FLOWS IST-2001-32125 Deliverable No:D14 (k-d) 色- x() X) 区-d Figure 2.1:MIMO-DFE scheme Layered space-time architecture(BLAST) In [AD01].a al fra finite-length MIMO MMSE fo cha s was deri e f MIMO chan- In [LP02],MIMO rchit ectures based on channe ation have been proposed and inves V-BLAST ed successive interference feed span of K taps. The feedforward section contains(K+1)nrnR taps,while the feedback Decisions on all data streams are fed back eedback b. [wi bilTR=Pi (2.19) where w FIR filters of length(K+1)and projecting (in the MMSE sense)ith data stream away from its own uncancelec vector pi is pi=Ef[y(k)-s:(k-d-1)]Tsi(k-d)} (2.20) and the matrix R is defined by R-E y() (2.21) 17 December 2003 Page 27
FLOWS IST-2001-32125 Deliverable No: D14 Figure 2.1: MIMO-DFE scheme. Layered space-time architecture (BLAST) Similarly to receivers in flat-fading channels, layered space-time receivers are of interest for frequency-selective channels [LP02, AD01]. In [AD01], a general framework for analysing finite-length MIMO MMSE equalisers for frequency-selective channels was derived. Such equalisers shorten the memory of MIMO channels and, in addition, perform noise whitening and multi-channel matched filtering. In [LP02], MIMO architectures based on channel equalisation have been proposed and investigated, such as MIMO decision-feedback equaliser (DFE) and ordered successive interference cancellers (OSICs) which are an extension of narrowband V-BLAST. The MIMO DFE architecture is shown in Fig.2.1. The MIMO-DFE consists of a feedforward FIR filter with a temporal span of Kf + 1 taps and a feedback FIR filter with a temporal span of Kb taps. The feedforward section contains (Kf + 1)nT nR taps, while the feedback section contains (Kb + 1)nT nT taps. Decisions on all (nT ) data streams are fed back into the detection process for each stream. Assuming that previous decisions are correct, the MMSE solution for feedforward wi and feedback bi taps for i th data stream is given by [wi bi] TR = pi (2.19) where wi = vec{Wi}, bi = vec{Bi}. The (Kf + 1) × nR matrix Wi contains feedforward taps connecting all receive antennas with one output of the feedforward section through 1-D FIR filters of length (Kf + 1) and projecting (in the MMSE sense) i th data stream away from its own uncanceled (precursor) ISI as well as the uncanceled co-channel interference (CCI) from the other data streams. The matrix Bi has dimensionality nT × Kb. The vector pi is defined by pi = E [y(k) − si(k − d − 1)]T s∗ i(k − d) (2.20) and the matrix R is defined by R = E y(k) −si(k − d − 1) y(k) −si(k − d − 1) H . (2.21) 17 December 2003 Page 27
FLOWS 1ST-2001-32125 Deliverable No:D14 x(5). ,-d x()E () Figure 2.2:MISO-DFE scheme. MSE Figure 2.3:Partly connected OSIC-DFE scheme 17 December 2003 Page 28
FLOWS IST-2001-32125 Deliverable No: D14 Figure 2.2: MISO-DFE scheme. Figure 2.3: Partly connected OSIC-DFE scheme. 17 December 2003 Page 28
FLOWS IST-2001-32125 Deliverable No:D14 MO-DF MO-DFE MSE substrea Figure 2.4:Fully connected OSIC-DFE scheme The partly connected OSIC-DFE scheme consists of nr successive stag s.the first two of which are sho n in Fig.2.3.At each data stream (in MMSE sense) th is DF用 stream is c estimate S(-d)for (k-d)thsymbol ofith data stream selected by an ordering mechanism. The fully coneted SIC-DFE scheme is shown inFig.oists ofstages ever,now DFE data strea m (in the ted an canceled out fro me. ted scheme. plicitly deco mechanism at every stage is retained. nesmsareymReeam The performance of the three schemes was evaluated by simulation for a TU channel [LP02] MIMO-DFE latt implementation ime equ tion e ha adv age in that the umber of tatively tect nance and con ture is sho structure.especially with a large delay spread omer numb receive antennas In [Ari].two possib coded layered space-time coding ac agheatherassumingindependeg EDGE and GSM.where multipath dispersion may last up to several symbol periods.Block error rate(BLER)compares with the"outage capacity",i.e..the probability that a specified 17 December 2003 Page 29
FLOWS IST-2001-32125 Deliverable No: D14 Figure 2.4: Fully connected OSIC-DFE scheme. Finally, the parameter d is the decision delay which is considered to be the same for all data streams. In a particular case of a single transmit antenna the MIMO DFE scheme transforms to the MISO DFE scheme as shown in Fig.2.2. The partly connected OSIC-DFE scheme consists of nT successive stages, the first two of which are shown in Fig.2.3. At each stage, the “best” data stream (in the MMSE sense) is extracted and detected by exploiting the MISO DFE scheme shown in Fig.2.2, and then this stream is cancelled out from the received signals. The output of every stage at time k is an estimate sˆi(k −d) for (k −d)th symbol of i th data stream selected by an ordering mechanism. The fully connected OSIC-DFE scheme is shown in Fig.2.4. It also consists of nT stages. However, now at each stage, the “best” data stream (in the MMSE sense) is extracted and detected by exploiting the MIMO DFE scheme shown in Fig.2.1, and then this stream is canceled out from the receive signals. As in the partially connected scheme, the nT stages are ordered so that data streams are extracted and decoded in sequence of decreasing MSE. Unlike the partially connected scheme, all nT data streams are implicitly decoded within every stage and their symbol decisions are fed back. However, only the stream selected by the ordering mechanism at every stage is retained. The performance of the three schemes was evaluated by simulation for a TU channel [LP02]. The OSIC receivers show significantly better BER performance in comparison with that of the MIMO-DFE scheme. Both the partially connected and fully connected OSIC receivers perform similarly in most cases. Since the latter receiver is more complicated, the partially connected OSIC receiver is preferable for practical implementation. In [ZM02] another layered space-time equalisation architecture is proposed, which utilises the MIMO delayed decision feedback sequence estimation (MIMO-DDFSE) instead of DFE at each stage of detection. This structure has an advantage in that the number of tentatively detected and output data streams can be variable at different stages, allowing tradeoffs between performance and complexity. This structure is shown to significantly outperform the MISODFE structure, especially with a large delay spread or a smaller number of receive antennas. In [Ari00], two possible coded layered space-time structures are considered: one applying coding across the multiple signal processing layers (LST-I), and the other assuming independent coding within each layer (LST-II). The system is assumed to have similar ISI situations as in EDGE and GSM, where multipath dispersion may last up to several symbol periods. Block error rate (BLER) compares with the ”outage capacity”, i.e., the probability that a specified 17 December 2003 Page 29
FLOWS 1ST-2001-32125 Deliverable No:D14 bit rate is ot by the chane Perfect nal coding and modulation sc e are considered B8n5Hwswrtrgh8eeemaana4B2olwmn3ea the Shannon 2 than the latter The coded lay ce-time a po annon ounda BLERout 2 of hichisuetrac Schemes with a feedback channel be acquired at the transmitter either if a feedback s present or th chann tran unction is t e in th th air DD ons(tha known at the receiveras welas the transmitter.problems of ptimal desi Optimal desi ce-time lin nd do ed in ISSB+02 The design is ba sed on an ntima symbols and G(decoder)of received samples. I he scheme operates a a block transmission system vectors y四 are el nrough a mapp bchane3luoe t the c gendeco used ru ows the pr and decoder to be y op sed.do tre quency-s he r ance provided by the designs can be used as a lower bound for comparison similar ated in application to MIMO transmission in HIPER- 1. where for every OFDM carrier a SISO-like link is o based on the SISO-Iik was reporte posed to use in order to provide data rates beyond the maximum rate of 54 Mbps. A system with linear precoding and decoding in application to DS/CDMA communication systems in the downlink is proposed in [CLMOO].In the syste m,adaptive tra nsversal filters are optimum transmitand receive fiters which maximise the found.The propose system has been shown to possess a better performance than a (1,1)Rake receiver te In MIMO OFDM S th fee data tted thro mp es th fee k infor ati this sche ame sig all sub in ach su ier.it is t sub-channels perform f the scheme is demonstrated for(4,4)and(4.6)systems in a multipath channe with perfect channel estimates. 17 December 2003 Page30
FLOWS IST-2001-32125 Deliverable No: D14 bit rate is not supported by the channel capacity. Perfect channel estimation is assumed. Convolutional coding and 8-PSK modulation scheme are considered. For TU and HT GSM profiles and nR = nT = 4, the performance at 10% BLER is within 3 dB of the Shannon bound for LST-I with six iterations, and within 4 dB for LST-II with two iterations. For a large number of transmit and receive antennas, coding across the layers provides a better performance than independent coding within each layer. However, with nR = nT = 2, the former is heavily affected by decisions errors and, therefore, provides a poorer performance than the latter. The coded layered space-time approaches can achieve a performance within about 3 dB of the Shannon bound at 10% BLER, about 2 dB of which is a loss due to practical coding schemes used. Schemes with a feedback channel In some scenarios, channel information can be acquired at the transmitter either if a feedback channel is present or the MIMO channel transfer function is the same in both directions (that is, reciprocity applies) and the transmitter and receiver operate in the TDD mode. Assuming the channel to be known at the receiver as well as the transmitter, problems of optimal designs can be formulated. Optimal designs for space-time linear precoders and decoders have been derived in [SSB+02]. The design is based on an optimal pair of linear transformations F (precoder) of transmit symbols and G (decoder) of received samples. The scheme operates as a block transmission system in which vectors of symbols are encoded and modulated through a linear mapping in space and time dimensions. The solutions convert the MIMO channel with memory into a set of parallel flat-fading subchannels. The designs exploit the channel eigendecomposition in constructing the optimal F and G, assuming the channel to be perfectly known. The approach used allows the precoder and decoder to be jointly optimised, does not treat frequency-selective and flat fading channel cases separately, and does not rely on the full rank of the matrices involved. The performance provided by the designs can be used as a lower bound for comparison with more realistic schemes. A similar approach was investigated in application to MIMO transmission in HIPERLAN/2 [MGI01], where for every OFDM subcarrier a SISO-like link is formed, based on the highest singular value of the channel matrix. Further development of this idea was reported in [MGI02] where several SISO-like channels with independent symbol streams has been proposed to use in order to provide data rates beyond the maximum rate of 54 Mbps. A system with linear precoding and decoding in application to DS/CDMA communication systems in the downlink is proposed in [CLM00]. In the system, adaptive transversal filters are used at the transmit and receive antennas. Assuming that the channel is estimated perfectly, optimum transmit and receive filters which maximise the overall SINR are found. The proposed system has been shown to possess a better performance than a (1,1) Rake receiver. In practical applications of multicarrier modulation, the subcarriers are each allocated the same power. The “water-pouring” approach shows that significant performance improvement can be obtained by adaptively allocating the information bits and power over the sub-channels. In MIMO OFDM systems with feedback, the data transmitted through the feedback channel can indicate an optimal bit allocation for different subcarriers and different transmit antennas. A simplified bit allocation scheme for V-BLAST OFDM MIMO system proposed in [NCM02] minimises the feedback information. In this scheme, all sub-channels, if used, apply the same signal constellation. This reduces the complexity in determining the detection order. However, in each subcarrier, it is possible that not all the spatial sub-channels are active. High performance of the scheme is demonstrated for (4,4) and (4,6) systems in a multipath channel with perfect channel estimates. 17 December 2003 Page 30
FLOWS 1ST-2001-32125 Deliverable No:D14 MIMO channel estimation the channel tra sfer fun ns fo he same eamble in the standard HIPERLAN/2.The first part of the data pack et is a preamble consisting of ntennas transm t th ate of the rans n is L will not be compatible with HIPERLAN/2 and IEEE 802.11a standards.The proposed method based on a yclic time-shitt nhsrplogglboanitegheseondantena; t ha are the same. chanetrmnsietne back to t -compens ates for the channel phase es at the with respect to the single transmit antenna system. nario 2.4 Time-varying channels 2.4.1 Model of time-varying MIMO channel with n nsmit antenna i to transmit antennas and antennas. nt e ante C at each moment are independent complex Gaussian random variables with variances o2=Eh()2}and Doppler spectrum Eth(k)hi(+n) （2.22) n=-o The symbol transmitted from antenna i at moment is denoted by si().Denote the total energy transmitted by all antennas as E()=s(k)2 (2.23) The signal observed at receive antennaj at the moment is given by 5=∑)s,因+n, (2.24) 1 where the additive whit and Gaussian with E{nj(k)ni(k)}= at eac y(k)=H(k)s()+n(k) (2.25) where H(k)is a ng x nr matrix with entries(k).the nr x 1 vector s(k)contains the transmitted symbols s,(k).the ng x 1 vector n()contains the noise samples nj(),and the nx 1 vector y(k)contains samples vi(A)at receive antennas. 17 December 2003 Page 31
FLOWS IST-2001-32125 Deliverable No: D14 MIMO channel estimation The paper [Sli02] proposes a channel estimation technique for OFDM based systems, such as HIPERLAN/2 and IEEE 802.11a, with transmitter diversity. This technique is able to extract the channel transfer functions for two transmitter antennas using the same preamble as defined in the standard HIPERLAN/2. The first part of the data packet is a preamble consisting of two identical OFDM symbols. If both the antennas transmit this packet, a receiver will only provide an estimate of the sum of the channel transfer functions of the different antennas and does not allow the estimation of individual transfer functions. One possible solution is to use orthogonal training sequences, different for each transmit antenna. However, such a solution will not be compatible with HIPERLAN/2 and IEEE 802.11a standards. The proposed method is based on a cyclic time-shift of the pilot signal transmitted by the second antenna. It has been shown that, in a case when for two neighbouring subcarriers the channel coefficients are the same, the two channel transfer functions can be separated. The receiver estimates the two channels and feeds the channel estimates back to the transmitter. The transmitter pre-compensates for the channel phases at each subcarrier and then makes the transmission. A considerable performance gain has been obtained by simulation for different indoor scenarios with respect to the single transmit antenna system. 2.4 Time-varying channels 2.4.1 Model of time-varying MIMO channel Now we describe a time-varying MIMO channel with nT transmit antennas and nR receive antennas. The tap gain from transmit antenna i to receive antenna j is denoted by hij . The antennas are separated far enough to ensure nT ·nR independent fading channels from transmit to receive antennas. The channel taps hij (k) at each moment k are independent zero-mean complex Gaussian random variables with variances σ2 h = E{|hij (k)| 2} and Doppler spectrum S(ω) = ∞ n=−∞ E{hij (k)h∗ ij (k + n)}ejωn. (2.22) The symbol transmitted from antenna i at moment k is denoted by si(k). Denote the total energy transmitted by all antennas as Es(k) = nT i=1 |si(k)| 2 . (2.23) The signal observed at receive antenna j at the moment k is given by yj (k) = nT i=1 hij (k)si(k) + nj (k) (2.24) where the additive noise nj (k) is white and Gaussian with E{nj (k)n∗ l(k)} = N0δjl, N0 is the noise spectral density. The average SNR at each receive antenna is γ = (E{Es(k)}σ2 h)/N0. In matrix notation, the channel model is represented as y(k) = H(k)s(k) + n(k) (2.25) where H(k) is a nR × nT matrix with entries hij (k), the nT × 1 vector s(k) contains the transmitted symbols si(k), the nR × 1 vector n(k) contains the noise samples nj (k), and the nT × 1 vector y(k) contains samples yj (k) at receive antennas. 17 December 2003 Page 31