See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/26438520

Geometry-based channel modelling of MIMO channels in comparison with

channel sounder measurements

Article in Advances in Radio Science · January 2004

DOI: 10.5194/ars-2-117-2004 · Source: DOAJ

CITATIONS READS

36 140

3 authors:

Giovanni Del Galdo Martin Haardt

Technische Universität Ilmenau Technische Universität Ilmenau
311 PUBLICATIONS 3,470 CITATIONS 530 PUBLICATIONS 14,800 CITATIONS

SEE PROFILE SEE PROFILE

Christian Schneider
Technische Universität Ilmenau
178 PUBLICATIONS 3,496 CITATIONS

SEE PROFILE

All content following this page was uploaded by Martin Haardt on 02 June 2014.

The user has requested enhancement of the downloaded file.

Advances in Radio Science (2004) 2: 117–126
© Copernicus GmbH 2004 Advances in
Radio Science

Geometry-based channel modelling of MIMO channels in

comparison with channel sounder measurements
G. Del Galdo, M. Haardt, and C. Schneider
Communications Research Laboratory, Ilmenau University of Technology, P.O. Box 100565, 98684 Ilmenau, Germany

Abstract. In this paper we propose a flexible geometry- at both ends of the link, i.e. a MIMO system (Multiple Input
based propagation model for wireless communications de- Multiple Output). In fact, if MR and MT are the number of
veloped at Ilmenau University of Technology. The IlmProp antennas at the receiver and at the transmitter, respectively,
comprises a geometrical representation of the environment the number of channels to be calculated grows to MR · MT .
surrounding the experiment and a precise representation of Furthermore a precise physical and geometrical description
the transmitting and receiving antennas. The IlmProp is of the objects would be necessary.
capable of simulating Multi-User MIMO scenarios and in- In order to reduce the complexity of the model the so
cludes a complete collection of tools to analyze the synthetic called GTD (Geometric Theory of Diffraction) can be em-
channels. In order to assess the potentials as well as the limits ployed (Correia, 2001). This approach postulates the exis-
of our channel simulator we reconstruct the scenario encoun- tence of direct, diffracted and reflected rays only. As the
tered in a recent measurement campaign at Ilmenau Univer- wavelength λ approaches zero, this approximation becomes
sity of Technology leading to synthetic data sets similar to increasingly accurate.
the ones actually measured. The measurements have been
collected with the RUSK MIMO multi-dimensional channel With a ray-tracing engine and a full 3D description of the
sounder. From the comparisons of the two channel matri- environment it is then possible to calculate all possible rays
ces it is possible to derive useful information to improve the linking the antennas. However with a GTD approach a com-
model itself and to better understand the physical origins of plete geometrical representation of the scenario is still re-
small-scale fading. In particular the effects of the different quired. A further approximation is to simplify the geometry
parameters on the synthetic channel have been studied in or- surrounding the antennas with a discrete number of reflec-
der to assess the sensibility of the model. This analysis shows tors, called scatterers; when only direct and reflected rays
that the correct positioning of a small number of scatterers are considered the computational complexity drops signifi-
is enough to achieve frequency selectiveness as well as spe- cantly. In this paper we present the IlmProp, a simplified
cific traits of the channel statistics. The size of the scattering model which generates channels which mimic most charac-
clusters, the number of scatterers per cluster, and the Rician teristics of measured channels. To this aim it is extremely im-
K-factor can be modified in order to tune the channel statis- portant to determine which features of the synthetic channel
tics at will. To obtain higher levels of time variance, moving are affected by the different parameters present in the model.
scatterers or time dependent reflection coefficients must be A relatively low computational complexity allows us to gen-
introduced. erate channels with all possible combinations of values for
the parameters. Then we analyze the data by finding strong
traits which link a particular parameter to a particular feature.
At the same time we are able to assess whether a parameter
1 Introduction has no effect on a specific feature.

The high complexity of the propagation phenomenon sets a This extensive search and its results are reported in Sect. 3.
vast challenge in the effort of modelling the wireless channel. Then we modeled the scenario found in one of the measure-
The most accurate approach would be to solve the Maxwell ment campaigns undertaken at Ilmenau University of Tech-
equations. However the computational effort required would nology. Section 4 reports the direct comparison between the
be prohibitive, especially if we considered multiple antennas synthetic and measured channels which led to a better under-
standing on the origins of the fast and slow fading processes.
Correspondence to: G. Del Galdo Section 5 deals in particular with the comparison of the phase
([email protected]) responses. Lastly, Sect. 6 concludes this paper.
118 G. Del Galdo et al.: Geometry-based channel modelling

Fig. 2. The geometrical representation of the model Mα . Two

scattering clusters and a strong LOS are visible. The transmitter (at
the top) moves on a linear path.

Fig. 1. In this example 2 Mobiles (M’s) surrounded by scatterers

move in the proximity of one fixed Base Station (BS).

2 The IlmProp: a geometry-based channel model

The IlmProp developed at TU Ilmenau relies entirely on a

3D geometric representation and has been designed to sim-
ulate a Multi-User-MIMO scenario. It has been simplified
to single-bounce reflections in order to reduce the compu-
Fig. 3. The geometrical representation of the model Mβ . The
tational complexity. Clusters of scatterers can be arbitrarily model is characterized by three scattering clusters; one of them sur-
positioned in the 3D reconstruction of the environment. Each rounds the transmitter.
scatterer corresponds to a single ray and is characterized by a
complex coefficient which, for simplicity, is independent of
the Direction of Arrival (DoA) and the Direction of Depar- reason more processing is required. On the contrary, in the
ture (DoD); but may vary in time with an arbitrary law. The frequency domain we can calculate the channel response for
coefficient determines the phase shift and power attenuation the specific frequency bins that we are interested in. The
introduced by the scatterer. choice of a specific Bello domain is at the end a matter of
The receiver thus receives a number of rays equal to the convenience since it is possible to derive the other represen-
tations via Fourier transforms. For instance, having H (f, t),
number of scatterers plus the Line Of Sight (LOS), which
i.e. the channel expressed in delay time and time, we can cal-
links the transmitter to the receiver directly. Figure 1 shows culate h(τ, t), the channel in frequency and time, as:
a simple example of this modelling approach. Such a sim-
plified model still requires a fairly high computational com-
Z
h(τ, t) = H (f, t)ej 2πf τ df . (1)
plexity because the total number of rays to be computed is:
MR · MT · Nscatt .
The advantage of a geometry-based model with respect to The amplitude and phase of each ray are calculated sepa-
a stochastic one is indeed the possibility to easily generate rately at the receiver. The phase of the i-th ray at a certain
a frequency selective time variant channel. In fact the fre- time snapshot to and frequency fo are calculated as follows:
quency selectiveness is caused by the multi-path effect while fo
the time variance is set off by the movement of the antennas ϕi (to , fo ) = 6 (e−j 2π c Li ) + θi , (2)
and of the scatterers as well as the varying scattering coef-
ficients with respect to time. In other words it well models where Li is the ray’s path-length and θi is the phase shift
a typical frequency selective time variant radio channel. In introduced by the scatterer. The amplitude depends simply
order to characterize such a channel we need to describe it in on the scattering coefficient, antenna gain, and path-loss (Van
one of the four possible two-dimensional domains named the Trees, 2002).
Bello domains. (Bello, 1963). Either {t, f }, {t, τ }, {fD , f }
or {fD , τ } where t, fD , f , and τ denote time, Doppler fre-
quency, frequency, and delay time, respectively. The most 3 An exhaustive investigation
convenient domain in which to compute the channel is cer-
tainly {t, f }. Having a geometrical representation of all ob- The first step in understanding how sensible the model is to
jects in time makes the choice of t over fD quite obvious.
the different parameters and which features of the synthetic
The choice of f over τ requires more explanation. Thanks
to the geometrical representation it is simple to calculate the channel are affected, is to model some typical scenarios and
path lengths for each ray present in the scenario. Assuming to generate channels with all possible values for the param-
that the velocity of light, c, is equal to the velocity of all rays, eters. We implemented two models, to which we will refer
we can calculate the delay time associated with every ray, i.e. as Mα and Mβ , seen in Figs. 2 and 3. Mα models a typi-
the time needed to reach the receiver. The delay times found cal scenario in which the areas around the transmitter and the
will not necessarily match the sampling grid chosen. For this receiver are free of scatterers while some are present in the
G. Del Galdo et al.: Geometry-based channel modelling 119

middle of the link. Mβ , on the other hand, models a scatter- 3.1 K-factor
ing cluster in the proximity of the transmitter. We generated
each channel for the following sampling grid: The K-factor affects several features of the channel. Firstly
it influences its variance with respect to all four domains. It
– A bandwidth B of 400 MHz at a center frequency of 1 is thus very interesting to assess this information quantita-
GHz sampled with a 1f = 1.3 MHz for a total of 301 tively computing the RMS delay spread τRMS , the coherence
samples. bandwidth (1f )c , the RMS Doppler spread fD,RMS and the
– A total of 200 time snapshots were taken, one every 10 coherence time (1t)c .
The RMS delay spread of the channel, τRMS , is derived
ms for a total duration of 2 s.
from the multipath intensity profile ψDe (τ ) which represents
The choices of these parameters determine the sampling the average power of the channel output as a function of de-
grids in the coupled domains. In fact if the complex base- lay time (Paulraj et al., 2003; Stüber, 1996).
band signal is sampled in the range [− B2 , B2 ], then 1τ , the vR
u τmax
(τ − τ̄ )2 ψDe (τ )dτ
resolution in delay time, is equal to B1 . In the same way the
u
τRMS = t 0 R τmax . (5)
bandwidth in the Doppler frequency, 2 · fD,max can be cal- 0 ψDe (τ )dτ
culated from the resolution in time 1t as:
τmax is the maximum delay spread while τ̄ is the average
1
fD,max = . (3) delay spread given by
2 · 1t
R τmax
τ ψDe (τ )dτ
– The resolution in delay time is 1τ = 2.5 ns, which τ̄ = R0 τmax . (6)
corresponds to a distance of less than a meter (75 cm). 0 ψDe (τ )dτ
We can successfully resolve paths (i.e. without alias- The coherence bandwidth (1f )c corresponds to the fre-
ing), which are up to 220 meters long. quency lag in which the channel’s autocorrelation function
reduces to 0.7 (Paulraj et al., 2003). As previously described,
– Doppler shifts up to fD,max = ±50 Hz are visible with-
the channel representations in the four domains are coupled
out aliasing.
through Fourier transforms. For this reasons (1f )c and τRMS
The transmitter moves with a uniform velocity of about 5 are connected as well. In fact the coherence bandwidth is in-
versely proportional to the RMS delay spread so that
m/s. Both transmitter and receiver employ a ULA (Uniform
Linear Array) with MR = MT = 4 antennas spaced by λ2 . (1f )c =
const1
. (7)
After several experiments with different geometries and τRMS
different configurations we narrowed down our extensive The same concepts are applied on the {fD , t} pair. The RMS
search to three dominant parameters: Doppler spread, fD,RMS , is derived in a similar way from
– Cluster size ψDo , the Doppler power spectrum, which corresponds to the
average power as a function of Doppler frequency fD .
– Number of scatterers per cluster sR
F (fD − f¯D )2 ψDo (fD )dfD
fD,RMS =
– Rician K-factor
R
F ψDo (fD )dfD
R
A fourth dominant parameter is certainly the modelling of f ψ (f )df
R D Do D D
f¯D = F (8)
the reflection coefficients, and more importantly their trend F ψDo (fD )dfD
in time. This leaves, however, too many degrees of freedom
since we could implement any arbitrary function to describe The integration domain, F, corresponds to the Doppler fre-
it in time. For this reason we left this important parameter quency interval where the Doppler power spectrum is non
to future and more specific investigations. In our simulations zero. The coherence time (1t)c is calculated in the same
the reflection coefficients are kept constant along all dimen- way as the coherence bandwidth. It corresponds to the time
sions, i.e. frequency and time. The Rician K-factor is defined lag at which the autocorrelation function drops to 0.7 (Paulraj
as et al., 2003). (1t)c indicates how time variant the channel is
and similarly to the {f, τ } pair
PLOS
K= , (4) const2
Pscatt (1t)c = (9)
fD,RMS
i.e. the power attributed to the Line Of Sight divided by the
power of the scatterers. Clearly it is a function of the path- Figure 4 shows typical trends for τRMS , fD,RMS , (1f )c , and
losses which on their side are determined by the geometry, (1t)c for different K-factors. Similar plots can be observed
the transmit power, and the reflection coefficients. We scale for both models Mα and Mβ for almost any cluster size and
the amplitudes of the reflection coefficients to obtain a de- number of scatterers. As K grows τRMS decreases, mean-
sired K-factor. ing that the channel becomes rapidly less frequency selec-
We then generate many channels with different values for tive. The same observation can be derived from the trend of
these three parameters and then analyze the data to find any (1f )c . fD,RMS and (1t)c show the same trends. For bigger
strong dependency. K’s the channel becomes increasingly less time variant. The
120 G. Del Galdo et al.: Geometry-based channel modelling

−8
x 10 60 60
2.5 8
50 50
2
6

fD,RMS
τRMS

40 40
1.5
4 30 30
1

0.5 2 20 20
0 10 20 30 40 0 10 20 30 40
7
x 10 K factor K factor 10 10
5 0.1
0 0
0.08 −140 −130 −120 −110 −100 −140 −130 −120 −110 −100
4
K factor=1 K factor=2.5
(∆f)c

c
(∆t)
0.06
3
0.04 60 60
first
2 0.02 50 50 second
0 10 20 30 40 0 10 20 30 40
third
K factor K factor
0.5 1 40 40 fourth
const1 = (∆t)c ⋅ fD,RMS

const2 = (∆f)c ⋅ τRMS

0.4 0.8 30 30

0.3 0.6 20 20

0.2 0.4 10 10
0.1 0.2
0 10 20 30 40 0 10 20 30 40 0 0
K factor K factor −140 −130 −120 −110 −100 −140 −130 −120 −110 −100
K factor=5 K factor=15

Fig. 4. τRMS , fD,RMS , (1f )c and (1t)c plotted for different K- Fig. 6. Eigenvalue distributions in dB as K grows for Mα .
factors.

ξ with the effective diversity order, as described in (Nabar,

0.8 0.8

0.7 0.7
2003).
Coefficient of Variation

0.6 0.6
In fact the effective diversity order, Ndiv , can be written as:
0.5 0.5 1
0.4 0.4
Ndiv = 2 , (11)
ξ
0.3 0.3

0.2 0.2 and it approaches infinity for a channel with only the LOS
0.8 0.9 1
K factor=0.5
1.1 1.2
9
0.8 0.9 1
K factor=5
1.1 1.2
9
component (AWGN channel) while Ndiv = MR · MT for
x 10 x 10
Rayleigh channels. The trend in ξ and consequently in Ndiv
0.8 0.8
can be observed for both models proposed and independently
0.7 0.7
from other parameters.
Coefficient of Variation

0.6 0.6
Another interesting analysis can be performed on the
0.5 0.5
statistics of the eigenvalues. Figure 6 shows their distribu-
0.4 0.4
tions for different K-factors for the model Mα . Eigenvalue
0.3 0.3
decompositions have been applied on the spatial correlation
0.2 0.2
matrices at the receiver R R calculated for each time snap-
0.8 0.9 1 1.1 1.2 0.8 0.9 1 1.1 1.2
K factor=10 9
x 10 K factor=20 9
x 10
shot. Then we generated the histograms to derive the am-
plitudes distribution on a dB scale. The spatial correlation
Fig. 5. The coefficient of variation ξ plotted against frequency f matrix R R at the receiver for the specific time to is:
for different K-factors. F
1 X
R R (to ) = H (:, :, i, to ) · H (:, :, i, to )H , (12)
F i=1
two constants const1 and const2 are plotted as well to show
that the inverse proportionality expressed in Eqs. (7) and (9) where F is the total number of frequency bins, H H is the
is a good approximation. Hermitian transpose of H and the channel H has dimensions
Another interesting measure of the variation of the chan- MR × MT × F × T . The MATLAB like notation H (:, :, i, to )
nel is to look at the fluctuation of kH k2F . A good index is denotes the two-dimensional channel matrix H ∈ CMR ×MT
the so called coefficient of variation (Nabar, 2003) which is for a specific frequency f = i and time t = to . Figure 6
commonly used as a quantitative measure of the spread (fluc- shows the distributions of the eigenvalues plotted in dB. The
tuation) of a random variable. The coefficient of variation ξ channels have been generated without noise and without any
is defined as: normalization, i.e. the absolute amplitudes of the eigenval-
q ues appear very low (around −120 dB). This has no physical
E{(kH k2F )2 } − (E{kH k2F })2 meaning and only the relative distances between the eigen-
ξ= . (10)
E{kH k2F } values should be taken into account. As the K-factor grows,
the Line Of Sight becomes increasingly stronger and the
We calculated ξ for all frequency bins to discover that for channel displays low rank traits. In other words the strongest
growing K’s the coefficient of variation becomes increasingly eigenvalue separates itself from the weaker ones as K in-
smaller (Fig. 5). This is actually expected if we associate creases.
G. Del Galdo et al.: Geometry-based channel modelling 121

3500 3500 3500 10 10 10

3000 3000 3000

Effective Diversity Order

8 8 8
2500 2500 2500

2000 2000 2000

6 6 6
1500 1500 1500

1000 1000 1000 4 4 4

500 500 500

0 0 0 2 2 2
0 2 4 0 2 4 0 2 4 0.8 1 1.2 0.8 1 1.2 0.8 1 1.2
K factor=0 −6 K factor=1 −6 K factor=2 −6 cluster size=0.5 x 109 cluster size=2.5 x 109 cluster size=5 x 109
x 10 x 10 x 10

3500 3500 3500 10 10 10

Effective Diversity Order

3000 3000 3000
8 8 8
2500 2500 2500

2000 2000 2000 6 6 6

1500 1500 1500

1000 1000 1000 4 4 4

500 500 500

2 2 2
0 0 0 0.8 1 1.2 0.8 1 1.2 0.8 1 1.2
0 2 4 0 2 4 0 2 4 cluster size=10 x 109 cluster size=15 x 109 cluster size=20 x 109
K factor=3 −6
x 10 K factor=5 −6
x 10 K factor=15 x 10−6

Fig. 7. Amplitude distributions of channels considering all frequen- Fig. 8. The effective diversity order Ndiv plotted against frequency
cies and time snapshots for different Rician K-factors for Mα . f for different cluster sizes (expressed in λ’s) for Mβ .

A higher K-factor inhibits the effects of the fast-fading. (cluster size = 0.5 λ), we see a slow moving trend which
This can be easily seen by plotting the distributions of the looks like the sum of a concave semicircle and a zero mean
amplitudes of the channel considering all frequencies and noise-like process which introduces fast variations. After
time snapshots (Fig. 7). many simulations we observed that the magnitude of the fast
moving Ndiv changed at every realization while the variance
3.2 Cluster size and number of scatterers of the slow moving Ndiv changed with the cluster diameter
with an undoubtedly direct proportionality. In other words,
While there exists a strong relationship between the different for bigger cluster sizes, Ndiv changes more rapidly along fre-
channel statistics and the Rician K-factor, there is no strict quency while this does not necessarily correspond to higher
relation with the parameters which characterize a scattering frequency selectiveness nor to greater time variance. The
cluster, namely the number of scatterers, the cluster size and number of scatterers affects the channel in a similar way but
the type of distribution of the scatterers inside the cluster. In with a much weaker dependency.
the first simulations we investigated whether the geometrical
distribution affected the channel. We generated clusters with
scatterers distributed uniformly on a sphere, on a circle on 4 A comparison with measurements
the horizontal plane (as seen in Figs. 2 and 3), and on the
surface of a cylinder or a sphere. All channels showed the In order to assess the results derived in Sect. 3 with an even
same trends and for this reason we continued with more de- more realistic scenario we modeled the surroundings of a
tailed investigations fixing the distribution of the scatterers measurement campaign undertaken at TU Ilmenau. Addi-
on a circle. We generated models characterized by cluster tionally the simulation allowed us to investigate the impor-
sizes with diameters ranging from λ2 to 20λ and with num- tance of other critical parameters of the model: the positions
bers of scatterers ranging from 1 up to 100 reflection points of the scatterers and the modelling of the reflection coeffi-
per cluster. cients. The first step was to determine whether a rough po-
We then investigated trends in the different channel fea- sitioning of the scatterers was enough to recreate the main
tures fixing all parameters as either the cluster size or the features which characterize the measured channel. To this
number of scatterers changed. All features but the effective end, we chose, among the many measurements gathered, a
diversity order did not show any strong dependency even channel with a strong LOS and characterized by a simple
though they were all affected. For instance, increasing the geometry so that the position of the scatterers could easily
size of the scattering clusters affects the frequency selective- be estimated. Figure 9 shows the bird’s eye view of the re-
ness of the channel, even though there is neither a direct nor constructed scenario. The receiver was a 4 × 8 polarization
an inverse proportionality between the two. Figure 8 shows sensitive patch-array URA (Uniform Rectangular Antenna)
the effective diversity order Ndiv plotted for different frequen- and it was positioned by a window on the second floor of
cies f and for different cluster diameters expressed in λ’s a building. The transmitter was a single 2 GHz omnidirec-
for the first model, Mα . Analyzing Ndiv along different fre- tional antenna mounted on a wheeled cart pulled by hand up
quencies we can recognize a slow-moving trend corrupted a road in the direction of a second building (Fig. 10). The
by faster variations. For instance, in the first plot of Fig. 8 transmitter path is represented by a continuous line in Fig. 9.
122 G. Del Galdo et al.: Geometry-based channel modelling

Fig. 10. The receiver (top circle) and the transmitting antenna (bot-
tom circle) mounted on a cart.

the IlmProp then the whole channel matrix must be scaled so

Fig. 9. Bird’s eye view of the synthetic reconstruction of the sce- that
nario where the measurements were taken. X X
kH M k2 = kH I k2 (13)
2 2

In the proximity of the building the cart was turned around where 2 spans the complete domain of all dimensions. In
and pulled in the opposite direction (towards the receiver) our simulations H I has been generated without noise. The
normalization and the adding of the noise must then be done
approximately on the same path. The whole measurement
concurrently. First we have to estimate the power of the noise
took approximately 1 min at a speed of 3 m/s. The measure- from the measurements. We assume constant AWGN (Addi-
ment device employed was the broadband real-time channel tive White Gaussian Noise) over all dimensions (delay time,
sounder RUSK MIMO from MEDAV (Thomä et al., 2001). time, antennas, etc.). An effective way to measure the noise
The following parameters characterize the measurements: floor in the measurements is to observe the channel matrix in
the time and delay time domain. If the path lengths are short
– A bandwidth B of 120 MHz at a center frequency of enough, so that the last echoes extinguish before the maxi-
1.95 GHz sampled with 1f = 300 kHz for a total of mum delay resolvable, we have a measurement of the noise
385 samples. without signal. If the number of samples is sufficient we can
use this data to estimate the power of the noise. The noisy
synthetic channel will then be
– A total of 6313 time snapshots were taken, one every
9.2 ms for a total of 58 s. H I,noisy = H I,noiseless + H w , (14)

where H w is a matrix whose elements are ZMCSCG (Zero

The same parameters were chosen for the synthetic IlmProp Mean Circular Symmetric Complex Gaussian) random num-
model so that the two data sets had same dimensions and bers with variance σ 2 (Paulraj et al., 2003).
sampling grids. However, in order to compare the measure- Equation (13) then becomes:
ments with the synthetic channels, adding the noise and per- X X
forming a proper normalization must be taken into account. kH M k2 − γ · σ 2 = kH I k2 , (15)
2 2
The need for a normalization is due to the fact that the
absolute powers sent by the antennas have not been consid- where γ is the total number of samples present in H , i.e.
ered. Only the relative difference in power between the LOS γ = MR · MT · T · F , where T and F are the number of time
and the scattering components have been modeled through snapshots and frequency bins, respectively.
the Rician K-factor which can be set at will. This simpli- The geometry has been roughly reconstructed with the
fied the modelling effort even though the results need post-
help of blueprints of the area and through visual inspection.
processing in order to be finally compared with the measure-
ments. The normalization is performed in such a way that Cars, buildings, light poles, and trash bins have been identi-
the total power received for the synthetic channel matches fied as the probable sources of scattering.
the one received in the measurements. Clearly both data Figure 11 shows the channel impulse responses in time t
sets must have the same number of samples, dimensions and and delay time τ . The channel has been gathered originally
more in general, the same sampling grids. If H M denotes the in the {t, f } domain; the CIRs in delay time have been com-
channel for the measurements and H I the one generated with puted via the Fourier transform described in Eq. (1). The
G. Del Galdo et al.: Geometry-based channel modelling 123

0 0
−30 −30

LOS
10 double−bounce −40 10 −40
echo

20 −50 −50
20
time [s]

time [s]
−60 −60
30 30

−70
−70
40 single−bounce
echo 40

−80
−80
50
50

−90
−90
0 0.5 1 1.5 2 2.5 3
delay time [µs] 0 0.5 1 1.5 2 2.5 3
delay time [µs]

Fig. 11. Channel Impulse Responses from the measurements in Fig. 12. The synthetically generated Channel Impulse Responses in
time t and delay time τ . The Line Of Sight LOS and echos are time t and delay time τ .
clearly visible. The grayscale represents amplitudes in a dB scale

−25
LOS is clearly visible and as expected its delay lag grows as
−30
the transmitter moves further away from the receiver. When
Amplitude [dB]

the cart is turned around and pulled back towards the re- −35

ceiver, the arrival delay time decreases. Many echoes are −40

visible as well. Note that as the transmitter moves most −45

echoes arrive earlier, meaning that their path length becomes −50
0 5 10 15 20 25
smaller. When the curve corresponding to one echo changes
in time with the same slope characterizing the Line Of Sight,
−25
then we can deduce that the path length of the echo changes
−30
with the same rate of the path length of the LOS. The strong
Amplitude [dB]

echo labeled double-bounce echo has the same trend as the −35

LOS just delayed by approximately 0.8 µs, which corre- −40

sponds to 240 m, the distance between the top and bottom −45

buildings doubled. For this reason it is very likely that this ar- −50
0 5 10 15 20 25
rival corresponds to the signal reflected by the bottom build- time [s]

ing first, and then by the top building, thus a doble-bounce

echo. Figure 12 shows the channel generated by our model. Fig. 13. Amplitudes of the CIRs for a specific frequency bin plot-
The trend of the LOS is extremely well reproduced. This ted in dB against time for the measurements (top) and the IlmProp
was, however, expected since the LOS is the only compo- model (bottom).
nent easily predictable once the position of the transmitter
in known. Most of the scattering contributions can be ob-
served in the synthetic data set as well, with the exception dependent on the Direction of Arrival (DoA) and Direction
of the double-bounce paths which have been omitted. An- of Departure (DoD). Figure 13 shows the fluctuations of the
alyzing the two data sets in the {t, τ } domain suggests im- amplitudes of the channel for a specific frequency fo plot-
provements in the modelling approach: in the measurements ted against time for the first 25 s. Fast-fading characterizes
we can observe that certain echo traces are stronger than oth- both signals even though the measurements appear somehow
ers. In the modeled channel this does not happen. This is a more variant. In order to assess this feature quantitatively we
result of the modelling of the reflection coefficients. In fact compute the multipath intensity profile ψDe (τ ), the Doppler
while their phase has been generated uniformly distributed in power spectrum ψDo (fD ) and the autocorrelation functions
the [0, 2π] range, the amplitude has been taken equal for all. in frequency acf(1f ) and time acf(1t) for both data sets.
The comparison suggests that the reflection coefficients need From their analysis we can obtain the RMS delay spread
a more sophisticated modelling. Furthermore, a closer look τRMS , the coherence bandwidth (1f )c , the RMS Doppler
on the measurements reveals that some echoes exist only in spread fD,RMS and the coherence time (1t)c (Figs. 14 and
certain time periods. This suggests that the amplitude of the 15). The RMS delay spread shows that the energy is simi-
reflection coefficients (perhaps their phases too) should be larly distributed in delay time τ . In the measurements two
 0.8
1.5

0.6

ψ Do(fD)
(f )

ψ De(τ)
Do D
1 1

ψ
124 0.4 G. Del Galdo et al.: Geometry-based channel modelling
0.5
0.2
−10 −10 −10 −10
x 10 x 10 x 10 x 10
2 1 2
0 0 0
4 5 −40 −20 0 20
0.8 40 0 1 2 3 4 −40 −20 0 20 40
−7
Doppler Frequency 1.5 delay time −7
x 10 x 10 Doppler Frequency
0.6

ψ Do(fD)
ψ Do(fD)

ψ De(τ)
ψ De(τ)

1 1 1
1 0.4 1 1
0.5
0.2
0.8 0.8 0.8
0 0 0 0
0 1 2 3 4 5 −40 −20 0 20 40 0 1 2 3 4 −40 −20 0 20 40
delay time Doppler Frequency delay time
0.6 −7 0.6 x 10
−7
0.6 Doppler Frequency

acf(∆f)

acf(∆t)
x 10
acf(∆t)

1 1 1 1
0.4 0.4 0.4
0.8 0.8 0.8 0.8

0.2 0.2 0.2

0.6 0.6 0.6 0.6

acf(∆f)

acf(∆t)
acf(∆f)

acf(∆t)

0.4 0.4 0 0.4 0.4 0

0 −4 −2 0 2 4 −0.2 −0.1 0 0.1 0.2 0.3
2 4 −0.1 −0.05 0 0.05 0.1
0.2 0.2 0.2 frequency 6 0.2 time
x 10
6 time x 10
0 0 0 0
−4 −2 0 2 4 −0.1 −0.05 0 0.05 0.1 −4 −2 0 2 4 −0.2 −0.1 0 0.1 0.2 0.3
frequency 6 time frequency x 10
6 time
x 10

path intensity profile ψDe (τ ), Doppler power Fig. 15. The multipath intensity profile ψDe (τ ), Doppler power
Fig. 14. The multipath
and the autocorrelation intensityin
functions profile ψDe (τ ), Dopplerspectrum
frequency power ψDo15.(fD
Fig. ) multipath
The and theintensity profile ψDe (τfunctions
autocorrelation ), Doppler power
in frequency
spectrum ψDo (fD ) and the autocorrelation functions in frequency spectrum ψDo (fD ) and the autocorrelation functions in frequency
cf(∆t) foracf(1f
the measurements
) and time acf(1t) for the measurements
acf(∆f ) acf(1f
and time) and acf(∆t)
time acf(1t)for theIlmProp
for the IlmPropmodelmodel

more spikes can be observed. This is due to the fact that two Q
specific (∆f
ce bandwidth echoes)have
c which confirms
a significantly strongerthatmagnitude com- echoes
have approximately the same coherence τRMS is any-
pared to other echoes. The RMS delay spread
how very similar suggesting that the two channels will have a
{t, fD } pair
similarthe two channels
frequency selectiveness.show
The RMS moredelay spread cor- LOS
e the Doppler
respondspower betweenψ
spectrum
to the distance theDo two(fvertical
D) lines plot-
received
ends, thetedautocorrelation
in Figs. 14 and 15. From the autocorrelation
function in the function in
the frequency domain acf(1f ) we can derive the coherence
s that thebandwidth
measured (1f channel is much
)c which confirms that themore
two channels have ϕLOS ϕR
he synthetic one. Furthermore this analy-
approximately the same coherence bandwidth. In the {t, f D }
pair the two channels show more differences. While the I
ation (9) Doppler
only represents
power spectrum a good
ψDo (fDempirical
) has very similar trends,
but not all
the cases. In Figure
autocorrelation function13 wetime
in the candomain
in shows that Fig. 16. When the amplitude of the echoes is small the phase of the
the measured channel Fig.the16. When
receivedthe amplitude
signal ϕR is a goodof the echoes
approximation is phase
of the smallof the
the LOSphase of the
the measurements have isa much
faster more time variant than
moving component only ϕ .
signal ϕR is a good approximation of the phase of the LOS
synthetic one. Furthermore this analysis proves that received Eq. (9) LOS
t, especially duringa the
only represents good first 10law,
empirical seconds.
valid in most butcomponent
not all only ϕLOS .
ed how we could
cases. In Fig.generate
13 we canmore time vari-
in fact recognize that the measure-
ments have a faster moving small-fading effect, especially the sum of the phase introduced by the LOS ϕLOS plus the
y positions,
duringthethesize ofs.the
first 10 scattering
We then investigated clus-
how we could gen- additional phase introduced by the echoes. The smaller the
er of scatterers
erate more pertime
cluster,
variancediscovering that the size of amplitude of the echoes is compared to the LOS component,
changing only positions,
clus- of λ
ment theinsmaller
the radial
will be direction
the differencecorresponds
ϕR − ϕLOS . As to thearadial
full rotation
ble. Ourthe scattering clusters, and the number of scatterers per
simulations showed that in or-
ter, discovering that this was not possible. Our simulations of 2π radians in the phase. If this wavelength has been sam-
distance between transmitter and receiver changes, the LOS
her ordershowed
of time variance
that in it is higher
order to achieve absolutely
order of time variance component, ϕLOS , will change accordingly. If the movement
pled with more than two samples per wavelength, thus ful-
it is absolutely
l time variant necessary
reflection to model time or
coefficients to reflection co- is slow enough and the phase-noise introduced by the echoes
variant
efficients or to introduce moving scatterers along the filling
time Shannon’s
is negligible,theorem,
it is possiblethe phase
to derive the can be efficiently
displacement from ’un-
scatterersdimension.
along the time dimension. the phase trend of the received signal. In fact, a movement
wrapped’, or in other words, absolute jumps greater than π
of λ in the radial direction corresponds to a full rotation of
are converted to their
2π radians 2π complement.
in the phase. If this wavelength Thehas spatial
been sam-displace-
5 A deeper look at the phase response pled with more than two samples
ment can be then derived from the phase simply per wavelength, thus multiplying
ful-
k at the Phase Response filling Shannon’s theorem, the phase can be efficiently ’un-
λ
Looking more closely at the phase of the receivedby the factor:
signal wrapped’,2π or .in other
Figure 17absolute
words, shows thegreater
jumps unwrapped
than π phase
sely at the
leadsphase of the
to interesting received
comparisons andsignal
remarks. Figurefor 16 everyarefrequency bin 2π
converted to their derived directly
complement. fromdisplace-
The spatial the measured
shows how the phase of the received signal (in its complex ment can be then derived from the phase simply multiplying
g comparisons and remarks. Figure 16
base-band representation) is constructed in a flat-fading fre-
channel analyzed in λSection
by the factor: 2π
4. Even though a different fluc-
. Figure 17 shows the unwrapped phase
ase of thequency
received
selectivesignal
channel.(inTheitsresulting
complex tuation
phase ϕR will be perturbs each frequency
for every frequency bin, the
bin derived directly from overall trend is con-
the measured
ntation) is constructed in a flat-fading fre- stant and shows a descending phase for half of the time and
hannel. The resulting phase ϕR will be a symmetrical ascending phase for the other half, in accor-
G. Del Galdo et al.: Geometry-based channel modelling 125

−500

−1000

phase [rad]
−1500

−2000

−2500

real
−3000 model
real LOS
model LOS

−3500
0 5 10 15 20 25 30 35 40 45 50
time [sec]

Fig. 17. Unwrapped phase trend derived from measurements plot- Fig. 18. Phase comparison between the measurements and its syn-
ted in radians against frequency f and time t. thetic reconstruction. The upper curves are obtained simply un-
wrapping the signal received at one specific frequency. The lower
curves represent the LOS component only.
channel analyzed in Sect. 4. Even though a different fluctua-
tion perturbs each frequency bin, the overall trend is constant
and shows a descending phase for half of the time and a sym- find increasingly less steep slopes. In other words the phase-
metrical ascending phase for the other half, in accordance noise introduced by the echoes is uniformly distributed, but
with the experiment’s geometry. The different slope is due due to the non-linearity of the unwrapper, the resulting error
to the different velocity. In fact the transmitter moved faster leads always to more moderate slopes.
in the second half of the experiment. If the channel were
flat-fading we would see the slow moving LOS component
only, absolutely equal for every frequency. It is interesting 6 Conclusions
to compare the overall phase trend with the LOS component
only. In case of the synthetic channel this is very easy to Our simulations show that for certain modelling applications
obtain. More complicated is the case of the measurements. a simple geometry-based single-bounce model can mimic
Once the channel has been transformed in the {t, τ } domain, most channel features with sufficient precision, among which
it is possible to identify the samples in delay time τ with the frequency selectiveness and the time variance are the
highest energy for every time snapshot. The changing phase most characterizing. A rough positioning of the scattering
of these samples against time approximates the phase trend clusters allows us to achieve the desired frequency selective-
introduced by the LOS component only. The comparison of ness as well as other specific traits of the channel statistics
the unwrapped LOS phase with the overall phase is shown in by tuning three basic parameters, i.e, the size of the scatter-
Fig. 18. The model and the measurements show very sim- ing clusters, the number of scattering coefficients, and the
ilar traits. The upper curves show the phase trend for one Rician K-factor. To obtain higher levels of time variance and
frequency bin derived directly from the unwrapping of the more realistic synthetic channels, moving scatterers and time
phase. The lower curves are derived, respectively, with the dependent reflection coefficients must be modeled as well.
LOS approximation for the measurements, and from a direct
calculation of the displacement for the model. In fact, for
Acknowledgements. The authors gratefully acknowledge the sup-
the synthetic data, the exact path of the transmitter is known.
port of MEDAV (www.channelsounder.de) and of the Electronic
From these coordinates it is possible to derive the radial dis- Measurements Engineering Laboratory at Ilmenau University of
placement and consequentially the phase. The phase trends Technology in performing the channel measurements.
match for the first 10–15 s, while they start to diverge signifi-
cantly around 25 s. This can be simply explained considering
that the Rician K-factor drops in the middle part of the exper-
References
iment. When the amplitude of the echoes is comparable with
the LOS, then the received vector will have a phase uniformly
Bello, P.: Characterization of randomly time-variant linear chan-
distributed distributed in the interval [0, 2π ]. The unwrapper nels, IEEE Trans. Comm. Syst., 11, 360–393, 1963.
will not find any trend and will jump from ascending to de- Correia, L. M.: Wireless Flexible Personalised Communications,
scending phases randomly. The resulting unwrapped signal Wiley, 2001.
will have a slope close to 0, that is horizontal. For this rea- Nabar, R. U.: Performance Analysis and Transmit Optimization for
son, as we pass from a higher K-factor to a smaller one we General MIMO Channels, Doctoral Dissertation, Stanford, 2003.
 View publication stats
126 G. Del Galdo et al.: Geometry-based channel modelling

Paulraj, A., Nabar, R., and Gore, D.: Introduction to space-time Thomä, R., Hampicke, D., Richter, A., Sommerkorn, G., and
wireless communications, Cambridge University Press, Cam- Trautwein, U.: MIMO vector channel sounder measurement
bridge, 2003. for smart antenna system evaluation, European Transaction on
Stüber, G.: Principles of Mobile Communication, Kluwer, Norwell, Telecommunications, Special issue on smart antennas, 12, 2001.
MA, 1996. Van Trees, H. L.: Optimum Array Processing: Part IV of Detec-
tion, Estimation, and Modulation Theory, John Wiley & Sons
Inc, New York, 2002.