Sie sind auf Seite 1von 11

792

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 3, MARCH 2011

Resource Allocation for the Parallel Relay Channel with Multiple Relays
Ka an Bakano lu, Student Member, IEEE, Stefano Tomasin, Member, IEEE, g g and Elza Erkip, Fellow, IEEE

AbstractA cooperative network where the transmission between two nodes is assisted by many half-duplex relays over parallel Gaussian channels is considered. The parallel channel model is suitable for a broadband system, such as orthogonal frequency division multiplexing or a block fading channel. For the decode-and-forward protocol, an optimization problem for joint power, time and subchannel allocation under per-node power constraints is formulated to maximize the total transmission rate between the source and the destination. To solve this optimization problem, rst the optimal power allocation for a given subchannel allocation is found. Then a greedy algorithm that jointly allocates subchannels and power is described. Finally, the time allocation is optimized by a numerical search procedure. The limiting case where the number of subchannels goes to innity is also studied. Numerical results reveal that the achieved rate for the innite number of subchannels is an upper bound for the nite subchannel case and the proposed greedy algorithm results in rates close to those for innite number of subchannels when the number of subchannels is sufciently large. Furthermore, most of the cooperative gains can be achieved by the use of a small number of relays. Index TermsResource allocation, parallel relay channel, user cooperation.

I. I NTRODUCTION OOPERATION among nodes of a wireless network is known to improve system performance [1], [2], [3], by implementing a distributed multiple input-multiple output (MIMO) system. For broadband systems orthogonal frequency division multiplexing (OFDM) provides a set of narrowband channels, which typically experience different fading levels. When multiple channels are present, allocating resources such as power among the various subchannels can provide a signicant performance improvement [4]. Similarly, a block fading channel can be modeled as a set of parallel channels, one for each fading state, and under channel state information at the transmitter (CSIT) power allocation over different fading states can improve the achievable rates in the ergodic case with innitely many fading blocks [5] as well as the outage probability when the number of fading blocks is nite [6]. An overview of resource allocation techniques applied to relay networks is provided in [7]. Optimal power allocation

for amplify and forward (AF) cooperation is investigated in [8] for fading channels. For the decode and forward (DF) strategy power and time allocation is optimized in a network with a single relay operating on a single at fading channel, with perfect CSIT at all nodes [10]. An analysis based on the outage probability for the case of partial CSIT and DF relaying strategy is proposed in [9]. In [10] the outage performance of a cooperative system is explored with a short-term power constraint for various relaying and resource allocation schemes. Assuming partial CSIT and DF relaying, time and power are optimized in [11], under a constraint on the average total system power, with the goal of either maximizing the delay-limited capacity or minimizing the outage probability. Ergodic capacity for a DF relay system has also been explored under various assumptions on the CSIT (see [11], [12] and references therein). Resource allocation for ergodic multiaccess relay channel is provided in [13] where the relay has an orthogonal channel to the source and uses DF. In [14], a broadband relay channel is analyzed where each link is composed of parallel independent Rayleigh fading channels while power and transmission time are dynamically allocated either to improve the delay-limited capacity or to decrease the outage probability. Power allocation is optimized in [15] for a network where all channels are orthogonal and a relay serves multiple source/destination nodes. The above referenced papers mostly deal with a single relay. Resource allocation is even more critical when multiple relays are available. In [16], [17], resource allocation for multiple relays is studied when relays are available in clusters around the source and the destination under the assumption of partial CSIT in the form of channel state amplitudes. A multi-relay network operating on a at-fading channel using DF and distributed space-time coding is considered in [18] where both power and time are optimized. Opportunistic AF relaying is investigated for multi-relay networks in [19] and [20]. The case of a network with a single source, a single destination and multiple relays is considered in [21], where all channels among nodes are orthogonal and pre-assigned. Resource allocation for a wireless multihop network with multiple sources, multiple relays and one destination with OFDM is studied in [22] and [23]. In [22], the end-to-end rate is maximized in a two hop scenario without the direct link. In [23] the single source, single destination and multiple relays with one relay per hop scenario is investigated in detail. In this paper we study resource allocation, namely subchannel, power and time allocation, in order to maximize the end-to-end rate for the parallel, half-duplex, single sourcedestination, multi-relay channel. We assume that the relays

Manuscript received November 12, 2009; revised May 2, 2010 and September 24, 2010; accepted September 28, 2010. The associate editor coordinating the review of this paper and approving it for publication was I.-M. Kim. K. Bakano lu and E. Erkip are with the Dept. of Electrical and Computer g Engineering, Polytechnic Institute of New York University, Brooklyn, NY 11201, USA (e-mail: kbakan01@students.poly.edu, elza@poly.edu). S. Tomasin is with the Department of Information Engineering, University of Padova, Padova 35131, Italy (e-mail: tomasin@dei.unipd.it). This work is partially supported by NSF grants No. 0520054, 0635177, 0905446. Digital Object Identier 10.1109/TWC.2011.011111.091682

c 1536-1276/11$25.00 2011 IEEE

BAKANOGLU et al.: RESOURCE ALLOCATION FOR THE PARALLEL RELAY CHANNEL WITH MULTIPLE RELAYS

793

employ DF. As discussed above, the parallel channel model can be used to analyze not only broadband channels with OFDM, but also narrowband channels with block fading [24]. The problem that we address can be thought of as a combination of downlink (source to relays) and uplink (relays to destination) resource allocation problem for multiuser parallel channels, which are both well investigated [24]. However, because of the DF protocol, we aim at maintaining rate matching at each relay, i.e. we bound the rate reaching the destination from a relay by the incoming rate at the relay. This constraint makes the resource allocation problem more challenging. Furthermore, our model also contains a direct link from the source to the destination. In this context, we formulate the optimal resource allocation problem taking into account some practical OFDM implementations, explained in detail in Section II, to maximize the end-to-end rate. We establish the optimal power allocation for a given subchannel assignment and we propose a greedy algorithm for end-toend rate maximization, which jointly allocates subchannels and power. The proposed algorithm can be implemented both in a centralized way with full CSIT at a single node and also in a distributed fashion based on local CSIT. As end-to-end rate turns out to be neither a convex nor a concave function of time sharing between the phases of transmission from the source and from the relays, we resort to a search procedure to optimize time allocation. A similar problem to ours is studied in [25] with total power constraint among the source and the relays. However, the solution in [25] is obtained under some limiting assumptions based on subchannel matching and the direct channel between the source and the destination is not utilized. A xed subchannel allocation is considered instead in [26] where time allocation is used to avoid outages in source/relay and relay/destination transmission. In order to build a benchmark for the nite subchannel case, we also study the limiting case where the number of subchannels is innite and the link gains follow an ergodic statistical model. This statistical nature simplies the resource allocation problem. The solution for this limiting case gives an achievable rate for the large bandwidth OFDM system and also a lower bound to the ergodic capacity of fading multirelay channel with CSIT, both under practical constraints of Section II. The remainder of this paper is organized as follows. In Section II the system model and the optimization problem are introduced. The optimization is solved in Section III for an nite number of subchannels and in Section IV for an innite number of subchannels. Numerical results and conclusions follow in Section V and VI. II. S YSTEM M ODEL A ND P ROBLEM F ORMULATION We consider a parallel relay channel with a source node S, a destination node D and relay nodes R , = 1, 2, . . . , , employing DF. After decoding, the relays send independent parity bits to enable cooperative coding [30]. We assume the relays only receive signals from the source, hence there is no inter-relay communication. Transmissions are performed over parallel subchannels, each of which is corrupted by independent, unit variance complex additive white Gaussian noise. This could for example model an OFDM system where each subchannel gets 1/ of the total bandwidth. Alternatively,

Fig. 1. Parallel relay channel with multiple relays. The gure illustrates one of the subchannels.

the model is suitable for a block fading channel where each subchannel represents a different fading state. For subchannel , , and , denote the power gain of the S R and the R D link, respectively, while , denotes the power gain of the S D link. This is illustrated in Fig. 1. We assume that only channel gains are known and the phase information is not available, hence our scenario is also applicable to phase fading as in [31]. We rst consider a centralized resource allocation scheme, where the source is aware of all channel gains and broadcasts to the relays and the destination the optimized power and time allocation strategy. This can be accomplished for example when the destination estimates the channel gains from relays and then forwards the estimates to the source through the relays, along with each relayss estimated channel gains from the source. We also discuss a possible distributed implementation of the algorithm in Section III-C where only local CSIT is required. Relays are assumed to be half-duplex and relaying in subchannel takes place in two phases comprising two time slots: phase 1, during which R listens, and phase 2 during which R transmits. Source and relay power constraints are S and R , = 1, 2, . . . , , respectively, which are to be satised over the transmission period. Source power S represents the average over phase 1 and 2. We make the following simplifying assumptions based on practical considerations. 1) Due to the lack of phase knowledge and the difculty of synchronizing the relays, cooperative beamforming is not used. Superposition coding at the source or simultaneous transmissions of relays on the same subcarrier are also excluded as they require complex multiuser detection techniques such as interference cancelation. 2) From 1) we obtain that at most one relay transmits on a given subchannel at any time and there is no advantage in having two or more relays receiving the same information on the same subcarrier from the source. 3) For a total unitary transmission time, phase 1 has duration , while phase 2 has duration 1. Although in general the duration of the two phases could be different for each relay and even for each subchannel, in practice it is difcult to have the same node transmitting and receiving at the same time on different subchannels. Therefore, we focus on the simpler case of same phase duration for all relays. 4) Some subchannels may be used for the whole time by the source only, in what we call direct transmission.

794

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 3, MARCH 2011 (1 ) , , R =1

A. Achievable Data Rates Let us dene the indicator function for relay and subchannel at phase 1 as { 1 subchannel allocated to R in phase 1 , = (1) 0 otherwise . Similarly, let , be the indicator function for relay and subchannel at phase 2. Lastly, let and be the indicator functions for direct transmission on subchannel in phase 1 and 2, respectively. The source power allocated to subchannel in phase 1 and 2 is and , respectively, and power allocated by R to subchannel in phase 2 is , . Note that variables with bars belong to phase 2. Dening the total S R data rate in phase 1 as , and and assuming the total R D data rate in phase 2 as Gaussian codebooks, we have

(7c) (7d) (7e)

, , , 0 , 0 1 + +
=1 =1

, 1 ,

{0, 1} , {0, 1} ,

, {0, 1} , (7f) , {0, 1} , .

, 1 ,

= =

, log2 (1 + , ) =1 1 , log2 (1 + , , ). =1

(2) (3)

Note that the factor 1/ accounts for the fact that each subchannel gets 1/ of the bandwidth (or the time). We assume that D overhears the transmission from the source to relays in phase 1 and uses cooperative combining to increase the data rate. Hence, we denote the mutual information collected at node D during phase 1 transmission to R as
, = , log2 (1 + , ). =1

(7g) In this problem, the objective function (7a) reects the fact that the maximum rate achieved in a DF relayed transmission is the minimum of the S R and the (S, R) D rates. We denote this constraint as rate-matching constraint, since optimization is achieved when the rate at which each relay receives is matched to the corresponding rate at the destination. Constraints (7b) and (7c) are power constraints for the source and the relays, respectively. The total transmit duration is unitary and half-duplex operation is assumed, as indicated in (7e). Lastly, constraints due to single node transmission/reception on each subchannel are reected by (7f)-(7g). III. O PTIMIZATION FOR F INITE N UMBER OF S UBCHANNELS The rate matching constraint couples phase 1 and phase 2, leading to a challenging problem that requires joint optimization of both phases. Moreover, the optimization is a mixed integer programming problem since it includes both set selection (subchannel assignment) and continuous variable (power and time) optimization. Lastly, since the target function is a minimum of two functions, the problem is neither convex nor concave, and it can be easily shown that it can not be solved in polynomial time. Therefore, we focus on a greedy suboptimal solution in two steps. In Section III-A, we rst nd optimal power allocation for given subchannel and time allocation. Then, based on this solution, in Section III-B a greedy algorithm, which jointly allocates subchannels and power, is proposed. Finally, we numerically optimize the time allocation. A. Power Allocation For Given Subchannel and Time Allocations In this section, we investigate the optimum power allocation that achieves the maximum end-to-end rate for a given subchannel and time allocation. In this case, in the optimization problem (7) we drop constraints (7e)-(7g), since now the maximization is only over the allocated power. In the following we will use the function ()+ = for > 0 and ()+ = 0 for < 0. Proposition 3.1: For a given subchannel allocation, , , , , and , and time allocation, , the optimal power allocation in phase 2 for R is found by waterlling over allocated subchannels, i.e. ( )+ 1 1 , = , , (8) ,

(4)

Lastly, for direct transmission the S D rates from phase 1 and phase 2 are
= log2 (1 + , ) =1

(5)

1 = log2 (1 + , ). =1

(6)

B. Rate Maximization Problem We aim at nding the optimum subchannel allocations, , , , , and , power allocations, , and , , and time allocation , which maximize the normalized endto-end achievable rate for a certain realization of the channel power gains. The resulting optimization problem can be formulated as Maximize
max + + =1 min( , , + )

(7a)

] 1 ( + , ) + (1 ) =1 =1 =1 (7b)

subject to

BAKANOGLU et al.: RESOURCE ALLOCATION FOR THE PARALLEL RELAY CHANNEL WITH MULTIPLE RELAYS

795

where is chosen to satisfy the power budget for R in (7c). Let , and , be the rates in (2)-(4) with optimal power allocation and , for = 1, . . . , , = 1, . . . , . Then the optimal power allocation of the source in phase 1 and phase 2 respectively are given by ( )+ 1 1 , , = 1 ( )+ 1 , , < + , , , = 1 = ( )+ 1 , > + , , , = 1 , (, )+ , = + , , , = 1 (9) and ( )+ 1 1 , for = 1, (10) = , where the constraints , and are chosen to satisfy source power constraint (7b) and 2 , 4, , , = (11) 2 2 1 1 + , = (12) + , , 1 , = . (13) , , , , Proof: See Appendix A. Proposition 3.1 states that the optimal power allocation for R is obtained by waterlling over the subchannels that are assigned to R . The source applies waterlling to the subchannels allocated to the direct transmission in both phases. The power allocation strategy on the remaining subchannels strictly depends on the maximum rate that R can forward and the source power budget, S . For in phase 2, < + , , the source will apply waterlling on the relayed subchannels, using gain , with an aim of increasing the S-R data rate, . For > + , , the source will apply waterlling on the subchannels using gains , with an aim of increasing S-D data rate, , . When rates are matched at R , the optimum source allocation is determined by the last case of (9). B. Greedy Algorithm for Joint Subchannel and Power Allocation We next propose a greedy subchannel assignment algorithm along with the corresponding optimal power allocation for a given time allocation. Let denote the end-to-end rate obtained by the optimal power allocation for a given subchannel assignment as described in Section III-A. Recall that each relay can only forward the minimum of phase 1 and phase 2 rates due to the DF assumption. Thus, a mismatch of phase 1 and phase 2 rates for the relays is not favorable. A good algorithm has to decrease the mismatch by allocating more resources to the bottleneck side of the transmission. Motivated by this observation, our greedy algorithm jointly allocates subchannels and powers. The main idea is to increase the end-to-end rate at each step of the algorithm by keeping the relay rate mismatch small.

TABLE I G REEDY A LGORITHM FOR J OINT S UBCHANNEL AND P OWER A LLOCATION

1) Initialization a) Set = {1, . . . , } and = {1, . . . , } which are the available subchannels in phase 1 and 2, respectively. b) Set , = 0, , = 0, = 0 and = 0 for all . 2) Until = and = a) Set = , = , = and (or) = ; b) Find = arg max , ; c) Find using when is temporarily allocated to direct transmission in phase 1, that is = 1. Set = ; d) Find = arg max , ; e) Find using when is temporarily allocated to direct transmission in phase 2, that is = 1. Set = ; f) For = 1 to i) Find = arg max , and = arg max , ; 1 ii) Find using when subchannel is temporarily allocated to relay in phase 1, that is , = 1; 2 iii) Find using when subchannel is temporarily allocated to relay in phase 2, that is , = 1; 3 iv) Find using when subchannels and are temporarily allocated to relay in phase 1 and phase 2, respectively, that is , = 1 and , = 1; 1 2 3 v) Find = max( , , ). The maximum suggests which phase(s) to allocate the additional subchannel to R in this intermediate step; vi) Based on the subchannel allocation achieving the maximum in step 2(f)v above, set = and (or) = . g) = arg max( ) as the relay which leads to the highest end-to-end rate; h) If + = max( + , ), update = 1; = 1 or i) If = max( + , ), update , = 1 and (or) , = 1; j) Based on the results of 2h and 2i, update = and = , or = and = . (or) The proposed greedy algorithm is reported in Table I and works as follows. First of all, we initialize the sets , collecting the available subchannels in phase 1 and phase 2 respectively. We then evaluate the end-to-end data rates when the channels with the highest gains among those available in phase 1 and (or) phase 2 are temporarily used either for direct transmission (steps 2c and 2e in Table I) or for relay transmission through each relay R (step 2f). We then select the conguration

796

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 3, MARCH 2011

providing the highest end-to-end rate, i.e. either direct or relayed transmission. For the relayed transmission, the relay that provides the highest end-to-end rate is chosen. When evaluating the end-to-end rate, power allocation is performed as described in Section III-A. The subchannels are allocated according to the selected conguration and the set of available subchannels is updated until all subchannels are allocated. To nd the optimal subchannel assignment, the brute-force method searches over all possible subchannel assignments and evaluates for each subchannel allocation, involving ( + 1)2 evaluations of . Thus the complexity of the optimum solution increases exponentially with the number of subchannels . For the greedy algorithm Step 2 involves 3 + 2 evaluations of and at least one subchannel is allocated to the corresponding node after each evaluation. Since the total number of subchannels is , the total number of step 2 evaluations is given by , where 0 1. Thus the proposed greedy algorithm evaluates ( )3+2 times. Therefore, its complexity increases polynomially with the number of subchannels . Although the complexity is exponential in the number of relays, this is signicantly lower than that of the brute-force method since the number of relays is typically much smaller than the number of subchannels, . Finally for the optimal time allocation among phase 1 and phase 2, we observe that the maximum rate achieved as a function of the allocated time is neither concave nor convex. This will be further illustrated in Section V. As a closed-form optimization of the time sharing seems infeasible, we resort to a numerical method, i.e. we enumerate possible values of in the interval [0, 1] with a sufciently tight step and select the value of that provides the maximum sum-rate. Results of this optimization are also provided in Section V. C. Distributed Implementation of the Greedy Algorithm We now consider a possible distributed implementation of the greedy algorithm, where subchannels and powers are allocated based on local channel knowledge. We assume local channel amplitude knowledge at each node, i.e. node S is aware of the S D and S R channel gains, while R knows R D channel gains. These CSITs can be easily obtained if the channel is used in a time-duplex fashion and reciprocity of the channel can be exploited, or by use of local feedback. The greedy algorithm of Table I is applied iteratively. We observe that the core of the GA algorithm is the solution of which can be split into two parts: solution of (8) and solution of (9)-(13). Solution of (8) can be found locally at each relay only with the knowledge of the allocated subchannels in phase 2 and the CSITs of the R D links. The solution of (9)-(13) instead requires both CSIT of the source and the maximum achievable rates in phase 2 ( ) for all relays, therefore, we propose to implement the second part at the source. In particular, the greedy algorithm of Table I is performed as follows by splitting each iteration into two actions. The rst action is performed by each relay , which computes locally both and using (8) and then sends the computed rate and subchannel index to the source. The second action is performed by the source, which applies step 2) of the GA to solve and then broadcasts the allocated subchannel

index in either or both of the phases and the new rates of phase 2 to the relays. IV. R ESOURCE A LLOCATION FOR AN I NFINITE N UMBER OF S UBCHANNELS In the following we study the optimization problem for an innite number of subchannels, that is and incorporate the statistical nature of the channel gains. We assume that the subchannel gains are i.i.d. both across relays and across subchannels. This model is applicable to a slow fading broadband frequency selective channel or to a at fast fading channel. In both scenarios, since there are innite subchannels or fading states, the channel statistics are fully observed. Hence, for the innite subchannel case the resource allocation can be solved based on channel statistics instead of a particular set of subchannel gains as in Section III. This is another motivation for considering the innite subchannels. In the innite subchannel case, dropping the subchannel indices, the realizations of the power gain of the S R and the R D link are denoted by and , respectively, while denotes the power gain of the S D link. With this notation the random vectors H = [1 , 2 , , , ] and H = [1 , 2 , , ] denote the power gain vectors for the source to relays links and relays to the destination links, respectively. The corresponding power allocations are denoted by random variables (H, H) and (H, H) for S and R , respectively. Similarly, subchannel allocations are random variables (H, H) and (H, H) for R in phases 1 and 2 respectively, and (H, H) and (H, H) for the source. As stated in the Introduction, this problem is a combination of downlink (phase 1) and uplink (phase 2) resource allocation problems. However, the general problem (7) is not a simple cascade of downlink and uplink problems, since phase 1 and phase 2 problems are intertwined due to cooperative combining at the destination. Therefore, phase 1 and phase 2 resource allocations have to be addressed jointly by solving the constrained nonlinear minimax problem in (7), hence two phases can not be separated into two different problems. In order to nd a closed form solution to innite subchannel case, we assume that in phase 1, the destination only listens to the subchannels allocated to direct transmission, but not those allocated to the relays. Hence, cooperative combining is not used. Then, the optimization problem (7) for becomes Maximize
max

+ +

=1

( ) min , , , + ,, (14a) )] (14b) (14c) (14d) (14e) (14f) + (1 ) subject to )

[ (( E +

=1

(1 )E[ ] R {1, . . . , } 0 , 0 for all 0 1 +


=1

1 and , {0, 1} for all

BAKANOGLU et al.: RESOURCE ALLOCATION FOR THE PARALLEL RELAY CHANNEL WITH MULTIPLE RELAYS =1

797

+ where

1 and , {0, 1} for all

(14g)

to the rst subproblem let and denote the optimized values for a given . Finally, rate matching at the relays is enforced on the maximization of the sum rate by solving
,

(, ) = E[ log2 (1 + )] , ( , ) ,, ( , ) (, ) , ( , )

(15) (16) (17) (18) (19)

= E[ log2 (1 + )] = E[ log2 (1 + )] = (1 )E[ log2 (1 + )]

max

=1

{ } min , ( , ), , ( , ) (22)
=1

subject to

= 1 and

=1

= 1.

= (1 )E[ log2 (1 + )]

All expectations E[] are taken with respect to (H, H) and , , , , , and are functions of (H, H). Explicit dependence on (H, H) is not shown to simplify the notation. A. No Direct Link As a rst step to solve (14), we start by analyzing the simpler case where we do not utilize direct link from the source to the destination. Hence, we set = 0, = 0, = 0, = 0 and = 0. The optimization problem (14) now becomes the combination of resource allocation problems in fading broadcast channel (BC) (phase 1, source to relays) and multiple access channel (MAC) (phase 2, relays to destination) under constraints of Section II and rate matching at relays. The ergodic capacities for BC and MAC (without constraints) are studied in [27] and [28], [29], respectively. In [27], the ergodic capacity region of an user fading BC is derived for code division, time division and frequency division. Power allocation for fading MAC is studied in [29] where the special case of symmetric users with equal rate requirements is solved, while the ergodic capacity region of fading MAC is found in [28]. We divide the optimization problem in (14) into three subproblems: 1) nding rate regions for phase 1 (BC) under (14f), 2) nding rate regions for phase 2 (MAC) under (14g) and 3) enforcing rate matching at relays. In the rst subproblem, for a given we characterize the boundary of the BC rate region subject to the condition that each subchannel serves only one relay (14f). Note that the boundary can be obtained by solving the following resource allocation problem
,

Note that for all (, ) the rate tuples { } (min 1, (1 , ), 1, (1 , 1 ) , . . . , { } min ( , ), ( , ) )


, ,

(23)

fall in the intersection of the phase 1 and phase 2 rate regions and include all the boundary points of the intersection. The optimization problem (22) nds the point on the matched region that maximizes the sum rate. For the solution of (20) and (21) we use the approach of [27], and decompose each problem into two sub-problems: 1) resource allocation among users for a certain channel state (H, H) 2) resource allocation among channel states. Proposition 4.1: The optimal solution of (20) for a given and can be written as ( )+ 1 , for = , = (24) for all = = 1, and = 0 ( ) where = arg max ( 1 )+ . () = log2 (1 + ) and is found by solving (14b) for = , and = . Proof: Details of the proof can be found in Appendix B. Proposition 4.2: The optimal solution of (21) for a given and can be written as { ( )+ 1 for = = (25) 0, for = = 1, and = 0 for all = , ( ) where = arg max ( 1 )+ . () = ) and , = 1, 2, . . . , are found by log2 (1 + solving (14c) jointly for all = 1, 2, . . . , with = , = . Proof: Details of the proof can be found in Appendix C. Remark 4.1: Lemma 4.1 suggests that S applies waterlling over the channels allocated to Rm where Rm is the relay leading to the maximum weighted rate in phase 1. In Lemma 4.2, each relay is used only when it results in the largest weighted rate. Furthermore, each relay applies waterlling over the used channel states. B. Direct Link Available In this section, we study the most general setup of (14). We will observe that the insights obtained from the special case of Section IV-A will be useful in solving the resource allocation problem. Similar to Section IV-A we treat the system as a cascade combination of BC and MAC. In order to incorporate the

max

=1

, ( , )

(20)

subject to (14b), (14d) and (14f) for all = 1 , . . . , ] [ such that 0 1, = {1, . . . , } and =1 = 1, [27]. For a xed , let and denote the optimized values according to (20). Dependence on is not shown for simplicity in the notation. Similarly, the second subproblem, namely nding the boundary of MAC for xed , under the condition that each relay transmits in a separate subchannel (14g) can be written as , ( , ) , (21) max
, =1

subject to (14c) and (14g) for all = [1 , . . . , ], with 0 1, = {1, . . . , } and =1 = 1. Similar

798

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 3, MARCH 2011

direct link, the destination and the source are treated as the virtual relay + 1 in phase 1 and phase 2, respectively. In phase 1 the source transmits to the destination (which acts as a virtual relay) as well as to the relays. In phase 2 the source acting as a virtual relay transmits together with the actual relays. Similar to Section IV-A, we divide the optimization problem (14) into three subproblems: 1) nding rate regions for phase 1 (BC) under (14f), 2) nding rate regions for phase 2 (MAC) under (14g) and 3) maximizing the sum rate under rate matching. Note that the source transmits in both phases and the source power has to be divided among the two phases. Thus we introduce a new parameter to be optimized, 0 1, where S is the power used in phase 1 and (1 )S is the power used in phase 2. In the rst subproblem, for a given we characterize the boundary of the BC rate region subject to the condition that each subchannel serves only one relay (14f). Note that the boundary can be characterized by solving the following resource allocation problem
, , max +1 (, ) + +1 =1 =1 , ( , ),

the following two modications: 1) the source is the ( + 1)th virtual relay in phase 2 and 2) the power budget of the ( + 1)th relay, which is in fact the source in phase 2, is (1 ) . However, problem (28) differs from (22) for the presence of the ( + 1)th virtual relay. Remark 4.2: Resource allocation problems (26), (27) and (28), boil down to problems (20), (21) and (22) of Section IV-A, when we set = 0 and = 0 or = 0. V. N UMERICAL R ESULTS To evaluate the performance of the proposed algorithms, without loss of generality we assume that the source and the destination are at unitary distance and the relays are located on the straight line connecting the source to the destination. Let d = [1 , . . . , ] denote the distance vector where is the distance between the source and relay . Each parallel subchannel gain is chosen independently according to Rayleigh distribution with mean given by pathloss model. We assume path loss with exponent 3. The end-to-end rate is optimized for each channel realization. The average end-to-end rate over 10000 channel realizations is computed. For innite number of subchannels, the statistical solution in Section IV is utilized. For the nite subchannel case, we compare the end-to-end rate obtained by the greedy algorithm (GA) of Table I with the following schemes. Direct transmission (DT), where the source transmits to the destination and the relays are not utilized. The optimum rates are found by waterlling [33] and denoted by . No direct link (NDL), where the source transmits only to the relays in phase 1 and only relays transmit to the destination in phase 2. The destination does not utilize the direct signal form the source in any fashion. The end to end rate, denoted by , is obtained by assuming , = 0 for all = 1, 2, . . . , in the proposed greedy algorithm of Table I. No direct link, suboptimal (NDL-S) case investigated in [25] where subchannels are allocated to phase 1 and phase 2 in pairs in each iteration of the proposed algorithm of [25] in addition to ignoring the direct link signal at the destination. The end to end rate achieved is denoted by . No cooperative combining (NCC), where the destination does not overhear any message from the subchannels allocated to the relays in phase 1, but receives information on the subchannels allocated to direct transmission. The end-to-end rate for this scheme is denoted by and evaluated by the proposed greedy algorithm by letting = 0 in the optimization problem (7). Greedy algorithm, t=0.5 (GA-H), which is obtained by the greedy algorithm without optimized time allocation. Specically we let phases 1 and 2 last each half of the entire transmission time ( = 0.5) regardless of the channel gains. The resulting rate is denoted by . Greedy algorithm, equal power allocation (GA-EP) is an equal power allocation scheme denoted by , where each node (including the source) equally splits the power among the available subchannels (hence Prop. 3.1

(26)

subject to E[( +

) ]

with (14d) and (14f) for all = [1 , . , +1 ] such that .. +1 0 1, = {1, . . . , + 1} and =1 = 1. For a xed , let , and denote the optimized values of the above problem. Similarly, the second subproblem, namely nding the boundary of MAC for xed , under the condition that each relay transmits in a separate subchannels (14g) can be written as
, ,

max +1 (, ) +

=1

, ( , ) (27)

subject to (1 )E[ ] (1 )S . . with (14c), (14d) and (14g) for all = [1 , . , +1 ], with +1 0 1, = {1, . . . , + 1} and =1 = 1. Similar to the rst subproblem, for a xed , let , , and denote the optimized values. For sum rate maximization with rate matching we solve
, , max ( , ) + ( , )

=1

( ) min , ( , ), , ( , ) (28)
+1 =1

subject to

+1 =1

= 1 and

= 1 and 0 1.

The solution to (26) can be easily found by Lemma 4.1 for + 1 relays with the following two modications: 1) the destination is the ( + 1)th virtual relay in phase 1 and 2) the source power budget in phase 1 is . Similarly, the solution of (27) can be found by Lemma 4.2 for + 1 relays with

BAKANOGLU et al.: RESOURCE ALLOCATION FOR THE PARALLEL RELAY CHANNEL WITH MULTIPLE RELAYS

799

20 18 16 14 12 10 8 6 4

RO, Optimum solution RGA, Greedy algo. RGAEP, Greedy algo., equal power R R
GAH NCC NDL

, Greedy algo., t=0.5

, No cooperative combining

5 Expected rate (bps per Hz)

Expected rate (bps per Hz)

, No direct link

RNDLS, No direct link, suboptimal RDT, Direct transmission

NCCI

, No cooperative combining, infinite subchannels

RNCC, No cooperative combining RNDLI, No direct link, infinite subchannels 1


2 0

RNDL, No direct link R


DT

, Direct transmission 8 N 10 12 14

10

15

20

25 P (dB)

30

35

40

45

50

Fig. 2. Total expected end-to-end rate as a function of the source and relay power = S = R where = 4 and = 2. The distance vector is d = [0.3, 0.5].

Fig. 3. Total expected end-to-end rate as a function of number of subchannels where = 2, = = 10 dB and d = [0.3, 0.5].

is not used) while subchannels are still allocated using the greedy algorithm of Table I and time allocation is optimized. Optimum solution (O) denoted by is the optimal solution of (7). Note that this optimal solution is computationally complex and the numerical solution is unfeasible for large number of relays and subchannels as discussed in Section III. However, we numerically evaluate the optimum solution for small number of subchannels and relays, namely = 4 and = 2, by searching all possible subchannel allocations and using the proposed optimal power allocation in Section III-A for each subchannel allocation. For the innite subchannel case, we evaluate the end to end rates for the following cases: No direct link (NDL-I) studied in Section IV-A. The end to end rate is denoted by . No cooperative combining (NCC-I) studied in Section IV-B. The end to end rate is denoted by . Fig. 2 shows the total end-to-end rate as function of the source and relay power S = R , with = 4 subchannels and = 2 relays. In order to compute the performance of O, is small however, we see similar trends for larger as well. The distance vector is d = [0.3, 0.5]. We observe that GA outperforms all other schemes except O since it optimally allocates power for a given subchannel allocation and subchannels are allocated by an iterative algorithm that matches the rate at each step. Moreover, most of the gain of the GA is achieved also by GA-EP without power optimization. On the other hand, time allocation plays an important role as we can see from the performance gap between GA and GAH. Fig. 2 suggests that the greedy approach performs very close to the optimal allocation, while requiring a much lower complexity. As SNR increases, schemes not using direct link, i.e. NLD and NDL-S, perform worse than the other schemes because of loss in spectral efciency. It is not surprising that for

very large SNR DT approaches to GA as the direct link is chosen more often due to its higher spectral efciency. The performance boost provided by cooperative combining can be observed by comparing GA and NCC. Although GA outperforms NCC and NDL, the latter schemes entail a simpler setup. Furthermore NDL outperforms NDL-S since the latter assigns subchannels as pairs by ignoring rate mismatch. Recall that in computing and we also use the greedy power and subchannel allocation algorithm of Table I, taking into account proper modications. The effect of the number of subchannels on the end-to-end rate as well as the innite subchannel case are illustrated in Fig. 3 where = 2, = = 10 dB and d = [0.3, 0.5]. The end-to-end rates and saturate quickly, after = 10. We have also evaluated the rate obtained for an innite number of subchannels, using the techniques described in Section IV. The performance of NCC and NDL is very close to NCC-I and NDL-I, respectively. The small gaps between NCC and NCC-I and NDL and NDL-I are due to the suboptimality of the proposed greedy algorithm. The end-to-end rate of various schemes as a function of number of relays is illustrated in Fig. 4 for S = R = 25 dB and = 16. All relays are located at = 0.5. For all schemes the end-to-end rate saturates for a small number of relays and the improvement over direct transmission is about 50% with only two relays for GA. The improvement is even more signicant for small SNR where direct transmission is poor. Similar observations can also be made for other operating SNRs suggesting that a small number of relays already provides most of the gain. As mentioned in Section III-B time allocation among phase 1 and phase 2 is done numerically. The effect of time allocation on GA can be seen in Fig. 5 for = 2 relays and = 16 subchannels. The source and the relay power budget are assumed to be S = R = 25 dB, respectively and the distance vector is d = [0.3, 0.5]. Here we consider the instantaneous end-to end rate achieved by the proposed greedy algorithm on a single set of random channel realization, which is denoted by onech . We observe that the rate is

800

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 3, MARCH 2011

12

10

Expected rate (bps per Hz)

Expected rate (bps per Hz)

R R

GA

, Greedy algo. , Greedy algo., equal power

GAEP

RGAH, Greedy algo., t=0.5 4 R


NCC

3 R 2 R R 1
GA

, No cooperative combining

, Greedy algo. , Greedy algo., equal power

RNDL, No direct link RNDLS, No direct link, suboptimal 2 R


DT

GAEP

RNCC, No cooperative combining


NDL

, Direct transmission

, No direct link

RNDLS, No direct link, suboptimal RDT, Direct transmission

6 M

10

12

14

0 0.1

0.2

0.3

0.4

0.5 t

0.6

0.7

0.8

0.9

Fig. 4. Total expected end-to-end rate as a function of number of relays where = = 25 dB, = 16 and = 0.5 for all relays.

Fig. 6. Total expected end-to-end rate when phase 1 duration is xed for all channel realizations, = 2, = 16, = = 10 dB and d = [0.3, 0.5].

VI. C ONCLUSION In this paper, we analyzed resource allocation in a wireless network with multiple relays and parallel channels. We formulated an optimization problem to maximize the end-to-end rate. For the nite subchannel case, we solved the problem in two steps. First, we established the optimal power allocation given subchannel allocation and time allocation. Then we proposed a greedy algorithm which allocates subchannels and power jointly and solved the time allocation numerically. Furthermore, we studied the limiting case with innite number of subchannels whose gains are chosen independently from a given distribution where we can obtain a solution based only on the channel statistics. Numerical results revealed that the rates for innite number of subchannels is an upper bound for the nite subchannel case and the proposed greedy algorithm results in rates close to those for innite number of subchannels when the number of subchannels is sufciently large. Furthermore, a small number of relays is shown to be enough to obtain signicant cooperation gains. A PPENDIX A P ROOF OF P ROPOSITION 3.1 First, we consider the optimum relay power allocation for a given set {, }. In (7a), each term in the summation is subject to a separate relay power constraint (7c), resulting in separate optimization problems max = (1 )
=1

5 Rate (bps per Hz)

3 Ronech, Greedy algo., one ch. reali. 2


GA

0 0.1

0.2

0.3

0.4

0.5 t

0.6

0.7

0.8

0.9

Fig. 5. Total instantaneous end-to-end rate evaluated by the greedy algorithm as a function of phase 1 duration where the corresponding instantaneous channel gains are randomly chosen and xed, = 2, = 16, = = 10 dB and d = [0.3, 0.5].

neither convex not concave and it is not continuous, since it is obtained by subchannel allocation, which is a discrete optimization. Therefore a numerical search is required to nd the maximum. However, some general considerations can be made on the average rate achieved by the various schemes, when is xed to a constant level for all channel realizations. This is shown in Fig. 6. When is small the rate is limited by phase 1 and indeed, by increasing , the rate in phase 1 (and the overall rate) increases. However, even when 0 the overall rate does not go to zero for schemes GA, GA-EP and NCC since we have the option of selecting direct transmission. At around = 0.6 the rate of the GA achieves its maximum. Note that this xed time allocation maximizes the average end-to-end rate, not the instantaneous rate. Lastly, note that for 1 the rate decreases again, as in this case the rate that can be achieved in phase 2 is limited.

, log2 (1 + , , )

(29)

subject to (7c) and (7d). Problem (29) is equivalent to the standard problem of point-to-point communication over parallel channels whose optimum solution is found by allocating power R according to the waterlling procedure [33] to the subchannels for which , = 1, leading to (8). The resulting optimal phase 2 rate of R can be written as in (3). The source power allocation problem strictly depends on rates achieved by relays in phase 2 since the source has to

BAKANOGLU et al.: RESOURCE ALLOCATION FOR THE PARALLEL RELAY CHANNEL WITH MULTIPLE RELAYS

801

allocate power according to phase 2 optimum rates of each relay. The optimization problem for source power allocation can be written as
max + + =1

where the expectation is taken with respect to the channel state. Solving (37) by forming the Lagrangian with multiplier , we obtain Proposition 4.1. A PPENDIX C P ROOF OF P ROPOSITION 4.2 The problem is decomposed into two parts: Part 1) Assuming R allocates (H) to channel power gains we determine which relay to assign H to, so that total H, weighted rate is maximized, i.e., max , =1 ( ) (38) = (H) subject to and constraints (14g) with = 0. Since can only be 0 or 1, it is easy to see that the solution to (38) is to assign the subchannel state to the relay maximizing ( (H)), i.e. = ( (H)), = 1 and = 0 = , (39) where = arg max ( (H)). Part 2) After we have found which relay transmits for a given H, we allocate the total relays power among different H so that total weighted rate averaged over all channel states as expressed in (21) is maximized. That is, we solve
(H)

(30a)

subject to and + , (30b) and under constraints (7b) and (7d). In order to solve (30), we compute the Lagrangian and using Karush-Kuhn-Tucker (KKT) conditions we have

, 1+,

1+, + = 0, = 1 ,
, 1+,

+ = 0, , = 1

, = 1 + , + = 0, = 1.

(31) (32)

The complementary slackness conditions and power constraints lead to [


, log2 (1 + , )] = 0 =1

(33)

[ , log2 (1 + , )] = 0 =1 (34)

max (1 )E[]

(40)

(7b) and (7d). Solving above equations for the source power allocation in phase 1 and phase 2 we obtain the rst equation of (9) and (10). Considering the optimal rates , and , , we obtain (9), since = 0 from complementary slackness condition (33) for the rst case, = 0 from complementary slackness condition (34) for the second case and 0 and 0 from complementary slackness conditions (33) and (34) for the third case. A PPENDIX B P ROOF OF PROPOSITION 4.1 As in [27], the maximization problem is decomposed into two parts: Part 1) Letting (H) be the total allocated power to a channel state H, we determine how to distribute it among relays, i.e. solve max
, =1

subject to (1 )E[ (H)] = R , where the expectation is taken with respect to channel state. We solve (40) by forming the Lagrangian with multiplier and we obtain Proposition 4.2. R EFERENCES
[1] A. Sendonaris, E. Erkip, and B. Aazhang, User cooperation diversity part I: system description, IEEE Trans. Commun., vol. 48, no. 11, pp. 19271938, Nov. 2003. [2] A. Sendonaris, E. Erkip, and B. Aazhang, User cooperation diversity part II: implementation aspects and performance analysis, IEEE Trans. Commun., vol. 48, no. 11, pp. 1939-1948, Nov. 2003. [3] J. N. Laneman, D. Tse, and G. W. Wornell, Cooperative diversity in wireless networks: efcient protocols and outage behaviour, IEEE Trans. Inf. Theory, vol. 50, no. 12, pp. 3062-3080, Dec. 2004. [4] B. S. Krongold, K. Ramchandran, and D. L. Jones, Computationally efcient optimal power allocation algorithm for multicarrier communication systems, IEEE Trans. Commun., vol. 48, no. 1, pp. 23-27, Jan. 2000. [5] A. Goldsmith and P. Varaiya, Capacity of fading channels with channel side information, IEEE Trans. Inf. Theory, vol. 43, no. 6, pp. 1986 1992, Nov. 1997. [6] G. Caire, G. Taricco, and E. Biglieri, Optimum power control over fading channels, IEEE Trans. Inf. Theory, vol. 45, no. 5, pp. 1468-1489, July 1999. [7] Y.-W. Hong, W.-J. Huang, F.-H. Chiu, and C.-C. J. Kuo, Cooperative communications in resource-constrained wireless networks, IEEE Signal Process. Mag., vol. 24, no. 3, pp. 47-57, May 2007. [8] X. Deng and A. M. Haimovich, Power allocation for cooperative relaying in wireless networks, IEEE Commun. Lett., vol. 9, no. 11, pp. 994-996, Nov. 2005. [9] M. N. Khormuji and E. G. Larsson, Finite-SNR analysis and optimization of decode-and-forward relaying over slow fading channels, IEEE Trans. Veh. Technol., vol. 58, no. 8, pp. 4292-4305, Oct. 2009. [10] E. G. Larsson and Y. Cao, Collaborative transmit diversity with adaptive radio resource and power allocation, IEEE Commun. Lett., vol. 9, no. 6, pp. 511-513, June 2005. [11] D. Gunduz and E. Erkip, Opportunistic cooperation by dynamic resource allocation, IEEE Trans. Wireless Commun., vol. 6, no. 4, pp. 1446-1454, Apr. 2007.

( ) subject to

=1

= (H) (35)

and constraints (14f) with = 0. Since can only be 0 or 1, it is easy to see that the solution to (35) is the allocation of (H) to the relay maximizing ( (H)), i.e.
= ( (H)) and = 1; = 0 = , (36)

where = arg max ( (H)). Part 2) After we have found how to allocate the source power (H) among relays for given H, we determine how to allocate the total source power among different H so that total weighted rate averaged over all channel states as expressed in (20) is maximized. That is, we solve
(H)

max E[] such that E[ (H)] S ,

(37)

802

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 3, MARCH 2011

[12] A. Host-Madsen and J. Zhang, Capacity bounds and power allocation for wireless relay channel, IEEE Trans. Inf. Theory, vol. 51, no. 6, pp. 2020-2040, June 2005. [13] L. Sankar, Y. Liang, N. Mandayam, and H. V. Poor, Opportunistic communications in fading multiaccess relay channels, under review, arXiv:0902.1220. [14] K. Bakanoglu, D. Gunduz, and E. Erkip, Dynamic resource allocation for the broadband relay channel, in Proc. Asilomar Conf. on Signals, Systems and Computers, Monterey, CA, Nov. 2007. [15] S. Serbetli and A. Yener, Relay assisted F/TDMA ad hoc networks: node classication, power allocation and relaying strategies, IEEE Trans. Commun., vol. 56, no. 6, pp. 937947, June 2008. [16] C. T. K. Ng and A. J. Goldsmith, Capacity and power allocation for transmitter and receiver cooperation in fading channels, in Proc. IEEE Int. Conf. on Commun. (ICC), Istanbul, Turkey, June 2006, pp. 37413746. [17] C. T. K. Ng and A. J. Goldsmith, The impact of CSI and power allocation on relay channel capacity and cooperation strategies, IEEE Trans. Wireless Commun., vol. 7, no. 12, pp. 5380-5389, Dec. 2008. [18] H. Ochiai, P. Mitran, and V. Tarokh, Variable rate two phase collaborative communication protocols for wireless networks, IEEE Trans. Inf. Theory, vol. 52, no. 9, pp. 4299-4313, Sep. 2006. [19] I. Hammerstrom, M. Kuhn, and A. Wittneben, Impact of relay gain allocation on the performance of cooperative diversity networks, in Proc. IEEE Vehic. Tech. Conf. (VTC), Los Angeles, CA, Sep. 2004, pp. 18151819. [20] Y. Zhao, R. Adve, and T. J. Lim, Improving amplify-and-forward relay networks: optimal power allocation versus selection, IEEE Trans. Wireless Commun., vol. 6, no. 8, pp. 3114-3123, Aug. 2007. [21] M. Chen, S. Serbetli, and A. Yener, Distributed power allocation strategies for parallel relay networks, IEEE Trans. Wireless Commun., vol. 7, no. 2, pp. 552561, Feb. 2008 [22] K. Bakanoglu, S. Tomasin, and E. Erkip, Resource allocation in wireless networks with multiple relays in Proc. Asilomar Conf. on Signals, Systems and Computers, Monterey, CA, Nov. 2008. [23] D. Chen and J. N. Laneman, Joint power and bandwidth allocation in multihop wireless networks, in Proc. IEEE Wireless Commun. and Networking Conf. (WCNC), Las Vegas, NV, pp. 990-995, Apr. 2008. [24] D. Tse and P. Viswanath, Fundamentals of Wireless Commmunications. Cambridge University Press, 2008. [25] H. X. Li, H. Yu, H. W. Luo, J. Guo, and C. Li, Dynamic subchannel and power allocation in OFDMA-based DF cooperative relay networks in Proc. IEEE Global Conf. on Commun. (GLOBECOM), Nov. 2008. [26] M. N. Khormuji and E. G. Larsson, Cooperative transmission based on decode-and-forward relaying with partial repetition coding, IEEE Trans. Wireless Commun., vol. 8, no. 4, pp. 1716-1725, Apr. 2009. [27] L. Li and A. Goldsmith, Capacity and optimal resource allocation for fading broadcast channelspart I: ergodic capacity, IEEE Trans. Inf. Theory, vol. 47, no. 3, pp. 1083-1102, Mar. 2001. [28] D. Tse and S. Hanly, Multiaccess fading channelspart I: polymatroid structure, optimal resource allocation and throughput capacities, IEEE Trans. Inf. Theory, vol. 44, no. 7, pp. 2796-2815, Nov. 1998. [29] R. Knopp and P. A. Humblet, Information capacity and power control in single-cell multiuser communications, in Proc. Int. Conf. on Commun. (ICC), Seattle, WA, June 1995. [30] A. Stefanov and E. Erkip, Cooperative coding for wireless networks, IEEE Trans. Commun., vol. 52, no. 9, pp. 14701476, Sep. 2004. [31] G. Kramer, M. Gastpar, and P. Gupta, Cooperative strategies and capacity theorems for relay networks, IEEE Trans. Inf. Theory, vol. 51, no. 7, pp. 3037-3063, Sep. 2005. [32] W. Yu and J. M. Ciof, FDMA capacity of Gaussian multiple-access channels with ISI, IEEE Trans. Commun., vol. 50, no. 1, pp. 102-111, Jan. 2002.

[33] T. M. Cover and J. A. Thomas, Elements of Information Theory. Wiley, 1991. Ka an Bakano lu (S06) received the B.S. degree g g in electrical and electronics engineering from the Middle East Technical University, Ankara, Turkey, in 2006. He is currently working toward his Ph.D. at Polytechnic Institute of New York University, Brooklyn. In 2008 and 2009, he was a summer researcher in the Corporate Research Lab, Thomson Inc. in Princeton, NJ. His research interests lie in the areas of information theory, communication theory, and wireless communications. Stefano Tomasin (S99, M03) received the Laurea degree and the Ph.D. degree in Telecommunications Engineering from the University of Padova, Italy, in 1999 and 2003, respectively. In the Academic year 19992000 he was on leave at the IBM Research Laboratory, Zurich, Switzerland, doing research on signal processing for magnetic recording systems. In the Academic year 20012002 he was on leave at Philips Research, Eindhoven, the Netherlands, studying multicarrier transmission for mobile applications. In the second half of 2004 he was visiting Qualcomm, San Diego (CA) doing research on receiver design for mobile cellular systems. Since 2005 he is assistant professor at University of Padova, Italy. In 2007 he has been visiting faculty at Polytechnic Institute of New York University, Brooklyn, working on cooperative communications. His current research interests include signal processing for wireless communications, access technologies for multiuser/multiantenna systems and cross-layer protocol design and evaluation. Elza Erkip (S93-M96-SM05) received the B.S. degree in electrical and electronics engineering from the Middle East Technical University, Ankara, Turkey, and the M.S. and Ph.D. degrees in electrical engineering from Stanford University, Stanford, CA. Currently, she is an associate professor of electrical and computer engineering at Polytechnic Institute of New York University, Brooklyn. In the past, she has held positions at Rice University, Houston, TX, and at Princeton University, Princeton, NJ. Her research interests are in information theory, communication theory, and wireless communications. Dr. Erkip received the National Science Foundation CAREER Award in 2001, the IEEE Communications Society Rice Paper Prize in 2004, and the ICC Communication Theory Symposium Best Paper Award in 2007. She coauthored a paper that received the ISIT Student Paper Award in 2007. She was a Finalist for The New York Academy of Sciences Blavatnik Awards for Young Scientists in 2010. Currently, she is an associate editor of the IEEE T RANSACTIONS ON I NFORMATION T HEORY and a Technical Co-Chair of WiOpt 2011. She was an associate editor of the IEEE T RANSACTIONS ON C OMMUNICATIONS during 2006-2009, a publications editor of the IEEE T RANSACTIONS ON I NFORMATION T HEORY during 2006-2009 and a guest editor of IEEE Signal Processing Magazine in 2007. She was the cochair of the GLOBECOM Communication Theory Symposium in 2009, the publications chair of ITW Taormina in 2009, the MIMO Communications and Signal Processing Technical Area chair of the Asilomar Conference on Signals, Systems, and Computers in 2007, and the technical program co-chair of the Communication Theory Workshop in 2006.

Das könnte Ihnen auch gefallen