FOURTH QUARTER 2003, VOLUME 5, NO.

www.comsoc.org/pubs/surveys

ERROR MODELING SCHEMES FOR

FADING CHANNELS IN
WIRELESS COMMUNICATIONS: A SURVEY
HAOWEI BAI, HONEYWELL AEROSPACE
MOHAMMED ATIQUZZAMAN, UNIVERSITY OF OKLAHOMA

ABSTRACT
Network system designers need to understand the error performance of wireless
mobile channels in order to improve the quality of communications by deploying
better modulation and coding schemes, and better network architectures. It is also
desirable to have an accurate and thoroughly reproducible error model, which would
allow network designers to evaluate a protocol or algorithm and its variations in a
controlled and repeatable way. However, the physical properties of radio
propagation, and the diversities of error environments in a wireless medium, lead to
complexity in modeling the error performance of wireless channels. This article
surveys the error modeling methods of fading channels in wireless communications,
and provides a novel user-requirement (researchers and designers) based approach
to classify the existing wireless error models.

M odeling the error performance of wireless channels

is usually a complex task, because the performance
of wireless channels inherently depends on radio
propagating modes, such as, line of sight (LOS) radiation,
reflections from a smooth surface, diffractions around a cor-
• Intersymbol interference (ISI): This is caused by delay
spread (the arrival of a transmitted symbol is delayed),
resulting in partial cancellation of the current symbol.
• Doppler shift: This is due to the relative velocities of the
transmitter and the receiver. Doppler shift causes fre-
ner, and scattering caused by an object with dimensions on quency shifts in the arriving signal, thereby complicating
the order of the wavelength. The radio link is highly variable the successful reception of the signal.
over short distances due to the statistical distribution of path • Multipath fading: This is caused by multipath propaga-
loss (PL) and the physical properties of propagation environ- tion of radio frequency (RF) signals between a transmit-
ments, thereby making it difficult to generalize the results of ter and a receiver. Multipath propagation can lead to
error performance analysis. fluctuations in the amplitude, phase, and angle of the sig-
Schemes to improve the reliability of wireless channels nal received at a receiver. We will describe multipath
range from innovative transport-layer protocols to robust fading in more detail later.
physical-layer schemes, including better modulation and cod- It is challenging for wireless network protocol developers
ing. The development and selection of the schemes are based to consider the large number of factors that affect the error
on the understanding of the statistical nature of errors. There- performance of wireless channels. Typically, if a wireless
fore, a good understanding of the nature of errors in wireless channel’s propagation characteristics are not specified, one
channels is critical in having a reliable wireless communication usually assumes that the signal attenuation versus distance
for upper-layer applications. Some of the main causes of bit behaves as if the propagation takes place over ideal free
errors, and consequently packet losses, in the widely deployed space [1]. The free space model treats the region between
CDMA wireless channel are described below. the transmitter and the receiver as free of all objects that
• Attenuation: This is due to a decrease in the intensity of might absorb or reflect radio frequency energy. It also
electromagnetic energy at the receiver (e.g., due to long assumes that the atmosphere is perfectly uniform and non-
distance), which leads to low signal-to-noise ratio (SNR). absorbing. However, in practice, the free space model is not

2 IEEE Communications Surveys & Tutorials • Fourth Quarter 2003

accurate enough to describe the performance of a real wire-
less mobile channel.
Three main physical phenomenon affect radio propagation
in a real-world scenario: reflection, diffraction, and scattering.
When electromagnetic radiation reflects off objects or diffracts
around objects, it can travel from the transmitter to the
receiver over multiple paths, giving rise to multipath propaga-
tion [2]. This can result in fluctuations in the received signal’s
amplitude, phase, and angle of arrival, giving rise to multipath
fading. System modeling and design, which mitigate the
effects of fading, are usually challenging [1]. Therefore, error
models of fading channels in wireless mobile communications
are very helpful in designing and evaluating the performance Reflection
of wireless networks and communication systems. Diffraction
The error performance of wireless channels is usually mod- Scattering
eled by capturing the statistical nature of the interaction
among reflected radio waves. The statistical calculations for
Bit Error Rate (BER), which is generally used to characterize ■ FIGURE 1. Three radio propagation mechanisms.
channel errors at the physical layer, is a well known practice.
From the perspective of higher layers, network protocol devel-
opers and algorithm designers are interested in block errors system error performance (typically measured by SNR) not
(packet errors), since most of the higher-layer applications only depends on the desired signal power, but also depends on
(running on top of link layers) exchange blocks of data the noise power. It is the objective of this article to present
between peers. For example, bit errors in a link-layer packet existing error modeling methods where both parameters are
may result in the loss of the entire packet; a single packet loss considered. Readers who are interested in wireless channel
within a message may lead to the loss of the entire message. propagation models are encouraged to read the article by
Therefore, it is desirable to have accurate packet-level error Neskovic, et al. [5], which presents a good overview of popular
models for wireless channels, which can be used by network prediction models based on two main categories, outdoor
protocol developers and network system engineers to simulate (macrocell and microcell) and indoor propagation models, and
and analyze the end-to-end performance at the packet level. It describes some useful algorithms to improve their accuracy.
has been observed empirically [3] that errors in wireless fad- The rest of this article is organized as follows. The nature
ing channels can be approximated by a two-state Markov pro- of radio propagation, and characteristics and models of fading
cess. In other words, a well designed channel may enter a channels, are described. Several fading-channel error model-
state where bursty errors occur for a small time interval. ing schemes are classified and introduced. Finally, concluding
Analytical models which can be used for the statistical cal- remarks are presented.
culation of BER at the physical layer, and error approxima-
tion at higher layers, are described in this survey. In addition
to analytical models, it is highly desirable to model wireless MULTIPATH FADING
networks in a thoroughly repeatable fashion. This is especially
important for people who must experiment with realistic Electromagnetic waves reflecting off objects or diffracting
channel parameters. However, the results from analytical around objects can result in the signal travelling over multiple
models are either inaccurate due to errors in the approxima- paths from the transmitter to the receiver. This phenemonon,
tion and simplification of environmental characteristics, or called multipath propagation, can cause fluctuations in the
unlikely to be reproduced due to complex and inefficient algo- received signal’s amplitude, phase, and angle of arrival, giving
rithms. This gives rise to empirical distribution-based models. rise to multipath fading. In this section, we briefly describe the
In this article, our objective is to survey the existing error various types and models of fading channels.
modeling methods for fading channels in wireless communica- There are three main mechanisms that impact radio propa-
tion systems, and classify them according to their modeling gation in wireless channels [6], as illustrated in Fig. 1.
approach, viz., analytical and empirical. In contrast to traditional Reflection, which may interfere constructively or destruc-
classification methods, our approach is based on differing tively at the receiver, occurs when an electromagnetic wave
requirements of wireless communication (and network) impinges on a smooth surface with very large dimensions
researchers and designers. The goal of this article is to help when compared to the wavelength of the radio wave.
researchers who need to analyze characteristics of wireless error Diffraction occurs when the path of the electromagnetic
channels, and network protocol designers who need to synthe- wave is obstructed by an impenetrable body of large dimen-
size wireless error channels for their simulations and evalua- sions as compared to the RF signal wavelength. This causes
tions, find appropriate error models for wireless fading channels. secondary waves to be formed behind the obstructing body,
We must emphasize that it is not the aim of this article to without any LOS path between the secondary waves. Diffrac-
present various wireless channel propagation models, but to tion, which is also called shadowing because the diffracted
describe some existing efforts on error (specifically, BER at field can reach the receiver even when shadowed by an
the physical layer, and packet loss rate at upper layers) model- impenetrable obstruction, explains how RF energy can travel
ing methods for wireless fading channels. The propagation in urban and rural environments without a LOS path.
models for wireless communication channels are used to pre- Scattering occurs when the radio channel contains objects
dict the path propagation loss (e.g., the Lee model [4]). Design of dimensions that are on the order (or less) of the electro-
engineers use an accurate estimation of path loss to help select magnetic wavelength, causing energy from a transmitter to be
the locations of base stations, and determine a proper frequen- radiated in many different directions. In urban environments,
cy plan. The field signal strength (or signal power) can be typical objects that cause scattering are lamp posts, street
derived from the estimation result of path loss. However, the signs and foliage.

IEEE Communications Surveys & Tutorials • Fourth Quarter 2003 3

 Environment Channel type

Mobile systems with no LOS path between transmitter and receiver antenna, Rayleigh [10].
propagation of reflected and refracted paths through troposphere and
ionosphere, ship-to-ship radio links [11].

Satellite links subject to strong ionospheric scintillation [12]. Nakagami-q (Hoyt) (spans range from one-sided
Gaussian (q = 0) to Rayleigh (q = 1)) [13].

Propagation paths consisting of one strong direct LOS component and many Nakagami-n (Rice) (spans range from Rayleigh (n = 0)
random weaker components — microcellular urban and suburban land to no fading (n = • )) [14].
mobile, picocellular indoor and factory environments [15].

Land mobile [16], indoor mobile multipath propagation as well as ionospheric Nakagami-m (spans range from one-sided Gaussian
radio links. m = 1/2), Rayleigh (m = 1) to no fading (m = • )) [17].

Terrain, buildings, trees — urban land mobile systems, land mobile satellite Log-normal shadowing [18].
systems [19].

Nakagami-m multipath fading superimposed on log-normal shadowing. Composite gamma/log-normal [18].

Congested downtown areas with slow-moving pedestrians and vehicles.
Also in land mobile systems subject to vegetative and/or urban shadowing [20].

Convex combination of unshadowed multipath and a composite multipath/ Combined (time-shared) shadowed/unshadowed [21].
shadowed fading. Land mobile satellite systems [21].

■ Table 1. Models that can be used to characterize various wireless environments.

The above three radio propagation mechanisms impact the Multipath fading arises from the constructive and destruc-
strength of the received signal in different ways. If there is a tive combination of randomly delayed reflected, scattered,
strong LOS between the transmitter and the receiver, diffrac- diffracted signal components. Based on the nature of the
tion and scattering are not the dominant factors in the propa- radio propagation environment, different mathematical mod-
gation of the radio waves. However, in the absence of a LOS els exist to describe the statistical behavior of the multipath
between the transmitter and the receiver, diffraction and scat- fading envelope.
tering become the dominant factors in the propagation. Typi- • The Rayleigh distribution [10] is used to model the prop-
cally, the received signal is a sum of the components arising agation environment where the mobile antenna receives
from the above three phenomena. The strength of the received a large number of reflected and scattered waves. Because
signal fluctuates rapidly with respect to time and the displace- of wave cancellation effects, the instantaneous received
ment of the transmitter and the receiver. power seen by a moving antenna becomes a random vari-
able, dependent on the location of the antenna.
LARGE-SCALE FADING AND SMALL-SCALE-FADING • The Nakagami-q distribution [13] is typically observed on
satellite links subjected to strong ionospheric scintilla-
Based on the distance over which a mobile moves, there are tion.
two different types of fading effects: large-scale fading and • The Nakagami-n distribution, known as the Rice distribu-
small-scale fading [1]. If the mobile moves away from the tion [14], is often used to model similar environments to
transmitter over a large distance, the received signal will Rayleigh fading channels, except that the set of reflected
experience large-scale signal variation. Large-scale fading and scattered waves are dominated by one strong compo-
represents the average signal power attenuation due to nent.
motion over large areas. The receiver is often represented • The Nakagami-m distribution [17] can be used to model
as being shadowed by prominent terrains, such as hills, fading-channel conditions that are more severe than the
forests, billboards, clumps of buildings, etc. Small-scale fad- Rayleigh distribution. It often gives the best fit to land-
ing refers to the dramatic changes in signal amplitude and mobile, indoor-mobile multipath propagation, as well as
phase that can be experienced as a result of small changes scintillating ionospheric radio links.
(as small as a half-wavelength) in the distance between the In terrestrial and satellite land-mobile systems, the link
transmitter and the receiver. When there are a large num- quality is also affected by slow variation of the mean signal
ber of reflective paths with no LOS signal components, the level, resulting from the effects of shadowing from terrain,
envelope of the received signal can be statistically described buildings, and trees. The shadowing can generally be mod-
by the Rayleigh distribution [7, 8]. If dominant non-fading eled by a log-normal distribution for various outdoor and
components exist, such as a LOS propagation path, the indoor environments [22]. If the receiver is able to average
small-scale fading envelope is Rice distributed [7, 8]. A out the fast multipath fading, the performance of mobile
mobile radio propagating over a large area will experience systems depends only on shadowing. However, in an envi-
both types of fading, i.e., small-scale fading superimposed ronment consisting of multipath fading superimposed on
on large-scale fading. shadowing, the receiver does not average out the fading
envelope. This scenario, called composite multipath/shad-
DIFFERENT TYPES OF FADING-CHANNEL MODELS owing, is typically observed in congested downtown areas
with slow moving pedestrians and vehicles [20]. A detailed
In this section, we present the different types of fading chan- discussion of this topic, and the corresponding probability
nels that are typical in communication environments, and the density functions of fading amplitude and SNR, can be
mathematical models that can be used to describe the chan- found in [23].
nels. Table 1 [9] shows various fading-channel models, classi-
fied by the environments to which they apply.

4 IEEE Communications Surveys & Tutorials • Fourth Quarter 2003

 ERROR MODELS OF FADING CHANNELS Proakis [25] derived a generic equation for evaluating the
BER of multichannel noncoherent and differentially coherent
With the proliferation of wireless networks, much research has reception of binary signals over multiple independent additive
focussed on improving the quality of fading channels. This has white Gaussian noise (AWGN) channels. Lindsey [26] devel-
led to a growing interest in characterizing the packet loss behav- oped a generic expression for the average BER of binary cor-
ior of fading channels. A number of error models have been pro- related FSK for multichannel reception over multiple
posed in the literature to characterize the loss behavior of fading independent Rician fading channels, where the strength of the
channels. The error models can be classified into two groups: scattered component is assumed to be constant for all the
analytical models and empirical distribution-based models. channels. Charash [27] performed the average BER perfor-
mance analysis for binary orthogonal FSK with multichannel
ANALYTICAL MODELS reception over multiple independent and identically distribut-
ed (i.i.d.) Nakagami-m fading channels. Weng et al. in [28]
Errors occurring in wireless channels are due to the diversity derived a closed form average BER expression for binary
of wireless connections and the complicated physical impair- DPSK with multichannel reception over multiple Nakagami-m
ments. As a result, it is difficult to generalize the mathemati- fading channels. In [29] Patenaude et al. extended the results
cal results from one specific domain to another. Therefore, of [27] and [28]. They derived a closed form expression for
analytical models are highly dependent on the characteriza- the average BER of binary orthogonal square-law detected
tion of error environments, such as fading channels. The tra- FSK and binary DPSK with multichannel reception over mul-
ditional metric used for characterizing channel errors at the tiple independent, but not necessarily i.i.d., Nakagami-m fad-
physical layer is the average BER. This gives rise to the first ing channels. Tjhung et al. [30] and Tanda [31] analyzed the
type of analytical modeling methods, called physical-layer ori- average BER for Differential Quadrature PSK (DQPSK) over
ented modeling, which deals with bit errors. Rician and Nakagami-m fading channels, respectively. Tellam-
It is expected that future wireless communications will bura et al. [32] presented an alternate unified BER analysis of
include image, video, and data applications. Accordingly, DQPSK over Rician and Nakagami-m fading channels.
higher-layer encoding algorithms for multimedia transmissions In contrast to the above approaches, Simon and Alouini [9]
will have to be carefully designed to overcome the impact of deploy an alternate form of the Marcum Q-function and the
errors on wireless channels. Higher-layer applications, running resulting alternate integral representation of the conditional
on top of a link layer, usually manage data transfer in blocks BER, as well as some well known Laplace transforms, to inde-
of multiple bits or symbols, and employ various block error pendently average over the PDF of each fading channel. This
detection and retransmission mechanisms. In this case, it is results in a useful generalized expression for the average BER
important to examine the effects of multipath fading-channel performance of noncoherent and differentially coherent com-
dynamics on block data throughput, delay, and queuing per- munication systems over AWGN and fading channels, thus
formance. This gives rise to the second type of analytical mod- unifying all the results mentioned above.
eling methods, called higher-layer oriented modeling, which is Figure 2 [9] shows a multilink channel model, which is
based on block errors. In the following subsections, we used to develop the generalized model for the determination
describe physical-layer oriented modeling and higher-layer of the BER performance. This model is general enough to
oriented modeling. include the cases where different independent channels are
not necessarily identically distributed, nor even distributed
Physical-Layer Oriented Modeling — This method aims at according to the same family of distributions, and systems that
determining the BER performance of a wireless communica- employ postdetection equal gain combining [18] (EGC). The
tion system over fading channels. Based on the specific chan- transmitted signal is received over L independent channels,
L
nel model and modulation/detection combination, the average each of them being a fading channel. {r l(t)} l=1 is a set of L
BER is obtained through statistical calculations. One of the received replicas of the signal, where l is the index of the
approaches is to approximate the probability density function channel, and a , q , and t are the random fading-channel
(PDF) of SNR at the receiver, and then average
the BER over that PDF. Much work on this mod-
eling approach has been reported in the litera- Transmitted AWGN
ture. signal r1(t)
In many applications, it is difficult to track the S(t) Delay t1
phase of the received signal. It is therefore impos-
sible to perform coherent detection, which a1 exp(-jq1) n1 (t)
requires reference in phase with the received car- r2(t)
Delay t2
rier signal. In such cases, the design of a commu- EGC
nication receiver must depend on either noncoherent Decision
noncoherent detection, where no attempt is a2 exp(-jq2) n2 (t) or
made to determine the phase of the incoming r3(t) diff-coherent
Delay t3 receiver
signal, or differentially coherent detection, where
the carrier phase of the previous signaling inter- n3 (t)
val can be used as a phase reference. In a ran- a3 exp(-jq3)
domly fading channel, it may be difficult to
establish and maintain a coherent reference.
rL(t)
Therefore, a noncoherent system or a differen- Delay tL
tially coherent system is more desirable. Readers
are encouraged to read [24] for more details. In nL(t)
the rest of this section, we briefly describe some aL exp(-jqL)
reported BER calculations for noncoherent and
differentially coherent detections. ■ FIGURE 2. Generalized fading channel.

IEEE Communications Surveys & Tutorials • Fourth Quarter 2003 5

 Detection type Modulation (signal set) Parameters of Marcum Q-Function

Noncoherent Equal energy, equiprobable correlated binary signals

(l = complex correlation coefficient) 1 - 1- | l |2 1 + 1- | l |2
h = 1, a = , b=
2 2
Equal energy, equiprobable uncorrelated binary signals, h = 1, a = 0, b = 1
e.g., BFSK
—
Differentially Binary phase-shift-keying (DPSK) h = 1, a = 0, b = ÷ 2
coherent
Quadrature phase-shift-keying (DQPSK) with Gray coding
h = 1, a = 2 - 2 , b = 2 + 2

■ Table 2. Special cases of generalized equation for specific modulation/detection schemes.

( )
amplitudes, phase, and delays, respectively. The first channel
(with index of 1) is assumed to be the reference channel Pb ( E;g ) = Q1 a g , b g
whose delay is 0. The fading amplitude, a l, of the lth channel
—
is a random variable with a mean squared value of a 2l which is Ê h ˆ È ( a2 + b2 )g ù (3)
-Á ˜ exp Í- ú I0 ( abg ).
denoted by W l . The probability distribution of the random Ë 1+h¯ ÍÎ 2 úû
variable a l is any of the family of distributions presented in
—
Table 1. Based on this channel model, the authors in [9] use For h =1, a =0, and b = ÷ 2 (Table 2), Eq. 3 can be reduced
alternate representations of Gaussian and Marcum Q-func- to the well known expressions for DPSK as reported by a
tions that are characteristic of error-probability expressions number of authors:
for differentially coherent and noncoherent forms of detection 1
to obtain the generalized BER expression as follows: Pb ( E; g ) DPSK = exp(- g ). (4)
2

(
È
Pb ( E;g ) = Q1 a g , b g - ÍÍ1 - ) l(
L - 1 2 Lr - 1 l
 l =r 0 h ) ù
ú Contributions made by hundreds of authors dealing with
(1 + h) 2 Lr - 1 ú BER probability performance over generalized AWGN and
ÍÎ úû fading channels, using alternate representations of Marcum
Q-functions, are now unified in a common framework [9]. The
È ( a2 + b2 )g ù 1 (1) coverage of this framework is broad enough to represent and
¥ exp Í- ú I0 ( abg ) +
ÍÎ 2 úû (1 + h)2 Lr - 1 describe almost all combinations of modulation/detection
types and fading-channel types.

( )h [ ( )]
È Lr
) (
2 Lr - 1 Lr - 1
Most wireless communication system designers make use
¥ ÍÂ Lr - 1 ¥ Ql a g , b g - Q1 a g , b g of commercially available tools. To determine the parameters
ÍÎl = 2 for a specific design module, they carry out many on-site mea-

[ ( )]
surements. However, this does not mean that the analytical
 ( Lr r- l )h
ù
) (
Lr
2 L -1 Lr – 1 + l
- ¥ Ql b g , a g - Q1 b g , a g ú, error models we presented here are not useful. The merger of
l =2 úû networking technology and wireless communications requires
a modern wireless network to be more mobile, capable of
where the function Q(·) is the generalized Marcum Q-function higher data rates, easy to configure and use, and more afford-
obtained by the authors in [9] and is given by able. Apparently, achieving such a near-optimal set of require-
ments requires careful balancing of trades. Simulation
È Ê 2 ù
1 • l x + a2 ˆ ú assessment of performance is the currently widely used
Ql (a,b) = Ú
x exp Í-Á ˜ I (ax) dx. (2) method to achieve this goal. Results presented in this article
l -1 b Í ÁË ˜ ú l -1
a Î
2 ¯û (e.g., BER expression of orthogonal DPSK over Nakagami-m
fading channel) could be easily plugged into the receiver’s
The I(·) function in Eq. 2 is the modified Bessel function BER calculation procedure of commercially available network
of the first kind and order (l – 1). The parameter l in Eq. 1 is simulation tools (e.g., OPNET) to facilitate the physical-layer
the channel index, and g = S l=1Lr
g l is the total instantaneous performance analysis.
SNR per bit. Although simulation plays an essential role in comparing
Typical values of h , a, and b, corresponding to specific competing design alternatives, simulation of wireless network
modulation/detection schemes, are shown in Table 2 [9]. Note systems, especially wireless links, is sometimes unreliable, which
that in all possible cases, a and b are independent of the fad- may lead to incorrect design choices. This is due to errors in
ing-channel model. For the case of single-channel reception, the modeling, including unwanted approximations and simplifi-
the value of L r in Eq. 1 is equal to one. Otherwise, L r ≥ 1, cations. Generally, higher accuracy in modeling requires more
which corresponds to the case of multichannel detection. complexity of computation. In this case, it is desirable to have
Details on the derivation of this generalized expression can be models with low computation complexity, while maintaining the
found in [9]. desired accuracy. The unified expression of BER performance
The effectiveness of the generalized BER equation (as for noncoherent and differentially coherent modulations over
given by Eq. 1) is shown by an example below. The well- generalized fading channels, which is presented above, is a
known conditional BER expression for orthogonal DPSK with good example of such models. The expression of average BER
single-channel reception is obtained by plugging in appropri- only involves a single finite-range integral whose integrand con-
ate parameters summarized in Table 2. tains only elementary functions. It can, therefore, be easily
For Lr = 1 (i.e., single-channel reception), the latter two computed numerically. Such contributions are very useful and
summations in Eq. 1 do not contribute. One can immediately helpful to refining the existing simulation models and tools,
obtain resulting in more accurate and fair design choices.

6 IEEE Communications Surveys & Tutorials • Fourth Quarter 2003

 p10

Higher-Layer Oriented Modeling — As mentioned earlier,

modeling of fading errors at higher layers aims at calculating p00 0 1 p11
the average block error rate (packet error rate). In the litera-
ture, most of the models are based on the assumption that
data packet transmissions are i.i.d. In addition, many coding
schemes and protocols were initially designed for i.i.d. chan- p01
nels. One may consider an apparent alternative solution to
this problem: the study of channels with memory by deploying ■ FIGURE 3. State transition diagram for an example simplified
some prediction techniques. Unfortunately, little work is avail- Gilbert model.
able on such models.
It has been shown that the special structure of Markov
approximation makes it naturally useful and tractable for this the range of the received SNR into a finite number of inter-
purpose [33]. The Markov chain assumes that an adequate vals, the FSMC model is constructed for Rayleigh fading
description of a system is given by a finite number of states. channels [37].
Each state is assigned a probability of the system being in As wireless network technology evolves, the design of
that state. For example, the typical movement of the stock communication networks becomes more complicated because
market could be considered as a simple two-sate model in most of the existing network protocols were designed for
terms of up and down movement of the index. The study of wireline networks. Analyzing and evaluating the perfor-
Markov approximation for fading channels dates back to the mance of existing protocols in wireless environments pro-
early work of Gilbert [34] and Elliott [35]. They built a two- vides many insights and possible solutions on how to adapt
state Markov channel known as the Gilbert-Elliott channel. existing network protocols to wireless environments. There-
In a simplified Gilbert model [36], the error probabilities in fore, accurate models of network dynamics such as packet
bad and good states are 1 and 0, respectively. Assuming 1 error probabilities is very essential to wireless network pro-
and 0 denote successful and erroneous transmission in a tocol design and development. The higher-layer oriented
given slot, the state transition diagram is shown in Fig. 3. modeling methods directly provide packet error probabilities
Accordingly, without the need to understand the complex physical-layer
schemes and long conditional probability calculations. One
Êp p01 ˆ may easily put such a higher-layer oriented model to a
P = Á 00 ˜ (5)
Ë p10 p11 ¯ prevalent network simulation tool (e.g., OPNET, or NS),
and evaluate (and furthermore, improve) the existing higher-
is the transition matrix for the packet error process. There- layer network protocols. For example, by simulating link-
fore, the probability of having packet errors is given by [33] layer error control protocols, such as automatic repeat
request (ARQ), one may be able to tune the existing param-
p01
e= (6) eters of the protocol and even find performance improve-
p10 + p01 ment schemes. By simulating TCP over wireless links with
random link errors generated by such a higher-layer error
This Markovian model for average packet error probability model, many TCP improvement algorithms have been pro-
can be easily extended for diversity. The performance can be posed during the past several years. These algorithms pro-
improved by using two (or more) suitably spaced antennas vide different ways for TCP to react differently in front of
over fading channels [33]. For example, if a diversity of order congestion errors and link corruption errors.
two (i.e., two antennas) is employed, and the signals received
at the two antennas fade independently, the channel can be EMPIRICAL DISTRIBUTION-BASED MODELS
modeled by three states: both channels are good (state 0), two
channels have different states (state 1), and both channels are A large number of measurements have been conducted by
bad (state 2). If the transition matrix of each channel has the telecommunication companies, research laboratories, and
same form as Eq. 5, the transition matrix of this three-state universities in order to determine reasonable fading-chan-
Markov chain can be written as [33] nel parameters (such as error rates) for wireless communi-
cation systems. These measurements show that
Ê p2 2 p01 p00 2 ˆ
p01 environmental factors, such as terrain, construction materi-
00
Á ˜ als, speed of pedestrians and vehicles, etc., have a direct
P = Á p10 p00 p11 p00 + p10 p01 p01 p11 ˜ . (7)
Á p2 impact on radio propagation characteristics. Therefore,
2 ˜
Ë 10 2 p10 p11 p11 ¯ although analytical modeling methods as described here are
sound in theory, in practice it is hard to determine the val-
In this three-state Markov model, states 0 and 1 correspond to ues of the parameters in these models, especially when we
successful transmission, while state 2 corresponds to a trans- want to build a specific model under a special environment.
mission failure. This gives rise to empirical distribution-based modeling
A binary symmetric channel (BSC) with a given crossover methods.
probability can be associated with each state so that the The development of empirical distribution-based models
channel quality for each state can be identified. The Gilbert- consists of three phases: data collection, statistical analysis
Elliott channel is a special case of this type of channel, and model construction, and model validation.
where the crossover probabilities of the BSC are 0 and 0.5, • In the data collection phase, a large number of statistical
respectively [37]. Each state represents a specific channel data are collected and recorded by network tracing or
quality that is either totally noisy or noiseless. When the measurements for many different scenarios.
channel quality varies dramatically, a two-state Gilbert- • Those statistical data are extracted in terms of interests,
Elliott model is not adequate. As a solution, a finite-state such as packet errors, and are modeled in the statistical
Markov channel (FSMC) model is proposed in [37] as an analysis and model construction phase. The model is
extension of the two-state Markov model. By partitioning built by fitting known probability distributions to data.

IEEE Communications Surveys & Tutorials • Fourth Quarter 2003 7

• In the model validation phase, the models are simulated SUMMARY
or analyzed, and validated. The models are refined to
make them consistent with the collected data and traces. Wireless error modeling methods have been surveyed and
An impressive example is presented in [3], where the loss classified in this article. The models that have been developed
behavior of the AT&T WaveLAN, an in-building wireless to approximate the loss behavior of transmission channels
interface, is characterized and modeled using empirical distri- have been classified into two groups: analytical models and
bution-based modeling methods. Though IEEE 802.11b has empirical distribution-based models. Since errors in wireless
been prevalent among wireless LAN technologies, we still channels are due to the diversity of wireless connections and
choose this example to show the usefulness and effectiveness the complicated error environment, analytical modeling meth-
of empirical distribution-based models, because as a result of ods are highly dependent on the characterization of error
conducting network tracing, the traditional method of evaluat- environments, such as fading channels. Although physical-
ing wireless network protocols with a uniform error model has layer oriented modeling methods and higher-layer oriented
been shown to be inaccurate by [3]. Furthermore, the authors modeling methods can be used to evaluate the BER perfor-
in [3] also reveal that the wireless error behavior of Wave- mance and packet error performance, respectively, it is highly
LAN cannot be accurately modeled by a simple two-state desirable that wireless errors be modeled in a thoroughly
Markov chain using analytical modeling methods. Instead, reproducible way. This is especially important for developers
another improved two-state Markov model, based on the dis- of network protocol and mobility algorithms who must experi-
tribution of the error length (defined to be the number of ment with realistic channel parameters. The empirical distri-
packets that are lost consecutively) and error-free length bution-based modeling approach, which alleviates the above
(defined to be the number of packets that are successfully problem, has been described in this article.
received between two adjacent bursts of errors [3]) of the
packet stream, has been shown to be more accurate. The key ACKNOWLEDGMENT
difference between the two lies in the probability distribution
of the error and error-free length. The simple two-state The authors would like to thank Dr. Martin Reisslein, Dr.
Markov model assumes that the error length and error-free John Daigle, and anonymous reviewers for their helpful sug-
length are geometrical distributed. However, results of the gestions.
empirical distribution-based modeling shows that the error
length is better described by a combination of two exponential
distributed segments, and the error-free length is better REFERENCES
described by a combination of three segments, two Pareto dis- [1] B. Sklar, “Rayleigh Fading Channels in Mobile Digital Commu-
tributions and one exponential distribution. nication Systems, Part I: Characterization,” IEEE Commun.
In addition to validating and refining the existing models, Mag., vol. 35, no. 9, Sept. 1997, pp. 136–46.
the empirical modeling method can also be used to find [2] D. Eckhardt and P. Steenkiste, “Measurement and Analysis of
inappropriate hypotheses that have been used in analytical the Error Characteristics of an In-Building Wireless Network,”
modeling in the past. This indicates limitations of some Proc. ACM SIGCOMM ’96, Stanford, CA, Aug. 1996, pp.
243–54.
widely used analytical models. The authors in [38] point out [3] G. T. Nguyen et al., “A Trace-Based Approach for Modeling
that the tracing data on wireless links is non-stationary in Wireless Channel Behavior,” Proc. Winter Simulation Conf.,
nature by performing error traces on a GSM network plat- Coronado, CA, Dec. 1996, pp. 597–604.
form. (The authors define a trace to be stationary when the [4] W. C. Y. Lee, Mobile Communications Design Fundamentals,
error statistics remain relatively constant over time.) They New York: John Wiley, 1993.
arrived at the above conclusion by applying the Run Test [5] A. Neskovic, N. Neskovic, and G. Paunovic, “Modern Approach-
[39] (a method used to test stationarity) to the tracing data. es in Modeling of Mobile Radio Systems Propagation Environ-
Many users have used the assumption of stationarity of ment,” IEEE Commun. Surveys, vol. 3rd Quarter, Sept. 2000,
Markov chains to model errors in wireless channels, includ- pp. 2–12.
[6] T. S. Rappaport, Wireless Communications, Upper Saddle River,
ing the Gilbert model introduced earlier. The authors show NJ: Prentice Hall, 1996.
the inaccuracy of the Gilbert model as a result of inappro- [7] F. Babich and G. Lombardi, “Statistical Analysis and Characteri-
priate stationarity assumptions. Furthermore, they propose a zation of the Indoor Propagation Channel,” IEEE Trans. Com-
new method, called Markov-based trace analysis (MTA), mun., vol. 48, no. 3, Mar. 2000, pp. 455–64.
which results in a more accurate error burst distribution [8] W. C. Jakes, Microwave Mobile Communications, New York:
model for wireless networks. MTA takes advantage of a John Wiley, 1974.
Markov process by decomposing a non-stationary trace into [9] M. K. Simon and M. S. Alouini, “A Unified Approach to the
a set of piece-wise stationary traces consisting of lossy and Performance Analysis of Digital Communication over General-
ized Fading Channels,” Proc. IEEE, vol. 86, no. 9, Sept. 1998,
error-free states. The error-free trace is a deterministic pro-
pp. 1860–77.
cess, whereas the lossy trace is a stationary random process [10] L. Rayleigh, “On the Resultant of a Large Number of Vibra-
that is modeled by a Markov process. tions of the Same Pitch and of Arbitrary Phase,” Philosophical
Many advantages and uses of empirical distribution-based Mag., vol. 27, no. 6, June 1889, pp. 460–69.
models have been discussed above. However, in order to build [11] H. B. James and P. I. Wells, “Some Tropospheric Scatter Prop-
an empirical model, many resources are required. These agation Measurements Near the Radio-Horizon,” Proc. IRE, vol.
include real communication networks, measurement equip- 43, Oct. 1955, pp. 1336–40.
ment, data tracing and recording tools (hardware and soft- [12] B. Chytil, “The Distribution of Amplitude Scintillation and the
ware), and data processing algorithms, which definitely lead to Conversion of Scintillation Indices,” J. Atmospheric and Terres-
high costs. The authors applying the MTA method in [38] trial Physics, vol. 29, no. 9, Sept. 1967, pp. 1175–77.
[13] R. S. Hoyt, “Probability Functions for the Modulus and Angle
required access to a GSM network platform. A study of wire- of the Normal Complex Variate,” Bell System Tech. J., vol. 26,
less communication for aircraft systems, performed by no. 4, Apr. 1967, pp. 318–59.
Aerospace Vehicle Systems Institute (AVSI), even required a [14] S. O. Rice, “Statistical Properties of a Sine Wave Plus Random
Boeing-777 airplane as part of their testbed. Their purpose of Noise,” Bell System Tech. J., vol. 27, no. 1, Jan. 1948, pp.
tracing was to build a model. 109–57.

8 IEEE Communications Surveys & Tutorials • Fourth Quarter 2003

[15] R. J. C. Bultitude, S. A. Mahmoud, and W. A. Sullivan, “A [37] H. S. Wang and N. Moayeri, “Finite-State Markov Channel: A
Comparison of Indoor Radio Propagation Characteristics at Useful Model for Radio Communication Channels,” IEEE Trans.
910 MHz and 1.75 GHz,” IEEE JSAC, vol. 7, no. 1, Jan. 1989, Vehic. Tech., vol. 44, no. 1, Feb. 1995, pp. 163–71.
pp. 20–30. [38] A. Konrad et al., “A Markov-Based Channel Model Algorithm
[16] T. Aulin, “Characteristics of a Digital Mobile Radio Channel,” for Wireless Networks,” Proc. 4th ACM Int’l. Wksp. Modeling,
IEEE Trans. Vehic. Tech., vol. VT-30, no. 5, May 1981, pp. 45–53. Analysis and Simulation of Wireless and Mobile System, Rome,
[17] M. Nakagami, Statistical Methods in Radio Wave Propaga- Italy, July 2001, pp. 28–36.
tion, Oxford, England: Pergamon, 1960, pp. 3–36. [39] J. Bendat and A. Piersol, Random Data: Analysis and Mea-
[18] G. L. Stuber, Principles of Mobile Communications, ch. sec. surement Procedures, New York: John Wiley, 1986.
2.4, Norwell, MA: Kluwer, 1996.
[19] T. S. Rappaport, S. Y. Seidel, and K. Takamizawa, “Statistical
Channel Impulse Response Models for Factory and Open Plan
BIOGRAPHIES
Building Radio Communication System Design,” IEEE Trans. H AOWEI B AI ([email protected]) received his M.Sc.
Commun., vol. 39, no. 5, May 1991, pp. 794–807. degree in electrical and computer engineering from the University
[20] M. J. Ho and G. L. Stuber, “Co-Channel Interference of Micro- of Dayton, Dayton, Ohio, and his B.E. degree from the Depart-
cellular Systems on Shadowed Nakagami Fading Channels,” ment of Information and Communication Engineering, Xi’an Jiao-
Proc. IEEE Vehic. Tech. Conf., Secaucus, NJ, May 1993, pp. 568–71. tong University, P.R. China. He is currently a research scientist at
[21] E. Lutz et al., “The Land Mobile Satellite Communication AES Technology Centers of Excellence, Honeywell Aerospace. He
Channel–Recording, Statistics, and Channel Model,” IEEE Trans. has previously been a research scientist at the Communication
Vehic. Tech., vol. VT-40, no. 5, May 1992, pp. 375–86. and Systems Architecture Lab of Honeywell Laboratories. Much of
[22] H. Suzuki, “A Statistical Model for Urban Multipath Propaga- his current research has focused on communication network
tion,” IEEE Trans. Commun., vol. COM-25, no. 7, July 1977, architectures and protocols, high-speed avionic data buses,
pp. 673–80. broadband wireless communication systems, QoS of next-genera-
[23] M. K. Simon and M. S. Alouini, “A Unified Approach to the tion Internet, TCP over wireless links, and SCTP over satellite links.
Probability of Error for Noncoherent and Differentially Coher- He has several publications on these areas. He is a member of
ent Modulations over Generalized Fading Channels,” IEEE IEEE and SAE. He serves as a committee member on SAE AS-1A
Trans. Commun., vol. 46, no. 12, Dec. 1998, pp. 1625–38. Avionic Networks Subcommittee, and referees for various techni-
[24] B. Sklar, Digital Communications Fundamentals and Applica- cal journals and IEEE conferences.
tions, Upper Saddle River, NJ: Prentice Hall, 1988.
[25] J. G. Proakis, “On the Probability of Error for Multichannel M OHAMMED A TIQUZZAMAN ([email protected]) received the M.Sc. and
Reception of Binary Signals,” IEEE Trans. Commun. Tech., vol. Ph.D. degrees in electrical engineering from the University of
COM-16, no. 2, Feb. 1968, pp. 68–71. Manchester, England. Currently he is faculty member in the
[26] W. C. Lindsey, “Error Probability for Ricean Fading Multichan- School of Computer Science at the University of Oklahoma. He
nel Reception of Binary and n-ary signals,” IEEE Trans. Info. has previously held faculty positions at the University of Day-
Theory, vol. IT-10, no. 10, Oct. 1964, pp. 339–50. ton, Dayton, Ohio, and Monash University, Australia. He is the
[27] U. Charash, “Reception through Nakagami Fading Multipath co-editor-in-chief of the journal Computer Communications,
Channels with Random Delay,” IEEE Trans. Commun., vol. and serves on the editorial board of IEEE Communications Mag-
COM-27, no. 4, Apr. 1979, pp. 657–70. azine, as well as the journals Telecommunications Systems,
[28] J. F. Weng and S. H. Leung, “Analysis of DPSK with Equal Wireless and Optical Networks, and Real Time Imaging. He has
Gain Combining in Nakagami Fading Channels,” IEEE Electron- guest-edited many special issues in various journals, including
ics Letters, vol. 33, no. 4, Apr. 1997, pp. 654–56. feature topics on switching and traffic management for multi-
[29] F. Patenaude, J. H. Lodge, and J. Y. Chouinard, “Error Proba- media in IEEE Communications Magazine; architecture, protocol
bility Expressions for Noncoherent Diversity in Nakagami Fad- and quality of service in the European Transactions on Telecom-
ing Channels,” Proc. IEEE Vehic. Tech. Conf., Phoenix, AZ, May munications; ATM switching and ATM networks in the Interna-
1997, pp. 1484–87. tional Journal of Computer Systems Science & Engineering; and
[30] T. T. Tjhung, C. Loo, and N. P. Secord, “BER Performance of parallel computing on clusters of workstations in Parallel Com-
DQPSK in Slow Rician Fading,” IEEE Electronics Letters, vol. 28, puting. He was the conference chair of the 2003 Workshop on
no. 8, Aug. 1992, pp. 1763–65. High Performance Switching and Routing, and the SPIE Quality
[31] M. Tanda, “Bit Error Rate of DQPSK Signals in Slow Nakagami of Service over Next-Generation Data Networks conference. He
Fading,” IEEE Electronics Letters, vol. 29, no. 3, Mar. 1993, pp. 431–32. has also served on the technical program committee of many
[32] C. Tellambura and V. K. Bhargava, “Unified Error Analysis of national and international conferences, including IEEE INFO-
DQPSK in Fading Channels,” IEEE Electronics Letters, vol. 30, COM, IEEE GLOBECOM, and the IEEE Annual Conference on
no. 12, Dec. 1994, pp. 2110–11. Local Computer Networks. He is a senior member of the IEEE.
[33] M. Zorzi and R. R. Rao, “Lateness of Probability of a Retrans- His current research interests are in quality of service for next-
mission Scheme for Error Control on a Two-State Markov generation Internet, broadband networks, multimedia over
Channel,” IEEE Trans. Commun., vol. 47, no. 10, Oct. 1999, high-speed networks, wireless and satellite networks, TCP/IP
pp. 1537–1548. over ATM, multiprocessor systems, and image processing. He is
[34] E. N. Gillbert, “Capacity of a Bursty-Noise Channel,” Bell Sys- the co-author of the book TCP/IP over ATM Networks (Artech
tem Tech. J., vol. 39, no. 9, Sept. 1960, pp. 1253–65. House). He has taught many short courses to industry in the
[35] E. O. Elliott, “Estimates of Error Rates for Codes on Bursty- area of computer and telecommunication networking. His
noise Channels,” Bell System Tech. J., vol. 42, no. 9, Sept. research has been supported by state and federal agencies
1963, pp. 1977–97. including NASA (USA), the Ohio Board of Regents (USA), and
[36] J. R. Yee and E. J. Weldon, “Evaluation of the Performance of DITARD (Australia). He has more than 100 refereed publications
Error Correcting Codes on a Gilbert Channel,” IEEE Trans. Com- in the above areas, most of which can be accessed at
mun., vol. 43, no. 8, Aug. 1995, pp. 2316–23. http://www.cs.ou.edu/~atiq

IEEE Communications Surveys & Tutorials • Fourth Quarter 2003 9