3808 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO.

12, DECEMBER 2012

Novel Adaptive Transmission Algorithms for
Free-Space Optical Links
Mehdi Karimi and Murat Uysal, Senior Member, IEEE
AbstractAtmospheric turbulence is a major degrading factor
in free-space optical (FSO) systems. Since turbulence-induced
fading results in a very slowly-varying channel, reliable feedback
is possible allowing the implementation of adaptive transmission
in practical FSO systems. In this paper, we propose adaptive
transmission algorithms in which transmit power and/or mod-
ulation size are varied according to the channel conditions.
The design of adaptive algorithms is formulated as optimization
problems of spectral efciency and average power consumption
at a targeted value of outage probability under peak power
constraints. The proposed adaptive schemes bring signicant
improvements over their non-adaptive counterparts.
Index TermsFree-space optical communications, adaptive
transmission, pulse position modulation, fading channel.
I. INTRODUCTION
O
PTICAL wireless communication systems offer signi-
cant technical and operational advantages such as higher
bandwidth capacity, robustness to electromagnetic interfer-
ence, a high degree of spatial connement (bringing virtually
unlimited reuse), and unregulated spectrum. Variations of
optical wireless technology can be employed for a number
of indoor and outdoor communication applications [1][3].
Outdoor terrestrial optical wireless links that will be also
focus of this paper are commonly referred to as free-space
optical (FSO) links. Such links can be used for a wide range
of applications including metropolitan area network (MAN)
extension, local area network (LAN)-to-LAN connection to
enable enterprise/campus connectivity, broadband access to
remote or underserved areas, temporary links for disaster
recovery and emergency response, ber optic back-up, cel-
lular backhaul, wireless video surveillance and monitoring,
high denition television (HDTV) transmission, quantum key
distribution among others.
Despite the wide potential application areas and undis-
putable advantages, there exist some major technical chal-
lenges that hinder the widespread deployment of FSO technol-
ogy. A major problem is the rather disappointing performance
Paper approved by W. Shieh, the Editor for Optical Transmission and
Switching of the IEEE Communications Society. Manuscript received August
19, 2011; revised March 3, 2012.
M. Karimi was with the Faculty of Engineering,

Ozye gin University,
34794, Istanbul, Turkey. He is now at the Department of Combinatorics and
Optimization, University of Waterloo, Waterloo, Canada N2L 3G1 (e-mail:
m7karimi@uwaterloo.ca).
M. Uysal is with the Faculty of Engineering,

Ozye gin University, 34794,
Istanbul, Turkey (e-mail: murat.uysal@ozyegin.edu.tr).
This work is supported by the Marie Curie International Reintegration
Grant (PIRG07-GA-2010-268318), the Turkish National Academy of Sciences
Distinguished Young Scientist Award Program (TUBA GEBIP), and the
Scientic and Technological Research Council of Turkey (TUBITAK) Grant
111E143.
Digital Object Identier 10.1109/TCOMM.2012.091012.110550
for long-range links in which turbulence-induced fading be-
comes a major performance limiting factor [4][6]. Signicant
research efforts have aimed to address how to overcome such
limitations and improve the reliability of FSO links. Towards
this purpose, various techniques have been proposed such
as multiple-input multiple-output (MIMO) communications
[7], [8], relay-assisted (cooperative) communications [9], [10],
adaptive optics [11], and media diversity [12], [13].
A common assumption in the current literature is open-
loop implementation in which transmitter has no knowledge
of the channel. Such open-loop designs are favorable in time-
varying channels where the feedback of channel estimates
becomes problematic. However, particularly for quasi-static
channels, reliable feedback is possible and available channel
state information (CSI) at the transmitter can be used to design
adaptive transmission schemes for signicant performance im-
provements. Atmospheric turbulence results in a very slowly-
varying fading in FSO systems. The channel coherence time
is about 1-100 ms, therefore fading remains constant over
hundreds of thousand up to millions of consecutive bits for
typical transmission rates [5]. Therefore, adaptive transmission
emerges as a promising solution for FSO systems. Further-
more, feedback information required in adaptive transmission
is relatively easy to implement in FSO systems. This is
because commercially available FSO units have full-duplex
(bi-directional) capabilities and a small portion of the large
available bandwidth can be allocated for feedback purposes
without much effect on data rates.
Adaptive transmission has been extensively investigated in
the context of wireless radio-frequency (RF) networks (see
[14] and the references therein) and recently applied to FSO
communication [15][17]. In [15], Anguita et.al. have consid-
ered channel coding schemes whose rates adaptively change
to accommodate the varying channel conditions. In [16],
Djordjevic has studied adaptive modulation and coding (AMC)
assuming M-ary pulse amplitude modulation (M-PAM) and
quasi-cyclic low density parity check codes (LDPC). However,
the work in [16] builds upon rather idealistic assumptions
such that the constellation size can be changed continuously
(i.e., M can take any real value) and ignores constraints on
the peak power. The latter is particularly important for FSO
applications where eye safety standards impose restrictions on
the peak of transmit power. In [17], Barsimantov and Nikulin
have investigated adaptive power/rate transmission (assuming
a xed constellation). This work considers only the path loss
and ignores the effect of atmospheric turbulence.
In this paper, we revisit the design problem of adaptive
FSO transmission over atmospheric turbulence channels under
0090-6778/12$31.00 c 2012 IEEE
KARIMI and UYSAL: NOVEL ADAPTIVE TRANSMISSION ALGORITHMS FOR FREE-SPACE OPTICAL LINKS 3809
the assumption of practical constellation (modulation) sizes
and average/peak power constraints. Specically, we consider
two cases: (1) Power adaptation with a xed modulation size
and (2) Joint power and modulation adaptation. We assume
intensity modulation/direct detection (IM/DD) with M-ary
pulse position modulation (M-PPM) where M can take only
integer values. We consider both weak and strong turbulence
regimes that are respectively modeled as log-normal and
gamma-gamma fading. We formulate the design of adaptive
transmission algorithms as the optimization of average power
consumption (P
A
) and the number of transmitted bits per
chip (BpC) at a targeted value of outage probability. The
optimization is carried out with respect to adaptive parameters
(i.e., transmit power and modulation size) under peak power
constraints. Our simulation results demonstrate that proposed
adaptive algorithms provide signicant performance improve-
ments particularly in the presence of strong turbulence.
The rest of the paper is organized as follows: In Section II,
we present the system model under consideration. In Section
III, we describe adaptive transmission scenarios, formulate the
associated optimization problems, and present the resulting
adaptive algorithms. In Section IV, we provide simulation
results to demonstrate the performance of proposed adaptive
transmission schemes. Section V concludes the paper.
II. CHANNEL AND TRANSMISSION MODELS
We assume an IM/DDFSO link with M-ary PPM over
atmospheric turbulence channel. In this section, we rst sum-
marize the channel characteristics, then introduce the signal
model under consideration.
A. Channel Model
We consider an aggregate channel model that takes into ac-
count the effects of turbulence-induced fading and propagation
loss. The channel coefcient is expressed as [18]
h = h
l
h
f
(1)
where h
l
and h
f
denote the propagation loss and the fading
coefcient, respectively. h
l
can be expressed as [18], [19]
h
l
= A(, d) e
d
(2)
where the rst term A(, d) in (2) represents the losses in the
transmitted power and is mainly a function of the link length
(d), and the optical wavelength (). Under the assumption
of a point source in free space, this is simply given by
[18] A
TX
A
RX
/(d)
2
where A
TX
and A
RX
are transmitter
aperture area and receiver aperture area, respectively. In real-
life environments and practical systems, one needs to further
consider transmitter/receiver isolation factor, optical system ef-
ciency, modulation extinction ratio, optical beam divergence,
geometric spreading loss, and lter losses. A comprehensive
loss model can be found in [20].
The second term in (2) is based on Beers-Lambert law
where represents the attenuation coefcient [19]. It is given
by [21] = (3.91/V )(/550)
q
where V is the visibility in
kilometers (km) and q is a parameter related to the visibility
expressed as
q =
_
0.585V
1/3
V < 6km
1.3 6km < V < 50km
(3)
For the modeling of h
f
, we consider two cases: In weak
turbulence conditions, log-normal model is typically used [5].
In this case, we have h
f
= exp(H) where H is a Gaussian
random variable with a variance of
2
H
. Under the assumption
of plane wave propagation and normalized fading coefcient,
i.e., E[h
f
] = 1 , the probability density function (pdf) of h
f
is given by
p(h
f
) =
1
h
f

2
H
exp
_

1
2
2
H
_
ln(h
f
) +

2
H
2
_
2
_
(4)
where
2
H
=
2
R
with Rytov variance
2
R
=
1.23(2/)
7/6
C
2
n
d
11/6
and refractive-index structure
constant C
2
n
. In medium and strong turbulence conditions,
Gamma-Gamma fading model becomes a better t [5]. In
this model, the pdf of h
f
is given by
p (h
f
) =
2(
1

2
)
(1+2)/2
(
1
) (
2
)
h
f

1
+
2
2
1
K
12
_
2
_

2
h
f
_
(5)
where K(.) is the modied Bessel function of the second
type and 1/
1
, 1/
2
are the variances of large and small
scale eddies, respectively. Under the assumption of plane wave
propagation, they are given by

1
=
_
exp
_
0.49
2
R
_
_
1 + 1.11
12/5
R
_
7/6
_
1
_
1

2
=
_
exp
_
0.51
2
R
_
_
1 + 0.69
12/5
R
_
5/6
_
1
_
1
(6)
The value of
2
R
species the level of turbulence; weak and
strong uctuations are associated with
2
R
< 1 and
2
R
> 1,
respectively.
B. Signal Model
We consider an IM/DD FSO system with M-ary PPM
where M = 2
m
with m denoting positive integers. PPM is a
form of orthogonal signaling in which information is encoded
by the position of an optical pulse [3]. The symbol duration is
divided into M parts (called chips) and a pulse is transmitted
in one of these chips according to the value of the symbol. We
assume that the pulse duration (i.e., chip duration) is constant
for all values of M and the symbol duration changes with M.
We also assume that the average power is equal for all values
of M (i.e., different modulation sizes). To maintain this, since
the chip duration is constant, the amplitude of the pulse is
increased by the increase of M.
To quantify the bandwidth usage, we adopt the number of
bits carried per chip time (BpC) instead of spectral efciency.
BpC is dened as the ratio of data rate over channel band-
width. Since bandwidth and chip duration are xed under our
assumptions, BpC becomes a more meaningful measure. Not-
ing that a BpC of 0.5 is achieved for both M = 2 and M = 4,
we use M = 4 as the minimum allowable modulation size as
this results in a better error rate performance. Therefore, in
our case, M is chosen from the set M = {4, 8, . . . , M
max
}
where M
max
= 2
mmax
is the maximum value of M that can be
chosen by the modulator, specied by the minimum acceptable
instantaneous value of BpC.
3810 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 12, DECEMBER 2012
Dene x
i
as a binary variable that takes the values of 0 or
1 depending on whether a pulse is sent or not during the i
th
chip. In a vector notation, the transmitted modulation signal
has the form of x = [x
1
, . . . , x
M
] in which only one of x
i
s
is equal to 1 (i.e., the location of the transmitted pulse). At
the receiver side, the electrical signal during the i
th
chip is
given by
y
i
= hP
t
x
i
+n, i= 1, 2,...,M (7)
where P
t
is the power of the transmitted pulse, h is the
channel gain dened in (1), is the optical-to-electrical
conversion coefcient, and n is a signal-independent zero-
mean Gaussian noise with variance
2
n
. Let y = [y
1
, . . . , y
M
]
denote the received vector over a symbol duration that consists
of M chips. The optimum detector compares y
i
s, nds the
maximum of them and therefore identies the location of the
chip containing the pulse.
III. ADAPTIVE FSO TRANSMISSION SCHEMES
A. Adaptive Parameters and Performance Measures
In our adaptive transmission scheme, transmit power and
modulation size are chosen as the system parameters that
will be adaptively varied according to the channel conditions.
Therefore, hereinafter, we denote the constellation size and
transmit power as a function of the fading channel coefcient
h, i.e., namely M(h) and P
t
(h).
In our work, we adopt three performance measures. The
rst one is the average transmit power denoted by P
A
. It
is calculated in terms of the pulse power as P
A
= P
t
/M. As
earlier noted, to maintain the same average power for different
modulation sizes (i.e., different values of M), the amplitude
of the pulse is increased by the increase of M. Life time of
the laser and eye safety standards impose some constraints on
the average and peak transmission power. These constraints
can be stated as
P
A
= E
_
P
t
(h)
M(h)
_
=
_

0
P
t
(h)
M(h)
p(h) dh P
A,max
(8)
P
t
(h) P
m
(9)
where P
A,max
and P
m
denote respectively the average and
peak bounds of the transmission power.
The second performance measure is the number of bits
carried per chip time (BpC) which was earlier dened in
Section II. It can be noted that in a non-adaptive system
with a xed M-PPM constellation, BpC is calculated as
= (log
2
M)/M.
The third performance measure is the outage probability.
Let P
th
denote the highest bit error probability (P
b
) that
is acceptable by the system for a xed signal-to-noise ratio
(SNR). An outage is said to occur if P
b
> P
th
and the outage
probability is calculated as P
out
= P{P
b
> P
th
}. To maintain
a reliable communication, the outage probability must be equal
to a pre-specied small value that will be denoted by PO in
this paper.
B. Adaptive Power and/or Modulation Algorithms
We aim to design two adaptive algorithms: 1) Adaptive
power (AP) algorithm with a xed M-PPM constellation, and
2) Adaptive power and modulation (APM) algorithm. In both
algorithms, M is allowed to take only discrete values. For each
scheme, our objective is to either i) maximize BpC under a
constraint on the average power, or ii) minimize the average
power P
A
for a specic value of BpC. This therefore leads
to four different adaptive algorithms that we name as AP-1,
AP-2, APM-1 and APM-2. In all four algorithms, the outage
probability is set equal to the desired value, i.e., P
out
= PO.
1) AP-1: In this scheme, we aim to maximize BpC by
adaptively changing the transmit power (under a constraint
on the average power) according to channel conditions while
keeping the outage probability equal to PO. First note that we
must have at least P
b
= P
th
to prevent outage. Let C
M
denote
the signal-to-noise ratio (SNR) corresponding to P
b
= P
th
. It
is obtained as
1
.
SNR

= A

hP
t
= C
M
A

=

N
(10)
In AP-1 scheme, P
t
is a function of h. Clearly, to satisfy
A

hP
t
= C
M
for low values of h, P
t
might take too high
values and violate the peak power limit. To avoid this problem,
we set P
t
equal to zero for sufciently small values of h (less
than a threshold ) as
P
t
(h, M) =
_
0 h <
CM
A

h
h
(11)
where P
t
(h, M) is the transmit power as a function of h and
M. Outage occurs for the case of h < and this can be
expressed as
P
out
AP
= P {h } =
_

0
p(h) dh (12)
The value of can be calculated from (12) setting P
out
AP
=
PO. Substituting P
t
(h = , M) from (11) in (9), we have
C
M
/(A

P
m
) . Hence, to achieve P
out
AP
= PO, there must
be at least one M value that satises C
M
/(A

P
m
) . C
M
is an increasing function of M, so if the value of PO is such
that in (12) satises C
Mmax
/(A

P
m
) , then all values of
M M fulll this condition. Note that for practical values of
M (i.e., M 2
10
), the change of (C
Mmax
C
4
)/(A

P
m
) with
respect to does not considerably affect the outage probability
given by (12). On the other hand, P
t
(h, M) dened in (11) is
a decreasing function of h. Therefore, for the case of h ,
C
Mmax
/(A

) P
m
guarantees the satisfaction of (9) for
all values of M. In conclusion, for either case, any value of
M M can be chosen.
To maximize BpC, we should nd the minimum value of
M (denoted by M

) that satises (8). Hence, from (8) and

(11), we have
M

= min
MM
M
s.t. P
A
=
CM
M
_
A

1
h
p(h) dh
f
_
P
A,max
.
(13)
1
See the Appendix for the derivation of C
M
KARIMI and UYSAL: NOVEL ADAPTIVE TRANSMISSION ALGORITHMS FOR FREE-SPACE OPTICAL LINKS 3811
Find the modulation
size from (13)
Find the transmit power
from (11)
Choose such that
max
/ ( ' )
M m
C A P F b
F
,max A
P
, '
t
P h M
' M
Set as the value of
such that
2
log M M C
Choose such that
max
/ ( ' )
M m
C A P F b
' B
yes
No
M M
2
' log B M M
No
solution
F
M
' M
( ) a ( ) b
Find the transmit power
from (11) , '
t
P h M
Fig. 1. (a) The owchart for AP-1 adaptive scheme. (b) The owchart for
AP-2 adaptive scheme.
Note that C
M
/M is a decreasing function of M. Substituting
M = M

in (11), we nd the transmit power as P

t
(h, M

)
for AP-1 scheme. The corresponding BpC is given by =
(log
2
M

)/M

. The algorithm owchart for AP-1 scheme is

shown in Fig. 1(a).
2) AP-2: In this scheme, for a xed BpC, we aim to
minimize the average power P
A
by changing the transmit
power while keeping the outage probability equal to PO.
Modulation size is xed and for each value of M, BpC is
uniquely specied by M log
2
M. Hence for AP-2 scheme,
only discrete values of B

= M log
2
M can be used as
an input. Following the same reasoning presented for AP-1,
transmit power for AP-2 is given by (11) where the threshold
is calculated from (12). However, it can be seen from (13) that
any positive integer value of M uniquely species the average
power P
A
and there is no degree of freedom. Hence, we do
not have to solve any optimization problem and by substituting
M = M

(solved from B

= M

log
2
M

) in (11), we can nd
the transmit power P
t
(h, M

) for AP-2 scheme. The algorithm

owchart for AP-2 scheme is found in Fig. 1(b).
3) APM-1: In this scheme, we aim to maximize BpC by
changing the transmit power and modulation size under a
constraint on the average power as well as outage prevention.
Similar to AP-1 scheme, the transmit power is a function
of h and is given by (11). However, different from AP-1,
C
M
(M) is now a function of h. The outage probability P
out
APM
can be calculated from (12). M is chosen from the set of
permissible values, according to h, to achieve the maximum
BpC. The permissible values of M are those from the set M
that satisfy the peak power constraint in (9). As discussed for
AP-1 scheme, we choose the upper limit of outage probability
PO such that the threshold satises C
Mmax
/(A

P
m
) ;
this makes all values of m M permissible.
Since M takes discrete values, this optimization problem
is an integer programming. To simplify it, we use a heuristic
reasoning as follows: In a non-adaptive transmission, xed
power level and modulation size must be chosen according to
the worst condition of the channel. However, in an adaptive
system, modulation size is chosen according to channel con-
ditions. Therefore, we can partition the interval [, ) into
subintervals as = h
mmax
h
mmax1
. . . h
2

h
1
= +. The chosen modulation size M(h) can be therefore
stated as
M(h) =
_
0, if h < = h
mmax
2
i
, if h
i
< h h
i1
(14)
Let P
t
(h, M(h)) denote the transmit power for APM-1
scheme that is calculated by substituting (14) in (11). Now, our
optimization problem reduces to a continuous one of nding
the vector h = [h
mmax
, h
mmax1
, . . . , h
2
] which would
yield the maximum value of average BpC. This optimization
problem can be stated as
max
h
(h)
s.t. 1. = h
mmax
h
mmax1
. . . h
2
2. P
A
(h) P
A,max
(15)
where (h) is the achieved average BpC and P
A
(h) is the
average transmitted power, that are respectively calculated as
(h) =
log
2
M
max
M
max
_
hmmax1
hmmax
p(h)dh
+
log
2
Mmax
2
1
2
M
max
_
hmmax2
hmmax1
p(h)dh
+ +
2
4
_
h1=
h2
p(h)dh, (16)
P
A
(h) = E
_
P
t
(h)
M (h)
_
=
C
Mmax
A

M
max
_
hmmax1
hmmax
p(h)
h
dh
+
C
Mmax/2
A

M
max
/2
_
hmmax2
hmmax1
p(h)
h
dh
+ +
C
4
4A

_

h2
p(h)
h
dh. (17)
To solve (15), Karush-Kuhn-Tucker (KKT) theorem can be
used [22]. According to KKT, if h is a local maximizer for
the optimization problem of (15), there exists R and

j
R, j {m
max
1, m
max
2, . . . , 2} such that

(h)
h
i
+
(P
A
(h) P
A,max
)
h
i
+

j
(h
j+1
h
j
)
h
i
= 0
i {m
max
1, m
max
2, . . . , 2} (18a)
0,
j
0 (18b)
(P
A
(h) P
A,max
) = 0,
j
(h
j+1
h
j
) = 0 (18c)
where and
j
s are the KKT multipliers. Eq. (18c) implies
that if the conditions P
A
(h) P
A,max
or (h
j+1
h
j
) 0
are inactive, then = 0 or
j
= 0. To nd the candidate
maximizer in (15), we consider three different cases based on
KKT conditions. Namely, we have
i) P
A
(h) = P
A,max
and h
j+1
< h
j
for all values of j,
ii) P
A
(h) < P
A,max
3812 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 12, DECEMBER 2012
iii) P
A
(h) = P
A,max
and h
j+1
= h
j
for at least one value
of j.
Each case will be separately investigated in the following:
i) In this case, we have P
A
(h) = P
A,max
(i.e., the rst
inequality constraint is active) and h
j+1
< h
j
for all values
of j. From (18c), we obtain
j
= 0 for all values of j. By
applying the rst KKT condition given by (18a) to (15) for
i {m
max
1, m
max
2, . . . , 2}, we obtain
_

i + 1
2
i+1
+
i
2
i
_
p(h
i
) +
_
C
2
i+1
A

2
i+1

C
2
i
A

2
i
_
p(h
i
)
h
i
= 0 (19)
which yields
h
i
=
(C
2
i+1 2C
2
i )
A

(1 i)
(20)
Note that (20) is completely independent of the probability
density function p(h). By substituting h
i
s from (20) in (17)
and solving P
A
(h) = P
A,max
numerically, we obtain and
therefore h
i
s can be computed from (20).
ii) In this case, we have P
A
(h) < P
A,max
which, using
(18c), results in = 0. By applying the rst KKT condition
to (15) for i {m
max
1, m
max
2, . . . , 3}, we obtain
_

i + 1
2
i+1
+
i
2
i
_
p(h
i
)
i
+
i1
= 0 (21)
where
i
s are KKT multipliers dened in(18a). The rst term
in (21) is always positive, so at least one of
i
or
i1
must
be nonzero. Eq. (18b) further indicates that
i
and
i1
are
non-negative. Hence, to satisfy the equality of (21),
i
must be
positive which, from (18c), results in h
i+1
= h
i
. Hence, we
conclude that = h
mmax
= h
mmax1
= . . . = h
2
. This means
that M = 4 is chosen throughout the range of h [, ) and
the second inequality constraint of (15) reduces to
C
4
4A

p(h)
h
dh < P
A,max
. (22)
It can be concluded that when the inequality in (22) is satised
or, in other words, P
A,max
is high enough, the adaptive
transmitter always chooses 4-PPM for h and the maximum
value of = 0.5 is achieved.
iii) In this case, we have P
A
(h) = P
A,max
and h
j+1
= h
j
for at least one j. Assume that we replace each h
j
by h
j
+j
for a small value of > 0. By this transformation, case (iii)
becomes similar to case (i). So, it can be easily argued that
the h derived for this case leads to a lower (h) than the
one derived for the rst case. (15) is used for the derivation
of h and M(h) is obtained from (14). The transmit power
P
t
(h, M(h)) is then calculated by substituting M(h) in (11).
The owchart for APM-1 scheme is given in Fig. 2.
4) APM-2: In this scheme, we aim to minimize the av-
erage power required to achieve a targeted value of BpC
by changing the transmit power and modulation size while
preventing outage. The corresponding optimization problem
can be formulated as
min
h
P
A
(h)
s.t. 1. = h
mmax
h
mmax1
. . . h
2
2. (h) = B

(23)
Choose such that
max
/ ( ' )
M m
C A P F b
F
Is (22)
satisfied?
Set M(h)=4
yes
Use (20) and
to find
,max
( )
A A
P h P
max
2
[ , , ]
m
h h h y
No
,max A
P
Find the modulation
size M(h) from (14)
Find the transmit
power
form (11)
, ( )
t
P h M h
Fig. 2. The owchart for APM-1 adaptive scheme.
' B
yes
No
min 2 min
' 0.5
log '
B
M M B
b
b
No solution
Choose such that
max
/ ( ' )
M m
C A P F b
F
Use (26) and
to find
( ) ' h B C
max
2
[ , , ]
m
h h h y
Find the modulation
size M(h) from (14)
Find the transmit
power
from (11)
, ( )
t
P h M h
Fig. 3. The owchart for APM-2 adaptive scheme.
where P
A
(h) and (h) are dened in (16) and B

denotes
the desired BpC. According to KKT conditions, for h to be
a local minimizer in the optimization problem of (23), there
exists R and
j
R, j {m
max
1, m
max
2, . . . , 2}
such that
P
A
(h)
h
i
+
((h) B

)
h
i
+

j
(h
j+1
h
j
)
h
i
= 0
i {m
max
1, m
max
2, . . . , 2} (24a)

j
0 (24b)

j
(h
j+1
h
j
) = 0 (24c)
Following similar steps as in the previous optimization prob-
lem and using the rst KKT condition (24a) for i {m
max

1, m
max
2, . . . , 2}, we can write
_
C
2
i+1
A

2
i+1

C
2
i
A

2
i
_
p(h
i
)
h
i
+
_
i + 1
2
i+1

i
2
i
_
p(h
i
) = 0 (25)
which yields
h
i
=
1

(C
2
i+1 2C
2
i )
A

(1 i)
. (26)
By comparison of (26) with (20) and dening the constant

= 1/, we realize that both optimization problems result

in the same form of h. By substituting h
i
s from (26) in (16)
and solving the second constraint of (23) (i.e., (h) = B

),
we obtain and therefore can compute h
i
(equivalently h)
from (26). Hence solving the resulting optimization problem
yields h which is then used to derive M(h) from (14). Finally,
the transmit power P
t
(h, M(h)) is calculated by substituting
M(h) in (11). The owchart for APM-2 scheme is given in
Fig. 3.
The practical implementation of the proposed algorithms
requires the feedback of the fading channel coefcient h.
KARIMI and UYSAL: NOVEL ADAPTIVE TRANSMISSION ALGORITHMS FOR FREE-SPACE OPTICAL LINKS 3813
Under the assumption of quasi-static fading model, we have to
feedback h only once in every coherence time (e.g., 1-100 ms).
It was shown in [23] that 5 bits per dimension are sufcient
for a precise quantization of the fading channel coefcient. In
[23], Lloyd Algorithm [24] was used for quantization and an
error-free feedback channel was assumed. In our case, consid-
ering further the turbulence characteristics of the FSO channel,
additional bits will be needed for channel coding to mitigate
the degrading effect of turbulence channels. Assuming a code
rate of 1/3 such as the turbo code used in the 4G LTE standard
[25], 35 = 15 bits are needed in the feedback channel. Even
with a much more powerful code, the number of feedback bits
will be less than a hundred in most typical conditions and this
is a negligible fraction of FSO transmission rates typically on
the order of hundreds of Mbps up to multiple Gbps.
IV. NUMERICAL RESULTS
In this section, we evaluate the performance of the pro-
posed adaptive transmission algorithms through Monte-Carlo
simulations. We assume a wavelength of = 1550nm, noise
standard deviation
n
= 10
4
, visibility V = 10km, trans-
mitter aperture width W
0
= 5cm, receiver aperture diameter
D
r
= 10cm, m
max
= 7, peak power limit P
M
= 5W, and the
desired probability of bit error P
th
= 10
6
.
Fig. 4 illustrates BpC versus the average transmitted power
for APM-1 and AP-1 algorithms over weak (log-normal)
turbulence. We assume d = 2000 and C
2
n
= 10
15
which,
according to the Rytov variance formula, result in
2
R
= 0.07.
The outage probability is set equal to PO = 10
6
. The
performance of non-adaptive transmission
2
is also included
as a benchmark. It can be observed from Fig. 4 that for
P
A
= 30dB, non-adaptive transmission achieves = 0.05.
This climbs up to = 0.15 and = 0.2 respectively for AP-1
and APM-1 schemes. By increase of P
A
to 20dB, further
increases to 0.5 for both adaptive strategies while it becomes
0.25 for the non-adaptive scheme.
In Fig. 5, we illustrate the performance of APM-1 and
AP-1 algorithms over strong turbulence modeled by gamma-
gamma fading. We assume d = 3000 and C
2
n
= 10
14
which
results in
2
R
= 1.55. The outage probability is kept equal
to PO = 10
4
. It can be observed from Fig. 5 that even for
P
A
= 20dB, non-adaptive transmission achieves = 0. That
is because the performance of the non-adaptive transmission
is greatly degraded by the increase of turbulence. On the
other hand, AP-1 and APM-1 algorithms achieve = 0.15
and = 0.35 respectively. This dramatic improvement is a
result of adaptive transmission which allocates the resources
properly based on the channel conditions and allows the
transmitter to change the power accordingly. Note that the
peak power constraint P
m
imposes a limit on the achievable
PO. This limit is more critical in strong turbulence regime
because the probability density function of h
f
is concentrated
around the near-zero values and the power cannot be increased
so much to compensate it. For example, PO cannot be
decreased lower than 10
4
for the parameters chosen in Fig. 5.
2
Calculation of BpC and transmit power for non-adaptive scheme is
provided in the Appendix.
40 35 30 25 20 15
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
P
A
(dB)
B
p
C


APM1
AP1
Nonadaptive
Fig. 4. BpC versus transmit power for log-normal turbulence channel with

2
R
= 0.07.
30 25 20 15 10 5 0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
P
A
(dB)
B
p
C


APM1
AP1
Nonadaptive
Fig. 5. BpC versus transmit power for gamma-gamma turbulence channel
with
2
R
= 1.55.
So, our adaptation strategies greatly improve the performance
over strong turbulence channel within the achievable outage
probability range. If desired, additional strategies such as
multiple transmitters can be used in conjunction with adaptive
transmission to further reduce the outage probability. The
comparison of Fig. 4 and Fig. 5 shows that there is not a
signicant difference between the performances of AP-1 and
APM-1 schemes over log-normal channel. However, APM-1
performs much better than AP-1 over gamma-gamma channel
(strong turbulence regime).
Fig. 6 illustrates the minimum average power versus the
targeted BpC over log-normal turbulence with
2
R
= 0.07. The
outage probability is kept equal to PO = 10
6
as in Fig. 4.
It can be observed from Fig. 6 that, to achieve = 0.25 in
weak turbulence, both the AP-2 and APM-2 schemes require
approximately 6 dB lower than the power required by the
non-adaptive scheme. Note that for discrete values of =
M log
2
M, both AP-2 and APM-2 result in approximately the
3814 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 12, DECEMBER 2012
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
40
35
30
25
20
15
BpC
P
A

(
d
B
)


APM2
AP2
Nonadaptive
Fig. 6. Minimum average power versus BpC for log-normal turbulence
channel with
2
R
= 0.07.
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
30
25
20
15
10
5
0
5
BpC
P
A

(
d
B
)


APM2
AP2
Nonadaptive
Fig. 7. Minimum average power versus BpC for gamma-gamma turbulence
channel with
2
R
= 1.55.
same P
A
. APM-2 can further achieve all the values of from
7/2
7
to 0.5.
Fig. 7 illustrates the minimum average power versus the
targeted BpC over gamma-gamma turbulence with
2
R
= 1.55.
The outage probability is kept equal to PO = 10
4
as
in Fig. 5. It can be observed from Fig. 6 that to achieve
= 0.25 in strong turbulence, the required P
A
is 23 dB
for APM-2 which is respectively 3 dB and 21 dB lower than
the power required by AP-2 and non-adaptive schemes. By
comparison of Fig. 6 and Fig. 7, it can be further noted that
the performance improvement of adaptive schemes is much
larger in strong turbulence regime.
V. CONCLUSIONS
In this paper, we have proposed adaptive FSO transmission
schemes taking into account practical modulation sizes and
average/peak power constraints. Specically, we have devel-
oped algorithms in which either transmit power is alone or
jointly with modulation size adapted according to the chan-
nel conditions. The adaptive algorithms have been obtained
through solving the optimization of performance measures
under consideration (namely, average power consumption and
the number of transmitted bits per chip) for a targeted value of
outage probability. Our simulation results over log-normal and
gamma-gamma atmospheric turbulence channels have demon-
strated that proposed adaptive algorithms bring signicant
performance improvements over non-adaptive schemes. While
power-only and joint power-modulation adaptive schemes
yield similar performance results in weak turbulence condi-
tions, the advantage of using the latter becomes obvious in
strong turbulence channel.
APPENDIX A
In this Appendix, we investigate the BpC and the transmit
power for the non-adaptive system to establish the base line
performance. The probability of error for non-adaptive system
with M-PPM modulation is given by [26]
P
b
=
M
2(M 1)
1

_
1 (1 Q(z))
M1
_
exp
_

1
2
_
z
hP
t

n
_
2
_
dz (27)
where Q(.) is the Gaussian Q-function. We must have at least
P
b
= P
th
to prevent outage in the transmission. As dened
in (10), let C
M
denote the SNR for which P
b
= P
th
C
M
depends on M and is calculated numerically from (27).
We can therefore write the outage probability for the non-
adaptive scheme as P
out
NA
= P (h < C
M
/(A

P
t
)). Under the
assumption of log-normal turbulence, by using (1) and (4) we
have
P
out
NA
= P
_
h = h
l
h
f
<
C
M
A

P
t
_
= P
_
H < ln
C
M
h
l
A

P
t
_
= Q
_
1

H
ln
_
C
M
h
l
A

P
t
_
+

H
2
_
(28)
For the Gamma-Gamma fading model, we have
P
out
NA
= P
_
h = h
l
h
f
<
C
M
A

P
t
_
=
_
C
M
h
l
A

P
t
0
p (h
f
) dh
f
(29)
where p(h
f
) is given by (5). (29) does not yield a closed-form
expression, but can be calculated numerically.
Since transmit power and modulation size are kept constant
in non-adaptive transmission, we obtain the BpC and transmit
power respectively as = log
2
M/M and P
t
= MP
A
. To
achieve the maximum value of BpC, the minimum value of
M is chosen as to satisfy P
out
NA
= PO.
REFERENCES
[1] H. Willebrand and B. Ghuman, Free Space Optics: Enabling Optical-
Connectivity in Todays Networks. Sams Publishing, 2002.
[2] D. Kedar and S. Arnon, Urban optical wireless communication net-
works:the main challenges and possible solutions, IEEE Commun.
Mag., vol. 42, no. 5, pp. 27, Jan. 2004.
KARIMI and UYSAL: NOVEL ADAPTIVE TRANSMISSION ALGORITHMS FOR FREE-SPACE OPTICAL LINKS 3815
[3] S. Hranilovic, Wireless Optical Communication Systems. Springer
Verlag, 2005.
[4] I. Kim, H. Hakakha, P. Adhikari, E. Korevaar, and A. Majumdar,
Scintillation reduction using multiple transmitters, in Proc. SPIE, vol.
2990, pp. 102113, Apr. 1997.
[5] L. Andrews and R. Phillips, Laser Beam Propagation through Random
Media, 2nd edition. SPIE Press, 2005.
[6] X. Zhu and J. Kahn, Free-space optical communication through atmo-
spheric turbulence channels, IEEE Trans. Commun., vol. 50, no. 8, pp.
12931300, Aug. 2002.
[7] S. M. Navidpour, M. Uysal, and M. Kavehrad, BER performance
of free-space optical transmission with spatial diversity, IEEE Trans.
Wireless Commun., vol. 6, no. 8, Aug. 2007.
[8] S. Wilson, M. Brandt-Pearce, Q. Cao, and J. Leveque, Free-space
optical MIMO transmission with Q-ary PPM, IEEE Trans. Commun.,
vol. 53, no. 8, pp. 14021412, Aug. 2005.
[9] G. K. Karagiannidis, T. A. Tsiftsis, and H. G. Sandalidis, Outage
probability of relayed free space optical communication systems, IEE
Electron. Lett., vol. 42, no. 17, 17 Aug. 2006.
[10] M. Safari and M. Uysal, Relay-assisted free-space optical communi-
cation, IEEE Trans. Wireless Commun., vol. 7, no. 12, pp. 54415449,
Dec. 2008.
[11] R. K. Tyson, Bit error rate for free space adaptive optics laser
communications, OSA J. Optical Society America A, vol. 19, no. 4,
pp. 753758, Apr. 2002.
[12] W. Zhang, S. Hranilovic, and C. Shi, Soft-switching hybrid FSO/RF
links using short-length raptor codes: design and implementation, IEEE
J. Sel. Areas Commun., vol. 27, no. 9, Dec. 2009.
[13] B. He and R. Schober, Bit-interleaved coded modulation for hybrid
RF/FSO systems, IEEE Trans. Commun., vol. 57, no. 12, pp. 3753
3763, Dec. 2009.
[14] A. Goldsmith, Wireless Communications, Cambridge University Press,
2005.
[15] J. A. Anguita, M. A. Neifeld, B. Hildner, and B. Vasic, Rateless coding
on experimental temporally correlated FSO channels, IEEE J. Lightw.
Technol., vol. 28, no. 7, pp. 9901002, Apr. 2010.
[16] I. B. Djordjevic, Adaptive modulation and coding for free-space optical
channels, IEEE/OSA J. Optical Commun. Netw., vol. 2, no. 5, pp. 221
229, May 2010.
[17] O. Barsimantov and N. N. Nikulin, Adaptive optimization of a free
space laser communication system under dynamic link attenuation,
IEEE/OSA J. Optical Commun. Netw., vol. 3, no. 3, pp. 219222, Mar.
2011.
[18] S. Karp, R. M. Gagliardi, S. E. Moran, and L. B. Stotts, Optical
Channels: Fibers, Clouds, Water and the Atmosphere. Plenum, 1988.
[19] I. Kim, J. Koontz, H. Hakakha, P. Adhikari, R. Stieger, C. Moursund,
M. Barclay, A. Stanford, R. Ruigrok, J. Schuster, et al., Measurement
of scintillation and link margin for the Terra Link laser communication
system, in Proc. SPIE, vol. 3232, pp. 100118, Jan. 1998.
[20] K. Kiasaleh, Performance analysis of free-space on-off-keying optical
communication systems impaired by turbulence, Proc. 2002 SPIE Free-
Space Laser Commun. Technol. XIV, vol. 4635.
[21] I. Kim, B. McArthur, and E. Korevaar, Comparison of laser beam
propagation at 785 nm and 1550 nm in fog and haze for optical wireless
communications, Proc. SPIE, vol. 4214, pp. 2637, Feb. 2001.
[22] E. K. P. Chong and S. H. Zak, An Introduction to Optimization, 3rd
edition. John Wiley and Sons, 2008.
[23] O. Amin, B. Gedik, and M. Uysal, Channel estimation for amplify-
and-forward relaying: cascaded against disintegrated estimators, IET
Commun., vol. 4, pp. 12071216, July 2010.
[24] Y. Linde, A. Buzo, and R. M. Gray, An algorithm for vector quantizer
design, IEEE Trans. Commun., vol. 28, no. 11, p. 8495, Jan. 1980.
[25] F. Khan, LTE for 4G Mobile Broadband: Air Interface Technologies and
Performance. Cambridge University Press, 2009.
[26] J. G. Proakis and M. Salehi, Digital Communications, 5th edition.
McGraw-Hill, 2008.
Mehdi Karimi was born in 1982. He received the
Ph.D. degree (with honors) in electrical engineering
from Sharif University of Technology, Tehran, Iran,
in 2010. He is currently pursuing the M.Sc. degree
in the department of combinatorics and optimization,
University of Waterloo, Canada.
From 2004 to 2010, he was a Research Assis-
tant at Advanced Communication Research Insti-
tute, Wireless Research Laboratory of the Electrical
Engineering Department in Sharif University. He
was a Postdoctoral Research Assistant in Ozyegin
University, Istanbul, Turkey, and the University of Waterloo, Canada, in 2011.
His research interests include continuous optimization, robust optimization,
optical communication systems, channel coding and information theory.
Murat Uysal was born in Istanbul, Turkey in 1973.
He received the B.Sc. and the M.Sc. degree in
electronics and communication engineering from
Istanbul Technical University, Istanbul, Turkey, in
1995 and 1998, respectively, and the Ph.D. degree in
electrical engineering from Texas A&M University,
College Station, Texas, in 2001. Dr. Uysal is cur-
rently an Associate Professor at Ozyegin University,
Istanbul, Turkey, where he leads the Communication
Theory and Technologies (CT&T) Research Group.
Prior to joining Ozyegin University, he was a tenured
Associate Professor at the University of Waterloo (Canada) where he still
holds an Adjunct Associate Professor position. Dr. Uysals research interests
are in the broad areas of communication theory and signal processing with a
particular emphasis on the physical layer aspects of wireless communication
systems in radio, acoustic, and optical frequency bands. He has authored more
than 160 journal and conference papers on these topics.
Dr. Uysal is a Senior IEEE member and an active contributor to his
professional society. He currently leads the EU COST Action OPTICWISE
which is a high-prole consolidated European scientic platform for in-
terdisciplinary research activities in the emerging area of optical wireless
communications. He serves on the editorial boards of IEEE TRANSACTIONS
ON COMMUNICATIONS, IEEE TRANSACTIONS ON VEHICULAR TECHNOL-
OGY, Wiley Wireless Communications and Mobile Computing (WCMC) Jour-
nal, and Wiley Transactions on Emerging Telecommunications Technologies
(ETT). In the past, he served as an Editor for IEEE TRANSACTIONS
ON WIRELESS COMMUNICATIONS (2003-2011), IEEE COMMUNICATIONS
LETTERS (2004-2011), Guest Co-Editor for WCMC Special Issue on MIMO
Communications (October 2004), and IEEE JOURNAL ON SELECTED AREAS
IN COMMUNICATIONS Special Issue on Optical Wireless Communications
(December 2009). Over the years, Dr. Uysal has served on the technical
program committee of more than 80 international conferences and workshops
in the communications area. He is currently serving as the Technical Program
Co-Chair of IEEE WCNC14. He was the Co-Chair of IEEE ICC07 Commu-
nication Theory Symposium, the Chair of CCECE08 Communications and
Networking Symposium, and the Tutorial Chair of IEEE WCNC11. Dr. Uysal
is the recipient of several awards including the Turkish Academy of Sciences
Distinguished Young Scientist Award, University of Waterloo Engineering
Research Excellence Award, and NSERC Discovery Accelerator Supplement
Award, among others.