Cooperative Network Coding Strategies for ... - Semantic Scholar

3 downloads 827 Views 892KB Size Report
material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work ... ducing dedicated relay nodes to help the transmission, one.
Cooperative Network Coding Strategies for Wireless Relay Networks with Backhaul IEEE Transactions on Communications, vol. 59, pp. 2502โ€“2514, Sep. 2011. c 2011 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this

material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

JINFENG DU, MING XIAO, MIKAEL SKOGLUND

Stockholm September 2011 School of Electrical Engineering and the ACCESS Linnaeus Center, Royal Institute of Technology (KTH), SE-100 44 Stockholm, Sweden IR-EE-KT 2011:024

2502

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 9, SEPTEMBER 2011

Cooperative Network Coding Strategies for Wireless Relay Networks with Backhaul Jinfeng Du, Student Member, IEEE, Ming Xiao, Member, IEEE, and Mikael Skoglund, Senior Member, IEEE

Index Termsโ€”Cooperative communication, network coding, relay, source cooperation, cut-set bound.

I. I NTRODUCTION

C

APACITY bounds and various cooperative strategies for three-node relaying networks (source-relay-sink, or two cooperative sources and one sink) have been studied in [1], [2]. The relay (or the other source) uses decode-and-forward (DF) or compress-and-forward (CF) to aid the transmission. Coding schemes have been investigated for multiple-access relay channels (MARC) [3], [4] involving multiple sources and a single destination, and for broadcast relay channels (BRC) [3], [5] where a single source transmits messages to multiple destinations. Recent results on capacity bounds for multiple-source multiple-destination relay networks, [6]โ€“[9] and references therein, have provided valuable insight into the benefits of relaying. Motivated by the MAC channel at the relay node where different messages mix up by nature, various network coding (NC) [10]โ€“[12] approaches, which essentially combine multiple messages together, can be introduced to boost the sum rate. For instance, in a relay-aided two-source two-sink multicast network, achievable rates for a full-duplex amplify-and-forward (AF) relay with linear NC (LNC) have been studied in [13], and in [8] the relay uses lattice codes for network coding. In [14] joint NC and physical layer coding is performed via lattice coding for the bi-directional Paper approved by E. Serpedin, the Editor for Synchronization and Sensor Networks of the IEEE Communications Society. Manuscript received August 28, 2010; revised February 18, 2011. This work was presented in part at IEEE ITW, Aug. 2010. This work was supported in part by the Swedish Governmental Agency for Innovation Systems (VINNOVA) and the Swedish Foundation for Strategic Research (SSF). The authors are with the School of Electrical Engineering and the ACCESS Linnaeus Center, Royal Institute of Technology, Stockholm, Sweden (e-mail: [email protected]; {ming.xiao, mikael.skoglund}@ee.kth.se). Digital Object Identifier 10.1109/TCOMM.2011.061311.100525

๐‘Š1

๐‘Š2

๐‘Œ1

๐’ฎ1 ๐‘‹1

Backhaul

Abstractโ€”We investigate cooperative network coding strategies for relay-aided two-source two-destination wireless networks with a backhaul connection between the source nodes. Each source multicasts information to all destinations using a shared relay. We study cooperative strategies based on different network coding schemes, namely, finite field and linear network coding, and lattice coding. To further exploit the backhaul connection, we also propose network coding based beamforming. We measure the performance in term of achievable rates over Gaussian channels, and observe significant gains over benchmark schemes. We derive the achievable rate regions for these schemes and find the cut-set bound for our system. We also show that the cut-set bound can be achieved by network coding based beamforming when the signal-to-noise ratios lie in the sphere defined by the source-relay and relay-destination channel gains.

๐’ฎ2

๐‘‹2

๐‘Œ๐‘Ÿ

โ„›

๐‘‹๐‘Ÿ

๐‘Œ2

๐’Ÿ1

๐’Ÿ2

Fig. 1. Two source nodes ๐’ฎ1 and ๐’ฎ2 , connected with backhaul, multicast information ๐‘Š1 and ๐‘Š2 respectively to both destinations ๐’Ÿ1 and ๐’Ÿ2 , with aid from a full-duplex relay node โ„›.

relay channel. The recently proposed noisy network coding scheme (Noisy NC) [15] for transmitting multiple sources over a general noisy network, has been shown to outperform the conventional CF scheme in the Gaussian two-way relay channel and the interference relay channel. Apart from introducing dedicated relay nodes to help the transmission, one can also utilize cooperative strategies among sources [16]โ€“ [21] and/or among destinations [20]โ€“[22] with the help of orthogonal conferencing channels. In this paper, we aim at evaluating achievable rate regions for various cooperative strategies when source cooperation and network coding are designed jointly with the relaying. More specifically, we focus on a relay-aided two-source twodestination multicast network with backhaul support, as shown in Fig. 1. Sources ๐’ฎ1 and ๐’ฎ2 multicast their own information (๐‘Š1 and ๐‘Š2 respectively) to geographically separated destinations ๐’Ÿ1 and ๐’Ÿ2 , with the help of a relay โ„›. This model arises, for example, in a wireless cellular downlink where two base stations multicast to two mobile terminals, one in each cell, with the help of a dedicated relay deployed at the common cell boundary. Since the base stations are connected through the (fiber or microwave) backhaul, more general network coding schemes can be used at the relay to cooperate with the sourcesโ€™ transmission. This model is interesting since it is a combination of relaying, MARC, BRC, source cooperation, and network coding. It can be extended to more general networks by tuning the channel gains within the range [0, โˆž). In this paper, we are interested in the scenario without cross channels between ๐’ฎ1 and ๐’Ÿ2 , or ๐’ฎ2 and ๐’Ÿ1 . While, in general, the signal from ๐’ฎ๐‘– would be heard also at ๐’Ÿ๐‘— , ๐‘— โˆ•= ๐‘–, our assumption can be motivated for example in scenarios where the cross links are too weak to be of any use, or are technically suppressed. In any case we consider any contribution directly from ๐’ฎ๐‘– at ๐’Ÿ๐‘— (๐‘— โˆ•= ๐‘–) not to be useful and therefore part of the noise. We also restrict our analysis to fixed channel gains, and we assume a full-duplex DF relay. Furthermore, any extensions of the cooperative NC strategies

c 2011 IEEE 0090-6778/11$25.00 โƒ

DU et al.: COOPERATIVE NETWORK CODING STRATEGIES FOR WIRELESS RELAY NETWORKS WITH BACKHAUL

developed in this paper to multiple sources and/or multiple relays are left to future work. The paper is organized as follows. The system model is introduced in Section II. For symmetric channel gains and highrate backhaul, various cooperative NC strategies are investigated in Section III, and a benchmark scheme together with the cut-set bound are presented in Section IV. Cooperative NC strategies for non-symmetric channel gains and for lowrate backhaul (i.e., partial transmitter cooperation) scenarios are discussed in Section V. Numerical results are presented in Section VI and concluding remarks in Section VII. Notation: Capital letter ๐‘‹ indicates a real valued random variable and ๐‘(๐‘‹) indicates its probability density/mass function. ๐‘‹ (๐‘›) denotes a vector of random variables of length ๐‘›, and with the ๐‘˜th component ๐‘‹[๐‘˜] (in general without emphasizing the (โ‹…)(๐‘›) ). ๐ผ(๐‘‹; ๐‘Œ ) denotes the mutual information between ๐‘‹ and ๐‘Œ , and ๐ถ(๐‘ฅ) = 12 log2 (1 + ๐‘ฅ) is the Gaussian capacity function. II. S YSTEM M ODEL To simplify our analysis, we first consider the symmetric channel gain scenario illustrated in Fig. 1 (๐‘›)

(๐‘›)

= ๐‘‹1

๐‘Œ1

(๐‘›) ๐‘Œ2 ๐‘Œ๐‘Ÿ(๐‘›)

= =

(๐‘›)

+ ๐‘๐‘‹๐‘Ÿ(๐‘›) + ๐‘1 ,

(๐‘›) (๐‘›) ๐‘‹2 + ๐‘๐‘‹๐‘Ÿ(๐‘›) + ๐‘2 , (๐‘›) (๐‘›) ๐‘Ž๐‘‹1 + ๐‘Ž๐‘‹2 + ๐‘๐‘Ÿ(๐‘›) ,

(1a) (1b) (1c)

where ๐‘Ž โ‰ฅ 0 is the normalized channel gain for the sourcerelay links and ๐‘ โ‰ฅ 0 for the relay-destination links. For (๐‘›) (๐‘›) (๐‘›) ๐‘– = 1, 2, ๐‘Ÿ, ๐‘‹๐‘– , ๐‘Œ๐‘– and ๐‘๐‘– are ๐‘›-dimensional transmitted signals, received signals, and noise, respectively, where ๐‘๐‘– [๐‘˜], ๐‘˜ = 1, ..., ๐‘› are i.i.d. Gaussian with zero-mean and unitvariance. The transmitted signals are subject to individual average power constraints, i.e., ๐‘›

1โˆ‘ 2 ๐‘‹๐‘– [๐‘˜] โ‰ค ๐‘ƒ๐‘– , ๐‘– = 1, 2, ๐‘Ÿ. ๐‘›

(2)

๐‘˜=1

Note that (1) implies simultaneously perfect synchronization at ๐’Ÿ1 , ๐’Ÿ2 , and โ„›, respectively. This assumption, although widely adopted in information-theoretic work, is optimistic in practice. In general, the results we obtain based on perfect synchronization will serve as upper bounds on any practical performance, and can be directly extended in the same way as in [2] to scenarios where constructive (co-phase) addition is not available. In practice the backhaul normally has much higher capacity and lower error rates than the forward wireless channels. Therefore, in our model the backhaul is assumed to be errorfree and of sufficiently high capacity (higher than the forward sum-rate). The case of a backhaul capacity smaller than the sum-rate will be discussed in Section V. With a high rate backhaul, our system is closely related to the MIMO relay channel scenario, as studied in [23], [24]. However the problems are not equivalent, and we emphasize the following three main differences between the system investigated in this paper and the MIMO relay scenario with a two-antenna source node. First, in our system each source/antenna is subject to an individual power constraint (2), while in the MIMO relay

2503

channel model a sum-power constraint is usually applied at the source node, which in general implies a larger achievable rate region. Second, in our system the relay combines the messages from the sources by performing NC rather than forwarding them separately through orthogonal channels. Last but not the least, the cooperative strategies proposed for high rate backhaul in Section III can be directly extended to the finite-rate backhaul scenario with the help of superposition coding or time-sharing strategies, as stated in Section V. III. C OOPERATIVE N ETWORK C ODING S TRATEGIES Similar to [1]โ€“[3], [6], source ๐’ฎ๐‘– , ๐‘– = 1, 2, divides its messages ๐‘Š๐‘– into ๐ต blocks ๐‘Š๐‘–,1 , . . . , ๐‘Š๐‘–,๐ต with ๐‘›๐‘…๐‘– bits each. The transmission is completed over ๐ต + 1 blocks. At the first block the two sources exchange ๐‘Š๐‘–,1 over the backhaul and also broadcast their own messages over the relay channels; in block ๐‘ก, source ๐’ฎ๐‘– exchanges ๐‘Š๐‘–,๐‘ก through the backhaul (๐‘›) and broadcasts its codeword ๐‘‹๐‘–,๐‘ก , which is a function of (๐‘Š๐‘–,๐‘ก , ๐‘Š1,๐‘กโˆ’1 , ๐‘Š2,๐‘กโˆ’1 ), over the channels; in block ๐ต + 1 only ๐‘Š๐‘–,๐ต is broadcasted. As each transmission is over ๐‘› channel uses, and assuming the backhaul is used for free, the ๐ต๐‘›๐‘…๐‘– bits per channel use, which converges overall rate is (๐ต+1)๐‘› to ๐‘…๐‘– when ๐ต goes to infinity. Three decoding protocols, namely successive decoding [1], backward decoding [25], and sliding-window decoding [26], have been summarized and extended to multiple-source or multiple-relay scenarios in [3]. We implement these protocols at relay/destination nodes depending on the cooperative NC strategy under consideration. Unless stated otherwise, random coding is used for encoding and joint-typicality is used for decoding. Each codeword is generated randomly in the memoryless fashion [27]: For transmitting messages in {๐‘Š } each of ๐‘›๐‘… bits, we create a codebook consisting of 2๐‘›๐‘… randomly and independently generated sequences {๐‘ˆ (๐‘›) }, each of ๐‘›-bit length, according to the distribution ฮ ๐‘›๐‘–=1 ๐‘(๐‘ข๐‘– ). We assign a codeword ๐‘ˆ (๐‘›) to a message ๐‘Š and associate them via an encoding function ๐‘ˆ (๐‘›) (๐‘Š ), omitting the explicit relation where appropriate. A. Finite-field Network Coding With DF (DF+FNC) At the end of block ๐‘ก โˆ’ 1, the relay decodes (๐‘›) (๐‘Š1,๐‘กโˆ’1 , ๐‘Š2,๐‘กโˆ’1 ) jointly from its received signal ๐‘Œ๐‘Ÿ,๐‘กโˆ’1 and then creates a new message ๐‘Š๐‘Ÿ,๐‘ก = ๐‘Š1,๐‘กโˆ’1 โŠ• ๐‘Š2,๐‘กโˆ’1 (bitwise GF(2) addition). If the lengths of ๐‘Š1,๐‘กโˆ’1 and ๐‘Š2,๐‘กโˆ’1 are not equal, i.e., ๐‘…1 โˆ•= ๐‘…2 , we can append zeros at the end of the shorter message. During block ๐‘ก, โ„› transmits ๐‘Š๐‘Ÿ,๐‘ก using an independent random codebook {๐‘ˆ (๐‘›) } of size 2๐‘›๐‘… (where ๐‘… = max(๐‘…1 , ๐‘…2 )), โˆš (๐‘›) ๐‘‹๐‘Ÿ,๐‘ก = ๐‘ƒ๐‘Ÿ ๐‘ˆ (๐‘›) (๐‘Š๐‘Ÿ,๐‘ก ). (3) ๐’ฎ1 and ๐’ฎ2 , on the other hand, transmit their information (๐‘›) via independent random codebooks {๐‘‰1 } of size 2๐‘›๐‘…1 and (๐‘›) ๐‘›๐‘…2 {๐‘‰2 } of size 2 , respectively. Since ๐‘Š1,๐‘กโˆ’1 and ๐‘Š2,๐‘กโˆ’1 are exchanged via the backhaul in block ๐‘ก โˆ’ 1, ๐’ฎ1 and ๐’ฎ2 also know ๐‘Š๐‘Ÿ,๐‘ก if decoding at โ„› is reliable. Therefore to exploit the possibility of coherent combining gain, ๐’ฎ1 and ๐’ฎ2 can coordinate their transmission with โ„› as follows, โˆš โˆš (๐‘›) (๐‘›) ๐‘‹1,๐‘ก = ๐›ผ1 ๐‘ƒ1 ๐‘‰1 (๐‘Š1,๐‘ก ) + (1 โˆ’ ๐›ผ1 )๐‘ƒ1 ๐‘ˆ (๐‘›) (๐‘Š๐‘Ÿ,๐‘ก ), (4a) โˆš โˆš (๐‘›) (๐‘›) ๐‘‹2,๐‘ก = ๐›ผ2 ๐‘ƒ2 ๐‘‰2 (๐‘Š2,๐‘ก ) + (1 โˆ’ ๐›ผ2 )๐‘ƒ2 ๐‘ˆ (๐‘›) (๐‘Š๐‘Ÿ,๐‘ก ), (4b)

2504

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 9, SEPTEMBER 2011

TABLE I I LLUSTRATION OF THE E NCODING AND D ECODING P ROCESS FOR DF+FNC, W ITH ๐‘Š๐‘Ÿ,๐‘ก = ๐‘Š1,๐‘กโˆ’1 โŠ• ๐‘Š2,๐‘กโˆ’1 , ๐‘Š๐‘Ÿ,1 = 1, AND ๐ต = 3 ๐‘ก= โ‡Œ ๐’ฎ1 transmits ๐’ฎ2 transmits โ„› transmits โ„› decodes ๐’Ÿ1 decodes recovers by โŠ•

1 โˆฃ 2 โˆฃ 3 โˆฃ 4 ๐‘Š1,1 โ‡”๐‘Š2,1 โˆฃ๐‘Š1,2 โ‡”๐‘Š2,2 โˆฃ๐‘Š1,3 โ‡”๐‘Š2,3 โˆฃ / (๐‘Š1,1 , 1) โˆฃ(๐‘Š1,2 , ๐‘Š๐‘Ÿ,2 )โˆฃ(๐‘Š1,3 , ๐‘Š๐‘Ÿ,3 )โˆฃ(1, ๐‘Š๐‘Ÿ,4 ) (๐‘Š2,1 , 1) โˆฃ(๐‘Š2,2 , ๐‘Š๐‘Ÿ,2 )โˆฃ(๐‘Š2,3 , ๐‘Š๐‘Ÿ,3 )โˆฃ(1, ๐‘Š๐‘Ÿ,4 ) 1 โˆฃ ๐‘Š๐‘Ÿ,2 โˆฃ ๐‘Š๐‘Ÿ,3 โˆฃ ๐‘Š๐‘Ÿ,4 ๐‘Š1,1 , ๐‘Š2,1 โˆฃ ๐‘Š1,2 , ๐‘Š2,2 โˆฃ ๐‘Š1,3 , ๐‘Š2,3 โˆฃ / ๐‘Š1,1 โˆฃ ๐‘Š1,2 , ๐‘Š๐‘Ÿ,2 โˆฃ ๐‘Š1,3 , ๐‘Š๐‘Ÿ,3 โˆฃ ๐‘Š๐‘Ÿ,4 / โˆฃ ๐‘Š2,1 โˆฃ ๐‘Š2,2 โˆฃ ๐‘Š2,3

Without the backhaul, ๐’ฎ1 and ๐’ฎ2 cannot know/estimate ๐‘Š๐‘Ÿ and therefore cannot cooperate with โ„›, i.e. ๐›ผ1 = ๐›ผ2 = 1. Hence, no coherent combining gain can be achieved.

B. Linear Network Coding With DF (DF+LNC)

When LNC is used in the signal domain, โ„› essentially performs superposition coding. The scheme presented here is a natural extension of the one in Theorem 1 of [6] which is designed for transmitting both private and common messages via the interference relay channel (IFRC). In our case, only where 0 โ‰ค ๐›ผ1 , ๐›ผ2 โ‰ค 1 are power allocation parameters. The common messages are transmitted (i.e., multicast). Unlike received signals are therefore in [6] where each source can only cooperate with node โ„› (๐‘›) โˆš โˆš (๐‘›) โˆš (๐‘›) (๐‘›) (๐‘›) ๐‘Œ1,๐‘ก = ๐›ผ1 ๐‘ƒ1 ๐‘‰1 +( (1 โˆ’ ๐›ผ1 )๐‘ƒ1 +๐‘ ๐‘ƒ๐‘Ÿ )๐‘ˆ + ๐‘1,๐‘ก , (5a) regarding its own message in ๐‘‹๐‘Ÿ , the two source nodes can in our case cooperate to transmit both messages, thanks โˆš โˆš (๐‘›) โˆš (๐‘›) (๐‘›) ๐‘Œ2,๐‘ก = ๐›ผ2 ๐‘ƒ2 ๐‘‰2 +( (1 โˆ’ ๐›ผ2 )๐‘ƒ2 +๐‘ ๐‘ƒ๐‘Ÿ )๐‘ˆ (๐‘›) + ๐‘2,๐‘ก , (5b) to the backhaul. We first generate two independent random โˆš โˆš โˆš (๐‘›) (๐‘›) (๐‘›) (๐‘›) codebooks {๐‘ˆ1 } of size 2๐‘›๐‘…1 and {๐‘ˆ2 } of size 2๐‘›๐‘…2 . At ๐‘Œ๐‘Ÿ,๐‘ก =๐‘Ž( (1โˆ’๐›ผ1 )๐‘ƒ1 + (1โˆ’๐›ผ2 )๐‘ƒ2 )๐‘ˆ (๐‘›) +๐‘Ž ๐›ผ1 ๐‘ƒ1 ๐‘‰1 โˆš the end of block ๐‘ก โˆ’ 1, โ„› decodes (๐‘Š1,๐‘กโˆ’1 , ๐‘Š2,๐‘กโˆ’1 ) and then (๐‘›) (๐‘›) +๐‘Ž ๐›ผ2 ๐‘ƒ2 ๐‘‰2 + ๐‘๐‘Ÿ,๐‘ก . (5c) (๐‘›) (๐‘›) picks up codewords ๐‘ˆ1 (๐‘Š1,๐‘กโˆ’1 ) and ๐‘ˆ2 (๐‘Š2,๐‘กโˆ’1 ) from the Successive decoding is implemented at both the relay two codebooks respectively, and transmits the superposition of and the two destination nodes: assuming ๐‘Š1,๐‘กโˆ’1 has been these in block ๐‘ก with power allocation parameter 0 โ‰ค ๐›ผ๐‘Ÿ โ‰ค 1 โˆš โˆš successfully decoded by ๐’Ÿ1 , at the end of block ๐‘ก, ๐’Ÿ1 (๐‘›) (๐‘›) (๐‘›) (๐‘›) recovers (๐‘Š1,๐‘ก , ๐‘Š๐‘Ÿ,๐‘ก ) jointly from ๐‘Œ1,๐‘ก , and then retrieves ๐‘‹๐‘Ÿ,๐‘ก = ๐›ผ๐‘Ÿ ๐‘ƒ๐‘Ÿ ๐‘ˆ1 (๐‘Š1,๐‘กโˆ’1 ) + (1โˆ’๐›ผ๐‘Ÿ )๐‘ƒ๐‘Ÿ ๐‘ˆ2 (๐‘Š2,๐‘กโˆ’1 ). ๐‘Š2,๐‘กโˆ’1 = ๐‘Š๐‘Ÿ,๐‘ก โŠ• ๐‘Š1,๐‘กโˆ’1 . This approach is also used for (๐‘›) (๐‘›) For each codeword ๐‘ˆ1 (๐‘Š1,๐‘กโˆ’1 ), we generate an indepen๐’Ÿ2 . The relay โ„› decodes jointly (๐‘Š1,๐‘ก , ๐‘Š2,๐‘ก ) from ๐‘Œ๐‘Ÿ,๐‘ก by (๐‘›) first cancelling out ๐‘ˆ (๐‘›) . The encoding/decoding process is dent codebook {๐‘‰1 } of size 2๐‘›๐‘…1 , and then use this codebook to encode the new message ๐‘Š1,๐‘ก . We denote the selected illustrated in Table I. (๐‘›) codeword for ๐‘Š1,๐‘ก given ๐‘Š1,๐‘กโˆ’1 as ๐‘‰1 (๐‘Š1,๐‘ก , ๐‘Š1,๐‘กโˆ’1 ). Proposition 1: The achievable rate region for DF+FNC is (๐‘›) the union over all (๐‘…1 , ๐‘…2 ) satisfying Similarly we choose ๐‘‰2 (๐‘Š2,๐‘ก , ๐‘Š2,๐‘กโˆ’1 ) for ๐‘Š2,๐‘ก . With power allocation parameters 0 โ‰ค ๐›ผโ€ฒ๐‘– , ๐›ผโ€ฒโ€ฒ๐‘– โ‰ค 1, ๐‘– = 1, 2 to cooperate { } โˆš โˆš with โ„›, the transmitted signal at ๐’ฎ1 and ๐’ฎ2 are therefore 2 2 ๐‘…1 < min ๐ถ(๐‘Ž ๐›ผ1 ๐‘ƒ1 ), ๐ถ(๐›ผ1 ๐‘ƒ1 ), ๐ถ(( (1โˆ’๐›ผ2 )๐‘ƒ2 +๐‘ ๐‘ƒ๐‘Ÿ ) ) , { } โˆš โˆš ๐‘…2 < min ๐ถ(๐‘Ž2 ๐›ผ2 ๐‘ƒ2 ), ๐ถ(๐›ผ2 ๐‘ƒ2 ), ๐ถ(( (1โˆ’๐›ผ1 )๐‘ƒ1 +๐‘ ๐‘ƒ๐‘Ÿ )2 ) , ) { ( โˆš ๐‘…1 + ๐‘…2 < min ๐ถ ๐‘ƒ1 +๐‘2 ๐‘ƒ๐‘Ÿ +2๐‘ (1โˆ’๐›ผ1 )๐‘ƒ1 ๐‘ƒ๐‘Ÿ , (6) )} ( โˆš 2 , ๐ถ(๐‘Ž ๐›ผ1 ๐‘ƒ1 +๐‘Ž2 ๐›ผ2 ๐‘ƒ2 ), ๐ถ ๐‘ƒ2 +๐‘2 ๐‘ƒ๐‘Ÿ +2๐‘ (1โˆ’๐›ผ2 )๐‘ƒ2 ๐‘ƒ๐‘Ÿ

where the union is taken over 0 โ‰ค ๐›ผ1 , ๐›ผ2 โ‰ค 1. Proof: The proof can be found in Appendix A. The constraint on ๐‘…1 corresponds to the condition that ๐‘Š1 can be decoded reliably at โ„› and ๐’Ÿ1 , and that the NC message ๐‘Š๐‘Ÿ can be decoded at ๐’Ÿ2 , and similarly for ๐‘…2 and ๐‘…1 + ๐‘…2 . Note that our scheme is similar to the strategy in [8]: ๐’Ÿ1 recovers ๐‘Š1 from the direct link and ๐‘Š๐‘Ÿ from the โ„›โ€“๐’Ÿ1 link, and then retrieves ๐‘Š2 based on the observation of ๐‘Š1 and ๐‘Š๐‘Ÿ . But there are two main differences: finite-field NC rather than lattice coding is used; both source nodes know ๐‘Š๐‘Ÿ thanks to the backhaul and therefore they cooperate with โ„› to get a coherent combining gain. Corollary 1: For the symmetric scenario with ๐‘ƒ1 =๐‘ƒ2 = ๐‘ƒ๐‘Ÿ = ๐‘ƒ and ๐‘…1 =๐‘…2 =๐‘…, rate ๐‘… is achievable by DF+FNC if โˆš { } ๐ถ(2๐‘Ž2 ๐‘ƒ ๐›ผ) ๐ถ((1+๐‘2 +2๐‘ 1โˆ’๐›ผ)๐‘ƒ ) , ๐‘…< max min ๐ถ(๐›ผ๐‘ƒ ), . 0โ‰ค๐›ผโ‰ค1 2 2 (7)

Proof: The result follows straightforwardly from (6) by setting ๐›ผ1 = ๐›ผ2 = ๐›ผ.

(๐‘›)

โˆš โˆš (๐‘›) (๐‘›) โˆš (๐‘›) ๐›ผโ€ฒ1 ๐‘ƒ1 ๐‘ˆ1 + ๐›ผโ€ฒโ€ฒ1 ๐‘ƒ1 ๐‘ˆ2 + (1โˆ’๐›ผโ€ฒ1 โˆ’๐›ผโ€ฒโ€ฒ1 )๐‘ƒ1 ๐‘‰1 , โˆš โˆš โˆš (๐‘›) (๐‘›) (๐‘›) = ๐›ผโ€ฒ2 ๐‘ƒ2 ๐‘ˆ2 + ๐›ผโ€ฒโ€ฒ2 ๐‘ƒ2 ๐‘ˆ1 + (1โˆ’๐›ผโ€ฒ2 โˆ’๐›ผโ€ฒโ€ฒ2 )๐‘ƒ2 ๐‘‰2 .

๐‘‹1,๐‘ก = (๐‘›)

๐‘‹2,๐‘ก

The received signals at the destinations and the relay are โˆš โˆš โˆš (๐‘›) (๐‘›) (1โˆ’๐›ผโ€ฒ1 โˆ’๐›ผโ€ฒโ€ฒ1 )๐‘ƒ1 ๐‘‰1 + ( ๐›ผโ€ฒ1 ๐‘ƒ1 +๐‘ ๐›ผ๐‘Ÿ ๐‘ƒ๐‘Ÿ )๐‘ˆ1 โˆš โˆš (๐‘›) (๐‘›) +( ๐›ผโ€ฒโ€ฒ1 ๐‘ƒ1 + ๐‘ (1 โˆ’ ๐›ผ๐‘Ÿ )๐‘ƒ๐‘Ÿ )๐‘ˆ2 + ๐‘1 , โˆš โˆš โˆš (๐‘›) (๐‘›) (๐‘›) ๐‘Œ2 = (1โˆ’๐›ผโ€ฒ2 โˆ’๐›ผโ€ฒโ€ฒ2 )๐‘ƒ2 ๐‘‰2 + ( ๐›ผโ€ฒโ€ฒ2 ๐‘ƒ2 +๐‘ ๐›ผ๐‘Ÿ ๐‘ƒ๐‘Ÿ )๐‘ˆ1 โˆš โˆš (๐‘›) (๐‘›) +( ๐›ผโ€ฒ2 ๐‘ƒ2 + ๐‘ (1 โˆ’ ๐›ผ๐‘Ÿ )๐‘ƒ๐‘Ÿ )๐‘ˆ2 + ๐‘2 , [โˆš (๐‘›) โˆš (๐‘›) ๐‘Œ๐‘Ÿ(๐‘›) = ๐‘Ž (1โˆ’๐›ผโ€ฒ1 โˆ’๐›ผโ€ฒโ€ฒ1 )๐‘ƒ1 ๐‘‰1 + (1โˆ’๐›ผโ€ฒ2 โˆ’๐›ผโ€ฒโ€ฒ2 )๐‘ƒ2 ๐‘‰2 โˆš โˆš โˆš โˆš (๐‘›) (๐‘›) (๐‘›) +( ๐›ผโ€ฒ1 ๐‘ƒ1 + ๐›ผโ€ฒโ€ฒ2 ๐‘ƒ2 )๐‘ˆ1 +( ๐›ผโ€ฒโ€ฒ1 ๐‘ƒ1 + ๐›ผโ€ฒ2 ๐‘ƒ2 )๐‘ˆ2 ]+๐‘๐‘Ÿ . (8) (๐‘›)

๐‘Œ1

=

The decoding follows directly from [6]: the relay performs successive decoding and the destinations use backward de(๐‘›) coding. โ„› decodes (๐‘Š1,๐‘ก , ๐‘Š2,๐‘ก ) reliably from ๐‘Œ๐‘Ÿ,๐‘ก at the end of block ๐‘ก. ๐’Ÿ1 and ๐’Ÿ2 start decoding when transmission is finished. At block ๐ต + 1, no new message is transmitted and the received signal at ๐’Ÿ1 (๐’Ÿ2 ) only depends on (๐‘Š1,๐ต , ๐‘Š2,๐ต ). After decoding (๐‘Š1,๐ต , ๐‘Š2,๐ต ) successfully, only ๐‘Š1,๐ตโˆ’1 ( (๐‘›) (๐‘›) ๐‘Š2,๐ตโˆ’1 ) is unknown in ๐‘Œ1,๐ต (๐‘Œ2,๐ต ), and we repeat this process backwards until all messages are recovered.

DU et al.: COOPERATIVE NETWORK CODING STRATEGIES FOR WIRELESS RELAY NETWORKS WITH BACKHAUL

Proposition 2: The achievable rate region for DF+LNC is given by { ๐‘…1 < min ๐ถ(๐‘Ž2 ๐‘ƒ1 (1 โˆ’ ๐›ผโ€ฒ1 โˆ’ ๐›ผโ€ฒโ€ฒ1 )), ) ( โˆš ๐ถ (1โˆ’๐›ผโ€ฒโ€ฒ1 )๐‘ƒ1 + ๐‘2 ๐›ผ๐‘Ÿ ๐‘ƒ๐‘Ÿ + 2๐‘ ๐›ผโ€ฒ1 ๐›ผ๐‘Ÿ ๐‘ƒ1 ๐‘ƒ๐‘Ÿ , ( )} โˆš ๐ถ ๐›ผโ€ฒโ€ฒ2 ๐‘ƒ2 + ๐‘2 ๐›ผ๐‘Ÿ ๐‘ƒ๐‘Ÿ + 2๐‘ ๐›ผโ€ฒโ€ฒ2 ๐›ผ๐‘Ÿ ๐‘ƒ2 ๐‘ƒ๐‘Ÿ , { 2 โ€ฒ โ€ฒโ€ฒ ๐‘…2 < min ๐ถ(๐‘Ž ๐‘ƒ2 (1 โˆ’ ๐›ผ2 โˆ’ ๐›ผ2 )), ( ) โˆš ๐ถ (1โˆ’๐›ผโ€ฒโ€ฒ2 )๐‘ƒ2 +๐‘2 (1โˆ’๐›ผ๐‘Ÿ )๐‘ƒ๐‘Ÿ +2๐‘ ๐›ผโ€ฒ2 (1โˆ’๐›ผ๐‘Ÿ )๐‘ƒ2 ๐‘ƒ๐‘Ÿ , ( )} โˆš ๐ถ ๐›ผโ€ฒโ€ฒ1 ๐‘ƒ1 + ๐‘2 (1โˆ’๐›ผ๐‘Ÿ )๐‘ƒ๐‘Ÿ + 2๐‘ ๐›ผโ€ฒโ€ฒ1 (1โˆ’๐›ผ๐‘Ÿ )๐‘ƒ1 ๐‘ƒ๐‘Ÿ , { ๐‘…1 +๐‘…2 < min ๐ถ(๐‘Ž2 (1โˆ’๐›ผโ€ฒ1 โˆ’๐›ผโ€ฒโ€ฒ1 )๐‘ƒ1 +๐‘Ž2 (1โˆ’๐›ผโ€ฒ2 โˆ’๐›ผโ€ฒโ€ฒ2 )๐‘ƒ2 ), [โˆš ]) ( โˆš โˆš ๐›ผโ€ฒ1 ๐›ผ๐‘Ÿ + ๐›ผโ€ฒโ€ฒ1 (1โˆ’๐›ผ๐‘Ÿ ) , ๐ถ ๐‘ƒ1 +๐‘2 ๐‘ƒ๐‘Ÿ +2๐‘ ๐‘ƒ1 ๐‘ƒ๐‘Ÿ ( [โˆš ])} โˆš โˆš ๐ถ ๐‘ƒ2 +๐‘2 ๐‘ƒ๐‘Ÿ +2๐‘ ๐‘ƒ2 ๐‘ƒ๐‘Ÿ ๐›ผโ€ฒโ€ฒ2 ๐›ผ๐‘Ÿ + ๐›ผโ€ฒ2 (1โˆ’๐›ผ๐‘Ÿ ) , (9) with the union taken over all 0 โ‰ค ๐›ผ๐‘Ÿ , ๐›ผโ€ฒ1 , ๐›ผโ€ฒโ€ฒ1 , ๐›ผโ€ฒ2 , ๐›ผโ€ฒโ€ฒ2 โ‰ค 1, with ๐›ผโ€ฒ1 + ๐›ผโ€ฒโ€ฒ1 โ‰ค 1, ๐›ผโ€ฒ2 + ๐›ผโ€ฒโ€ฒ2 โ‰ค 1. Proof: The proof can be found in Appendix B. The constraint on ๐‘…1 refers to the condition that ๐‘Š1 can be decoded successfully at โ„›, ๐’Ÿ1 , and ๐’Ÿ2 , respectively, and similarly for ๐‘…2 and ๐‘…1 + ๐‘…2 . Corollary 2: For the symmetric scenario, the following equal rate constraints apply { โˆš ๐‘… < โ€ฒ maxโ€ฒโ€ฒ min ๐ถ((๐›ผโ€ฒโ€ฒ + 12 ๐‘2 +๐‘ 2๐›ผโ€ฒโ€ฒ )๐‘ƒ ), ๐›ผ โ‰ฅ0, ๐›ผ โ‰ฅ0 0โ‰ค๐›ผโ€ฒ +๐›ผโ€ฒโ€ฒ โ‰ค1

( ) โˆš ( ) ๐ถ (1โˆ’๐›ผโ€ฒโ€ฒ + 12 ๐‘2 +๐‘ 2๐›ผโ€ฒ )๐‘ƒ , 12 ๐ถ 2๐‘Ž2 ๐‘ƒ (1โˆ’๐›ผโ€ฒ โˆ’๐›ผโ€ฒโ€ฒ ) , ( )} โˆš โˆš 1 2 โ€ฒ +๐‘ 2๐›ผโ€ฒโ€ฒ )๐‘ƒ ๐ถ (1+๐‘ +๐‘ 2๐›ผ . (10) 2 Proof: Follows from (9) directly by setting ๐›ผโ€ฒ1 = ๐›ผโ€ฒ2 = ๐›ผ , ๐›ผโ€ฒโ€ฒ1 = ๐›ผโ€ฒโ€ฒ2 = ๐›ผโ€ฒโ€ฒ , and ๐›ผ๐‘Ÿ = 1/2. Without backhaul, ๐‘‹๐‘Ÿ would only be partially known by the source nodes, i.e., ๐›ผโ€ฒโ€ฒ1 = ๐›ผโ€ฒโ€ฒ2 = 0. โ€ฒ

C. Physical Layer Network Coding by Lattice Coding In contrast to Section III-A where โ„› first decodes (๐‘Š1 , ๐‘Š2 ) and then encodes into a joint NC message ๐‘Š๐‘Ÿ , the relay can (๐‘›) decode the NC message directly from ๐‘Œ๐‘Ÿ by using lattice encoding at the sources and lattice decoding at the relay, as in [8], [14] where only the case of symmetric powers is considered. We propose a protocol based on superposition of a lattice code and a random code to be able to handle the case of non-symmetric powers. Without loss of generality, we assume that ๐‘ƒ1 โ‰ค๐‘ƒ2 (hence ๐‘…1 โ‰ค๐‘…2 ). ๐’ฎ2 splits its message โ€ฒ โ€ฒโ€ฒ โ€ฒ , ๐‘Š2,๐‘ก ], where ๐‘Š2,๐‘ก has the same ๐‘Š2,๐‘ก into two parts [๐‘Š2,๐‘ก length as ๐‘Š1,๐‘ก . ๐’ฎ1 encodes ๐‘Š1,๐‘ก based on a nested lattice code [28], and we denote the corresponding transmitted code(๐‘›) โ€ฒ using the same nested word by ๐‘‰1 (๐‘Š1,๐‘ก ). ๐’ฎ2 encodes ๐‘Š2,๐‘ก lattice code as ๐’ฎ1 , denoting the corresponding codeword by (๐‘›) โ€ฒ โ€ฒโ€ฒ ๐‘‰2 (๐‘Š2,๐‘ก ), and encodes ๐‘Š2,๐‘ก using a random codebook (๐‘›) (๐‘›) (๐‘›) ๐‘›(๐‘…2 โˆ’๐‘…1 ) {๐‘‰3 } of size 2 . Note that codewords ๐‘‰1 and ๐‘‰2 are independent even though they are generated by the same nested lattice code, since the dither vectors used at ๐’ฎ1 and ๐’ฎ2 โ€ฒโ€ฒ are independent [8], [28]. The relay, after decoding ๐‘Š2,๐‘กโˆ’1 via a single-user joint-typicality decoder and the NC message

2505

โ€ฒ using a lattice decoder, encodes all these new ๐‘Š1,๐‘กโˆ’1 โŠ•๐‘Š2,๐‘กโˆ’1 messages by using an independent random codebook {๐‘ˆ (๐‘›) } of size 2๐‘›๐‘…2 , โˆš (๐‘›) โ€ฒ โ€ฒโ€ฒ ๐‘‹๐‘Ÿ,๐‘ก = ๐‘ƒ๐‘Ÿ ๐‘ˆ (๐‘›) (๐‘Š1,๐‘กโˆ’1 โŠ• ๐‘Š2,๐‘กโˆ’1 , ๐‘Š2,๐‘กโˆ’1 ).

Since ๐‘Š1,๐‘กโˆ’1 and ๐‘Š2,๐‘กโˆ’1 are known both at ๐’ฎ1 and ๐’ฎ2 โ€ฒ โ€ฒโ€ฒ thanks to the backhaul, ๐‘ˆ (๐‘›) (๐‘Š1,๐‘กโˆ’1 โŠ•๐‘Š2,๐‘กโˆ’1 , ๐‘Š2,๐‘กโˆ’1 ) is also known. Therefore ๐’ฎ1 and ๐’ฎ2 cooperate with โ„› as follows โˆš โˆš (๐‘›) (๐‘›) ๐‘‹1,๐‘ก = ๐›ฟ๐‘‰1 (๐‘Š1,๐‘ก ) + ๐‘ƒ1 โˆ’ ๐›ฟ๐‘ˆ (๐‘›) , (11) โˆš โˆš (๐‘›) โˆš (๐‘›) (๐‘›) โ€ฒ โ€ฒโ€ฒ (๐‘›) ๐‘‹2,๐‘ก = ๐›ฟ๐‘‰2 (๐‘Š2,๐‘ก ) + ๐œ–๐‘‰3 (๐‘Š2,๐‘ก ) + ๐‘ƒ2 โˆ’๐›ฟโˆ’๐œ–๐‘ˆ , where 0 โ‰ค ๐›ฟ โ‰ค ๐‘ƒ1 and 0 โ‰ค ๐œ– โ‰ค ๐‘ƒ2 โˆ’๐›ฟ are the allocated power to transmit the new messages. The corresponding received signals at the relay and destinations are ) โˆš ( (๐‘›) โˆš (๐‘›) (๐‘›) (๐‘›) + ๐‘Ž ๐œ–๐‘‰3 ๐‘Œ๐‘Ÿ,๐‘ก =๐‘Ž ๐›ฟ ๐‘‰1 +๐‘‰2 (โˆš ) โˆš (๐‘›) +๐‘Ž ๐‘ƒ1 โˆ’๐›ฟ + ๐‘ƒ2 โˆ’๐›ฟโˆ’๐œ– ๐‘ˆ (๐‘›) + ๐‘๐‘Ÿ,๐‘ก , โˆš ) โˆš (๐‘›) (โˆš (๐‘›) (๐‘›) ๐‘ƒ1 โˆ’๐›ฟ + ๐‘ ๐‘ƒ๐‘Ÿ ๐‘ˆ (๐‘›) + ๐‘1,๐‘ก , (12) ๐‘Œ1,๐‘ก = ๐›ฟ๐‘‰1 + โˆš (๐‘›) โˆš (๐‘›) (โˆš โˆš ) (๐‘›) (๐‘›) (๐‘›) ๐‘Œ2,๐‘ก = ๐›ฟ๐‘‰2 + ๐œ–๐‘‰3 + ๐‘ƒ2 โˆ’๐›ฟโˆ’๐œ–+๐‘ ๐‘ƒ๐‘Ÿ ๐‘ˆ +๐‘2,๐‘ก . ๐’Ÿ1 performs successive decoding: at the end of block ๐‘ก, (๐‘›) โ€ฒ โ€ฒโ€ฒ ๐’Ÿ1 decodes (๐‘Š1,๐‘กโˆ’1 โŠ• ๐‘Š2,๐‘กโˆ’1 , ๐‘Š2,๐‘กโˆ’1 ) from ๐‘Œ1,๐‘ก by joint โ€ฒ by using ๐‘Š1,๐‘กโˆ’1 which has typicality and recovers ๐‘Š2,๐‘กโˆ’1 been recovered successfully from block ๐‘กโˆ’1; after cancelling out ๐‘ˆ (๐‘›) the new information ๐‘Š1,๐‘ก can be decoded. This approach is also used for ๐’Ÿ2 . Proposition 3: Using lattice coding, an achievable rate region is given by { } ๐‘…1 < min ๐ถ(โˆ’1/2 + ๐‘Ž2 ๐›ฟ), ๐ถ(๐›ฟ) , { } ๐‘…2 < min ๐ถ(โˆ’1/2 + ๐‘Ž2 ๐›ฟ + ๐‘Ž2 ๐œ–/2), ๐ถ(๐›ฟ + ๐œ–) , { ( ) โˆš ๐‘…1 + ๐‘…2 < min ๐ถ ๐‘ƒ1 + ๐‘2 ๐‘ƒ๐‘Ÿ + 2๐‘ ๐‘ƒ๐‘Ÿ (๐‘ƒ1 โˆ’ ๐›ฟ) , ( )} โˆš ๐ถ ๐‘ƒ2 + ๐‘2 ๐‘ƒ๐‘Ÿ + 2๐‘ ๐‘ƒ๐‘Ÿ (๐‘ƒ2 โˆ’๐›ฟโˆ’๐œ–) , (13) with the union taken over 0 โ‰ค ๐›ฟ โ‰ค ๐‘ƒ1 and 0 โ‰ค ๐œ– โ‰ค ๐‘ƒ2 โˆ’ ๐›ฟ. Proof: The proof can be found in Appendix C. The first term in ๐‘…1 (๐‘…2 ) refers to the decoding constraint at โ„› for the nested lattice code. Corollary 3: For the symmetric scenario, the achievable rate region is { ๐‘… < max min ๐ถ(๐›ผ๐‘ƒ ), ๐ถ(โˆ’1/2 + ๐‘Ž2 ๐‘ƒ ๐›ผ), 0โ‰ค๐›ผโ‰ค1 (( ) )} โˆš 1 1 + ๐‘2 + 2๐‘ 1 โˆ’ ๐›ผ ๐‘ƒ . (14) 2๐ถ Proof: The result follows straightforwardly from (14) by setting ๐œ– = 0 and ๐›ฟ = ๐‘ƒ ๐›ผ. Without backhaul, the NC message would not be known at the sources, i.e., ๐›ฟ = ๐‘ƒ1 and ๐œ– = ๐‘ƒ2 โˆ’ ๐‘ƒ1 . D. Network (DF+NBF)

Coding

Based

Beam-forming

With

DF

To further exploit the available coherent combining (beamforming) gain [1]โ€“[3] at the sinks, we propose a new strategy that performs NC at both ๐’ฎ1 and ๐’ฎ2 but not at the relay (decreasing the complexity at โ„›). We refer to this scheme as

2506

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 9, SEPTEMBER 2011

NC based beam-forming (NBF) since the signals transmitted at ๐’ฎ1 , ๐’ฎ2 and โ„› are formed in a beamforming-like fashion. NBF requires ๐ต+2 blocks in total: (๐‘Š1,๐‘กโˆ’1 , ๐‘Š2,๐‘กโˆ’1 ) are exchanged via the backhaul during block ๐‘กโˆ’1; at block ๐‘ก the NC message ๐‘Š๐‘ก = ๐‘“ (๐‘Š1,๐‘กโˆ’1 , ๐‘Š2,๐‘กโˆ’1 ) is transmitted; at block ๐‘ก+1, ๐‘Š๐‘ก is transmitted by โ„›. The relay transmits ๐‘Š๐‘กโˆ’1 using a random codebook {๐‘ˆ (๐‘›) } of size 2๐‘›(๐‘…1 +๐‘…2 ) . For each codeword ๐‘ˆ (๐‘›) (๐‘Š๐‘กโˆ’1 ), we generate an independent random codebook {๐‘‰ (๐‘›) } of size 2๐‘›(๐‘…1 +๐‘…2 ) , and then use it to encode the new message ๐‘Š๐‘ก . We denote the selected codeword for ๐‘Š๐‘ก given ๐‘Š๐‘กโˆ’1 as ๐‘‰ (๐‘›) (๐‘Š๐‘ก , ๐‘Š๐‘กโˆ’1 ). At block ๐‘ก, the transmitted signals are โˆš (๐‘›) ๐‘‹๐‘Ÿ,๐‘ก = ๐‘ƒ๐‘Ÿ ๐‘ˆ (๐‘›) (๐‘Š๐‘กโˆ’1 ), (15) โˆš โˆš (๐‘›) (๐‘›) (๐‘›) ๐‘‹1,๐‘ก = ๐›ผ1 ๐‘ƒ1 ๐‘‰ (๐‘Š๐‘ก , ๐‘Š๐‘กโˆ’1 ) + (1โˆ’๐›ผ1 )๐‘ƒ1 ๐‘ˆ (๐‘Š๐‘กโˆ’1 ), โˆš โˆš (๐‘›) ๐‘‹2,๐‘ก = ๐›ผ2 ๐‘ƒ2 ๐‘‰ (๐‘›) (๐‘Š๐‘ก , ๐‘Š๐‘กโˆ’1 ) + (1โˆ’๐›ผ2 )๐‘ƒ2 ๐‘ˆ (๐‘›) (๐‘Š๐‘กโˆ’1 ), where 0 โ‰ค ๐›ผ1 , ๐›ผ2 โ‰ค 1 are power allocation parameters. The corresponding received signals are โˆš โˆš โˆš (๐‘›) (๐‘›) ๐‘Œ1,๐‘ก = ๐›ผ1 ๐‘ƒ1 ๐‘‰ (๐‘›) +(๐‘ ๐‘ƒ๐‘Ÿ + (1โˆ’๐›ผ1 )๐‘ƒ1 )๐‘ˆ (๐‘›) +๐‘1,๐‘ก , โˆš โˆš โˆš (๐‘›) (๐‘›) ๐‘Œ2,๐‘ก = ๐›ผ2 ๐‘ƒ2 ๐‘‰ (๐‘›) +(๐‘ ๐‘ƒ๐‘Ÿ + (1โˆ’๐›ผ2 )๐‘ƒ2 )๐‘ˆ (๐‘›) +๐‘2,๐‘ก , ) (โˆš โˆš (๐‘›) ๐›ผ1 ๐‘ƒ1 + ๐›ผ2 ๐‘ƒ2 ๐‘‰ (๐‘›) (16) ๐‘Œ๐‘Ÿ,๐‘ก =๐‘Ž ) (โˆš โˆš (๐‘›) (1โˆ’๐›ผ1 )๐‘ƒ1 + (1โˆ’๐›ผ2 )๐‘ƒ2 ๐‘ˆ (๐‘›) + ๐‘๐‘Ÿ,๐‘ก . +๐‘Ž The decoding process is similar as in the other cooperative strategies: the relay performs successive decoding and the destinations utilize backward decoding. Proposition 4: The achievable rate region for NBF is defined by ) { ( โˆš ๐‘…1 + ๐‘…2 < min ๐ถ ๐‘ƒ1 + ๐‘2 ๐‘ƒ๐‘Ÿ + 2๐‘ (1 โˆ’ ๐›ผ1 )๐‘ƒ1 ๐‘ƒ๐‘Ÿ , ( ) โˆš ๐ถ ๐‘ƒ2 + ๐‘2 ๐‘ƒ๐‘Ÿ + 2๐‘ (1 โˆ’ ๐›ผ2 )๐‘ƒ2 ๐‘ƒ๐‘Ÿ , ( ( ))} โˆš ๐ถ ๐‘Ž2 ๐›ผ1 ๐‘ƒ1 + ๐›ผ2 ๐‘ƒ2 + 2 ๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 , (17) with the union taken over the power allocation parameters 0 โ‰ค ๐›ผ1 , ๐›ผ2 โ‰ค 1. Proof: Since ๐’ฎ1 and ๐’ฎ2 transmit the same NC message ๐‘Š๐‘ก , the achievable sum-rate can be split arbitrarily between them. Therefore in the NBF strategy only the constraints for the sum-rate matter. Following similar arguments as in Appendix A, the sum-rate constraint (55c) still holds here. By (๐‘›) applying successive decoding to ๐‘Œ๐‘Ÿ,๐‘ก and backward decoding (๐‘›) (๐‘›) to ๐‘Œ1,๐‘ก and ๐‘Œ2,๐‘ก , the jointly Gaussian distributed random variables (๐‘‰ (๐‘›) , ๐‘ˆ (๐‘›) ) will translate (55c) into (17). The terms in (17) indicate the constraints at ๐’Ÿ1 , ๐’Ÿ2 , and โ„›, respectively. Corollary 4: For the symmetric scenario, the achievable rate region is defined by { (( ) )} โˆš ๐‘… < max min 12 ๐ถ(4๐‘Ž2 ๐‘ƒ ๐›ผ), 12 ๐ถ 1+๐‘2 +2๐‘ 1โˆ’๐›ผ ๐‘ƒ . 0โ‰ค๐›ผโ‰ค1

(18) Proof: The result follows straightforwardly from (17) by setting ๐›ผ1 = ๐›ผ2 = ๐›ผ. Without the backhaul, this strategy is impossible.

IV. B ENCHMARK S CHEMES AND C UT- SET B OUND To evaluate the performance of the cooperative NC strategies presented in Section III, we consider two benchmark schemes, namely the non-NC based time-sharing relay scheme and the non-DF based noisy NC scheme [15]. We also derive the cut-set bound [27] for our scenario. A. Time Sharing Relay With DF (DF+TD) In contrast to the orthogonal scheme described in [6] for the case of the IFRC, ๐’ฎ1 and ๐’ฎ2 here cooperate with โ„› to convey both messages. We first generate two independent random (๐‘›) (๐‘›) codebooks {๐‘ˆ1 } of size 2๐‘›๐‘…1 and {๐‘ˆ2 } of size 2๐‘›๐‘…2 , and they will be used by โ„› to help ๐’ฎ1 and ๐’ฎ2 , respectively. For (๐‘›) each codeword in {๐‘ˆ2 }, we generate an independent random (๐‘›) codebook {๐‘‰1 } of size 2๐‘›๐‘…1 , and then use it to encode the new message at ๐’ฎ1 . Similarly we generate a random codebook (๐‘›) (๐‘›) {๐‘‰2 } of size 2๐‘›๐‘…2 for each codeword in {๐‘ˆ1 }. During block ๐‘ก, ๐‘Š1,๐‘ก and ๐‘Š2,๐‘ก are exchanged via the backhaul, and the transmission during block ๐‘ก is divided into two parts. During the first part of block ๐‘ก, the transmitted signals are โˆš (๐‘›) (๐‘›) (๐‘›) ๐‘‹2,๐‘ก = 0, ๐‘‹๐‘Ÿ,๐‘ก1 = ๐‘ƒ๐‘Ÿ ๐‘ˆ2 (๐‘Š2,๐‘กโˆ’1 ), โˆš 1 โˆš (๐‘›) (๐‘›) (๐‘›) 1) ๐‘ˆ2 (๐‘Š2,๐‘กโˆ’1 ), ๐‘‹1,๐‘ก1 = ๐›ผ1๐›ฝ๐‘ƒ1 ๐‘‰1 (๐‘Š1,๐‘ก , ๐‘Š2,๐‘กโˆ’1 )+ ๐‘ƒ1 (1โˆ’๐›ผ ๐›ฝ where 0 โ‰ค ๐›ผ1 โ‰ค 1 is the power allocation parameter and 0 โ‰ค ๐›ฝ โ‰ค 1 is the time sharing parameter. Transmission power (๐‘›) ๐‘ƒ1 /๐›ฝ is used in ๐‘‹1,๐‘ก1 to meet the power constraint (2). The received signals are โˆš (๐‘›) (๐‘›) (๐‘›) (๐‘›) (๐‘›) ๐‘Œ2,๐‘ก1 = ๐‘๐‘‹๐‘Ÿ,๐‘ก1 + ๐‘2,๐‘ก1 = ๐‘ ๐‘ƒ๐‘Ÿ ๐‘ˆ2 + ๐‘2,๐‘ก1 , โˆš โˆš (๐‘›) (๐‘›) (๐‘›) (๐‘›) 1) ๐‘ˆ2 + ๐‘๐‘Ÿ,๐‘ก1 , (19) ๐‘Œ๐‘Ÿ,๐‘ก1 = ๐‘Ž ๐›ผ1๐›ฝ๐‘ƒ1 ๐‘‰1 + ๐‘Ž ๐‘ƒ1 (1โˆ’๐›ผ ๐›ฝ (โˆš ) โˆš โˆš (๐‘›) (๐‘›) (๐‘›) (๐‘›) ๐‘ƒ1 (1โˆ’๐›ผ1 ) ๐‘Œ1,๐‘ก1 = ๐›ผ1๐›ฝ๐‘ƒ1 ๐‘‰1 + +๐‘ ๐‘ƒ๐‘Ÿ ๐‘ˆ2 + ๐‘1,๐‘ก1 . ๐›ฝ The relay decodes ๐‘Š1,๐‘ก given ๐‘Š2,๐‘กโˆ’1 and then encodes it (๐‘›) to ๐‘ˆ1 (๐‘Š1,๐‘ก ). During the remaining part of block ๐‘ก, the transmitted signals are โˆš (๐‘›) (๐‘›) (๐‘›) ๐‘‹1,๐‘ก2 = 0, ๐‘‹๐‘Ÿ,๐‘ก2 = ๐‘ƒ๐‘Ÿ ๐‘ˆ1 (๐‘Š1,๐‘ก ), โˆš โˆš (๐‘›) (๐‘›) (๐‘›) 2) 2 ๐‘ƒ2 ๐‘‰2 (๐‘Š2,๐‘ก , ๐‘Š1,๐‘ก )+ ๐‘ƒ2 (1โˆ’๐›ผ ๐‘ˆ1 (๐‘Š1,๐‘ก ). ๐‘‹2,๐‘ก2 = ๐›ผ1โˆ’๐›ฝ 1โˆ’๐›ฝ The corresponding received signals are โˆš (๐‘›) (๐‘›) (๐‘›) (๐‘›) (๐‘›) ๐‘Œ1,๐‘ก2 = ๐‘๐‘‹๐‘Ÿ,๐‘ก2 + ๐‘1,๐‘ก2 = ๐‘ ๐‘ƒ๐‘Ÿ ๐‘ˆ1 + ๐‘1,๐‘ก2 , โˆš โˆš (๐‘›) (๐‘›) (๐‘›) (๐‘›) 2) 2 ๐‘ƒ2 ๐‘‰2 + ๐‘Ž ๐‘ƒ2 (1โˆ’๐›ผ ๐‘ˆ1 + ๐‘2,๐‘ก2 , (20) ๐‘Œ๐‘Ÿ,๐‘ก2 = ๐‘Ž ๐›ผ1โˆ’๐›ฝ 1โˆ’๐›ฝ ( ) โˆš โˆš โˆš (๐‘›) (๐‘›) (๐‘›) (๐‘›) (1โˆ’๐›ผ2 )๐‘ƒ2 2 ๐‘ƒ2 ๐‘Œ2,๐‘ก2 = ๐›ผ1โˆ’๐›ฝ ๐‘‰2 + + ๐‘ ๐‘ƒ๐‘Ÿ ๐‘ˆ1 + ๐‘2,๐‘ก2 . 1โˆ’๐›ฝ At the end of block ๐‘ก, โ„› decodes ๐‘Š2,๐‘ก given ๐‘Š1,๐‘ก , and ๐’Ÿ1 can retrieve (๐‘Š1,๐‘ก , ๐‘Š2,๐‘กโˆ’1 ) reliably using sliding-window decoding based on the received signals during block ๐‘ก. Similarly, after the first part of block ๐‘ก + 1, ๐’Ÿ2 can decode (๐‘Š2,๐‘ก , ๐‘Š1,๐‘ก ) reliably based on signals received from the first part of block ๐‘ก + 1 and the second part of block ๐‘ก.

DU et al.: COOPERATIVE NETWORK CODING STRATEGIES FOR WIRELESS RELAY NETWORKS WITH BACKHAUL

Proposition 5: The achievable rate region for this time sharing strategy is defined by { 2 ๐‘…1 < min ๐›ฝ๐ถ( ๐›ผ1 ๐‘Ž๐›ฝ ๐‘ƒ1 ), ๐›ฝ๐ถ( ๐›ผ1๐›ฝ๐‘ƒ1 ) + (1โˆ’๐›ฝ)๐ถ(๐‘2 ๐‘ƒ๐‘Ÿ ), ( )} โˆš ๐‘ƒ2 (1โˆ’๐›ฝ)๐ถ ๐‘2 ๐‘ƒ๐‘Ÿ + 1โˆ’๐›ฝ + 2๐‘ (1โˆ’๐›ผ2 )๐‘ƒ2 ๐‘ƒ๐‘Ÿ /(1โˆ’๐›ฝ) , { ๐‘Ž2 ๐‘ƒ2 2 ๐‘ƒ2 ), (1โˆ’๐›ฝ)๐ถ( ๐›ผ1โˆ’๐›ฝ )+๐›ฝ๐ถ(๐‘2 ๐‘ƒ๐‘Ÿ ), ๐‘…2 < min (1โˆ’๐›ฝ)๐ถ( ๐›ผ21โˆ’๐›ฝ ( )} โˆš ๐›ฝ๐ถ ๐‘2 ๐‘ƒ๐‘Ÿ + ๐‘ƒ๐›ฝ1 + 2๐‘ ๐‘ƒ1 ๐‘ƒ๐‘Ÿ (1โˆ’๐›ผ1 )/๐›ฝ , ๐‘…1 +๐‘…2 < min{

(21) ( ) โˆš (1โˆ’๐›ฝ)๐ถ(๐‘2 ๐‘ƒ๐‘Ÿ )+๐›ฝ๐ถ ๐‘2 ๐‘ƒ๐‘Ÿ + ๐‘ƒ๐›ฝ1 +2๐‘ ๐‘ƒ1 ๐‘ƒ๐‘Ÿ (1โˆ’๐›ผ1 )/๐›ฝ , ( )} โˆš ๐‘ƒ2 ๐‘ƒ๐‘Ÿ (1โˆ’๐›ผ2 ) 2 ๐‘ƒ2 2 , ๐›ฝ๐ถ(๐‘ ๐‘ƒ๐‘Ÿ ) + (1โˆ’๐›ฝ)๐ถ ๐‘ ๐‘ƒ๐‘Ÿ + 1โˆ’๐›ฝ +2๐‘ 1โˆ’๐›ฝ

with the union taken over all 0 โ‰ค ๐›ผ1 , ๐›ผ2 โ‰ค 1 and the time sharing parameter 0 โ‰ค ๐›ฝ โ‰ค 1. Proof: The proof follows immediately from Appendix B by applying the Gaussian condition and noting the dependence stated in (19) and (20). Constraints in ๐‘…1 (๐‘…2 ) correspond to the condition of successful decoding of ๐‘Š1 (๐‘Š2 ) at โ„›, ๐’Ÿ1 (๐’Ÿ2 ), and ๐’Ÿ2 (๐’Ÿ1 ), respectively. Constraints in ๐‘…1 + ๐‘…2 refer to successful decoding at ๐’Ÿ1 and ๐’Ÿ2 . By setting ๐›ผ1 = ๐›ผ2 = ๐›ผ and ๐›ฝ = 1/2, (21) can be translated into the symmetric rate constraint { ( ( )) โˆš ๐‘… < max min 12 ๐ถ ๐‘ƒ 2+๐‘2 +2๐‘ 2โˆ’2๐›ผ , 0โ‰ค๐›ผโ‰ค1 ( ) 1 1 2 2 2 2 ๐ถ (2๐›ผ+๐‘ +2๐›ผ๐‘ ๐‘ƒ )๐‘ƒ , 2 ๐ถ(2๐‘Ž ๐‘ƒ ๐›ผ), (( ) )]} [ โˆš 1 2 2+๐‘2 +2๐‘ 2โˆ’2๐›ผ ๐‘ƒ . (22) 4 ๐ถ(๐‘ ๐‘ƒ )+๐ถ Without backhaul, sources can only encode over their own messages. Therefore we have ๐›ผ1 = ๐›ผ2 = 1 and the first term in (22) reduces to 12 ๐ถ(๐‘2 ๐‘ƒ ). B. Noisy Network Coding (Noisy NC) The basic principle of noisy NC, as described in [15], is to convey a โ€œsuper messageโ€ ๐ต times, each time using an independent codebook and letting ๐ตโ†’โˆž, before the destination(s) can successfully decode the message. Therefore it is not clear how collaboration via the finite-rate backhaul can be implemented, since it requires a ๐ตโ†’โˆž times higher backhaul rate to exchange the super message before transmission starts. On the other hand, the backhaul provides orthogonal (i.e., out-ofband) conferencing bit-pipes between two source nodes. How to extend the noisy NC scheme [15], originally designed for relay networks with co-channel (i.e., in-band) transmission, so as to optimally utilize the rate-limited backhaul is interesting but out of the scope of this paper. The achievable rate region for noisy NC (without backhaul collaboration) can be specialized from Theorem 1 of [15] to the multicast relay network in Fig. 1 as follows ๐‘…1 < min{๐ผ(๐‘‹1 ; ๐‘Œห†๐‘Ÿ ๐‘Œ1 โˆฃ๐‘‹2 ๐‘‹๐‘Ÿ ๐‘„), ๐ผ(๐‘‹1 ; ๐‘Œห†๐‘Ÿ ๐‘Œ2 โˆฃ๐‘‹2 ๐‘‹๐‘Ÿ ๐‘„), ๐ผ(๐‘‹1 ๐‘‹๐‘Ÿ ; ๐‘Œ1 โˆฃ๐‘‹2 ๐‘„) โˆ’ ๐ผ(๐‘Œ๐‘Ÿ ; ๐‘Œห†๐‘Ÿ โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ๐‘Œ1 ๐‘„), ๐ผ(๐‘‹1 ๐‘‹๐‘Ÿ ; ๐‘Œ2 โˆฃ๐‘‹2 ๐‘„) โˆ’ ๐ผ(๐‘Œ๐‘Ÿ ; ๐‘Œห†๐‘Ÿ โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ๐‘Œ2 ๐‘„)}, ๐‘…2 < min{๐ผ(๐‘‹2 ; ๐‘Œห†๐‘Ÿ ๐‘Œ1 โˆฃ๐‘‹1 ๐‘‹๐‘Ÿ ๐‘„), ๐ผ(๐‘‹2 ; ๐‘Œห†๐‘Ÿ ๐‘Œ2 โˆฃ๐‘‹1 ๐‘‹๐‘Ÿ ๐‘„), ๐ผ(๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ1 โˆฃ๐‘‹1 ๐‘„) โˆ’ ๐ผ(๐‘Œ๐‘Ÿ ; ๐‘Œห†๐‘Ÿ โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ๐‘Œ1 ๐‘„),

2507

๐ผ(๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ2 โˆฃ๐‘‹1 ๐‘„) โˆ’ ๐ผ(๐‘Œ๐‘Ÿ ; ๐‘Œห†๐‘Ÿ โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ๐‘Œ2 ๐‘„)}, ๐‘…1 +๐‘…2 < min{๐ผ(๐‘‹1 ๐‘‹2 ; ๐‘Œห†๐‘Ÿ ๐‘Œ1 โˆฃ๐‘‹๐‘Ÿ ๐‘„), ๐ผ(๐‘‹1 ๐‘‹2 ; ๐‘Œห†๐‘Ÿ ๐‘Œ2 โˆฃ๐‘‹๐‘Ÿ ๐‘„), ๐ผ(๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ1 โˆฃ๐‘„) โˆ’ ๐ผ(๐‘Œ๐‘Ÿ ; ๐‘Œห†๐‘Ÿ โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ๐‘Œ1 ๐‘„), ๐ผ(๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ2 โˆฃ๐‘„) โˆ’ ๐ผ(๐‘Œ๐‘Ÿ ; ๐‘Œห†๐‘Ÿ โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ๐‘Œ2 ๐‘„)},

(23)

where ๐‘Œห†๐‘Ÿ is the compressed version of ๐‘Œ๐‘Ÿ , ๐‘„ is the timesharing random variable, and the joint probability can be ๐‘ฆ๐‘Ÿ โˆฃ๐‘ฅ๐‘Ÿ , ๐‘ฆ๐‘Ÿ , ๐‘ž). By partitioned as ๐‘(๐‘ž)๐‘(๐‘ฅ1 โˆฃ๐‘ž)๐‘(๐‘ฅ2 โˆฃ๐‘ž)๐‘(๐‘ฅ๐‘Ÿ โˆฃ๐‘ž)๐‘(ห† setting ๐‘„ = โˆ… and ๐‘Œห†๐‘Ÿ = ๐‘Œ๐‘Ÿ + ๐‘ห† with ๐‘ห† โˆผ ๐‘ (0, ๐œŽ 2 ), and applying (1) and (2), the achievable rate region in (23) is simplified to ๐‘…1 < min{๐ถ(๐‘Ž2 ๐‘ƒ1 /(1 + ๐œŽ 2 )), ๐ถ(๐‘2 ๐‘ƒ๐‘Ÿ ) โˆ’ ๐ถ(1/๐œŽ 2 )}, ๐‘…2 < min{๐ถ(๐‘Ž2 ๐‘ƒ2 /(1 + ๐œŽ 2 )), ๐ถ(๐‘2 ๐‘ƒ๐‘Ÿ ) โˆ’ ๐ถ(1/๐œŽ 2 )}, ๐‘…1 + ๐‘…2 < min{๐ถ(๐‘ƒ1 +๐‘Ž2 (๐‘ƒ1 +๐‘ƒ2 +๐‘ƒ1 ๐‘ƒ2 )/(1+๐œŽ 2 )), (24) ๐ถ(๐‘ƒ2 + ๐‘Ž2 (๐‘ƒ1 + ๐‘ƒ2 + ๐‘ƒ1 ๐‘ƒ2 )/(1+๐œŽ 2 )), ๐ถ(๐‘ƒ1 + ๐‘2 ๐‘ƒ๐‘Ÿ )โˆ’๐ถ(1/๐œŽ 2 ), ๐ถ(๐‘ƒ2 + ๐‘2 ๐‘ƒ๐‘Ÿ )โˆ’๐ถ(1/๐œŽ 2 )}, with the union taken over all ๐œŽ 2 > 0. Note that redundant terms have been removed from ๐‘…1 and ๐‘…2 . For the symmetric scenario, the achievable rate region is ๐‘Ž2 (2๐‘ƒ +๐‘ƒ 2 ) 1 ), 2 ๐ถ(๐‘ƒ + 1+๐œŽ2 2 1 2 1 2 ๐ถ(๐‘ ๐‘ƒ )โˆ’๐ถ(1/๐œŽ ), 2 ๐ถ(๐‘ƒ + ๐‘ ๐‘ƒ ) โˆ’ 2 ๐ถ(1/๐œŽ 2 )}. 2

๐‘Ž ๐‘ƒ ๐‘… < min{๐ถ( 1+๐œŽ 2 ),

(25)

C. Cut-Set Bound By the cut-set bound [27], the maximum achievable sum rate from the source nodes to any of the destinations can be no larger than the minimum of the mutual information flows across all possible cuts, maximized over a joint distribution for the transmitted signals. In our case, the cut-set bound between the two sources and each of the sink for the network in Fig. 1 can be derived based on four cuts, as demonstrated in Fig. 2, as follows (the dimension super script (๐‘›) is suppressed to simplify the notation) ๐‘…1 +๐‘…2 โ‰ค ๐ถcut-set =

sup

๐‘(๐‘‹1 ,๐‘‹2 ,๐‘‹๐‘Ÿ )

min{

1 1 ๐ผ(๐‘‹1 ๐‘‹2 ; ๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ ), ๐ผ(๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ1 ), ๐‘› ๐‘› } 1 1 ๐ผ(๐‘‹1 ๐‘‹2 ; ๐‘Œ2 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ ), ๐ผ(๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ2 ) , ๐‘› ๐‘›

(26)

where ๐‘‹1 , ๐‘‹2 and ๐‘‹๐‘Ÿ are potentially correlated. As suggested in [29], to find the exact cut-set bound ๐ถcut-set , we will first find an upper bound ๐ถupp โ‰ฅ ๐ถcut-set based on the technique used in [1], and then find a lower bound ๐ถcut-set, G โ‰ค ๐ถcut-set by restricting the source distribution to Gaussian, and finally show that ๐ถcut-set, G = ๐ถupp . Following the conventional notation for the differential entropy โ„Ž(๐‘‹) of a continuous valued random variable ๐‘‹, the mutual information corresponding to cut 2 can be written as ๐ผ(๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ1 ) = โ„Ž(๐‘Œ1 ) โˆ’ โ„Ž(๐‘Œ1 โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ) ๐‘› = โ„Ž(๐‘Œ1 ) โˆ’ โ„Ž(๐‘1 ) = โ„Ž(๐‘Œ1 ) โˆ’ log(2๐œ‹๐‘’), 2

(27)

2508

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 9, SEPTEMBER 2011

๐’ฎ1 ๐‘‹1

๐‘Š1

Cut 1

Backhaul

๐‘Œ๐‘Ÿ ๐’ฎ2 ๐‘‹ 2

๐‘Š2

Cut 2 ๐‘Œ1 โ„›

๐’Ÿ1

โˆš 1 ๐ผ(๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ1 ) โ‰ค ๐ถ(๐‘ƒ1 +๐‘2 ๐‘ƒ๐‘Ÿ +2๐‘ (1โˆ’๐›ผ1 )๐‘ƒ1 ๐‘ƒ๐‘Ÿ ), ๐‘› โˆš 1 ๐ผ(๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ2 ) โ‰ค ๐ถ(๐‘ƒ2 +๐‘2 ๐‘ƒ๐‘Ÿ +2๐‘ (1โˆ’๐›ผ2 )๐‘ƒ2 ๐‘ƒ๐‘Ÿ ). (37) ๐‘›

๐‘‹๐‘Ÿ Cut 4 ๐‘Œ2

Cut 3

Now, substituting (36) and (30) into (29), and applying the same approach also to cut 4, we get

๐’Ÿ2

Fig. 2. The sum multicast capacity is bounded by the cut-set bound based on the four cuts shown in the figure.

For cut 1 we have ๐ผ(๐‘‹1 ๐‘‹2 ; ๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ ) = โ„Ž(๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ )โˆ’โ„Ž(๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ) (38) = โ„Ž(๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ ) โˆ’ โ„Ž(๐‘Œ1 โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ) โˆ’ โ„Ž(๐‘Œ๐‘Ÿ โˆฃ๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ )

From the maximum entropy lemma [27], we get โ„Ž(๐‘Œ1 ) โ‰ค

๐‘› โˆ‘ ๐‘–=1

โ„Ž(๐‘Œ1,๐‘– ) โ‰ค

๐‘› โˆ‘ 1 ๐‘–=1

2

log(2๐œ‹๐‘’Var[๐‘Œ1,๐‘– ]),

(28)

where the second equality is achieved when ๐‘Œ1,๐‘– is Gaussian distributed. Hence ๐‘› 1 1โˆ‘1 ๐ผ(๐‘‹1 ๐‘‹2 ๐‘‹๐‘Ÿ ; ๐‘Œ1 ) โ‰ค log(Var(๐‘Œ1,๐‘– )) ๐‘› ๐‘› ๐‘–=1 2 ๐‘›

โ‰ค

1 1โˆ‘ Var(๐‘Œ1,๐‘– )), log( 2 ๐‘› ๐‘–=1

(29)

where the last steps follow from Jensenโ€™s inequality. Furthermore, we have Var(๐‘Œ1,๐‘– ) = Var(๐‘‹1,๐‘– + ๐‘๐‘‹๐‘Ÿ,๐‘– ) + 1 โ‰ค 1 + ๐ธ[(๐‘‹1,๐‘– + ๐‘๐‘‹๐‘Ÿ,๐‘– )2 ] 2 2 = 1 + ๐ธ[๐‘‹1,๐‘– ] + ๐‘2 ๐ธ[๐‘‹๐‘Ÿ,๐‘– ] + 2๐‘๐ธ[๐‘‹1,๐‘– ๐‘‹๐‘Ÿ,๐‘– ], (30)

with equality for ๐ธ[๐‘‹๐‘– ] = 0. Also, using the Cauchyโ€“Schwarz inequality we get

๐‘› ๐‘› ๐‘› โˆ‘ โˆ‘ 1 โˆ‘ 1 2 1 ๐ธ[๐‘‹1,๐‘– ๐‘‹๐‘Ÿ,๐‘– ] โ‰ค โŽท ๐ธ[๐‘‹๐‘Ÿ,๐‘– ] ๐ธ[(๐ธ(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– ))2 ]. ๐‘› ๐‘–=1 ๐‘› ๐‘–=1 ๐‘› ๐‘–=1

(31)

As in [1], we can introduce an auxiliary variable ๐›ผ1 โˆˆ [0, 1] such that ๐‘› 1โˆ‘ ๐ธ[(๐ธ(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– ))2 ] = (1 โˆ’ ๐›ผ1 )๐‘ƒ1 . (32) ๐‘› ๐‘–=1

= โ„Ž(๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ )โˆ’โ„Ž(๐‘1 )โˆ’โ„Ž(๐‘๐‘Ÿ ) = โ„Ž(๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ )โˆ’๐‘› log(2๐œ‹๐‘’) ๐‘› ๐‘› ( ) 1โˆ‘ 1โˆ‘ โ‰ค log (2๐œ‹๐‘’)2 โˆฃ๐‘ฒ ๐‘– โˆฃ โˆ’ ๐‘› log(2๐œ‹๐‘’) = log (โˆฃ๐‘ฒ ๐‘– โˆฃ) , 2 ๐‘–=1 2 ๐‘–=1 where the second equality in (38) comes from the fact that ๐‘Œ1 and ๐‘Œ๐‘Ÿ are independent given (๐‘‹1 , ๐‘‹2 , ๐‘‹๐‘Ÿ ) and the inequality is due to the maximum entropy lemma [27], with equality achieved by joint Gaussian distributed (๐‘Œ1,๐‘– , ๐‘Œ๐‘Ÿ,๐‘– ) with conditional covariance matrices ๐‘ฒ ๐‘– defined by [ ] ๐ธ[Var(๐‘Œ1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐ธ[Cov(๐‘Œ1,๐‘– , ๐‘Œ๐‘Ÿ,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐‘ฒ๐‘–= , (39) ๐ธ[Cov(๐‘Œ1,๐‘– , ๐‘Œ๐‘Ÿ,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐ธ[Var(๐‘Œ๐‘Ÿ,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] where ๐ธ[Var(๐‘Œ1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] = 1 + ๐ธ[Var(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )], ๐ธ[Cov(๐‘Œ1,๐‘– , ๐‘Œ๐‘Ÿ,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] = ๐‘Ž๐ธ[Cov(๐‘‹1,๐‘– , ๐‘‹1,๐‘– +๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] = ๐‘Ž(๐ธ[Var(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )]+๐ธ[Cov(๐‘‹1,๐‘– , ๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )]), (40) ๐ธ[Var(๐‘Œ๐‘Ÿ,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] = 1 + ๐‘Ž2 ๐ธ [Var(๐‘‹1,๐‘– +๐‘‹2,๐‘– )โˆฃ๐‘‹๐‘Ÿ,๐‘– ] = 1 + ๐‘Ž2 (๐ธ[Var(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] + ๐ธ[Var(๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] + 2๐ธ[Cov(๐‘‹1,๐‘– , ๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )]). (41) Obviously, the covariance matrices ๐‘ฒ ๐‘– are positive semidefinite. Since the function log โˆฃ๐‘ฒโˆฃ is concave [30], we can thus bound the throughput of cut 1 as follows ๐‘›

1โˆ‘1 1 ๐ผ(๐‘‹1 ๐‘‹2 ; ๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ ) โ‰ค log (โˆฃ๐‘ฒ ๐‘– โˆฃ) ๐‘› 2 ๐‘–=1 ๐‘› ) ( ๐‘› 1 โˆ‘  1   โ‰ค log  ๐‘ฒ๐‘– . ๐‘›  2

Then by applying the power constraint defined in (2), we have ๐‘› ๐‘› 1โˆ‘ 1โˆ‘ 2 ๐ธ[Var(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] = (๐ธ[๐‘‹1,๐‘– ]โˆ’๐ธ[(๐ธ(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– ))2 ]) ๐‘› ๐‘–=1 ๐‘› ๐‘–=1

โ‰ค ๐›ผ1 ๐‘ƒ1 .

(33)

Similarly, we can introduce ๐›ผ2 โˆˆ [0, 1] such that ๐‘› 1โˆ‘ ๐ธ[(๐ธ(๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– ))2 ] = (1 โˆ’ ๐›ผ2 )๐‘ƒ2 , ๐‘› ๐‘–=1 ๐‘› 1โˆ‘ ๐ธ[Var(๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] โ‰ค ๐›ผ2 ๐‘ƒ2 . ๐‘› ๐‘–=1

(34)

By substituting (32) into (31) we get ๐‘›

โˆš 1โˆ‘ ๐ธ[๐‘‹1,๐‘– ๐‘‹๐‘Ÿ,๐‘– ] โ‰ค (1 โˆ’ ๐›ผ1 )๐‘ƒ1 ๐‘ƒ๐‘Ÿ . ๐‘› ๐‘–=1

It is then straightforward to show that for the inner term in (42) we can get

๐‘› ๐‘›

โˆ‘

1 + ๐‘Ž2 โˆ‘

1

๐‘ฒ ๐‘– = 1 + ๐ธ[Var(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )]

๐‘›

๐‘› ๐‘–=1 ๐‘–=1 +

(35)

(36)

(42)

๐‘–=1

๐‘› ๐‘› ๐‘Ž2 โˆ‘ 2๐‘Ž2 โˆ‘ ๐ธ[Var(๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] + ๐ธ[Cov(๐‘‹1,๐‘– , ๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐‘› ๐‘–=1 ๐‘› ๐‘–=1 ) ( ๐‘› ๐‘› 1โˆ‘ 1โˆ‘ 2 +๐‘Ž ๐ธ[Var(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐ธ[Var(๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐‘› ๐‘–=1 ๐‘› ๐‘–=1 )2 ( ๐‘› 1โˆ‘ 2 โˆ’๐‘Ž ๐ธ[Cov(๐‘‹1,๐‘– , ๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] . (43) ๐‘› ๐‘–=1

DU et al.: COOPERATIVE NETWORK CODING STRATEGIES FOR WIRELESS RELAY NETWORKS WITH BACKHAUL

โˆš As Cov(๐‘‹1,๐‘– , ๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )=๐œ™๐‘– Var(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )Var(๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– ), where โˆฃ๐œ™๐‘– โˆฃ โ‰ค 1 is the correlation coefficient, we have ๐‘›

1โˆ‘ ๐ธ[Cov(๐‘‹1,๐‘– , ๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐‘› ๐‘–=1   ๐‘› ๐‘› 1 โˆ‘ 1โˆ‘ โ‰คโŽท ๐œ™๐‘– ๐ธ[Var(๐‘‹1,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐œ™๐‘– ๐ธ[Var(๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] ๐‘› ๐‘–=1 ๐‘› ๐‘–=1 โˆš โ‰ค ๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 , (44) where the first inequality is due to the Cauchyโ€“Schwarz inequality and the last step is given by (33) and (35). Given that โˆฃ๐œ™๐‘– โˆฃ โ‰ค 1, we can introduce another auxiliary variable 0 โ‰ค ๐œŒ โ‰ค 1 such that ๐‘›

โˆš 1โˆ‘ ๐ธ[Cov(๐‘‹1,๐‘– , ๐‘‹2,๐‘– โˆฃ๐‘‹๐‘Ÿ,๐‘– )] = ๐œŒ ๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 . ๐‘› ๐‘–=1

(45)

Now, by substituting (33), (35) and (45) into (43), the bound (42) can be translated into 1 ๐ผ(๐‘‹1 ๐‘‹2 ; ๐‘Œ1 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ ) โ‰ค ๐ถ(๐›ผ1 ๐‘ƒ1 (46) ๐‘› ) โˆš +๐‘Ž2 (๐›ผ1 ๐‘ƒ1 +๐›ผ2 ๐‘ƒ2 +2๐œŒ ๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 +(1โˆ’๐œŒ2 )๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 ) . Similarly, we can bound the throughput of cut 3 as follows 1 ๐ผ(๐‘‹1 ๐‘‹2 ; ๐‘Œ2 ๐‘Œ๐‘Ÿ โˆฃ๐‘‹๐‘Ÿ ) โ‰ค ๐ถ(๐›ผ2 ๐‘ƒ2 (47) ๐‘› ) โˆš +๐‘Ž2 (๐›ผ2 ๐‘ƒ2 +๐›ผ1 ๐‘ƒ1 +2๐œŒ ๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 +(1โˆ’๐œŒ2 )๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 ) . By combining the individual bounds defined by (37), (46) and (47), the cut-set bound ๐ถcut-set in (26) can be upper bounded by ๐ถupp , as defined in (48) at the bottom of this page. On the other hand, we can also lower bound ๐ถcut-set by ๐ถcut-set, G , obtained by restricting ๐‘(๐‘‹1 , ๐‘‹2 , ๐‘‹๐‘Ÿ ) in (26) to be Gaussian. We partition the Gaussian variables ๐‘‹1 , ๐‘‹2 and ๐‘‹๐‘Ÿ as follows โˆš ๐‘ƒ๐‘Ÿ ๐‘ˆ, (49a) โˆš โˆš โˆš ๐‘‹1 = (1โˆ’๐œŒ)๐›ผ1 ๐‘ƒ1 ๐‘†1 + ๐œŒ๐›ผ1 ๐‘ƒ1 ๐‘‰ + (1โˆ’๐›ผ1 )๐‘ƒ1 ๐‘ˆ, (49b) โˆš โˆš โˆš ๐‘‹2 = (1โˆ’๐œŒ)๐›ผ2 ๐‘ƒ2 ๐‘†2 + ๐œŒ๐›ผ2 ๐‘ƒ2 ๐‘‰ + (1โˆ’๐›ผ2 )๐‘ƒ2 ๐‘ˆ, (49c) ๐‘‹๐‘Ÿ =

where ๐‘†1 , ๐‘†2 , ๐‘‰ , and ๐‘ˆ are ๐‘›-dimensional independent Gaussian random vectors with zero-mean and unit-variance. 0 โ‰ค ๐›ผ1 , ๐›ผ2 , ๐œŒ โ‰ค 1 are auxiliary variables introduced to represent the potential correlation among ๐‘‹1 , ๐‘‹2 and ๐‘‹๐‘Ÿ due

๐ถupp =

2509

to cooperation. The received signals are then โˆš โˆš โˆš ๐‘Œ1 = (1โˆ’๐œŒ)๐›ผ1 ๐‘ƒ1 ๐‘†1 + (๐‘ ๐‘ƒ๐‘Ÿ + (1โˆ’๐›ผ1 )๐‘ƒ1 )๐‘ˆ โˆš + ๐œŒ๐›ผ1 ๐‘ƒ1 ๐‘‰ + ๐‘1 , (50a) โˆš โˆš โˆš ๐‘Œ2 = (1โˆ’๐œŒ)๐›ผ2 ๐‘ƒ2 ๐‘†2 + (๐‘ ๐‘ƒ๐‘Ÿ + (1โˆ’๐›ผ2 )๐‘ƒ2 )๐‘ˆ โˆš + ๐œŒ๐›ผ2 ๐‘ƒ2 ๐‘‰ + ๐‘2 , (50b) โˆš โˆš โˆš ๐‘Œ๐‘Ÿ = ๐‘Ž (1โˆ’๐œŒ)๐›ผ1 ๐‘ƒ1 ๐‘†1 +๐‘Ž( (1โˆ’๐›ผ1 )๐‘ƒ1 + (1โˆ’๐›ผ2 )๐‘ƒ2 )๐‘ˆ โˆš โˆš โˆš +๐‘Ž (1โˆ’๐œŒ)๐›ผ2 ๐‘ƒ2 ๐‘†2 +๐‘Ž( ๐œŒ๐›ผ1 ๐‘ƒ1 + ๐œŒ๐›ผ2 ๐‘ƒ2 )๐‘‰ +๐‘๐‘Ÿ . (50c) By substituting (50) into (27) and (38), we can derive from (26) ๐‘›

๐ถcut-set, G =

sup

0โ‰ค๐›ผ1 ,๐›ผ2 ,๐œŒโ‰ค1

min

1 โˆ‘ {log(Var(๐‘Œ1,๐‘– )), 2๐‘› ๐‘–=1

log(Var(๐‘Œ2,๐‘– )), log (โˆฃ๐‘ฒ 1,๐‘– โˆฃ) , log (โˆฃ๐‘ฒ 2,๐‘– โˆฃ)} , (51) where for ๐‘– = 1, ..., ๐‘›, we have

โˆš Var(๐‘Œ1,๐‘– ) = 1 + ๐‘ƒ1 + ๐‘2 ๐‘ƒ๐‘Ÿ + 2๐‘ (1 โˆ’ ๐›ผ1 )๐‘ƒ1 ๐‘ƒ๐‘Ÿ , โˆš Var(๐‘Œ2,๐‘– ) = 1 + ๐‘ƒ2 + ๐‘2 ๐‘ƒ๐‘Ÿ + 2๐‘ (1 โˆ’ ๐›ผ2 )๐‘ƒ2 ๐‘ƒ๐‘Ÿ , โˆš โˆฃ๐‘ฒ 1,๐‘– โˆฃ = 1+(1+๐‘Ž2 )๐›ผ1 ๐‘ƒ1 +๐‘Ž2 ๐›ผ2 ๐‘ƒ2 +2๐‘Ž2 ๐œŒ ๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 +๐‘Ž2 (1โˆ’๐œŒ2 )๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 , โˆš โˆฃ๐‘ฒ 2,๐‘– โˆฃ = 1+(1+๐‘Ž2 )๐›ผ2 ๐‘ƒ2 +๐‘Ž2 ๐›ผ2 ๐‘ƒ2 +2๐‘Ž2 ๐œŒ ๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2

(52)

+๐‘Ž2 (1โˆ’๐œŒ2 )๐›ผ1 ๐›ผ2 ๐‘ƒ1 ๐‘ƒ2 . By substituting (52) into (51) we get ๐ถcut-set, G which actually equals to ๐ถupp as defined in (48), i.e., ๐ถcut-set, G = ๐ถupp . Recall that ๐ถcut-set, G โ‰ค ๐ถcut-set โ‰ค ๐ถupp , we can finally conclude that ๐ถcut-set = ๐ถupp , i.e., the capacity upper bound defined in (48) is actually the cut-set bound. For the symmetric scenario where ๐‘ƒ1 = ๐‘ƒ2 = ๐‘ƒ๐‘Ÿ = ๐‘ƒ , by setting ๐›ผ = ๐›ผ1 = ๐›ผ2 , the cut-set bound defined in (48) can be translated to the following constraint { โˆš )) 1 ( ( ๐ถ ๐‘ƒ 1 + ๐‘2 + 2๐‘ 1 โˆ’ ๐›ผ , ๐‘… < sup min 2 0โ‰ค๐›ผ,๐œŒโ‰ค1 } ) 1 ( 2 2 2 2 2 ๐ถ ๐‘ƒ [(1 + 2๐‘Ž )๐›ผ + 2๐‘Ž ๐œŒ๐›ผ + ๐‘Ž (1 โˆ’ ๐œŒ )๐›ผ ๐‘ƒ ] . (53) 2 D. Achievability of the Cut-Set Bound by DF+NBF Proposition 6: In the symmetric scenario where ๐‘ƒ1 = ๐‘ƒ2 = ๐‘ƒ๐‘Ÿ = ๐‘ƒ and ๐‘…1 = ๐‘…2 = ๐‘…, DF+NBF can achieve the cutset bound, i.e. (18) and (53) are equivalent, if and only if (๐‘Ž2 , ๐‘2 , ๐‘ƒ ) satisfy { 4๐‘Ž2 > max{2, 1 + ๐‘2 } 8๐‘Ž2 (2๐‘Ž2 โˆ’1) (54) โˆš 2 2 . 01+๐‘2 , i.e., there should exist two variables 0 โ‰ค ๐›ผโˆ— , ๐œŒ โ‰ค 1 such that โˆš (66a) 4๐‘Ž2 ๐›ผโˆ— =1 + ๐‘2 + 2๐‘ 1 โˆ’ ๐›ผโˆ— , 4๐‘Ž2 ๐›ผโˆ— =๐›ผโˆ— + ๐‘Ž2 (2๐›ผโˆ— + 2๐›ผโˆ— ๐œŒ + (1 โˆ’ ๐œŒ2 )(๐›ผโˆ— )2 ๐‘ƒ ).

(66b)

By subtracting 1 + ๐‘2 from both sides of (66a) and then taking square, we have 16๐‘Ž4 (๐›ผโˆ— )2 โˆ’ ๐›ผโˆ— (8๐‘Ž2 + 8๐‘Ž2 ๐‘2 โˆ’ 4๐‘2 ) + (1 โˆ’ ๐‘2 )2 = 0, which has only one true root for (66a) (must satisfy 4๐‘Ž2 ๐›ผโˆ— > 1 + ๐‘2 ) โˆš 2๐‘Ž2 (1 + ๐‘2 ) โˆ’ ๐‘2 + (4๐‘Ž2 โˆ’ ๐‘2 )(4๐‘Ž2 โˆ’ ๐‘Ž)๐‘2 โˆ— ๐›ผ = . (67) 8๐‘Ž4 From (66b) we get โˆš ๐œŒ๐›ผโˆ— = 1/๐‘ƒ + ๐›ผโˆ— /(๐‘Ž2 ๐‘ƒ ) + (๐›ผโˆ— โˆ’ 1/๐‘ƒ )2 , or โˆš ๐œŒ๐›ผโˆ— = 1/๐‘ƒ โˆ’ ๐›ผโˆ— /(๐‘Ž2 ๐‘ƒ ) + (๐›ผโˆ— โˆ’ 1/๐‘ƒ )2 . Since 0 โ‰ค ๐œŒ๐›ผโˆ— โ‰ค ๐›ผโˆ— , the first root is obviously a false root and therefore omitted. To make the second root satisfy the constraint, we must have โˆš 0 โ‰ค 1/๐‘ƒ โˆ’ ๐›ผโˆ— /(๐‘Ž2 ๐‘ƒ ) + (๐›ผโˆ— โˆ’ 1/๐‘ƒ )2 โ‰ค ๐›ผโˆ— . The second inequality is self-evident, and the first inequality requires 1 1 ๐‘Ž2 > 1/2 and ๐›ผโˆ— โ‰ค (2 โˆ’ 2 ). (68) ๐‘ƒ ๐‘Ž We therefore conclude from (67) and (68), given 4๐‘Ž2 > 1 + ๐‘2 and ๐‘ƒ > 0, that (66) holds if and only if 4๐‘Ž2 > 2 and โˆš 2๐‘Ž2 (1 + ๐‘2 ) โˆ’ ๐‘2 + (4๐‘Ž2 โˆ’ ๐‘2 )(4๐‘Ž2 โˆ’ ๐‘Ž)๐‘2 1 1 โ‰ค (2 โˆ’ 2 ). 4 8๐‘Ž ๐‘ƒ ๐‘Ž Combined with the finding that ๐‘”NBF = ๐‘”cut-set is impossible for 4๐‘Ž2 โ‰ค 1+๐‘2, we can conclude that ๐‘”NBF = ๐‘”cut-set , i.e. (18) and (53) are identical, if and only if (๐‘Ž2 , ๐‘2 , ๐‘ƒ ) satisfies (54). ACKNOWLEDGMENTS The authors would like to thank the anonymous reviewers and the associate editor for their suggestions that helped improve the quality and presentation of the paper. R EFERENCES [1] T. M. Cover and A. El Gamal, โ€œCapacity theorems for the relay channel,โ€ IEEE Trans. Inf. Theory, vol. 25, pp. 572โ€“584, Sep. 1979. [2] A. Hรธst-Madsen and J. Zhang, โ€œCapacity bounds and power allocation for wireless relay channels,โ€ IEEE Trans. Inf. Theory, vol. 51, pp. 2020โ€“ 2040, June 2005. [3] G. Kramer, M. Gastpar, and P. Gupta, โ€œCooperative strategies and capacity theorems for relay networks,โ€ IEEE Trans. Inf. Theory, vol. 51, pp. 3037โ€“3063, Sep. 2005. [4] G. Kramer and A. J. van Wijngaarden, โ€œOn the white Gaussian multipleaccess relay channel,โ€ in Proc. IEEE ISIT, Jan. 2000. [5] Y. Liang and G. Kramer, โ€œRate regions for relay broadcast channels,โ€ IEEE Trans. Inf. Theory, vol. 53, pp. 3517โ€“3535, Oct. 2007. [6] O. Sahin and E. Erkip, โ€œAchievable rates for the Gaussian interference relay channel,โ€ in Proc. IEEE GLOBECOM, Nov. 2007. [7] D. Gรผndรผz, O. Simeone, A. J. Goldsmith, H. V. Poor, and S. Shamai, โ€œRelaying simultaneous multicast messages,โ€ in Proc. IEEE ITW, June 2009.

2513

[8] D. Gรผndรผz, O. Simeone, A. J. Goldsmith, H. V. Poor, and S. Shamai, โ€œMultiple multicasts with the help of a relay,โ€ IEEE Trans. Inf. Theory, vol. 56, pp. 6142โ€“6158, Dec. 2010. [9] A. Avestimehr and T. Ho, โ€œApproximate capacity of the symmetric halfduplex Gaussian butterfly network,โ€ in Proc. IEEE ITW, June 2009. [10] S. R. Li, R. W. Yeung, and N. Cai, โ€œLinear network coding,โ€ IEEE Trans. Inf. Theory, vol. 49, no. 2, pp. 371โ€“381, Feb. 2003. [11] R. Koetter and M. Mรฉdard, โ€œAn algebraic approach to network coding,โ€ IEEE/ACM Trans. Netw., vol. 11, pp. 782โ€“795, Oct. 2003. [12] S. Katti, H. Rahul, W. Hu, D. Katabi, M. Mรฉdard, and J. Crowcroft, โ€œXORs in the air: practical wireless network coding,โ€ IEEE/ACM Trans. Netw., vol. 16, pp. 497โ€“510, June 2008 [13] S. Katti, I. Mariยดc, A. J. Goldsmith, D. Katabi, and M. Mรฉdard, โ€œJoint relaying and network coding in wireless networks,โ€ in Proc. IEEE ISIT, June 2007. [14] M. P. Wilson, K. Narayanan, H. D. Pfister, and A. Sprintson, โ€œJoint physical layer coding and network coding for bi-directional relaying,โ€ IEEE Trans. Inf. Theory, vol. 56, no. 11, pp. 5641โ€“5654, Nov. 2010. [15] S. H. Lim, Y.-H. Kim, A. El Gamal, and S.-Y. Chung, โ€œNoisy network coding,โ€ IEEE Trans. Inf. Theory, vol. 57, no. 5, pp. 3132โ€“3152, May 2011. [16] F. M. J. Willems, โ€œThe discrete memoryless multiple-access channel with partially cooperating encoders,โ€ IEEE Trans. Inf. Theory, vol. 29, pp. 441โ€“445, May 1983. [17] I. Mariยดc, R. Yates, and G. Kramer, โ€œCapacity of interference channels with partial transmitter cooperation,โ€ IEEE Trans. Inf. Theory, vol. 53, pp. 3536โ€“3548, Oct. 2007. [18] S. I. Bross, A. Lapidoth, and M. A. Wigger, โ€œThe Gaussian MAC with conferencing encoders,โ€ in Proc. IEEE ISIT, July 2008. [19] N. Devroye, P. Mitran, and V. Tarokh, โ€œAchievable rates in cognitive radio channels,โ€ IEEE Trans. Inf. Theory, vol. 52, pp. 1813โ€“1827, May 2006. [20] C. T. K Ng, N. Jindal, A. J. Goldsmith, and U. Mitra, โ€œCapacity gain from two-transmitter and two-receiver cooperation,โ€ IEEE Trans. Inf. Theory, vol. 53, pp. 3822โ€“3827, Oct. 2007. [21] O. Simeone, D. Gรผndรผz, H. V. Poor, A. J. Goldsmith, and S. Shamai, โ€œCompound multiple-access channels with partial cooperation,โ€ IEEE Trans. Inf. Theory, vol. 55, pp. 2425โ€“2441, June 2009. [22] J. N. Laneman, D. N. C. Tse, and G. W. Wornell, โ€œCooperative diversity in wireless networks: efficient protocols and outage behavior,โ€ IEEE Trans. Inf. Theory, vol. 50, pp. 3062โ€“3080, Dec. 2004. [23] B. Wang, J. Zhang, and A. Hรธst-Madsen, โ€œOn the capacity of MIMO relay channels,โ€ IEEE Trans. Inf. Theory, vol. 51, pp. 29โ€“43, Jan. 2005. [24] S. Simoens, O. Muรฑoz-Medina, J. Vidal, and A. del Coso, โ€œOn the Gaussian MIMO relay channel with full channel state information,โ€ IEEE Trans. Signal Process., vol. 57, pp. 3588โ€“3599, Sep. 2009. [25] A. B. Carleial, โ€œMultiple-access channels with different generalized feedback signals,โ€ IEEE Trans. Inf. Theory, vol. 28, pp. 841โ€“850, Nov. 1982. [26] F. M. J. Willems, โ€œInformation theoretical results for the discrete memoryless multiple access channel,โ€ Ph.D. thesis, Katholieke Univ. Leuven, Oct. 1982. [27] T. M. Cover and J. A. Thomas, Elements of Information Theory. Wiley, 2006. [28] U. Erez and R. Zamir, โ€œAchieving 1/2 log (1+ SNR) on the AWGN channel with lattice encoding and decoding,โ€ IEEE Trans. Inf. Theory, vol. 50, pp. 2293โ€“2314, Oct. 2004. [29] A. El Gamal and Y.-H. Kim, Lecture Notes on Network Information Theory, chap. 17. Available: http://arxiv.org/abs/1001.3404 [30] T. M. Cover and J. A. Thomas, โ€œDeterminant inequalities via information theory,โ€ SIAM J. Matrix Analysis Applications, vol. 9, pp. 384โ€“392, July 1998. [31] B. Rimoldi and R. Urbanke, โ€œA rate-splitting approach to the Gaussian multiple-access channel,โ€ IEEE Trans. Inf. Theory, vol. 42, pp. 364โ€“375, Mar. 1996.

2514

Jinfeng Du (Sโ€™07) received his B.Eng. degree in electronic information engineering from the University of Science and Technology of China (USTC), Hefei, China, in 2004, his M.Sc. degree in electronic engineering, and his Tekn. Lic. degree in electronics and computer systems, both from the Royal Institute of Technology (KTH), Stockholm, Sweden, in 2006 and 2008, respectively. He is currently working for his Ph.D. degree in telecommunications at KTH. He received the best paper award from IC-WCSP, Suzhou, October 2010, the โ€œHans Werthรฉn Grantโ€ from the Royal Swedish Academy of Engineering Science (IVA) in March 2011, and a grant from Ericssonโ€™s Research Foundation in May 2011. Ming Xiao (Sโ€™02-Mโ€™07) was born in the SiChuan Province, P. R. China, on May 22nd, 1975. He received bachelor and master degrees in engineering from the University of Electronic Science and Technology of China, ChengDu, in 1997 and 2002, respectively. He received his Ph.D degree from Chalmers University of Technology, Sweden, in November 2007. From 1997 to 1999, he worked as a network and software assistant engineer for ChinaTelecom. From 2000 to 2002, he also held a position in the SiChuan Communications Administration. Since November 2007, he has been with the ACCESS Linnaeus Center, School of Electrical Engineering, Royal Institute of Technology, Sweden, where he is currently an assistant professor. He received the โ€œChinese Government Award for Outstanding Self-Financed Students Studying Aboradโ€ in 2007. He received the โ€œHans Werthรฉn Grantโ€ from the Royal Swedish Academy of Engineering Science (IVA) in March 2006, and a grant from Ericssonโ€™s Research Foundation in 2010.

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 9, SEPTEMBER 2011

Mikael Skoglund (Sโ€™93-Mโ€™97-SMโ€™04) received the Ph.D. degree in 1997 from Chalmers University of Technology, Sweden. In 1997, he joined the Royal Institute of Technology (KTH), Stockholm, Sweden, where he was appointed to the Chair in Communication Theory in 2003. At KTH, he heads the Communication Theory Lab and he is the Assistant Dean for Electrical Engineering. Dr. Skoglundโ€™s research interests are in the theoretical aspects of wireless communications. He has worked on problems in source-channel coding, coding and transmission for wireless communications, Shannon theory, and statistical signal processing. He has authored some 220 scientific papers, including papers that have received awards, invited conference presentations, and papers ranking as highly cited according to the ISI Essential Science Indicators. He has also consulted for industry, and he holds six patents. Dr. Skoglund has served on numerous technical program committees for IEEE conferences. During 2003โ€“2008, he was an Associate Editor with the IEEE T RANSACTIONS ON C OMMUNICATIONS and he is presently on the editorial board for the IEEE T RANSACTIONS ON I NFORMATION T HEORY.