IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO.

3, MARCH 2010 1
Rateless Coding for Hybrid Free-Space Optical and
Radio-Frequency Communication
Ali AbdulHussein, Anand Oka, Member, IEEE, Trung Thanh Nguyen, Student Member, IEEE,
and Lutz Lampe, Senior Member, IEEE
AbstractFree-space optical (FSO) transmission systems en-
able high-speed communication with relatively small deployment
costs. However, FSO suffers a critical disadvantage, namely
susceptibility to fog, smoke, and conditions alike. A possible
solution to this dilemma is the use of hybrid systems employing
FSO and radio frequency (RF) transmission. In this paper we
propose the application of a rateless coded automatic repeat-
request scheme for such hybrid FSO/RF systems. The advantages
of our approach are (a) the full utilization of available FSO
and RF channel resources at any time, regardless of FSO or
RF channel conditions and temporal variations, and (b) no
need for a-priori rate selection at the transmitter. In order to
substantiate these claims, we establish the pertinent capacity
limits for hybrid FSO/RF transmission and present simulation
results for transmission with off-the-shelf Raptor codes, which
achieve realized rates close to these limits under a wide range of
channel conditions. We also show that in conditions of strong
atmospheric turbulence, rateless coding is advantageous over
xed-rate coding with rate adaptation at the transmitter.
Index TermsFree-space optical (FSO), hybrid FSO/RF trans-
mission, rateless coding, channel capacity.
I. INTRODUCTION
Free space optical (FSO) communication systems have
received renewed interest due to their low deployment costs
and potential use in high-throughput applications like last mile
access [1]. The main drawback of FSO systems is their limited
and unpredictable availability due to atmospheric conditions
[2], [3]. To combat the deterioration of signal quality due
to adverse atmospheric conditions, various diversity methods
have been proposed, including the use of temporal and spatial
diversity, e.g. [4][7]. However, only media diversity schemes
are able to cope with extreme weather conditions like fog
or snow which due to scattering cause attenuation of the
order of a hundred dB per kilometer. The most prominent
media diversity scheme, which retains the above-mentioned
advantages of FSO systems, is the use of a license-free radio
frequency (RF) channel in conjunction with the FSO channel,
e.g., [1][3], [8], [9]. The practical heuristic for such hybrid
Copyright c 2010 IEEE. Personal use of this material is permitted. How-
ever, permission to use this material for any other purposes must be obtained
from the IEEE by sending a request to [email protected].
Manuscript received August 14, 2008; revised July 5 and October 29, 2009.
This work was supported by the National Sciences and Engineering Research
Council (NSERC) of Canada.
The authors have been with the Deptartment of Electrical and Com-
puter Engineering, University of British Columbia, Vancouver, Canada. Ali
AbdulHussein is now with Silvertip Telematics Inc., Vancouver, Canada
and Anand Oka is now with Research in Motion, Waterloo, Canada.
(e-mail: [email protected], [email protected], [email protected],
[email protected]).
FSO/RF systems is that fog and rain, which drastically affect
FSO and RF respectively, but only insignicantly vice versa,
rarely occur simultaneously [2], [3].
Commercially available hybrid solutions simply use the RF
link as a hot-standby backup for the FSO link, to be used only
when the FSO channel is inoperative [1]. Recently, Vangala
and Pishro-Nik [8], [9] have presented a coding scheme for
the overall hybrid channel based on non-uniform punctured
low-density parity check (LDPC) codes. This scheme relies
on knowledge of the instantaneous channel conditions at the
transmitter and the proper adjustment of code rates for FSO
and RF transmission. In this letter, we present an alternative
approach to joint FSO/RF channel coding using a practical
implementation of rateless Fountain codes [10].
1
Our scheme
falls into the class of hybrid automatic repeat-request (HARQ)
with incremental redundancy coding [14], and as such does
not require rate adjustment prior to transmission. This is
a clear distinction relative to rate-adaptive coding schemes
such as the one proposed in [8], [9]. Our motivation for the
application of rateless coding HARQ is twofold. First, it is
a very elegant approach for transmission over channels with
varying quality such as FSO/RF channels, since it requires
neither a bank of codes with various rates nor explicit code
selection. Hence, other considerations aside, one may opt
for this architecture simply due to its ease of operation.
Second, no rate mismatch due to outdated or imprecise channel
estimation can occur. This is a distinct performance advantage
over schemes with code-rate selection [8], [9], which fail
if the channel quality varies notably from one codeword to
another. In the context of hybrid FSO/RF systems, these
variations may occur especially due to temporal uctuation
in the received optical signal strength due to atmospheric
turbulence. HARQ requires reliable feedback from the receiver
in the form of negative or positive acknowledgements, and thus
the availability of a robust reverse channel for such messages
needs to be assumed. We note that this assumption is similar
to the feedback requirements for explicit rate adjustment at
the transmitter, since the information content of any feedback
signal can be made commensurate with the transmitter-side
uncertainty about the instantaneous rate supported by the
channel.
1
Parallel to our work, Zhang et al. [11] developed a soft-switching scheme
for hybrid FSO/RF links using rateless codes similar to ours, with a focus on
practical implementation aspects. Also, in course of the review of this letter,
references [12], [13] were brought to our attention. They suggest the use of
Fountain codes for a parallel multirate FSO transmission method. However,
the use of Fountain codes for hybrid FSO/RF systems is not considered in
[12], [13].
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 3, MARCH 2010 2
RF Channel
FSO Channel
RF
Demodulator
FSO
Photodector
Modulator
RF (QAM)
FSO (OOK)
xFSO()
xRF()
Modulator
M
u
lt
ip
le
x
e
r
B
it
m
e
t
r
ic
s
data, vk
Encoder
Rateless
ci
D
e
m
u
lt
ip
le
x
e
r
i BP
Decoder
vk
rRF()
rFSO()
Fig. 1. Block diagram of the proposed coded hybrid FSO/RF transmission
system.
It has been proven in [15] that todays most powerful
rateless codes, so-called Raptor codes [16], are not universal
for channels other than the erasure channel. Hence, a question
to be answered is whether there is a single rateless code
design that performs well for the variety of channel conditions
experienced in hybrid FSO/RF channels. This is particularly
relevant since FSO and RF channels are characterized by
dissimilar transmission statistics. In this letter, we provide
evidence that the answer is in the afrmative. To this end,
we derive the pertinent modulation-constrained capacity limits
for the hybrid FSO/RF channel and show through simulations
that a single moderate-length Raptor code design enables us
to closely approach these limits under a variety of channel
conditions. Our results are in line with the ndings in [15],
[17], [18] that Raptor codes designed for the binary erasure
channel perform remarkably well for binary symmetric and
(block-fading) additive white Gaussian noise (AWGN) chan-
nels too. Based on the capacity expression we further show
that in the strong-turbulence regime, the adjustment of code
rate prior to transmission will lead to rate loss and codeword
outage. Hence, the proposed rateless coding approach provides
performance advantages over xed-rate coding schemes as
presented in [8], [9] for hybrid FSO/RF systems.
The remainder of this letter is organized as follows. Sec-
tion II introduces the transmission model and presents the
proposed coding scheme. The pertinent capacity limits are
established in Section III, and simulation results for different
channel conditions are presented and discussed in Section IV.
Section V summarizes the paper.
II. TRANSMISSION MODEL AND PROPOSED CODING
SCHEME
In this section, we rst introduce the channel model and
signal schemes used for hybrid FSO/RF transmission and then
we describe the proposed coding scheme. The block diagram
of the coded hybrid transmission system is shown on Figure 1.
A. Transmission Model
1) Transmitter: The considered hybrid system consists of
two transmitter-receiver pairs, one for FSO transmission and
one for RF transmission. Both transmitters receive a bit
stream from a binary encoder, which is specied in detail in
Section II-B. At symbol time index , the RF transmitter maps
m
RF
= log
2
(M) binary symbols to an M-ary quadrature
amplitude modulation (M-QAM) signal point x
RF
() using
Gray mapping. The QAM constellation is typically non-
binary, i.e., M > 2, for bandwidth-efcient transmission.
The signal elements x
RF
() are sent with a baud rate of
1/T
RF
symbols per second using conventional pulse shapes
such as a root-raised cosine. The FSO transmitter employs
intensity modulation with on-off keying (OOK), which is the
prevalent modulation format for FSO communication. The
signal elements are x
FSO
() {0, 1} with FSO-symbol-time
index and the optical signal intensity is P
FSO
during the
on-period and zero during the off-period, each of which is of
length T
FSO
. Obviously, one data bit is represented by one
OOK symbol, i.e., m
FSO
= 1.
2) Channel and Receiver: The FSO receiver applies di-
rect detection. To formulate the transmission model, we
adopt the often used photon-count model for direct detection
with an ideal photodetector where the th detector output
r
FSO
() has a Poisson count distribution with mean parameter
g
FSO
()K
s
+K
b
if x
FSO
() = 1 and K
b
if x
FSO
() = 0. The
constants K
s
and K
b
are given by K
s
= P
FSO
T
FSO
/(hf)
and K
b
= P
b
T/(hf), where is the efciency of the pho-
todetector, f denotes the center frequency of the transmission,
h is Plancks constant, and P
b
represents the power from
background radiation incident on the photodetector [19], [6].
The variable g
FSO
() is the FSO channel gain for the th
transmission.
The RF system uses a line-of-sight link in the millimeter
wave band, which is most suitable for a powerful RF back-up
with data rates comparable to FSO. The RF receiver consists
of a classical matched-lter front-end followed by baud-rate
sampling. The sampled and phase-compensated receiver output
in the equivalent complex baseband at RF-symbol time is
written as
r
RF
() = g
RF
()x
RF
() + n
RF
() , (1)
where g
RF
() denotes the real-valued channel gain, n
RF
()
is complex-valued AWGN with variance P
n
= N
0
/T
RF
and
N
0
is the one-sided noise power spectral density.
The FSO/RF receiver passes the received samples r
FSO
()
and r
RF
() along with the parameters (g
FSO
()K
s
, K
b
) for
the FSO channel, and (g
RF
(), P
n
) for the RF channel to the
bit-metric calculator (see Figure 1).
3) Channel Statistics: The FSO channel gain g
FSO
()
includes the effects of signal attenuation and random irradiance
uctuation due to atmospheric turbulence. Signal attenuation
is strongly weather dependent and, since weather conditions
change very slowly with respect to symbol rate, can safely be
assumed to be quasi static. Signal fading due to atmospheric
turbulence, i.e., scintillation, occur on a relatively shorter time
scale. In this regard, the related literature makes different
assumptions about fading rate, ranging from quasi-static to
notable temporal changes within a few symbol intervals, cf.
e.g. [20][22].
It is known, from physical principles, that the optical
scintillation process I(t) due to random variations in the
refractive index of air is well modeled to be a stationary
process with a Gamma-Gamma marginal probability density
function (PDF) in all regimes of scintillation (weak to strong).
This model results from the observation [23] that I(t) is a
multiplication of two independent random processes I
1
(t) and
I
2
(t), corresponding to small and large scale turbulence,
I(t) = I
1
(t) I
2
(t) (2)
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 3, MARCH 2010 3
and these individual processes are accurately modeled by
Gamma marginal PDFs. Similarly, again from physical argu-
ments, the autocovariance function (ACF) B
1
() of I
1
(t) and
B
2
() of I
2
(t) (and hence B() of I(t)) can be derived in a
closed form. Both the marginal PDF and the ACF of the small
and large scale scintillation are functions of the wavelength ,
the path-length L and the structure parameter C
2
n
. In addition,
the ACFs also are a function of the wind-speed v. Please refer
to [23] for more details.
In this work, we adopt the scintillation model according
to (2) with the process I(t) matching the marginal PDF and
ACF given in [23, Chapter 3.4]. We note that specifying the
marginal PDF and ACF does not uniquely specify the statis-
tical properties of I(t). However, since no reliable expression
is known for any higher order statistics, this appears to be the
most reasonable approach. Applying sampling at baud rate
1/T
FSO
we obtain the discrete-time scintillation process I()
and thus the FSO channel gain g
FSO
().
With regards to the millimeter wave RF channel, we observe
from e.g. [24] that the gain g
RF
() will remain practically
constant for several seconds and thus the assumption of a
quasi-static RF channel whose gain does not change during
the transmission of (at least) one codeword is validated.
B. Proposed Coding Scheme
1) Encoder and Multiplexing: The hybrid FSO/RF system
employs a single encoder of a binary rateless code. More
specically, we apply Raptor codes proposed in [16], whose
encoder consists of two stages. The rst stage encodes the
k-bit input vector [v
1
, v
2
, . . . , v
k
] into a k

-bit intermediate
codevector [v

1
, v

2
, . . . , v

k
] using a rate-k/k

linear binary
block code, for example an LDPC code. In the second stage,
a Luby transform (LT) [10] encoder converts the k

-bit LDPC
codeword into a semi-innite stream of Raptor-code bits c
i
,
i = 1, 2, . . ..
The code bits are then demultiplexed into two bit-streams,
one entering the RF QAM modulator and the other one being
input to the FSO OOK modulator. To be more specic, let us
dene the positive, mutually prime integers n
FSO
and n
RF
as
n
FSO
n
RF
=
m
FSO
/T
FSO
m
RF
/T
RF
. (3)
Then, out of a block of n
tot
= n
FSO
+ n
RF
bits
[c
i
, c
i+1
, . . . , c
i+ntot1
], the rst n
FSO
bits are passed to
the FSO modulator and the remaining n
RF
bits are RF
modulated. The formulation in (3) makes the mild assumption
that m
FSO
T
RF
/(m
RF
T
FSO
) can be expressed as a rational
number.
Encoding of the message [v
1
, v
2
, . . . , v
k
] is terminated
when a positive acknowledgment of successful decoding has
been received.
2) Metric Computation and Demultiplexing: Based on the
received samples and the channel parameters, decoding met-
rics in the form of log-likelihood ratios are generated.
For FSO transmission with the channel and detector model
described in Section II-A the log-likelihood ratio (LLR) asso-
ciated with x
FSO
() can be written as [25, Eq. (4.5-4)]

FSO
() = r
FSO
() log
_
1 +
g
FSO
()K
s
K
b
_
g
FSO
()K
s
.
(4)
For RF transmission the LLR for the th bit mapped to x
RF
()
is given by (cf. [26, Eq. (7)])

RF
() = log
_

xX
,0
exp
_

|rRF()gRF()x|
2
Pn
_
_
log
_

xX
,1
exp
_

|rRF()gRF()x|
2
Pn
_
_
,
(5)
1 m
RF
, where X
,b
denotes the subset of the QAM
signal constellation for which the th bit of the label is equal
to b {0, 1}. According to the demultiplexing of code bits
described above, groups of n
FSO
LLRs
FSO
() and n
RF
LLRs

RF
() are multiplexed into one stream of LLRs
i
,
i = 1, 2, . . ., which then is forwarded to the decoder (see
Figure 1).
3) Decoder: The decoder for the rateless code collects LLRs

i
progressively and a rst decoding attempt is made when it is
believed that sufcient information for successful decoding has
been collected. A good estimate for the number n of collected
LLRs is given by
n

i=1
C
H,i
= k(1 + ) , (6)
where C
H,i
denotes the constrained capacity of the hybrid
FSO/RF channel as dened in Section III below, and > 0
is an overhead factor which accounts for the suboptimality of
the (nite-length) Raptor code with respect to capacity.
Decoding is performed using well-known belief propagation
(BP) on the factor graph of the Raptor code. The variable
nodes c
i
are initialized with the LLRs
i
, 1 i n,
and extrinsic likelihood messages are exchanged between
the variable and parity-check nodes. If the decoded message
[ v
1
, v
2
, . . . , v
k
] after BP decoding is deemed unreliable,
which could be determined by either making use of a cyclic
redundancy check (CRC) code or estimating the bit-error rate
(BER) based on the nal LLRs [27], a new decoding attempt is
scheduled after a new batch of n
inc
samples has been received.
The value of n
inc
will depend on the update schedule applied
for BP decoding. Complexity-efcient update schedules have
been presented in [28], [29].
Once the decoded message [ v
1
, v
2
, . . . , v
k
] is deemed
correct, decoding is terminated and an acknowledgment signal
is transmitted to the encoder.
4) Remark: We assume that modulation and signalling rate,
and thus n
FSO
and n
RF
are constant. However, one may
opt for adaptive modulation and even adaptive baud rate as
a function of the channel conditions. In this case, rateless
coding as proposed can still be used in tandem with these
to achieve a ne-grained adaptation to the channel conditions
(with granularity of n
inc
bits).
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 3, MARCH 2010 4
III. CONSTRAINED CAPACITY OF THE HYBRID FSO/RF
CHANNEL
In order to benchmark the proposed coding scheme, we now
specify the information-theoretic limit for the hybrid FSO/RF
channel. To this end, we rst note that the FSO and RF
modulation formats are not optimized for maximal mutual
information and assumed constant regardless of the channel
condition. This is not a particularly limiting assumption since
(a) practical FSO use OOK modulation and (b) as long as the
size M of the QAM constellation used for RF transmission
is chosen such that m
RF
is at least 1 bit above the Shannon
capacity of the RF channel, hardly any gains are achievable
with a larger alphabet size [30]. Hence, choosing a reasonably
large constellation size, say 64QAM, is sufciently optimal.
Secondly, we note that we are not interested in the ergodic
capacity of the hybrid channel due to the typically large
coherence times of the RF and FSO scintillation processes (see
Section IV-B). Instead, we are interested in the capacity given
the sequences of channel gains [g
FSO
(
s
), . . . , g
FSO
(
e
)] and
[g
RF
(
s
), . . . , g
RF
(
e
)] during the transmission of n bits,
where n = (
e

s
+ 1)m
FSO
+ (
e

s
+ 1)m
RF
and
(
e

s
+ 1)/(
e

s
+ 1) = T
RF
/T
FSO
.
Let us consider the binary channel from c
i
to
i
(see
Figure 1), and denote by
i
and
i
the FSO and RF channel
symbol-time indices corresponding to the transmission of c
i
.
Then, since the FSO and RF channels are used in parallel, the
constrained hybrid-channel capacity can be written as
C
H,i
=
1
n
FSO
+ n
RF
_
n
FSO
m
FSO
C
FSO
(g
FSO
(
i
))
+
n
RF
m
RF
C
RF
(g
RF
(
i
))
_
(7)
with unit [bit/binary channel use], where C
FSO
(g
FSO
())
and C
RF
(g
RF
()) respectively denote the constellation-
constrained capacities of the FSO and RF channel in bit per
FSO and RF channel use for a given channel gain. These are
given by (cf. [26, Eq. (14)])
C
FSO
(g
FSO
()) =
1 E
_
log
2
_
1 + exp
_
(1)
xFSO()

FSO
()
___
(8)
C
RF
(g
RF
()) =
m
RF

mRF

=1
E
_
log
2
_
1 + exp
_
(1)
b

RF
()
__
,(9)
where b

denotes the th bit in the label of x

RF
(), and
expectation E{} is with respect to the joint probabilities
p(r
FSO
(), x
FSO
()) and p(r
RF
(), b

), respectively. We note
that neither C
FSO
(g
FSO
()) nor C
RF
(g
RF
()) lend them-
selves for closed-form expressions, and hence they are evalu-
ated using Monte Carlo integration.
IV. SIMULATION RESULTS
In this section, we present numerical results and simulative
evidence that (a) the proposed coded hybrid FSO/RF scheme
performs close to the information-theoretic limits and (b)
rateless coded transmission has advantages over xed-rate
schemes with rate selection for hybrid FSO/RF systems.
A. Simulation Setup
1) System Parameters: The considered hybrid system em-
ploys 64QAM modulation for RF transmission, i.e., m
RF
= 6,
which is practically optimal as long as C
RF
(g
RF
()) 5
bit/(RF channel use) [30]. As specied in Section II-A1, OOK
modulation is used for FSO signalling and thus m
FSO
= 1.
The FSO baud interval is T
FSO
= 10
8
s and the RF baud
interval is adjusted to T
RF
=
mRF
mFSO
T
FSO
= 6T
FSO
such that
n
RF
= n
FSO
= 1 (see Eq. (3)). Furthermore, we assume a
wavelength of 1550 nm, a photodetector efciency of = 0.5,
and a normalized background radiation of P
b
T = 170 dBJ
[6], which yields the average background photon count K
b
=
39. As described in Section II-A3, the RF channel is assumed
quasi-static, i.e., g
RF
() = g
RF
, while the FSO channel gain
g
FSO
() varies according to the scintillation process I(t).
For simulations we apply a Raptor code consisting of a
rate-0.95 regular LDPC code and an LT code generated using
the degree distribution from [16, Table I, 2nd column]. The
information word length is chosen as k = 9500. Decoding is
continued until the correct codeword has been found, which
emulates the use of a CRC outer code. The number n of
code bits required for successful decoding is recorded and
the realized rate is obtained as R = k/n. The achievable
rate according to the capacity limit C
H,i
in (7) is given by

R = k/ n, where n is dened through

n

i=1
C
H,i
= k . (10)
The size of decoding increments n
inc
is set to 0.5% of the
expected word length n and BP decoding is started afresh after
each newly received batch. For termination of transmission
after successful decoding, we assume the availability of a
reliable feedback channel.
2) A Note on the Generation of the Scintillation Process:
Realizations of the scintillation process I(t) are generated by
rst individually generating I
1
(t) and I
2
(t) and then using
Eq. (2). To match the specied marginal PDFs and ACFs for
I
1
(t) and I
2
(t), the simulated processes are generated by using
the randomized Markov chain (MC) algorithm suggested in
[31]. The set of transition probabilities of the MC needs to
be adjusted (in a one time off-line operation) by solving a
non-linear optimization problem to match the ACF. We have
used a commercially available non-linear optimization routine
which can practically handle up to 16 MC states and found
that the resulting solution gave a sufciently good match for
the task at hand.
B. Results and Discussion
First, Figure 2 shows the PDF and ACF of the FSO
scintillation process I(t) for an FSO link of L = 1000 m,
a wind-speed of v = 5 m/s, and different structure parameters
C
2
n
. The corresponding Rytov variances
2
1
and scintillation
indices S.I., which are common measures for the strength of
irradiance uctuations, are given in the caption of Figure 2
and represent strong to weak turbulence scenarios, cf. [23].
Also included are the statistics for the simulated process for
C
2
n
= 10
13
m
2/3
. We observe that the channel coherence
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 3, MARCH 2010 5
10
1
10
0
10
1
10
2
0
0.05
0.1
0.15
0.2
0.25


10
4
10
3
10
2
0
0.2
0.4
0.6
0.8
1
True process
Simulated Process
C
2
n
= 10
12
m
2/3
C
2
n
= 10
12
m
2/3
C
2
n
= 10
13
m
2/3
C
2
n
= 10
14
m
2/3
C
2
n
= 10
15
m
2/3
I
B
(

)
/

2I

p
(
I
)

[seconds]
Fig. 2. Statistics of the true gamma-gamma scintillation process I(t), as
well as the simulated process, at various values of the structure parameter C
2
n
.
Left: PDF of I. Right: ACF of I(t) normalized with respect to its variance

2
I
. Path-length L = 1000 meters, wavelength = 1550 nm, wind-velocity
v = 5.0 m/s, E[I] = 10. (Rytov variance
2
1
, scintillation index S.I.) =
{(19.9, 1.48), (1.99, 0.99), (0.199, 0.173), (0.019, 0.018)} respectively for
C
2
n
= {10
12
, 10
13
, 10
14
, 10
15
} m
2/3
.
time is on the order of milliseconds and decreases for relatively
strong turbulence scenarios. More signicant though are the
differences in the PDF, which indicates the occurrence of
deep fades in the strong turbulence regime. It can further be
seen that the statistics of the simulated process are a good
match to the true statistics. For the following performance
results we assume the above parameters L and v, and that
C
2
n
= 10
13
m
2/3
and thus relatively strong turbulence with
(
2
1
, S.I.) = (1.99, 0.99).
Figure 3 compares the rate R achieved with rateless cod-
ing and the achievable rate

R according to equation (10)
for the hybrid FSO/RF system and 1000 transmitted code-
words. The FSO channel supports an average rate of
E{C
FSO
(g
FSO
())} = 0.3 bit/(FSO channel use), while the
quasi-static RF channel gain is adjusted such that C
RF
=
{2.0, 6.0} bit/(RF channel use). We observe that the rateless
coded system performs consistently close to the capacity limit,
regardless of the instantaneous FSO or RF channel gains and
thus achievable rate. In particular, the overhead (see (6))
remains between about 10-20 % in all cases and thus, the same
rateless code design performs well for very different mixtures
and qualities of FSO and RF channels. We also observe the
boost in realized and achievable rate due to a stronger RF
channel, which will be further elaborated below.
The ability of the rateless coding scheme to adapt the
instantaneous channel conditions is further highlighted in
Figure 4, which shows for 200 transmitted codewords that
the required codeword length n follows the lower limit n
from equation (10). The results in Figures 3 and 4 clearly
support the idea of rateless coding to achieve close-to-capacity
performance in hybrid FSO/RF systems. In particular, the use
of a single degree distribution, taken from [16, Table I, 2nd
column], is justied by the consistency with which the capacity
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Achievable rate [bit/binary channel use] >
R
e
a
l
i
z
e
d

r
a
t
e

[
b
i
t
/
b
i
n
a
r
y

c
h
a
n
n
e
l

u
s
e
]

>


C
RF
=2 bit/(RF channel use)
C
RF
=6 bit/(RF channel use)
Fig. 3. Scatter plot of realized rate R versus achievable rate

R according
to equation (10). 1000 codewords have been transmitted. FSO channel with
(
2
1
, S.I.) = (1.99, 0.99) and E{C
FSO
(g
FSO
())} = 0.3 bit/(FSO channel
use). Quasi-static RF channel with C
RF
= {2.0, 6.0} bit/(RF channel use).
0 20 40 60 80 100 120 140 160 180 200
1
2
3
4
5
6
7
x 10
4
index of transmitted codeword >
n
u
m
b
e
r

o
f

c
o
d
e
d

b
i
t
s

u
n
t
i
l

s
u
c
c
e
s
s
f
u
l

d
e
c
o
d
i
n
g

>


rateless code
lower limit
Fig. 4. Number of transmitted coded bits until successful decoding for
200 codeword transmissions. Simulated codeword length n using rateless
codes and lower limit n from equation (10). FSO channel with (
2
1
, S.I.) =
(1.99, 0.99) and E{C
FSO
(g
FSO
())} = 0.3 bit/(FSO channel use). Quasi-
static RF channel with C
RF
= 2.0 bit/(RF channel use).
limit is approached. Similar observations have been made in
[13, 14] for transmission over AWGN channels. This renders
rateless coding a very attractive approach as no adjustment of
the code structure is necessary.
The effect of the RF back-up, used concurrently with the
FSO link operating in strong turbulence conditions, is illus-
trated in Figure 5, where we plot the measured complementary
cumulative density function (CCDF) of n for k = 9500 and
RF channels with different capacities. It can be seen that
the scintillation process I(t) and the resulting variation of
g
FSO
() leads to large uctuations of n, which in turn makes
the achievable throughput a random variable. The integration
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 3, MARCH 2010 6
10
5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1


Complementary CDF
Mean value
CRF = 0.0
CRF = 0.6
CRF = 1.2
CRF = 1.3
x
CRF = 2.4
P
r
(
n
>
x
)

Fig. 5. Complementary CDF of n from equation (10) with k = 9500

for various values of C
RF
(unit is bit/(RF channel use)). FSO channel with
(
2
1
, S.I.) = (1.99, 0.99) and E{C
FSO
(g
FSO
())} = 0.3 bit/(FSO channel
use).
of the RF channel through joint coding moderates (dampens)
the throughput uctuations, with more pronounced effect for
large RF channel capacity. In particular, in accordance with
information theory, the mean of the distribution does decrease
monotonically as the RF capacity increases. The observation
that the curve for C
RF
= 0 intersects with curves for C
RF
> 0
is not a contradiction, but reects the larger variance of the
required number of received samples n for smaller C
RF
.
Finally, it is interesting to consider the change in required
codeword length, and thus code rate, for successively trans-
mitted codewords. For this purpose, Figure 6 plots the CCDF
of n, which is the difference of two successive values of n,
for k = 9500 and RF channels with different capacities. We
observe that notable variations in required codeword length
from one codeword to the next occur with high probability,
especially when C
RF
is low. This implies that coding schemes
using xed-rate codes with a-priori rate adjustment based
on the prior received word, as in [8], [9], will be affected
by a signicant rate mismatch. That is, either a rate loss is
suffered or outages occur. In contrast, as conrmed by our
simulations, the proposed rateless coded system is able to
seamlessly adapt to these changes. To provide one concrete,
quantitative example, we consider a xed-rate coding scheme
for the scenario with C
RF
= 1.2 bit/(RF channel use) from
Figures 5 and 6. The code rate for the ith codeword is chosen
as R
i
= (1 + )k/ n
i1
, where n
i1
is the achievable rate
from (10) for the transmission of the (i 1)st codeword. That
is, the receiver provides the transmitter with most up-to-date
channel information before transmission. is a margin to
account for channel variations. Then, even when assuming that
a decoding error occurs only if R
i
> k/ n
i
(which is a highly
optimistical assumption for this scheme), we obtain outage
rates of 17% for = 10% and 7% for = 20%. We can thus
conclude that the rateless coding scheme is not only an elegant
approach to utilize resources in hybrid FSO/RF systems, but
10
3
10
4
10
5
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1


Complementary CDF
Mean value
x
CRF = 1.8
CRF = 1.2
CRF = 0.6
CRF = 0.0
CRF = 2.4
P
r
(

n
>
x
)

Fig. 6. Complementary CDF of the successive difference n of n

from equation (10) with k = 9500 for various values of C
RF
(unit is
bit/(RF channel use)). FSO channel with (
2
1
, S.I.) = (1.99, 0.99) and
E{C
FSO
(g
FSO
())} = 0.3 bit/(FSO channel use).
also enjoys a distinct throughput advantage over coding with
a-priori rate selection in strong turbulence situations.
V. CONCLUSIONS
We have proposed the application of rateless codes for
hybrid FSO/RF transmission systems. Hybrid FSO/RF systems
combine the best of two worlds to render the system perfor-
mance robust to short and long-term variations of the FSO and
RF transmission channel. This is critical for the adoption of
FSO as a reliable high-speed access technology. The distinct
feature of the proposed scheme is that it (a) enables the
realization of the potential advantages due to parallel FSO and
RF channels without the need for redesign or reconguration
of the transmitter-side coding or modulation and (b) adapts
seamlessly to the changes in rate supported by the channel.
We have established the pertinent information-theoretic limits,
which show that coding schemes with transmitter side rate
selection may suffer from rate loss or outages depending on
the rate of channel uctuation. Simulative evidence has been
provided that the proposed rateless coding scheme, using an
off-the-shelf Raptor code, well approaches the information-
theoretic limits regardless of channel conditions.
REFERENCES
[1] S. Bloom and W. Hartley, The last-mile solution: Hybrid FSO radio,
in Whitepaper, AirFiber Inc., May 2002.
[2] I. I. Kim and E. Korevaar, Availability of free space optics (FSO) and
hybrid FSO/RF systems, in Proc. SPIE, Optical Wireless Communica-
tions IV, vol. 4530, Denver, CO, USA, Aug. 2001, pp. 8495.
[3] J. Pan, M. Evans, T. Euler, H. Johnson, and F. DeNap, Free-space
optical communications: Opportunities and challenges, a carriers per-
spective, in Proc. SPIE, Wireless and Mobile Communications II, vol.
4911, 2002, pp. 5872.
[4] V. Chan, Coding for the turbulent atmospheric optical channel, IEEE
Trans. Commun., vol. 30, pp. 269275, Jan. 1982.
[5] E. Lee and V. Chan, Optical communication over the clear turbulent
atmospheric channel using diversity, IEEE J. Select. Areas Commun.,
vol. 22, pp. 18961906, Nov. 2004.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 3, MARCH 2010 7
[6] S. Wilson, M. Brandt-Pearce, Q. Cao, and J. Leveque, III, Free-space
optical MIMO transmission with Q-ary PPM, IEEE Trans. Commun.,
vol. 53, no. 8, pp. 1402 1412, Aug. 2005.
[7] M. Safari and M. Uysal, Do we really need OSTBCs for free-space
optical communication with direct detection? IEEE Trans. Wireless
Commun., vol. 7, no. 11, pp. 44454448, November 2008.
[8] S. Vangala and H. Pishro-Nik, Optimal hybrid RF-wireless optical
communication for maximum efciency and reliability, in Proc. 41st
Annual Conference on Information Sciences and Systems (CISS), Mar.
2007, pp. 684689.
[9] , A highly reliable FSO/RF communication system using efcient
codes, in Proc. IEEE Global Telecommunications Conference (GLOBE-
COM), Washington, DC, USA, Nov. 2007, pp. 22322236.
[10] M. Luby, LT codes, in IEEE Symp. on Foundations of Comp. Sc.
(FOCS), Vancouver, BC, Canada, Nov. 2002, pp. 271280.
[11] W. Zhang, S. Hranilovic, and C. Chi, Soft-switching hybrid FSO/RF
links using short-length raptor codes: Design and implementation, IEEE
J. Select. Areas Commun., vol. 27, no. 9, pp. 16981708, Dec. 2009.
[12] S. Navidpour, B. Hamzeh, and M. Kavehrad, Multirate fractal free
space optical communications using Fountain codes, in Proc. SPIE,
Optical Transm. Syst. and Equipm,. for WDM Netw., vol. 6012, 2005,
pp. U1U7.
[13] M. Kavehrad, S. Navidpour, and S. Lee, Fractal transmission in a hybrid
RF and wireless optical link; a reliable way to beam bandwidth in a 3D-
global grid, in Proc. SPIE, Broadband Access Commun. Techn., vol.
6390, 2006, pp. E1E11.
[14] C. Lott, O. Milenkovic, and E. Soljanin, Hybrid ARQ: Theory, state
of the art and future directions, IEEE Information Theory Workshop on
Information Theory for Wireless Networks, pp. 15, July 2007.
[15] O. Etesami and A. Shokrollahi, Raptor codes on binary memoryless
symmetric channels, IEEE Trans. Inform. Theory, vol. 52, no. 5, pp.
2033 2051, May 2006.
[16] A. Shokrollahi, Raptor codes, IEEE Trans. Inform. Theory, vol. 52,
no. 6, pp. 2551 2567, Jun. 2006.
[17] R. Palanki and J. Yedidia, Rateless codes on noisy channels,
in Proc. IEEE International Symposium on Information Theory
(ISIT), Chicago, IL, USA, June-July 2004, p. 37, (see also
www.merl.com/papers/TR2003-124/).
[18] J. Castura and Y. Mao, Rateless coding over fading channels, IEEE
Commun. Lett., vol. 10, no. 1, pp. 4648, Jan. 2006.
[19] R. Gagliardi and S. Karp, Optical Communications. New York: John
Wiley & Sons, 1976.
[20] H. Henniger, F. David, D. Giggenbach, and C. Rapp, Evaluation of FEC
for the atmposheric optical IM/DD channel, in Proc. SPIE, Free-Space
Laser Communications XV, vol. 4975, 2003, pp. 111.
[21] X. Zhu, J. Kahn, and J. Wang, Mitigation of turbulence-induced
scintillation noise in free-space optical links using temporal-domain
detection techniques, IEEE Photon. Technol. Lett., vol. 15, no. 4, pp.
623625, April 2003.
[22] M. Uysal, S. Navidpour, and J. Li, Error rate performance of coded free-
space optical links over strong turbulence channels, IEEE Commun.
Lett., vol. 8, no. 10, pp. 635637, Oct. 2004.
[23] L. Andrews, R. Phillips, and C. Hopen, Laser Beam Scintillation with
Applications. Bellingham, WA: SPIE Press, 2001.
[24] F. Medeiros Filho, D. Jayasuriya, R. Cole, and C. Helmis, Spectral
density of millimeter wave amplitude scintillations in an absorption
region, IEEE Trans. Antennas Propagat., vol. 31, no. 4, pp. 672676,
Jul 1983.
[25] S. Dolinar, J. Hamkins, B. Moision, and V. Vilnrotter, Optical modula-
tion and coding, in Deep Space Optical Communications, H. Hemmati,
Ed. Wiley-Interscience, Apr. 2006.
[26] G. Caire, G. Taricco, and E. Biglieri, Bit-interleaved coded modula-
tion, IEEE Trans. Inform. Theory, vol. 44, no. 3, pp. 927946, May
1998.
[27] P. Hoeher, Adaptive modulation and channel coding using reliability
information, in Proc. of 5th International OFDM-Workshop, Hamburg,
Germany, Sep. 2000, pp. 14.114.4.
[28] K. Hu, J. Castura, and Y. Mao, Performance-complexity tradeoffs of
Raptor codes over Gaussian channels, IEEE Commun. Lett., vol. 11,
no. 4, pp. 343 345, Apr. 2007.
[29] A. AbdulHussein, A. Oka, and L. Lampe, Decoding with early termi-
nation for Raptor codes, IEEE Commun. Lett., vol. 12, no. 6, pp. 444
446, Jun. 2008.
[30] L. Ozarow and A. Wyner, On the capacity of the Gaussian channel with
a nite number of input levels, IEEE Trans. Inform. Theory, vol. 36,
no. 11, pp. 14261428, Nov. 1990.
[31] A. S. Rodionov, H. Choo, H. Y. Youn, and T. M. Chung, Generation
of Pseudo Random Processes with a Given Marginal Distribution and
Autocorrelation Function, Intl. Journal of Computer Mathematics,
vol. 78, no. 4, pp. 521537, Jan. 2001.