Você está na página 1de 11

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.


IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, ACCEPTED FOR PUBLICATION 1

Cyclic Space–Frequency Filtering for


BICM–OFDM Systems with Multiple Co–located
or Distributed Transmit Antennas
Harry Z. B. Chen, Student Member, IEEE, and Robert Schober, Senior Member, IEEE

Abstract— In this paper, we introduce cyclic space–frequency relay nodes to form a virtual multiple–input multiple–output
(CSF) filtering for orthogonal frequency division multiplexing (MIMO) system [3]. In this paper, we refer to the former
(OFDM) systems using bit–interleaved coded modulation (BICM) and the latter scenarios as transmission with co–located and
and multiple transmit antennas for fading mitigation, and discuss
the extension to orthogonal frequency division multiple access distributed antennas, respectively. In the distributed case, we
(OFDMA) systems. CSF filtering is a simple form of space– assume a two–hop protocol with decode–and–forward relays
frequency coding (SFC) and may be viewed as a generalization [3]. A particularly simple and efficient form of SFC, which
of cyclic delay diversity (CDD) where the cyclic delays are is applicable to both co–located [4], [5] and distributed [6]
replaced by CSF filters. CSF is applicable to both traditional transmit antennas, is cyclic delay diversity (CDD). CDD is
multiple–input multiple–output systems with co–located transmit
antennas and cooperative diversity systems with decode–and– typically combined with bit–interleaved coded modulation
forward relaying and distributed transmit antennas. Similar to (BICM) [7] which is capable of exploiting the additional
CDD and in contrast to other SFC schemes, CSF filtering does frequency diversity created by CDD and widely used in
not require any changes to the receiver compared to single– existing wireless standards [8].
antenna transmission. Based on the asymptotic pairwise error
probability of the overall system we derive an optimization In this paper, we introduce a novel transmit diversity
criterion for the CSF filters for co–located and distributed scheme for BICM–OFDM systems with multiple co–located
transmit antennas, respectively. We show that the optimum CSF
filters are independent of the interleaver if they do not exceed a or distributed transmit antennas. The proposed scheme is
certain length. If this length is exceeded, the adopted interleaver referred to as cyclic space–frequency (CSF) filtering and may
has to be carefully taken into account in the CSF filter design. be viewed as a generalization of CDD where the cyclic
For several special cases we derive closed–form solutions for the delays are replaced by CSF filters. Unlike in conventional
optimum CSF filters and for the general case we provide various SFC schemes, the CSF filters are not defined over a finite
CSF filter design methods. Our simulation results show that CSF
filtering can achieve significant performance gains over existing field but over the field of complex numbers. Assuming BICM
CDD schemes. at the transmitter and Viterbi decoding at the receiver, we
Index Terms— BICM-OFDM, space-frequency coding (SFC),
develop optimization criteria for CSF filtering with co–located
cyclic space-frequency (CSF) filtering, co–located and distributed and distributed transmit antennas, respectively. Based on these
transmit antennas. criteria we present closed–form solutions for the optimum CSF
filters for several special cases and numerical methods for
I. I NTRODUCTION efficient CSF filter design for the general case, where closed–
form solutions do not exist. Furthermore, we show that BICM–
RTHOGONAL frequency division multiplexing
O (OFDM) converts a frequency–selective fading channel
into a number of parallel flat fading channels. The negative
OFDM with CSF filtering can exploit the full spatial and
spectral diversity of wireless systems with co–located or dis-
tributed antennas. We also discuss the extension of our results
effects of flat fading can be reduced with multiple transmit
to orthogonal frequency division multiple access (OFDMA)
and multiple receive antennas and application of some form
systems. Related Work: There is a large body of literature
of space–frequency coding (SFC), cf. e.g. [1], [2]. Thereby,
on finite field based SFC schemes for co–located, e.g. [1],
the multiple transmit antennas may either belong to the
[9]–[11], and distributed, e.g. [12]–[16], transmit antennas. In
transmitting user (source node) or to other users who act as
contrast to these schemes, the proposed CSF filtering scheme,
Manuscript received December 16, 2007; revised March 17, 2008; accepted if properly designed, does not require any changes to the
May 28, 2008. The associate editor coordinating the review of this paper and receiver compared to single–antenna transmission. This is a
approving it for publication is A. G. I. Fabregas.
The completion of this research was made possible thanks to Bell Canada’s major advantage in networks where devices with different
support through its Bell University Laboratories R&D program and the numbers of antennas have to coexist and facilitates the upgrade
National Science and Engineering Research Council of Canada (Grant STPG of existing wireless systems. For example, to provide transmit
350451). This work was presented in part at the IEEE Vehicular Technology
Conference (VTC–Fall), September 2007, Baltimore, and in part at VTC– diversity in the downlink only the access points have to be
Spring, May 2008, Singapore. equipped with CSF filters and multiple antennas but the user
H. Z. Chen and R. Schober are with the Department of Electrical and devices do not have to be modified. Recently, several attempts
Computer Engineering, University of British columbia, Vancouver, BC V6T
1Z4, Canada (e-mail: {zhibingc, rschober}@ece.ubc.ca). have been made to optimize CDD for co–located antennas.
Digital Object Identifier 10.1109/T-WC.2008.071441 In [8], CDD with large cyclic delays is advocated and the
1536-1276/08$25.00 
c 2008 IEEE

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, ACCEPTED FOR PUBLICATION

1
s1 [t]
g1 CP
2
3 ck  X[k]
4

...

...
...
S 7 8 D
CC Π Mapper IFFT x[t]
sNT [t]
5 9
a) g NT CP
6 10

Fig. 1. Example of a simple two–phase transmission with N = 10 potentially


cooperating relay nodes. NT = 4 relay nodes (S = {1, 3, 9, 10}) ck  X[k] x[t] sn [t]
successfully decode the packet received from the source node (S) and forward CC Π Mapper IFFT gn CP
it to the destination node (D).
b)

interleaver is optimized accordingly. Precoding for CDD was Fig. 2. Block diagram of a) transmitter with NT co–located antennas
employing BICM–OFDM and CSF filtering, and b) transmitter of nth relay
proposed in [17] and optimized for uncoded transmission. node employing BICM–OFDM and CSF filtering.
The idea of replacing the cyclic delays with cyclic filters
has been proposed independently in [18].1 However, [18]
advocates a relatively complex simulation–based approach to Distributed Antennas: We consider a network with one
filter optimization whereas we derive a simple closed–form source node, one destination node, and N relay nodes. Each
optimality criterion. The optimization of CDD for cooperative relay node is assigned a number n ∈ N , N  {1, 2, · · · , N }.
OFDM transmission with distributed antennas was discussed Furthermore, the source and relay nodes are equipped with a
in [5]. We finally note that there are several related papers single antenna, whereas the destination node has NR antennas.
which optimize linear DD and/or linear space–time filtering We adopt a commonly used two–phase protocol [3]. During
for uncoded single–carrier transmission with co–located [19], the first phase, the source node broadcasts a data packet to the
[20] and distributed [21], [22] antennas. relay nodes, and there is no transmission from the source to
Organization: In Section II, the system models for co– the destination node. During the second phase (relay phase),
located and distributed CSF filtering are introduced. The only those NT ≤ N relay nodes which can successfully
optimization of the CSF filters when the set of transmit decode the packet received from the source in the first phase
antennas is known and not known is presented in Sections forward the packet to the destination node, cf. Fig. 1. The relay
III and IV, respectively. In Section V, simulation results are nodes participating in the second phase of transmission are
provided, and Section VI concludes this paper. collected in the set of active nodes S = {n1 , . . . , nNT } ⊆ N ,
Notation: In this paper, E{·}, [·]T , (·)∗ , [·]H , ⊗, ||·||, det(·), |S| = NT . We consider both coordinated and uncoordinated
and tr(·) denote statistical expectation, transposition, complex relaying. In coordinated relaying, the transmitting relay nodes
conjugation, Hermitian transposition, the Kronecker product know S (i.e., they know which other nodes are active),
of two matrices, the L2 –norm of a vector, the determinant of whereas in uncoordinated relaying, the relay nodes do not
a matrix, and the trace of a matrix, respectively. r(X) and know S. Uncoordinated relaying requires less overhead since
λi (X), 1 ≤ i ≤ X, with λi (X) ≥ λi+1 (X), refer to the rank the active relay nodes do not have to communicate with each
and the ordered eigenvalues of X × X matrix X, respectively. other.
Furthermore, I X and 0X are the X×X identity matrix and the
X dimensional all–zeros column vector, respectively. x and
x denote the smallest integer larger than x and the closest B. Transmitter
integer to x, respectively. Finally, (·)X and x[t]  y[t] denote Block diagrams of the proposed transmitter structure for co–
modulo X reduction and circular convolution of x[t] and y[t], located and distributed antennas are shown in Figs. 2a) and
respectively. b), respectively.
Co–located Antennas: The bit stream is encoded with a
II. S YSTEM M ODEL convolutional encoder (CC) which has rate R and free distance
In this section, we introduce the considered network, trans- dfree . The K log2 (M ) coded bits ck , 1 ≤ k  ≤ K log2 (M )
mitter, channel, and receiver models. are interleaved and mapped onto symbols X[k], 1 ≤ k ≤ K,
taken from an M –ary symbol constellation X such as M –ary
A. Network Model phase-shift keying (M –PSK) or M –ary quadrature amplitude
modulation (M –QAM). The effect of the interleaver Π can be
We consider a BICM–OFDM system with K sub–carriers,
modeled by k  → (k, i) where k  denotes the original index
NR receive antennas, and NT co–located or distributed trans-
of coded bit ck , and k and i denote the index of symbol X[k]
mit antennas.
and the position of ck in the label of X[k], respectively. We
Co–located Antennas: In this case, we assume a simple
assume that the interleaver is chosen such that coded bits ck1
point–to–point transmission where transmitter and receiver are
and ck2 with |k1 − k2 | ≤ dfree are mapped onto different
equipped with NT and NR antennas, respectively.
symbols which are transmitted over different sub–carriers.
1 Both the conference version of the submitted paper and [18] were This is necessary to exploit the full diversity offered by the
presented at IEEE VTC–Fall 2007. channel, cf. [2]. Next, an inverse fast Fourier transform (IFFT)
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
CHEN et al.: CYCLIC SPACE–FREQUENCY FILTERING FOR BICM–OFDM SYSTEMS 3

is performed which converts frequency–domain symbols X[k], correlation may be introduced by transmitter and receiver
1 ≤ k ≤ K, into time–domain symbols x[t], 1 ≤ t ≤ K. filtering. For co–located transmit antennas the off–diagonal
At transmit antenna n, 1 ≤ n ≤ NT , the symbols x[t] are elements of ΦS will be in general non–zero due to spatial
passed through CSF filter g n  [gn [0] gn [1] . . . gn [Lg − 1]]T correlation. In contrast, for distributed transmit antennas ΦS
of length Lg ≤ K. The vector g  [g T1 . . . g TNT ]T containing will be a diagonal matrix but its main diagonal elements may
all CSF filter coefficients is normalized to ||g||2 = 1. The CSF not be identical because of non–identical relay–destination
filter output at antenna n is given by distances.
Lg −1 For both co–located and distributed transmit antennas we

sn [t] = gn [t]  x[t] = gn [ν]x[(t − ν)K ]. (1) assume the channel to be constant for the duration of at least
ν=0 one OFDM symbol and the autocorrelation matrix to have full
rank, i.e., r(ΦR ) = NR , r(ΦS ) = NT , and r(ΦL ) = L.
We note that conventional CDD with delay D may be in-
terpreted as a special case of √ the CSF filtering operation
described in (1) with gn [ν] = 1/ NT for ν = (n − 1)D and D. Receiver Processing
gn [ν] = 0 otherwise. Before the symbols sn [t] are transmitted At the receiver, CSF filtering does not require any modi-
by antenna n, a cyclic prefix (CP) is added. We note that the fications compared to CDD and single–antenna transmission.
CSF filtering in (1) has no implications for the required length Assuming that the CP has a length of at least L − 1, after
of the CP which – similar to single–antenna transmission removal of the CP and FFT the received signal at antenna r,
and CDD – is determined only by the length of the channel 1 ≤ r ≤ NR , can be expressed as
impulse response (CIR). Furthermore, similar to CDD, and
in contrast to other SFC schemes, only one IFFT has to Rr [k] = Heq,r [k]X[k] + Nr [k], 1 ≤ k ≤ K, (2)
be performed at the transmitter. However, CSF filtering has
a higher complexity than CDD since it requires NT cyclic where the channel coefficient Heq,r [k] is given by Heq,r [k] =
 L−1 −j2πkl/K
filtering operations whereas for CDD only NT cyclic delays n∈S Hnr [k]Gn[k], with Hnr [k]  l=0 hnr [l]e
Lg −1 −j2πkl/K
are necessary. and Gn [k]  l=0 gn [l]e , and Nr [k] is addi-
Distributed Antennas: The processing at active node n ∈ S tive white Gaussian noise (AWGN) with variance σn2 
is identical to the processing at antenna n in the co–located E{|Nr [k]|2 }. Next, optimum maximum ratio combining of
antenna case, i.e., the transmit signal (before the CP is added) the received signals in each sub–carrier is performed. The
is given by (1). All active relay nodes employ identical CCs, combined signal
interleavers, and mappers but in general different CSF filters
g n . In case of coordinated relaying the CSFs g n , n ∈ S, are 
NR

R[k] = Heq,r [k]Rr [k], 1 ≤ k ≤ K, (3)
optimized for the given set S, cf. Section III. In contrast, for r=1
uncoordinated relaying, the CSF filters g n , 1 ≤ n ≤ N , are
optimized for all nodes in N and do not depend on S (which is then used for bit metric calculation for the coded bit ck
is assumed to be unknown), cf. Section IV. We adopt the [7], [2]
power constraint ||g S ||2 = 1, where g S  [g Tn1 . . . g TnN ]T ,  
NR 2 
T  
for coordinated relaying. For uncoordinated relaying a stricter λi (R[k], ck ) = min R[k] − |Heq,r [k]|2 X  , (4)
X∈Xci
per–node power constraint ||g n ||2 = 1/NT , n ∈ S, is adopted k r=1
since S is not known.
where Xbi
denotes the subset of all symbols X ∈ X whose
For the following, we formally define S  {1, 2, . . . , NT }
label has the value b ∈ {0, 1} in position i. The bit metrics
for the co–located antenna case, which enables a unified
are de–interleaved and Viterbi decoded [7], [2].
treatment of the channel models and the receiver processing
Remark: Eq. (2) shows that CSF filtering transforms the
for CSF filtering with co–located and distributed transmit
MIMO channel into an equivalent single–input multiple–
antennas.
output (SIMO) channel with frequency response Heq,r [k] for
receive antenna r, 1 ≤ r ≤ NR . This corresponds to effective
C. Channel Model SIMO CIRs heq,r [l] of length Leq = L + Lg − 1. Note that,
We consider a frequency–selective channel with CIR in general, Heq,r [k] will contain more spectral degrees of
hnr [l], 0 ≤ l ≤ L − 1, of length L between transmit freedom than Hnr [k] because of the CSF filtering. The re-
antenna n and receive antenna r. For convenience we col- ceiver can directly estimate the equivalent overall CIRs heq,r [l]
lect the corresponding CIR coefficients in vector hnr  using one of the numerous existing techniques for channel
[hnr [0] . . . hnr [L − 1]]T . The CIR coefficients are modeled estimation in SIMO OFDM systems, cf. e.g. [23]. Direct
as possibly correlated zero–mean complex Gaussian random estimation of heq,r [l] is particularly convenient for distributed
variables (Rayleigh fading). antennas since the receiver does not have to know the set of
We define the vector containing all CIR coefficients as active nodes S. However, even for co–located antennas it may
hS  [hTn1 1 hTn2 1 . . . hTnN NR ]T . For the channel autocor- be advisable to directly estimate the NR Leq coefficients of the
T
relation matrix we adopt the popular Kronecker model, i.e., equivalent SIMO channel instead of the NT NR L coefficients
Φ  E{hS hS } = ΦR ⊗ ΦS ⊗ ΦL , where ΦR , ΦS , and
H
of the MIMO channel as long as Leq ≤ NT L. Thus, as far
ΦL denote the receive antenna, the transmit antenna, and as channel estimation is concerned, small Lg are preferable
the temporal autocorrelation matrices, respectively. Temporal since they result in short equivalent CIRs.
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, ACCEPTED FOR PUBLICATION

E. Extension to OFDMA be upper bounded as


⎛ ⎞
Similar to CDD [6], [8], CSF filtering can be easily 2 
1 d
extended to orthogonal frequency division multiple access P (c → ĉ|hS , g) ≤ exp ⎝− min |Heq [k]|2 ⎠ , (6)
(OFDMA) systems. In the following, we briefly point out the 2 4σn2
k,dfree
necessary modifications.
where dmin denotes the minimum Euclidean distance of the
Co–located Antennas: For co–located antennas, the data
signal constellation X and the sum in the argument of the
streams of U users are separately encoded, interleaved,
exponential function in (6) is over dfree distinct sub–carriers
mapped, and assigned to mutually exclusive sets of K/U
K  {k1 , k2 , . . . , kdfree }.2 For the following, it is helpful to
sub–carriers. Subsequently a joint IFFT is applied to all K
rewrite this sum as
frequency domain symbols, cf. [8, Fig. 6]. The resulting time– 
NR  
2
domain sequence x[t], 1 ≤ t ≤ K, is applied to the inputs of    
 
|Heq [k]| =2
 w [k]Gn hnr 
H
the NT CSF filters as shown in Fig. 2a).  
k,dfree k,dfree r=1 n∈S
Distributed Antennas: In the distributed case, U source
nodes communicate with up to U destination nodes at the same = hHS (I NR ⊗ G W G)hS ,
H
(7)
time using mutually exclusive sets of sub–carriers [6]. The nth . . . ej2πk(Leq −1)/K ]T , W 
 w[k] H [1 e
j2πk/K
where
relay node tries to decode the received packets of all U users, w[k]w [k], G  [Gn1 . . . GnNT ], and Gn is an
and forwards the successfully decoded packets to the desti- k,dfree
nation node(s) using OFDMA. Thereby, the packets received Leq × L column–circular matrix with vector [g Tn 0TL−1 ]T in
from different sources are separately encoded, interleaved, its first column. Using (7) in (6) and averaging the PEP with
mapped, and assigned to mutually exclusive sets of K/U sub– respect to hS yields
carriers. The sub–carriers of unsuccessfully decoded source
nodes remain unmodulated. After the IFFT, CSF filtering is 1 d2
P (c → ĉ|g) ≤ det I NR Leq + min
applied as shown in Fig. 2b). 2 4σn2
The received signals Rr [k], 1 ≤ r ≤ NR , belonging to −1
different users/source nodes are processed separately using (ΦR ⊗ W G(ΦS ⊗ ΦL )GH ) . (8)
the standard procedure outlined in Section II-D. The CSF
filter optimization for BICM–OFDM presented in Sections Introducing the definition M  G(ΦS ⊗ ΦL )GH and
III and IV is – with minor modifications – also applicable assuming σn2 1, we can rewrite (8) as
to BICM–OFDMA. We will point out these modifications
1 −r(W M )
wherever necessary. P (c → ĉ|g) ≤ [det(ΦR )]
2
 r(W M ) −NR  −d
 d2min
III. CSF F ILTER O PTIMIZATION FOR K NOWN S λi (W M ) , (9)
i=1
4σn2
In this section, we assume that S is known, which applies
to co–located antennas and coordinated relaying. A unified where d  NR r(W M ) denotes the diversity gain. As-
treatment is possible by adopting again S = {1, 2, . . . , NT } suming dfree ≤ K and Leq ≤ K, 3 we obtain r(W ) =
for the co–located antenna case. We assume that the correla- min{dfree , Leq }, cf. e.g. [2, Appendix], and r(M ) =
tion matrices ΦS and ΦL are known and optimize the CSF min{Leq , NT L}. For the rank of the product W M the
filters accordingly. For simplicity of notation, in this section, inequality
we replace g S with g, since S is fixed and known. r(W ) + r(M ) − Leq ≤ r(W M ) ≤ min{r(W ), r(M )}
(10)
A. Optimization Criterion holds [24], where the upper bound is achievable if W and/or
M are not singular. We conclude from (9) that the achievable
The proposed optimization criterion is based on the worst– diversity gain with CSF filtering is
case pairwise error probability (PEP) of the considered trans-
mission system. From (3) we observe that for a given sub– d = NR · min{dfree , Leq , NT L}. (11)
carrier and a given channel realization the overall system The coding gain of BICM–OFDM with CSF filtering can be
including maximum ratio combining is equivalent to a single– defined as
input single–output (SISO) system with signal–to–noise ratio  r(W M ) 1/r(W M )
(SNR) SNR[k] = |Heq [k]|2 /σn2 , 1 ≤ k ≤ K, where 
C(g)  λi (W M ) , (12)

NR  
2

NR   i=1
 
|Heq [k]|2  |Heq,r [k]|2 =  Hnr [k]Gn [k] . (5) cf. (9). To minimize the asymptotic PEP in (9), the CSF filters
 
r=1 r=1 n∈S have to simultaneously maximize the coding gain C(g) and
Therefore, we can exploit the results presented in [2, Section the diversity gain d. Note that (11) implies that for given
IV], where the PEP of BICM–OFDM was recently derived. In
2
2 Wenote that in OFDMA the possible sets K are user dependent.
particular, exploiting [2, Eq. (7)] and Q(x) ≤ 0.5e−x /2 the OFDMA is employed, we have to assume dfree ≤ K/U and Leq ≤
3 If
PEP of two codewords c and ĉ conditioned on hS and g can K/U , respectively.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
CHEN et al.: CYCLIC SPACE–FREQUENCY FILTERING FOR BICM–OFDM SYSTEMS 5

dfree , NT , and L the maximum possible diversity gain can also have to be cyclic, cf. Section III-A, f (g) can be bounded
be exploited with CSF filters of length Lg,min = min{(NT − as
 Leq Leq
1)L + 1, dfree − L + 1}. Increasing Lg (and consequently Leq ) L
further does not improve d but may increase C(g) and will f (g) ≤ λi (ΦS ⊗ ΦL ) (15)
Leq i=1
significantly impact the design of the optimal g. Therefore,
we consider two different cases for Lg in the following. for any g with ||g||2 = 1. This (in general not achievable)
Case 1 [Lg ≤ Lg,min ]: In this case, Leq ≤ bound will be used in the following to establish optimality of
min{dfree , NT L} and d = NR Leq are valid and we can CSF filters for certain special cases.
rewrite the coding gain as C(g) = (det(W ) det(M ))1/Leq ,
which implies that we can optimize the interleaver (which C. Closed–Form Solution for Special Cases
influences W ) and the CSF filters separately. In particular, for In general, a closed–form solution for the optimum CSF
a given interleaver it suffices to maximize the cost function filter vector g opt does not seem to exist. However, there are
f (g)  det(M ) = det(G(ΦS ⊗ ΦL )GH ) (13) three special cases for which closed–form solutions do exist.
Case 1 [L = 1]: For L = 1 (frequency–nonselective
subject to the power constraint ||g||2 = 1. channel) matrix G does not have a circular structure and is
Case 2 [Lg > Lg,min ]: In this case, we obtain Leq > given by G = [g n1 . . . g nN ]. Therefore, the results from
T
min{dfree , NT L} and d = NR min{dfree , NT L}, and (12) Section III-B can be directly used to obtain the optimum
cannot be further simplified. Therefore, the interleaver and matrix Gopt = √1 U H Lg , where the unitary NT × Lg matrix
Lg
the CSF filters have to be optimized jointly. Based on
U Lg contains the EVs of ΦS corresponding to the Lg largest
(10) it can be shown that for min{dfree , NT L} < Leq ≤
eigenvalues. The optimum CSF filter vectors are simply given
max{dfree , NT L} maximum–rank W and M guarantee a
by the columns of Gopt and achieve the upper bound in (15).
diversity order of d = NR min{dfree , NT L}. This means
Case 2 [Lg = (NT −1)L+1]: In this case, Leq = NT L and
if we choose to optimize/design W and M separately, we
are guaranteed to achieve full diversity but the coding gain
conventional CDD √ (C–CDD) [19] with delay D = L results
in gnm [ν] = 1/ NT for ν = (m − 1)L and gnm [ν] = 0
may not be maximized. On the other hand, for Leq >
otherwise, 1 ≤ m ≤ NT , i.e., G = √N1
I NT L . Therefore, we
max{dfree , NT L} we obtain from (10) NR (NT L + dfree − T
obtain for C–CDD
Leq ) ≤ d ≤ NR min{dfree , NT L}, which means that in this
case maximum–rank W and M do not even guarantee full 1
f (g) = NT L det(ΦS ⊗ ΦL ), (16)
diversity. This latter case was considered for the CDD scheme NT
in [8], where it was indeed found that the interleaver and the which is identical to the  upper bound in (15) for this case
NT L
cyclic delays have to be designed jointly. since det(ΦS ⊗ ΦL ) = i=1 λi (ΦS ⊗ ΦL ). This shows
Considering the above discussion, we will concentrate on that, surprisingly, C–CDD is optimum for Lg = (NT −1)L+1
Case 1 (Lg ≤ Lg,min ) in the remainder of this section. This independent of the correlation matrix ΦS ⊗ ΦL . Note that we
allows us to optimize the CSF filters independent from the implicitly have assumed that Lg,min is limited by NT L and not
interleaver, which is a major advantage. Note that this choice by dfree , i.e., dfree ≥ NT L. Furthermore, Lg < (NT − 1)L + 1
does not compromise the diversity gain since Lg = Lg,min ex- may be preferable in practice since short equivalent channels
ploits the maximum diversity gain d = NR min{dfree , NT L} facilitate channel estimation. If dfree < NT L and/or Lg <
of the BICM–OFDM system. Choosing Lg > Lg,min yields (NT − 1)L + 1, C–CDD is not optimum and the CSF filters
additional coding gains which, however, tend to be small if S have to be optimized.
is known, cf. discussion of Fig. 6. Case 3 [Lg = (κ−1)L+1, κ ∈ {1, 2, . . . , NT }]: We define
the matrix whose columns are the CSF filter vectors and the
B. Upper Bound on Cost Function f (g) EV of ΦS corresponding to the nth largest eigenvalue λn (ΦS )
as G̃  [g n1 g n2 . . . g nN ] and un , respectively, and choose
In this section, we derive an upper bound on f (g). For this T

purpose we assume that G is a general Leq × NT L matrix G̃ as


(without a pre–defined structure), which only has to fulfill 1  T
G̃ = √ u1 0NT ×(L−1) u2 0NT ×(L−1) . . . uκ . (17)
the power constraint tr{GH G} = L. For clarity we denote κ
the corresponding cost function by f˜(G) = det(G(ΦS ⊗ It can be shown that the CSF filters defined by (17) yield a
ΦL )GH ). It can be shown that the maximum value of f˜(G) cost function of
is given by  κ L
1 
 Leq 
Leq f (g) = κL (det(ΦL ))
κ
λi (ΦS ) . (18)
˜ L κ
f (Gopt ) = λi (ΦS ⊗ ΦL ) (14) i=1
Leq i=1 On the other hand, with Leq = κL we obtain from (15)

with optimum matrix Gopt = LLeq U H Leq , where the unitary 1 
κL

NT L × Leq matrix U Leq contains the eigenvectors (EVs) of f (g) ≤ κL λi (ΦS ⊗ ΦL ). (19)
κ i=1
ΦS ⊗ ΦL corresponding to the Leq largest eigenvalues.
Since the CSF filter matrix G does not only have to fulfill Furthermore, λi (ΦS ⊗ΦL ) = λn (ΦS )λl (ΦL ), 1 ≤ i ≤ NT L,
a power constraint but the sub–matrices Gnm , 1 ≤ m ≤ NT , 1 ≤ n ≤ NT , 1 ≤ l ≤ L holds [25]. Therefore, for
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, ACCEPTED FOR PUBLICATION

TABLE I
1 ≤ κ < NT , f (g) in (18) achieves the upper bound in
G RADIENT ALGORITHMS (GA S ) FOR CALCULATION OF THE OPTIMUM
(19) if and only if λκ (ΦS )λL (ΦL ) ≥ λκ+1 (ΦS )λ1 (ΦL )
CSF FILTERS FOR KNOWN AND UNKNOWN S, RESPECTIVELY. T HE
is true. In this case, the columns of matrix G̃ contain the
ELEMENTS OF THE GRADIENT VECTORS ∂f (g)/∂g∗ AND
coefficients of the optimum CSF filters. We note that our
∂f2,K (gS )/∂g∗S ARE GIVEN IN (22) AND (30), RESPECTIVELY. T HE
results in Section V suggest that this design is also optimum
ADAPTATION CONSTANT μ[i] HAS TO BE OPTIMIZED EXPERIMENTALLY.
if λκ (ΦS )λL (ΦL ) < λκ+1 (ΦS )λ1 (ΦL ). However, in this
T HE TERMINATION CONSTANT  HAS A SMALL VALUE ( E . G .  = 10−6 ). i
case, we cannot prove optimality since the upper bound is not
DENOTES THE ITERATION .
achieved. In contrast, for κ = NT , (19) is always identical to
the upper bound regardless of the eigenvalues of ΦS and ΦL , Known S
i.e., the CSF filters according to (17) are optimum. 1 Initialization: Let i = 0 and initialize the CSF filter vector g[i]
with e.g. O–CDD or EV–CSF filters.
D. CSF Filter Design for the General Case 2 Update CSF filter vector:
∂f (g[i])
Due to the particular structure of G it does not seem to be g̃[i + 1] = g[i] + μ[i] .
∂g∗ [i]
possible to find a closed–form solution for g in general. Fur- 3 Normalization: g[i + 1] = g̃[i + 1]/||g̃[i + 1]||.
thermore, since f (g) is not a concave function, the application 4 Termination: If |f (g[i + 1]) − f (g[i])|/f (g[i + 1]) < , goto 5,
of well–known standard tools from the convex optimization otherwise increment i → i + 1 and goto 2.
literature does not seem to be possible. Therefore, we consider 5 End: g[i + 1] is the desired CSF filter vector.
three numerical design methods in the following. Unknown S
1) O–CDD:√ For optimized CDD (O–CDD), we choose 1 Initialization: Let i = 0 and generate a random initial set of
gn [ν] = 1/ NT for ν = Dn < Lg and gn [ν] = 0 otherwise, complex vectors gn [i], 1 ≤ n ≤ N , with ||gn [i]||2 = 1/Na .
n ∈ S. The delays Dn are optimized according to 2 Find worst set of nodes Smin [i] and the worst set of sub–
{Dopt,n1 , . . . , Dopt,nNT } = argmax {f (g)}, carriers Kmin [i]:
(20) 
Dn , n∈S {Smin [i], Kmin [i]} = argmin f2,K (gS ) .
S, K
|S|=Na
which requires a brute force search over all LN g
T
possible
3 Update CSF filter vector:
delay combinations 0 ≤ Dn < Lg , n ∈ S. Similar to ∂f2,Kmin [i] (gSmin [i] [i])
conventional (unoptimized) CDD, for O–CDD the transmit g̃Smin [i] [i + 1] = gSmin [i] [i] + μ[i]
∂g∗S [i]
signal only has to be delayed and multiplications are not  min [i]
4 Normalization: gn [i + 1] = g̃n [i + 1]/( Na ||g̃n [i + 1]||),
necessary. However, compared to conventional CDD, O–CDD
n ∈ Smin [i]
offers more flexibility in the choice of the delays.
5 Termination:
2) EV–CSF Filters: Motivated by Case 3 in Section III-C
If |f2,Kmin [i] (gSmin [i] [i + 1]) − f2,Kmin [i−1] (gSmin [i−1] [i])|/
we propose the following (in general suboptimum) CSF filter
f2,Kmin [i] (gSmin [i] [i + 1]) < , goto 6, otherwise increment
matrix
i → i + 1 and goto 2.
1  T End: gn [i + 1], 1 ≤ n ≤ N , is the desired set Ggrad .
G̃ = √ u1 0NT ×D1 u2 0NT ×D2 . . . 0NT ×Dκ−1 uκ , 6
κ
(21)
where κ  (Lg − 1)/L + 1 and un , 1 ≤ n ≤ κ, denotes
the EVs corresponding to the κ largest eigenvalues of ΦS . random initializations of the algorithm lead to the same f (g)
Furthermore, the Dn , 0 ≤ Dn ≤ L − 1, 1 ≤ n ≤ κ − 1, after the algorithm has converged, cf. also Fig. 3.
κ−1 Remark: We note that for co–located antennas the complex-
meet the constraint n=1 Dn + κ = Lg and are optimized
for maximization of f (g) in a similar fashion as the delays in ity of the proposed CSF design methods (e.g. speed of conver-
(20). If the conditions outlined in Case 3 in Section III-C are gence of the GA) is of minor importance. Since the statistical
met, the EV–CSF filter design in (21) is optimum, otherwise properties of the channel are typically constant, the CSF filters
it is suboptimum. can be pre–computed off–line. For coordinated relaying the
3) GA–CSF Filters: If g is not required to have a CSF filters have to be updated whenever S changes. For this
particular structure, the gradient algorithm (GA) in Ta- purpose, the CSF filters may be pre–computed and stored for
ble I (known S) can be used to optimize g directly. all possible S or calculated on–line before data transmission
The elements of the gradient vector ∂f (g)/∂g ∗  starts.
[∂f (g)/∂gn1 [0] ∂f (g)/∂gn1 [1] . . . ∂f (g)/∂gnNT [Lg − 1]]H
required for the adaptation in Table I are given by IV. CSF F ILTER O PTIMIZATION FOR U NKNOWN S
∂f (g) In this section, we assume that N , ΦL , and all possible
= det (M ) tr M −1 E nν (ΦS ⊗ ΦL )GH , (22) correlation matrices ΦS are known for CSF filter optimization.
∂gn [ν]
However, the set of antennas (nodes) S participating in the
where the entries of Leq × NT L matrix E nν are equal to current transmission are not known (uncoordinated relaying).
1 at those positions where G has entries gn [ν] and zero Thus, the relay nodes n ∈ N can only be assigned fixed,
otherwise. Because of its relatively involved form, we cannot pre–optimized CSF filters g n which are independent of the
prove convergence of the GA to the global optimum of f (g). current S. The set of all filters g n , n ∈ N , is denoted by
However, our extensive simulations have shown that different G  {g 1 , g 2 , . . . , g N }. Under these circumstances the goal
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
CHEN et al.: CYCLIC SPACE–FREQUENCY FILTERING FOR BICM–OFDM SYSTEMS 7

of the CSF filter optimization is to design sets G of CSF filters where ρS  NT g H n1 g n2 denotes the correlation between g n1
which achieve a good performance regardless of which nodes and g n2 . Therefore, in this case, (25), (26) is equivalent to
are active. In this section, for clarity we add the subscript S minimizing the maximum correlation between any two vectors
to g, G, and M to emphasize their dependence on the set in G. The maximum correlation can be bounded as [26]
of active nodes. Furthermore, we add subscript K to matrix  
W to emphasize its dependence on set K which contains the N −2 2
ρmax  max{|ρS |} ≥ max , 1− . (28)
sub–carrier indices corresponding to dfree consecutive outputs S 2(N − 1) N
of the convolutional encoder. Therefore, an upper bound for the minimum of the cost
function over all sets S is given by
A. Optimization Problem
1
To make the problem more manageable, we assume for f1,min  min{f1 (g S )} ≤ NT det(ΦS )
S NT
optimization of G that the number of active nodes NT = Na is   
fixed and known. In a practical transmission, the actual number N 4 1
min , 1− . (29)
of active nodes NT may differ from the design parameter 2(N − 1) N N
Na , of course. If we exclude the non–diversity case NT = 1, Sets of vectors with minimum maximum correlation ρmax
NT = 2 may be considered as the worst–case scenario and (and consequently maximum f1,min ) are called Grassmannian
more active nodes can only increase the diversity of the frames [27]. For N ≤ L2g = 4 these Grassmannian frames
channel. Therefore, it is reasonable to adopt Na = 2 for CSF achieve the bounds in (28) and (29) with equality. For the
filter design. construction of the corresponding optimum sets for N = 3
Similar to Section III the optimization is based on the and N = 4 we refer to [27]. For N > L2g the bounds in
PEP. More precisely, our goal is to minimize the maximum (28) and (29) are not achievable and Grassmannian frames are
asymptotic PEP in (9) over all possible sets S with |S| = Na . difficult to construct. In this case, numerical design methods
The diversity gain d is also limited by (11) if S is unknown, are an attractive alternative, see [27] and next subsection.
i.e., the maximum diversity gains for known and unknown S
are identical. Furthermore, to maximize the coding gain (12),
C. CSF Filter Design for General Case
depending on Lg , we use the cost functions
In this section, we propose suboptimum and optimum CSF
f1 (g S ) = det(GS (ΦS ⊗ ΦL )GH
S ), Lg ≤ Lg,min , (23) filter design methods if S is not known.
and 1) Distributed C–CDD: In distributed C–CDD, the delay
at node n is equal to Dn = (n − 1)(Lg − 1)/(N − 1),
f2 (g S ) = det(I Leq +γW K GS (ΦS ⊗ΦL )GH S ), Lg > Lg,min , 1 ≤ n ≤ N . For the special case Lg = K this choice
(24) for the CDD delays was proposed in [6] for an amplify–
where γ can be interpreted as an SNR (cf. (8)). For γ 1 and–forward protocol. However, with the decode–and–forward
f2 (g S ) is equivalent to the coding gain C(g S ) in (12) but protocol considered in this paper, distributed C–CDD fails
is easier to optimize. Similar to the case when S is known, to capture the full diversity of the channel in general. The
choosing Lg > Lg,min complicates the optimization since the minimum delay difference between two nodes is ΔD 
properties of the interleaver have to be taken into account via min2≤n≤N {Dn − Dn−1 }. Therefore, assuming for example
W K . However, unlike when S is known, when S is unknown, that the diversity gain is not limited by dfree and NT = 2, it
substantial gains are possible with Lg > Lg,min (cf. Fig. 6). is easy to see that distributed C–CDD cannot exploit the full
Therefore, we also consider Lg > Lg,min in this section. Based diversity of the channel if L > ΔD.
upon the above discussion, we formally state our optimization 2) GA–CSF Filters: Similar to the case of known S,
problem as a GA can be used to optimize G based on f1 (g S ) and
 
f2 (g S ), respectively. We refer to the corresponding set of CSF
Gopt = argmax min fν (g S ) , ν ∈ 1, 2,(25)
G S, |S|=Na filter vectors as Ggrad . The elements of the gradient vector
subject to ||g n || = 1/Na , 1 ≤ n ≤ N,
2
(26) ∂f1 (g S )/∂g ∗S are already given in (22), and the elements of
the gradient vector ∂f2 (g S )/∂g ∗S can be obtained as
i.e., the optimum set of CSF filters Gopt maximizes the
∂f2,K (g S ) 
minimum coding gain for a given Lg . −1
= γf2,K (g S ) tr I Leq + γW K M S
∂gn [ν]

B. Upper Bound and Closed–Form Solution for Special Case W K E nν (ΦS ⊗ ΦL )GH S , n ∈ S, 0 ≤ ν < Leq , (30)
In this section, we concentrate on the case Lg ≤ Lg,min . We
note that (15) is also an upper bound for f1 (g S ), g n ∈ Gopt , where we have added the subscript “K” in f2,K (g S ) to
n ∈ S. However, this bound may be very loose in general. In emphasize its dependence on K.
the following, we will show that a tighter bound and a closed– The GA with f2,K (g S ) as cost function is given in Table I.
form solution to (25), (26) exist for Na = NT = 2, L = 1, Similar to the case where S is known, we cannot prove global
and Lg = 2. For this special case, f1 (g S ) can be re–written convergence of the proposed algorithm. However, in our (ex-
as tensive) simulations the sets Ggrad obtained with the algorithm
1 for different random initializations always resulted in similar
f1 (g S ) = NT det(ΦS ) (1 − |ρS |2 ), (27)
NT performances (cf. also Fig. 6). If OFDMA is employed, the
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, ACCEPTED FOR PUBLICATION

1 1
search for the worst set of sub–carriers Kmin [i] has to be ρL = 0.0
GA-CSF
EV-CSF ρL = 0.6
GA-CSF
EV-CSF
performed over the U sets of sub–carriers corresponding to 0.9 O-CDD 0.9 O-CDD

the U users. On the other hand, if Lg ≤ Lg,min , f2,K (g S ) can 0.8 0.8
be replaced by f1 (g S ) in the GA in Table I, which becomes
0.7 0.7
independent of K. Therefore, the algorithm can be simplified
and in Step 2 only the worst set of nodes Smin [i] but not the 0.6 0.6

worst set of sub–carriers has to be found, which reduces the 0.5 0.5
complexity of the algorithm considerably.
0.4 0.4
Remark: We would like to point out that the speed of con-

α
vergence and complexity of the GA are of minor importance. 0.3 0.3
If the positions of the source, relay, and destination nodes are 0.2 0.2
fixed, the channel statistics are time–invariant and the set of
CSF filters has to computed only once. If the nodes are mobile, 0.1 0.1

the channel statistics will change and the CSF filters should 0 0
5 10 15 5 10 15
be updated accordingly. a)
Lg
b)
Lg
Finally, we point out that for the special case S = N
the optimization problems for known and unknown S are Fig. 3. Ratio α  f (g)/f˜(Gopt ) vs. Lg for GA–CSF, EV–CSF, and
identical. Therefore, the GA in Table I for unknown S with O– CDD filters, respectively. NT = 4, ρ = 0.9, and L = 5.
S = N can be used to find the optimum CSF filters for known
S and a given interleaver if Lg > Lg,min , cf. Fig. 6.
3) R–CSF Filters: If nodes frequently enter and leave with the O–CDD solution. The adaptation step size of the GA
the network, random CSF (R–CSF) filters are an attractive was μ[i + 1] = λμ[i], where μ[0] = 0.1 and λ = 0.999.
alternative to deterministic CSFs. In this case, each relay node First we compare the CSF filter designs from Section III-D.
generates an R–CSF filter before it encodes and forwards In particular, we consider the cost function f (g) normalized
the received message to the destination node. If the R–CSF with the upper bound f˜(Gopt ) in (14). The resulting ratio
filters are taken independently from some suitable distribution, α  f (g)/f˜(Gopt ) is shown in Figs. 3a) and b) for ρL = 0.0
matrix GS will have maximum rank for all possible S with and ρL = 0.6, respectively, as a function of Lg for GA–
probability one. Therefore, for Lg ≤ Lg,min R–CSF filters CSF, EV–CSF, and O–CDD filters. NT = 4, ρ = 0.9, and
can exploit the full available diversity. Note that this is not L = 5 are valid. As expected from the results in Section
necessarily true if Lg > Lg,min , cf. Case 2 in Section III-A. III-C, for Lg = (NT − 1)L + 1 = 16 EV–CSF filters and
In this paper, we consider R–CSF filters whose coefficients O–CDD, which is equivalent to C–CDD with D = L = 5 in
are uniformly distributed on a complex hypersphere. Similar this case, achieve the upper bound, i.e., α = 1. Furthermore,
random designs have been proposed for distributed space–time for ρL = 0, λκ (ΦS )λL (ΦL ) ≥ λκ+1 (ΦS )λ1 (ΦL ) holds for
codes in [28]. κ = {1, 2, 3}. Therefore, the EV–CSF filters also achieve the
4) R–CDD: Similar to the R–CSF filters in 3), each upper bound for Lg = (κ−1)L+1, κ = {1, 2, 3}, cf. Sections
relay node n may simply apply a random delay Dn ∈ III-C, III-D, Fig. 3a). The fact that the GA–CSF filters also
{0, 1, . . . , Lg − 1} before data transmission. This scheme achieve the upper bound in this case indicates that the GA
is referred to as R–CDD. Since two relay nodes will pick indeed finds the optimum solution. For ρL = 0.6 and κ =
the same delay with non–zero probability, R–CDD does not {2, 3}, λκ (ΦS )λL (ΦL ) < λκ+1 (ΦS )λ1 (ΦL ). Therefore, the
achieve the maximum diversity gain of the BICM–OFDM EV–CSF filters cannot achieve the upper bound, cf. Case 3
system. in Section III-C. However, the fact that the EV–CSF filters
achieve the same performance as the GA–CSF filters suggests
that the EV–CSF filters are still optimum. Finally, we note that
V. S IMULATION R ESULTS GA–CSF and EV–CSF filters outperform O–CDD for all Lg
In this section, we present simulation results for CSF filter- except for the extreme cases Lg = 1 and Lg = 16.
ing with co–located and distributed antennas, respectively. For In Fig. 4, we consider the BER of BICM–OFDM with GA–
all bit error rate (BER) results shown, we adopted the quasi– CSF filtering vs. Eb /N0 (Eb : received energy per information
standard convolutional code (171, 133)8 with rate R = 1/2, bit, N0 : power spectral density of underlying continuous–time
constraint length 7, and free distance dfree = 10. Since the passband noise process) and compare it with that of the CDD
CSF design is not affected by the number of receive antennas, scheme proposed by Bauch and Malik (BM–CDD) in [8] and
we assume NR = 1 throughout this section. with BICM–OFDM concatenated with Alamouti’s space–time
block code (STBC). We assume BPSK, K = 64, NT = 2,
ρL = 0, ρ = {0.0, 0.9}, and L = {1, 5}. For CSF filtering
A. Co–located Antennas full–diversity–achieving filter lengths of Lg = 2 and Lg = 6
In this section, we consider co–located antennas with are assumed for L = 1 and L = 5, respectively. Furthermore,
|i−j|
[ΦS ]i,j = ρ|i−j| and [ΦL ]i,j = ρL , where ρ and ρL denote 8×8 and 16×8 block interleavers are adopted for CSF filtering
the spatial and temporal correlation coefficients, respectively. and the STBC scheme, respectively. For BM–CDD, which
The GA–CSF filters employed in this section were generated uses a delay of D = K/2 = 32, we show results for the
with the GA given in Table I (known S), which was initialized 8 × 8 interleaver and without interleaving. For L = 1 BM–
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
CHEN et al.: CYCLIC SPACE–FREQUENCY FILTERING FOR BICM–OFDM SYSTEMS 9

GA-CSF
BM-CDD, D = 32, no interl.
In Fig. 5, we show the BER of user 1 of a BICM–OFDMA
10
−1 BM-CDD, D = 32, interl.
Alamouti’s STBC
system with GA–CSF filtering, O–CDD, BM–CDD, and C–
CDD. We assume U = 4, K = 512, 16–QAM, NT = 4, L =
L=1
5, ρL = 0, ρ = {0.6, 0.9}, and a 32 × 16 block interleaver.
10
−2 ρ = 0.9 User u ∈ {1, 2, 3, 4} is assigned sub–carriers u + 4k, 0 ≤
k ≤ 127. Since NT L > dfree , the diversity gain is limited by
the code. Therefore, for the GA–CSF filters and the O–CDD
10
−3 filters we adopted Lg = Lg,min = 6. In contrast, for BM–
BER

ρ=0 CDD and C–CDD we chose D = K/(NT U ) = 32 [8] and


L=5 D = L = 5 [21], respectively, resulting in Lg > Lg,min in
ρ = 0.9
10
−4 both cases. Fig. 5 shows that the large delays of BM–CDD
ρ=0 and C–CDD are not beneficial if the diversity gain is limited
by the code. BM–CDD even suffers from a loss in diversity
10
−5 (see Case 2 in Section III-A for an explanation). For ρ = 0.6
0 2 4 6 8 10 12 14 16
the proposed O–CDD scheme with optimized delays yields
Eb /N0 [dB]
performance gains of 1.8 dB and 7 dB compared to C–CDD
Fig. 4. BER of BICM–OFDM with GA–CSF filtering, BM–CDD [8], and and BM–CDD at BER = 10−5 , respectively. An additional
Alamouti’s STBC vs. Eb /N0 . OFDM, K = 64, BPSK, NT = 2, and ρL = gain of 0.3 dB is possible with GA–CSF filtering. We note
0. GA–CSF filtering: 8 × 8 block interleaver, Lg = 2 for L = 1, Lg = 6
for L = 5. BM–CDD: 8 × 8 block interleaver and no interleaver. Alamouti’s
that the performance of both C–CDD and BM–CDD could be
STBC: 16 × 8 block interleaver. Convolutional code with (171, 133)8 , R = improved by optimizing the interleaver for the given delays.4
1/2, and dfree = 10.
B. Distributed Antennas
GA-CSF, L g = 6, ρ = 0.6 We assume that the source–relay and the relay–destination
GA-CSF, L g = 6, ρ = 0.9
10
−1 C-CDD, D = 5, ρ = 0.6 channels are independent and identically distributed (i.i.d.),
C-CDD, D = 5, ρ = 0.9
O-CDD, L g = 6, ρ = 0.6 respectively. This implies that the distances between the relay
O-CDD, L g = 6, ρ = 0.9
BM-CDD, D = 32, ρ = 0.6
nodes are much smaller than the source–relay and the relay–
10
−2
BM-CDD, D = 32, ρ = 0.9 destination distances, respectively. Furthermore, we assume
perfect synchronization throughout this section.5 The GA–
CSF filters were optimized with the GA specified in Table
10
−3 I (unknown S) with Na = 2, γ = 15, and adaptation step
BER

size μ[i + 1] = λμ[i], where μ[0] = 0.1 and λ = 0.9999. We


considered a system with K = 128, 16–QAM, and a 16 × 32
10
−4 block interleaver throughout this section.
In Fig. 6, we consider the case L = 1 and NT = 2 and
compare the minimum coding gain Cmin  minS {C(g S )}
10
−5 of sets of GA–CSF filters of various lengths Lg with the
0 2 4 6 8 10 12 14 16 18 20 22
Eb /N0 [dB]
corresponding upper bounds for Lg = 2 and ! known and
unknown ! S which are given by Cb,k  12 det(W ) and
Fig. 5. BER of BICM–OFDMA with GA–CSF filtering, O–CDD, BM– Cb,u  f1,min det(W ), respectively, cf. (12), (15), (29).
CDD [8], and C–CDD [21] vs. Eb /N0 . U = 4, K = 512, 16–QAM, NT = For Lg = Lg,min = 2 the GA–CSF filters achieve the upper
4, L = 5, ρL = 0, 32 × 16 block interleaver. Convolutional code with
(171, 133)8 , R = 1/2, and dfree = 10. bound Cb,u for N ≤ L2g = 4 as expected from Section IV-B.
For N > 4 the upper bound cannot be achieved in principle,
but the minimum coding gain of the GA–CSF filters closely
follows the upper bound, which confirms the effectiveness
CDD without interleaving achieves the same performance as
of the proposed GA. For Lg > Lg,min = 2 considerable
CSF filtering. As explained in [8], for L = 1 the interleaver
additional performance gains are possible for large N . For
is not beneficial to BM–CDD and leads even to a small
N = 2 S = N holds and the second GA in Table I calculates
performance degradation. In contrast, for L = 5 interleaving
the GA–CSF filters for known S and Lg > Lg,min , cf. Section
and no interleaving is preferable for BM–CDD if ρ = 0.0 and
IV-C. Fig. 6 shows that for known S (i.e., N = 2) the
ρ = 0.9, respectively, which underlines the importance of a
performance gains achievable with Lg > Lg,min are small
channel–dependent joint optimization of interleaver and delay
which also validates our focus on Lg ≤ Lg,min in Section III.
in BM–CDD. For L = 5 and ρ = 0.9 CSF filtering yields a
In Fig. 7, we show the BER of BICM–OFDM with GA–
gain of more than 2 dB at BER = 10−5 even if interleaving
CSF, R–CSF, R–CDD, and C–CDD [6] filtering (as defined in
is not applied in BM–CDD. While for L = 1 Alamouti’s
code and GA–CSF achieve a similar performance, for L = 5 4 We have not included results for the scheme in [18] in Figs. 4 and 5 since

Alamouti’s code yields a performance gain compared to GA– the simulation–based CSF filter optimization outlined in [18] is difficult to
CSF. Note, however, that unlike CSF filtering Alamouti’s code reproduce and the designs given in [18] were optimized for a convolutional
turbo code, whereas we consider convolutional codes.
requires a change to the receiver structure compared to single– 5 Additional results, which are not shown here because of space limitation,
antenna transmission. show that both CSF and CDD are robust against timing errors.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, ACCEPTED FOR PUBLICATION

5.5
−1 NT =1
GA-CSF, Lg =5 10
GA-CSF, Lg =4 C-CDD, Lg =K
5 R-CSF, Lg =6
GA-CSF, Lg =3
GA-CSF, Lg =2 GA-CSF, Lg =6
4.5 Cb,k , Lg =2 GA-CSF, known S, Lg =6
Cu,k , Lg =2
4 −2
10
3.5

3
Cmin

BER
2.5 −3
10
p = 0.5
2
p = 0.4
1.5
p = 0.6
1 10
−4
2 3 4 5 7 10 15 20 25 30 6 8 10 12 14 16 18 20 22 24 26
N
Eb /N0 [dB]

Fig. 6. Minimum coding gain Cmin of GA–CSF filtering for various Lg ,


upper bound Cb,k for known S and Lg = 2, and upper bound Cu,k for Fig. 8. BER at destination node for two–hop transmission with GA–
unknown S and Lg = 2 vs. N . L = 1, NT = Na = 2, K = 128, CSF, R–CSF, and C–CDD [6] filtering when S is unknown vs. Eb /N0 . For
16–QAM, and 16 × 32 block interleaver. comparison the BER of GA–CSF filtering when S is known is also shown.
Each relay node listens to source node with probability p. N = 10, L = 1,
Na = 2, K = 128, 16–QAM, and 16 × 32 block interleaver. Convolutional
code with (171, 133)8 , R = 1/2, and dfree = 10.
GA-CSF, L g = 2
GA-CSF, L g = 5
−1
10 R-CSF, L g = 2
R-CSF, L g = 5
R-CDD, L g = 2 In Fig. 8, we show the BER at the destination node for a
R-CDD, L g = 5
C-CDD, L g = K two–hop transmission with N = 10 and L = 1. The energy
10
−2 GA-CSF, known S, L g = 2
GA-CSF, known S, L g = 5 per received bit at the relays in the first hop is identical
NT = 1
to the energy per received bit Eb at the destination and
each relay listens to the source with probability p. p < 1
−3
10 may be caused by power saving protocols or devices that
BER

are permanently turned off. Only listening relays which can


successfully decode the packet (corresponding to one OFDM
−4
10 symbol) received from the source participate in the second
phase of transmission using GA–CSF, R–CSF, or C–CDD [6]
filtering. If no relay can decode the source packet, the packet
−5
10
0 5 10 15 20 25 30 35 40 is retransmitted. Note that this decoding failure is not included
Eb /N0 [dB] in the BER curves in Fig. 8. For comparison, we also show
the BER of GA–CSF filtering with known S and for the case
Fig. 7. BER of BICM–OFDM with GA–CSF, R–CSF, R–CDD, and C–
CDD [6] filtering when S is unknown vs. Eb /N0 . For comparison the BER
where only a single relay retransmits the packet (NT = 1). In
of GA–CSF filtering when S is known is also shown. N = 10, L = 1, contrast to Fig. 7, the number of active relay nodes NT is now
NT = Na = 2, K = 128, 16–QAM, and 16 × 32 block interleaver. a random variable and only a single node will be active with a
Convolutional code with (171, 133)8 , R = 1/2, and dfree = 10.
certain probability. For high SNR the latter case dominates the
BER which explains why asymptotically all BER curves are
parallel to the curve for NT = 1. However, CSF filtering and
Section IV-C) when S is not known for the second phase of CDD still achieve significant performance gains compared to
transmission. We assume that NT = Na = 2 out of N = 10 NT = 1. For GA–CSF filtering the performance loss caused
relay nodes forward the (error–free) decoded packet to the by the lack of knowledge of S is remarkably small (i.e., less
destination node over a flat fading channel (L = 1) and the than one dB) for all considered cases.
BER is averaged over all possible relay node combinations.
For comparison we also show the performance of GA–CSF
VI. C ONCLUSIONS
filtering when S is known. Fig. 7 shows that CSF filtering
with optimized filters (GA–CSF) and random filters (R–CSF) In this paper, we have proposed CSF filtering for BICM–
achieves full diversity. As expected from Fig. 6, increasing Lg OFDM systems with multiple co–located or distributed trans-
from 2 to 5 yields significant performance gains in both cases. mit antennas. CSF filtering may be viewed as an extension
For Lg = 5 GA–CSF with unknown S closely approaches the of well–known CDD where the cyclic delays are replaced by
performance of GA–CSF with known S. Both R–CDD and CSF filters. Similar to CDD and unlike other SFC schemes,
C–CDD achieve only diversity gain one since, in R–CDD, CSF filtering requires only one IFFT at the transmitter and
two active relays will select the same delay with non–zero no modifications at the receiver compared to single–antenna
probability and, in C–CDD, Lg > Lg,min is valid in general transmission. Based on the PEP of the overall system, we have
but the interleaver is not optimized accordingly. derived optimization criteria for the CSF filters when the set of
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
CHEN et al.: CYCLIC SPACE–FREQUENCY FILTERING FOR BICM–OFDM SYSTEMS 11

transmitting antennas is known and not known, respectively. [16] H. Mheidat, M. Uysal, and N. Al-Dhahir, “Equalization Techniques
For certain special cases we have provided closed–form so- for Distributed Space–Time Block Codes With Amplify–and–Forward
Relaying,” IEEE Trans. Signal Processing, vol. 55, pp. 1839–1852, May
lutions for the resulting optimization problems, and we have 2007.
proposed various CSF filter design methods for the general [17] Z. Hong and L. Thibault, “Bandwidth Efficient Space-Frequency
case. We have shown that the proposed CSF filter designs Codes Based on Cyclic Delay Diversity,” in Proc. IEEE Ve-
hic. Techn. Conf. (VTC), Sep. 2006.
are independent of the interleaver if the CSF filters do not [18] W. Hillery, T. Krauss, B. Mondal, T. Thomas, and F. Vook, “Finite Im-
exceed a certain length Lg,min . If this length is exceeded, the pulse Response Cyclic Shift Transmit Diversity for Broadband Mobile
interleaver has to be taken into account for CSF filter design OFDM,” in Proc. IEEE Vehic. Techn. Conf. (VTC), Oct. 2007.
[19] D. Gore, S. Sandhu, and A. Paulraj, “Delay Diversity Codes
to ensure full diversity. Our simulation results have shown for Frequency Selective Channels,” in Proc. IEEE Intern. Com-
that CSF filtering can achieve significant performance gains mun. Conf. (ICC), New York, May 2002, pp. 1949–1953.
over various existing CDD designs at the expense of a small [20] T. Hehn, R. Schober, and W. H. Gerstacker, “Optimized Delay Diver-
increase in complexity. sity for Frequency-Selective Fading Channels,” IEEE Trans. Wireless
Commun., vol. 4, pp. 2289–2298, Sep. 2005.
[21] H. El Gamal and D. Aktas, “Distributed Space-Time Filtering for Coop-
erative Wireless Networks,” in Proc. IEEE Global Telecommun. Conf.,
R EFERENCES Dec. 2003.
[22] S. Yiu and R. Schober, “Optimized Distributed Space–Time Filtering,”
[1] H. Bölcskei and A. Paulraj, “Space–Frequency Coded Broadband IEEE Trans. Wireless Commun., vol. 6, pp. 982–992, Mar. 2007.
OFDM Systems,” in Proc. IEEE Wireless Commun. and Networking [23] X. Cai and G. Giannakis, “Error Probability Minimizing Pilots for
Conf. (WCNC), Chicago (IL), Sep. 2000, pp. 1–6. OFDM With M –PSK Modulation Over Rayleigh-Fading Channels,”
IEEE Trans. Veh. Technol., vol. 53, pp. 146–155, Jan. 2004.
[2] E. Akay and E. Ayanoglu, “Achieving Full Frequency and Space
[24] R. Horn and C. Johnson, Matrix Analysis. Cambridge: Cambridge
Diversity in Wireless Systems via BICM, OFDM, STBC, and Viterbi
University Press, 1999.
Decoding,” IEEE Trans. Commun., vol. 54, pp. 2164–2172, Dec. 2006.
[25] T. Moon and W. Stirling, Mathematical Methods and Algorithms for
[3] J. Laneman and G. Wornell, “Distributed Space–Time Block Coded Signal Processing. New York: Prentice Hall, 2000.
Protocols for Exploiting Cooperative Diversity in Wireless Networks,” [26] P. Xia, S. Zhou, and G. Giannakis, “Achieving the Welch Bound With
IEEE Trans. Inform. Theory, vol. 49, pp. 2415–2425, Oct. 2003. Difference Sets,” IEEE Trans. Inform. Theory, vol. 51, pp. 1900–1907,
[4] A. Dammann and S. Kaiser, “Standard Conformable Antenna Diversity May 2005.
Techniques for OFDM Systems and its Application to the DVB-T [27] T. Strohmer and R. Heath Jr., “Grassmannian Frames with Applications
System,” in Proc. IEEE Global Telecommun. Conf. (Globecom), 2001, to Coding and Communications,” Applied and Computational Harmonic
pp. 3100–3105. Analysis, vol. 14, pp. 257–275, 2003.
[5] J. Tan and G. Stüber, “Multicarrier Delay Diversity Modulation for [28] B. Mergen and A. Scaglione, “Randomized Space–Time Coding for Dis-
MIMO Systems,” IEEE Trans. Wireless Commun., vol. 3, pp. 1756– tributed Cooperative Communication,” IEEE Trans. Signal Processing,
1762, Sep. 2004. vol. 55, pp. 5003–5017, Oct. 2007.
[6] P. Tarasak and Y. Lee, “Joint Cooperative Diversity and Scheduling
in OFDMA Relay Systems,” in Proc. IEEE Wireless Commun. and Harry Z.B. Chen (S’03) received the B.E. degree from Huazhong University
Networking Conf. (WCNC), Hong Kong, Mar. 2007. of Science and Technology, China, in 1990, the M.E. degree from University
[7] G. Caire, G. Taricco, and E. Biglieri, “Bit–Interleaved Coded Modula- of Science and Technology of China, China, in 2000, and the M.A.Sc degree
tion,” IEEE Trans. Inform. Theory, vol. 44, pp. 927–946, May 1998. from the University of British Columbia (UBC), Canada, in 2004, respectively,
[8] G. Bauch and J. Malik, “Cyclic Delay Diversity with Bit-Interleaved all in electrical engineering. He is currently pursuing the Ph.D. degree at UBC.
Coded Modulation in Orthogonal Frequency Division Multiple Access,” His research interests include MIMO-OFDM systems, cooperative diversity
IEEE Trans. Wireless Commun., vol. 5, no. 8, pp. 2092–2100, Aug. systems, beamforming, and space-frequency coding theory.
2006.
[9] Z. Liu and G. Giannakis, “Space–Frequency Coded OFDM over Robert Schober (M’01, SM’07) was born in Neuendettelsau, Germany, in
Frequency–Selective Fading Channels,” IEEE Trans. Signal Processing, 1971. He received the Diplom (Univ.) and the Ph.D. degrees in electrical
vol. 50, pp. 2465–2476, Oct. 2002. engineering from the University of Erlangen-Nuermberg in 1997 and 2000,
[10] W. Su, Z. Safar, and K. Liu, “Full–Rate Full–Diversity Space–Frequency respectively. From May 2001 to April 2002 he was a Postdoctoral Fellow
Codes With Optimum Coding Advantage,” IEEE Trans. Inform. Theory, at the University of Toronto, Canada, sponsored by the German Academic
vol. 51, pp. 229–249, Jan. 2005. Exchange Service (DAAD). Since May 2002 he has been with the University
[11] H. Lu and M.-C. Chiu, “Construction of Asymptotically Optimal Space– of British Columbia (UBC), Vancouver, Canada, where he is currently an
Frequency Codes for MIMO–OFDM Systems,” IEEE Trans. Inform. Associate Professor and Canada Research Chair (Tier II) in Wireless Com-
Theory, vol. 53, pp. 1676–1688, May 2007. munications. His research interests fall into the broad areas of Communication
[12] S. Yatawatta and A. Petropulu, “A Multiuser OFDM System With User Theory, Wireless Communications, and Statistical Signal Processing.
Cooperation,” in 38th Asilomar Conf. Signals, Syst., Comput., Pacific Dr. Schober received the 2002 Heinz Maier-Leibnitz Award of the German
Grove, CA, Nov. 2004. Science Foundation (DFG), the 2004 Innovations Award of the Vodafone
[13] P. Anghel and M. Kaveh, “Exact Symbol Error Probability of a Cooper- Foundation for Research in Mobile Communications, the 2006 UBC Killam
ative Network in a Rayleigh-Fading Environment,” IEEE Trans. Wireless Research Prize, and the 2007 Wilhelm Friedrich Bessel Research Award
Commun., vol. 3, pp. 1416–1421, Sep. 2004. of the Alexander von Humboldt Foundation. In addition, he received best
[14] L. Yu and A. Stefanov, “Cooperative Space-Time Coding for MIMO paper awards from the German Information Technology Society (ITG), the
OFDM Systems,” in Proc. IEEE Military Commun. Conf., Oct. 2005. European Association for Signal, Speech and Image Processing (EURASIP),
[15] Y. Li, W. Zhang, and X.-G. Xia, “Distributive High-Rate Full-Diversity IEEE ICUWB 2006, the International Zurich Seminar on Broadband Commu-
Space-Frequency Codes for Asynchronous Cooperative Communica- nications, and European Wireless 2000. Dr. Schober is also the Area Editor for
tions,” in Proc. IEEE Intern. Symp. Inform. Theory, Jul. 2006. Modulation and Signal Design for the IEEE Transactions on Communications.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:58 from IEEE Xplore. Restrictions apply.

Você também pode gostar