Professional Documents
Culture Documents
net/publication/3160078
CITATIONS READS
94 709
2 authors, including:
T.s. Ng
The University of Hong Kong
267 PUBLICATIONS 4,508 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by T.s. Ng on 25 November 2014.
URL http://hdl.handle.net/10722/42844
Abstract—This paper proposes a mathematical model for simu- channel (e.g., [3]–[8]), techniques for simulating an fading
m m
lating a Nakagami- fading channel, < 1. It is shown that the m channel for are relatively few [8], [9].
fading process can be expressed as a product of a square-root-beta In Section II-A, we propose a mathematical model that en-
process and a complex Gaussian process. Extension of the proposed
model to simulate correlated diversity channels is also given. Nu- ables one to generate a complex random process for simulating
merical results demonstrate that statistical properties of the sam- an fading channel, . Emphasis is given to appli-
ples generated from the proposed model are close to the required cations of the proposed model for HF communications. Since
ones. the original fading channel model [2] does not provide a
Index Terms—Fading channels, HF channels, simulation, sto- phase characterization, we assume that the channel phase shift is
chastic processes. uniformly distributed over . This assumption seems rea-
sonable for an fading channel where the channel fading is
more severe than Rayleigh fading. It is shown that the complex
I. INTRODUCTION
random process used for simulating an fading channel,
work of Braun and Dersch [10], Dersch and Rüegg [11] have parts. Multiplying both sides of (3) by gives .
developed a model for simulating Nakagami- fading channels This expression implies that can be represented by
for urban mobile communications. The model is, however, lim-
ited only for the case . (5)
The rest of the paper is organized as follows. In Section where i) is a WSS nonnegative random process and
III, details on the implementation of the proposed simulation follows a standard beta distribution with parameters and
model are given. Numerical examples illustrating the use of the , and ii) , being independent of , is a zero-mean WSS
proposed model follow in Section IV. Finally, conclusions are complex Gaussian process with uncorrelated real and imaginary
drawn in Section V. parts and with an autocorrelation function
(9)
knowledge of the correlation matrix that charac- ated by a known method. Let and
terizes the correlation among the diversity channels, where . Then
and is real symmetric. To simulate
correlated channels, it is required to generate correlated com- (15)
plex processes , each of which has the prop-
where is a factor of such that .
erties of specified in Section II-A. Note that
To generate , the matrix must exist, which re-
quires that is nonnegative definite. That is, for any
(11) with at least one of which is nonzero, the
condition needs to be satisfied.
and that is nonnegative definite ([14], p. 295). Applying the Substituting (13) into this requirement yields
proposed model (5) to the situation under consideration, we
have (16)
TABLE I
R (1t)=R (0) VALUES AGAINST (1t) FOR 0:5 m < 1. NOTE THAT R (0) = m
is most often band-limited to at most . It is rec- it is recommended that may be modified by the
ommended that a convenient choice of is selected, following steps.
where a) Set wherever computed by (20)
is negative.
(18) b) Set for .
c) Scale the resultant such that
vanishes or is negligible for . .
3) Compute from by (26a) and (26b) Although autocorrelation properties of the generated
or with the aid of Table I, which lists the deviate from the desired ones as a result of this modifica-
values against . Then compute by (7). tion, numerical results of Section IV demonstrate that this
To ensure that can be realized, is required deviation is small. The modified is used to com-
to satisfy that and for all pute new values of .
, where 4) Based on the selected or , generate Gaussian-
distributed samples by a known method. A com-
(19) putationally efficient method is preferred. Then generate
from the generated by (9). In (9), although
is the power spectrum of . If the selected a closed-form expression of is available only
cannot yield 's that satisfy for (see Appendix B), for other values of , nu-
, another choice of merical routines are easily implemented for computation
should be made, and Step 2) should be repeated. How- of this function.
ever, selection of that satisfies for all 5) Generate complex-Gaussian distributed samples
is rather difficult. Appendix C shows that can be based on the modified or the new values of
computed by (20), shown at the bottom of the page. For . Again, an efficient technique is preferred.
a given , it is possible that the computed 6) Compute by (5).
is negative for some , although the contribution of this To simulate correlated diversity channels, one is required
negative part to the total is usually not significant. to generate where each of
In addition, computed by (20) is not limited for has a Doppler power spectrum given
. Instead of carrying out the difficult task to by (17). A sequence of equidistant discrete-time samples
find a particular that makes for all , and is first
(20)
otherwise
218 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 48, NO. 2, FEBRUARY 2000
generated, followed by interpolation to obtain the value of sinusoids method. Parameters of the sum of sinusoids method
for an arbitrary time . The procedure for generating used in generating and were computed based
follows from the one for generating described above, on the procedure of [8].1 The value of was computed
except that an additional step is required to compute from and by (5). After 's were gener-
from the generated by (15). In applying (15), different ated, , were computed by
factorization methods can be used. However, Cholesky fac- piecewise-constant interpolation. A total of samples,
torization is preferred. It gives a lower triangular matrix so i.e., , where ,
that some of the steps in multiplication and addition can be were finally obtained.
eliminated in simulation. This reduces the computation cost. Fig. 1(a) and (b) plot the cumulative distribution functions
(cdf’s) of the amplitude and phase, respectively, computed from
IV. NUMERICAL EXAMPLES the generated samples. The cdf of the dis-
tribution with and , and that of the uniform dis-
A. Example 1 tribution for the range , are also plotted for comparison
A realization of was generated based on the procedure purposes. It is apparent that the distributions of the generated
given in Section III. For illustration purposes, we assumed that samples fit closely to the desired ones. Fig. 1(c) plots the auto-
and . A double-Gaussian spectrum was correlation function computed from the generated samples and
used the one for the double-Gaussian spectrum. It is shown that the
generated autocorrelation function is close to the desired one.
The result indicates that despite the need to truncate the original
, the loss of accuracy in the autocorrelation property of
the generated samples is small.
B. Example 2
(21) In this example, we consider simulating a dual-branch
diversity channel with correlated fading by generating
where and we arbitrarily selected using (12). It is assumed that the correlation be-
. This value of was chosen in order to ensure that tween the two branches is and that the
is negligible for , so that the Doppler power fading parameter is . These two values are arbitrarily
spectrum under consideration can be considered band-limited. selected for illustration purpose. Furthermore, we assume that
The double-Gaussian spectrum was selected because it is ap- and is given by (21) with .
propriate for modeling Doppler power spectra of HF channels We generated 10 samples of by generating
[1]. Although a symmetric spectrum is used here for simplicity, and where and
realistic spectra for HF channels are usually asymmetric [1]. . Generation of each of the indepen-
An oversampling ratio of was employed. dent time series and followed the same
We first computed , where procedure as in generating in Example 1. The same
was arbitrarily chosen. For the sake of convenience, , given by (22) was used in generating and
we selected . As a result of this particular choice of ,
0.20% of the original needed to be truncated. The
was generated by (15). Generation of each of
the time series and also followed the
procedure of generating in Example 1. The time
(22) series was obtained by substituting the generated
and into (12). After the time series
where was used so that is negligible for was generated, , were
. Based on the selected and values of computed by piecewise-constant interpolation. A total of
, we computed and . Then samples of , were obtained.
was computed from values by (20). For Fig. 2(a)–(c) plot the cdf’s of amplitude, phase, and auto-
and at frequencies where computed values were nega- correlation functions, respectively, computed from each set of
tive, we set . The resultant was scaled so generated samples . The desired
that . It was found that 3.26% of cdf’s and autocorrelation functions are also plotted. It is ap-
the original were truncated. A total of 10 discrete-time parent that computed autocorrelation functions and cdf’s are
samples were generated by the following procedure. The
Gaussian-distributed time series was generated by an ef-
ficient technique based on Rice's sum of sinusoids [8]. The time 1One of the reviewers pointed out that statistical properties of fy (nT )g and
series was generated from the time series by the fw(nT )g generated by the sum of sinusoids method can be verified analyti-
nonlinear transformation given by (9). The complex Gaussian cally by the closed-form expressions of [8], so that the autocorrelation and phase
distribution of the generated fz (nT )g can be verified before z (nT )'s are actu-
time series was generated by separately generating the ally generated. This advantage enables more rapid verification of the simulation
in-phase and quadrature-phase components based on the sum of accuracy for the selected set of parameters.
YIP AND NG: A SIMULATION MODEL FOR NAKAGAMI- FADING CHANNELS 219
(a) (b)
(c)
Fig. 1. Statistics computed from the generated samples of z (t) in Example 1: (a) cdf of amplitude, (b) cdf of phase, and (c) autocorrelation function. The desired
cdf’s and autocorrelation function are also plotted.
close to the desired ones. We also computed the correlation ma- generating a complex Gaussian process and a square-root-beta
trix from the generated time series and random process. Application of the proposed model to simulate
and found that correlated fading channels has also been demonstrated. In
numerical examples, it has been shown that statistical properties
of the samples generated from the use of the proposed model are
close to the required ones.
(a) (b)
(c)
Fig. 2. Statistics computed from the generated samples of z (t) and z (t) in Example 2: (a) cdf’s of amplitude, (b) cdf’s of phase, and (c) cdf’s of autocorrelation
functions. The desired cdf’s and autocorrelation function are also plotted.
(24)
respectively, where
is the incomplete beta function. Note that in the special case of (26a)
for . It is easy to In special cases of , it is easy to show that
show that is a strictly increasing function for
and is bounded by . If we specify that ,
the inverse of , gives a value within for ,
then exists and is unique. Let be a WSS zero-mean (26b)
YIP AND NG: A SIMULATION MODEL FOR NAKAGAMI- FADING CHANNELS 221
Numerical integration of (26a) is made easy by using the trans- [3] P. Hoeher, “A statistical discrete-time model for the WSSUS multipath
formation and channel,” IEEE Trans. Veh. Technol., vol. 41, pp. 461–468, Nov. 1992.
[4] P. M. Crespo and J. Jiménez, “Computer simulation of radio channels
followed by another transformation and using a harmonic decomposition technique,” IEEE Trans. Veh. Technol.,
. It follows that (26a) can be simplified to vol. 44, pp. 414–419, Aug. 1995.
[5] K. W. Yip and T. S. Ng, “Efficient simulation of digital transmission over
WSSUS channels,” IEEE Trans. Commun., vol. 43, pp. 2907–2913, Dec.
1995.
[6] K. W. Yip and T. S. Ng, “Karhunen-Loève expansion of the WSSUS
channel output and its application to efficient simulation,” IEEE J. Se-
lect. Areas Commun., vol. 15, pp. 640–646, May 1997.
[7] J. F. Mastrangelo, J. J. Lemmon, L. E. Vogler, J. A. Hoffmeyer, L. E.
Pratt, and C. J. Behm, “A new wideband high frequency channel simu-
lation system,” IEEE Trans. Commun., vol. 45, pp. 26–34, Jan. 1997.
The inner and outer integrals can be readily computed by the [8] M. Pätzold, U. Killat, F. Laue, and Y. Li, “On the statistical properties
of deterministic simulation models for mobile fading channels,” IEEE
Gauss–Laguerre rule and the Simpson's rule, respectively. Table Trans. Veh. Technol., vol. 47, pp. 254–269, Feb. 1998.
I lists the computed values against for [9] H. Kong and E. Shwedyk, “A hidden Markov model (HMM)-based
. Note that . MAP receiver for Nakagami fading channels,” in Proc. 1995 IEEE Int.
Symp. Information Theory, Whistler, BC, Canada, Sept. 17–22, 1995,
p. 210.
APPENDIX C [10] W. R. Braun and U. Dersch, “A physical mobile radio channel model,”
PROOF OF (20) IEEE Trans. Veh. Technol., vol. 40, pp. 472–482, May 1991.
[11] U. Dersch and R. J. Rüegg, “Simulation of the time and frequency se-
Consider lective outdoor mobile radio channel,” IEEE Trans. Veh. Technol., vol.
42, pp. 338–344, Aug. 1993.
[12] N. L. Johnson, S. Kotz, and N. Balakrishnan, Continuous Univariate
(27) Distributions, 2nd ed. New York: Wiley, 1995, vol. 2.
[13] M. Grigoriu, Applied Non-Gaussian Processes. Englewood Cliffs, NJ:
Prentice-Hall, 1995.
[14] H. Cramér, Mathematical Method of Statistics. Princeton, NJ:
It follows that , Princeton Univ. Press, 1946.
so that