Useful Data Model for Antenna Array Signal Processing
Contributed by Haibing Yang, inspired by lecture notes taken
from Alle-Jan van der Veen
The properties of the physical channel
for the communication system can be described by a channel model with a large
set of parameters. Though useful for generating simulated data, it is not always
a suitable model for the purposes of signal processing. For example, if the
angle spreads within a cluster are large, thus the parametrization in terms
of directions is not possible. Hence, some simplified data models may be meaningful
and developed for the purpose of further array
signal processing. Here two simplified data models are described:
I-MIMO model and FIR-MIMO model.
Instantaneous Multi-Input Multi-Output model (I-MIMO) is a generic
model for source separation, which is valid only when the
delay spread of the dominant rays is much smaller than the inverse bandwidth
of the signals, i.e., for narrowband signals, in line-of-sight situations or in
scenarios where there is only local scattering.
In this model, it's assumed that d independent source signals s1(t),
···, sd(t) are transmitted
from d independent sources at different locations. If the delay spread
is small, then according to the vector
channel model for antenna array, the noise-free output signal of M
antenna elements in the receiver is a simple linear combinations of these signals:
|
x(t)=a1(t)s1(t)+···+adsd(t), |
which can be rewritten in matrix form:
|
x(t)=As(t), A=[ |
|
], s(t)= |
é
ê
ê
ê
ê
ë |
|
ù
ú
ú
ú
ú
û |
|
Suppose we sample with a period T and set T=1. If N samples
are considered, the data matrix is written as
where
and
This data model is called I-MIMO model.
For the multipath propagation model, each
column vector ai in matrix A
is the summation of the direction vector a(qij)
by different fading amplitude hij,
i.e.,
|
a |
i= |
|
a(qij)hij=[ |
|
] |
é
ê
ê
ê
ê
ê
ë |
|
ù
ú
ú
ú
ú
ú
û |
|
where ri is the number of rays for
the i-th source.
To get a compact structure, we define
|
|
= |
|
|
|
B |
= |
|
|
|
J |
= |
é
ê
ê
ê
ê
ê
ê
ë |
|
ù
ú
ú
ú
ú
ú
ú
û |
|
|
(1) |
where r=åi=1dri
is the total number of d sources and 1m
denotes an m× 1 vector consisting of 1's. Finally, the noise-free I-MIMO
data model can be written as
Finite impulse response multi-input multi-output data model (FIR-MIMO)
assumes that the channel impulse
responses can be simply described by the multichannel finite impulse response
(FIR) filters, which are time invariant. The derivation of FIR-MIMO data model
is described as follows.
Assume again there are d independent source signals s1(t),
···, sd(t) and sj(t)
is defined as the combination of infinite symbols, i.e.,
where g(t) represents the pulse shape function and the symbol period
T is set to 1 for the convenience of notation. If they are received by
M antennas through a convolutive channel, then the received signal without
noise is
where `*' denotes the operation of convolution, and
|
x(t)= |
é
ê
ê
ê
ê
ë |
|
ù
ú
ú
ú
ú
û |
, H(t)= |
é
ê
ê
ê
ê
ë |
h11(t) |
··· |
h1d(t) |
·
·
· |
|
·
·
· |
hM1(t) |
··· |
hMd(t) |
|
ù
ú
ú
ú
ú
û |
, s(t)= |
é
ê
ê
ê
ê
ë |
|
ù
ú
ú
ú
ú
û |
. |
|
Here it's assumed that each channel hij(t)
associated to the j-th source and the i-th antenna is a FIR filter
with integer length at most Lj symbols,
i.e.,
|
hij(t)=0,
tÏ[0,Lj)
(i=1, ···, M; j=1, ···, d). |
With the assumption above, we notice that there are at most Lj
consecutive symbols of signal sj(t)
take effects in x(t) at any given moment. The maximal channel length
among all sources is denoted by L=max Lj
and for the simplicity of the exposition, we set the length of all the channels
as L.
Suppose we sample the signal xi(t)
during N symbol periods, at a rate of P times the symbol rate, then
the data matrix X collecting all samples can be written as
|
X=[ |
|
]= |
é
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ë |
x(0) |
x(1) |
··· |
x(N-1) |
|
|
|
· |
·
·
· |
|
|
·
·
· |
|
· |
··· |
|
|
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
û |
|
|
where the vector xk contains the
MP spatial and temporal samples taken during the k-th symbol interval.
Based on the FIR assumption above, the data matrix X has a factorization
where the sampled channel matrix
|
H |
= |
é
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ë |
H(0) |
H(1) |
··· |
H(L-1) |
|
· |
|
· |
·
·
· |
|
|
·
·
· |
|
· |
··· |
|
|
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
û |
|
, |
|
(2) |
and the transmitted symbol matrix
|
SL |
= |
é
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ë |
s0 |
s1 |
·
·
· |
sN-2 |
sN-1 |
s-1 |
s0 |
·
·
· |
·
·
· |
sN-2 |
·
·
· |
·
·
· |
·
·
· |
·
·
· |
·
·
· |
s-L+1 |
·
·
· |
·
·
· |
·
·
· |
sN-L |
|
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
û |
|
|
(3) |
where the data vector si=s(i)
is d-dimensional. We observe that the transmitted symbol matrix
SL has a block-Toeplitz structure,
which is a consequence of the time invariance of the channel. (Note: The channel
matrix H can be rearranged into the discrete channel matrix and they are equivalent).
By exploiting the data matrix above, a linear equalizer can be realized by multiplying
a vector w to generate the output y=wHX.
More generally, to develop a space-time equalizer with a length of m symbol
periods, we write
|
Xm= |
é
ê
ê
ê
ê
ê
ê
ê
ë |
x0 |
x1 |
··· |
xN-m |
x1 |
x2 |
··· |
·
·
· |
·
·
· |
·
·
· |
·
·
· |
xN-2 |
xm-1 |
xN-2 |
··· |
xN-1 |
|
ù
ú
ú
ú
ú
ú
ú
ú
û |
|
|
By using this data matrix, a general space-time linear equalizer can be written
as y=wH Xm,
which combines mP snapshots of M antennas. Furthermore, the data
matrix can be factorized into
|
Xm=
Hm
SL+m-1= |
é
ê
ê
ê
ê
ê
ë |
|
ù
ú
ú
ú
ú
ú
û |
é
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ê
ë |
sm-1 |
·
·
· |
sN-2 |
sN-1 |
·
·
· |
·
·
· |
·
·
· |
sN-2 |
s-L+2 |
s-L+3 |
·
·
· |
·
·
· |
s-L+1 |
s-L+2 |
·
·
· |
sN-L-m+1 |
|
ù
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
ú
û |
|
where Hm
has size mMP× d(L+m-1) and the m shifts of
H to the left are each over d positions. It's observed that
Hm has a block-Hankel structure
and SL-m+1
block Toeplitz matrix. A necessary condition for space-time equalization is that
the matrix Hm
is tall, which results in
where M is the number of antennas, P is the oversampling factor,
d is the source number, m is the considered symbol number and L
is the length of the channel impulse response.
3 Summation
Until now we have introduced two noise-free data models: I-MIMO and FIR MIMO.
The I-MIMO model is generally valid for linear time-invariant channels, whereas
the FIR MIMO modle is the consequence of the vector
channel model. These models have some structural properties and are very
suitable for the purpose of array signal processing. More detailed introduction
for these data models can be found in [1], [2], [3].
[1] A.J. Paulraj and C.B. Papadias, "Space-time processing for
wireless communications," IEEE Signal Proc. Mag., vol. 14, pp.
49-83, Nov. 1997.
[2] B. Ottersten, "Array processing for wireless communications,"
in Proc. IEEE workshop on Stat. Signal Array Proc., (Corfu), pp.
466-473, June 1996.
[3] H. Liu and G.Xu, "Multiuser blind channel estimation and spatial channel
pre-equalization," in Proc. IEEE ICASSP, (Detroit), pp. 1756-1759
vol. 3, May 1995.