You are on page 1of 5

Distributed and Layered Codes for Relaying

Gerhard Kramer Bell Laboratories, Lucent Technologies 600 Mountain Ave, Murray Hill, NJ 07974 USA

Abstract Block-Markov coding and modulation methods for relay channels are presented that are motivated by coding methods for multi-input, multi-output (MIMO) channels. A partial decode-and-forward strategy is shown to achieve high rates. Using this strategy, protocols employing one, two, and three codes are discussed and compared. Low-density parity-check (LDPC) codes are designed and simulated for a protocol related to Diagonal Bell-Labs Layered Space-Time (D-BLAST).

II. D ECODE - AND -F ORWARD A memoryless relay channel [2] is dened by the conditional probability distribution PY2 Y3 |X1 X2 (a, b|c, d), (1) where a Y2 , b Y3 , c X1 , d X2 , Y2 and Y3 are the relay and destination channel outputs, respectively, and X1 and X2 are the source and relay channel inputs, respectively. Several coding strategies for relay channels are known including:

I. I NTRODUCTION Block-Markov superposition coding was developed for the multi-access channel (MAC) with feedback in [1]. The method has since been applied to many other multi-user channels with noisy or noise-free feedback, e.g., the relay channel [2], the MAC with generalized feedback [3, 4], the multi-access relay channel [5], the broadcast relay channel [6], to name a few. We here address similarities between Block-Markov coding for relay channels and diagonally layered protocols for multi-input, multi-output (MIMO) channels, i.e., Diagonal Bell-Labs Layered Space-Time (D-BLAST) [7]. There is a direct relation between MIMO communication and relaying: consider a MIMO channel and consider the rst MIMO channel input as the input of a source node, and the remaining MIMO channel inputs as inputs of relays that happen to be colocated with the source node. One nds that DBLAST encoding is precisely a Block-Markov superposition coding scheme for full-duplex relays. We use this insight to adapt coding strategies for MIMO communication to relay channels. For example, we suggest coding protocols for distributed bit-interleaved coded modulation (BICM), distributed D-BLAST, and distributed Vertical-BLAST (V-BLAST). We remark that some of the methods we consider have been studied by other authors. The variation across papers involves different channel models (no/ergodic/quasistatic fading), different channel state information (CSI) scenarios (CSI at the transmitter and/or receiver, no CSI), different device models (perdevice/network power constraints, full/half-duplex), and different codes (convolutional [8,9], space-time [10], turbo [11], lowdensity parity-check (LDPC) codes [12, 13]). Our aim here is to categorize several MIMO coded modulation methods, and to study how one might adapt these for relay communication. This paper is organized as follows. Sec. II describes several coding strategies, Sec. III describes channel models, and Sec. IV describes device models. Sec. V presents protocol designs for relaying that are motivated by MIMO methods. We further design LDPC codes with extrinsic information transfer (EXIT) charts and simulate their performance.

Amplify-and-forward (AF): the relay amplies the most recent Y2 , i.e., each relay channel input is some function of a past channel output. More generally, the relay transmits some function of a small number of past Y2 . Decode-and-forward (DF): the relay decodes the source message, re-encodes it, and transmits the resulting codeword. Compress-and-forward (CF): the relay quantizes, compresses, and channel encodes a string of Y2 and transmits the resulting values digitally to the sink. Variations and combinations of the above. For example, a partial decode-and-forward (PDF) strategy has the source split the message into two parts, use superposition encoding to transmit these two parts, and has the relay decode only one of the two parts (see [2, Thm. 7]).

We will here consider only DF strategies and their variations, of which there are several types (see [14, Sec. I.A]). For example, Carleials regular encoding/sliding window decoding DF strategy achieves the rate R = max min {I (X1 ; Y2 |X2 ), I (X1 X2 ; Y3 )} .
PX1 X2

(2)

As explained above, the block Markov superposition encoding scheme used to achieve (2) has a diagonally layered structure that is basically the same as D-BLAST encoding. However, for half-duplex devices we improve rates by using Carleials PDF strategy [3] that achieves R = max min{I (U ; Y2 |X2 ) + I (X1 ; Y3 |U X2 ),
PU X1 X2

I (X1 X2 ; Y3 )}

(3)

where U [X1 , X2 ] [Y2 , Y3 ] forms a Markov chain. Observe that in (3) we have I (X1 X2 ; Y3 ) = I (U X2 ; Y2 ) + I (X1 ; Y3 |U X2 ). (4)

The PDF rate (3) is thus the sum of a DF rate (2) with U replacing X1 and a single-hop rate rate I (X1 ; Y3 |U X2 ).

1424401321/05/$20.002005IEEE

1752

1 d Source Relay Destination

Fig. 1. A linear geometry.

powers. Wireless devices usually operate under a half-duplex constraint that one can model by replacing (5) with (see [14] and [16]) 12 , if X 2 = 0 H12 , H /2 X 1 + Z 2 d12 (10) Y2 = 0, if X 2 = 0. Alternatively, we introduce a mode M2 that takes on the values L and T for listen and talk, respectively. (The source is assumed to always talk and the destination to always listen.) This mode can be considered to be part of the relays channel input so that (3) becomes R=
P U X 1 X 2 M2

III. C HANNEL M ODEL We consider additive white Gaussian noise (AWGN) channels with Rayleigh fading and CSI at the receivers only. This scenario seems to capture the fundamental issues concerning coding and modulation while avoiding other (important) issues such as training, non-coherence, or water-lling optimization. The channel is dened by (see [14, 15]) Y2 = Y3 = H12 , H12
/2 d12

max

min{

X1 + Z2 H13
/2 d13

(5) H23
/2 d23

I (U ; Y 2 |X 2 M2 ) + I (X 1 ; Y 3 |U X 2 M2 ), I (X 1 X 2 ; Y 3 |M2 ) + I (M2 ; Y 3 ) }

(11)

H13 , H23 ,

X1 +

X2 + Z3

(6)

where X t , t = 1, 2, and Y t and Z t , t = 2, 3, are complex column vectors of length nt , Hst is a complex nt ns fading matrix, dst is the distance between nodes s and t, and is an attenuation exponent (e.g., = 2 for free space propagation). The Z t have statistically independent, proper, complex, Gaussian, zero-mean, unit variance entries and are statistically independent of each other and all the X t and Hst . We further suppose that Hst is statistically independent of X t , t = 1, 2, Z t , t = 2, 3, and all other fading matrices. Rayleigh fading has Hst that have statistically independent, proper, complex, Gaussian, zero-mean, unit variance entries. We will consider the linear geometry depicted in Fig. 1 where d13 = 1. We further consider primarily d12 1. IV. D EVICE M ODELS Let X ti be the channel input of device (or node) t at time i. The transmitting nodes often have per device and block power constraints
n i=1

where U is a column vector of length n1 . Note that if M2 is known ahead of time by the destination then one loses the gain I (M2 ; Y 3 ) above (see [16]). On the other hand, this gain might be difcult to realize because the relay must switch rapidly between M2 = L and M2 = T . For simplicity, we will ignore this gain here (see also [17, 18]). Let V be a column vector of length n1 and let I be an appropriately sized identity matrix. We choose U , V , and X 2 to be statistically independent, proper, complex, Gaussian, zero-mean, and having covariance matrices (M2 )P I , (1 (M2 ))P1 I , and P2 I , respectively, where 0 (M2 ) 1 (note that (9) prevents using power control across modes). We further choose X 1 = U + V . The resulting expressions in (11) with the model dened by (6) and (10) are I (U ; Y 2 |X 2 , M2 = L) =
h 1

p(h) log I +

(1 (L))P1 P1 hh I + hh d d 12 12

dh (12)

I (X 1 ; Y 3 |U X 2 , M2 = m2 ) = p(h) log I +
h

E[ X ti ]/n Pt ,

t = 1, 2

(7)

(1 (m2 ))P1 hh dh d 13 P1 hh dh d 13 P1 P2 hh + h h dh dh d d 13 23

(13)

I (X 1 X 2 ; Y 3 |M2 = L) =
h

where X 2 = X X and X is the complex-conjugate transpose of X . Alternatively, one might use the network constraint
n

p(h) log 1 +

(14)

I (X 1 X 2 ; Y 3 |M2 = T ) =
h,h

E[ X 1i
i=1

+ X 2i ]/n P1 + P2

(8)

) log I + p(h)p(h

(15)

or, perhaps, the symbol constraints E[ X ti ] Pt ,


2

t = 1, 2,

i = 1, 2, . . . , n.

(9)

We consider only (9) to avoid power control optimization. The model dened by (5) and (6) lets the relay transmit and receive at the same time in the same frequency band. This is often not possible due to large differences in transmit and receive

) are Gaussian fading distributions (h where the p(h) and p(h are matrices in general). Note that for d12 d13 it is and h best to choose (L) = 1 and (T ) = 0. Moreover, this distribution is basically the same as using the strategy depicted in Fig. 2 where X 1 has the same distribution irrespective of M2 . It therefore remains to optimize PM2 (). In fact, we shall avoid this optimization and consider only PM2 (L) = PM2 (T ) = 1/2.

1753

Block 1 Source Relay

Block 2

Block 3

Block 4

2 1.8 1.6 1x1x1 channel, 1/2duplex Rayleigh fading, =4 QPSK, E /N =3 dB


s o

x1(w1) 0

x1(w2) x2(w1)

x1(w3) 0

x1(w4) x2(w3)

PDF (relay talks 1/2 of time)

Fig. 2. A PDF strategy for half-duplex relays.

1.4

Rate [bits/use]

1.2 1 0.8 0.6 0.4 0.2 0 1 0.75 0.5 0.25 0 0.25 0.5 0.75 1 multihopping (optimum duplexing) relay off

V. MIMO C ODING Consider a MIMO channel with M transmit and N receive antennas. We compare the MIMO methods listed in Table I (APP refers to a posteriori probability). Direct Mapping (see [19]): use one code with rate R c and length nc and map the coded bits directly onto the modulation signal set. For instance, for quaternary phase-shift keying (QPSK) parse the coded bits into blocks of length 2M and map these blocks onto an M -antenna QPSK symbol using Gray mappings. BICM (see [20]): use one code with rate Rc and length nc , interleave the coded bits, and then map them onto the modulation signal set as above. BICM with Inner Space-Time Codes (see [21]): use one code with rate Rc and length nc , interleave the coded bits, and then map them onto a space-time code. D-BLAST (see [7]): use one code with rate Rc and length nc and map nc /M of the coded bits onto the rst antenna symbol in a rst block, another nc /M of these bits onto a second antenna symbol in a second block, and so forth until the M th block. The symbols corresponding to the entire codeword of length nc are called a layer. Perform similar steps with other codewords, but successively shift the mappings by one block for every codeword. V-BLAST (see [7]): use M codes with rates Rc (m), m = 1, 2, . . . , M , and lengths nc /M . The symbols corresponding to each codeword of length nc /M are called a layer. Map the coded bits of codeword m onto antenna m, m = 1, 2, . . . , M . V-BLAST encoding is basically the same as multi-level coding or generalized concatenated coding. VI. B LOCK -M ARKOV C ODING We adapt the above MIMO methods for the PDF protocol depicted in Fig. 2. We refer to the relay channel as n1 n2 n3 based on the number of device antennas. We consider two cases with QPSK modulation. A 1 1 1 setup with P1 = P2 = 2 (or Es /N0 = 3dB ). The PDF rates are shown in Fig. 3 as a function of d. Also shown are the no-relay rate (R 1.13 bits per use) and the traditional multihopping rates with optimized listen and transmit times. Observe that PDF achieves substantial rate gains over both no-relay transmission and traditional multihopping. For instance, the points marked with in Fig. 3 are (d, R) = (0.25, 1.0) and (d, R) = (0.25, 1.5). Note that the multihopping curve is well below the relay off curve, and that the PDF curve is at near d = 0.25. This happens because the source-to-relay link capacity is

Distance d
Fig. 3. PDF rates for a 1 1 1 setup.
1.4 1x1x2 channel, 1/2duplex Rayleigh fading, =4 QPSK, Es/No=6 dB PDF (relay talks 1/2 of time)

1.2

Rate [bits/use]

0.8

0.6 relay off 0.4 multihopping (optimum duplexing)

0.2

0 1

0.75

0.5

0.25

0.25

0.5

0.75

Distance d
Fig. 4. PDF rates for a 1 1 2 setup.

almost saturated at the maximum QPSK rate of 2 bits per use. One should therefore use a larger modulation signal set, e.g. 8-PSK, for the odd-numbered blocks in Fig. 2. A 1 1 2 setup with P1 = P2 = 0.25 (or Es /N0 = 6dB ). The PDF rates are shown in Fig. 4. The gure also shows the no-relay rate (R 0.54 bits per use) and the traditional multihopping rates with optimized listen and transmit times. The points marked with in Fig. 4 are (d, R) = (0.25, 0.5) and (d, R) = (0.25, 1).

A. Direct Mapping and BICM with One Code Consider using one encoder with rate Rc that generates one codeword for every pair of blocks. We map the coded bits onto the modulation signal sets at the source and relay, either without or with bit interleaving. The relay decodes the message bits after having received the rst block of outputs from the

1754

TABLE I
C OMPARISON OF MIMO C ODED M ODULATION M ETHODS

Strategy Direct Mapping BICM BICM with Inner Space-Time Codes D-BLAST V-BLAST

Advantages use one code achieve ergodic & quasistatic inform. rates use one code designed for AWGN channels use one code designed for AWGN channels few detector updates (at EXIT) use one code designed for AWGN channels achieve ergodic & quasistatic inform. rates few soft detector updates (at EXIT) use codes designed for AWGN channels achieve ergodic inform. rates few soft detector updates (at EXIT)

Disadvantages custom code designs required many APP detector updates (steep EXIT) lose information rates many APP detector updates (steep EXIT) lose information rates (unless orthogonal) complex APP detection (unless orthogonal) interference cancellation increased delay (reduced reliability) error propagation lose quasistatic inform. rates (many codes) interference cancellation increased delay (reduced reliability)

source. This approach gives good reliability on the combined source/relay-to-destination link. However, early decoding at the relay severely restricts the code rate Rc . For example, if n1 = n2 and both the source and relay use QPSK, then we require Rc < 1/3 because the relay has seen only 1/3 of the potential received symbols at the time of decoding. One can try to x this problem by adjusting the modulation signal sets and the amount time that the relay listens and talks. However, there is a lack of exibility. A second criticism is that a strategy with one code is a pure DF scheme where the relay decodes all of the message bits. Such an approach is generally suboptimal for half-duplex channels. We will not pursue it further here. B. V-BLAST An adaptation of V-BLAST encoding to half-duplex relaying is to use three different codes: one for each of the three boxes in the rst two blocks in Fig. 2. For example, consider the 1 1 1 rate point (d, R) = (0.25, 1.5) in Fig. 4. We can achieve this point by using the PDF strategy in Fig. 2 with three rate 1/2 codes. However, we must decode the messages wb , b odd, by using the combined graph of both codes that carry this message. In fact, by using a common rate for these two codes and decoding them jointly we are using precisely Carleials regular encoding/sliding window decoding strategy (see [3, 14]). One can, of course, also choose different rates for these codes without any conceptual changes to the protocol or theory. We have not implemented this approach here. We instead focus on an approach inspired by D-BLAST. C. D-BLAST An adaptation of D-BLAST encoding to half-duplex relaying is to use two different codes: one for each of the two messages in the rst two blocks in Fig. 2. The relay must decode w1 after having received only the rst block of outputs from the source. We therefore require Rc < 1/2 for the w1 encoder (assuming that n1 = n2 and the source and relay use the same signal set). This constraint can be serious but it is somewhat less stringent than in Sec. VI-A because only part of the source message must

be decoded at the relay. Note also that the D-BLAST approach to half-duplex relaying does not suffer from error propagation, as opposed to the full-duplex case. Code design is usually done by using density evolution [22] or EXIT charts [23]. We use the latter approach and design irregular low-density parity-check (LDPC) codes with the curvetting procedure described in [19]. The coded bits are mapped to QPSK symbols via the Gray mapping. The decoder uses the standard graph representation of an LDPC code with variable nodes on the left and check nodes on the right. The left and right nodes are connected by edges whose nodes are chosen with a random permutation that avoids 2-cycles. The decoder iterates 60 times between the left and right nodes by using an a posteriori probability (APP) decoder. Consider the 1 1 2 setup of Fig. 4 and d = 0.25 (this scenario is also considered in [13]). Consider R = 1/2 without a relay. We design an LDPC code with rate Rc = 1/4 and length nc = 8, 000 that has an (single-antenna, no fading, BPSK) AWGN decoding threshold of Eb /N0 = 0.4 dB which is about 0.3 dB from capacity. The resulting frame error rates (FER) are shown on the right in Fig. 5. Observe that the code operates within 1.5 dB of capacity at an FER of 103 . The extra loss (as compared to 0.3 dB for the single-antenna case) can be attributed to the short code length and the fading. Consider next R = 1. We design an LDPC code with rate Rc = 3/8 and length nc = 16, 000 that has an (single-antenna, no fading, BPSK) AWGN decoding threshold of Eb /N0 = 0.1 dB which is about 0.45 dB from capacity. The encoding and decoding procedure is as follows (see Figure 2 and [13]). In the odd-numbered blocks b = 1, 3, 5, . . . , the source transmits 4000 QPSK symbols (or 8,000 of the 16,000 codeword bits) by using the rate Rc = 3/8 LDPC code. After every odd-numbered block b, the relay decodes the information bits of the Rc = 3/8 code from this block. Note that the relay has received only half of this codewords symbols. In the even-numbered blocks b = 2, 4, 6, . . . , the source transmits using the rate Rc = 1/4 code described above. In the even-numbered blocks, the relay encodes the information bits decoded from the previous block by using the

1755

10

Frame Error Rate (FER)

10

2x2 distr. DBLAST R=3/4 (Rc=3/8) n =16,000 c 1 detector activation relay decodes early

1x2 no relay R=1/2 (Rc=1/4) nc=8,000

10

d=0.25, =4 LDPC codes random edge perm. 60 iterations

R=1/2 Capacity 10
3

R=3/4 Capacity

6.5

5.5

4.5

SignaltoNoise Ratio ES/NO [dB]


Fig. 5. PDF frame error rates for the 1 1 2 scenario.

Rc = 3/8 encoder and transmits the last 4000 QPSK symbols of this codeword (or the last 8,000 of the 16,000 codeword bits). After every even-numbered block, the sink decodes the information bits of the rate Rc = 3/8 code. The sink performs only one detector activation per codeword (we remark that multiple detector activations would improve the performance a little [19]). The sink cancels the interference caused by the symbols of the Rc = 3/8 code from the even-numbered blocks. After every even-numbered block, the sink decodes the information bits of the Rc = 1/4 code. The overall rate is R = 2(3/8) + 2(1/4)(1/2) = 1 bit per use, where the leading factors 2 are due to the QPSK modulation. There are three decoding steps to consider. The FER of the relay decoding step is not shown in Figure 5 because it lies far to the left of the other two curves. The FER of the sink decoding the information bits from the Rc = 3/8 code is shown as the left curve in Figure 5 (labeled 2 2 distr. D-BLAST). The FER of the sink decoding the information bits from the Rc = 1/4 code is the same as the case where there is no relay, and is the curve on the right in Figure 5. We see that the dominating FER is in both cases (without and with a relay) due to the direct link from the source to the sink. The reliability of the two schemes is therefore the same. However, the PDF scheme doubles the rate. R EFERENCES
[1] T. M. Cover and C. S. K. Leung, An achievable rate region for the multiple-access channel with feedback, IEEE Trans. Inform. Theory, vol. 27, no. 3, pp. 292298, May 1981. [2] T. M. Cover and A. A. El Gamal, Capacity theorems for the relay channel, IEEE Trans. Inf. Theory, vol. 25, pp. 572584, Sept. 1979.

[3] A. B. Carleial, Multiple-access channels with different generalized feedback signals, IEEE Trans. Inf. Theory, vol. 28, no. 6, pp. 841850, Nov. 1982. [4] F. M. J. Willems, Informationtheoretical Results for the Discrete Memoryless Multiple Access Channel. Doctor in de Wetenschappen Proefschrift, Katholieke Universiteit Leuven, Leuven, Belgium, Oct. 1982. [5] L. Sankaranarayanan, G. Kramer, and N. B. Mandayam, Capacity theorems for the multiple-access relay channel, Proc. 42nd Annu. Allerton Conf. Commun., Control, Computing, Monticello, IL, Sep./Oct. 2004, pp. 1782-1791. [6] Y. Liang, Multiuser Communications with Relaying and User Cooperation. Ph.D. thesis, Univ. Illinois at Urbana-Champaign, Urbana, IL, Aug. 2005. [7] G. J. Foschini, Layered spacetime architecture for wireless communication in a fading environment when using multi element antennas, Bell Labs. Tech. J., vol. 1, no. 2, pp. 4159, 1996. [8] T. E. Hunter and A. Nosratinia, Cooperation diversity through coding, Proc. IEEE Int. Symp. Inf. Theory, Lausanne, Switzerland, p. 220, June 30-July 5, 2002. [9] B. Zhao and M. C. Valenti, Distributed turbo coded diversity for relay channel, Electron. Lett., vol. 39, no. 10, pp. 786787, May 2003. [10] J. N. Laneman, Cooperative Diversity in Wireless Networks: Algorithms and Architectures. Ph.D. Thesis, Massachusetts Institute of Technology, Cambridge, MA, Aug. 2002. [11] Z. Zhang and T. M. Duman, Capacity-approaching turbo coding and iterative decoding for relay channels, IEEE Trans. Commun., vol. 53, no. 11, pp. 1895-1905, Nov. 2005. [12] M. A. Khojastepour, N. Ahmed, and B. Aazhang, Code design for the relay channel and factor graph decoding, Proc. 38th Annu. Conf. Signals, Systems, Computers, Asilomar, Pacic Grove, CA, pp. 2000-2004, Nov. 7-10, 2004. [13] G. Kramer, Communication strategies and coding for relaying, Proc. IMA Summer Program for Wireless Communication, Univ. Minnesota, Minneapolis, MN, June 22-July 1, 2005, unpublished. [14] G. Kramer, M. Gastpar, and P. Gupta, Cooperative strategies and capacity theorems for relay networks, IEEE Trans. Inf. Theory, vol. 51, no. 9, pp. 3037-3063, Sept. 2005. [15] B. Wang, J. Zhang, and A. Hst-Madsen, On the capacity of MIMO relay channels, IEEE Trans. Inf. Theory, vol. 51, no. 1, pp. 29-43, Jan. 2005. [16] G. Kramer, Models and theory for relay channels with receive constraints, Proc. 42nd Annu. Allerton Conf. Commun., Control, Computing, Monticello, IL, Sep./Oct. 2004, pp. 1312-1321. [17] A. Hst-Madsen, On the capacity of wireless relaying, Proc. IEEE Vehic. Techn. Conf., VTC 2002 Fall, Vancouver, BC, vol. 3, pp. 13331337, Sept. 2428, 2002. [18] M. A. Khojastepour, A. Sabharwal, and B. Aazhang, On the capacity of cheap relay networks, Proc. 37th Annu. Conf. Inform. Sci. Sys. (CISS), Baltimore, MD, March 1214, 2003. [19] S. ten Brink, G. Kramer, and A. Ashikhmin Design of lowdensity parity-check codes for modulation and detection, IEEE Trans. Commun., vol. 52, no. 4, pp. 670-678, April 2004. [20] G. Caire, G. Taricco, E. Biglieri, Bit-interleaved coded modulation, IEEE Trans. Inf. Theory, vol. 44, no. 3, pp. 927-946, May 1998. [21] J. Hou, P. H. Siegel, and L. B. Milstein, Design of multi-input, multi-output systems based on low-density parity-check codes, IEEE Trans. Commun., vol. 53, no. 4, pp. 601-611, April 2005. [22] T. J. Richardson, A. Shokrollahi, and R. L. Urbanke, Design of capacity-approaching low-density parity-check codes, IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. 619637, Feb. 2001. [23] S. ten Brink, Convergence of iterative decoding, Electron. Lett., vol. 35, no. 10, pp. 806808, May 1999.

1756

You might also like