You are on page 1of 11

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 51, NO.

7, JULY 2005

2625

A Generalized Framework for Distributed Power Control in Wireless Networks


Chi Wan Sung, Member, IEEE, and Kin-Kwong Leung, Member, IEEE
AbstractMost power control algorithms that aim at hitting a signal-to-interference ratio (SIR) target fall within Yates framework. However, for delay-tolerable applications, it is unnecessary to maintain the SIR at a certain level all the time. To maximize throughput, one should increase ones power when the interference level is low, and the information transmission rate is adjusted accordingly by adaptive modulation and coding techniques. This approach is called opportunistic communications. In this paper, we generalize Yates result and establish a new framework, which is applicable to systems supporting opportunistic communications and with heterogeneous service requirements. Simulation results show that our proposed algorithm yields signicant improvement in throughput when compared with the conventional target tracking approach. Index TermsPower control, distributed algorithms, iterative functions, convergence, opportunistic communications, wireless networks.

I. INTRODUCTION OWER control is an important component of the design of a wireless system, as the capacity of which depends heavily on how mutual interference among different users is managed. Since Zanders early work on power control [19], there have been many research studies on developing distributed power control algorithms. For second-generation cellular systems, supporting voice connections is the major task. Thus, the objective of power control is mainly to meet given signal-to-interference ratio (SIR) requirements of individual users. For this purpose, many power control algorithms have been developed (e.g., [1], [4]). To analyze the convergence of these algorithms, Yates established a unied framework in 1995 [18]a seminal work in the area of power control. This framework has recently been generalized by Leung et al. [9] such that some variations of discrete algorithms (e.g., [12]) are also included. Due to the possible proliferation of wireless Internet, power control should be designed to support both data and voice services. For data services, it is no longer needed to maintain a certain SIR throughout a connection. Since data services generally can tolerate a larger delay, it is possible to schedule

Manuscript received June 16, 2003; revised September 20, 2004. This work was supported in part by a grant from the City University of Hong Kong (Project 7001454). C. W. Sung is with the Department of Computer Engineering and Information Technology, City University of Hong Kong, Hong Kong (e-mail: itcwsung@cityu.edu.hk). K.-K. Leung is with the Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, BC V6T 1Z4, Canada (e-mail: kkleung@ece.ubc.ca). Communicated by D. N. C. Tse, Associate Editor for Communications. Digital Object Identier 10.1109/TIT.2005.850045

the transmission of different users according to their channel quality. This idea is explored in [8]. It is shown that to maximize system throughput of a single-cell code-division multiple-access (CDMA) system, only one user should transmit at a time; the one with largest instantaneous channel gain should transmit. Since wireless channel is inherently time varying, the system may exploit the channel uctuation by transmitting information opportunistically when the signal strength is strong. This concept is named opportunistic communications, and is further investigated from an information-theoretic viewpoint in [6], [15], [16]. Opportunistic communications can be put into practice. An example is the proportional fair scheduling algorithm for Qualcomms high data rate (HDR) system [7]. Furthermore, it has potential applications in wireless ad hoc networks [5]. For wireless systems with a single access point, the power control problem reduces to a scheduling one, since the optimal solution allows only one user to transmit. However, this property does not hold for a general wireless network with multiple transmitters and receivers, such as cellular systems and wireless ad hoc networks. In this paper, we use distributed power control to implement opportunistic communications. A user increases his power to transmit more information when the channel gain is large and interference is low. Information rate can be adjusted accordingly using adaptive modulation and coding techniques. Since there is no centralized coordinator to regulate the transmit power of the users, it is important to ensure that the control algorithm is stable when each user executes it independently. To put it another way, we have to nd conditions that ensure convergence of a power control algorithm. It should be noted that Yates framework does not apply to this situation, since it requires the power of a user to increase when the interference is largethe so-called monotonicity property. In view of this limitation, we generalize Yates result and establish a new framework, which can be applied to a larger class of algorithms, including those that are suitable for opportunistic communications. The rest of the paper is organized as follows. In Section II, we present the idea of opportunistic power control. In Section III, we consider a new class of iterative functions which is specifically designed for opportunistic communications. In Sections IV and V, a new framework for power control is established. In Section VI, we consider asynchronous update on the power vector, which allows the users to perform power adjustments at different times and with outdated information. In Section VII, we derive conditions that ensure the existence of xed points. In Section VIII, we prove a property of the xed point and show that the xed point of the opportunistic algorithm does possess

0018-9448/$20.00 2005 IEEE

2626

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 51, NO. 7, JULY 2005

the property of opportunism. In Section IX, we present our simulation results. In Section X, we discuss how our approach is connected to the game-theoretic one. Finally, we draw a conclusion in Section XI. II. OPPORTUNISTIC POWER CONTROL mobile users. Let Consider a wireless system with , where is the transmit power of user . It is sometimes useful to focus on the powers of users other to denote the vector than the th one. We use . Conventional power control aims at nding a suitable power vector so that the SIR requirement of all users can be met. Let be the SIR of user and his target SIR. Denote the link gain between the transmitter of mobile and the receiver . Let be the noise power at the receiver of of mobile by can be expressed as follows: mobile . Then (1) where (2)

Recently, Xiao et al. proposed a utility-based power control (UBPC) method [17], whose iterative function takes the following form: for (5)

It is similar to (3) except that a xed is replaced by , . In other words, the target which is a decreasing function of SIR decreases for a worse transmission environment. Under certain conditions, UBPC is shown to be standard and falls within Yates framework. The algorithm being standard implies that the , for power iteration of user is an increasing function of otherwise the monotonicity condition is violated. In the light of opportunistic communications, it is reasonable to go one step further; instead of just decreasing the target SIR, a user may decrease even his power if he is facing a hostile transmission environment such as strong interference or deep fading. We call this strategy opportunistic power control, and is formalized as follows. Denition 1: A power control algorithm for user is said to is a decreasing be opportunistic if the iterative function . function of the effective interference For example, one may keep the sum of the signal power and the effective interference to a constant (6) This approach is similar to water-lling in a fading channel. The associated iterative function is (7)

is the effective interference of mobile . If all users adjust their powers so as to hit their corresponding targets, then we have the following set of iterative equations: for (3)

This iterative algorithm is a particular case of the FoschiniMiljanic algorithm [4]. To ensure system stability, it is important to know whether the power vector converges to a xed point. Yates framework [18] is able to answer this question. In Yates framework, the power vector is adjusted by an iterative function1 (4) . An iterative function is where called standard if it satises the following two conditions.2 , then 1) Monotonicity: If 2) Scalability: For all . Yates shows that if a xed point exists, then the point is unique. Moreover, the power vector converges to the xed point given any initial power vector. It can be veried that the iterative equation in (3) is standard; hence Yates framework applies.

which is opportunistic by denition. However, this design has a drawback. Consider a two-user single-cell system. Assume that and all link gains are equal to . Then is a xed point if it satises (8) So, the xed point is not unique and various problems may arise. For example, if user 1 transmits his data before user 2, then user , whereas user 2 can 1 will occupy the channel using transmit nothing, resulting in an unfair situation. An alternative design is to keep the product of the signal power and the effective interference to a constant (9) We call the target signalinterference product (SIP) of user . The associated iterative function becomes (10)

1In Yates framework, I (p) is called an interference function, since it represents the effective interference that the users must overcome. However, in this work, we do not attach to it any physical meaning. It is regarded only as an iterative function that adjusts the power vector. 2The positivity condition, an additional condition in the original denition of standard iterative function [18], can be shown as a consequence of monotonicity and scalability [9].

which is also opportunistic. In this paper, special emphasis will be given to this iterative function, which we call the opportunistic power control algorithm. Our power control framework to be developed is a spin-off from an investigation of this iterative function. We are interested in it for the following reasons.

SUNG AND LEUNG: A GENERALIZED FRAMEWORK FOR DISTRIBUTED POWER CONTROL IN WIRELESS NETWORKS

2627

1) In a time-division system, the transmitter can determine its transmission order to multiple receivers based on the magnitude of their channel gains. If higher priorities are assigned to those who have better channel conditions, this is called opportunistic scheduling. However, this method requires a central decision maker. On the contrary, opportunistic power control is amenable to distributed implementation. 2) This iterative function has a unique xed point. Furthermore, given any initial power vector, the iterations always converge to that xed point. 3) Using this iterative function, system throughput can be increased signicantly when compared with the conventional target tracking approach. The rst property is clear since the effective interference of user can be estimated at her receiver. Therefore, it can be applied to wireless systems such as multiple-cell mobile phone systems and mobile ad hoc networks, which do not have central decision makers. The second property is a direct consequence of our power control framework. It is as nice as we can hope for a distributed algorithm. The third property is very important from an engineering viewpoint. In our simulation model, we use the Shannon capacity formula to measure the throughput (11) For the case where the link gains are time varying, we simply over a long period of time to obtain the average average performance. Although the opportunistic algorithm does not guarantee optimality, it yields very good performance when compared with target tracking. This will be shown by simulation toward the end of the paper.

Theorem 2: If is type-II standard and has a xed point, then that xed point is unique. be distinct xed points. Obviously, Proof: Let and . Since is standard, both of them are xed points of its xed point, if exists, is unique [18]. A contradiction. Theorem 3: If is type-II standard and has a xed point , then for any initial power vector , the power vector converges to . is standard, both Proof: Since and converge to . Hence, converges to .

In future wireless networks, users may have heterogenous service requirements. It is natural to ask whether the power vector converges when some of their iterative functions are standard while the others are type-II standard. For ease of presentation, we call the vector function wide-sense standard. Denition 2: of its components, is said to be wide-sense standard if each , is either standard or type-II standard.

In the next section, we show that the wide-sense standard function belongs to a more general class of iteration functions called two-sided scalable functions. Convergence property will be proven in the sequel. IV. TWO-SIDED SCALABILITY We now present a new class of iterative functions, which enthat satises compasses Yates as a special case. Consider the following condition. Two-Sided Scalability: For all , implies

III. TYPE-II STANDARD FUNCTIONS Given an opportunistic iterative function, we are interested in the question whether the system is stable. Yates results cannot is not standard. In fact, an opportunistic apply since algorithm cannot be standard as it violates the monotonicity condition. with its comWe note that the iterative vector function ponents given in (10) satises the following two conditions. 1) Type-II Monotonicity: If 2) Type-II Scalability: For all , then . . By scalability, and An iterative function that satises this condition is said to be two-sided scalable. Now we consider how two-sided scalable functions relate to wide-sense standard functions. Proposition 4: A wide-sense standard function sided scalable. is standard. Let . If Proof: Suppose , then by monotonicity is two-

(12)

We call an iterative function satisfying these two conditions type-II standard. Next we examine whether a type-II standard algorithm converges or not. For ease of presentation, we denote as . the composite function Theorem 1: If is type-II standard, then is standard. satises monotonicity. It reProof: It is obvious that satises scalability. Let ; then mains to prove that . By type-II monotonicity, . . Hence, the By type-II scalability, is proven: . scalability of

Hence, we have (13) Now suppose is type-II standard. Again let , then by type-II monotonicity . If

(14)

2628

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 51, NO. 7, JULY 2005

By type-II scalability, . Again we have

and

only the case dened: (15)

. Therefore, the following function is well

(22) By two-sided scalability, we immediately have the following. Lemma 7: If , then . .

As a result,

is two-sided scalable.

The following example shows that the converse is not true. Consider the iterative function (16) which is two-sided scalable. However, it is neither standard nor type-II standard, since the monotonicity properties do not hold. Another useful construction of two-sided scalable functions and , is as follows. Given two iterative functions, dene (17) and (18) where the minimum or maximum operation is applied componentwise. Proposition 5: If and are two-sided scalable, then and are also two-sided scalable. , where , we have Proof: Given (19) and (20) Hence, (21) Similarly, is also two-sided scalable.

This helps us to establish the continuity of

Lemma 8 (Continuity): is continuous at for all . Proof: It sufces to show that each of the functions is continuous at . Given , there such that exists (23) Consider the open box contains the point . By two-sided scalability, . Continuity follows. , which maps into

Now we are ready to prove our main results: the uniqueness of the xed point and the convergence of the power vector. is two-sided scalable and a xed point Theorem 9: If exists, then that xed point is unique. and be xed points of . Then Proof: Let . By Lemma 7, . Theorem 10: If is two-sided scalable and a xed point exists, then the power vector converges to . as . By Lemma 7, is Proof: Denote is lower-bounded by , it monotonically decreasing. Since converges to a limit, say . Note that for all We prove our assertion by showing that must be equal to . ; we prove this by contradiction. Let be Suppose . Since , we have ; hence for all . Consider the function (24) is compact, the Since is continuous (by Lemma 8) and is also compact. Hence, a minimum exists in image . By Lemma 7, ; we have for all . As a result, for all , which contrabeing lower bounded. dicts VI. ASYNCHRONOUS POWER CONTROL As in Yates framework, we examine an asynchronous version of the iterative algorithm using the totally asynchronous model of Bertsekas and Tsitsiklis [2]. be the power vector at time . We Let assume that the base station to which mobile belongs may not have access to the most recent values of the components of . Let be the time corresponding to the th component available . If user adjusts to mobile at time . Note that its power at time , that adjustment is based on the vector (25)

This construction method is relevant to this scenario. Suppose user is connected to two base stations. She is given a choice between two target powers and from the two base stations, respectively, both of them are two-sided scalable. If she makes either the maximum or the minimum of her power choices, the resultant iterative function is still two-sided scalable. V. CONVERGENCE THEOREM In this section, we investigate the convergence of two-sided iterative functions. Lemmas 68 state some properties of twosided scalable iterative functions. First, we note that two-sided scalability implies positivity. Lemma 6 (Positivity): Proof: By denition, for all . for .

As a consequence, except possibly when . If for some , we may relabel the index by letting the . Without loss of generality, we consider initial vector be

SUNG AND LEUNG: A GENERALIZED FRAMEWORK FOR DISTRIBUTED POWER CONTROL IN WIRELESS NETWORKS

2629

Without loss of generality, we assume that the set of times at which one or more components of are updated is the discrete . Let be the set of times at which user set adjusts his power. Given the sets , the asynchronous iterative algorithm is dened by (26)

where the elements of

is dened as (29)

and

is the vector with elements (30)

therwise.

The general solution of the difference equation is (31) It is well known that the vector will converge if the dominant eigenvalue of , is less than (e.g., [4], [13]). This is called , of the users are the feasibility condition. If the target SIR, will be greater than . In that case, we have too large, then the following solution for large : (32) where is the eigenvector associated with and is a constant depending on the initial vector and the gain matrix. Thus, the distance measure will converge to the dominant eigenvalue . On the other hand, the power vector goes to innity. Therefore, given a two-sided scalable iterative function, it is important to determine whether a xed point exists. Brouwers xed-point theorem is found to be useful. Theorem 13 (Brouwers Fixed-Point Theorem [3]): Let be compact and convex and a continuous such that . function. Then there exists The following result gives a sufcient condition for the existence of a xed point. Theorem 14: Given a two-sided scalable iterative function , if there exists such that for all , then a xed point exists. . Since is two-sided Proof: Let scalable, by Lemma 8, it is continuous in . Since is compact and convex, Brouwers xed-point theorem implies that a xed point exists. This result can be applied to the case where power constraints are imposed, which is common in practical wireless systems. Suppose the power vector is constrained to lie within the box . Dene the constrained iterative function by (33) By repeatedly applying Proposition 5, one can show that is also two-sided scalable. By Theorem 14, a xed point exists; hence, the power vector converges. Note that Theorem 14 does not apply if there is no minimum power constraint. More generally, when one or more of the commay not be ponents of the lower bound are equal to zero, is concontinuous in , since Lemma 8 only guarantees . With the additional condition that tinuous in is continuous at every , then a xed point exists. Theorem 15: Given a two-sided scalable iterative function , if is continuous and for all , then a xed point exists.

for every mobile , which We assume guarantees that old information is eventually purged from the system. For details, see [2]. Convergence of the asynchronous power control algorithm will be established by applying the following result. Theorem 11 (Asynchronous Convergence Theorem [2]): If there is a sequence of nonempty sets with for all satisfying the following two conditions: 1) (Synchronous Convergence Condition) For all and . If is a sequence such that for all , then every limit point of is a xed point of . 2) (Box Condition) For every , there exist sets such that . and the initial solution estimate belongs to the set , is a xed point of . then every limit point of Note that and in the above statement correspond to and in the context of our work. To apply this theorem, we have to identify a suitable sequence of sets , which is crucial to the proof of the following theorem. Theorem 12: If a xed point exists, then the power vector converges to under the totally asynchronous model. as ; therefore, Proof: As before, denote we have . We dene (27) For all , the set satises the Box Condition. The twosided scalability implies that the sequence is monotonfor all . In the ically decreasing; hence, converges to . proof of Theorem 10, we have shown that such that for all Hence, any sequence must have a unique limit point , which is the xed point of . Since the initial power vector satises , the Asynchronous Convergence Theorem implies convergence to the xed point . VII. EXISTENCE OF THE FIXED POINT The convergence of a two-sided scalable iterative function is established under the condition that a xed point exists. In general, this may not be the case. Although Lemma 7 shows is decreasing and thus that the distance function converges, the power vector may not converge. Take the FoschiniMiljanic algorithm as an example. In matrix form, it can be expressed as follows: (28)

2630

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 51, NO. 7, JULY 2005

For wide-sense standard iterative functions, the existence of upper bound is sufcient to ensure that a xed point exists. We remark that a similar result for standard functions can also be found in [18]. Theorem 16: Given a wide-sense standard iterative function , if for all , then a xed point exists. as follows. If Proof: Construct a lower bound for is standard, then let . Otherwise, let . By positivity, is strictly greater than . By monotonicity and type-II monotonicity, is bounded below by . Consider , which is compact and convex. the set is a continuous mapping from this set to itself, by Since Brouwers theorem, a xed point exists. For type-II standard iterative functions, is always upperbounded by due to type-II monotonicity. Hence. we have the following result. Corollary 17: Given a type-II standard iterative function , if is nite, then a xed point exists. Corollary 17 can be applied to a system in which all terminals employ the opportunistic algorithm in (10). Note that equals and is nite for nonzero noise. As a consequence, a xed point exists and the system is always stable. Contrary to target tracking, the opportunistic approach does not require admission control. The system is always stable no matter how many users are admitted. In general, however, the boundedness of a two-sided scalable function does not imply the existence of xed points. An example is shown in the Appendix. to For easy reference, we summarize the conditions for have xed points as follows. 1) is two-sided scalable, upper- and lower-bounded by some positive vectors and . 2) is two-sided scalable, continuous and upperbounded. is wide-sense standard and upper-bounded. 3) is type-II standard and is nite. 4) VIII. A PROPERTY OF THE FIXED POINT In this section, we investigate how the xed point reacts to changes in the iterative function. In particular, we are interested in determining whether the xed point exhibits opportunism if the opportunistic algorithm is used. . Consider a certain realization of the link gain matrix The opportunistic power control algorithm in (10) is used, and be the xed point. Suppose that the link gain of the uplet link channel between user and her serving base station is in, while all other link gains remain creased by a factor of changes to , unchanged. When this occurs, not only to if user is connected to the same base stabut also tion as user . As a result, is increased by and is decreased for all other users connected to that base station. Denote the new xed point by . It is interesting to know whether . To tackle this problem, we formulate it in a more general setting. Two iterative functions and are dened. Both of

them are two-sided scalable, and they are related to each other as follows: (34) and for (35) where and is a function of satisfying (36) This setting is relevant to th problem at hand, since we may let for user connected to the same base station as user otherwise. Our question is answered by the following theorem. to Theorem 18: If the xed points of and and , respectively, then . exist and are equal (37)

Before proving this theorem, we construct a new iterative function and prove a lemma. Dene such that (38) Consider the sequence , where (39) (40) converges to Note that with xed point . Lemma 19: Let , since is two-sided scalable . We have

for all . Proof: We prove by mathematical induction. The state. Assume that it is true for ment is obviously true for , that is, (41) The th component of the statement holds for , since (42) Consider the th component, where (43) (44) (45)

SUNG AND LEUNG: A GENERALIZED FRAMEWORK FOR DISTRIBUTED POWER CONTROL IN WIRELESS NETWORKS

2631

Fig. 1. Convergence of power levels in a four-user systema comparison between the Foschini-Miljanic algorithm ( = 0:2) and the opportunistic algorithm ( = 1).

Denote the vector

by

. Then

Due to the induction hypothesis and (36) (46) By the two-sided scalability of (47) This completes the proof. Now we are ready to prove Theorem 18 Proof of Theorem 18: Suppose , that is, for some ; we prove by contradiction. By Lemma 19

Theorem 18 states that if user becomes more aggressive, she will be able to transmit at a higher power, provided that other users change their behaviors to a lesser extent. An example is the increase of link gain for user , as we described in the beginning of this section. Another application relates to the setting of the target level. When a user raises her target SIP, she will transmit at a higher power after the iterations converge. Thus, the target SIP may be used as a control parameter for resource allocation. The larger its value, the higher the transmit power at the xed point. IX. SIMULATION First we examine the convergence of the FoschiniMiljanic algorithm and the opportunistic algorithm. We consider a fouruser single-cell system. We assume the link gains of the users and . The noise power is are xed and equal to equal to one. Initially, all users transmit with unity power. For the FoschiniMiljanic algorithm, all users have the same target SIR . For the opportunistic algorithm, all users have the same target SIP . Fig. 1 shows the evolution of SIR for the four users. The graphs for these two algorithms are shown in Fig. 1(a) and (b), respectively. The power and hence the SIR of all users converge within 12 iterations. For opportunistic power control, the SIRs of the users converge to different values. It is because the instantaneous link gains of the users are different. A user having a larger gain gets a

(48) By the two-sided scalability of (49) By (34), we have

which contradicts with the rst inequality in (49).

2632

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 51, NO. 7, JULY 2005

Fig. 2.

Evolution of power levels in a two-user system.

higher SIR, which is the spirit of opportunistic communications. This issue will be further discussed in the next experiment. Next we use a simple two-user single-cell system to illustrate the idea of opportunistic communications. Both users experience Rayleigh fading. Flat Doppler spectrum with maximum Doppler frequency of 20 Hz is assumed, which models a typical indoor wireless environment. Large-scale fading effects such as path loss and shadowing are ignored. Noise power is normalized to . We simulate the system for 1000 power control iterations. Each iteration corresponds to 1 ms. Fig. 2 shows the change of link gains and the power evolution of the two users from iteration 100 to 300. A comparison and between the FoschiniMiljanic algorithm is made. Since the aim of the opportunistic algorithm the FoschiniMiljanic algorithm is to maintain a constant SIR, the power levels of the users are somehow reciprocal of their corresponding link gains. On the contrary, the opportunistic algorithm does just the opposite. Roughly speaking, a user transmits less power when she experiences a deep fade. She transmits more when her link gain is relatively large. The advantage of this design is that much larger throughput can to so that be obtained. In this example, we set the FoschiniMijanic algorithm consumes the same amount of power as the opportunistic algorithm. However, the throughput

obtained by the opportunistic algorithm is six times larger than that obtained by the FoschiniMiljanic algorithm. Finally, we evaluate the capacity improvement by opportunistic power control. We consider a ten-user single-cell system. Ten different scenarios are generated. In each scenario, the large-scale component of the link gains from the mobiles to the base station are normalized to one. On top of the large-scale effect, we assume that each link gain has a small-scale component due to Rayleigh fading. As before, at Doppler spectrum with maximum Doppler frequency of 20 Hz is assumed. Noise power is normalized to . For each scenario, we simulate the system for 5000 power control iterations. The performance result is averaged over the ten scenarios. Fig. 3 shows the total throughput against average power consumption for the FoschiniMiljanic algorithm and the opportunistic algorithm. For the FoschiniMiljanic algorithm, the curve is obtained by adjusting the target SIR, . The larger the value of , the larger the throughput but the more power is consumed. For the opportunistic algorithm, the curve is obtained by adjusting the target SIP, . The larger the value of , the more power is consumed and the larger the throughput. From the graph, we see that opportunistic power control dramatically increases the system capacity. For the same power consumption, the throughput is increased by about 2.8 times.

SUNG AND LEUNG: A GENERALIZED FRAMEWORK FOR DISTRIBUTED POWER CONTROL IN WIRELESS NETWORKS

2633

Fig. 3.

The tradeoff between throughput and power.

X. CONNECTION WITH THE GAME-THEORETIC APPROACH In recent years, game theory has been applied to the power control problems [10], [11], [14], [17]. In a power control game, each user determines his own power in order to maximize his . Formally, a power control game can be utility function stated as follows: for (50)

In this paper, we propose the opportunistic algorithm for throughput maximization. The iterative function may be viewed as the best response of this utility function (53) This function may be interpreted as each user trying to maximize the square root of his SIR with a linear cost on power usage. Alternatively, one may simply regard the opportunistic iterative function as the best response of the game in which every user keeps her SIP to a constant, that is, (54) This shows that the correspondence between the iterative function and the utility function is not unique. The same is true the other way around. Given a utility function, there are many ways to dene an appropriate iterative funcrepresents the tion other than the best response. Suppose is twobest response of user , and the vector function sided scalable. In order to avoid large uctuations on transmit with power, user may average her current transmit power the best response . The iterative function after this averaging operation may be expressed as (55) where . Suppose that (56)

For example, the traditional target tracking problem may be posed in the game-theoretic framework by dening (51) , the value of that maximizes is . This is called the best response of user . Therefore, for this game, we may dene the iterative function as Given (52) This is precisely the FoschiniMiljanic algorithm. This example illustrates that given a utility function, one can simply dene an iterative function as the best response. Our framework provides a valuable tool for analyzing the convergence. One can simply check whether the best response function is two-sided scalable. This methodology is used in [17]; the best response function is shown to be standard, and thus Yates framework applies.

2634

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 51, NO. 7, JULY 2005

Since

is two-sided scalable (57)

Consider a two-dimensional iterative function. Its rst component is dened as follows: (61)

By (56) and (57), we have (58) Hence, is also two-sided scalable. Therefore, two-sided scalability is preserved after power averaging. In game theory, there is an important solution concept called Nash equilibrium. Denition 3: A power vector every user is a Nash equilibrium if, for Dene and for . The second component of the iterative function is dened as follows:

(62)

It is obvious that the xed point is a Nash equilibrium if the iterative function is dened as the best response function. The same is true when power averaging is applied. XI. CONCLUSION In this paper, we present a new framework for distributed power control, which generalizes Yates result. We have shown that if an iterative function is two-sided scalable, the resultant power vector always converges, provided that a xed point exists. Sufcient conditions for the existence of xed point are described. The convergence is guaranteed for both synchronous and asynchronous systems, which is very desirable from a practical viewpoint. In Yates framework, the monotonicity property requires a user to increase his power when other users increase their powers. This restriction is removed in the new framework. Consequently, it may be applied to systems designed for opportunistic communications. We have obtained an algorithm for opportunistic power control. Its convergence is guaranteed by our framework. Simulation results show that it yields signicant improvement in throughput when compared with the traditional target-tracking approach. We hope that this work provides a guideline for designing more sophisticated power control algorithms in the future. APPENDIX EXAMPLE: A BOUNDED TWO-SIDED SCALABLE FUNCTION WITHOUT FIXED POINT In this appendix, we exhibit a bounded two-sided scalable function that has no xed point. Dene two classes of functions with parameters and as follows: (59) (60) Note that they have the properties that . and

It can be veried that is two-sided scalable. Furthermore, is upper-bounded by . Let the initial vector be . It is obvious that decreases monotonically and converges to whereas oscillates between and . If a xed point exists, Theorem 10 implies the convergence of the power vector. Now the power vector does not converge. Hence, there is no xed point. ACKNOWLEDGMENT The authors would like to thank the Associate Editor and the reviewers for their helpful and stimulating comments. REFERENCES
[1] N. Bambos, S. C. Chen, and G. Pottie, Radio link admission algorithms for wireless networks with power control and active link quality protection, in IEEE Proc. INFOCOM95, Boston, MA, Apr. 1995, pp. 97104. [2] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computation. Englewood Cliffs, NJ: Prentice-Hall, 1989. [3] K. C. Border, Fixed Point Theorems with Applications to Economics and Game Theory. Cambridge, U.K.: Cambridge Univ. Press, 1985. [4] G. J. Foschini and Z. Miljanic, A simple distributed autonomous power control algorithm and its convergence, IEEE Trans. Veh. Technol., vol. 42, pp. 641646, Nov. 1993. [5] M. Grossglauser and D. N. C. Tse, Mobility increases the capacity of adhoc wireless networks, IEEE/ACM Trans. Netw., vol. 10, no. 4, pp. 477486, Aug. 2002. [6] S. Hanly and D. N. C. Tse, Multi-access fading channels: Part II: Delay-limited capacities, IEEE Trans. Inf. Theory, vol. 44, no. 7, pp. 28162831, Nov. 1998. [7] A. Jalali, R. Padovani, and R. Pankaj, Data throughput of CDMA-HDR a high efciency high data rate personal communication wireless system, in Proc. IEEE Vehicular Tehnology Conf., Tokyo, Japan, May 2000, pp. 18541858. [8] R. Knopp and P. A. Humblet, Information capacity and power control in single-cell multiuser communications, in IEEE Proc. int. Conf. Communications (ICC95), Seattle, WA, Jun. 1995, pp. 1822. [9] K. K. Leung, C. W. Sung, W. S. Wong, and T. M. Lok, Convergence theorem for a general class of power control algorithms, IEEE Trans. Commun., vol. 52, no. 9, pp. 15661574, Sep. 2004. [10] C. U. Saraydar, N. B. Mandayam, and D. J. Goodman, Pricing and power control in a multicell wireless data networks, IEEE J. Select. Areas Commun., vol. 19, no. 10, pp. 18831892, Oct. 2001. [11] , Efcient power control via pricing in wireless data networks, IEEE Trans. Commun., vol. 50, no. 2, pp. 291303, Feb. 2002. [12] C. W. Sung and W. S. Wong, A distributed xed-step power control algorithm with quantization and active link quality protection, IEEE Trans. Veh. Technol., vol. 48, no. 2, pp. 553562, Mar. 1999. [13] , Mathematical aspects of the power control problem in mobile communication systems, in Lectures on Systems, Control and Information: Lectures at the Morningside Center of Mathematics, L. Guo and S. S.-T. Yau, Eds. Providence, RI: AMS/International Press, 2000, pp. 127172.

SUNG AND LEUNG: A GENERALIZED FRAMEWORK FOR DISTRIBUTED POWER CONTROL IN WIRELESS NETWORKS

2635

, A noncooperative power control game for multirate CDMA data networks, IEEE Trans. Wireless Commun., vol. 2, no. 1, pp. 186194, Jan. 2003. [15] D. N. C. Tse and S. Hanly, Multi-access fading channels: Part I: Polymatroid structure, optimal resource allocation, and throughput capacities, IEEE Trans. Inf. Theory, vol. 44, no. 7, pp. 27962815, Nov. 1998. [16] P. Viswanath, D. N. C. Tse, and R. Laroia, Opportunistic beamforming using dumb antennas, IEEE Trans. Inf. Theory, vol. 48, no. 6, pp. 12771294, Jun. 2002.

[14]

[17] M. Xiao, N. B. Shroff, and E. K. P. Chong, A utility-based power-control scheme in wireless cellular systems, IEEE/ACM Trans. Netw., vol. 11, no. 2, pp. 210221, Apr. 2003. [18] R. D. Yates, A framework for uplink power control in cellular radio systems, IEEE J. Select Areas Commun., vol. 13, no. 7, pp. 13411347, Sep. 1995. [19] J. Zander, Performance of optimum transmitter power control in cellular radio systems, IEEE Trans. Veh. Technol., vol. 41, no. 1, pp. 5762, Feb. 1992.

You might also like