On Measuring Memory Length of the Error

Rate Process in Wireless Channels

Muhammad U. Ilyas, Student Member IEEE, and Hayder Radha, Senior Member, IEEE

AbstractBit errors are orders of magnitude more frequent in

wireless channels than in wired channels. Design of network
protocols and other architectural components of wireless
networking systems entail a better understanding of the process
that introduces errors into transmissions. A key characteristic of
the error process operating on a channel is its memory length
which determines the degree of similarity and clustering of
errors. In this paper we demonstrate the inadequacy of
correlation coefficient based analysis used in determining memory
length of bit level errors to the packet-level Bit Error Rate (BER)
process. To overcome this problem we propose an alternate means
of measuring channel memory, called the Relative Mutual
Information (RMI) and a lower complexity variant, the Pairwise
RMI. We demonstrate the use of RMI and Pairwise RMI to sets
of Residual Bit-error Traces collected in IEEE 802.15.4 Low
Rate-Wireless Personal Area Network (LR-WPAN) and IEEE
802.11b Wireless Local Area Network (WLAN) channels.
Index TermsChannel memory, Memory length, Relative
mutual information, Correlation coefficient, Residual bit-error
traces, IEEE 802.11, IEEE 802.15.4.

I. INTRODUCTION

OMMUNICATION channels in the real world are not

perfect and are prone to introduce errors into
transmissions. Errors are orders of magnitude more
frequent in wireless channels than in wired channels. Design of
network protocols and other architectural components of
wirelessly networked communication systems entail a better
understanding of the error process that affects transmissions.
The memory length of the error process is an important
parameter of interest that has to be taken into consideration in
the formulation of channel models. In particular, channel
models that take into account the persistence of errors in
wireless channels, such as those based on Markov chains,
require information about the memory length of the error
process. The error process exhibits memory for two reasons;
1. Slow fading: It affects average channel conditions on
Manuscript received February 22, 2008. This work was supported in part
by NSF Award CNS-0721550, NSF Award CCF 0728996, NSF Award CCF0515253, and NSF Award CNS-0430436.
M. U. Ilyas is with the Department of Electrical and Computer
Engineering, Michigan State University, East Lansing, MI 48824 USA
(phone: 517-355-3769; fax: 517-353-1980; e-mail: [email protected]).
H. Radha, is with the Department of Electrical and Computer Engineering,
Michigan State University, East Lansing, MI 48824 USA (e-mail:
[email protected]).

978-1-4244-2247-0/08/$25.00 2008 IEEE.

large timescales of the order of tens to thousands of

seconds.
2. Fast fading: It determines channel conditions on very
small timescales, of the order of fractions of seconds.
Previous works measuring memory length of the bit-level
error process relied on the correlation coefficient based
analysis which worked well enough. But unlike previous bitlevel channel models such as [2],[3],[4],[5], this paper is
concerned with the packet-level process called the Bit Error
Rate (BER) process, denoted by . The BER process is
defined as a series of measurements of the rate at which each
packet has been subjected to bit-errors. The nth measurement in
a realization of is denoted by [n ] and is computed as
shown in (1).
Number of bits in n th packet received with errors
[n ] =
(1)
Total number of bits received in n th packet
In this paper we are concerned with determining the
memory length of errors due to the effects of fast fading and
short bursts of interference. As we will demonstrate, the
determination of this memory length is complicated by several
factors which include slow fading and periodic interference. At
large time scales, slow fading produces non-stationary
behavior in the BER process. An obvious approach would be
to preprocess the BER process by de-trending and
normalization [6]. But these trends change over time, at
unpredictable, irregular intervals.
Like some prior works that measured memory length of
errors, this paper too depends on bit-level, binary signals,
called Residual1 Bit-error Traces (short: error trace or trace),
representing the positions of bit-errors in received
transmissions. The concept of error traces is very simple and is
explained easily enough. An error trace consists of maps of
positions of bit error of all packets collected in the course of a
trace collection session. Conceptually, for a single packet such
a bitmap of errors is obtained by comparing a transmitted
packet (free of errors) with its received version (may contain
errors and failed the Cyclic Redundancy Check). Thus, the
sequence of BERs computed for each packet in a trace
constitutes the BER process.
Channel memory can be loosely defined as the duration for
which the packet-level BER persists in a channel. In this paper
we apply three different methods to identify memory length in
1

The term residual errors was coined by Zorzi and Rao in their seminal
work [1] on wireless error analysis and modeling. Residual errors are the
errors observed at the MAC layer after physical layer processing.

1262

the packet-level error process. These include the correlation

coefficient, and two variants of Relative Mutual Information
(RMI), a new information theoretic measure we propose here.
As we demonstrate, the RMI based analysis is relatively
invariant to non-stationary behavior. We apply these measures
to two sets of error traces. The first consists of traces of IEEE
802.15.4 Low Rate-Wireless Personal Area Networks (LRWPAN) links operating in different channels [9]. The second
consists of IEEE 802.11b Wireless Local Area Network
(WLAN) [8] link traces at data rates of 2, 5.5 and 11Mbps. For
each WLAN data rate, traces are available for four different
throughput rates of 500, 750, 900 and 1024Kbps. We will not
concern ourselves with the exact methodology of collecting
traces. The configuration details and equipment setup used for
collection of IEEE 802.15.4 traces are available in Ilyas and
Radhas work [1], and for IEEE 802.11b traces in Karande,
Khayam, Cho, Misra, Radha, Kim, and Hongs work [7].
The rest of this paper is organized as follows. Section II
demonstrates the use of correlation coefficient in detecting the
memory length of the packet-level BER process. Section III
describes the use of RMI in measuring the memory length of
the BER process observed in different channels of IEEE
802.15.4 LR-WPANs and IEEE 802.11b WLANs at different
throughput rates. Section IV describes a lower complexity
alternative to RMI, called the Pairwise RMI and demonstrates
its use on the same traces as in the previous section. Section V
concludes the paper.
II. MEMORY LENGTH MEASUREMENT BY CORRELATION
COEFFICIENT
A. Description: Correlation Function
Let X be a random process consisting of random variables
X , X , X , , X , . The corresponding measurements in a
0 1 2

N
random process are denoted by x 0 , x 1, x 2 , , x n , . X (m ) is a
m time unit delayed (right-shifted) version of X . To measure
memory length we compute the correlation coefficient of X
0.11
Moderate
BER Interval

High BER
Interval

defined as,
RX (m ) =

Cov X , X (m )

X X ( m )

Here Cov X , X (m )

E (X X ) X (m ) X (m )

.
X X (m )

and E the expectation function. This function of m is also

known as the correlation function or correlogram. The value of

m after which the correlation function RX (m ) becomes

insignificant and drops below threshold Rt is the memory
length M , i.e.

M = min (RX (m ) < Rt ) .

B. Discussion

There is no clear consensus on what the value of Rt should

be. In the simplest of cases the smallest value of m for which
the correlation function is 0 or close to it is taken as a
measure of M . Unfortunately, the effects of slow channel
fades and interference means that average channel conditions
can remain steady for very long stretches of time. Thus, in
traces collected over long periods measurements resemble
their nearby values for rather long durations of time for which
average channel conditions remain steady. To illustrate the
long term steadiness of channel conditions, Figure 1 shows a
plot of the BER observed in a trace collected from an IEEE
802.15.4 link operating in channel 25. To make the average
trends more visible at large timescales the BER process has
been processed by a 600 point averaging filter. There are
clearly three regions visible in the plot. For approximately the
first 1500sec the average BER remains at a moderate level,
then it increases to a high level for another 1000sec before it
settles to a low level for the remainder of the trace. These long
term effects complicate correlogram analysis and the

Low BER Interval

MC-11
MC-12
MC-13
MC-14
MC-15
MC-16
MC-17
MC-18
MC-19
MC-20
MC-21
MC-22
MC-23
MC-24
MC-25
MC-26

0.4

0.07
0.06
RB(t*10)

0.3

0.05
0.04
0.03

0.2
0.1

0.02
0

0.01
2000

(3)

IEEE 802.15.4

1000

(2)

denotes the covariance of X and

0.08

X (m ) , X the standard deviation of X , X the mean of X ,

0.09

Averaged BER

3000

4000

5000

Figure 1. BER process observed in a trace of an IEEE 802.15.4 link

operating in channel 25. The averaged BER values plotted here were
obtained from an averaging filter of length 600, corresponding to a 1
minute interval.

0.8
0.6
0.4
0.2
0

10

20
30
time (sec)

a.

6000

time (sec)

IEEE 802.11b

RB(k)

0.1

and X (m ) for a range of values of m . Thus, the correlation

coefficients denoted by RX (m ) are a function of m and are

40

50

50

100

150
k

200

250

300

b.

Figure 2. Correlogram functions, a) for 16 different IEEE 802.15.4 traces

spanning all of its 16 channels, and b) for 41 different IEEE 802.11b traces at
2, 5.5 and 11Mbps link speeds, each at throughput rates of 500, 750, 900 and
1024Kbps .

1263

determination of a threshold Rt . However, there is a 95%

III. RELATIVE MUTUAL INFORMATION

significance level that is frequently used as a rule of thumb to

A. Shannon Information Measures

For random variables X and Y , with Probability Density
Functions (PDF) pX (x ) and pY (y ) and joint PDF pXY (x , y ) ,
the mutual information I (X ;Y ) is defined as,

determine whether M is zero (successive measurements in

X are independent) or non-zero (successive measurements in

X are not independent). For a signal of N consecutive
measurements the 95% significance range is defined as
range 1 m N 4 lie within this specified range, then
consecutive measurements are deemed independent, leaving

M = 0 .
This is amply demonstrated by the correlogram functions in
Figure 2.a for 16 different IEEE 802.15.4 traces (labeled MC11 through MC-26) spanning all of its 16 frequency channels,
and Figure 2.b for 41 different IEEE 802.11b traces spanning
link speeds 2, 5.5 and 11Mbps, each at throughput rates of
500, 750, 900 and 1024Kbps. The length of each of the traces
is truncated to a uniform N = 100, 000 points, making the
95% significance range 0.0063, +0.0063 . Clearly, by this
measure all channel traces exhibit memory to some degree. To
determine just how much memory several rules of thumb have
been used in literature. These include;
1. Rt = min (R (i ) = 0) , i.e. correlation coefficient falls
i

close to insignificance/ zero. In all our analyzed traces this

has rarely been the case.
2. Rt = 0.1 R (1) , i.e. correlation coefficient drops to less
than 10% of the coefficient at lag 1. By this standard, the
memory length of the BER process will range in the tens
to thousands of seconds.

3. Rt = min R (i ) R ( j ) < : j > i, 0 ,

i

i.e.

correlation coefficient becomes steady and subsequent

changes between consecutive values are within a very
small value . Periodic interference, most likely from
IEEE 802.11b beacon frames, causes periodic spikes in
the correlogram functions rendering this criterion useless.
Furthermore, there is no clear interpretation of the value
of correlation coefficients and the degree of predictability
of one measurement on another.
According to all these selection criteria for Rt

, memory

lengths of all these traces lie in the range of tens to hundreds of

seconds. Clearly, that rules out correlogram analysis to
measure memory length of fast fading effects.

I (X ;Y ) =

2 / N . If 95% of correlation coefficients for m in the

pXY (x, y ) log2

x ,y :pXY (x ,y )> 0

pXY (x , y )
.
pX (x ) pY (y )

(4)

Mutual information [11] can also be understood more

intuitively in terms of entropy as,
I (X ;Y ) = H (X ) + H (Y ) H (X ,Y ) .
More generally, mutual information is defined between two
sets of random variables. If random variables X and Y in (4)
are replaced by sets of random variables X1, X 2 , , X n and
Y1 ,Y2 , ,YM , respectively, then pX (x ) ,
pXY (x , y ) are replaced by respective

pX X
1

...XN

(x , x ,..., x )
1

and pY Y ...Y
1 2

pY (y ) and
joint PDFs

(y , y ,..., y ) ,
1

joint
PDF
of
all
random
pX1X2 ...XNY1Y2 ...YM (x 1, x 2 ,..., x N , y1 , y2 ,..., yM ) .

and the
variables

B. Description: Relative Mutual Information

We are essentially faced with the challenge of evaluating;
How much information a random variable Y can provide
about another random variable X ;
While at the same time providing a measure of the
remaining uncertainty about X .
The Shannon mutual information I (X ;Y ) achieves the first
goal. However, mutual information is not restricted to a fixed
range. Thus, an evaluation of I (X ;Y ) does not give us a
sense of how much information about X or Y remains
unknown. We use Relative Mutual Information (RMI),
previously used by [10], denoted by RMI (X ;Y ) ;

RMI (X ;Y ) =

I (X ;Y )
H (X )

(5)

Note that while I (X ;Y ) is a symmetric measure, by

definition
RMI
is
non-symmetric,
i.e.
RMI (X ;Y ) RMI (Y ; X ) because H (X ) H (Y ) . Since

I (X ;Y ) min (H (X ), H (Y )) , the RMIs value is limited to

the interval [0,1] . An RMI close to 1 implies Y contains most
of the information contained in X , leaving little uncertainty,
while an RMI close to 0 implies the opposite.
In the current context we replace X by the BER process
(0)
, and replace Y by the BER process of preceding packets
(1) , i.e. a one-right shifted version of (0) . Then

RMI (0) ; (1)

measures the amount of information that a

packets BER [n ] shares, on average, with the following

packets BER [n 1] . A natural extension of this measure

1264

would be to include more than just the immediately following

packet, but include any arbitrary number m such that,
RMI ((0) ; (1) , (2) , , ( m ) ) =

I ((0) ; (1) , (2) , , ( m ) )

H ((0) )

BER becomes smaller than , i.e.

M ( ) = max ( RMI (1, m 1) RMI (1, m) )
m

Since we are assuming the BER process to be wide sense

stationary, the RMI becomes a function of m , the number of
immediately following measurements. We use the abbreviated
notation RMI (1, m ) for the RMI in (6). Recall that
RMI (1, m ) is a function of the Shannon mutual information
(and hence the joint PDF) of BER processes through (m ) .
Ideally, once m exceeds the channels memory length M ,
become independent of (m ) and the Shannon mutual
information (4) will drop to zero, producing a zero RMI in (5).
But as in the case of correlogram based analysis, the presence
of slow fading complicates interpretation of RMI. Therefore,
for the duration of the slow fade a packets BER will remain
weakly correlated with its followers. Hence, after m > M ,
this translates into a slowly increasing RMI (1, m ) for
successive values of m until it finally becomes 1. However,
for the problem at hand we are not particularly interested in
slow fades but the memory length of fast fades. Therefore, we
define channel memory length M as the time it takes for the
RMI (1, m ) function to rise to a level after which it grows
only slowly. Then channel memory M is determined by
computing the RMI function for increasing values of m and
setting M equal to the largest lag m for which the amount of
additional RMI provided by including the (m + 1) delayed

Thus the memory length M ( ) is a function of , the

operating channels 11, 13, 15, 17, 25 and 26 for a range of lag
values m . The middle figure plots the RMI (1, m )
functions of the same traces for the same range of lags. The
bottom figure plots M ( ) the memory length as a function of
increments .
Similarly, the top row in the matrix of Figure 5, Figure 5.a
plots the same three quantities, RMI (1, m ) , RMI (1, m )
(defined as RMI (1, m 1) RMI (1, m ) ) and M ( ) for
different traces collected on IEEE 802.11b channels. The
leftmost figure plots these quantities on 2Mbps at throughputs
of 500, 750, 900 and 1024Kbps. Figure 5.b and Figure 5.c
repeat the same plots for the same range of throughput rates
for traces collected on links operating at 5.5 and 11Mbps,
respectively.

IEEE 802.15.4

RMIB(1,m)

RMIB(1,m)

0.6
0.4
0.2
10

15

20

25

30

35

0.2

0.1
0.05
5

10

15

25

30

0.2

35

40

10
0.1

0.15

15

0.2

0.25

25

30

35

40

25

30

35

40

0.2
0.15
0.1
0.05
5

10

15

20

Lag m

30
20
10
0

0.3

20

Lag m

40

20

0.05

10

0.25

Lag m

30

0.4

MB()

MB()

20

0.6

0.15

40

MC-11
MC-13
MC-15
MC-17
MC-25
MC-26

0.8

40

Lag m

RMIB(1,m)

RMIB(1,m)

MC-11
MC-13
MC-15
MC-17
MC-25
MC-26

0.8

(7)

significant RMI increment threshold. That leaves us with the

choice an appropriate value of . RMI is understood to be the
fraction of Shannon information of one random variable
contained collectively in another set of random variables.
Thus, in terms of RMI the memory length of the BER process
can then be understood as the lag m after which every
subsequent BER measurement will contribute less than
100 % of new information about (0) . The top figure in
Figure 3 is the RMI (1, m ) of traces of IEEE 802.15.4 links

IEEE 802.15.4

0.25

m [1, ]

(6)

0.05

0.1

0.15

0.2

0.25

0.3

a.
Figure 3. (from top to bottom). [Top] RMI (1, m ) of BER process observed in

b.
Figure 4. (from top to bottom). [Top] RMI (m ) of BER process observed in

a trace of IEEE 802.15.4 links operating in channels 11, 13, 15, 17, 25 and 26
for lag m varying from 1 through 40. [Middle] RMI (1, m ) of BER

a trace of IEEE 802.15.4 links operating in channels 11, 13, 15, 17, 25 and 26
for lag m varying from 1 through 40. [Middle] RMI (m ) of BER process

process for the same channel traces. [Bottom] The memory length M plotted

for the same channel traces. [Bottom] The memory length M plotted as a

as a function of , the increments in RMI (1, m ) .

function of , the increments in RMI (m ) .

1265

this methodology especially unsuitable for applications in

which this computation is to be performed online. We
repeatedly performed this computation offline on different data
sets, and found that even on a mid-level server class machine
the computation of the RMI function based on 50,000 data
points for m ranging from 1 to 40 takes more than 30
minutes.

C. Discussion
The feature that makes RMI attractive for use as a tool for
memory measurement is not the fact that provides an
unequivocal measurement of the BER process memory length.
Like the correlation coefficient before it, the RMI depends on
the subjective selection of a threshold that determines the
cutoff between significance and insignificance. Rather, its
strength lies in the fact that there is a clear interpretation of the
threshold (in this case ) in information theoretic terms, e.g.
= 0.15 implies that on average, an M delayed

IV. PAIRWISE RMI BASED MEMORY MEASUREMENT

A. Description: Pairwise RMI

In view of the fact that the RMI is difficult and slow to
compute, we evaluate another measure for memory length that
allows a similarly intuitive interpretation for selecting the
cutoff threshold, but without the high computationally
complexity. Therefore, we take a different approach for the
determination of M . Instead of relying on RMI as computed
in (6), we go back to its definition in (5) and use

measurement contains at least 15% information about the

current measurement. Another advantage of using RMI over
the commonly used correlation coefficient is that the changing
trends operating on the timescales of slow fades that cause
non-stationary behavior of the BER process do not complicate
the analysis.
However, there is a major practical challenge to the use of
RMI in this manner. The computation of RMI (1, m ) is
based on an m + 1 dimensional joint PDF of all processes
(0) , (1) , (2) ,, ( m ) . Populating such a high dimensional
PDF requires large data set even for moderate values of m .
Collecting data points to fill a modest 10 dimensional PDF
takes significant time, especially when considering that
802.15.4 is a low rate communication standard. This makes
2 Mbps

30

35

40

45

0.15
0.1
0.05
15

20

25

30

35

40

45

50

20
10
0.06

0.08

0.1

20

25

0.12

0.14

0.16

10

15

20

25

0.18

RMIB(m)

RMIB(m)

0.1
0.05
30

35

40

45

0.06

0.08

0.1

30

35

40

45

10

10

15

20

25

30

35

40

45

15

20

0.08

25

0.1

0.12

0.14

0.16

30

35

40

45

0.1

0.12

50

10
0.1

0.15

0.2

0.25

0.3

0.35

0.4

500Kbps
750Kbps
900Kbps
1024Kbps

RMIB(1,m)

0.2

0.1
0.05
5

10

15

20

25

30

35

40

45

50

30

35

40

45

50

0.2
0.1
0
5

10

15

20

0.14

0.16

0.18

0.2

0.22

25

Lag m

40
30
20
10
0

0.08

45

0.15

MB()

10
0.06

40

20

50

20

0.04

35

30

50

30

0.02

30

0.25

Lag m

40

25

Lag m

20

50

40

50

-0.2
15

45

c.

-0.1

10

40

0.1

35

Lag m

0.05

0.2

MB()

10
0.06

0.22

500Kbps
750Kbps
900Kbps
1024Kbps

50

20

0.04

0.2

30

IEEE 802.11b at 11Mbps

0.1

50

30

0.02

0.18

0.2

Lag m

40

0.16

0.3

RMIB(m)

RMIB(m)

MB()

RMI(m)

50

0.14

25

0.1

50

Lag m

-0.1

25

0.12

20

0.2

10
0.04

15

0.3

50

20

0.02

10

Lag m

Lag m

Lag m

20

45

30

50

0.1

15

5
0.4

b.
500Kbps
750Kbps
900Kbps
1024Kbps

10

40

0.2

IEEE 802.11b at 5.5Mbps

0.2

35

0.4

0.15

25

30

0.6

50

40

0.25

20

45

0.1

a.

15

40

0.05

IEEE 802.11b at 2Mbps

10

35

0.15

30

0.2

MB()

30

0.04

15

50

40

0.02

10

Lag m

Lag m

50

0.2

500Kbps
750Kbps
900Kbps
1024Kbps

0.8

MB()

10

0.4

50

Lag m

0.6

RMI B(m)

25

500Kbps
750Kbps
900Kbps
1024Kbps

0.8

RMIB(m)

RMIB(1,m)

20

11 Mbps
IEEE 802.11b at 11Mbps

RMIB(1,m)

0.2

MB()

RMI(1,m)

RMIB(1,m)

0.4

RMIB(1,m)

RMIB(1,m)

500Kbps
750Kbps
900Kbps
1024Kbps

0.6

15

RMI (1, m ) , we compute an RMI function. However, this

time around the RMI is computed between BER processes and
its m steps delayed version. We distinguish this RMI from the
previous one by calling it a Pairwise RMI and denote it by
RMI (m ) . Again, a non-zero value of the RMI (m )

IEEE 802.11b at 5.5Mbps

0.8

10

5.5 Mbps

IEEE 802.11b at 2Mbps

RMI (0) ; (m ) = RMI (m ) . As in the previous use of

0.05

0.1

0.15

d.

e.

f.

0.2

0.25

Figure 5. Plots of RMI (1, m ) , RMI (1, m ) and M ( ) at 500, 750, 900 and 1024Kbps for IEEE 802.11b WLAN links operating at, a) 2Mbps, b)
5.5Mbps, and c) 11Mbps. Plots of RMI (m ) , RMI (m ) and M () at 500, 750, 900 and 1024Kbps for IEEE 802.11b WLAN links operating at, d)
2Mbps, e) 5.5Mbps, and f) 11Mbps.

1266

indicates the presence of some common information between

packets separated by m packets. Recall that RMI (m ) is a
function of the Shannon mutual information (and hence the
joint PDF) of BER processes of m separated packets.
Therefore, due to the effects of slow fading discussed in the
preceding section even after m exceeds the true value of M ,
the value of RMI (m ) will not fall to zero. We conclude that
the determination of M using the Pairwise RMI function be
made as,
M () = max ( RMI (m 1) RMI (m) )
m
.
(8)
m [1, ]
Thus memory length is defined as the maximum lag m
between measurements for which RMI (m) (defined as
RMI (m 1) RMI (m ) ), the variation in RMI from

RMI (m 1) to RMI (m) , remains larger than . As in the

case of RMI, in the case of Pairwise RMI too the memory
length M () is a function of the threshold . Once the lag

the steadiness of the Pairwise RMI. The Pairwise RMI

function RMI (m) is considered steady once changes between
successive values become less than 100 % .
V. CONCLUSIONS
In this paper we highlighted the inadequacy of using the
traditional correlogram function to measure the memory length
of the BER process that operates on transmission in wireless
channels. The use of the correlogram requires the selection of
a threshold that is difficult to justify or interpret. We offer as
an alternative the RMI, and information theoretic measure
based on the Shannon mutual information. Although high in
computational complexity, the threshold value chosen to
discriminate between significant and insignificant levels of
mutual information has an information theoretic interpretation.
RMI also does not seem to suffer from any complications
arising due to periodic interference. From RMI we derive
Pairwise RMI, a simplified, less complex measure for
measuring memory length. Pairwise RMI is not completely
immune from the effects of periodic interference like RMI.

m approaches M () the information contained in later

measurements will approach a steady value. The fact that this

value is not completely zero even for packets separated by
large lags, is because of slow fading even distant packets will
provide some information due to similar average channel
conditions.
Figure 4 shows three plots for the same IEEE 802.15.4
traces used in the demonstration of RMI in the previous
section; At the top we show the Pairwise RMI functions for
traces collected from links operating in channels 11, 13, 15, 17
25 and 26. The middle plot shows the successive decrements
in Pairwise RMI with increasing lag values m . The bottom
figure plots the memory length M () as a function of , as
defined in equation (8). Figure 5.d plots

RMI (m) and

RMI (m) ,

M () for four different traces of IEEE

802.11b WLAN links operating at 2 Mbps at throughput rates

of 500, 750, 900 and 1024Kbps. Figure 5.e and Figure 5.f
repeat these same plots for links operating at 5.5 and 11Mbps,
respectively.
B. Discussion
The demonstration of using Pairwise RMI shows that it is
significantly easier to compute than RMI. The memory length
M () is a function of which is interpreted as a measure of

REFERENCES
[1]

M. Zorzi and R.R. Rao, On the statistics of block errors in bursty

channels, IEEE Transactions on Communications, vol. 11, 1997.
[2] S.A. Khayam, S. Karande, H. Radha, and D. Loguinov, "Performance
analysis and modeling of errors and losses over 802.11 b LANs for
high-bit-rate real-time multimedia," Signal Processing: Image
Communication, vol. 18, pp. 575-595, 2003.
[3] S.A. Khayam and H. Radha, "Linear-Complexity Models for Wireless
MAC-to-MAC Channels," Wireless Networks, vol. 11, pp. 543-555,
2005.
[4] S.A. Khayam and H. Radha, "Constant-complexity models for wireless
channels," in Proceedings of the 25th Annual Joint Conference on the
IEEE Computer and Communications Societies (INFOCOM06),
Barcelona, Spain, 2006.
[5] M.U. Ilyas, and H. Radha, "Measurement based analysis and modeling
of the error process in IEEE 802.15.4 LR-WPANs," in Proceedings of
the 27th Annual Joint Conference on the IEEE Computer and
Communications Societies (INFOCOM08), Apr 15-17, 2008.
[6] G. P. Williams, Chaos Theory Tamed: Joseph Henry Press, 1997.
[7] S. Karande, S. A. Khayam, Y. Cho, K. Misra, H. Radha, J. Kim, and J.
W. Hong, On Channel State Inference and Prediction Using
Observable Variables in 802.11b Network, IEEE International
Conference on Communications (ICC07), pp. 4554-4559, 2007.
[8] ANSI/IEEE Std 802, Part 11: Wireless LAN Medium Access Control
(MAC) and Physical Layer (PHY) Specifications, 1999.
[9] ANSI/IEEE Std 802, Part 15.4: Low Rate Wireless Personal Area
Networks, Sep 8, 2006.
[10] R. K. Colwell, and D. J. Futuyma, "On the Measurement of Niche
Breadth and Overlap," Ecology, vol. 52, pp. 567-576, 1971.
[11] T. M. Cover and J. A. Thomas, Elements of Information Theory: WileyInterscience New York, 2006.

1267