomputed if Table 1: Maximization of the Sum of Utilities (MSU) Algorithm
there is any user j for which β 2j = 0. Then, user j is removed from 1. Set K = {1, . . . , K}.
the active set K , thus the jth row is eliminated from H. Therefore, 
H with
2. Build matrix the users in the set K , and compute
the solution in (16) shall be computed again. By construction, it −1 
αk2 = 1/ HHH , ∀k ∈ K .
is clear that this scheme will provide a lower BER than the MMB, k,k
but the drawback is that for the sake of the collective revenue, some 3. If the condition in (24) is satisfied, go to step 5.
users might not even be allocated for transmission. 4. Otherwise, select the user ks : mink αk2 , and remove it from
the active set, K = K − ks . Go to step 2.
3.3 Maximum Sum of Utilities (MSU)
5. Compute the power for the users in K according to (19).
Differently to other papers, e.g. [1], [2], or [3], we solve a utility- For the users not in K , set βk2 = 0.
based downlink power control using convex optimization [10]. Al-
though the game is generally competitive [4], the MSU here is anal-
ogous to a refereed game in which a cooperative strategy is sought, Applying the KKT conditions [10], we obtain that the solution
see [2] and references therein. Whereas in the refereed game the βk2 might be in the set
AP would tell the terminals the uplink power, in this case the AP
allocates a certain power to the users for downlink transmission2 .  
σ 2 NE
With minor modifications to the utility function given e.g. in βk2 ∈ 0, 2 γ , ∀k ∈ K . (23)
[2], the utility perceived by the kth user can be expressed as αk
 L  
α2β 2 L
1 − BER
c1 1 − exp(−c2 σk 2 k ) If the power were unbounded, the utility maximization would
uk = = , (17) yield the same performance as the MMB, since all the users would
βk2 βk2 get the same equilibrium SNR given by γ NE . However, since the
in which, in agreement with e.g. [5], the FSR in the numerator has power is limited, either the user is allocated at a point such that its
been slightly modified. By dividing the BER by the constant c1 we own utility is maximized or it is not scheduled. The question is
guarantee that if βk2 = 0 ⇒ uk = 0, and that the utility tends to 0 which users will not be allocated for transmission.
as the power goes to infinity, i.e. limβ 2 →∞ uk = 0. If we had not If we add the power factors βk2 obtained in (19) as if all the users

k
were active, the total power is σ 2 γ NE tr (HHH )−1 . Therefore, the
proceeded so, at null power, βk2 = 0, the utility would be infinity,
and the terminal would choose not to transmit. This modification MSU problem serves all the users with the SNR of the NE, γ NE , if3
does not have a deep impact in the trend of the FSR [1].   P /σ 2
In the literature for the uplink power control, each user max- tr (HHH )−1 ≤ NE .
T
(24)
imizes its own utility. Then, if we derive the utility in (17), the γ
solution is the βk2 whose equilibrium SNR γk∗ satisfies
In any other case, we should decide which users are allocated
exp(−c2 γk∗ )(1 + Lc2 γk∗ ) − 1 = 0, (18) null power. It can be easily seen that if we substitute the equilib-
so that the power allocation factors can be obtained using (4) as rium βk2 obtained in (19) in the utility function in (17), the utility
k = kNE αk , where kNE is a constant.
for the kth user at the NE is uNE 2
σ2 ∗
βk2 = γ = c. (19) Therefore, the user with lower αk (worst channel) is the selected
2
αk2 candidate to be allocated null power, since it is the user that penal-
izes the performance of the rest of the users.
This point c is a maximum of the utility function in (17). There-
With these issues, we summarize in Table 1 the algorithm that
fore, since there exists a point c such that uk is non-decreasing for
yields the highest sum of utilities. First, it tries to allocate all the
t ≤ c, and non-increasing for t > c, the function in (17) is quasi-
users, but if the problem is not feasible, the best strategy is to re-
concave [10]. Moreover, this point constitutes a Nash Equilibrium
move the user with worst channel, see step 4 in Table 1. Step 5
(NE) for the uplink power control game, which is taken as a bench-
mark in existing literature, e.g. [5] and [2]. Therefore, we denote reflects (23). Note that if a user is allocated null power, the αk2 shall
be recomputed because they increase when less users are served.
the equilibrium SNR as γ NE = γk∗ . A NE is a point where no user
can increase its own utility function by changing its own transmitted In most cases, ∑k∈K βk2 < PT because the power is determined by
power, given the transmitted power from the other users [13]. (19). The idea behind is that using more or less power than the NE
Since we focus on the downlink, the AP shall distribute the lim- for any user would imply a lower utility, which is not the objective.
ited instantaneous power among the users in the cell. This consti-
tutes a difference with respect to existing literature, see e.g. [12] 4. SIMULATIONS
and references therein. For this multiuser communication, the AP We have an AP provided with Q = 6 antennas, which tries to serve
has several alternatives involving fairness issues [9]. In this paper, K = 6 active users in the cell. The SNR in the figures refers to
the AP wishes to maximize the sum of utilities of all the users in the ratio PT /σ 2 , and the range is from 4 dB to 28 dB in steps of 4
the cell, which means that the total perceived satisfaction would be dB. We assume 2 bits of symbol without loss of generality, thus the
maximum. Since the objective function (sum of utilities) is quasi- constants in (5) are c1 = 0.2 and c2 = 1.6/3. We have conducted
concave because it is obtained by a sum of quasi-concave func- simulations to evaluate the utility, the BER, and the power.
tions [10], minus a sum of quasi-concave functions is quasi-convex. First, we plot in Figure 1 the sum of utilities for all users
Therefore, we can formulate the optimization in convex form as in the cell with respect to the SNR. It is clear that the technique
min −
βk2
∑ uk (20) based on the maximization of the utility yields the best results
compared to the UPA, the MMB, and the MSB. Moreover, these
k∈K
techniques always use the total available power, whereas the MSU
s.t. ∑ βk2 − PT ≤ 0, (21) only uses the following percentages of total transmitted power
k∈K [89.24, 78.10, 73.41, 68.17, 55.59, 42.27, 30.83] for the simulated
−βk2 ≤ 0, ∀k ∈ K . (22) SNR. Therefore, we can conclude that the utility is maximized
while the used power is the lowest among the studied methods.
2 Note that if the system is TDMA/TDD and the channel is quasi-static,

the same power could be used for the uplink 3 In this case, we say that the problem is feasible.
 Sum of utilities for the inner−cell users, Q=K=6
We plot in Figure 2 the sum of BER vs. the SNR for the pro- 3
10
posed methods. The BER is set to 0.5 if the users do not transmit. MSU
MMB
The MSB yields the optimum performance since it is designed for MSB
that purpose. It is important to see that the maximization of the 2
10
UPA

sum of utilities does not yield a good performance in terms of BER.

Finally, this SDMA system based on ZF fully exploits the multi-

Sum of utilities in logarithmic scale

plexing gain because it is serving the maximum number of users, 1

i.e. K = Q, but the diversity gain is penalized [14]. The authors 10

show in [9] that even when K = Q − 1, the BER decreases in more

than one magnitude order for moderate SNR. 0
As final comments, note that the alternatives based on game 10

theory, e.g. pricing [5] or repeated games [12], would increase the
utility while reducing even more the power. These options are stud-
ied in order to overcome the Pareto deficiency of the NE. Briefly, a −1
10

Pareto optimum point means that no user can increase its own utility
without decreasing the utility obtained by other users [13]. To the
best of our knowledge, it is not shown in existing papers how the −2
10
0 5 10 15 20 25 30
BER (or FSR) performance degrades, see e.g. [2]. Therefore, con- SNR [dB]

straints on the SNR or on the BER should be added to the problem Figure 1: Sum of utilities for the inner-cell users.
in order to fulfill the real traffic requirements from the users.
Sum of BER for the inner−cell users, Q=K=6
0
10
5. DISCUSSION AND CONCLUSIONS MSU
MMB
UPA
In this paper, we have compared the utility-based power control MSB
with some schemes based on the BER, namely the minimum sum
BER and the minimum maximum BER. We have solved the prob-
lems using convex optimization, and results have shown that the −1
10

maximization of utility does not yield a good performance in terms

Sum of BER

of BER, even compared to the classical uniform power allocation.

To the best of our knowledge, BER performance was not shown
in previous papers developing a game-theoretic formulation of the
power control in CDMA. In any case, game theory provides an at- −2
10
tractive mathematical framework, and concepts such as pricing can
be useful for future communication systems.
The pricing factor can be set by the AP in order to force the
terminal to transmit at a certain power level in the uplink. For in-
stance, we can assume that each selfish terminal wishes to maximize −3

the following modified utility function ũk = uk − ck βk2 , where ck is 10

0 5 10 15
SNR [dB]
20 25 30

a different pricing factor for each user. The pricing ck would be

chosen by the AP in a way such that when the terminal optimizes Figure 2: Sum of BER for the inner-cell users.
individually ũk with respect to βk2 , the selected power would be the
one previously computed by the AP in order to optimize a certain [7] R.W. Heath, M. Airy, and A.J. Paulraj, “Multiuser Diversity
cost function. For instance, the AP could select among the UPA, for MIMO Wireless Systems with Linear Receivers,” in Proc.
the MSB, or the MMB. Since the AP has all the necessary informa- of the Thirty-Fifth Asilomar Conference on Signals, Systems
tion and computational capabilities, it can communicate the pricing and Computers, Nov. 2001.
value to the terminals, so that the power allocation is computed in a [8] D. Bartolome, D.P. Palomar, A.I. Perez-Neira, “Real-Time
distributed manner. However, note that some information is needed Scheduling for Wireless Multiuser MISO Systems under
at the terminals, which shall be provided by the AP. Different Fairness Criteria,” in Proceedings of the Interna-
tional Symposium on Signal Processing and its Applications
REFERENCES (ISSPA), July 2003.
[9] D. Bartolome and A.I. Perez-Neira, “Spatial Scheduling in
[1] D. Goodman and N. Mandayam, “Power control for wireless
Multiuser Wireless Systems: From Power Allocation to Ad-
data,” IEEE Personal Commun. Mag., vol. 7, no. 2, Apr. 2000.
mission Control,” submitted to IEEE Trans. on Wireless Com-
[2] A.B. MacKenzie and S.B. Wicker, “Game theory and the de- mun., October 2003.
sign of self-configuring, adaptive wireless networks,” IEEE
[10] S. Boyd and L. Vanderberghe, Convex optimization. Avail-
Commun. Mag., vol. 39, no. 11, Nov. 2001.
able online at http://www.stanford.edu/ boyd/bv cvxbook.pdf,
[3] E. Altman and Z. Altman, “S-modular games and power 2003.
control in wireless networks,” IEEE Trans. Automat. Contr.,
[11] S.T. Chung and A.J. Goldsmith, “Degrees of Freedom in
vol. 48, no. 5, May 2003.
Adaptive Modulation: A Unified View,” IEEE Trans. Com-
[4] W. Yu, G. Ginis, and J.M. Cioffi, “Distributed Multiuser mun., vol. 49, no. 9, pp. 1561–1571, Sept. 2001.
Power Control for Digital Subscriber Lines,” IEEE J. Select.
[12] A.B. MacKenzie and S.B. Wicker, “A Repeated Game Ap-
Areas Commun., vol. 20, no. 5, June 2002.
proach to Distributed Power Control in CDMA Wireless Data
[5] C.U. Saraydar, N.B. Mandayam, and D.J. Goodman, “Effi- Networks,” in revision for IEEE Trans. on Wireless Commun.
cient power control via pricing in wireless data networks,”
[13] D. Fudenberg and J. Tirole, Game Theory, Cambridge, MA:
IEEE Trans. Commun., vol. 50, no. 2, Feb. 2002.
MIT Press, Ed., 1991.
[6] D. Bartolome, A. Pascual Iserte, A.I. Perez-Neira, “Spatial
[14] L. Zheng and D.N.C. Tse, “Diversity and Multiplexing: A
Scheduling Algorithms for Wireless Systems,” in Proc. of In-
fundamental Tradeoff in Multiple-Antenna Channels,” IEEE
ternational Conference on Acoustics, Speech, and Signal Pro-
Trans. Inform. Theory, vol. 49, no. 5, May 2003.
cessing (ICASSP), Apr. 2003.