Você está na página 1de 108

Performance Analysis of

Multipulse PPM on MIMO


Free-Space Optical Channels
A Thesis
Presented to
the Faculty of the School of Engineering and Applied Science
University of Virginia
In Partial Fulllment
of the Requirements for the Degree
Master of Science
Electrical Engineering
by
Michael L. Baedke
August 2004
Approval Sheet
This thesis is submitted in partial fulllment of the
requirements for the degree of
Master of Science Electrical Engineering
Author
This thesis has been read and approved by the examining Committee:
Dissertation advisor
Accepted for the School of Engineering and Applied Science:
Dean, School of Engineering and
Applied Science
August 2004
Acknowledgements
Over the course of the past few years, there have been many people who helped to make
this point in my education possible. In particular, I would like to thank my advisor, Professor
Wilson, whose dedication to learning and contagious passion for the eld has been an inspiration
to me. Thank you for taking a personal interest in sharing your knowledge and providing me
with the best education imaginable.
I would also like to thank my wife, Alison for her unfailing love and support. Thank you
for putting your dreams on hold this past year so I could follow mine, and thank you for always
believing in me and encouraging me to be my best.
Thanks also goes to my parents for always offering their nancial and emotional support
over the past few years. Being able to concentrate fully on our educational goals would have
been impossible without a safety net to rely on whenever we needed it.
Finally, I would like to thank Dr. Guess and Dr. Brandt-Pearce, who also happen to be on
my committee, for their time and assistance this past year. You each made time for me whenever
I needed your help, and Ive really enjoyed getting to know both of you.
i
Abstract
Free-space-optics (FSO) has emerged as a technology that has the potential to bridge the
last-mile gap that separates homes and businesses from high speed access to the Internet. In
FSOsystems, information is transmitted between two points by modulating a light source, much
like with traditional ber optic communication. However, FSO is a wireless technology in that
it operates via line-of-sight, transmitting the data through the air, potentially over distances on
the order of 1 km.
Its main advantages over other competing technologies are that it can provide high speed
access into the Gbps range, it operates in an unregulated frequency domain, and unlike ber
optic systems or other wired services, it avoids the need for trenching, which is much slower
and more costly.
The main challenge to FSO systems is the atmosphere itself. The systems must be designed
such that the potentially harsh atmospheric effects of turbulence and aerosol scattering can be
mitigated. Traditional solutions include increasing the link margin (supplying excess power to
overcome potentially deep fades), and keeping the link distances small. However, these solu-
tions are limited: Link margins become prohibitively difcult (often impossible) to maintain
during deep fades, and keeping link distances small is often impractical in real-world imple-
mentations.
In this thesis, we study a method to combat link fading based on the multiple-input, multiple-
output (MIMO) approach that has seen much success in the RF domain. By using multiple
transmitters and receivers spaced sufciently far from one another, we are able to create mul-
tiple, uncorrelated paths over which to send the data. The probability that all of the paths are
simultaneously faded is much lower than when only relying on a single path from transmitter
to receiver, as with a single-input, single-output (SISO) system.
For our analysis of the MIMO FSO system, we explore multiple pulse position modulation
ii
iii
(MPPM). This is a modulation technique where the duration of a signal is divided into Q slots,
and the laser array is pulsed simultaneously during w of them, creating
_
Q
w
_
possible patterns.
Traditional PPM is a special case of MPPM, where w = 1. We show that MPPM is superior to
PPM with respect to bandwidth efciency (and maximized when w =
Q
2
), and exhibits supe-
rior symbol error performance when the system is peak-power-limited. PPM exhibits superior
symbol error performance when the system is average-power-limited.
In this thesis, we develop the maximum likelihood detectors for the system operating in the
perfect photon counting (Poisson) and thermal-noise-limited (Gaussian) regimes. We demon-
strate that for non-faded channels, having multiple receivers improves symbol error perfor-
mance due to the increase in receiver aperture size. We also demonstrate that for faded chan-
nels, performance gains are seen for increases in the number of transmitters and receivers. Full
transmitter and receiver diversity is observable by analyzing the Rayleigh fading case.
We also analyze the information-theoretic channel capacity of the Poisson regime by look-
ing at the ergodic (average) capacity and the outage probability (probability of the instantaneous
capacity dropping below some set threshold). We see from these results that the maximum ca-
pacity is achieved at lower power levels for MIMO and SIMO systems over non-faded channels,
and for MIMO, SIMO, and MISO systems over faded channels. The outage probability curves
are steeper and show the benecial effect of transmitter and/or receiver diversity. Full diversity
is once again observable by looking at the outage probability curves in the Rayleigh fading case.
We conclude that MIMO system design is a technique that improves MPPM FSO sys-
tem performance under various fading cases, and that full transmitter and receiver diversity
is achievable in this system.
Contents
1 Introduction 1
1.1 What is free-space optics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The history of free-space optical communication . . . . . . . . . . . . . . . . 1
1.3 Uses of free-space optics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 Military communication . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.2 Satellite and deep-space communication . . . . . . . . . . . . . . . . . 3
1.3.3 The last-mile solution . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 System overview 6
2.1 Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Source and channel encoders . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Multiple pulse position modulator . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 The transmitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 The channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5.1 Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5.2 Aerosol scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5.3 Fading models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.6 The receiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6.1 p-i-n photodiodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6.2 Avalanche photodiodes . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6.3 Bandwidth and noise considerations in p-i-n and APD receivers . . . . 17
2.6.3.1 Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6.3.2 Shot noise . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6.3.3 Background noise . . . . . . . . . . . . . . . . . . . . . . . 18
iv
v
2.6.3.4 Thermal noise . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6.3.5 Excess APD noise . . . . . . . . . . . . . . . . . . . . . . . 18
2.7 MPPM demodulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.8 Source decoder, channel decoder, and retrieved information . . . . . . . . . . 19
3 MIMO applied to FSO systems using MPPM - background, system model and
denitions 20
3.1 Research on MIMO and MPPM FSO communications systems . . . . . . . . . 20
3.1.1 MIMO in wireless systems . . . . . . . . . . . . . . . . . . . . . . . . 20
3.1.1.1 Application of MIMO concepts to free-space optics . . . . . 22
3.1.2 Multiple pulse position modulation (MPPM) . . . . . . . . . . . . . . 22
3.1.3 Error probability - Gaussian vs. Poisson . . . . . . . . . . . . . . . . . 23
3.2 Power comparisons between PPM and MPPM . . . . . . . . . . . . . . . . . . 24
3.3 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.1 MPPM signaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.2 Transmitter array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3.3 Receiver array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3.4 Channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3.5 Detector and observable . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.5.1 Poisson Regime . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.5.2 Gaussian Regime . . . . . . . . . . . . . . . . . . . . . . . 29
4 Maximum likelihood (ML) detection 33
4.1 ML Detection for the poisson regime . . . . . . . . . . . . . . . . . . . . . . 33
4.1.1 Case 1: ML detection with no background and no fading . . . . . . . . 34
4.1.2 Case 2: ML detection with no background and fading . . . . . . . . . . 36
4.1.3 Case 3: ML detection with background radiation and no fading . . . . . 36
4.1.4 Case 4: Background radiation and fading . . . . . . . . . . . . . . . . 37
4.2 General ML detection in the Gaussian regime . . . . . . . . . . . . . . . . . . 38
4.2.1 Thermal noise dominates over shot noise . . . . . . . . . . . . . . . . 40
4.2.1.1 No fading present . . . . . . . . . . . . . . . . . . . . . . . 41
4.2.1.2 Fading present . . . . . . . . . . . . . . . . . . . . . . . . . 41
vi
4.2.2 Shot noise dominates over thermal noise . . . . . . . . . . . . . . . . . 41
4.2.2.1 Fading present . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.2.2 No fading present . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.3 Shot and thermal noise are not dominant . . . . . . . . . . . . . . . . . 42
4.2.3.1 Fading Present . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2.3.2 No fading present . . . . . . . . . . . . . . . . . . . . . . . 43
5 Error analysis of MIMO FSO system using MPPM 44
5.1 Error analysis in the Poisson regime . . . . . . . . . . . . . . . . . . . . . . . 44
5.1.1 No background radiation . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.1.1.1 No background, no fading . . . . . . . . . . . . . . . . . . . 46
5.1.1.2 No background, Rayleigh fading . . . . . . . . . . . . . . . 47
5.1.1.3 No background, log-normal fading . . . . . . . . . . . . . . 49
5.1.2 Error probability in the presence of background radiation . . . . . . . . 52
5.1.2.1 Background radiation, no fading . . . . . . . . . . . . . . . 52
5.1.2.2 Background radiation, Rayleigh fading . . . . . . . . . . . . 57
5.1.2.3 Background radiation, log-normal fading . . . . . . . . . . . 58
5.2 Error probability in the Gaussian regime . . . . . . . . . . . . . . . . . . . . . 58
5.2.0.4 No fading . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2.0.5 Rayleigh fading . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.0.6 Log-normal fading . . . . . . . . . . . . . . . . . . . . . . . 65
6 Capacity of the MIMO FSO system using MPPM in the Poisson regime 68
6.1 No background radiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.1.1 No background, no fading . . . . . . . . . . . . . . . . . . . . . . . . 70
6.1.2 No background, Rayleigh fading . . . . . . . . . . . . . . . . . . . . . 72
6.1.3 No background, log-normal fading . . . . . . . . . . . . . . . . . . . . 74
6.2 Background radiation and no fading . . . . . . . . . . . . . . . . . . . . . . . 77
6.2.1 Background radiation, no fading . . . . . . . . . . . . . . . . . . . . . 80
6.2.2 Background radiation, Rayleigh fading . . . . . . . . . . . . . . . . . 82
6.2.3 Background radiation, log-normal fading . . . . . . . . . . . . . . . . 85
vii
7 Conclusions 88
List of Figures
2.1 FSO system block diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Output light power vs. input drive current for all three most common light
sources [1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Probability density functions for Rayleigh and log-normal distributions. . . . . 14
2.4 A typical FSO transceiver [2]. . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Simplied detector circuit employing a p-i-n photodiode. . . . . . . . . . . . . 16
2.6 Simplied model of an APD and integrator. . . . . . . . . . . . . . . . . . . . 17
3.1 The MIMO concept. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Relative energy efciencies vs. Q for w {1, Q/2}. . . . . . . . . . . . . . 26
3.3 Norton equivalent noise model. . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 Poisson and Gaussian p.d.f.s with equal means and variances of = 200. . . . 31
3.5 Poisson and Gaussian p.d.f.s with equal means and variances of = 2. . . . . 32
5.1 Symbol error probability vs. average power with no fading and no background
radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Symbol error probability vs. peak power with no fading and no background
radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.3 Symbol error probability vs. average power with Rayleigh fading, no back-
ground radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.4 Symbol error probability vs. peak power with Rayleigh fading, no background
radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.5 Symbol error probability vs. average power with log-normal fading (S.I. = 1.0),
no background radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . 51
viii
ix
5.6 Symbol error probability vs. peak power with log-normal fading (S.I. = 1.0), no
background radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . 51
5.7 An example of a denite error. . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.8 An example of an indenite error (where the receiver chooses incorrectly from
the 3 possible modulator symbols). . . . . . . . . . . . . . . . . . . . . . . . 54
5.9 Symbol error probability vs. average power with no fading, P
b
T
b
= 170 dbJ,
and Q = 8 dashed and solid lines overlap. . . . . . . . . . . . . . . . . . . . 56
5.10 Symbol error probability vs. peak power with no fading, P
b
T
b
= 170 dbJ, and
Q = 8 dashed and solid lines overlap. . . . . . . . . . . . . . . . . . . . . . 56
5.11 Symbol error probability vs. average power with Rayleigh fading, P
b
T
b
=
170 dbJ, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.12 Symbol error probability vs. peak power with Rayleigh fading, P
b
T
b
= 170
dbJ, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.13 Symbol error probability vs. average power with Log-normal fading, S.I. = 1.0,
P
b
T
b
= 170 dbJ, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.14 Symbol error probability vs. peak power with Log-normal fading, S.I. = 1.0,
P
b
T
b
= 170 dbJ, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.15 Symbol error probability vs. average power with no fading, Q = 8, R = 100
, T
0
= 290 K, P
b
T
b
= 170 dbJ, and R
b
= 100 Mbps. . . . . . . . . . . . . 63
5.16 Symbol error probability vs. peak power with no fading, Q = 8, R = 100 ,
T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps. . . . . . . . . . . . . . . . 64
5.17 Symbol error probability vs. average power with Rayleigh fading, Q = 8,
R = 100 , T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps. . . . . . . . . 64
5.18 Symbol error probability vs. peak power with Rayleigh fading, Q = 8, R = 100
, T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps. . . . . . . . . . . . . . 65
5.19 Symbol error probability vs. signal power with Rayleigh fading using equal
gain combining (EGC) or optimal gain combining (OGC). Q = 8, R = 100 ,
T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps. . . . . . . . . . . . . . . . 66
5.20 Symbol error probability vs. average power with log-normal fading, S.I. = 1.0,
Q = 8, R = 100 , T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps. . . . . 66
x
5.21 Symbol error probability vs. peak power with log-normal fading, S.I. = 1.0,
Q = 8, R = 100 , T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps. . . . . 67
5.22 Symbol error probability vs. signal power with log-normal fading using equal
gain combining (EGC) or optimal gain combining (OGC). Q = 8, R = 100 ,
T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps. . . . . . . . . . . . . . . . 67
6.1 Ergodic capacity vs. average power with no fading, no background radiation,
and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.2 Ergodic capacity vs. peak power with no fading, no background radiation, and
Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.3 50% outage probability vs. average power with no fading, no background radi-
ation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.4 50% outage probability vs. peak power with Rayleigh fading, no background
radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.5 Ergodic capacity vs. average power with Rayleigh fading, no background radi-
ation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.6 Ergodic capacity vs. peak power with Rayleigh fading, no background radia-
tion, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.7 50% outage probability vs. average power with Rayleigh fading, no background
radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.8 50% outage probability vs. peak power with Rayleigh fading, no background
radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.9 Ergodic capacity vs. average power with log-normal fading, no background
radiation, Q = 8, and S.I. = 1.0. . . . . . . . . . . . . . . . . . . . . . . . . 76
6.10 Ergodic capacity vs. peak power with log-normal fading, no background radia-
tion, Q = 8, and S.I. = 1.0. . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.11 50% outage probability vs. average power with log-normal fading, no back-
ground radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.12 50% outage probability vs. peak power with log-normal fading, no background
radiation, and Q = 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.13 Ergodic capacity vs. average power with no fading, Q = 8, and P
b
T = 170
dBJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
xi
6.14 Ergodic capacity vs. peak power with no fading, Q = 8, and P
b
T = 170 dBJ. 81
6.15 50% outage probability vs. average power with no fading, Q = 8, and P
b
T =
170 dbJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.16 50% outage probability vs. peak power with no fading, Q = 8, and P
b
T =
170 dBJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.17 Ergodic capacity vs. average power with Rayleigh fading, Q = 8, and P
b
T =
170 dBJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.18 Ergodic capacity vs. peak power with Rayleigh fading, Q = 8, and P
b
T =
170 dBJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.19 50% outage probability vs. average power with Rayleigh fading, Q = 8, and
P
b
T = 170 dbJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.20 50% outage probability vs. peak power with Rayleigh fading, Q = 8, and
P
b
T = 170 dBJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.21 Ergodic capacity vs. average power with log-normal fading, Q = 8, and P
b
T =
170 dBJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.22 Ergodic capacity vs. peak power with log-normal fading, Q = 8, and P
b
T =
170 dBJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.23 50% outage probability vs. average power with log-normal fading, Q = 8, and
P
b
T = 170 dbJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.24 50% outage probability vs. peak power with log-normal fading, Q = 8, and
P
b
T = 170 dBJ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
xii
List of Symbols
a Substitution for a
nm
in integration.
A M N matrix representing the path gains from each receiver to each
transmitter.
a
nm
One realization of the individual path gain from transmitter m to receiver
n; this quantity is squared to nd power gain.
b, g, i, j, k, l Generic variables used to index summations and for other tasks.
C Information theoretic capacity of a channel, measured in bits per channel
use.
C
d
Junction capacitance of a photodetector.
d
0
Correlation distance, measured in meters.
E
s
Energy per modulator symbol contributed by the entire transmit array to
one receiver element.
f Frequency of transmitted signal. For a 1550 nm laser diode, this would be
approximately 2 10
14
Hz.
F(M) Excess noise factor in an avalanche photodiode.
G Average gain of an avalanche photodiode.
h Plancks constant, approximately equal to 6.6 10
34
.
H(X) Entropy of X, measured in bits.
L Link length, measured in meters.
m Denotes a single transmitter. m {1, 2, ..., M}
M Number of transmitters in the system.
n Denotes a single receiver. n {1, 2, ..., N}
N Number of receivers in the system.
P Peak on power observed at the receiver.
P
ave
On power observed at the receiver, averaged over the duration of the
modulator symbol.
P
b
Background power observed at the receiver.
P
def
Probability of a denite error; occurs when one or more off slots have a
higher count rate than on slots.
xiii
P
indef
Probability of an indenite error; occurs when one or more off slots
have a count rate equal to one or more on slots, with no off slots having
a count rate greater than any on slot.
P
peak
Also used to denote peak on power at the receiver.
Q Number of slots per MPPM symbol.
Q
i
on
The set of w on slots for a MPPM symbol.
Q
i
off
The set of Qw off slots for a MPPM symbol.
R
b
Bit rate of system.
S.I. Scintillation index. A measure of strength of fading in the log-normal
fading model.
T Duration of one slot of a pulse position modulation symbol, measured in
seconds.
T
b
Duration of a single bit. Inverse is the bit rate.
T
s
Duration of the entire pulse position modulation symbol, measured in
seconds.
w Number of on slots per MPPM symbol.

X The estimate of the receiver; the modulator symbol chosen by the receiver
as the most probable symbol received.
Z A N Q matrix representing the received observations by all N detectors
over all Q slots.
Z
nq
Observable at output of integrator at receiver n at slot q.
Probability of a deep fade detrimentally affecting a link path from
transmitter m to detector n.

n
Dened as (
on,n
+
off
)/
off
Quantum efciency of detector; the average number of electron-hole pairs
generated per incident photon
Wavelength of the transmitted signal; we have used 1550 nm for our
analysis.

on
Poisson parameter denoting average number of photoelectrons observed
during an on slot, summed over all N photodetectors.

on,n
Poisson parameter denoting average number of photoelectrons observed
during an on slot at a single detector n.
xiv

off
Poisson parameter denoting average number of photoelectrons
observed during an off slot at a single detector n.

on
Used for the Gaussian regime, as the average number of photoelectrons
observed during an on slot, summed over all N photodetectors.

on,n
Used for the Gaussian regime, as the average number of photoelectrons
observed during an on slot at detector n.

off
Used for the Gaussian regime, as the average number of photoelectrons
observed during an off slot at detector n.

2
X
Denotes the mean of the log-normal fading variable.

2
Used for the Gaussian regime, when thermal noise dominates, as the
variance of the number of photoelectrons observed during an on or
off slot.

2
on
Used for the Gaussian regime, when thermal noise is not dominant, as
the variance of the number of photoelectrons observed during an
on slot.

2
off
Used for the Gaussian regime, when thermal noise is not dominant, as
the variance of the number of photoelectrons observed during an
off slot.

2
X
Denotes the variance of the log-normal fading variable.

0
Correlation time, measured in seconds.

pdf
(k, ) Probability that a Poisson random variable with parameter equals k.

cdf
(k, ) Probability that a Poisson random variable with parameter is less than or
equal to k.
Chapter 1
Introduction
1.1 What is free-space optics?
Free-space optics is a method of communication that involves using a light source, usually
a laser, to transmit information through space or the atmosphere to a receiver. It is similar to
traditional ber-optics communication, in that it uses light to communicate information, but
the difference is the medium through which the information travels. Fiber-optics, as the name
implies, uses a ber to carry the light wave from transmitter to receiver. Free-space optics,
however, relies on a line-of-sight approach. As a result, the medium could be a near-vacuum,
as it is for satellite-to-satellite communication, or it could be the atmosphere, which includes
atmospheric and other natural obstructions that come into the path of the light.
1.2 The history of free-space optical communication
Although free-space optical (FSO) communication as we know it originated in the 1970s
[3], the history of optical communication through free-space using light really began when
signal res were used to send messages across long distances. Paul Reveres lanterns, as well
as manually operated lanterns on ships are examples of early FSO communication systems [4].
The rst major technological advancement in FSO communication, that moved us past the
ages of signal res, happened in 1880, when Alexander Graham Bell invented the photophone.
The device could send intensity-modulated sunlight over a distance of a few hundred feet. How-
1
2
ever, there were no major advancements in FSO communication for nearly a century, until the
laser was invented in the 1960s [4].
Within a few months of the invention of the laser, there was interest in using the technology
to communicate through the atmosphere. Bell Labs engineers brought an early ruby laser to the
top of a microwave tower at Murray Hill, NJ, and pointed it at a large screen 40 km away. A
colleague watched the screen for signs of the red pulses coming through, but not many pulses
made it. Although the results were poor, the engineers were able to make a spot as large as a
dining table glow like a replace at a distance of 2-6 km [5].
FSO communication, which also sometimes goes by the names optical wireless communi-
cation, lasercom (patented by McDonnell Douglas in the 1980s), wireless optical commu-
nication, and ber-free optical communication, really started to take off in the 1980s. Fund-
ing was increased in the United States and Europe as governments tried to plan for next gen-
eration communication technologies for air-to-air, satellite-to-submarine, air-to-satellite, and
satellite-to-satellite links [6].
There was also an interest for non-military communication, as well, but after the rst few
trials of FSO communication, interest began to decline. One reason is that communicating
through optical bers seemed so far superior to communicating through the atmosphere [5].
A lack of reliable components was also always an issue with early optical systems. Almost
every single component had to be developed, including laser sources, detectors, high-speed
electronics, and high accuracy pointing components to name a few. Because of these technical
difculties, as well as nancial and political issues as well, one-by-one, almost all early FSO
communications programs met with an early demise [6].
However, in the past few years, mainly due to a need to solve the last-mile problem,
an increased interest has been seen in non-military uses of FSO communication [5]. In fact,
Acampora cites a study that predicts that the FSO communications industry could grow from
an annual 120 million dollars in 2000 to more than 2 billion dollars annually by 2006 [3].
3
1.3 Uses of free-space optics
1.3.1 Military communication
Free-space optics is an attractive method of communication for military applications. The
reason is security. Using traditional radio frequency (RF) communications methods, eaves-
dropping on a conversation is much easier than with free-space optics, since the RF waves are
transmitted over a large area. This makes it possible to receive the signal while in the vicinity
of the system, although it is still necessary to demodulate and decode it. A FSO link, on the
other hand, has a very narrow beam divergence [4], typically milliradians, so the only way to
intercept the signal is to be in the path of transmission.
1.3.2 Satellite and deep-space communication
Free-space optics has several advantages that make it well suited to satellite communication.
First, it can provide high data-rate communication links between satellites at geosynchronous
distances and beyond. Second, it has several attributes that are superior to traditional RF com-
munication methods.
Traditionally, satellite communication has been accomplished using microwaves, but these
systems are bulky and expensive [4]. Free-space optics has the advantage of being much smaller
and far more inexpensive, which is an enormous asset to any space vehicle. For this reason,
NASA is developing a deep-space optical communication transceiver in its X2000 program,
also known as the Advanced Deep-Space Systems Development Program [7]. Early in the
X2000 program, NASA plans to support tens of kilobits per second of data from the Mars range
[8]. It also plans on building the rst of two 10 m class ground receiving telescopes by 2008
[7].
The International Space Station (ISS) Engineering Research and Technology Development
program is sponsoring the development of a high data rate FSO transmitter from the low-earth-
orbit range (on board the ISS) [7], that is projected to be able to support a data rate of 2.5 Gbps.
FSO communications appears to be the technology that will meet the needs of future space
ventures, including near-earth, solar-system, and interstellar missions [8].
4
1.3.3 The last-mile solution
Probably one of the most compelling and timely uses of free-space optics is to provide a
solution to the last-mile problem (or rst-mile problem, [9] depending on your perspective).
The multibillion-dollar optical ber backbone that was built to provide high-speed broadband
access to ofces and homes has come up less than one mile short for 9 of 10 US businesses
with 100 or fewer workers. As a result, only 2-5% of the ber network is actually being used
today [3]. Most businesses and homes are currently connected to the ber backbone using
traditional copper wires, which do not possess the gigabit-per-second capacity required to carry
bandwidth-intensive applications [3].
Laying ber optic cable to each home and business that needs broadband access would be
the ideal solution to this problem, but it is slow and expensive. The process can take 6-12
months, and can cost anywhere from $100,000 to $500,000 per mile[3], with up to 85% of the
cost due to trenching and installation [2]. Free-space optics, on the other hand, can be up and
running in a few days, and costs 1/3 to 1/10 of the cost of a ber installation [3]. Trenching
also causes trafc jams, displaces trees, and can destroy historical areas. For these reasons,
Washington D.C. is considering a moratorium on ber trenching [2].
FSO communication is thought by many experts to have the best chance at succeeding over
other ber-free technologies (like DSL, microwave radio, etc.) at bridging the last-mile gap
[3]. In addition to the cost and speed benets, it has a greater potential because it operates
in an unlicensed band (which is an enormous cost benet as well), its scalable (unlike RF
networks [9]), and it can be set up in a mesh conguration to carry full duplex gigabit-per-
second communications around a city town or region [3].
One challenge that faces FSO communications systems is building sway. Because of the
very narrow beamwidths possible from the lasers, very small changes in the position of either
the transmitter or receiver can cause the laser beam to miss its target. The two possible solutions
to this problemare to increase the beamdivergence at the receiver (which also reduces the power
density) or, for especially tall buildings or narrow beamwidths, to use an active tracking system;
mirrors continually adjust to keep the beam centered on the target.
The biggest challenge facing free-space optics in terrestrial applications, which is also a
major focus of this thesis, is overcoming limitations caused by the atmosphere. Bad weather,
especially thick fog, can severely attenuate the signal before it reaches the receiver [3]. In fact,
5
weather is the reason links in non-desert regions are often kept to 200-500m to ensure carrier-
class availability (99.999% availability) [2].
In this thesis we analyze an approach called spatial diversity which attempts to overcome
these atmospheric difculties. This approach employs multiple transmitters and/or receivers
simultaneously sending and/or receiving the information. The idea is to keep the transmitters
and receivers sufciently far from one another (which is a surprisingly small distance, as we
will see), such that all of the individual paths from transmitters to receivers would have to be
simultaneously faded (a much lower probability event) in order to degrade system performance.
Chapter 2
System overview
In this section, we will describe the free-space optical link from beginning to end. The
physical devices, channel, and noise and other disturbances will be considered, however in this
section we are most concerned with the hardware used to construct such a system, and a better
understanding of the physical phenomena that affect both the hardware and the channel to make
them non-ideal.
An overall block diagram depicting the FSO system is shown below in Figure 2.1.
Figure 2.1: FSO system block diagram.
6
7
2.1 Information
For the purpose of this paper, we are going to assume that the information to be transmitted
is already in binary format. This could be any kind of data including, but not limited to, Internet
and intranet trafc, multimedia applications (including streaming audio and video), and le
transfers or data exchanges of any kind.
2.2 Source and channel encoders
From an information-theoretic point of view, the raw information described in Section 2.1
contains natural redundancies that can be removed to make the system more efcient. This is
known as compression or source coding. The channel encoder then adds intelligent redun-
dancies back into the stream of data that make the system more resistant to errors. More detail
on these topics can be found in texts by Wilson [10] and Cover and Thomas [11]. Upper limits
on the performance of channel coding will be investigated in Chapter 6.
2.3 Multiple pulse position modulator
For this system, we are focusing on multiple pulse position modulation (MPPM), which is
a intensity modulation technique. This is, of course, not the only possible way to build this
system. Since a laser is, in effect, an optical oscillator [5], any modulation that is possible
with RF communication is also possible with optical communication (including coherent mod-
ulation/demodulation techniques). The drawback, however, comes from the fact that coherent
techniques like phase or frequency modulation are far more complex and expensive to build
[12].
The analysis and a detailed description of MPPM will be completed in Section 3.1.2. Until
then, it sufces to say that the modulator can take a certain number of bits (from the channel
encoder), and map them to a single MPPM symbol that will be sent across the channel by the
transmitter.
8
2.4 The transmitter
In Figure 2.1, the transmitter and modulator are depicted as being separate entities, but there
are actually different ways to construct this. It is possible for the transmitter to be constantly on,
and then be modulated as it is passed on to the channel, or the laser can be directly modulated
in one step.
Here we consider the transmitter to be the light source, which simply has the task of sending
light over the channel.
There are three different types of light sources that are commonly used in free-space optics:
Light Emitting Diode (LED): LEDs can produce light in the 800-900 nm band, they are
cheap, and they can produce radiation with low current drive levels. However, they have
limited output powers (1-10 mW), there is more frequency spreading than the other light
sources, and the light tends to be incoherent and unfocused [1].
Laser: Lasers have power outputs of 0.1-1 W, but are much bulkier than LEDs. The
laser is an optical cavity lled with light amplication material and mirrored facets at
each end. When the cavity lases, an initiated optical eld crosses back and forth in a
self-sustaining reaction. A small aperture in one of the mirrored facets allows some of the
energy to escape as radiated light. In the linear range of operation (see Figure 2.2), lasers
are unstable, so they are usually operated as continuous-wave devices at peak power [1].
Laser Diode: Like LEDs, laser diodes are semiconductor junction devices [1], but they
operate more like lasers with reecting etched substrates which act like small reectors
(like the reectors in the laser). Laser diodes are small, rugged, and very power-efcient.
They require more drive current than LEDs, but also generate more power. A laser diode
produces about a hundred milliwatts of useable optical power [4] with a more focused
beam than with LEDs [13].
All of the three light sources have the same output power characteristics, shown in Figure
2.2. From this, one can see a distinct linear region of operation, where an increase in input
current would result in a proportional increase in output light power.
The wavelength chosen for FSO systems usually falls near one of two wavelengths, 850 nm
or 1550 nm. The shorter of the two wavelengths is cheaper and is favored for shorter distances.
9
Figure 2.2: Output light power vs. input drive current for all three most common light sources
[1].
The 1550 nm light source is favored for longer distances since it has an allowed power that
is two orders of magnitude higher than at 850 nm [2]. These power limits are determined by
the American National Standards Institute (ANSI) Z136.1 Safety Standard [9]. The reason for
the higher allowed power is that laser-tissue interaction is very dependent on wavelength. The
cornea and lens are transparent to visible wavelengths (such as 850 nm) so the power can reach
the retina at the back of the eye. At 1550 nm retinal absorption is much lower, since the power
is absorbed mostly by the lens and cornea before it can reach the retina. The power at 1550 nm
is not unlimited, however, since it can still cause photo-keratitis and cataracts at higher levels
[14].
The 1550 nm wavelength is also preferred since more photons per watt of power arrive for
longer wavelengths, and therefore more photocurrent is produced per watt of incident power for
equal efciency devices[4].
For the remainder of this thesis, we will assume the laser to be an ideal, innite bandwidth
light source. For the simulations in Chapters 5, 6, and 5.2 we will make use of the 1550 nm
wavelength.
2.5 The channel
In a terrestrial free-space optical link, the channel is simply the atmosphere plus any other
disturbances through which the optical signal will pass. This is a very important component of
10
our system, since the channel is often the limiting factor for how long the link can be.
The atmospheric channel is uncontrolled in that the designers have no way of preventing
obstructions and other disturbances from coming between the transmitter and receiver. The
engineer will attempt put the system in a location where it is unlikely for obstructions to occur,
but it is always possible for a bird, for example, to temporarily pass through the beam. However,
in a packet-switched network, short duration interruptions are easily handled by retransmitting
the data [9].
A more serious threat is the atmosphere itself. Zhu and Kahn classify atmospheric effects on
the FSO channel into two categories, atmospheric turbulence and aerosol scattering [12]. These
are discussed further below.
2.5.1 Turbulence
Atmospheric turbulence is also known as scintillation. Even on a clear day, there are con-
tinual variations in the intensity of the light at the receiver due to inhomogeneities in the tem-
perature and pressure of the atmosphere. The Kolmogorov turbulence model is often used to
describe atmospheric turbulence [12, 15] and predicts that changes in the air temperature as
small as 1 degree Kelvin can cause refractive index changes as large as several parts per mil-
lion [15]. These pockets of air with different refractive indices, or eddies, act like time-varying
prisms [2] whose size ranges from a few millimeters to a few meters [12], and whose time scale
is related to wind speed [16] among other things.
These eddies cause the light to diffract along the path to the receiver in a time-varying
manner, affecting the intensity of the light. This phenomenon is visible to the naked eye by
watching the stars twinkle at night, or by watching the horizon shimmer on a hot day [2].
The effect of scintillation on a FSO communications link can be a wandering beam when
the eddies are bigger than the beam diameter and move the beam completely off target [2],
uctuating power at the receiver [9], and changes in the phase of the received light wave [12].
For weak turbulence, the intensity of the received signal is a random variable best approximated
by a log-normal distribution [13, 16]. This model is described further in Section 2.5.3.
To describe turbulence-induced fading, we do so using parameters in the spatial and tem-
poral domains. The rst useful parameter is the correlation length, which we call d
0
. This is
simply the distance for which the intensity of a light wave at two points in the atmosphere is
11
essentially uncorrelated. This distance can be approximated by d
0

L, where is the wave-


length of the transmitted wave, and L is the length of the FSO link [13]. This approximation
is valid for most FSO communication systems using visible or infrared lasers for link lengths
ranging from a few hundred meters to a few kilometers [12], and is approximately 1-10 cm for
most terrestrial links [13]. The importance of correlation distance will become evident as we
talk about spatial diversity as a method of mitigating the effect of turbulence on FSO links.
The second useful parameter is the correlation time, which we call
0
. When observing a
single point in the atmosphere at two different times,
0
represents the amount of time between
observations for which the atmospheric parameters are uncorrelated. The time scale for scintil-
lation is about the time it takes a volume of air the size of the beam to move across the path, and
is therefore related to wind speed [16]. Typical values for terrestrial links are 1-10 ms [13].
Correlation time is important to our discussion in order to justify spatial diversity as a
method of mitigating block fading. At the transmission rates desirable for a FSO system
(2.5 Gbps for example), a deep fade that could last 1-10 ms could potentially affect 2.5 to
25 megabits of data. The normal approach to counteract block fades is to interleave the data
before coding, but this is an unattractive solution due to the enormous size of the interleaver that
would be necessary to be effective [15]. Spatial diversity, which will be discussed in Section
3.1.1 is a method that avoids the need for such large interleavers.
2.5.2 Aerosol scattering
The most detrimental atmospheric phenomenon that affects FSO links is fog, which is clas-
sied as aerosol scattering. According to Acampora, susceptibility to fog has slowed the com-
mercial development of free-space optics, since it so severely limits the range of a FSO link
[3].
The exact amount of signal attenuation caused by fog varies with its density. Acampora
states that the link might lose 90% of its power for every 50 meters in moderately dense fog
[3]. This translates into a loss of 200 dB/km. Other sources give ranges in attenuation from 16
dB/km in light fog [9] to 300 dB/km in dense fog [17].
There are various ways to combat link fade due to fog. One such method is to simply
increase the power, also known as increasing the link margin. The link margin is simply extra
transmit power that is in excess of what is normally needed to communicate. The only problem
12
with increasing the link margin is that power levels are limited for any system, both because
of eye safety as well as practical limitations in the system itself. For moderately dense fog,
increasing the link power by a large amount, 60 dB (a factor of one million) for example, would
still only allow for an extra 300 m in link length.
Fading can also be mitigated by making the link length as small as possible. Longer links
can be accommodated by arranging the transmitters and receivers in a mesh-topology. In an
urban setting, the mesh could jump from building-to-building or house-to-house, so that the
signal propagates only over shorter distances and has multiple paths to reach any point in the
network [3].
The wavelength of the link also affects the links susceptibility to fog. Future FSO systems
will most likely take advantage of the long wavelength infrared range (LWIR) spectrum(8m <
< 14m), also known as the night-vision spectrum. LWIR systems are called all-weather
systems because they are 10-20 times less sensitive to fog, rain, smog, and other atmospheric
disturbances. These wavelengths are also far less dangerous to eye safety so allowable power
levels are higher than those for the 0.7-1.55 m range. [18].
Point-to-point microwave radio is an alternative to free-space optics that is immune to fog.
However, this technology requires spectrum licensing, which is a major disadvantage when
compared to FSO systems [3].
Effectively overcoming challenges imposed by foggy weather for any particular FSO link
would most likely involve a combination of the aforementioned solutions. Using spatial diver-
sity to combat atmospheric effects, which is the focus of this thesis, can be incorporated into
almost any well-designed system that also uses link margin, a mesh topology, and LWIR lasers.
The effect of also using spatial diversity is to introduce yet one more weapon in the arsenal of
the communications engineer to combat link fading.
2.5.3 Fading models
There are two widely used models for fading. One is the log-normal distribution, the other
is the Rayleigh distribution. In all of the fading cases, we keep the expected path gain E[A
2
]
equal to one, in order to make fair comparisons between them and the non-fading case.
The log-normal distribution is very often used in the literature to describe atmospheric tur-
bulence as experienced in FSO systems. In the log-normal distribution, the amplitude of the
13
path gain is a random variable A where A = e
X
and X is normal with mean
X
and vari-
ance
2
X
. By denition, the logarithm of A follows a normal distribution. The optical intensity
I = A
2
is also log-normally distributed.
The p.d.f. for A is
f
A
(a) =
1
(2
2
X
)
1/2
a
exp((log
e
a
X
)
2
/2
2
X
), a > 0 (2.1)
In order to keep the mean path intensity unity, i.e. E[A
2
] = 1, it can be shown that
X
=

2
X
. For the log-normal distribution, we make use of a parameter called the scintillation index,
dened as
S.I. =
E[A
4
]
E
2
[A
2
]
1 (2.2)
This quantity is proportional to the degree of fading, as seen in Figure 2.3, and can be related
to the variance
2
X
by S.I. = e
4
2
X
1. Typical values appearing in the literature for S.I. are in
the range 0.4-1.0.
They Rayleigh distribution is used less often in the literature than log-normal fading to
analyze FSO systems, but has some nice properties that make it an attractive model to use. First
of all, the Rayleigh fading case exhibits deeper fading than log-normal fading because of the
higher concentration of low-amplitude path amplitudes (see Figure 2.3). Second, with Rayleigh
fading, the diversity order of the MIMO system becomes apparent when analyzing the slopes
of the curves for symbol error probability.
In Rayleigh fading, the amplitude of the path gain follows a Rayleigh distribution. The
wavelength of the light is modeled to be large compared with the size of the scatterer, such that
the composite eld is produced by a large number of non-dominating scatterers, each contribut-
ing random optical phase upon arrival at the detector. The central limit theorem then gives a
complex Gaussian eld, whose amplitude is Rayleigh:
f
A
(a) = 2ae
a
2
, a > 0 (2.3)
where we have normalized so that E[A
2
] = 1. The random intensity I = A
2
is a one-sided
exponential random variable, whose density function is heavily concentrated at low (deeply
faded) values. The scintillation index for the Rayleigh situation is 1, though the distribution
is quite different from the log-normal case, especially in the small-amplitude tail, as shown in
Figure 2.3 [19].
14
0 0.5 1 1.5 2 2.5 3 3.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
a
f
A
(
a
)
Lognormal fading, S.I. = 0.4
Lognormal fading, S.I. = 0.6
Lognormal fading, S.I. = 1.0
Rayleigh fading
Figure 2.3: Probability density functions for Rayleigh and log-normal distributions.
We will make use of both of these fading models in Chapters 5 and 6 when we discuss the
effect of fading on error probability and channel capacity.
2.6 The receiver
Once the transmitted signal passes through the atmosphere, it must be collected and mea-
sured by the receiver. As we mentioned before, both coherent and noncoherent detection
schemes are possible, but for complexity and cost reasons noncoherent (or direct detection)
is preferred. For this thesis, we will only consider noncoherent systems.
In noncoherent optical signal detection, the detectors rely on the photoelectric effect inci-
dent photons are absorbed by the detector and free-carriers are generated and can be measured.
This is a probabilistic phenomenon, since it is possible for a photon to pass through the pho-
todetector without generating any free-carriers. However, in a well-designed photodetector, the
probability of an incident photon causing a free-carrier is high [4].
There are two models that we will use in our analysis of the system. In the ideal photon
counting model, we assume no thermal noise is present, and the system is capable of counting
current blips that occur as each photoelectron is produced. Integrating the photocurrent over
15
a certain period of time (called a slot) is equivalent to counting the current blips.
In the Gaussian model, we assume that zero mean, additive white Gaussian noise (AWGN)
is added to the generated photocurrent. We still integrate over a slot, and the integration process
should, on average, remove the noise power from the observable.
Figure 2.4 shows what a FSO transceiver (receiver and transmitter in one unit) might look
like, showing other components also present in many FSO systems.
Figure 2.4: A typical FSO transceiver [2].
According to Alexander, photodetectors fall into one of four categories: photomultipliers,
photoconductors, photodiodes, and avalanche photodiodes [4]. There are numerous congura-
tions and variations in these four categories, so we will concentrate on the two most popular
detectors in optical receivers for communication, p-i-n photodiodes and avalanche photodiodes.
2.6.1 p-i-n photodiodes
A p-i-n photodiode is made up of a p-type, and an n-type layer of semiconductor, separated
by an intrinsic layer (hence the name p-i-n photodiode). The p-type layer is made to be very thin,
so incident photons can pass directly through to the intrinsic region where they can generate
electron-hole pairs. Any pairs that are generated are quickly swept into the p- and n-type layers
where they contribute to the photocurrent.
16
The p-i-n photodiode has a quantum efciency associated with it that depends on the re-
ectivity of the p-type layer, the absorption length of the intrinsic region, and the length of
the depletion region [4]. The quantum efciency is often denoted by and is a measure of
the average number of electron-hole pairs generated per incident photon. In a practical p-i-n
photodiode, ranges from 0.3 to 0.95 [20].
A simplied model of the p-i-n photodiode with its biasing voltage and integrator is shown
in Figure 2.5 below.
Figure 2.5: Simplied detector circuit employing a p-i-n photodiode.
2.6.2 Avalanche photodiodes
An avalanche photodiode (APD) is constructed similarly to a p-i-n photodiode. In some
models, there is a second p-type layer between the intrinsic layer and the n-type layer (the layers
are p-i-p-n). Incident photons still generate electron-hole pairs, but now there is an avalanche
effect each free electron and/or hole has the potential to create more free electrons and/or
holes as it traverses the gain region (the extra p-type layer and part of the n-type layer). Each
newly created electron or hole can then repeat the process until all carriers have exited the gain
region.
This avalanche process creates multiple carriers for every incident photon. This increase in
the number of carriers is known as the APD gain. It is the ratio of observable photocurrent at the
APD terminals to the internal photocurrent before multiplication [4], and is a random variable
with mean G. A simplied model of an APD is shown in Figure 2.6.
17
Figure 2.6: Simplied model of an APD and integrator.
2.6.3 Bandwidth and noise considerations in p-i-n and APD receivers
2.6.3.1 Bandwidth
An ideal photodetector would be noiseless and have an innite bandwidth. An actual pho-
todetector, however, has neither of these attributes due to the physical qualities of the photode-
tector itself and the accompanying electronics that take part in the detection process.
A photodetector has a junction capacitance C
d
associated with it that is proportional to the
aperture size. Increasing the capacitance of the detector also increases the time-constant which
lowers the bandwidth. Therefore there is a tradeoff between increasing the eld-of-view (FOV)
for a detector and the detectors bandwidth.
The time-constant, and consequently the bandwidth is also dependent on the load resistor
(R in Figures 2.5 and 2.6). Kedar and Arnon give an estimate for the data rate based on these
parameters [21]
R
b

1
2RC
d
(2.4)
which is simply the inverse of the time constant converted to frequency in Hz. Clearly, increas-
ing the capacitance (by increasing the aperture size) or increasing the resistance decreases the
data rate.
2.6.3.2 Shot noise
There are also many sources of noise that must be considered when doing analysis on pho-
todetection circuits. The rst source we consider is optical shot noise, which occurs because
18
of the randomness of the creation of photoelectrons. We adopt the semi-classical view of pho-
todetection, in which light arrives as a wave, and produces a stream of photoelectrons from
the detector. The number of photoelectrons produced during a slot time can be described by
a Poisson random variable with a mean and variance . This variance in the generation of
photoelectrons can be seen as noise, which can affect the probability of error.
2.6.3.3 Background noise
In a terrestrial free-space optical system, there is also a very good chance that background
noise will enter the receiver along with the signal. Sources of background noise include the sun
and articial lighting, and can enter the receivers aperture directly or by reecting off of other
surfaces. One way of minimizing background noise is by blocking other sources of radiation so
that light can enter the receiver only from approximately the direction of the transmitter.
Another method for minimizing background noise is to use a frequency selective lter in
front of the receiver to only allow a narrow band around the center frequency of the laser. The
potential drawback to this method is that it reduces the strength of the received signal.
2.6.3.4 Thermal noise
Thermal noise is also called Johnson noise, and is a result of thermally induced random
uctuations in the charge carriers in a resistive element. Thermal noise is technically present in
any semiconductor where thermally induced charge carriers can be present, which even includes
the photodetector itself, but it is only signicant in the load resistor (see Figures 2.5 and 2.6)
whose resistance is higher than the other sources.
Both the p-i-n photodiode and APD are affected by thermal noise, but it is more detrimental
for the p-i-n photodiode. The APD has internal amplication that can be seen as a low-noise
amplier, whereas the p-i-n photodiode relies completely on the circuitry for amplication.
2.6.3.5 Excess APD noise
The APD has the advantage of internal multiplication to raise the overall SNR, but it does so
in spite of the excess noise factor F(M), caused by the random nature of the gain mechanism.
F(M) depends on the semiconductor material, the average gain of the APD, the ratio of the
ionization coefcients for electrons and holes, and is largest in devices where both holes and
19
electrons produce ionizing collisions [4]. The SNR after the avalanche process for the APD is
reduced by multiplying the SNR before amplication by F(M)
1
.
2.7 MPPM demodulator
The MPPM demodulator has the task of taking the electronic signals delivered by the re-
ceiver and deciding which of the MPPM symbols was sent. This is a non-trivial task, and the
maximum likelihood decision metric is investigated further in Chapter 4.
2.8 Source decoder, channel decoder, and retrieved informa-
tion
Once the MPPM demodulator has decided which MPPM symbol was sent, the source and
channel decoders perform the inverse operations of the source and channel encoders. If the
symbol was chosen correctly, the retrieved information matches the input information to the
system. If the symbol was chosen incorrectly, the error may be detected and possibly even
corrected (depending on the error and the coding scheme). It is also possible that the error could
remain undetected and the retrieved information would not exactly match input information, i.e.
bit errors would occur. These aspects are studied in detail in Section 5.
Chapter 3
MIMO applied to FSO systems using
MPPM - background, system model and
denitions
In this chapter we look at the application of multiple-input, multiple-output (MIMO) tech-
niques to the FSO system using multiple pulse position modulation (MPPM).
3.1 Research on MIMO and MPPM FSO communications
systems
3.1.1 MIMO in wireless systems
Multiple Input Multiple Output (MIMO) systems have recently emerged as one of the most
signicant breakthroughs in modern communications. The idea behind MIMO systems can be
explained quite simply. At both the transmitter and receiver end, the system employs multiple
antennas. The effect of MIMO approaches is that the signals at the transmitter and receiver can
be combined such that the bit error rate or the data rate (in bits/sec) is improved. In the wireless
RF domain this is done at no extra cost of spectrum only added hardware and complexity [22].
The MIMO concept is depicted in Figure 3.1.
The advantage from the MIMO setup can be utilized through two different concepts, spatial
multiplexing and spatial diversity.
20
21
Figure 3.1: The MIMO concept.
In spatial multiplexing, such as with the BLAST technique, the incoming high-rate data is
decomposed into M independent data streams, and sent to all M antennas to be transmitted
simultaneously over the channel. The receiver array, having learned the mixing channel matrix
through training sequences, can identify each of the individual data streams and recombine
them to retrieve the original message. The result is that the spectral efciency improves; the
transmitter array can send at a new data rate M times faster than with a single antenna [22].
The second benecial concept is spatial diversity. In MIMO systems, the MN path gains
from each transmitter to each receiver can be described in a M N matrix form (see Figure
3.1). If the antennas are situated far enough apart, the paths can be considered decorrelated
and the effects of random fading caused by multipath or other phenomena can be mitigated.
The improvement of a MIMO system is directly related to the number of decorrelated antenna
elements, also known as the diversity order, whose maximum is MN [22].
There has been much research into the performance advantages of MIMO systems. The
consensus is that MIMO design can increase the capacity and decrease the bit error rate over a
single input, single output (SISO), MISO, or SIMO system with a given power and bandwidth.
The results of this research have focused heavily on RF systems, and the reader is directed
to [11] or [22] (among many possibilites) for more detail.
22
3.1.1.1 Application of MIMO concepts to free-space optics
Free-space optical communications systems can also benet from spatial diversity, as has
been shown in [15, 2325] . Although a well-designed FSO link will not suffer from traditional
multipath effects (except for diffuse FSO systems), atmospheric fading (the most serious prob-
lem facing FSO systems) can be mitigated through MIMO design. The key is to place the lasers
and photodetectors sufciently far from one another to ensure with a high degree of probability
that each of the MN path gains are independent.
The distance that each laser or photodetector should be placed away from one another can
be calculated from the correlation distance d
0
(see Section 2.5.1). This is the distance for which
two points in the atmosphere are uncorrelated and can be approximated by d
0

L, where
is the wavelength of the transmitted wave, and L is the length of the FSO link [13].
As an example, if the length of the FSO link is 1 km and the wavelength is 1550 nm, the
correlation distance would be approximately 4 cm. Therefore, keeping the lasers and photode-
tectors separated by at least 4 cm would ensure with a high degree of probability that the MN
path gains are independent. This small of a separation distance is perhaps surprising but
it is small enough for MIMO to be considered for terrestrial FSO links, where the laser and
photodetector arrays would be placed on rooftops or even behind windows.
The result of keeping the MN path gains independent is seen by considering the probability
that a path gain a
nm
is sufciently small, such that the signal falls behind the background level.
If we call the probability of this event , then the probability of a deep fade detrimentally
affecting all of the paths for one realization of the path gain matrix A is
MN
. Therefore, the
system has the potential to achieve the diversity order MN.
3.1.2 Multiple pulse position modulation (MPPM)
The capacity and performance of PPM and MPPM has been studied in detail in [2633].
We will rst investigate the attributes of PPM signaling, and then transition to MPPM.
In pulse position modulation (PPM), an observable time period (the duration of a modulator
symbol, T) is divided up into slots each having a duration of T
s
. A symbol is represented by
sending a pulse in only one of the time slots. If there are Q slots in a symbol, then there are
consequently Q possible symbols, each representing up to log
2
Q bits of information. PPM is a
23
Q-ary orthogonal signaling scheme.
To increase the throughput of a PPM system, it is necessary to increase Q, which decreases
the pulsewidth [28]. This is an attractive solution for a number of reasons. First, decreasing the
width of the slot also decreases the number of background photons that will be received [29],
since we are assuming the background count rate remains the same. Second, the probability of
symbol error for noncoherent detection of M-ary orthogonal signals decreases for an increasing
number of symbols for a xed energy per bit, E
b
[10].
Decreasing the slot width, however, has its limitations. Namely, an increase in the required
bandwidth, implying more thermal noise. For high data rate applications, MPPM is a more
attractive alternative [34]. Multipulse transmission works with the same concept as PPM, but
instead of having only one on slot, there are w of Q slots that can be on for each modulator
symbol, giving
_
Q
w
_
possible symbols. The bandwidth efciency, dened as the number of bits
that can be transmitted per is superior in MPPM, and as we will see later, the MPPM system
has an improved performance when the system is peak-power-limited.
This was studied in detail by Atkin and Fung in [33]. In their analysis, they compared
different schemes with similar bandwidths, and found that MPPM can outperform standard
PPM in coded and uncoded systems. Our analysis differs from theirs in a few ways. We allow
all
_
Q
w
_
symbols to remain in the set, whereas they would limit the symbol set size to a power
of two, which is a logical limitation to place on the modulation scheme. The result is that their
error analysis was limited to an upper bound on error probability. By allowing all
_
Q
w
_
symbols
to remain in the set, we preserve symmetry in the problem, and can often obtain closed form
expressions for error probability. We also consider MIMO techniques overlayed with MPPM,
whereas one of their focuses was on Reed-Solomon coding of a MPPM system.
3.1.3 Error probability - Gaussian vs. Poisson
In optical detection the Gaussian approximation is often used to analyze systems that em-
ploy APDs as detectors [35, 36]. The properties of APDs and the accuracy of the Gaussian
approximation were established in the early 1970s in work by McIntyre, Conradi, and Webb
[3739], and the approximation is used to take many factors specic to APDs into consider-
ation. However, we are concerned with detection using p-i-n photodiodes, and details of the
operation of APDs is beyond the scope of this thesis.
24
Instead, when we speak of the Gaussian approximation, we are either interested in approx-
imating the Poisson point process at the output of the photodetector when the signal and back-
ground power levels are both large, or when additive white Gaussian noise (thermal noise) is
introduced by the receiver.
3.2 Power comparisons between PPM and MPPM
Care must be taken to fairly compare system performance as w varies. The reason for this is
that increasing w with a constant peak signaling power (the on power during a slot) increases
the average power consumption for the system. Also, increasing w while keeping the symbol
duration T
s
constant would increase the bit rate.
To address the rst concern, we will need to make a distinction between peak-power-limited
systems, and average-power-limited systems. In a peak-power-limited system, the signal power
in an on slot is limited to P
peak
, regardless of how many on slots there are. This means the
total received optical energy per symbol is equal to the peak power multiplied by the duration
of the on slots.
E
s
= P
peak
Tw (3.1)
In an average-power-limited system, the average power in all Q of the slots (both on and
off) observed over the duration of the symbol must remain constant, and the optical energy
per symbol is equal to
E
s
= P
ave
TQ (3.2)
Therefore, combining (3.1) and (3.2), we can state that the relationship between average
power and peak power is
P
ave
= P
peak
w
Q
(3.3)
P
ave
is easily recognized as P
peak
times the duty cycle of a symbol.
To consider the second concern, we observe that the bit rate of the system is related to w.
The bit time T
b
multiplied by the number of bits per modulator symbol is equal to the symbol
time,
T
s
= T
b
log
2
_
Q
w
_
(3.4)
25
and the slot time is therefore
T =
T
b
log
2
_
Q
w
_
Q
(3.5)
We can address both of these concerns by dening a general optical energy parameter PT
with which to compare systems (used in plots in Chapters 5 and 6), where P is the signal-
ing power and T is the slot duration. Peak and average power are handled by the following
conversions:
PT = P
ave
T
b
log
2
_
Q
w
_
w
or PT = P
peak
T
b
log
2
_
Q
w
_
Q
(3.6)
For a better understanding of the effect of average or peak power limitations on system
performance, we can dene a relative energy efciency for equal asymptotic performance, based
on (3.6) as the multiplier on PT
b
:

ave
=
log
2
_
Q
w
_
w
bits/slot (3.7)

peak
=
log
2
_
Q
w
_
Q
bits/slot (3.8)
which shows that for a given (Q, w) pair, multipulse is more efcient than single pulse in the
peak-power-limited system. In fact, for the peak-power-limited system, the relative energy
efciency is at a maximum when w = Q/2. The efciencies as a function of Q are shown
below in Figure 3.2.
It is important to be careful in interpreting this plot. The rst thing that the plots reveal is
that for a given PT
b
, the average-power-limited system will always outperform the peak-power-
limited system. The interpretation of this goes back to (3.3). Notice that if we are given P, and
interpret it as average power, the corresponding peak would be Q/w times greater than if P
were interpreted as peak power.
Also interesting to note is that for the peak-power-limited system, the MPPM system has
a superior relative energy efciency, and in an average-power-limited system, standard PPM is
superior.
It is also interesting to note from this plot that the efciency is unbounded for the average-
power-limited system where w = 1 (unlike all of the other curves). From a logical standpoint,
this makes sense; as Q increases, the efciency increases at a rate of log
2
Q. However, what
26
0 5 10 15 20 25 30
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Q
R
e
l
a
t
i
v
e

E
n
e
r
g
y

E
f
f
i
c
i
e
n
c
y

ave
, w=1

ave
, w=Q/2

peak
, w=Q/2

peak
, w=1
(Q,w) = (8,1)
(Q,w) = (8,4)
3.00
1.53
0.766
0.375
Figure 3.2: Relative energy efciencies vs. Q for w {1, Q/2}.
is not seen from this plot is that for the average power to remain constant as Q becomes large,
the peak power must also become large. Another way to think of this is that all of the power is
concentrated in a single slot the on slot is constantly becoming narrower as Q grows, which
squeezes the peak power (as well as the bandwidth) up toward some large value. Because of
power and bandwidth limitations in any real system, the advantages of this phenomenon become
impractical to implement for large values of Q. The zig-zag trajectories are due to the fact that
w = Q/2 only changes at even values of Q.
Another important advantage of the MPPM system is its spectral efciency. We can dene
this to be , which is measured in bps per Hertz. If we state that the bandwidth is inversely
proportional to the slot duration T, is calculated as
=
R
b
1/T
(3.9)
= R
b
T
b
log
2
_
Q
w
_
Q
(3.10)
= R
b
log
2
_
Q
w
_
R
b
Q
(3.11)
=
log
2
_
Q
w
_
Q
(3.12)
27
where R
b
is the bit rate of the system.
Since
_
Q
w
_
is maximized when w = Q/2, we can state that the spectral efciency of
MPPM is also maximized for w = Q/2. As Q becomes large, the spectral efciency of
MPPM approaches one for the w = Q/2 system, and approaches zero for the w = 1 system.
This can also be seen in Figure 3.2, since the spectral efciency is equivalent to the relative
energy efciency for the peak-power-limited case,
peak
. Although bandwidth conservation is
not of particular interest in FSO communications, since it operates on unregulated spectrum, the
spectral efciency is more of a measure of the necessary electronics speed to operate at a given
bit rate. Alternatively, the spectral efciency shows us that for a given bandwidth (or minimum
slot duration allowed by the electronics), the MPPM system will allow the system designer to
transmit at a higher bit rate. This is one of the most attractive features of MPPM for high data
rate communications.
3.3 System model
3.3.1 MPPM signaling
We send binary information using multipulse pulse-position modulation, where each mod-
ulator symbol is of duration T
s
seconds and is comprised of Q slots, each T seconds long. The
symbols are created by turning the laser on for w of the Q slots. This results in
_
Q
w
_
different
symbols which we assume are equiprobable. Each symbol can represent log
2
_
Q
w
_
bits of infor-
mation. For simplication of the analysis, we ignore the fact that
_
Q
w
_
may not be a power of
two.
We will often revert to the notation that an on slot is denoted by a one and an off slot
by a zero. For example, a Q = 8, w = 4 symbol, where the rst w slots are on can be
denoted by [1 1 1 1 0 0 0 0]. We also make use of the notation that Q
i
on
denotes the set of
on slots for modulator symbol i, and Q
i
off
is the set of off slots. In our example above, for
q Q
i
on
, q = {0, 1, 2, 3} and for q Q
i
off
, q = {4, 5, 6, 7}.
28
3.3.2 Transmitter array
For our system, the transmitters are modeled to have innite bandwidth, so it is assumed
possible to have the laser on at some constant power for the duration of a slot. Each laser is also
completely off (full extinction) for the entire duration of any off slot.
The number of lasers is denoted by M. In order to compare systems differing in the number
of lasers, we divide the transmitter power at each laser by M so the power delivered by the entire
transmitter array is constant as we vary M. We also assume that the lasers are noncoherent,
without any special precautions. The wavelength we have chosen in simulations for the system
is 1550 nm, which corresponds to a frequency of 2 10
14
Hz.
3.3.3 Receiver array
Each of the N receivers is assumed to be perfect-photon-counting devices, also with innite
bandwidth. The arrival of the signal and background photons at the receiver is modeled as a
Poisson process, where the number of photons arriving at detector n during a single slot time
has a mean and variance of
on,n
for on slots and
off
for off slots.
We also assume perfect synchronization, so signal photons are only received during on
slots, and off slots can only contain background photons.
3.3.4 Channel
The channel was described in detail in Section 2.5. We assume that the M transmitters and
N receivers are placed sufciently far from one another such that each of the individual paths
from transmitter to receiver is independent.
The intensity gain along each path from transmitter m to receiver n is denoted by a
2
nm
, and
it is a random variable following the distributions described in Section 2.5.3. We denote a single
realization of all of the amplitude path gains as a M N matrix, A.
The signal received at detector n is a composite of the signals received from all of the
M transmitters simultaneously. Therefore, at each detector n, the signal power received is
proportional to

m
a
2
nm
, assuming all transmitters send with the same power.
29
3.3.5 Detector and observable
3.3.5.1 Poisson Regime
We rst focus on the Poisson regime, which is applicable when the background count rate
is low and the variance in the observable at detector n for slot q is Z
nq
due to thermal noise in
the amplier is small. Each of the N receivers is modeled according to Figure 2.5, and consists
of a photodetector and integrator. The photodetector is analyzed using the semi-classical view
of photodetection. The optical wave is received, and the output is a ow of photoelectrons that
obey Poisson statistics over any slot interval. The observable at detector n and slot q is depicted
as Z
nq
, and has a mean on slot count
on,n
, and a mean off slot count
o
ff. These are given
by

on,n
=
PT
hfM

m
a
2
nm
+
P
b
T
hf
(3.13)
and

off
=
P
b
T
hf
, n = 1, ..., N (3.14)
respectively, where P
b
is the power received due to background radiation, and is an efciency
factor for the detector, dened as the ratio of generated photoelectrons to incident photons.
3.3.5.2 Gaussian Regime
For this case, it is convenient to model the noise using a Norton equivalent circuit, as in
Figure 3.3.
Figure 3.3: Norton equivalent noise model.
30
In the gure, I is the mean current, equal to the average number of photoelectrons created
by the photodetector times the charge associated with each photoelectron, q, giving
I =
_
_
_
_
P
hfM

m
a
2
nm
+
P
b
hf
_
q, signal present;
_
Pb
hf
_
q, no signal present.
(3.15)
The voltage across the (noiseless) resistor R is equal to IR. Since the mean values of the
optical shot noise i
s
(t) and thermal noise i
t
(t) are equal to zero, and the mean of an on or off
slot is assumed to be constant over the slot time T, the output of the integrator Z has a mean
equal to
E[Z] =
_
_
_
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
Rq, signal present;
_
PbT
hf
_
Rq, no signal present.
(3.16)
The generation of photoelectrons is a Poisson point process, so the variance in the photo-
electron count during an on or off slot is equal to (3.13) and (3.14). The voltage at the input
to the integrator caused by this optical shot noise current is this value multiplied by a constant
qR. Since a random variable X with a variance of
2
x
times a constant k has a new variance of
k
2

2
x
, we can state that the variance of the output due to optical shot noise is equal to
V ar[Z]
shot
=
_
_
_
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
R
2
q
2
, signal present;
_
PbT
hf
_
R
2
q
2
, no signal present.
(3.17)
The variance in the thermal noise current i
t
(t) can be described by its spectral noise density
S
t
(f) = 2kT
0
/R, where T
0
is the absolute temperature of the resistor R [40]. Multiplying this
current by the resistance again means multiplying its variance by R
2
. Taking the integral over
T seconds gives
V ar[Z]
thermal
= 2kT
0
TR (3.18)
Since the thermal and shot noise processes are independent, we can add them together to get
the variance at the output of the integrator. Putting everything together, we have that the output
of the integrator is a random variable Z described by
E[Z] =
_
_
_
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
Rq, signal present;
_
PbT
hf
_
Rq, no signal present.
(3.19)
31
and
V ar[Z] =
_
_
_
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
R
2
q
2
+ 2kT
0
TR, signal present;
_
PbT
hf
_
R
2
q
2
+ 2kT
0
TR, no signal present.
(3.20)
At this point, we may choose to use (3.19) and (3.20) in a Gaussian approximation to model
Z. However, we must be careful to determine when this approximation is valid.
The thermal noise is modeled as a Gaussian random variable. Therefore, if its variance is
much larger than the optical shot noise variance, the rst term may be neglected in (3.20), and
Z may be accurately approximated by a Gaussian random variable.
Also, if the level of background power P
b
is high enough, the p.d.f. for the count of photo-
electrons, which is Poisson in nature, becomes increasingly Gaussian in shape (see Figure 3.4).
When this is added to the Gaussian noise, the resulting p.d.f. may also be modeled as Gaus-
sian. This fact is also true of high signal power P, but high background power P
b
is a sufcient
condition for the Gaussian approximation to hold, and necessary when thermal noise is low.
Figure 3.4: Poisson and Gaussian p.d.f.s with equal means and variances of = 200.
In the case where the background power P
b
and thermal noise power levels are both low,
the Gaussian approximation ceases to yield accurate results. The Poisson distribution would
32
be poorly approximated by a Gaussian distribution (see Figure 3.5). As a result, we must let
the model for the distribution of the off slots remain as Poisson. Thus, the convolution (due
to adding independent random variables) of the Gaussian and Poisson p.d.f.s would no longer
have a Gaussian shape. Instead, for low mean count rates, the convolution of the two would
create a p.d.f. having many non-overlapping Gaussian shapes, each centered and weighted at
the different integer values given by the Poisson p.d.f., which would be poorly approximated by
a Gaussian distribution.
Figure 3.5: Poisson and Gaussian p.d.f.s with equal means and variances of = 2.
Chapter 4
Maximum likelihood (ML) detection
In this section, the general ML detection rules are derived for MIMO FSO detection in both
the Poisson and Gaussian regimes.
4.1 ML Detection for the poisson regime
Let Z
nq
be the photoelectron counts for detector n and slot q, and Z = {Z
nq
, n = 1, ..., N, q =
0, ..., Q 1} represent the set of received observations. If we send one of
_
Q
w
_
binary patterns
represented by X
i
, then the ML decision is

X = arg max
X
i
f(Z|X
i
) (4.1)
Using the denitions of
on,n
and
off
from Section 3.3.5, and recognizing that Z
nq
at each of
the N detectors and Q slots is independent, the conditional distribution of the N Q random
matrix Z can be written as a N Q-fold product over all of the individual elements in Z. These
elements, which are represented as Z
nq
, are conditioned on whether they are in an on slot, or
an off slot.

X = arg max
X
i

qQ
(i)
on
exp(
on,n
)(
on,n
)
Z
nq
Z
nq
!

qQ
(i)
off
exp(
off
)(
off
)
Z
nq
Z
nq
!
(4.2)
Since the Z
nq
! terms in the denominator are invariant to X
i
, they can be removed without
affecting the outcome of the ML decision. Thus
33
34

X = arg max
X
i

qQ
(i)
on
exp(
on,n
)(
on,n
)
Z
nq

qQ
(i)
off
exp(
off
)(
off
)
Z
nq
(4.3)
If we assume that the number of on slots for all modulator symbols is equal to w, the
exponential terms are equal to exp (w
on,n
) and exp ((Qw)
b
) for all X
i
and can also be
eliminated, leaving

X = arg max
X
i

qQ
(i)
on
(
on,n
)
Z
nq

qQ
(i)
off
(
off
)
Z
nq
(4.4)
Next, we take the logarithm of the entire quantity to nd the log-likelihood function.

X = arg max
X
i

n
_

qQ
(i)
on
ln((
on,n
)
Z
nq
) +

qQ
(i)
off
ln((
off
)
Z
nq
)
_

_
(4.5)
which can be rewritten as

X = arg max
X
i

n
_

qQ
(i)
on
Z
nq
ln(
on,n
) +

qQ
(i)
off
Z
nq
ln(
off
)
_

_
(4.6)
= arg max
X
i

n
_
_

qQ
(i)
on
Z
nq
ln(
on,n
) +

all q
Z
nq
ln(
off
)

qQ
(i)
on
Z
nq
ln(
off
)
_
_
(4.7)
= arg max
X
i

n
_
_

qQ
(i)
on
Z
nq
ln
_

on,n

off
_
_
_
(4.8)
Therefore, the ML detector would make a decision based on a weighted sum over the on
slots.
What will be interesting to note in the following subsections is that of the four cases we
will consider in the Poisson regime, the optimal detector simplies to nding the w largest
column sums with equal gain combining for three of them. The exception to this rule is the case
where background radiation and fading are both present, and the optimal detector searches for
a weighted sum over the on and off slots.
4.1.1 Case 1: ML detection with no background and no fading
For this case, we start with (4.6), and express
on,n
and
off
explicitly:
35

X = arg max
X
i

n
_

qQ
(i)
on
Z
nq
ln(
on,n
) +

qQ
(i)
off
Z
nq
ln(
off
)
_
(4.9)
= arg max
X
i

n
_

qQ
(i)
on
Z
nq
ln
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
+

qQ
(i)
off
Z
nq
ln
_
P
b
T
hf
_
_
(4.10)
For the cases where no background radiation is present, P
b
= 0 and Z
nq
= 0 for all off
slots. Since
lim
x0
+
x log(x) = 0 (4.11)
(4.10) can be restated as

X = arg max
X
i

qQ
(i)
on
Z
nq
ln
_
PT
hfM

m
a
2
nm
_
(4.12)
In the no fading case, all of the fading variables A are equal to one, meaning that the
ln
_
PT
hfM

m
a
2
nm
_
term is constant for all X
i
leading to

X = arg max
X
i

qQ
(i)
on
Z
nq
(4.13)
= arg max
X
i

qQ
(i)
on

n
Z
nq
(4.14)
If we dene

n
Z
nq
= S
q
, which we will refer to as a column sum,

X = arg max
X
i

qQ
(i)
on
S
q
(4.15)
which shows that the optimal detector simply searches for the w largest column sums. In the
case where ties occur and there are more than w column sums that could provide an optimal
solution, the detector must resort to making a random choice among all of the equally likely
codewords.
36
4.1.2 Case 2: ML detection with no background and fading
With no background and fading, we know that any slot that has a non-zero count must be
an on slot, and any slot that has a zero count may be an off slot. Intuitively, we can easily
state that if w slots have non-zero counts at one or more detectors, there is no ambiguity in the
transmitted symbol and the system will err with zero probability.
Therefore, no outside information is required to make the maximum likelihood decision.
Any scheme involving weighting each column sum differently based on the current realization
of A cannot assist in the outcome of the decision, so the optimal detector simply searches for
the w largest (or simply non-zero) column sums.
4.1.3 Case 3: ML detection with background radiation and no fading
In this case, we again start with

X = arg max
X
i

n
_

qQ
(i)
on
Z
nq
ln
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
+

qQ
(i)
off
Z
nq
ln
_
P
b
T
hf
__
(4.16)
Here, the

m
a
2
nm
term is equal to M, and bias terms can be eliminated, yielding

X = arg max
X
i

n
_

qQ
(i)
on
Z
nq
ln (P +P
b
) +

qQ
(i)
off
Z
nq
ln P
b
_
(4.17)
Here, we can make the observation that logarithms are monotonic increasing functions, and
P and P
b
are both positive numbers or zero. Therefore we can state that ln(P +P
b
) ln P
b
for
all P and P
b
. If we make a substitution and let = ln(P
b
) and + = ln(P + P
b
) (so that
0 is the difference between the two logarithmic terms), we can rewrite (4.17) as

X = arg max
X
i

n
_
( + )

qQ
(i)
on
Z
nq
+

qQ
(i)
off
Z
nq
_
(4.18)
= arg max
X
i

n
_

qQ
(i)
on
Z
nq
+

all q
Z
nq
_
(4.19)
37
Since the

all q
Z
nq
term is independent of X
i
, it can be eliminated. is also a scale
factor which is independent of X
i
, leaving the following ML decision rule:

X = arg max
X
i

qQ
(i)
on
Z
nq
(4.20)
= arg max
X
i

qQ
(i)
on

n
Z
nq
(4.21)
= arg max
X
i

qQ
(i)
on
S
q
(4.22)
This once again shows that the optimal detector simply searches for the w largest column
sums.
4.1.4 Case 4: Background radiation and fading
When background radiation and fading are both present, we again start with

X = arg max
X
i

n
_

qQ
(i)
on
Z
nq
ln
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
+

qQ
(i)
off
Z
nq
ln
_
P
b
T
hf
__
(4.23)
Eliminating bias terms leaves

X = arg max
X
i

n
_

qQ
(i)
on
Z
nq
ln
_
P
M

m
a
2
nm
+P
b
_
+

qQ
(i)
off
Z
nq
ln P
b
_
(4.24)
If we again employ the same method applied to the previous case, we can state that
ln [(P/M)

m
a
2
nm
+P
b
] ln P
b
for all P, P
b
, M, and a
nm
. Therefore, we can make the
following substitutions: Let = ln P
b
and +
n
= ln [(P/M)

m
a
2
nm
+P
b
], such that

n
0 is the difference between the two logarithmic terms, and is dependent on n through the
fading distribution A.
We can then rewrite the ML decision rule as
38

X = arg max
X
i

n
_
( +
n
)

qQ
(i)
on
Z
nq
+

qQ
(i)
off
Z
nq
_
(4.25)
= arg max
X
i

n
_
( +
n
)

qQ
(i)
on
Z
nq
+

qQ
(i)
off
Z
nq
_
(4.26)
= arg max
X
i

n
_

qQ
(i)
on
Z
nq
+

all q
Z
nq
_
(4.27)
The last summation term is independent of X
i
and can be eliminated, which leaves

X = arg max
X
i

n
_

qQ
(i)
on
Z
nq
_
(4.28)
where
n
= ln [(P/M)

m
a
2
nm
+P
b
]ln P
b
. This shows that the optimal detector will perform
a weighted sum of the on slots.
A reasonable approximation to this detector is to perform equal gain combining and search
for the w largest column sums, treating this case like the other 3 cases. This is only slightly
suboptimal, as was shown in [19], and is the method that was chosen for the analysis in Chapters
5 and 6.
4.2 General ML detection in the Gaussian regime
Assuming the necessary conditions are met for the Gaussian approximation to be justied
(see Section 3.3.5.2), we can start to develop the general ML detector in the Gaussian regime.
For now, we will not assume that thermal noise or shot noise is dominant, and we will
assume the background and signal power levels are such that the Gaussian approximation is
justied. This will allow us to develop a more general ML detector. More specic cases will be
considered in the following subsections.
Based on the analysis in Section 3.3.5.2 we can state that the observable Z is a Gaussian
random variable with parameters

on,n
=
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
Rq (4.29)
39

off
=
_
P
b
T
hf
_
Rq (4.30)

2
on,n
= (
on,n
) Rq + 2kT
0
TR (4.31)

2
off
= (
off
) Rq + 2kT
0
TR (4.32)
Similar to the development from the Poisson regime, we can claim that N Q elements in
Z are independent, and therefore the distribution of Z is simply a N Q-fold product over all
detectors and slots of the (Gaussian) distribution of each Z
nq
, which are conditioned on q Q
on
or q Q
off
:

X = arg max
X
i
N

n=1

qQ
(i)
On
_
1
_
2
2
on,n
e

(Z
nq

on,n
)
2
2
2
on,n
_

qQ
(i)
Off
_
_
1
_
2
2
off
e

(Z
nq

off
)
2
2
2
off
_
_
(4.33)
Eliminating the 1/

2 scale factors and taking the logarithm gives

X = arg max
X
i
N

n=1
_

qQ
(i)
On
_

(Z
nq

on,n
)
2
2
2
on,n
ln(
on,n
)
_
+

qQ
(i)
Off
_

(Z
nq

off
)
2
2
2
off
ln(
off
)
_
_
(4.34)
= arg max
X
i
N

n=1
_

qQ
(i)
On
_
(Z
2
nq
+ 2
on,n
Z
nq

2
on,n
)
2
2
on,n
ln(
on,n
)
_
+

qQ
(i)
Off
_
(Z
2
nq
+ 2
off
Z
nq

2
off
)
2
2
off
ln(
off
)
_
_
(4.35)
The
2
on,n
,
2
off
, ln(
on,n
), and ln(
off
) terms are constant for all X
i
, so
40

X = arg max
X
i
N

n=1
_

qQ
(i)
On
_
2
on,n
Z
nq
Z
2
nq
2
2
on,n
_
+

qQ
(i)
Off
_
2
off
Z
nq
Z
2
nq
2
2
off
_
_

_
(4.36)
= arg max
X
i
N

n=1
_

qQ
(i)
On
_
2
on,n
Z
nq
Z
2
nq
2
2
on,n
_
+

qQ
(i)
All
_
2
off
Z
nq
Z
2
nq
2
2
off
_

qQ
(i)
On
_
2
off
Z
nq
Z
2
nq
2
2
off
__
(4.37)
= arg max
X
i
N

n=1
_

qQ
(i)
On
_
2
on,n
Z
nq
Z
2
nq
2
2
on,n

2
off
Z
nq
Z
2
nq
2
2
off
__
(4.38)
= arg max
X
i
N

n=1
_

qQ
(i)
On
_
_

on,n

2
off

off

2
on,n
_
Z
nq
+
_

2
on,n

2
off
2
_
Z
2
nq
__
(4.39)
We will use this as the starting point for the different cases in the Gaussian regime.
4.2.1 Thermal noise dominates over shot noise
When the thermal noise is the dominant noise in the system, we can neglect the variance
due to the shot noise, and state that the slot count at each receiver, Z
nq
follows a Gaussian
distribution with mean and variance of

on,n
=
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
Rq, signal present (4.40)

off
=
_
P
b
T
hf
_
Rq, no signal present (4.41)

2
= 2kT
0
TR, either case (4.42)
Therefore, to determine the ML decision in the Gaussian regime, we start with (4.39), but
let the variance terms be dominated by 2kT
0
TR. This results in

X = arg max
X
i
N

n=1
(
on,n

off
) Z
nq
(4.43)
41
After making the appropriate substitutions, eliminating the scale factors gives the ML de-
tector as

X = arg max
X
i
N

n=1
_

qQ
(i)
On
_

m
a
2
nm
_
Z
nq
_

_
(4.44)
4.2.1.1 No fading present
When no fading is present (and regardless of the presence of background radiation),

m
a
2
nm
=
M. This becomes a scale factor that can be eliminated, giving

X = arg max
X
i

qQ
(i)
on
Z
nq
(4.45)
= arg max
X
i

qQ
(i)
on

n
Z
nq
(4.46)
= arg max
X
i

qQ
(i)
on
S
q
(4.47)
which shows that the optimal detector searches for the largest column sums. This is an iden-
tical result to the Poisson regime when no fading was present, where the optimal detector also
searched for the w largest column sums.
4.2.1.2 Fading present
With fading present (and regardless of the presence of background radiation), (4.44) is irre-
ducible and gives the optimal detector, which searches for the largest weighted column sums,
dependent on the fading distribution A.
4.2.2 Shot noise dominates over thermal noise
For this case, we will start with (4.39), and simply plug in the mean and variance expressions
given by

on,n
=
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
Rq (4.48)
42

off
=
_
P
b
T
hf
_
Rq (4.49)

2
on,n
= (
on,n
) Rq (4.50)

2
off
= (
off
) Rq (4.51)
4.2.2.1 Fading present
After plugging in the appropriate means and variances into (4.39) and eliminating the scale
factors, the ML detector is

X = arg max
X
i
N

n=1
_

qQ
(i)
On
_

m
a
2
nm
_
Z
2
nq
_

_
(4.52)
This differs from the Poisson regime, where the optimal detector looks for weighted column
sums. Here, the optimal detector will square the observable at each detector before multiplying
by the weighting coefcient and summing over detectors and on slots.
4.2.2.2 No fading present
With no fading present,

m
a
2
nm
= M, and just as when thermal noise was dominant, the
optimal detector becomes

X = arg max
X
i

qQ
(i)
on
Z
2
nq
(4.53)
= arg max
X
i

qQ
(i)
on

n
Z
2
nq
(4.54)
This also differs from the Poisson regime since the optimal detector will square the observ-
able at each detector, and then look for the w largest column sums (as opposed to simply
looking for the w largest column sums in the Poisson regime).
4.2.3 Shot and thermal noise are not dominant
In this case, we use the means and variances dened earlier as
43

on,n
=
_
PT
hfM

m
a
2
nm
+
P
b
T
hf
_
Rq (4.55)

off
=
_
P
b
T
hf
_
Rq (4.56)

2
on,n
= (
on,n
) Rq + 2kT
0
TR (4.57)

2
off
= (
off
) Rq + 2kT
0
TR (4.58)
4.2.3.1 Fading Present
Plugging these values into (4.39) and eliminating scale factors terms gives us

X = arg max
X
i

n
_
_

qQ
(i)
on
_

m
a
2
nm
_
_
(2kT
0
T) Z
nq
+
_
q
2
_
Z
2
nq
_
_
_
(4.59)
This result is similar to other fading results in the Gaussian and Poisson regimes, since it
shows the optimal detector to be a weighted sum based on the fading distribution, A.
In most instances, one would expect (q/2)Z
2
nq
to negligible, in which case 2kT
0
T becomes
a scale factor. In these instances, the optimal detector is identical to the optimal detector for the
other fading cases in the Gaussian regime.
4.2.3.2 No fading present
When no fading is present, (4.60) becomes

X = arg max
X
i

n
_
_

qQ
(i)
on
_
(2kT
0
T) Z
nq
+
_
q
2
_
Z
2
nq
_
_
_
(4.60)
Again, when (q/2)Z
2
nq
is negligible, 2kT
0
T becomes a scale factor and the optimal detector
is simply searching for the w largest column sums.
Chapter 5
Error analysis of MIMO FSO system
using MPPM
In this chapter, we will perform symbol error probability analysis of the MIMO MPPM
FSO system in both the Poisson and Gaussian regimes. For simplicity of the analysis, we will
assume equal gain combining for all cases. This is optimal in three of the four cases under
consideration, and was shown to be a very good approximation for the Poisson regime in the
case where background radiation and atmospheric fading are both present [19]. We also show
this to be a very good approximation in the Gaussian regime in Section 5.2.
5.1 Error analysis in the Poisson regime
For the Poisson regime, we consider a number of possible scenarios that are combinations
of the following: no fading, Rayleigh fading, log-normal fading, background radiation, and no
background radiation.
5.1.1 No background radiation
For the case of no background radiation, the w column sums with the largest counts will be
the w on slots in the transmitted symbol X
j
unless i > 0 of them are zero (due to quantum
effects in the photodetector).
With i of the column sums equaling zero, the detector may have partial information (or no
44
45
information if i = w), with which it can make a guess at the transmitted symbol. The receiver
must decide among
_
Qw+i
i
_
equally probable symbols. This quantity can be interpreted by
noting that this is the number of ways the receiver can place the i non-received signal pulses in
Qw +i vacant slots. The detector errs with probability
t(Q, w, i) =
_
Qw+i
i
_
1
_
Qw+i
i
_ (5.1)
As described in Section 3.3.5, for on slots, we can model the photoelectron slot count Z
nq
at each of the N receivers as a Poisson process with a Poisson parameter of
on,n
and
off
for
on and off slots. With equal gain combining, the sum over all receivers for each on slot is
also a Poisson random variable (conditional on A) with a parameter of

on
=
PT
hfM
N

n=1
M

m=1
a
2
nm
(5.2)
If we dene S
q
=

N
n=1
Z
nq
, an on slot will have zero photoelectrons with probability
P(S
q
= 0) = p = e

on
(5.3)
Treating each column count as an independent random variable, we get
P[i of w columns = 0] =
w

i=1
_
w
i
_
p
i
(1 p)
wi
(5.4)
Therefore, we can state for the no background case, that the error probability conditioned
on the fading variables A is
P
s|A
=
w

i=1
t(Q, w, i)
_
w
i
_
p
i
(1 p)
wi
(5.5)
We can take the (1 p)
wi
term and rewrite it using the binomial expansion:
(1 p)
wi
=
wi

l=0
(1)
l
_
w i
l
_
p
l
(5.6)
This gives the probability expression
P
s|A
=
w

i=1
t(Q, w, i)
_
w
i
_
p
i
wi

l=0
(1)
l
_
w i
l
_
p
l
(5.7)
46
which is easily rewritten
P
s|A
=
w

i=1
wi

l=0
(1)
l
_
Qw+i
i
_
1
_
Qw+i
i
_
_
w
i
__
w i
l
_
p
(i+l)
(5.8)
This is the probability of symbol error, conditioned on no background radiation and the
fading variables A. In each of the next three subsections we will use different distributions of
A, and show how the system performs.
5.1.1.1 No background, no fading
In the non-fading case, a
2
nm
= 1 for all n and m. Plugging this value into (5.2), (5.3), and
(5.8) gives

on
=
PTN
hf
(5.9)
P(S
q
= 0) = p = e

PTN
hf
(5.10)
and
P
s|no fading
=
w

i=1
wi

l=0
(1)
l
_
Qw+i
i
_
1
_
Qw+i
i
_
_
w
i
__
w i
l
_
e

PTN(i+l)
hf
(5.11)
respectively.
Observe that the average number of photoelectron counts and the probability of error are
independent of the number of lasers M in the non-fading case. This is because the M path
gains from each laser to each detector are unity, and we kept the total transmit power constant
by dividing the transmit power at each laser by M. However, performance does depend on the
number of receivers because the total aperture size increases by a factor of N.
For consistency, we could have removed that dependency on receivers by also decreasing the
total aperture size by N (as done by Shin and Chan in [15]), but in a realistic setting increasing
the number of receivers will also increase the aperture size, so our choice is justied.
As can be seen by Figures 5.1 and 5.2, no change in performance is observable for varying
M. However, a 6 dB gain in performance is seen in the cases where N is equal to four. This
can be attributed to the exponent of p being increased by a factor of four.
47
It is also important to note that multipulse PPM outperforms standard PPM only in a peak-
power-limited system. This makes intuitive sense, since average-power-limited systems would
require that the total signal (on) power be distributed among the w on slots in the average-
power-limited system, as w increases, each on slot has a lower peak power, which increases
the probability that a zero is observed at the detector during an on slot.
210 200 190 180 170 160 150
10
12
10
10
10
8
10
6
10
4
10
2
10
0
PT
b
,dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.1: Symbol error probability vs. average power with no fading and no background
radiation, and Q = 8.
5.1.1.2 No background, Rayleigh fading
For the Rayleigh fading case, we refer back to Section 2.5.3, which discusses different
fading models. (2.3) gives the p.d.f. for Rayleigh fading, and is repeated here for convenience:
f
A
(a
nm
) = 2a
nm
e
a
2
nm
, a
nm
> 0 (5.12)
We assume that each realization of a channel fading gain a
nm
follows this distribution, and
therefore the expected error probability can be found by averaging (5.8) with respect to the
Rayleigh distribution for all MN channel paths.
48
210 200 190 180 170 160 150
10
12
10
10
10
8
10
6
10
4
10
2
10
0
PT
b
,dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.2: Symbol error probability vs. peak power with no fading and no background radia-
tion, and Q = 8.
P
s|Rayleigh fading
=
N

n=1
M

m=1
_

0
w

i=1
wi

l=0
_
Qw+i
i
_
1
_
Qw+i
i
_
_
w
i
__
w i
l
_

(1)
l
2a
nm
e
[PTa
2
nm
(i+l)]/hfMQ
da
nm
(5.13)
=
w

i=1
wi

l=0
_
Qw+i
i
_
1
_
Qw+i
i
_
_
w
i
__
w i
l
_
(1)
l

m=1
N

n=1
_

0
2a
nm
e
[PTa
2
nm
(i+l)]/hfMQ
da
nm
(5.14)
This is a NM-fold integration, but each term is identical so it can be rewritten
P
s|Rayleigh fading
=
w

i=1
wi

l=0
_
Qw+i
i
_
1
_
Qw+i
i
_
_
w
i
__
w i
l
_
(1)
l

__

0
2ae
[PTa
2
(i+l)]/hfMQ
da
_
MN
(5.15)
and applying the solution to this integral gives the nal closed-form result
P
s|Rayleigh fading
=
w

i=1
wi

l=0
_
Qw+i
i
_
1
_
Qw+i
i
_
_
w
i
__
w i
l
_
(1)
l
_
1
1 +
PT(i+l)
hfMQ
_
MN
(5.16)
49
This is plotted in Figures 5.3 and 5.4. Since the error probability drops by a factor of 10
MN
for every 10 dB increase in signal power, we claim that the system achieves a diversity equal
to MN. This shows a performance advantage to increasing the number of lasers as well as the
number of receivers. However, comparing the M = 1, N = 4 to the M = 4, N = 1 curves
for constant w shows that for a given diversity, increasing the number of receivers has more
of a benecial effect than increasing the number of transmitters, due to the increase in overall
receiver aperture size.
210 200 190 180 170 160 150 140
10
12
10
10
10
8
10
6
10
4
10
2
10
0
PT
b
, dBJ
P
s
M=1 N=1, W=1
M=4 N=1, W=1
M=1 N=4, W=1
M=4 N=4, W=1
M=1 N=1, W=4
M=4 N=1, W=4
M=1 N=4, W=4
M=4 N=4, W=4
Figure 5.3: Symbol error probability vs. average power with Rayleigh fading, no background
radiation, and Q = 8.
5.1.1.3 No background, log-normal fading
With log-normal fading, the channel path gains a
nm
follow the distribution in (2.1), repeated
here for convenience
f
A
(a
nm
) =
1
(2
2
X
)
1/2
a
nm
exp((log
e
a
nm

X
)
2
/2
2
X
), a
nm
> 0 (5.17)
where
X
=
2
X
and S.I. = E[A
4
]/E
2
[A
2
] 1 = e
4
2
X
1 [0.4, 1.0].
The symbol error probability could be found by averaging over the log-normal distribution:
50
210 200 190 180 170 160 150 140
10
12
10
10
10
8
10
6
10
4
10
2
10
0
PT
b
, dBJ
P
s
M=1 N=1, W=1
M=4 N=1, W=1
M=1 N=4, W=1
M=4 N=4, W=1
M=1 N=1, W=4
M=4 N=1, W=4
M=1 N=4, W=4
M=4 N=4, W=4
Figure 5.4: Symbol error probability vs. peak power with Rayleigh fading, no background
radiation, and Q = 8.
P
s|Lognormal fading
=
N

n=1
N

m=1
_

0
w

i=1
wi

l=0
_
Qw+i
i
_
1
_
Qw+i
i
_
_
w
i
__
w i
l
_
(1)
l

e
[PT(i+l)]/hfMQ
1
(2
2
X
)
1/2
a
nm
e
((lna
nm

X
)
2
/2
2
X
)
da
nm
(5.18)
=
w

i=1
wi

l=0
_
Qw+i
i
_
1
_
Qw+i
i
_
_
w
i
__
w i
l
_
(1)
l

_
_

0
e
[PT(i+l)]/hfMQ
1
(2
2
X
)
1/2
a
e
((lna
X
)
2
/2
2
X
)
da
_
MN
(5.19)
However, unlike the Rayleigh fading case, the resulting integral is not easily solved in closed
form. Instead, we can choose from a number of methods that still give reasonable accuracy,
including numerical integration, importance sampling, and Monte Carlo simulation.
The method chosen for this thesis is numerical integration. The log-normal density is sam-
pled for the values 0 to 10.0 in step sizes of 10
5
, and using a S.I. of 1.0. The resulting plots
can be seen in Figures 5.5 and 5.6.
Clearly, the log-normal fading case causes a degradation in system performance compared
to the non-fading case, although not as severe as with Rayleigh fading. Most notable are the
51
200 195 190 185 180 175 170 165 160 155 150 145
10
12
10
10
10
8
10
6
10
4
10
2
10
0
PT
b
,dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.5: Symbol error probability vs. average power with log-normal fading (S.I. = 1.0), no
background radiation, and Q = 8.
200 195 190 185 180 175 170 165 160 155 150 145
10
12
10
10
10
8
10
6
10
4
10
2
10
0
PT
b
,dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.6: Symbol error probability vs. peak power with log-normal fading (S.I. = 1.0), no
background radiation, and Q = 8.
52
effects of transmitter and receiver diversity. Just as in the Rayleigh fading case, a considerable
performance gain is also achievable by increasing only the number of lasers. However, the
increased aperture size resulting from adding receivers is also visible, showing that receiver
diversity is even more effective than transmitter diversity.
5.1.2 Error probability in the presence of background radiation
The analysis of error probability for the case of background radiation present is more com-
plicated than for no background radiation. In the no background case, errors only occurred
when an on slot was received as a zero at all N receivers. In the background case, the off
slots have the potential to have an equal or higher count than any of the on slots, in which case
the receiver could pick the wrong modulator symbol.
For this reason, a closed-form solution, even for the non-fading case is an intractable prob-
lem. For the non-fading case, an innite-summation expression is possible. For both fading
cases, however, some sort of Monte Carlo simulation is necessary.
5.1.2.1 Background radiation, no fading
To obtain the innite-summation solution we rst classify the possible symbol error types
into two categories. The rst are errors caused when one or more of the off slots has a higher
count than one or more of the on slots. In this case, the receiver will certainly make an error.
We will refer to this as a denite error and use the symbol P
def
. This scenario is depicted in
Figure 5.7.
We will start by analyzing denite errors. When the receiver receives a vector Zrepresenting
the Q slots summed over all N photodetectors, w of them were originally sent as on slots, and
Q w as off slots. Of the on slots, there will be i {1, . . . , w} of them that will have the
lowest slot count u. If j {1, . . . , Q w} of the off slots have a value (labeled v in Figure
5.7) which is greater than u, a denite error has occurred.
Letting {S
on,l
, l = 1, ..., w} denote the column sums for on slots, and {S
off,l
, l = 1, ..., Q
w} denote the column sums for off slots, we can state this for all possible values of u as
53
Figure 5.7: An example of a denite error.
P
def
=

u=0
P
_
(|{l : S
on,l
= u}| = i) (|{l : S
on,l
> u}| = w i) (5.20)
(|{l : S
off,l
> u}| = j) (|{l : S
off,l
u}| = Qw j)
_
(5.21)
=

u=0
w

i=1
_
w
i
_
P(S
on
= u)
i
P(S
on
> u)
wi

Qw

j=1
_
Qw
j
_
P(S
off
> u)
j
P(S
off
u)
Qwj
(5.22)
For simplicity, we will make the following symbolic substitutions regarding Poisson random
variables with parameter :

pdf
(k, ) = P(slot count = k|) =

k
e

k!
, k = 0, 1, 2... (5.23)

cdf
(k, ) = P(slot count k|) =
k

b=0

b
e

b!
, k = 0, 1, 2... (5.24)
(being careful to note that
cdf
and
pdf
are both equal to zero for k < 0). If we let
on
=
PT

m
a
2
nm
/hfM +NP
b
T/hf and
off
= NP
b
T/hf,
54
P
def
=

u=0
w

i=1
_
w
i
_

pdf
(u,
on
)
i
(1
cdf
(u,
on
))
wi

Qw

j=1
_
Qw
j
_
(1
cdf
(u,
off
))
j

cdf
(u,
off
)
Qwj
(5.25)
The second error type, which we will call an indenite error (P
indef
), happens when one or
more of the off slots has the same count as one or more of the lowest-valued on slots. We
make the distinction of lowest-valued, because an off slot tying with an on slot whose count
is higher than any of the other on slots will cause a denite error. Indenite errors (depicted in
Figure 5.8) cause symbol errors a fraction of the time, since the receiver makes a correct guess
at the modulator symbol at least some of the time. The probability of symbol error is the sum
of denite and indenite error probabilities.
Figure 5.8: An example of an indenite error (where the receiver chooses incorrectly from the
3 possible modulator symbols).
Similar to the development for denite errors, we can dene an expression for indenite
errors. Here, we are concerned about g off slots with count u when the i lowest-valued on
slots also have a count of u. This is expressed as
55
P
indef
=

u=0
_
g+i
i
_
1
_
g+i
i
_ P
_
(|{l : S
on,l
= u}| = i) (|{l : S
on,l
> u}| = w i)
(|{l : S
off,l
= u}| = g) (|{l : S
on,l
< u}| = Qw g)
_
(5.26)
=

u=0
w

i=1
_
w
i
_

pdf
(u,
on
)
i
(1
cdf
(u,
on
))
wi

Qw

g=1
_
g+i
i
_
1
_
g+i
i
_
_
Qw
g
_

pdf
(u,
off
)
g

cdf
(u 1,
off
)
Qwg
(5.27)
where the
__
g+i
i
_
1
_
/
_
g+i
i
_
term represents the probability of the receiver choosing the wrong
symbol if i on slots have the same count as g off slots, and the remaining w i slots have
the highest counts. The full expression is then
P
s
= P
def
+P
indef
(5.28)
A reasonable numerical approximation to this can be obtained from any mathematical soft-
ware package. We take advantage of the built-in Matlab functions poisspdf and poisscdf which
correspond directly to what we dened as
pdf
and
cdf
respectively. To avoid an innite sum-
mation, feedback can be used in the summation-loop to determine when the probability of error
has asymptotically approached the nal value to within a certain threshold (1% of the nal value
in plots generated for this thesis).
This advantage of this method over the Monte Carlo approach is reasonable accuracy for
error probabilities as low as 10
12
without the need for large numbers of trials. The result of
this simulation is shown in Figures 5.9 and 5.10.
Looking at the plots, we again notice that there is no advantage to increasing M, since this
is a non-fading situation. The advantage to increasing the number of receivers is caused by the
increased aperture size at the receiver.
It is interesting to note that with 170 dBJ of background radiation present, the advantage
for having N = 4 over N = 1 drops from 6 dB in the no background case to 4 dB, showing that
background radiation cuts into the advantage provided by increasing the receiver aperture size.
Also, comparing Figures 5.9 and 5.10 to Figures 5.1 and 5.2 respectively, background radiation
shifts the curves to the right by amounts ranging from 4 to 7 dB, and causes them to drop off
56
210 200 190 180 170 160 150
10
12
10
10
10
8
10
6
10
4
10
2
10
0
PT
b
,dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.9: Symbol error probability vs. average power with no fading, P
b
T
b
= 170 dbJ, and
Q = 8 dashed and solid lines overlap.
210 200 190 180 170 160 150
10
12
10
10
10
8
10
6
10
4
10
2
10
0
PT
b
,dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.10: Symbol error probability vs. peak power with no fading, P
b
T
b
= 170 dbJ, and
Q = 8 dashed and solid lines overlap.
57
more steeply than in the no background radiation case. This suggests a signal power threshold
for reliable communications caused by the presence of background radiation.
5.1.2.2 Background radiation, Rayleigh fading
When background radiation and fading are present, the innite summation expression from
the previous subsection could be combined with a Monte Carlo approach to obtain the desired
error probability. A large number of fading variables could be drawn according to a fading
distribution, plugged-in to (5.28), and averaged to achieve accurate results for values of the
error probability much lower than possible with a pure Monte Carlo approach. However, this
is a prohibitively slow process due to the large number of fading variables combined with the
innite summation.
For results in the two fading cases, we resort to simple Monte Carlo simulation, generating a
large number of Poisson and Rayleigh (or log-normal) fading variables, and dividing the number
of errors that occur by the number of trials.
As Figures 5.11 and 5.12 show, the system achieves diversity equal to M N (measurable
for M N 4) in the presence of Rayleigh fading, even with background present.
210 200 190 180 170 160 150
10
6
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
, dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.11: Symbol error probability vs. average power with Rayleigh fading, P
b
T
b
= 170
dbJ, and Q = 8.
58
210 200 190 180 170 160 150
10
6
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
, dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.12: Symbol error probability vs. peak power with Rayleigh fading, P
b
T
b
= 170 dbJ,
and Q = 8.
5.1.2.3 Background radiation, log-normal fading
Results for the log-normal fading case were also obtained using Monte Carlo simulation.
Unfortunately, due to the large number of trials that would be necessary, the results are only
accurate to symbol probabilities of 10
5
.
Once again, we notice the advantage to transmitter and receiver diversity, with the greatest
advantage being gained at the receiver due to the increased aperture size.
5.2 Error probability in the Gaussian regime
We will now investigate the probability of error in the Gaussian regime. In this regime, we
are dealing with continuous random variables, which changes our analysis to a certain degree
from the Poisson regime.
Again, we will assume equal gain combining, such that our detector will look for the w slots
with the highest observable,

n
Z
nq
= S
q
. An error occurs if the highest-valued off slot has
a higher count rate than the lowest-valued on slot. We will ignore tie-break situations, since
we are dealing with continuous random variables where ties are zero-probability events.
Assuming that modulator symbol X
i
is sent, we dene
59
210 200 190 180 170 160 150
10
6
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.13: Symbol error probability vs. average power with Log-normal fading, S.I. = 1.0,
P
b
T
b
= 170 dbJ, and Q = 8.
210 200 190 180 170 160 150
10
6
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.14: Symbol error probability vs. peak power with Log-normal fading, S.I. = 1.0,
P
b
T
b
= 170 dbJ, and Q = 8.
60
U = min
qQ
i
on
(S
q
) (5.29)
V = max
qQ
i
off
(S
q
) (5.30)
where S
q
=

N
n=1
Z
nq
. Therefore, the conditional probability of error is
P
s|A
= P(V > U|A) (5.31)
If we write the joint density function of U and V as f(u, v), we can state that
P(V > U) =
_

__

u
f
UV
(u, v)dv
_
du (5.32)
Since U and V are independent, this simplies to
P(V > U) =
_

_

u
f
U
(u)f
V
(v)dvdu
=
_

f
U
(u)
__

u
f
V
(v)dv
_
du (5.33)
If we state the c.d.f. of V as F
V
(u) = P(V u) =
_
u

f
V
(u)du,
then
[1 F
V
(u)] =
_

u
f
V
(v)dv (5.34)
and
P(V > U) =
_

f
U
(u) [1 F
V
(u)] du (5.35)
We are rst interested in determining the density f
U
(u). We do so by rst looking at the
c.d.f., F
U
(u), whose derivative will give us the desired p.d.f. Similar to the development in
Papoulis [40], we can write F
U
(u) = P(S
1
= s
1
u

S
2
= s
2
u

...

S
w
= s
w
u),
where S
1
...S
w
are the column sums in the w on slots.
When w = 2, for example, F
U
(u) can be expressed as
F
U
(u) = F
S
1
(u) +F
S
2
(u) F
S
1
S
2
(u, u) (5.36)
which is recognized as the sum of the individual c.d.f.s minus the joint c.d.f.
61
To convert this to a p.d.f. as required in (5.35), we simply take the derivative of the respective
c.d.f.s. For the last term, the chain rule is required, giving
f
U
(u) =
dF
U
(u)
dv
= f
S
1
(u) + f
S
2
(u) [F
S
1
(u)f
S
2
(u) + F
S
2
(u)f
S
1
(u)] (5.37)
= f
S
1
(u) [1 F
S
1
(u)] + f
S
2
(u) [1 F
S
2
(u)] (5.38)
Since S
1
and S
2
are i.i.d., both being in the set of on slots, we can write them as S
on
instead, and simplify (5.38):
f
U
(u) = 2f
S
on
(u) [1 F
S
on
(u)] (5.39)
Extending this to arbitrary w results in
f
U
(u) = wf
S
on
(u)
_
1 F
S
on
(u)
w1

(5.40)
Since we are dealing with Gaussian density functions, the F
S
on
(u)
w1
term can be expressed
in terms of a Qfunction, and then rewritten using the binomial expansion as
F
S
on
(u)
w1
= P(S
on
u)
w1
(5.41)
= [1 Q
S
on
(u)]
w1
(5.42)
=
w1

i=0
_
w 1
i
_
(1)
i
Q
S
on
(u)
i
(5.43)
By isolating the i = 0 term in the summation, we can pull out a 1, and eliminate the 1-
from the expression for f
U
(u):
f
U
(u) = wf
S
on
(u)
_
1 F
S
on
(u)
w1

(5.44)
= wf
S
on
(u)
_
1
w1

i=0
(1)
i
_
w 1
i
_
Q
S
on
(u)
i
_
(5.45)
= wf
S
on
(u)
_
1 1
w1

i=1
(1)
i
_
w 1
i
_
Q
S
on
(u)
i
_
(5.46)
= wf
S
on
(u)
_

w1

i=1
(1)
i
_
w 1
i
_
Q
S
on
(u)
i
_
(5.47)
62
Now focusing our attention on the 1 F
V
(u) term in (5.35), it can be rewritten as
1 F
V
(u) = 1 F
S
b
(u)
Qw
= 1 [1 Q
S
b
(u)]
Qw
(5.48)
where S
b
is the column sum at an off slot receiving only background radiation, if present. The
Qw in the exponent is due to each of the Qw off slots being independent. The right-most
term can be rewritten by once again using the binomial expansion formula:
1 [1 Q
S
b
(u)]
Qw
= 1
Qw

l=0
(1)
l
_
Qw
l
_
Q
S
b
(u)
l
(5.49)
To eliminate the 1- term the l = 0 term is expressed explicitly:
1 [1 Q
S
b
(u)]
Qw
= 1
Qw

l=0
(1)
l
_
Qw
l
_
Q
S
b
(u)
l
(5.50)
= 1 1
Qw

l=1
(1)
l
_
Qw
l
_
Q
S
b
(u)
l
(5.51)
=
Qw

l=1
(1)
l
_
Qw
l
_
Q
S
b
(u)
l
(5.52)
Putting everything together, we get
P
s
=
_

wf
S
on
(u)
_

w1

i=1
(1)
i
_
w 1
i
_
Q
S
on
(u)
i
_

Qw

l=1
(1)
l
_
Qw
l
_
Q
S
b
(u)
l
_
du (5.53)
which is a general expression for the probability of symbol error for all of the cases in the
Gaussian regime. Implicit in this expression is that the mean, variance, background noise,
number of lasers, number of receivers, number of on and off slots, and fading will all affect
the probability of symbol error through their respective Qfunctions.
5.2.0.4 No fading
For the non-fading case, we can simply take (5.53) and perform a numerical integration over
a large range of values for u large enough to thoroughly encompass the tail regions of the on
and off Gaussian p.d.f.s.
63
Values were chosen for the load resistor, R, the system temperature, T
0
, and the bit rate R
b
and the integral was performed, yielding Figures 5.15 and 5.16. The curves exhibit the same
general shape as in the Poisson regime
1
. Again, the multipulse system outperforms the single
pulse system only in the peak-power-limited system. The effect of N = 4 is also observed as
approximately 6 dB of improved performance.
90 85 80 75 70 65 60 55 50
10
12
10
10
10
8
10
6
10
4
10
2
10
0
P
ave
,dBW
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.15: Symbol error probability vs. average power with no fading, Q = 8, R = 100 ,
T
0
= 290 K, P
b
T
b
= 170 dbJ, and R
b
= 100 Mbps.
5.2.0.5 Rayleigh fading
With Rayleigh fading, the integral would need to be averaged M N times for all possible
fading paths from transmitter to receiver. This is prohibitively difcult, and instead, a simple
Monte Carlo approach using equal gain combining was chosen. As can be seen from the slopes
of the curves in Figures 5.17 and 5.18, full transmitter and receiver diversity is achieved in the
Gaussian regime as well.
Optimal gain combining, using the ML detector developed in Chapter 4 gives only slightly
improved results for faded channels where N > 1. This is shown for selected cases with
1
Note here, however, that the error probability is plotted versus received power P in dBW, as opposed to PT
b
in dBJ for the Poisson regime.
64
90 85 80 75 70 65 60 55 50
10
12
10
10
10
8
10
6
10
4
10
2
10
0
P
peak
,dBW
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.16: Symbol error probability vs. peak power with no fading, Q = 8, R = 100 ,
T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps.
90 85 80 75 70 65 60 55 50 45 40
10
6
10
5
10
4
10
3
10
2
10
1
10
0
P
ave
,dBW,ECG
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.17: Symbol error probability vs. average power with Rayleigh fading, Q = 8, R = 100
, T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps.
65
90 85 80 75 70 65 60 55 50 45 40
10
6
10
5
10
4
10
3
10
2
10
1
10
0
P
peak
,dBW,EGC
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.18: Symbol error probability vs. peak power with Rayleigh fading, Q = 8, R = 100
, T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps.
Rayleigh fading in Figure 5.22.
5.2.0.6 Log-normal fading
Again, with log-normal fading, Monte Carlo simulation was chosen, and the results are
shown belowin Figures 5.20 and 5.21. The MIMOsystemclearly exhibits superior performance
to the SISO system, as we have come to see.
Figure 5.22 shows that the optimal detector only slightly outperforms equal gain combining
in the log-normal fading case.
66
90 85 80 75 70 65 60 55 50 45 40
10
6
10
5
10
4
10
3
10
2
10
1
10
0
P
,
dBW
P
s
M=1,N=1,w=1, P
ave
, EGC
M=1,N=1,w=1, P
ave
, OGC
M=1,N=4,w=1, P
peak
, EGC
M=1,N=4,w=1, P
peak
, OGC
M=4,N=1,w=4, P
ave
, EGC
M=4,N=1,w=4, P
ave
, OGC
M=4,N=4,w=4, P
peak
, EGC
M=4,N=4,w=4, P
peak
, OGC
Figure 5.19: Symbol error probability vs. signal power with Rayleigh fading using equal gain
combining (EGC) or optimal gain combining (OGC). Q = 8, R = 100 , T
0
= 290 K,
P
b
= 90 dBW, and R
b
= 100 Mbps.
90 85 80 75 70 65 60 55 50 45 40
10
6
10
5
10
4
10
3
10
2
10
1
10
0
P
ave
,dBW
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.20: Symbol error probability vs. average power with log-normal fading, S.I. = 1.0,
Q = 8, R = 100 , T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps.
67
90 85 80 75 70 65 60 55 50 45 40
10
6
10
5
10
4
10
3
10
2
10
1
10
0
P
peak
,dBW
P
s
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 5.21: Symbol error probability vs. peak power with log-normal fading, S.I. = 1.0,
Q = 8, R = 100 , T
0
= 290 K, P
b
= 90 dBW, and R
b
= 100 Mbps.
90 85 80 75 70 65 60 55 50 45 40
10
6
10
5
10
4
10
3
10
2
10
1
10
0
P
,
dBW
P
s
M=1,N=1,w=1, P
ave
, EGC
M=1,N=1,w=1, P
ave
, OGC
M=1,N=4,w=1, P
peak
, EGC
M=1,N=4,w=1, P
peak
, OGC
M=4,N=1,w=4, P
ave
, EGC
M=4,N=1,w=4, P
ave
, OGC
M=4,N=4,w=4, P
peak
, EGC
M=4,N=4,w=4, P
peak
, OGC
Figure 5.22: Symbol error probability vs. signal power with log-normal fading using equal
gain combining (EGC) or optimal gain combining (OGC). Q = 8, R = 100 , T
0
= 290 K,
P
b
= 90 dBW, and R
b
= 100 Mbps.
Chapter 6
Capacity of the MIMO FSO system using
MPPM in the Poisson regime
In this chapter, we investigate the information-theoretic capacity for the MIMO MPPM FSO
channel. This is broken down into different cases using the same assumptions about the system
in the Poisson regime as before in Chapter 5.
6.1 No background radiation
For the channel capacity of the multipulse case, we rst start by considering the mutual
information between Z, the random array of counts for the Q slots, and the input symbol X:
I(X; Z) = H(X) H(X|Z) (6.1)
For each modulator symbol X
j
, w slots will be on slots, and Q w will be off. There is
a probability that all N of the detectors will generate zero photoelectrons for i of the w slots,
in which case the detector must choose from
_
Qw+i
i
_
equally probable symbols. Since we are
assuming no background radiation, uncertainty exists only when i > 0 of the w column sums
for on slots equal zero at the detector.
Since this is a symmetric channel, mutual information is maximized when the input symbols
are equiprobable, and the capacity is given by
68
69
C = H(X)
w

i=1
H(X|i out of w column counts = 0)
P(i out of w column counts = 0) (6.2)
= log
2
_
Q
w
_

i=1
log
2
_
Qw +i
i
_
P(i out of w column counts = 0) (6.3)
We repeat the argument developed in Chapter 5, that with equal gain combining, the sum
over all receivers for each on slot is a Poisson random variable with a parameter of
on
=
PT
hfM

N
n=1

M
m=1
a
2
nm
, and an on slot will have zero photoelectrons with probability P(S
q
=
0) = p = e

on
.
Treating each slot count as an independent random variable, we get
P(i out of w column counts = 0) =
w

i=1
_
w
i
_
p
i
(1 p)
wi
(6.4)
and thus
H(X|Z) =
w

i=1
log
2
_
Qw +i
i
__
w
i
_
p
i
(1 p)
wi
(6.5)
The last term can be expanded using the binomial expansion formula, which after simpli-
cation gives
H(X|Z) =
w

i=1
wi

l=0
(1)
l
log
2
_
Qw +i
i
__
w
i
__
w i
l
_
p
i+l
(6.6)
Replacing p gives the general formula for capacity. Observe that the capacity, C depends on
fading, A:
C
A
= log
2
_
Q
w
_

i=1
wi

l=0
(1)
l
log
2
_
Qw +i
i
_

_
w
i
__
w i
l
_
e
PT

M
m=1

N
n=1
a
2
nm
(i+l)/hfM
(6.7)
There are two different metrics we will use when analyzing the capacity of this system: er-
godic capacity and outage probability. The ergodic capacity is simply the time average capacity
of a channel, whereas the outage probability is simply the probability that a single realization
of the channel will cause the instantaneous capacity to drop below some threshold (one-half of
70
the maximum attainable capacity in our analysis). If a system designer is coding at a rate equal
to this threshold, and the instantaneous capacity is below the threshold, then it is not possible to
drive the error probability arbitrarily low during that instant.
The ergodic capacity, averaged over the fading distribution f
A
(a
nm
) (which we will do in
the following sections), can be expressed as such:
E
A
[C] = log
2
_
Q
w
_

w

i=1
wi

l=0
(1)
l
log
2
_
Qw +i
i
__
w
i
__
w i
l
_

m=1
N

n=1
_

0
f
A
(a
nm
)e
PTa
2
nm
(i+l)/hfM
da
nm
(6.8)
Outage capacity is computed differently for the various cases, and will be discussed in each
individual section.
6.1.1 No background, no fading
In the non-fading case,

M
m=1

N
n=1
a
2
nm
= MN, and ergodic capacity can be expressed as
E
A
[C] = log
2
_
Q
w
_

i=1
wi

l=0
(1)
l
log
2
_
Qw +i
i
__
w
i
__
w i
l
_
e
PTN(i+l)/hf
(6.9)
Here, just as with the error analysis, there is no advantage to increasing the number of
lasers in the absence of fading. However, performance improves with increasing N, due to the
increase in the overall aperture size. The ergodic capacity is shown in Figures 6.1 and 6.2. For
consistency with the analysis and plots developed in Chapter 5, this is plotted for a Q = 8
system with different values of M, N, and w all in {1,4}.
In the plots for ergodic capacity, we can notice that at lowpower, the capacity asymptotically
approaches zero, and as power is increased, the expected capacity for the channel asymptotically
approaches log
2
_
Q
w
_
, which is 3.00 and approximately 6.13 bits per channel use for the (Q, w) =
(8, 1) and (8, 4) systems, respectively.
For the systemto experience an outage as we are dening it, its instantaneous capacity would
have to drop below the coding rate, which we are allowing to be 0.5 log
2
_
Q
w
_
. For the non-fading
case,

M
m=1

N
n=1
a
2
nm
= MN which removes the conditioning from (6.8). In fact, capacity is
deterministic in this case, and is equivalent to (6.9). Therefore, we can look at Figures 6.1 and
71
220 215 210 205 200 195 190 185 180 175 170
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,

b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.1: Ergodic capacity vs. average power with no fading, no background radiation, and
Q = 8.
220 215 210 205 200 195 190 185 180 175 170
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,

b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.2: Ergodic capacity vs. peak power with no fading, no background radiation, and
Q = 8.
72
6.2 and state that the outage probability curve is simply an inverted unit step function, equal to
one for power levels when the capacity is below 0.5 log
2
_
Q
w
_
and equal to zero for power levels
when capacity is above 0.5 log
2
_
Q
w
_
. This is shown in Figures 6.3 and 6.4. One can note the 6
dB advantage to having N = 4 versus N = 1, due to increased aperture size.
220 215 210 205 200 195 190 185 180 175 170
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.3: 50% outage probability vs. average power with no fading, no background radiation,
and Q = 8.
6.1.2 No background, Rayleigh fading
To nd the ergodic capacity with Rayleigh fading, we let f
A
(a
nm
) be the Rayleigh p.d.f. and
perform the M N integrations over all values of a
nm
:
E
A
[C] = log
2
_
Q
w
_

i=1
wi

l=0
log
2
_
Qw +i
i
__
w
i
__
w i
l
_
(1)
l

n=1
M

m=1
_

e
P
r
T

M
m=1

N
n=1
a
2
nm
(i+l)/hfMQ
2a
nm
e
a
2
nm
da
nm
(6.10)
Just as in the development for Chapter 5, we assume the a
nm
terms are all independent, so
we can replace them with a generic fading variable a, and rewrite in the following form:
73
220 215 210 205 200 195 190 185 180 175 170
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.4: 50% outage probability vs. peak power with Rayleigh fading, no background radi-
ation, and Q = 8.
E
A
[C] = log
2
_
Q
w
_

i=1
wi

l=0
log
2
_
Qw +i
i
__
w
i
__
w i
l
_

(1)
l
__

0
2ae
[PTa
2
(i+l)+1]/hfM
da
_
MN
(6.11)
This then integrates to give the following closed-form expression for the ergodic capacity
with respect to Rayleigh fading:
E
A
[C] = log
2
_
Q
w
_

i=1
wi

l=0
log
2
_
Qw +i
i
__
w
i
__
w i
l
_

(1)
l
_
1
1 + PT(i +l)/hfM
_
MN
(6.12)
This is plotted in Figures 6.5 and 6.6. As can be seen from the plots, transmitter and receiver
diversity are benecial in a fading environment. It is also interesting to note that while the
(M, N) = (4, 1) system shows a superior performance to the (1, 1) system, the (4, 4) system
is only slightly better than the (1, 4) system. This suggests that diversity is very helpful in a
fading environment to achieve a high capacity, but that adding diversity at both ends of the link
74
yields a diminishing return in terms of channel capacity. Just as in the performance analysis, the
plots also indicate that if one were choosing between multiple transmitters or multiple receivers,
it would be more advantageous to use the extra resources at the receiver, where the increased
aperture size can be exploited (assuming the total power at the transmitter array is xed with
respect to the number of transmitters).
220 215 210 205 200 195 190 185 180 175 170
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.5: Ergodic capacity vs. average power with Rayleigh fading, no background radiation,
and Q = 8.
The outage probability curves for the Rayleigh fading case, shown in Figures 6.7 and 6.8, are
important in that they once again show full transmitter and receiver diversity in their respective
slopes.
6.1.3 No background, log-normal fading
For log-normal fading, the ergodic capacity analysis is remarkably similar to the Rayleigh
fading case. We replace the log-normal fading distribution, (2.1) for f
A
(a
nm
) in (6.8), and
perform the M N-fold integration. Unfortunately, when assessing the ergodic capacity, the
closed-form solution is not as easily found as in the Rayleigh case. However, just as with
Section 5.1.1.3, we can solve the integral numerically to obtain the desired results. The result
of this procedure is shown in Figures 6.9 and 6.10.
75
220 215 210 205 200 195 190 185 180 175 170
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.6: Ergodic capacity vs. peak power with Rayleigh fading, no background radiation,
and Q = 8.
195 190 185 180 175 170 165 160 155
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.7: 50% outage probability vs. average power with Rayleigh fading, no background
radiation, and Q = 8.
76
190 185 180 175 170 165 160 155
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.8: 50% outage probability vs. peak power with Rayleigh fading, no background radi-
ation, and Q = 8.
220 215 210 205 200 195 190 185 180 175 170
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.9: Ergodic capacity vs. average power with log-normal fading, no background radia-
tion, Q = 8, and S.I. = 1.0.
77
220 215 210 205 200 195 190 185 180 175 170
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.10: Ergodic capacity vs. peak power with log-normal fading, no background radiation,
Q = 8, and S.I. = 1.0.
Once again, we can see that transmitter and receiver diversity is advantageous with respect
to ergodic capacity, with the same general result as with Rayleigh fading: Diversity can be
achieved on both ends of the FSO link, but increasing the number of transmitters alone provides
more improvement in ergodic capacity than increasing the number of receivers alone, due to the
increased aperture size.
The outage probability plots showthe very important result that both transmitter and receiver
diversity is achievable in the presence of log-normal fading. A system designer using a MIMO
FSO system, and employing a rate R = 0.5 log
2
_
Q
w
_
code could reduce the probability of outage
to a desired level with far less power than with a comparable SISO system.
6.2 Background radiation and no fading
When doing an analysis on the channel capacity with background radiation present, the
problem is complicated even further. In the situation with no background, uncertainty in X
could only occur if i of the on slots at the transmitter are received as off slots at the receiver.
In the background radiation case, uncertainty exists to a certain extent in every received symbol,
78
195 190 185 180 175 170 165 160
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.11: 50% outage probability vs. average power with log-normal fading, no background
radiation, and Q = 8.
190 185 180 175 170 165 160 155
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.12: 50% outage probability vs. peak power with log-normal fading, no background
radiation, and Q = 8.
79
since each of the column sums has the potential of being non-zero, regardless of which symbol
was sent.
Again, the channel is symmetric, so we can make all symbol probabilities equal to achieve
capacity.
We rst use an alternative, but equivalent, denition for mutual information:
I(X; Z) =

z
p(x)p(z|x) log
2
_
p(z|x)
p(z)
_
(6.13)
By symmetry of the channel, we can rewrite the conditional probability of z for any of
the modulator symbols, X = x
0
for example. Therefore, we can eliminate the summation
(averaging) over all of the symbols, giving
I(X; Z) =

z
p(z|x
0
) log
2
_
p(z|x
0
)
p(z)
_
(6.14)
Upon inspection, this is simply the formula for
E
Z|X=x
0
_
log
2
_
p(z|x
0
)
p(z)
__
(6.15)
Therefore, to evaluate the capacity of the channel with background radiation, we can per-
form a Monte-Carlo simulation for log
2
_
p(z|x
0
)
p(z)
_
and use the following equation:
E[C] = E
Z|X=x
0
_
log
2
_
p(z|x
0
)
p(z)
__
(6.16)
To perform the simulation, it is necessary to rst look at the probability distribution for Z
given x
0
. If we dene Q
0
to be the set of on slots in modulator symbol x
0
, and Q
C
0
to be the
complement of that set signifying the off slots, we can state that for symbol x
0
, the probability
distribution is
p(Z|X = x
0
) =
N

n=1
_
_

qQ
O
(
on,n
+
off
)
Z
nq
e
(
on,n
+
off
)
Z
nq
!

qQ
C
O
(
off
)
Z
nq
e
(
off
)
Z
nq
!
_
_
(6.17)
Similarly, the distribution for Z given any equiprobable symbol x
i
with weight w is
p(Z) =

allx
i
1
_
Q
w
_P(Z|X = x
i
) (6.18)
80
p(Z) =
1
_
Q
w
_

allx
i

N
n=1
_

qQ
i
(
on,n
+
off
)
Z
nq
e
W
on,n

qQ
C
i

Z
nq
off
e
Q
off
_

allq
Z
nq
!
(6.19)
Therefore,
p(Z)
p(Z|X = x
0
)
=
1
_
Q
w
_

allx
i

N
n=1
_

qQ
i
(
on,n
+
off
)
Z
nq

qQ
C
i

Z
nq
off
_

N
n=1
_

qQ
O
(
on,n
+
off
)
Z
nq

qQ
C
O

Z
nq
off
_ (6.20)
or equivalently,
p(Z)
p(Z|X = x
0
)
=
1
_
Q
w
_

allx
i

N
n=1
_
(
on,n
+
off
)

qQ
i
Z
nq
(
off
)
(

all q
Z
nq

qQ
i
Z
nq)
_

N
n=1
_
(
on,n
+
off
)

qQ
0
Z
nq
(
off
)
(

all q
Z
nq

qQ
0
Z
nq)
_
(6.21)
We can dene a new parameter
n
= (
on,n
+
off
)/
off
and cancel like terms, to give
p(Z)
p(Z|X = x
0
)
=
1
_
Q
w
_

allx
i
N

n=1
_
(
n
)
(

qQ
i
Z
nq

qQ
0
Z
nq)
_
(6.22)
Plugging this into the ergodic capacity expression gives
E[C] = E
Z|X=x
0
_
log
2
_
p(z|X = x
0
)
p(z)
__
(6.23)
= E
Z|X=x
0
_
log
2
_
Q
w
_
log
2

allx
i
N

n=1
_
(
n
)
(

qQ
i
Z
nq

qQ
0
Z
nq)
_
_
(6.24)
= log
2
_
Q
w
_
E
Z|X=x
0
_
log
2

allx
i
N

n=1
_
(
n
)
(

qQ
i
Z
nq

qQ
0
Z
nq)
_
_
(6.25)
From this, it becomes easy to devise a programming strategy for simulating the terms in
(6.22) for each trial.
6.2.1 Background radiation, no fading
Figures 6.13 and 6.14 show the ergodic capacity for the non-fading case when background
radiation is present. As expected, increasing M shows no improvement in ergodic capacity, but
increasing N does, just as with the no background case.
81
200 195 190 185 180 175 170 165 160
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.13: Ergodic capacity vs. average power with no fading, Q = 8, and P
b
T = 170 dBJ.
200 195 190 185 180 175 170 165 160
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.14: Ergodic capacity vs. peak power with no fading, Q = 8, and P
b
T = 170 dBJ.
82
Unlike the no background case, the outage probability of the background and no fading case
is no longer a deterministic function. Instantaneous capacity can drop below the threshold due
to the probability of not receiving an on pulse in an on slot, and also due to off slots having
non-zero counts. This gives the recognizable waterfall curves in Figures 6.15 and 6.16. Again,
the advantage is clearly in the N = 4 system, independent of M.
195 190 185 180 175 170 165 160 155
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.15: 50% outage probability vs. average power with no fading, Q = 8, and P
b
T =
170 dbJ.
6.2.2 Background radiation, Rayleigh fading
Figures 6.17 and 6.18 show the ergodic capacity when Rayleigh fading and background
radiation are both present. Again, we notice the gains attributable to transmitter and receiver
diversity, as well as to the increase in aperture size at the receiver, showing the superiority of
MIMO systems.
When Rayleigh fading and background radiation are present, the outage probability plots
are similar in appearance to the corresponding symbol error probability plots from Chapter 5.
Here, we again see evidence of full transmitter and receiver diversity when looking at the slopes
of the lines for the different systems.
83
195 190 185 180 175 170 165 160 155
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.16: 50% outage probability vs. peak power with no fading, Q = 8, and P
b
T = 170
dBJ.
200 195 190 185 180 175 170 165 160
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.17: Ergodic capacity vs. average power with Rayleigh fading, Q = 8, and P
b
T =
170 dBJ.
84
200 195 190 185 180 175 170 165 160
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.18: Ergodic capacity vs. peak power with Rayleigh fading, Q = 8, and P
b
T = 170
dBJ.
195 190 185 180 175 170 165 160 155
10
6
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.19: 50% outage probability vs. average power with Rayleigh fading, Q = 8, and
P
b
T = 170 dbJ.
85
195 190 185 180 175 170 165 160 155
10
6
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.20: 50% outage probability vs. peak power with Rayleigh fading, Q = 8, and P
b
T =
170 dBJ.
6.2.3 Background radiation, log-normal fading
For log-normal fading with background radiation, the ergodic capacity and outage probabil-
ity curves are once again indicative of the strength of the MIMO system over its SISO counter-
part. We see a diversity advantage of increasing the number of transmitters and receivers.
86
200 195 190 185 180 175 170 165 160
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.21: Ergodic capacity vs. average power with log-normal fading, Q = 8, and P
b
T =
170 dBJ.
200 195 190 185 180 175 170 165 160
0
1
2
3
4
5
6
7
PT
b
,dBJ
E
[
C
]
,
b
i
t
s
/
c
h
a
n
n
e
l

u
s
e
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.22: Ergodic capacity vs. peak power with log-normal fading, Q = 8, and P
b
T = 170
dBJ.
87
195 190 185 180 175 170 165 160 155
10
6
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.23: 50% outage probability vs. average power with log-normal fading, Q = 8, and
P
b
T = 170 dbJ.
195 190 185 180 175 170 165 160 155
10
6
10
5
10
4
10
3
10
2
10
1
10
0
PT
b
,dBJ
P
o
u
t
M=1,N=1,w=1
M=4,N=1,w=1
M=1,N=4,w=1
M=4,N=4,w=1
M=1,N=1,w=4
M=4,N=1,w=4
M=1,N=4,w=4
M=4,N=4,w=4
Figure 6.24: 50% outage probability vs. peak power with log-normal fading, Q = 8, and
P
b
T = 170 dBJ.
Chapter 7
Conclusions
In this thesis, we have analyzed MIMO MPPM FSO systems and developed the ML de-
tection schemes for the Poisson and Gaussian regimes. For the Poisson regime, we see that
the optimal detector in three of the four cases simply combines the received signal from all N
receivers equally, and chooses the symbol with the w highest slot counts. In the case where
background radiation and fading are both present, the optimal detector maximizes a weighted
sum over the on and off slots, but as we saw from previous research, equal gain combining
achieves essentially equal performance. For the Gaussian regime, the ML detector is also equal
gain combining for all but one scenario: when the system is thermal noise limited and fading is
present.
We have shown that MPPM used with a MIMO FSO system is a robust and effective way
to mitigate the negative effects of atmospheric turbulence. In all but the case where there is no
fading, increasing the number of lasers (MISO) resulted in better error performance, with the
ergodic capacity approaching its maximum value at lower power levels, and a steeper outage
probability, when compared to a SISO system.
In all cases, including non-fading, we demonstrated that increasing the number of receivers
(SIMO) resulted in similar improvements in system performance compared with the MISO sys-
tem. When compared to a MISO system, the SIMO system displayed an overall improvement
in all cases due to the increased aperture size at the receiver.
The MIMO system performed the best in all of the fading cases, and equivalent to SIMO
in the non-fading case. In the case of Rayleigh fading, we were able to ascertain from the
symbol error probability and outage probability plots, that the system achieves full transmitter
88
89
and receiver diversity. The improvements in symbol error probability seen in the Poisson regime
were echoed in the Gaussian regime, with full transmitter diversity being observable in the
presence of Rayleigh fading.
The use of MPPM was a good choice for this system. In the case where average power is
limited, traditional PPM (w = 1) exhibited better error performance than MPPM (w > 1) due
to the distribution of signal power across the w on slots. However, in the case where peak power
is limited, MPPM is the clear winner. MPPM also has the advantage of being more bandwidth
efcient for increasing w until w =
Q
2
, consuming less unit bandwidth per given bit rate.
There are many possibilities for future work on this topic. Implementation of a decoder
for modulator symbol sets where
_
Q
w
_
is not a power of 2 is a particularly interesting design
challenge which should be addressed. Also, study of different channel coding schemes applied
to the MPPM system, with the goal of achieving coding gain on the channel, is of interest and
should be investigated further.
Bibliography
[1] R. M. Gagliardi and S. Karp, Optical Communications. John Wiley and Sons, second ed., 1995.
[2] H. Willebrand and B. Ghuman, Fiber optics without the ber, IEEE Spectrum, August 2001.
[3] A. Acampora, Last mile by laser, Scientic American, June 17 2003.
[4] S. B. Alexander, Optical Communication Receiver Design, vol. T22. SPIE Optical Engineering
Press, 1997.
[5] R. Dettmer, A ray of light [free space optical transmission], IEE Review, vol. 47, pp. 3233,
March 2001.
[6] D. Begley, Free-space laser communications: a historical perspective, Lasers and Electro-Optics
Society, vol. 2, pp. 391392, 2002.
[7] H. Hemmati, Free-space optical communications program at JPL, IEEE Lasers and Electro-
Optics Society 1999 12th Annual Meeting. LEOS 99, vol. 1, pp. 106107, November 1999.
[8] H. Hemmati, Status of free-space optical communications program at JPL, 2000 IEEE Aerospace
Conference Proceedings, vol. 3, pp. 101105, March 2000.
[9] I. S. C. Davis and S. Milner, Flexible optical wireless links and networks, IEEE Communications
Magazine, vol. 41, pp. 5157, March 2003.
[10] S. Wilson, Digital Modulation and Coding. Prentice-Hall, 1996.
[11] T. Cover and J. Thomas, Elements of Information Theory. New York: Wiley, 1991.
[12] X. Zhu and J. Kahn, Free-space optical communication through atmospheric turbulence channels,
IEEE Trans. on Communications, no. 2, pp. 12931300, 2003.
[13] X. Zhu and J. Kahn, Communication techniques to mitigate atmospheric turbulence in free-space
optical links, The 16th Annual Meeting of the IEEE Lasers and Electro-Optics Society, 2003. LEOS
2003, vol. 1, pp. 8990, October 2003.
[14] J. Alwan, Eye safety and wireless optical networks, Website of AirFiber, pp. 110, April 2001.
90
91
[15] E. Shin and V. Chan, Optical communication over the turbulent atmospheric channel using spatial
diversity, IEEE Conference on Communication, pp. 20552060, 2002.
[16] S. Bloom, The physics of free-space optics, Website of AirFiber, pp. 122, December 2001.
[17] Anonymous, Link range - how far will FSO work?. [2004 July 22] Available at http://www.
freespaceoptics.com/Free Space Optics Link Range.html.
[18] Z. Jaksic and Z. Djuric, Extraction photodiodes with auger suppression for all-weather free-space
optical communication, 6th International Conference on Telecommunications in Modern Satellite,
Cable and Broadcasting Service, 2003. TELSIKS 2003.
[19] S. Wilson, M. Brandt-Pearce, Q. Cao, and M. Baedke, Optical MIMO transmission with multi-
pulse PPM, Submitted to IEEE Journal on Selected Areas in Communication, 2004.
[20] G. Keiser, Optical Fiber Communications. McGraw-Hill, third ed., 2000.
[21] D. Kedar and S. Arnon, Optical wireless communication through fog in the presence of pointing
errors, Applied Optics, vol. 42, pp. 49464954, August 2003.
[22] D. Gesbert, M. Sha, D. Shiu, P. Smith, and A. Naguib, From theory to practice: An overview of
MIMO space-time coded wireless systems, IEEE Journal on Selected Areas in Communications,
vol. 21, pp. 281302, April 2003.
[23] M. Srinivasan and V. Vilnrotter, Avalanche photodiode arrays for optical communications re-
ceivers, JPL TMO Progress Report, vol. 42-144, February 15 2001.
[24] X. Zhu and J. Kahn, Maximum-likelihood spatial-diversity reception on correlated turbulent free-
space optical channels, IEEE Global Telecommunications Conference, 2000. GLOBECOM 00,
vol. 2, pp. 12371241, November-December 2000.
[25] S. Haas, J. Shapiro, and V. Tarokh, Space-time codes for wireless optical channels, IEEE Inter-
national Symposium on Information Theory, p. 244, 2001.
[26] J. Pierce, Optical channels: Practical limits with photon counting, IEEE Trans. on Communica-
tions, vol. 26, pp. 18191820, December 1978.
[27] J. Lesh, Capacity limit of the noiseless, energy efcient optical PPM channel, IEEE Trans. on
Communications, vol. 31, pp. 546548, April 1983.
[28] C. Georghiades, Modulation and coding for throughput-efcient optical systems, IEEE Trans. on
Information Theory, vol. 40, pp. 13131326, September 1994.
[29] S. Karp and R. Gagliardi, The design of a pulse-position modulated optical communication sys-
tem, IEEE Trans. on Communication Technology, vol. COM-17, pp. 670676, December 1969.
92
[30] R. Gagliardi and S. Karp, M-ary Poisson detection and optical communications, IEEE Trans. on
Communications Technology, vol. COM-17, pp. 208216, April 1969.
[31] R. Cryan and R. Unwin, Optical space communications employing pulse position modulation,
IEE Colloquium on Advanced Modulation and Coding Techniques for Satellite Communications,
pp. 7/17/5, January 1992.
[32] J. Lesh, Capacity limit of the noiseless, energy-efcient optical PPM channel, IEEE Trans. on
Communications, vol. 31, pp. 546548, April 1983.
[33] G. Atkin and K. Fung, Coded multipulse modulation in optical communication systems, IEEE
Trans. on Communications, vol. 42, pp. 574582, February-April 1994.
[34] H. Shalaby, Maximum achievable throughputs for uncoded OPPM and MPPM in optical direct-
detection channels, Journal of Lightwave Technology, vol. 13, pp. 21212128, November 1995.
[35] F. Davidson and X. Sun, Gaussian approximation versus nearly exact performance analysis of
optical communication systems with PPM signaling and APD receivers, IEEE Trans. on Commu-
nications, vol. 36, pp. 11851192, Nov 1988.
[36] M. Srinivasan and V. Vilnrotter, Symbol-error probabilities for pulse-position modulation sig-
naling with an avalanche photodiode receiver and Gaussian thermal noise, JPL TMO Progress
Report, vol. 42-134, August 15 1998.
[37] R. J. McIntyre, The distribution of gains in uniformly multiplying avalanche photodiodes: The-
ory, IEEE Transactions on Electron Devices, vol. ED-19, pp. 703713, June 1972.
[38] J. Conradi, The distribution of gains in uniformly multiplying avalanche photodiodes: Experimen-
tal, IEEE Transactions on Electron Devices, vol. ED-19, pp. 713718, June 1972.
[39] P. Webb, R. J. McIntyre, and J. Conradi, Properties of avalanche photodiodes, RCA Review,
vol. 35, pp. 234278, June 1974.
[40] A. Papoulis, Probability, Random Variables, and Stochastic Processes. McGraw-Hill, 3 ed., 1991.

Você também pode gostar