=
=
L
l
l l
H H
1
) ( ) ( (3)
where L is the number of taps (time bins) of the channel model, ) ( H is the MxN
matrix of the channel impulse responses.
n
s
Tx
Processin
g
h
1,1
h
1,2
Rx
Processin
g
y
h
2,2
h
2,1
Channel
H
12
For a 2x2 MIMO system the channel matrix is
MxN
C H ) (
(
=
22 21
12 11
h h
h h
H (4)
MxN
MN
l
l
H ] [ = is a complex matrix which describes the linear transformation
between two considered antenna arrays at delay
l
and
l
MN
is the complex
transmission coefficient from antenna M at the transmitter to antenna N at the
receiver. The complex transmission coefficients are assumed to be zero mean
complex Gaussian and have the same average power
l
p . The coefficients are
independent from one time delay to another. The correlation between different pairs
of the complex transmission coefficients is presented in Section 3.5. For all models,
both physical and nonphysical, the correlation coefficients are computed using
mathematical formulas from Appendix C.
The models presented in this diploma work are classified in different ways. But before
explaining model structure, the reader should have some knowledge of different
classifications in the area of channel modeling.
Wideband Models vs. Narrowband Models: the MIMO channel models can be divided
into the wideband models and the narrowband models directly by considering the
bandwidth of the system. The wideband models treat the propagation channel as
frequency selective, which means that different frequency subchannels have different
channel response. On the other hand, the narrowband models assume that the channel
has frequency nonselective fading and therefore the channel has the same response
over the entire system bandwidth.
Field Measurements vs. Scatter Models: to model the MIMO channel, one approach
is to measure the MIMO channel responses through field measurements. Some
important characteristics of the MIMO channel can be obtained by investigating the
recorded data and the MIMO channel model can be modelled to have similar
characteristics. Models based on MIMO channel measurements were reported in [4].
An alternative approach is to postulate a model (usually involving distributed scatters)
that attempts to capture the channel characteristics. Such a model can often illustrate
the essential characteristics of the MIMO channel as long as the constructed scattering
environment is reasonable. It is the environment of scatters that is in detail studied
here.
Nonphysical Models vs. Physical Models: the MIMO channel models can be divided
into the nonphysical and physical models. The nonphysical models describe MIMO
channel via statistical characteristics obtained from the measured data. Another
category is the physical models that are based on parameter setup and theoretical
results. In general, these models choose some crucial physical parameters to describe
the MIMO propagation channels. Some typical parameters include Angel of Arrival
(AoA), Angle of Departure (AoD), carrier frequency, antenna spacing.
13
3 Theory, Part I
In this Section, the summary of literature study is presented. The theory behind the
MIMO channel models is needed to understand MIMO channel properties and Matlab
program [3]. The theory section is also needed to be able to implement MIMO
channel matrix H into ADS (see Section 4) and to be able to compare theoretical
results against measured and collected data (see Section 5).
3.1 Physical MIMO channel model
A primitive physical channel model is illustrated in Figure 2. A major characteristic
of this model is that it does not rely on a geometrical description of the environment
under study. It is described by power profiles, spatial correlation functions and fading
characteristics.
Figure 2. Primitive channel model
The main input parameters required for the model are the shape of the power delay
spectrum, the fading characteristics and the spatial correlation functions at the
transmitter and receiver ends, explained further down in this section.
In order to capture path correlation, a general abstract scattering model is considered,
based on multicluster model. All the scatters are divided into groups, which are
called clusters of scatters. Each of the cluster corresponds to one multipath, see Figure
3. This model represents the physical model D which is further depicted in Figure 5 of
Section 3.4.
FIR filter
(L taps)
Fading characteristics
Spatial correlation
mapping matrix
Steering matrix
Power delay
profile
R
Rx
R
Tx
Radiation
Pattern
Tx
Rx
S/P P/S
14
Figure 3. Scatter cluster example
In Figure 3, the scatters are represented with arrival times. It can be seen that the
scatters which are closer to Rx have faster arrival time compared to those that are
placed far away. This artefact of motion is further studied in Section 3.8 as Doppler
shift.
There has been a use of AF models, presented in [5]. These models are representative
for small environments, such as residential homes and small offices, i.e. indoor
environments. Model F represents larger space either indoor or outdoor. Appendix A
lists the tables of all 6 models (AF) with its model tap delays, corresponding power,
Azimuth Spread (AS), Angle of Arrival/Departure (AoA/AoD).
Table 1 summarizes the channel model parameters.
Model
Environment
LOS/NLOS
K (dB)
RMS delay
spread (ns)
Number
of
clusters
A Flat fading NLOS  0 1 tap
B Residential LOS/NLOS 10
(first tap
only)
15 2
C Residential/
Small Office
LOS/NLOS 3
(first tap
only)
30 2
D Typical Office NLOS  50 3
E Large Office NLOS  100 4
F Large Space/
(Indoors and
Outdoors)
NLOS  150 6
Table 1. Channel model parameters
LO
Cluster 2
Tx Antennas
Rx Antennas
Cluster 3
Cluster 1
Arrival Times
t
15
3.2 Nonphysical MIMO channel model
A statistical model for wireless indoor MIMO channel is conducted at Victoria
University in Melbourne by Jason Gao and Michael Faulkner. The measured data is
characterised and presented as a nonphysical model with scatter and wideband
assumptions considering NLOS Rayleigh fading channel.
The nonphysical models provide accurate channel characterization for the
environments under study. On the other hand, they give limited insight to the
propagation characteristics of the MIMO channel and depend on the measurement
equipment, e.g. the bandwidth, the configuration and aperture of the arrays, the height
and response of transmit and receive antennas in the measurement.
In order to capture path correlation, a general abstract scattering model is assumed,
based on onering model. This model was first employed by [6]. It is an abstract
model, which is nonsitespecific and mathematically derivable. Yet, it captures some
of the key aspects of the environment under investigation.
Figure 4. Onering model
A circular disc (with radius R ) of uniformly distributed scatters S is placed around the
mobile unit. The channel parameter h_{NM } connecting transmit element M and
receive element N is geometrically constraint. The base station (BS) is assumed to be
elevated and therefore not obstructed by local scattering, while the mobile station
(MS) is surrounded by scatters. Figure 4 illustrates this scenario where Tx is antenna
element at the BS, Rx is the antenna element at the MS. D is the distance between the
BS and MS. R is the radius of the ring of scatters, is the AoA at the BS, is the AS
at BS. Denote the effective scatter on the ring by S() and let be the angle between
the scatter and the array at the MS. In the model, it is assumed that S() is uniformly
distributed over all angles and i.i.d. . It is further assumed that each ray is reflected
only once and that all rays reach the receiver array with the same power [6]. For this
reason, the channel coefficients are modelled by a zero mean complex Gaussian
random variable.
The path correlation between elements of transmit and receive antenna arrays is
considered and determined by antenna spacing, angle of departure/arrival
(AoD/AoA). When determining the delay power profile the SalehValenzuelas model
[7] was considered, see Section 3.4. The Doppler effects are not considered. As such,
the model is applicable to situations where during data transmission, the transmission
channels can be assumed stationary in time domain. Appendix B lists onering model
BS MS
16
with 6 clusters and data of Azimuth Spread (AS) and Angle of Arrival/Departure
(AoA/AoD) as a uniform distribution over all angles.
Following sections characterize the properties of both physical and nonphysical
models. Their mathematical derivation is the same, except that onering model does
not include Doppler effect.
3.3 MIMO Matrix Formulation
The MIMO channel matrix H for each tap, at one instance of time, in the AF delay
profile models can be separated into a fixed (constant, LOS) matrix and a variable
Rayleigh matrix. For the case of the onering model there is only a part of the
Rayleigh matrix since LOS component is not included.
For 2 x 2 MIMO system, the channel matrix H [5] is:
=


.

\

+
+
+
=
v F
H
K
H
K
K
P H
1
1
1
(5)




.

\

(
(
(
(
+
+
(
(
(
(
+
=
22 21
12 11
2
1
2
1
1
1
10
1
10
1
1
22 21
12 11
X X
X X
K
e e
e e
K
K
P
j j
j j
where X
NM
(Nth receiving and Mth transmitting antenna) are correlated zeromean,
unit variance, complex Gaussian random variables as coefficients of the variable
(Rayleigh) matrix H
v
, exp(j
NM
) are the elements of fixed matrix H
F
, K is the Rician
Kfactor, and P is the power of each tap.
To correlate the X
NM
elements of the matrix X [5], the Kronecker product of the
transmit and receive correlation matrices is preformed:
      { } 
iid Rx Tx
H R R X
2 / 1
= (6)
where R
Tx
and R
Rx
are the receive and transmit correlation matrices, respectively, and
H
iid
is a vector (only here, otherwise it is a matrix) of independent zero mean, unit
variance, complex Gaussian random variables, and
   
   
RxMN Rx
TxNM Tx
R
R
=
= (7)
where
TxNM
are the complex correlation coefficients between Nth and Mth
transmitting antennas, and
RxMN
are the complex correlation coefficients between M
th and Nth receiving antennas. Section 3.5 describes in detail the relationship of
correlation coefficients.
17
3.4 Power Delay profile
When determining the power delay profile (PDP) the SalehValenzuelas model was
used [7]. This model is based on indoor measurement results where it was found that
received signal rays (due to multipath) arrive in clusters.
The mathematical representation of the received signal amplitude
kl
is a Rayleigh
distributed random variable with a meansquare value that obeys a double exponential
decay law
/ / 2 2
) 0 , 0 (
kl l
e e
T
kl
= (8)
where ) 0 , 0 (
2
represents the average power of the first arrival of the first cluster, T
l
represents the arrival time of the l
th
cluster, and
kl
is the arrival time of the k
th
arrival
within the l
th
cluster, relative to T
l
. The parameters and determine the intercluster
signal level rate of decay and the intracluster rate of decay, respectively. The rates of
the cluster and ray arrivals can be determined using exponential rate laws
) (
1
1
)  (
=
l l
T T
l l
e T T p (9)
) (
, 1
1
)  (
=
l l
T T
l k kl
e p
(10)
where is the cluster arrival rate and is the ray arrival rate.
Figure 5 shows Model D delay profile with clusters outlined by exponential decay
(straight line on a logscale)., see Appendix A.
Figure 5. Model D, Power Delay Profile
Source: V. Erceg, Indoor MIMO WLAN Channel Models, 2003.
50 0 5
10
15
20
25
30
35
40
0
5
1
1
2
2
3
Delay in nanosec.
dB
Cluster 1
Cluster 2
Cluster 3
18
3.5 Correlation functions and PAS distribution
A crosscorrelation function between the waves impinging on two antenna elements
mostly depends on the Power Azimuth Spectrum (PAS), the antenna spacing and on
the radiation pattern of the antenna elements. In the following, MIMO channel model
is presented with the uniform linear array (ULA) and with the omnidirectional
antenna elements.
The PAS distribution in physical models, consider three types of different
distributions, Uniform [7][8], Gaussian [7][8] and Laplacian [9], see Appendix C.
The nonphysical model (onering model) is modelled using only Laplacian PAS
distribution.
For these three distributions, the envelope correlation coefficient is computed as a
function of the normalised distance, using the angel of incidence and the Azimuth
Spread (AS) as indexing variables. All three distributions consider multicluster
model. Which makes them valid to use for all models considered here. It is assumed
that the spatial correlation function at the Rx is independent of n. This is reasonable
assumption provided that all antennas at the Tx are closely colocated and have the
same radiation pattern, so they illuminate the same surrounding scatters and therefore
also generate the same PAS at the Rx, i.e. the same spatial correlation function.
Figure 6. Impinging waves from Tx to Rx
Using the notations of [9], with
d
standing for the normalised distance between
elements, where d is the element spacing and the wavelength, and D=2
d
, one can
easily derive the crosscorrelation function between the real and imaginary parts of
the complex baseband signals received at two omnidirectional antennas separated by
the distance d.
For ULA [7, 8] the complex correlation coefficient at the linear antenna array is
expressed as field
f
(D) and envelope
e
(D):
2
2
) ( ) ( ) ( ) ( D jR D R D D
XY XX f e
+ = = (11)
where / 2 d D = , and R
XX
and R
XY
are the crosscorrelation functions between the
real parts (equal to the crosscorrelation function between the imaginary parts) and
between the real part and imaginary part, respectively, with
Dr = D sin
D = 2d/
s
s
Rx
Tx
19
d PAS D D
XX
R ) ( ) sin cos( ) ( (12)
and
d PAS D D
XY
R ) ( ) sin sin( ) ( (13)
The correlation properties of the fading, and the values of the symmetrical correlation
matrices R
Tx
and R
Rx
, are completely defined by the PAS and its standard deviation.
For the 2 x 2 MIMO channel, transmit and receive correlation matrices are expressed
as:
(
=
1
* 1
21
12
Tx
Tx
Tx
R
and
(
=
1
* 1
21
12
Rx
Rx
Rx
R
(14)
Further, the spatial correlation properties of a MIMO system uses Kronecker product
of the spatial correlation matrices R
Tx
and R
Rx
to define total correlation
1
) ( ) (
Rx Tx
H
R R H vec H vec R = = (15)
whose elements are correlation coefficients, where (.)
H
is the conjugate transpose and
the vec(.) operator rearranges the 2 x 2 matrix H into a column vector of size 4 x 1.
Such a model has been experimentally validated in [12].
Cluster and tap PAS shape follow Laplacian distribution [9].
The angle of arrival statistics within a cluster were found to closely match the
Laplacian distribution [13, 14, 15]
/ 2
2
1
) (
= e p (16)
where is the standard deviation of the PAS (which corresponds to the numerical
value of AS). The Laplacian distribution is shown in Figure 7 [14] (a typical
simulated distribution within a cluster, with AS = 30
o
).
1
The structure of the Kronecker product depends whether one wants to simulate a downlink
transmission (as presented here) or uplink
Tx Rx
R R R =
20
Figure 7. Laplacian distribution with AS=30 [14]
The Laplacian function exhibits the sharp peak in the LOS direction and is confined
within [180, 180].
For the case of the physical models, it was found in [13, 14] that the cluster mean
AoA and AoD have a random uniform distribution over all angles [0, 2]. Due to the
central limit theorem, when the number of scatters becomes large, the channel
coefficient of matrix H is Gaussian distributed.
For the case of the onering model, the references [9, 10] show that the AS increases
with decreasing distance between BS and MS, provided that this distance is still much
greater than the radius of the circle within the scatters surrounding the MS are placed,
and that the assumption of scatters distributed around BS holds. Consequently, for a
same element separation distance D, BS elements are less correlated than MS once as
they experience a greater AS [10]. On the other hand, it is shown in [9, 11] that the
height of the MS has also an influence on the AS, as the spreading increases with
decreasing antenna height.
3.6 Tap Time and Angle Dependence
The channel impulse response as a function of time and angle is a separable function,
h, [16, 17]
) ( ) ( ) , ( h t h t h = (17)
where t is time and is an angle.
21
3.7 Fading Multipath Channels
Multipath fading occurs when the transmitted signal reaches the receiver via multiple
paths with different delays and attenuation (or amplification).
Characteristics of a multipath medium is:
Time spread introduced in the signal that is transmitted through the channel, and
Time variations in the structure of the medium.
Multipath fading is broadly divided into two categories: large (or slow) and small (or
fast) scale fading. Large scale fading refers to mean path loss averaged over several
wavelengths, whereas small scale fading refers to dramatic changes in amplitude and
phase when the receiver moves by as little as half a wavelength.
There are several probability distributions that can be considered in attempting to
model the statistical characteristics of the fading channel. When there are a large
number of scatters in the channel that contribute to the signal at the receiver, as is the
case in BF models and onering model, application of the central limit theorem leads
to a Gaussian process model for the channel impulse response. It is assumed that the
process is zeromean, that gives, the envelope of the channel response at any time
instant has a Rayleigh probability distribution and the phase is uniformly distributed
in the interval (0, 2). That is
2
2
0
2
2
0
0
) (
r
e
r
r p
= for 0
0
r (25)
where r
0
is the magnitude of the multipath signal.
Rayleigh fading is considered a worstcase scenario of fading.
In the frequency domain, the reciprocal of the multipath spread is a measure of the
coherence bandwidth of the channel. It is the range of the frequencies within which
the signal impairments of the channel done not vary significantly, see Figure 8 c). In
other words, if two signals are sent that are more than f
d
apart form each other in
frequency, they will experience different channel conditions. Coherence bandwidth is
denoted as
m
d
T
f
1
(26)
where f
d
denotes coherence bandwidth and T
m
is delaytime of the multipath spread.
When a signal is transmitted through the channel, if f
d
is small in comparison to the
bandwidth of the transmitted signal, the channel is said to be frequencyselective. In
this case, the signal is severely distorted by the channel and wideband signal could see
a large variation in received power over its bandwidth. On the other hand, if f
d
is large
in comparison with the bandwidth of the transmitted signal, the channel is said to be
frequencynonselective.
22
In the time domain, the time variations are evidenced as a Doppler broadening, further
explained in Section 3.8. A slow changing channel (slow fading) has large coherence
time T
d
or, equivalently, a small Doppler spread, see Figure 8 a). T
d
denotes the time
period over which the channels impulse response is highly correlated.
d
d
B
T
1
(27)
where B
d
is a Doppler spread of the channel.
Slow fading is desired fading since the channel conditions are stable and predictable
during the time that the symbol is transmitted. Counter to the slow fading there is fast
fading where fading conditions are coming and going faster than the symbols are
being transmitted.
The fading characteristics are defined by shaping a Doppler spectrum, correlation
functions and multipath intensity profile. Figure 8 below, illustrates relationship
between Doppler spectrum and its Spacedtime correlation function, Scattering
function of the channel and Multipath intensity profile. Scattering function in Figure 8
b) provides a measure of the average power output of the channel as a function of the
time delay and the Doppler frequency .
Figure 8. Relationship among the correlation functions and power spectra
Source: John G. Proakis, Digital Communications, 2001
Multipath intensity profile Doppler power spectrum
a) b) c)
23
For the physical models AF following is valid: In order to avoid frequency selective
fading, the transmission rate is set to be less than the coherence bandwidth of the
channel. And in order to reduce the distortion caused by fast fading, it is important to
set the transmission rate to be more than the channelfading rate.
The simulations results of the fading are only presented for the case of the frequency
nonselective and slow fading, see Section 3.9. In WLAN technology, the OFDM
modulation is used for IEEE 802.11a standard. OFDM avoids frequency selective
fading by breaking the carrier signal into subcarriers with lower bit rates and thereby
longer symbol duration. The simulation results are presented in Section 3.9.
24
3.8 Doppler Spectrum
Assumption in this thesis was made as; both transmitter and receiver are stationary.
Many may think that the impact of Doppler spectrum should not then have any impact
on the channel model. However there are many clusters and their scatter rays in the
channel. Their motion, reflection and diffraction give rise to the time variant nature of
the channel.
The expression of the Doppler spectrum has been derived by Clarke [18]. It is well
known that the Doppler spectrum lies within a [f
D
, f
D
] bandwidth, where f
D
is the
maximal Doppler shift/spread, defined as
v
f
D
= (18)
where v stands for the velocity of the movement. (f
D
should not be confused with f
d
which is coherence bandwidth of the channel in frequency domain). Time variations
of the channel are evidenced as a Doppler shift of a spectral line. The metric used to
measure the impairment caused by the time variant nature of channel is how rapidly
the channel fades.
The main parameter of the Doppler spectrum is heavily depending on simulation
assumptions. On the other hand, in the physical models the shape of the Doppler
spectrum can be derived knowing the PAS and the radiation pattern of the receiving
antenna [19]. Then the classical Ushaped, see Figure 9 below, Clarkes Doppler
spectrum expression is:
2
1
1 ) (
(


.

\

D
f
f
f P (19)
This formula (19) applies to all physical models, AE, as both communication ends
are surrounded by scatters.
Figure 9. Doppler spread/shift
Source: John G. Proakis, Digital Communications, 2001
The Doppler power spectrum is plotted above in the Figure 9 where
B D m
f f f = (f
D
Doppler spread and f
B
spectral broadening).
25
3.8.1 Power Delay Spectrum
The Power Delay Spectrum (PDS) has been widely studied as part of the timedomain
characterization of wireless radio channels. In accordance with [20], the PDS is
accurately modeled by a onesided exponential decaying function
=
0
) (
D
t
e
t P
t>0 (20)
where
D
represents the delay spectrum (DS). Values of DS have been proposed for
the different environments [12]. For our models, the interesting value is for the indoor
environments, which value lies between 35100ns [17].
3.8.2 Fluorescent Lights
Effects of fluorescent lights on signal fading characteristics for indoor channel were
presented in [21]. The presence of fluorescent lamps creates an environment where
reflections are being introduced, thus creating a fast changing electromagnetic
environment. This is an effect that can yield significant variations of received signal
power. This effect is included in physical models D and E by modulating several taps
in order to artificially introduce an AM modulation by
( ) { }
2
0
exp (4 (2 1) )
l m l
l
g t A j l f t
=
= + +
(21)
where
( ) g t The modulating function
l
A Relative harmonic amplitudes
m
f The main AC frequency
l
A series of i.i.d. phase RV's ~ [0, 2 ) U
t Time
The interferer to carrier energy ratio is selected using the following random variable:
2
I
X
C
= (22)
where
2
~ (0.0203,0.0107 ) X G (23)
In equation (23), the first figure is the mean of the Gaussian, and the second is the
variance (the standard deviation squared).
otherwise
26
Figure 10 shows the cumulative distribution function (CDF) of the modelled I/C
(interferencetocarrier ration) in green, and the measured experimental results in
blue. This plot shows good agreement with the measured I/C.
Figure 10. CDF of modelled I/C vs. measured I/C.
Source: V. Erceg, Indoor MIMO WLAN Channel Models, 2003
For the models D and E, 3 taps are being modulated by the modulating function ( ) g t
in accordance with the drawn I/C. The time value of each one of these coefficients is
as follows:
'( ) ( )(1 ( )) c t c t g t = + (24)
where
( ) c t Original tap value
'( ) c t Modified tap value
( ) g t The modulating function
Normalization constant
The value of is determined such that the total modulation energy (modulation in
the modulated taps, compared to the entire channel response) matches the drawn
random I/C. The following taps are modulated
Model Cluster Tap numbers
D 2 2,4,6
E 1 3,5,7
Table 2. Tap modulation of models D and E
27
3.9 MIMO channel properties
In this section, MIMO channel properties (explained in Section 3.13.8) are
represented using the graphical plots of Matlab program distributed and written by L.
Schumacher (together with AAUCsys, FUNDINFO, and project IST200030148 I
METRA) and is publicly available, see [3].
Before illustrating MIMO properties such as channel impulse response, power
profiles, correlation coefficient and Doppler spectra, the capacity equation for MIMO
channels is presented. Note that the simulation results only consider a physically
based channel model C. However, the simulation results of nonphysical model are
presented and compared to the applicable physical model in Section 5.
3.9.1 Capacity
The capacity of a channel depends completely on the channel realization, noise, and
transmitted signal power. Capacity equation is [1] [2]:
(
+ = ) det( log
*
2
N
SNR
C
M EP
b/s/Hz (28)
where (*) means transposeconjugate, H is the MxN channel matrix and I
M
is un
identity matrix. In the case of two transmitter antennas and two receiver antennas the
matrix of channel is 2x2 rank.
Capacity grows linearly with m=min (M,N) [12] rather than logarithmically as was
described by Foschini [2] and Telatar [1].
The capacity gain is highly dependent on the multipath richness in the radio channel,
since a fully correlated MIMO channel only offers one subchannel, while a
completely decorrelated channel offers multiple subchannels, depending on the
antenna configuration. In this thesis the physical models are assumed partially
correlated/decorrelated channels, since that is the case in practice.
It has been demonstrated that increasing the number of antennas in the both ends
results in a rapid increase in theoretical capacity [12].
28
3.9.2 Simulation results
In this section the simulation results of MIMO channel are presented. For the sake of
simplicity, only simulation graphs of the model case C is included. This simulation
results, of Matlab program [3], have been used to verify the implementation results in
Section 4.
There are 5 types of graphs included to illustrate characteristics of the MIMO
channel. The plots represent impulse response of the channel matrix H, Power Delay
profile (PDP) of the H channel matrix, CDF of the taps, spatial correlation functions
for number of paths and Doppler spectra. In the Section 5 this simulated
characteristics are compared to measured data conducted at Victoria University in
Melbourne, Australia. While here the comparison has been done to the literature
references, either the Rayleigh distribution (CDF) or the desired curve (PDP,
correlation and Doppler spectrum).
The following pages present the plots of 2x2 MIMO channel, model C, using the
Matlab program [3]. The carrier frequency used is 5.25 GHz (f
d
= 6 Hz) with one
wavelength (= 6 cm) spacing at the transmitter and half wavelength spacing at the
receiver. 16 384 blocks of 32 samples have been simulated at 0 speed (note:
Stationary Receiver), and a sample has been stored per block. Dashed red
curves/markers correspond to the reference values, whereas the blue curves/markers
are the outcome of the simulation. In the Doppler plot, the green curve represents the
Welch periodogram [22, p. 256].
The match between reference curves and simulation results using Matlab program [3]
is satisfactory. The achieved tap power distribution in Figure 11 b) fit the PDPs
defined as Rayleigh distribution in [23].
In Figure 12, the spatial correlation coefficient of the simulated impulse response
match the mathematical formulas presented in Section 3.5.
In Figure 13 and 14, the red vertical lines are drawn at
d
f . The upper blue line is set
at the maximum of Doppler spectrum, and lower blue line lies 10 dB below. Ideally,
the Doppler spectrum should meet the crossing of the red and blue lines. This would
be the case, if the jitter would be removed from the sampled spectra presented in [24].
Figure 11. Channel model C: a) Impulse response, b) PDP
a) b)
29
Figure 12. Channel model C: a) CDF, b) Spatial correlation of the firs six taps
Figure 13. Channel model C: a) Spatial correlation of the six middle taps, b) Spatial
correlation of the last two taps
a) b)
a) b)
30
Figure 14. Channel model C: a) Doppler spectra of the first six taps, b) Doppler spectra of the
middle six taps
Figure 15. Channel model C: Doppler spectra of the last two taps
a) b)
31
4 Implementation results
In this section the implementation results of MIMO channel model are presented.
These results only consider implementation of physical models AF.
4.1 Background
Seen from Acreos point of view, there are two possible implementation tools: Matlab
6.5 and Advanced Design Systems (ADS). Both have its advantages and
disadvantages.
It is obvious that the easiest method to implement the channel model is to use Matlab
6.5, since the program written by L. Schumacher [3], is for Matlab. However, all other
work done on algorithms, transmission schemes and detection methods here at Acreo
is done in ADS. So, for Acreo it would be best to even have the channel model
implemented in ADS. For this solution there are three alternatives. The first
alternative is to use a Matlab program written by L. Schumacher [3] and translate it
into the C++ programmable language. The second alternative is to implement a
Matlab block which links between Matlab 6.5 and ADS. The third alternative is to use
the Matlab 6.5 that generates channel matrix H and save it as an ASCII file. This
ASCII file can then be read from ADS and used as a covariance channel matrix H.
4.2 Translation of Matlab code into C++
Matlab 6.5 Converter was used to translate the whole Matlab program [3] written by
L. Schumacher, which consisted of approx. 20 files each 12 A4 pages long, into C++
code. The translation went without any problems since behind Matlab code lies a C
language. The Matlab converter even linked the files and their functions together. To
be able to use C++ in ADS, the requirement is to build a specific structure of C++ and
use a Make files during the compilation. ADS Ptolemy manual was used and the
structure of C++ got a new appearance, see Figure 16.
After modifying the C++ codes, it was realized that calling the functions and their
parameters is not as smooth as in Matlab. In the Matlab program there is a main
function, which calls other functions together with their multiple parameters of 2D
and even 3D matrices. This main function generates the 4D channel matrix H that is
needed. However, to be able to generate the channel matrix H, the calculations of
fading matrices and spatial correlation needs to be done. The call of these could not
be done successfully in C++. This appeared to be very time consuming, instead some
other alternatives were studied, explained further down.
32
Ptolemy example of C++ code
defstar { // Body of the program
name {example_MIMO}
domain {SDF}
desc { MIMO channel models}
version {Source: Indoor MIMO WLAN channel model,
Date: 20030919}
author {Aida Botonji}
location {MIMO}
inmulti { // Defines multiple inputs
name {Fading_Type}
type {string}
desc {indicats the nature of the Doppler}
}
output { // Defines output
name {H}
type {FLOAT_MATRIX}
desc {4D matrix}
}
defstate { // Defines Parameters
name {ID}
type {enum}
default {"D"}
desc {IEEE 802.11 case to be simulated}
enumlist {A, B, C, D, E, F}
//////////Initialisation///////////////////////
setup { //Parameter conversion
} // end of setup
code {
//Initialisation of size variables
//Parameters assessment
//Largescale fading
//Setup of the iteration process
} //end of code
/////////Main loop//////////////////////////////
go { //Beginning of simulation
//Defines H channel matrix
}//end of go
////////End of simulation//////////////////////
wrapup {
} //end of wrapup
}//end of defstar
Figure 16. Ptolemy example of C++ code
33
4.3 ADS schematic with Matlab block
Following presents schematics from ADS for the second implementation alternative.
The implementation method described here is to link Matlab 6.5 together with ADS
using predefined ADS block. This means to have the Matlab program written by L.
Schumacher [3], as it is and use ADS block which links the Matlab 6.5 with the ADS
and generates desired channel matrix H, see Figure 17.
This implementation method had a big disadvantage. That is, the simulation took long
time, 30h. Even the generation of the matrices is not as smooth as in Matlab. In
comparison to Matlab where one could work with as big matrices as wonted, in ADS
this was not possible. ADS is limited to 3D matrices. This means that generation of
the channel matrix H, which is of size 4D, has been carried out with errors. The
schematics in Figure 17 generated H matrix without a imaginary part. This of course
does not give a right image of the channel matrix H which is a complex matrix.
Luckily an updated version of ADS is coming after New Years Eve 2003 and
improvements have been made on the simulation speed when using Matlab together
with ADS. Also, a generation of bigger matrices than 3D is going to be possible. This
schematic was left as it was and some other alternatives were explored, described
further down.
Figure 17 depicts ADS schematic. In the schematic, the Matlab block opens,
internally from ADS, a Matlab program written by L. Schumacher [3], and performs
all calculations using Matlab. Then it generates a desired output, in this case channel
matrix H, to ADS. The complex channel matrix H is stored in a block called
NumericSink. This block can show only data on the screen or it can plot desired
form of graphs, such as linear, Smithchart, and so on.
Each ADS schematic requires a controller called DF in to order perform
simulations. This block indicates the simulation setup. In reality this schematic
consists of 14 inputs. But here are only 4 input parameters illustrated, for the sake of
simplicity.
Figure 17. ADS with Matlab block
Matlab block
NumericSink
Controller
Input
parameters
34
4.4 ADS schematic with ReadFile block
Figure 18 depicts channel model implemented with delays and matrix H multiplied
together with incoming data, arbitrary bits. Before using this schematic the generation
of matrix H must be done in the Matlab 6.5 using the Matlab program written by L.
Schumacher [3]. The channel matrix H is then stored into the ASCII file. This ASCII
file is read from ADS using a ReadFile block. Channel matrix H is then multiplied
first with arbitrary bits and then with the power delay in order to separate data in time.
Figure 18. ADS with ReadFile block
Arbitrary bits
Delay Channel H
ReadFile
35
4.5 Channel model implemented in ADS
This section presents a channel model implemented in ADS.
Figure 19 illustrates the channel model with the channel matrix H and tap delays. This
model is based on the Jakes model from 1974 [6]. Here is matrix H multiplied with
tap delays explained in Section 3.4. The delay profile determines the frequency non
selective nature of the channel. Delay profile is taken directly from the table in
Appendix A and varies from model to model. Other parameters such as impact of
fading, tap power, correlation and generation of actual channel matrix H is done by
the Matlab program written by L. Schumacher [3]. When using this model for
simulations in ADS, the user is required to use one of the above two alternatives to
get the channel matrix H for the desired channel model.
This schematic (together with the second alternative of the generation of the channel
matrix H in Section 4.4) was used to confirm if the implementation is successful or
not. The comparison between the graphs that ADS generated and those that Matlab
6.5 generated in Section 3.9.2 were identical. This means that the MIMO channel
implemented in ADS gives satisfactory properties comparing to the theoretical
presentation of MIMO channel in Section 3.13.8.
The number of delays depends on number of taps a channel model has. When
multiplication is performed, the channel is summed and used as an input to the
antenna array at the receiver.
Figure 19. Channel model in ADS with channel matrix H and power delays
Delays
Channel H
Arbitrary bits
36
5 Comparison results
In this section the comparison results between the measured data conducted at
Victoria University in Melbourne by Jason Gao and Michael Falukner and theoretical
data generated by the Matlab program written by L. Schumacher [3] are presented.
When investigating and comparing the physical models against onering model the
conclusion is drawn that the best suitable model for measured data is a case F model
since model F includes 6 clusters, same as for onering model.
The following graphs consider only onering model compared against the F model.
The MIMO channel impulse responses are generated using Matlab 6.5. The Fourier
transform is used to transform the simulated impulse responses into the frequency
domain, see Figure 20 and 21.
Figure 20 represents impulse response simulated for case F channel model. The
impulse response is an averaged sum for absolute magnitude of channel matrix H (10

3
= 40 dB).
Figure 21 illustrates impulse response for the onering model. The graph represents
impulse response that is accounted for hardwareinduced errors, such as insertion loss
and leakage in the switches and cabling.
It can be seen that the magnitude for the onering model is much lower than for the
model F. This may be due to the interference of people moving around in the
corridors at the time when the data was collected. But also since impulse response for
model F is averaged over time while onering model is measured for a certain time
period.
Figure 20. Impulse response of covariance channel matrix H for channel model F
Impulse response of h
11
Impulse response of h
12
Impulse response of h
21
Impulse response of h
22
37
Figure 21. Impulse response and phase of onering model
It is known from theoretical study [25] that the lower the spatial correlation within the
indoor environment, the greater the achieved capacity. The spatial power correlation
coefficient for the onering is not treated at the BS since antennas separated with 1.5
is decorrelated. Model F assumes partly correlated Tx and Rx antennas because of 0.5
1 antenna separation. Antenna spacing on the order of 0.4 0.6 is adequate for
independent fading [26], see also Section 6.4.
Table 3 represents the spatial correlation coefficients for the channel model F and
onering model at the MS antennas. Numbers in brackets are theoretical results for the
channel model F while numbers without brackets are estimated correlation
coefficients for the onering model. It can be seen that deviation is marginal.
Cluster R
12

2
R
21

2
1 1.00 (1.00) 1.00 (1.00)
2 0.86 (0.82) 0.55 (0.54)
3 096 (0.97) 0.88 (0.90)
4 0.74 (0.78) 0.39 (0.45)
5 0.87 (0.90) 0.65 (0.69)
6 0.67 (0.68) 0.22 (0.25)
Table 3. Spatial correlation coefficients of model F and onering model
The Doppler power spectrum in the onering model is not included due to stationary
realization of MS and BS antennas. It follows that the comparison results does not
consider Doppler power spectrum.
The reader should notice that since the data presented here considers different
scenarios, the power correlation coefficients and impulse response vary. The data for
onering model is collected for smaller rooms and for a 12m long corridor. Model F
considers much bigger space than the space where onering model was measured.
There may be several reasons why the simulated and measured results do not match
perfectly: insufficient amount of statistics, Wide Sense Stationary (WSS) assumption
not 100% fulfilled, This needs further investigations.
38
6 Theory, Part II
This section presents the second part of literature study. It is explained how the
channel H is transmitted and received and how the algorithms for detection should
be developed.
Focus of MIMO systems is to combine the signals at the receiver in such a way that
the quality, bit error rate BER and data rate, is dramatically improved. To achieve
these improvements, both transmitter and receiver must be designed in a special way.
The transmission of the channel H at the transmitter and algorithms for detection and
decoding at the receiver are discussed in this section.
6.1 Spatial multiplexing scheme
The wireless channel constitutes a hostile propagation medium, which suffers from
fading (caused by destructive addition of multipath components) and interference
from other users. Diversity is a powerful technique to combat fading and interference.
Spatial diversity has become very popular in recent years since it can be provided
without loss in spectral efficiency, see section 6.4. Because of its popularity and
spectral efficiency, here is only transmission of spatial multiplexing considered.
Spatial multiplexing transmits independent parallel data streams through multiple
antennas at both transmitter and receiver, see Figure 22.
Figure 22. Spatial multiplexing for 2x2 MIMO system
In a spatial multiplexing system the data stream to be transmitted is demultiplexed
into lower rate streams which are then simultaneously sent from the transmitter
antennas after coding and modulation. As soon as signals leave antennas at the
transmitter they are mixed together in the wireless channel, since they use the same
frequency spectrum. Each receiver antenna observes a superposition of the
transmitted signals. The receiver then separates them into constituent data streams and
remultiplexes them to recover the original data steam. This occurs for 2x2 MIMO
system, as two unknowns are resolved from a linear system of two equations. Clearly,
the separation step determines the computational complexity of the receiver.
In the following, both transmitter and receiver architecture is discussed.
Tx1
Tx 2
Rx1
Rx2
Spatial
Demulti
plexing
Spatial
Multi
plexing
39
6.2 Transmitter
There are two alternatives for the transmitting channel. The first alternative is that the
transmitter has partial or full knowledge of the channel and uses this knowledge to
increase diversity gain. The second alternative is that the transmitter has no
knowledge of the channel and uses a coding technique to achieve diversity. In both
categories it is presumed that the receiver has knowledge of the channel in order to
separate data from the multiple spatial channels. This information is obtained by
sending training or pilot symbols to estimate the channel.
For the first alternative, the transmitter has to gain knowledge of the channel. There
are two possibilities, (1) knowledge of the channel at the receiver is turned around and
used at the transmitter (no feedback is necessary) or (2) knowledge of the channel at
the receiver is fed back to the transmitter using a handshaking protocol.
When the transmitter has no knowledge of the channel an algorithm should be
selected accordingly. A transmitter encodes the bits over space and frequency and
transmits these bits over multiple spatial channels. The receiver then separates the
symbols from the channel and decodes the bits, see Figure 23. In the following, it is
assumed that the transmitter has no knowledge of the channel since that is the usual
case.
Transmitter IFFT &
Add CP
802.11a Encoder Space Modulation Spatial
& puncturing Frequency (BPSK, QPSK Mapping
Interleaver 16&64QAM)
IFFT &
Add CP
Receiver OFDMModulator
FFT &
Remove
CP
Antenna Soft Space Depuncture
selection VBLAST decision Frequency &
FFT & Interleaver Decoder(V)
Remove
CP
OFDMDemodulator
Figure 23. MIMO system without knowledge of the channel at the transmitter
Source: Intel Technology Journal, HighThroughput Wireless LAN Air Interface
Interleaving and coding at the transmitter can be used to reduce the carriertonoise
(C/N) ration requirement for accurate detection. Coding provides redundancy by
sending multiple copies of the signal on orthogonal symbol paths. Interleaving
provides robustness to the link by spreading errors out in time, avoiding large
amounts of contiguous data loss that would cut a wireless link.
As illustrated in Figure 23, assuming N subcarriers (or tones) the individual data
streams are passed through OFDM modulator, which performs an IFFT on blocks of
length N followed by a parallel to serial conversion. A cyclic prefix (CP) containing a
40
copy of the parallel to serial converted output of the Npoint IFFT is then prepended.
The resulting OFDM symbols are launched simultaneously from the individual
transmitting antennas. The CP is essentially a guard interval which serves to eliminate
interference between OFDM symbols and turns linear convolution into circular
convolution such that the channel is diagonalized by the FFT. In the receiver the
individual signals are passed through OFDM demodulator which first discard the CP
and then perform an Npoint FFT. The outputs of the OFDM demodulator are finally
separated and decoded.
Transmitting and receiving diversity are both similar and different in many ways.
While receiving diversity needs merely multiple antennas which fade independently,
and is independent of coding/modulation schemes, transmitting diversity needs
special modulation/coding schemes in order to be effective. Also, receive diversity
provides array gain, whereas transmitting diversity does not provide array gain when
the channel is unknown in the transmitter.
6.3 Receiver
In the following a detection and decoding algorithm at the receiver are presented. The
receiver components are illustrated in the above Figure 23.
Detection algorithm, so called VBLAST, is most promising one to be used in future
realizations of MIMO systems. VBLAST does not jointly decode all the transmit
signals, it first decodes the strongest signal then it subtracts this strongest signal
from the received signal, proceed to decode the strongest signal of the remaining
transmit signal, and so on. The optimum detection order in such a nulling and
constellation strategy is from the strongest to the weakest signal. Assuming that the
channel H is known, the main steps of the VBLAST algorithm can be summarized as
follows:
Nulling: An estimate of the strongest transmit signal is obtained by nulling out
all the weaker transmit signals, say using zero forcing (ZF) criterion. ZF
basically inverts the channel transfer matrix H. Then the transmitted data
symbol vector s is obtained as r H s
1
= .
Slicing: The estimated signal is detected to obtain the data bits.
Cancellation: These data bits are remodulated and the channel is applied to
estimate its vector signal contribution at the receiver. The resulting vector is
then subtracted from the received signal vector and the algorithm returns to
the nulling step until all transmit signals are decoded.
For a more in depth treatment of the VBLAST algorithm the interested reader is
referred to [27, 28].
The optimum decoding method is maximum likelihood (ML) where the decoder
compares all possible combinations of symbols which could have been transmitted
with what is observed.
The ML decoder yields the best performance in terms of error rate. However, this
decoder also has the highest computational complexity which moreover exhibits
exponential growth in the number of transmit antennas. For 2x2 MIMO system ML
decoder works very well.
41
Received signal is expressed as:
n s H r + = (29)
where s is a codeword expressed in vector form and n is a noise.
Optimum ML decoding rule is:
=
=
M
m
m m
s
s H r s
1
2
=
)
`
+
=
c
N
k k k
k k
k U U
Q PAS
1 , 0
, 0
,
)] ( [
)] ( [
) (
where () is the step function and N
c
is the number of clusters.
The first step is to normalise the PAS such that it can be regarded as a probability distribution.
The constants Q
U,k
are derived such that PAS
U
() fulfils the requirements of a probability
distribution function:
=
+
= =
c
k mk
mk
N
k
k U U
d Q d PAS
1
,
0
0
1 ) (
where stands for the halfdomain definition of the PAS (domain assumed symmetric).
Derivation of probability distribution leads then to
=
=
c
N
k
k k U
Q
1
,
1 2
The crosscorrelation function for realreal and imaginaryimaginary parts are derived in the
same way:
] [ ] [ ) ( ) ( y y E x x E D R D R
YY XX
= = =
The crosscorrelation function between the realreal part is:
) 2 sin( ) 2 cos(
2
) (
4 ) ( ) ( ) sin cos( ) (
1
, 0
2
1
, 0 , k
m
k
m
N
k
k U U XX
m m
m
D J
Q D J d PAS D D R
c
+ = =
= =
where J
m
is the Bessel function of the first kind and m
th
order.
The crosscorrelation function between realimaginary part is:
] [ ] [ ) ( ) ( x y E y x E D R D R
YX XY
= = =
] ) 1 2 sin[( ] ) 1 2 sin[(
1 2
) (
4 ) ( ) sin sin( ) (
0
, 0
) 1 2 (
1
, , k
m
k
m
N
k
k U U XY
m m
m
D J
Q d PAS D D R
c
+
+
= =
=
+
=
From these resolutions, both field
f
(D) and envelope
e
(D) correlation coefficient are defined
as:
2
2
) ( ) ( ) ( ) ( D jR D R D D
XY XX f e
+ = =
57
Truncated Gaussian PAS
Truncated Gaussian PAS model is defined as:
)
`
+
=
c
k G
N
k k k
k k
k G
k G
G
e
Q
PAS
1 , 0
, 0
2
) (
,
,
)] ( [
)] ( [
2
) (
,
2
2
0
The normalisation constant Q
G,k
are derived such that
1
2
,
,
=


.

\

=
c
N
m k
k G
k G
erf Q
Using above definition of Truncated Gaussian PAS and normalisation condition the cross
correlation functions are easily derived:
(
(
(
(
(
(


.

\



.

\

+ =
= =
2
2
2
2
) 2 exp( ) 2 cos( ) ( ) ( ) (
,
,
,
,
2
,
2
1
, 0 2
1
, 0 ,
k G
k G
k
k G
k G
k
k G
m
k m
N
k
k G G XX
jm erf
jm erf
m m D J Q D J D R
c
(
(
(
(
(
(


.

\

+


.

\

+
+ =
=
+
=
)
2
1
( 2
2
)
2
1
( 2
2
)
2
1
( 2 exp ] ) 1 2 sin[( ) ( ) (
,
,
,
,
2
,
2
1
, 0 ) 1 2 (
1
, ,
m j erf
m j erf
m m D J Q D R
k G
k G
k
k G
k G
k
k G
m
k m
N
k
k G G XY
c
58
Truncated Laplacian PAS
Truncated Laplacian PAS model is defined as:
)
`
+
=
c
k L
N
k k k
k k
k
k L
k L
L
e
Q
PAS
1 , 0
, 0
, 2
,
,
)] ( [
)] ( [
2
) (
,
2
0
The normalisation condition is given by
1
2
exp 1
,
,
=
(
(


.

\

=
c
N
m k k L
k
k L
Q
The crosscorrelation function is then given by:
(
(
(
(
(


.

\

+
+


.

\

+ =
= =
) 2 cos(
2
) 2 sin( 2
2
exp
2
) 2 cos(
) 2 (
2
) (
2
4 ) ( ) (
,
, ,
1
, 0
2
2
,
2
1
,
,
0 ,
k
k L
k
k L
k
k L
m
k
k L
m
N
k
k L
k L
L XX
m m m
m
m
D J
Q
D J D R
c
 
   
(
(
(
(
(
+ + + +


.

\

+
+ +


.

\

=
=
+
=
k
k L
k
k L
k
k L
m
k
k L
m
N
k
k L
k L
L XY
m m m
m
m
D J Q
D R
c
) 1 2 ( cos
2
) 1 2 ( sin ) 1 2 (
2
exp
2
) 1 2 ( sin
) 1 2 (
2
) (
2
4 ) (
,
, ,
0
, 0
2
2
,
) 1 2 (
1
,
,
,