Linear Scalable Dispersion Codes for Frequency Selective Channels

Marc Kuhn, Ingmar Hammerstroem, and Armin Wittneben

Swiss Federal Institute of Technology (ETH) Zurich
Communication Technology Laboratory, CH-8092 Zurich, Switzerland
Email:{hammerstroem, kuhn, wittneben}@nari.ee.ethz.ch

Abstract— We investigate a combination of linear scalable 1

space-time block codes with OFDM and present a systematic s(1)
Inter− S/P OFDM
code design. This approach leads to space-frequency coded LSD Code leaver Converter Modulator
MIMO-OFDM with a variable tradeoff between improving
link reliability (by using diversity techniques) and increas- Data
ing data rate (using spatial multiplexing). The proposed cod- M
Encoder
ing scheme achieves spatial, temporal and frequency diversity Inter− S/P OFDM
and is able to utilize the potential of rich arrays (antenna ar- leaver Converter Modulator
s(M)
rays with a large number of antennas) with reasonable (and
scalable) decoding complexity. Fig. 1. System block diagram of a space-frequency coded OFDM
transmitter
I. I NTRODUCTION
In this paper we present a space-frequency signal-
Future wireless communication systems will have ing scheme for frequency selective channels. The
to support high data rates at a high link reliability. scheme is based on the class of high rate space-time
In a rich scattering environment this can be done us- codes presented in [6], [5]. The resulting space-
ing Multiple Input Multiple Output (MIMO) signal- frequency coded MIMO-OFDM exploits frequency
ing techniques. There are two basic coding methods, diversity, is highly adaptive and scalable and able to
namely space-time coding to improve link reliabil- trade spatial multiplexing gain for diversity gain.
ity and spatial multiplexing to increase spectral effi-
ciency. II. C HANNEL M ODEL
Space-time codes [1] combat the fading effects of
wireless multipath channels by utilizing the diversity We will consider a multi-antenna system with M
of the communication channel given for example by Tx antennas and N Rx antennas. The following chan-
the use of an antenna array at the transmitter (Tx) nel model for a time-varying wideband MIMO chan-
and/or at the receiver (Rx). Such methods aim to nel is used:
XL
achieve transmit diversity without channel state infor-
mation (CSI) at the Tx. H(τ ) = Hi · δ(τ − τi ) (1)
i=1
Due to the use of multiple antennas at the Tx and
the Rx there is an increase in channel capacity. In an The MIMO channel is modelled as the sum of sev-
uncorrelated Rayleigh fading environment the capac- eral delayed random fading components (tapped de-
ity grows linearly as the number of Tx and Rx anten- lay line model). H(τ ) is a complex (N × M ) matrix.
nas grow simultaneously [2]. This increase in capac- In this paper we assume that Hi with i ∈ [1, 2, .., L]
ity is possible without an increase of the bandwidth. are (N × M ) Rayleigh fading matrices; i.e. the el-
Spatial multiplexing exploits this offered capacity by ements of these matrices are uncorrelated circularly
breaking up the data stream into parallel substreams symmetric complex Gaussian random variables with
which are then transmitted simultaneously on individ- zero mean and variance σi2 (path gains σi2 are given
ual antennas [3]. It exists a tradeoff between spatial by the power delay profile of the channel). Different
multiplexing and transmit diversity [4], [5]. taps are assumed to be uncorrelated.
Because future wireless communication systems
III. C ODING S CHEME
(e.g. next generation wireless LANs) will have to
support high data rates the required system bandwidth A coded OFDM communication system with NS
will increase. Wideband channels provide an extra de- OFDM subcarriers (OFDM tones), M Tx and N Rx
gree of diversity (frequency diversity), which can be antennas is considered. Fig. 1 shows the transmitter
exploited through the use of Orthogonal Frequency of such a system (complex baseband representation).
Division Multiplexing (OFDM) with an appropriate The data encoding is based on the linear scalable dis-
coding across the OFDM tones. persion (LSD) codes presented in [6], [5].
A. Short review of LSD space-time coding Outer Code Inner Code
a b bν sν
Originally the high rate LSD space-time coding Reshape
R Cν
scheme is designed for flat fading channels [6]; it con- (NC x 1) (NU x NL)
NI NC NU M
sists of two concatenated but decoupled linear block
codes: the time variant inner code and the time invari- sν [1] 1
Interl., S/P Conv.
ant outer code. The outer code is optimized for di- OFDM Mod.
sν 1
versity performance in fast fading channels. The time
variant inner code is optimized with respect to the out-
sν [M] M
age channel capacity available for the outer code; due M
Interl., S/P Conv.
to the time variance the inner code transforms a block OFDM Mod.
1
fading channel in a fast fading channel to enable high
diversity gains of the outer code. In [6] efficient code Fig. 2. Symbol discrete model of the LSD encoding scheme
matrices are presented. No a priori channel knowl-
edge is required at the Tx. Fig. 2 shows the combination of the LSD codes
The possible tradeoff between diversity and spatial with OFDM and multiple Tx antennas. As in the case
multiplexing is demonstrated in [5] for Rayleigh and of the space-time codes the LSD coding scheme con-
Ricean flat fading; the more spatial subchannels are sists of two concatenated but decoupled linear block
used (for a given number of Tx and Rx antennas) the codes: the outer code R and the inner code Cν ; the
higher the rate, but the symbol error probability in- inner code is frequency variant (a different Matrix Cν
creases too, because the diversity gain decreases. So, for each OFDM subcarrier ν) instead of time variant
using LSD codes it is possible to trade diversity gain in the space-time coding case. In this paper we as-
(link reliability) for data rate in a very flexible way. sume that the Tx has no a priori CSI.
The input symbol vector a, consisting of NI in-
formation symbols, is multiplied with the (NC × NI )
B. Space-frequency coded MIMO-OFDM outer code matrix R to form the transmit symbol vec-
In this work we derive a MIMO signaling scheme tor b. The dimensions of the code matrix determine
NI
for frequency selective channels. Because the LSD the code rate to rC = N C
. Note that a simple modifi-
codes are originally designed as space-time codes for cation of the code rate is possible by adding/deleting
flat fading channels, OFDM is used to convert the fre- columns of the outer code matrix. Because of the in-
quency selective channel into a set of NS independent terdependence of code rate and required decoder com-
frequency flat subchannels. OFDM is realized by an plexity this feature is highly useful in networks with
IFFT; the OFDM modulators shown in Fig. 1 include heterogeneous node capabilities [7]. Throughout this
a cyclic extension of the OFDM symbol in a guard in- paper code rate rC = 1 is used. Thereafter the trans-
terval. The channel is is considered constant over one mit symbol vector b of dimension (NC × 1) is re-
OFDM symbol. shaped into a (NU × NL ) matrix. The columns of this
The symbols transmitted on the ν-th OFDM sub- matrix are the consecutive NL = N NU input vectors bν
C

carrier are given by the column vector sν = of the linear inner code Cν ; a different Cν is used
(sν [1], sν [2], ..., sν [M ])T ; the complex scalar sν [i] de- for each OFDM subcarrier. Coding across the tones
notes the symbol transmitted by the i-th antenna. In is required to exploit the inherent frequency diversity
one timestep the antenna i transmits one OFDM sym- of the frequency selective channel. We first consider
bol that is derived by frequency interleaving, serial- the situation that the outer coding is done across all
parallel conversion and IFFT of the vector s(i) = OFDM subcarriers, i.e. NL = NS .
(sν=1 [i], s2 [i], ..., sNS [i])T (Fig. 1). The M elements of the vector sν are transmitted
The (N × 1) receive vector rν of subcarrier ν fol- by the M OFDM modulators in Fig. 2. The system
lows to transmits NU ≤ rank (H) symbols over one OFDM
√ subcarrier in one timestep. We refer to NU as the
rν = SNR · Hν · sν + wν (2)
number of spatial subchannels to be used for spatial
where the (N × M ) matrix Hν is the discrete Fourier multiplexing. The remaining spatial dimensions can
transform of the MIMO channel H at the frequency be used by the code achieving an additional diversity
ν, SNR is the average signal-to-noise-ratio at each Rx gain.
antenna and wν is complex-valued AWGN at the ν-th Due to the code concatenation we have the possi-
subcarrier. bility to optimize two decoupled linear block codes.
 1) Coding for pure transmit diversity (MISO case): At the frequency ν the M elements of the vector sν are
In the MISO case the receiver uses only one Rx an- transmitted over M antennas; due to this inner code cν
tenna, N = 1. The channel matrix Hν becomes a row every element is a phase-shifted variation of the scalar
vector hνH with M elements. For only one Rx an- bν .
tenna no spatial multiplexing is possible, i.e. NU = 1 b) Design of the outer code: A suitable cost
and NL = NC . The matrix Cν and the vector bν be- function for the outer code is the fading averaged pair-
come a column vector cν and a scalar bν , respectively. wise probability (PEP) of message error. The PEP of
The vectors cν (ν ∈ [1, 2, ..., NL ]) are of dimension two different input symbol vectors a(1) and a(2) is the
(M × 1) and are used to drive the Tx antenna array probability that a(1) is transmitted and the Rx decides
(normalized to maintain the total transmit power). erroneously in favor of a(2) :
After linear signal processing of the received signal
vector of subcarrier ν - given by (2) - follows the esti- P [a(1) → a(2) ] = P [â = â(2) |a(1) ] (7)
mation of bν to (superscript H : conjugate transpose): The inner code and the MISO channel (i.e. the re-
√
b̂ν = SN R · cH H shaped equivalent SISO channel zν ) are modelled as a
ν · hν · hν · cν · bν + wν (3)
Rayleigh fast fading channel x with variance σx2 ; the
The scalar equivalent fading coefficients are matrix Dx = diag(x) is a diagonal matrix of dimen-
zν = cH H sion (NC ×NC ) with the elements of vector x on the
ν · hν · hν · cν (4)
main diagonal:
Without CSI at the Tx the inner code vector is inde-
pendent of the current realization of the channel vec- r = Dx R · a + w = Dx · b + w (8)
tor. If the elements of hν are random variables, zν
with b = R · a (Fig. 2).
is a random variable for any ν, when viewed over all For a given channel realization Dx the PEP is given
channel realizations. If the channel is not frequency by [8]
selective (hν = h), a frequency invariant inner code s 
cν = c does not allow to exploit a Tx diversity gain; d2ED 
(1)
P [â = â (2)
|a , Dx ] = P [b(1)
x → b(2)
x |Dx ] = Q 2
since in this case follows from (4) that all NC transmit 2σw
symbols bν of one OFDM symbol are affected by the (9)
with σw2 is the variance of AWGN w, d
same fading coefficient z. Therefore no Tx diversity ED = k∆bx k
(1)
gain can be achieved. But if cν is frequency variant the Euclidian distance of bx = Dx R · a(1) and
(2)
(e.g. two different subcarrier use two different Tx an- bx = Dx R · a(2) . All pairs (i, j) of input symbol
tennas), different transmit symbols bν are affected by vectors leading to the same ∆bx show the same PEP.
2
different fading variables. A diversity gain due to the Using the relation Q(x) ≤ 12 · e−x /2 we get an upper
Tx antenna array is possible (even if the channel is bound 2
not frequency selective); but it requires an appropriate 1 − d4σED 2
P [∆b, Dx ] ≤ · e w (10)
outer code that has to be optimized for diversity gain 2
in fast fading. One element ∆bx [i] of the vector ∆bx has the vari-
An equivalent situation was found for the optimiza- ance λi = σx2 · |∆b[i]|2 , the corresponding eigenvalue
tion of the space-time LSD codes for flat fading in [6]. of the matrix Λbx bx . Averaging over the Gaussian dis-
As shown there we can optimize the inner code with tributed random realizations of ∆bx we get an upper
respect to the outage capacity of the channel available bound for the fading averaged PEP for a given ∆b:
for the outer code. An appropriate objective function " #
NC
for the outer code is the fading averaged pairwise error 1 Y 1
probability (PEP). P [∆b] ≤ · λi
(11)
2 +1
i=1 4·σ 2 w
a) Design of the inner code: According to [6]
for Rayleigh fading channels h a good set of inner Clearly a repetition code of rate N1C minimizes
coding vectors is given by the first M columns of the the objective function, the fading averaged PEP, it
(NL ×NL ) discrete Fourier matrix F: achieves full diversity in fast fading; but the spectral
cν = F[ν, 1 : M ]T (5) efficiency is low. Therefore spectral efficiency is a
reasonable constraint for the optimization: a high di-
with versity gain should be reached at code rate rC = 1.
1 To require R to be unitary is another sensible con-
F[k, l] = √ · e−j·2π(l−1)(k−1)/NL . (6)
NL straint. So the Euclidean distance and thus the error
performance on an AWGN channel is maintained; this 0
10
4−QAM: [NC, NU, N, M] = [32, 1, 1, 2]

allows a low complexity decoder under AWGN con-

ditions (and in Ricean fading environments with large −1
10
Ricean K-factor).
L=
In [6] the choice to meet these constraints is a com- −2
1
10
plex and unitary coding matrix based on the impulse

SER
response of a cyclic chirp filter. The columns are −3 AWGN 5
2

found by cyclic shifting of the first column R[:, 1], 10

that is given by the Fourier transform of the vector m

−4 15
10
j·2πc·(i−1)2 /NC
2
m[i] = e . (12) 32

The optimal value of the parameter c is found as 3 4 5 6 7 8

E / N [dB]
9 10 11 12
b 0
follows: For each of the V = v NC (v is size of the
input alphabet) transmit symbol vectors b(i) that vec- Fig. 3. Combination of Tx diversity and frequency diversity
tor b(j) (j 6= i) is determined that leads to the highest
fading averaged PEP. Then we determine that value of This is possible by using k LSD encoders (instead of
c that minimizes the maximum of all V possible PEPs. one as in Fig. 1). Then each encoder uses only NkS
2) Coding for spatial multiplexing in combina- OFDM subcarriers. In this case s(i) (i ∈ [1, 2, ..., M ])
tion with transmit diversity (MIMO case): In the in Fig. 1 consists of k stacked vectors, one from each
MIMO case it is possible to multiplex independent encoder.
data streams and transmit them over spatial subchan- Space-time-frequency coding is useful if in addi-
nels. If there are more Tx antennas than used spa- tion to the spatial diversity of the MIMO system also
tial subchannels, inner and outer code have to be de- time and frequency diversity are available. In this case
signed to use spatial multiplexing in combination with the coding is done across m OFDM subcarriers and
Tx diversity. The number of used spatial subchannels across n symbols on each subcarrier, i.e. NL = m · n.
should be adaptive in a very flexible way, because this For m = NS - only one LSD encoder is used - this
allows a flexible tradeoff between diversity to increase means that e.g. s(ν=1) and s(ν=m+1) are transmitted
link reliability and spatial multiplexing to increase the over the same subcarrier; time diversity is exploited
data rate. This tradeoff is controlled by the inner code. by the outer code. Altogether the system is able to use
We use the same outer code as in the MISO case. frequency, spatial and temporal diversity.
Again the inner code is optimized with respect to
the outage capacity of the channel available for the IV. D ECODER
outer code. Using spatial multiplexing an efficient
choice of Cν is given by [6] As appropriate decoder for the LSD codes in [10]
a suboptimal reduced complexity ISI decoder (MAP-
Cν = diag(cν ) · M = Dcν · M (13) DFE) is presented. The ISI results from interfering
spatial subchannels and from the optimized diversity
The NL vectors cν are defined in (5) and (6). The performance of the outer code in fading. This de-
matrix M is a (M × NU ) unitary chirp matrix. The coder is feasible for large block length NC showing
elements of M are determined as described in section a reasonable and scalable complexity. For the LSD
III-B.1.b for the outer code. The adaption to a differ- codes it performs better and shows a lower complex-
ent number NU of spatial subchannels can simply be ity as the V-BLAST decoder (MMSE-DFE) according
done by adding/deleting columns of M. to [11]; alternatively other interference compensation
Outer coding across all OFDM subcarriers may algorithms (e.g. Zero-Forcing, MMSE equalization
lead to large block lengths NI and therefore to high etc.) can be used.
decoder complexities. But usually the distance be-
tween two neighboring OFDM subcarriers is small
and therefore their fading may be highly correlated. V. P ERFORMANCE R ESULTS
Hence, the frequency diversity actually available in In the following we will analyze the performance of
the channel is limited. Design criteria for space- the combination of LSD codes and OFDM based on
frequency codes are derived in [9]. It is more effective simulations for different scenarios. In all simulations
to code across subcarriers that are widely separated. we use the MAP-DFE [10] as decoder and consider a
 10
0 4−QAM: pure Spatial Multiplexing, NC = 30
simulated system (NC = 30) uses 6 Tx and 6 Rx an-
tennas and two different numbers of spatial subchan-
(M, N, NU) = (3, 3, 3)
10
−1
nels; for NU = 6 follows NL = 5 (coding across 5
tones); NU = 3 leads to NL = 10 (coding across
−2 L=1
10
10 tones). The higher the number NU of used spa-
tial subchannels, the higher the data rate, but the SER
SER

−3 L=10
10
increases too. In addition to the spatial Rx and Tx
10
−4
(M, N, NU) = (6, 6, 6) diversity gain a frequency diversity gain is achieved
L=1
(shown in Fig. 5 for L = NL ).
−5 L=5
10
VI. C ONCLUSIONS
−6
10
−4 −2 0 2 4 6 The considered combination of LSD codes and
E / N [dB]
b 0 OFDM leads to space-frequency - or space-time-
Fig. 4. Combination of spatial multiplexing and frequency di-
frequency - coded MIMO OFDM. The code construc-
versity tion allows to exploit spatial, temporal and frequency
4−QAM: (N , N, M) = (30, 6, 6)
diversity. A rich tradeoff between spatial multiplexing
0 C
10 to increase the data rate and diversity to improve the
−1 NU=6, L = 1
link reliability is possible while using a large number
10
of antennas with reasonable complexity.
−2
NU=6, L = 5
10
R EFERENCES
SER

−3
10
NU=3, L = 10 [1] V. Tarokh, N. Seshadri, and A.R. Calderbank, “Space-time
codes for high data rate wireless communication: Perfor-
−4 AWGN mance criterion and code construction,” IEEE Trans. In-
10
form. Theory, vol. 44, no. 2, pp. 744–765, Mar. 1998.
[2] I. Emre Telatar, “Capacity of multi-antenna Gaussian chan-
−5
10 NU=3, L = 1
nels,” Tech. Rep., AT&T Bell Laboratories, 1995.
[3] G. J. Foschini, “Layered space-time architecture for wire-
−6
10
−4 −3 −2 −1 0 1 2 3 4
less communication in a fading environment when using
Eb / N0 [dB] multi-element antennas,” Bell System Techn. J., vol. 1, no.
2, pp. 41–59, Autumn 1996.
Fig. 5. Combination of spatial multiplexing, Tx/Rx diversity and [4] Lizhong Zheng and David N. C. Tse, “Diversity and multi-
frequency diversity plexing: A fundamental tradeoff in multiple-antenna chan-
nels,” IEEE Trans. Inform. Theory, vol. 49, no. 5, pp. 1073–
1096, May 2003.
uniform power delay profile (path gains σi2 = 1 for
[5] Marc Kuhn, Ingmar Hammerstroem, and Armin Wittneben,
i = 1, 2, ...L). “Linear scalable space-time codes: Tradeoff between spatial
Eb
Fig. 3 shows the symbol error rate (SER) over N 0
multiplexing and transmit diversity,” in Proc. SPAWC, 2003.
per Rx antenna for the MISO case (NC = NL = 32), [6] Armin Wittneben and Marc Kuhn, “A new concatenated
linear high rate space-time block code,” in Proc. 55th IEEE
with and without additional frequency diversity; the Veh. Tech. Conf., May 6–9, 2002, vol. 1, pp. 289–293.
coding is done across 32 OFDM subcarriers. For L = [7] Marc Kuhn and Armin Wittneben, “Low complexity decod-
1 only spatial Tx diversity (M = 2) is achieved; with ing of a class of linear space-time block codes by subspace
an increasing number L of (symbol discrete) chan- partitioning,” in Proc. 56th IEEE Veh. Tech. Conf., Septem-
ber 24–28, 2002, vol. 4, pp. 2182–2186.
nel taps the SER decreases. The system clearly ex-
[8] J. G. Proakis, Digital Communications, John Wiley & Sons,
ploits the frequency diversity of the frequency selec- first edition, 2002.
tive MIMO channel. In the pure spatial multiplexing [9] H. Bölcskei and A. J. Paulraj, “Space-frequency coded
MIMO case (without spatial Tx diversity) the codes broadband OFDM systems,” in Proc. IEEE Wirel. Comm.
and Netw. Conf. (WCNC), Chicago, IL, September 23–28,
are able to achieve an additional frequency diversity
2000, pp. 1–6.
gain. Fig. 4 demonstrates this for two different sce- [10] Marc Kuhn and Armin Wittneben, “A new scalable decoder
narios: (i) NC = 30, M = N = NU = 3, NL = 10 for linear block codes with intersymbol interference,” in
with L = 1 and L = 10 (coding across 10 subcarri- Proc. 55th IEEE Veh. Tech. Conf., May 6–9, 2002, vol. 4,
ers). (ii) NC = 30, M = N = NU = 6, NL = 5 with pp. 1795–1799.
[11] G.D. Golden, G.J. Foschini, R.A. Valenzuela, and R.W.
L = 1 and L = 5 (coding across 5 subcarriers). Wolniansky, “Detection algorithm and initial laboratory re-
An example for the possible tradeoff between spa- sults using V-BLAST space-time communication architec-
tial multiplexing and diversity is shown in Fig. 5. The ture,” Electron. Lett., vol. 35, no. 1, pp. 14–15, Jan. 1999.