Abstract—In this contribution we propose new techniques for d0 [m] dˆ0 [m]
multi-user MISO broadcasting with Filter Bank based Multicarrier d1 [m] AFB dˆ1 [m]
(FBMC) systems. FBMC is a prominent alternative to OFDM .. SFB Channel ..
. +Eq .
that can achieve higher bandwidth efficiency since no cyclic dM−1 [m] dˆM−1 [m]
prefix is required. However, when extending FBMC to a multi- WGN
antenna system, approaches that are used in OFDM can not be
directly applied due to its special structure and the existence Fig. 1. Single user FBMC System
of inter-carrier and inter-symbol interference. We will consider a
downlink scenario for the transmission from a multi-antenna base
station to several single-antenna mobile stations. To this end, a different approach that is not restricted in terms of the numbers
Minimum Mean Square Error (MMSE) based multi-tap precoder of antennas is presented in [2]. For scenarios where the channel
and equalizer will be designed. Our method is of particular
can no longer be assumed flat, however, different approaches
interest for highly frequency selective channels.
are required. In [3] an equalizer design is proposed for a SIMO
and MIMO case with no precoding operation at the transmitter.
I. I NTRODUCTION If the complexity is to be shared between the transmitter and
receiver, a suboptimal linear approach is given in [4] for a joint
Multicarrier systems are used in a wide range of current
transmitter and receiver design while at least one of those is
communication systems such as Wi-Fi, DVB, DSL or LTE.
constrained to be a single-tap filter.
The most prominent modulation scheme certainly is Orthogo-
nal Frequency Division Multiplexing (OFDM), which can be For multi-user systems the number of contributions is fairly
efficiently implemented using the Discrete Fourier Transform limited. The authors of [5] propose an SDMA approach for
(DFT). The problem of frequency selectivity in transmission the MISO broadcast channel based on Tomlinson-Harashima
channels can be coped with by introducing a cyclic prefix, Precoding. However, the proposed technique is again only
restoring orthogonality between subcarriers. This, however, valid for channels with high coherence bandwidth. In this
leads to redundant transmissions and with that to a loss of contribution we will devise a new multi-user MISO broad-
spectral efficiency, especially in highly frequency selective casting approach for highly frequency selective channels. We
channels. Filter Bank based Multicarrier (FBMC) systems will propose two different MMSE approaches. In the first one,
present an alternative modulation where this disadvantage can we will provide a closed form solution for a system that has all
be overcome. Specific pulses that are longer than the symbol the complexity at the transmitter side by the use of precoders
period and well localized in both time and frequency are used only. In the second, we will extend this approach to a joint
instead of the rectangular pulses in OFDM. A cyclic prefix transmitter and receiver design by an iterative technique.
is no longer required. The increase of efficiency, however,
comes with the price of higher complexity in transmitter and The paper is organized as follows: In Section II we will
receiver design and implementation of a multi-tap equalizer is introduce the system and subchannel model. In Section III the
now required as opposed to the one-tap equalizer in OFDM. MMSE-based precoder design will be introduced, followed by
Further advantages of the more complex design are stronger the iterative transmitter and receiver design in Section IV.
stop-band attenuation due to the filter specifications and re- Both approaches are evaluated in Section V by numerical
laxed synchronization requirements compared to OFDM. The simulations. Finally, conclusions are drawn in Section VI.
latter is particularly interesting when multi-user systems are
employed. II. S YSTEM M ODEL
The majority of recent contributions regarding MIMO An abstract model of a single-user SISO FBMC system is
FBMC consider point-to-point communication. In this context shown in Fig. 1. The M complex valued QAM input signals
a frequent assumption is that the channel’s frequency response dk [m] at a symbol rate of 1/Ts are combined by the Synthesis
is approximately flat within one subcarrier and only differs Filter Bank (SFB) into one signal at a sampling rate of M/Ts
slightly at adjacent subchannels. For this scenario, the authors and transmitted via a frequency selective AWGN channel. On
of [1] have proposed a spatial multiplexing scheme for a the receiver side the signals are again separated by the Analysis
system with two transmit and receive antennas (2x2). A Filter Bank (AFB) including an equalization step to account
xk−1 [n] x
dk−1 [m] Ok−1 M hk−1
2
xk+1 [n] x
dk+1 [m] Ok+1 M hk+1 η[r]
2
x
M Ts
dk [m] ℜ{•} 2 2 xk [n] Bk
x s1k [n]
M
jℑ{•} 2 b1,1
k
are commonly used for the design of SFB and AFB to allow an
efficient implementation based on modified DFT filter banks b1,2
k
as shown in [6]. The pulse shaping filters at subcarrier k are s2k [n]
x2k [n] b2,2
therefore given by k
..
2π LP − 1 .
hk [r] = hP [r] exp j k r − , r = 0, . . . , LP − 1,
M 2 bN
k
t ,2
XU
where ν is the delay of the whole system and ck and uk cor- +σx2 ξkj,T Ψs,T s j
k,k Ψk,k ξk + Rη̂k j,(R) }, (13)
respond to the interference introduced in adjacent subcarriers s=1
s6=j
and interference introduced in other users, respectively, given
by where Rη̂j,(R) corresponds to the covariance matrix of the real
k
part of the noise received at user j. For this result we have
|ck |2 = |ℑ{xj,T j j 2 j,T j j 2
k Hk−1,k bk }| + |ℑ{xk Hk+1,k bk }| assumed that all input signals are independently and identically
U distributed
X
2
|uk | = |ℑ{xj,T s j 2 j,T s j 2
k Hk−1,k bk }| + |ℜ{xk Hk,k bk }|
i σ 2 I
x B+Q−1 , if s = t and k = l,
h
s=1 E xsk [n]xt,H
l [n] =
s6=j 0B+Q−1 , otherwise,
(14)
+ |ℑ{xj,T
k H s
bj 2
k+1,k k }| . (9)
and σx2 = 0.5σd2 with σd2 being the variance of the QAM
Whenever the interference at the subchannel of interest k is input symbols djk [n] The above expression can be optimized
considered, the real part of the interference is taken. Due to by taking its derivative with respect to bjk and setting it to zero,
alternating structure of the OQAM staggering from subchannel which yields
to subchannel (see Section II), we have to take the imaginary
part if one of the neighbouring subcarriers k±1 is taken into
account. For the simulations in Section V we have used a U
X
fixed delay ν. In general the delay can also be optimized to ξ̂kj = Φs,T s s,T s
k−1,k Φk−1,k + Ψk,k Ψk,k
minimize the MSE. This, however, is beyond the scope of this s=1
contribution. !−1
+ Φs,T s
k+1,k Φk+1,k Ψj,T
k,k eν . (15)
To find linear expressions for the real and imaginary parts
of the interference terms in Eqs. (6) and (9) it is practical to
redefine the input vectors xjk and and channel matrices Hl,k s
The vector eν is a unit vector with a one at position ν which
as follows corresponds to the precoder delay. Simulations have shown that
depending on design parameters such as number of antennas,
xj,T s j j,T s j
k Hl,k bk = x̃k H̃l,k bk (10) number of users or length of the precoders, the matrix
U
X
where x̃jk no longer contains the OQAM symbols but is a Φs,T s s,T s s,T s
k−1,k Φk−1,k + Ψk,k Ψk,k + Φk+1,k Φk+1,k
purely real vector. Basically, all imaginary j ’s (see (4)) have s=1
s
been shifted into the matrix H̃l,k by multiplying every second
s might be rank deficient. In this case, however, simulations
row of Hl,k with these j ’s. To simplify notation, we further
have shown that using a Moore-Penrose inverse yields good
define
results. This can be done by calculating the singular value
h
s,(R) s,(I)
i decomposition (SVD) and only inverting its singular values
Ψsl,k = H̃l,k −H̃l,k , above a certain threshold. All other inverses are set to zero. The
h
s,(I) s,(R)
i threshold itself has to be chosen carefully to achieve optimal
Φsl,k = H̃l,k H̃l,k , performance.
" #
j,(R)
bk If we want to limit the transmitted signal power we can
ξkj = j,(I) , (11)
bk additionally introduce a power constraint given by
b̂j,T j
k,opt b̂k,opt = 1. (16)
and we can find the following relations
We can satisfy this condition by normalizing the result of (15)
ℜ{xj,T s j j,T s j
k Hl,k bk } = x̃k Ψl,k ξk ,
which yields
ℑ{xj,T s j j,T s j
k Hl,k bk } = x̃k Φl,k ξk . (12) j 1
ξ̂k,opt =q ξ̂kj (17)
j,T j
ξ̂k ξ̂k
Plugging (12) into (6) and (9) and then into (8) and evaluating
Obviously, in this case we need to perform a real-valued signal at user j on subchannel k prior to equalization is given
scaling at the receiver to adjust the power level. The receiver by
output x̂jk [n] therefore has to be multiplied with the scalar Nt k+1 U
X X X
q ykj [n] = hi,j
k,l [n] ∗ bi,s s j
l [n] ∗ xl [n] + η̂k [n]. (22)
γk = ξ̂kj,T ξˆkj .
j
(18) i=1 l=k−1 s=1
In the following approach we extend the previous design pi,j i,j i,s
k,l [n] = hk,l [n] ∗ bl [n]. (24)
method by dividing the complexity between the transmitter Again we switch to a matrix-vector notation where we define
and receiver. In particular, we will modify the precoder only the vector ykj ∈ CLeq consisting of Leq consecutive output
slightly and additionally design an equalizer that takes care symbols
of residual interference left by the precoder. The result can
then be fed back to design a new precoder, starting the second U
X k+1
X X Nt
iteration step. An initial equalizer is required for the first ykj [n] = i,j s
Pk,l xl [n] + Γk η j [r], (25)
iteration. A possible choice is a simple delay equal to half s=1 l=k−1 i=1
of the equalizer length. i,j
where Pk,l ∈ CLe q×Leq +Q+B−2 is the convolution matrix
corresponding to pi,j k,l . This signal is now filtered with the
A. Precoder Design
equalizer gkj ∈ CLeq and we again shift the imaginary units
In contrast to the precoder design in Section III, the j from the input signal xjk into the convolution matrix by
equalizer now has to be taken into account. We do that by multiplying every second column of Pk,l i,j
with j leaving the
slightly modifying our subchannel model from (2) to resulting x̃sl as a purely real vector:
Nt Xk+1 U
!
Nt
X X ( U k+1
!)
j i,j i,s
x̂k [n] = ql,k [n] ∗ s
bl [n] ∗ xl [n] + η̃kj [n], j j,H
X X X i,j s j
α̂k [n] = ℜ gk P̃k,l x̃l [n] + Γk η [r] .
i=1 l=k−1 s=1 s=1 l=k−1 i=1
(19) (26)
with We can now formulate the optimization problem which is given
i,j by
ql,k [n] = gkj [n] ∗ hi,j
l,k [n] (20) h i
ĝkj = arg min E α̂jk [m] − αjk [m − ν] . (27)
being the convolution of the total transmission chain and the j
gk
equalizer. η̃kj [n] = gkj [n] ∗ η̂kj [n] corresponds to the received
noise convolved with the equalizer. If we again change to Compared to (8) we do not have to add additional interference
matrix-vector notation and take the real part of the received terms. They are already embedded in the received signal
signal we get α̂jk [m]. Again we require a linear expression for taking the
real part in (26). Accordingly we now define the new vector
k+1 U
X X and matrices
α̂jk [m] = ℜ{xs,T j s j
l [n]Qk,l bl } + ℜ{η̃k [n]}. (21) "
i,j,(R)
#
l=k−1 s=1 i,j P̃k,l
Θk,l = i,j,(I) ,
P̃k,l
The convolution matrix is now represented by Qjk,l ∈ h iT
j,(R),T j,(I),T
CQ+Leq +B−2×B with Leq being the length of the equalizer. The ϑjk = gk gk . (28)
only differences with (6) are that we have replaced the matrix
j
Hk,l by Qjk,l which includes the equalization and we have Similar to Section III we can show that
obtained a different noise vector which, however, is irrelevant ℜ{gkj,H P̃k,l
i,j s
x̃l [n]} = ϑj,T i,j s
for the optimization. The remaining precoder optimization is k Θk,l x̃l [n]. (29)
therefore equivalent to that in Section III. If we now plug (29) into (26) and then into (27) we find the
optimal equalizer by taking the derivative of (27) with respect
B. Equalizer Design to ϑjk and set it to zero yielding
U k+1
!−1
The proposed equalizer is similar to the one described in j
X X j j,T
ση2j
[8] which is intended for use in a SISO system. Here, we are ϑ̂k = Θk,l Θk,l + 2 Rη̂j,(R) Θjk,k eν , (30)
s=1
σd k
adapting it for use in a multi-user MISO scenario. The received l=k−1
where
0
10
Nt
X
Θjk,k = Θi,j
k,l . (31)
i=1
−1
10
Again we have assumed that the input signals are i.i.d. (see
(14)). The noise covariance matrix Rη̂j,(R) is given by
k
" #
2 (R) (I)
ση ′ ′ T
BER
′ Γk Γk 10
−2
−2
10
those used for data transmission. Random 16-QAM symbols
were used and we additionally assumed perfect channel state
information. As a figure of merit we calculate the Bit Error
Ratio (BER) for 200 randomly generated channel realizations 10
−3
it
increased from 1 to 5. Incrementing B any further, however,
−1
does not lead to any additional improvement. To measure the 10
10
−2
4x4 precoder only: B=5 first approach yields a closed-form solution for the precoder-
4x4 iterative: B=3, L =5
eq only design, the second iterative approach allows for a joint
4x3 precoder only: B=5
4x3 iterative: B=3, Leq=5 precoder and equalizer optimization. Both approaches show
−3 4x2 precoder only: B=5
good results in a wide SNR range, and the iterative design
10
4x2 iterative: B=3, L =5
eq outperforms the precoder-only design in low and medium
Eb /N0 regimes.
−4
10
−20 −15 −10 −5 0 5 10 15 20 25 30
ACKNOWLEDGMENT
E /N
b 0
This work is partially supported by the European project
EMPhAtiC (ICT- 318362).
Fig. 10. Comparison of iterative and precoder only design for different
number of users and Nt = 4
R EFERENCES
[1] M. El Tabach, J.-P. Javaudin, and M. Helard, “Spatial data multiplexing
in the low Eb /N0 regime the lowest BER is already achieved over OFDM/OQAM modulations,” in Communications, 2007. ICC ’07.
after only one iteration. Further improvement with additional IEEE International Conference on, 2007, pp. 4201–4206.
iterations can only be reached in the high Eb /N0 regime. [2] M. Payaró, A. Pascual-Iserte, and M. Najar, “Performance comparison
Although not shown in this picture, the same is true for the between FBMC and OFDM in MIMO systems under channel uncer-
tainty,” in Wireless Conference (EW), 2010 European, 2010, pp. 1023–
4x2 and the 4x4 systems. 1030.
Finally, the simulation results for precoder-only and it- [3] T. Ihalainen, A. Ikhlef, J. Louveaux, and M. Renfors, “Channel equaliza-
erative approaches are compared for all scenarios in Fig. tion for multi-antenna FBMC/OQAM receivers,” Vehicular Technology,
IEEE Transactions on, vol. 60, no. 5, pp. 2070–2085, 2011.
10. In general, the performance of both designs deteriorates
[4] M. Caus and A. Perez-Neira, “Transmitter-receiver designs for highly
when increasing the number of users. This however is not frequency selective channels in MIMO FBMC systems,” Signal Process-
unexpected since the number of interference terms that have ing, IEEE Transactions on, vol. 60, no. 12, pp. 6519–6532, 2012.
to be minimized increases with additional users. A direct [5] M. Caus and A. I. Perez-Neira, “SDMA for filterbank with Tomlinson
comparison between precoder-only and the iterative approach Harashima precoding,” in Communications (ICC), 2013 IEEE Interna-
in the same scenario shows that in all cases the BER can be tional Conference on, 2013, pp. 4571–4575.
improved in a wide Eb /N0 range by the computationally more [6] T. Karp and N. Fliege, “Modified DFT filter banks with perfect recon-
complex iterative technique. While the overall performance of struction,” Circuits and Systems II: Analog and Digital Signal Processing,
IEEE Transactions on, vol. 46, no. 11, pp. 1404–1414, 1999.
the 4x4 case is rather poor we can still observe a considerable
[7] B. Saltzberg, “Performance of an efficient parallel data transmission
improvement by additionally using the equalizers. In the 4x2 system,” Communication Technology, IEEE Transactions on, vol. 15,
system, a small gain can still be seen but it is rather insignif- no. 6, pp. 805–811, 1967.
icant. It is also interesting to take a look at the complexity in [8] D. Waldhauser, L. Baltar, and J. Nossek, “MMSE subcarrier equalization
terms of “number of taps per user” which amounts to 20 and 17 for filter bank based multicarrier systems,” in Signal Processing Advances
for the precoder-only and the iterative approaches, respectively. in Wireless Communications, 2008. SPAWC 2008. IEEE 9th Workshop
The result suggests that the more complex iterative precoder on, 2008, pp. 525–529.
design allows for a reduced complexity in the implemented [9] L. Hentilä, P. Kyösti, M. Käske, M. Narandzic, and M. Alatossava.
(2007, December) MATLAB implementation of the WINNER
system while achieving the same performance. phase II channel model ver1.1. [Online]. Available: https://www.ist-
winner.org/phase 2 model.html