IEEE PHOTONICS TECHNOLOGY LETTERS, VOL. 29, NO.

9, MAY 1, 2017 727

Optimization of LDPC Codes for PIN-Based OOK

FSO Communication Systems
Jun Ao, Jiwei Liang, Chunbo Ma, Guixing Cao, Cong Li, and Yufei Shen

Abstract— For middle/short-range terrestrial high-speed free- of LDPC codes. Using an optimization scheme combining
space optical (FSO) communication systems employing ON–OFF four different search algorithms, we optimized irregular LDPC
keying (OOK) and PIN photodiode receivers, the noise distrib- codes for these systems. Simulation results indicate that the
ution is symbol-dependent, and the symmetry of the underlying
communication channel cannot be assumed. Therefore, this letter performance of the codes is very close to asymptotic theoret-
employs the asymmetric density evolution technique and an ical bounds.
optimization scheme combining four different search algorithms
to optimize low-density parity-check (LDPC) codes for these II. O N -O FF K EYING W ITH PIN P HOTODIODE
FSO systems. Compared with the conventional LDPC codes that O UTPUT M ODELLING
are designed assuming symbols-independent symmetric channels
the optimized codes are much closer to the Shannon limit for For OOK, the average numbers of photons k̄ x respectively
OOK modulated binary non-symmetric channels. absorbed by a PIN photodiode receiver for symbols x = 0
Index Terms— Free-space optical communication, channel and x = 1 over the duration of the bit interval (slot) Ts can be
coding. expressed as

I. I NTRODUCTION Ts
R
k̄ x = [Pb (t) + x · Ps (t)]dt, x = 0 or 1. (1)
S IGNIFICANT efforts have been made to apply on-off
keying (OOK) and PIN photodiode receivers to middle/
short-range terrestrial high-speed free-space optical (FSO)
e
0

Here, Ps (t) and Pb (t) (W/m2 ) respectively represent the opti-

communication systems due to the abundant spectrum and low cal signal intensity and the intensity of background radiation
cost of PIN photodiode receivers. However, system perfor- incident on the PIN photodiode at time t, R is the responsivity
mance can be affected severely by various noise sources such of the detector, and e = 1.6×10−19C is the electron charge.
as background optical, shot, and thermal noise. Consequently, The electrons output by a PIN photodiode derive from two
error correction coding (ECC) is applied to improve the per- sources: the electrons excited by signal and background light,
formance of FSO communication systems. Low-density parity- and the electrons produced by the dark current and indepen-
check (LDPC) codes are considered the best ECC known today dent thermal noise of follow-on electronics. The total slot
and approach the Shannon limit very closely with acceptable statistic y given k̄ x absorbed photons is approximated by a
encoding complexity [1]. Gaussian process of mean m x and variance σx2 [3]
For FSO communication systems employing OOK and
PIN photodiode receivers, the noise distribution is symbol- m x = k̄ x + i s · T /e (2)
  
dependent. As a result, the symmetry of the underlying com- σx2 = k̄ x + 2B · Ts · i s Ts /e + 2kb T · Ts / R f · e 2
(3)
munication channel, which is fundamental for conventional
LDPC code design, cannot be assumed. Therefore, this letter where i s is the dark current of the detector, T is the equivalent
considers the optimization of LDPC codes for FSO communi- noise temperature in Kelvin, Bis the noise bandwidth, kb =
cation systems employing OOK and PIN photodiode receivers, 1.38 × 10−23 J/K is the Boltzmann constant and R f is the
where the symbol-dependent noise distribution is addressed load resistance. Therefore, the channel is memoryless and is
by adopting asymmetric density evolution [2] for modelling specified by the conditional probability density function
of OOK signal behaviour with belief-propagation decoding 1 −(y − m x )2
p(y|x) =  exp( ) (4)
Manuscript received December 21, 2016; revised February 9, 2017; 2πσx2 2σx2
accepted March 8, 2017. Date of publication March 15, 2017; date
of current version April 13, 2017. This work was supported by Generically, the channel output alphabet is equal to the decoder
the program for National Natural Science Foundation of China under input alphabet. Let u 0 = ln p(x=0|y)
p(x=1|y) be the initial log-likelihood
Grant 61167006. (Corresponding author: Jiwei Liang.) ratio (LLR) of the transmitted OOK symbol x conditioned on
J. Ao, J. Liang, and C. Ma are with the Key Laboratory of Cognitive
Radio and Information Processing, Ministry of Education, Guilin University of the slot statistic y. Assuming that 0 and 1 are equally likely,
Electronic Technology, Gulling 541004, China (e-mail: [email protected]; by Bayes’ rule we have [4]
[email protected]; [email protected]).
p (x = 0|y) p (y|x = 0)
G. Cao, C. Li, and Y. Shen are with the Department of Satellite Communi- u 0 = ln = ln
cation, China Academy of Space Technology, Beijing 100094, China (e-mail: p (x = 1|y) p (y|x = 1)
[email protected]; [email protected]; [email protected]).  
Color versions of one or more of the figures in this letter are available σ1 − (y − m 0 )2 (y − m 1 )2
= ln + + = LLR (y) (5)
online at http://ieeexplore.ieee.org. σ0 2σ02 2σ12
Digital Object Identifier 10.1109/LPT.2017.2682269
1041-1135 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
728 IEEE PHOTONICS TECHNOLOGY LETTERS, VOL. 29, NO. 9, MAY 1, 2017

So the probability density p0x of the initial LLR, which as

⎛ ⎞
is also the initial average density for messages emitted from 0 ∞
variable nodes, can be expressed as follows 1⎜ ⎟
pel = ⎝ p0l (du l ) + p1l (du l )⎠ (9)
    2
− L L R −1 u 0 − m 2 u l =−∞ u l =0
1 x
p0x =  exp
2πσx2 2σx2 For any integrable function h, we have
      
d L L R −1 u 0 h u l p0l d u l
· (6)
du 0   
p y l |x = 0
= E x=0 h ln  l
p y |x = 1
III. A SYMMETRIC AND C ODEWORD -AVERAGED    
p y |x = 0
l p y l |x = 0
D ENSITY E VOLUTION = E x=1  h ln  l
p y l |x = 1 p y |x = 1
For p0x , as defined in (6), it is easily verified that     
l
p00 = p10 , indicating that the symmetry condition assumed for = eu h u l p1l d u l (10)
the analysis of conventional LDPC codes cannot be satisfied.
To circumvent this problem, the asymmetric density evolution  ul
Assuming h u l = e− 2 , then
for irregular codes C (λ, ρ) which considers a uniform average  
density of the messages over all valid tree-satisfying and cycle- eu /2 p1l du l = e−u /2 p0l du l
l l
(11)
free codes with projections being on bit x is adopted [2]. Given
a degree distribution pair (λ, ρ), the variable and check node By (11) and the perfect projection assumption, the Chernoff
degree distribution pairs are specified by λ (x) = λk x k−1 bound of the error probability for different symbols x are
k
andρ (x) = ρk x k−1 . More precisely, λk and ρk denote the equivalent, and can therefore be expressed as an averaged
k bound
fraction of edges connecting to degree k variable and check   1 
nodes respectively.
 And
 the
 design rate of codes R is equal to C B Pl = C B P0l + C B P1l
2 
1 1
R = 1 − 0 λ (x) d x / 0 ρ (x) d x . ul
   
= C B P0 = C B P1 = e− 2 pl d u l
l l
The asymmetric density evolution formula for C (λ, ρ) can
be expressed as follows −u l
= E(exp( )) ≥ pel (12)
2
plx = p 0x ⊗ λ(q xl−1 )   
where pl := p0l + p1l /2. It was proved in [2] that if
p0l−1 + p1l−1   
qxl−1 =  −1 (ρ(( )) e−u /2 pl d u l is symmetric with respect tou l = 0, the
l
2
Chernoff bound of pel is a nonincreasing function of l and will
pl−1 − p1l−1
+(−1)x ρ(( 0 ))) (7) always converge to a certain value, which  might be zero. With
2 0  
the regularity assumption that R esu p 0 d u 0 < ∞ for all s
Here, plx and q xl are the average probability densities for in neighbourhood of zero, the sufficient and necessary stability
messages emitted from variable nodes and check nodes at conditions for the average fraction of incorrect messages
the l-th iteration respectively, is the density transformation pel under asymmetric density evolution are summarized as
function. follows:  
The stability of asymmetric density evolution on BP decod- Sufficiency: Ifλ2 ρ  (1) C B P0 < 1 then there exists a
ε= 
 ε(λ(ρ (1), C B P  ),ε >l 0, such that if for some
constant 0
ing algorithm should be proved to analyze the convergence of 
bit error probability. In fact, it is hard to prove the convergence l0 , C B P < ε, then lim
l 0
 l→∞  C B P = 0.
of bit error probability directly. So we focus on the Chernoff Necessity: Ifλ2 ρ  (1) C B P 0 > 1, then liml→∞ pel > 0.
upper bound of bits error probability. If we can prove the
Chernoff upper bound of bits error probability is convergent, IV. O PTIMIZATION S CHEME
the bit error probability will be convergent. Thus the stability Four search algorithms, including differential evolution
of asymmetric density evolution on BP decoding algorithm is (DE), genetic algorithm (GA), particle swarm optimization
proved. By Bayes’ rule, we have (PSO) and simulated annealing (SA), were combined to
   l  search for optimized degree distribution pairs that yield the
P x = 0|y l P y |x = 0 minimum possible signal to noise ratio (SNR) threshold for
u l = ln  = ln  (8)
P x = 1|y l P y l |x = 1 a given bit error rate (BER) [5]. And SNR is defined as
 S N R = (m 1 − m 0 )2 /σ02 [6]. The search process begins with
where P x|y l is the conditional distribution function of p0x corresponding to an initial SNR threshold. The degree
x assuming y l is messages coming from the channel and the distribution pairs at a predetermined code rate are randomly
connected edges at the l-th iteration [2]. Therefore, the average generated and partitioned into four populations, which are
fraction of incorrect message after l iterations can be defined applied to the four search algorithms, respectively. Using the
AO et al.: OPTIMIZATION OF LDPC CODES FOR PIN-BASED OOK FSO COMMUNICATION SYSTEMS 729

Fig. 2. Evaluation of the BER performance for different degree distribution

pairs under density evolution as a function of SNR.

TABLE I
SNR T HRESHOLD A RE C OMPUTED AT A BER OF
10−6 OF D IFFERENT D EGREE D ISTRIBUTIONS
Fig. 1. Structure of optimization scheme employing DE, GA, PSO and SA.

density evolution in (7) to calculate the error probability as a

cost function, the four algorithms execute search and update
operations in parallel. After a predetermined number of updat-
ing iterations, the best updated degree distribution pair which
has the lowest BER, is inserted into updated populations and
the four search algorithms exchange populations randomly.
If an updated degree distribution pair exhibits a BER less than
the benchmark, we reduce the SNR threshold and restart the
search process. The structure of the optimization scheme is
illustrated in Fig. 1.
In order to accelerate the search for good degree distribution
pair, the search spaces of the nonzero variable node degrees are much closer to the Shannon limit [11]. Fig. 2 shows the
are limited to (2, 3, 4, 5) , (2, 3, 4, 10) and (2, 3, 10, 20) ,and evaluate results of the BER performance for different degree
nonzero check node degrees are limited to (6, 7) , (7, 8) distribution pairs under density evolution as a function of SNR.
and (7, 8)respectively. The resulting degree distribution pairs The maximum allowed number of iterations was 2000. When
of rate-1/2 codes for the FSO communication systems employ- the decoding came to an end with 40 decreases in BER of
ing on-off keying (OOK) and PIN photodiode receivers are less than 1 × 10−12 per iteration, about 180 - 400 iterations
shown below were necessary. As shown in Fig. 2, the BER performances of
the degree distribution pairs in 802.11n and 802.16e standards
Degree I : are the worst. The profiles in [10] have waterfall regions at

λ(x) = 0.1362x + 0.435x 2 +0.12475x 3 + 0.416325x 4 lower SNR than proposed degree distribution pairs, but they
have high error floors. Although, the performance curve of the
ρ(x) = 0.352x 5 + 0.648x 6
degree distribution pair in the CCSDS Optical Uplink standard
Degree II : does not show the error floor in Fig.2, its waterfall region

λ(x) = 0.122x + 0.4006x 2 +0.076352x 3 + 0.401048x 9 appears at higher SNR than the proposed degree distribution
ρ(x) = 0.175341x 6 + 0.824659x 7 pairs and profiles in [10].
As shown in Table I and Fig.2, the threshold values precisely
Degree III :
 predict the asymptotic performance as the block length of the
λ(x) = 0.122x +0.4606x 2 +0.365056x 9 +0.052344x 19 LDPC codes approaches infinity. We are also interested in
ρ(x) = 0.175341x 6 + 0.824659x 7 the performance of optimized irregular LDPC codes when
(13) finite block size is considered. In the following results,
we consider four types of rate-1/2 LDPC codes. The first
The SNR thresholds of the above proposed degree distribu- one, Degree I-Degree III with code length 648, is constructed
tion pairs, the SNR thresholds of those obtained by the CCSDS according to the degree distribution pairs presented in (13).
Optical Uplink [7], 802.16e [8] and 802.11n [9] standards The second one, Profile I - Profile IV with code length
designed in accordance with symbols-independent symmetric 648, is constructed according to the degree distribution pairs
channels at rate 1/2, and the SNR thresholds of the degree optimized for the Poisson channel, as shown in [10]. The third
distribution pairs obtained in Poisson channel [10] are listed one, 802.16e (576) with code length 576 and 802.11n (648)
in Table I. The information in the table verifies that our results with code length 648, are constructed according to degree
730 IEEE PHOTONICS TECHNOLOGY LETTERS, VOL. 29, NO. 9, MAY 1, 2017

rate 1/2 were designed by adopting asymmetric density evo-

lution and an optimization scheme combining four different
search algorithms. Simulation results verified that the proposed
degree distribution pairs performed very well compared with
three conventional LDPC codes designed in accordance with
symbols-independent symmetric channels and codes designed
in Poisson channel, and the optimized codes were much
closer to the Shannon limit for OOK modulated binary non-
symmetric channels. Future work will take turbulence into
account to optimize codes not only for middle/short-range but
also long-range terrestrial FSO communication systems.

Fig. 3. BER with respect to the SNR for different LDPC codes of code R EFERENCES
length 648 constructed using different degree distribution pairs at 1/2 rate
with code lengths indicated within parentheses. [1] T. J. Richardson, M. A. Shokrollahi, and R. L. Urbanke, “Design of
capacity-approaching irregular low-density parity-check codes,” IEEE
Trans. Inf. Theory, vol. 47, no. 2, pp. 619–637, Feb. 2001.
distribution pairs presented in [8] and [9] respectively. The [2] C.-C. Wang, S. R. Kulkarni, and H. V. Poor, “Density evolution for
last one, CCSDS Optical Uplink (512) with code length 512, asymmetric memoryless channels,” IEEE Trans. Inf. Theory, vol. 51,
is constructed according to a degree distribution pair presented no. 12, pp. 4216–4236, Dec. 2005.
[3] M. N. Khan, “Importance of noise models in FSO communications,”
in [7]. To ensure the fairness of BER performance comparison, EURASIP J. Wireless Commun. Netw., vol. 2014, no. 102, pp. 1–10,
all these codes are constructed using the progressive edge Jun. 2014.
growth (PEG) algorithm [12]. [4] F. Xu, M.-A. Khalighi, and S. Bourennane, “Impact of different noise
sources on the performance of PIN- and APD-based FSO receivers,” in
Fig. 3 compares the Monte-Carlo simulation results of these Proc. IEEE Int. Conf. Telecommun., Aug. 2011, pp. 211–218.
four kinds of LDPC codes on the OOK and PIN photodiode [5] R. P. S. Oldenhuis, “Trajectory optimization for a mission to the solar
photon-detecting channel. As the curves show, Degree I, bow shock and minor planets,” M.S. thesis, Dept. Astrodyn. Space
Missions, Delft Univ. Technol., Delft, The Netherlands, May 2010.
Degree II and Degree III codes all outperform the rest of [6] H. Hemmati, Ed., Deep Space Optical Communications, vol. 11.
codes for high values of the SNR and the slope of BER Hoboken, NJ, USA: Wiley, 2006.
curves of these three codes are the steepest. Note that the [7] High Photon Efficiency Optical Communications Coding & Modulation,
CCSDS, Washington, DC, USA, Sep. 2015, pp. 14–20.
simulation results are consistent with the evaluation of the [8] Draft IEEE Standard for Local Metropolitan Area Networks Wireless
BER performances for different degree distribution pairs under Access Systems. Part 16: Air Interface for Fixed and Mobile Broadband,
density evolution as shown in Fig. 2. It is shown that the first IEEE Standard P802.16e/D12, IEEE, New York, NY, USA, Oct. 2005.
[9] Draft IEEE Standard for Local Metropolitan Networks—Specific
type codes exhibit longer waterfall regions compared to second Requirements. Part 11: Wireless LAN Medium Access Control (MAC),
and third types of LDPC codes. Although the waterfall region and Physical Layer (PHY) Specifications: Enhancements for Higher
of CCSDS Optical Uplink (512) codes is longest, it occurs Throughput, IEEE Standard P802.11n/D10, IEEE, New York, NY, USA,
Mar. 2006.
much later (at higher SNR) than the first and second type [10] T. Xie, S. G. Wilson, M. Brandt-Pearce, and C. L. Brown, “LDPC code
codes. design for OOK modulated Poisson optical channels,” in Proc. IEEE
47th Annu. Conf. Inf. Sci. Syst. (CISS), Mar. 2013, pp. 1–6.
[11] J. A. Anguita, I. B. Djordjevic, M. A. Neifeld, and B. V. Vasic, “Shannon
V. C ONCLUSION capacities and error-correction codes for optical atmospheric turbulent
Optimized degree distribution pairs of LDPC codes for channels,” J. Opt. Netw., vol. 4, no. 9, pp. 586–601, Aug., 2005.
[12] H. Xiao and A. H. Banihashemi, “Improved progressive-edge-
middle/short-range terrestrial high-speed FSO communication growth (PEG) construction of irregular LDPC codes,” IEEE Commun.
systems employing OOK and PIN photodiode receivers at Lett., vol. 8, no. 12, pp. 715–717, Dec. 2004.