You are on page 1of 4

1476 IEEE COMMUNICATIONS LETTERS, VOL. 19, NO.

9, SEPTEMBER 2015

A Construction of Physical-Layer Systematic


Raptor Codes Based on Protographs
Shiuan-Hao Kuo, Huang-Chang Lee, Yeong-Luh Ueng, Member, IEEE, and Mao-Chao Lin, Member, IEEE

Abstract—A construction of physical-layer Raptor (PLR) codes R0 = 6/7 or 8/10. Vi+1 is constructed by adding new variable
based on protographs is proposed. We first set up a sequence of nodes with check sums to the existing nodes of Vi , where the
signal-to-noise ratios (SNRs). For each SNR, we employ a modified check sums should yield the lowest decodable SNR for Vi+1 .
protograph EXIT analysis to find the extra check nodes in addition
The proposed PLR code employs the ideas of pEXIT in
to the existing check nodes so as to meet the threshold required
by the outer low-density parity check code. PLR codes which are [10] and incremental design in [11]. The pEXIT for LDPC
capacity-approaching for a wide SNR range can be obtained using codes can be modified for applications to the systematic PLR
the proposed construction. codes by adding channel values to check nodes of the inner
LT code in computing the mutual information. For the PLR
Index Terms—Protograph, physical-layer Raptor code.
code design, we use a protograph outer LDPC code V with
rate 0.95. There is a mutual information (MI) threshold IV
I. I NTRODUCTION which allows asymptotically error-free performances for V. A

L T CODE [1] is a coding scheme, which was originally


proposed to cope with package loss. With a low-density
parity-check (LDPC) [2] code as precoding to the LT code, we
modified pEXIT analysis for the protograph LT code C is used
to calculate IC which is the MI obtained from the message
passing between the variable nodes in the systematic part and
have a Raptor code [3]. A physical-layer Raptor (PLR) code the check nodes of the LT code. Suppose that at a certain SNR
which is designed for physical-layer channels can be considered γi , IC ≥ IV is obtained. We then need to find the least amount
as a continuous version of incremental redundancy (IR) based of extra check nodes so that IC ≥ IV still holds for a lower
hybrid automatic repeat request (HARQ) [4]. SNR γi+1 . In this way, we would like to search for a PLR
For the PLR code, a conventional design is to optimize its code with best throughput performance for a wide SNR range.
output degree distribution. Optimization for a certain signal-to- To reduce the search load, we consider a simple procedure
noise ratio (SNR) is studied in [5], [6]. Adaptive schemes for to construct PLR codes with regular check node degrees first.
non-systematic LT-accumulator code [7] and non-systematic Then, a somewhat more complicated procedure for searching
Raptor codes [8] that can work well over a wide SNR range are PLR codes with irregular check node degrees is proposed.
designed by incrementally adjusting output degree distributions Numerical results show that capacity-approaching throughput
according to the channel condition. efficiency can be achieved in a wide SNR range.
In this letter, we resort to the concept of LDPC codes using
protographs [9] for designing PLR codes. A protograph is a II. S OME P RELIMINARIES
small graph prototype of Tanner graph. By applying a copy- The PLR code V◦C is a serially concatenated code with an
and-permute process to the protograph called lifting, large outer LDPC code V and an inner LT code C. Each information
derived graphs of various sizes can be obtained. It is indicated block with k bits is firstly fed to the LDPC encoder to yield n
in [9] that lifting process with circulant permutation is favorable coded bits of V, which are also referred as the intermediate bits
to efficient hardware implementations of associated codes. of V◦C. For a systematic PLR code, the first n code bits of C
In [10], a protograph EXIT (pEXIT) method is developed are also the n code bits of V, while the other code bits of C will
for performance analysis of protograph LDPC codes. In [11], be referred as the output bits or check bits of C. To obtain an LT
a punctured-node protograph-based Raptor-like (PN-PBRL) output (check) bit co with a predefined degree do , the encoder
LDPC code has been constructed through an incremental design will select do intermediate bits and co is the binary sum of these
without considering degree distribution, where a sequence of do intermediate bits. The LT encoder will generate output bits
protograph LDPC codes V0 , V1 , V2 , . . . with decreasing rates continuously until an ACK is received.
are extended from a rate-R0 protograph mother code V0 where For a PLR code, the decoding capability of the outer LDPC
code V is marked by a threshold IV . For a PLR code, it is desired
Manuscript received February 4, 2015; accepted June 15, 2015. Date of to find the least number of check nodes so that the iterative
publication June 23, 2015; date of current version September 4, 2015. This decoding of C can obtain the threshold
 IV . Conventionally, an
work was supported by the Ministry of Science and Technology of the ROC output degree distribution (x) = d d xd is used to describe
under Grant MOST 103-2221-E-002-087-MY3 and MOST 103-2221-E-007-
025-MY3. The associate editor coordinating the review of this paper and the LT code C, where d indicates the probability of an LT
approving it for publication was Z. Ma. output bit having degree d. In this letter, the construction of C
S.-H. Kuo and M.-C. Lin are with the Graduate Institute of Communi- is designed from a different approach for which the information
cation Engineering, National Taiwan University, Taipei 106, Taiwan (e-mail:
d96942022@ntu.edu.tw; mclin@ntu.edu.tw).
of (x) is not needed.
H.-C. Lee and Y.-L. Ueng are with the Department of Electrical Engineering
and the Institute of Communications Engineering, National Tsing Hua Uni- III. P ROTOGRAPH -BASED R APTOR C ODES
versity, Hsinchu 30013, Taiwan (e-mail: akula.lee@gmail.com; ylueng@ee.
nthu.edu.tw). Fig. 1 illustrates the protograph for the proposed proto-
Digital Object Identifier 10.1109/LCOMM.2015.2448678 graph Raptor (PR) code, where πij is a shorthand for the
1558-2558 © 2015 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.
KUO et al.: A CONSTRUCTION OF PHYSICAL-LAYER SYSTEMATIC RAPTOR CODES BASED ON PROTOGRAPHS 1477

and Icv (i, j) = 0 if bij = 0. The above equations are evaluated


iteratively until the maximum number of iterations is reached.
The MI of decoded LLR at the ith pVN is the same as in [10],

Iv (i) = J bis J −1 (Icv (i, s)) + σ02 . (3)
s

Finally, we define the average MI of LT decoder  for transmis-


sion up to the th pCN to be ĪC (B , γ ) = K1B Iv (i).

Fig. 1. Protograph of the proposed PR code, where vi , 1 ≤ i ≤ KB , denotes the IV. C ONSTRUCTION OF C HECK -R EGULAR PR C ODES
pVN and cj , 1 ≤ j ≤ NB , denotes the pCN of the LT code.
We first investigate the construction of the check-regular LT
copy-and-permute procedure. A KB × NB base generator matrix codes. The B matrix is constructed in a column-by-column
B = [bij] = [b1 , b2 , · · · , bNB ] is constructed for the LT code, manner. Suppose that the B−1 have been determined. To obtain
where KB is the total number of the protograph variable nodes the th column of B, we randomly generate Q b̃ vectors. Let
(pVNs) and the entry bij denotes the number of edges con- B̃ = [b1 , · · · , b−1 , b̃ ]. The pEXIT technique is applied to
necting the ith pVN and the jth protograph check node (pCN). these Q B̃ matrices to determine the best candidate b which
 B
The degree of the jth pCN is K maximizes the average decoded MI ĪC (B̃ , γi ).
i=1 bij . The lifting operation is
processed by replacing each element in the base matrix with The pEXIT analysis is first applied to B̃ at a relatively high
a q-by-q permutation matrix that connects q variable nodes SNR γ1 . Once the average decoded MI at γi , ĪC (B , γi ) ≥ IV ,
(intermediate nodes) and q check nodes (output nodes). In this where B = [b1 , · · · , b ], the following columns should be
way, the q copies of the protograph are interconnected. The designed for a lower SNR γi+1 . Suppose that ĪC (B , γi−1 ) ≥ IV
derived graph is q times larger than the original protograph but only if  ≥ pi − 1 and ĪC (B , γi ) ≥ IV only if  ≥ pi+1 − 1.
preserving the same rate and degree distributions. The number The columns bpi , · · · , bpi+1 −1 are the extra columns deter-
of variable nodes of C is qKB . We set bij ∈ {0, 1, 2}. Hence, mined by the pEXIT analysis under SNR γi . In other words,
parallel edges are allowed in the protograph as shown in Fig. 1. the extra number of pCNs required to achieve IV under SNR γi
However, there will be no parallel edges after lifting. Using is pi = pi+1 − pi . The design target is to keep pi as small as
B = [b1 , · · · , b−1 , b ], we can generate  pCN (q check possible for all i. The design process is as follows.
nodes), where 0 ≤  ≤ NB . Since the LT code is systematic, its Protograph Raptor (PR) Algorithm:
code rate is RC, = KB /( + KB ). The overall rate of the PR Step 1 Choose KB , NB , IV and an SNR set {γ1 , γ2 , · · · }, where
code is nk RC,  . The maximum length of C is q(KB + NB ). γi > γi+1 for all i. Initialize the SNR index i to 1 and
To determine the achievable performance of the LT code using the column index  to 1.
B , we used a protograph EXIT (pEXIT) analysis modified from Step 2 Randomly generate Q b̃ vectors and obtain their as-
that used in [10] for LDPC codes. For the Gaussian noise chan- sociated B̃ = [b1 , · · · , b−1 , b̃ ] matrices. For each
nel with noise variance σ 2 , we define SNR as γ = Es /σ 2 , where candidate B̃ , use pEXIT to evaluate its decoded
Es is the energy per code symbol. Let J(σ02 ) [8] be the mutual
MI ĪC (B̃ , γi ). Let b = arg max ĪC (B̃ , γi ) and B =
information (MI) between the transmitted equally-likely binary b̃
signal and the received LLR (log-likelihood ratio), where σ02 = [b1 , · · · , b ].
4γ is the variance of the received LLR. As in [10], the MI for up- Step 3 If ĪC (B , γi ) ≥ IV , set i = i + 1 so as to switch the SNR
dating the message passing from the ith pVN to the jth pCN is from γi to γi+1 . Otherwise, the SNR value remains the
⎛ same. If  < NB , increase  by 1 and go to Step 2 to
 determine the next column. Otherwise, the algorithm is
Ivc (i, j) = J ⎝ bis J −1 (Icv (i, s)) completed with B = BNB .
s =j
⎞ We now compare the results of the check-regular PR codes
to the conventional systematic Raptor codes over the Gaussian
+ (bij − 1)J −1 (Icv (i, j)) + σ02 ⎠, if bij = 0, (1) channel optimized for a single SNR. The outer code V is a
protograph LDPC code with rate nk = 0.95. For the conven-
tional Raptor codes, we follow the optimization method in [6] to
and Ivc (i, j) = 0 if bij = 0. Since the LT code bits are associated obtain three output degree distributions, 1 (x) = 0.5420x30 +
to the check nodes, the MI for updating the message passing 0.4580x31, 2 (x) = 0.7128x13 + 0.2783x14 + 0.0089x100 and
from the jth pCN to the ith pVN in [10] must include the 3 (x) = 0.1585x7 + 0.8412x8 + 0.0002x100, which are opti-
channel values and hence is modified as mized based on SNR 6 dB, 4 dB and 2 dB, respectively. For
⎛ the PR codes, we devised three codes respectively with regular
 output degrees 31, 15 and 7 using the parameters of KB = 100,
Icv (i, j) = 1 − J ⎝ bsj J −1 (1 − Ivc (s, j)) + (bij − 1) NB = 300, Q = 2000, IV = 0.975, 20 pEXIT iterations and
s =i {γ1 , γ2 , · · · } = {6, 5.5, 5, 4.5, · · · } in decibel.

  The efficiencies (average throughput/channel capacity, as
· J −1 (1 − Ivc (i, j)) + J −1 1 − J σ02 ⎠, if bij = 0, (2) defined in [8]) of the proposed PR codes when q = 20 and the
conventional randomly connected Raptor codes in IR-HARQ
1478 IEEE COMMUNICATIONS LETTERS, VOL. 19, NO. 9, SEPTEMBER 2015

Fig. 2. Efficiencies (throughput/capacity) of conventionally optimized Raptor Fig. 3. Efficiencies (throughput/capacity). Iteration numbers: 20 and 100 for
code, PR code, IPR code and IR-HARQ scheme with RC-LDPC. The informa- the IPR code and the PN-PBRL code; 20, 50, and 100 for the ANR code [8].
tion block size is k = 1900 for all codes.

systems are shown in Fig. 2. In both codes the maximum Step 4 If ĪC (Bd , γi ) ≥ IV for one or more d values, select
iteration of LT code is 20 and the maximum iteration of LDPC d as the d which has the largest ĪC (Bd , γi ), and if
code is 60. When the decoding fails, the receiver attempts to there is still a tie regarding ĪC (Bd , γi ), select d as the
decode again after collecting 60 more LT code bits. We see that 
smallest d. Set pi+1 =  + 1, set bj equal to bdj for
there exists a check-regular PR code which has slightly better
j = pi , · · · , , and set i = i + 1 so as to switch the SNR
throughput performance than the conventional Raptor code that
from γi to γi+1 . Otherwise, the SNR value remains the
is optimized for each given SNR.
same. If  < NB , increase  by 1 and go to Step 2 to
determine the next column. Otherwise, B = BNB .
V. C ONSTRUCTION OF I RREGULAR PR C ODES
Fig. 2 shows the efficiency performance of the IPR code,
We have seen that the PR code with a fixed output degree which uses the same settings as the PR codes and the possible
is good only for a restricted SNR range. It is clear that if we degrees are D = {7, 15, 31}. It shows that the IPR code extend
relieve the restriction on the random b̃ vectors for PR codes, the capacity-approaching region of PR codes as expected. For
better throughput efficiency may be obtained. For an irregular comparison, Fig. 2 also shows the performance of IR-HARQ
PR (IPR) code, each column in B has a degree from the multiple scheme using rate-compatible (RC) LDPC codes. We use a
values in a set D. RC-LDPC code scheme that has a larger applicable rate range
To search for good IPR codes with the irregular check-node based on the puncturing method presented in [12]. The mother
degrees, simply modifying Step 2 of the regular PR algorithm codes in consideration are rate 0.5 and 0.8 regular PEG LDPC
by replacing Q candidates for b by Q|D| candidates is not codes with k = 1900. A length 2000 punctured codeword is
enough. To achieve a small pi , we may consider an algorithm firstly transmitted, followed by 60 punctured bits every retrans-

which uses B̃ = [b1 , · · · , bpi −1 , b̃pi , b̃pi +1 , · · · , b̃ ] with mission. If the entire LDPC codeword has been transmitted,
(Q|D|)−pi+1 candidates for b̃pi , b̃pi +1 , · · · , b̃ in selecting the it repeats the LDPC coded bit until a maximum of 8000 bits
∗ are sent. The repeated bits are combined using maximal ratio
best ĪC (B̃ , γi ) that achieve IV under SNR γi with smallest .
However, the search load is too much. Instead, we use a less- combining at the receiver. The LDPC codes are decoded with
d 60 iterations. Clearly, these RC-LDPC codes can not match the
greedy algorithm in which, for each d ∈ D, there are Q b̃
d d IPR code in the efficiency performance.
vectors which are used to select the best ĪC (B̃ , γi ), where B̃ = We are also interested in the performance when lifting the
d d
[b1 , · · · , bpi −1 , bdpi , · · · , bd−1 , b̃ ], and bdpi , · · · , bd−1 , b̃ all IPR code to a longer code. Fig. 3 illustrates the efficiencies of
have weight d. The search space has ( − pi + 1)Q|D| candi- the IPR code when q = 100 (k = 9500) comparing to the 6-step
dates. The algorithm is as follows. adaptive non-systematic Raptor (ANR) code in [8] with the
Irregular Protograph Raptor (IPR) Algorithm: same size. Under the same iterations, the proposed IPR code
outperforms the ANR code in the presented SNR range, espe-
Step 1 Choose KB , NB , IV and an SNR set {γ1 , γ2 , · · · }, where cially in high SNR value.
γi > γi+1 for all i. Let D be the set of possible degrees. In [11], a punctured-node protograph-based Raptor-like (PN-
Initialize i = 1,  = 1 and pi = 1. PBRL) code is constructed by using a base parity check matrix
d
Step 2 For a degree d ∈ D, randomly generate Q b̃ vectors
d HV0 0
and obtain the associated B̃ matrices. For each can- HB, = (4)
HLT, I
didate, use pEXIT to evaluate the average decoded MI
d d
ĪC (B̃ , γi ). Select bd = arg max ĪC (B̃ , γi ). where HV0 and HB, are the base parity check matrices of
d
b̃ the LDPC code V0 and V respectively, HLT, represents the
Step 3 Repeat Step 2 by (|D| − 1) times to search all available connections between the  incremented pCNs and the pVNs in
degrees. An associated matrix Bd = [b1 , · · · , bpi −1 , V0 . Density evolution is used to find the best HLT, which yields
bdpi , · · · , bd ] is obtained for each d ∈ D. the lowest decodable SNR. In this way, a capacity approaching
KUO et al.: A CONSTRUCTION OF PHYSICAL-LAYER SYSTEMATIC RAPTOR CODES BASED ON PROTOGRAPHS 1479

than 0 dB. For hardware implementation, a protograph based


code with circulant permutation can be agreeable to simple
routing for both encoder and decoder. Therefore, the IPR code
is more hardware-friendly with less decoding complexity than
the ANR code. The average input degree for the PN-PBRL
code is also shown in Fig. 4. The IPR code has slightly more
complexity than the PN-PBRL code. This result may come
from the condition that the IPR code is designed to achieve
good efficiency in the SNR regime that starts from a high value
Fig. 4. The average input degree (edge/input bit) of proposed IPR code (LT
part), the adaptive nonsystematic Raptor code (LT part) in [8] and the PN-PBRL and hence a significant portion of high output degrees is needed.
code with R0 = 6/7 in IR-HARQ scheme.

RC-LDPC code with R0 = 6/7 or 8/10 have been constructed. VII. C ONCLUSION
Fig. 3 shows the efficiency performance of a IR-HARQ scheme We introduce a protograph-based systematic Raptor code. The
using PN-PBRL code with R0 = 6/7 and k = 9504 under 20 design of the proposed code employs a modified pEXIT analysis
and 100 iterations. The PN-PBRL code with large iteration to construct a sequence of incrementally enlarged base generator
numbers can be more efficient if the SNR matches its rate range. matrices for gradually decreasing SNR. Compared to the PN-
For a hardware-friendly application, the presented IPR code is PBRL code with R0 = 6/7 in [11] which is constructed through
constructed with 20 pEXIT iterations and thus its connections a sequence of incrementally enlarged base parity-check matrices
are selected to perform well under 20 iterations. Hence, accord- for gradually decreasing coding rates, the proposed IPR code has
ing to the efficiency performances, the IPR code is similar or the advantage in throughput efficiency at high SNR and suffers
superior to the PN-PBRL code under 20 iterations for a wide some loss if the number of iterations is high. Compared to the
SNR range and is inferior to the PN-PBRL code under 100 (or ANR code in [8] which is designed for an optimized output de-
more) iterations at some SNR range. The comparison of BER gree distribution, the proposed IPR code has the advantages of
performances for the IPR code and the PN-PBRL code has smaller edge count, smaller needed iterations and the hardware-
similar trend. Due to space limitation, the BER performances friendly structure inherited from the protograph design.
are not provided here.
A larger search space in the code generation can increase the
probability of finding a code with better BER performance. For R EFERENCES
the ANR code in [8], the search space for output degree dis- [1] M. Luby, “LT codes,” in Proc. 43rd Annu. IEEE Symp. Found. Comput.
Sci., 2002, pp. 271–280.
tribution is {ωd : d ∈ D = {1, 2, 3, · · · , 9, 10, 15, 20, 25, · · · , [2] D. J. C. MacKay, “Good error-correcting codes based on very sparse ma-
75, 80}}. For the PN-PBRL code with R0 = 6/7 in [11], the trices,” IEEE Trans. Inf. Theory, vol. 45, no. 2, pp. 399–431, Mar. 1999.
search space for each row in HLT, is the set of all the rows with [3] A. Shokrollahi, “Raptor codes,” IEEE Trans. Inf. Theory, vol. 52, no. 6,
weight d, where d ∈ D, and D ⊇ {3, 4, 5, 6}. For the PR code, pp. 2551–2567, Jun. 2006.
[4] E. Soijanin, N. Varnica, and P. Whiting, “Punctured vs rateless codes
the search space for each column in B̃ is a set of size Q = for hybrid ARQ,” in Proc. ITW, Punta del Este, Uruguay, Mar. 2006,
2000 which contains columns of a given weight d. For the IPR pp. 155–159.
code, the size of search space is further multiplied by |D|, where [5] O. Etesami and A. Shokrollahi, “Raptor codes on binary memoryless sym-
metric channels,” IEEE Trans. Inf. Theory, vol. 52, no. 5, pp. 2033–2051,
D = {7, 15, 31}. For any of the ANR code, PN-PBRL code and May 2006.
the IPR code, there is still room for generating a better code by [6] Z. Cheng, J. Castura, and Y. Mao, “On the design of raptor codes for
expanding the current search space. However, the IPR code has binary-input Gaussian channels,” IEEE Trans. Commun., vol. 57, no. 11,
more room for expanding the search space since in our current pp. 3269–3277, Nov. 2009.
[7] N. Bonello, R. Zhang, S. Chen, and L. Hanzo, “Reconfigurable rateless
search we use the least |D| and a less-greedy algorithm. codes,” IEEE Trans. Wireless Commun., vol. 8, no. 11, pp. 5592–5600,
Nov. 2009.
[8] S.-H. Kuo, Y. L. Guan, S.-K. Lee, and M.-C. Lin, “A design of physical-
VI. D ECODING C OMPLEXITY layer raptor codes for wide SNR ranges,” IEEE Commun. Lett., vol. 18,
no. 3, pp. 491–494, Mar. 2014.
For the decoder, its computational complexity can be eval- [9] D. Divsalar, S. Dolinar, C. R. Jones, and K. Andrews, “Capacity-
uated by the number of required iterations and the number approaching protograph codes,” IEEE J. Sel. Areas Commun., vol. 27,
of edges which implies the complexity for each iteration. We no. 6, pp. 876–888, Aug. 2009.
compare the average LT code input degrees of the proposed [10] G. Liva and M. Chiani, “Protograph LDPC codes design based on EXIT
analysis,” in Proc. IEEE GLOBECOM, Washington, DC, USA, Nov. 2007,
IPR code and the ANR codes [8]. From Fig. 4, we see that pp. 3250–3254.
the edge counts for both codes are vaguely equal at low rate. [11] T.-Y. Chen, K. Vakilinia, D. Divsalar, and R. D. Wesel, “Protograph-
The superiority of the IPR code at high SNR comes from based raptor-like LDPC codes,” IEEE Trans. Commun., vol. 63, no. 5,
its systematic design. In Fig. 3, the efficiency performance pp. 1522–1532, May 2015.
[12] R. Asvadi and A. H. Banihashemi, “A rate-compatible puncturing scheme
obtained by the ANR through 50 iterations is still inferior to that for finite-length LDPC codes,” IEEE Commun. Lett., vol. 17, no. 1,
obtained by the IPR code under 20 iterations for SNR greater pp. 147–150, Jan. 2013.

You might also like