
1680 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 3, NO. 5, SEPTEMBER 2004
where is the -space sampled
version of
, and is the -space
sampled version of
.
With the statistical properties of the discrete-time channel co-
efficients
and the additive noises , the MIMO
channel input–output can be fully characterized in the discrete-
time domain with high computational efficiency and no loss of
information. Details are given in Section III.
B. MIMO Channel Assumptions
We have two assumptions on the continuous-time physical
channel of wideband MIMO wireless systems.
Assumption 1: The
th subchannel of a
MIMO system is a wide-sense stationary uncorrelated scat-
tering (WSSUS) [18], [19] Rayleigh fading channel with a zero
mean and autocorrelation given by
(7)
where
is the conjugate operator, is the maximum
Doppler frequency, and
is the power delay profile with
.
It is important to note that this assumption is commonly
employed for SISO channels in the literature [13], [15]–[17]
and in wireless standards documents [22], [23] for both
TDMA-based global system for mobile (GSM) communi-
cations, Enhanced Data rate for Global Evolution (EDGE)
systems, and code-division multiple-access (CDMA)-based
universal mobile telecommunications systems (UMTS) and
cdma2000 systems. Moreover, the power delay profile
is
often assumed to be discrete and is given by
(8)
where
is the number of total resolvable paths and is the
power of the
th path with delay . For example, the typical
urban (TU), hilly terrain (HT), and equalization test (EQ) pro-
files for GSM and EDGE systems [22] as well as the pedestrian
and vehicular profiles for channel A and channel B of cdma2000
and UMTS systems [23] have all been defined as discrete de-
layed Rayleigh fading paths, and almost all the path delays
are not an integer multiple of their system’s symbol period
(or chip period for CDMA systems).
This assumption implies that the fades of all the subchannels
are identically distributed. However, it does not require them to
be statistically independent. This implies that the subchannels
are not necessarily i.i.d., which was commonly assumed in the
literature for MIMO channels.
Assumption 2: The space selectivity or (spatial correlation)
between the
th subchannel and the th
subchannel
is given by
(9)
where
is the receive correlation coefficient between re-
ceive antennas
and with , and
is the transmit correlation coefficient between transmit
antennas
and with .
Assumption 2 is a straightforward extension of the MIMO
Rayleigh flat fading case in [24] to the MIMO WSSUS mul-
tipath Rayleigh fading case. It implies three subassumptions as
explained in [24] and cited as follows: 1) the transmit correlation
between the fading from transmit antennas
and to the same
receive antenna does not depend on the receive antenna; 2) the
receive correlation between the fading from a transmit antenna
to receive antennas
and does not depend on the transmit
antenna; and 3) the correlation between the fading of two dis-
tinct transmit-receive antenna pairs is the product of the cor-
responding transmit correlation and receive correlation. These
three subassumptions are actually the “Kronecker correlation”
assumption used in the literature, and they are quite accurate and
commonly used for MIMO Rayleigh fading channels [4], [25].
However, it should be pointed out that the third subassumption
may not be extended to Rice fading MIMO channels [26].
It is noted here that the spatial correlation coefficients
and are determined by the spatial arrangements of the
transmit and receive antennas, and the angle of arrival, the an-
gular spread, etc. They can be calculated by mathematical for-
mulas [4], [25] or obtained from experimental data.
III. D
ISCRETE-TIME MIMO C
HANNEL MODEL
In this section, we present a discrete-time model for triply se-
lective MIMO Rayleigh fading channels, then we investigate the
statistical properties of this MIMO channel in the discrete-time
domain. These statistics are further used to build a computation-
ally efficient discrete-time MIMO channel simulator, which is
equivalent to its counterpart in the continuous-time domain in
terms of various statistic measures.
A. Discrete-Time Channel Model
It is known that the total number of
-spaced discrete-time
channel coefficients
is determined by the maximum
delay spread of the physical fading channel
and the
time durations of the transmit filter and receive filter, which
are usually infinite in theory to maintain limited frequency
bandwidth. Therefore,
is normally a time-varying
noncausal filter with infinite impulse response (IIR).
1
However,
in practice, the time-domain tails of the transmit and receive
filters are designed to fall off rapidly. Thus, the amplitudes of
the channel coefficients
will decrease quickly with
increasing
. When the power (or squared amplitude) of a
coefficient is smaller than a predefined threshold, for example,
0.01% of the total power of its corresponding subchannel, it
has very little impact on the output signal and thus can be
discarded. Therefore, the time-varying noncausal IIR channel
can be truncated to a finite impulse response (FIR) channel.
Without loss of generality, we assume that the coefficient index
is in the range of , where and are nonnegative
1
It should be noted that the noncausality of
h
(
k;l
)
is induced by
the effects of the transmit filter and receive filter, while the physical CIR
g
(
t;
)
is always causal.