Energy-Efficient Antenna Sharing and Relaying for Wireless Networks J. Nicholas Laneman and Gregory W. Wonell Research Laboratory of Electronics Massachusetts Institute of Technology Cambridge, MA 02139 USA I. INTRODUCTION Relaying information over several point-to-point communi- cation links is @ basic building block of communication net- ‘works. Such relaying is utilized in wired and wireless networks to chive higher network connoctvity (broader coverage), ef- ficient utilization of resources such as power and bandwidth, better economies of scale in the cost of long-haul transmis- sions (through traffic aggregation), interoperability among net- works, and more easily manageable, hierarchical network ar- chitectures. In wireless networks, direct transmission between widely separated radios can be very expensive in terms of transmit- ted power required for reliable communication. High-power transmissions lead to faster battery drain (shorter network life) 2s well as increased interference at nearby radios. As alterna tives to direct transmission, there are two basic and frequently- employed examples of relayed transmission for wireless net- Works. In cellular settings, for example, newworks provide con- nectivity between low-power mobiles by providing focal con- nections to high-power basestations that are relayed via a line basestaion network. In sensor networks, and military b Uleield communication networks in general, the use of wireline infrastructure is often precluded and the radios may be sub- stantially power constrained; fr these ad-hoc or peer-to-peer networks, transmissions can be relayed wirelessly. As these ‘examples suggest, relayed transmission enlists two or more a- dios to perform multiple transmissions. The end-to-end trans- ‘missions potentially incur higher delay, but because the indi- ‘vidual transmissions are over shorter distances (in the wireless case), oF over high-quality cabling (in the wireline case), the ‘This work hasbeen spporedin par by ARL Federated Labs unde Cooper ave Apreement No DAALDI- 96-2000 andy NSF unde Grant No. CCR ‘979303 well ar though an NSF Grate Reseach Fellows 0-7803-6596-#/00/810.00 © 2000 IEEE ig. 1. Example hero (uber for which relaying protocols, and ‘special ancnna sang o ives, protocol canbe otvaed and “evloped. Inet are the tamed signa mand, te eee Siena, and and he ao separations power requirements for reliable communication can be much Tower. The basic relaying protocols described above are con- structed from the sequential use of point-to-point links, where the links are essentially viewed atthe network protocol layer; however, more general approaches are possible that involve the coordination of both the direct and relayed transmissions, at the network and lower protocol layers, and correspond to sce- narios to which the classical relay channel model (1] applies. In this paper, we develop energy-efficient relaying protocols ‘that create and exploit spatial diversity to combat fading due to ‘multipath propagation, a particularly severe form of interfer ence experienced in wireless networks. ‘oillustratethe main concepis, we consider the simple wire- less network depicted in Fig. 1. We focus specifically on rans- missions from radio 1, called the source, to radio 3, called the destination, with the possibility of employing radio 2 as a re- lay. At the physical layer, the destination receives potentially useful signals from all transmitters that are active, and may ‘combine multiple transmissions of the same signal to reduce variations in performance caused by signal fading, a technique refered to broadly as spatial diversity combining [2). We refer to this form of spatial diversity as antenna sharing, in con- ‘rast tothe curently more conventional forms of spatial diver- sity 3], because the radios essentially share their antennas and ‘other resources to create a “virtual array” through distributed transmission and signal processing. ‘After developing a mathematical mode! in Section II forthe network in Fig. 1, we scratch the surface of the rich set of de-sign issues and options that arise in the context of antenna shar- ing and relaying for wireless nerworks, Section IH casts the basic relaying protocols, referred 1 as singlehop and multihop ‘transmission, respectively, into our framework, and explores a ‘number of possibilities for antenna sharing protocols, in terms of what signals the source and relay jointly wansmit as well as how the relay and destination jointly process signals. Per- formance comparisons, and simulation results in Section IV, suggest that antenna sharing transmission protocols are capa- ble of overcoming the noisy channels between the distributed radio antennas to achieve diversity gain and outperform sin- glehop and multihop transmission ina variety of scenarios of interest I, System MopEL, Inour mode! forthe three-radio wireless network depicted in Fig. 1, narrowband transmissions suffer the effects of path loss and flat fading as arse in e.., slow-frequency-hop networks ‘Our analysis focuses on the case of slow fading to isolate the benefits of spatial diversity alone; however, we emphasize at the outset that our results extend naturally othe kinds of highly ‘mobile scenarios in which faster fading is encountered. Our baseband-equivalent, diserete-time channel model for the network in Fig. 1 consists of two subchannels, onhogonal in, eg, adjacent time slots or frequencies. This decomposition is necessary because practical limitations in radio implemen- tation prevent the relay from simultaneously transmitting and receiving on the same channel. On the first subchannel, the source transmits a sequence x(n], with average sample energy 1, and the relay and destination receive signal yaln) = a12 VE mln) + zfn], ay valr] = a1 VE. mile] + zaln), @ respectively. On the second subchannel, the relay transmits 2 sequence xj[n], with average sample energy <1, and the destination receives! y5lrd = 20 VEs able) + 23m @ loss and state fading on transmissions from radio ito radio j,&; isthe transmitted en- ergy of radio i and z)(n} and z4(n] model additive receiver noise and other forms of interference. ‘Statistically, we model the fading coefficients ay as zer0- ‘mean, mutually independent complex jointly Gaussian ran- dom variables with variances 03,,, and we model the additive noises zj[n] and z5[n} as zero-mean, mutually independeat, ‘white complex jointly Gaussian sequences with variance Nj ‘We define the signal-to-noise rato (SNR) in each received nal as 7.3 © |ai,[? &/Nos under the Rayleigh fading model, "We employ he nttion()" to dsingih he signal the second sab tanal from hase onthe fst: he fading coeicients 2 athe ame be ‘use the subchanals are assume tobe adjacent andthe Tading is at arous ffeaency pn va oe mo the SNRs are independent exponential random variables with expected values ¥,; © Bltij] = 03, £:/No. IIL, TRANSMISSION PROTOCOLS ‘Within the physical layer framework described in Section Tl, ‘we examine several protocols that support transmission be- tween the source and destination. Each protocol consists of| 1 source modulation format, a relay processing/modulation scheme, and a destination receiver structure, For simplicity of exposition, we treat coherently detected, constant-modulus binary transmissions, so that the source transmitted signal x;[n] is white and takes values 29 and 21 with equal probability. To enable coherent detection, the re- lay and destination recsivers must frst obtain, via training so- {quences in the protocol headers, accurate estimates of the link fading coefficients; in several scenarios, the destination also utilizes an estimate of 72. We assume these estimates are perfect in our preliminary analysis ‘All of our destination receiver structures can be imple- ‘mented as shown in Fig. 2. This “combiner” can be viewed as a generalized matched-filter, or maximum-ratio combiner, suitably modified to fit the protocol. As we will see, quai- {ative comparisons among the various transmission protocols can be made by examining their respective weights w and w/ as well a their mappings f(). A. Singlehop Transmission Singlehop transmission, often referred to as singlehop rout- ing in the ad-hoc networking community [4], consists of direct ‘transmission between the source and destination radios. In this, ‘ase, the source transmits x; n], the relay transmits x(n) ie, nothing, and the destination processes only (2). ‘Minimum probability of error (MPE) detection corresponds toconditional MPE detection for each value ofthe fading co- efficient a3. Since the input symbols are equally likely, con- ditional MPE detection corresponds to conditional maximum- likelihood (ML) detection; this can be implemented by the combiner in Fig. with any mapping f(-) and weights aver No” ‘kes the destination ignores y(n). Since the equivalent channel is conditionally Gaussian with SNR 7y,s, the conditional etror oprobability for singlehop transmission can be obtained from standard Gaussian results (2) Panne = (li=vna) © where Q(t) = he O° e~** Ids, and p isa constant depend ing upon the modulation format. For example, coberently- detected BPSK has p = —1, while coherently-detected FSK hhas p = 0. The average error performance of singlehop trans- tission, Pax, follows by averaging (5) over the exponential probability density function for 75; the esult can be approx {mated for large (average) SNR by (2) Poa ha>h © a Kis ‘where Kis another constant depending upon the modulation format. For example, coherently-detected BPSK has I = 4, ‘while coherently detected FSK has K B. Multihop Transmission ‘The basic wireless relaying protocol qualitatively described in Seetion Lis called multi-hop routing inthe ad-hoe network ing community 4. Multi-hop transmission in our framework can be viewed as cascading singlehop transmission between the source and relay with singlchop transmission between the relay and destination, Specifically, the source transmits xs, and the relay forms an estimate %[n] from (1) The relay trans- mits this estimate as xin} = 5 [71]. Finally, the destination forms an estimate [7 1 of x;[n—1] from (3), The sample delay accounts for processing and (relative) propagation delay through the relay. ‘As we will develop, ML detection of x(n) atthe relay is preferable. We examine two destination receivers for multihop transmission. The first forms ML estimates ofthe relays rans- mitted signal x(n, and is useful for developing average error performance bounds. The second makes ML estimates of the souree transmitted sequence x [7-1]. B.I ML Detection of x3[n} Conditional ML. detetion of x(n] comesponds tothe sngle- hhop ML detector from Section HI-A, with the rolls of y[—1] and y4(n] swapped. Specially, the conditional ML detector ‘an be implemented as the combiner in Fig. 2 with 0, + £0) I the relay decision process can be modeled as a binary symmetric channel (BSC) with crossover probability « depend- ing upon the SNR 7,2, the conditional error probability in es- timating [=I] at the combiner specified by (7) can be upper bounded by o AavB No Patttnn amas SF Poth ® ‘where the first tem arises from the event thatthe relay makes decision error, and the second term arises from the event that the destination makes a decision error given that the relay does ‘ot. The result (8) suggests that MPE detection of xi[n] atthe relay is preferable. In this case, ¢ = Psi, and the average error performance can be approximated for large SNR by [2] 1 Pa St ay Taharl ‘As we will see in Section IV, this bound is tight in several regimes of interest. B2 ML Detection of s1[n—1] Conditional ML detection of {7 ~ 1] atthe destination is Somewhat more involved, but can also be implemented a8 a combiner in the form of Fig. 2. Again assuming the relay d- cision process can be modeled asa BSC with crossover prob- ability, some algebra shows thatthe destination conditional ML detector of x,f0—1] has Sav |. co ‘The key step in obtaining (10) lies in the expanding the lkeli- hood plana x) by averaging over whether or not the relay makes a decision ero, i.e, (194 ln% = 20) teplyileaasi = 21), fore larly for = = 21. ‘The results (10) fl- lows after substitution ofthe conditional Gaussian likelihoods, taking the log-likelihood ratio, and algebraic simplifications. Limiting arguments indicat, and Fig. 3 exhibits, that the rapping /() im (10) essentially “lips” its input to the values lnje/(1 ~ 6] and is approximately near between these ex- teemes foe small 1. For ¢-< 1/2, the mappings in (7) and (10) satisty f(t) > 0 for t > 0 and f(t) < 0 fort < 0; hence, their symbol estimates for uncoded transmissions wil be iden. tical. Consequently, as we have seen previously, ML detection at the relay is preferable, andthe average ercor performance Of this multihop protocol should also be well approximated by (©). Finally, we observe that as ¢ > 1/2 the mapping f(t) in (0) goes to, andif the destination fies ¢ = 0in the detector, die, it does not explicil take into account the uncertainty of| the relay decisions, (10) reduces to (7). ‘Although apparently iclevant for wncoded multihop trans- ‘mission, we will seein Section II-C that the clipping property ‘of f()in(10) s signiticantin the context of diversity transmis- sion. More generally, while clipping the matched-iler output ‘may be irrelevant foruncoded transmissions, tis important for ‘coded systems (with symbol-by-symibol detection employed at the relay), because it imits the contribution of any one sym- bot’ log-likelihood, e.g. branch metic ina Viterbi algorithm, to the sum ofthe log-liketihoods, eg, path metric ina Viterbi algorithm, $0)ig. 3. Combiner mapping f(t) fom (10). Socesively higher dashed ares Gore > Ojeomespondio = 10°" 10", respecte) For ‘comparison thes curve coesponds othe linear mapping (0) = . Diversity Transmission with Decoding Relay (Our first diversity transmission protocol combines singlehop ‘and multihop transmission to ereate and exploit spatial diver- sity. Specifically, our protocol for diversity transmission with a decoding relay consists of the following. The source transmits {nl to both the relay and destination on the subchannel (1) ‘and (2). The relay forms an estimate of x(n] from its received ‘signal yo[n], and transmits this estimate, delayed by one sam- pile to account for processing and (relative) propagation delay, as xj{n] on subchannel (3). As throughout Section II-B, we assume the relay decision process can be modeled as a BSC ‘with crossover probability ¢, and based upon reasoning similar {0 that discussion, we employ ML deteston at the relay. ‘The destination estimates x(n — 1] from both its received signals (2) and (3). When suitably combined, the chances of| both signals exhibiting deep fading is reduced; therein les the diversity benefit. The challenge inthis setting isto design ade- tector that can overcome the effets of uncertainty inthe relay decisions and still exploit the available spati diversity CAL ML Detection of x:{2—1] Combining the results of Section I-A and Section II-B, conditional ML detection of x,[n—1] from both (2) and (3) can bbe implemented as the combiner in Fig.2 with Ha VEL os a et(l-oet +9, (uy Here the clipping effect of f(t) in (11) is more important than itwas for uncoded multihop transmission. The noalinearty in J(t) increasingly reduces, with increasing ¢, the contribution ofthe diversity branch through the relay. Ifthe destination assumes the relay decisions are always cor- rect, then (11) with ¢ ~» 0 becomes a conventional maximum goto A) =t (2 ‘This combiner, though mismatched in genera, performs rea sonaly wel for small. Similar to the bound in Section I 1, we can upper bound the average eror performance ofthe detector corespondng to (12), again using large SNR approx imation from (21, by TunTiaFas> L 3) "The first term in (13) arises from the event that the relay makes a decision error, and the second term arses from the event that the destination makes a decision error given thatthe relay does not, corresponding to a conventional transmit antenna diversity scenario [3]- This bound is only useful for approximating the performance of the ML detector (11) in channel environments for which 7, 2 is especially large, ¢g., when the relay is very close to the source, (C2. Maximum SNR Detector ‘Asan alternative design criterion, we determine the receiver ‘that maximizes the SNR of the slicer input. To arrive at this ‘max SNR receiver, we examine the relay decision 1 = m1 +e, where isa random variable capturing the effects of decision errors. A few calculations yield ifm itm, Elen) = {a == B= (1 — les — 20P? Letting % = my + Blefn), and @ = e ~ Bfely), the relay estimate 5 = 5 + & can be viewed as equally-likely symbols ‘drawn from the constant-magnitude constellation (oar tez0, (1 deo ters os plu an additive nose that is uncortelate with, having mean Zero and variance ¢2. Thus, the two signa received by the destination maybe written as, Yoln—1] = a9 VE miln—1) + zaln—1), ln] = a2 VEs (Hn—1] + @n—1)) + z5ln. Observing that (21 ~ #0) = (1 ~ 2)(21 ~ 20), the maximum 'SNR destination receiver, a matched-filter for (15), canbe it~ plemented a the combiner in Fig. 2 with Vill jess E02 +N" as 06)Examining w' in (16) more closely, we see that it consists of the maximum-ratio combiner weight win (12) followed by the linear mapping (-29 50) t an For ¢ 1/2, (17) goes to 2er0 indicating thatthe maximum SNR detector ignores the received signal y4 just as the ML, ‘detector defined by (11). Like the ML detector (16) converges to the maximum-raio combiner (12) for ¢ 0 ‘We conclude this section by noting that [5] develops results similar to (11) and (16) in the context of cellular networks. Specifically, the linear detector in [5] corresponds to the com= biner in Fig. 2 with HO =t the parameter ) is chosen numerically to minimize the condi- ‘ional error probability of ths linear detector. D. Diversity Transmission with Amplifying Relay In the previous section, we explored several destination de- algorithms assuming the relay employed ML detection. If we constrain the relay to employ linear processing, ie, am- plifying, alternative transmission protocols result, We might expect this constraint to induce excessive noise amplification, ‘but, as the simulation results in Section IV suggest, a desti- nation ML detector designed for an amplifying relay can be {quite competitive, and perhaps even outperform the tansmis- sion protocols from the previous section, when the relay is close to the destination. ur protocol for diversity transmission with an amplifying relay consists ofthe following. The source transmits x:[n] to both the relay and destination on the subchannel (1) and (2) ‘The relay transmits an amplified (and delayed) version ofits received sequence, Le, x{[n] = 8ya{n—I] on the subchannel {@). To decode symbol xj[n~1], the destination processes its ‘two received signals valn—1] = aramin—1] + 2[n-1], ln] = 0,38 (a1,2m[n—1] + za[n—I) + 20} ‘The destination conditional ML detector of x [n—1] from (18) ‘can be implemented as the combiner in Fig, 2 with wie fal ta VEL as) as 50 ‘To satisfy its output power constraint, the relay amplifier can ‘operate at a maximum gain satisfying & (aPE, + Ne" var 20) ‘here we allow the gain to depend upon the fading realization 41,2 from the source to the relay. Substituting (20) into (19), ‘we see thatthe channel is conditionally Gaussian with SNR that can be manipulated into the form 71,3 + Yeas Where an error probability can be readily computed using standard Gaussian results, yielding, Posraamanns=@(Y@— iat a])» 2 [Note thatthe conditional error probability (22) exhibits a sum of SNRs as we might expect ina diversity scenario. Examining (21), we see that na + @) Since sand 72,9 are independent exponential random vari- ables in our model, their minimum is also exponential with expected value satisfying ea < in # min (7.3, 723) Fain = Tig + Tas (24) analogous toa parallel combination of resistances in circuit theory. Since Qt) is decreasing int, (23) gives Porinamamne2@(Yi=aliarimel) — @9 Finally, averaging (25) over the exponential density functions for 7.3. we obtain a lower bound on the average error per- {ormance of divesity transmission with an amplifying relay. Using the large SNR approximations from [2], we obtain 3 Pou? a 26) Fig Tana >1 In addition to the lower bound provided by (26), we can esti- ‘mate the average eror performance Pa by computing sample averages of independent realizations of (22), or by Monte Carlo simulation ofthe system. TV. PERFORMANCE SIMULATIONS. To compare performance of the transmission protocols, we ‘examine a network with coordinates normalized by the dis- tance di. Between the source and destination radios. I these coordinates, the source canbe located at (0,0), andthe des nation canbe located at (1,0), without los of generality. Due {0 space considerations, we limit our scope to scenarios with the elay located at (,0} for Oe, the lay is very close to the source; = 1/2, Le, the relay is halfway between the source and destination; and / = 1, ie., the relay is very close tothe destination “The fading variances o2,, can be assigned wsing wireless path-loss models based on the network geometry [6]; here, weutilize models ofthe form o2,, cd, where is the dis- tance from mobile ito mobile, and » isa constant whose value, as estimated from field experiments, les in the range 3 (6) TheodeeS Rappapo, Wireless Communications: Principe and Prac tee, Pence Hale Upper ea River, New Jee, 1996 2 Bese Singeop verge SNR (4B) Fig. 4 Simulate perfomance ofthe transmission protocols for = 4 and ‘onmalized geome with the rely Toewed a (01,0), ce close o he r ae eel Ti Sinsleop Avenge SNR (8) Fe. 5. Simulated perfomance ofthe transmission protocols for = 4 aad ‘normalized geometries wih the ray located t (05,0), alfa Be ‘ween the source and destin, wt Average itor Pomanse Solenp Avenge SNR (a8) Fig. 6 Simulted performance ofthe transmission protocols for» = 4 and eemaledgeometies with there) lose at (09,0), Le clas 0 he ‘exinaion,