RADIOENGINEERING, VOL. 19, NO.

4, DECEMBER 2010 695

Estimation of the Radio Channel Parameters

using the SAGE Algorithm
Susana MOTA.1, Maura OUTEIRAL GARCIA 2, Armando ROCHA 1, Fernando PEREZ-FONTAN 2
1
Dept. of Electronics, Telecommunications and Informatics, Instituto de Telecomunicações, University of Aveiro
Campus Universitário de Santiago, 3800 – 193 Aveiro, Portugal
2
Dept. of Signal Theory and Communications, University of Vigo, 36310 Vigo, Spain

smota@ua.pt, mau.outeiral@ua.pt, arocha@av.it.pt, fpfontan@tsc.uvigo.es

Abstract. This paper presents the problem of estimating waves in mobile radio environments. In particular, the
the parameters of a given number of superimposed signals, SAGE (Space-Alternating Generalized Expectation-
as is the case of the received signal in wireless com- Maximization) algorithm simplifies the complex multi-
munications. Based on the description of the received dimensional optimization problem, such as estimating the
signal in the frequency domain, one version of the SAGE parameters of these waves in a multipath propagation
(Space-Alternating Generalized Expectation- environment, to several separate one-dimensional
Maximization) algorithm is presented, allowing the optimization processes which can be performed
estimation, for each impinging ray, the delay, azimuth, sequentially. This algorithm, derived in its general form in
elevation and complex amplitude. Ray retrieval results are [4], is an extension to the Expectation-Maximization (EM)
presented in synthetic channels, using data generated with algorithm [5] and has been used in areas like image
the extended Saleh-Valenzuela (ESV) model, and also in reconstruction [6].
real channels.
In the context of array signal processing, comparative
convergence studies of the EM and SAGE algorithms
applied to Angle of Arrival (AoA) estimation may be
Keywords found in [7] using synthetic data, and in [8] using measured
sonar data. Concerning the wireless communications
Parameter estimation, SAGE algorithm, radio channel context, the SAGE algorithm has been used for joint delay,
measurements, multipath components. azimuth and Doppler frequency estimation in time-variant
channels [9], [10], as well as, for joint delay, azimuth and
elevation estimation in time-invariant channels [11], [12].
1. Introduction In this work, we present one version of the SAGE
algorithm in the frequency domain, allowing the estimation
Research, design and analysis of advanced wireless of the delay, azimuth, elevation and complex amplitude for
communications systems require a detailed understanding a given number of superimposed electromagnetic waves.
of the electromagnetic wave propagation phenomena in Estimation accuracy and reliability is investigated using
mobile radio environments. System performance both synthetic and measured data.
evaluation yields representative results only if the
underlying channel models reflect the most relevant The paper is organized as follows. To start, in Section
features of the physical channel. An important effort has 2, some definitions, notations and the signal and channel
been devoted to the channel’s directional information: how models are established. Then, in Section 3, EM based
many are the most important waves arriving at the antenna, estimation and the SAGE algorithm are presented. Ray
from where and with what amplitude and delays do they retrieval results in synthetic channels and also in real
come [1], [2], [3]. The development of these new channel channels appear, respectively, in Section 4 and Section 5.
models relies on a realistic characterization of the Finally in Section 6 the conclusions are presented.
probability distribution of several parameters, requires
extensive measurements on a wide range of scenarios and
validation against theoretical predictions. Therefore, 2. Signal Model
suitable and computationally effi-cient processing tools
In the considered underlying model [12], a finite
need to be employed in order to extract the parameters of
number, L, of plane waves are impinging at the receiver
interest from the measured data.
antenna array with M elements and the channel is assumed
A few high-resolution algorithms have been proposed time-invariant. The channel impulse response at the mth
and used to estimate the parameters of the impinging antenna element can be expressed as
696 S. MOTA, M. OUTEIRAL GARCIA, A. ROCHA, F. PEREZ-FONTAN, ESTIMATION OF THE RADIO CHANNEL PARAMETERS ...

L
 2π  to express the maxima of the log-likelihood function used
hm t ,  ,       exp j rm , e ,     t     (1)
 1  λ  by the MLE. Even taking into consideration that the values
of the complex amplitudes may be expressed as a function
where:  represents the time delay,  the incidence of the other parameters, the procedure to obtain the MLE
azimuth,  the incidence elevation (measured with respect represents a L 3-D non-linear optimization process [9].
to the horizontal plane) and  the complex amplitude of
the -th wave; λ denotes the wavelength and , the scalar The EM algorithm [5] has been developed to address
product; rm is a row vector containing the m-th antenna the MLE problem when a part of the observations is
element coordinates and missing or suppressed. It is based on two key concepts: the
complete data (unobservable) and the incomplete data
e ,    cos  cos  , cos  sin  , sin   (2)
T
(observable), allowing the decomposition of the above
procedure in L 3-D optimization procedures to estimate the
is the unit vector in IR3 pointing toward the direction waves’ parameters, which may be performed separately
defined by  and , where  denotes matrix transposition.
T
and in parallel. Each 3-D optimization procedure aims to
In (1), the expression obtain the parameters of a given wave only. In our
problem, a possible choice for the complete data set is the
 2π  contribution of each individual wave to the channel
cm  ,     exp j rm , e ,     (3)
 λ  transfer function, corrupted by a fraction of the additive
noise, i.e.,
accounts for the phase shift suffered by the th wave due to
a small difference in the traveled distance to reach the mth X   f   S  f ;      N  f  (8)
antenna element. The vector
where, μ, =1,…L, must satisfy     1 . The vector
c ,    c1  ,  ,..., cM  ,   (4)
T
containing the parameters of the th wave, θ, constitutes
one parameter subset. On the other hand, the measured
is the so called array steering vector.
(observed) channel transfer function, H ( f ) , represents the
Defining      ,  , ,    as being the vector which incomplete data set.
contains the parameters of the -th wave, the contribution
of this wave to the M impulse responses may be expressed
as 3.2 Description of the SAGE Algorithm
ht ;     h1 t ;   ,..., hM t ;       c ,    t     . (5)
T The SAGE algorithm may be viewed as an extension
of the EM algorithm: each one of the SAGE iterations is, in
Alternatively, in the frequency domain, the channel fact, an EM iteration to update just a subset of the
transfer matrix across the array, possibly corrupted with components of θ, maintaining the parameters of the other
noise is given by components fixed at their previous values. It replaces the L
L 3-D parallel optimization procedures, used in the EM al-
H  f ;      c ,    exp j2π  f   N  f  (6)
 1
gorithm, by a serial optimization approach. As a result ac-
cording to [4] and [9], in comparison to the EM algorithm,
with N  f  denoting a M-dimensional vector of complex
SAGE algorithm presents faster convergence and lower
white Gaussian noise and   1 ,...,  L  . The contribution complexity. Although, according to [7] this may not be
of the th wave to the channel transfer function is denoted always true and the faster convergence can only be
as guaranteed if certain conditions are fulfilled. Again, the
S  f ;      c ,    exp j2π  f  . (7) complete data set is chosen to be the contribution of each
wave to the channel transfer function as given in (8) but
choosing μ=1. The L complete data sets, X  ( f ) , =1,…L,
are independent, therefore the components X  ' ,    are
3. Estimation of Superimposed Signals of no importance for the estimation of θ. The log-
using the SAGE Algorithm likelihood function of θ, given an observation
X   f   X obs  f  , is
3.1 Maximum-Likelihood Estimation and the
EM Algorithm    
   ; X obs  2   S H  f ;  X   f  df    S H  f ;   df  (9)
2

The problem to solve is the estimation of channel where  H

 denotes the Hermitian operator. The MLE of θ
parameters, i.e., to obtain the L components of vector θ. is given as
The maximum likelihood estimation (MLE) of this vector
is computationally prohibitive owing to its high dimension ˆ  X  arg max ; X 
 ML
obs
 
obs

(10)
when L is large and also because there is no closed formula 
RADIOENGINEERING, VOL. 19, NO. 4, DECEMBER 2010 697

where ˆ is the vector that contains the estimate of the th 3.3 Initialization of the SAGE Algorithm
wave parameters. Beginning with the pre-initial setting ˆ  0,...,0, the
Inserting (7) in (10) and approximating the integral by initial estimates for each =1,…L, are obtained according
a sample summation, we may write to

ˆ ,ˆ , ˆ  X   arg max{ | z( , ,  ; X 

  
M

   ML
obs

obs
 ) |} ˆ  arg max  Xˆ  f ;ˆ exp j2πf   , (18)
[ , ,  ]
(11)   m 1 

ˆ  ML X obs   1

MN

z ˆ , ˆ , ˆ  X  ; X 
ML
obs

obs

(ˆ, ˆ)  arg max{ | z (ˆ, ,  ; Xˆ  ( f ;ˆ)) | }
[ ,  ]
(17)

with N the number of samples in the frequency domain and and (17) to obtain ˆ  .

 
z  ,  ,  ; X obs  c H  ,  X obs  f  exp j2πf  . (12) In (18) the term inside the summation expresses
frequency correlation. It is used as a method to obtain the
Since all signals are superimposed on the available initial delay estimate since at this point ˆ and ˆ  are
(measured) signal, X  ( f ) cannot be observed, so we may unknown. The 2-D optimization in (19) is used instead of
 
try to obtain its estimate, Xˆ  f ;ˆ , given the observation (15) and (16) because according to [11] assuming ˆ  0
of H ( f ) and the previous estimate ˆ of  . This can be may cause an erroneous azimuth estimation.
done by removing the contribution of all waves, except the
th wave, from the observation, i.e.,
4. Results using Synthetic Data
   
L
Xˆ  f ;ˆ  H  f    S f ;ˆ . (13)
Synthetic data have been generated using the ex-
 1
   tended Saleh-Valenzuela (ESV) model. The ESV model
Additional complexity reduction may be achieved, characterizes complex amplitude, time of arrival (ToA),
angle of arrival (AoA) [1] and angle of departure (AoD)
within the SAGE algorithm framework, by further
[13] for each multipath component (MPC). This model
decomposing the optimization procedure. Each subset θ is assumes that rays (or MPCs) arrive at the receiver in
split into three subsets: [τ,α], [ ,α] and [β,α] and the clusters and also that they have different statistical
MLE is obtained for the parameters in each subset while distributions for each of the parameters. If we discard de
maintaining the parameters in other sets fixed. As already AoD, like in [1], the impulse response may be described by
mentioned, the MLE of α may be expressed as a function h ,    kl exp j kl    Tl   kl     l  kl  (20)
of [τ, ,β], so that the 3-D optimization procedure in (11) l k

reduces to 3 1-D optimization procedures. The update where index l indicates the cluster number and index k
procedures needed to obtain a new estimate for the indicates the echo number within a cluster.
parameters of the th wave, ˆ , given the previous The amplitude, αkl, follows a zero-mean Complex-
estimates of all waves, ˆ , can then be written as Normal distribution whose deviation is described by an
exponential decay with two time constants: one of them is
ˆ  arg max{ | z ( , ˆ, ˆ; Xˆ  ( f ;ˆ)) | } , (14) associated to the clusters (Γ) and the other to the rays

within a cluster (γ). The AoA follows a Laplacian dis-
ˆ  arg max{ | z (ˆ,  , ˆ; Xˆ  ( f ;ˆ)) | } , (15) tribution around a nominal cluster angle that is assumed to
 be uniformly distributed in the interval [0,2π]. The ToA is
modeled by a Poisson process with two arrival rates: again,
ˆ  arg max{ | z (ˆ, ˆ,  ; Xˆ  ( f ;ˆ)) | } , (16)
one is associated to the clusters and other to the rays.


1 We aim to investigate the SAGE capability to retrieve

ˆ   z (ˆ, ˆ, ˆ; Xˆ  ( f ;ˆ)) | } . (17)
the superimposed signals, that is: how the number of MPCs
MN
in the channel and their relative power influence the quality
Like in the original EM algorithm, the SAGE basic of the solution obtained. Thus, we have generated several
iteration stage that updates the parameters of a given wave sets of data with different combinations for number of
comprises two steps: the E-step (Expectation), given in clusters (NC) – number of rays in each cluster (NR) and
(13), aims to obtain the expected complete data set; and the also with different combinations for the time constants
M-step (Maximization), given in (11), performs the controlling the power decay (Γ and γ). The arrival rates
estimation of each parameter of the considered wave. were the same for all sets. We have selected three of these
698 S. MOTA, M. OUTEIRAL GARCIA, A. ROCHA, F. PEREZ-FONTAN, ESTIMATION OF THE RADIO CHANNEL PARAMETERS ...

sets to present and the corresponding parameters, used in

the ESV model, can be found in Tab. 1. Set Name NC NR Γ (ns) γ (ns)
ch3 5 10 60 20
This data is then used to obtain the matrix of channel ch4 3 5 90 30
transfer functions which is computed according to (6) but ch9 5 10 120 40
without the added noise. We assumed: 801 frequency Tab. 1. Parameters for the ESV model used to generate the data sets.
samples in a 200 MHz bandwidth centered on 2 GHz and
a uniform rectangular array with 11 x 11 elements spaced
by 0.5 λ in both dimensions. Some results may be observed
in Fig. 1, Fig. 2 and Fig. 3.

FD-SAGE ESTIMATES: total number of rays

0
FD-SAGE Estimates
Synthetic Data
-5

-10
Amplitude(dB)

-15

-20

-25

-30

-35
-20 0 20 40 60 80 100 120 140
Delay(ns)

Fig. 1. SAGE retrieval results for a channel with 15 rays (ch4), moderate power decay and 15 estimates requested. Left: Generated impulse
response and SAGE retrieval in time domain. Center: Generated impulse response. Right: SAGE retrieved impulse response. – Both
in time and azimuth domains.

FD-SAGE ESTIMATES: total number of rays

0
FD-SAGE Estimates
-5 Synthetic Data

-10

-15
Amplitude(dB)

-20

-25

-30

-35

-40
0 50 100 150 200
Delay(ns)

Fig. 2. SAGE retrieval results for a channel with 50 rays (ch9), moderate power decay and 15 estimates requested. Left: Generated impulse
response and SAGE retrieval in time domain. Center: Generated impulse response. Right: SAGE retrieved impulse response. – Both
in time and azimuth domains.

FD-SAGE ESTIMATES: total number of rays

0
FD-SAGE Estimates
Synthetic Data
-10

-20
Amplitude(dB)

-30

-40

-50

-60

-70
0 50 100 150 200 250
Delay(ns)

Fig. 3. SAGE retrieval results for a channel with 50 rays (ch3), pronounced power decay and 15 estimates requested. Left: Generated impulse
response and SAGE retrieval in time domain. Centre: Generated impulse response. Right: SAGE retrieved impulse response. – Both
in time and azimuth domains.
RADIOENGINEERING, VOL. 19, NO. 4, DECEMBER 2010 699

Observing the impulse responses of the sets presented

here, we may conclude that in two of them – Fig. 1 and
Fig. 2 – although the number of MPCs is very different,
they present almost the same amplitude range (from 0 dB
to about -30 dB) and in Fig. 3 the amplitude range is wider
(from 0 dB to about -60 dB). Thus, we have classified the
first two cases as “moderate” and the last one as
“pronounced” power decay.
From the results shown we can conclude that if the
channel presents a small number of rays and moderate
power decay (Fig. 1), the SAGE algorithm is able to Fig. 4. Block diagram of the measurement system.
retrieve a good estimate for almost all the rays. In this
The frequency response of the time-invariant channel
particular case, notice that only the last two rays, which are
simultaneously the weakest ones, were not correctly was measured at M=11x11 positions spaced by /2 in both
estimated. Instead, two inexistent and low amplitude rays dimensions. The RF bandwidth was 200 MHz centered at
are placed near the first and more powerful rays. Average 2 GHz. An elevated disconic antenna was used at the
delay and delay spread for the considered channel and the transmitter (Tx) location and a /4 monopole antenna was
respective SAGE retrieval have been computed and com- used at the receiver (Rx) location. The monopole was
pared. For this case errors were less than 1 % for both moved along the M positions of the virtual rectangular
parameters, showing that, despite the failure in the esti- antenna array. Fig. 5 shows the outdoor scenarios where
mation of those two rays, the retrieved waves represent the measurements presented below have taken place. The
a good description of this channel. lines represent buildings with heights varying from 10 m to
15 m. The Tx height was about 7.5 m and the Rx height
As the number of rays in the channel increases, the was about 1.3 m.
number of rays whose estimate is missed increases and
In order to obtain the parameters’ estimates for a giv-
therefore, the number of false rays retrieved also increases.
en number of waves, the matrix of frequency responses
In the case of Fig. 2, although the retrieved rays still
measured was used as input to the SAGE algorithm.
provide a good description of the channel (average delay
Aiming the investigation on the algorithm’s behavior, we
error of 1 % and delay spread error about 2.4 %), we have
must have a previous idea of the channel response to be
missed the estimate for 17 rays (out of a total of 50) and,
able to judge the retrieved results. Therefore, in order to
obviously, there are 17 false echoes in the solution given
make such judgment, ray-tracing simulations were per-
by the algorithm. However, as the power decay becomes
formed for the measurement scenarios and compared with
more pronounced the phenomenon also becomes more
the algorithm’s results. This also provides us with the
severe and thus, more evident. In the case of Fig. 3, the
necessary insight to use the algorithm in situations in
number of missed rays has increased to 24 (almost 50 % of
which we do not have a clear idea on the channel’s
the rays in the channel). Nevertheless, in general, the
response.
missed rays are the weakest ones so that the impact is not
as bad as one could expect at first glance (for this case, Nevertheless, we must bear in mind that results from
average delay error is 3.2 % and delay spread error about ray tracing simulations may not be completely perfect. As
6 %). a matter of fact, they depend on a rigorous description of
the scenario dimensions and properties, the latter not
always easy to obtain. Also, the more accurate are the
results desired, the more complex and time consuming the
5. Results on Real Channels simulation process will be, because it has to consider
The measurement system, shown schematically in several physical phenomena (direct ray, reflection,
Fig. 4, consists of a 2D positioning device, driven by diffraction, transmission). In fact, if we compare the
stepper motors, equipped with one antenna, a vector net- simulated time domain impulse response of the channel
work analyzer (VNA) and a personal computer that is used (obtained by ray tracing tools) with the amplitude averaged
to control the equipment, acquire and save experimental time domain impulse response given by the measurements
data. The VNA is used to measure the frequency response (obtained from the measured frequency response by
of the time-invariant channel at the M locations of a virtual applying an IFFT), sometimes we detect some simulated
rectangular antenna array. The computer controls the strong echoes at delays that are not present in the measured
positioning device and the VNA through the use of data and vice-versa.
a commercial stepper motor control card and a GPIB The number of MPCs may be estimated, from ob-
interface, respectively. The software needed to control all served data, by applying well-known information theoretic
the equipment, acquire and save experimental data was criteria, namely, the Akaike information criterion (AIC)
implemented in LabVIEW. and the minimum description length (MDL) [14], for
700 S. MOTA, M. OUTEIRAL GARCIA, A. ROCHA, F. PEREZ-FONTAN, ESTIMATION OF THE RADIO CHANNEL PARAMETERS ...

which several performance studies and improvements have the number of rays to be requested from the SAGE
been reported [15], [16], [17], since these were first pro- algorithm, L, has been manually chosen by performing
posed. Nevertheless, one must take into account the be- several attempts (trying different values) and analyzing the
havior presented by the SAGE algorithm in the previous averaged time domain impulse response of the measured
section, showing that, even in the absence of noise, the channel, the simulations and the output results of the
algorithm misses some of the rays undergoing longer SAGE algorithm. Results may be observed in Fig. 6 and
delays and supplies, in their place, false rays. As a result, Fig. 7.

40 90

35 80

70
30
y y
x 60 x
25
50
(m)

20

(m)
40
15
30

10
20 Rx Tx

5 Rx
10

Tx
0 0
0 20 40 60 80 0 20 40 60 80 100
(m) (m)

Fig. 5. Layouts of the measurement scenarios – Left: Ens5. Right: Ens6.

Impulsive response
-60
Mesuared impulsive response
FD-SAGE Estimates
-65

-70

-75
IR(dB)

-80

-85

-90

-95

-100

80 100 120 140 160 180 200 220 240 260 280 300
Delay(ns)

Fig. 6. Results for Ens5 with L = 21. Left: Measured channel and SAGE retrieved impulse responses (time domain). Centre: SAGE retrieved
impulse response (time-azimuth domain). Right: Ray tracing simulations impulse response (time-azimuth domain).

Impulsive response
-60
Mesuared impulsive response
-65 FD-SAGE Estimates

-70

-75

-80
IR(dB)

-85

-90

-95

-100

-105
0 50 100 150 200 250 300 350 400 450
Delay(ns)

Fig. 7. Results for Ens6 with L = 23. Left: Measured channel and SAGE retrieved impulse responses (time domain). Centre: SAGE retrieved
impulse response (time-azimuth domain). Right: Ray tracing simulations impulse response (time-azimuth domain).
RADIOENGINEERING, VOL. 19, NO. 4, DECEMBER 2010 701

The results in Fig. 6 and Fig. 7 were obtained by References

requesting the SAGE algorithm to provide estimates for 21
and 23 rays, respectively. We were able to retrieve almost [1] SPENCER, Q. H., JEFFS, B. D., JENSEN, M. A.,
all the contributions that are no more than about 25 dB SWINDLEHURST, A. L. Modeling the statistical time and
angle of arrival characteristics of an indoor multipath channel.
below the strongest component. Again, we had computed IEEE Journal on Selected Areas in Communications, 2000, vol.
the average delay and delay spread, using the meas- 18, no. 3, p. 347 – 360.
urements and the corresponding SAGE retrieval as well.
[2] STEINBAUER, M., MOLISCH, A. F., BONEK, E. The double-
When using the averaged time domain impulse response directional radio channel. IEEE Antennas and Propagation
obtained from measurements, the computation has con- Magazine, 2001, vol. 43, no. 4, p. 51 – 63.
sidered only the components which are no more than 25 dB [3] CHONG, C. C. et al. A new statistical wideband spatio-
below the highest peak. Averaged delays from SAGE temporal channel model for 5-GHz band WLAN systems. IEEE
retrieved data were within a maximum error of 0.3 % in Journal on Selected Areas in Communications, 2003, vol. 21,
agreement with the measured data, while for delay spreads, no. 2, p. 139 – 150.
for the case of Fig. 6, we obtained an error of 5.8 % and for [4] FESSLER, J. A., HERO, A. O. Space-alternating generalized
the case of Fig. 7, an error of 9.0 %, which are both expectation-maximization algorithm. IEEE Transactions on
reasonable. Signal Processing, 1994, vol. 42, no. 10, p. 2664 – 2677.
[5] FEDER, M., WEINSTEIN, E. Parameter-estimation of super-
imposed signals using the EM algorithm. IEEE Transactions on
Acoustics Speech and Signal Processing, 1988, vol. 36, no. 4,
6. Conclusions p. 477 – 489.
[6] FESSLER, J. A., HERO, A. O. Penalized maximum-likelihood
This paper has dealt with the problem of estimating
image-reconstruction using space-alternating generalized EM
the parameters of a given number of superimposed signals algorithms. IEEE Transactions on Image Processing, 1995,
using the SAGE algorithm. Ray retrieval results are vol. 4, no. 10, p. 1417 – 1429.
presented both in synthetic channels, using data generated [7] CHUNG, P. J., BOHME, J. F. Comparative convergence
with the ESV model, and also in real channels. analysis of EM and SAGE algorithms in DOA estimation. IEEE
Transactions on Signal Processing, 2001, vol. 49, no. 12,
From results in synthetic channels, we concluded that p. 2940 – 2949.
if the channel presents a small number of MPCs and
[8] CHUNG, P. J., BOHME, J. F. Experimental study of the EM
moderate power decay, the SAGE algorithm is able to and SAGE algorithms with application to sonar data. In Sensor
produce a good estimate for almost all MPCs. However, Array and Multichannel Signal Processing Workshop
sometimes, the algorithm misses some of the most delayed Proceedings. Rosslyn (USA), 2002, p. 77 – 81.
MPCs and provides, in their place, some false MPCs. As [9] FLEURY, B. H., TSCHUDIN, M., HEDDERGOTT, R.,
the number of MPCs in the channel and the power decay DAHLHAUS, D., PEDERSEN, K. I. Channel parameter
increases, the number of MPCS whose estimate is missed estimation in mobile radio environments using the SAGE
increases and therefore, the number of false MPCs re- algorithm. IEEE Journal on Selected Areas in Communications,
1999, vol. 17, no. 3, p. 434 – 450.
trieved also increases. In general the missed MPCs are the
weakest ones and thus, the impact is not as critical as one [10] CHONG, C. C. et al. Joint detection-estimation of directional
channel parameters using the 2-D frequency domain SAGE
could expect. algorithm with serial interference cancellation. In ICC 2002,
IEEE International Conference on Communications, vol. 2.
The SAGE algorithm presents, in real channels, New York (USA), 2002, p. 906 – 910.
a similar behavior to the one presented with synthetic data.
As a result, the number of rays that is requested from the [11] TSCHUDIN, M., HEDDERGOTT, R., TRUFFER, P. Vali-
dation of a high resolution measurement technique for
algorithm must be carefully chosen. We must be aware that estimating the parameters of impinging waves in indoor
some of the outputted MPCs may be false and, if we do not environments. In Ninth IEEE International Symposium on
have previous knowledge of the channel properties, we Personal, Indoor and Mobile Radio Communications. Boston
may not distinguish them easily. Nevertheless, false rays (USA), 1998, p. 1411 – 1416.
are likely to show parameters very similar to their [12] MOTA, S., ROCHA, A. Parameter estimation of synthetic and
neighbors, as if they were repeated. real radio channels using the SAGE algorithm. In 13th IEEE
International Symposium on Personal, Indoor and Mobile
Radio Communications Proceedings - Sailing the Waves of the
Wireless Oceans. Lisboa (Portugal), 2002, p. 634 – 638.
[13] WALLACE, J. W., JENSEN, M. A. Modeling the indoor
Acknowledgements MIMO wireless channel. IEEE Transactions on Antennas and
Propagation, 2002, vol. 50, no. 5, p. 591 – 599.
The authors would like to thank to the technicians [14] WAX, M., KAILATH, T. Detection of signals by information
from the Instituto de Telecomunicações for help with the theoretic criteria. IEEE Transactions on Acoustics Speech and
measurements. Signal Processing, 1985, vol. 33, no. 2, p. 387 – 392.
702 S. MOTA, M. OUTEIRAL GARCIA, A. ROCHA, F. PEREZ-FONTAN, ESTIMATION OF THE RADIO CHANNEL PARAMETERS ...

channel propagation, mainly concerning, directional

[15] ZHANG, Q. T., WONG, K. M., YIP, P. C., REILLY, J. P. channel characterization and clustering algorithms.
Statistical analysis of the performance of information theoretic
criteria in the detection of the number of signals in array Armando ROCHA was born in Laguaira, Venezuela. He
processing. IEEE Transactions on Acoustics, Speech and Signal obtained his degree in telecommunication and electronics
Processing, 1989, vol. 37, no. 10, p. 1557 – 1567. Engineering in 1984 from the University of Aveiro and his
[16] FISHLER, E., GROSMANN, M., MESSER, H. Detection of PhD from the same university in 1996. He is an auxiliary
signals by information theoretic criteria: general asymptotic professor at the Department of Electronics, Telecom-
performance analysis. IEEE Transactions on Signal Processing, munications and Informatics at the same university. He
2002, vol. 50, no. 5, p. 1027 – 1036.
lectures in wave propagation, guided propagation, antennas
[17] NADLER, B. Nonparametric detection of signals by and cellular planning. He has been involved in Earth-
information theoretic criteria: performance analysis and an Satellite microwave propagation activities developing
improved estimator. IEEE Transactions on Signal Processing,
2010, vol. 58, no. 5, p. 2746 – 2756. beacon receivers, doing propagation measurements and
channel modeling. He was in Management Committee
Member of EU COST 280 and currently participates with
the same role in COST IC0802. He also participated in
About Authors … several public funded projects.
Fernando PEREZ FONTAN was born in Villagarcía de
Susana MOTA was born in Anadia, Portugal. She
Arosa, Spain. He obtained his degree in telecom-
received her B.Sc. (5 years course) and M.Sc. degrees from
munication engineering in 1982 from the Technical
the University of Aveiro, Portugal, both in electronics and
University of Madrid and his PhD in 1992 from the same
telecommunications, in 1999 and 2003, respectively.
university. After working in industry since 1984 he became
Currently she is working towards the Ph.D. degree at the
an assistant professor at the University of Vigo in 1988. In
same university where she is also an invited lecture at the
1993 he became an associate professor and in 1999 a full
Department of Electronics, Telecommunications and
professor at the Signal Theory and Communications
Informatics. Her research interests are in wireless com-
Department of the University of Vigo. He lectures in Ra-
munications, directional channel characterization and
diocommunication Systems, especially in terrestrial fixed
MIMO channel modeling. She is member of EU COST
and mobile system related topics. He is the author of
IC0802 action.
a number of books and journal papers and has been the
Maura OUTEIRAL GARCIA was born in Boiro, Spain. leader in several projects funded by public and private
During this work, she was an ERASMUS student on entities. He participates in ITU-R WG3 on Propagation
a leave from University of Vigo, Spain, working towards Modeling. He has been a Management Committee Member
the M.Sc. degree in telecommunications, at University of of EU' COST 255, 280 and 297, and currently participates
Aveiro, Portugal. She has been involved in wireless with the same role in COST IC0802.