View metadata, citation and similar papers at [Link].

uk brought to you by CORE

provided by Ghent University Academic Bibliography

Statistics of Multipath Component Clustering in an

Office Environment
E. Tanghe∗ , W. Joseph∗† , M. Lienard† , A. Nasr† , P. Stefanut† , L. Martens∗ , and P. Degauque†
∗ Ghent
University / IBBT, Dept. of Information Technology
Gaston Crommenlaan 8 box 201, B-9050 Ghent, Belgium, e-mail: [Link]@[Link]
† University of Lille, IEMN, Group TELICE

Bldg. P3, F 59655 Villeneuve d’Ascq, France, e-mail: [Link]@[Link]

Abstract—In this paper, directional MIMO measurements in II. M EASUREMENTS

an indoor office environment are presented. A 5-D ESPRIT
estimation algorithm is used to extract parameters associated The measurement setup for the MIMO measurements is
with discrete propagation paths, such as their azimuth of arrival, shown in Fig. 1. A network analyzer is used to measure the
azimuth of departure, delay, and power. The estimated path
parameters are grouped into clusters using the statistical K- complex channel frequency response for a set of transmitting
power-means algorithm. Statistical distributions are determined and receiving antenna positions. The channel is probed at 1601
for the path parameters within individual clusters and for their evenly spaced frequency points in a range from 3 GHz to
change between clusters. To validate the distributional choices, 3.5 GHz. As transmitting (Tx) and receiving antenna (Rx),
the goodness-of-fit to the proposed distributions is verified using broadband omnidirectional biconical antennas with a nominal
statistical hypothesis tests with sufficient power.
gain of 1 dBi are used. To be able to perform measurements for
large Tx-Rx separations, one port of the network analyzer is
I. I NTRODUCTION connected to the Tx through an RF/optical link with an optical
fiber of length 500 m. The RF signal sent into the Tx and the
The last decade, the demand for high throughput wireless RF signal coming from the Rx are both amplified using an
communication has increased enormously. To meet the ever amplifier with an average gain of 37 dB.
increasing requirements for reliable communication with high Measurements are performed using a virtual MIMO array.
throughput, novel wireless technologies have to be consi- The virtual array is created by moving the antennas to prede-
dered. A promising approach to increase wireless capaci- fined positions along rails in two directions in the horizontal
ty is to exploit the spatial structure of wireless channels plane using stepper motors. Both Tx and Rx are moved along
through multiple-input multiple-output (MIMO) techniques. 10 by 4 virtual uniform rectangular arrays (URAs), and are
High throughput MIMO specifications are already being inclu- polarized vertically and positioned at a height of 1.80 m during
ded in wireless standards, most notably IEEE 802.11n, IEEE measurements. The URA elements are spaced 4.29 cm apart,
802.16e, and 3GPP Long Term Evolution (LTE). which is equal to half a wavelength at 3.5 GHz in order to
The potential benefits of implementing MIMO are highly avoid spatial aliasing. The stepper motor controllers, as well
dependent on the sort of propagation environment. Therefore, as the network analyzer, are controlled by a personal computer
the development of propagation channel models is indispensa- (PC). At each of the 1600 (10 × 4 × 10 × 4) combinations of
ble. In this paper, the geometry-based stochastic type of MIMO Tx and Rx positioning along the URAs, the network analyser
channel model is considered. This kind of model presents measures the S21 scattering parameter ten times (i.e., 10 time
a statistical distribution for the propagation path parameters
(e.g., their direction of arrival, direction of departure, delay,
etc.). Geometry-based stochastic channel models use propa-
gation path clusters in their description: paths with similar PC
propagation parameters are grouped into clusters. An example
of this type of channel model is the COST 273 model [1]. Network analyzer

This work investigates the statistics of path powers, azi- RF to optical RF RF

muths of arrival (AoA), azimuths of departure (AoD), and
delays in an indoor office environment. For this, MIMO Optical fiber Amplifier

channel sounding measurements with a virtual antenna ar- Amplifier Tx Rx

ray are carried out on a typical office floor. Parameters of RF
1.80 m

1.80 m

Optical to RF RF RF
propagation paths are extracted from measurement data and
are subsequently grouped into clusters using an automatic
clustering algorithm. In this paper, statistical distributions are
provided for the clustered propagation path parameters. Fig. 1. Measurement setup
observations). are postprocessed in the delay domain by considering the
The measurements are carried out on the first floor of an power delay profile (PDP). For a typical PDP, power is
office building. Fig. 2 presents a floor plan of the measurement concentrated at small delays while at large delays only the
environment, along with some relevant dimensions. Most inner noise floor remains. In our measurements, the noise floor is
walls are plasterboard. Fig. 2 also shows locations of the Tx set to the power of the MPC with the largest delay. Following,
and Rx during measurements. A total of 9 MIMO measure- all MPCs with power less than the noise floor plus a noise
ments are performed, their Tx and Rx locations indicated threshold of 6 dB are omitted from further analysis. Fig. 3(a)
by couples of Txi and Rxi (i = 1, . . . , 9). Measurements shows a scatter plot of detected MPCs versus their AoA, AoD,
are executed in both line-of-sight (LoS) and non line-of-sight and delay for measurement location 7 (nLoS). The power on
(nLoS) conditions: measurement locations 1, 5, and 6 are LoS. a dB-scale of each MPC is indicated by a color.
B. Clustering of specular paths
Y
For our data, automatic joint clustering of AoA, AoD, and
AoA
or AoD
delay is performed using the K-power-means algorithm [4].
Tx1 Tx3
X
Rx7 Tx4 Tx5 The K-power-means algorithm result is in agreement with the
COST 273 definition of a cluster as a set of MPCs with similar
Tx6 8.1 m propagation characteristics [1]. Because some parameters for
clustering are circular, multipath component distance (MCD)
Rx6 Rx5 1.9 m
is used as the distance measure for clustering [4].
4.2 m
For each measurement location, the number of clusters for
the K-power-means algorithm is varied between 2 and 10.
Rx1 Rx2 Rx4 Rx9 Rx3 Rx8 Tx7 Tx2 The optimal number of clusters is selected using the Kim-
Tx8 Tx9
Parks index [5]. The number of clusters according to Kim-
Parks index varies from 3 to 8 between measurement locations,
57.9 m
and for all MIMO measurements combined, a total of 45
Fig. 2. Floor plan of the measurement environment with Tx and Rx locations clusters are found (16 clusters from LoS and 29 clusters from
nLoS measurements). Next, to ease the statistical analysis,
clearly outlying MPCs are removed from each cluster using
III. DATA PROCESSING the shapeprune algorithm detailed in [4]. Fig. 3(b) shows
A. Extraction of specular paths clustering results for measurement location 7 (nLoS). MPCs
The power, azimuth of arrival (AoA), azimuth of departure grouped into different clusters are shown with different marker
(AoD), and delay parameters of propagation paths or multipath shapes and colors (in total 4 clusters).
components (MPCs) are extracted from measurement data IV. S IGNAL MODEL
using the 5-D unitary ESPRIT (Estimation of Signal Parame-
For the analysis of the within-cluster and between-cluster
ters via Rotational Invariance Techniques) algorithm [2]. The
propagation path parameters, the following basic signal model
coordinate system with respect to which AoA and AoD are
is used. For one of the measurement locations, the complex
defined is shown in Fig. 2.
received envelope h φA , φD , τ is written as function of the


URAs allow easy application of the spatial smoothing
propagation path parameters: φA denotes the AoA, φD the
technique to increase the number of observations [3]. We
AoD, and τ is the path delay. The use of MPC clusters is
choose sub-URAs with dimensions 2/3 of the length in each
reflected in the complex envelope’s notation:
direction of the original 10 by 4 URA, i.e., 7 by 3 sub-URAs.
In total at both link ends, 64 different 7 by 3 sub-URAs nC X
P,c n

X
h φA , φD , τ = Ac,k · δ φA − ΦA


can be found, thereby increasing the number of observations c,k
by a factor of 64. Together with the previously mentioned c=1 k=1
· δ φD − ΦD


10 time observations, the total number of available observa- c,k · δ (τ − Tc,k ) (1)
tions per measurement location is 640. From the measured
In (1), nC is the number of clusters and nP,c is the number
frequency points, 10 equally spaced frequencies are selected
of MPCs within cluster c. For the k-th propagation path in
from 3.5 GHz down for use with the ESPRIT algorithm.
cluster c, Ac,k is its received complex amplitude, ΦA c,k and
The considered constant spacing between these frequencies
is 4 MHz. With this choice, the maximum resolvable path ΦDc,k are its AoA and AoD, and T c,k is its delay. δ (.) denotes
length is 75 m, which is expected to be large enough to the Dirac delta function. We also define h Pc,k asi the power
2
limit possible aliasing in the delay domain. Summarizing, 5-D of path k in cluster c, i.e., Pc,k = E |Ac,k | where the
unitary ESPRIT is applied to 640 observations of a 5-D vector expectation operator E [·] is taken over all 640 observations of
space of size 7 × 3 × 7 × 3 × 10. Ac,k . To allow statistical analysis of propagation parameters
The ESPRIT algorithm is used to estimate the 100 most of all measurement locations collectively, the dependence of
strongest paths from measurement data. The estimated MPCs power Pc,k and delay Tc,k on distance is removed. Power is
 Power [dB]
each of the clusters. The following sections will work towards
−70
a statistical description of the between-cluster and within-
90
−75 cluster propagation parameters.
80 V. S TATISTICAL DISTRIBUTIONS PER CLUSTER
Delay [ns]

−80
This section discusses the statistical distributions of Pc,k ,
70 −85
ΦA D
c,k , Φc,k , and Tc,k within each cluster. The proposed distri-
−90 butions are location-scale distributions: they are parameterized
60
450 by a location parameter, which determines the distribution’s
−95
360
270
location or shift, and a scale parameter, which determines the
270 180 −100 distribution’s dispersion or spread.
180 90
0
90 −90 −105 A. Power Pc,k
AoD [°] AoA [°]
A natural model for the fading of MPC powers Pc,k in
(a)
cluster c is the lognormal fading model. For cluster c, it is
investigated if the samples Pc,k on a dB-scale could originate
from a normal distribution. This normal distribution is parame-
90 terized by the mean µc (location parameter) and the standard
deviation σc (scale parameter) of Pc,k in dB.
80
Delay [ns]

Composite normality of Pc,k [dB] is assessed with a few

statistical tests in literature such as the Anderson-Darling (AD)
70 test, the Shapiro-Wilk (SW) test, and the Henze-Zirkler (HZ)
test. Multiple tests for normality are executed as no uniformly
60
450 most powerful test exists against all possible alternative dis-
360 tributions. Of the 45 clusters in this measurement campaign,
270
270 180 normality of Pc,k [dB] is retained at the 5% significance level
180 90 for 39, 38, and 40 clusters with the AD, SW, and HZ tests,
0
AoD [°] 90 −90 respectively. For the 45 clusters, average p-values are 0.38
AoA [°]
(AD), 0.43 (SW), and 0.44 (HZ). Concluding, normality for
(b) Pc,k [dB] is assumed in the following, as the majority of
Fig. 3. MPC scatter plot (a) and clustering (b) for measurement location 7 clusters pass the different goodness-of-fit tests.
B. Azimuths of arrival ΦA D
c,k and departure Φc,k

rescaled such that the total received MPC power equals one In literature, various distributions are proposed for the
and the origin of the delay axis is set to coincide with the first azimuth angles ΦA D
c,k and Φc,k within a certain cluster c, among
arriving MPC. Assuming larger values of c or k mean later which the normal distribution and the Laplacian distribution.
arriving paths: Additionally, we consider the von Mises distribution. The von
nC nP,c
Mises distribution can be thought of as an analogue of the
normal distribution for circular data. For the AoAs ΦA c,k in
X X
Pc,k = 1 and T1,1 = 0 ns (2)
c=1 k=1
cluster
 c, the von Mises
 probability density function (pdf)
pvM ΦA c,k ; α A
c , κ A
c is given as:
Each of the propagation path parameters Pc,k , ΦA D
c,k , Φc,k ,
and Tc,k are split up into a between-cluster and a within-cluster
  
exp κA A
c cos Φc,k − αc
A
part as follows: pvM ΦA A A


c,k ; αc , κc = (4)
2πI0 (κA
c )
Pc,k = pc pc,k ΦA A A
c,k = φc + φc,k
(3) In (4), I0 (·) is the modified Bessel function of the zeroth order.
ΦD D D
c,k = φc + φc,k Tc,k = τc + τc,k
The two parameters that characterize the von Mises pdf are
In (3), the parameters pc , φA D
c , φc , and τc denote between-
αcA , the circular mean of ΦA c,k (location parameter), and κc ,
A
A
cluster propagation parameters, and are representative for which is a measure of concentration of Φc,k angles around
the location of each cluster in the power/AoA/AoD/delay αcA (scale parameter). For the von Mises pdf of AoDs ΦD c,k
parameter space. Also in (3), pc,k , φA D
c,k , φc,k , and τc,k are in cluster c, an expression analogous to (4) can be written.
within-cluster propagation parameters. The within-cluster pa- The most fit distribution is determined by performing simple
rameters can be seen as the deviations of individual paths likelihood ratio tests (LRTs): the statistical distribution which
from the cluster’s location as dictated by the between-cluster renders the largest likelihood is most appropriate for describing
parameters. The within-cluster parameters are therefore fully the azimuth angle statistics for that cluster. For the 45 clusters
determined by the spread of power, AoA, AoD, and delay in in this measurement campaign, all LRTs decided in favor
of the von Mises distribution for both ΦA D
c,k and Φc,k . We distributional parameter (which could change between these
therefore conclude that the von Mises distribution is most fit two circumstances). The premise of a uniform distribution is
for describing the statistics of azimuth angles within clusters. validated through a statistical hypothesis test, namely Rao’s
spacing test for uniformity. For both the 45 cluster mean AoAs
C. Delay Tc,k φA D
c and the 45 cluster mean AoDs φc , Rao’s spacing test
Delays Tc,k within cluster c are modeled according the retained the null hypothesis of a uniform distribution at the
principle laid out by the well-known, cluster-based Saleh- 5% significance level (p-values of 0.67 and 0.14, respectively).
Valuenzuela (SV) model [6]. Herein, the waiting time bet- 2) Cluster onset τc : We adopt the Saleh-Valenzuela model
ween the arrival of two consecutive MPCs within a certain for the between-cluster delay: the waiting time between the
cluster is modeled by an exponential distribution. For the onsets τc − τc−1 of two consecutively arriving clusters is
MPCs in cluster c (assuming the delays are ordered such modeled by an exponential distribution [6]. This exponential
that Tc,1 < Tc,2 < . . . < Tc,nP,c ), the exponential pdf distribution is fully parameterized by the mean of waiting
pexp (Tc,k | Tc,k−1 ; λc ) as function of the delay Tc,k of the times τc − τc−1 .
k-th MPC, given that the (k − 1)-th MPC arrived at known Under the assumption of an exponential distribution, it is
delay Tc,k−1 , is written as: first investigated if the mean waiting time between clusters
1

Tc,k − Tc,k−1
 differs between LoS and nLoS measurements. This done by
pexp (Tc,k | Tc,k−1 ; λc ) = exp − (5) executing the two-sample Anderson-Darling (AD) test, which
λc λc
assesses if τc − τc−1 grouped according to LoS or nLoS could
In (5), the exponential distribution has the parameter λc which both originate from the same statistical distribution. This test
corresponds to the mean waiting time between consecutive results in a p-value of 0.04, which is borderline significant
MPCs in cluster c (scale parameter). An additional distributi- at the 5% level and prompts us to distinguish between LoS
onal parameter θc is defined as the delay of the first arriving and nLoS. Next, for LoS and nLoS separately, composite
path in cluster c, i.e., θc = Tc,1 (location parameter). exponentiality of τc − τc−1 is verified using the one-sample
The plausibility of an exponential distribution for the arrival AD test. An exponential distribution is accepted for both LoS
times Tc,k is then validated by executing an Anderson-Darling and nLoS at the 5% significance level (p-values of 0.13 and
(AD) goodness-of-fit test for composite exponentiality. For the 0.12, respectively). The mean of waiting times τc − τc−1 is
45 clusters in the measurement campaign, the minimum, aver- estimated at 2.30 ns for LoS and 1.21 ns for nLoS.
age, and maximum p-values associated with the AD test are 3) Cluster mean power pc : Significant correlation is found
equal to 0.06, 0.40, and 0.92, respectively. This means that, at between cluster mean power pc and cluster onset τc : Spear-
the 5% significance level, all 45 clusters retain exponentiality. man’s rank correlation coefficient is equal to −0.80 for LoS
and −0.58 for nLoS, both are significant at the strict 1% level
VI. S TATISTICS OF THE DISTRIBUTIONAL PARAMETERS
with p-values of 1.8 · 10−4 for LoS and 9.7 · 10−4 for nLoS.
This section models the between-cluster and within-cluster The Saleh-Valenzuela model proposes a linear decrease of the
propagation parameters as defined in (3). The propagation average pc of MPC powers in dB with the cluster onset τc in
parameters are fully determined by the distributional para- ns [6]:
meters of the location-scale distributions of Section V. In
the following, the between-cluster propagation parameters are pc = a0 + a1 · τc + a2 · Dc + a3 · τc · Dc + ǫc (8)
identified with the location parameters of these distributions, In the linear model (8), pc is made dependent on τc and the
i.e., for cluster c: dummy variable Dc . The value of Dc is one for clusters stem-
φA A
φD D ming from LoS measurements and is zero for nLoS clusters.
c , αc c , αc τc , θ c pc , µc (6)
As such, Dc accounts for possible changes in the intercept and
The within-cluster propagation parameters are characterized slope of (8) between LoS and nLoS situations. Furthermore, a0
by the scale parameters of the distributions, i.e., for the MPCs through a3 are regression parameters, and the term ǫc denotes
in cluster c: the model’s error for cluster c and is generally assumed to
be zero-mean normally distributed. The regression parameters
φA A
c,k → κc φD D
c,k → κc τc,k → λc pc,k → σc (7)
in (8) are estimated using a backward elimination procedure:
In the following, the statistics of the distributional pa- simple t-tests are carried out on a0 through a3 to determine
rameters are discussed. In this section, distinction is made which of these regression parameters can assumed to be zero at
between distributional parameters originating from LoS and the 5% significance level. The backward elimination procedure
nLoS measurements. resulted in the following estimated regression parameters:

A. Location parameters (between-cluster) a0 = −20.14 a1 = −0.81 a2 = 0 a3 = 0 (9)

1) Cluster angular means φA D
c and φc : The suitability The standard deviation of ǫc in (8) is estimated at 4.72 dB.
of a uniform distribution in (−π, π] for modeling φA c and The coefficient of determination of the fitted model is equal
φD
c is investigated. No distinction is made between LoS and to 0.42. In (9), it is noted that the regression parameters a2
nLoS, as the uniform distribution is not parameterized by any and a3 associated with the dummy variable Dc are assumed
to be zero by the backward elimination procedure. This means statistical distribution. A two-sample AD test on λc grouped
that the form of the exponential and power law models is not according to LoS or nLoS results in a p-value of 0.19,
significantly different between LoS and nLoS measurements. indicating no significant difference between LoS and nLoS
Fig. 4 shows a scatter plot of pc versus τc along with the fitted at the 5% level. Next, normality for log (λc ) without making
linear model. distinction between LoS and nLoS is considered: AD, SW, and
HZ hypothesis tests accepted normality at the 5% level with
−10
p-values of 0.13, 0.21, and 0.13, respectively. We therefore
assume a normal distribution for log (λc ): the sample mean
−15 and sample standard deviation of log (λc ) are equal to 0.03
and 0.35, respectively.
−20 3) Cluster standard deviation of power σc : For σc (in dB),
a two-sample AD test decides there is no significant change in
p [dB]

the statistical distribution of this parameter between LoS and

−25
nLoS measurements (p-value of 0.34). Normality for log (σc )
c

is assessed with the AD, SW, and HZ hypothesis tests, all of

−30 which accepted normality at the 5% level (p-values of 0.61,
−20.14 −0.81 ⋅ τ [ns]
c
0.78, and 0.41, respectively). The sample mean and sample
−35
standard deviation of log (σc ) are equal to 0.88 and 0.14,
respectively.

−40 VII. C ONCLUSIONS

0 2 4 6 8 10 12 14 16 18
τ [ns] In this paper, the statistics of propagation path parameters
c
including their azimuth of arrival, azimuth of departure, delay,
Fig. 4. Scatter plot of pc versus τc and fitted linear model and power, are determined in an indoor office environment.
Path parameters are grouped into clusters. Statistical distri-
butions and correlations are determined for the path parame-
B. Scale parameters (within-cluster) ters within individual clusters and for their change between
To our knowledge, no examples of possible statistical dis- clusters. As validation for the distributional choices, statistical
tributions for the scale parameters exist in literature. We will goodness-of-fit tests are used.
therefore use the entropy-maximizing normal distribution to ACKNOWLEDGMENT
model these parameters. As the scale parameters can only take
on positive values, they are first log-transformed to match the W. Joseph is a Post-Doctoral Fellow of the FWO-V (Re-
support of the normal distribution (i.e., any positive or non- search Foundation - Flanders). This work was carried out
positive number). within the frame of CISIT (International Campus on Safety
1) Cluster angular concentrations κA D and Intermodality in Transportation) and with the support of
c and κc : It is first
investigated if the statistical distribution of κc (κD
A the FEDER funds, the French Ministry of research and the
c ) differs
significantly between LoS and nLoS measurements. For this, Region Nord Pas-de-Calais (France).
the two-sample Anderson-Darling (AD) test is used on obtai- R EFERENCES
ned values of κA D
c (κc ) grouped according to LoS or nLoS. For
A D [1] L. M. Correia, Mobile Broadband Multimedia Networks - Techniques,
both κc and κc , this test detects no difference between LoS Models, and Tools for 4G, 1st ed. Elsevier Ltd., 2006.
and nLoS distributions at the 5% significance level (p-values [2] M. Haardt, “Efficient One-, Two-, and Multidimensional High-Resolution
of 0.16 and 0.20, respectively). Without making distinction Array Signal Processing,” Ph.D. dissertation, Technische Universität
München, Shaker Verlag GmbH, Aachen, DE, 1996.
between
LoS and nLoS,
the assumptions of normality for [3] T. J. Shan, M. Wax, and T. Kailath, “On Spatial Smoothing for Direction-
log κA c and log κD c are validated using the Anderson- of-Arrival Estimation of Coherent Signals,” IEEE Transactions on Acous-
Darling (AD), Shapiro-Wilk (SW), and Henze-Zirkler (HZ) tics, Speech, and Signal Processing, vol. 33, no. 4, pp. 806–811, August
1985.
tests. For log κA


c , all three tests accepted normality at the [4] N. Czink, P. Cera, J. Salo, E. Bonek, J.-P. Nuutinen, and J. Ylitalo,
5% level with p-values of 0.37 (AD), 0.46 (SW), and 0.31 “A Framework for Automatic Clustering of Parametric MIMO Channel
(HZ). The Data Including Path Powers,” in IEEE Vehicular Technology Conference,

sample mean and sample standard deviation of Montréal, CA, September 2006, pp. 1–5.
log κA c are equal to 0.50 and 0.33, respectively. Furthermore, [5] D.-J. Kim, Y.-W. Park, and D.-J. Park, “A Novel Validity Index for
normality is also accepted for log κD


c with p-values of Determination of the Optimal Number of Clusters,” IEICE Transactions
0.09 (AD), 0.14 (SW), and 0.59 (HZ). The sample mean on Information and Systems, vol. E84-D, no. 2, pp. 281–285, February
2001.
and standard deviation of log κD


c equal 0.36 and 0.32, [6] A. A. M. Saleh and R. A. Valenzuela, “A Statistical Model for Indoor
respectively. Multipath Propagation,” IEEE Journal on Selected Areas in Communica-
2) Cluster mean waiting time between MPCs λc : It is tions, vol. 5, no. 2, pp. 128–137, February 1987.
first assessed whether λc (in ns) originating from LoS or
nLoS measurements could have been drawn from the same