Static Al

INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMS
Int. J. Commun. Syst. 2005; 18:267–284

Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/dac.703
Statistical impulse response models for indoor optical

wireless channels
J. B. Carruthers1,n,y and S. M. Carroll2

1
Department of Electrical and Computer Engineering, Boston University, Boston, MA, U.S.A.
2
IBM Global Services, New York, NY, U.S.A.
SUMMARY
The indoor optical wireless channel is significantly different than the radio channel. Statistical propagation
models developed for the radio channel, which characterize path loss, shadowing, and multipath fading, do
not apply. In this paper, we investigate statistical modelling for the indoor optical wireless channel through
the examination of the characteristics of a large set of channel impulse responses. The channel responses
are generated using an estimation method based on geometrical modelling of indoor environments together
with an iterative technique for calculating multiple reflections. We confirm prior studies showing that
channels with line-of-sight paths must be modelled separately from those fully diffuse channels with no
such path. We show that the distribution of the channel gain in dB for the LOS component follows a
modified gamma distribution, and the channel gain in dB for LOS channels including all reflections follows
a modified Rayleigh distribution for most transmitter–receiver distances. Similar results are described for
distributions of channel gains of diffuse channels and for distributions of rms delay spreads. Finally, we
describe a method for generating a statistically realistic impulse response given any transmitter–receiver
separation in an indoor environment. Copyright # 2005 John Wiley & Sons, Ltd.
KEY WORDS: optical wireless communications; channel simulation; statistical channel modelling
1. INTRODUCTION
High-quality wireless access to information, networks, and computing resources by users of

portable computing and communication devices is driving recent activity in indoor optical
communication [1–4]. High-quality access is achieved via links with low delay, high data rates,
and reliable performance, and accurate characterization of the channel is essential to
understanding the performance limits and design issues for optical wireless links.
Electromagnetic waves at optical frequencies exhibit markedly different propagation behavior
than those at radio or microwave frequencies. At optical frequencies, most building surfaces are
n
Correspondence to: Jeffrey B. Carruthers, Department of Electrical and Computer Engineering, Boston University,
8 St. Mary’s Street, Boston, MA 02215, U.S.A.
y
E-mail: [email protected]
Contract/grant sponsor: U.S. National Science Foundation; contract/grant number: ECS-9876149
Received 3 June 2004

Revised 11 September 2004
Copyright # 2005 John Wiley & Sons, Ltd. Accepted 20 September 2004
268 J. B. CARRUTHERS AND S. M. CARROLL
opaque, which generally limits the propagation of light to the transmitter’s room. Furthermore,
for most surfaces, the reflected light wave is diffusely reflected (as from a matte surface) rather
than specularly reflected (as from a mirrored surface).
These differences, as well as fundamental differences in the transmitting and receiving devices,
have led researchers to develop channel and communication concepts for optical wireless
systems and channels. In particular, characterization for optical wireless channels has been done
by a variety of methods at different levels. Basic system models were developed in References
[5, 6]. Measurement studies [7–9] have validated the basic diffuse reflection model and have
shown the importance of the orientation of the transmitter and receiver as well as the
importance of shadowing. Statistical models of channel characteristics [10] have attempted to
make sense of the important factors illustrated in the above measurement studies. In addition to
extensive channel characterization by measurement in Reference [9], the authors use a statistical
model based on the gamma distribution to model path loss variations in LOS channels as
determined by angle deviations. They observed variations of a few decibels, whereas we see
variations of tens of decibels since we consider all channels covering the same distance rather
than a single configuration with an angle variation. In Reference [11], the authors develop a
physical channel model with emphasis on the relationship between the LOS and diffuse
components; their goal is to allow for swift calculation of path loss and multipath in a particular
configuration, rather than to characterize the variation of those parameters.
In the present paper, we develop statistical models for characterization of the impulse
response of indoor optical wireless channels. The model development is made possible by recent
advances in impulse response calculation which allow fast and accurate computation of
thousands of impulse responses and which are not limited to including only two or three
reflection orders [12–14]. The method presented in Reference [14] uses a concise matrix
formulation for the problem but does not include shadowing effects.
In the next section, we describe models for characterizing the properties of transmitters,
receivers, and reflecting surfaces within the indoor environment and present an efficient method
for impulse response calculation. In Section 3, we present the results of a study of propagation
characteristics for a large ensemble of channels in a variety of rooms. In Section 4, we use these
results to develop statistical distributions for important channel characteristics. Conclusions are
presented in Section 5.
2. CHANNEL ESTIMATION METHOD
We model optical wireless channels formed by a transmitter and receiver placed inside a
reflective environment. The transmitter or source S is a laser diode or a light-emitting diode
transmitting a signal XðtÞ using intensity modulation (IM). The signal received by receiver Ri
when source S is transmitting is YðtÞ; the current from the photodiode:
YðtÞ ¼ rXðtÞ * hðtÞ þ NðtÞ ð1Þ
where * denotes convolution, hðtÞ is the impulse response of the channel between source S and
receiver Ri ; r is the photodiode responsivity, and NðtÞ is noise at the receiver. This baseband
impulse response for IM/DD communication [6] is fixed and completely determined for a given
set of source properties S; receiver properties Ri ; and environment properties E; and hence we
will write hðtÞ more specifically as hE ðt; S; Ri Þ:
Copyright # 2005 John Wiley & Sons, Ltd. Int. J. Commun. Syst. 2005; 18:267–284
STATISTICAL MODELS OF INDOOR OPTICAL CHANNELS 269
We provide an outline of the impulse response calculation method developed in References

[12, 13]. The calculation involves decomposition into bounces, discretization into facets, and
finally multi-receiver iteration.
All transmitted light arriving at the receiver has undergone a definite number of reflections or
bounces. Hence, we can decompose the impulse response hE ðt; S; Ri Þ as
X
1
hE ðt; S; Ri Þ ¼ hEðkÞ ðt; S; Ri Þ ð2Þ
k¼0
where hEðkÞ ðt; S; Ri Þ is the impulse response due to signal light undergoing exactly k bounces
during its path from the source S to the receiver Ri :
The line of sight impulse response hð0Þ E ðt; S; Ri Þ is given by

ð0Þ Ari gðyi Þ Di
hE ðt; S; Ri Þ ¼ Vð~ rri ; EÞTðfij Þ
rs ;~ d t ð3Þ
D2i c
where Di ¼ j~ rs ~
rri j is the distance between the source and the receiver. The visibility function
rs ;~
Vð~ rri ; EÞ is 1 when the LOS path between S and Ri is unobstructed, and is zero otherwise.
Now, the k-bounce response can be calculated using the ðk 1Þ-bounce response using
Z
hðkÞ
E ðt; S; R i Þ ¼ rder hEðk1Þ ðt; S; der Þ * hð0Þ s
E ðt; de ; Ri Þ ð4Þ
E
where the integral is over all surfaces in E and r is the surface reflectivity function. The surfaces
act as receivers with
8
< cosðyÞ if 04y4p=2
gðyÞ ¼
:
0 otherwise
and as first-order Lambertian transmitters. Since jjhEðkÞ ðt; S; Ri Þjj1 ! 0 as k ! 1 since r51
everywhere, we can estimate hE ðt; S; Ri Þ using (2) by considering only the first M bounces. Hence
X
M
hE ðt; S; Ri Þ hðkÞ
E ðt; S; Ri Þ ð5Þ
k¼0
Excellent approximations can be obtained for M ranging from 3 to 10, as discussed in

References [12, 13].
The integration in (4) is approximated by representing each face Fi at a spatial partitioning
factor P; i.e. each face is divided into small elements of size 1=P 1=P m2 : Hence, we estimate
hðkÞ
E ðt; S; Ri Þ using
X
N
hðkÞ
E ðt; S; Ri Þ rern hEðk1Þ ðt; S; ern Þ * hð0Þ s
E ðt; en ; Ri Þ ð6Þ
n¼1
where ern and esn represent element n acting as a receiver and a source, respectively.
The key step is to consider the surface facets as receivers. We apply (6) with Ri ¼ erm to obtain
X
N
hðkÞ r
E ðt; S; em Þ rern hEðk1Þ ðt; S; ern Þ * hð0Þ s r
E ðt; en ; em Þ
n¼1
X
N
¼ amn hEðk1Þ ðt tmn ; S; erm Þ ð7Þ
n¼1
where
rerm Tðfmn Þgðymn Þ
amn ¼ Vð~
resn ;~
rerm ; EÞ ð8Þ
P2 D2mn
and tmn ¼ Dmn =c: The quantities fmn ; ymn ; and Dmn are the receiver’s angle to the source, the
source’s angle to the receiver, and the source-to-receiver distance, respectively, for the source esn
and the receiver erm :
We note that evaluation of these N equations of (7) for k 1 allows for iteration to k: A key
observation is that the previously calculated impulse responses hEðk1Þ ðt; S; ern Þ do not depend on
the properties of the individual receiver, and thus can be computed once, independently of the
number of receivers involved. Hence, the only calculation required for each receiver is the
‘collection stage’ given by (6).
The final calculation stage is to combine the k-bounce impulse responses into an estimate of
the overall impulse response, as shown in (5).
3. CHANNEL ENSEMBLES
Using this channel estimation tool, called IrSimIt [15], we will investigate propagation
characteristics for a large ensemble of transmitter and receiver locations and orientations in a
suite of rooms. In each case, we consider the impulse response between a single transmitter and a
single receiver. More than eighty thousand impulse responses were calculated in total, arising
from different room sizes, transmitter locations and orientations, and receiver locations and
orientations.
3.1. Configuration
We create models of empty rooms ranging from small offices to large classrooms and conference
rooms and evaluate channel characteristics. Each room is divided into 27 equal-sized cubes and
one transmitter is placed uniformly within each cube. As discussed above, IrSimIt allows
multiple receivers for each transmitter without significantly increased simulation time. We place
ten receivers uniformly distributed over the sphere of radius D centered at the transmitter,
rejecting any receiver placement that causes a receiver to be outside the room. The reflectivities
of the walls are assumed to be 0.9, the ceiling 0.8 and the floor 0.2. These values are typical of
predominantly white rooms; lower values should be considered for furnished indoor
environments.
The transmitter radiant intensity pattern is Lambertian. The receiver field-of-view (FOV) is
set at p=2 and the area Ari ¼ 104 m2 : For each transmitter location, 10 different transmitter
orientations are simulated. These are combined with 10 receiver positions, for a total of one
0.15 0.15
0.82563
Channel impulse response (1/s)

Channel impulse response (1/s)
0.1 0.1
0.05 0.05
0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
(a) Time (ns) (b) Time (ns)
Figure 1. Typical impulse responses for a transmitter and receiver separated by 0:8 m in a 4 6 m2 room:
(a) a channel containing a LOS path; and (b) a channel with no LOS path.
hundred cases at each transmitter location. Each of these one hundred cases uses a random,
spherically uniform receiver orientation. The orientation angle of the transmitter and receivers is
determined by two random variables, the elevation and azimuth. The azimuth is a uniform
random variable between ½0; 2p; and the cosine of the elevation is a uniform random variable
between [0,1]. We assume all transmitters are pointing somewhere into the room to eliminate
calculations when a transmitter is facing the wall. The range of distances ðDÞ evaluated between
the transmitter and receivers will vary depending on room size. The ceiling height is 3 m in all
rooms considered.
Since the rooms are empty, all combinations of transmitter and receiver will result in some
received power. The impulse responses from the simulations are evaluated and two pieces of
data are collected, channel gain and root-mean-square delay spread. Based on the relationship
between rms delay spread and multipath power penalty for optical wireless channels [10, 16], we
choose to predict and model the multipath in the impulse response through the rms delay
spread. Typical impulse responses for a channel with a LOS path and a channel without a LOS
path can be seen in Figure 1. These agree with previously reported impulse responses [6, 17].
3.2. Channel gain

Channel gain is defined as the ratio between the received power and the transmitted power. The
channel gain in dB is equal to the received power in dBW when 1 W is transmitted. Channel
gain is the single most important feature of an optical wireless channel, as it determines the
achievable signal-to-noise ratio for fixed transmitter powers and is important regardless of the
data rate or modulation scheme employed [1].
Figure 2(a) shows typical channel gain distributions of the data collected from IrSimIt.
Although only data for a 4 6 m2 room is shown, all rooms measured experienced similar
trends. When analyzing the histograms, we notice that two distinct curves exist. We hypothesize
that having a LOS component in the channel may be causing this effect. Hence in Figure 2(a) the
channels containing a LOS component are highlighted. As D increases, the channel gain
distribution of the LOS channels merges with that of the channels containing no LOS path. As
450 350
LOS Mean LOS Mean
400 Diffuse Mean Diffuse Mean
All Channels − 56.7 300 All Channels
LOS Channels 443 LOS Channels
350
250
Number of Samples
− 48.9
Number of Samples
300 1.13
200 5.38
250
200 150
150
100
100
50 50
0 0
− 100 − 90 − 80 − 70 − 60 − 50 − 40 − 30 −20 0 2 4 6 8 10 12 14 16 18
(a) Channel Gain (dB) (b) Delay Spread (ns)
Figure 2. Channel histograms for Dj ¼ 0:8 m in a 4 6 m2 room. Channels with a LOS component are
highlighted: (a) channel gain; and (b) RMS delay spread.
shown in (3), the LOS path hð0Þ 2

E ðÞ is inversely proportional to D ; causing the LOS component to
become less prominent at large values of D: The distributions of the non-LOS channels fall in a
similar channel gain range for all D:
3.3. Multipath dispersion

It is important not only to look at channel gain but also the rms delay spread Srms of the
channel, which is calculated using the impulse response hðtÞ according to
"R #1=2
ðt mÞ2 h2 ðtÞ dt
Srms ¼ R ð9Þ
h2 ðtÞ dt
where m; the mean delay, is calculated by

Z Z
2 2
m¼ th ðtÞ dt h ðtÞ dt ð10Þ
The delay spread will be increasingly important for higher data rates [1]. For example, when
Srms ¼ 0:2=R; where R is the data rate, the power penalty for on-off keying modulation is about
2 dB for typical channels. Thus, delay spreads of 4 ns cause power penalties of 2 dB at a data
rate of 50 Mb=s: Schemes such as PPM that employ narrow pulses are even more susceptible to
delay spread effects.
Figure 2(b) shows typical rms delay spread distributions of the data collected from IrSimIt.
Notice here that a large majority of the small delay spread values have a LOS channel. At
D ¼ 0:1 m the delay spread (Srms ) is 0:09 ns on average for all different room sizes. The delay
spread is especially low here and at D ¼ 0:2 m because the power received from the LOS
component is so strong that additional power from the first and second bounce are virtually
null. As distance increases the initial power received from the LOS path falls off, making any
additional multi-path components more effective.
4. STATISTICAL MODELLING
We develop a model for impulse response characterization which requires only knowledge of the
distance separating the transmitter and receiver and room parameters. Since different models
will be used for LOS and diffuse channel, we first develop a method to determine if a transmitter
and receiver have a LOS path. We then present models for LOS channel gain, diffuse channel
gain, LOS delay spread and diffuse delay spread.
4.1. LOS path existence

To determine LOS path existence, we develop the following simple model. Assume the
transmitter and receiver are inside a room with infinite walls and are placed at a separation
distance D ¼ 1: This creates an area, equal to the surface area of a sphere, which could contain
the transmitter SAsphere ¼ 4p: Assuming random (uniform) and independent orientation of the
transmitter and receiver, then each must be able to ‘see’ the other in order for a LOS path to
exist. Thus
Pr½LOS path ¼ Pr½Rx sees Tx Pr½Tx sees Rx ð11Þ
The surface area of the sphere visible to the receiver with half-angle field of view FOV is

2 FOV
SAFOV ¼ 4p sin ð12Þ
2
The probability that the receiver can see the transmitter is the ratio of SAFOV to the entire
area SAsphere ; hence

FOV
Pr½Rx sees Tx ¼ sin2 ð13Þ
2
Similarly the probability that the transmitter is transmitting radiation in the receiver’s
direction is based on the half-angle transmission angle (tang ), which we have fixed as a
Lambertian making tang ¼ p=2:
t 1
ang
Pr½Tx sees Rx ¼ sin2 ¼ ð14Þ
2 2
and so

1 2 FOV
Pr½LOS path ¼ sin ð15Þ
2 2
Since a FOV ¼ p=2 was chosen for our experiment, the expected fraction of channels
containing an LOS path is 25%.
4.2. Models of LOS path channel gain

The LOS impulse response for a receiver Ri in an environment E is given by
hð0Þ
E ðt; S; Ri Þ ¼ TðfÞAðyÞ=D
2
ð16Þ
where S is the transmitter or source. The effective optical collection area at incident angles of y is
AðyÞ ¼ Ar gðyÞ where gðyÞ is the receiver optical gain function and Ar is the total optical
collection area. The distance separating the transmitter and receiver is D: The radiant intensity
pattern of the source at angle f is given by

nþ1
TðfÞ ¼ cosn ðfÞ ð17Þ
2p
for Lambertian order n: We restrict our attention to Lambertian order n ¼ 1; so
TðfÞ ¼ cosðfÞ=p: The receiver optical gain for bare detectors is gðyÞ ¼ cosðyÞ: Now (16) can
be re-written as:
Ar
hð0Þ
E ðt; S; Ri Þ ¼ cosðfÞ cosðyÞ ð18Þ
pD2
We hypothesize that we can use (18) as a basis to model LOS characteristics for both the
mean and distributions of LOS channels based on D: In this section, probability distributions
are compared with the histograms of the collected data. First we assign the random variable V
to the distribution of cosðyÞ; and W to the distribution of cosðfÞ: The uniform random
orientation of the transmitter and receiver results in V and W i.i.d. with unif [0,1] distributions,
i.e.
8
< 1 05x41
fW ðxÞ ¼ fV ðxÞ ¼ ð19Þ
:
0 else
Next, we declare U ¼ VW and so the channel gain in dB, Y; is given by

Ar
Y ¼ 10 log10 U
D2 p
¼ 10 log10 ðUÞ þ Kmg
2
where Kmg ¼ 10 log10 ðAr =D pÞ:
Now, since [18]:
Z 1
fU ðuÞ ¼ fV;W ðv; u=vÞjvj1 dv ð20Þ
1
and for our situation, when u > v we have fV;W ðv; u=vÞ ¼ 0 and when u5v we have
fV;W ðv; u=vÞ ¼ 1; then
8R1
< u 1v dv ¼ lnðuÞ 05u41
fU ðuÞ ¼ ð21Þ
:
0 else
Hence,
d d
fY ðyÞ ¼ PðY4yÞ ¼ P U410ðyKmg Þ=10 ð22Þ
dy dy
ðyKmgÞ
Z 10 10
d
¼ ln u du ð23Þ
dy 0
8 ðKmg yÞ
>
ln10 2
< ðKmg yÞ10 10 ; y4Kmg
10
¼ ð24Þ
>
:
0 else
− 25
4x4 All Bounces
− 30 6x6 All Bounces
8x8 All Bounces
− 35 10x10 All Bounces
12x12 All Bounces
− 40 4x4 LOS Only
Channel Gain (dB)

6x6 LOS Only
− 45 8x8 LOS Only
10x10 LOS Only
− 50 12x12 LOS Only
Curve Fit
− 55
− 60
− 65
− 70
− 75
0 1 2 3 4 5 6 7 8 9 10
Distance (m)
Figure 3. The mean channel gain of LOS channels for various rooms fitted with ideal curve.
It turns out that the distribution of Y is a simple transformation of a gamma distribution. We

can write
Y ¼ Kmg Z ð25Þ
where Z is Gðl; tÞ distributed, defined as
1 t t1 lz
fZ ðzÞ ¼ lz e ð26Þ
GðtÞ
with l ¼ lnð10Þ=10 and t ¼ 2 and GðtÞ is the gamma function [18]. The mean and variance of Y
are given by
20
mY ¼ Kmg t=l ¼ Kmg ð27Þ
ln 10
and
200
s2Y ¼ t=l2 ¼ ð28Þ
ðln 10Þ2
We test the effectiveness of this distribution in predicting mean channel gain as a function of
distance, as shown in Figure 3. The line generated from (27) is an excellent fit for the LOS only
data. However, when the contribution of other bounces is included we see that the curve
underestimates the sample data by up to 7 dB:
4.3. Evaluating channel gain models

We now evaluate not only the mean channel gain but the distribution of channel gains at a
particular distance. In this regard, the modified gamma distribution is limited in that it has a
fixed variance, and hence a fixed shape. To address this issue we have looked at a number of
alternatives. The most promising of these is the modified Rayleigh distribution. We say that Y
has a modified Rayleigh distribution when

Y ¼ Kmr R ð29Þ
where R is a Rayleigh distribution given by
2
=2s2
fR ðrÞ ¼ sr2 er r50 ð30Þ
This distribution has two parameters that can be used for fitting: s and Kmr :
To evaluate the rooms for all D we use the w2 goodness test [19]. The w2 -test is used to test the
hypothesis that a function FðxÞ is the distribution of the sample population x1 ; . . . ; xn : If the
sample population deviates too much from FðxÞ then we shall reject the hypothesis. We will
perform the test on all groups of data for our distribution defined in (24). For comparison, the
modified Rayleigh distribution is also evaluated because of its similar shape and characteristics.
Both distributions require that only one parameter be estimated. We bin our data in 20 bins; this
is approximately 1 bin/dBW. According to w2 theory [20] there are K r 1 degrees of
freedom, where K is the number of bins and r is the number of unknown parameters. This
means we have 18 degrees of freedom.
The sample deviation is calculated according to:
XK
ðbj ej Þ2
w20 ¼ ð31Þ
j¼1
ej
where bj is the number of sample values in an interval Ij and

ej ¼ npj ð32Þ
where n is the number of samples in the population and pj is the probability that a random
variable X assumes any value in interval Ij :
Figure 4(a) shows the w2 -test results for the modified gamma distribution performed on only
the LOS component of LOS channel gain. We see that for all cases except the 12 12 m2 room
at D ¼ 0:2 m; we do not reject the hypothesis that the samples fit a modified gamma
103 10 3
α = 0.05
α = 0.005
Average − Gamma
Average − Rayleigh
Deviation
Deviation
10 2 10 2
α = 0.05
α = 0.005
4x4 Gamma
6x6 Gamma
8x8 Gamma
10x10 Gamma
12x12 Gamma 1
1 10
10
0 1 2 3 4 5 6 0 1 2 3 4 5 6
(a) Distance (m) (b) Distance (m)
Figure 4. (a) w2 -test results for a sample of rooms. For these rooms the modified gamma distribution was
fit to the LOS component of the LOS channel gain histograms. (b) The average w2 -test results for the
modified gamma and Rayleigh distributions fit to the LOS component of the LOS channel gain.
120 120
100 100
Number of Samples
Number of Samples
80 80
60 60
40 40
20 20
0 0
−100 −90 −80 −70 −60 −50 −40 −100 −90 −80 −70 −60 −50 −40
(a) Channel Gain (dB) (b) Channel Gain (dB)
Figure 5. Example distribution curves plotted atop histograms of LOS channel gain data, LOS path only.
The data shown represents an 8 8 m2 room and D ¼ 1:6 m: (a) modified gamma distribution;
and (b) modified Rayleigh distribution.
distribution. Only a sample of all the rooms evaluated is shown in Figure 4(a). For all other
rooms no additional points lie above the a ¼ 0:005 threshold. From this we can conclude that
our modified gamma distribution is a good estimate of LOS path channel gain.
Figure 4(b) shows w2 -test results, averaged over all rooms, for both the modified gamma and
Rayleigh distributions on the LOS component of LOS channels. As one can see the modified
Rayleigh distribution is no where near the deviation required for hypothesis acceptance. Other
distributions were attempted and yielded worse results than the modified Rayleigh.
Examples of our sample data plotted with the distribution curves can be seen in Figure 5.
Examining the modified gamma curve one notices how good the fit looks even with the naked
eye. The w2 deviation value for this curve is 25.66, which is less than 37.61 (deviation for
a ¼ 0:005). The value of Kmg for the modified gamma distribution for this particular example is
derived from the value of D; Kmg ¼ 10 log10 ð104 =ðD2 pÞÞ ¼ 49:05: This gives us a mean of
57:73 dB and variance of 37.72. The histogram for this case has a mean of 57:68 dB; and
variance of 40.07. Examining the modified Rayleigh distribution it is clear why this is not a good
fit. In order to match the mean and variance of the sample data the curve is too wide to be a
good fit. The w2 deviation value for this case using the modified Rayleigh distribution is 78.55,
higher than 37.61.
4.4. Models of LOS channel gain with diffuse components

Next we will attempt to fit the LOS channel including the diffuse components. The w2 -test results
of the modified Rayleigh distribution for a sample of all rooms evaluated can be seen in
Figure 6. For most rooms when D50:4 m the modified Rayleigh distribution passes the
hypothesis test.
Figure 7 shows the modified gamma and Rayleigh distributions evaluated for LOS channels
and averaged over all rooms. The modified Rayleigh fits the data best for D50:4 m; for reasons
discussed above. For D50:4 m the modified gamma distribution still holds because the LOS
103
α = 0.05
α = 0.005
4x4 Rayleigh
6x6 Rayleigh
8x8 Rayleigh
10x10 Rayleigh
102 12x12 Rayleigh
Deviation
101
0 1 2 3 4 5 6
Distance (m)
Figure 6. The deviation results for the w2 -test for the modified Rayleigh distribution. Results are shown for
select rooms for simplicity. Both the LOS and diffuse components are evaluated.
10 3
α = 0.05
α = 0.005
Average − Gamma
Deviation
10 2 Average − Rayleigh
10 1
0 1 2 3 4 5 6
Distance (m)
Figure 7. The deviation results for the w2 -test for the modified Rayleigh and gamma distributions. Results
are averaged over all rooms. Both the LOS and diffuse components are evaluated here.
component is dominant at close distances. The values of D50:4 m is also the range where (27)
fits the overall distribution of the mean LOS channel gain.
In Figure 7, for the modified Rayleigh curve, 77% of all samples lie below 40.0 (the deviation
threshold for a ¼ 0:005 and 20 degrees of freedom). When you count only those samples for
60 60
50 50
Number of Samples
Number of Samples
40 40
30 30
20 20
10 10
0 0
− 75 −70 − 65 − 60 − 55 − 50 −45 −75 −70 −65 −60 −55 −50 −45
(a) Channel Gain (dB) (b) Channel Gain (dB)
Figure 8. Example distribution curves plotted atop histograms of LOS channel gain data,
LOS and diffuse path, for an 8 8 m2 room and D ¼ 1:6 m: (a) modified gamma distribution;
and (b) modified Rayleigh distribution.
D50:4 m; this rises to 90% below the threshold. On average the larger rooms show poorer
results because the additional power is weaker than smaller rooms. For the modified gamma
distribution, 86% of samples lie below the threshold for D50:4 m:
Examples of the modified Rayleigh and gamma distribution curves plotted with the sample
data can be seen in Figure 8. It is clear that the modified Rayleigh distribution is the better
choice in this case. The shape of the distribution changes with the additional bounces. The width
of the histogram is smaller than seen in Figure 5. Our modified gamma distribution has a fixed
variance, unlike the modified Rayleigh distribution. This is why the modified Rayleigh
distribution is superior when the diffuse components are added. The modified gamma
distribution is too wide with additional bounces. The modified Rayleigh has a deviation of
23.46, the a ¼ 0:005 threshold is 40.00. The deviation for the modified gamma distribution is
186.81, which is much higher than the a ¼ 0:005 deviation threshold. The histogram has a mean
of 54:3 dB and a variance of 11.37 for this case. The modified Rayleigh distribution
parameters s2 and Kmr ; were chosen so that mean and variance match identically.
From these results we can conclude that the modified gamma distribution is the best
probabilistic model for small values of D50:4 m: When D50:4 m the modified Rayleigh
distribution will be used to model the LOS channel gain distribution.
4.5. Models for diffuse channels and RMS delay spread

Following a similar methodology to that used here, models for the channel gain of diffuse
channels and the rms delay spread of LOS and diffuse channels are developed in Reference [21].
In contrast to the model of LOS channel gain, we were unable to develop other successful
physical arguments based on channel structure to inform our remaining choices for statistical
models. Nevertheless, good empirical models were found, as summarized below:
* the shifted lognormal distribution was the best fit for the observed distributions of diffuse
channel gains (see Figure 9).
* the gamma distribution and the lognormal distribution are both excellent fits for LOS
delay spreads, with the lognormal being preferred for small D (see Figure 10).
* the gamma distribution was the best found for diffuse delay spreads, but no distribution
was found that passed the w2 fit test. However, the gamma distribution model was able to
predict the power penalty due to rms delay spread to within 1 dB (see Figure 11).
500
Lognormal
450 Gaussian
Data
400
350
Number of Samples
300
250
200
150
100
50
0
− 90 − 85 − 80 −75 −70 −65 −60 −55 −50
Channel Gain (dB)
Figure 9. Diffuse channel gain data for a 12 12 m2 room with D ¼ 6:0 m:
400
350
300
Number of Samples
250
200
150
100
50
0
0 2 4 6 8 10 12 14 16 18
Delay Spread (ns)
Figure 10. LOS delay spread data for an 8 8 m2 room with D ¼ 1:8 m:
The distribution is gamma with t ¼ 0:72:
350
300
250
Number of Samples
200
150
100
50
0
0 5 10 15 20 25 30 35
Delay Spread (ns)
Figure 11. Diffuse delay spread for a 12 12 m2 room at dtr ¼ 6:0 m: The distribution is gamma.
* varying the reflectivity of the walls did not affect the distribution fit tests described above.
* in modelled rooms with typical configurations of furniture, the same distributions model
the channel gain and rms delay spread as in empty rooms.
4.6. Generating a channel impulse response

In summary, the following method can be used to generate a statistically realistic impulse
response hðtÞ for an indoor optical wireless channel:
1. Determine the transmitter/receiver separation D and the room size.
2. Based on the FOV of the receiver, use (15) to determine if the channel is LOS or diffuse.
3. Choose the mean channel gain in dB from an appropriate curve fit, based on the room
size and D (see Figures 12 and 13).
4. Determine the actual channel gain by applying either the modified gamma, Rayleigh or
lognormal distribution.
5. Determine the mean rms delay spread and use that to choose an actual rms delay spread
from the appropriate distribution.
6. If an actual impulse response shape is required, one can use the previously determined
gain and rms delay spread to determine the impulse response [10]
6a6 Hð0Þ
hðtÞ ¼
uðtÞ ð33Þ
ðt þ aÞ7
pffiffiffiffiffiffiffiffiffiffiffiffiffi
where the delay spread Srms is given by a ¼ 12 11=13Srms and Hð0Þ is the channel gain
in linear units (i.e. Hð0Þ ¼ 10Y=10 ).
Details of the curve fit parameters can be found in Reference [21] and are available as data
files in Reference [22].
− 25
4x4 Room
− 30 6x6 Room
8x8 Room
− 35 10x10 Room
12x12 Room
− 40
Channel Gain (dB)

− 45
− 50
− 55
− 60
− 65
− 70
0 1 2 3 4 5 6 7 8 9 10
Distance (m)
Figure 12. Mean channel gain for LOS channels with curve fits.
− 52
4x4 Room
4x6 Room
− 54 4x8 Room
6x6 Room
− 56 6x8 Room
8x8 Room
− 58 8x10 Room
Channel Gain (dB)
8x12 Room
10x10 Room
− 60 10x12 Room
12x12 Room
− 62
− 64
− 66
− 68
− 70
0 1 2 3 4 5 6
Distance (m)
Figure 13. Mean channel gain for diffuse channel with curve fits.
5. CONCLUSIONS
We have outlined a method for statistical modelling for the indoor optical wireless channel
through the examination of the characteristics of a large set of channel impulse responses. The
channel responses are generated using an estimation method based on geometrical modelling of
indoor environments together with an iterative technique for calculating multiple reflections. We
confirm prior studies showing that channels with line-of-sight paths must be modelled separately
from those fully diffuse channels with no such path. We show that the distribution of the
channel gain in dB for the LOS component follows a modified gamma distribution, and the
channel gain in dB for LOS channels including all reflections follows a modified Rayleigh
distribution for most transmitter–receiver distances. Similar results are described for
distributions of channel gains of diffuse channels and for distributions of rms delay spreads.
Finally, we describe a method for generating a statistically realistic impulse response given any
transmitter–receiver separation in an indoor environment.
ACKNOWLEDGEMENTS
This work was supported by National Science Foundation under grant No. ECS-9876149.
REFERENCES
1. Kahn JM, Barry JR. Wireless infrared communications. Proceedings of the IEEE 1997; 85(2):265–298.
2. Heatley DJT, Wisely DR, Neild I, Cochrane P. Optical wireless: The story so far. IEEE Communications Magazine
1998; (12):72–82.
3. Infrared Data Association, http://www.irda.org
4. IEEE 80211 Standard Working Group. Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specifications. IEEE: New York, 1997.
5. Gfeller FR, Bapst UH. Wireless in-house data communication via diffuse infrared radiation. Proceedings of the
IEEE 1979; 67(11):1474–1486.
6. Kahn JM, Krause WJ, Carruthers JB. Experimental characterization of non-directed indoor infrared channels.
IEEE Transactions on Communications 1995; 43(2–4):1613–1623.
7. Hashemi H, Yun G, Kavehrad M, Behbahani F, Galko P. Frequency response measurements of the wireless indoor
channel at infrared optics. International Zurich Seminar on Digital Communications 1994; (3):273–284.
8. Hashemi H, Yun G, Kavehrad M, Behbahani F, Galko P. Indoor propagation measurements at infrared frequencies
for wireless local area networks applications. IEEE Transactions on Vehicular Technology 1994; 43(3):562–576.
9. Pakravan MR, Kavehrad M, Hashemi H. Indoor wireless infrared channel characterization by measurements. IEEE
Transactions on Vehicular Technology 2001; 50(4):1053–1073.
10. Carruthers JB, Kahn JM. Modeling of nondirected wireless infrared channels. IEEE Transactions on
Communications 1997; 45(10):1260–1268.
11. Jungnickel V, Pohl V, No. nnig S, von Helmolt C. A physical model of the wireless infrared communication channel.
IEEE Journal on Selected Areas in Communications 2002; 20(3):631–640.
12. Carruthers JB, Kannan P. Iterative site-based modeling for wireless infrared channels. IEEE Transactions on
Antennas and Propagation 2002; 50(5):759–765.
13. Carruthers JB, Carroll SM, Kannan P. Propagation modelling for indoor optical wireless communications using fast
multi-receiver channel estimation. IEE Proc-Optoelectronics 2003; 150(10):473–481.
14. Alqudah YA, Kavehrad M. MIMO characterization of indoor wireless optical link using a diffuse-transmission
configuration. IEEE Transactions on Communications 2003; 51(9):1553–1560.
15. Carruthers JB, Kannan P. IrSimIt. http://issbuedu/bwc/irsimit
16. Carruthers JB, Kahn JM. Angle diversity for nondirected wireless infrared communication. IEEE Transactions on
Communications 2000; 48(6):960–969.
17. Barry JR, Kahn JM, Krause WJ, Lee EA, Messerschmitt DG. Simulation of multipath impulse response for indoor
wireless optical channels. IEEE Journal on Selected Areas in Communications 1993; 11(3):367–379.
18. Grimmett GR, Stirzaker DR. Probability and Random Processes. Oxford University Press: New York, 1982.
19. Ayyub BM, McCuen RH. Probability, Statistics & Reliability for Engineers. CRC Press: Boca Raton, 1997.
20. Kreyszig E. Advanced Engineering Mathematics (6th edn). Wiley: New York, 1988.
21. Carroll SM. Modeling Wireless Infrared Communications for Network Simulation. Ph.D. Dissertation, Boston
University, 2004.
22. Carroll SM, Carruthers JB. http://iss.bu.edu/bwc/irmodel
AUTHORS’ BIOGRAPHIES
Jeffrey B. Carruthers received the BEng degree in computer systems engineering

from Carleton University, Ottawa, Canada, in 1990, and the MS and PhD degrees in
electrical engineering from the University of California, Berkeley in 1993 and 1997,
respectively. He has been with the Department of Electrical and Computer
Engineering at Boston University since 1997, and is currently an Associate
Professor. Dr Carruthers received the US National Science Foundation CAREER
Award in 1999. His research interests include wireless infrared communications,
wireless networking, and digital communications.
Sarah M. Carroll received a BS in Electrical Engineering from the University of

Notre Dame, Notre Dame, IN in 1999 and a PhD in Electrical Engineering from
Boston University, Boston, MA in 2003. She is currently an Advisory IT Specialist
with IBM Global Services, New York, NY. Dr Carroll received the Presidential
University Graduate Fellowship to attend Boston University and the James L.
Massey award whilst at Notre Dame. Her research interests include wireless infrared
communications, ad hoc networking and digital communications. She is a member of
IEEE Communications Society.

Static Al

Uploaded by

Copyright:

Available Formats

Static Al

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Static Al

Uploaded by

Copyright:

Available Formats

INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMS

Int. J. Commun. Syst. 2005; 18:267–284

Statistical impulse response models for indoor optical

J. B. Carruthers1,n,y and S. M. Carroll2

High-quality wireless access to information, networks, and computing resources by users of

Contract/grant sponsor: U.S. National Science Foundation; contract/grant number: ECS-9876149

Received 3 June 2004

2. CHANNEL ESTIMATION METHOD

We provide an outline of the impulse response calculation method developed in References

Excellent approximations can be obtained for M ranging from 3 to 10, as discussed in

Channel impulse response (1/s)

3.2. Channel gain

shown in (3), the LOS path hð0Þ 2

3.3. Multipath dispersion

where m; the mean delay, is calculated by

4.1. LOS path existence

4.2. Models of LOS path channel gain

pattern of the source at angle f is given by

Channel Gain (dB)

It turns out that the distribution of Y is a simple transformation of a gamma distribution. We

4.3. Evaluating channel gain models

has a modiﬁed Rayleigh distribution when

where bj is the number of sample values in an interval Ij and

4.4. Models of LOS channel gain with diﬀuse components

4.5. Models for diﬀuse channels and RMS delay spread

Figure 9. Diﬀuse channel gain data for a 12 12 m2 room with D ¼ 6:0 m:

4.6. Generating a channel impulse response

Channel Gain (dB)

Jeﬀrey B. Carruthers received the BEng degree in computer systems engineering

Sarah M. Carroll received a BS in Electrical Engineering from the University of

You might also like